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)
927 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
928 * counted as CACHE even if it's on ANON LRU.
931 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
934 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
937 if (PageTransHuge(page))
938 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
941 /* pagein of a big page is an event. So, ignore page size */
943 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
945 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
946 nr_pages = -nr_pages; /* for event */
949 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
955 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
957 struct mem_cgroup_per_zone *mz;
959 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
960 return mz->lru_size[lru];
964 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
965 unsigned int lru_mask)
967 struct mem_cgroup_per_zone *mz;
969 unsigned long ret = 0;
971 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
974 if (BIT(lru) & lru_mask)
975 ret += mz->lru_size[lru];
981 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
982 int nid, unsigned int lru_mask)
987 for (zid = 0; zid < MAX_NR_ZONES; zid++)
988 total += mem_cgroup_zone_nr_lru_pages(memcg,
994 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
995 unsigned int lru_mask)
1000 for_each_node_state(nid, N_MEMORY)
1001 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1005 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1006 enum mem_cgroup_events_target target)
1008 unsigned long val, next;
1010 val = __this_cpu_read(memcg->stat->nr_page_events);
1011 next = __this_cpu_read(memcg->stat->targets[target]);
1012 /* from time_after() in jiffies.h */
1013 if ((long)next - (long)val < 0) {
1015 case MEM_CGROUP_TARGET_THRESH:
1016 next = val + THRESHOLDS_EVENTS_TARGET;
1018 case MEM_CGROUP_TARGET_SOFTLIMIT:
1019 next = val + SOFTLIMIT_EVENTS_TARGET;
1021 case MEM_CGROUP_TARGET_NUMAINFO:
1022 next = val + NUMAINFO_EVENTS_TARGET;
1027 __this_cpu_write(memcg->stat->targets[target], next);
1034 * Check events in order.
1037 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1040 /* threshold event is triggered in finer grain than soft limit */
1041 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1042 MEM_CGROUP_TARGET_THRESH))) {
1044 bool do_numainfo __maybe_unused;
1046 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1047 MEM_CGROUP_TARGET_SOFTLIMIT);
1048 #if MAX_NUMNODES > 1
1049 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1050 MEM_CGROUP_TARGET_NUMAINFO);
1054 mem_cgroup_threshold(memcg);
1055 if (unlikely(do_softlimit))
1056 mem_cgroup_update_tree(memcg, page);
1057 #if MAX_NUMNODES > 1
1058 if (unlikely(do_numainfo))
1059 atomic_inc(&memcg->numainfo_events);
1065 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1068 * mm_update_next_owner() may clear mm->owner to NULL
1069 * if it races with swapoff, page migration, etc.
1070 * So this can be called with p == NULL.
1075 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1078 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1080 struct mem_cgroup *memcg = NULL;
1085 * Because we have no locks, mm->owner's may be being moved to other
1086 * cgroup. We use css_tryget() here even if this looks
1087 * pessimistic (rather than adding locks here).
1091 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1092 if (unlikely(!memcg))
1094 } while (!css_tryget(&memcg->css));
1100 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1101 * ref. count) or NULL if the whole root's subtree has been visited.
1103 * helper function to be used by mem_cgroup_iter
1105 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1106 struct mem_cgroup *last_visited)
1108 struct cgroup_subsys_state *prev_css, *next_css;
1110 prev_css = last_visited ? &last_visited->css : NULL;
1112 next_css = css_next_descendant_pre(prev_css, &root->css);
1115 * Even if we found a group we have to make sure it is
1116 * alive. css && !memcg means that the groups should be
1117 * skipped and we should continue the tree walk.
1118 * last_visited css is safe to use because it is
1119 * protected by css_get and the tree walk is rcu safe.
1121 * We do not take a reference on the root of the tree walk
1122 * because we might race with the root removal when it would
1123 * be the only node in the iterated hierarchy and mem_cgroup_iter
1124 * would end up in an endless loop because it expects that at
1125 * least one valid node will be returned. Root cannot disappear
1126 * because caller of the iterator should hold it already so
1127 * skipping css reference should be safe.
1130 if ((next_css == &root->css) ||
1131 ((next_css->flags & CSS_ONLINE) && css_tryget(next_css)))
1132 return mem_cgroup_from_css(next_css);
1134 prev_css = next_css;
1141 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1144 * When a group in the hierarchy below root is destroyed, the
1145 * hierarchy iterator can no longer be trusted since it might
1146 * have pointed to the destroyed group. Invalidate it.
1148 atomic_inc(&root->dead_count);
1151 static struct mem_cgroup *
1152 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1153 struct mem_cgroup *root,
1156 struct mem_cgroup *position = NULL;
1158 * A cgroup destruction happens in two stages: offlining and
1159 * release. They are separated by a RCU grace period.
1161 * If the iterator is valid, we may still race with an
1162 * offlining. The RCU lock ensures the object won't be
1163 * released, tryget will fail if we lost the race.
1165 *sequence = atomic_read(&root->dead_count);
1166 if (iter->last_dead_count == *sequence) {
1168 position = iter->last_visited;
1171 * We cannot take a reference to root because we might race
1172 * with root removal and returning NULL would end up in
1173 * an endless loop on the iterator user level when root
1174 * would be returned all the time.
1176 if (position && position != root &&
1177 !css_tryget(&position->css))
1183 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1184 struct mem_cgroup *last_visited,
1185 struct mem_cgroup *new_position,
1186 struct mem_cgroup *root,
1189 /* root reference counting symmetric to mem_cgroup_iter_load */
1190 if (last_visited && last_visited != root)
1191 css_put(&last_visited->css);
1193 * We store the sequence count from the time @last_visited was
1194 * loaded successfully instead of rereading it here so that we
1195 * don't lose destruction events in between. We could have
1196 * raced with the destruction of @new_position after all.
1198 iter->last_visited = new_position;
1200 iter->last_dead_count = sequence;
1204 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1205 * @root: hierarchy root
1206 * @prev: previously returned memcg, NULL on first invocation
1207 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1209 * Returns references to children of the hierarchy below @root, or
1210 * @root itself, or %NULL after a full round-trip.
1212 * Caller must pass the return value in @prev on subsequent
1213 * invocations for reference counting, or use mem_cgroup_iter_break()
1214 * to cancel a hierarchy walk before the round-trip is complete.
1216 * Reclaimers can specify a zone and a priority level in @reclaim to
1217 * divide up the memcgs in the hierarchy among all concurrent
1218 * reclaimers operating on the same zone and priority.
1220 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1221 struct mem_cgroup *prev,
1222 struct mem_cgroup_reclaim_cookie *reclaim)
1224 struct mem_cgroup *memcg = NULL;
1225 struct mem_cgroup *last_visited = NULL;
1227 if (mem_cgroup_disabled())
1231 root = root_mem_cgroup;
1233 if (prev && !reclaim)
1234 last_visited = prev;
1236 if (!root->use_hierarchy && root != root_mem_cgroup) {
1244 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1245 int uninitialized_var(seq);
1248 int nid = zone_to_nid(reclaim->zone);
1249 int zid = zone_idx(reclaim->zone);
1250 struct mem_cgroup_per_zone *mz;
1252 mz = mem_cgroup_zoneinfo(root, nid, zid);
1253 iter = &mz->reclaim_iter[reclaim->priority];
1254 if (prev && reclaim->generation != iter->generation) {
1255 iter->last_visited = NULL;
1259 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1262 memcg = __mem_cgroup_iter_next(root, last_visited);
1265 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1270 else if (!prev && memcg)
1271 reclaim->generation = iter->generation;
1280 if (prev && prev != root)
1281 css_put(&prev->css);
1287 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1288 * @root: hierarchy root
1289 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1291 void mem_cgroup_iter_break(struct mem_cgroup *root,
1292 struct mem_cgroup *prev)
1295 root = root_mem_cgroup;
1296 if (prev && prev != root)
1297 css_put(&prev->css);
1301 * Iteration constructs for visiting all cgroups (under a tree). If
1302 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1303 * be used for reference counting.
1305 #define for_each_mem_cgroup_tree(iter, root) \
1306 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1308 iter = mem_cgroup_iter(root, iter, NULL))
1310 #define for_each_mem_cgroup(iter) \
1311 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1313 iter = mem_cgroup_iter(NULL, iter, NULL))
1315 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1317 struct mem_cgroup *memcg;
1320 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1321 if (unlikely(!memcg))
1326 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1329 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1337 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1340 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1341 * @zone: zone of the wanted lruvec
1342 * @memcg: memcg of the wanted lruvec
1344 * Returns the lru list vector holding pages for the given @zone and
1345 * @mem. This can be the global zone lruvec, if the memory controller
1348 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1349 struct mem_cgroup *memcg)
1351 struct mem_cgroup_per_zone *mz;
1352 struct lruvec *lruvec;
1354 if (mem_cgroup_disabled()) {
1355 lruvec = &zone->lruvec;
1359 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1360 lruvec = &mz->lruvec;
1363 * Since a node can be onlined after the mem_cgroup was created,
1364 * we have to be prepared to initialize lruvec->zone here;
1365 * and if offlined then reonlined, we need to reinitialize it.
1367 if (unlikely(lruvec->zone != zone))
1368 lruvec->zone = zone;
1373 * Following LRU functions are allowed to be used without PCG_LOCK.
1374 * Operations are called by routine of global LRU independently from memcg.
1375 * What we have to take care of here is validness of pc->mem_cgroup.
1377 * Changes to pc->mem_cgroup happens when
1380 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1381 * It is added to LRU before charge.
1382 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1383 * When moving account, the page is not on LRU. It's isolated.
1387 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1389 * @zone: zone of the page
1391 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1393 struct mem_cgroup_per_zone *mz;
1394 struct mem_cgroup *memcg;
1395 struct page_cgroup *pc;
1396 struct lruvec *lruvec;
1398 if (mem_cgroup_disabled()) {
1399 lruvec = &zone->lruvec;
1403 pc = lookup_page_cgroup(page);
1404 memcg = pc->mem_cgroup;
1407 * Surreptitiously switch any uncharged offlist page to root:
1408 * an uncharged page off lru does nothing to secure
1409 * its former mem_cgroup from sudden removal.
1411 * Our caller holds lru_lock, and PageCgroupUsed is updated
1412 * under page_cgroup lock: between them, they make all uses
1413 * of pc->mem_cgroup safe.
1415 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1416 pc->mem_cgroup = memcg = root_mem_cgroup;
1418 mz = page_cgroup_zoneinfo(memcg, page);
1419 lruvec = &mz->lruvec;
1422 * Since a node can be onlined after the mem_cgroup was created,
1423 * we have to be prepared to initialize lruvec->zone here;
1424 * and if offlined then reonlined, we need to reinitialize it.
1426 if (unlikely(lruvec->zone != zone))
1427 lruvec->zone = zone;
1432 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1433 * @lruvec: mem_cgroup per zone lru vector
1434 * @lru: index of lru list the page is sitting on
1435 * @nr_pages: positive when adding or negative when removing
1437 * This function must be called when a page is added to or removed from an
1440 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1443 struct mem_cgroup_per_zone *mz;
1444 unsigned long *lru_size;
1446 if (mem_cgroup_disabled())
1449 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1450 lru_size = mz->lru_size + lru;
1451 *lru_size += nr_pages;
1452 VM_BUG_ON((long)(*lru_size) < 0);
1456 * Checks whether given mem is same or in the root_mem_cgroup's
1459 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1460 struct mem_cgroup *memcg)
1462 if (root_memcg == memcg)
1464 if (!root_memcg->use_hierarchy || !memcg)
1466 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1469 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1470 struct mem_cgroup *memcg)
1475 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1480 bool task_in_mem_cgroup(struct task_struct *task,
1481 const struct mem_cgroup *memcg)
1483 struct mem_cgroup *curr = NULL;
1484 struct task_struct *p;
1487 p = find_lock_task_mm(task);
1489 curr = try_get_mem_cgroup_from_mm(p->mm);
1493 * All threads may have already detached their mm's, but the oom
1494 * killer still needs to detect if they have already been oom
1495 * killed to prevent needlessly killing additional tasks.
1498 curr = mem_cgroup_from_task(task);
1500 css_get(&curr->css);
1506 * We should check use_hierarchy of "memcg" not "curr". Because checking
1507 * use_hierarchy of "curr" here make this function true if hierarchy is
1508 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1509 * hierarchy(even if use_hierarchy is disabled in "memcg").
1511 ret = mem_cgroup_same_or_subtree(memcg, curr);
1512 css_put(&curr->css);
1516 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1518 unsigned long inactive_ratio;
1519 unsigned long inactive;
1520 unsigned long active;
1523 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1524 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1526 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1528 inactive_ratio = int_sqrt(10 * gb);
1532 return inactive * inactive_ratio < active;
1535 #define mem_cgroup_from_res_counter(counter, member) \
1536 container_of(counter, struct mem_cgroup, member)
1539 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1540 * @memcg: the memory cgroup
1542 * Returns the maximum amount of memory @mem can be charged with, in
1545 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1547 unsigned long long margin;
1549 margin = res_counter_margin(&memcg->res);
1550 if (do_swap_account)
1551 margin = min(margin, res_counter_margin(&memcg->memsw));
1552 return margin >> PAGE_SHIFT;
1555 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1558 if (!css_parent(&memcg->css))
1559 return vm_swappiness;
1561 return memcg->swappiness;
1565 * memcg->moving_account is used for checking possibility that some thread is
1566 * calling move_account(). When a thread on CPU-A starts moving pages under
1567 * a memcg, other threads should check memcg->moving_account under
1568 * rcu_read_lock(), like this:
1572 * memcg->moving_account+1 if (memcg->mocing_account)
1574 * synchronize_rcu() update something.
1579 /* for quick checking without looking up memcg */
1580 atomic_t memcg_moving __read_mostly;
1582 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1584 atomic_inc(&memcg_moving);
1585 atomic_inc(&memcg->moving_account);
1589 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1592 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1593 * We check NULL in callee rather than caller.
1596 atomic_dec(&memcg_moving);
1597 atomic_dec(&memcg->moving_account);
1602 * 2 routines for checking "mem" is under move_account() or not.
1604 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1605 * is used for avoiding races in accounting. If true,
1606 * pc->mem_cgroup may be overwritten.
1608 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1609 * under hierarchy of moving cgroups. This is for
1610 * waiting at hith-memory prressure caused by "move".
1613 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1615 VM_BUG_ON(!rcu_read_lock_held());
1616 return atomic_read(&memcg->moving_account) > 0;
1619 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1621 struct mem_cgroup *from;
1622 struct mem_cgroup *to;
1625 * Unlike task_move routines, we access mc.to, mc.from not under
1626 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1628 spin_lock(&mc.lock);
1634 ret = mem_cgroup_same_or_subtree(memcg, from)
1635 || mem_cgroup_same_or_subtree(memcg, to);
1637 spin_unlock(&mc.lock);
1641 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1643 if (mc.moving_task && current != mc.moving_task) {
1644 if (mem_cgroup_under_move(memcg)) {
1646 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1647 /* moving charge context might have finished. */
1650 finish_wait(&mc.waitq, &wait);
1658 * Take this lock when
1659 * - a code tries to modify page's memcg while it's USED.
1660 * - a code tries to modify page state accounting in a memcg.
1661 * see mem_cgroup_stolen(), too.
1663 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1664 unsigned long *flags)
1666 spin_lock_irqsave(&memcg->move_lock, *flags);
1669 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1670 unsigned long *flags)
1672 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1675 #define K(x) ((x) << (PAGE_SHIFT-10))
1677 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1678 * @memcg: The memory cgroup that went over limit
1679 * @p: Task that is going to be killed
1681 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1684 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1686 /* oom_info_lock ensures that parallel ooms do not interleave */
1687 static DEFINE_MUTEX(oom_info_lock);
1688 struct mem_cgroup *iter;
1694 mutex_lock(&oom_info_lock);
1697 pr_info("Task in ");
1698 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1699 pr_info(" killed as a result of limit of ");
1700 pr_cont_cgroup_path(memcg->css.cgroup);
1705 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1706 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1707 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1708 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1709 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1710 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1711 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1712 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1713 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1714 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1715 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1716 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1718 for_each_mem_cgroup_tree(iter, memcg) {
1719 pr_info("Memory cgroup stats for ");
1720 pr_cont_cgroup_path(iter->css.cgroup);
1723 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1724 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1726 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1727 K(mem_cgroup_read_stat(iter, i)));
1730 for (i = 0; i < NR_LRU_LISTS; i++)
1731 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1732 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1736 mutex_unlock(&oom_info_lock);
1740 * This function returns the number of memcg under hierarchy tree. Returns
1741 * 1(self count) if no children.
1743 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1746 struct mem_cgroup *iter;
1748 for_each_mem_cgroup_tree(iter, memcg)
1754 * Return the memory (and swap, if configured) limit for a memcg.
1756 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1760 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1763 * Do not consider swap space if we cannot swap due to swappiness
1765 if (mem_cgroup_swappiness(memcg)) {
1768 limit += total_swap_pages << PAGE_SHIFT;
1769 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1772 * If memsw is finite and limits the amount of swap space
1773 * available to this memcg, return that limit.
1775 limit = min(limit, memsw);
1781 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1784 struct mem_cgroup *iter;
1785 unsigned long chosen_points = 0;
1786 unsigned long totalpages;
1787 unsigned int points = 0;
1788 struct task_struct *chosen = NULL;
1791 * If current has a pending SIGKILL or is exiting, then automatically
1792 * select it. The goal is to allow it to allocate so that it may
1793 * quickly exit and free its memory.
