1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata = 0;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct mem_cgroup_reclaim_iter {
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup *last_visited;
154 /* scan generation, increased every round-trip */
155 unsigned int generation;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone {
162 struct lruvec lruvec;
163 unsigned long lru_size[NR_LRU_LISTS];
165 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
167 struct rb_node tree_node; /* RB tree node */
168 unsigned long long usage_in_excess;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup *memcg; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node {
176 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone {
185 struct rb_root rb_root;
189 struct mem_cgroup_tree_per_node {
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
193 struct mem_cgroup_tree {
194 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
199 struct mem_cgroup_threshold {
200 struct eventfd_ctx *eventfd;
205 struct mem_cgroup_threshold_ary {
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries[0];
214 struct mem_cgroup_thresholds {
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary *primary;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary *spare;
226 struct mem_cgroup_eventfd_list {
227 struct list_head list;
228 struct eventfd_ctx *eventfd;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event {
236 * memcg which the event belongs to.
238 struct mem_cgroup *memcg;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx *eventfd;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event)(struct mem_cgroup *memcg,
253 struct eventfd_ctx *eventfd, const char *args);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event)(struct mem_cgroup *memcg,
260 struct eventfd_ctx *eventfd);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t *wqh;
268 struct work_struct remove;
271 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
272 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css;
288 * the counter to account for memory usage
290 struct res_counter res;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure;
296 * the counter to account for mem+swap usage.
298 struct res_counter memsw;
301 * the counter to account for kernel memory usage.
303 struct res_counter kmem;
305 * Should the accounting and control be hierarchical, per subtree?
308 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
312 atomic_t oom_wakeups;
315 /* OOM-Killer disable */
316 int oom_kill_disable;
318 /* set when res.limit == memsw.limit */
319 bool memsw_is_minimum;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds;
330 /* For oom notifier event fd */
331 struct list_head oom_notify;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock;
347 struct mem_cgroup_stat_cpu __percpu *stat;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base;
353 spinlock_t pcp_counter_lock;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list. per-memcg */
361 struct list_head memcg_slab_caches;
362 /* Not a spinlock, we can take a lot of time walking the list */
363 struct mutex slab_caches_mutex;
364 /* Index in the kmem_cache->memcg_params->memcg_caches array */
368 int last_scanned_node;
370 nodemask_t scan_nodes;
371 atomic_t numainfo_events;
372 atomic_t numainfo_updating;
375 /* List of events which userspace want to receive */
376 struct list_head event_list;
377 spinlock_t event_list_lock;
379 struct mem_cgroup_per_node *nodeinfo[0];
380 /* WARNING: nodeinfo must be the last member here */
383 /* internal only representation about the status of kmem accounting. */
385 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
386 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
389 #ifdef CONFIG_MEMCG_KMEM
390 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
392 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
395 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
397 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
400 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
403 * Our caller must use css_get() first, because memcg_uncharge_kmem()
404 * will call css_put() if it sees the memcg is dead.
407 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
408 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
411 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
413 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
414 &memcg->kmem_account_flags);
418 /* Stuffs for move charges at task migration. */
420 * Types of charges to be moved. "move_charge_at_immitgrate" and
421 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
424 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
425 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
429 /* "mc" and its members are protected by cgroup_mutex */
430 static struct move_charge_struct {
431 spinlock_t lock; /* for from, to */
432 struct mem_cgroup *from;
433 struct mem_cgroup *to;
434 unsigned long immigrate_flags;
435 unsigned long precharge;
436 unsigned long moved_charge;
437 unsigned long moved_swap;
438 struct task_struct *moving_task; /* a task moving charges */
439 wait_queue_head_t waitq; /* a waitq for other context */
441 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
442 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
445 static bool move_anon(void)
447 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
450 static bool move_file(void)
452 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
456 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
457 * limit reclaim to prevent infinite loops, if they ever occur.
459 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
460 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
463 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
464 MEM_CGROUP_CHARGE_TYPE_ANON,
465 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
466 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
470 /* for encoding cft->private value on file */
478 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
479 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
480 #define MEMFILE_ATTR(val) ((val) & 0xffff)
481 /* Used for OOM nofiier */
482 #define OOM_CONTROL (0)
485 * Reclaim flags for mem_cgroup_hierarchical_reclaim
487 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
488 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
489 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
490 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
493 * The memcg_create_mutex will be held whenever a new cgroup is created.
494 * As a consequence, any change that needs to protect against new child cgroups
495 * appearing has to hold it as well.
497 static DEFINE_MUTEX(memcg_create_mutex);
499 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
501 return s ? container_of(s, struct mem_cgroup, css) : NULL;
504 /* Some nice accessors for the vmpressure. */
505 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
508 memcg = root_mem_cgroup;
509 return &memcg->vmpressure;
512 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
514 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
517 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
519 return (memcg == root_mem_cgroup);
523 * We restrict the id in the range of [1, 65535], so it can fit into
526 #define MEM_CGROUP_ID_MAX USHRT_MAX
528 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
531 * The ID of the root cgroup is 0, but memcg treat 0 as an
532 * invalid ID, so we return (cgroup_id + 1).
534 return memcg->css.cgroup->id + 1;
537 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
539 struct cgroup_subsys_state *css;
541 css = css_from_id(id - 1, &memory_cgrp_subsys);
542 return mem_cgroup_from_css(css);
545 /* Writing them here to avoid exposing memcg's inner layout */
546 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
548 void sock_update_memcg(struct sock *sk)
550 if (mem_cgroup_sockets_enabled) {
551 struct mem_cgroup *memcg;
552 struct cg_proto *cg_proto;
554 BUG_ON(!sk->sk_prot->proto_cgroup);
556 /* Socket cloning can throw us here with sk_cgrp already
557 * filled. It won't however, necessarily happen from
558 * process context. So the test for root memcg given
559 * the current task's memcg won't help us in this case.
561 * Respecting the original socket's memcg is a better
562 * decision in this case.
565 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
566 css_get(&sk->sk_cgrp->memcg->css);
571 memcg = mem_cgroup_from_task(current);
572 cg_proto = sk->sk_prot->proto_cgroup(memcg);
573 if (!mem_cgroup_is_root(memcg) &&
574 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
575 sk->sk_cgrp = cg_proto;
580 EXPORT_SYMBOL(sock_update_memcg);
582 void sock_release_memcg(struct sock *sk)
584 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
585 struct mem_cgroup *memcg;
586 WARN_ON(!sk->sk_cgrp->memcg);
587 memcg = sk->sk_cgrp->memcg;
588 css_put(&sk->sk_cgrp->memcg->css);
592 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
594 if (!memcg || mem_cgroup_is_root(memcg))
597 return &memcg->tcp_mem;
599 EXPORT_SYMBOL(tcp_proto_cgroup);
601 static void disarm_sock_keys(struct mem_cgroup *memcg)
603 if (!memcg_proto_activated(&memcg->tcp_mem))
605 static_key_slow_dec(&memcg_socket_limit_enabled);
608 static void disarm_sock_keys(struct mem_cgroup *memcg)
613 #ifdef CONFIG_MEMCG_KMEM
615 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
616 * The main reason for not using cgroup id for this:
617 * this works better in sparse environments, where we have a lot of memcgs,
618 * but only a few kmem-limited. Or also, if we have, for instance, 200
619 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
620 * 200 entry array for that.
622 * The current size of the caches array is stored in
623 * memcg_limited_groups_array_size. It will double each time we have to
626 static DEFINE_IDA(kmem_limited_groups);
627 int memcg_limited_groups_array_size;
630 * MIN_SIZE is different than 1, because we would like to avoid going through
631 * the alloc/free process all the time. In a small machine, 4 kmem-limited
632 * cgroups is a reasonable guess. In the future, it could be a parameter or
633 * tunable, but that is strictly not necessary.
635 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
636 * this constant directly from cgroup, but it is understandable that this is
637 * better kept as an internal representation in cgroup.c. In any case, the
638 * cgrp_id space is not getting any smaller, and we don't have to necessarily
639 * increase ours as well if it increases.
641 #define MEMCG_CACHES_MIN_SIZE 4
642 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
645 * A lot of the calls to the cache allocation functions are expected to be
646 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
647 * conditional to this static branch, we'll have to allow modules that does
648 * kmem_cache_alloc and the such to see this symbol as well
650 struct static_key memcg_kmem_enabled_key;
651 EXPORT_SYMBOL(memcg_kmem_enabled_key);
653 static void disarm_kmem_keys(struct mem_cgroup *memcg)
655 if (memcg_kmem_is_active(memcg)) {
656 static_key_slow_dec(&memcg_kmem_enabled_key);
657 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
660 * This check can't live in kmem destruction function,
661 * since the charges will outlive the cgroup
663 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
666 static void disarm_kmem_keys(struct mem_cgroup *memcg)
669 #endif /* CONFIG_MEMCG_KMEM */
671 static void disarm_static_keys(struct mem_cgroup *memcg)
673 disarm_sock_keys(memcg);
674 disarm_kmem_keys(memcg);
677 static void drain_all_stock_async(struct mem_cgroup *memcg);
679 static struct mem_cgroup_per_zone *
680 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
682 VM_BUG_ON((unsigned)nid >= nr_node_ids);
683 return &memcg->nodeinfo[nid]->zoneinfo[zid];
686 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
691 static struct mem_cgroup_per_zone *
692 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
694 int nid = page_to_nid(page);
695 int zid = page_zonenum(page);
697 return mem_cgroup_zoneinfo(memcg, nid, zid);
700 static struct mem_cgroup_tree_per_zone *
701 soft_limit_tree_node_zone(int nid, int zid)
703 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
706 static struct mem_cgroup_tree_per_zone *
707 soft_limit_tree_from_page(struct page *page)
709 int nid = page_to_nid(page);
710 int zid = page_zonenum(page);
712 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
716 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
717 struct mem_cgroup_per_zone *mz,
718 struct mem_cgroup_tree_per_zone *mctz,
719 unsigned long long new_usage_in_excess)
721 struct rb_node **p = &mctz->rb_root.rb_node;
722 struct rb_node *parent = NULL;
723 struct mem_cgroup_per_zone *mz_node;
728 mz->usage_in_excess = new_usage_in_excess;
729 if (!mz->usage_in_excess)
733 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
735 if (mz->usage_in_excess < mz_node->usage_in_excess)
738 * We can't avoid mem cgroups that are over their soft
739 * limit by the same amount
741 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
744 rb_link_node(&mz->tree_node, parent, p);
745 rb_insert_color(&mz->tree_node, &mctz->rb_root);
750 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
751 struct mem_cgroup_per_zone *mz,
752 struct mem_cgroup_tree_per_zone *mctz)
756 rb_erase(&mz->tree_node, &mctz->rb_root);
761 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
762 struct mem_cgroup_per_zone *mz,
763 struct mem_cgroup_tree_per_zone *mctz)
765 spin_lock(&mctz->lock);
766 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
767 spin_unlock(&mctz->lock);
771 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
773 unsigned long long excess;
774 struct mem_cgroup_per_zone *mz;
775 struct mem_cgroup_tree_per_zone *mctz;
776 int nid = page_to_nid(page);
777 int zid = page_zonenum(page);
778 mctz = soft_limit_tree_from_page(page);
781 * Necessary to update all ancestors when hierarchy is used.
782 * because their event counter is not touched.
784 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
785 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
786 excess = res_counter_soft_limit_excess(&memcg->res);
788 * We have to update the tree if mz is on RB-tree or
789 * mem is over its softlimit.
791 if (excess || mz->on_tree) {
792 spin_lock(&mctz->lock);
793 /* if on-tree, remove it */
795 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
797 * Insert again. mz->usage_in_excess will be updated.
798 * If excess is 0, no tree ops.
800 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
801 spin_unlock(&mctz->lock);
806 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
809 struct mem_cgroup_per_zone *mz;
810 struct mem_cgroup_tree_per_zone *mctz;
812 for_each_node(node) {
813 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
814 mz = mem_cgroup_zoneinfo(memcg, node, zone);
815 mctz = soft_limit_tree_node_zone(node, zone);
816 mem_cgroup_remove_exceeded(memcg, mz, mctz);
821 static struct mem_cgroup_per_zone *
822 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
824 struct rb_node *rightmost = NULL;
825 struct mem_cgroup_per_zone *mz;
829 rightmost = rb_last(&mctz->rb_root);
831 goto done; /* Nothing to reclaim from */
833 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
835 * Remove the node now but someone else can add it back,
836 * we will to add it back at the end of reclaim to its correct
837 * position in the tree.
839 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
840 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
841 !css_tryget(&mz->memcg->css))
847 static struct mem_cgroup_per_zone *
848 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
850 struct mem_cgroup_per_zone *mz;
852 spin_lock(&mctz->lock);
853 mz = __mem_cgroup_largest_soft_limit_node(mctz);
854 spin_unlock(&mctz->lock);
859 * Implementation Note: reading percpu statistics for memcg.
861 * Both of vmstat[] and percpu_counter has threshold and do periodic
862 * synchronization to implement "quick" read. There are trade-off between
863 * reading cost and precision of value. Then, we may have a chance to implement
864 * a periodic synchronizion of counter in memcg's counter.
866 * But this _read() function is used for user interface now. The user accounts
867 * memory usage by memory cgroup and he _always_ requires exact value because
868 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
869 * have to visit all online cpus and make sum. So, for now, unnecessary
870 * synchronization is not implemented. (just implemented for cpu hotplug)
872 * If there are kernel internal actions which can make use of some not-exact
873 * value, and reading all cpu value can be performance bottleneck in some
874 * common workload, threashold and synchonization as vmstat[] should be
877 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
878 enum mem_cgroup_stat_index idx)
884 for_each_online_cpu(cpu)
885 val += per_cpu(memcg->stat->count[idx], cpu);
886 #ifdef CONFIG_HOTPLUG_CPU
887 spin_lock(&memcg->pcp_counter_lock);
888 val += memcg->nocpu_base.count[idx];
889 spin_unlock(&memcg->pcp_counter_lock);
895 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
898 int val = (charge) ? 1 : -1;
899 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
902 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
903 enum mem_cgroup_events_index idx)
905 unsigned long val = 0;
909 for_each_online_cpu(cpu)
910 val += per_cpu(memcg->stat->events[idx], cpu);
911 #ifdef CONFIG_HOTPLUG_CPU
912 spin_lock(&memcg->pcp_counter_lock);
913 val += memcg->nocpu_base.events[idx];
914 spin_unlock(&memcg->pcp_counter_lock);
920 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
922 bool anon, int nr_pages)
925 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
926 * counted as CACHE even if it's on ANON LRU.
929 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
932 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
935 if (PageTransHuge(page))
936 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
939 /* pagein of a big page is an event. So, ignore page size */
941 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
943 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
944 nr_pages = -nr_pages; /* for event */
947 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
951 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
953 struct mem_cgroup_per_zone *mz;
955 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
956 return mz->lru_size[lru];
960 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
961 unsigned int lru_mask)
963 struct mem_cgroup_per_zone *mz;
965 unsigned long ret = 0;
967 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
970 if (BIT(lru) & lru_mask)
971 ret += mz->lru_size[lru];
977 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
978 int nid, unsigned int lru_mask)
983 for (zid = 0; zid < MAX_NR_ZONES; zid++)
984 total += mem_cgroup_zone_nr_lru_pages(memcg,
990 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
991 unsigned int lru_mask)
996 for_each_node_state(nid, N_MEMORY)
997 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1001 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1002 enum mem_cgroup_events_target target)
1004 unsigned long val, next;
1006 val = __this_cpu_read(memcg->stat->nr_page_events);
1007 next = __this_cpu_read(memcg->stat->targets[target]);
1008 /* from time_after() in jiffies.h */
1009 if ((long)next - (long)val < 0) {
1011 case MEM_CGROUP_TARGET_THRESH:
1012 next = val + THRESHOLDS_EVENTS_TARGET;
1014 case MEM_CGROUP_TARGET_SOFTLIMIT:
1015 next = val + SOFTLIMIT_EVENTS_TARGET;
1017 case MEM_CGROUP_TARGET_NUMAINFO:
1018 next = val + NUMAINFO_EVENTS_TARGET;
1023 __this_cpu_write(memcg->stat->targets[target], next);
1030 * Check events in order.
1033 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1036 /* threshold event is triggered in finer grain than soft limit */
1037 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1038 MEM_CGROUP_TARGET_THRESH))) {
1040 bool do_numainfo __maybe_unused;
1042 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1043 MEM_CGROUP_TARGET_SOFTLIMIT);
1044 #if MAX_NUMNODES > 1
1045 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1046 MEM_CGROUP_TARGET_NUMAINFO);
1050 mem_cgroup_threshold(memcg);
1051 if (unlikely(do_softlimit))
1052 mem_cgroup_update_tree(memcg, page);
1053 #if MAX_NUMNODES > 1
1054 if (unlikely(do_numainfo))
1055 atomic_inc(&memcg->numainfo_events);
1061 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1064 * mm_update_next_owner() may clear mm->owner to NULL
1065 * if it races with swapoff, page migration, etc.
1066 * So this can be called with p == NULL.
1071 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1074 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1076 struct mem_cgroup *memcg = NULL;
1081 * Because we have no locks, mm->owner's may be being moved to other
1082 * cgroup. We use css_tryget() here even if this looks
1083 * pessimistic (rather than adding locks here).
1087 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1088 if (unlikely(!memcg))
1090 } while (!css_tryget(&memcg->css));
1096 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1097 * ref. count) or NULL if the whole root's subtree has been visited.
1099 * helper function to be used by mem_cgroup_iter
1101 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1102 struct mem_cgroup *last_visited)
1104 struct cgroup_subsys_state *prev_css, *next_css;
1106 prev_css = last_visited ? &last_visited->css : NULL;
1108 next_css = css_next_descendant_pre(prev_css, &root->css);
1111 * Even if we found a group we have to make sure it is
1112 * alive. css && !memcg means that the groups should be
1113 * skipped and we should continue the tree walk.
1114 * last_visited css is safe to use because it is
1115 * protected by css_get and the tree walk is rcu safe.
1117 * We do not take a reference on the root of the tree walk
1118 * because we might race with the root removal when it would
1119 * be the only node in the iterated hierarchy and mem_cgroup_iter
1120 * would end up in an endless loop because it expects that at
1121 * least one valid node will be returned. Root cannot disappear
1122 * because caller of the iterator should hold it already so
1123 * skipping css reference should be safe.
1126 if ((next_css == &root->css) ||
1127 ((next_css->flags & CSS_ONLINE) && css_tryget(next_css)))
1128 return mem_cgroup_from_css(next_css);
1130 prev_css = next_css;
1137 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1140 * When a group in the hierarchy below root is destroyed, the
1141 * hierarchy iterator can no longer be trusted since it might
1142 * have pointed to the destroyed group. Invalidate it.
1144 atomic_inc(&root->dead_count);
1147 static struct mem_cgroup *
1148 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1149 struct mem_cgroup *root,
1152 struct mem_cgroup *position = NULL;
1154 * A cgroup destruction happens in two stages: offlining and
1155 * release. They are separated by a RCU grace period.
1157 * If the iterator is valid, we may still race with an
1158 * offlining. The RCU lock ensures the object won't be
1159 * released, tryget will fail if we lost the race.
1161 *sequence = atomic_read(&root->dead_count);
1162 if (iter->last_dead_count == *sequence) {
1164 position = iter->last_visited;
1167 * We cannot take a reference to root because we might race
1168 * with root removal and returning NULL would end up in
1169 * an endless loop on the iterator user level when root
1170 * would be returned all the time.
