1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata = 0;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct mem_cgroup_reclaim_iter {
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup *last_visited;
154 /* scan generation, increased every round-trip */
155 unsigned int generation;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone {
162 struct lruvec lruvec;
163 unsigned long lru_size[NR_LRU_LISTS];
165 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
167 struct rb_node tree_node; /* RB tree node */
168 unsigned long long usage_in_excess;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup *memcg; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node {
176 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone {
185 struct rb_root rb_root;
189 struct mem_cgroup_tree_per_node {
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
193 struct mem_cgroup_tree {
194 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
199 struct mem_cgroup_threshold {
200 struct eventfd_ctx *eventfd;
205 struct mem_cgroup_threshold_ary {
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries[0];
214 struct mem_cgroup_thresholds {
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary *primary;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary *spare;
226 struct mem_cgroup_eventfd_list {
227 struct list_head list;
228 struct eventfd_ctx *eventfd;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event {
236 * memcg which the event belongs to.
238 struct mem_cgroup *memcg;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx *eventfd;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event)(struct mem_cgroup *memcg,
253 struct eventfd_ctx *eventfd, const char *args);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event)(struct mem_cgroup *memcg,
260 struct eventfd_ctx *eventfd);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t *wqh;
268 struct work_struct remove;
271 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
272 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css;
288 * the counter to account for memory usage
290 struct res_counter res;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure;
296 * the counter to account for mem+swap usage.
298 struct res_counter memsw;
301 * the counter to account for kernel memory usage.
303 struct res_counter kmem;
305 * Should the accounting and control be hierarchical, per subtree?
308 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
312 atomic_t oom_wakeups;
315 /* OOM-Killer disable */
316 int oom_kill_disable;
318 /* set when res.limit == memsw.limit */
319 bool memsw_is_minimum;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds;
330 /* For oom notifier event fd */
331 struct list_head oom_notify;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock;
347 struct mem_cgroup_stat_cpu __percpu *stat;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base;
353 spinlock_t pcp_counter_lock;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list. per-memcg */
361 struct list_head memcg_slab_caches;
362 /* Not a spinlock, we can take a lot of time walking the list */
363 struct mutex slab_caches_mutex;
364 /* Index in the kmem_cache->memcg_params->memcg_caches array */
368 int last_scanned_node;
370 nodemask_t scan_nodes;
371 atomic_t numainfo_events;
372 atomic_t numainfo_updating;
375 /* List of events which userspace want to receive */
376 struct list_head event_list;
377 spinlock_t event_list_lock;
379 struct mem_cgroup_per_node *nodeinfo[0];
380 /* WARNING: nodeinfo must be the last member here */
383 /* internal only representation about the status of kmem accounting. */
385 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
386 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
389 #ifdef CONFIG_MEMCG_KMEM
390 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
392 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
395 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
397 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
400 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
403 * Our caller must use css_get() first, because memcg_uncharge_kmem()
404 * will call css_put() if it sees the memcg is dead.
407 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
408 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
411 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
413 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
414 &memcg->kmem_account_flags);
418 /* Stuffs for move charges at task migration. */
420 * Types of charges to be moved. "move_charge_at_immitgrate" and
421 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
424 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
425 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
429 /* "mc" and its members are protected by cgroup_mutex */
430 static struct move_charge_struct {
431 spinlock_t lock; /* for from, to */
432 struct mem_cgroup *from;
433 struct mem_cgroup *to;
434 unsigned long immigrate_flags;
435 unsigned long precharge;
436 unsigned long moved_charge;
437 unsigned long moved_swap;
438 struct task_struct *moving_task; /* a task moving charges */
439 wait_queue_head_t waitq; /* a waitq for other context */
441 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
442 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
445 static bool move_anon(void)
447 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
450 static bool move_file(void)
452 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
456 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
457 * limit reclaim to prevent infinite loops, if they ever occur.
459 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
460 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
463 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
464 MEM_CGROUP_CHARGE_TYPE_ANON,
465 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
466 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
470 /* for encoding cft->private value on file */
478 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
479 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
480 #define MEMFILE_ATTR(val) ((val) & 0xffff)
481 /* Used for OOM nofiier */
482 #define OOM_CONTROL (0)
485 * Reclaim flags for mem_cgroup_hierarchical_reclaim
487 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
488 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
489 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
490 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
493 * The memcg_create_mutex will be held whenever a new cgroup is created.
494 * As a consequence, any change that needs to protect against new child cgroups
495 * appearing has to hold it as well.
497 static DEFINE_MUTEX(memcg_create_mutex);
499 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
501 return s ? container_of(s, struct mem_cgroup, css) : NULL;
504 /* Some nice accessors for the vmpressure. */
505 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
508 memcg = root_mem_cgroup;
509 return &memcg->vmpressure;
512 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
514 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
517 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
519 return (memcg == root_mem_cgroup);
523 * We restrict the id in the range of [1, 65535], so it can fit into
526 #define MEM_CGROUP_ID_MAX USHRT_MAX
528 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
531 * The ID of the root cgroup is 0, but memcg treat 0 as an
532 * invalid ID, so we return (cgroup_id + 1).
534 return memcg->css.cgroup->id + 1;
537 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
539 struct cgroup_subsys_state *css;
541 css = css_from_id(id - 1, &memory_cgrp_subsys);
542 return mem_cgroup_from_css(css);
545 /* Writing them here to avoid exposing memcg's inner layout */
546 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
548 void sock_update_memcg(struct sock *sk)
550 if (mem_cgroup_sockets_enabled) {
551 struct mem_cgroup *memcg;
552 struct cg_proto *cg_proto;
554 BUG_ON(!sk->sk_prot->proto_cgroup);
556 /* Socket cloning can throw us here with sk_cgrp already
557 * filled. It won't however, necessarily happen from
558 * process context. So the test for root memcg given
559 * the current task's memcg won't help us in this case.
561 * Respecting the original socket's memcg is a better
562 * decision in this case.
565 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
566 css_get(&sk->sk_cgrp->memcg->css);
571 memcg = mem_cgroup_from_task(current);
572 cg_proto = sk->sk_prot->proto_cgroup(memcg);
573 if (!mem_cgroup_is_root(memcg) &&
574 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
575 sk->sk_cgrp = cg_proto;
580 EXPORT_SYMBOL(sock_update_memcg);
582 void sock_release_memcg(struct sock *sk)
584 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
585 struct mem_cgroup *memcg;
586 WARN_ON(!sk->sk_cgrp->memcg);
587 memcg = sk->sk_cgrp->memcg;
588 css_put(&sk->sk_cgrp->memcg->css);
592 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
594 if (!memcg || mem_cgroup_is_root(memcg))
597 return &memcg->tcp_mem;
599 EXPORT_SYMBOL(tcp_proto_cgroup);
601 static void disarm_sock_keys(struct mem_cgroup *memcg)
603 if (!memcg_proto_activated(&memcg->tcp_mem))
605 static_key_slow_dec(&memcg_socket_limit_enabled);
608 static void disarm_sock_keys(struct mem_cgroup *memcg)
613 #ifdef CONFIG_MEMCG_KMEM
615 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
616 * The main reason for not using cgroup id for this:
617 * this works better in sparse environments, where we have a lot of memcgs,
618 * but only a few kmem-limited. Or also, if we have, for instance, 200
619 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
620 * 200 entry array for that.
622 * The current size of the caches array is stored in
623 * memcg_limited_groups_array_size. It will double each time we have to
626 static DEFINE_IDA(kmem_limited_groups);
627 int memcg_limited_groups_array_size;
630 * MIN_SIZE is different than 1, because we would like to avoid going through
631 * the alloc/free process all the time. In a small machine, 4 kmem-limited
632 * cgroups is a reasonable guess. In the future, it could be a parameter or
633 * tunable, but that is strictly not necessary.
635 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
636 * this constant directly from cgroup, but it is understandable that this is
637 * better kept as an internal representation in cgroup.c. In any case, the
638 * cgrp_id space is not getting any smaller, and we don't have to necessarily
639 * increase ours as well if it increases.
641 #define MEMCG_CACHES_MIN_SIZE 4
642 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
645 * A lot of the calls to the cache allocation functions are expected to be
646 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
647 * conditional to this static branch, we'll have to allow modules that does
648 * kmem_cache_alloc and the such to see this symbol as well
650 struct static_key memcg_kmem_enabled_key;
651 EXPORT_SYMBOL(memcg_kmem_enabled_key);
653 static void disarm_kmem_keys(struct mem_cgroup *memcg)
655 if (memcg_kmem_is_active(memcg)) {
656 static_key_slow_dec(&memcg_kmem_enabled_key);
657 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
660 * This check can't live in kmem destruction function,
661 * since the charges will outlive the cgroup
663 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
666 static void disarm_kmem_keys(struct mem_cgroup *memcg)
669 #endif /* CONFIG_MEMCG_KMEM */
671 static void disarm_static_keys(struct mem_cgroup *memcg)
673 disarm_sock_keys(memcg);
674 disarm_kmem_keys(memcg);
677 static void drain_all_stock_async(struct mem_cgroup *memcg);
679 static struct mem_cgroup_per_zone *
680 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
682 VM_BUG_ON((unsigned)nid >= nr_node_ids);
683 return &memcg->nodeinfo[nid]->zoneinfo[zid];
686 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
691 static struct mem_cgroup_per_zone *
692 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
694 int nid = page_to_nid(page);
695 int zid = page_zonenum(page);
697 return mem_cgroup_zoneinfo(memcg, nid, zid);
700 static struct mem_cgroup_tree_per_zone *
701 soft_limit_tree_node_zone(int nid, int zid)
703 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
706 static struct mem_cgroup_tree_per_zone *
707 soft_limit_tree_from_page(struct page *page)
709 int nid = page_to_nid(page);
710 int zid = page_zonenum(page);
712 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
716 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
717 struct mem_cgroup_per_zone *mz,
718 struct mem_cgroup_tree_per_zone *mctz,
719 unsigned long long new_usage_in_excess)
721 struct rb_node **p = &mctz->rb_root.rb_node;
722 struct rb_node *parent = NULL;
723 struct mem_cgroup_per_zone *mz_node;
728 mz->usage_in_excess = new_usage_in_excess;
729 if (!mz->usage_in_excess)
733 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
735 if (mz->usage_in_excess < mz_node->usage_in_excess)
738 * We can't avoid mem cgroups that are over their soft
739 * limit by the same amount
741 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
744 rb_link_node(&mz->tree_node, parent, p);
745 rb_insert_color(&mz->tree_node, &mctz->rb_root);
750 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
751 struct mem_cgroup_per_zone *mz,
752 struct mem_cgroup_tree_per_zone *mctz)
756 rb_erase(&mz->tree_node, &mctz->rb_root);
761 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
762 struct mem_cgroup_per_zone *mz,
763 struct mem_cgroup_tree_per_zone *mctz)
765 spin_lock(&mctz->lock);
766 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
767 spin_unlock(&mctz->lock);
771 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
773 unsigned long long excess;
774 struct mem_cgroup_per_zone *mz;
775 struct mem_cgroup_tree_per_zone *mctz;
776 int nid = page_to_nid(page);
777 int zid = page_zonenum(page);
778 mctz = soft_limit_tree_from_page(page);
781 * Necessary to update all ancestors when hierarchy is used.
782 * because their event counter is not touched.
784 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
785 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
786 excess = res_counter_soft_limit_excess(&memcg->res);
788 * We have to update the tree if mz is on RB-tree or
789 * mem is over its softlimit.
791 if (excess || mz->on_tree) {
792 spin_lock(&mctz->lock);
793 /* if on-tree, remove it */
795 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
797 * Insert again. mz->usage_in_excess will be updated.
798 * If excess is 0, no tree ops.
800 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
801 spin_unlock(&mctz->lock);
806 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
809 struct mem_cgroup_per_zone *mz;
810 struct mem_cgroup_tree_per_zone *mctz;
812 for_each_node(node) {
813 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
814 mz = mem_cgroup_zoneinfo(memcg, node, zone);
815 mctz = soft_limit_tree_node_zone(node, zone);
816 mem_cgroup_remove_exceeded(memcg, mz, mctz);
821 static struct mem_cgroup_per_zone *
822 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
824 struct rb_node *rightmost = NULL;
825 struct mem_cgroup_per_zone *mz;
829 rightmost = rb_last(&mctz->rb_root);
831 goto done; /* Nothing to reclaim from */
833 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
835 * Remove the node now but someone else can add it back,
836 * we will to add it back at the end of reclaim to its correct
837 * position in the tree.
839 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
840 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
841 !css_tryget(&mz->memcg->css))
847 static struct mem_cgroup_per_zone *
848 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
850 struct mem_cgroup_per_zone *mz;
852 spin_lock(&mctz->lock);
853 mz = __mem_cgroup_largest_soft_limit_node(mctz);
854 spin_unlock(&mctz->lock);
859 * Implementation Note: reading percpu statistics for memcg.
861 * Both of vmstat[] and percpu_counter has threshold and do periodic
862 * synchronization to implement "quick" read. There are trade-off between
863 * reading cost and precision of value. Then, we may have a chance to implement
864 * a periodic synchronizion of counter in memcg's counter.
866 * But this _read() function is used for user interface now. The user accounts
867 * memory usage by memory cgroup and he _always_ requires exact value because
868 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
869 * have to visit all online cpus and make sum. So, for now, unnecessary
870 * synchronization is not implemented. (just implemented for cpu hotplug)
872 * If there are kernel internal actions which can make use of some not-exact
873 * value, and reading all cpu value can be performance bottleneck in some
874 * common workload, threashold and synchonization as vmstat[] should be
877 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
878 enum mem_cgroup_stat_index idx)
884 for_each_online_cpu(cpu)
885 val += per_cpu(memcg->stat->count[idx], cpu);
886 #ifdef CONFIG_HOTPLUG_CPU
887 spin_lock(&memcg->pcp_counter_lock);
888 val += memcg->nocpu_base.count[idx];
889 spin_unlock(&memcg->pcp_counter_lock);
895 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
898 int val = (charge) ? 1 : -1;
899 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
902 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
903 enum mem_cgroup_events_index idx)
905 unsigned long val = 0;
909 for_each_online_cpu(cpu)
910 val += per_cpu(memcg->stat->events[idx], cpu);
911 #ifdef CONFIG_HOTPLUG_CPU
912 spin_lock(&memcg->pcp_counter_lock);
913 val += memcg->nocpu_base.events[idx];
914 spin_unlock(&memcg->pcp_counter_lock);
920 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
922 bool anon, int nr_pages)
925 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
926 * counted as CACHE even if it's on ANON LRU.
929 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
932 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
935 if (PageTransHuge(page))
936 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
939 /* pagein of a big page is an event. So, ignore page size */
941 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
943 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
944 nr_pages = -nr_pages; /* for event */
947 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
951 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
953 struct mem_cgroup_per_zone *mz;
955 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
956 return mz->lru_size[lru];
960 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
961 unsigned int lru_mask)
963 struct mem_cgroup_per_zone *mz;
965 unsigned long ret = 0;
967 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
970 if (BIT(lru) & lru_mask)
971 ret += mz->lru_size[lru];
977 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
978 int nid, unsigned int lru_mask)
983 for (zid = 0; zid < MAX_NR_ZONES; zid++)
984 total += mem_cgroup_zone_nr_lru_pages(memcg,
990 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
991 unsigned int lru_mask)
996 for_each_node_state(nid, N_MEMORY)
997 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1001 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1002 enum mem_cgroup_events_target target)
1004 unsigned long val, next;
1006 val = __this_cpu_read(memcg->stat->nr_page_events);
1007 next = __this_cpu_read(memcg->stat->targets[target]);
1008 /* from time_after() in jiffies.h */
1009 if ((long)next - (long)val < 0) {
1011 case MEM_CGROUP_TARGET_THRESH:
1012 next = val + THRESHOLDS_EVENTS_TARGET;
1014 case MEM_CGROUP_TARGET_SOFTLIMIT:
1015 next = val + SOFTLIMIT_EVENTS_TARGET;
1017 case MEM_CGROUP_TARGET_NUMAINFO:
1018 next = val + NUMAINFO_EVENTS_TARGET;
1023 __this_cpu_write(memcg->stat->targets[target], next);
1030 * Check events in order.