1795 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1796 set_thread_flag(TIF_MEMDIE);
1800 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1801 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1802 for_each_mem_cgroup_tree(iter, memcg) {
1803 struct css_task_iter it;
1804 struct task_struct *task;
1806 css_task_iter_start(&iter->css, &it);
1807 while ((task = css_task_iter_next(&it))) {
1808 switch (oom_scan_process_thread(task, totalpages, NULL,
1810 case OOM_SCAN_SELECT:
1812 put_task_struct(chosen);
1814 chosen_points = ULONG_MAX;
1815 get_task_struct(chosen);
1817 case OOM_SCAN_CONTINUE:
1819 case OOM_SCAN_ABORT:
1820 css_task_iter_end(&it);
1821 mem_cgroup_iter_break(memcg, iter);
1823 put_task_struct(chosen);
1828 points = oom_badness(task, memcg, NULL, totalpages);
1829 if (!points || points < chosen_points)
1831 /* Prefer thread group leaders for display purposes */
1832 if (points == chosen_points &&
1833 thread_group_leader(chosen))
1837 put_task_struct(chosen);
1839 chosen_points = points;
1840 get_task_struct(chosen);
1842 css_task_iter_end(&it);
1847 points = chosen_points * 1000 / totalpages;
1848 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1849 NULL, "Memory cgroup out of memory");
1852 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1854 unsigned long flags)
1856 unsigned long total = 0;
1857 bool noswap = false;
1860 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1862 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1865 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1867 drain_all_stock_async(memcg);
1868 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1870 * Allow limit shrinkers, which are triggered directly
1871 * by userspace, to catch signals and stop reclaim
1872 * after minimal progress, regardless of the margin.
1874 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1876 if (mem_cgroup_margin(memcg))
1879 * If nothing was reclaimed after two attempts, there
1880 * may be no reclaimable pages in this hierarchy.
1889 * test_mem_cgroup_node_reclaimable
1890 * @memcg: the target memcg
1891 * @nid: the node ID to be checked.
1892 * @noswap : specify true here if the user wants flle only information.
1894 * This function returns whether the specified memcg contains any
1895 * reclaimable pages on a node. Returns true if there are any reclaimable
1896 * pages in the node.
1898 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1899 int nid, bool noswap)
1901 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1903 if (noswap || !total_swap_pages)
1905 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1910 #if MAX_NUMNODES > 1
1913 * Always updating the nodemask is not very good - even if we have an empty
1914 * list or the wrong list here, we can start from some node and traverse all
1915 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1918 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1922 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1923 * pagein/pageout changes since the last update.
1925 if (!atomic_read(&memcg->numainfo_events))
1927 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1930 /* make a nodemask where this memcg uses memory from */
1931 memcg->scan_nodes = node_states[N_MEMORY];
1933 for_each_node_mask(nid, node_states[N_MEMORY]) {
1935 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1936 node_clear(nid, memcg->scan_nodes);
1939 atomic_set(&memcg->numainfo_events, 0);
1940 atomic_set(&memcg->numainfo_updating, 0);
1944 * Selecting a node where we start reclaim from. Because what we need is just
1945 * reducing usage counter, start from anywhere is O,K. Considering
1946 * memory reclaim from current node, there are pros. and cons.
1948 * Freeing memory from current node means freeing memory from a node which
1949 * we'll use or we've used. So, it may make LRU bad. And if several threads
1950 * hit limits, it will see a contention on a node. But freeing from remote
1951 * node means more costs for memory reclaim because of memory latency.
1953 * Now, we use round-robin. Better algorithm is welcomed.
1955 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1959 mem_cgroup_may_update_nodemask(memcg);
1960 node = memcg->last_scanned_node;
1962 node = next_node(node, memcg->scan_nodes);
1963 if (node == MAX_NUMNODES)
1964 node = first_node(memcg->scan_nodes);
1966 * We call this when we hit limit, not when pages are added to LRU.
1967 * No LRU may hold pages because all pages are UNEVICTABLE or
1968 * memcg is too small and all pages are not on LRU. In that case,
1969 * we use curret node.
1971 if (unlikely(node == MAX_NUMNODES))
1972 node = numa_node_id();
1974 memcg->last_scanned_node = node;
1979 * Check all nodes whether it contains reclaimable pages or not.
1980 * For quick scan, we make use of scan_nodes. This will allow us to skip
1981 * unused nodes. But scan_nodes is lazily updated and may not cotain
1982 * enough new information. We need to do double check.
1984 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1989 * quick check...making use of scan_node.
1990 * We can skip unused nodes.
1992 if (!nodes_empty(memcg->scan_nodes)) {
1993 for (nid = first_node(memcg->scan_nodes);
1995 nid = next_node(nid, memcg->scan_nodes)) {
1997 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2002 * Check rest of nodes.
2004 for_each_node_state(nid, N_MEMORY) {
2005 if (node_isset(nid, memcg->scan_nodes))
2007 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2014 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2019 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2021 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2025 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2028 unsigned long *total_scanned)
2030 struct mem_cgroup *victim = NULL;
2033 unsigned long excess;
2034 unsigned long nr_scanned;
2035 struct mem_cgroup_reclaim_cookie reclaim = {
2040 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2043 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2048 * If we have not been able to reclaim
2049 * anything, it might because there are
2050 * no reclaimable pages under this hierarchy
2055 * We want to do more targeted reclaim.
2056 * excess >> 2 is not to excessive so as to
2057 * reclaim too much, nor too less that we keep
2058 * coming back to reclaim from this cgroup
2060 if (total >= (excess >> 2) ||
2061 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2066 if (!mem_cgroup_reclaimable(victim, false))
2068 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2070 *total_scanned += nr_scanned;
2071 if (!res_counter_soft_limit_excess(&root_memcg->res))
2074 mem_cgroup_iter_break(root_memcg, victim);
2078 #ifdef CONFIG_LOCKDEP
2079 static struct lockdep_map memcg_oom_lock_dep_map = {
2080 .name = "memcg_oom_lock",
2084 static DEFINE_SPINLOCK(memcg_oom_lock);
2087 * Check OOM-Killer is already running under our hierarchy.
2088 * If someone is running, return false.
2090 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2092 struct mem_cgroup *iter, *failed = NULL;
2094 spin_lock(&memcg_oom_lock);
2096 for_each_mem_cgroup_tree(iter, memcg) {
2097 if (iter->oom_lock) {
2099 * this subtree of our hierarchy is already locked
2100 * so we cannot give a lock.
2103 mem_cgroup_iter_break(memcg, iter);
2106 iter->oom_lock = true;
2111 * OK, we failed to lock the whole subtree so we have
2112 * to clean up what we set up to the failing subtree
2114 for_each_mem_cgroup_tree(iter, memcg) {
2115 if (iter == failed) {
2116 mem_cgroup_iter_break(memcg, iter);
2119 iter->oom_lock = false;
2122 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2124 spin_unlock(&memcg_oom_lock);
2129 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2131 struct mem_cgroup *iter;
2133 spin_lock(&memcg_oom_lock);
2134 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2135 for_each_mem_cgroup_tree(iter, memcg)
2136 iter->oom_lock = false;
2137 spin_unlock(&memcg_oom_lock);
2140 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2142 struct mem_cgroup *iter;
2144 for_each_mem_cgroup_tree(iter, memcg)
2145 atomic_inc(&iter->under_oom);
2148 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2150 struct mem_cgroup *iter;
2153 * When a new child is created while the hierarchy is under oom,
2154 * mem_cgroup_oom_lock() may not be called. We have to use
2155 * atomic_add_unless() here.
2157 for_each_mem_cgroup_tree(iter, memcg)
2158 atomic_add_unless(&iter->under_oom, -1, 0);
2161 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2163 struct oom_wait_info {
2164 struct mem_cgroup *memcg;
2168 static int memcg_oom_wake_function(wait_queue_t *wait,
2169 unsigned mode, int sync, void *arg)
2171 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2172 struct mem_cgroup *oom_wait_memcg;
2173 struct oom_wait_info *oom_wait_info;
2175 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2176 oom_wait_memcg = oom_wait_info->memcg;
2179 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2180 * Then we can use css_is_ancestor without taking care of RCU.
2182 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2183 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2185 return autoremove_wake_function(wait, mode, sync, arg);
2188 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2190 atomic_inc(&memcg->oom_wakeups);
2191 /* for filtering, pass "memcg" as argument. */
2192 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2195 static void memcg_oom_recover(struct mem_cgroup *memcg)
2197 if (memcg && atomic_read(&memcg->under_oom))
2198 memcg_wakeup_oom(memcg);
2201 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2203 if (!current->memcg_oom.may_oom)
2206 * We are in the middle of the charge context here, so we
2207 * don't want to block when potentially sitting on a callstack
2208 * that holds all kinds of filesystem and mm locks.
2210 * Also, the caller may handle a failed allocation gracefully
2211 * (like optional page cache readahead) and so an OOM killer
2212 * invocation might not even be necessary.
2214 * That's why we don't do anything here except remember the
2215 * OOM context and then deal with it at the end of the page
2216 * fault when the stack is unwound, the locks are released,
2217 * and when we know whether the fault was overall successful.
2219 css_get(&memcg->css);
2220 current->memcg_oom.memcg = memcg;
2221 current->memcg_oom.gfp_mask = mask;
2222 current->memcg_oom.order = order;
2226 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2227 * @handle: actually kill/wait or just clean up the OOM state
2229 * This has to be called at the end of a page fault if the memcg OOM
2230 * handler was enabled.
2232 * Memcg supports userspace OOM handling where failed allocations must
2233 * sleep on a waitqueue until the userspace task resolves the
2234 * situation. Sleeping directly in the charge context with all kinds
2235 * of locks held is not a good idea, instead we remember an OOM state
2236 * in the task and mem_cgroup_oom_synchronize() has to be called at
2237 * the end of the page fault to complete the OOM handling.
2239 * Returns %true if an ongoing memcg OOM situation was detected and
2240 * completed, %false otherwise.
2242 bool mem_cgroup_oom_synchronize(bool handle)
2244 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2245 struct oom_wait_info owait;
2248 /* OOM is global, do not handle */
2255 owait.memcg = memcg;
2256 owait.wait.flags = 0;
2257 owait.wait.func = memcg_oom_wake_function;
2258 owait.wait.private = current;
2259 INIT_LIST_HEAD(&owait.wait.task_list);
2261 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2262 mem_cgroup_mark_under_oom(memcg);
2264 locked = mem_cgroup_oom_trylock(memcg);
2267 mem_cgroup_oom_notify(memcg);
2269 if (locked && !memcg->oom_kill_disable) {
2270 mem_cgroup_unmark_under_oom(memcg);
2271 finish_wait(&memcg_oom_waitq, &owait.wait);
2272 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2273 current->memcg_oom.order);
2276 mem_cgroup_unmark_under_oom(memcg);
2277 finish_wait(&memcg_oom_waitq, &owait.wait);
2281 mem_cgroup_oom_unlock(memcg);
2283 * There is no guarantee that an OOM-lock contender
2284 * sees the wakeups triggered by the OOM kill
2285 * uncharges. Wake any sleepers explicitely.
2287 memcg_oom_recover(memcg);
2290 current->memcg_oom.memcg = NULL;
2291 css_put(&memcg->css);
2296 * Currently used to update mapped file statistics, but the routine can be
2297 * generalized to update other statistics as well.
2299 * Notes: Race condition
2301 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2302 * it tends to be costly. But considering some conditions, we doesn't need
2303 * to do so _always_.
2305 * Considering "charge", lock_page_cgroup() is not required because all
2306 * file-stat operations happen after a page is attached to radix-tree. There
2307 * are no race with "charge".
2309 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2310 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2311 * if there are race with "uncharge". Statistics itself is properly handled
2314 * Considering "move", this is an only case we see a race. To make the race
2315 * small, we check mm->moving_account and detect there are possibility of race
2316 * If there is, we take a lock.
2319 void __mem_cgroup_begin_update_page_stat(struct page *page,
2320 bool *locked, unsigned long *flags)
2322 struct mem_cgroup *memcg;
2323 struct page_cgroup *pc;
2325 pc = lookup_page_cgroup(page);
2327 memcg = pc->mem_cgroup;
2328 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2331 * If this memory cgroup is not under account moving, we don't
2332 * need to take move_lock_mem_cgroup(). Because we already hold
2333 * rcu_read_lock(), any calls to move_account will be delayed until
2334 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2336 if (!mem_cgroup_stolen(memcg))
2339 move_lock_mem_cgroup(memcg, flags);
2340 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2341 move_unlock_mem_cgroup(memcg, flags);
2347 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2349 struct page_cgroup *pc = lookup_page_cgroup(page);
2352 * It's guaranteed that pc->mem_cgroup never changes while
2353 * lock is held because a routine modifies pc->mem_cgroup
2354 * should take move_lock_mem_cgroup().
2356 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2359 void mem_cgroup_update_page_stat(struct page *page,
2360 enum mem_cgroup_stat_index idx, int val)
2362 struct mem_cgroup *memcg;
2363 struct page_cgroup *pc = lookup_page_cgroup(page);
2364 unsigned long uninitialized_var(flags);
2366 if (mem_cgroup_disabled())
2369 VM_BUG_ON(!rcu_read_lock_held());
2370 memcg = pc->mem_cgroup;
2371 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2374 this_cpu_add(memcg->stat->count[idx], val);
2378 * size of first charge trial. "32" comes from vmscan.c's magic value.
2379 * TODO: maybe necessary to use big numbers in big irons.
2381 #define CHARGE_BATCH 32U
2382 struct memcg_stock_pcp {
2383 struct mem_cgroup *cached; /* this never be root cgroup */
2384 unsigned int nr_pages;
2385 struct work_struct work;
2386 unsigned long flags;
2387 #define FLUSHING_CACHED_CHARGE 0
2389 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2390 static DEFINE_MUTEX(percpu_charge_mutex);
2393 * consume_stock: Try to consume stocked charge on this cpu.
2394 * @memcg: memcg to consume from.
2395 * @nr_pages: how many pages to charge.
2397 * The charges will only happen if @memcg matches the current cpu's memcg
2398 * stock, and at least @nr_pages are available in that stock. Failure to
2399 * service an allocation will refill the stock.
2401 * returns true if successful, false otherwise.
2403 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2405 struct memcg_stock_pcp *stock;
2408 if (nr_pages > CHARGE_BATCH)
2411 stock = &get_cpu_var(memcg_stock);
2412 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2413 stock->nr_pages -= nr_pages;
2414 else /* need to call res_counter_charge */
2416 put_cpu_var(memcg_stock);
2421 * Returns stocks cached in percpu to res_counter and reset cached information.
2423 static void drain_stock(struct memcg_stock_pcp *stock)
2425 struct mem_cgroup *old = stock->cached;
2427 if (stock->nr_pages) {
2428 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2430 res_counter_uncharge(&old->res, bytes);
2431 if (do_swap_account)
2432 res_counter_uncharge(&old->memsw, bytes);
2433 stock->nr_pages = 0;
2435 stock->cached = NULL;
2439 * This must be called under preempt disabled or must be called by
2440 * a thread which is pinned to local cpu.
2442 static void drain_local_stock(struct work_struct *dummy)
2444 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2446 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2449 static void __init memcg_stock_init(void)
2453 for_each_possible_cpu(cpu) {
2454 struct memcg_stock_pcp *stock =
2455 &per_cpu(memcg_stock, cpu);
2456 INIT_WORK(&stock->work, drain_local_stock);
2461 * Cache charges(val) which is from res_counter, to local per_cpu area.
2462 * This will be consumed by consume_stock() function, later.
2464 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2466 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2468 if (stock->cached != memcg) { /* reset if necessary */
2470 stock->cached = memcg;
2472 stock->nr_pages += nr_pages;
2473 put_cpu_var(memcg_stock);
2477 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2478 * of the hierarchy under it. sync flag says whether we should block
2479 * until the work is done.
2481 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2485 /* Notify other cpus that system-wide "drain" is running */
2488 for_each_online_cpu(cpu) {
2489 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2490 struct mem_cgroup *memcg;
2492 memcg = stock->cached;
2493 if (!memcg || !stock->nr_pages)
2495 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2497 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2499 drain_local_stock(&stock->work);
2501 schedule_work_on(cpu, &stock->work);
2509 for_each_online_cpu(cpu) {
2510 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2511 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2512 flush_work(&stock->work);
2519 * Tries to drain stocked charges in other cpus. This function is asynchronous
2520 * and just put a work per cpu for draining localy on each cpu. Caller can
2521 * expects some charges will be back to res_counter later but cannot wait for
2524 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2527 * If someone calls draining, avoid adding more kworker runs.
2529 if (!mutex_trylock(&percpu_charge_mutex))
2531 drain_all_stock(root_memcg, false);
2532 mutex_unlock(&percpu_charge_mutex);
2535 /* This is a synchronous drain interface. */
2536 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2538 /* called when force_empty is called */
2539 mutex_lock(&percpu_charge_mutex);
2540 drain_all_stock(root_memcg, true);
2541 mutex_unlock(&percpu_charge_mutex);
2545 * This function drains percpu counter value from DEAD cpu and
2546 * move it to local cpu. Note that this function can be preempted.
2548 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2552 spin_lock(&memcg->pcp_counter_lock);
2553 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2554 long x = per_cpu(memcg->stat->count[i], cpu);
2556 per_cpu(memcg->stat->count[i], cpu) = 0;
2557 memcg->nocpu_base.count[i] += x;
2559 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2560 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2562 per_cpu(memcg->stat->events[i], cpu) = 0;
2563 memcg->nocpu_base.events[i] += x;
2565 spin_unlock(&memcg->pcp_counter_lock);
2568 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2569 unsigned long action,
2572 int cpu = (unsigned long)hcpu;
2573 struct memcg_stock_pcp *stock;
2574 struct mem_cgroup *iter;
2576 if (action == CPU_ONLINE)
2579 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2582 for_each_mem_cgroup(iter)
2583 mem_cgroup_drain_pcp_counter(iter, cpu);
2585 stock = &per_cpu(memcg_stock, cpu);
2591 /* See __mem_cgroup_try_charge() for details */
2593 CHARGE_OK, /* success */
2594 CHARGE_RETRY, /* need to retry but retry is not bad */
2595 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2596 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2599 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2600 unsigned int nr_pages, unsigned int min_pages,
2603 unsigned long csize = nr_pages * PAGE_SIZE;
2604 struct mem_cgroup *mem_over_limit;
2605 struct res_counter *fail_res;
2606 unsigned long flags = 0;
2609 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2612 if (!do_swap_account)
2614 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2618 res_counter_uncharge(&memcg->res, csize);
2619 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2620 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2622 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2624 * Never reclaim on behalf of optional batching, retry with a
2625 * single page instead.