1172 if (position && position != root &&
1173 !css_tryget(&position->css))
1179 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1180 struct mem_cgroup *last_visited,
1181 struct mem_cgroup *new_position,
1182 struct mem_cgroup *root,
1185 /* root reference counting symmetric to mem_cgroup_iter_load */
1186 if (last_visited && last_visited != root)
1187 css_put(&last_visited->css);
1189 * We store the sequence count from the time @last_visited was
1190 * loaded successfully instead of rereading it here so that we
1191 * don't lose destruction events in between. We could have
1192 * raced with the destruction of @new_position after all.
1194 iter->last_visited = new_position;
1196 iter->last_dead_count = sequence;
1200 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1201 * @root: hierarchy root
1202 * @prev: previously returned memcg, NULL on first invocation
1203 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1205 * Returns references to children of the hierarchy below @root, or
1206 * @root itself, or %NULL after a full round-trip.
1208 * Caller must pass the return value in @prev on subsequent
1209 * invocations for reference counting, or use mem_cgroup_iter_break()
1210 * to cancel a hierarchy walk before the round-trip is complete.
1212 * Reclaimers can specify a zone and a priority level in @reclaim to
1213 * divide up the memcgs in the hierarchy among all concurrent
1214 * reclaimers operating on the same zone and priority.
1216 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1217 struct mem_cgroup *prev,
1218 struct mem_cgroup_reclaim_cookie *reclaim)
1220 struct mem_cgroup *memcg = NULL;
1221 struct mem_cgroup *last_visited = NULL;
1223 if (mem_cgroup_disabled())
1227 root = root_mem_cgroup;
1229 if (prev && !reclaim)
1230 last_visited = prev;
1232 if (!root->use_hierarchy && root != root_mem_cgroup) {
1240 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1241 int uninitialized_var(seq);
1244 int nid = zone_to_nid(reclaim->zone);
1245 int zid = zone_idx(reclaim->zone);
1246 struct mem_cgroup_per_zone *mz;
1248 mz = mem_cgroup_zoneinfo(root, nid, zid);
1249 iter = &mz->reclaim_iter[reclaim->priority];
1250 if (prev && reclaim->generation != iter->generation) {
1251 iter->last_visited = NULL;
1255 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1258 memcg = __mem_cgroup_iter_next(root, last_visited);
1261 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1266 else if (!prev && memcg)
1267 reclaim->generation = iter->generation;
1276 if (prev && prev != root)
1277 css_put(&prev->css);
1283 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1284 * @root: hierarchy root
1285 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1287 void mem_cgroup_iter_break(struct mem_cgroup *root,
1288 struct mem_cgroup *prev)
1291 root = root_mem_cgroup;
1292 if (prev && prev != root)
1293 css_put(&prev->css);
1297 * Iteration constructs for visiting all cgroups (under a tree). If
1298 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1299 * be used for reference counting.
1301 #define for_each_mem_cgroup_tree(iter, root) \
1302 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1304 iter = mem_cgroup_iter(root, iter, NULL))
1306 #define for_each_mem_cgroup(iter) \
1307 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1309 iter = mem_cgroup_iter(NULL, iter, NULL))
1311 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1313 struct mem_cgroup *memcg;
1316 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1317 if (unlikely(!memcg))
1322 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1325 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1333 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1336 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1337 * @zone: zone of the wanted lruvec
1338 * @memcg: memcg of the wanted lruvec
1340 * Returns the lru list vector holding pages for the given @zone and
1341 * @mem. This can be the global zone lruvec, if the memory controller
1344 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1345 struct mem_cgroup *memcg)
1347 struct mem_cgroup_per_zone *mz;
1348 struct lruvec *lruvec;
1350 if (mem_cgroup_disabled()) {
1351 lruvec = &zone->lruvec;
1355 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1356 lruvec = &mz->lruvec;
1359 * Since a node can be onlined after the mem_cgroup was created,
1360 * we have to be prepared to initialize lruvec->zone here;
1361 * and if offlined then reonlined, we need to reinitialize it.
1363 if (unlikely(lruvec->zone != zone))
1364 lruvec->zone = zone;
1369 * Following LRU functions are allowed to be used without PCG_LOCK.
1370 * Operations are called by routine of global LRU independently from memcg.
1371 * What we have to take care of here is validness of pc->mem_cgroup.
1373 * Changes to pc->mem_cgroup happens when
1376 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1377 * It is added to LRU before charge.
1378 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1379 * When moving account, the page is not on LRU. It's isolated.
1383 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1385 * @zone: zone of the page
1387 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1389 struct mem_cgroup_per_zone *mz;
1390 struct mem_cgroup *memcg;
1391 struct page_cgroup *pc;
1392 struct lruvec *lruvec;
1394 if (mem_cgroup_disabled()) {
1395 lruvec = &zone->lruvec;
1399 pc = lookup_page_cgroup(page);
1400 memcg = pc->mem_cgroup;
1403 * Surreptitiously switch any uncharged offlist page to root:
1404 * an uncharged page off lru does nothing to secure
1405 * its former mem_cgroup from sudden removal.
1407 * Our caller holds lru_lock, and PageCgroupUsed is updated
1408 * under page_cgroup lock: between them, they make all uses
1409 * of pc->mem_cgroup safe.
1411 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1412 pc->mem_cgroup = memcg = root_mem_cgroup;
1414 mz = page_cgroup_zoneinfo(memcg, page);
1415 lruvec = &mz->lruvec;
1418 * Since a node can be onlined after the mem_cgroup was created,
1419 * we have to be prepared to initialize lruvec->zone here;
1420 * and if offlined then reonlined, we need to reinitialize it.
1422 if (unlikely(lruvec->zone != zone))
1423 lruvec->zone = zone;
1428 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1429 * @lruvec: mem_cgroup per zone lru vector
1430 * @lru: index of lru list the page is sitting on
1431 * @nr_pages: positive when adding or negative when removing
1433 * This function must be called when a page is added to or removed from an
1436 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1439 struct mem_cgroup_per_zone *mz;
1440 unsigned long *lru_size;
1442 if (mem_cgroup_disabled())
1445 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1446 lru_size = mz->lru_size + lru;
1447 *lru_size += nr_pages;
1448 VM_BUG_ON((long)(*lru_size) < 0);
1452 * Checks whether given mem is same or in the root_mem_cgroup's
1455 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1456 struct mem_cgroup *memcg)
1458 if (root_memcg == memcg)
1460 if (!root_memcg->use_hierarchy || !memcg)
1462 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1465 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1466 struct mem_cgroup *memcg)
1471 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1476 bool task_in_mem_cgroup(struct task_struct *task,
1477 const struct mem_cgroup *memcg)
1479 struct mem_cgroup *curr = NULL;
1480 struct task_struct *p;
1483 p = find_lock_task_mm(task);
1485 curr = try_get_mem_cgroup_from_mm(p->mm);
1489 * All threads may have already detached their mm's, but the oom
1490 * killer still needs to detect if they have already been oom
1491 * killed to prevent needlessly killing additional tasks.
1494 curr = mem_cgroup_from_task(task);
1496 css_get(&curr->css);
1502 * We should check use_hierarchy of "memcg" not "curr". Because checking
1503 * use_hierarchy of "curr" here make this function true if hierarchy is
1504 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1505 * hierarchy(even if use_hierarchy is disabled in "memcg").
1507 ret = mem_cgroup_same_or_subtree(memcg, curr);
1508 css_put(&curr->css);
1512 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1514 unsigned long inactive_ratio;
1515 unsigned long inactive;
1516 unsigned long active;
1519 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1520 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1522 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1524 inactive_ratio = int_sqrt(10 * gb);
1528 return inactive * inactive_ratio < active;
1531 #define mem_cgroup_from_res_counter(counter, member) \
1532 container_of(counter, struct mem_cgroup, member)
1535 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1536 * @memcg: the memory cgroup
1538 * Returns the maximum amount of memory @mem can be charged with, in
1541 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1543 unsigned long long margin;
1545 margin = res_counter_margin(&memcg->res);
1546 if (do_swap_account)
1547 margin = min(margin, res_counter_margin(&memcg->memsw));
1548 return margin >> PAGE_SHIFT;
1551 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1554 if (!css_parent(&memcg->css))
1555 return vm_swappiness;
1557 return memcg->swappiness;
1561 * memcg->moving_account is used for checking possibility that some thread is
1562 * calling move_account(). When a thread on CPU-A starts moving pages under
1563 * a memcg, other threads should check memcg->moving_account under
1564 * rcu_read_lock(), like this:
1568 * memcg->moving_account+1 if (memcg->mocing_account)
1570 * synchronize_rcu() update something.
1575 /* for quick checking without looking up memcg */
1576 atomic_t memcg_moving __read_mostly;
1578 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1580 atomic_inc(&memcg_moving);
1581 atomic_inc(&memcg->moving_account);
1585 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1588 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1589 * We check NULL in callee rather than caller.
1592 atomic_dec(&memcg_moving);
1593 atomic_dec(&memcg->moving_account);
1598 * 2 routines for checking "mem" is under move_account() or not.
1600 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1601 * is used for avoiding races in accounting. If true,
1602 * pc->mem_cgroup may be overwritten.
1604 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1605 * under hierarchy of moving cgroups. This is for
1606 * waiting at hith-memory prressure caused by "move".
1609 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1611 VM_BUG_ON(!rcu_read_lock_held());
1612 return atomic_read(&memcg->moving_account) > 0;
1615 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1617 struct mem_cgroup *from;
1618 struct mem_cgroup *to;
1621 * Unlike task_move routines, we access mc.to, mc.from not under
1622 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1624 spin_lock(&mc.lock);
1630 ret = mem_cgroup_same_or_subtree(memcg, from)
1631 || mem_cgroup_same_or_subtree(memcg, to);
1633 spin_unlock(&mc.lock);
1637 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1639 if (mc.moving_task && current != mc.moving_task) {
1640 if (mem_cgroup_under_move(memcg)) {
1642 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1643 /* moving charge context might have finished. */
1646 finish_wait(&mc.waitq, &wait);
1654 * Take this lock when
1655 * - a code tries to modify page's memcg while it's USED.
1656 * - a code tries to modify page state accounting in a memcg.
1657 * see mem_cgroup_stolen(), too.
1659 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1660 unsigned long *flags)
1662 spin_lock_irqsave(&memcg->move_lock, *flags);
1665 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1666 unsigned long *flags)
1668 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1671 #define K(x) ((x) << (PAGE_SHIFT-10))
1673 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1674 * @memcg: The memory cgroup that went over limit
1675 * @p: Task that is going to be killed
1677 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1680 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1682 /* oom_info_lock ensures that parallel ooms do not interleave */
1683 static DEFINE_MUTEX(oom_info_lock);
1684 struct mem_cgroup *iter;
1690 mutex_lock(&oom_info_lock);
1693 pr_info("Task in ");
1694 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1695 pr_info(" killed as a result of limit of ");
1696 pr_cont_cgroup_path(memcg->css.cgroup);
1701 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1702 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1703 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1704 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1705 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1706 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1707 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1708 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1709 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1710 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1711 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1712 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1714 for_each_mem_cgroup_tree(iter, memcg) {
1715 pr_info("Memory cgroup stats for ");
1716 pr_cont_cgroup_path(iter->css.cgroup);
1719 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1720 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1722 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1723 K(mem_cgroup_read_stat(iter, i)));
1726 for (i = 0; i < NR_LRU_LISTS; i++)
1727 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1728 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1732 mutex_unlock(&oom_info_lock);
1736 * This function returns the number of memcg under hierarchy tree. Returns
1737 * 1(self count) if no children.
1739 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1742 struct mem_cgroup *iter;
1744 for_each_mem_cgroup_tree(iter, memcg)
1750 * Return the memory (and swap, if configured) limit for a memcg.
1752 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1756 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1759 * Do not consider swap space if we cannot swap due to swappiness
1761 if (mem_cgroup_swappiness(memcg)) {
1764 limit += total_swap_pages << PAGE_SHIFT;
1765 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1768 * If memsw is finite and limits the amount of swap space
1769 * available to this memcg, return that limit.
1771 limit = min(limit, memsw);
1777 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1780 struct mem_cgroup *iter;
1781 unsigned long chosen_points = 0;
1782 unsigned long totalpages;
1783 unsigned int points = 0;
1784 struct task_struct *chosen = NULL;
1787 * If current has a pending SIGKILL or is exiting, then automatically
1788 * select it. The goal is to allow it to allocate so that it may
1789 * quickly exit and free its memory.
1791 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1792 set_thread_flag(TIF_MEMDIE);
1796 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1797 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1798 for_each_mem_cgroup_tree(iter, memcg) {
1799 struct css_task_iter it;
1800 struct task_struct *task;
1802 css_task_iter_start(&iter->css, &it);
1803 while ((task = css_task_iter_next(&it))) {
1804 switch (oom_scan_process_thread(task, totalpages, NULL,
1806 case OOM_SCAN_SELECT:
1808 put_task_struct(chosen);
1810 chosen_points = ULONG_MAX;
1811 get_task_struct(chosen);
1813 case OOM_SCAN_CONTINUE:
1815 case OOM_SCAN_ABORT:
1816 css_task_iter_end(&it);
1817 mem_cgroup_iter_break(memcg, iter);
1819 put_task_struct(chosen);
1824 points = oom_badness(task, memcg, NULL, totalpages);
1825 if (!points || points < chosen_points)
1827 /* Prefer thread group leaders for display purposes */
1828 if (points == chosen_points &&
1829 thread_group_leader(chosen))
1833 put_task_struct(chosen);
1835 chosen_points = points;
1836 get_task_struct(chosen);
1838 css_task_iter_end(&it);
1843 points = chosen_points * 1000 / totalpages;
1844 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1845 NULL, "Memory cgroup out of memory");
1848 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1850 unsigned long flags)
1852 unsigned long total = 0;
1853 bool noswap = false;
1856 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1858 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1861 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1863 drain_all_stock_async(memcg);
1864 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1866 * Allow limit shrinkers, which are triggered directly
1867 * by userspace, to catch signals and stop reclaim
1868 * after minimal progress, regardless of the margin.
1870 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1872 if (mem_cgroup_margin(memcg))
1875 * If nothing was reclaimed after two attempts, there
1876 * may be no reclaimable pages in this hierarchy.
1885 * test_mem_cgroup_node_reclaimable
1886 * @memcg: the target memcg
1887 * @nid: the node ID to be checked.
1888 * @noswap : specify true here if the user wants flle only information.
1890 * This function returns whether the specified memcg contains any
1891 * reclaimable pages on a node. Returns true if there are any reclaimable
1892 * pages in the node.
1894 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1895 int nid, bool noswap)
1897 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1899 if (noswap || !total_swap_pages)
1901 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1906 #if MAX_NUMNODES > 1
1909 * Always updating the nodemask is not very good - even if we have an empty
1910 * list or the wrong list here, we can start from some node and traverse all
1911 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1914 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1918 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1919 * pagein/pageout changes since the last update.
1921 if (!atomic_read(&memcg->numainfo_events))
1923 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1926 /* make a nodemask where this memcg uses memory from */
1927 memcg->scan_nodes = node_states[N_MEMORY];
1929 for_each_node_mask(nid, node_states[N_MEMORY]) {
1931 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1932 node_clear(nid, memcg->scan_nodes);
1935 atomic_set(&memcg->numainfo_events, 0);
1936 atomic_set(&memcg->numainfo_updating, 0);
1940 * Selecting a node where we start reclaim from. Because what we need is just
1941 * reducing usage counter, start from anywhere is O,K. Considering
1942 * memory reclaim from current node, there are pros. and cons.
1944 * Freeing memory from current node means freeing memory from a node which
1945 * we'll use or we've used. So, it may make LRU bad. And if several threads
1946 * hit limits, it will see a contention on a node. But freeing from remote
1947 * node means more costs for memory reclaim because of memory latency.
1949 * Now, we use round-robin. Better algorithm is welcomed.
1951 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1955 mem_cgroup_may_update_nodemask(memcg);
1956 node = memcg->last_scanned_node;
1958 node = next_node(node, memcg->scan_nodes);
1959 if (node == MAX_NUMNODES)
1960 node = first_node(memcg->scan_nodes);
1962 * We call this when we hit limit, not when pages are added to LRU.
1963 * No LRU may hold pages because all pages are UNEVICTABLE or
1964 * memcg is too small and all pages are not on LRU. In that case,
1965 * we use curret node.
1967 if (unlikely(node == MAX_NUMNODES))
1968 node = numa_node_id();
1970 memcg->last_scanned_node = node;
1975 * Check all nodes whether it contains reclaimable pages or not.
1976 * For quick scan, we make use of scan_nodes. This will allow us to skip
1977 * unused nodes. But scan_nodes is lazily updated and may not cotain
1978 * enough new information. We need to do double check.
1980 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1985 * quick check...making use of scan_node.
1986 * We can skip unused nodes.
1988 if (!nodes_empty(memcg->scan_nodes)) {
1989 for (nid = first_node(memcg->scan_nodes);
1991 nid = next_node(nid, memcg->scan_nodes)) {
1993 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1998 * Check rest of nodes.
2000 for_each_node_state(nid, N_MEMORY) {
2001 if (node_isset(nid, memcg->scan_nodes))
2003 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2010 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2015 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2017 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2021 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2024 unsigned long *total_scanned)
2026 struct mem_cgroup *victim = NULL;
2029 unsigned long excess;
2030 unsigned long nr_scanned;
2031 struct mem_cgroup_reclaim_cookie reclaim = {
2036 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2039 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2044 * If we have not been able to reclaim
2045 * anything, it might because there are
2046 * no reclaimable pages under this hierarchy
2051 * We want to do more targeted reclaim.
2052 * excess >> 2 is not to excessive so as to
2053 * reclaim too much, nor too less that we keep
2054 * coming back to reclaim from this cgroup
2056 if (total >= (excess >> 2) ||
2057 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2062 if (!mem_cgroup_reclaimable(victim, false))
2064 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2066 *total_scanned += nr_scanned;
2067 if (!res_counter_soft_limit_excess(&root_memcg->res))
2070 mem_cgroup_iter_break(root_memcg, victim);
2074 #ifdef CONFIG_LOCKDEP
2075 static struct lockdep_map memcg_oom_lock_dep_map = {
2076 .name = "memcg_oom_lock",
2080 static DEFINE_SPINLOCK(memcg_oom_lock);
2083 * Check OOM-Killer is already running under our hierarchy.
2084 * If someone is running, return false.
2086 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2088 struct mem_cgroup *iter, *failed = NULL;
2090 spin_lock(&memcg_oom_lock);
2092 for_each_mem_cgroup_tree(iter, memcg) {
2093 if (iter->oom_lock) {
2095 * this subtree of our hierarchy is already locked
2096 * so we cannot give a lock.
2099 mem_cgroup_iter_break(memcg, iter);
2102 iter->oom_lock = true;
2107 * OK, we failed to lock the whole subtree so we have
2108 * to clean up what we set up to the failing subtree
2110 for_each_mem_cgroup_tree(iter, memcg) {
2111 if (iter == failed) {
2112 mem_cgroup_iter_break(memcg, iter);
2115 iter->oom_lock = false;
2118 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2120 spin_unlock(&memcg_oom_lock);
2125 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2127 struct mem_cgroup *iter;
2129 spin_lock(&memcg_oom_lock);
2130 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2131 for_each_mem_cgroup_tree(iter, memcg)
2132 iter->oom_lock = false;
2133 spin_unlock(&memcg_oom_lock);
2136 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2138 struct mem_cgroup *iter;
2140 for_each_mem_cgroup_tree(iter, memcg)
2141 atomic_inc(&iter->under_oom);
2144 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2146 struct mem_cgroup *iter;
2149 * When a new child is created while the hierarchy is under oom,
2150 * mem_cgroup_oom_lock() may not be called. We have to use
2151 * atomic_add_unless() here.