1033 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1036 /* threshold event is triggered in finer grain than soft limit */
1037 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1038 MEM_CGROUP_TARGET_THRESH))) {
1040 bool do_numainfo __maybe_unused;
1042 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1043 MEM_CGROUP_TARGET_SOFTLIMIT);
1044 #if MAX_NUMNODES > 1
1045 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1046 MEM_CGROUP_TARGET_NUMAINFO);
1050 mem_cgroup_threshold(memcg);
1051 if (unlikely(do_softlimit))
1052 mem_cgroup_update_tree(memcg, page);
1053 #if MAX_NUMNODES > 1
1054 if (unlikely(do_numainfo))
1055 atomic_inc(&memcg->numainfo_events);
1061 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1064 * mm_update_next_owner() may clear mm->owner to NULL
1065 * if it races with swapoff, page migration, etc.
1066 * So this can be called with p == NULL.
1071 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1074 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1076 struct mem_cgroup *memcg = NULL;
1081 * Page cache insertions can happen withou an
1082 * actual mm context, e.g. during disk probing
1083 * on boot, loopback IO, acct() writes etc.
1086 memcg = root_mem_cgroup;
1088 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1089 if (unlikely(!memcg))
1090 memcg = root_mem_cgroup;
1092 } while (!css_tryget(&memcg->css));
1098 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1099 * ref. count) or NULL if the whole root's subtree has been visited.
1101 * helper function to be used by mem_cgroup_iter
1103 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1104 struct mem_cgroup *last_visited)
1106 struct cgroup_subsys_state *prev_css, *next_css;
1108 prev_css = last_visited ? &last_visited->css : NULL;
1110 next_css = css_next_descendant_pre(prev_css, &root->css);
1113 * Even if we found a group we have to make sure it is
1114 * alive. css && !memcg means that the groups should be
1115 * skipped and we should continue the tree walk.
1116 * last_visited css is safe to use because it is
1117 * protected by css_get and the tree walk is rcu safe.
1119 * We do not take a reference on the root of the tree walk
1120 * because we might race with the root removal when it would
1121 * be the only node in the iterated hierarchy and mem_cgroup_iter
1122 * would end up in an endless loop because it expects that at
1123 * least one valid node will be returned. Root cannot disappear
1124 * because caller of the iterator should hold it already so
1125 * skipping css reference should be safe.
1128 if ((next_css == &root->css) ||
1129 ((next_css->flags & CSS_ONLINE) && css_tryget(next_css)))
1130 return mem_cgroup_from_css(next_css);
1132 prev_css = next_css;
1139 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1142 * When a group in the hierarchy below root is destroyed, the
1143 * hierarchy iterator can no longer be trusted since it might
1144 * have pointed to the destroyed group. Invalidate it.
1146 atomic_inc(&root->dead_count);
1149 static struct mem_cgroup *
1150 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1151 struct mem_cgroup *root,
1154 struct mem_cgroup *position = NULL;
1156 * A cgroup destruction happens in two stages: offlining and
1157 * release. They are separated by a RCU grace period.
1159 * If the iterator is valid, we may still race with an
1160 * offlining. The RCU lock ensures the object won't be
1161 * released, tryget will fail if we lost the race.
1163 *sequence = atomic_read(&root->dead_count);
1164 if (iter->last_dead_count == *sequence) {
1166 position = iter->last_visited;
1169 * We cannot take a reference to root because we might race
1170 * with root removal and returning NULL would end up in
1171 * an endless loop on the iterator user level when root
1172 * would be returned all the time.
1174 if (position && position != root &&
1175 !css_tryget(&position->css))
1181 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1182 struct mem_cgroup *last_visited,
1183 struct mem_cgroup *new_position,
1184 struct mem_cgroup *root,
1187 /* root reference counting symmetric to mem_cgroup_iter_load */
1188 if (last_visited && last_visited != root)
1189 css_put(&last_visited->css);
1191 * We store the sequence count from the time @last_visited was
1192 * loaded successfully instead of rereading it here so that we
1193 * don't lose destruction events in between. We could have
1194 * raced with the destruction of @new_position after all.
1196 iter->last_visited = new_position;
1198 iter->last_dead_count = sequence;
1202 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1203 * @root: hierarchy root
1204 * @prev: previously returned memcg, NULL on first invocation
1205 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1207 * Returns references to children of the hierarchy below @root, or
1208 * @root itself, or %NULL after a full round-trip.
1210 * Caller must pass the return value in @prev on subsequent
1211 * invocations for reference counting, or use mem_cgroup_iter_break()
1212 * to cancel a hierarchy walk before the round-trip is complete.
1214 * Reclaimers can specify a zone and a priority level in @reclaim to
1215 * divide up the memcgs in the hierarchy among all concurrent
1216 * reclaimers operating on the same zone and priority.
1218 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1219 struct mem_cgroup *prev,
1220 struct mem_cgroup_reclaim_cookie *reclaim)
1222 struct mem_cgroup *memcg = NULL;
1223 struct mem_cgroup *last_visited = NULL;
1225 if (mem_cgroup_disabled())
1229 root = root_mem_cgroup;
1231 if (prev && !reclaim)
1232 last_visited = prev;
1234 if (!root->use_hierarchy && root != root_mem_cgroup) {
1242 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1243 int uninitialized_var(seq);
1246 int nid = zone_to_nid(reclaim->zone);
1247 int zid = zone_idx(reclaim->zone);
1248 struct mem_cgroup_per_zone *mz;
1250 mz = mem_cgroup_zoneinfo(root, nid, zid);
1251 iter = &mz->reclaim_iter[reclaim->priority];
1252 if (prev && reclaim->generation != iter->generation) {
1253 iter->last_visited = NULL;
1257 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1260 memcg = __mem_cgroup_iter_next(root, last_visited);
1263 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1268 else if (!prev && memcg)
1269 reclaim->generation = iter->generation;
1278 if (prev && prev != root)
1279 css_put(&prev->css);
1285 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1286 * @root: hierarchy root
1287 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1289 void mem_cgroup_iter_break(struct mem_cgroup *root,
1290 struct mem_cgroup *prev)
1293 root = root_mem_cgroup;
1294 if (prev && prev != root)
1295 css_put(&prev->css);
1299 * Iteration constructs for visiting all cgroups (under a tree). If
1300 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1301 * be used for reference counting.
1303 #define for_each_mem_cgroup_tree(iter, root) \
1304 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1306 iter = mem_cgroup_iter(root, iter, NULL))
1308 #define for_each_mem_cgroup(iter) \
1309 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1311 iter = mem_cgroup_iter(NULL, iter, NULL))
1313 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1315 struct mem_cgroup *memcg;
1318 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1319 if (unlikely(!memcg))
1324 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1327 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1335 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1338 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1339 * @zone: zone of the wanted lruvec
1340 * @memcg: memcg of the wanted lruvec
1342 * Returns the lru list vector holding pages for the given @zone and
1343 * @mem. This can be the global zone lruvec, if the memory controller
1346 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1347 struct mem_cgroup *memcg)
1349 struct mem_cgroup_per_zone *mz;
1350 struct lruvec *lruvec;
1352 if (mem_cgroup_disabled()) {
1353 lruvec = &zone->lruvec;
1357 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1358 lruvec = &mz->lruvec;
1361 * Since a node can be onlined after the mem_cgroup was created,
1362 * we have to be prepared to initialize lruvec->zone here;
1363 * and if offlined then reonlined, we need to reinitialize it.
1365 if (unlikely(lruvec->zone != zone))
1366 lruvec->zone = zone;
1371 * Following LRU functions are allowed to be used without PCG_LOCK.
1372 * Operations are called by routine of global LRU independently from memcg.
1373 * What we have to take care of here is validness of pc->mem_cgroup.
1375 * Changes to pc->mem_cgroup happens when
1378 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1379 * It is added to LRU before charge.
1380 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1381 * When moving account, the page is not on LRU. It's isolated.
1385 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1387 * @zone: zone of the page
1389 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1391 struct mem_cgroup_per_zone *mz;
1392 struct mem_cgroup *memcg;
1393 struct page_cgroup *pc;
1394 struct lruvec *lruvec;
1396 if (mem_cgroup_disabled()) {
1397 lruvec = &zone->lruvec;
1401 pc = lookup_page_cgroup(page);
1402 memcg = pc->mem_cgroup;
1405 * Surreptitiously switch any uncharged offlist page to root:
1406 * an uncharged page off lru does nothing to secure
1407 * its former mem_cgroup from sudden removal.
1409 * Our caller holds lru_lock, and PageCgroupUsed is updated
1410 * under page_cgroup lock: between them, they make all uses
1411 * of pc->mem_cgroup safe.
1413 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1414 pc->mem_cgroup = memcg = root_mem_cgroup;
1416 mz = page_cgroup_zoneinfo(memcg, page);
1417 lruvec = &mz->lruvec;
1420 * Since a node can be onlined after the mem_cgroup was created,
1421 * we have to be prepared to initialize lruvec->zone here;
1422 * and if offlined then reonlined, we need to reinitialize it.
1424 if (unlikely(lruvec->zone != zone))
1425 lruvec->zone = zone;
1430 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1431 * @lruvec: mem_cgroup per zone lru vector
1432 * @lru: index of lru list the page is sitting on
1433 * @nr_pages: positive when adding or negative when removing
1435 * This function must be called when a page is added to or removed from an
1438 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1441 struct mem_cgroup_per_zone *mz;
1442 unsigned long *lru_size;
1444 if (mem_cgroup_disabled())
1447 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1448 lru_size = mz->lru_size + lru;
1449 *lru_size += nr_pages;
1450 VM_BUG_ON((long)(*lru_size) < 0);
1454 * Checks whether given mem is same or in the root_mem_cgroup's
1457 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1458 struct mem_cgroup *memcg)
1460 if (root_memcg == memcg)
1462 if (!root_memcg->use_hierarchy || !memcg)
1464 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1467 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1468 struct mem_cgroup *memcg)
1473 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1478 bool task_in_mem_cgroup(struct task_struct *task,
1479 const struct mem_cgroup *memcg)
1481 struct mem_cgroup *curr = NULL;
1482 struct task_struct *p;
1485 p = find_lock_task_mm(task);
1487 curr = get_mem_cgroup_from_mm(p->mm);
1491 * All threads may have already detached their mm's, but the oom
1492 * killer still needs to detect if they have already been oom
1493 * killed to prevent needlessly killing additional tasks.
1496 curr = mem_cgroup_from_task(task);
1498 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 - try charging a memcg
2662 * @memcg: memcg to charge
2663 * @nr_pages: number of pages to charge
2664 * @oom: trigger OOM if reclaim fails
2666 * Returns 0 if @memcg was charged successfully, -EINTR if the charge
2667 * was bypassed to root_mem_cgroup, and -ENOMEM if the charge failed.
2669 static int mem_cgroup_try_charge(struct mem_cgroup *memcg,
2671 unsigned int nr_pages,
2674 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2675 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2678 if (mem_cgroup_is_root(memcg))
2681 * Unlike in global OOM situations, memcg is not in a physical
2682 * memory shortage. Allow dying and OOM-killed tasks to
2683 * bypass the last charges so that they can exit quickly and
2684 * free their memory.
2686 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2687 fatal_signal_pending(current)))
2690 if (unlikely(task_in_memcg_oom(current)))
2693 if (gfp_mask & __GFP_NOFAIL)
2696 if (consume_stock(memcg, nr_pages))
2700 bool invoke_oom = oom && !nr_oom_retries;
2702 /* If killed, bypass charge */
2703 if (fatal_signal_pending(current))
2706 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2707 nr_pages, invoke_oom);
2711 case CHARGE_RETRY: /* not in OOM situation but retry */
2714 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2716 case CHARGE_NOMEM: /* OOM routine works */
2717 if (!oom || invoke_oom)
2722 } while (ret != CHARGE_OK);
2724 if (batch > nr_pages)
2725 refill_stock(memcg, batch - nr_pages);
2729 if (!(gfp_mask & __GFP_NOFAIL))
2736 * mem_cgroup_try_charge_mm - try charging a mm
2737 * @mm: mm_struct to charge
2738 * @nr_pages: number of pages to charge
2739 * @oom: trigger OOM if reclaim fails
2741 * Returns the charged mem_cgroup associated with the given mm_struct or
2742 * NULL the charge failed.
2744 static struct mem_cgroup *mem_cgroup_try_charge_mm(struct mm_struct *mm,
2746 unsigned int nr_pages,
2750 struct mem_cgroup *memcg;
2753 memcg = get_mem_cgroup_from_mm(mm);
2754 ret = mem_cgroup_try_charge(memcg, gfp_mask, nr_pages, oom);
2755 css_put(&memcg->css);
2757 memcg = root_mem_cgroup;
2765 * Somemtimes we have to undo a charge we got by try_charge().
2766 * This function is for that and do uncharge, put css's refcnt.
2767 * gotten by try_charge().
2769 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2770 unsigned int nr_pages)
2772 if (!mem_cgroup_is_root(memcg)) {
2773 unsigned long bytes = nr_pages * PAGE_SIZE;
2775 res_counter_uncharge(&memcg->res, bytes);
2776 if (do_swap_account)
2777 res_counter_uncharge(&memcg->memsw, bytes);
2782 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2783 * This is useful when moving usage to parent cgroup.
2785 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2786 unsigned int nr_pages)
2788 unsigned long bytes = nr_pages * PAGE_SIZE;
2790 if (mem_cgroup_is_root(memcg))
2793 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2794 if (do_swap_account)
2795 res_counter_uncharge_until(&memcg->memsw,
2796 memcg->memsw.parent, bytes);
2800 * A helper function to get mem_cgroup from ID. must be called under
2801 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2802 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2803 * called against removed memcg.)
2805 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2807 /* ID 0 is unused ID */
2810 return mem_cgroup_from_id(id);
2813 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2815 struct mem_cgroup *memcg = NULL;
2816 struct page_cgroup *pc;
2820 VM_BUG_ON_PAGE(!PageLocked(page), page);
2822 pc = lookup_page_cgroup(page);
2823 lock_page_cgroup(pc);
2824 if (PageCgroupUsed(pc)) {
2825 memcg = pc->mem_cgroup;
2826 if (memcg && !css_tryget(&memcg->css))
2828 } else if (PageSwapCache(page)) {
2829 ent.val = page_private(page);
2830 id = lookup_swap_cgroup_id(ent);
2832 memcg = mem_cgroup_lookup(id);
2833 if (memcg && !css_tryget(&memcg->css))
2837 unlock_page_cgroup(pc);
2841 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2843 unsigned int nr_pages,
2844 enum charge_type ctype,
2847 struct page_cgroup *pc = lookup_page_cgroup(page);
2848 struct zone *uninitialized_var(zone);
2849 struct lruvec *lruvec;
2850 bool was_on_lru = false;
2853 lock_page_cgroup(pc);
2854 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2856 * we don't need page_cgroup_lock about tail pages, becase they are not
2857 * accessed by any other context at this point.
2861 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2862 * may already be on some other mem_cgroup's LRU. Take care of it.
2865 zone = page_zone(page);
2866 spin_lock_irq(&zone->lru_lock);
2867 if (PageLRU(page)) {
2868 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2870 del_page_from_lru_list(page, lruvec, page_lru(page));
2875 pc->mem_cgroup = memcg;
2877 * We access a page_cgroup asynchronously without lock_page_cgroup().
2878 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2879 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2880 * before USED bit, we need memory barrier here.
2881 * See mem_cgroup_add_lru_list(), etc.
2884 SetPageCgroupUsed(pc);
2888 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2889 VM_BUG_ON_PAGE(PageLRU(page), page);
2891 add_page_to_lru_list(page, lruvec, page_lru(page));
2893 spin_unlock_irq(&zone->lru_lock);
2896 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2901 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2902 unlock_page_cgroup(pc);
2905 * "charge_statistics" updated event counter. Then, check it.
2906 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2907 * if they exceeds softlimit.
2909 memcg_check_events(memcg, page);
2912 static DEFINE_MUTEX(set_limit_mutex);
2914 #ifdef CONFIG_MEMCG_KMEM
2915 static DEFINE_MUTEX(activate_kmem_mutex);
2917 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2919 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2920 memcg_kmem_is_active(memcg);
2924 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2925 * in the memcg_cache_params struct.