2627 if (nr_pages > min_pages)
2628 return CHARGE_RETRY;
2630 if (!(gfp_mask & __GFP_WAIT))
2631 return CHARGE_WOULDBLOCK;
2633 if (gfp_mask & __GFP_NORETRY)
2634 return CHARGE_NOMEM;
2636 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2637 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2638 return CHARGE_RETRY;
2640 * Even though the limit is exceeded at this point, reclaim
2641 * may have been able to free some pages. Retry the charge
2642 * before killing the task.
2644 * Only for regular pages, though: huge pages are rather
2645 * unlikely to succeed so close to the limit, and we fall back
2646 * to regular pages anyway in case of failure.
2648 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2649 return CHARGE_RETRY;
2652 * At task move, charge accounts can be doubly counted. So, it's
2653 * better to wait until the end of task_move if something is going on.
2655 if (mem_cgroup_wait_acct_move(mem_over_limit))
2656 return CHARGE_RETRY;
2659 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2661 return CHARGE_NOMEM;
2665 * __mem_cgroup_try_charge() does
2666 * 1. detect memcg to be charged against from passed *mm and *ptr,
2667 * 2. update res_counter
2668 * 3. call memory reclaim if necessary.
2670 * In some special case, if the task is fatal, fatal_signal_pending() or
2671 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2672 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2673 * as possible without any hazards. 2: all pages should have a valid
2674 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2675 * pointer, that is treated as a charge to root_mem_cgroup.
2677 * So __mem_cgroup_try_charge() will return
2678 * 0 ... on success, filling *ptr with a valid memcg pointer.
2679 * -ENOMEM ... charge failure because of resource limits.
2680 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2682 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2683 * the oom-killer can be invoked.
2685 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2687 unsigned int nr_pages,
2688 struct mem_cgroup **ptr,
2691 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2692 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2693 struct mem_cgroup *memcg = NULL;
2697 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2698 * in system level. So, allow to go ahead dying process in addition to
2701 if (unlikely(test_thread_flag(TIF_MEMDIE)
2702 || fatal_signal_pending(current)))
2705 if (unlikely(task_in_memcg_oom(current)))
2708 if (gfp_mask & __GFP_NOFAIL)
2712 * We always charge the cgroup the mm_struct belongs to.
2713 * The mm_struct's mem_cgroup changes on task migration if the
2714 * thread group leader migrates. It's possible that mm is not
2715 * set, if so charge the root memcg (happens for pagecache usage).
2718 *ptr = root_mem_cgroup;
2720 if (*ptr) { /* css should be a valid one */
2722 if (mem_cgroup_is_root(memcg))
2724 if (consume_stock(memcg, nr_pages))
2726 css_get(&memcg->css);
2728 struct task_struct *p;
2731 p = rcu_dereference(mm->owner);
2733 * Because we don't have task_lock(), "p" can exit.
2734 * In that case, "memcg" can point to root or p can be NULL with
2735 * race with swapoff. Then, we have small risk of mis-accouning.
2736 * But such kind of mis-account by race always happens because
2737 * we don't have cgroup_mutex(). It's overkill and we allo that
2739 * (*) swapoff at el will charge against mm-struct not against
2740 * task-struct. So, mm->owner can be NULL.
2742 memcg = mem_cgroup_from_task(p);
2744 memcg = root_mem_cgroup;
2745 if (mem_cgroup_is_root(memcg)) {
2749 if (consume_stock(memcg, nr_pages)) {
2751 * It seems dagerous to access memcg without css_get().
2752 * But considering how consume_stok works, it's not
2753 * necessary. If consume_stock success, some charges
2754 * from this memcg are cached on this cpu. So, we
2755 * don't need to call css_get()/css_tryget() before
2756 * calling consume_stock().
2761 /* after here, we may be blocked. we need to get refcnt */
2762 if (!css_tryget(&memcg->css)) {
2770 bool invoke_oom = oom && !nr_oom_retries;
2772 /* If killed, bypass charge */
2773 if (fatal_signal_pending(current)) {
2774 css_put(&memcg->css);
2778 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2779 nr_pages, invoke_oom);
2783 case CHARGE_RETRY: /* not in OOM situation but retry */
2785 css_put(&memcg->css);
2788 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2789 css_put(&memcg->css);
2791 case CHARGE_NOMEM: /* OOM routine works */
2792 if (!oom || invoke_oom) {
2793 css_put(&memcg->css);
2799 } while (ret != CHARGE_OK);
2801 if (batch > nr_pages)
2802 refill_stock(memcg, batch - nr_pages);
2803 css_put(&memcg->css);
2808 if (!(gfp_mask & __GFP_NOFAIL)) {
2813 *ptr = root_mem_cgroup;
2818 * Somemtimes we have to undo a charge we got by try_charge().
2819 * This function is for that and do uncharge, put css's refcnt.
2820 * gotten by try_charge().
2822 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2823 unsigned int nr_pages)
2825 if (!mem_cgroup_is_root(memcg)) {
2826 unsigned long bytes = nr_pages * PAGE_SIZE;
2828 res_counter_uncharge(&memcg->res, bytes);
2829 if (do_swap_account)
2830 res_counter_uncharge(&memcg->memsw, bytes);
2835 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2836 * This is useful when moving usage to parent cgroup.
2838 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2839 unsigned int nr_pages)
2841 unsigned long bytes = nr_pages * PAGE_SIZE;
2843 if (mem_cgroup_is_root(memcg))
2846 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2847 if (do_swap_account)
2848 res_counter_uncharge_until(&memcg->memsw,
2849 memcg->memsw.parent, bytes);
2853 * A helper function to get mem_cgroup from ID. must be called under
2854 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2855 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2856 * called against removed memcg.)
2858 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2860 /* ID 0 is unused ID */
2863 return mem_cgroup_from_id(id);
2866 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2868 struct mem_cgroup *memcg = NULL;
2869 struct page_cgroup *pc;
2873 VM_BUG_ON_PAGE(!PageLocked(page), page);
2875 pc = lookup_page_cgroup(page);
2876 lock_page_cgroup(pc);
2877 if (PageCgroupUsed(pc)) {
2878 memcg = pc->mem_cgroup;
2879 if (memcg && !css_tryget(&memcg->css))
2881 } else if (PageSwapCache(page)) {
2882 ent.val = page_private(page);
2883 id = lookup_swap_cgroup_id(ent);
2885 memcg = mem_cgroup_lookup(id);
2886 if (memcg && !css_tryget(&memcg->css))
2890 unlock_page_cgroup(pc);
2894 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2896 unsigned int nr_pages,
2897 enum charge_type ctype,
2900 struct page_cgroup *pc = lookup_page_cgroup(page);
2901 struct zone *uninitialized_var(zone);
2902 struct lruvec *lruvec;
2903 bool was_on_lru = false;
2906 lock_page_cgroup(pc);
2907 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2909 * we don't need page_cgroup_lock about tail pages, becase they are not
2910 * accessed by any other context at this point.
2914 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2915 * may already be on some other mem_cgroup's LRU. Take care of it.
2918 zone = page_zone(page);
2919 spin_lock_irq(&zone->lru_lock);
2920 if (PageLRU(page)) {
2921 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2923 del_page_from_lru_list(page, lruvec, page_lru(page));
2928 pc->mem_cgroup = memcg;
2930 * We access a page_cgroup asynchronously without lock_page_cgroup().
2931 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2932 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2933 * before USED bit, we need memory barrier here.
2934 * See mem_cgroup_add_lru_list(), etc.
2937 SetPageCgroupUsed(pc);
2941 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2942 VM_BUG_ON_PAGE(PageLRU(page), page);
2944 add_page_to_lru_list(page, lruvec, page_lru(page));
2946 spin_unlock_irq(&zone->lru_lock);
2949 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2954 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2955 unlock_page_cgroup(pc);
2958 * "charge_statistics" updated event counter. Then, check it.
2959 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2960 * if they exceeds softlimit.
2962 memcg_check_events(memcg, page);
2965 static DEFINE_MUTEX(set_limit_mutex);
2967 #ifdef CONFIG_MEMCG_KMEM
2968 static DEFINE_MUTEX(activate_kmem_mutex);
2970 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2972 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2973 memcg_kmem_is_active(memcg);
2977 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2978 * in the memcg_cache_params struct.
2980 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2982 struct kmem_cache *cachep;
2984 VM_BUG_ON(p->is_root_cache);
2985 cachep = p->root_cache;
2986 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2989 #ifdef CONFIG_SLABINFO
2990 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2992 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2993 struct memcg_cache_params *params;
2995 if (!memcg_can_account_kmem(memcg))
2998 print_slabinfo_header(m);
3000 mutex_lock(&memcg->slab_caches_mutex);
3001 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
3002 cache_show(memcg_params_to_cache(params), m);
3003 mutex_unlock(&memcg->slab_caches_mutex);
3009 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3011 struct res_counter *fail_res;
3012 struct mem_cgroup *_memcg;
3015 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3020 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3021 &_memcg, oom_gfp_allowed(gfp));
3023 if (ret == -EINTR) {
3025 * __mem_cgroup_try_charge() chosed to bypass to root due to
3026 * OOM kill or fatal signal. Since our only options are to
3027 * either fail the allocation or charge it to this cgroup, do
3028 * it as a temporary condition. But we can't fail. From a
3029 * kmem/slab perspective, the cache has already been selected,
3030 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3033 * This condition will only trigger if the task entered
3034 * memcg_charge_kmem in a sane state, but was OOM-killed during
3035 * __mem_cgroup_try_charge() above. Tasks that were already
3036 * dying when the allocation triggers should have been already
3037 * directed to the root cgroup in memcontrol.h
3039 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3040 if (do_swap_account)
3041 res_counter_charge_nofail(&memcg->memsw, size,
3045 res_counter_uncharge(&memcg->kmem, size);
3050 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3052 res_counter_uncharge(&memcg->res, size);
3053 if (do_swap_account)
3054 res_counter_uncharge(&memcg->memsw, size);
3057 if (res_counter_uncharge(&memcg->kmem, size))
3061 * Releases a reference taken in kmem_cgroup_css_offline in case
3062 * this last uncharge is racing with the offlining code or it is
3063 * outliving the memcg existence.
3065 * The memory barrier imposed by test&clear is paired with the
3066 * explicit one in memcg_kmem_mark_dead().
3068 if (memcg_kmem_test_and_clear_dead(memcg))
3069 css_put(&memcg->css);
3073 * helper for acessing a memcg's index. It will be used as an index in the
3074 * child cache array in kmem_cache, and also to derive its name. This function
3075 * will return -1 when this is not a kmem-limited memcg.
3077 int memcg_cache_id(struct mem_cgroup *memcg)
3079 return memcg ? memcg->kmemcg_id : -1;
3082 static size_t memcg_caches_array_size(int num_groups)
3085 if (num_groups <= 0)
3088 size = 2 * num_groups;
3089 if (size < MEMCG_CACHES_MIN_SIZE)
3090 size = MEMCG_CACHES_MIN_SIZE;
3091 else if (size > MEMCG_CACHES_MAX_SIZE)
3092 size = MEMCG_CACHES_MAX_SIZE;
3098 * We should update the current array size iff all caches updates succeed. This
3099 * can only be done from the slab side. The slab mutex needs to be held when
3102 void memcg_update_array_size(int num)
3104 if (num > memcg_limited_groups_array_size)
3105 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3108 static void kmem_cache_destroy_work_func(struct work_struct *w);
3110 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3112 struct memcg_cache_params *cur_params = s->memcg_params;
3114 VM_BUG_ON(!is_root_cache(s));
3116 if (num_groups > memcg_limited_groups_array_size) {
3118 struct memcg_cache_params *new_params;
3119 ssize_t size = memcg_caches_array_size(num_groups);
3121 size *= sizeof(void *);
3122 size += offsetof(struct memcg_cache_params, memcg_caches);
3124 new_params = kzalloc(size, GFP_KERNEL);
3128 new_params->is_root_cache = true;
3131 * There is the chance it will be bigger than
3132 * memcg_limited_groups_array_size, if we failed an allocation
3133 * in a cache, in which case all caches updated before it, will
3134 * have a bigger array.
3136 * But if that is the case, the data after
3137 * memcg_limited_groups_array_size is certainly unused
3139 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3140 if (!cur_params->memcg_caches[i])
3142 new_params->memcg_caches[i] =
3143 cur_params->memcg_caches[i];
3147 * Ideally, we would wait until all caches succeed, and only
3148 * then free the old one. But this is not worth the extra
3149 * pointer per-cache we'd have to have for this.
3151 * It is not a big deal if some caches are left with a size
3152 * bigger than the others. And all updates will reset this
3155 rcu_assign_pointer(s->memcg_params, new_params);
3157 kfree_rcu(cur_params, rcu_head);
3162 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3163 struct kmem_cache *root_cache)
3167 if (!memcg_kmem_enabled())
3171 size = offsetof(struct memcg_cache_params, memcg_caches);
3172 size += memcg_limited_groups_array_size * sizeof(void *);
3174 size = sizeof(struct memcg_cache_params);
3176 s->memcg_params = kzalloc(size, GFP_KERNEL);
3177 if (!s->memcg_params)
3181 s->memcg_params->memcg = memcg;
3182 s->memcg_params->root_cache = root_cache;
3183 INIT_WORK(&s->memcg_params->destroy,
3184 kmem_cache_destroy_work_func);
3186 s->memcg_params->is_root_cache = true;
3191 void memcg_free_cache_params(struct kmem_cache *s)
3193 kfree(s->memcg_params);
3196 void memcg_register_cache(struct kmem_cache *s)
3198 struct kmem_cache *root;
3199 struct mem_cgroup *memcg;
3202 if (is_root_cache(s))
3206 * Holding the slab_mutex assures nobody will touch the memcg_caches
3207 * array while we are modifying it.
3209 lockdep_assert_held(&slab_mutex);
3211 root = s->memcg_params->root_cache;
3212 memcg = s->memcg_params->memcg;
3213 id = memcg_cache_id(memcg);
3215 css_get(&memcg->css);
3219 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3220 * barrier here to ensure nobody will see the kmem_cache partially
3226 * Initialize the pointer to this cache in its parent's memcg_params
3227 * before adding it to the memcg_slab_caches list, otherwise we can
3228 * fail to convert memcg_params_to_cache() while traversing the list.
3230 VM_BUG_ON(root->memcg_params->memcg_caches[id]);
3231 root->memcg_params->memcg_caches[id] = s;
3233 mutex_lock(&memcg->slab_caches_mutex);
3234 list_add(&s->memcg_params->list, &memcg->memcg_slab_caches);
3235 mutex_unlock(&memcg->slab_caches_mutex);
3238 void memcg_unregister_cache(struct kmem_cache *s)
3240 struct kmem_cache *root;
3241 struct mem_cgroup *memcg;
3244 if (is_root_cache(s))
3248 * Holding the slab_mutex assures nobody will touch the memcg_caches
3249 * array while we are modifying it.
3251 lockdep_assert_held(&slab_mutex);
3253 root = s->memcg_params->root_cache;
3254 memcg = s->memcg_params->memcg;
3255 id = memcg_cache_id(memcg);
3257 mutex_lock(&memcg->slab_caches_mutex);
3258 list_del(&s->memcg_params->list);
3259 mutex_unlock(&memcg->slab_caches_mutex);
3262 * Clear the pointer to this cache in its parent's memcg_params only
3263 * after removing it from the memcg_slab_caches list, otherwise we can
3264 * fail to convert memcg_params_to_cache() while traversing the list.
3266 VM_BUG_ON(!root->memcg_params->memcg_caches[id]);
3267 root->memcg_params->memcg_caches[id] = NULL;
3269 css_put(&memcg->css);
3273 * During the creation a new cache, we need to disable our accounting mechanism
3274 * altogether. This is true even if we are not creating, but rather just
3275 * enqueing new caches to be created.
3277 * This is because that process will trigger allocations; some visible, like
3278 * explicit kmallocs to auxiliary data structures, name strings and internal
3279 * cache structures; some well concealed, like INIT_WORK() that can allocate
3280 * objects during debug.
3282 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3283 * to it. This may not be a bounded recursion: since the first cache creation
3284 * failed to complete (waiting on the allocation), we'll just try to create the
3285 * cache again, failing at the same point.
3287 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3288 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3289 * inside the following two functions.
3291 static inline void memcg_stop_kmem_account(void)
3293 VM_BUG_ON(!current->mm);
3294 current->memcg_kmem_skip_account++;
3297 static inline void memcg_resume_kmem_account(void)
3299 VM_BUG_ON(!current->mm);
3300 current->memcg_kmem_skip_account--;
3303 static void kmem_cache_destroy_work_func(struct work_struct *w)
3305 struct kmem_cache *cachep;
3306 struct memcg_cache_params *p;
3308 p = container_of(w, struct memcg_cache_params, destroy);
3310 cachep = memcg_params_to_cache(p);
3313 * If we get down to 0 after shrink, we could delete right away.
3314 * However, memcg_release_pages() already puts us back in the workqueue
3315 * in that case. If we proceed deleting, we'll get a dangling
3316 * reference, and removing the object from the workqueue in that case
3317 * is unnecessary complication. We are not a fast path.