2153 for_each_mem_cgroup_tree(iter, memcg)
2154 atomic_add_unless(&iter->under_oom, -1, 0);
2157 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2159 struct oom_wait_info {
2160 struct mem_cgroup *memcg;
2164 static int memcg_oom_wake_function(wait_queue_t *wait,
2165 unsigned mode, int sync, void *arg)
2167 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2168 struct mem_cgroup *oom_wait_memcg;
2169 struct oom_wait_info *oom_wait_info;
2171 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2172 oom_wait_memcg = oom_wait_info->memcg;
2175 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2176 * Then we can use css_is_ancestor without taking care of RCU.
2178 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2179 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2181 return autoremove_wake_function(wait, mode, sync, arg);
2184 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2186 atomic_inc(&memcg->oom_wakeups);
2187 /* for filtering, pass "memcg" as argument. */
2188 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2191 static void memcg_oom_recover(struct mem_cgroup *memcg)
2193 if (memcg && atomic_read(&memcg->under_oom))
2194 memcg_wakeup_oom(memcg);
2197 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2199 if (!current->memcg_oom.may_oom)
2202 * We are in the middle of the charge context here, so we
2203 * don't want to block when potentially sitting on a callstack
2204 * that holds all kinds of filesystem and mm locks.
2206 * Also, the caller may handle a failed allocation gracefully
2207 * (like optional page cache readahead) and so an OOM killer
2208 * invocation might not even be necessary.
2210 * That's why we don't do anything here except remember the
2211 * OOM context and then deal with it at the end of the page
2212 * fault when the stack is unwound, the locks are released,
2213 * and when we know whether the fault was overall successful.
2215 css_get(&memcg->css);
2216 current->memcg_oom.memcg = memcg;
2217 current->memcg_oom.gfp_mask = mask;
2218 current->memcg_oom.order = order;
2222 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2223 * @handle: actually kill/wait or just clean up the OOM state
2225 * This has to be called at the end of a page fault if the memcg OOM
2226 * handler was enabled.
2228 * Memcg supports userspace OOM handling where failed allocations must
2229 * sleep on a waitqueue until the userspace task resolves the
2230 * situation. Sleeping directly in the charge context with all kinds
2231 * of locks held is not a good idea, instead we remember an OOM state
2232 * in the task and mem_cgroup_oom_synchronize() has to be called at
2233 * the end of the page fault to complete the OOM handling.
2235 * Returns %true if an ongoing memcg OOM situation was detected and
2236 * completed, %false otherwise.
2238 bool mem_cgroup_oom_synchronize(bool handle)
2240 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2241 struct oom_wait_info owait;
2244 /* OOM is global, do not handle */
2251 owait.memcg = memcg;
2252 owait.wait.flags = 0;
2253 owait.wait.func = memcg_oom_wake_function;
2254 owait.wait.private = current;
2255 INIT_LIST_HEAD(&owait.wait.task_list);
2257 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2258 mem_cgroup_mark_under_oom(memcg);
2260 locked = mem_cgroup_oom_trylock(memcg);
2263 mem_cgroup_oom_notify(memcg);
2265 if (locked && !memcg->oom_kill_disable) {
2266 mem_cgroup_unmark_under_oom(memcg);
2267 finish_wait(&memcg_oom_waitq, &owait.wait);
2268 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2269 current->memcg_oom.order);
2272 mem_cgroup_unmark_under_oom(memcg);
2273 finish_wait(&memcg_oom_waitq, &owait.wait);
2277 mem_cgroup_oom_unlock(memcg);
2279 * There is no guarantee that an OOM-lock contender
2280 * sees the wakeups triggered by the OOM kill
2281 * uncharges. Wake any sleepers explicitely.
2283 memcg_oom_recover(memcg);
2286 current->memcg_oom.memcg = NULL;
2287 css_put(&memcg->css);
2292 * Currently used to update mapped file statistics, but the routine can be
2293 * generalized to update other statistics as well.
2295 * Notes: Race condition
2297 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2298 * it tends to be costly. But considering some conditions, we doesn't need
2299 * to do so _always_.
2301 * Considering "charge", lock_page_cgroup() is not required because all
2302 * file-stat operations happen after a page is attached to radix-tree. There
2303 * are no race with "charge".
2305 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2306 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2307 * if there are race with "uncharge". Statistics itself is properly handled
2310 * Considering "move", this is an only case we see a race. To make the race
2311 * small, we check mm->moving_account and detect there are possibility of race
2312 * If there is, we take a lock.
2315 void __mem_cgroup_begin_update_page_stat(struct page *page,
2316 bool *locked, unsigned long *flags)
2318 struct mem_cgroup *memcg;
2319 struct page_cgroup *pc;
2321 pc = lookup_page_cgroup(page);
2323 memcg = pc->mem_cgroup;
2324 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2327 * If this memory cgroup is not under account moving, we don't
2328 * need to take move_lock_mem_cgroup(). Because we already hold
2329 * rcu_read_lock(), any calls to move_account will be delayed until
2330 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2332 if (!mem_cgroup_stolen(memcg))
2335 move_lock_mem_cgroup(memcg, flags);
2336 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2337 move_unlock_mem_cgroup(memcg, flags);
2343 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2345 struct page_cgroup *pc = lookup_page_cgroup(page);
2348 * It's guaranteed that pc->mem_cgroup never changes while
2349 * lock is held because a routine modifies pc->mem_cgroup
2350 * should take move_lock_mem_cgroup().
2352 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2355 void mem_cgroup_update_page_stat(struct page *page,
2356 enum mem_cgroup_stat_index idx, int val)
2358 struct mem_cgroup *memcg;
2359 struct page_cgroup *pc = lookup_page_cgroup(page);
2360 unsigned long uninitialized_var(flags);
2362 if (mem_cgroup_disabled())
2365 VM_BUG_ON(!rcu_read_lock_held());
2366 memcg = pc->mem_cgroup;
2367 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2370 this_cpu_add(memcg->stat->count[idx], val);
2374 * size of first charge trial. "32" comes from vmscan.c's magic value.
2375 * TODO: maybe necessary to use big numbers in big irons.
2377 #define CHARGE_BATCH 32U
2378 struct memcg_stock_pcp {
2379 struct mem_cgroup *cached; /* this never be root cgroup */
2380 unsigned int nr_pages;
2381 struct work_struct work;
2382 unsigned long flags;
2383 #define FLUSHING_CACHED_CHARGE 0
2385 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2386 static DEFINE_MUTEX(percpu_charge_mutex);
2389 * consume_stock: Try to consume stocked charge on this cpu.
2390 * @memcg: memcg to consume from.
2391 * @nr_pages: how many pages to charge.
2393 * The charges will only happen if @memcg matches the current cpu's memcg
2394 * stock, and at least @nr_pages are available in that stock. Failure to
2395 * service an allocation will refill the stock.
2397 * returns true if successful, false otherwise.
2399 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2401 struct memcg_stock_pcp *stock;
2404 if (nr_pages > CHARGE_BATCH)
2407 stock = &get_cpu_var(memcg_stock);
2408 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2409 stock->nr_pages -= nr_pages;
2410 else /* need to call res_counter_charge */
2412 put_cpu_var(memcg_stock);
2417 * Returns stocks cached in percpu to res_counter and reset cached information.
2419 static void drain_stock(struct memcg_stock_pcp *stock)
2421 struct mem_cgroup *old = stock->cached;
2423 if (stock->nr_pages) {
2424 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2426 res_counter_uncharge(&old->res, bytes);
2427 if (do_swap_account)
2428 res_counter_uncharge(&old->memsw, bytes);
2429 stock->nr_pages = 0;
2431 stock->cached = NULL;
2435 * This must be called under preempt disabled or must be called by
2436 * a thread which is pinned to local cpu.
2438 static void drain_local_stock(struct work_struct *dummy)
2440 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2442 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2445 static void __init memcg_stock_init(void)
2449 for_each_possible_cpu(cpu) {
2450 struct memcg_stock_pcp *stock =
2451 &per_cpu(memcg_stock, cpu);
2452 INIT_WORK(&stock->work, drain_local_stock);
2457 * Cache charges(val) which is from res_counter, to local per_cpu area.
2458 * This will be consumed by consume_stock() function, later.
2460 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2462 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2464 if (stock->cached != memcg) { /* reset if necessary */
2466 stock->cached = memcg;
2468 stock->nr_pages += nr_pages;
2469 put_cpu_var(memcg_stock);
2473 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2474 * of the hierarchy under it. sync flag says whether we should block
2475 * until the work is done.
2477 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2481 /* Notify other cpus that system-wide "drain" is running */
2484 for_each_online_cpu(cpu) {
2485 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2486 struct mem_cgroup *memcg;
2488 memcg = stock->cached;
2489 if (!memcg || !stock->nr_pages)
2491 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2493 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2495 drain_local_stock(&stock->work);
2497 schedule_work_on(cpu, &stock->work);
2505 for_each_online_cpu(cpu) {
2506 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2507 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2508 flush_work(&stock->work);
2515 * Tries to drain stocked charges in other cpus. This function is asynchronous
2516 * and just put a work per cpu for draining localy on each cpu. Caller can
2517 * expects some charges will be back to res_counter later but cannot wait for
2520 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2523 * If someone calls draining, avoid adding more kworker runs.
2525 if (!mutex_trylock(&percpu_charge_mutex))
2527 drain_all_stock(root_memcg, false);
2528 mutex_unlock(&percpu_charge_mutex);
2531 /* This is a synchronous drain interface. */
2532 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2534 /* called when force_empty is called */
2535 mutex_lock(&percpu_charge_mutex);
2536 drain_all_stock(root_memcg, true);
2537 mutex_unlock(&percpu_charge_mutex);
2541 * This function drains percpu counter value from DEAD cpu and
2542 * move it to local cpu. Note that this function can be preempted.
2544 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2548 spin_lock(&memcg->pcp_counter_lock);
2549 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2550 long x = per_cpu(memcg->stat->count[i], cpu);
2552 per_cpu(memcg->stat->count[i], cpu) = 0;
2553 memcg->nocpu_base.count[i] += x;
2555 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2556 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2558 per_cpu(memcg->stat->events[i], cpu) = 0;
2559 memcg->nocpu_base.events[i] += x;
2561 spin_unlock(&memcg->pcp_counter_lock);
2564 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2565 unsigned long action,
2568 int cpu = (unsigned long)hcpu;
2569 struct memcg_stock_pcp *stock;
2570 struct mem_cgroup *iter;
2572 if (action == CPU_ONLINE)
2575 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2578 for_each_mem_cgroup(iter)
2579 mem_cgroup_drain_pcp_counter(iter, cpu);
2581 stock = &per_cpu(memcg_stock, cpu);
2587 /* See __mem_cgroup_try_charge() for details */
2589 CHARGE_OK, /* success */
2590 CHARGE_RETRY, /* need to retry but retry is not bad */
2591 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2592 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2595 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2596 unsigned int nr_pages, unsigned int min_pages,
2599 unsigned long csize = nr_pages * PAGE_SIZE;
2600 struct mem_cgroup *mem_over_limit;
2601 struct res_counter *fail_res;
2602 unsigned long flags = 0;
2605 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2608 if (!do_swap_account)
2610 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2614 res_counter_uncharge(&memcg->res, csize);
2615 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2616 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2618 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2620 * Never reclaim on behalf of optional batching, retry with a
2621 * single page instead.
2623 if (nr_pages > min_pages)
2624 return CHARGE_RETRY;
2626 if (!(gfp_mask & __GFP_WAIT))
2627 return CHARGE_WOULDBLOCK;
2629 if (gfp_mask & __GFP_NORETRY)
2630 return CHARGE_NOMEM;
2632 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2633 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2634 return CHARGE_RETRY;
2636 * Even though the limit is exceeded at this point, reclaim
2637 * may have been able to free some pages. Retry the charge
2638 * before killing the task.
2640 * Only for regular pages, though: huge pages are rather
2641 * unlikely to succeed so close to the limit, and we fall back
2642 * to regular pages anyway in case of failure.
2644 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2645 return CHARGE_RETRY;
2648 * At task move, charge accounts can be doubly counted. So, it's
2649 * better to wait until the end of task_move if something is going on.
2651 if (mem_cgroup_wait_acct_move(mem_over_limit))
2652 return CHARGE_RETRY;
2655 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2657 return CHARGE_NOMEM;
2661 * __mem_cgroup_try_charge() does
2662 * 1. detect memcg to be charged against from passed *mm and *ptr,
2663 * 2. update res_counter
2664 * 3. call memory reclaim if necessary.
2666 * In some special case, if the task is fatal, fatal_signal_pending() or
2667 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2668 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2669 * as possible without any hazards. 2: all pages should have a valid
2670 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2671 * pointer, that is treated as a charge to root_mem_cgroup.
2673 * So __mem_cgroup_try_charge() will return
2674 * 0 ... on success, filling *ptr with a valid memcg pointer.
2675 * -ENOMEM ... charge failure because of resource limits.
2676 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2678 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2679 * the oom-killer can be invoked.
2681 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2683 unsigned int nr_pages,
2684 struct mem_cgroup **ptr,
2687 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2688 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2689 struct mem_cgroup *memcg = NULL;
2693 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2694 * in system level. So, allow to go ahead dying process in addition to
2697 if (unlikely(test_thread_flag(TIF_MEMDIE)
2698 || fatal_signal_pending(current)))
2701 if (unlikely(task_in_memcg_oom(current)))
2704 if (gfp_mask & __GFP_NOFAIL)
2708 * We always charge the cgroup the mm_struct belongs to.
2709 * The mm_struct's mem_cgroup changes on task migration if the
2710 * thread group leader migrates. It's possible that mm is not
2711 * set, if so charge the root memcg (happens for pagecache usage).
2714 *ptr = root_mem_cgroup;
2716 if (*ptr) { /* css should be a valid one */
2718 if (mem_cgroup_is_root(memcg))
2720 if (consume_stock(memcg, nr_pages))
2722 css_get(&memcg->css);
2724 struct task_struct *p;
2727 p = rcu_dereference(mm->owner);
2729 * Because we don't have task_lock(), "p" can exit.
2730 * In that case, "memcg" can point to root or p can be NULL with
2731 * race with swapoff. Then, we have small risk of mis-accouning.
2732 * But such kind of mis-account by race always happens because
2733 * we don't have cgroup_mutex(). It's overkill and we allo that
2735 * (*) swapoff at el will charge against mm-struct not against
2736 * task-struct. So, mm->owner can be NULL.
2738 memcg = mem_cgroup_from_task(p);
2740 memcg = root_mem_cgroup;
2741 if (mem_cgroup_is_root(memcg)) {
2745 if (consume_stock(memcg, nr_pages)) {
2747 * It seems dagerous to access memcg without css_get().
2748 * But considering how consume_stok works, it's not
2749 * necessary. If consume_stock success, some charges
2750 * from this memcg are cached on this cpu. So, we
2751 * don't need to call css_get()/css_tryget() before
2752 * calling consume_stock().
2757 /* after here, we may be blocked. we need to get refcnt */
2758 if (!css_tryget(&memcg->css)) {
2766 bool invoke_oom = oom && !nr_oom_retries;
2768 /* If killed, bypass charge */
2769 if (fatal_signal_pending(current)) {
2770 css_put(&memcg->css);
2774 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2775 nr_pages, invoke_oom);
2779 case CHARGE_RETRY: /* not in OOM situation but retry */
2781 css_put(&memcg->css);
2784 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2785 css_put(&memcg->css);
2787 case CHARGE_NOMEM: /* OOM routine works */
2788 if (!oom || invoke_oom) {
2789 css_put(&memcg->css);
2795 } while (ret != CHARGE_OK);
2797 if (batch > nr_pages)
2798 refill_stock(memcg, batch - nr_pages);
2799 css_put(&memcg->css);
2804 if (!(gfp_mask & __GFP_NOFAIL)) {
2809 *ptr = root_mem_cgroup;
2814 * Somemtimes we have to undo a charge we got by try_charge().
2815 * This function is for that and do uncharge, put css's refcnt.
2816 * gotten by try_charge().
2818 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2819 unsigned int nr_pages)
2821 if (!mem_cgroup_is_root(memcg)) {
2822 unsigned long bytes = nr_pages * PAGE_SIZE;
2824 res_counter_uncharge(&memcg->res, bytes);
2825 if (do_swap_account)
2826 res_counter_uncharge(&memcg->memsw, bytes);
2831 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2832 * This is useful when moving usage to parent cgroup.
2834 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2835 unsigned int nr_pages)
2837 unsigned long bytes = nr_pages * PAGE_SIZE;
2839 if (mem_cgroup_is_root(memcg))
2842 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2843 if (do_swap_account)
2844 res_counter_uncharge_until(&memcg->memsw,
2845 memcg->memsw.parent, bytes);
2849 * A helper function to get mem_cgroup from ID. must be called under
2850 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2851 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2852 * called against removed memcg.)
2854 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2856 /* ID 0 is unused ID */
2859 return mem_cgroup_from_id(id);
2862 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2864 struct mem_cgroup *memcg = NULL;
2865 struct page_cgroup *pc;
2869 VM_BUG_ON_PAGE(!PageLocked(page), page);
2871 pc = lookup_page_cgroup(page);
2872 lock_page_cgroup(pc);
2873 if (PageCgroupUsed(pc)) {
2874 memcg = pc->mem_cgroup;
2875 if (memcg && !css_tryget(&memcg->css))
2877 } else if (PageSwapCache(page)) {
2878 ent.val = page_private(page);
2879 id = lookup_swap_cgroup_id(ent);
2881 memcg = mem_cgroup_lookup(id);
2882 if (memcg && !css_tryget(&memcg->css))
2886 unlock_page_cgroup(pc);
2890 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2892 unsigned int nr_pages,
2893 enum charge_type ctype,
2896 struct page_cgroup *pc = lookup_page_cgroup(page);
2897 struct zone *uninitialized_var(zone);
2898 struct lruvec *lruvec;
2899 bool was_on_lru = false;
2902 lock_page_cgroup(pc);
2903 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2905 * we don't need page_cgroup_lock about tail pages, becase they are not
2906 * accessed by any other context at this point.
2910 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2911 * may already be on some other mem_cgroup's LRU. Take care of it.
2914 zone = page_zone(page);
2915 spin_lock_irq(&zone->lru_lock);
2916 if (PageLRU(page)) {
2917 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2919 del_page_from_lru_list(page, lruvec, page_lru(page));
2924 pc->mem_cgroup = memcg;
2926 * We access a page_cgroup asynchronously without lock_page_cgroup().
2927 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2928 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2929 * before USED bit, we need memory barrier here.
2930 * See mem_cgroup_add_lru_list(), etc.
2933 SetPageCgroupUsed(pc);
2937 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2938 VM_BUG_ON_PAGE(PageLRU(page), page);
2940 add_page_to_lru_list(page, lruvec, page_lru(page));
2942 spin_unlock_irq(&zone->lru_lock);
2945 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2950 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2951 unlock_page_cgroup(pc);
2954 * "charge_statistics" updated event counter. Then, check it.
2955 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2956 * if they exceeds softlimit.
2958 memcg_check_events(memcg, page);
2961 static DEFINE_MUTEX(set_limit_mutex);
2963 #ifdef CONFIG_MEMCG_KMEM
2964 static DEFINE_MUTEX(activate_kmem_mutex);
2966 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2968 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2969 memcg_kmem_is_active(memcg);
2973 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2974 * in the memcg_cache_params struct.