2927 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2929 struct kmem_cache *cachep;
2931 VM_BUG_ON(p->is_root_cache);
2932 cachep = p->root_cache;
2933 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2936 #ifdef CONFIG_SLABINFO
2937 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2939 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2940 struct memcg_cache_params *params;
2942 if (!memcg_can_account_kmem(memcg))
2945 print_slabinfo_header(m);
2947 mutex_lock(&memcg->slab_caches_mutex);
2948 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2949 cache_show(memcg_params_to_cache(params), m);
2950 mutex_unlock(&memcg->slab_caches_mutex);
2956 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2958 struct res_counter *fail_res;
2961 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2965 ret = mem_cgroup_try_charge(memcg, gfp, size >> PAGE_SHIFT,
2966 oom_gfp_allowed(gfp));
2967 if (ret == -EINTR) {
2969 * mem_cgroup_try_charge() chosed to bypass to root due to
2970 * OOM kill or fatal signal. Since our only options are to
2971 * either fail the allocation or charge it to this cgroup, do
2972 * it as a temporary condition. But we can't fail. From a
2973 * kmem/slab perspective, the cache has already been selected,
2974 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2977 * This condition will only trigger if the task entered
2978 * memcg_charge_kmem in a sane state, but was OOM-killed during
2979 * mem_cgroup_try_charge() above. Tasks that were already
2980 * dying when the allocation triggers should have been already
2981 * directed to the root cgroup in memcontrol.h
2983 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2984 if (do_swap_account)
2985 res_counter_charge_nofail(&memcg->memsw, size,
2989 res_counter_uncharge(&memcg->kmem, size);
2994 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2996 res_counter_uncharge(&memcg->res, size);
2997 if (do_swap_account)
2998 res_counter_uncharge(&memcg->memsw, size);
3001 if (res_counter_uncharge(&memcg->kmem, size))
3005 * Releases a reference taken in kmem_cgroup_css_offline in case
3006 * this last uncharge is racing with the offlining code or it is
3007 * outliving the memcg existence.
3009 * The memory barrier imposed by test&clear is paired with the
3010 * explicit one in memcg_kmem_mark_dead().
3012 if (memcg_kmem_test_and_clear_dead(memcg))
3013 css_put(&memcg->css);
3017 * helper for acessing a memcg's index. It will be used as an index in the
3018 * child cache array in kmem_cache, and also to derive its name. This function
3019 * will return -1 when this is not a kmem-limited memcg.
3021 int memcg_cache_id(struct mem_cgroup *memcg)
3023 return memcg ? memcg->kmemcg_id : -1;
3026 static size_t memcg_caches_array_size(int num_groups)
3029 if (num_groups <= 0)
3032 size = 2 * num_groups;
3033 if (size < MEMCG_CACHES_MIN_SIZE)
3034 size = MEMCG_CACHES_MIN_SIZE;
3035 else if (size > MEMCG_CACHES_MAX_SIZE)
3036 size = MEMCG_CACHES_MAX_SIZE;
3042 * We should update the current array size iff all caches updates succeed. This
3043 * can only be done from the slab side. The slab mutex needs to be held when
3046 void memcg_update_array_size(int num)
3048 if (num > memcg_limited_groups_array_size)
3049 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3052 static void kmem_cache_destroy_work_func(struct work_struct *w);
3054 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3056 struct memcg_cache_params *cur_params = s->memcg_params;
3058 VM_BUG_ON(!is_root_cache(s));
3060 if (num_groups > memcg_limited_groups_array_size) {
3062 struct memcg_cache_params *new_params;
3063 ssize_t size = memcg_caches_array_size(num_groups);
3065 size *= sizeof(void *);
3066 size += offsetof(struct memcg_cache_params, memcg_caches);
3068 new_params = kzalloc(size, GFP_KERNEL);
3072 new_params->is_root_cache = true;
3075 * There is the chance it will be bigger than
3076 * memcg_limited_groups_array_size, if we failed an allocation
3077 * in a cache, in which case all caches updated before it, will
3078 * have a bigger array.
3080 * But if that is the case, the data after
3081 * memcg_limited_groups_array_size is certainly unused
3083 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3084 if (!cur_params->memcg_caches[i])
3086 new_params->memcg_caches[i] =
3087 cur_params->memcg_caches[i];
3091 * Ideally, we would wait until all caches succeed, and only
3092 * then free the old one. But this is not worth the extra
3093 * pointer per-cache we'd have to have for this.
3095 * It is not a big deal if some caches are left with a size
3096 * bigger than the others. And all updates will reset this
3099 rcu_assign_pointer(s->memcg_params, new_params);
3101 kfree_rcu(cur_params, rcu_head);
3106 char *memcg_create_cache_name(struct mem_cgroup *memcg,
3107 struct kmem_cache *root_cache)
3109 static char *buf = NULL;
3112 * We need a mutex here to protect the shared buffer. Since this is
3113 * expected to be called only on cache creation, we can employ the
3114 * slab_mutex for that purpose.
3116 lockdep_assert_held(&slab_mutex);
3119 buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
3124 cgroup_name(memcg->css.cgroup, buf, NAME_MAX + 1);
3125 return kasprintf(GFP_KERNEL, "%s(%d:%s)", root_cache->name,
3126 memcg_cache_id(memcg), buf);
3129 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3130 struct kmem_cache *root_cache)
3134 if (!memcg_kmem_enabled())
3138 size = offsetof(struct memcg_cache_params, memcg_caches);
3139 size += memcg_limited_groups_array_size * sizeof(void *);
3141 size = sizeof(struct memcg_cache_params);
3143 s->memcg_params = kzalloc(size, GFP_KERNEL);
3144 if (!s->memcg_params)
3148 s->memcg_params->memcg = memcg;
3149 s->memcg_params->root_cache = root_cache;
3150 INIT_WORK(&s->memcg_params->destroy,
3151 kmem_cache_destroy_work_func);
3152 css_get(&memcg->css);
3154 s->memcg_params->is_root_cache = true;
3159 void memcg_free_cache_params(struct kmem_cache *s)
3161 if (!s->memcg_params)
3163 if (!s->memcg_params->is_root_cache)
3164 css_put(&s->memcg_params->memcg->css);
3165 kfree(s->memcg_params);
3168 void memcg_register_cache(struct kmem_cache *s)
3170 struct kmem_cache *root;
3171 struct mem_cgroup *memcg;
3174 if (is_root_cache(s))
3178 * Holding the slab_mutex assures nobody will touch the memcg_caches
3179 * array while we are modifying it.
3181 lockdep_assert_held(&slab_mutex);
3183 root = s->memcg_params->root_cache;
3184 memcg = s->memcg_params->memcg;
3185 id = memcg_cache_id(memcg);
3188 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3189 * barrier here to ensure nobody will see the kmem_cache partially
3195 * Initialize the pointer to this cache in its parent's memcg_params
3196 * before adding it to the memcg_slab_caches list, otherwise we can
3197 * fail to convert memcg_params_to_cache() while traversing the list.
3199 VM_BUG_ON(root->memcg_params->memcg_caches[id]);
3200 root->memcg_params->memcg_caches[id] = s;
3202 mutex_lock(&memcg->slab_caches_mutex);
3203 list_add(&s->memcg_params->list, &memcg->memcg_slab_caches);
3204 mutex_unlock(&memcg->slab_caches_mutex);
3207 void memcg_unregister_cache(struct kmem_cache *s)
3209 struct kmem_cache *root;
3210 struct mem_cgroup *memcg;
3213 if (is_root_cache(s))
3217 * Holding the slab_mutex assures nobody will touch the memcg_caches
3218 * array while we are modifying it.
3220 lockdep_assert_held(&slab_mutex);
3222 root = s->memcg_params->root_cache;
3223 memcg = s->memcg_params->memcg;
3224 id = memcg_cache_id(memcg);
3226 mutex_lock(&memcg->slab_caches_mutex);
3227 list_del(&s->memcg_params->list);
3228 mutex_unlock(&memcg->slab_caches_mutex);
3231 * Clear the pointer to this cache in its parent's memcg_params only
3232 * after removing it from the memcg_slab_caches list, otherwise we can
3233 * fail to convert memcg_params_to_cache() while traversing the list.
3235 VM_BUG_ON(root->memcg_params->memcg_caches[id] != s);
3236 root->memcg_params->memcg_caches[id] = NULL;
3240 * During the creation a new cache, we need to disable our accounting mechanism
3241 * altogether. This is true even if we are not creating, but rather just
3242 * enqueing new caches to be created.
3244 * This is because that process will trigger allocations; some visible, like
3245 * explicit kmallocs to auxiliary data structures, name strings and internal
3246 * cache structures; some well concealed, like INIT_WORK() that can allocate
3247 * objects during debug.
3249 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3250 * to it. This may not be a bounded recursion: since the first cache creation
3251 * failed to complete (waiting on the allocation), we'll just try to create the
3252 * cache again, failing at the same point.
3254 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3255 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3256 * inside the following two functions.
3258 static inline void memcg_stop_kmem_account(void)
3260 VM_BUG_ON(!current->mm);
3261 current->memcg_kmem_skip_account++;
3264 static inline void memcg_resume_kmem_account(void)
3266 VM_BUG_ON(!current->mm);
3267 current->memcg_kmem_skip_account--;
3270 static void kmem_cache_destroy_work_func(struct work_struct *w)
3272 struct kmem_cache *cachep;
3273 struct memcg_cache_params *p;
3275 p = container_of(w, struct memcg_cache_params, destroy);
3277 cachep = memcg_params_to_cache(p);
3280 * If we get down to 0 after shrink, we could delete right away.
3281 * However, memcg_release_pages() already puts us back in the workqueue
3282 * in that case. If we proceed deleting, we'll get a dangling
3283 * reference, and removing the object from the workqueue in that case
3284 * is unnecessary complication. We are not a fast path.
3286 * Note that this case is fundamentally different from racing with
3287 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3288 * kmem_cache_shrink, not only we would be reinserting a dead cache
3289 * into the queue, but doing so from inside the worker racing to
3292 * So if we aren't down to zero, we'll just schedule a worker and try
3295 if (atomic_read(&cachep->memcg_params->nr_pages) != 0)
3296 kmem_cache_shrink(cachep);
3298 kmem_cache_destroy(cachep);
3301 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3303 if (!cachep->memcg_params->dead)
3307 * There are many ways in which we can get here.
3309 * We can get to a memory-pressure situation while the delayed work is
3310 * still pending to run. The vmscan shrinkers can then release all
3311 * cache memory and get us to destruction. If this is the case, we'll
3312 * be executed twice, which is a bug (the second time will execute over
3313 * bogus data). In this case, cancelling the work should be fine.
3315 * But we can also get here from the worker itself, if
3316 * kmem_cache_shrink is enough to shake all the remaining objects and
3317 * get the page count to 0. In this case, we'll deadlock if we try to
3318 * cancel the work (the worker runs with an internal lock held, which
3319 * is the same lock we would hold for cancel_work_sync().)
3321 * Since we can't possibly know who got us here, just refrain from
3322 * running if there is already work pending
3324 if (work_pending(&cachep->memcg_params->destroy))
3327 * We have to defer the actual destroying to a workqueue, because
3328 * we might currently be in a context that cannot sleep.
3330 schedule_work(&cachep->memcg_params->destroy);
3333 int __kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3335 struct kmem_cache *c;
3339 * If the cache is being destroyed, we trust that there is no one else
3340 * requesting objects from it. Even if there are, the sanity checks in
3341 * kmem_cache_destroy should caught this ill-case.
3343 * Still, we don't want anyone else freeing memcg_caches under our
3344 * noses, which can happen if a new memcg comes to life. As usual,
3345 * we'll take the activate_kmem_mutex to protect ourselves against
3348 mutex_lock(&activate_kmem_mutex);
3349 for_each_memcg_cache_index(i) {
3350 c = cache_from_memcg_idx(s, i);
3355 * We will now manually delete the caches, so to avoid races
3356 * we need to cancel all pending destruction workers and
3357 * proceed with destruction ourselves.
3359 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3360 * and that could spawn the workers again: it is likely that
3361 * the cache still have active pages until this very moment.
3362 * This would lead us back to mem_cgroup_destroy_cache.
3364 * But that will not execute at all if the "dead" flag is not
3365 * set, so flip it down to guarantee we are in control.
3367 c->memcg_params->dead = false;
3368 cancel_work_sync(&c->memcg_params->destroy);
3369 kmem_cache_destroy(c);
3371 if (cache_from_memcg_idx(s, i))
3374 mutex_unlock(&activate_kmem_mutex);
3378 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3380 struct kmem_cache *cachep;
3381 struct memcg_cache_params *params;
3383 if (!memcg_kmem_is_active(memcg))
3386 mutex_lock(&memcg->slab_caches_mutex);
3387 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3388 cachep = memcg_params_to_cache(params);
3389 cachep->memcg_params->dead = true;
3390 schedule_work(&cachep->memcg_params->destroy);
3392 mutex_unlock(&memcg->slab_caches_mutex);
3395 struct create_work {
3396 struct mem_cgroup *memcg;
3397 struct kmem_cache *cachep;
3398 struct work_struct work;
3401 static void memcg_create_cache_work_func(struct work_struct *w)
3403 struct create_work *cw = container_of(w, struct create_work, work);
3404 struct mem_cgroup *memcg = cw->memcg;
3405 struct kmem_cache *cachep = cw->cachep;
3407 kmem_cache_create_memcg(memcg, cachep);
3408 css_put(&memcg->css);
3413 * Enqueue the creation of a per-memcg kmem_cache.
3415 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3416 struct kmem_cache *cachep)
3418 struct create_work *cw;
3420 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3422 css_put(&memcg->css);
3427 cw->cachep = cachep;
3429 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3430 schedule_work(&cw->work);
3433 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3434 struct kmem_cache *cachep)
3437 * We need to stop accounting when we kmalloc, because if the
3438 * corresponding kmalloc cache is not yet created, the first allocation
3439 * in __memcg_create_cache_enqueue will recurse.
3441 * However, it is better to enclose the whole function. Depending on
3442 * the debugging options enabled, INIT_WORK(), for instance, can
3443 * trigger an allocation. This too, will make us recurse. Because at
3444 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3445 * the safest choice is to do it like this, wrapping the whole function.
3447 memcg_stop_kmem_account();
3448 __memcg_create_cache_enqueue(memcg, cachep);
3449 memcg_resume_kmem_account();
3452 * Return the kmem_cache we're supposed to use for a slab allocation.
3453 * We try to use the current memcg's version of the cache.
3455 * If the cache does not exist yet, if we are the first user of it,
3456 * we either create it immediately, if possible, or create it asynchronously
3458 * In the latter case, we will let the current allocation go through with
3459 * the original cache.
3461 * Can't be called in interrupt context or from kernel threads.
3462 * This function needs to be called with rcu_read_lock() held.
3464 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3467 struct mem_cgroup *memcg;
3468 struct kmem_cache *memcg_cachep;
3470 VM_BUG_ON(!cachep->memcg_params);
3471 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3473 if (!current->mm || current->memcg_kmem_skip_account)
3477 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3479 if (!memcg_can_account_kmem(memcg))
3482 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3483 if (likely(memcg_cachep)) {
3484 cachep = memcg_cachep;
3488 /* The corresponding put will be done in the workqueue. */
3489 if (!css_tryget(&memcg->css))
3494 * If we are in a safe context (can wait, and not in interrupt
3495 * context), we could be be predictable and return right away.
3496 * This would guarantee that the allocation being performed
3497 * already belongs in the new cache.
3499 * However, there are some clashes that can arrive from locking.
3500 * For instance, because we acquire the slab_mutex while doing
3501 * kmem_cache_dup, this means no further allocation could happen
3502 * with the slab_mutex held.
3504 * Also, because cache creation issue get_online_cpus(), this
3505 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3506 * that ends up reversed during cpu hotplug. (cpuset allocates
3507 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3508 * better to defer everything.
3510 memcg_create_cache_enqueue(memcg, cachep);
3516 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3519 * We need to verify if the allocation against current->mm->owner's memcg is
3520 * possible for the given order. But the page is not allocated yet, so we'll
3521 * need a further commit step to do the final arrangements.