3319 * Note that this case is fundamentally different from racing with
3320 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3321 * kmem_cache_shrink, not only we would be reinserting a dead cache
3322 * into the queue, but doing so from inside the worker racing to
3325 * So if we aren't down to zero, we'll just schedule a worker and try
3328 if (atomic_read(&cachep->memcg_params->nr_pages) != 0)
3329 kmem_cache_shrink(cachep);
3331 kmem_cache_destroy(cachep);
3334 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3336 if (!cachep->memcg_params->dead)
3340 * There are many ways in which we can get here.
3342 * We can get to a memory-pressure situation while the delayed work is
3343 * still pending to run. The vmscan shrinkers can then release all
3344 * cache memory and get us to destruction. If this is the case, we'll
3345 * be executed twice, which is a bug (the second time will execute over
3346 * bogus data). In this case, cancelling the work should be fine.
3348 * But we can also get here from the worker itself, if
3349 * kmem_cache_shrink is enough to shake all the remaining objects and
3350 * get the page count to 0. In this case, we'll deadlock if we try to
3351 * cancel the work (the worker runs with an internal lock held, which
3352 * is the same lock we would hold for cancel_work_sync().)
3354 * Since we can't possibly know who got us here, just refrain from
3355 * running if there is already work pending
3357 if (work_pending(&cachep->memcg_params->destroy))
3360 * We have to defer the actual destroying to a workqueue, because
3361 * we might currently be in a context that cannot sleep.
3363 schedule_work(&cachep->memcg_params->destroy);
3366 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3367 struct kmem_cache *s)
3369 struct kmem_cache *new = NULL;
3370 static char *tmp_path = NULL, *tmp_name = NULL;
3371 static DEFINE_MUTEX(mutex); /* protects tmp_name */
3373 BUG_ON(!memcg_can_account_kmem(memcg));
3377 * kmem_cache_create_memcg duplicates the given name and
3378 * cgroup_name for this name requires RCU context.
3379 * This static temporary buffer is used to prevent from
3380 * pointless shortliving allocation.
3382 if (!tmp_path || !tmp_name) {
3384 tmp_path = kmalloc(PATH_MAX, GFP_KERNEL);
3386 tmp_name = kmalloc(NAME_MAX + 1, GFP_KERNEL);
3387 if (!tmp_path || !tmp_name)
3391 cgroup_name(memcg->css.cgroup, tmp_name, NAME_MAX + 1);
3392 snprintf(tmp_path, PATH_MAX, "%s(%d:%s)", s->name,
3393 memcg_cache_id(memcg), tmp_name);
3395 new = kmem_cache_create_memcg(memcg, tmp_path, s->object_size, s->align,
3396 (s->flags & ~SLAB_PANIC), s->ctor, s);
3398 new->allocflags |= __GFP_KMEMCG;
3402 mutex_unlock(&mutex);
3406 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3408 struct kmem_cache *c;
3411 if (!s->memcg_params)
3413 if (!s->memcg_params->is_root_cache)
3417 * If the cache is being destroyed, we trust that there is no one else
3418 * requesting objects from it. Even if there are, the sanity checks in
3419 * kmem_cache_destroy should caught this ill-case.
3421 * Still, we don't want anyone else freeing memcg_caches under our
3422 * noses, which can happen if a new memcg comes to life. As usual,
3423 * we'll take the activate_kmem_mutex to protect ourselves against
3426 mutex_lock(&activate_kmem_mutex);
3427 for_each_memcg_cache_index(i) {
3428 c = cache_from_memcg_idx(s, i);
3433 * We will now manually delete the caches, so to avoid races
3434 * we need to cancel all pending destruction workers and
3435 * proceed with destruction ourselves.
3437 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3438 * and that could spawn the workers again: it is likely that
3439 * the cache still have active pages until this very moment.
3440 * This would lead us back to mem_cgroup_destroy_cache.
3442 * But that will not execute at all if the "dead" flag is not
3443 * set, so flip it down to guarantee we are in control.
3445 c->memcg_params->dead = false;
3446 cancel_work_sync(&c->memcg_params->destroy);
3447 kmem_cache_destroy(c);
3449 mutex_unlock(&activate_kmem_mutex);
3452 struct create_work {
3453 struct mem_cgroup *memcg;
3454 struct kmem_cache *cachep;
3455 struct work_struct work;
3458 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3460 struct kmem_cache *cachep;
3461 struct memcg_cache_params *params;
3463 if (!memcg_kmem_is_active(memcg))
3466 mutex_lock(&memcg->slab_caches_mutex);
3467 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3468 cachep = memcg_params_to_cache(params);
3469 cachep->memcg_params->dead = true;
3470 schedule_work(&cachep->memcg_params->destroy);
3472 mutex_unlock(&memcg->slab_caches_mutex);
3475 static void memcg_create_cache_work_func(struct work_struct *w)
3477 struct create_work *cw;
3479 cw = container_of(w, struct create_work, work);
3480 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3481 css_put(&cw->memcg->css);
3486 * Enqueue the creation of a per-memcg kmem_cache.
3488 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3489 struct kmem_cache *cachep)
3491 struct create_work *cw;
3493 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3495 css_put(&memcg->css);
3500 cw->cachep = cachep;
3502 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3503 schedule_work(&cw->work);
3506 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3507 struct kmem_cache *cachep)
3510 * We need to stop accounting when we kmalloc, because if the
3511 * corresponding kmalloc cache is not yet created, the first allocation
3512 * in __memcg_create_cache_enqueue will recurse.
3514 * However, it is better to enclose the whole function. Depending on
3515 * the debugging options enabled, INIT_WORK(), for instance, can
3516 * trigger an allocation. This too, will make us recurse. Because at
3517 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3518 * the safest choice is to do it like this, wrapping the whole function.
3520 memcg_stop_kmem_account();
3521 __memcg_create_cache_enqueue(memcg, cachep);
3522 memcg_resume_kmem_account();
3525 * Return the kmem_cache we're supposed to use for a slab allocation.
3526 * We try to use the current memcg's version of the cache.
3528 * If the cache does not exist yet, if we are the first user of it,
3529 * we either create it immediately, if possible, or create it asynchronously
3531 * In the latter case, we will let the current allocation go through with
3532 * the original cache.
3534 * Can't be called in interrupt context or from kernel threads.
3535 * This function needs to be called with rcu_read_lock() held.
3537 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3540 struct mem_cgroup *memcg;
3541 struct kmem_cache *memcg_cachep;
3543 VM_BUG_ON(!cachep->memcg_params);
3544 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3546 if (!current->mm || current->memcg_kmem_skip_account)
3550 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3552 if (!memcg_can_account_kmem(memcg))
3555 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3556 if (likely(memcg_cachep)) {
3557 cachep = memcg_cachep;
3561 /* The corresponding put will be done in the workqueue. */
3562 if (!css_tryget(&memcg->css))
3567 * If we are in a safe context (can wait, and not in interrupt
3568 * context), we could be be predictable and return right away.
3569 * This would guarantee that the allocation being performed
3570 * already belongs in the new cache.
3572 * However, there are some clashes that can arrive from locking.
3573 * For instance, because we acquire the slab_mutex while doing
3574 * kmem_cache_dup, this means no further allocation could happen
3575 * with the slab_mutex held.
3577 * Also, because cache creation issue get_online_cpus(), this
3578 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3579 * that ends up reversed during cpu hotplug. (cpuset allocates
3580 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3581 * better to defer everything.
3583 memcg_create_cache_enqueue(memcg, cachep);
3589 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3592 * We need to verify if the allocation against current->mm->owner's memcg is
3593 * possible for the given order. But the page is not allocated yet, so we'll
3594 * need a further commit step to do the final arrangements.
3596 * It is possible for the task to switch cgroups in this mean time, so at
3597 * commit time, we can't rely on task conversion any longer. We'll then use
3598 * the handle argument to return to the caller which cgroup we should commit
3599 * against. We could also return the memcg directly and avoid the pointer
3600 * passing, but a boolean return value gives better semantics considering
3601 * the compiled-out case as well.
3603 * Returning true means the allocation is possible.
3606 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3608 struct mem_cgroup *memcg;
3614 * Disabling accounting is only relevant for some specific memcg
3615 * internal allocations. Therefore we would initially not have such
3616 * check here, since direct calls to the page allocator that are marked
3617 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3618 * concerned with cache allocations, and by having this test at
3619 * memcg_kmem_get_cache, we are already able to relay the allocation to
3620 * the root cache and bypass the memcg cache altogether.
3622 * There is one exception, though: the SLUB allocator does not create
3623 * large order caches, but rather service large kmallocs directly from
3624 * the page allocator. Therefore, the following sequence when backed by
3625 * the SLUB allocator:
3627 * memcg_stop_kmem_account();
3628 * kmalloc(<large_number>)
3629 * memcg_resume_kmem_account();
3631 * would effectively ignore the fact that we should skip accounting,
3632 * since it will drive us directly to this function without passing
3633 * through the cache selector memcg_kmem_get_cache. Such large
3634 * allocations are extremely rare but can happen, for instance, for the
3635 * cache arrays. We bring this test here.
3637 if (!current->mm || current->memcg_kmem_skip_account)
3640 memcg = try_get_mem_cgroup_from_mm(current->mm);
3643 * very rare case described in mem_cgroup_from_task. Unfortunately there
3644 * isn't much we can do without complicating this too much, and it would
3645 * be gfp-dependent anyway. Just let it go
3647 if (unlikely(!memcg))
3650 if (!memcg_can_account_kmem(memcg)) {
3651 css_put(&memcg->css);
3655 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3659 css_put(&memcg->css);
3663 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3666 struct page_cgroup *pc;
3668 VM_BUG_ON(mem_cgroup_is_root(memcg));
3670 /* The page allocation failed. Revert */
3672 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3676 pc = lookup_page_cgroup(page);
3677 lock_page_cgroup(pc);
3678 pc->mem_cgroup = memcg;
3679 SetPageCgroupUsed(pc);
3680 unlock_page_cgroup(pc);
3683 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3685 struct mem_cgroup *memcg = NULL;
3686 struct page_cgroup *pc;
3689 pc = lookup_page_cgroup(page);
3691 * Fast unlocked return. Theoretically might have changed, have to
3692 * check again after locking.
3694 if (!PageCgroupUsed(pc))
3697 lock_page_cgroup(pc);
3698 if (PageCgroupUsed(pc)) {
3699 memcg = pc->mem_cgroup;
3700 ClearPageCgroupUsed(pc);
3702 unlock_page_cgroup(pc);
3705 * We trust that only if there is a memcg associated with the page, it
3706 * is a valid allocation
3711 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3712 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3715 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3718 #endif /* CONFIG_MEMCG_KMEM */
3720 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3722 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3724 * Because tail pages are not marked as "used", set it. We're under
3725 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3726 * charge/uncharge will be never happen and move_account() is done under
3727 * compound_lock(), so we don't have to take care of races.
3729 void mem_cgroup_split_huge_fixup(struct page *head)
3731 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3732 struct page_cgroup *pc;
3733 struct mem_cgroup *memcg;
3736 if (mem_cgroup_disabled())
3739 memcg = head_pc->mem_cgroup;
3740 for (i = 1; i < HPAGE_PMD_NR; i++) {
3742 pc->mem_cgroup = memcg;
3743 smp_wmb();/* see __commit_charge() */
3744 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3746 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3749 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3752 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3753 struct mem_cgroup *to,
3754 unsigned int nr_pages,
3755 enum mem_cgroup_stat_index idx)
3757 /* Update stat data for mem_cgroup */
3759 __this_cpu_sub(from->stat->count[idx], nr_pages);
3760 __this_cpu_add(to->stat->count[idx], nr_pages);
3765 * mem_cgroup_move_account - move account of the page
3767 * @nr_pages: number of regular pages (>1 for huge pages)
3768 * @pc: page_cgroup of the page.
3769 * @from: mem_cgroup which the page is moved from.
3770 * @to: mem_cgroup which the page is moved to. @from != @to.
3772 * The caller must confirm following.
3773 * - page is not on LRU (isolate_page() is useful.)
3774 * - compound_lock is held when nr_pages > 1
3776 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3779 static int mem_cgroup_move_account(struct page *page,
3780 unsigned int nr_pages,
3781 struct page_cgroup *pc,
3782 struct mem_cgroup *from,
3783 struct mem_cgroup *to)
3785 unsigned long flags;
3787 bool anon = PageAnon(page);
3789 VM_BUG_ON(from == to);
3790 VM_BUG_ON_PAGE(PageLRU(page), page);
3792 * The page is isolated from LRU. So, collapse function
3793 * will not handle this page. But page splitting can happen.
3794 * Do this check under compound_page_lock(). The caller should
3798 if (nr_pages > 1 && !PageTransHuge(page))
3801 lock_page_cgroup(pc);
3804 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3807 move_lock_mem_cgroup(from, &flags);
3809 if (!anon && page_mapped(page))
3810 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3811 MEM_CGROUP_STAT_FILE_MAPPED);
3813 if (PageWriteback(page))
3814 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3815 MEM_CGROUP_STAT_WRITEBACK);
3817 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3819 /* caller should have done css_get */
3820 pc->mem_cgroup = to;
3821 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3822 move_unlock_mem_cgroup(from, &flags);
3825 unlock_page_cgroup(pc);
3829 memcg_check_events(to, page);
3830 memcg_check_events(from, page);
3836 * mem_cgroup_move_parent - moves page to the parent group
3837 * @page: the page to move
3838 * @pc: page_cgroup of the page
3839 * @child: page's cgroup
3841 * move charges to its parent or the root cgroup if the group has no
3842 * parent (aka use_hierarchy==0).
3843 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3844 * mem_cgroup_move_account fails) the failure is always temporary and
3845 * it signals a race with a page removal/uncharge or migration. In the
3846 * first case the page is on the way out and it will vanish from the LRU
3847 * on the next attempt and the call should be retried later.
3848 * Isolation from the LRU fails only if page has been isolated from
3849 * the LRU since we looked at it and that usually means either global
3850 * reclaim or migration going on. The page will either get back to the
3852 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3853 * (!PageCgroupUsed) or moved to a different group. The page will
3854 * disappear in the next attempt.
3856 static int mem_cgroup_move_parent(struct page *page,
3857 struct page_cgroup *pc,
3858 struct mem_cgroup *child)
3860 struct mem_cgroup *parent;
3861 unsigned int nr_pages;
3862 unsigned long uninitialized_var(flags);
3865 VM_BUG_ON(mem_cgroup_is_root(child));
3868 if (!get_page_unless_zero(page))
3870 if (isolate_lru_page(page))
3873 nr_pages = hpage_nr_pages(page);
3875 parent = parent_mem_cgroup(child);
3877 * If no parent, move charges to root cgroup.
3880 parent = root_mem_cgroup;
3883 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3884 flags = compound_lock_irqsave(page);
3887 ret = mem_cgroup_move_account(page, nr_pages,
3890 __mem_cgroup_cancel_local_charge(child, nr_pages);
3893 compound_unlock_irqrestore(page, flags);
3894 putback_lru_page(page);
3902 * Charge the memory controller for page usage.
3904 * 0 if the charge was successful
3905 * < 0 if the cgroup is over its limit
3907 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3908 gfp_t gfp_mask, enum charge_type ctype)
3910 struct mem_cgroup *memcg = NULL;
3911 unsigned int nr_pages = 1;
3915 if (PageTransHuge(page)) {
3916 nr_pages <<= compound_order(page);
3917 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3919 * Never OOM-kill a process for a huge page. The
3920 * fault handler will fall back to regular pages.
3925 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3928 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3932 int mem_cgroup_newpage_charge(struct page *page,
3933 struct mm_struct *mm, gfp_t gfp_mask)
3935 if (mem_cgroup_disabled())
3937 VM_BUG_ON_PAGE(page_mapped(page), page);
3938 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3940 return mem_cgroup_charge_common(page, mm, gfp_mask,
3941 MEM_CGROUP_CHARGE_TYPE_ANON);
3945 * While swap-in, try_charge -> commit or cancel, the page is locked.
3946 * And when try_charge() successfully returns, one refcnt to memcg without
3947 * struct page_cgroup is acquired. This refcnt will be consumed by
3948 * "commit()" or removed by "cancel()"
3950 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3953 struct mem_cgroup **memcgp)
3955 struct mem_cgroup *memcg;
3956 struct page_cgroup *pc;
3959 pc = lookup_page_cgroup(page);
3961 * Every swap fault against a single page tries to charge the
3962 * page, bail as early as possible. shmem_unuse() encounters
3963 * already charged pages, too. The USED bit is protected by
3964 * the page lock, which serializes swap cache removal, which
3965 * in turn serializes uncharging.
3967 if (PageCgroupUsed(pc))
3969 if (!do_swap_account)
3971 memcg = try_get_mem_cgroup_from_page(page);
3975 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3976 css_put(&memcg->css);
3981 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3987 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3988 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3991 if (mem_cgroup_disabled())
3994 * A racing thread's fault, or swapoff, may have already
3995 * updated the pte, and even removed page from swap cache: in
3996 * those cases unuse_pte()'s pte_same() test will fail; but
3997 * there's also a KSM case which does need to charge the page.
3999 if (!PageSwapCache(page)) {
4002 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4007 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4010 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4012 if (mem_cgroup_disabled())
4016 __mem_cgroup_cancel_charge(memcg, 1);
4020 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4021 enum charge_type ctype)
4023 if (mem_cgroup_disabled())
4028 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4030 * Now swap is on-memory. This means this page may be
4031 * counted both as mem and swap....double count.
4032 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4033 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4034 * may call delete_from_swap_cache() before reach here.