2976 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2978 struct kmem_cache *cachep;
2980 VM_BUG_ON(p->is_root_cache);
2981 cachep = p->root_cache;
2982 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2985 #ifdef CONFIG_SLABINFO
2986 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2988 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2989 struct memcg_cache_params *params;
2991 if (!memcg_can_account_kmem(memcg))
2994 print_slabinfo_header(m);
2996 mutex_lock(&memcg->slab_caches_mutex);
2997 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2998 cache_show(memcg_params_to_cache(params), m);
2999 mutex_unlock(&memcg->slab_caches_mutex);
3005 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3007 struct res_counter *fail_res;
3008 struct mem_cgroup *_memcg;
3011 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3016 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3017 &_memcg, oom_gfp_allowed(gfp));
3019 if (ret == -EINTR) {
3021 * __mem_cgroup_try_charge() chosed to bypass to root due to
3022 * OOM kill or fatal signal. Since our only options are to
3023 * either fail the allocation or charge it to this cgroup, do
3024 * it as a temporary condition. But we can't fail. From a
3025 * kmem/slab perspective, the cache has already been selected,
3026 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3029 * This condition will only trigger if the task entered
3030 * memcg_charge_kmem in a sane state, but was OOM-killed during
3031 * __mem_cgroup_try_charge() above. Tasks that were already
3032 * dying when the allocation triggers should have been already
3033 * directed to the root cgroup in memcontrol.h
3035 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3036 if (do_swap_account)
3037 res_counter_charge_nofail(&memcg->memsw, size,
3041 res_counter_uncharge(&memcg->kmem, size);
3046 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3048 res_counter_uncharge(&memcg->res, size);
3049 if (do_swap_account)
3050 res_counter_uncharge(&memcg->memsw, size);
3053 if (res_counter_uncharge(&memcg->kmem, size))
3057 * Releases a reference taken in kmem_cgroup_css_offline in case
3058 * this last uncharge is racing with the offlining code or it is
3059 * outliving the memcg existence.
3061 * The memory barrier imposed by test&clear is paired with the
3062 * explicit one in memcg_kmem_mark_dead().
3064 if (memcg_kmem_test_and_clear_dead(memcg))
3065 css_put(&memcg->css);
3069 * helper for acessing a memcg's index. It will be used as an index in the
3070 * child cache array in kmem_cache, and also to derive its name. This function
3071 * will return -1 when this is not a kmem-limited memcg.
3073 int memcg_cache_id(struct mem_cgroup *memcg)
3075 return memcg ? memcg->kmemcg_id : -1;
3078 static size_t memcg_caches_array_size(int num_groups)
3081 if (num_groups <= 0)
3084 size = 2 * num_groups;
3085 if (size < MEMCG_CACHES_MIN_SIZE)
3086 size = MEMCG_CACHES_MIN_SIZE;
3087 else if (size > MEMCG_CACHES_MAX_SIZE)
3088 size = MEMCG_CACHES_MAX_SIZE;
3094 * We should update the current array size iff all caches updates succeed. This
3095 * can only be done from the slab side. The slab mutex needs to be held when
3098 void memcg_update_array_size(int num)
3100 if (num > memcg_limited_groups_array_size)
3101 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3104 static void kmem_cache_destroy_work_func(struct work_struct *w);
3106 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3108 struct memcg_cache_params *cur_params = s->memcg_params;
3110 VM_BUG_ON(!is_root_cache(s));
3112 if (num_groups > memcg_limited_groups_array_size) {
3114 struct memcg_cache_params *new_params;
3115 ssize_t size = memcg_caches_array_size(num_groups);
3117 size *= sizeof(void *);
3118 size += offsetof(struct memcg_cache_params, memcg_caches);
3120 new_params = kzalloc(size, GFP_KERNEL);
3124 new_params->is_root_cache = true;
3127 * There is the chance it will be bigger than
3128 * memcg_limited_groups_array_size, if we failed an allocation
3129 * in a cache, in which case all caches updated before it, will
3130 * have a bigger array.
3132 * But if that is the case, the data after
3133 * memcg_limited_groups_array_size is certainly unused
3135 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3136 if (!cur_params->memcg_caches[i])
3138 new_params->memcg_caches[i] =
3139 cur_params->memcg_caches[i];
3143 * Ideally, we would wait until all caches succeed, and only
3144 * then free the old one. But this is not worth the extra
3145 * pointer per-cache we'd have to have for this.
3147 * It is not a big deal if some caches are left with a size
3148 * bigger than the others. And all updates will reset this
3151 rcu_assign_pointer(s->memcg_params, new_params);
3153 kfree_rcu(cur_params, rcu_head);
3158 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3159 struct kmem_cache *root_cache)
3163 if (!memcg_kmem_enabled())
3167 size = offsetof(struct memcg_cache_params, memcg_caches);
3168 size += memcg_limited_groups_array_size * sizeof(void *);
3170 size = sizeof(struct memcg_cache_params);
3172 s->memcg_params = kzalloc(size, GFP_KERNEL);
3173 if (!s->memcg_params)
3177 s->memcg_params->memcg = memcg;
3178 s->memcg_params->root_cache = root_cache;
3179 INIT_WORK(&s->memcg_params->destroy,
3180 kmem_cache_destroy_work_func);
3182 s->memcg_params->is_root_cache = true;
3187 void memcg_free_cache_params(struct kmem_cache *s)
3189 kfree(s->memcg_params);
3192 void memcg_register_cache(struct kmem_cache *s)
3194 struct kmem_cache *root;
3195 struct mem_cgroup *memcg;
3198 if (is_root_cache(s))
3202 * Holding the slab_mutex assures nobody will touch the memcg_caches
3203 * array while we are modifying it.
3205 lockdep_assert_held(&slab_mutex);
3207 root = s->memcg_params->root_cache;
3208 memcg = s->memcg_params->memcg;
3209 id = memcg_cache_id(memcg);
3211 css_get(&memcg->css);
3215 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3216 * barrier here to ensure nobody will see the kmem_cache partially
3222 * Initialize the pointer to this cache in its parent's memcg_params
3223 * before adding it to the memcg_slab_caches list, otherwise we can
3224 * fail to convert memcg_params_to_cache() while traversing the list.
3226 VM_BUG_ON(root->memcg_params->memcg_caches[id]);
3227 root->memcg_params->memcg_caches[id] = s;
3229 mutex_lock(&memcg->slab_caches_mutex);
3230 list_add(&s->memcg_params->list, &memcg->memcg_slab_caches);
3231 mutex_unlock(&memcg->slab_caches_mutex);
3234 void memcg_unregister_cache(struct kmem_cache *s)
3236 struct kmem_cache *root;
3237 struct mem_cgroup *memcg;
3240 if (is_root_cache(s))
3244 * Holding the slab_mutex assures nobody will touch the memcg_caches
3245 * array while we are modifying it.
3247 lockdep_assert_held(&slab_mutex);
3249 root = s->memcg_params->root_cache;
3250 memcg = s->memcg_params->memcg;
3251 id = memcg_cache_id(memcg);
3253 mutex_lock(&memcg->slab_caches_mutex);
3254 list_del(&s->memcg_params->list);
3255 mutex_unlock(&memcg->slab_caches_mutex);
3258 * Clear the pointer to this cache in its parent's memcg_params only
3259 * after removing it from the memcg_slab_caches list, otherwise we can
3260 * fail to convert memcg_params_to_cache() while traversing the list.
3262 VM_BUG_ON(!root->memcg_params->memcg_caches[id]);
3263 root->memcg_params->memcg_caches[id] = NULL;
3265 css_put(&memcg->css);
3269 * During the creation a new cache, we need to disable our accounting mechanism
3270 * altogether. This is true even if we are not creating, but rather just
3271 * enqueing new caches to be created.
3273 * This is because that process will trigger allocations; some visible, like
3274 * explicit kmallocs to auxiliary data structures, name strings and internal
3275 * cache structures; some well concealed, like INIT_WORK() that can allocate
3276 * objects during debug.
3278 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3279 * to it. This may not be a bounded recursion: since the first cache creation
3280 * failed to complete (waiting on the allocation), we'll just try to create the
3281 * cache again, failing at the same point.
3283 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3284 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3285 * inside the following two functions.
3287 static inline void memcg_stop_kmem_account(void)
3289 VM_BUG_ON(!current->mm);
3290 current->memcg_kmem_skip_account++;
3293 static inline void memcg_resume_kmem_account(void)
3295 VM_BUG_ON(!current->mm);
3296 current->memcg_kmem_skip_account--;
3299 static void kmem_cache_destroy_work_func(struct work_struct *w)
3301 struct kmem_cache *cachep;
3302 struct memcg_cache_params *p;
3304 p = container_of(w, struct memcg_cache_params, destroy);
3306 cachep = memcg_params_to_cache(p);
3309 * If we get down to 0 after shrink, we could delete right away.
3310 * However, memcg_release_pages() already puts us back in the workqueue
3311 * in that case. If we proceed deleting, we'll get a dangling
3312 * reference, and removing the object from the workqueue in that case
3313 * is unnecessary complication. We are not a fast path.
3315 * Note that this case is fundamentally different from racing with
3316 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3317 * kmem_cache_shrink, not only we would be reinserting a dead cache
3318 * into the queue, but doing so from inside the worker racing to
3321 * So if we aren't down to zero, we'll just schedule a worker and try
3324 if (atomic_read(&cachep->memcg_params->nr_pages) != 0)
3325 kmem_cache_shrink(cachep);
3327 kmem_cache_destroy(cachep);
3330 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3332 if (!cachep->memcg_params->dead)
3336 * There are many ways in which we can get here.
3338 * We can get to a memory-pressure situation while the delayed work is
3339 * still pending to run. The vmscan shrinkers can then release all
3340 * cache memory and get us to destruction. If this is the case, we'll
3341 * be executed twice, which is a bug (the second time will execute over
3342 * bogus data). In this case, cancelling the work should be fine.
3344 * But we can also get here from the worker itself, if
3345 * kmem_cache_shrink is enough to shake all the remaining objects and
3346 * get the page count to 0. In this case, we'll deadlock if we try to
3347 * cancel the work (the worker runs with an internal lock held, which
3348 * is the same lock we would hold for cancel_work_sync().)
3350 * Since we can't possibly know who got us here, just refrain from
3351 * running if there is already work pending
3353 if (work_pending(&cachep->memcg_params->destroy))
3356 * We have to defer the actual destroying to a workqueue, because
3357 * we might currently be in a context that cannot sleep.
3359 schedule_work(&cachep->memcg_params->destroy);
3362 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3363 struct kmem_cache *s)
3365 struct kmem_cache *new = NULL;
3366 static char *tmp_path = NULL, *tmp_name = NULL;
3367 static DEFINE_MUTEX(mutex); /* protects tmp_name */
3369 BUG_ON(!memcg_can_account_kmem(memcg));
3373 * kmem_cache_create_memcg duplicates the given name and
3374 * cgroup_name for this name requires RCU context.
3375 * This static temporary buffer is used to prevent from
3376 * pointless shortliving allocation.
3378 if (!tmp_path || !tmp_name) {
3380 tmp_path = kmalloc(PATH_MAX, GFP_KERNEL);
3382 tmp_name = kmalloc(NAME_MAX + 1, GFP_KERNEL);
3383 if (!tmp_path || !tmp_name)
3387 cgroup_name(memcg->css.cgroup, tmp_name, NAME_MAX + 1);
3388 snprintf(tmp_path, PATH_MAX, "%s(%d:%s)", s->name,
3389 memcg_cache_id(memcg), tmp_name);
3391 new = kmem_cache_create_memcg(memcg, tmp_path, s->object_size, s->align,
3392 (s->flags & ~SLAB_PANIC), s->ctor, s);
3394 new->allocflags |= __GFP_KMEMCG;
3398 mutex_unlock(&mutex);
3402 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3404 struct kmem_cache *c;
3407 if (!s->memcg_params)
3409 if (!s->memcg_params->is_root_cache)
3413 * If the cache is being destroyed, we trust that there is no one else
3414 * requesting objects from it. Even if there are, the sanity checks in
3415 * kmem_cache_destroy should caught this ill-case.
3417 * Still, we don't want anyone else freeing memcg_caches under our
3418 * noses, which can happen if a new memcg comes to life. As usual,
3419 * we'll take the activate_kmem_mutex to protect ourselves against
3422 mutex_lock(&activate_kmem_mutex);
3423 for_each_memcg_cache_index(i) {
3424 c = cache_from_memcg_idx(s, i);
3429 * We will now manually delete the caches, so to avoid races
3430 * we need to cancel all pending destruction workers and
3431 * proceed with destruction ourselves.
3433 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3434 * and that could spawn the workers again: it is likely that
3435 * the cache still have active pages until this very moment.
3436 * This would lead us back to mem_cgroup_destroy_cache.
3438 * But that will not execute at all if the "dead" flag is not
3439 * set, so flip it down to guarantee we are in control.
3441 c->memcg_params->dead = false;
3442 cancel_work_sync(&c->memcg_params->destroy);
3443 kmem_cache_destroy(c);
3445 mutex_unlock(&activate_kmem_mutex);
3448 struct create_work {
3449 struct mem_cgroup *memcg;
3450 struct kmem_cache *cachep;
3451 struct work_struct work;
3454 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3456 struct kmem_cache *cachep;
3457 struct memcg_cache_params *params;
3459 if (!memcg_kmem_is_active(memcg))
3462 mutex_lock(&memcg->slab_caches_mutex);
3463 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3464 cachep = memcg_params_to_cache(params);
3465 cachep->memcg_params->dead = true;
3466 schedule_work(&cachep->memcg_params->destroy);
3468 mutex_unlock(&memcg->slab_caches_mutex);
3471 static void memcg_create_cache_work_func(struct work_struct *w)
3473 struct create_work *cw;
3475 cw = container_of(w, struct create_work, work);
3476 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3477 css_put(&cw->memcg->css);
3482 * Enqueue the creation of a per-memcg kmem_cache.
3484 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3485 struct kmem_cache *cachep)
3487 struct create_work *cw;
3489 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3491 css_put(&memcg->css);
3496 cw->cachep = cachep;
3498 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3499 schedule_work(&cw->work);
3502 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3503 struct kmem_cache *cachep)
3506 * We need to stop accounting when we kmalloc, because if the
3507 * corresponding kmalloc cache is not yet created, the first allocation
3508 * in __memcg_create_cache_enqueue will recurse.
3510 * However, it is better to enclose the whole function. Depending on
3511 * the debugging options enabled, INIT_WORK(), for instance, can
3512 * trigger an allocation. This too, will make us recurse. Because at
3513 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3514 * the safest choice is to do it like this, wrapping the whole function.
3516 memcg_stop_kmem_account();
3517 __memcg_create_cache_enqueue(memcg, cachep);
3518 memcg_resume_kmem_account();
3521 * Return the kmem_cache we're supposed to use for a slab allocation.
3522 * We try to use the current memcg's version of the cache.
3524 * If the cache does not exist yet, if we are the first user of it,
3525 * we either create it immediately, if possible, or create it asynchronously
3527 * In the latter case, we will let the current allocation go through with
3528 * the original cache.
3530 * Can't be called in interrupt context or from kernel threads.
3531 * This function needs to be called with rcu_read_lock() held.
3533 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3536 struct mem_cgroup *memcg;
3537 struct kmem_cache *memcg_cachep;
3539 VM_BUG_ON(!cachep->memcg_params);
3540 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3542 if (!current->mm || current->memcg_kmem_skip_account)
3546 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3548 if (!memcg_can_account_kmem(memcg))
3551 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3552 if (likely(memcg_cachep)) {
3553 cachep = memcg_cachep;
3557 /* The corresponding put will be done in the workqueue. */
3558 if (!css_tryget(&memcg->css))
3563 * If we are in a safe context (can wait, and not in interrupt
3564 * context), we could be be predictable and return right away.
3565 * This would guarantee that the allocation being performed
3566 * already belongs in the new cache.
3568 * However, there are some clashes that can arrive from locking.
3569 * For instance, because we acquire the slab_mutex while doing
3570 * kmem_cache_dup, this means no further allocation could happen
3571 * with the slab_mutex held.
3573 * Also, because cache creation issue get_online_cpus(), this
3574 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3575 * that ends up reversed during cpu hotplug. (cpuset allocates
3576 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3577 * better to defer everything.
3579 memcg_create_cache_enqueue(memcg, cachep);
3585 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3588 * We need to verify if the allocation against current->mm->owner's memcg is
3589 * possible for the given order. But the page is not allocated yet, so we'll
3590 * need a further commit step to do the final arrangements.
3592 * It is possible for the task to switch cgroups in this mean time, so at
3593 * commit time, we can't rely on task conversion any longer. We'll then use
3594 * the handle argument to return to the caller which cgroup we should commit
3595 * against. We could also return the memcg directly and avoid the pointer
3596 * passing, but a boolean return value gives better semantics considering
3597 * the compiled-out case as well.
3599 * Returning true means the allocation is possible.
3602 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3604 struct mem_cgroup *memcg;
3610 * Disabling accounting is only relevant for some specific memcg
3611 * internal allocations. Therefore we would initially not have such
3612 * check here, since direct calls to the page allocator that are marked
3613 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3614 * concerned with cache allocations, and by having this test at
3615 * memcg_kmem_get_cache, we are already able to relay the allocation to
3616 * the root cache and bypass the memcg cache altogether.
3618 * There is one exception, though: the SLUB allocator does not create
3619 * large order caches, but rather service large kmallocs directly from
3620 * the page allocator. Therefore, the following sequence when backed by
3621 * the SLUB allocator:
3623 * memcg_stop_kmem_account();
3624 * kmalloc(<large_number>)
3625 * memcg_resume_kmem_account();
3627 * would effectively ignore the fact that we should skip accounting,
3628 * since it will drive us directly to this function without passing
3629 * through the cache selector memcg_kmem_get_cache. Such large
3630 * allocations are extremely rare but can happen, for instance, for the
3631 * cache arrays. We bring this test here.
3633 if (!current->mm || current->memcg_kmem_skip_account)
3636 memcg = try_get_mem_cgroup_from_mm(current->mm);
3639 * very rare case described in mem_cgroup_from_task. Unfortunately there
3640 * isn't much we can do without complicating this too much, and it would
3641 * be gfp-dependent anyway. Just let it go
3643 if (unlikely(!memcg))
3646 if (!memcg_can_account_kmem(memcg)) {
3647 css_put(&memcg->css);
3651 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3655 css_put(&memcg->css);
3659 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3662 struct page_cgroup *pc;
3664 VM_BUG_ON(mem_cgroup_is_root(memcg));
3666 /* The page allocation failed. Revert */
3668 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3672 pc = lookup_page_cgroup(page);
3673 lock_page_cgroup(pc);
3674 pc->mem_cgroup = memcg;
3675 SetPageCgroupUsed(pc);
3676 unlock_page_cgroup(pc);
3679 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3681 struct mem_cgroup *memcg = NULL;
3682 struct page_cgroup *pc;
3685 pc = lookup_page_cgroup(page);
3687 * Fast unlocked return. Theoretically might have changed, have to
3688 * check again after locking.
3690 if (!PageCgroupUsed(pc))
3693 lock_page_cgroup(pc);
3694 if (PageCgroupUsed(pc)) {
3695 memcg = pc->mem_cgroup;
3696 ClearPageCgroupUsed(pc);
3698 unlock_page_cgroup(pc);
3701 * We trust that only if there is a memcg associated with the page, it
3702 * is a valid allocation
3707 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3708 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3711 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3714 #endif /* CONFIG_MEMCG_KMEM */
3716 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3718 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3720 * Because tail pages are not marked as "used", set it. We're under
3721 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3722 * charge/uncharge will be never happen and move_account() is done under
3723 * compound_lock(), so we don't have to take care of races.
