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
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
68 #include <net/tcp_memcontrol.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
76 EXPORT_SYMBOL(memory_cgrp_subsys);
78 #define MEM_CGROUP_RECLAIM_RETRIES 5
79 static struct mem_cgroup *root_mem_cgroup __read_mostly;
81 /* Whether the swap controller is active */
82 #ifdef CONFIG_MEMCG_SWAP
83 int do_swap_account __read_mostly;
85 #define do_swap_account 0
88 static const char * const mem_cgroup_stat_names[] = {
97 static const char * const mem_cgroup_events_names[] = {
104 static const char * const mem_cgroup_lru_names[] = {
113 * Per memcg event counter is incremented at every pagein/pageout. With THP,
114 * it will be incremated by the number of pages. This counter is used for
115 * for trigger some periodic events. This is straightforward and better
116 * than using jiffies etc. to handle periodic memcg event.
118 enum mem_cgroup_events_target {
119 MEM_CGROUP_TARGET_THRESH,
120 MEM_CGROUP_TARGET_SOFTLIMIT,
121 MEM_CGROUP_TARGET_NUMAINFO,
124 #define THRESHOLDS_EVENTS_TARGET 128
125 #define SOFTLIMIT_EVENTS_TARGET 1024
126 #define NUMAINFO_EVENTS_TARGET 1024
128 struct mem_cgroup_stat_cpu {
129 long count[MEM_CGROUP_STAT_NSTATS];
130 unsigned long events[MEMCG_NR_EVENTS];
131 unsigned long nr_page_events;
132 unsigned long targets[MEM_CGROUP_NTARGETS];
135 struct reclaim_iter {
136 struct mem_cgroup *position;
137 /* scan generation, increased every round-trip */
138 unsigned int generation;
142 * per-zone information in memory controller.
144 struct mem_cgroup_per_zone {
145 struct lruvec lruvec;
146 unsigned long lru_size[NR_LRU_LISTS];
148 struct reclaim_iter iter[DEF_PRIORITY + 1];
150 struct rb_node tree_node; /* RB tree node */
151 unsigned long usage_in_excess;/* Set to the value by which */
152 /* the soft limit is exceeded*/
154 struct mem_cgroup *memcg; /* Back pointer, we cannot */
155 /* use container_of */
158 struct mem_cgroup_per_node {
159 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
163 * Cgroups above their limits are maintained in a RB-Tree, independent of
164 * their hierarchy representation
167 struct mem_cgroup_tree_per_zone {
168 struct rb_root rb_root;
172 struct mem_cgroup_tree_per_node {
173 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
176 struct mem_cgroup_tree {
177 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
180 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
182 struct mem_cgroup_threshold {
183 struct eventfd_ctx *eventfd;
184 unsigned long threshold;
188 struct mem_cgroup_threshold_ary {
189 /* An array index points to threshold just below or equal to usage. */
190 int current_threshold;
191 /* Size of entries[] */
193 /* Array of thresholds */
194 struct mem_cgroup_threshold entries[0];
197 struct mem_cgroup_thresholds {
198 /* Primary thresholds array */
199 struct mem_cgroup_threshold_ary *primary;
201 * Spare threshold array.
202 * This is needed to make mem_cgroup_unregister_event() "never fail".
203 * It must be able to store at least primary->size - 1 entries.
205 struct mem_cgroup_threshold_ary *spare;
209 struct mem_cgroup_eventfd_list {
210 struct list_head list;
211 struct eventfd_ctx *eventfd;
215 * cgroup_event represents events which userspace want to receive.
217 struct mem_cgroup_event {
219 * memcg which the event belongs to.
221 struct mem_cgroup *memcg;
223 * eventfd to signal userspace about the event.
225 struct eventfd_ctx *eventfd;
227 * Each of these stored in a list by the cgroup.
229 struct list_head list;
231 * register_event() callback will be used to add new userspace
232 * waiter for changes related to this event. Use eventfd_signal()
233 * on eventfd to send notification to userspace.
235 int (*register_event)(struct mem_cgroup *memcg,
236 struct eventfd_ctx *eventfd, const char *args);
238 * unregister_event() callback will be called when userspace closes
239 * the eventfd or on cgroup removing. This callback must be set,
240 * if you want provide notification functionality.
242 void (*unregister_event)(struct mem_cgroup *memcg,
243 struct eventfd_ctx *eventfd);
245 * All fields below needed to unregister event when
246 * userspace closes eventfd.
249 wait_queue_head_t *wqh;
251 struct work_struct remove;
254 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
255 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
258 * The memory controller data structure. The memory controller controls both
259 * page cache and RSS per cgroup. We would eventually like to provide
260 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
261 * to help the administrator determine what knobs to tune.
264 struct cgroup_subsys_state css;
266 /* Accounted resources */
267 struct page_counter memory;
268 struct page_counter memsw;
269 struct page_counter kmem;
271 /* Normal memory consumption range */
275 unsigned long soft_limit;
277 /* vmpressure notifications */
278 struct vmpressure vmpressure;
280 /* css_online() has been completed */
284 * Should the accounting and control be hierarchical, per subtree?
290 atomic_t oom_wakeups;
293 /* OOM-Killer disable */
294 int oom_kill_disable;
296 /* protect arrays of thresholds */
297 struct mutex thresholds_lock;
299 /* thresholds for memory usage. RCU-protected */
300 struct mem_cgroup_thresholds thresholds;
302 /* thresholds for mem+swap usage. RCU-protected */
303 struct mem_cgroup_thresholds memsw_thresholds;
305 /* For oom notifier event fd */
306 struct list_head oom_notify;
309 * Should we move charges of a task when a task is moved into this
310 * mem_cgroup ? And what type of charges should we move ?
312 unsigned long move_charge_at_immigrate;
314 * set > 0 if pages under this cgroup are moving to other cgroup.
316 atomic_t moving_account;
317 /* taken only while moving_account > 0 */
318 spinlock_t move_lock;
319 struct task_struct *move_lock_task;
320 unsigned long move_lock_flags;
324 struct mem_cgroup_stat_cpu __percpu *stat;
326 * used when a cpu is offlined or other synchronizations
327 * See mem_cgroup_read_stat().
329 struct mem_cgroup_stat_cpu nocpu_base;
330 spinlock_t pcp_counter_lock;
332 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
333 struct cg_proto tcp_mem;
335 #if defined(CONFIG_MEMCG_KMEM)
336 /* Index in the kmem_cache->memcg_params.memcg_caches array */
338 bool kmem_acct_activated;
339 bool kmem_acct_active;
342 int last_scanned_node;
344 nodemask_t scan_nodes;
345 atomic_t numainfo_events;
346 atomic_t numainfo_updating;
349 /* List of events which userspace want to receive */
350 struct list_head event_list;
351 spinlock_t event_list_lock;
353 struct mem_cgroup_per_node *nodeinfo[0];
354 /* WARNING: nodeinfo must be the last member here */
357 #ifdef CONFIG_MEMCG_KMEM
358 bool memcg_kmem_is_active(struct mem_cgroup *memcg)
360 return memcg->kmem_acct_active;
364 /* Stuffs for move charges at task migration. */
366 * Types of charges to be moved.
368 #define MOVE_ANON 0x1U
369 #define MOVE_FILE 0x2U
370 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
372 /* "mc" and its members are protected by cgroup_mutex */
373 static struct move_charge_struct {
374 spinlock_t lock; /* for from, to */
375 struct mem_cgroup *from;
376 struct mem_cgroup *to;
378 unsigned long precharge;
379 unsigned long moved_charge;
380 unsigned long moved_swap;
381 struct task_struct *moving_task; /* a task moving charges */
382 wait_queue_head_t waitq; /* a waitq for other context */
384 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
385 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
389 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
390 * limit reclaim to prevent infinite loops, if they ever occur.