4036 if (do_swap_account && PageSwapCache(page)) {
4037 swp_entry_t ent = {.val = page_private(page)};
4038 mem_cgroup_uncharge_swap(ent);
4042 void mem_cgroup_commit_charge_swapin(struct page *page,
4043 struct mem_cgroup *memcg)
4045 __mem_cgroup_commit_charge_swapin(page, memcg,
4046 MEM_CGROUP_CHARGE_TYPE_ANON);
4049 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4052 struct mem_cgroup *memcg = NULL;
4053 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4056 if (mem_cgroup_disabled())
4058 if (PageCompound(page))
4061 if (!PageSwapCache(page))
4062 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4063 else { /* page is swapcache/shmem */
4064 ret = __mem_cgroup_try_charge_swapin(mm, page,
4067 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4072 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4073 unsigned int nr_pages,
4074 const enum charge_type ctype)
4076 struct memcg_batch_info *batch = NULL;
4077 bool uncharge_memsw = true;
4079 /* If swapout, usage of swap doesn't decrease */
4080 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4081 uncharge_memsw = false;
4083 batch = ¤t->memcg_batch;
4085 * In usual, we do css_get() when we remember memcg pointer.
4086 * But in this case, we keep res->usage until end of a series of
4087 * uncharges. Then, it's ok to ignore memcg's refcnt.
4090 batch->memcg = memcg;
4092 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4093 * In those cases, all pages freed continuously can be expected to be in
4094 * the same cgroup and we have chance to coalesce uncharges.
4095 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4096 * because we want to do uncharge as soon as possible.
4099 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4100 goto direct_uncharge;
4103 goto direct_uncharge;
4106 * In typical case, batch->memcg == mem. This means we can
4107 * merge a series of uncharges to an uncharge of res_counter.
4108 * If not, we uncharge res_counter ony by one.
4110 if (batch->memcg != memcg)
4111 goto direct_uncharge;
4112 /* remember freed charge and uncharge it later */
4115 batch->memsw_nr_pages++;
4118 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4120 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4121 if (unlikely(batch->memcg != memcg))
4122 memcg_oom_recover(memcg);
4126 * uncharge if !page_mapped(page)
4128 static struct mem_cgroup *
4129 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4132 struct mem_cgroup *memcg = NULL;
4133 unsigned int nr_pages = 1;
4134 struct page_cgroup *pc;
4137 if (mem_cgroup_disabled())
4140 if (PageTransHuge(page)) {
4141 nr_pages <<= compound_order(page);
4142 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
4145 * Check if our page_cgroup is valid
4147 pc = lookup_page_cgroup(page);
4148 if (unlikely(!PageCgroupUsed(pc)))
4151 lock_page_cgroup(pc);
4153 memcg = pc->mem_cgroup;
4155 if (!PageCgroupUsed(pc))
4158 anon = PageAnon(page);
4161 case MEM_CGROUP_CHARGE_TYPE_ANON:
4163 * Generally PageAnon tells if it's the anon statistics to be
4164 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4165 * used before page reached the stage of being marked PageAnon.
4169 case MEM_CGROUP_CHARGE_TYPE_DROP:
4170 /* See mem_cgroup_prepare_migration() */
4171 if (page_mapped(page))
4174 * Pages under migration may not be uncharged. But
4175 * end_migration() /must/ be the one uncharging the
4176 * unused post-migration page and so it has to call
4177 * here with the migration bit still set. See the
4178 * res_counter handling below.
4180 if (!end_migration && PageCgroupMigration(pc))
4183 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4184 if (!PageAnon(page)) { /* Shared memory */
4185 if (page->mapping && !page_is_file_cache(page))
4187 } else if (page_mapped(page)) /* Anon */
4194 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4196 ClearPageCgroupUsed(pc);
4198 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4199 * freed from LRU. This is safe because uncharged page is expected not
4200 * to be reused (freed soon). Exception is SwapCache, it's handled by
4201 * special functions.
4204 unlock_page_cgroup(pc);
4206 * even after unlock, we have memcg->res.usage here and this memcg
4207 * will never be freed, so it's safe to call css_get().
4209 memcg_check_events(memcg, page);
4210 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4211 mem_cgroup_swap_statistics(memcg, true);
4212 css_get(&memcg->css);
4215 * Migration does not charge the res_counter for the
4216 * replacement page, so leave it alone when phasing out the
4217 * page that is unused after the migration.
4219 if (!end_migration && !mem_cgroup_is_root(memcg))
4220 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4225 unlock_page_cgroup(pc);
4229 void mem_cgroup_uncharge_page(struct page *page)
4232 if (page_mapped(page))
4234 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4236 * If the page is in swap cache, uncharge should be deferred
4237 * to the swap path, which also properly accounts swap usage
4238 * and handles memcg lifetime.
4240 * Note that this check is not stable and reclaim may add the
4241 * page to swap cache at any time after this. However, if the
4242 * page is not in swap cache by the time page->mapcount hits
4243 * 0, there won't be any page table references to the swap
4244 * slot, and reclaim will free it and not actually write the
4247 if (PageSwapCache(page))
4249 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4252 void mem_cgroup_uncharge_cache_page(struct page *page)
4254 VM_BUG_ON_PAGE(page_mapped(page), page);
4255 VM_BUG_ON_PAGE(page->mapping, page);
4256 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4260 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4261 * In that cases, pages are freed continuously and we can expect pages
4262 * are in the same memcg. All these calls itself limits the number of
4263 * pages freed at once, then uncharge_start/end() is called properly.
4264 * This may be called prural(2) times in a context,
4267 void mem_cgroup_uncharge_start(void)
4269 current->memcg_batch.do_batch++;
4270 /* We can do nest. */
4271 if (current->memcg_batch.do_batch == 1) {
4272 current->memcg_batch.memcg = NULL;
4273 current->memcg_batch.nr_pages = 0;
4274 current->memcg_batch.memsw_nr_pages = 0;
4278 void mem_cgroup_uncharge_end(void)
4280 struct memcg_batch_info *batch = ¤t->memcg_batch;
4282 if (!batch->do_batch)
4286 if (batch->do_batch) /* If stacked, do nothing. */
4292 * This "batch->memcg" is valid without any css_get/put etc...
4293 * bacause we hide charges behind us.
4295 if (batch->nr_pages)
4296 res_counter_uncharge(&batch->memcg->res,
4297 batch->nr_pages * PAGE_SIZE);
4298 if (batch->memsw_nr_pages)
4299 res_counter_uncharge(&batch->memcg->memsw,
4300 batch->memsw_nr_pages * PAGE_SIZE);
4301 memcg_oom_recover(batch->memcg);
4302 /* forget this pointer (for sanity check) */
4303 batch->memcg = NULL;
4308 * called after __delete_from_swap_cache() and drop "page" account.
4309 * memcg information is recorded to swap_cgroup of "ent"
4312 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4314 struct mem_cgroup *memcg;
4315 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4317 if (!swapout) /* this was a swap cache but the swap is unused ! */
4318 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4320 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4323 * record memcg information, if swapout && memcg != NULL,
4324 * css_get() was called in uncharge().
4326 if (do_swap_account && swapout && memcg)
4327 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4331 #ifdef CONFIG_MEMCG_SWAP
4333 * called from swap_entry_free(). remove record in swap_cgroup and
4334 * uncharge "memsw" account.
4336 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4338 struct mem_cgroup *memcg;
4341 if (!do_swap_account)
4344 id = swap_cgroup_record(ent, 0);
4346 memcg = mem_cgroup_lookup(id);
4349 * We uncharge this because swap is freed.
4350 * This memcg can be obsolete one. We avoid calling css_tryget
4352 if (!mem_cgroup_is_root(memcg))
4353 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4354 mem_cgroup_swap_statistics(memcg, false);
4355 css_put(&memcg->css);
4361 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4362 * @entry: swap entry to be moved
4363 * @from: mem_cgroup which the entry is moved from
4364 * @to: mem_cgroup which the entry is moved to
4366 * It succeeds only when the swap_cgroup's record for this entry is the same
4367 * as the mem_cgroup's id of @from.
4369 * Returns 0 on success, -EINVAL on failure.
4371 * The caller must have charged to @to, IOW, called res_counter_charge() about
4372 * both res and memsw, and called css_get().
4374 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4375 struct mem_cgroup *from, struct mem_cgroup *to)
4377 unsigned short old_id, new_id;
4379 old_id = mem_cgroup_id(from);
4380 new_id = mem_cgroup_id(to);
4382 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4383 mem_cgroup_swap_statistics(from, false);
4384 mem_cgroup_swap_statistics(to, true);
4386 * This function is only called from task migration context now.
4387 * It postpones res_counter and refcount handling till the end
4388 * of task migration(mem_cgroup_clear_mc()) for performance
4389 * improvement. But we cannot postpone css_get(to) because if
4390 * the process that has been moved to @to does swap-in, the
4391 * refcount of @to might be decreased to 0.
4393 * We are in attach() phase, so the cgroup is guaranteed to be
4394 * alive, so we can just call css_get().
4402 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4403 struct mem_cgroup *from, struct mem_cgroup *to)
4410 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4413 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4414 struct mem_cgroup **memcgp)
4416 struct mem_cgroup *memcg = NULL;
4417 unsigned int nr_pages = 1;
4418 struct page_cgroup *pc;
4419 enum charge_type ctype;
4423 if (mem_cgroup_disabled())
4426 if (PageTransHuge(page))
4427 nr_pages <<= compound_order(page);
4429 pc = lookup_page_cgroup(page);
4430 lock_page_cgroup(pc);
4431 if (PageCgroupUsed(pc)) {
4432 memcg = pc->mem_cgroup;
4433 css_get(&memcg->css);
4435 * At migrating an anonymous page, its mapcount goes down
4436 * to 0 and uncharge() will be called. But, even if it's fully
4437 * unmapped, migration may fail and this page has to be
4438 * charged again. We set MIGRATION flag here and delay uncharge
4439 * until end_migration() is called
4441 * Corner Case Thinking
4443 * When the old page was mapped as Anon and it's unmap-and-freed
4444 * while migration was ongoing.
4445 * If unmap finds the old page, uncharge() of it will be delayed
4446 * until end_migration(). If unmap finds a new page, it's
4447 * uncharged when it make mapcount to be 1->0. If unmap code
4448 * finds swap_migration_entry, the new page will not be mapped
4449 * and end_migration() will find it(mapcount==0).
4452 * When the old page was mapped but migraion fails, the kernel
4453 * remaps it. A charge for it is kept by MIGRATION flag even
4454 * if mapcount goes down to 0. We can do remap successfully
4455 * without charging it again.
4458 * The "old" page is under lock_page() until the end of
4459 * migration, so, the old page itself will not be swapped-out.
4460 * If the new page is swapped out before end_migraton, our
4461 * hook to usual swap-out path will catch the event.
4464 SetPageCgroupMigration(pc);
4466 unlock_page_cgroup(pc);
4468 * If the page is not charged at this point,
4476 * We charge new page before it's used/mapped. So, even if unlock_page()
4477 * is called before end_migration, we can catch all events on this new
4478 * page. In the case new page is migrated but not remapped, new page's
4479 * mapcount will be finally 0 and we call uncharge in end_migration().
4482 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4484 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4486 * The page is committed to the memcg, but it's not actually
4487 * charged to the res_counter since we plan on replacing the
4488 * old one and only one page is going to be left afterwards.
4490 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4493 /* remove redundant charge if migration failed*/
4494 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4495 struct page *oldpage, struct page *newpage, bool migration_ok)
4497 struct page *used, *unused;
4498 struct page_cgroup *pc;
4504 if (!migration_ok) {
4511 anon = PageAnon(used);
4512 __mem_cgroup_uncharge_common(unused,
4513 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4514 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4516 css_put(&memcg->css);
4518 * We disallowed uncharge of pages under migration because mapcount
4519 * of the page goes down to zero, temporarly.
4520 * Clear the flag and check the page should be charged.
4522 pc = lookup_page_cgroup(oldpage);
4523 lock_page_cgroup(pc);
4524 ClearPageCgroupMigration(pc);
4525 unlock_page_cgroup(pc);
4528 * If a page is a file cache, radix-tree replacement is very atomic
4529 * and we can skip this check. When it was an Anon page, its mapcount
4530 * goes down to 0. But because we added MIGRATION flage, it's not
4531 * uncharged yet. There are several case but page->mapcount check
4532 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4533 * check. (see prepare_charge() also)
4536 mem_cgroup_uncharge_page(used);
4540 * At replace page cache, newpage is not under any memcg but it's on
4541 * LRU. So, this function doesn't touch res_counter but handles LRU
4542 * in correct way. Both pages are locked so we cannot race with uncharge.
4544 void mem_cgroup_replace_page_cache(struct page *oldpage,
4545 struct page *newpage)
4547 struct mem_cgroup *memcg = NULL;
4548 struct page_cgroup *pc;
4549 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4551 if (mem_cgroup_disabled())
4554 pc = lookup_page_cgroup(oldpage);
4555 /* fix accounting on old pages */
4556 lock_page_cgroup(pc);
4557 if (PageCgroupUsed(pc)) {
4558 memcg = pc->mem_cgroup;
4559 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4560 ClearPageCgroupUsed(pc);
4562 unlock_page_cgroup(pc);
4565 * When called from shmem_replace_page(), in some cases the
4566 * oldpage has already been charged, and in some cases not.
4571 * Even if newpage->mapping was NULL before starting replacement,
4572 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4573 * LRU while we overwrite pc->mem_cgroup.
4575 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4578 #ifdef CONFIG_DEBUG_VM
4579 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4581 struct page_cgroup *pc;
4583 pc = lookup_page_cgroup(page);
4585 * Can be NULL while feeding pages into the page allocator for
4586 * the first time, i.e. during boot or memory hotplug;
4587 * or when mem_cgroup_disabled().
4589 if (likely(pc) && PageCgroupUsed(pc))
4594 bool mem_cgroup_bad_page_check(struct page *page)
4596 if (mem_cgroup_disabled())
4599 return lookup_page_cgroup_used(page) != NULL;
4602 void mem_cgroup_print_bad_page(struct page *page)
4604 struct page_cgroup *pc;
4606 pc = lookup_page_cgroup_used(page);
4608 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4609 pc, pc->flags, pc->mem_cgroup);
4614 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4615 unsigned long long val)
4618 u64 memswlimit, memlimit;
4620 int children = mem_cgroup_count_children(memcg);
4621 u64 curusage, oldusage;
4625 * For keeping hierarchical_reclaim simple, how long we should retry
4626 * is depends on callers. We set our retry-count to be function
4627 * of # of children which we should visit in this loop.
4629 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4631 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4634 while (retry_count) {
4635 if (signal_pending(current)) {
4640 * Rather than hide all in some function, I do this in
4641 * open coded manner. You see what this really does.
4642 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4644 mutex_lock(&set_limit_mutex);
4645 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4646 if (memswlimit < val) {
4648 mutex_unlock(&set_limit_mutex);
4652 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4656 ret = res_counter_set_limit(&memcg->res, val);
4658 if (memswlimit == val)
4659 memcg->memsw_is_minimum = true;
4661 memcg->memsw_is_minimum = false;
4663 mutex_unlock(&set_limit_mutex);
4668 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4669 MEM_CGROUP_RECLAIM_SHRINK);
4670 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4671 /* Usage is reduced ? */
4672 if (curusage >= oldusage)
4675 oldusage = curusage;
4677 if (!ret && enlarge)
4678 memcg_oom_recover(memcg);
4683 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4684 unsigned long long val)
4687 u64 memlimit, memswlimit, oldusage, curusage;
4688 int children = mem_cgroup_count_children(memcg);
4692 /* see mem_cgroup_resize_res_limit */
4693 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4694 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4695 while (retry_count) {
4696 if (signal_pending(current)) {
4701 * Rather than hide all in some function, I do this in
4702 * open coded manner. You see what this really does.
4703 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4705 mutex_lock(&set_limit_mutex);
4706 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4707 if (memlimit > val) {
4709 mutex_unlock(&set_limit_mutex);
4712 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4713 if (memswlimit < val)
4715 ret = res_counter_set_limit(&memcg->memsw, val);
4717 if (memlimit == val)
4718 memcg->memsw_is_minimum = true;
4720 memcg->memsw_is_minimum = false;
4722 mutex_unlock(&set_limit_mutex);
4727 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4728 MEM_CGROUP_RECLAIM_NOSWAP |
4729 MEM_CGROUP_RECLAIM_SHRINK);
4730 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4731 /* Usage is reduced ? */
4732 if (curusage >= oldusage)
4735 oldusage = curusage;
4737 if (!ret && enlarge)
4738 memcg_oom_recover(memcg);
4742 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4744 unsigned long *total_scanned)
4746 unsigned long nr_reclaimed = 0;
4747 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4748 unsigned long reclaimed;
4750 struct mem_cgroup_tree_per_zone *mctz;
4751 unsigned long long excess;
4752 unsigned long nr_scanned;
4757 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4759 * This loop can run a while, specially if mem_cgroup's continuously
4760 * keep exceeding their soft limit and putting the system under
4767 mz = mem_cgroup_largest_soft_limit_node(mctz);
4772 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4773 gfp_mask, &nr_scanned);
4774 nr_reclaimed += reclaimed;
4775 *total_scanned += nr_scanned;
4776 spin_lock(&mctz->lock);
4779 * If we failed to reclaim anything from this memory cgroup
4780 * it is time to move on to the next cgroup
4786 * Loop until we find yet another one.
4788 * By the time we get the soft_limit lock
4789 * again, someone might have aded the
4790 * group back on the RB tree. Iterate to
4791 * make sure we get a different mem.
4792 * mem_cgroup_largest_soft_limit_node returns
4793 * NULL if no other cgroup is present on
4797 __mem_cgroup_largest_soft_limit_node(mctz);
4799 css_put(&next_mz->memcg->css);
4800 else /* next_mz == NULL or other memcg */
4804 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4805 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4807 * One school of thought says that we should not add
4808 * back the node to the tree if reclaim returns 0.