3523 * It is possible for the task to switch cgroups in this mean time, so at
3524 * commit time, we can't rely on task conversion any longer. We'll then use
3525 * the handle argument to return to the caller which cgroup we should commit
3526 * against. We could also return the memcg directly and avoid the pointer
3527 * passing, but a boolean return value gives better semantics considering
3528 * the compiled-out case as well.
3530 * Returning true means the allocation is possible.
3533 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3535 struct mem_cgroup *memcg;
3541 * Disabling accounting is only relevant for some specific memcg
3542 * internal allocations. Therefore we would initially not have such
3543 * check here, since direct calls to the page allocator that are marked
3544 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3545 * concerned with cache allocations, and by having this test at
3546 * memcg_kmem_get_cache, we are already able to relay the allocation to
3547 * the root cache and bypass the memcg cache altogether.
3549 * There is one exception, though: the SLUB allocator does not create
3550 * large order caches, but rather service large kmallocs directly from
3551 * the page allocator. Therefore, the following sequence when backed by
3552 * the SLUB allocator:
3554 * memcg_stop_kmem_account();
3555 * kmalloc(<large_number>)
3556 * memcg_resume_kmem_account();
3558 * would effectively ignore the fact that we should skip accounting,
3559 * since it will drive us directly to this function without passing
3560 * through the cache selector memcg_kmem_get_cache. Such large
3561 * allocations are extremely rare but can happen, for instance, for the
3562 * cache arrays. We bring this test here.
3564 if (!current->mm || current->memcg_kmem_skip_account)
3567 memcg = get_mem_cgroup_from_mm(current->mm);
3569 if (!memcg_can_account_kmem(memcg)) {
3570 css_put(&memcg->css);
3574 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3578 css_put(&memcg->css);
3582 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3585 struct page_cgroup *pc;
3587 VM_BUG_ON(mem_cgroup_is_root(memcg));
3589 /* The page allocation failed. Revert */
3591 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3595 pc = lookup_page_cgroup(page);
3596 lock_page_cgroup(pc);
3597 pc->mem_cgroup = memcg;
3598 SetPageCgroupUsed(pc);
3599 unlock_page_cgroup(pc);
3602 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3604 struct mem_cgroup *memcg = NULL;
3605 struct page_cgroup *pc;
3608 pc = lookup_page_cgroup(page);
3610 * Fast unlocked return. Theoretically might have changed, have to
3611 * check again after locking.
3613 if (!PageCgroupUsed(pc))
3616 lock_page_cgroup(pc);
3617 if (PageCgroupUsed(pc)) {
3618 memcg = pc->mem_cgroup;
3619 ClearPageCgroupUsed(pc);
3621 unlock_page_cgroup(pc);
3624 * We trust that only if there is a memcg associated with the page, it
3625 * is a valid allocation
3630 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3631 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3634 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3637 #endif /* CONFIG_MEMCG_KMEM */
3639 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3641 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3643 * Because tail pages are not marked as "used", set it. We're under
3644 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3645 * charge/uncharge will be never happen and move_account() is done under
3646 * compound_lock(), so we don't have to take care of races.
3648 void mem_cgroup_split_huge_fixup(struct page *head)
3650 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3651 struct page_cgroup *pc;
3652 struct mem_cgroup *memcg;
3655 if (mem_cgroup_disabled())
3658 memcg = head_pc->mem_cgroup;
3659 for (i = 1; i < HPAGE_PMD_NR; i++) {
3661 pc->mem_cgroup = memcg;
3662 smp_wmb();/* see __commit_charge() */
3663 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3665 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3668 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3671 * mem_cgroup_move_account - move account of the page
3673 * @nr_pages: number of regular pages (>1 for huge pages)
3674 * @pc: page_cgroup of the page.
3675 * @from: mem_cgroup which the page is moved from.
3676 * @to: mem_cgroup which the page is moved to. @from != @to.
3678 * The caller must confirm following.
3679 * - page is not on LRU (isolate_page() is useful.)
3680 * - compound_lock is held when nr_pages > 1
3682 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3685 static int mem_cgroup_move_account(struct page *page,
3686 unsigned int nr_pages,
3687 struct page_cgroup *pc,
3688 struct mem_cgroup *from,
3689 struct mem_cgroup *to)
3691 unsigned long flags;
3693 bool anon = PageAnon(page);
3695 VM_BUG_ON(from == to);
3696 VM_BUG_ON_PAGE(PageLRU(page), page);
3698 * The page is isolated from LRU. So, collapse function
3699 * will not handle this page. But page splitting can happen.
3700 * Do this check under compound_page_lock(). The caller should
3704 if (nr_pages > 1 && !PageTransHuge(page))
3707 lock_page_cgroup(pc);
3710 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3713 move_lock_mem_cgroup(from, &flags);
3715 if (!anon && page_mapped(page)) {
3716 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3718 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3722 if (PageWriteback(page)) {
3723 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3725 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3729 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3731 /* caller should have done css_get */
3732 pc->mem_cgroup = to;
3733 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3734 move_unlock_mem_cgroup(from, &flags);
3737 unlock_page_cgroup(pc);
3741 memcg_check_events(to, page);
3742 memcg_check_events(from, page);
3748 * mem_cgroup_move_parent - moves page to the parent group
3749 * @page: the page to move
3750 * @pc: page_cgroup of the page
3751 * @child: page's cgroup
3753 * move charges to its parent or the root cgroup if the group has no
3754 * parent (aka use_hierarchy==0).
3755 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3756 * mem_cgroup_move_account fails) the failure is always temporary and
3757 * it signals a race with a page removal/uncharge or migration. In the
3758 * first case the page is on the way out and it will vanish from the LRU
3759 * on the next attempt and the call should be retried later.
3760 * Isolation from the LRU fails only if page has been isolated from
3761 * the LRU since we looked at it and that usually means either global
3762 * reclaim or migration going on. The page will either get back to the
3764 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3765 * (!PageCgroupUsed) or moved to a different group. The page will
3766 * disappear in the next attempt.
3768 static int mem_cgroup_move_parent(struct page *page,
3769 struct page_cgroup *pc,
3770 struct mem_cgroup *child)
3772 struct mem_cgroup *parent;
3773 unsigned int nr_pages;
3774 unsigned long uninitialized_var(flags);
3777 VM_BUG_ON(mem_cgroup_is_root(child));
3780 if (!get_page_unless_zero(page))
3782 if (isolate_lru_page(page))
3785 nr_pages = hpage_nr_pages(page);
3787 parent = parent_mem_cgroup(child);
3789 * If no parent, move charges to root cgroup.
3792 parent = root_mem_cgroup;
3795 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3796 flags = compound_lock_irqsave(page);
3799 ret = mem_cgroup_move_account(page, nr_pages,
3802 __mem_cgroup_cancel_local_charge(child, nr_pages);
3805 compound_unlock_irqrestore(page, flags);
3806 putback_lru_page(page);
3813 int mem_cgroup_charge_anon(struct page *page,
3814 struct mm_struct *mm, gfp_t gfp_mask)
3816 unsigned int nr_pages = 1;
3817 struct mem_cgroup *memcg;
3820 if (mem_cgroup_disabled())
3823 VM_BUG_ON_PAGE(page_mapped(page), page);
3824 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3827 if (PageTransHuge(page)) {
3828 nr_pages <<= compound_order(page);
3829 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3831 * Never OOM-kill a process for a huge page. The
3832 * fault handler will fall back to regular pages.
3837 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, nr_pages, oom);
3840 __mem_cgroup_commit_charge(memcg, page, nr_pages,
3841 MEM_CGROUP_CHARGE_TYPE_ANON, false);
3846 * While swap-in, try_charge -> commit or cancel, the page is locked.
3847 * And when try_charge() successfully returns, one refcnt to memcg without
3848 * struct page_cgroup is acquired. This refcnt will be consumed by
3849 * "commit()" or removed by "cancel()"
3851 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3854 struct mem_cgroup **memcgp)
3856 struct mem_cgroup *memcg = NULL;
3857 struct page_cgroup *pc;
3860 pc = lookup_page_cgroup(page);
3862 * Every swap fault against a single page tries to charge the
3863 * page, bail as early as possible. shmem_unuse() encounters
3864 * already charged pages, too. The USED bit is protected by
3865 * the page lock, which serializes swap cache removal, which
3866 * in turn serializes uncharging.
3868 if (PageCgroupUsed(pc))
3870 if (do_swap_account)
3871 memcg = try_get_mem_cgroup_from_page(page);
3873 memcg = get_mem_cgroup_from_mm(mm);
3874 ret = mem_cgroup_try_charge(memcg, mask, 1, true);
3875 css_put(&memcg->css);
3877 memcg = root_mem_cgroup;
3885 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3886 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3888 if (mem_cgroup_disabled()) {
3893 * A racing thread's fault, or swapoff, may have already
3894 * updated the pte, and even removed page from swap cache: in
3895 * those cases unuse_pte()'s pte_same() test will fail; but
3896 * there's also a KSM case which does need to charge the page.
3898 if (!PageSwapCache(page)) {
3899 struct mem_cgroup *memcg;
3901 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3907 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3910 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3912 if (mem_cgroup_disabled())
3916 __mem_cgroup_cancel_charge(memcg, 1);
3920 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3921 enum charge_type ctype)
3923 if (mem_cgroup_disabled())
3928 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3930 * Now swap is on-memory. This means this page may be
3931 * counted both as mem and swap....double count.
3932 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3933 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3934 * may call delete_from_swap_cache() before reach here.
3936 if (do_swap_account && PageSwapCache(page)) {
3937 swp_entry_t ent = {.val = page_private(page)};
3938 mem_cgroup_uncharge_swap(ent);
3942 void mem_cgroup_commit_charge_swapin(struct page *page,
3943 struct mem_cgroup *memcg)
3945 __mem_cgroup_commit_charge_swapin(page, memcg,
3946 MEM_CGROUP_CHARGE_TYPE_ANON);
3949 int mem_cgroup_charge_file(struct page *page, struct mm_struct *mm,
3952 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3953 struct mem_cgroup *memcg;
3956 if (mem_cgroup_disabled())
3958 if (PageCompound(page))
3961 if (PageSwapCache(page)) { /* shmem */
3962 ret = __mem_cgroup_try_charge_swapin(mm, page,
3966 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3970 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3973 __mem_cgroup_commit_charge(memcg, page, 1, type, false);
3977 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3978 unsigned int nr_pages,
3979 const enum charge_type ctype)
3981 struct memcg_batch_info *batch = NULL;
3982 bool uncharge_memsw = true;
3984 /* If swapout, usage of swap doesn't decrease */
3985 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3986 uncharge_memsw = false;
3988 batch = ¤t->memcg_batch;
3990 * In usual, we do css_get() when we remember memcg pointer.
3991 * But in this case, we keep res->usage until end of a series of
3992 * uncharges. Then, it's ok to ignore memcg's refcnt.
3995 batch->memcg = memcg;
3997 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3998 * In those cases, all pages freed continuously can be expected to be in
3999 * the same cgroup and we have chance to coalesce uncharges.
4000 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4001 * because we want to do uncharge as soon as possible.
4004 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4005 goto direct_uncharge;
4008 goto direct_uncharge;
4011 * In typical case, batch->memcg == mem. This means we can
4012 * merge a series of uncharges to an uncharge of res_counter.
4013 * If not, we uncharge res_counter ony by one.
4015 if (batch->memcg != memcg)
4016 goto direct_uncharge;
4017 /* remember freed charge and uncharge it later */
4020 batch->memsw_nr_pages++;
4023 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4025 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4026 if (unlikely(batch->memcg != memcg))
4027 memcg_oom_recover(memcg);
4031 * uncharge if !page_mapped(page)
4033 static struct mem_cgroup *
4034 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4037 struct mem_cgroup *memcg = NULL;
4038 unsigned int nr_pages = 1;
4039 struct page_cgroup *pc;
4042 if (mem_cgroup_disabled())
4045 if (PageTransHuge(page)) {
4046 nr_pages <<= compound_order(page);
4047 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
4050 * Check if our page_cgroup is valid
4052 pc = lookup_page_cgroup(page);
4053 if (unlikely(!PageCgroupUsed(pc)))
4056 lock_page_cgroup(pc);
4058 memcg = pc->mem_cgroup;
4060 if (!PageCgroupUsed(pc))
4063 anon = PageAnon(page);
4066 case MEM_CGROUP_CHARGE_TYPE_ANON:
4068 * Generally PageAnon tells if it's the anon statistics to be
4069 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4070 * used before page reached the stage of being marked PageAnon.
4074 case MEM_CGROUP_CHARGE_TYPE_DROP:
4075 /* See mem_cgroup_prepare_migration() */
4076 if (page_mapped(page))
4079 * Pages under migration may not be uncharged. But
4080 * end_migration() /must/ be the one uncharging the
4081 * unused post-migration page and so it has to call
4082 * here with the migration bit still set. See the
4083 * res_counter handling below.
4085 if (!end_migration && PageCgroupMigration(pc))
4088 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4089 if (!PageAnon(page)) { /* Shared memory */
4090 if (page->mapping && !page_is_file_cache(page))
4092 } else if (page_mapped(page)) /* Anon */
4099 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4101 ClearPageCgroupUsed(pc);
4103 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4104 * freed from LRU. This is safe because uncharged page is expected not
4105 * to be reused (freed soon). Exception is SwapCache, it's handled by
4106 * special functions.
4109 unlock_page_cgroup(pc);
4111 * even after unlock, we have memcg->res.usage here and this memcg
4112 * will never be freed, so it's safe to call css_get().
4114 memcg_check_events(memcg, page);
4115 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4116 mem_cgroup_swap_statistics(memcg, true);
4117 css_get(&memcg->css);
4120 * Migration does not charge the res_counter for the
4121 * replacement page, so leave it alone when phasing out the
4122 * page that is unused after the migration.
4124 if (!end_migration && !mem_cgroup_is_root(memcg))
4125 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4130 unlock_page_cgroup(pc);
4134 void mem_cgroup_uncharge_page(struct page *page)
4137 if (page_mapped(page))
4139 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4141 * If the page is in swap cache, uncharge should be deferred
4142 * to the swap path, which also properly accounts swap usage
4143 * and handles memcg lifetime.
4145 * Note that this check is not stable and reclaim may add the
4146 * page to swap cache at any time after this. However, if the
4147 * page is not in swap cache by the time page->mapcount hits
4148 * 0, there won't be any page table references to the swap
4149 * slot, and reclaim will free it and not actually write the
4152 if (PageSwapCache(page))
4154 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4157 void mem_cgroup_uncharge_cache_page(struct page *page)
4159 VM_BUG_ON_PAGE(page_mapped(page), page);
4160 VM_BUG_ON_PAGE(page->mapping, page);
4161 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4165 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4166 * In that cases, pages are freed continuously and we can expect pages
4167 * are in the same memcg. All these calls itself limits the number of
4168 * pages freed at once, then uncharge_start/end() is called properly.
4169 * This may be called prural(2) times in a context,
4172 void mem_cgroup_uncharge_start(void)
4174 current->memcg_batch.do_batch++;
4175 /* We can do nest. */
4176 if (current->memcg_batch.do_batch == 1) {
4177 current->memcg_batch.memcg = NULL;
4178 current->memcg_batch.nr_pages = 0;
4179 current->memcg_batch.memsw_nr_pages = 0;
4183 void mem_cgroup_uncharge_end(void)
4185 struct memcg_batch_info *batch = ¤t->memcg_batch;
4187 if (!batch->do_batch)
4191 if (batch->do_batch) /* If stacked, do nothing. */
4197 * This "batch->memcg" is valid without any css_get/put etc...
4198 * bacause we hide charges behind us.
4200 if (batch->nr_pages)
4201 res_counter_uncharge(&batch->memcg->res,
4202 batch->nr_pages * PAGE_SIZE);
4203 if (batch->memsw_nr_pages)
4204 res_counter_uncharge(&batch->memcg->memsw,
4205 batch->memsw_nr_pages * PAGE_SIZE);
4206 memcg_oom_recover(batch->memcg);
4207 /* forget this pointer (for sanity check) */
4208 batch->memcg = NULL;
4213 * called after __delete_from_swap_cache() and drop "page" account.
4214 * memcg information is recorded to swap_cgroup of "ent"
4217 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4219 struct mem_cgroup *memcg;
4220 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4222 if (!swapout) /* this was a swap cache but the swap is unused ! */
4223 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4225 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4228 * record memcg information, if swapout && memcg != NULL,
4229 * css_get() was called in uncharge().