392 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
393 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
396 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
397 MEM_CGROUP_CHARGE_TYPE_ANON,
398 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
399 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
403 /* for encoding cft->private value on file */
411 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
412 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
413 #define MEMFILE_ATTR(val) ((val) & 0xffff)
414 /* Used for OOM nofiier */
415 #define OOM_CONTROL (0)
418 * The memcg_create_mutex will be held whenever a new cgroup is created.
419 * As a consequence, any change that needs to protect against new child cgroups
420 * appearing has to hold it as well.
422 static DEFINE_MUTEX(memcg_create_mutex);
424 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
426 return s ? container_of(s, struct mem_cgroup, css) : NULL;
429 /* Some nice accessors for the vmpressure. */
430 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
433 memcg = root_mem_cgroup;
434 return &memcg->vmpressure;
437 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
439 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
442 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
444 return (memcg == root_mem_cgroup);
448 * We restrict the id in the range of [1, 65535], so it can fit into
451 #define MEM_CGROUP_ID_MAX USHRT_MAX
453 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
455 return memcg->css.id;
459 * A helper function to get mem_cgroup from ID. must be called under
460 * rcu_read_lock(). The caller is responsible for calling
461 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
462 * refcnt from swap can be called against removed memcg.)
464 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
466 struct cgroup_subsys_state *css;
468 css = css_from_id(id, &memory_cgrp_subsys);
469 return mem_cgroup_from_css(css);
472 /* Writing them here to avoid exposing memcg's inner layout */
473 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
475 void sock_update_memcg(struct sock *sk)
477 if (mem_cgroup_sockets_enabled) {
478 struct mem_cgroup *memcg;
479 struct cg_proto *cg_proto;
481 BUG_ON(!sk->sk_prot->proto_cgroup);
483 /* Socket cloning can throw us here with sk_cgrp already
484 * filled. It won't however, necessarily happen from
485 * process context. So the test for root memcg given
486 * the current task's memcg won't help us in this case.
488 * Respecting the original socket's memcg is a better
489 * decision in this case.
492 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
493 css_get(&sk->sk_cgrp->memcg->css);
498 memcg = mem_cgroup_from_task(current);
499 cg_proto = sk->sk_prot->proto_cgroup(memcg);
500 if (!mem_cgroup_is_root(memcg) &&
501 memcg_proto_active(cg_proto) &&
502 css_tryget_online(&memcg->css)) {
503 sk->sk_cgrp = cg_proto;
508 EXPORT_SYMBOL(sock_update_memcg);
510 void sock_release_memcg(struct sock *sk)
512 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
513 struct mem_cgroup *memcg;
514 WARN_ON(!sk->sk_cgrp->memcg);
515 memcg = sk->sk_cgrp->memcg;
516 css_put(&sk->sk_cgrp->memcg->css);
520 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
522 if (!memcg || mem_cgroup_is_root(memcg))
525 return &memcg->tcp_mem;
527 EXPORT_SYMBOL(tcp_proto_cgroup);
531 #ifdef CONFIG_MEMCG_KMEM
533 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
534 * The main reason for not using cgroup id for this:
535 * this works better in sparse environments, where we have a lot of memcgs,
536 * but only a few kmem-limited. Or also, if we have, for instance, 200
537 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
538 * 200 entry array for that.
540 * The current size of the caches array is stored in memcg_nr_cache_ids. It
541 * will double each time we have to increase it.
543 static DEFINE_IDA(memcg_cache_ida);
544 int memcg_nr_cache_ids;
546 /* Protects memcg_nr_cache_ids */
547 static DECLARE_RWSEM(memcg_cache_ids_sem);
549 void memcg_get_cache_ids(void)
551 down_read(&memcg_cache_ids_sem);
554 void memcg_put_cache_ids(void)
556 up_read(&memcg_cache_ids_sem);
560 * MIN_SIZE is different than 1, because we would like to avoid going through
561 * the alloc/free process all the time. In a small machine, 4 kmem-limited
562 * cgroups is a reasonable guess. In the future, it could be a parameter or
563 * tunable, but that is strictly not necessary.
565 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
566 * this constant directly from cgroup, but it is understandable that this is
567 * better kept as an internal representation in cgroup.c. In any case, the
568 * cgrp_id space is not getting any smaller, and we don't have to necessarily
569 * increase ours as well if it increases.
571 #define MEMCG_CACHES_MIN_SIZE 4
572 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
575 * A lot of the calls to the cache allocation functions are expected to be
576 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
577 * conditional to this static branch, we'll have to allow modules that does
578 * kmem_cache_alloc and the such to see this symbol as well
580 struct static_key memcg_kmem_enabled_key;
581 EXPORT_SYMBOL(memcg_kmem_enabled_key);
583 #endif /* CONFIG_MEMCG_KMEM */
585 static struct mem_cgroup_per_zone *
586 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
588 int nid = zone_to_nid(zone);
589 int zid = zone_idx(zone);
591 return &memcg->nodeinfo[nid]->zoneinfo[zid];
594 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
599 static struct mem_cgroup_per_zone *
600 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
602 int nid = page_to_nid(page);
603 int zid = page_zonenum(page);
605 return &memcg->nodeinfo[nid]->zoneinfo[zid];
608 static struct mem_cgroup_tree_per_zone *
609 soft_limit_tree_node_zone(int nid, int zid)
611 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
614 static struct mem_cgroup_tree_per_zone *
615 soft_limit_tree_from_page(struct page *page)
617 int nid = page_to_nid(page);
618 int zid = page_zonenum(page);
620 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
623 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
624 struct mem_cgroup_tree_per_zone *mctz,
625 unsigned long new_usage_in_excess)
627 struct rb_node **p = &mctz->rb_root.rb_node;
628 struct rb_node *parent = NULL;
629 struct mem_cgroup_per_zone *mz_node;
634 mz->usage_in_excess = new_usage_in_excess;
635 if (!mz->usage_in_excess)
639 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
641 if (mz->usage_in_excess < mz_node->usage_in_excess)
644 * We can't avoid mem cgroups that are over their soft
645 * limit by the same amount
647 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
650 rb_link_node(&mz->tree_node, parent, p);
651 rb_insert_color(&mz->tree_node, &mctz->rb_root);
655 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
656 struct mem_cgroup_tree_per_zone *mctz)
660 rb_erase(&mz->tree_node, &mctz->rb_root);
664 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
665 struct mem_cgroup_tree_per_zone *mctz)
669 spin_lock_irqsave(&mctz->lock, flags);
670 __mem_cgroup_remove_exceeded(mz, mctz);
671 spin_unlock_irqrestore(&mctz->lock, flags);
674 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
676 unsigned long nr_pages = page_counter_read(&memcg->memory);
677 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
678 unsigned long excess = 0;
680 if (nr_pages > soft_limit)
681 excess = nr_pages - soft_limit;
686 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
688 unsigned long excess;
689 struct mem_cgroup_per_zone *mz;
690 struct mem_cgroup_tree_per_zone *mctz;
692 mctz = soft_limit_tree_from_page(page);
694 * Necessary to update all ancestors when hierarchy is used.
695 * because their event counter is not touched.
697 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
698 mz = mem_cgroup_page_zoneinfo(memcg, page);
699 excess = soft_limit_excess(memcg);
701 * We have to update the tree if mz is on RB-tree or
702 * mem is over its softlimit.