4809 * But our reclaim could return 0, simply because due
4810 * to priority we are exposing a smaller subset of
4811 * memory to reclaim from. Consider this as a longer
4814 /* If excess == 0, no tree ops */
4815 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4816 spin_unlock(&mctz->lock);
4817 css_put(&mz->memcg->css);
4820 * Could not reclaim anything and there are no more
4821 * mem cgroups to try or we seem to be looping without
4822 * reclaiming anything.
4824 if (!nr_reclaimed &&
4826 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4828 } while (!nr_reclaimed);
4830 css_put(&next_mz->memcg->css);
4831 return nr_reclaimed;
4835 * mem_cgroup_force_empty_list - clears LRU of a group
4836 * @memcg: group to clear
4839 * @lru: lru to to clear
4841 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4842 * reclaim the pages page themselves - pages are moved to the parent (or root)
4845 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4846 int node, int zid, enum lru_list lru)
4848 struct lruvec *lruvec;
4849 unsigned long flags;
4850 struct list_head *list;
4854 zone = &NODE_DATA(node)->node_zones[zid];
4855 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4856 list = &lruvec->lists[lru];
4860 struct page_cgroup *pc;
4863 spin_lock_irqsave(&zone->lru_lock, flags);
4864 if (list_empty(list)) {
4865 spin_unlock_irqrestore(&zone->lru_lock, flags);
4868 page = list_entry(list->prev, struct page, lru);
4870 list_move(&page->lru, list);
4872 spin_unlock_irqrestore(&zone->lru_lock, flags);
4875 spin_unlock_irqrestore(&zone->lru_lock, flags);
4877 pc = lookup_page_cgroup(page);
4879 if (mem_cgroup_move_parent(page, pc, memcg)) {
4880 /* found lock contention or "pc" is obsolete. */
4885 } while (!list_empty(list));
4889 * make mem_cgroup's charge to be 0 if there is no task by moving
4890 * all the charges and pages to the parent.
4891 * This enables deleting this mem_cgroup.
4893 * Caller is responsible for holding css reference on the memcg.
4895 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4901 /* This is for making all *used* pages to be on LRU. */
4902 lru_add_drain_all();
4903 drain_all_stock_sync(memcg);
4904 mem_cgroup_start_move(memcg);
4905 for_each_node_state(node, N_MEMORY) {
4906 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4909 mem_cgroup_force_empty_list(memcg,
4914 mem_cgroup_end_move(memcg);
4915 memcg_oom_recover(memcg);
4919 * Kernel memory may not necessarily be trackable to a specific
4920 * process. So they are not migrated, and therefore we can't
4921 * expect their value to drop to 0 here.
4922 * Having res filled up with kmem only is enough.
4924 * This is a safety check because mem_cgroup_force_empty_list
4925 * could have raced with mem_cgroup_replace_page_cache callers
4926 * so the lru seemed empty but the page could have been added
4927 * right after the check. RES_USAGE should be safe as we always
4928 * charge before adding to the LRU.
4930 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4931 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4932 } while (usage > 0);
4935 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4937 lockdep_assert_held(&memcg_create_mutex);
4939 * The lock does not prevent addition or deletion to the list
4940 * of children, but it prevents a new child from being
4941 * initialized based on this parent in css_online(), so it's
4942 * enough to decide whether hierarchically inherited
4943 * attributes can still be changed or not.
4945 return memcg->use_hierarchy &&
4946 !list_empty(&memcg->css.cgroup->children);
4950 * Reclaims as many pages from the given memcg as possible and moves
4951 * the rest to the parent.
4953 * Caller is responsible for holding css reference for memcg.
4955 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4957 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4958 struct cgroup *cgrp = memcg->css.cgroup;
4960 /* returns EBUSY if there is a task or if we come here twice. */
4961 if (cgroup_has_tasks(cgrp) || !list_empty(&cgrp->children))
4964 /* we call try-to-free pages for make this cgroup empty */
4965 lru_add_drain_all();
4966 /* try to free all pages in this cgroup */
4967 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4970 if (signal_pending(current))
4973 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4977 /* maybe some writeback is necessary */
4978 congestion_wait(BLK_RW_ASYNC, HZ/10);
4983 mem_cgroup_reparent_charges(memcg);
4988 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4991 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4993 if (mem_cgroup_is_root(memcg))
4995 return mem_cgroup_force_empty(memcg);
4998 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5001 return mem_cgroup_from_css(css)->use_hierarchy;
5004 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5005 struct cftype *cft, u64 val)
5008 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5009 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5011 mutex_lock(&memcg_create_mutex);
5013 if (memcg->use_hierarchy == val)
5017 * If parent's use_hierarchy is set, we can't make any modifications
5018 * in the child subtrees. If it is unset, then the change can
5019 * occur, provided the current cgroup has no children.
5021 * For the root cgroup, parent_mem is NULL, we allow value to be
5022 * set if there are no children.
5024 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5025 (val == 1 || val == 0)) {
5026 if (list_empty(&memcg->css.cgroup->children))
5027 memcg->use_hierarchy = val;
5034 mutex_unlock(&memcg_create_mutex);
5040 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5041 enum mem_cgroup_stat_index idx)
5043 struct mem_cgroup *iter;
5046 /* Per-cpu values can be negative, use a signed accumulator */
5047 for_each_mem_cgroup_tree(iter, memcg)
5048 val += mem_cgroup_read_stat(iter, idx);
5050 if (val < 0) /* race ? */
5055 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5059 if (!mem_cgroup_is_root(memcg)) {
5061 return res_counter_read_u64(&memcg->res, RES_USAGE);
5063 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5067 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5068 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5070 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5071 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5074 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5076 return val << PAGE_SHIFT;
5079 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
5082 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5087 type = MEMFILE_TYPE(cft->private);
5088 name = MEMFILE_ATTR(cft->private);
5092 if (name == RES_USAGE)
5093 val = mem_cgroup_usage(memcg, false);
5095 val = res_counter_read_u64(&memcg->res, name);
5098 if (name == RES_USAGE)
5099 val = mem_cgroup_usage(memcg, true);
5101 val = res_counter_read_u64(&memcg->memsw, name);
5104 val = res_counter_read_u64(&memcg->kmem, name);
5113 #ifdef CONFIG_MEMCG_KMEM
5114 /* should be called with activate_kmem_mutex held */
5115 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
5116 unsigned long long limit)
5121 if (memcg_kmem_is_active(memcg))
5125 * We are going to allocate memory for data shared by all memory
5126 * cgroups so let's stop accounting here.
5128 memcg_stop_kmem_account();
5131 * For simplicity, we won't allow this to be disabled. It also can't
5132 * be changed if the cgroup has children already, or if tasks had
5135 * If tasks join before we set the limit, a person looking at
5136 * kmem.usage_in_bytes will have no way to determine when it took
5137 * place, which makes the value quite meaningless.
5139 * After it first became limited, changes in the value of the limit are
5140 * of course permitted.
5142 mutex_lock(&memcg_create_mutex);
5143 if (cgroup_has_tasks(memcg->css.cgroup) || memcg_has_children(memcg))
5145 mutex_unlock(&memcg_create_mutex);
5149 memcg_id = ida_simple_get(&kmem_limited_groups,
5150 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
5157 * Make sure we have enough space for this cgroup in each root cache's
5160 err = memcg_update_all_caches(memcg_id + 1);
5164 memcg->kmemcg_id = memcg_id;
5165 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
5166 mutex_init(&memcg->slab_caches_mutex);
5169 * We couldn't have accounted to this cgroup, because it hasn't got the
5170 * active bit set yet, so this should succeed.
5172 err = res_counter_set_limit(&memcg->kmem, limit);
5175 static_key_slow_inc(&memcg_kmem_enabled_key);
5177 * Setting the active bit after enabling static branching will
5178 * guarantee no one starts accounting before all call sites are
5181 memcg_kmem_set_active(memcg);
5183 memcg_resume_kmem_account();
5187 ida_simple_remove(&kmem_limited_groups, memcg_id);
5191 static int memcg_activate_kmem(struct mem_cgroup *memcg,
5192 unsigned long long limit)
5196 mutex_lock(&activate_kmem_mutex);
5197 ret = __memcg_activate_kmem(memcg, limit);
5198 mutex_unlock(&activate_kmem_mutex);
5202 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5203 unsigned long long val)
5207 if (!memcg_kmem_is_active(memcg))
5208 ret = memcg_activate_kmem(memcg, val);
5210 ret = res_counter_set_limit(&memcg->kmem, val);
5214 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5217 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5222 mutex_lock(&activate_kmem_mutex);
5224 * If the parent cgroup is not kmem-active now, it cannot be activated
5225 * after this point, because it has at least one child already.
5227 if (memcg_kmem_is_active(parent))
5228 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
5229 mutex_unlock(&activate_kmem_mutex);
5233 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5234 unsigned long long val)
5238 #endif /* CONFIG_MEMCG_KMEM */
5241 * The user of this function is...
5244 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5247 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5250 unsigned long long val;
5253 type = MEMFILE_TYPE(cft->private);
5254 name = MEMFILE_ATTR(cft->private);
5258 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5262 /* This function does all necessary parse...reuse it */
5263 ret = res_counter_memparse_write_strategy(buffer, &val);
5267 ret = mem_cgroup_resize_limit(memcg, val);
5268 else if (type == _MEMSWAP)
5269 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5270 else if (type == _KMEM)
5271 ret = memcg_update_kmem_limit(memcg, val);
5275 case RES_SOFT_LIMIT:
5276 ret = res_counter_memparse_write_strategy(buffer, &val);
5280 * For memsw, soft limits are hard to implement in terms
5281 * of semantics, for now, we support soft limits for
5282 * control without swap
5285 ret = res_counter_set_soft_limit(&memcg->res, val);
5290 ret = -EINVAL; /* should be BUG() ? */
5296 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5297 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5299 unsigned long long min_limit, min_memsw_limit, tmp;
5301 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5302 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5303 if (!memcg->use_hierarchy)
5306 while (css_parent(&memcg->css)) {
5307 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5308 if (!memcg->use_hierarchy)
5310 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5311 min_limit = min(min_limit, tmp);
5312 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5313 min_memsw_limit = min(min_memsw_limit, tmp);
5316 *mem_limit = min_limit;
5317 *memsw_limit = min_memsw_limit;
5320 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5322 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5326 type = MEMFILE_TYPE(event);
5327 name = MEMFILE_ATTR(event);
5332 res_counter_reset_max(&memcg->res);
5333 else if (type == _MEMSWAP)
5334 res_counter_reset_max(&memcg->memsw);
5335 else if (type == _KMEM)
5336 res_counter_reset_max(&memcg->kmem);
5342 res_counter_reset_failcnt(&memcg->res);
5343 else if (type == _MEMSWAP)
5344 res_counter_reset_failcnt(&memcg->memsw);
5345 else if (type == _KMEM)
5346 res_counter_reset_failcnt(&memcg->kmem);
5355 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5358 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5362 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5363 struct cftype *cft, u64 val)
5365 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5367 if (val >= (1 << NR_MOVE_TYPE))
5371 * No kind of locking is needed in here, because ->can_attach() will
5372 * check this value once in the beginning of the process, and then carry
5373 * on with stale data. This means that changes to this value will only
5374 * affect task migrations starting after the change.
5376 memcg->move_charge_at_immigrate = val;
5380 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5381 struct cftype *cft, u64 val)
5388 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5392 unsigned int lru_mask;
5395 static const struct numa_stat stats[] = {
5396 { "total", LRU_ALL },
5397 { "file", LRU_ALL_FILE },
5398 { "anon", LRU_ALL_ANON },
5399 { "unevictable", BIT(LRU_UNEVICTABLE) },
5401 const struct numa_stat *stat;
5404 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5406 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5407 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5408 seq_printf(m, "%s=%lu", stat->name, nr);
5409 for_each_node_state(nid, N_MEMORY) {
5410 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5412 seq_printf(m, " N%d=%lu", nid, nr);
5417 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5418 struct mem_cgroup *iter;
5421 for_each_mem_cgroup_tree(iter, memcg)
5422 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5423 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5424 for_each_node_state(nid, N_MEMORY) {
5426 for_each_mem_cgroup_tree(iter, memcg)
5427 nr += mem_cgroup_node_nr_lru_pages(
5428 iter, nid, stat->lru_mask);
5429 seq_printf(m, " N%d=%lu", nid, nr);
5436 #endif /* CONFIG_NUMA */
5438 static inline void mem_cgroup_lru_names_not_uptodate(void)
5440 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5443 static int memcg_stat_show(struct seq_file *m, void *v)
5445 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5446 struct mem_cgroup *mi;
5449 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5450 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5452 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5453 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5456 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5457 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5458 mem_cgroup_read_events(memcg, i));
5460 for (i = 0; i < NR_LRU_LISTS; i++)
5461 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5462 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5464 /* Hierarchical information */
5466 unsigned long long limit, memsw_limit;
5467 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5468 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5469 if (do_swap_account)
5470 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5474 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5477 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5479 for_each_mem_cgroup_tree(mi, memcg)
5480 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5481 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5484 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5485 unsigned long long val = 0;
5487 for_each_mem_cgroup_tree(mi, memcg)
5488 val += mem_cgroup_read_events(mi, i);
5489 seq_printf(m, "total_%s %llu\n",
5490 mem_cgroup_events_names[i], val);
5493 for (i = 0; i < NR_LRU_LISTS; i++) {
5494 unsigned long long val = 0;
5496 for_each_mem_cgroup_tree(mi, memcg)
5497 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5498 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5501 #ifdef CONFIG_DEBUG_VM
5504 struct mem_cgroup_per_zone *mz;
5505 struct zone_reclaim_stat *rstat;
5506 unsigned long recent_rotated[2] = {0, 0};
5507 unsigned long recent_scanned[2] = {0, 0};
5509 for_each_online_node(nid)
5510 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5511 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5512 rstat = &mz->lruvec.reclaim_stat;
5514 recent_rotated[0] += rstat->recent_rotated[0];
5515 recent_rotated[1] += rstat->recent_rotated[1];
5516 recent_scanned[0] += rstat->recent_scanned[0];
5517 recent_scanned[1] += rstat->recent_scanned[1];
5519 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5520 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5521 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5522 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5529 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5532 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5534 return mem_cgroup_swappiness(memcg);
5537 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5538 struct cftype *cft, u64 val)
5540 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5541 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5543 if (val > 100 || !parent)
5546 mutex_lock(&memcg_create_mutex);
5548 /* If under hierarchy, only empty-root can set this value */
5549 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5550 mutex_unlock(&memcg_create_mutex);
5554 memcg->swappiness = val;
5556 mutex_unlock(&memcg_create_mutex);
5561 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5563 struct mem_cgroup_threshold_ary *t;
5569 t = rcu_dereference(memcg->thresholds.primary);
5571 t = rcu_dereference(memcg->memsw_thresholds.primary);
5576 usage = mem_cgroup_usage(memcg, swap);
5579 * current_threshold points to threshold just below or equal to usage.
5580 * If it's not true, a threshold was crossed after last
5581 * call of __mem_cgroup_threshold().
5583 i = t->current_threshold;
5586 * Iterate backward over array of thresholds starting from
5587 * current_threshold and check if a threshold is crossed.
5588 * If none of thresholds below usage is crossed, we read
5589 * only one element of the array here.
5591 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5592 eventfd_signal(t->entries[i].eventfd, 1);
5594 /* i = current_threshold + 1 */
5598 * Iterate forward over array of thresholds starting from
5599 * current_threshold+1 and check if a threshold is crossed.
5600 * If none of thresholds above usage is crossed, we read
5601 * only one element of the array here.