3725 void mem_cgroup_split_huge_fixup(struct page *head)
3727 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3728 struct page_cgroup *pc;
3729 struct mem_cgroup *memcg;
3732 if (mem_cgroup_disabled())
3735 memcg = head_pc->mem_cgroup;
3736 for (i = 1; i < HPAGE_PMD_NR; i++) {
3738 pc->mem_cgroup = memcg;
3739 smp_wmb();/* see __commit_charge() */
3740 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3742 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3745 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3748 * mem_cgroup_move_account - move account of the page
3750 * @nr_pages: number of regular pages (>1 for huge pages)
3751 * @pc: page_cgroup of the page.
3752 * @from: mem_cgroup which the page is moved from.
3753 * @to: mem_cgroup which the page is moved to. @from != @to.
3755 * The caller must confirm following.
3756 * - page is not on LRU (isolate_page() is useful.)
3757 * - compound_lock is held when nr_pages > 1
3759 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3762 static int mem_cgroup_move_account(struct page *page,
3763 unsigned int nr_pages,
3764 struct page_cgroup *pc,
3765 struct mem_cgroup *from,
3766 struct mem_cgroup *to)
3768 unsigned long flags;
3770 bool anon = PageAnon(page);
3772 VM_BUG_ON(from == to);
3773 VM_BUG_ON_PAGE(PageLRU(page), page);
3775 * The page is isolated from LRU. So, collapse function
3776 * will not handle this page. But page splitting can happen.
3777 * Do this check under compound_page_lock(). The caller should
3781 if (nr_pages > 1 && !PageTransHuge(page))
3784 lock_page_cgroup(pc);
3787 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3790 move_lock_mem_cgroup(from, &flags);
3792 if (!anon && page_mapped(page)) {
3793 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3795 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3799 if (PageWriteback(page)) {
3800 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3802 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3806 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3808 /* caller should have done css_get */
3809 pc->mem_cgroup = to;
3810 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3811 move_unlock_mem_cgroup(from, &flags);
3814 unlock_page_cgroup(pc);
3818 memcg_check_events(to, page);
3819 memcg_check_events(from, page);
3825 * mem_cgroup_move_parent - moves page to the parent group
3826 * @page: the page to move
3827 * @pc: page_cgroup of the page
3828 * @child: page's cgroup
3830 * move charges to its parent or the root cgroup if the group has no
3831 * parent (aka use_hierarchy==0).
3832 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3833 * mem_cgroup_move_account fails) the failure is always temporary and
3834 * it signals a race with a page removal/uncharge or migration. In the
3835 * first case the page is on the way out and it will vanish from the LRU
3836 * on the next attempt and the call should be retried later.
3837 * Isolation from the LRU fails only if page has been isolated from
3838 * the LRU since we looked at it and that usually means either global
3839 * reclaim or migration going on. The page will either get back to the
3841 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3842 * (!PageCgroupUsed) or moved to a different group. The page will
3843 * disappear in the next attempt.
3845 static int mem_cgroup_move_parent(struct page *page,
3846 struct page_cgroup *pc,
3847 struct mem_cgroup *child)
3849 struct mem_cgroup *parent;
3850 unsigned int nr_pages;
3851 unsigned long uninitialized_var(flags);
3854 VM_BUG_ON(mem_cgroup_is_root(child));
3857 if (!get_page_unless_zero(page))
3859 if (isolate_lru_page(page))
3862 nr_pages = hpage_nr_pages(page);
3864 parent = parent_mem_cgroup(child);
3866 * If no parent, move charges to root cgroup.
3869 parent = root_mem_cgroup;
3872 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3873 flags = compound_lock_irqsave(page);
3876 ret = mem_cgroup_move_account(page, nr_pages,
3879 __mem_cgroup_cancel_local_charge(child, nr_pages);
3882 compound_unlock_irqrestore(page, flags);
3883 putback_lru_page(page);
3891 * Charge the memory controller for page usage.
3893 * 0 if the charge was successful
3894 * < 0 if the cgroup is over its limit
3896 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3897 gfp_t gfp_mask, enum charge_type ctype)
3899 struct mem_cgroup *memcg = NULL;
3900 unsigned int nr_pages = 1;
3904 if (PageTransHuge(page)) {
3905 nr_pages <<= compound_order(page);
3906 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3908 * Never OOM-kill a process for a huge page. The
3909 * fault handler will fall back to regular pages.
3914 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3917 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3921 int mem_cgroup_newpage_charge(struct page *page,
3922 struct mm_struct *mm, gfp_t gfp_mask)
3924 if (mem_cgroup_disabled())
3926 VM_BUG_ON_PAGE(page_mapped(page), page);
3927 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3929 return mem_cgroup_charge_common(page, mm, gfp_mask,
3930 MEM_CGROUP_CHARGE_TYPE_ANON);
3934 * While swap-in, try_charge -> commit or cancel, the page is locked.
3935 * And when try_charge() successfully returns, one refcnt to memcg without
3936 * struct page_cgroup is acquired. This refcnt will be consumed by
3937 * "commit()" or removed by "cancel()"
3939 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3942 struct mem_cgroup **memcgp)
3944 struct mem_cgroup *memcg;
3945 struct page_cgroup *pc;
3948 pc = lookup_page_cgroup(page);
3950 * Every swap fault against a single page tries to charge the
3951 * page, bail as early as possible. shmem_unuse() encounters
3952 * already charged pages, too. The USED bit is protected by
3953 * the page lock, which serializes swap cache removal, which
3954 * in turn serializes uncharging.
3956 if (PageCgroupUsed(pc))
3958 if (!do_swap_account)
3960 memcg = try_get_mem_cgroup_from_page(page);
3964 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3965 css_put(&memcg->css);
3970 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3976 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3977 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3980 if (mem_cgroup_disabled())
3983 * A racing thread's fault, or swapoff, may have already
3984 * updated the pte, and even removed page from swap cache: in
3985 * those cases unuse_pte()'s pte_same() test will fail; but
3986 * there's also a KSM case which does need to charge the page.
3988 if (!PageSwapCache(page)) {
3991 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3996 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3999 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4001 if (mem_cgroup_disabled())
4005 __mem_cgroup_cancel_charge(memcg, 1);
4009 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4010 enum charge_type ctype)
4012 if (mem_cgroup_disabled())
4017 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4019 * Now swap is on-memory. This means this page may be
4020 * counted both as mem and swap....double count.
4021 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4022 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4023 * may call delete_from_swap_cache() before reach here.
4025 if (do_swap_account && PageSwapCache(page)) {
4026 swp_entry_t ent = {.val = page_private(page)};
4027 mem_cgroup_uncharge_swap(ent);
4031 void mem_cgroup_commit_charge_swapin(struct page *page,
4032 struct mem_cgroup *memcg)
4034 __mem_cgroup_commit_charge_swapin(page, memcg,
4035 MEM_CGROUP_CHARGE_TYPE_ANON);
4038 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4041 struct mem_cgroup *memcg = NULL;
4042 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4045 if (mem_cgroup_disabled())
4047 if (PageCompound(page))
4050 if (!PageSwapCache(page))
4051 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4052 else { /* page is swapcache/shmem */
4053 ret = __mem_cgroup_try_charge_swapin(mm, page,
4056 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4061 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4062 unsigned int nr_pages,
4063 const enum charge_type ctype)
4065 struct memcg_batch_info *batch = NULL;
4066 bool uncharge_memsw = true;
4068 /* If swapout, usage of swap doesn't decrease */
4069 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4070 uncharge_memsw = false;
4072 batch = ¤t->memcg_batch;
4074 * In usual, we do css_get() when we remember memcg pointer.
4075 * But in this case, we keep res->usage until end of a series of
4076 * uncharges. Then, it's ok to ignore memcg's refcnt.
4079 batch->memcg = memcg;
4081 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4082 * In those cases, all pages freed continuously can be expected to be in
4083 * the same cgroup and we have chance to coalesce uncharges.
4084 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4085 * because we want to do uncharge as soon as possible.
4088 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4089 goto direct_uncharge;
4092 goto direct_uncharge;
4095 * In typical case, batch->memcg == mem. This means we can
4096 * merge a series of uncharges to an uncharge of res_counter.
4097 * If not, we uncharge res_counter ony by one.
4099 if (batch->memcg != memcg)
4100 goto direct_uncharge;
4101 /* remember freed charge and uncharge it later */
4104 batch->memsw_nr_pages++;
4107 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4109 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4110 if (unlikely(batch->memcg != memcg))
4111 memcg_oom_recover(memcg);
4115 * uncharge if !page_mapped(page)
4117 static struct mem_cgroup *
4118 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4121 struct mem_cgroup *memcg = NULL;
4122 unsigned int nr_pages = 1;
4123 struct page_cgroup *pc;
4126 if (mem_cgroup_disabled())
4129 if (PageTransHuge(page)) {
4130 nr_pages <<= compound_order(page);
4131 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
4134 * Check if our page_cgroup is valid
4136 pc = lookup_page_cgroup(page);
4137 if (unlikely(!PageCgroupUsed(pc)))
4140 lock_page_cgroup(pc);
4142 memcg = pc->mem_cgroup;
4144 if (!PageCgroupUsed(pc))
4147 anon = PageAnon(page);
4150 case MEM_CGROUP_CHARGE_TYPE_ANON:
4152 * Generally PageAnon tells if it's the anon statistics to be
4153 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4154 * used before page reached the stage of being marked PageAnon.
4158 case MEM_CGROUP_CHARGE_TYPE_DROP:
4159 /* See mem_cgroup_prepare_migration() */
4160 if (page_mapped(page))
4163 * Pages under migration may not be uncharged. But
4164 * end_migration() /must/ be the one uncharging the
4165 * unused post-migration page and so it has to call
4166 * here with the migration bit still set. See the
4167 * res_counter handling below.
4169 if (!end_migration && PageCgroupMigration(pc))
4172 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4173 if (!PageAnon(page)) { /* Shared memory */
4174 if (page->mapping && !page_is_file_cache(page))
4176 } else if (page_mapped(page)) /* Anon */
4183 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4185 ClearPageCgroupUsed(pc);
4187 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4188 * freed from LRU. This is safe because uncharged page is expected not
4189 * to be reused (freed soon). Exception is SwapCache, it's handled by
4190 * special functions.
4193 unlock_page_cgroup(pc);
4195 * even after unlock, we have memcg->res.usage here and this memcg
4196 * will never be freed, so it's safe to call css_get().
4198 memcg_check_events(memcg, page);
4199 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4200 mem_cgroup_swap_statistics(memcg, true);
4201 css_get(&memcg->css);
4204 * Migration does not charge the res_counter for the
4205 * replacement page, so leave it alone when phasing out the
4206 * page that is unused after the migration.
4208 if (!end_migration && !mem_cgroup_is_root(memcg))
4209 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4214 unlock_page_cgroup(pc);
4218 void mem_cgroup_uncharge_page(struct page *page)
4221 if (page_mapped(page))
4223 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4225 * If the page is in swap cache, uncharge should be deferred
4226 * to the swap path, which also properly accounts swap usage
4227 * and handles memcg lifetime.
4229 * Note that this check is not stable and reclaim may add the
4230 * page to swap cache at any time after this. However, if the
4231 * page is not in swap cache by the time page->mapcount hits
4232 * 0, there won't be any page table references to the swap
4233 * slot, and reclaim will free it and not actually write the
4236 if (PageSwapCache(page))
4238 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4241 void mem_cgroup_uncharge_cache_page(struct page *page)
4243 VM_BUG_ON_PAGE(page_mapped(page), page);
4244 VM_BUG_ON_PAGE(page->mapping, page);
4245 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4249 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4250 * In that cases, pages are freed continuously and we can expect pages
4251 * are in the same memcg. All these calls itself limits the number of
4252 * pages freed at once, then uncharge_start/end() is called properly.
4253 * This may be called prural(2) times in a context,
4256 void mem_cgroup_uncharge_start(void)
4258 current->memcg_batch.do_batch++;
4259 /* We can do nest. */
4260 if (current->memcg_batch.do_batch == 1) {
4261 current->memcg_batch.memcg = NULL;
4262 current->memcg_batch.nr_pages = 0;
4263 current->memcg_batch.memsw_nr_pages = 0;
4267 void mem_cgroup_uncharge_end(void)
4269 struct memcg_batch_info *batch = ¤t->memcg_batch;
4271 if (!batch->do_batch)
4275 if (batch->do_batch) /* If stacked, do nothing. */
4281 * This "batch->memcg" is valid without any css_get/put etc...
4282 * bacause we hide charges behind us.
4284 if (batch->nr_pages)
4285 res_counter_uncharge(&batch->memcg->res,
4286 batch->nr_pages * PAGE_SIZE);
4287 if (batch->memsw_nr_pages)
4288 res_counter_uncharge(&batch->memcg->memsw,
4289 batch->memsw_nr_pages * PAGE_SIZE);
4290 memcg_oom_recover(batch->memcg);
4291 /* forget this pointer (for sanity check) */
4292 batch->memcg = NULL;
4297 * called after __delete_from_swap_cache() and drop "page" account.
4298 * memcg information is recorded to swap_cgroup of "ent"
4301 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4303 struct mem_cgroup *memcg;
4304 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4306 if (!swapout) /* this was a swap cache but the swap is unused ! */
4307 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4309 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4312 * record memcg information, if swapout && memcg != NULL,
4313 * css_get() was called in uncharge().
4315 if (do_swap_account && swapout && memcg)
4316 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4320 #ifdef CONFIG_MEMCG_SWAP
4322 * called from swap_entry_free(). remove record in swap_cgroup and
4323 * uncharge "memsw" account.
4325 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4327 struct mem_cgroup *memcg;
4330 if (!do_swap_account)
4333 id = swap_cgroup_record(ent, 0);
4335 memcg = mem_cgroup_lookup(id);
4338 * We uncharge this because swap is freed.
4339 * This memcg can be obsolete one. We avoid calling css_tryget
4341 if (!mem_cgroup_is_root(memcg))
4342 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4343 mem_cgroup_swap_statistics(memcg, false);
4344 css_put(&memcg->css);
4350 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4351 * @entry: swap entry to be moved
4352 * @from: mem_cgroup which the entry is moved from
4353 * @to: mem_cgroup which the entry is moved to
4355 * It succeeds only when the swap_cgroup's record for this entry is the same
4356 * as the mem_cgroup's id of @from.
4358 * Returns 0 on success, -EINVAL on failure.
4360 * The caller must have charged to @to, IOW, called res_counter_charge() about
4361 * both res and memsw, and called css_get().
4363 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4364 struct mem_cgroup *from, struct mem_cgroup *to)
4366 unsigned short old_id, new_id;
4368 old_id = mem_cgroup_id(from);
4369 new_id = mem_cgroup_id(to);
4371 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4372 mem_cgroup_swap_statistics(from, false);
4373 mem_cgroup_swap_statistics(to, true);
4375 * This function is only called from task migration context now.
4376 * It postpones res_counter and refcount handling till the end
4377 * of task migration(mem_cgroup_clear_mc()) for performance
4378 * improvement. But we cannot postpone css_get(to) because if
4379 * the process that has been moved to @to does swap-in, the
4380 * refcount of @to might be decreased to 0.
4382 * We are in attach() phase, so the cgroup is guaranteed to be
4383 * alive, so we can just call css_get().
4391 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4392 struct mem_cgroup *from, struct mem_cgroup *to)
4399 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4402 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4403 struct mem_cgroup **memcgp)
4405 struct mem_cgroup *memcg = NULL;
4406 unsigned int nr_pages = 1;
4407 struct page_cgroup *pc;
4408 enum charge_type ctype;
4412 if (mem_cgroup_disabled())
4415 if (PageTransHuge(page))
4416 nr_pages <<= compound_order(page);
4418 pc = lookup_page_cgroup(page);
4419 lock_page_cgroup(pc);
4420 if (PageCgroupUsed(pc)) {
4421 memcg = pc->mem_cgroup;
4422 css_get(&memcg->css);
4424 * At migrating an anonymous page, its mapcount goes down
4425 * to 0 and uncharge() will be called. But, even if it's fully
4426 * unmapped, migration may fail and this page has to be
4427 * charged again. We set MIGRATION flag here and delay uncharge
4428 * until end_migration() is called
4430 * Corner Case Thinking
4432 * When the old page was mapped as Anon and it's unmap-and-freed
4433 * while migration was ongoing.
4434 * If unmap finds the old page, uncharge() of it will be delayed
4435 * until end_migration(). If unmap finds a new page, it's
4436 * uncharged when it make mapcount to be 1->0. If unmap code
4437 * finds swap_migration_entry, the new page will not be mapped
4438 * and end_migration() will find it(mapcount==0).
4441 * When the old page was mapped but migraion fails, the kernel
4442 * remaps it. A charge for it is kept by MIGRATION flag even
4443 * if mapcount goes down to 0. We can do remap successfully
4444 * without charging it again.
4447 * The "old" page is under lock_page() until the end of
4448 * migration, so, the old page itself will not be swapped-out.
4449 * If the new page is swapped out before end_migraton, our
4450 * hook to usual swap-out path will catch the event.
4453 SetPageCgroupMigration(pc);
4455 unlock_page_cgroup(pc);
4457 * If the page is not charged at this point,
4465 * We charge new page before it's used/mapped. So, even if unlock_page()
4466 * is called before end_migration, we can catch all events on this new
4467 * page. In the case new page is migrated but not remapped, new page's
4468 * mapcount will be finally 0 and we call uncharge in end_migration().
4471 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4473 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4475 * The page is committed to the memcg, but it's not actually
4476 * charged to the res_counter since we plan on replacing the
4477 * old one and only one page is going to be left afterwards.
4479 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4482 /* remove redundant charge if migration failed*/
4483 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4484 struct page *oldpage, struct page *newpage, bool migration_ok)
4486 struct page *used, *unused;
4487 struct page_cgroup *pc;
4493 if (!migration_ok) {
4500 anon = PageAnon(used);
4501 __mem_cgroup_uncharge_common(unused,
4502 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4503 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4505 css_put(&memcg->css);
4507 * We disallowed uncharge of pages under migration because mapcount
4508 * of the page goes down to zero, temporarly.
4509 * Clear the flag and check the page should be charged.
4511 pc = lookup_page_cgroup(oldpage);
4512 lock_page_cgroup(pc);
4513 ClearPageCgroupMigration(pc);
4514 unlock_page_cgroup(pc);
4517 * If a page is a file cache, radix-tree replacement is very atomic
4518 * and we can skip this check. When it was an Anon page, its mapcount
4519 * goes down to 0. But because we added MIGRATION flage, it's not
4520 * uncharged yet. There are several case but page->mapcount check
4521 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4522 * check. (see prepare_charge() also)
4525 mem_cgroup_uncharge_page(used);
4529 * At replace page cache, newpage is not under any memcg but it's on
4530 * LRU. So, this function doesn't touch res_counter but handles LRU
4531 * in correct way. Both pages are locked so we cannot race with uncharge.