704 if (excess || mz->on_tree) {
707 spin_lock_irqsave(&mctz->lock, flags);
708 /* if on-tree, remove it */
710 __mem_cgroup_remove_exceeded(mz, mctz);
712 * Insert again. mz->usage_in_excess will be updated.
713 * If excess is 0, no tree ops.
715 __mem_cgroup_insert_exceeded(mz, mctz, excess);
716 spin_unlock_irqrestore(&mctz->lock, flags);
721 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
723 struct mem_cgroup_tree_per_zone *mctz;
724 struct mem_cgroup_per_zone *mz;
728 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
729 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
730 mctz = soft_limit_tree_node_zone(nid, zid);
731 mem_cgroup_remove_exceeded(mz, mctz);
736 static struct mem_cgroup_per_zone *
737 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
739 struct rb_node *rightmost = NULL;
740 struct mem_cgroup_per_zone *mz;
744 rightmost = rb_last(&mctz->rb_root);
746 goto done; /* Nothing to reclaim from */
748 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
750 * Remove the node now but someone else can add it back,
751 * we will to add it back at the end of reclaim to its correct
752 * position in the tree.
754 __mem_cgroup_remove_exceeded(mz, mctz);
755 if (!soft_limit_excess(mz->memcg) ||
756 !css_tryget_online(&mz->memcg->css))
762 static struct mem_cgroup_per_zone *
763 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
765 struct mem_cgroup_per_zone *mz;
767 spin_lock_irq(&mctz->lock);
768 mz = __mem_cgroup_largest_soft_limit_node(mctz);
769 spin_unlock_irq(&mctz->lock);
774 * Implementation Note: reading percpu statistics for memcg.
776 * Both of vmstat[] and percpu_counter has threshold and do periodic
777 * synchronization to implement "quick" read. There are trade-off between
778 * reading cost and precision of value. Then, we may have a chance to implement
779 * a periodic synchronizion of counter in memcg's counter.
781 * But this _read() function is used for user interface now. The user accounts
782 * memory usage by memory cgroup and he _always_ requires exact value because
783 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
784 * have to visit all online cpus and make sum. So, for now, unnecessary
785 * synchronization is not implemented. (just implemented for cpu hotplug)
787 * If there are kernel internal actions which can make use of some not-exact
788 * value, and reading all cpu value can be performance bottleneck in some
789 * common workload, threashold and synchonization as vmstat[] should be
792 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
793 enum mem_cgroup_stat_index idx)
799 for_each_online_cpu(cpu)
800 val += per_cpu(memcg->stat->count[idx], cpu);
801 #ifdef CONFIG_HOTPLUG_CPU
802 spin_lock(&memcg->pcp_counter_lock);
803 val += memcg->nocpu_base.count[idx];
804 spin_unlock(&memcg->pcp_counter_lock);
810 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
811 enum mem_cgroup_events_index idx)
813 unsigned long val = 0;
817 for_each_online_cpu(cpu)
818 val += per_cpu(memcg->stat->events[idx], cpu);
819 #ifdef CONFIG_HOTPLUG_CPU
820 spin_lock(&memcg->pcp_counter_lock);
821 val += memcg->nocpu_base.events[idx];
822 spin_unlock(&memcg->pcp_counter_lock);
828 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
833 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
834 * counted as CACHE even if it's on ANON LRU.
837 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
840 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
843 if (PageTransHuge(page))
844 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
847 /* pagein of a big page is an event. So, ignore page size */
849 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
851 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
852 nr_pages = -nr_pages; /* for event */
855 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
858 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
860 struct mem_cgroup_per_zone *mz;
862 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
863 return mz->lru_size[lru];
866 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
868 unsigned int lru_mask)
870 unsigned long nr = 0;
873 VM_BUG_ON((unsigned)nid >= nr_node_ids);
875 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
876 struct mem_cgroup_per_zone *mz;
880 if (!(BIT(lru) & lru_mask))
882 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
883 nr += mz->lru_size[lru];
889 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
890 unsigned int lru_mask)
892 unsigned long nr = 0;
895 for_each_node_state(nid, N_MEMORY)
896 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
900 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
901 enum mem_cgroup_events_target target)
903 unsigned long val, next;
905 val = __this_cpu_read(memcg->stat->nr_page_events);
906 next = __this_cpu_read(memcg->stat->targets[target]);
907 /* from time_after() in jiffies.h */
908 if ((long)next - (long)val < 0) {
910 case MEM_CGROUP_TARGET_THRESH:
911 next = val + THRESHOLDS_EVENTS_TARGET;
913 case MEM_CGROUP_TARGET_SOFTLIMIT:
914 next = val + SOFTLIMIT_EVENTS_TARGET;
916 case MEM_CGROUP_TARGET_NUMAINFO:
917 next = val + NUMAINFO_EVENTS_TARGET;
922 __this_cpu_write(memcg->stat->targets[target], next);
929 * Check events in order.
932 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
934 /* threshold event is triggered in finer grain than soft limit */
935 if (unlikely(mem_cgroup_event_ratelimit(memcg,
936 MEM_CGROUP_TARGET_THRESH))) {
938 bool do_numainfo __maybe_unused;
940 do_softlimit = mem_cgroup_event_ratelimit(memcg,
941 MEM_CGROUP_TARGET_SOFTLIMIT);
943 do_numainfo = mem_cgroup_event_ratelimit(memcg,
944 MEM_CGROUP_TARGET_NUMAINFO);
946 mem_cgroup_threshold(memcg);
947 if (unlikely(do_softlimit))
948 mem_cgroup_update_tree(memcg, page);
950 if (unlikely(do_numainfo))
951 atomic_inc(&memcg->numainfo_events);
956 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
959 * mm_update_next_owner() may clear mm->owner to NULL
960 * if it races with swapoff, page migration, etc.
961 * So this can be called with p == NULL.
966 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
969 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
971 struct mem_cgroup *memcg = NULL;
976 * Page cache insertions can happen withou an
977 * actual mm context, e.g. during disk probing
978 * on boot, loopback IO, acct() writes etc.
981 memcg = root_mem_cgroup;
983 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
984 if (unlikely(!memcg))
985 memcg = root_mem_cgroup;
987 } while (!css_tryget_online(&memcg->css));
993 * mem_cgroup_iter - iterate over memory cgroup hierarchy
994 * @root: hierarchy root
995 * @prev: previously returned memcg, NULL on first invocation
996 * @reclaim: cookie for shared reclaim walks, NULL for full walks
998 * Returns references to children of the hierarchy below @root, or
999 * @root itself, or %NULL after a full round-trip.
1001 * Caller must pass the return value in @prev on subsequent
1002 * invocations for reference counting, or use mem_cgroup_iter_break()
1003 * to cancel a hierarchy walk before the round-trip is complete.
1005 * Reclaimers can specify a zone and a priority level in @reclaim to
1006 * divide up the memcgs in the hierarchy among all concurrent
1007 * reclaimers operating on the same zone and priority.
1009 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1010 struct mem_cgroup *prev,
1011 struct mem_cgroup_reclaim_cookie *reclaim)
1013 struct reclaim_iter *uninitialized_var(iter);
1014 struct cgroup_subsys_state *css = NULL;
1015 struct mem_cgroup *memcg = NULL;
1016 struct mem_cgroup *pos = NULL;
1018 if (mem_cgroup_disabled())
1022 root = root_mem_cgroup;
1024 if (prev && !reclaim)
1027 if (!root->use_hierarchy && root != root_mem_cgroup) {
1036 struct mem_cgroup_per_zone *mz;
1038 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1039 iter = &mz->iter[reclaim->priority];
1041 if (prev && reclaim->generation != iter->generation)
1045 pos = READ_ONCE(iter->position);
1047 * A racing update may change the position and
1048 * put the last reference, hence css_tryget(),
1049 * or retry to see the updated position.