4231 if (do_swap_account && swapout && memcg)
4232 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4236 #ifdef CONFIG_MEMCG_SWAP
4238 * called from swap_entry_free(). remove record in swap_cgroup and
4239 * uncharge "memsw" account.
4241 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4243 struct mem_cgroup *memcg;
4246 if (!do_swap_account)
4249 id = swap_cgroup_record(ent, 0);
4251 memcg = mem_cgroup_lookup(id);
4254 * We uncharge this because swap is freed.
4255 * This memcg can be obsolete one. We avoid calling css_tryget
4257 if (!mem_cgroup_is_root(memcg))
4258 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4259 mem_cgroup_swap_statistics(memcg, false);
4260 css_put(&memcg->css);
4266 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4267 * @entry: swap entry to be moved
4268 * @from: mem_cgroup which the entry is moved from
4269 * @to: mem_cgroup which the entry is moved to
4271 * It succeeds only when the swap_cgroup's record for this entry is the same
4272 * as the mem_cgroup's id of @from.
4274 * Returns 0 on success, -EINVAL on failure.
4276 * The caller must have charged to @to, IOW, called res_counter_charge() about
4277 * both res and memsw, and called css_get().
4279 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4280 struct mem_cgroup *from, struct mem_cgroup *to)
4282 unsigned short old_id, new_id;
4284 old_id = mem_cgroup_id(from);
4285 new_id = mem_cgroup_id(to);
4287 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4288 mem_cgroup_swap_statistics(from, false);
4289 mem_cgroup_swap_statistics(to, true);
4291 * This function is only called from task migration context now.
4292 * It postpones res_counter and refcount handling till the end
4293 * of task migration(mem_cgroup_clear_mc()) for performance
4294 * improvement. But we cannot postpone css_get(to) because if
4295 * the process that has been moved to @to does swap-in, the
4296 * refcount of @to might be decreased to 0.
4298 * We are in attach() phase, so the cgroup is guaranteed to be
4299 * alive, so we can just call css_get().
4307 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4308 struct mem_cgroup *from, struct mem_cgroup *to)
4315 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4318 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4319 struct mem_cgroup **memcgp)
4321 struct mem_cgroup *memcg = NULL;
4322 unsigned int nr_pages = 1;
4323 struct page_cgroup *pc;
4324 enum charge_type ctype;
4328 if (mem_cgroup_disabled())
4331 if (PageTransHuge(page))
4332 nr_pages <<= compound_order(page);
4334 pc = lookup_page_cgroup(page);
4335 lock_page_cgroup(pc);
4336 if (PageCgroupUsed(pc)) {
4337 memcg = pc->mem_cgroup;
4338 css_get(&memcg->css);
4340 * At migrating an anonymous page, its mapcount goes down
4341 * to 0 and uncharge() will be called. But, even if it's fully
4342 * unmapped, migration may fail and this page has to be
4343 * charged again. We set MIGRATION flag here and delay uncharge
4344 * until end_migration() is called
4346 * Corner Case Thinking
4348 * When the old page was mapped as Anon and it's unmap-and-freed
4349 * while migration was ongoing.
4350 * If unmap finds the old page, uncharge() of it will be delayed
4351 * until end_migration(). If unmap finds a new page, it's
4352 * uncharged when it make mapcount to be 1->0. If unmap code
4353 * finds swap_migration_entry, the new page will not be mapped
4354 * and end_migration() will find it(mapcount==0).
4357 * When the old page was mapped but migraion fails, the kernel
4358 * remaps it. A charge for it is kept by MIGRATION flag even
4359 * if mapcount goes down to 0. We can do remap successfully
4360 * without charging it again.
4363 * The "old" page is under lock_page() until the end of
4364 * migration, so, the old page itself will not be swapped-out.
4365 * If the new page is swapped out before end_migraton, our
4366 * hook to usual swap-out path will catch the event.
4369 SetPageCgroupMigration(pc);
4371 unlock_page_cgroup(pc);
4373 * If the page is not charged at this point,
4381 * We charge new page before it's used/mapped. So, even if unlock_page()
4382 * is called before end_migration, we can catch all events on this new
4383 * page. In the case new page is migrated but not remapped, new page's
4384 * mapcount will be finally 0 and we call uncharge in end_migration().
4387 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4389 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4391 * The page is committed to the memcg, but it's not actually
4392 * charged to the res_counter since we plan on replacing the
4393 * old one and only one page is going to be left afterwards.
4395 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4398 /* remove redundant charge if migration failed*/
4399 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4400 struct page *oldpage, struct page *newpage, bool migration_ok)
4402 struct page *used, *unused;
4403 struct page_cgroup *pc;
4409 if (!migration_ok) {
4416 anon = PageAnon(used);
4417 __mem_cgroup_uncharge_common(unused,
4418 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4419 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4421 css_put(&memcg->css);
4423 * We disallowed uncharge of pages under migration because mapcount
4424 * of the page goes down to zero, temporarly.
4425 * Clear the flag and check the page should be charged.
4427 pc = lookup_page_cgroup(oldpage);
4428 lock_page_cgroup(pc);
4429 ClearPageCgroupMigration(pc);
4430 unlock_page_cgroup(pc);
4433 * If a page is a file cache, radix-tree replacement is very atomic
4434 * and we can skip this check. When it was an Anon page, its mapcount
4435 * goes down to 0. But because we added MIGRATION flage, it's not
4436 * uncharged yet. There are several case but page->mapcount check
4437 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4438 * check. (see prepare_charge() also)
4441 mem_cgroup_uncharge_page(used);
4445 * At replace page cache, newpage is not under any memcg but it's on
4446 * LRU. So, this function doesn't touch res_counter but handles LRU
4447 * in correct way. Both pages are locked so we cannot race with uncharge.
4449 void mem_cgroup_replace_page_cache(struct page *oldpage,
4450 struct page *newpage)
4452 struct mem_cgroup *memcg = NULL;
4453 struct page_cgroup *pc;
4454 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4456 if (mem_cgroup_disabled())
4459 pc = lookup_page_cgroup(oldpage);
4460 /* fix accounting on old pages */
4461 lock_page_cgroup(pc);
4462 if (PageCgroupUsed(pc)) {
4463 memcg = pc->mem_cgroup;
4464 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4465 ClearPageCgroupUsed(pc);
4467 unlock_page_cgroup(pc);
4470 * When called from shmem_replace_page(), in some cases the
4471 * oldpage has already been charged, and in some cases not.
4476 * Even if newpage->mapping was NULL before starting replacement,
4477 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4478 * LRU while we overwrite pc->mem_cgroup.
4480 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4483 #ifdef CONFIG_DEBUG_VM
4484 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4486 struct page_cgroup *pc;
4488 pc = lookup_page_cgroup(page);
4490 * Can be NULL while feeding pages into the page allocator for
4491 * the first time, i.e. during boot or memory hotplug;
4492 * or when mem_cgroup_disabled().
4494 if (likely(pc) && PageCgroupUsed(pc))
4499 bool mem_cgroup_bad_page_check(struct page *page)
4501 if (mem_cgroup_disabled())
4504 return lookup_page_cgroup_used(page) != NULL;
4507 void mem_cgroup_print_bad_page(struct page *page)
4509 struct page_cgroup *pc;
4511 pc = lookup_page_cgroup_used(page);
4513 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4514 pc, pc->flags, pc->mem_cgroup);
4519 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4520 unsigned long long val)
4523 u64 memswlimit, memlimit;
4525 int children = mem_cgroup_count_children(memcg);
4526 u64 curusage, oldusage;
4530 * For keeping hierarchical_reclaim simple, how long we should retry
4531 * is depends on callers. We set our retry-count to be function
4532 * of # of children which we should visit in this loop.
4534 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4536 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4539 while (retry_count) {
4540 if (signal_pending(current)) {
4545 * Rather than hide all in some function, I do this in
4546 * open coded manner. You see what this really does.
4547 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4549 mutex_lock(&set_limit_mutex);
4550 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4551 if (memswlimit < val) {
4553 mutex_unlock(&set_limit_mutex);
4557 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4561 ret = res_counter_set_limit(&memcg->res, val);
4563 if (memswlimit == val)
4564 memcg->memsw_is_minimum = true;
4566 memcg->memsw_is_minimum = false;
4568 mutex_unlock(&set_limit_mutex);
4573 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4574 MEM_CGROUP_RECLAIM_SHRINK);
4575 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4576 /* Usage is reduced ? */
4577 if (curusage >= oldusage)
4580 oldusage = curusage;
4582 if (!ret && enlarge)
4583 memcg_oom_recover(memcg);
4588 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4589 unsigned long long val)
4592 u64 memlimit, memswlimit, oldusage, curusage;
4593 int children = mem_cgroup_count_children(memcg);
4597 /* see mem_cgroup_resize_res_limit */
4598 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4599 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4600 while (retry_count) {
4601 if (signal_pending(current)) {
4606 * Rather than hide all in some function, I do this in
4607 * open coded manner. You see what this really does.
4608 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4610 mutex_lock(&set_limit_mutex);
4611 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4612 if (memlimit > val) {
4614 mutex_unlock(&set_limit_mutex);
4617 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4618 if (memswlimit < val)
4620 ret = res_counter_set_limit(&memcg->memsw, val);
4622 if (memlimit == val)
4623 memcg->memsw_is_minimum = true;
4625 memcg->memsw_is_minimum = false;
4627 mutex_unlock(&set_limit_mutex);
4632 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4633 MEM_CGROUP_RECLAIM_NOSWAP |
4634 MEM_CGROUP_RECLAIM_SHRINK);
4635 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4636 /* Usage is reduced ? */
4637 if (curusage >= oldusage)
4640 oldusage = curusage;
4642 if (!ret && enlarge)
4643 memcg_oom_recover(memcg);
4647 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4649 unsigned long *total_scanned)
4651 unsigned long nr_reclaimed = 0;
4652 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4653 unsigned long reclaimed;
4655 struct mem_cgroup_tree_per_zone *mctz;
4656 unsigned long long excess;
4657 unsigned long nr_scanned;
4662 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4664 * This loop can run a while, specially if mem_cgroup's continuously
4665 * keep exceeding their soft limit and putting the system under
4672 mz = mem_cgroup_largest_soft_limit_node(mctz);
4677 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4678 gfp_mask, &nr_scanned);
4679 nr_reclaimed += reclaimed;
4680 *total_scanned += nr_scanned;
4681 spin_lock(&mctz->lock);
4684 * If we failed to reclaim anything from this memory cgroup
4685 * it is time to move on to the next cgroup
4691 * Loop until we find yet another one.
4693 * By the time we get the soft_limit lock
4694 * again, someone might have aded the
4695 * group back on the RB tree. Iterate to
4696 * make sure we get a different mem.
4697 * mem_cgroup_largest_soft_limit_node returns
4698 * NULL if no other cgroup is present on
4702 __mem_cgroup_largest_soft_limit_node(mctz);
4704 css_put(&next_mz->memcg->css);
4705 else /* next_mz == NULL or other memcg */
4709 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4710 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4712 * One school of thought says that we should not add
4713 * back the node to the tree if reclaim returns 0.
4714 * But our reclaim could return 0, simply because due
4715 * to priority we are exposing a smaller subset of
4716 * memory to reclaim from. Consider this as a longer
4719 /* If excess == 0, no tree ops */
4720 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4721 spin_unlock(&mctz->lock);
4722 css_put(&mz->memcg->css);
4725 * Could not reclaim anything and there are no more
4726 * mem cgroups to try or we seem to be looping without
4727 * reclaiming anything.
4729 if (!nr_reclaimed &&
4731 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4733 } while (!nr_reclaimed);
4735 css_put(&next_mz->memcg->css);
4736 return nr_reclaimed;
4740 * mem_cgroup_force_empty_list - clears LRU of a group
4741 * @memcg: group to clear
4744 * @lru: lru to to clear
4746 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4747 * reclaim the pages page themselves - pages are moved to the parent (or root)
4750 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4751 int node, int zid, enum lru_list lru)
4753 struct lruvec *lruvec;
4754 unsigned long flags;
4755 struct list_head *list;
4759 zone = &NODE_DATA(node)->node_zones[zid];
4760 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4761 list = &lruvec->lists[lru];
4765 struct page_cgroup *pc;
4768 spin_lock_irqsave(&zone->lru_lock, flags);
4769 if (list_empty(list)) {
4770 spin_unlock_irqrestore(&zone->lru_lock, flags);
4773 page = list_entry(list->prev, struct page, lru);
4775 list_move(&page->lru, list);
4777 spin_unlock_irqrestore(&zone->lru_lock, flags);
4780 spin_unlock_irqrestore(&zone->lru_lock, flags);
4782 pc = lookup_page_cgroup(page);
4784 if (mem_cgroup_move_parent(page, pc, memcg)) {
4785 /* found lock contention or "pc" is obsolete. */
4790 } while (!list_empty(list));
4794 * make mem_cgroup's charge to be 0 if there is no task by moving
4795 * all the charges and pages to the parent.
4796 * This enables deleting this mem_cgroup.
4798 * Caller is responsible for holding css reference on the memcg.
4800 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4806 /* This is for making all *used* pages to be on LRU. */
4807 lru_add_drain_all();
4808 drain_all_stock_sync(memcg);
4809 mem_cgroup_start_move(memcg);
4810 for_each_node_state(node, N_MEMORY) {
4811 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4814 mem_cgroup_force_empty_list(memcg,
4819 mem_cgroup_end_move(memcg);
4820 memcg_oom_recover(memcg);
4824 * Kernel memory may not necessarily be trackable to a specific
4825 * process. So they are not migrated, and therefore we can't
4826 * expect their value to drop to 0 here.
4827 * Having res filled up with kmem only is enough.
4829 * This is a safety check because mem_cgroup_force_empty_list
4830 * could have raced with mem_cgroup_replace_page_cache callers
4831 * so the lru seemed empty but the page could have been added
4832 * right after the check. RES_USAGE should be safe as we always
4833 * charge before adding to the LRU.
4835 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4836 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4837 } while (usage > 0);
4840 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4842 lockdep_assert_held(&memcg_create_mutex);
4844 * The lock does not prevent addition or deletion to the list
4845 * of children, but it prevents a new child from being
4846 * initialized based on this parent in css_online(), so it's
4847 * enough to decide whether hierarchically inherited
4848 * attributes can still be changed or not.
4850 return memcg->use_hierarchy &&
4851 !list_empty(&memcg->css.cgroup->children);
4855 * Reclaims as many pages from the given memcg as possible and moves
4856 * the rest to the parent.
4858 * Caller is responsible for holding css reference for memcg.
4860 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4862 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4863 struct cgroup *cgrp = memcg->css.cgroup;
4865 /* returns EBUSY if there is a task or if we come here twice. */
4866 if (cgroup_has_tasks(cgrp) || !list_empty(&cgrp->children))
4869 /* we call try-to-free pages for make this cgroup empty */
4870 lru_add_drain_all();
4871 /* try to free all pages in this cgroup */
4872 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4875 if (signal_pending(current))
4878 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4882 /* maybe some writeback is necessary */
4883 congestion_wait(BLK_RW_ASYNC, HZ/10);
4888 mem_cgroup_reparent_charges(memcg);
4893 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4896 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4898 if (mem_cgroup_is_root(memcg))
4900 return mem_cgroup_force_empty(memcg);
4903 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4906 return mem_cgroup_from_css(css)->use_hierarchy;
4909 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4910 struct cftype *cft, u64 val)
4913 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4914 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4916 mutex_lock(&memcg_create_mutex);
4918 if (memcg->use_hierarchy == val)
4922 * If parent's use_hierarchy is set, we can't make any modifications
4923 * in the child subtrees. If it is unset, then the change can
4924 * occur, provided the current cgroup has no children.
4926 * For the root cgroup, parent_mem is NULL, we allow value to be
4927 * set if there are no children.