5603 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5604 eventfd_signal(t->entries[i].eventfd, 1);
5606 /* Update current_threshold */
5607 t->current_threshold = i - 1;
5612 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5615 __mem_cgroup_threshold(memcg, false);
5616 if (do_swap_account)
5617 __mem_cgroup_threshold(memcg, true);
5619 memcg = parent_mem_cgroup(memcg);
5623 static int compare_thresholds(const void *a, const void *b)
5625 const struct mem_cgroup_threshold *_a = a;
5626 const struct mem_cgroup_threshold *_b = b;
5628 if (_a->threshold > _b->threshold)
5631 if (_a->threshold < _b->threshold)
5637 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5639 struct mem_cgroup_eventfd_list *ev;
5641 list_for_each_entry(ev, &memcg->oom_notify, list)
5642 eventfd_signal(ev->eventfd, 1);
5646 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5648 struct mem_cgroup *iter;
5650 for_each_mem_cgroup_tree(iter, memcg)
5651 mem_cgroup_oom_notify_cb(iter);
5654 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5655 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5657 struct mem_cgroup_thresholds *thresholds;
5658 struct mem_cgroup_threshold_ary *new;
5659 u64 threshold, usage;
5662 ret = res_counter_memparse_write_strategy(args, &threshold);
5666 mutex_lock(&memcg->thresholds_lock);
5669 thresholds = &memcg->thresholds;
5670 else if (type == _MEMSWAP)
5671 thresholds = &memcg->memsw_thresholds;
5675 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5677 /* Check if a threshold crossed before adding a new one */
5678 if (thresholds->primary)
5679 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5681 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5683 /* Allocate memory for new array of thresholds */
5684 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5692 /* Copy thresholds (if any) to new array */
5693 if (thresholds->primary) {
5694 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5695 sizeof(struct mem_cgroup_threshold));
5698 /* Add new threshold */
5699 new->entries[size - 1].eventfd = eventfd;
5700 new->entries[size - 1].threshold = threshold;
5702 /* Sort thresholds. Registering of new threshold isn't time-critical */
5703 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5704 compare_thresholds, NULL);
5706 /* Find current threshold */
5707 new->current_threshold = -1;
5708 for (i = 0; i < size; i++) {
5709 if (new->entries[i].threshold <= usage) {
5711 * new->current_threshold will not be used until
5712 * rcu_assign_pointer(), so it's safe to increment
5715 ++new->current_threshold;
5720 /* Free old spare buffer and save old primary buffer as spare */
5721 kfree(thresholds->spare);
5722 thresholds->spare = thresholds->primary;
5724 rcu_assign_pointer(thresholds->primary, new);
5726 /* To be sure that nobody uses thresholds */
5730 mutex_unlock(&memcg->thresholds_lock);
5735 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5736 struct eventfd_ctx *eventfd, const char *args)
5738 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5741 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5742 struct eventfd_ctx *eventfd, const char *args)
5744 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5747 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5748 struct eventfd_ctx *eventfd, enum res_type type)
5750 struct mem_cgroup_thresholds *thresholds;
5751 struct mem_cgroup_threshold_ary *new;
5755 mutex_lock(&memcg->thresholds_lock);
5757 thresholds = &memcg->thresholds;
5758 else if (type == _MEMSWAP)
5759 thresholds = &memcg->memsw_thresholds;
5763 if (!thresholds->primary)
5766 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5768 /* Check if a threshold crossed before removing */
5769 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5771 /* Calculate new number of threshold */
5773 for (i = 0; i < thresholds->primary->size; i++) {
5774 if (thresholds->primary->entries[i].eventfd != eventfd)
5778 new = thresholds->spare;
5780 /* Set thresholds array to NULL if we don't have thresholds */
5789 /* Copy thresholds and find current threshold */
5790 new->current_threshold = -1;
5791 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5792 if (thresholds->primary->entries[i].eventfd == eventfd)
5795 new->entries[j] = thresholds->primary->entries[i];
5796 if (new->entries[j].threshold <= usage) {
5798 * new->current_threshold will not be used
5799 * until rcu_assign_pointer(), so it's safe to increment
5802 ++new->current_threshold;
5808 /* Swap primary and spare array */
5809 thresholds->spare = thresholds->primary;
5810 /* If all events are unregistered, free the spare array */
5812 kfree(thresholds->spare);
5813 thresholds->spare = NULL;
5816 rcu_assign_pointer(thresholds->primary, new);
5818 /* To be sure that nobody uses thresholds */
5821 mutex_unlock(&memcg->thresholds_lock);
5824 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5825 struct eventfd_ctx *eventfd)
5827 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5830 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5831 struct eventfd_ctx *eventfd)
5833 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5836 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5837 struct eventfd_ctx *eventfd, const char *args)
5839 struct mem_cgroup_eventfd_list *event;
5841 event = kmalloc(sizeof(*event), GFP_KERNEL);
5845 spin_lock(&memcg_oom_lock);
5847 event->eventfd = eventfd;
5848 list_add(&event->list, &memcg->oom_notify);
5850 /* already in OOM ? */
5851 if (atomic_read(&memcg->under_oom))
5852 eventfd_signal(eventfd, 1);
5853 spin_unlock(&memcg_oom_lock);
5858 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5859 struct eventfd_ctx *eventfd)
5861 struct mem_cgroup_eventfd_list *ev, *tmp;
5863 spin_lock(&memcg_oom_lock);
5865 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5866 if (ev->eventfd == eventfd) {
5867 list_del(&ev->list);
5872 spin_unlock(&memcg_oom_lock);
5875 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5877 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5879 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5880 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5884 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5885 struct cftype *cft, u64 val)
5887 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5888 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5890 /* cannot set to root cgroup and only 0 and 1 are allowed */
5891 if (!parent || !((val == 0) || (val == 1)))
5894 mutex_lock(&memcg_create_mutex);
5895 /* oom-kill-disable is a flag for subhierarchy. */
5896 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5897 mutex_unlock(&memcg_create_mutex);
5900 memcg->oom_kill_disable = val;
5902 memcg_oom_recover(memcg);
5903 mutex_unlock(&memcg_create_mutex);
5907 #ifdef CONFIG_MEMCG_KMEM
5908 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5912 memcg->kmemcg_id = -1;
5913 ret = memcg_propagate_kmem(memcg);
5917 return mem_cgroup_sockets_init(memcg, ss);
5920 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5922 mem_cgroup_sockets_destroy(memcg);
5925 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5927 if (!memcg_kmem_is_active(memcg))
5931 * kmem charges can outlive the cgroup. In the case of slab
5932 * pages, for instance, a page contain objects from various
5933 * processes. As we prevent from taking a reference for every
5934 * such allocation we have to be careful when doing uncharge
5935 * (see memcg_uncharge_kmem) and here during offlining.
5937 * The idea is that that only the _last_ uncharge which sees
5938 * the dead memcg will drop the last reference. An additional
5939 * reference is taken here before the group is marked dead
5940 * which is then paired with css_put during uncharge resp. here.
5942 * Although this might sound strange as this path is called from
5943 * css_offline() when the referencemight have dropped down to 0
5944 * and shouldn't be incremented anymore (css_tryget would fail)
5945 * we do not have other options because of the kmem allocations
5948 css_get(&memcg->css);
5950 memcg_kmem_mark_dead(memcg);
5952 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5955 if (memcg_kmem_test_and_clear_dead(memcg))
5956 css_put(&memcg->css);
5959 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5964 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5968 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5974 * DO NOT USE IN NEW FILES.
5976 * "cgroup.event_control" implementation.
5978 * This is way over-engineered. It tries to support fully configurable
5979 * events for each user. Such level of flexibility is completely
5980 * unnecessary especially in the light of the planned unified hierarchy.
5982 * Please deprecate this and replace with something simpler if at all
5987 * Unregister event and free resources.
5989 * Gets called from workqueue.
5991 static void memcg_event_remove(struct work_struct *work)
5993 struct mem_cgroup_event *event =
5994 container_of(work, struct mem_cgroup_event, remove);
5995 struct mem_cgroup *memcg = event->memcg;
5997 remove_wait_queue(event->wqh, &event->wait);
5999 event->unregister_event(memcg, event->eventfd);
6001 /* Notify userspace the event is going away. */
6002 eventfd_signal(event->eventfd, 1);
6004 eventfd_ctx_put(event->eventfd);
6006 css_put(&memcg->css);
6010 * Gets called on POLLHUP on eventfd when user closes it.
6012 * Called with wqh->lock held and interrupts disabled.
6014 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
6015 int sync, void *key)
6017 struct mem_cgroup_event *event =
6018 container_of(wait, struct mem_cgroup_event, wait);
6019 struct mem_cgroup *memcg = event->memcg;
6020 unsigned long flags = (unsigned long)key;
6022 if (flags & POLLHUP) {
6024 * If the event has been detached at cgroup removal, we
6025 * can simply return knowing the other side will cleanup
6028 * We can't race against event freeing since the other
6029 * side will require wqh->lock via remove_wait_queue(),
6032 spin_lock(&memcg->event_list_lock);
6033 if (!list_empty(&event->list)) {
6034 list_del_init(&event->list);
6036 * We are in atomic context, but cgroup_event_remove()
6037 * may sleep, so we have to call it in workqueue.
6039 schedule_work(&event->remove);
6041 spin_unlock(&memcg->event_list_lock);
6047 static void memcg_event_ptable_queue_proc(struct file *file,
6048 wait_queue_head_t *wqh, poll_table *pt)
6050 struct mem_cgroup_event *event =
6051 container_of(pt, struct mem_cgroup_event, pt);
6054 add_wait_queue(wqh, &event->wait);
6058 * DO NOT USE IN NEW FILES.
6060 * Parse input and register new cgroup event handler.
6062 * Input must be in format '<event_fd> <control_fd> <args>'.
6063 * Interpretation of args is defined by control file implementation.
6065 static int memcg_write_event_control(struct cgroup_subsys_state *css,
6066 struct cftype *cft, char *buffer)
6068 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6069 struct mem_cgroup_event *event;
6070 struct cgroup_subsys_state *cfile_css;
6071 unsigned int efd, cfd;
6078 efd = simple_strtoul(buffer, &endp, 10);
6083 cfd = simple_strtoul(buffer, &endp, 10);
6084 if ((*endp != ' ') && (*endp != '\0'))
6088 event = kzalloc(sizeof(*event), GFP_KERNEL);
6092 event->memcg = memcg;
6093 INIT_LIST_HEAD(&event->list);
6094 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
6095 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
6096 INIT_WORK(&event->remove, memcg_event_remove);
6104 event->eventfd = eventfd_ctx_fileget(efile.file);
6105 if (IS_ERR(event->eventfd)) {
6106 ret = PTR_ERR(event->eventfd);
6113 goto out_put_eventfd;
6116 /* the process need read permission on control file */
6117 /* AV: shouldn't we check that it's been opened for read instead? */
6118 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6123 * Determine the event callbacks and set them in @event. This used
6124 * to be done via struct cftype but cgroup core no longer knows
6125 * about these events. The following is crude but the whole thing
6126 * is for compatibility anyway.
6128 * DO NOT ADD NEW FILES.
6130 name = cfile.file->f_dentry->d_name.name;
6132 if (!strcmp(name, "memory.usage_in_bytes")) {
6133 event->register_event = mem_cgroup_usage_register_event;
6134 event->unregister_event = mem_cgroup_usage_unregister_event;
6135 } else if (!strcmp(name, "memory.oom_control")) {
6136 event->register_event = mem_cgroup_oom_register_event;
6137 event->unregister_event = mem_cgroup_oom_unregister_event;
6138 } else if (!strcmp(name, "memory.pressure_level")) {
6139 event->register_event = vmpressure_register_event;
6140 event->unregister_event = vmpressure_unregister_event;
6141 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
6142 event->register_event = memsw_cgroup_usage_register_event;
6143 event->unregister_event = memsw_cgroup_usage_unregister_event;
6150 * Verify @cfile should belong to @css. Also, remaining events are
6151 * automatically removed on cgroup destruction but the removal is
6152 * asynchronous, so take an extra ref on @css.
6154 cfile_css = css_tryget_from_dir(cfile.file->f_dentry->d_parent,
6155 &memory_cgrp_subsys);
6157 if (IS_ERR(cfile_css))
6159 if (cfile_css != css) {
6164 ret = event->register_event(memcg, event->eventfd, buffer);
6168 efile.file->f_op->poll(efile.file, &event->pt);
6170 spin_lock(&memcg->event_list_lock);
6171 list_add(&event->list, &memcg->event_list);
6172 spin_unlock(&memcg->event_list_lock);
6184 eventfd_ctx_put(event->eventfd);
6193 static struct cftype mem_cgroup_files[] = {
6195 .name = "usage_in_bytes",
6196 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6197 .read_u64 = mem_cgroup_read_u64,
6200 .name = "max_usage_in_bytes",
6201 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6202 .trigger = mem_cgroup_reset,
6203 .read_u64 = mem_cgroup_read_u64,
6206 .name = "limit_in_bytes",
6207 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6208 .write_string = mem_cgroup_write,
6209 .read_u64 = mem_cgroup_read_u64,
6212 .name = "soft_limit_in_bytes",
6213 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6214 .write_string = mem_cgroup_write,
6215 .read_u64 = mem_cgroup_read_u64,
6219 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6220 .trigger = mem_cgroup_reset,
6221 .read_u64 = mem_cgroup_read_u64,
6225 .seq_show = memcg_stat_show,
6228 .name = "force_empty",
6229 .trigger = mem_cgroup_force_empty_write,
6232 .name = "use_hierarchy",
6233 .flags = CFTYPE_INSANE,
6234 .write_u64 = mem_cgroup_hierarchy_write,
6235 .read_u64 = mem_cgroup_hierarchy_read,
6238 .name = "cgroup.event_control", /* XXX: for compat */
6239 .write_string = memcg_write_event_control,
6240 .flags = CFTYPE_NO_PREFIX,
6244 .name = "swappiness",
6245 .read_u64 = mem_cgroup_swappiness_read,
6246 .write_u64 = mem_cgroup_swappiness_write,
6249 .name = "move_charge_at_immigrate",
6250 .read_u64 = mem_cgroup_move_charge_read,
6251 .write_u64 = mem_cgroup_move_charge_write,
6254 .name = "oom_control",
6255 .seq_show = mem_cgroup_oom_control_read,
6256 .write_u64 = mem_cgroup_oom_control_write,
6257 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6260 .name = "pressure_level",
6264 .name = "numa_stat",
6265 .seq_show = memcg_numa_stat_show,
6268 #ifdef CONFIG_MEMCG_KMEM
6270 .name = "kmem.limit_in_bytes",
6271 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6272 .write_string = mem_cgroup_write,
6273 .read_u64 = mem_cgroup_read_u64,
6276 .name = "kmem.usage_in_bytes",
6277 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6278 .read_u64 = mem_cgroup_read_u64,
6281 .name = "kmem.failcnt",
6282 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6283 .trigger = mem_cgroup_reset,
6284 .read_u64 = mem_cgroup_read_u64,
6287 .name = "kmem.max_usage_in_bytes",
6288 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6289 .trigger = mem_cgroup_reset,
6290 .read_u64 = mem_cgroup_read_u64,
6292 #ifdef CONFIG_SLABINFO
6294 .name = "kmem.slabinfo",
6295 .seq_show = mem_cgroup_slabinfo_read,
6299 { }, /* terminate */
6302 #ifdef CONFIG_MEMCG_SWAP
6303 static struct cftype memsw_cgroup_files[] = {
6305 .name = "memsw.usage_in_bytes",
6306 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6307 .read_u64 = mem_cgroup_read_u64,
6310 .name = "memsw.max_usage_in_bytes",
6311 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6312 .trigger = mem_cgroup_reset,
6313 .read_u64 = mem_cgroup_read_u64,
6316 .name = "memsw.limit_in_bytes",
6317 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6318 .write_string = mem_cgroup_write,
6319 .read_u64 = mem_cgroup_read_u64,
6322 .name = "memsw.failcnt",
6323 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6324 .trigger = mem_cgroup_reset,
6325 .read_u64 = mem_cgroup_read_u64,
6327 { }, /* terminate */
6330 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6332 struct mem_cgroup_per_node *pn;
6333 struct mem_cgroup_per_zone *mz;
6334 int zone, tmp = node;
6336 * This routine is called against possible nodes.
6337 * But it's BUG to call kmalloc() against offline node.
6339 * TODO: this routine can waste much memory for nodes which will
6340 * never be onlined. It's better to use memory hotplug callback
6343 if (!node_state(node, N_NORMAL_MEMORY))
6345 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6349 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6350 mz = &pn->zoneinfo[zone];
6351 lruvec_init(&mz->lruvec);
6352 mz->usage_in_excess = 0;
6353 mz->on_tree = false;
6356 memcg->nodeinfo[node] = pn;
6360 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6362 kfree(memcg->nodeinfo[node]);
6365 static struct mem_cgroup *mem_cgroup_alloc(void)
6367 struct mem_cgroup *memcg;
6370 size = sizeof(struct mem_cgroup);
6371 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6373 memcg = kzalloc(size, GFP_KERNEL);
6377 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6380 spin_lock_init(&memcg->pcp_counter_lock);
6389 * At destroying mem_cgroup, references from swap_cgroup can remain.
6390 * (scanning all at force_empty is too costly...)
6392 * Instead of clearing all references at force_empty, we remember
6393 * the number of reference from swap_cgroup and free mem_cgroup when
6394 * it goes down to 0.
6396 * Removal of cgroup itself succeeds regardless of refs from swap.
6399 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6403 mem_cgroup_remove_from_trees(memcg);
6406 free_mem_cgroup_per_zone_info(memcg, node);
6408 free_percpu(memcg->stat);
6411 * We need to make sure that (at least for now), the jump label
6412 * destruction code runs outside of the cgroup lock. This is because
6413 * get_online_cpus(), which is called from the static_branch update,
6414 * can't be called inside the cgroup_lock. cpusets are the ones
6415 * enforcing this dependency, so if they ever change, we might as well.
6417 * schedule_work() will guarantee this happens. Be careful if you need
6418 * to move this code around, and make sure it is outside
6421 disarm_static_keys(memcg);
6426 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6428 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6430 if (!memcg->res.parent)
6432 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6434 EXPORT_SYMBOL(parent_mem_cgroup);
6436 static void __init mem_cgroup_soft_limit_tree_init(void)
6438 struct mem_cgroup_tree_per_node *rtpn;
6439 struct mem_cgroup_tree_per_zone *rtpz;
6440 int tmp, node, zone;
6442 for_each_node(node) {
6444 if (!node_state(node, N_NORMAL_MEMORY))
6446 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6449 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6451 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6452 rtpz = &rtpn->rb_tree_per_zone[zone];
6453 rtpz->rb_root = RB_ROOT;
6454 spin_lock_init(&rtpz->lock);
6459 static struct cgroup_subsys_state * __ref
6460 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6462 struct mem_cgroup *memcg;
6463 long error = -ENOMEM;
6466 memcg = mem_cgroup_alloc();
6468 return ERR_PTR(error);
6471 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6475 if (parent_css == NULL) {
6476 root_mem_cgroup = memcg;
6477 res_counter_init(&memcg->res, NULL);
6478 res_counter_init(&memcg->memsw, NULL);
6479 res_counter_init(&memcg->kmem, NULL);
6482 memcg->last_scanned_node = MAX_NUMNODES;
6483 INIT_LIST_HEAD(&memcg->oom_notify);
6484 memcg->move_charge_at_immigrate = 0;
6485 mutex_init(&memcg->thresholds_lock);
6486 spin_lock_init(&memcg->move_lock);
6487 vmpressure_init(&memcg->vmpressure);
6488 INIT_LIST_HEAD(&memcg->event_list);
6489 spin_lock_init(&memcg->event_list_lock);
6494 __mem_cgroup_free(memcg);
6495 return ERR_PTR(error);
6499 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6501 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6502 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6504 if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6510 mutex_lock(&memcg_create_mutex);
6512 memcg->use_hierarchy = parent->use_hierarchy;
6513 memcg->oom_kill_disable = parent->oom_kill_disable;
6514 memcg->swappiness = mem_cgroup_swappiness(parent);
6516 if (parent->use_hierarchy) {
6517 res_counter_init(&memcg->res, &parent->res);
6518 res_counter_init(&memcg->memsw, &parent->memsw);
6519 res_counter_init(&memcg->kmem, &parent->kmem);
6522 * No need to take a reference to the parent because cgroup
6523 * core guarantees its existence.