1051 } while (pos && !css_tryget(&pos->css));
1058 css = css_next_descendant_pre(css, &root->css);
1061 * Reclaimers share the hierarchy walk, and a
1062 * new one might jump in right at the end of
1063 * the hierarchy - make sure they see at least
1064 * one group and restart from the beginning.
1072 * Verify the css and acquire a reference. The root
1073 * is provided by the caller, so we know it's alive
1074 * and kicking, and don't take an extra reference.
1076 memcg = mem_cgroup_from_css(css);
1078 if (css == &root->css)
1081 if (css_tryget(css)) {
1083 * Make sure the memcg is initialized:
1084 * mem_cgroup_css_online() orders the the
1085 * initialization against setting the flag.
1087 if (smp_load_acquire(&memcg->initialized))
1097 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1099 css_get(&memcg->css);
1105 * pairs with css_tryget when dereferencing iter->position
1114 reclaim->generation = iter->generation;
1120 if (prev && prev != root)
1121 css_put(&prev->css);
1127 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1128 * @root: hierarchy root
1129 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1131 void mem_cgroup_iter_break(struct mem_cgroup *root,
1132 struct mem_cgroup *prev)
1135 root = root_mem_cgroup;
1136 if (prev && prev != root)
1137 css_put(&prev->css);
1141 * Iteration constructs for visiting all cgroups (under a tree). If
1142 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1143 * be used for reference counting.
1145 #define for_each_mem_cgroup_tree(iter, root) \
1146 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1148 iter = mem_cgroup_iter(root, iter, NULL))
1150 #define for_each_mem_cgroup(iter) \
1151 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1153 iter = mem_cgroup_iter(NULL, iter, NULL))
1155 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1157 struct mem_cgroup *memcg;
1160 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1161 if (unlikely(!memcg))
1166 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1169 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1177 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1180 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1181 * @zone: zone of the wanted lruvec
1182 * @memcg: memcg of the wanted lruvec
1184 * Returns the lru list vector holding pages for the given @zone and
1185 * @mem. This can be the global zone lruvec, if the memory controller
1188 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1189 struct mem_cgroup *memcg)
1191 struct mem_cgroup_per_zone *mz;
1192 struct lruvec *lruvec;
1194 if (mem_cgroup_disabled()) {
1195 lruvec = &zone->lruvec;
1199 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1200 lruvec = &mz->lruvec;
1203 * Since a node can be onlined after the mem_cgroup was created,
1204 * we have to be prepared to initialize lruvec->zone here;
1205 * and if offlined then reonlined, we need to reinitialize it.
1207 if (unlikely(lruvec->zone != zone))
1208 lruvec->zone = zone;
1213 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1215 * @zone: zone of the page
1217 * This function is only safe when following the LRU page isolation
1218 * and putback protocol: the LRU lock must be held, and the page must
1219 * either be PageLRU() or the caller must have isolated/allocated it.
1221 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1223 struct mem_cgroup_per_zone *mz;
1224 struct mem_cgroup *memcg;
1225 struct lruvec *lruvec;
1227 if (mem_cgroup_disabled()) {
1228 lruvec = &zone->lruvec;
1232 memcg = page->mem_cgroup;
1234 * Swapcache readahead pages are added to the LRU - and
1235 * possibly migrated - before they are charged.
1238 memcg = root_mem_cgroup;
1240 mz = mem_cgroup_page_zoneinfo(memcg, page);
1241 lruvec = &mz->lruvec;
1244 * Since a node can be onlined after the mem_cgroup was created,
1245 * we have to be prepared to initialize lruvec->zone here;
1246 * and if offlined then reonlined, we need to reinitialize it.
1248 if (unlikely(lruvec->zone != zone))
1249 lruvec->zone = zone;
1254 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1255 * @lruvec: mem_cgroup per zone lru vector
1256 * @lru: index of lru list the page is sitting on
1257 * @nr_pages: positive when adding or negative when removing
1259 * This function must be called when a page is added to or removed from an
1262 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1265 struct mem_cgroup_per_zone *mz;
1266 unsigned long *lru_size;
1268 if (mem_cgroup_disabled())
1271 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1272 lru_size = mz->lru_size + lru;
1273 *lru_size += nr_pages;
1274 VM_BUG_ON((long)(*lru_size) < 0);
1277 bool mem_cgroup_is_descendant(struct mem_cgroup *memcg, struct mem_cgroup *root)
1281 if (!root->use_hierarchy)
1283 return cgroup_is_descendant(memcg->css.cgroup, root->css.cgroup);
1286 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1288 struct mem_cgroup *task_memcg;
1289 struct task_struct *p;
1292 p = find_lock_task_mm(task);
1294 task_memcg = get_mem_cgroup_from_mm(p->mm);
1298 * All threads may have already detached their mm's, but the oom
1299 * killer still needs to detect if they have already been oom
1300 * killed to prevent needlessly killing additional tasks.
1303 task_memcg = mem_cgroup_from_task(task);
1304 css_get(&task_memcg->css);
1307 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1308 css_put(&task_memcg->css);
1312 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1314 unsigned long inactive_ratio;
1315 unsigned long inactive;
1316 unsigned long active;
1319 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1320 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1322 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1324 inactive_ratio = int_sqrt(10 * gb);
1328 return inactive * inactive_ratio < active;
1331 bool mem_cgroup_lruvec_online(struct lruvec *lruvec)
1333 struct mem_cgroup_per_zone *mz;
1334 struct mem_cgroup *memcg;
1336 if (mem_cgroup_disabled())
1339 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1342 return !!(memcg->css.flags & CSS_ONLINE);
1345 #define mem_cgroup_from_counter(counter, member) \
1346 container_of(counter, struct mem_cgroup, member)
1349 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1350 * @memcg: the memory cgroup
1352 * Returns the maximum amount of memory @mem can be charged with, in
1355 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1357 unsigned long margin = 0;
1358 unsigned long count;
1359 unsigned long limit;
1361 count = page_counter_read(&memcg->memory);
1362 limit = READ_ONCE(memcg->memory.limit);
1364 margin = limit - count;
1366 if (do_swap_account) {
1367 count = page_counter_read(&memcg->memsw);
1368 limit = READ_ONCE(memcg->memsw.limit);
1370 margin = min(margin, limit - count);
1376 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1379 if (mem_cgroup_disabled() || !memcg->css.parent)
1380 return vm_swappiness;
1382 return memcg->swappiness;
1386 * A routine for checking "mem" is under move_account() or not.