4533 void mem_cgroup_replace_page_cache(struct page *oldpage,
4534 struct page *newpage)
4536 struct mem_cgroup *memcg = NULL;
4537 struct page_cgroup *pc;
4538 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4540 if (mem_cgroup_disabled())
4543 pc = lookup_page_cgroup(oldpage);
4544 /* fix accounting on old pages */
4545 lock_page_cgroup(pc);
4546 if (PageCgroupUsed(pc)) {
4547 memcg = pc->mem_cgroup;
4548 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4549 ClearPageCgroupUsed(pc);
4551 unlock_page_cgroup(pc);
4554 * When called from shmem_replace_page(), in some cases the
4555 * oldpage has already been charged, and in some cases not.
4560 * Even if newpage->mapping was NULL before starting replacement,
4561 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4562 * LRU while we overwrite pc->mem_cgroup.
4564 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4567 #ifdef CONFIG_DEBUG_VM
4568 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4570 struct page_cgroup *pc;
4572 pc = lookup_page_cgroup(page);
4574 * Can be NULL while feeding pages into the page allocator for
4575 * the first time, i.e. during boot or memory hotplug;
4576 * or when mem_cgroup_disabled().
4578 if (likely(pc) && PageCgroupUsed(pc))
4583 bool mem_cgroup_bad_page_check(struct page *page)
4585 if (mem_cgroup_disabled())
4588 return lookup_page_cgroup_used(page) != NULL;
4591 void mem_cgroup_print_bad_page(struct page *page)
4593 struct page_cgroup *pc;
4595 pc = lookup_page_cgroup_used(page);
4597 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4598 pc, pc->flags, pc->mem_cgroup);
4603 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4604 unsigned long long val)
4607 u64 memswlimit, memlimit;
4609 int children = mem_cgroup_count_children(memcg);
4610 u64 curusage, oldusage;
4614 * For keeping hierarchical_reclaim simple, how long we should retry
4615 * is depends on callers. We set our retry-count to be function
4616 * of # of children which we should visit in this loop.
4618 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4620 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4623 while (retry_count) {
4624 if (signal_pending(current)) {
4629 * Rather than hide all in some function, I do this in
4630 * open coded manner. You see what this really does.
4631 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4633 mutex_lock(&set_limit_mutex);
4634 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4635 if (memswlimit < val) {
4637 mutex_unlock(&set_limit_mutex);
4641 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4645 ret = res_counter_set_limit(&memcg->res, val);
4647 if (memswlimit == val)
4648 memcg->memsw_is_minimum = true;
4650 memcg->memsw_is_minimum = false;
4652 mutex_unlock(&set_limit_mutex);
4657 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4658 MEM_CGROUP_RECLAIM_SHRINK);
4659 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4660 /* Usage is reduced ? */
4661 if (curusage >= oldusage)
4664 oldusage = curusage;
4666 if (!ret && enlarge)
4667 memcg_oom_recover(memcg);
4672 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4673 unsigned long long val)
4676 u64 memlimit, memswlimit, oldusage, curusage;
4677 int children = mem_cgroup_count_children(memcg);
4681 /* see mem_cgroup_resize_res_limit */
4682 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4683 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4684 while (retry_count) {
4685 if (signal_pending(current)) {
4690 * Rather than hide all in some function, I do this in
4691 * open coded manner. You see what this really does.
4692 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4694 mutex_lock(&set_limit_mutex);
4695 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4696 if (memlimit > val) {
4698 mutex_unlock(&set_limit_mutex);
4701 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4702 if (memswlimit < val)
4704 ret = res_counter_set_limit(&memcg->memsw, val);
4706 if (memlimit == val)
4707 memcg->memsw_is_minimum = true;
4709 memcg->memsw_is_minimum = false;
4711 mutex_unlock(&set_limit_mutex);
4716 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4717 MEM_CGROUP_RECLAIM_NOSWAP |
4718 MEM_CGROUP_RECLAIM_SHRINK);
4719 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4720 /* Usage is reduced ? */
4721 if (curusage >= oldusage)
4724 oldusage = curusage;
4726 if (!ret && enlarge)
4727 memcg_oom_recover(memcg);
4731 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4733 unsigned long *total_scanned)
4735 unsigned long nr_reclaimed = 0;
4736 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4737 unsigned long reclaimed;
4739 struct mem_cgroup_tree_per_zone *mctz;
4740 unsigned long long excess;
4741 unsigned long nr_scanned;
4746 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4748 * This loop can run a while, specially if mem_cgroup's continuously
4749 * keep exceeding their soft limit and putting the system under
4756 mz = mem_cgroup_largest_soft_limit_node(mctz);
4761 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4762 gfp_mask, &nr_scanned);
4763 nr_reclaimed += reclaimed;
4764 *total_scanned += nr_scanned;
4765 spin_lock(&mctz->lock);
4768 * If we failed to reclaim anything from this memory cgroup
4769 * it is time to move on to the next cgroup
4775 * Loop until we find yet another one.
4777 * By the time we get the soft_limit lock
4778 * again, someone might have aded the
4779 * group back on the RB tree. Iterate to
4780 * make sure we get a different mem.
4781 * mem_cgroup_largest_soft_limit_node returns
4782 * NULL if no other cgroup is present on
4786 __mem_cgroup_largest_soft_limit_node(mctz);
4788 css_put(&next_mz->memcg->css);
4789 else /* next_mz == NULL or other memcg */
4793 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4794 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4796 * One school of thought says that we should not add
4797 * back the node to the tree if reclaim returns 0.
4798 * But our reclaim could return 0, simply because due
4799 * to priority we are exposing a smaller subset of
4800 * memory to reclaim from. Consider this as a longer
4803 /* If excess == 0, no tree ops */
4804 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4805 spin_unlock(&mctz->lock);
4806 css_put(&mz->memcg->css);
4809 * Could not reclaim anything and there are no more
4810 * mem cgroups to try or we seem to be looping without
4811 * reclaiming anything.
4813 if (!nr_reclaimed &&
4815 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4817 } while (!nr_reclaimed);
4819 css_put(&next_mz->memcg->css);
4820 return nr_reclaimed;
4824 * mem_cgroup_force_empty_list - clears LRU of a group
4825 * @memcg: group to clear
4828 * @lru: lru to to clear
4830 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4831 * reclaim the pages page themselves - pages are moved to the parent (or root)
4834 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4835 int node, int zid, enum lru_list lru)
4837 struct lruvec *lruvec;
4838 unsigned long flags;
4839 struct list_head *list;
4843 zone = &NODE_DATA(node)->node_zones[zid];
4844 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4845 list = &lruvec->lists[lru];
4849 struct page_cgroup *pc;
4852 spin_lock_irqsave(&zone->lru_lock, flags);
4853 if (list_empty(list)) {
4854 spin_unlock_irqrestore(&zone->lru_lock, flags);
4857 page = list_entry(list->prev, struct page, lru);
4859 list_move(&page->lru, list);
4861 spin_unlock_irqrestore(&zone->lru_lock, flags);
4864 spin_unlock_irqrestore(&zone->lru_lock, flags);
4866 pc = lookup_page_cgroup(page);
4868 if (mem_cgroup_move_parent(page, pc, memcg)) {
4869 /* found lock contention or "pc" is obsolete. */
4874 } while (!list_empty(list));
4878 * make mem_cgroup's charge to be 0 if there is no task by moving
4879 * all the charges and pages to the parent.
4880 * This enables deleting this mem_cgroup.
4882 * Caller is responsible for holding css reference on the memcg.
4884 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4890 /* This is for making all *used* pages to be on LRU. */
4891 lru_add_drain_all();
4892 drain_all_stock_sync(memcg);
4893 mem_cgroup_start_move(memcg);
4894 for_each_node_state(node, N_MEMORY) {
4895 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4898 mem_cgroup_force_empty_list(memcg,
4903 mem_cgroup_end_move(memcg);
4904 memcg_oom_recover(memcg);
4908 * Kernel memory may not necessarily be trackable to a specific
4909 * process. So they are not migrated, and therefore we can't
4910 * expect their value to drop to 0 here.
4911 * Having res filled up with kmem only is enough.
4913 * This is a safety check because mem_cgroup_force_empty_list
4914 * could have raced with mem_cgroup_replace_page_cache callers
4915 * so the lru seemed empty but the page could have been added
4916 * right after the check. RES_USAGE should be safe as we always
4917 * charge before adding to the LRU.
4919 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4920 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4921 } while (usage > 0);
4924 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4926 lockdep_assert_held(&memcg_create_mutex);
4928 * The lock does not prevent addition or deletion to the list
4929 * of children, but it prevents a new child from being
4930 * initialized based on this parent in css_online(), so it's
4931 * enough to decide whether hierarchically inherited
4932 * attributes can still be changed or not.
4934 return memcg->use_hierarchy &&
4935 !list_empty(&memcg->css.cgroup->children);
4939 * Reclaims as many pages from the given memcg as possible and moves
4940 * the rest to the parent.
4942 * Caller is responsible for holding css reference for memcg.
4944 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4946 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4947 struct cgroup *cgrp = memcg->css.cgroup;
4949 /* returns EBUSY if there is a task or if we come here twice. */
4950 if (cgroup_has_tasks(cgrp) || !list_empty(&cgrp->children))
4953 /* we call try-to-free pages for make this cgroup empty */
4954 lru_add_drain_all();
4955 /* try to free all pages in this cgroup */
4956 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4959 if (signal_pending(current))
4962 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4966 /* maybe some writeback is necessary */
4967 congestion_wait(BLK_RW_ASYNC, HZ/10);
4972 mem_cgroup_reparent_charges(memcg);
4977 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4980 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4982 if (mem_cgroup_is_root(memcg))
4984 return mem_cgroup_force_empty(memcg);
4987 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4990 return mem_cgroup_from_css(css)->use_hierarchy;
4993 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4994 struct cftype *cft, u64 val)
4997 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4998 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5000 mutex_lock(&memcg_create_mutex);
5002 if (memcg->use_hierarchy == val)
5006 * If parent's use_hierarchy is set, we can't make any modifications
5007 * in the child subtrees. If it is unset, then the change can
5008 * occur, provided the current cgroup has no children.
5010 * For the root cgroup, parent_mem is NULL, we allow value to be
5011 * set if there are no children.
5013 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5014 (val == 1 || val == 0)) {
5015 if (list_empty(&memcg->css.cgroup->children))
5016 memcg->use_hierarchy = val;
5023 mutex_unlock(&memcg_create_mutex);
5029 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5030 enum mem_cgroup_stat_index idx)
5032 struct mem_cgroup *iter;
5035 /* Per-cpu values can be negative, use a signed accumulator */
5036 for_each_mem_cgroup_tree(iter, memcg)
5037 val += mem_cgroup_read_stat(iter, idx);
5039 if (val < 0) /* race ? */
5044 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5048 if (!mem_cgroup_is_root(memcg)) {
5050 return res_counter_read_u64(&memcg->res, RES_USAGE);
5052 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5056 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5057 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5059 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5060 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5063 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5065 return val << PAGE_SHIFT;
5068 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
5071 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5076 type = MEMFILE_TYPE(cft->private);
5077 name = MEMFILE_ATTR(cft->private);
5081 if (name == RES_USAGE)
5082 val = mem_cgroup_usage(memcg, false);
5084 val = res_counter_read_u64(&memcg->res, name);
5087 if (name == RES_USAGE)
5088 val = mem_cgroup_usage(memcg, true);
5090 val = res_counter_read_u64(&memcg->memsw, name);
5093 val = res_counter_read_u64(&memcg->kmem, name);
5102 #ifdef CONFIG_MEMCG_KMEM
5103 /* should be called with activate_kmem_mutex held */
5104 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
5105 unsigned long long limit)
5110 if (memcg_kmem_is_active(memcg))
5114 * We are going to allocate memory for data shared by all memory
5115 * cgroups so let's stop accounting here.
5117 memcg_stop_kmem_account();
5120 * For simplicity, we won't allow this to be disabled. It also can't
5121 * be changed if the cgroup has children already, or if tasks had
5124 * If tasks join before we set the limit, a person looking at
5125 * kmem.usage_in_bytes will have no way to determine when it took
5126 * place, which makes the value quite meaningless.
5128 * After it first became limited, changes in the value of the limit are
5129 * of course permitted.
5131 mutex_lock(&memcg_create_mutex);
5132 if (cgroup_has_tasks(memcg->css.cgroup) || memcg_has_children(memcg))
5134 mutex_unlock(&memcg_create_mutex);
5138 memcg_id = ida_simple_get(&kmem_limited_groups,
5139 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
5146 * Make sure we have enough space for this cgroup in each root cache's
5149 err = memcg_update_all_caches(memcg_id + 1);
5153 memcg->kmemcg_id = memcg_id;
5154 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
5155 mutex_init(&memcg->slab_caches_mutex);
5158 * We couldn't have accounted to this cgroup, because it hasn't got the
5159 * active bit set yet, so this should succeed.
5161 err = res_counter_set_limit(&memcg->kmem, limit);
5164 static_key_slow_inc(&memcg_kmem_enabled_key);
5166 * Setting the active bit after enabling static branching will
5167 * guarantee no one starts accounting before all call sites are
5170 memcg_kmem_set_active(memcg);
5172 memcg_resume_kmem_account();
5176 ida_simple_remove(&kmem_limited_groups, memcg_id);
5180 static int memcg_activate_kmem(struct mem_cgroup *memcg,
5181 unsigned long long limit)
5185 mutex_lock(&activate_kmem_mutex);
5186 ret = __memcg_activate_kmem(memcg, limit);
5187 mutex_unlock(&activate_kmem_mutex);
5191 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5192 unsigned long long val)
5196 if (!memcg_kmem_is_active(memcg))
5197 ret = memcg_activate_kmem(memcg, val);
5199 ret = res_counter_set_limit(&memcg->kmem, val);
5203 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5206 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5211 mutex_lock(&activate_kmem_mutex);
5213 * If the parent cgroup is not kmem-active now, it cannot be activated
5214 * after this point, because it has at least one child already.
5216 if (memcg_kmem_is_active(parent))
5217 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
5218 mutex_unlock(&activate_kmem_mutex);
5222 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5223 unsigned long long val)
5227 #endif /* CONFIG_MEMCG_KMEM */
5230 * The user of this function is...
5233 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5236 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5239 unsigned long long val;
5242 type = MEMFILE_TYPE(cft->private);
5243 name = MEMFILE_ATTR(cft->private);
5247 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5251 /* This function does all necessary parse...reuse it */
5252 ret = res_counter_memparse_write_strategy(buffer, &val);
5256 ret = mem_cgroup_resize_limit(memcg, val);
5257 else if (type == _MEMSWAP)
5258 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5259 else if (type == _KMEM)
5260 ret = memcg_update_kmem_limit(memcg, val);
5264 case RES_SOFT_LIMIT:
5265 ret = res_counter_memparse_write_strategy(buffer, &val);
5269 * For memsw, soft limits are hard to implement in terms
5270 * of semantics, for now, we support soft limits for
5271 * control without swap
5274 ret = res_counter_set_soft_limit(&memcg->res, val);
5279 ret = -EINVAL; /* should be BUG() ? */
5285 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5286 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5288 unsigned long long min_limit, min_memsw_limit, tmp;
5290 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5291 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5292 if (!memcg->use_hierarchy)
5295 while (css_parent(&memcg->css)) {
5296 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5297 if (!memcg->use_hierarchy)
5299 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5300 min_limit = min(min_limit, tmp);
5301 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5302 min_memsw_limit = min(min_memsw_limit, tmp);
5305 *mem_limit = min_limit;
5306 *memsw_limit = min_memsw_limit;
5309 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5311 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5315 type = MEMFILE_TYPE(event);
5316 name = MEMFILE_ATTR(event);
5321 res_counter_reset_max(&memcg->res);
5322 else if (type == _MEMSWAP)
5323 res_counter_reset_max(&memcg->memsw);
5324 else if (type == _KMEM)
5325 res_counter_reset_max(&memcg->kmem);
5331 res_counter_reset_failcnt(&memcg->res);
5332 else if (type == _MEMSWAP)
5333 res_counter_reset_failcnt(&memcg->memsw);
5334 else if (type == _KMEM)
5335 res_counter_reset_failcnt(&memcg->kmem);
5344 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5347 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5351 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5352 struct cftype *cft, u64 val)
5354 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5356 if (val >= (1 << NR_MOVE_TYPE))
5360 * No kind of locking is needed in here, because ->can_attach() will
5361 * check this value once in the beginning of the process, and then carry
5362 * on with stale data. This means that changes to this value will only
5363 * affect task migrations starting after the change.
5365 memcg->move_charge_at_immigrate = val;
5369 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5370 struct cftype *cft, u64 val)
5377 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5381 unsigned int lru_mask;
5384 static const struct numa_stat stats[] = {
5385 { "total", LRU_ALL },
5386 { "file", LRU_ALL_FILE },
5387 { "anon", LRU_ALL_ANON },
5388 { "unevictable", BIT(LRU_UNEVICTABLE) },
5390 const struct numa_stat *stat;
5393 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5395 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5396 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5397 seq_printf(m, "%s=%lu", stat->name, nr);
5398 for_each_node_state(nid, N_MEMORY) {
5399 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5401 seq_printf(m, " N%d=%lu", nid, nr);
5406 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5407 struct mem_cgroup *iter;
5410 for_each_mem_cgroup_tree(iter, memcg)
5411 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5412 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5413 for_each_node_state(nid, N_MEMORY) {
5415 for_each_mem_cgroup_tree(iter, memcg)
5416 nr += mem_cgroup_node_nr_lru_pages(
5417 iter, nid, stat->lru_mask);
5418 seq_printf(m, " N%d=%lu", nid, nr);
5425 #endif /* CONFIG_NUMA */
5427 static inline void mem_cgroup_lru_names_not_uptodate(void)
5429 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5432 static int memcg_stat_show(struct seq_file *m, void *v)
5434 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5435 struct mem_cgroup *mi;
5438 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5439 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5441 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5442 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5445 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5446 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5447 mem_cgroup_read_events(memcg, i));
5449 for (i = 0; i < NR_LRU_LISTS; i++)
5450 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5451 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5453 /* Hierarchical information */
5455 unsigned long long limit, memsw_limit;
5456 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5457 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5458 if (do_swap_account)
5459 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5463 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5466 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5468 for_each_mem_cgroup_tree(mi, memcg)
5469 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5470 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5473 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5474 unsigned long long val = 0;
5476 for_each_mem_cgroup_tree(mi, memcg)
5477 val += mem_cgroup_read_events(mi, i);
5478 seq_printf(m, "total_%s %llu\n",
5479 mem_cgroup_events_names[i], val);
5482 for (i = 0; i < NR_LRU_LISTS; i++) {
5483 unsigned long long val = 0;
5485 for_each_mem_cgroup_tree(mi, memcg)
5486 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5487 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5490 #ifdef CONFIG_DEBUG_VM
5493 struct mem_cgroup_per_zone *mz;
5494 struct zone_reclaim_stat *rstat;
5495 unsigned long recent_rotated[2] = {0, 0};
5496 unsigned long recent_scanned[2] = {0, 0};
5498 for_each_online_node(nid)
5499 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5500 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5501 rstat = &mz->lruvec.reclaim_stat;
5503 recent_rotated[0] += rstat->recent_rotated[0];
5504 recent_rotated[1] += rstat->recent_rotated[1];
5505 recent_scanned[0] += rstat->recent_scanned[0];
5506 recent_scanned[1] += rstat->recent_scanned[1];
5508 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5509 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5510 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5511 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5518 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5521 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5523 return mem_cgroup_swappiness(memcg);
5526 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5527 struct cftype *cft, u64 val)
5529 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5530 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5532 if (val > 100 || !parent)
5535 mutex_lock(&memcg_create_mutex);
5537 /* If under hierarchy, only empty-root can set this value */
5538 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5539 mutex_unlock(&memcg_create_mutex);
5543 memcg->swappiness = val;
5545 mutex_unlock(&memcg_create_mutex);
5550 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5552 struct mem_cgroup_threshold_ary *t;
5558 t = rcu_dereference(memcg->thresholds.primary);
5560 t = rcu_dereference(memcg->memsw_thresholds.primary);
5565 usage = mem_cgroup_usage(memcg, swap);
5568 * current_threshold points to threshold just below or equal to usage.