4929 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4930 (val == 1 || val == 0)) {
4931 if (list_empty(&memcg->css.cgroup->children))
4932 memcg->use_hierarchy = val;
4939 mutex_unlock(&memcg_create_mutex);
4945 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4946 enum mem_cgroup_stat_index idx)
4948 struct mem_cgroup *iter;
4951 /* Per-cpu values can be negative, use a signed accumulator */
4952 for_each_mem_cgroup_tree(iter, memcg)
4953 val += mem_cgroup_read_stat(iter, idx);
4955 if (val < 0) /* race ? */
4960 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4964 if (!mem_cgroup_is_root(memcg)) {
4966 return res_counter_read_u64(&memcg->res, RES_USAGE);
4968 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4972 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4973 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4975 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4976 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4979 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4981 return val << PAGE_SHIFT;
4984 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
4987 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4992 type = MEMFILE_TYPE(cft->private);
4993 name = MEMFILE_ATTR(cft->private);
4997 if (name == RES_USAGE)
4998 val = mem_cgroup_usage(memcg, false);
5000 val = res_counter_read_u64(&memcg->res, name);
5003 if (name == RES_USAGE)
5004 val = mem_cgroup_usage(memcg, true);
5006 val = res_counter_read_u64(&memcg->memsw, name);
5009 val = res_counter_read_u64(&memcg->kmem, name);
5018 #ifdef CONFIG_MEMCG_KMEM
5019 /* should be called with activate_kmem_mutex held */
5020 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
5021 unsigned long long limit)
5026 if (memcg_kmem_is_active(memcg))
5030 * We are going to allocate memory for data shared by all memory
5031 * cgroups so let's stop accounting here.
5033 memcg_stop_kmem_account();
5036 * For simplicity, we won't allow this to be disabled. It also can't
5037 * be changed if the cgroup has children already, or if tasks had
5040 * If tasks join before we set the limit, a person looking at
5041 * kmem.usage_in_bytes will have no way to determine when it took
5042 * place, which makes the value quite meaningless.
5044 * After it first became limited, changes in the value of the limit are
5045 * of course permitted.
5047 mutex_lock(&memcg_create_mutex);
5048 if (cgroup_has_tasks(memcg->css.cgroup) || memcg_has_children(memcg))
5050 mutex_unlock(&memcg_create_mutex);
5054 memcg_id = ida_simple_get(&kmem_limited_groups,
5055 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
5062 * Make sure we have enough space for this cgroup in each root cache's
5065 err = memcg_update_all_caches(memcg_id + 1);
5069 memcg->kmemcg_id = memcg_id;
5070 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
5071 mutex_init(&memcg->slab_caches_mutex);
5074 * We couldn't have accounted to this cgroup, because it hasn't got the
5075 * active bit set yet, so this should succeed.
5077 err = res_counter_set_limit(&memcg->kmem, limit);
5080 static_key_slow_inc(&memcg_kmem_enabled_key);
5082 * Setting the active bit after enabling static branching will
5083 * guarantee no one starts accounting before all call sites are
5086 memcg_kmem_set_active(memcg);
5088 memcg_resume_kmem_account();
5092 ida_simple_remove(&kmem_limited_groups, memcg_id);
5096 static int memcg_activate_kmem(struct mem_cgroup *memcg,
5097 unsigned long long limit)
5101 mutex_lock(&activate_kmem_mutex);
5102 ret = __memcg_activate_kmem(memcg, limit);
5103 mutex_unlock(&activate_kmem_mutex);
5107 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5108 unsigned long long val)
5112 if (!memcg_kmem_is_active(memcg))
5113 ret = memcg_activate_kmem(memcg, val);
5115 ret = res_counter_set_limit(&memcg->kmem, val);
5119 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5122 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5127 mutex_lock(&activate_kmem_mutex);
5129 * If the parent cgroup is not kmem-active now, it cannot be activated
5130 * after this point, because it has at least one child already.
5132 if (memcg_kmem_is_active(parent))
5133 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
5134 mutex_unlock(&activate_kmem_mutex);
5138 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5139 unsigned long long val)
5143 #endif /* CONFIG_MEMCG_KMEM */
5146 * The user of this function is...
5149 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5152 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5155 unsigned long long val;
5158 type = MEMFILE_TYPE(cft->private);
5159 name = MEMFILE_ATTR(cft->private);
5163 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5167 /* This function does all necessary parse...reuse it */
5168 ret = res_counter_memparse_write_strategy(buffer, &val);
5172 ret = mem_cgroup_resize_limit(memcg, val);
5173 else if (type == _MEMSWAP)
5174 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5175 else if (type == _KMEM)
5176 ret = memcg_update_kmem_limit(memcg, val);
5180 case RES_SOFT_LIMIT:
5181 ret = res_counter_memparse_write_strategy(buffer, &val);
5185 * For memsw, soft limits are hard to implement in terms
5186 * of semantics, for now, we support soft limits for
5187 * control without swap
5190 ret = res_counter_set_soft_limit(&memcg->res, val);
5195 ret = -EINVAL; /* should be BUG() ? */
5201 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5202 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5204 unsigned long long min_limit, min_memsw_limit, tmp;
5206 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5207 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5208 if (!memcg->use_hierarchy)
5211 while (css_parent(&memcg->css)) {
5212 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5213 if (!memcg->use_hierarchy)
5215 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5216 min_limit = min(min_limit, tmp);
5217 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5218 min_memsw_limit = min(min_memsw_limit, tmp);
5221 *mem_limit = min_limit;
5222 *memsw_limit = min_memsw_limit;
5225 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5227 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5231 type = MEMFILE_TYPE(event);
5232 name = MEMFILE_ATTR(event);
5237 res_counter_reset_max(&memcg->res);
5238 else if (type == _MEMSWAP)
5239 res_counter_reset_max(&memcg->memsw);
5240 else if (type == _KMEM)
5241 res_counter_reset_max(&memcg->kmem);
5247 res_counter_reset_failcnt(&memcg->res);
5248 else if (type == _MEMSWAP)
5249 res_counter_reset_failcnt(&memcg->memsw);
5250 else if (type == _KMEM)
5251 res_counter_reset_failcnt(&memcg->kmem);
5260 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5263 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5267 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5268 struct cftype *cft, u64 val)
5270 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5272 if (val >= (1 << NR_MOVE_TYPE))
5276 * No kind of locking is needed in here, because ->can_attach() will
5277 * check this value once in the beginning of the process, and then carry
5278 * on with stale data. This means that changes to this value will only
5279 * affect task migrations starting after the change.
5281 memcg->move_charge_at_immigrate = val;
5285 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5286 struct cftype *cft, u64 val)
5293 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5297 unsigned int lru_mask;
5300 static const struct numa_stat stats[] = {
5301 { "total", LRU_ALL },
5302 { "file", LRU_ALL_FILE },
5303 { "anon", LRU_ALL_ANON },
5304 { "unevictable", BIT(LRU_UNEVICTABLE) },
5306 const struct numa_stat *stat;
5309 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5311 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5312 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5313 seq_printf(m, "%s=%lu", stat->name, nr);
5314 for_each_node_state(nid, N_MEMORY) {
5315 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5317 seq_printf(m, " N%d=%lu", nid, nr);
5322 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5323 struct mem_cgroup *iter;
5326 for_each_mem_cgroup_tree(iter, memcg)
5327 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5328 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5329 for_each_node_state(nid, N_MEMORY) {
5331 for_each_mem_cgroup_tree(iter, memcg)
5332 nr += mem_cgroup_node_nr_lru_pages(
5333 iter, nid, stat->lru_mask);
5334 seq_printf(m, " N%d=%lu", nid, nr);
5341 #endif /* CONFIG_NUMA */
5343 static inline void mem_cgroup_lru_names_not_uptodate(void)
5345 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5348 static int memcg_stat_show(struct seq_file *m, void *v)
5350 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5351 struct mem_cgroup *mi;
5354 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5355 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5357 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5358 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5361 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5362 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5363 mem_cgroup_read_events(memcg, i));
5365 for (i = 0; i < NR_LRU_LISTS; i++)
5366 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5367 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5369 /* Hierarchical information */
5371 unsigned long long limit, memsw_limit;
5372 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5373 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5374 if (do_swap_account)
5375 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5379 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5382 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5384 for_each_mem_cgroup_tree(mi, memcg)
5385 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5386 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5389 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5390 unsigned long long val = 0;
5392 for_each_mem_cgroup_tree(mi, memcg)
5393 val += mem_cgroup_read_events(mi, i);
5394 seq_printf(m, "total_%s %llu\n",
5395 mem_cgroup_events_names[i], val);
5398 for (i = 0; i < NR_LRU_LISTS; i++) {
5399 unsigned long long val = 0;
5401 for_each_mem_cgroup_tree(mi, memcg)
5402 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5403 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5406 #ifdef CONFIG_DEBUG_VM
5409 struct mem_cgroup_per_zone *mz;
5410 struct zone_reclaim_stat *rstat;
5411 unsigned long recent_rotated[2] = {0, 0};
5412 unsigned long recent_scanned[2] = {0, 0};
5414 for_each_online_node(nid)
5415 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5416 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5417 rstat = &mz->lruvec.reclaim_stat;
5419 recent_rotated[0] += rstat->recent_rotated[0];
5420 recent_rotated[1] += rstat->recent_rotated[1];
5421 recent_scanned[0] += rstat->recent_scanned[0];
5422 recent_scanned[1] += rstat->recent_scanned[1];
5424 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5425 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5426 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5427 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5434 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5437 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5439 return mem_cgroup_swappiness(memcg);
5442 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5443 struct cftype *cft, u64 val)
5445 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5446 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5448 if (val > 100 || !parent)
5451 mutex_lock(&memcg_create_mutex);
5453 /* If under hierarchy, only empty-root can set this value */
5454 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5455 mutex_unlock(&memcg_create_mutex);
5459 memcg->swappiness = val;
5461 mutex_unlock(&memcg_create_mutex);
5466 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5468 struct mem_cgroup_threshold_ary *t;
5474 t = rcu_dereference(memcg->thresholds.primary);
5476 t = rcu_dereference(memcg->memsw_thresholds.primary);
5481 usage = mem_cgroup_usage(memcg, swap);
5484 * current_threshold points to threshold just below or equal to usage.
5485 * If it's not true, a threshold was crossed after last
5486 * call of __mem_cgroup_threshold().
5488 i = t->current_threshold;
5491 * Iterate backward over array of thresholds starting from
5492 * current_threshold and check if a threshold is crossed.
5493 * If none of thresholds below usage is crossed, we read
5494 * only one element of the array here.
5496 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5497 eventfd_signal(t->entries[i].eventfd, 1);
5499 /* i = current_threshold + 1 */
5503 * Iterate forward over array of thresholds starting from
5504 * current_threshold+1 and check if a threshold is crossed.
5505 * If none of thresholds above usage is crossed, we read
5506 * only one element of the array here.
5508 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5509 eventfd_signal(t->entries[i].eventfd, 1);
5511 /* Update current_threshold */
5512 t->current_threshold = i - 1;
5517 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5520 __mem_cgroup_threshold(memcg, false);
5521 if (do_swap_account)
5522 __mem_cgroup_threshold(memcg, true);
5524 memcg = parent_mem_cgroup(memcg);
5528 static int compare_thresholds(const void *a, const void *b)
5530 const struct mem_cgroup_threshold *_a = a;
5531 const struct mem_cgroup_threshold *_b = b;
5533 if (_a->threshold > _b->threshold)
5536 if (_a->threshold < _b->threshold)
5542 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5544 struct mem_cgroup_eventfd_list *ev;
5546 list_for_each_entry(ev, &memcg->oom_notify, list)
5547 eventfd_signal(ev->eventfd, 1);
5551 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5553 struct mem_cgroup *iter;
5555 for_each_mem_cgroup_tree(iter, memcg)
5556 mem_cgroup_oom_notify_cb(iter);
5559 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5560 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5562 struct mem_cgroup_thresholds *thresholds;
5563 struct mem_cgroup_threshold_ary *new;
5564 u64 threshold, usage;
5567 ret = res_counter_memparse_write_strategy(args, &threshold);
5571 mutex_lock(&memcg->thresholds_lock);
5574 thresholds = &memcg->thresholds;
5575 else if (type == _MEMSWAP)
5576 thresholds = &memcg->memsw_thresholds;
5580 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5582 /* Check if a threshold crossed before adding a new one */
5583 if (thresholds->primary)
5584 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5586 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5588 /* Allocate memory for new array of thresholds */
5589 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5597 /* Copy thresholds (if any) to new array */
5598 if (thresholds->primary) {
5599 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5600 sizeof(struct mem_cgroup_threshold));
5603 /* Add new threshold */
5604 new->entries[size - 1].eventfd = eventfd;
5605 new->entries[size - 1].threshold = threshold;
5607 /* Sort thresholds. Registering of new threshold isn't time-critical */
5608 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5609 compare_thresholds, NULL);
5611 /* Find current threshold */
5612 new->current_threshold = -1;
5613 for (i = 0; i < size; i++) {
5614 if (new->entries[i].threshold <= usage) {
5616 * new->current_threshold will not be used until
5617 * rcu_assign_pointer(), so it's safe to increment
5620 ++new->current_threshold;
5625 /* Free old spare buffer and save old primary buffer as spare */
5626 kfree(thresholds->spare);
5627 thresholds->spare = thresholds->primary;
5629 rcu_assign_pointer(thresholds->primary, new);
5631 /* To be sure that nobody uses thresholds */
5635 mutex_unlock(&memcg->thresholds_lock);
5640 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5641 struct eventfd_ctx *eventfd, const char *args)
5643 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5646 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5647 struct eventfd_ctx *eventfd, const char *args)
5649 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5652 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5653 struct eventfd_ctx *eventfd, enum res_type type)
5655 struct mem_cgroup_thresholds *thresholds;
5656 struct mem_cgroup_threshold_ary *new;
5660 mutex_lock(&memcg->thresholds_lock);
5662 thresholds = &memcg->thresholds;
5663 else if (type == _MEMSWAP)
5664 thresholds = &memcg->memsw_thresholds;
5668 if (!thresholds->primary)
5671 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5673 /* Check if a threshold crossed before removing */
5674 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5676 /* Calculate new number of threshold */
5678 for (i = 0; i < thresholds->primary->size; i++) {
5679 if (thresholds->primary->entries[i].eventfd != eventfd)
5683 new = thresholds->spare;
5685 /* Set thresholds array to NULL if we don't have thresholds */
5694 /* Copy thresholds and find current threshold */
5695 new->current_threshold = -1;
5696 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5697 if (thresholds->primary->entries[i].eventfd == eventfd)
5700 new->entries[j] = thresholds->primary->entries[i];
5701 if (new->entries[j].threshold <= usage) {
5703 * new->current_threshold will not be used
5704 * until rcu_assign_pointer(), so it's safe to increment
5707 ++new->current_threshold;
5713 /* Swap primary and spare array */
5714 thresholds->spare = thresholds->primary;
5715 /* If all events are unregistered, free the spare array */
5717 kfree(thresholds->spare);
5718 thresholds->spare = NULL;
5721 rcu_assign_pointer(thresholds->primary, new);
5723 /* To be sure that nobody uses thresholds */
5726 mutex_unlock(&memcg->thresholds_lock);
5729 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5730 struct eventfd_ctx *eventfd)
5732 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5735 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5736 struct eventfd_ctx *eventfd)
5738 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5741 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5742 struct eventfd_ctx *eventfd, const char *args)
5744 struct mem_cgroup_eventfd_list *event;
5746 event = kmalloc(sizeof(*event), GFP_KERNEL);
5750 spin_lock(&memcg_oom_lock);
5752 event->eventfd = eventfd;
5753 list_add(&event->list, &memcg->oom_notify);
5755 /* already in OOM ? */
5756 if (atomic_read(&memcg->under_oom))
5757 eventfd_signal(eventfd, 1);
5758 spin_unlock(&memcg_oom_lock);
5763 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5764 struct eventfd_ctx *eventfd)
5766 struct mem_cgroup_eventfd_list *ev, *tmp;
5768 spin_lock(&memcg_oom_lock);
5770 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5771 if (ev->eventfd == eventfd) {
5772 list_del(&ev->list);
5777 spin_unlock(&memcg_oom_lock);
5780 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5782 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5784 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5785 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5789 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5790 struct cftype *cft, u64 val)
5792 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5793 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5795 /* cannot set to root cgroup and only 0 and 1 are allowed */
5796 if (!parent || !((val == 0) || (val == 1)))
5799 mutex_lock(&memcg_create_mutex);
5800 /* oom-kill-disable is a flag for subhierarchy. */
5801 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5802 mutex_unlock(&memcg_create_mutex);
5805 memcg->oom_kill_disable = val;
5807 memcg_oom_recover(memcg);
5808 mutex_unlock(&memcg_create_mutex);
5812 #ifdef CONFIG_MEMCG_KMEM
5813 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5817 memcg->kmemcg_id = -1;
5818 ret = memcg_propagate_kmem(memcg);
5822 return mem_cgroup_sockets_init(memcg, ss);
5825 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5827 mem_cgroup_sockets_destroy(memcg);
5830 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5832 if (!memcg_kmem_is_active(memcg))
5836 * kmem charges can outlive the cgroup. In the case of slab
5837 * pages, for instance, a page contain objects from various
5838 * processes. As we prevent from taking a reference for every
5839 * such allocation we have to be careful when doing uncharge
5840 * (see memcg_uncharge_kmem) and here during offlining.