1388 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1389 * moving cgroups. This is for waiting at high-memory pressure
1392 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1394 struct mem_cgroup *from;
1395 struct mem_cgroup *to;
1398 * Unlike task_move routines, we access mc.to, mc.from not under
1399 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1401 spin_lock(&mc.lock);
1407 ret = mem_cgroup_is_descendant(from, memcg) ||
1408 mem_cgroup_is_descendant(to, memcg);
1410 spin_unlock(&mc.lock);
1414 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1416 if (mc.moving_task && current != mc.moving_task) {
1417 if (mem_cgroup_under_move(memcg)) {
1419 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1420 /* moving charge context might have finished. */
1423 finish_wait(&mc.waitq, &wait);
1430 #define K(x) ((x) << (PAGE_SHIFT-10))
1432 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1433 * @memcg: The memory cgroup that went over limit
1434 * @p: Task that is going to be killed
1436 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1439 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1441 /* oom_info_lock ensures that parallel ooms do not interleave */
1442 static DEFINE_MUTEX(oom_info_lock);
1443 struct mem_cgroup *iter;
1446 mutex_lock(&oom_info_lock);
1450 pr_info("Task in ");
1451 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1452 pr_cont(" killed as a result of limit of ");
1454 pr_info("Memory limit reached of cgroup ");
1457 pr_cont_cgroup_path(memcg->css.cgroup);
1462 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1463 K((u64)page_counter_read(&memcg->memory)),
1464 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1465 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1466 K((u64)page_counter_read(&memcg->memsw)),
1467 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1468 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1469 K((u64)page_counter_read(&memcg->kmem)),
1470 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1472 for_each_mem_cgroup_tree(iter, memcg) {
1473 pr_info("Memory cgroup stats for ");
1474 pr_cont_cgroup_path(iter->css.cgroup);
1477 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1478 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1480 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1481 K(mem_cgroup_read_stat(iter, i)));
1484 for (i = 0; i < NR_LRU_LISTS; i++)
1485 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1486 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1490 mutex_unlock(&oom_info_lock);
1494 * This function returns the number of memcg under hierarchy tree. Returns
1495 * 1(self count) if no children.
1497 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1500 struct mem_cgroup *iter;
1502 for_each_mem_cgroup_tree(iter, memcg)
1508 * Return the memory (and swap, if configured) limit for a memcg.
1510 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1512 unsigned long limit;
1514 limit = memcg->memory.limit;
1515 if (mem_cgroup_swappiness(memcg)) {
1516 unsigned long memsw_limit;
1518 memsw_limit = memcg->memsw.limit;
1519 limit = min(limit + total_swap_pages, memsw_limit);
1524 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1527 struct mem_cgroup *iter;
1528 unsigned long chosen_points = 0;
1529 unsigned long totalpages;
1530 unsigned int points = 0;
1531 struct task_struct *chosen = NULL;
1533 mutex_lock(&oom_lock);
1536 * If current has a pending SIGKILL or is exiting, then automatically
1537 * select it. The goal is to allow it to allocate so that it may
1538 * quickly exit and free its memory.
1540 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1541 mark_oom_victim(current);
1545 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL, memcg);
1546 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1547 for_each_mem_cgroup_tree(iter, memcg) {
1548 struct css_task_iter it;
1549 struct task_struct *task;
1551 css_task_iter_start(&iter->css, &it);
1552 while ((task = css_task_iter_next(&it))) {
1553 switch (oom_scan_process_thread(task, totalpages, NULL,
1555 case OOM_SCAN_SELECT:
1557 put_task_struct(chosen);
1559 chosen_points = ULONG_MAX;
1560 get_task_struct(chosen);
1562 case OOM_SCAN_CONTINUE:
1564 case OOM_SCAN_ABORT:
1565 css_task_iter_end(&it);
1566 mem_cgroup_iter_break(memcg, iter);
1568 put_task_struct(chosen);
1573 points = oom_badness(task, memcg, NULL, totalpages);
1574 if (!points || points < chosen_points)
1576 /* Prefer thread group leaders for display purposes */
1577 if (points == chosen_points &&
1578 thread_group_leader(chosen))
1582 put_task_struct(chosen);
1584 chosen_points = points;
1585 get_task_struct(chosen);
1587 css_task_iter_end(&it);
1591 points = chosen_points * 1000 / totalpages;
1592 oom_kill_process(chosen, gfp_mask, order, points, totalpages,
1593 memcg, NULL, "Memory cgroup out of memory");
1596 mutex_unlock(&oom_lock);
1599 #if MAX_NUMNODES > 1
1602 * test_mem_cgroup_node_reclaimable
1603 * @memcg: the target memcg
1604 * @nid: the node ID to be checked.
1605 * @noswap : specify true here if the user wants flle only information.
1607 * This function returns whether the specified memcg contains any
1608 * reclaimable pages on a node. Returns true if there are any reclaimable
1609 * pages in the node.
1611 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1612 int nid, bool noswap)
1614 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1616 if (noswap || !total_swap_pages)
1618 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1625 * Always updating the nodemask is not very good - even if we have an empty
1626 * list or the wrong list here, we can start from some node and traverse all
1627 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1630 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1634 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1635 * pagein/pageout changes since the last update.
1637 if (!atomic_read(&memcg->numainfo_events))
1639 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1642 /* make a nodemask where this memcg uses memory from */
1643 memcg->scan_nodes = node_states[N_MEMORY];
1645 for_each_node_mask(nid, node_states[N_MEMORY]) {
1647 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1648 node_clear(nid, memcg->scan_nodes);
1651 atomic_set(&memcg->numainfo_events, 0);
1652 atomic_set(&memcg->numainfo_updating, 0);
1656 * Selecting a node where we start reclaim from. Because what we need is just
1657 * reducing usage counter, start from anywhere is O,K. Considering
1658 * memory reclaim from current node, there are pros. and cons.
1660 * Freeing memory from current node means freeing memory from a node which
1661 * we'll use or we've used. So, it may make LRU bad. And if several threads
1662 * hit limits, it will see a contention on a node. But freeing from remote
1663 * node means more costs for memory reclaim because of memory latency.
1665 * Now, we use round-robin. Better algorithm is welcomed.
1667 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1671 mem_cgroup_may_update_nodemask(memcg);
1672 node = memcg->last_scanned_node;
1674 node = next_node(node, memcg->scan_nodes);
1675 if (node == MAX_NUMNODES)
1676 node = first_node(memcg->scan_nodes);
1678 * We call this when we hit limit, not when pages are added to LRU.
1679 * No LRU may hold pages because all pages are UNEVICTABLE or
1680 * memcg is too small and all pages are not on LRU. In that case,
1681 * we use curret node.
1683 if (unlikely(node == MAX_NUMNODES))
1684 node = numa_node_id();
1686 memcg->last_scanned_node = node;
1690 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1696 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1699 unsigned long *total_scanned)
1701 struct mem_cgroup *victim = NULL;
1704 unsigned long excess;
1705 unsigned long nr_scanned;
1706 struct mem_cgroup_reclaim_cookie reclaim = {
1711 excess = soft_limit_excess(root_memcg);
1714 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1719 * If we have not been able to reclaim
1720 * anything, it might because there are
1721 * no reclaimable pages under this hierarchy
1726 * We want to do more targeted reclaim.
1727 * excess >> 2 is not to excessive so as to
1728 * reclaim too much, nor too less that we keep
1729 * coming back to reclaim from this cgroup
1731 if (total >= (excess >> 2) ||
1732 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1737 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1739 *total_scanned += nr_scanned;
1740 if (!soft_limit_excess(root_memcg))
1743 mem_cgroup_iter_break(root_memcg, victim);
1747 #ifdef CONFIG_LOCKDEP
1748 static struct lockdep_map memcg_oom_lock_dep_map = {
1749 .name = "memcg_oom_lock",
1753 static DEFINE_SPINLOCK(memcg_oom_lock);
1756 * Check OOM-Killer is already running under our hierarchy.
1757 * If someone is running, return false.
1759 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1761 struct mem_cgroup *iter, *failed = NULL;
1763 spin_lock(&memcg_oom_lock);
1765 for_each_mem_cgroup_tree(iter, memcg) {
1766 if (iter->oom_lock) {
1768 * this subtree of our hierarchy is already locked
1769 * so we cannot give a lock.