5569 * If it's not true, a threshold was crossed after last
5570 * call of __mem_cgroup_threshold().
5572 i = t->current_threshold;
5575 * Iterate backward over array of thresholds starting from
5576 * current_threshold and check if a threshold is crossed.
5577 * If none of thresholds below usage is crossed, we read
5578 * only one element of the array here.
5580 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5581 eventfd_signal(t->entries[i].eventfd, 1);
5583 /* i = current_threshold + 1 */
5587 * Iterate forward over array of thresholds starting from
5588 * current_threshold+1 and check if a threshold is crossed.
5589 * If none of thresholds above usage is crossed, we read
5590 * only one element of the array here.
5592 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5593 eventfd_signal(t->entries[i].eventfd, 1);
5595 /* Update current_threshold */
5596 t->current_threshold = i - 1;
5601 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5604 __mem_cgroup_threshold(memcg, false);
5605 if (do_swap_account)
5606 __mem_cgroup_threshold(memcg, true);
5608 memcg = parent_mem_cgroup(memcg);
5612 static int compare_thresholds(const void *a, const void *b)
5614 const struct mem_cgroup_threshold *_a = a;
5615 const struct mem_cgroup_threshold *_b = b;
5617 if (_a->threshold > _b->threshold)
5620 if (_a->threshold < _b->threshold)
5626 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5628 struct mem_cgroup_eventfd_list *ev;
5630 list_for_each_entry(ev, &memcg->oom_notify, list)
5631 eventfd_signal(ev->eventfd, 1);
5635 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5637 struct mem_cgroup *iter;
5639 for_each_mem_cgroup_tree(iter, memcg)
5640 mem_cgroup_oom_notify_cb(iter);
5643 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5644 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5646 struct mem_cgroup_thresholds *thresholds;
5647 struct mem_cgroup_threshold_ary *new;
5648 u64 threshold, usage;
5651 ret = res_counter_memparse_write_strategy(args, &threshold);
5655 mutex_lock(&memcg->thresholds_lock);
5658 thresholds = &memcg->thresholds;
5659 else if (type == _MEMSWAP)
5660 thresholds = &memcg->memsw_thresholds;
5664 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5666 /* Check if a threshold crossed before adding a new one */
5667 if (thresholds->primary)
5668 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5670 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5672 /* Allocate memory for new array of thresholds */
5673 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5681 /* Copy thresholds (if any) to new array */
5682 if (thresholds->primary) {
5683 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5684 sizeof(struct mem_cgroup_threshold));
5687 /* Add new threshold */
5688 new->entries[size - 1].eventfd = eventfd;
5689 new->entries[size - 1].threshold = threshold;
5691 /* Sort thresholds. Registering of new threshold isn't time-critical */
5692 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5693 compare_thresholds, NULL);
5695 /* Find current threshold */
5696 new->current_threshold = -1;
5697 for (i = 0; i < size; i++) {
5698 if (new->entries[i].threshold <= usage) {
5700 * new->current_threshold will not be used until
5701 * rcu_assign_pointer(), so it's safe to increment
5704 ++new->current_threshold;
5709 /* Free old spare buffer and save old primary buffer as spare */
5710 kfree(thresholds->spare);
5711 thresholds->spare = thresholds->primary;
5713 rcu_assign_pointer(thresholds->primary, new);
5715 /* To be sure that nobody uses thresholds */
5719 mutex_unlock(&memcg->thresholds_lock);
5724 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5725 struct eventfd_ctx *eventfd, const char *args)
5727 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5730 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5731 struct eventfd_ctx *eventfd, const char *args)
5733 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5736 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5737 struct eventfd_ctx *eventfd, enum res_type type)
5739 struct mem_cgroup_thresholds *thresholds;
5740 struct mem_cgroup_threshold_ary *new;
5744 mutex_lock(&memcg->thresholds_lock);
5746 thresholds = &memcg->thresholds;
5747 else if (type == _MEMSWAP)
5748 thresholds = &memcg->memsw_thresholds;
5752 if (!thresholds->primary)
5755 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5757 /* Check if a threshold crossed before removing */
5758 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5760 /* Calculate new number of threshold */
5762 for (i = 0; i < thresholds->primary->size; i++) {
5763 if (thresholds->primary->entries[i].eventfd != eventfd)
5767 new = thresholds->spare;
5769 /* Set thresholds array to NULL if we don't have thresholds */
5778 /* Copy thresholds and find current threshold */
5779 new->current_threshold = -1;
5780 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5781 if (thresholds->primary->entries[i].eventfd == eventfd)
5784 new->entries[j] = thresholds->primary->entries[i];
5785 if (new->entries[j].threshold <= usage) {
5787 * new->current_threshold will not be used
5788 * until rcu_assign_pointer(), so it's safe to increment
5791 ++new->current_threshold;
5797 /* Swap primary and spare array */
5798 thresholds->spare = thresholds->primary;
5799 /* If all events are unregistered, free the spare array */
5801 kfree(thresholds->spare);
5802 thresholds->spare = NULL;
5805 rcu_assign_pointer(thresholds->primary, new);
5807 /* To be sure that nobody uses thresholds */
5810 mutex_unlock(&memcg->thresholds_lock);
5813 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5814 struct eventfd_ctx *eventfd)
5816 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5819 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5820 struct eventfd_ctx *eventfd)
5822 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5825 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5826 struct eventfd_ctx *eventfd, const char *args)
5828 struct mem_cgroup_eventfd_list *event;
5830 event = kmalloc(sizeof(*event), GFP_KERNEL);
5834 spin_lock(&memcg_oom_lock);
5836 event->eventfd = eventfd;
5837 list_add(&event->list, &memcg->oom_notify);
5839 /* already in OOM ? */
5840 if (atomic_read(&memcg->under_oom))
5841 eventfd_signal(eventfd, 1);
5842 spin_unlock(&memcg_oom_lock);
5847 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5848 struct eventfd_ctx *eventfd)
5850 struct mem_cgroup_eventfd_list *ev, *tmp;
5852 spin_lock(&memcg_oom_lock);
5854 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5855 if (ev->eventfd == eventfd) {
5856 list_del(&ev->list);
5861 spin_unlock(&memcg_oom_lock);
5864 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5866 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5868 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5869 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5873 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5874 struct cftype *cft, u64 val)
5876 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5877 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5879 /* cannot set to root cgroup and only 0 and 1 are allowed */
5880 if (!parent || !((val == 0) || (val == 1)))
5883 mutex_lock(&memcg_create_mutex);
5884 /* oom-kill-disable is a flag for subhierarchy. */
5885 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5886 mutex_unlock(&memcg_create_mutex);
5889 memcg->oom_kill_disable = val;
5891 memcg_oom_recover(memcg);
5892 mutex_unlock(&memcg_create_mutex);
5896 #ifdef CONFIG_MEMCG_KMEM
5897 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5901 memcg->kmemcg_id = -1;
5902 ret = memcg_propagate_kmem(memcg);
5906 return mem_cgroup_sockets_init(memcg, ss);
5909 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5911 mem_cgroup_sockets_destroy(memcg);
5914 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5916 if (!memcg_kmem_is_active(memcg))
5920 * kmem charges can outlive the cgroup. In the case of slab
5921 * pages, for instance, a page contain objects from various
5922 * processes. As we prevent from taking a reference for every
5923 * such allocation we have to be careful when doing uncharge
5924 * (see memcg_uncharge_kmem) and here during offlining.
5926 * The idea is that that only the _last_ uncharge which sees
5927 * the dead memcg will drop the last reference. An additional
5928 * reference is taken here before the group is marked dead
5929 * which is then paired with css_put during uncharge resp. here.
5931 * Although this might sound strange as this path is called from
5932 * css_offline() when the referencemight have dropped down to 0
5933 * and shouldn't be incremented anymore (css_tryget would fail)
5934 * we do not have other options because of the kmem allocations
5937 css_get(&memcg->css);
5939 memcg_kmem_mark_dead(memcg);
5941 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5944 if (memcg_kmem_test_and_clear_dead(memcg))
5945 css_put(&memcg->css);
5948 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5953 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5957 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5963 * DO NOT USE IN NEW FILES.
5965 * "cgroup.event_control" implementation.
5967 * This is way over-engineered. It tries to support fully configurable
5968 * events for each user. Such level of flexibility is completely
5969 * unnecessary especially in the light of the planned unified hierarchy.
5971 * Please deprecate this and replace with something simpler if at all
5976 * Unregister event and free resources.
5978 * Gets called from workqueue.
5980 static void memcg_event_remove(struct work_struct *work)
5982 struct mem_cgroup_event *event =
5983 container_of(work, struct mem_cgroup_event, remove);
5984 struct mem_cgroup *memcg = event->memcg;
5986 remove_wait_queue(event->wqh, &event->wait);
5988 event->unregister_event(memcg, event->eventfd);
5990 /* Notify userspace the event is going away. */
5991 eventfd_signal(event->eventfd, 1);
5993 eventfd_ctx_put(event->eventfd);
5995 css_put(&memcg->css);
5999 * Gets called on POLLHUP on eventfd when user closes it.
6001 * Called with wqh->lock held and interrupts disabled.
6003 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
6004 int sync, void *key)
6006 struct mem_cgroup_event *event =
6007 container_of(wait, struct mem_cgroup_event, wait);
6008 struct mem_cgroup *memcg = event->memcg;
6009 unsigned long flags = (unsigned long)key;
6011 if (flags & POLLHUP) {
6013 * If the event has been detached at cgroup removal, we
6014 * can simply return knowing the other side will cleanup
6017 * We can't race against event freeing since the other
6018 * side will require wqh->lock via remove_wait_queue(),
6021 spin_lock(&memcg->event_list_lock);
6022 if (!list_empty(&event->list)) {
6023 list_del_init(&event->list);
6025 * We are in atomic context, but cgroup_event_remove()
6026 * may sleep, so we have to call it in workqueue.
6028 schedule_work(&event->remove);
6030 spin_unlock(&memcg->event_list_lock);
6036 static void memcg_event_ptable_queue_proc(struct file *file,
6037 wait_queue_head_t *wqh, poll_table *pt)
6039 struct mem_cgroup_event *event =
6040 container_of(pt, struct mem_cgroup_event, pt);
6043 add_wait_queue(wqh, &event->wait);
6047 * DO NOT USE IN NEW FILES.
6049 * Parse input and register new cgroup event handler.
6051 * Input must be in format '<event_fd> <control_fd> <args>'.
6052 * Interpretation of args is defined by control file implementation.
6054 static int memcg_write_event_control(struct cgroup_subsys_state *css,
6055 struct cftype *cft, char *buffer)
6057 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6058 struct mem_cgroup_event *event;
6059 struct cgroup_subsys_state *cfile_css;
6060 unsigned int efd, cfd;
6067 efd = simple_strtoul(buffer, &endp, 10);
6072 cfd = simple_strtoul(buffer, &endp, 10);
6073 if ((*endp != ' ') && (*endp != '\0'))
6077 event = kzalloc(sizeof(*event), GFP_KERNEL);
6081 event->memcg = memcg;
6082 INIT_LIST_HEAD(&event->list);
6083 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
6084 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
6085 INIT_WORK(&event->remove, memcg_event_remove);
6093 event->eventfd = eventfd_ctx_fileget(efile.file);
6094 if (IS_ERR(event->eventfd)) {
6095 ret = PTR_ERR(event->eventfd);
6102 goto out_put_eventfd;
6105 /* the process need read permission on control file */
6106 /* AV: shouldn't we check that it's been opened for read instead? */
6107 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6112 * Determine the event callbacks and set them in @event. This used
6113 * to be done via struct cftype but cgroup core no longer knows
6114 * about these events. The following is crude but the whole thing
6115 * is for compatibility anyway.
6117 * DO NOT ADD NEW FILES.
6119 name = cfile.file->f_dentry->d_name.name;
6121 if (!strcmp(name, "memory.usage_in_bytes")) {
6122 event->register_event = mem_cgroup_usage_register_event;
6123 event->unregister_event = mem_cgroup_usage_unregister_event;
6124 } else if (!strcmp(name, "memory.oom_control")) {
6125 event->register_event = mem_cgroup_oom_register_event;
6126 event->unregister_event = mem_cgroup_oom_unregister_event;
6127 } else if (!strcmp(name, "memory.pressure_level")) {
6128 event->register_event = vmpressure_register_event;
6129 event->unregister_event = vmpressure_unregister_event;
6130 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
6131 event->register_event = memsw_cgroup_usage_register_event;
6132 event->unregister_event = memsw_cgroup_usage_unregister_event;
6139 * Verify @cfile should belong to @css. Also, remaining events are
6140 * automatically removed on cgroup destruction but the removal is
6141 * asynchronous, so take an extra ref on @css.
6143 cfile_css = css_tryget_from_dir(cfile.file->f_dentry->d_parent,
6144 &memory_cgrp_subsys);
6146 if (IS_ERR(cfile_css))
6148 if (cfile_css != css) {
6153 ret = event->register_event(memcg, event->eventfd, buffer);
6157 efile.file->f_op->poll(efile.file, &event->pt);
6159 spin_lock(&memcg->event_list_lock);
6160 list_add(&event->list, &memcg->event_list);
6161 spin_unlock(&memcg->event_list_lock);
6173 eventfd_ctx_put(event->eventfd);
6182 static struct cftype mem_cgroup_files[] = {
6184 .name = "usage_in_bytes",
6185 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6186 .read_u64 = mem_cgroup_read_u64,
6189 .name = "max_usage_in_bytes",
6190 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6191 .trigger = mem_cgroup_reset,
6192 .read_u64 = mem_cgroup_read_u64,
6195 .name = "limit_in_bytes",
6196 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6197 .write_string = mem_cgroup_write,
6198 .read_u64 = mem_cgroup_read_u64,
6201 .name = "soft_limit_in_bytes",
6202 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6203 .write_string = mem_cgroup_write,
6204 .read_u64 = mem_cgroup_read_u64,
6208 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6209 .trigger = mem_cgroup_reset,
6210 .read_u64 = mem_cgroup_read_u64,
6214 .seq_show = memcg_stat_show,
6217 .name = "force_empty",
6218 .trigger = mem_cgroup_force_empty_write,
6221 .name = "use_hierarchy",
6222 .flags = CFTYPE_INSANE,
6223 .write_u64 = mem_cgroup_hierarchy_write,
6224 .read_u64 = mem_cgroup_hierarchy_read,
6227 .name = "cgroup.event_control", /* XXX: for compat */
6228 .write_string = memcg_write_event_control,
6229 .flags = CFTYPE_NO_PREFIX,
6233 .name = "swappiness",
6234 .read_u64 = mem_cgroup_swappiness_read,
6235 .write_u64 = mem_cgroup_swappiness_write,
6238 .name = "move_charge_at_immigrate",
6239 .read_u64 = mem_cgroup_move_charge_read,
6240 .write_u64 = mem_cgroup_move_charge_write,
6243 .name = "oom_control",
6244 .seq_show = mem_cgroup_oom_control_read,
6245 .write_u64 = mem_cgroup_oom_control_write,
6246 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6249 .name = "pressure_level",
6253 .name = "numa_stat",
6254 .seq_show = memcg_numa_stat_show,
6257 #ifdef CONFIG_MEMCG_KMEM
6259 .name = "kmem.limit_in_bytes",
6260 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6261 .write_string = mem_cgroup_write,
6262 .read_u64 = mem_cgroup_read_u64,
6265 .name = "kmem.usage_in_bytes",
6266 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6267 .read_u64 = mem_cgroup_read_u64,
6270 .name = "kmem.failcnt",
6271 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6272 .trigger = mem_cgroup_reset,
6273 .read_u64 = mem_cgroup_read_u64,
6276 .name = "kmem.max_usage_in_bytes",
6277 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6278 .trigger = mem_cgroup_reset,
6279 .read_u64 = mem_cgroup_read_u64,
6281 #ifdef CONFIG_SLABINFO
6283 .name = "kmem.slabinfo",
6284 .seq_show = mem_cgroup_slabinfo_read,
6288 { }, /* terminate */
6291 #ifdef CONFIG_MEMCG_SWAP
6292 static struct cftype memsw_cgroup_files[] = {
6294 .name = "memsw.usage_in_bytes",
6295 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6296 .read_u64 = mem_cgroup_read_u64,
6299 .name = "memsw.max_usage_in_bytes",
6300 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6301 .trigger = mem_cgroup_reset,
6302 .read_u64 = mem_cgroup_read_u64,
6305 .name = "memsw.limit_in_bytes",
6306 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6307 .write_string = mem_cgroup_write,
6308 .read_u64 = mem_cgroup_read_u64,
6311 .name = "memsw.failcnt",
6312 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6313 .trigger = mem_cgroup_reset,
6314 .read_u64 = mem_cgroup_read_u64,
6316 { }, /* terminate */
6319 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6321 struct mem_cgroup_per_node *pn;
6322 struct mem_cgroup_per_zone *mz;
6323 int zone, tmp = node;
6325 * This routine is called against possible nodes.
6326 * But it's BUG to call kmalloc() against offline node.
6328 * TODO: this routine can waste much memory for nodes which will
6329 * never be onlined. It's better to use memory hotplug callback
6332 if (!node_state(node, N_NORMAL_MEMORY))
6334 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6338 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6339 mz = &pn->zoneinfo[zone];
6340 lruvec_init(&mz->lruvec);
6341 mz->usage_in_excess = 0;
6342 mz->on_tree = false;
6345 memcg->nodeinfo[node] = pn;
6349 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6351 kfree(memcg->nodeinfo[node]);
6354 static struct mem_cgroup *mem_cgroup_alloc(void)
6356 struct mem_cgroup *memcg;
6359 size = sizeof(struct mem_cgroup);
6360 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6362 memcg = kzalloc(size, GFP_KERNEL);
6366 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6369 spin_lock_init(&memcg->pcp_counter_lock);
6378 * At destroying mem_cgroup, references from swap_cgroup can remain.
6379 * (scanning all at force_empty is too costly...)
6381 * Instead of clearing all references at force_empty, we remember
6382 * the number of reference from swap_cgroup and free mem_cgroup when
6383 * it goes down to 0.
6385 * Removal of cgroup itself succeeds regardless of refs from swap.
6388 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6392 mem_cgroup_remove_from_trees(memcg);
6395 free_mem_cgroup_per_zone_info(memcg, node);
6397 free_percpu(memcg->stat);
6400 * We need to make sure that (at least for now), the jump label
6401 * destruction code runs outside of the cgroup lock. This is because
6402 * get_online_cpus(), which is called from the static_branch update,
6403 * can't be called inside the cgroup_lock. cpusets are the ones
6404 * enforcing this dependency, so if they ever change, we might as well.