5842 * The idea is that that only the _last_ uncharge which sees
5843 * the dead memcg will drop the last reference. An additional
5844 * reference is taken here before the group is marked dead
5845 * which is then paired with css_put during uncharge resp. here.
5847 * Although this might sound strange as this path is called from
5848 * css_offline() when the referencemight have dropped down to 0
5849 * and shouldn't be incremented anymore (css_tryget would fail)
5850 * we do not have other options because of the kmem allocations
5853 css_get(&memcg->css);
5855 memcg_kmem_mark_dead(memcg);
5857 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5860 if (memcg_kmem_test_and_clear_dead(memcg))
5861 css_put(&memcg->css);
5864 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5869 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5873 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5879 * DO NOT USE IN NEW FILES.
5881 * "cgroup.event_control" implementation.
5883 * This is way over-engineered. It tries to support fully configurable
5884 * events for each user. Such level of flexibility is completely
5885 * unnecessary especially in the light of the planned unified hierarchy.
5887 * Please deprecate this and replace with something simpler if at all
5892 * Unregister event and free resources.
5894 * Gets called from workqueue.
5896 static void memcg_event_remove(struct work_struct *work)
5898 struct mem_cgroup_event *event =
5899 container_of(work, struct mem_cgroup_event, remove);
5900 struct mem_cgroup *memcg = event->memcg;
5902 remove_wait_queue(event->wqh, &event->wait);
5904 event->unregister_event(memcg, event->eventfd);
5906 /* Notify userspace the event is going away. */
5907 eventfd_signal(event->eventfd, 1);
5909 eventfd_ctx_put(event->eventfd);
5911 css_put(&memcg->css);
5915 * Gets called on POLLHUP on eventfd when user closes it.
5917 * Called with wqh->lock held and interrupts disabled.
5919 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
5920 int sync, void *key)
5922 struct mem_cgroup_event *event =
5923 container_of(wait, struct mem_cgroup_event, wait);
5924 struct mem_cgroup *memcg = event->memcg;
5925 unsigned long flags = (unsigned long)key;
5927 if (flags & POLLHUP) {
5929 * If the event has been detached at cgroup removal, we
5930 * can simply return knowing the other side will cleanup
5933 * We can't race against event freeing since the other
5934 * side will require wqh->lock via remove_wait_queue(),
5937 spin_lock(&memcg->event_list_lock);
5938 if (!list_empty(&event->list)) {
5939 list_del_init(&event->list);
5941 * We are in atomic context, but cgroup_event_remove()
5942 * may sleep, so we have to call it in workqueue.
5944 schedule_work(&event->remove);
5946 spin_unlock(&memcg->event_list_lock);
5952 static void memcg_event_ptable_queue_proc(struct file *file,
5953 wait_queue_head_t *wqh, poll_table *pt)
5955 struct mem_cgroup_event *event =
5956 container_of(pt, struct mem_cgroup_event, pt);
5959 add_wait_queue(wqh, &event->wait);
5963 * DO NOT USE IN NEW FILES.
5965 * Parse input and register new cgroup event handler.
5967 * Input must be in format '<event_fd> <control_fd> <args>'.
5968 * Interpretation of args is defined by control file implementation.
5970 static int memcg_write_event_control(struct cgroup_subsys_state *css,
5971 struct cftype *cft, char *buffer)
5973 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5974 struct mem_cgroup_event *event;
5975 struct cgroup_subsys_state *cfile_css;
5976 unsigned int efd, cfd;
5983 efd = simple_strtoul(buffer, &endp, 10);
5988 cfd = simple_strtoul(buffer, &endp, 10);
5989 if ((*endp != ' ') && (*endp != '\0'))
5993 event = kzalloc(sizeof(*event), GFP_KERNEL);
5997 event->memcg = memcg;
5998 INIT_LIST_HEAD(&event->list);
5999 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
6000 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
6001 INIT_WORK(&event->remove, memcg_event_remove);
6009 event->eventfd = eventfd_ctx_fileget(efile.file);
6010 if (IS_ERR(event->eventfd)) {
6011 ret = PTR_ERR(event->eventfd);
6018 goto out_put_eventfd;
6021 /* the process need read permission on control file */
6022 /* AV: shouldn't we check that it's been opened for read instead? */
6023 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6028 * Determine the event callbacks and set them in @event. This used
6029 * to be done via struct cftype but cgroup core no longer knows
6030 * about these events. The following is crude but the whole thing
6031 * is for compatibility anyway.
6033 * DO NOT ADD NEW FILES.
6035 name = cfile.file->f_dentry->d_name.name;
6037 if (!strcmp(name, "memory.usage_in_bytes")) {
6038 event->register_event = mem_cgroup_usage_register_event;
6039 event->unregister_event = mem_cgroup_usage_unregister_event;
6040 } else if (!strcmp(name, "memory.oom_control")) {
6041 event->register_event = mem_cgroup_oom_register_event;
6042 event->unregister_event = mem_cgroup_oom_unregister_event;
6043 } else if (!strcmp(name, "memory.pressure_level")) {
6044 event->register_event = vmpressure_register_event;
6045 event->unregister_event = vmpressure_unregister_event;
6046 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
6047 event->register_event = memsw_cgroup_usage_register_event;
6048 event->unregister_event = memsw_cgroup_usage_unregister_event;
6055 * Verify @cfile should belong to @css. Also, remaining events are
6056 * automatically removed on cgroup destruction but the removal is
6057 * asynchronous, so take an extra ref on @css.
6059 cfile_css = css_tryget_from_dir(cfile.file->f_dentry->d_parent,
6060 &memory_cgrp_subsys);
6062 if (IS_ERR(cfile_css))
6064 if (cfile_css != css) {
6069 ret = event->register_event(memcg, event->eventfd, buffer);
6073 efile.file->f_op->poll(efile.file, &event->pt);
6075 spin_lock(&memcg->event_list_lock);
6076 list_add(&event->list, &memcg->event_list);
6077 spin_unlock(&memcg->event_list_lock);
6089 eventfd_ctx_put(event->eventfd);
6098 static struct cftype mem_cgroup_files[] = {
6100 .name = "usage_in_bytes",
6101 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6102 .read_u64 = mem_cgroup_read_u64,
6105 .name = "max_usage_in_bytes",
6106 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6107 .trigger = mem_cgroup_reset,
6108 .read_u64 = mem_cgroup_read_u64,
6111 .name = "limit_in_bytes",
6112 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6113 .write_string = mem_cgroup_write,
6114 .read_u64 = mem_cgroup_read_u64,
6117 .name = "soft_limit_in_bytes",
6118 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6119 .write_string = mem_cgroup_write,
6120 .read_u64 = mem_cgroup_read_u64,
6124 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6125 .trigger = mem_cgroup_reset,
6126 .read_u64 = mem_cgroup_read_u64,
6130 .seq_show = memcg_stat_show,
6133 .name = "force_empty",
6134 .trigger = mem_cgroup_force_empty_write,
6137 .name = "use_hierarchy",
6138 .flags = CFTYPE_INSANE,
6139 .write_u64 = mem_cgroup_hierarchy_write,
6140 .read_u64 = mem_cgroup_hierarchy_read,
6143 .name = "cgroup.event_control", /* XXX: for compat */
6144 .write_string = memcg_write_event_control,
6145 .flags = CFTYPE_NO_PREFIX,
6149 .name = "swappiness",
6150 .read_u64 = mem_cgroup_swappiness_read,
6151 .write_u64 = mem_cgroup_swappiness_write,
6154 .name = "move_charge_at_immigrate",
6155 .read_u64 = mem_cgroup_move_charge_read,
6156 .write_u64 = mem_cgroup_move_charge_write,
6159 .name = "oom_control",
6160 .seq_show = mem_cgroup_oom_control_read,
6161 .write_u64 = mem_cgroup_oom_control_write,
6162 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6165 .name = "pressure_level",
6169 .name = "numa_stat",
6170 .seq_show = memcg_numa_stat_show,
6173 #ifdef CONFIG_MEMCG_KMEM
6175 .name = "kmem.limit_in_bytes",
6176 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6177 .write_string = mem_cgroup_write,
6178 .read_u64 = mem_cgroup_read_u64,
6181 .name = "kmem.usage_in_bytes",
6182 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6183 .read_u64 = mem_cgroup_read_u64,
6186 .name = "kmem.failcnt",
6187 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6188 .trigger = mem_cgroup_reset,
6189 .read_u64 = mem_cgroup_read_u64,
6192 .name = "kmem.max_usage_in_bytes",
6193 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6194 .trigger = mem_cgroup_reset,
6195 .read_u64 = mem_cgroup_read_u64,
6197 #ifdef CONFIG_SLABINFO
6199 .name = "kmem.slabinfo",
6200 .seq_show = mem_cgroup_slabinfo_read,
6204 { }, /* terminate */
6207 #ifdef CONFIG_MEMCG_SWAP
6208 static struct cftype memsw_cgroup_files[] = {
6210 .name = "memsw.usage_in_bytes",
6211 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6212 .read_u64 = mem_cgroup_read_u64,
6215 .name = "memsw.max_usage_in_bytes",
6216 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6217 .trigger = mem_cgroup_reset,
6218 .read_u64 = mem_cgroup_read_u64,
6221 .name = "memsw.limit_in_bytes",
6222 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6223 .write_string = mem_cgroup_write,
6224 .read_u64 = mem_cgroup_read_u64,
6227 .name = "memsw.failcnt",
6228 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6229 .trigger = mem_cgroup_reset,
6230 .read_u64 = mem_cgroup_read_u64,
6232 { }, /* terminate */
6235 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6237 struct mem_cgroup_per_node *pn;
6238 struct mem_cgroup_per_zone *mz;
6239 int zone, tmp = node;
6241 * This routine is called against possible nodes.
6242 * But it's BUG to call kmalloc() against offline node.
6244 * TODO: this routine can waste much memory for nodes which will
6245 * never be onlined. It's better to use memory hotplug callback
6248 if (!node_state(node, N_NORMAL_MEMORY))
6250 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6254 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6255 mz = &pn->zoneinfo[zone];
6256 lruvec_init(&mz->lruvec);
6257 mz->usage_in_excess = 0;
6258 mz->on_tree = false;
6261 memcg->nodeinfo[node] = pn;
6265 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6267 kfree(memcg->nodeinfo[node]);
6270 static struct mem_cgroup *mem_cgroup_alloc(void)
6272 struct mem_cgroup *memcg;
6275 size = sizeof(struct mem_cgroup);
6276 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6278 memcg = kzalloc(size, GFP_KERNEL);
6282 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6285 spin_lock_init(&memcg->pcp_counter_lock);
6294 * At destroying mem_cgroup, references from swap_cgroup can remain.
6295 * (scanning all at force_empty is too costly...)
6297 * Instead of clearing all references at force_empty, we remember
6298 * the number of reference from swap_cgroup and free mem_cgroup when
6299 * it goes down to 0.
6301 * Removal of cgroup itself succeeds regardless of refs from swap.
6304 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6308 mem_cgroup_remove_from_trees(memcg);
6311 free_mem_cgroup_per_zone_info(memcg, node);
6313 free_percpu(memcg->stat);
6316 * We need to make sure that (at least for now), the jump label
6317 * destruction code runs outside of the cgroup lock. This is because
6318 * get_online_cpus(), which is called from the static_branch update,
6319 * can't be called inside the cgroup_lock. cpusets are the ones
6320 * enforcing this dependency, so if they ever change, we might as well.
6322 * schedule_work() will guarantee this happens. Be careful if you need
6323 * to move this code around, and make sure it is outside
6326 disarm_static_keys(memcg);
6331 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6333 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6335 if (!memcg->res.parent)
6337 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6339 EXPORT_SYMBOL(parent_mem_cgroup);
6341 static void __init mem_cgroup_soft_limit_tree_init(void)
6343 struct mem_cgroup_tree_per_node *rtpn;
6344 struct mem_cgroup_tree_per_zone *rtpz;
6345 int tmp, node, zone;
6347 for_each_node(node) {
6349 if (!node_state(node, N_NORMAL_MEMORY))
6351 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6354 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6356 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6357 rtpz = &rtpn->rb_tree_per_zone[zone];
6358 rtpz->rb_root = RB_ROOT;
6359 spin_lock_init(&rtpz->lock);
6364 static struct cgroup_subsys_state * __ref
6365 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6367 struct mem_cgroup *memcg;
6368 long error = -ENOMEM;
6371 memcg = mem_cgroup_alloc();
6373 return ERR_PTR(error);
6376 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6380 if (parent_css == NULL) {
6381 root_mem_cgroup = memcg;
6382 res_counter_init(&memcg->res, NULL);
6383 res_counter_init(&memcg->memsw, NULL);
6384 res_counter_init(&memcg->kmem, NULL);
6387 memcg->last_scanned_node = MAX_NUMNODES;
6388 INIT_LIST_HEAD(&memcg->oom_notify);
6389 memcg->move_charge_at_immigrate = 0;
6390 mutex_init(&memcg->thresholds_lock);
6391 spin_lock_init(&memcg->move_lock);
6392 vmpressure_init(&memcg->vmpressure);
6393 INIT_LIST_HEAD(&memcg->event_list);
6394 spin_lock_init(&memcg->event_list_lock);
6399 __mem_cgroup_free(memcg);
6400 return ERR_PTR(error);
6404 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6406 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6407 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6409 if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6415 mutex_lock(&memcg_create_mutex);
6417 memcg->use_hierarchy = parent->use_hierarchy;
6418 memcg->oom_kill_disable = parent->oom_kill_disable;
6419 memcg->swappiness = mem_cgroup_swappiness(parent);
6421 if (parent->use_hierarchy) {
6422 res_counter_init(&memcg->res, &parent->res);
6423 res_counter_init(&memcg->memsw, &parent->memsw);
6424 res_counter_init(&memcg->kmem, &parent->kmem);
6427 * No need to take a reference to the parent because cgroup
6428 * core guarantees its existence.
6431 res_counter_init(&memcg->res, NULL);
6432 res_counter_init(&memcg->memsw, NULL);
6433 res_counter_init(&memcg->kmem, NULL);
6435 * Deeper hierachy with use_hierarchy == false doesn't make
6436 * much sense so let cgroup subsystem know about this
6437 * unfortunate state in our controller.
6439 if (parent != root_mem_cgroup)
6440 memory_cgrp_subsys.broken_hierarchy = true;
6442 mutex_unlock(&memcg_create_mutex);
6444 return memcg_init_kmem(memcg, &memory_cgrp_subsys);
6448 * Announce all parents that a group from their hierarchy is gone.
6450 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6452 struct mem_cgroup *parent = memcg;
6454 while ((parent = parent_mem_cgroup(parent)))
6455 mem_cgroup_iter_invalidate(parent);
6458 * if the root memcg is not hierarchical we have to check it
6461 if (!root_mem_cgroup->use_hierarchy)
6462 mem_cgroup_iter_invalidate(root_mem_cgroup);
6465 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6467 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6468 struct mem_cgroup_event *event, *tmp;
6469 struct cgroup_subsys_state *iter;
6472 * Unregister events and notify userspace.
6473 * Notify userspace about cgroup removing only after rmdir of cgroup
6474 * directory to avoid race between userspace and kernelspace.