6526 res_counter_init(&memcg->res, NULL);
6527 res_counter_init(&memcg->memsw, NULL);
6528 res_counter_init(&memcg->kmem, NULL);
6530 * Deeper hierachy with use_hierarchy == false doesn't make
6531 * much sense so let cgroup subsystem know about this
6532 * unfortunate state in our controller.
6534 if (parent != root_mem_cgroup)
6535 memory_cgrp_subsys.broken_hierarchy = true;
6537 mutex_unlock(&memcg_create_mutex);
6539 return memcg_init_kmem(memcg, &memory_cgrp_subsys);
6543 * Announce all parents that a group from their hierarchy is gone.
6545 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6547 struct mem_cgroup *parent = memcg;
6549 while ((parent = parent_mem_cgroup(parent)))
6550 mem_cgroup_iter_invalidate(parent);
6553 * if the root memcg is not hierarchical we have to check it
6556 if (!root_mem_cgroup->use_hierarchy)
6557 mem_cgroup_iter_invalidate(root_mem_cgroup);
6560 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6562 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6563 struct mem_cgroup_event *event, *tmp;
6564 struct cgroup_subsys_state *iter;
6567 * Unregister events and notify userspace.
6568 * Notify userspace about cgroup removing only after rmdir of cgroup
6569 * directory to avoid race between userspace and kernelspace.
6571 spin_lock(&memcg->event_list_lock);
6572 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6573 list_del_init(&event->list);
6574 schedule_work(&event->remove);
6576 spin_unlock(&memcg->event_list_lock);
6578 kmem_cgroup_css_offline(memcg);
6580 mem_cgroup_invalidate_reclaim_iterators(memcg);
6583 * This requires that offlining is serialized. Right now that is
6584 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6586 css_for_each_descendant_post(iter, css)
6587 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
6589 mem_cgroup_destroy_all_caches(memcg);
6590 vmpressure_cleanup(&memcg->vmpressure);
6593 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6595 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6597 * XXX: css_offline() would be where we should reparent all
6598 * memory to prepare the cgroup for destruction. However,
6599 * memcg does not do css_tryget() and res_counter charging
6600 * under the same RCU lock region, which means that charging
6601 * could race with offlining. Offlining only happens to
6602 * cgroups with no tasks in them but charges can show up
6603 * without any tasks from the swapin path when the target
6604 * memcg is looked up from the swapout record and not from the
6605 * current task as it usually is. A race like this can leak
6606 * charges and put pages with stale cgroup pointers into
6610 * lookup_swap_cgroup_id()
6612 * mem_cgroup_lookup()
6615 * disable css_tryget()
6618 * reparent_charges()
6619 * res_counter_charge()
6622 * pc->mem_cgroup = dead memcg
6625 * The bulk of the charges are still moved in offline_css() to
6626 * avoid pinning a lot of pages in case a long-term reference
6627 * like a swapout record is deferring the css_free() to long
6628 * after offlining. But this makes sure we catch any charges
6629 * made after offlining:
6631 mem_cgroup_reparent_charges(memcg);
6633 memcg_destroy_kmem(memcg);
6634 __mem_cgroup_free(memcg);
6638 /* Handlers for move charge at task migration. */
6639 #define PRECHARGE_COUNT_AT_ONCE 256
6640 static int mem_cgroup_do_precharge(unsigned long count)
6643 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6644 struct mem_cgroup *memcg = mc.to;
6646 if (mem_cgroup_is_root(memcg)) {
6647 mc.precharge += count;
6648 /* we don't need css_get for root */
6651 /* try to charge at once */
6653 struct res_counter *dummy;
6655 * "memcg" cannot be under rmdir() because we've already checked
6656 * by cgroup_lock_live_cgroup() that it is not removed and we
6657 * are still under the same cgroup_mutex. So we can postpone
6660 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6662 if (do_swap_account && res_counter_charge(&memcg->memsw,
6663 PAGE_SIZE * count, &dummy)) {
6664 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6667 mc.precharge += count;
6671 /* fall back to one by one charge */
6673 if (signal_pending(current)) {
6677 if (!batch_count--) {
6678 batch_count = PRECHARGE_COUNT_AT_ONCE;
6681 ret = __mem_cgroup_try_charge(NULL,
6682 GFP_KERNEL, 1, &memcg, false);
6684 /* mem_cgroup_clear_mc() will do uncharge later */
6692 * get_mctgt_type - get target type of moving charge
6693 * @vma: the vma the pte to be checked belongs
6694 * @addr: the address corresponding to the pte to be checked
6695 * @ptent: the pte to be checked
6696 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6699 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6700 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6701 * move charge. if @target is not NULL, the page is stored in target->page
6702 * with extra refcnt got(Callers should handle it).
6703 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6704 * target for charge migration. if @target is not NULL, the entry is stored
6707 * Called with pte lock held.
6714 enum mc_target_type {
6720 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6721 unsigned long addr, pte_t ptent)
6723 struct page *page = vm_normal_page(vma, addr, ptent);
6725 if (!page || !page_mapped(page))
6727 if (PageAnon(page)) {
6728 /* we don't move shared anon */
6731 } else if (!move_file())
6732 /* we ignore mapcount for file pages */
6734 if (!get_page_unless_zero(page))
6741 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6742 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6744 struct page *page = NULL;
6745 swp_entry_t ent = pte_to_swp_entry(ptent);
6747 if (!move_anon() || non_swap_entry(ent))
6750 * Because lookup_swap_cache() updates some statistics counter,
6751 * we call find_get_page() with swapper_space directly.
6753 page = find_get_page(swap_address_space(ent), ent.val);
6754 if (do_swap_account)
6755 entry->val = ent.val;
6760 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6761 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6767 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6768 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6770 struct page *page = NULL;
6771 struct address_space *mapping;
6774 if (!vma->vm_file) /* anonymous vma */
6779 mapping = vma->vm_file->f_mapping;
6780 if (pte_none(ptent))
6781 pgoff = linear_page_index(vma, addr);
6782 else /* pte_file(ptent) is true */
6783 pgoff = pte_to_pgoff(ptent);
6785 /* page is moved even if it's not RSS of this task(page-faulted). */
6786 page = find_get_page(mapping, pgoff);
6789 /* shmem/tmpfs may report page out on swap: account for that too. */
6790 if (radix_tree_exceptional_entry(page)) {
6791 swp_entry_t swap = radix_to_swp_entry(page);
6792 if (do_swap_account)
6794 page = find_get_page(swap_address_space(swap), swap.val);
6800 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6801 unsigned long addr, pte_t ptent, union mc_target *target)
6803 struct page *page = NULL;
6804 struct page_cgroup *pc;
6805 enum mc_target_type ret = MC_TARGET_NONE;
6806 swp_entry_t ent = { .val = 0 };
6808 if (pte_present(ptent))
6809 page = mc_handle_present_pte(vma, addr, ptent);
6810 else if (is_swap_pte(ptent))
6811 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6812 else if (pte_none(ptent) || pte_file(ptent))
6813 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6815 if (!page && !ent.val)
6818 pc = lookup_page_cgroup(page);
6820 * Do only loose check w/o page_cgroup lock.
6821 * mem_cgroup_move_account() checks the pc is valid or not under
6824 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6825 ret = MC_TARGET_PAGE;
6827 target->page = page;
6829 if (!ret || !target)
6832 /* There is a swap entry and a page doesn't exist or isn't charged */
6833 if (ent.val && !ret &&
6834 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6835 ret = MC_TARGET_SWAP;
6842 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6844 * We don't consider swapping or file mapped pages because THP does not
6845 * support them for now.
6846 * Caller should make sure that pmd_trans_huge(pmd) is true.
6848 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6849 unsigned long addr, pmd_t pmd, union mc_target *target)
6851 struct page *page = NULL;
6852 struct page_cgroup *pc;
6853 enum mc_target_type ret = MC_TARGET_NONE;
6855 page = pmd_page(pmd);
6856 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6859 pc = lookup_page_cgroup(page);
6860 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6861 ret = MC_TARGET_PAGE;
6864 target->page = page;
6870 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6871 unsigned long addr, pmd_t pmd, union mc_target *target)
6873 return MC_TARGET_NONE;
6877 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6878 unsigned long addr, unsigned long end,
6879 struct mm_walk *walk)
6881 struct vm_area_struct *vma = walk->private;
6885 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6886 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6887 mc.precharge += HPAGE_PMD_NR;
6892 if (pmd_trans_unstable(pmd))
6894 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6895 for (; addr != end; pte++, addr += PAGE_SIZE)
6896 if (get_mctgt_type(vma, addr, *pte, NULL))
6897 mc.precharge++; /* increment precharge temporarily */
6898 pte_unmap_unlock(pte - 1, ptl);
6904 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6906 unsigned long precharge;
6907 struct vm_area_struct *vma;
6909 down_read(&mm->mmap_sem);
6910 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6911 struct mm_walk mem_cgroup_count_precharge_walk = {
6912 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6916 if (is_vm_hugetlb_page(vma))
6918 walk_page_range(vma->vm_start, vma->vm_end,
6919 &mem_cgroup_count_precharge_walk);
6921 up_read(&mm->mmap_sem);
6923 precharge = mc.precharge;
6929 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6931 unsigned long precharge = mem_cgroup_count_precharge(mm);
6933 VM_BUG_ON(mc.moving_task);
6934 mc.moving_task = current;
6935 return mem_cgroup_do_precharge(precharge);
6938 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6939 static void __mem_cgroup_clear_mc(void)
6941 struct mem_cgroup *from = mc.from;
6942 struct mem_cgroup *to = mc.to;
6945 /* we must uncharge all the leftover precharges from mc.to */
6947 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6951 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6952 * we must uncharge here.
6954 if (mc.moved_charge) {
6955 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6956 mc.moved_charge = 0;
6958 /* we must fixup refcnts and charges */
6959 if (mc.moved_swap) {
6960 /* uncharge swap account from the old cgroup */
6961 if (!mem_cgroup_is_root(mc.from))
6962 res_counter_uncharge(&mc.from->memsw,
6963 PAGE_SIZE * mc.moved_swap);
6965 for (i = 0; i < mc.moved_swap; i++)
6966 css_put(&mc.from->css);
6968 if (!mem_cgroup_is_root(mc.to)) {
6970 * we charged both to->res and to->memsw, so we should
6973 res_counter_uncharge(&mc.to->res,
6974 PAGE_SIZE * mc.moved_swap);
6976 /* we've already done css_get(mc.to) */
6979 memcg_oom_recover(from);
6980 memcg_oom_recover(to);
6981 wake_up_all(&mc.waitq);
6984 static void mem_cgroup_clear_mc(void)
6986 struct mem_cgroup *from = mc.from;
6989 * we must clear moving_task before waking up waiters at the end of
6992 mc.moving_task = NULL;
6993 __mem_cgroup_clear_mc();
6994 spin_lock(&mc.lock);
6997 spin_unlock(&mc.lock);
6998 mem_cgroup_end_move(from);
7001 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7002 struct cgroup_taskset *tset)
7004 struct task_struct *p = cgroup_taskset_first(tset);
7006 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7007 unsigned long move_charge_at_immigrate;
7010 * We are now commited to this value whatever it is. Changes in this
7011 * tunable will only affect upcoming migrations, not the current one.
7012 * So we need to save it, and keep it going.
7014 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
7015 if (move_charge_at_immigrate) {
7016 struct mm_struct *mm;
7017 struct mem_cgroup *from = mem_cgroup_from_task(p);
7019 VM_BUG_ON(from == memcg);
7021 mm = get_task_mm(p);
7024 /* We move charges only when we move a owner of the mm */
7025 if (mm->owner == p) {
7028 VM_BUG_ON(mc.precharge);
7029 VM_BUG_ON(mc.moved_charge);
7030 VM_BUG_ON(mc.moved_swap);
7031 mem_cgroup_start_move(from);
7032 spin_lock(&mc.lock);
7035 mc.immigrate_flags = move_charge_at_immigrate;
7036 spin_unlock(&mc.lock);
7037 /* We set mc.moving_task later */
7039 ret = mem_cgroup_precharge_mc(mm);
7041 mem_cgroup_clear_mc();
7048 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7049 struct cgroup_taskset *tset)
7051 mem_cgroup_clear_mc();
7054 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
7055 unsigned long addr, unsigned long end,
7056 struct mm_walk *walk)
7059 struct vm_area_struct *vma = walk->private;
7062 enum mc_target_type target_type;
7063 union mc_target target;
7065 struct page_cgroup *pc;
7068 * We don't take compound_lock() here but no race with splitting thp
7070 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7071 * under splitting, which means there's no concurrent thp split,
7072 * - if another thread runs into split_huge_page() just after we
7073 * entered this if-block, the thread must wait for page table lock
7074 * to be unlocked in __split_huge_page_splitting(), where the main
7075 * part of thp split is not executed yet.
7077 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
7078 if (mc.precharge < HPAGE_PMD_NR) {
7082 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
7083 if (target_type == MC_TARGET_PAGE) {
7085 if (!isolate_lru_page(page)) {
7086 pc = lookup_page_cgroup(page);
7087 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
7088 pc, mc.from, mc.to)) {
7089 mc.precharge -= HPAGE_PMD_NR;
7090 mc.moved_charge += HPAGE_PMD_NR;
7092 putback_lru_page(page);
7100 if (pmd_trans_unstable(pmd))
7103 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7104 for (; addr != end; addr += PAGE_SIZE) {
7105 pte_t ptent = *(pte++);
7111 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7112 case MC_TARGET_PAGE:
7114 if (isolate_lru_page(page))
7116 pc = lookup_page_cgroup(page);
7117 if (!mem_cgroup_move_account(page, 1, pc,
7120 /* we uncharge from mc.from later. */
7123 putback_lru_page(page);
7124 put: /* get_mctgt_type() gets the page */
7127 case MC_TARGET_SWAP:
7129 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7131 /* we fixup refcnts and charges later. */
7139 pte_unmap_unlock(pte - 1, ptl);
7144 * We have consumed all precharges we got in can_attach().
7145 * We try charge one by one, but don't do any additional
7146 * charges to mc.to if we have failed in charge once in attach()
7149 ret = mem_cgroup_do_precharge(1);
7157 static void mem_cgroup_move_charge(struct mm_struct *mm)
7159 struct vm_area_struct *vma;
7161 lru_add_drain_all();
7163 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7165 * Someone who are holding the mmap_sem might be waiting in
7166 * waitq. So we cancel all extra charges, wake up all waiters,
7167 * and retry. Because we cancel precharges, we might not be able
7168 * to move enough charges, but moving charge is a best-effort
7169 * feature anyway, so it wouldn't be a big problem.
7171 __mem_cgroup_clear_mc();
7175 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7177 struct mm_walk mem_cgroup_move_charge_walk = {
7178 .pmd_entry = mem_cgroup_move_charge_pte_range,
7182 if (is_vm_hugetlb_page(vma))
7184 ret = walk_page_range(vma->vm_start, vma->vm_end,
7185 &mem_cgroup_move_charge_walk);
7188 * means we have consumed all precharges and failed in
7189 * doing additional charge. Just abandon here.
7193 up_read(&mm->mmap_sem);
7196 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7197 struct cgroup_taskset *tset)
7199 struct task_struct *p = cgroup_taskset_first(tset);
7200 struct mm_struct *mm = get_task_mm(p);
7204 mem_cgroup_move_charge(mm);
7208 mem_cgroup_clear_mc();
7210 #else /* !CONFIG_MMU */
7211 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7212 struct cgroup_taskset *tset)
7216 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7217 struct cgroup_taskset *tset)
7220 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7221 struct cgroup_taskset *tset)
7227 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7228 * to verify sane_behavior flag on each mount attempt.
7230 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7233 * use_hierarchy is forced with sane_behavior. cgroup core
7234 * guarantees that @root doesn't have any children, so turning it
7235 * on for the root memcg is enough.
7237 if (cgroup_sane_behavior(root_css->cgroup))
7238 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7241 struct cgroup_subsys memory_cgrp_subsys = {
7242 .css_alloc = mem_cgroup_css_alloc,
7243 .css_online = mem_cgroup_css_online,
7244 .css_offline = mem_cgroup_css_offline,
7245 .css_free = mem_cgroup_css_free,
7246 .can_attach = mem_cgroup_can_attach,
7247 .cancel_attach = mem_cgroup_cancel_attach,
7248 .attach = mem_cgroup_move_task,
7249 .bind = mem_cgroup_bind,
7250 .base_cftypes = mem_cgroup_files,
7254 #ifdef CONFIG_MEMCG_SWAP
7255 static int __init enable_swap_account(char *s)
7257 if (!strcmp(s, "1"))
7258 really_do_swap_account = 1;
7259 else if (!strcmp(s, "0"))
7260 really_do_swap_account = 0;
7263 __setup("swapaccount=", enable_swap_account);
7265 static void __init memsw_file_init(void)
7267 WARN_ON(cgroup_add_cftypes(&memory_cgrp_subsys, memsw_cgroup_files));
7270 static void __init enable_swap_cgroup(void)
7272 if (!mem_cgroup_disabled() && really_do_swap_account) {
7273 do_swap_account = 1;
7279 static void __init enable_swap_cgroup(void)
7285 * subsys_initcall() for memory controller.
7287 * Some parts like hotcpu_notifier() have to be initialized from this context
7288 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7289 * everything that doesn't depend on a specific mem_cgroup structure should
7290 * be initialized from here.
7292 static int __init mem_cgroup_init(void)
7294 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7295 enable_swap_cgroup();
7296 mem_cgroup_soft_limit_tree_init();
7300 subsys_initcall(mem_cgroup_init);