6406 * schedule_work() will guarantee this happens. Be careful if you need
6407 * to move this code around, and make sure it is outside
6410 disarm_static_keys(memcg);
6415 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6417 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6419 if (!memcg->res.parent)
6421 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6423 EXPORT_SYMBOL(parent_mem_cgroup);
6425 static void __init mem_cgroup_soft_limit_tree_init(void)
6427 struct mem_cgroup_tree_per_node *rtpn;
6428 struct mem_cgroup_tree_per_zone *rtpz;
6429 int tmp, node, zone;
6431 for_each_node(node) {
6433 if (!node_state(node, N_NORMAL_MEMORY))
6435 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6438 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6440 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6441 rtpz = &rtpn->rb_tree_per_zone[zone];
6442 rtpz->rb_root = RB_ROOT;
6443 spin_lock_init(&rtpz->lock);
6448 static struct cgroup_subsys_state * __ref
6449 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6451 struct mem_cgroup *memcg;
6452 long error = -ENOMEM;
6455 memcg = mem_cgroup_alloc();
6457 return ERR_PTR(error);
6460 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6464 if (parent_css == NULL) {
6465 root_mem_cgroup = memcg;
6466 res_counter_init(&memcg->res, NULL);
6467 res_counter_init(&memcg->memsw, NULL);
6468 res_counter_init(&memcg->kmem, NULL);
6471 memcg->last_scanned_node = MAX_NUMNODES;
6472 INIT_LIST_HEAD(&memcg->oom_notify);
6473 memcg->move_charge_at_immigrate = 0;
6474 mutex_init(&memcg->thresholds_lock);
6475 spin_lock_init(&memcg->move_lock);
6476 vmpressure_init(&memcg->vmpressure);
6477 INIT_LIST_HEAD(&memcg->event_list);
6478 spin_lock_init(&memcg->event_list_lock);
6483 __mem_cgroup_free(memcg);
6484 return ERR_PTR(error);
6488 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6490 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6491 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6493 if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6499 mutex_lock(&memcg_create_mutex);
6501 memcg->use_hierarchy = parent->use_hierarchy;
6502 memcg->oom_kill_disable = parent->oom_kill_disable;
6503 memcg->swappiness = mem_cgroup_swappiness(parent);
6505 if (parent->use_hierarchy) {
6506 res_counter_init(&memcg->res, &parent->res);
6507 res_counter_init(&memcg->memsw, &parent->memsw);
6508 res_counter_init(&memcg->kmem, &parent->kmem);
6511 * No need to take a reference to the parent because cgroup
6512 * core guarantees its existence.
6515 res_counter_init(&memcg->res, NULL);
6516 res_counter_init(&memcg->memsw, NULL);
6517 res_counter_init(&memcg->kmem, NULL);
6519 * Deeper hierachy with use_hierarchy == false doesn't make
6520 * much sense so let cgroup subsystem know about this
6521 * unfortunate state in our controller.
6523 if (parent != root_mem_cgroup)
6524 memory_cgrp_subsys.broken_hierarchy = true;
6526 mutex_unlock(&memcg_create_mutex);
6528 return memcg_init_kmem(memcg, &memory_cgrp_subsys);
6532 * Announce all parents that a group from their hierarchy is gone.
6534 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6536 struct mem_cgroup *parent = memcg;
6538 while ((parent = parent_mem_cgroup(parent)))
6539 mem_cgroup_iter_invalidate(parent);
6542 * if the root memcg is not hierarchical we have to check it
6545 if (!root_mem_cgroup->use_hierarchy)
6546 mem_cgroup_iter_invalidate(root_mem_cgroup);
6549 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6551 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6552 struct mem_cgroup_event *event, *tmp;
6553 struct cgroup_subsys_state *iter;
6556 * Unregister events and notify userspace.
6557 * Notify userspace about cgroup removing only after rmdir of cgroup
6558 * directory to avoid race between userspace and kernelspace.
6560 spin_lock(&memcg->event_list_lock);
6561 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6562 list_del_init(&event->list);
6563 schedule_work(&event->remove);
6565 spin_unlock(&memcg->event_list_lock);
6567 kmem_cgroup_css_offline(memcg);
6569 mem_cgroup_invalidate_reclaim_iterators(memcg);
6572 * This requires that offlining is serialized. Right now that is
6573 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6575 css_for_each_descendant_post(iter, css)
6576 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
6578 mem_cgroup_destroy_all_caches(memcg);
6579 vmpressure_cleanup(&memcg->vmpressure);
6582 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6584 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6586 * XXX: css_offline() would be where we should reparent all
6587 * memory to prepare the cgroup for destruction. However,
6588 * memcg does not do css_tryget() and res_counter charging
6589 * under the same RCU lock region, which means that charging
6590 * could race with offlining. Offlining only happens to
6591 * cgroups with no tasks in them but charges can show up
6592 * without any tasks from the swapin path when the target
6593 * memcg is looked up from the swapout record and not from the
6594 * current task as it usually is. A race like this can leak
6595 * charges and put pages with stale cgroup pointers into
6599 * lookup_swap_cgroup_id()
6601 * mem_cgroup_lookup()
6604 * disable css_tryget()
6607 * reparent_charges()
6608 * res_counter_charge()
6611 * pc->mem_cgroup = dead memcg
6614 * The bulk of the charges are still moved in offline_css() to
6615 * avoid pinning a lot of pages in case a long-term reference
6616 * like a swapout record is deferring the css_free() to long
6617 * after offlining. But this makes sure we catch any charges
6618 * made after offlining:
6620 mem_cgroup_reparent_charges(memcg);
6622 memcg_destroy_kmem(memcg);
6623 __mem_cgroup_free(memcg);
6627 /* Handlers for move charge at task migration. */
6628 #define PRECHARGE_COUNT_AT_ONCE 256
6629 static int mem_cgroup_do_precharge(unsigned long count)
6632 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6633 struct mem_cgroup *memcg = mc.to;
6635 if (mem_cgroup_is_root(memcg)) {
6636 mc.precharge += count;
6637 /* we don't need css_get for root */
6640 /* try to charge at once */
6642 struct res_counter *dummy;
6644 * "memcg" cannot be under rmdir() because we've already checked
6645 * by cgroup_lock_live_cgroup() that it is not removed and we
6646 * are still under the same cgroup_mutex. So we can postpone
6649 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6651 if (do_swap_account && res_counter_charge(&memcg->memsw,
6652 PAGE_SIZE * count, &dummy)) {
6653 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6656 mc.precharge += count;
6660 /* fall back to one by one charge */
6662 if (signal_pending(current)) {
6666 if (!batch_count--) {
6667 batch_count = PRECHARGE_COUNT_AT_ONCE;
6670 ret = __mem_cgroup_try_charge(NULL,
6671 GFP_KERNEL, 1, &memcg, false);
6673 /* mem_cgroup_clear_mc() will do uncharge later */
6681 * get_mctgt_type - get target type of moving charge
6682 * @vma: the vma the pte to be checked belongs
6683 * @addr: the address corresponding to the pte to be checked
6684 * @ptent: the pte to be checked
6685 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6688 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6689 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6690 * move charge. if @target is not NULL, the page is stored in target->page
6691 * with extra refcnt got(Callers should handle it).
6692 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6693 * target for charge migration. if @target is not NULL, the entry is stored
6696 * Called with pte lock held.
6703 enum mc_target_type {
6709 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6710 unsigned long addr, pte_t ptent)
6712 struct page *page = vm_normal_page(vma, addr, ptent);
6714 if (!page || !page_mapped(page))
6716 if (PageAnon(page)) {
6717 /* we don't move shared anon */
6720 } else if (!move_file())
6721 /* we ignore mapcount for file pages */
6723 if (!get_page_unless_zero(page))
6730 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6731 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6733 struct page *page = NULL;
6734 swp_entry_t ent = pte_to_swp_entry(ptent);
6736 if (!move_anon() || non_swap_entry(ent))
6739 * Because lookup_swap_cache() updates some statistics counter,
6740 * we call find_get_page() with swapper_space directly.
6742 page = find_get_page(swap_address_space(ent), ent.val);
6743 if (do_swap_account)
6744 entry->val = ent.val;
6749 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6750 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6756 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6757 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6759 struct page *page = NULL;
6760 struct address_space *mapping;
6763 if (!vma->vm_file) /* anonymous vma */
6768 mapping = vma->vm_file->f_mapping;
6769 if (pte_none(ptent))
6770 pgoff = linear_page_index(vma, addr);
6771 else /* pte_file(ptent) is true */
6772 pgoff = pte_to_pgoff(ptent);
6774 /* page is moved even if it's not RSS of this task(page-faulted). */
6775 page = find_get_page(mapping, pgoff);
6778 /* shmem/tmpfs may report page out on swap: account for that too. */
6779 if (radix_tree_exceptional_entry(page)) {
6780 swp_entry_t swap = radix_to_swp_entry(page);
6781 if (do_swap_account)
6783 page = find_get_page(swap_address_space(swap), swap.val);
6789 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6790 unsigned long addr, pte_t ptent, union mc_target *target)
6792 struct page *page = NULL;
6793 struct page_cgroup *pc;
6794 enum mc_target_type ret = MC_TARGET_NONE;
6795 swp_entry_t ent = { .val = 0 };
6797 if (pte_present(ptent))
6798 page = mc_handle_present_pte(vma, addr, ptent);
6799 else if (is_swap_pte(ptent))
6800 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6801 else if (pte_none(ptent) || pte_file(ptent))
6802 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6804 if (!page && !ent.val)
6807 pc = lookup_page_cgroup(page);
6809 * Do only loose check w/o page_cgroup lock.
6810 * mem_cgroup_move_account() checks the pc is valid or not under
6813 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6814 ret = MC_TARGET_PAGE;
6816 target->page = page;
6818 if (!ret || !target)
6821 /* There is a swap entry and a page doesn't exist or isn't charged */
6822 if (ent.val && !ret &&
6823 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6824 ret = MC_TARGET_SWAP;
6831 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6833 * We don't consider swapping or file mapped pages because THP does not
6834 * support them for now.
6835 * Caller should make sure that pmd_trans_huge(pmd) is true.
6837 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6838 unsigned long addr, pmd_t pmd, union mc_target *target)
6840 struct page *page = NULL;
6841 struct page_cgroup *pc;
6842 enum mc_target_type ret = MC_TARGET_NONE;
6844 page = pmd_page(pmd);
6845 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6848 pc = lookup_page_cgroup(page);
6849 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6850 ret = MC_TARGET_PAGE;
6853 target->page = page;
6859 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6860 unsigned long addr, pmd_t pmd, union mc_target *target)
6862 return MC_TARGET_NONE;
6866 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6867 unsigned long addr, unsigned long end,
6868 struct mm_walk *walk)
6870 struct vm_area_struct *vma = walk->private;
6874 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6875 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6876 mc.precharge += HPAGE_PMD_NR;
6881 if (pmd_trans_unstable(pmd))
6883 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6884 for (; addr != end; pte++, addr += PAGE_SIZE)
6885 if (get_mctgt_type(vma, addr, *pte, NULL))
6886 mc.precharge++; /* increment precharge temporarily */
6887 pte_unmap_unlock(pte - 1, ptl);
6893 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6895 unsigned long precharge;
6896 struct vm_area_struct *vma;
6898 down_read(&mm->mmap_sem);
6899 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6900 struct mm_walk mem_cgroup_count_precharge_walk = {
6901 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6905 if (is_vm_hugetlb_page(vma))
6907 walk_page_range(vma->vm_start, vma->vm_end,
6908 &mem_cgroup_count_precharge_walk);
6910 up_read(&mm->mmap_sem);
6912 precharge = mc.precharge;
6918 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6920 unsigned long precharge = mem_cgroup_count_precharge(mm);
6922 VM_BUG_ON(mc.moving_task);
6923 mc.moving_task = current;
6924 return mem_cgroup_do_precharge(precharge);
6927 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6928 static void __mem_cgroup_clear_mc(void)
6930 struct mem_cgroup *from = mc.from;
6931 struct mem_cgroup *to = mc.to;
6934 /* we must uncharge all the leftover precharges from mc.to */
6936 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6940 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6941 * we must uncharge here.
6943 if (mc.moved_charge) {
6944 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6945 mc.moved_charge = 0;
6947 /* we must fixup refcnts and charges */
6948 if (mc.moved_swap) {
6949 /* uncharge swap account from the old cgroup */
6950 if (!mem_cgroup_is_root(mc.from))
6951 res_counter_uncharge(&mc.from->memsw,
6952 PAGE_SIZE * mc.moved_swap);
6954 for (i = 0; i < mc.moved_swap; i++)
6955 css_put(&mc.from->css);
6957 if (!mem_cgroup_is_root(mc.to)) {
6959 * we charged both to->res and to->memsw, so we should
6962 res_counter_uncharge(&mc.to->res,
6963 PAGE_SIZE * mc.moved_swap);
6965 /* we've already done css_get(mc.to) */
6968 memcg_oom_recover(from);
6969 memcg_oom_recover(to);
6970 wake_up_all(&mc.waitq);
6973 static void mem_cgroup_clear_mc(void)
6975 struct mem_cgroup *from = mc.from;
6978 * we must clear moving_task before waking up waiters at the end of
6981 mc.moving_task = NULL;
6982 __mem_cgroup_clear_mc();
6983 spin_lock(&mc.lock);
6986 spin_unlock(&mc.lock);
6987 mem_cgroup_end_move(from);
6990 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6991 struct cgroup_taskset *tset)
6993 struct task_struct *p = cgroup_taskset_first(tset);
6995 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6996 unsigned long move_charge_at_immigrate;
6999 * We are now commited to this value whatever it is. Changes in this
7000 * tunable will only affect upcoming migrations, not the current one.
7001 * So we need to save it, and keep it going.
7003 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
7004 if (move_charge_at_immigrate) {
7005 struct mm_struct *mm;
7006 struct mem_cgroup *from = mem_cgroup_from_task(p);
7008 VM_BUG_ON(from == memcg);
7010 mm = get_task_mm(p);
7013 /* We move charges only when we move a owner of the mm */
7014 if (mm->owner == p) {
7017 VM_BUG_ON(mc.precharge);
7018 VM_BUG_ON(mc.moved_charge);
7019 VM_BUG_ON(mc.moved_swap);
7020 mem_cgroup_start_move(from);
7021 spin_lock(&mc.lock);
7024 mc.immigrate_flags = move_charge_at_immigrate;
7025 spin_unlock(&mc.lock);
7026 /* We set mc.moving_task later */
7028 ret = mem_cgroup_precharge_mc(mm);
7030 mem_cgroup_clear_mc();
7037 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7038 struct cgroup_taskset *tset)
7040 mem_cgroup_clear_mc();
7043 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
7044 unsigned long addr, unsigned long end,
7045 struct mm_walk *walk)
7048 struct vm_area_struct *vma = walk->private;
7051 enum mc_target_type target_type;
7052 union mc_target target;
7054 struct page_cgroup *pc;
7057 * We don't take compound_lock() here but no race with splitting thp
7059 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7060 * under splitting, which means there's no concurrent thp split,
7061 * - if another thread runs into split_huge_page() just after we
7062 * entered this if-block, the thread must wait for page table lock
7063 * to be unlocked in __split_huge_page_splitting(), where the main
7064 * part of thp split is not executed yet.
7066 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
7067 if (mc.precharge < HPAGE_PMD_NR) {
7071 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
7072 if (target_type == MC_TARGET_PAGE) {
7074 if (!isolate_lru_page(page)) {
7075 pc = lookup_page_cgroup(page);
7076 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
7077 pc, mc.from, mc.to)) {
7078 mc.precharge -= HPAGE_PMD_NR;
7079 mc.moved_charge += HPAGE_PMD_NR;
7081 putback_lru_page(page);
7089 if (pmd_trans_unstable(pmd))
7092 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7093 for (; addr != end; addr += PAGE_SIZE) {
7094 pte_t ptent = *(pte++);
7100 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7101 case MC_TARGET_PAGE:
7103 if (isolate_lru_page(page))
7105 pc = lookup_page_cgroup(page);
7106 if (!mem_cgroup_move_account(page, 1, pc,
7109 /* we uncharge from mc.from later. */
7112 putback_lru_page(page);
7113 put: /* get_mctgt_type() gets the page */
7116 case MC_TARGET_SWAP:
7118 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7120 /* we fixup refcnts and charges later. */
7128 pte_unmap_unlock(pte - 1, ptl);
7133 * We have consumed all precharges we got in can_attach().
7134 * We try charge one by one, but don't do any additional
7135 * charges to mc.to if we have failed in charge once in attach()
7138 ret = mem_cgroup_do_precharge(1);
7146 static void mem_cgroup_move_charge(struct mm_struct *mm)
7148 struct vm_area_struct *vma;
7150 lru_add_drain_all();
7152 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7154 * Someone who are holding the mmap_sem might be waiting in
7155 * waitq. So we cancel all extra charges, wake up all waiters,
7156 * and retry. Because we cancel precharges, we might not be able
7157 * to move enough charges, but moving charge is a best-effort
7158 * feature anyway, so it wouldn't be a big problem.
7160 __mem_cgroup_clear_mc();
7164 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7166 struct mm_walk mem_cgroup_move_charge_walk = {
7167 .pmd_entry = mem_cgroup_move_charge_pte_range,
7171 if (is_vm_hugetlb_page(vma))
7173 ret = walk_page_range(vma->vm_start, vma->vm_end,
7174 &mem_cgroup_move_charge_walk);
7177 * means we have consumed all precharges and failed in
7178 * doing additional charge. Just abandon here.
7182 up_read(&mm->mmap_sem);
7185 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7186 struct cgroup_taskset *tset)
7188 struct task_struct *p = cgroup_taskset_first(tset);
7189 struct mm_struct *mm = get_task_mm(p);
7193 mem_cgroup_move_charge(mm);
7197 mem_cgroup_clear_mc();
7199 #else /* !CONFIG_MMU */
7200 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7201 struct cgroup_taskset *tset)
7205 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7206 struct cgroup_taskset *tset)
7209 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7210 struct cgroup_taskset *tset)
7216 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7217 * to verify sane_behavior flag on each mount attempt.
7219 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7222 * use_hierarchy is forced with sane_behavior. cgroup core
7223 * guarantees that @root doesn't have any children, so turning it
7224 * on for the root memcg is enough.
7226 if (cgroup_sane_behavior(root_css->cgroup))
7227 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7230 struct cgroup_subsys memory_cgrp_subsys = {
7231 .css_alloc = mem_cgroup_css_alloc,
7232 .css_online = mem_cgroup_css_online,
7233 .css_offline = mem_cgroup_css_offline,
7234 .css_free = mem_cgroup_css_free,
7235 .can_attach = mem_cgroup_can_attach,
7236 .cancel_attach = mem_cgroup_cancel_attach,
7237 .attach = mem_cgroup_move_task,
7238 .bind = mem_cgroup_bind,
7239 .base_cftypes = mem_cgroup_files,
7243 #ifdef CONFIG_MEMCG_SWAP
7244 static int __init enable_swap_account(char *s)
7246 if (!strcmp(s, "1"))
7247 really_do_swap_account = 1;
7248 else if (!strcmp(s, "0"))
7249 really_do_swap_account = 0;
7252 __setup("swapaccount=", enable_swap_account);
7254 static void __init memsw_file_init(void)
7256 WARN_ON(cgroup_add_cftypes(&memory_cgrp_subsys, memsw_cgroup_files));
7259 static void __init enable_swap_cgroup(void)
7261 if (!mem_cgroup_disabled() && really_do_swap_account) {
7262 do_swap_account = 1;
7268 static void __init enable_swap_cgroup(void)
7274 * subsys_initcall() for memory controller.
7276 * Some parts like hotcpu_notifier() have to be initialized from this context
7277 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7278 * everything that doesn't depend on a specific mem_cgroup structure should
7279 * be initialized from here.
7281 static int __init mem_cgroup_init(void)
7283 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7284 enable_swap_cgroup();
7285 mem_cgroup_soft_limit_tree_init();
7289 subsys_initcall(mem_cgroup_init);