6476 spin_lock(&memcg->event_list_lock);
6477 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6478 list_del_init(&event->list);
6479 schedule_work(&event->remove);
6481 spin_unlock(&memcg->event_list_lock);
6483 kmem_cgroup_css_offline(memcg);
6485 mem_cgroup_invalidate_reclaim_iterators(memcg);
6488 * This requires that offlining is serialized. Right now that is
6489 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6491 css_for_each_descendant_post(iter, css)
6492 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
6494 mem_cgroup_destroy_all_caches(memcg);
6495 vmpressure_cleanup(&memcg->vmpressure);
6498 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6500 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6502 * XXX: css_offline() would be where we should reparent all
6503 * memory to prepare the cgroup for destruction. However,
6504 * memcg does not do css_tryget() and res_counter charging
6505 * under the same RCU lock region, which means that charging
6506 * could race with offlining. Offlining only happens to
6507 * cgroups with no tasks in them but charges can show up
6508 * without any tasks from the swapin path when the target
6509 * memcg is looked up from the swapout record and not from the
6510 * current task as it usually is. A race like this can leak
6511 * charges and put pages with stale cgroup pointers into
6515 * lookup_swap_cgroup_id()
6517 * mem_cgroup_lookup()
6520 * disable css_tryget()
6523 * reparent_charges()
6524 * res_counter_charge()
6527 * pc->mem_cgroup = dead memcg
6530 * The bulk of the charges are still moved in offline_css() to
6531 * avoid pinning a lot of pages in case a long-term reference
6532 * like a swapout record is deferring the css_free() to long
6533 * after offlining. But this makes sure we catch any charges
6534 * made after offlining:
6536 mem_cgroup_reparent_charges(memcg);
6538 memcg_destroy_kmem(memcg);
6539 __mem_cgroup_free(memcg);
6543 /* Handlers for move charge at task migration. */
6544 #define PRECHARGE_COUNT_AT_ONCE 256
6545 static int mem_cgroup_do_precharge(unsigned long count)
6548 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6549 struct mem_cgroup *memcg = mc.to;
6551 if (mem_cgroup_is_root(memcg)) {
6552 mc.precharge += count;
6553 /* we don't need css_get for root */
6556 /* try to charge at once */
6558 struct res_counter *dummy;
6560 * "memcg" cannot be under rmdir() because we've already checked
6561 * by cgroup_lock_live_cgroup() that it is not removed and we
6562 * are still under the same cgroup_mutex. So we can postpone
6565 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6567 if (do_swap_account && res_counter_charge(&memcg->memsw,
6568 PAGE_SIZE * count, &dummy)) {
6569 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6572 mc.precharge += count;
6576 /* fall back to one by one charge */
6578 if (signal_pending(current)) {
6582 if (!batch_count--) {
6583 batch_count = PRECHARGE_COUNT_AT_ONCE;
6586 ret = mem_cgroup_try_charge(memcg, GFP_KERNEL, 1, false);
6588 /* mem_cgroup_clear_mc() will do uncharge later */
6596 * get_mctgt_type - get target type of moving charge
6597 * @vma: the vma the pte to be checked belongs
6598 * @addr: the address corresponding to the pte to be checked
6599 * @ptent: the pte to be checked
6600 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6603 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6604 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6605 * move charge. if @target is not NULL, the page is stored in target->page
6606 * with extra refcnt got(Callers should handle it).
6607 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6608 * target for charge migration. if @target is not NULL, the entry is stored
6611 * Called with pte lock held.
6618 enum mc_target_type {
6624 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6625 unsigned long addr, pte_t ptent)
6627 struct page *page = vm_normal_page(vma, addr, ptent);
6629 if (!page || !page_mapped(page))
6631 if (PageAnon(page)) {
6632 /* we don't move shared anon */
6635 } else if (!move_file())
6636 /* we ignore mapcount for file pages */
6638 if (!get_page_unless_zero(page))
6645 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6646 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6648 struct page *page = NULL;
6649 swp_entry_t ent = pte_to_swp_entry(ptent);
6651 if (!move_anon() || non_swap_entry(ent))
6654 * Because lookup_swap_cache() updates some statistics counter,
6655 * we call find_get_page() with swapper_space directly.
6657 page = find_get_page(swap_address_space(ent), ent.val);
6658 if (do_swap_account)
6659 entry->val = ent.val;
6664 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6665 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6671 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6672 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6674 struct page *page = NULL;
6675 struct address_space *mapping;
6678 if (!vma->vm_file) /* anonymous vma */
6683 mapping = vma->vm_file->f_mapping;
6684 if (pte_none(ptent))
6685 pgoff = linear_page_index(vma, addr);
6686 else /* pte_file(ptent) is true */
6687 pgoff = pte_to_pgoff(ptent);
6689 /* page is moved even if it's not RSS of this task(page-faulted). */
6691 /* shmem/tmpfs may report page out on swap: account for that too. */
6692 if (shmem_mapping(mapping)) {
6693 page = find_get_entry(mapping, pgoff);
6694 if (radix_tree_exceptional_entry(page)) {
6695 swp_entry_t swp = radix_to_swp_entry(page);
6696 if (do_swap_account)
6698 page = find_get_page(swap_address_space(swp), swp.val);
6701 page = find_get_page(mapping, pgoff);
6703 page = find_get_page(mapping, pgoff);
6708 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6709 unsigned long addr, pte_t ptent, union mc_target *target)
6711 struct page *page = NULL;
6712 struct page_cgroup *pc;
6713 enum mc_target_type ret = MC_TARGET_NONE;
6714 swp_entry_t ent = { .val = 0 };
6716 if (pte_present(ptent))
6717 page = mc_handle_present_pte(vma, addr, ptent);
6718 else if (is_swap_pte(ptent))
6719 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6720 else if (pte_none(ptent) || pte_file(ptent))
6721 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6723 if (!page && !ent.val)
6726 pc = lookup_page_cgroup(page);
6728 * Do only loose check w/o page_cgroup lock.
6729 * mem_cgroup_move_account() checks the pc is valid or not under
6732 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6733 ret = MC_TARGET_PAGE;
6735 target->page = page;
6737 if (!ret || !target)
6740 /* There is a swap entry and a page doesn't exist or isn't charged */
6741 if (ent.val && !ret &&
6742 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6743 ret = MC_TARGET_SWAP;
6750 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6752 * We don't consider swapping or file mapped pages because THP does not
6753 * support them for now.
6754 * Caller should make sure that pmd_trans_huge(pmd) is true.
6756 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6757 unsigned long addr, pmd_t pmd, union mc_target *target)
6759 struct page *page = NULL;
6760 struct page_cgroup *pc;
6761 enum mc_target_type ret = MC_TARGET_NONE;
6763 page = pmd_page(pmd);
6764 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6767 pc = lookup_page_cgroup(page);
6768 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6769 ret = MC_TARGET_PAGE;
6772 target->page = page;
6778 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6779 unsigned long addr, pmd_t pmd, union mc_target *target)
6781 return MC_TARGET_NONE;
6785 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6786 unsigned long addr, unsigned long end,
6787 struct mm_walk *walk)
6789 struct vm_area_struct *vma = walk->private;
6793 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6794 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6795 mc.precharge += HPAGE_PMD_NR;
6800 if (pmd_trans_unstable(pmd))
6802 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6803 for (; addr != end; pte++, addr += PAGE_SIZE)
6804 if (get_mctgt_type(vma, addr, *pte, NULL))
6805 mc.precharge++; /* increment precharge temporarily */
6806 pte_unmap_unlock(pte - 1, ptl);
6812 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6814 unsigned long precharge;
6815 struct vm_area_struct *vma;
6817 down_read(&mm->mmap_sem);
6818 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6819 struct mm_walk mem_cgroup_count_precharge_walk = {
6820 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6824 if (is_vm_hugetlb_page(vma))
6826 walk_page_range(vma->vm_start, vma->vm_end,
6827 &mem_cgroup_count_precharge_walk);
6829 up_read(&mm->mmap_sem);
6831 precharge = mc.precharge;
6837 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6839 unsigned long precharge = mem_cgroup_count_precharge(mm);
6841 VM_BUG_ON(mc.moving_task);
6842 mc.moving_task = current;
6843 return mem_cgroup_do_precharge(precharge);
6846 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6847 static void __mem_cgroup_clear_mc(void)
6849 struct mem_cgroup *from = mc.from;
6850 struct mem_cgroup *to = mc.to;
6853 /* we must uncharge all the leftover precharges from mc.to */
6855 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6859 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6860 * we must uncharge here.
6862 if (mc.moved_charge) {
6863 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6864 mc.moved_charge = 0;
6866 /* we must fixup refcnts and charges */
6867 if (mc.moved_swap) {
6868 /* uncharge swap account from the old cgroup */
6869 if (!mem_cgroup_is_root(mc.from))
6870 res_counter_uncharge(&mc.from->memsw,
6871 PAGE_SIZE * mc.moved_swap);
6873 for (i = 0; i < mc.moved_swap; i++)
6874 css_put(&mc.from->css);
6876 if (!mem_cgroup_is_root(mc.to)) {
6878 * we charged both to->res and to->memsw, so we should
6881 res_counter_uncharge(&mc.to->res,
6882 PAGE_SIZE * mc.moved_swap);
6884 /* we've already done css_get(mc.to) */
6887 memcg_oom_recover(from);
6888 memcg_oom_recover(to);
6889 wake_up_all(&mc.waitq);
6892 static void mem_cgroup_clear_mc(void)
6894 struct mem_cgroup *from = mc.from;
6897 * we must clear moving_task before waking up waiters at the end of
6900 mc.moving_task = NULL;
6901 __mem_cgroup_clear_mc();
6902 spin_lock(&mc.lock);
6905 spin_unlock(&mc.lock);
6906 mem_cgroup_end_move(from);
6909 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6910 struct cgroup_taskset *tset)
6912 struct task_struct *p = cgroup_taskset_first(tset);
6914 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6915 unsigned long move_charge_at_immigrate;
6918 * We are now commited to this value whatever it is. Changes in this
6919 * tunable will only affect upcoming migrations, not the current one.
6920 * So we need to save it, and keep it going.
6922 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6923 if (move_charge_at_immigrate) {
6924 struct mm_struct *mm;
6925 struct mem_cgroup *from = mem_cgroup_from_task(p);
6927 VM_BUG_ON(from == memcg);
6929 mm = get_task_mm(p);
6932 /* We move charges only when we move a owner of the mm */
6933 if (mm->owner == p) {
6936 VM_BUG_ON(mc.precharge);
6937 VM_BUG_ON(mc.moved_charge);
6938 VM_BUG_ON(mc.moved_swap);
6939 mem_cgroup_start_move(from);
6940 spin_lock(&mc.lock);
6943 mc.immigrate_flags = move_charge_at_immigrate;
6944 spin_unlock(&mc.lock);
6945 /* We set mc.moving_task later */
6947 ret = mem_cgroup_precharge_mc(mm);
6949 mem_cgroup_clear_mc();
6956 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6957 struct cgroup_taskset *tset)
6959 mem_cgroup_clear_mc();
6962 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6963 unsigned long addr, unsigned long end,
6964 struct mm_walk *walk)
6967 struct vm_area_struct *vma = walk->private;
6970 enum mc_target_type target_type;
6971 union mc_target target;
6973 struct page_cgroup *pc;
6976 * We don't take compound_lock() here but no race with splitting thp
6978 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6979 * under splitting, which means there's no concurrent thp split,
6980 * - if another thread runs into split_huge_page() just after we
6981 * entered this if-block, the thread must wait for page table lock
6982 * to be unlocked in __split_huge_page_splitting(), where the main
6983 * part of thp split is not executed yet.
6985 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6986 if (mc.precharge < HPAGE_PMD_NR) {
6990 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6991 if (target_type == MC_TARGET_PAGE) {
6993 if (!isolate_lru_page(page)) {
6994 pc = lookup_page_cgroup(page);
6995 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6996 pc, mc.from, mc.to)) {
6997 mc.precharge -= HPAGE_PMD_NR;
6998 mc.moved_charge += HPAGE_PMD_NR;
7000 putback_lru_page(page);
7008 if (pmd_trans_unstable(pmd))
7011 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7012 for (; addr != end; addr += PAGE_SIZE) {
7013 pte_t ptent = *(pte++);
7019 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7020 case MC_TARGET_PAGE:
7022 if (isolate_lru_page(page))
7024 pc = lookup_page_cgroup(page);
7025 if (!mem_cgroup_move_account(page, 1, pc,
7028 /* we uncharge from mc.from later. */
7031 putback_lru_page(page);
7032 put: /* get_mctgt_type() gets the page */
7035 case MC_TARGET_SWAP:
7037 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7039 /* we fixup refcnts and charges later. */
7047 pte_unmap_unlock(pte - 1, ptl);
7052 * We have consumed all precharges we got in can_attach().
7053 * We try charge one by one, but don't do any additional
7054 * charges to mc.to if we have failed in charge once in attach()
7057 ret = mem_cgroup_do_precharge(1);
7065 static void mem_cgroup_move_charge(struct mm_struct *mm)
7067 struct vm_area_struct *vma;
7069 lru_add_drain_all();
7071 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7073 * Someone who are holding the mmap_sem might be waiting in
7074 * waitq. So we cancel all extra charges, wake up all waiters,
7075 * and retry. Because we cancel precharges, we might not be able
7076 * to move enough charges, but moving charge is a best-effort
7077 * feature anyway, so it wouldn't be a big problem.
7079 __mem_cgroup_clear_mc();
7083 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7085 struct mm_walk mem_cgroup_move_charge_walk = {
7086 .pmd_entry = mem_cgroup_move_charge_pte_range,
7090 if (is_vm_hugetlb_page(vma))
7092 ret = walk_page_range(vma->vm_start, vma->vm_end,
7093 &mem_cgroup_move_charge_walk);
7096 * means we have consumed all precharges and failed in
7097 * doing additional charge. Just abandon here.
7101 up_read(&mm->mmap_sem);
7104 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7105 struct cgroup_taskset *tset)
7107 struct task_struct *p = cgroup_taskset_first(tset);
7108 struct mm_struct *mm = get_task_mm(p);
7112 mem_cgroup_move_charge(mm);
7116 mem_cgroup_clear_mc();
7118 #else /* !CONFIG_MMU */
7119 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7120 struct cgroup_taskset *tset)
7124 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7125 struct cgroup_taskset *tset)
7128 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7129 struct cgroup_taskset *tset)
7135 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7136 * to verify sane_behavior flag on each mount attempt.
7138 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7141 * use_hierarchy is forced with sane_behavior. cgroup core
7142 * guarantees that @root doesn't have any children, so turning it
7143 * on for the root memcg is enough.
7145 if (cgroup_sane_behavior(root_css->cgroup))
7146 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7149 struct cgroup_subsys memory_cgrp_subsys = {
7150 .css_alloc = mem_cgroup_css_alloc,
7151 .css_online = mem_cgroup_css_online,
7152 .css_offline = mem_cgroup_css_offline,
7153 .css_free = mem_cgroup_css_free,
7154 .can_attach = mem_cgroup_can_attach,
7155 .cancel_attach = mem_cgroup_cancel_attach,
7156 .attach = mem_cgroup_move_task,
7157 .bind = mem_cgroup_bind,
7158 .base_cftypes = mem_cgroup_files,
7162 #ifdef CONFIG_MEMCG_SWAP
7163 static int __init enable_swap_account(char *s)
7165 if (!strcmp(s, "1"))
7166 really_do_swap_account = 1;
7167 else if (!strcmp(s, "0"))
7168 really_do_swap_account = 0;
7171 __setup("swapaccount=", enable_swap_account);
7173 static void __init memsw_file_init(void)
7175 WARN_ON(cgroup_add_cftypes(&memory_cgrp_subsys, memsw_cgroup_files));
7178 static void __init enable_swap_cgroup(void)
7180 if (!mem_cgroup_disabled() && really_do_swap_account) {
7181 do_swap_account = 1;
7187 static void __init enable_swap_cgroup(void)
7193 * subsys_initcall() for memory controller.
7195 * Some parts like hotcpu_notifier() have to be initialized from this context
7196 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7197 * everything that doesn't depend on a specific mem_cgroup structure should
7198 * be initialized from here.
7200 static int __init mem_cgroup_init(void)
7202 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7203 enable_swap_cgroup();
7204 mem_cgroup_soft_limit_tree_init();
7208 subsys_initcall(mem_cgroup_init);