1772 mem_cgroup_iter_break(memcg, iter);
1775 iter->oom_lock = true;
1780 * OK, we failed to lock the whole subtree so we have
1781 * to clean up what we set up to the failing subtree
1783 for_each_mem_cgroup_tree(iter, memcg) {
1784 if (iter == failed) {
1785 mem_cgroup_iter_break(memcg, iter);
1788 iter->oom_lock = false;
1791 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1793 spin_unlock(&memcg_oom_lock);
1798 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1800 struct mem_cgroup *iter;
1802 spin_lock(&memcg_oom_lock);
1803 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1804 for_each_mem_cgroup_tree(iter, memcg)
1805 iter->oom_lock = false;
1806 spin_unlock(&memcg_oom_lock);
1809 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1811 struct mem_cgroup *iter;
1813 for_each_mem_cgroup_tree(iter, memcg)
1814 atomic_inc(&iter->under_oom);
1817 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1819 struct mem_cgroup *iter;
1822 * When a new child is created while the hierarchy is under oom,
1823 * mem_cgroup_oom_lock() may not be called. We have to use
1824 * atomic_add_unless() here.
1826 for_each_mem_cgroup_tree(iter, memcg)
1827 atomic_add_unless(&iter->under_oom, -1, 0);
1830 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1832 struct oom_wait_info {
1833 struct mem_cgroup *memcg;
1837 static int memcg_oom_wake_function(wait_queue_t *wait,
1838 unsigned mode, int sync, void *arg)
1840 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1841 struct mem_cgroup *oom_wait_memcg;
1842 struct oom_wait_info *oom_wait_info;
1844 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1845 oom_wait_memcg = oom_wait_info->memcg;
1847 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1848 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1850 return autoremove_wake_function(wait, mode, sync, arg);
1853 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1855 atomic_inc(&memcg->oom_wakeups);
1856 /* for filtering, pass "memcg" as argument. */
1857 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1860 static void memcg_oom_recover(struct mem_cgroup *memcg)
1862 if (memcg && atomic_read(&memcg->under_oom))
1863 memcg_wakeup_oom(memcg);
1866 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1868 if (!current->memcg_oom.may_oom)
1871 * We are in the middle of the charge context here, so we
1872 * don't want to block when potentially sitting on a callstack
1873 * that holds all kinds of filesystem and mm locks.
1875 * Also, the caller may handle a failed allocation gracefully
1876 * (like optional page cache readahead) and so an OOM killer
1877 * invocation might not even be necessary.
1879 * That's why we don't do anything here except remember the
1880 * OOM context and then deal with it at the end of the page
1881 * fault when the stack is unwound, the locks are released,
1882 * and when we know whether the fault was overall successful.
1884 css_get(&memcg->css);
1885 current->memcg_oom.memcg = memcg;
1886 current->memcg_oom.gfp_mask = mask;
1887 current->memcg_oom.order = order;
1891 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1892 * @handle: actually kill/wait or just clean up the OOM state
1894 * This has to be called at the end of a page fault if the memcg OOM
1895 * handler was enabled.
1897 * Memcg supports userspace OOM handling where failed allocations must
1898 * sleep on a waitqueue until the userspace task resolves the
1899 * situation. Sleeping directly in the charge context with all kinds
1900 * of locks held is not a good idea, instead we remember an OOM state
1901 * in the task and mem_cgroup_oom_synchronize() has to be called at
1902 * the end of the page fault to complete the OOM handling.
1904 * Returns %true if an ongoing memcg OOM situation was detected and
1905 * completed, %false otherwise.
1907 bool mem_cgroup_oom_synchronize(bool handle)
1909 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1910 struct oom_wait_info owait;
1913 /* OOM is global, do not handle */
1917 if (!handle || oom_killer_disabled)
1920 owait.memcg = memcg;
1921 owait.wait.flags = 0;
1922 owait.wait.func = memcg_oom_wake_function;
1923 owait.wait.private = current;
1924 INIT_LIST_HEAD(&owait.wait.task_list);
1926 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1927 mem_cgroup_mark_under_oom(memcg);
1929 locked = mem_cgroup_oom_trylock(memcg);
1932 mem_cgroup_oom_notify(memcg);
1934 if (locked && !memcg->oom_kill_disable) {
1935 mem_cgroup_unmark_under_oom(memcg);
1936 finish_wait(&memcg_oom_waitq, &owait.wait);
1937 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1938 current->memcg_oom.order);
1941 mem_cgroup_unmark_under_oom(memcg);
1942 finish_wait(&memcg_oom_waitq, &owait.wait);
1946 mem_cgroup_oom_unlock(memcg);
1948 * There is no guarantee that an OOM-lock contender
1949 * sees the wakeups triggered by the OOM kill
1950 * uncharges. Wake any sleepers explicitely.
1952 memcg_oom_recover(memcg);
1955 current->memcg_oom.memcg = NULL;
1956 css_put(&memcg->css);
1961 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1962 * @page: page that is going to change accounted state
1964 * This function must mark the beginning of an accounted page state
1965 * change to prevent double accounting when the page is concurrently
1966 * being moved to another memcg:
1968 * memcg = mem_cgroup_begin_page_stat(page);
1969 * if (TestClearPageState(page))
1970 * mem_cgroup_update_page_stat(memcg, state, -1);
1971 * mem_cgroup_end_page_stat(memcg);
1973 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1975 struct mem_cgroup *memcg;
1976 unsigned long flags;
1979 * The RCU lock is held throughout the transaction. The fast
1980 * path can get away without acquiring the memcg->move_lock
1981 * because page moving starts with an RCU grace period.
1983 * The RCU lock also protects the memcg from being freed when
1984 * the page state that is going to change is the only thing
1985 * preventing the page from being uncharged.
1986 * E.g. end-writeback clearing PageWriteback(), which allows
1987 * migration to go ahead and uncharge the page before the
1988 * account transaction might be complete.
1992 if (mem_cgroup_disabled())
1995 memcg = page->mem_cgroup;
1996 if (unlikely(!memcg))
1999 if (atomic_read(&memcg->moving_account) <= 0)
2002 spin_lock_irqsave(&memcg->move_lock, flags);
2003 if (memcg != page->mem_cgroup) {
2004 spin_unlock_irqrestore(&memcg->move_lock, flags);
2009 * When charge migration first begins, we can have locked and
2010 * unlocked page stat updates happening concurrently. Track
2011 * the task who has the lock for mem_cgroup_end_page_stat().
2013 memcg->move_lock_task = current;
2014 memcg->move_lock_flags = flags;
2020 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2021 * @memcg: the memcg that was accounted against
2023 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
2025 if (memcg && memcg->move_lock_task == current) {
2026 unsigned long flags = memcg->move_lock_flags;
2028 memcg->move_lock_task = NULL;
2029 memcg->move_lock_flags = 0;
2031 spin_unlock_irqrestore(&memcg->move_lock, flags);
2038 * mem_cgroup_update_page_stat - update page state statistics
2039 * @memcg: memcg to account against
2040 * @idx: page state item to account
2041 * @val: number of pages (positive or negative)
2043 * See mem_cgroup_begin_page_stat() for locking requirements.
2045 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2046 enum mem_cgroup_stat_index idx, int val)
2048 VM_BUG_ON(!rcu_read_lock_held());
2051 this_cpu_add(memcg->stat->count[idx], val);
2055 * size of first charge trial. "32" comes from vmscan.c's magic value.
2056 * TODO: maybe necessary to use big numbers in big irons.
2058 #define CHARGE_BATCH 32U
2059 struct memcg_stock_pcp {
2060 struct mem_cgroup *cached; /* this never be root cgroup */
2061 unsigned int nr_pages;
2062 struct work_struct work;
2063 unsigned long flags;
2064 #define FLUSHING_CACHED_CHARGE 0
2066 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2067 static DEFINE_MUTEX(percpu_charge_mutex);
2070 * consume_stock: Try to consume stocked charge on this cpu.
2071 * @memcg: memcg to consume from.
2072 * @nr_pages: how many pages to charge.
2074 * The charges will only happen if @memcg matches the current cpu's memcg
2075 * stock, and at least @nr_pages are available in that stock. Failure to
2076 * service an allocation will refill the stock.
2078 * returns true if successful, false otherwise.
2080 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2082 struct memcg_stock_pcp *stock;
2085 if (nr_pages > CHARGE_BATCH)
2088 stock = &get_cpu_var(memcg_stock);
2089 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2090 stock->nr_pages -= nr_pages;
2093 put_cpu_var(memcg_stock);
2098 * Returns stocks cached in percpu and reset cached information.
2100 static void drain_stock(struct memcg_stock_pcp *stock)
2102 struct mem_cgroup *old = stock->cached;
2104 if (stock->nr_pages) {
2105 page_counter_uncharge(&old->memory, stock->nr_pages);
2106 if (do_swap_account)
2107 page_counter_uncharge(&old->memsw, stock->nr_pages);
2108 css_put_many(&old->css, stock->nr_pages);
2109 stock->nr_pages = 0;
2111 stock->cached = NULL;
2115 * This must be called under preempt disabled or must be called by
2116 * a thread which is pinned to local cpu.
2118 static void drain_local_stock(struct work_struct *dummy)
2120 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2122 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2126 * Cache charges(val) to local per_cpu area.
2127 * This will be consumed by consume_stock() function, later.
2129 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2131 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2133 if (stock->cached != memcg) { /* reset if necessary */
2135 stock->cached = memcg;
2137 stock->nr_pages += nr_pages;
2138 put_cpu_var(memcg_stock);
2142 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2143 * of the hierarchy under it.
2145 static void drain_all_stock(struct mem_cgroup *root_memcg)
2149 /* If someone's already draining, avoid adding running more workers. */
2150 if (!mutex_trylock(&percpu_charge_mutex))
2152 /* Notify other cpus that system-wide "drain" is running */
2155 for_each_online_cpu(cpu) {
2156 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2157 struct mem_cgroup *memcg;
2159 memcg = stock->cached;
2160 if (!memcg || !stock->nr_pages)
2162 if (!mem_cgroup_is_descendant(memcg, root_memcg))
2164 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2166 drain_local_stock(&stock->work);
2168 schedule_work_on(cpu, &stock->work);
2173 mutex_unlock(&percpu_charge_mutex);
2177 * This function drains percpu counter value from DEAD cpu and
2178 * move it to local cpu. Note that this function can be preempted.
2180 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2184 spin_lock(&memcg->pcp_counter_lock);
2185 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2186 long x = per_cpu(memcg->stat->count[i], cpu);
2188 per_cpu(memcg->stat->count[i], cpu) = 0;
2189 memcg->nocpu_base.count[i] += x;
2191 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2192 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2194 per_cpu(memcg->stat->events[i], cpu) = 0;
2195 memcg->nocpu_base.events[i] += x;
2197 spin_unlock(&memcg->pcp_counter_lock);
2200 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2201 unsigned long action,
2204 int cpu = (unsigned long)hcpu;
2205 struct memcg_stock_pcp *stock;
2206 struct mem_cgroup *iter;
2208 if (action == CPU_ONLINE)
2211 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2214 for_each_mem_cgroup(iter)
2215 mem_cgroup_drain_pcp_counter(iter, cpu);
2217 stock = &per_cpu(memcg_stock, cpu);
2222 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2223 unsigned int nr_pages)
2225 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2226 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2227 struct mem_cgroup *mem_over_limit;
2228 struct page_counter *counter;
2229 unsigned long nr_reclaimed;
2230 bool may_swap = true;
2231 bool drained = false;
2234 if (mem_cgroup_is_root(memcg))
2237 if (consume_stock(memcg, nr_pages))
2240 if (!do_swap_account ||
2241 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2242 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2244 if (do_swap_account)
2245 page_counter_uncharge(&memcg->memsw, batch);
2246 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2248 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2252 if (batch > nr_pages) {
2258 * Unlike in global OOM situations, memcg is not in a physical
2259 * memory shortage. Allow dying and OOM-killed tasks to
2260 * bypass the last charges so that they can exit quickly and
2261 * free their memory.
2263 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2264 fatal_signal_pending(current) ||
2265 current->flags & PF_EXITING))
2268 if (unlikely(task_in_memcg_oom(current)))
2271 if (!(gfp_mask & __GFP_WAIT))
2274 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2276 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2277 gfp_mask, may_swap);
2279 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2283 drain_all_stock(mem_over_limit);
2288 if (gfp_mask & __GFP_NORETRY)
2291 * Even though the limit is exceeded at this point, reclaim
2292 * may have been able to free some pages. Retry the charge
2293 * before killing the task.
2295 * Only for regular pages, though: huge pages are rather
2296 * unlikely to succeed so close to the limit, and we fall back
2297 * to regular pages anyway in case of failure.
2299 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2302 * At task move, charge accounts can be doubly counted. So, it's
2303 * better to wait until the end of task_move if something is going on.
2305 if (mem_cgroup_wait_acct_move(mem_over_limit))
2311 if (gfp_mask & __GFP_NOFAIL)
2314 if (fatal_signal_pending(current))
2317 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2319 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2321 if (!(gfp_mask & __GFP_NOFAIL))
2327 css_get_many(&memcg->css, batch);
2328 if (batch > nr_pages)
2329 refill_stock(memcg, batch - nr_pages);
2330 if (!(gfp_mask & __GFP_WAIT))
2333 * If the hierarchy is above the normal consumption range,
2334 * make the charging task trim their excess contribution.
2337 if (page_counter_read(&memcg->memory) <= memcg->high)
2339 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
2340 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2341 } while ((memcg = parent_mem_cgroup(memcg)));
2346 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2348 if (mem_cgroup_is_root(memcg))
2351 page_counter_uncharge(&memcg->memory, nr_pages);
2352 if (do_swap_account)
2353 page_counter_uncharge(&memcg->memsw, nr_pages);
2355 css_put_many(&memcg->css, nr_pages);
2359 * try_get_mem_cgroup_from_page - look up page's memcg association
2362 * Look up, get a css reference, and return the memcg that owns @page.
2364 * The page must be locked to prevent racing with swap-in and page
2365 * cache charges. If coming from an unlocked page table, the caller
2366 * must ensure the page is on the LRU or this can race with charging.
2368 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2370 struct mem_cgroup *memcg;
2374 VM_BUG_ON_PAGE(!PageLocked(page), page);
2376 memcg = page->mem_cgroup;
2378 if (!css_tryget_online(&memcg->css))
2380 } else if (PageSwapCache(page)) {
2381 ent.val = page_private(page);
2382 id = lookup_swap_cgroup_id(ent);
2384 memcg = mem_cgroup_from_id(id);
2385 if (memcg && !css_tryget_online(&memcg->css))
2392 static void lock_page_lru(struct page *page, int *isolated)
2394 struct zone *zone = page_zone(page);
2396 spin_lock_irq(&zone->lru_lock);
2397 if (PageLRU(page)) {
2398 struct lruvec *lruvec;
2400 lruvec = mem_cgroup_page_lruvec(page, zone);
2402 del_page_from_lru_list(page, lruvec, page_lru(page));
2408 static void unlock_page_lru(struct page *page, int isolated)
2410 struct zone *zone = page_zone(page);
2413 struct lruvec *lruvec;
2415 lruvec = mem_cgroup_page_lruvec(page, zone);
2416 VM_BUG_ON_PAGE(PageLRU(page), page);
2418 add_page_to_lru_list(page, lruvec, page_lru(page));
2420 spin_unlock_irq(&zone->lru_lock);
2423 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2428 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2431 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2432 * may already be on some other mem_cgroup's LRU. Take care of it.
2435 lock_page_lru(page, &isolated);
2438 * Nobody should be changing or seriously looking at
2439 * page->mem_cgroup at this point:
2441 * - the page is uncharged
2443 * - the page is off-LRU
2445 * - an anonymous fault has exclusive page access, except for
2446 * a locked page table
2448 * - a page cache insertion, a swapin fault, or a migration
2449 * have the page locked
2451 page->mem_cgroup = memcg;
2454 unlock_page_lru(page, isolated);
2457 #ifdef CONFIG_MEMCG_KMEM
2458 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2459 unsigned long nr_pages)
2461 struct page_counter *counter;
2464 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2468 ret = try_charge(memcg, gfp, nr_pages);
2469 if (ret == -EINTR) {
2471 * try_charge() chose to bypass to root due to OOM kill or
2472 * fatal signal. Since our only options are to either fail
2473 * the allocation or charge it to this cgroup, do it as a
2474 * temporary condition. But we can't fail. From a kmem/slab
2475 * perspective, the cache has already been selected, by
2476 * mem_cgroup_kmem_get_cache(), so it is too late to change
2479 * This condition will only trigger if the task entered
2480 * memcg_charge_kmem in a sane state, but was OOM-killed
2481 * during try_charge() above. Tasks that were already dying
2482 * when the allocation triggers should have been already
2483 * directed to the root cgroup in memcontrol.h
2485 page_counter_charge(&memcg->memory, nr_pages);
2486 if (do_swap_account)
2487 page_counter_charge(&memcg->memsw, nr_pages);
2488 css_get_many(&memcg->css, nr_pages);
2491 page_counter_uncharge(&memcg->kmem, nr_pages);
2496 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2498 page_counter_uncharge(&memcg->memory, nr_pages);
2499 if (do_swap_account)
2500 page_counter_uncharge(&memcg->memsw, nr_pages);
2502 page_counter_uncharge(&memcg->kmem, nr_pages);
2504 css_put_many(&memcg->css, nr_pages);
2508 * helper for acessing a memcg's index. It will be used as an index in the
2509 * child cache array in kmem_cache, and also to derive its name. This function
2510 * will return -1 when this is not a kmem-limited memcg.
2512 int memcg_cache_id(struct mem_cgroup *memcg)
2514 return memcg ? memcg->kmemcg_id : -1;
2517 static int memcg_alloc_cache_id(void)
2522 id = ida_simple_get(&memcg_cache_ida,
2523 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2527 if (id < memcg_nr_cache_ids)
2531 * There's no space for the new id in memcg_caches arrays,
2532 * so we have to grow them.
2534 down_write(&memcg_cache_ids_sem);
2536 size = 2 * (id + 1);
2537 if (size < MEMCG_CACHES_MIN_SIZE)
2538 size = MEMCG_CACHES_MIN_SIZE;
2539 else if (size > MEMCG_CACHES_MAX_SIZE)
2540 size = MEMCG_CACHES_MAX_SIZE;
2542 err = memcg_update_all_caches(size);
2544 err = memcg_update_all_list_lrus(size);
2546 memcg_nr_cache_ids = size;
2548 up_write(&memcg_cache_ids_sem);
2551 ida_simple_remove(&memcg_cache_ida, id);
2557 static void memcg_free_cache_id(int id)
2559 ida_simple_remove(&memcg_cache_ida, id);
2562 struct memcg_kmem_cache_create_work {
2563 struct mem_cgroup *memcg;
2564 struct kmem_cache *cachep;
2565 struct work_struct work;
2568 static void memcg_kmem_cache_create_func(struct work_struct *w)
2570 struct memcg_kmem_cache_create_work *cw =
2571 container_of(w, struct memcg_kmem_cache_create_work, work);
2572 struct mem_cgroup *memcg = cw->memcg;
2573 struct kmem_cache *cachep = cw->cachep;
2575 memcg_create_kmem_cache(memcg, cachep);
2577 css_put(&memcg->css);
2582 * Enqueue the creation of a per-memcg kmem_cache.
2584 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2585 struct kmem_cache *cachep)
2587 struct memcg_kmem_cache_create_work *cw;
2589 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2593 css_get(&memcg->css);
2596 cw->cachep = cachep;
2597 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2599 schedule_work(&cw->work);
2602 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2603 struct kmem_cache *cachep)
2606 * We need to stop accounting when we kmalloc, because if the
2607 * corresponding kmalloc cache is not yet created, the first allocation
2608 * in __memcg_schedule_kmem_cache_create will recurse.
2610 * However, it is better to enclose the whole function. Depending on
2611 * the debugging options enabled, INIT_WORK(), for instance, can
2612 * trigger an allocation. This too, will make us recurse. Because at
2613 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2614 * the safest choice is to do it like this, wrapping the whole function.
2616 current->memcg_kmem_skip_account = 1;
2617 __memcg_schedule_kmem_cache_create(memcg, cachep);
2618 current->memcg_kmem_skip_account = 0;
2622 * Return the kmem_cache we're supposed to use for a slab allocation.
2623 * We try to use the current memcg's version of the cache.
2625 * If the cache does not exist yet, if we are the first user of it,
2626 * we either create it immediately, if possible, or create it asynchronously
2628 * In the latter case, we will let the current allocation go through with
2629 * the original cache.
2631 * Can't be called in interrupt context or from kernel threads.
2632 * This function needs to be called with rcu_read_lock() held.
2634 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2636 struct mem_cgroup *memcg;
2637 struct kmem_cache *memcg_cachep;
2640 VM_BUG_ON(!is_root_cache(cachep));
2642 if (current->memcg_kmem_skip_account)
2645 memcg = get_mem_cgroup_from_mm(current->mm);
2646 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2650 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2651 if (likely(memcg_cachep))
2652 return memcg_cachep;
2655 * If we are in a safe context (can wait, and not in interrupt
2656 * context), we could be be predictable and return right away.
2657 * This would guarantee that the allocation being performed
2658 * already belongs in the new cache.
2660 * However, there are some clashes that can arrive from locking.
2661 * For instance, because we acquire the slab_mutex while doing
2662 * memcg_create_kmem_cache, this means no further allocation
2663 * could happen with the slab_mutex held. So it's better to
2666 memcg_schedule_kmem_cache_create(memcg, cachep);
2668 css_put(&memcg->css);
2672 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2674 if (!is_root_cache(cachep))
2675 css_put(&cachep->memcg_params.memcg->css);
2679 * We need to verify if the allocation against current->mm->owner's memcg is
2680 * possible for the given order. But the page is not allocated yet, so we'll
2681 * need a further commit step to do the final arrangements.
2683 * It is possible for the task to switch cgroups in this mean time, so at
2684 * commit time, we can't rely on task conversion any longer. We'll then use
2685 * the handle argument to return to the caller which cgroup we should commit
2686 * against. We could also return the memcg directly and avoid the pointer
2687 * passing, but a boolean return value gives better semantics considering
2688 * the compiled-out case as well.
2690 * Returning true means the allocation is possible.
2693 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2695 struct mem_cgroup *memcg;
2700 memcg = get_mem_cgroup_from_mm(current->mm);
2702 if (!memcg_kmem_is_active(memcg)) {
2703 css_put(&memcg->css);
2707 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2711 css_put(&memcg->css);
2715 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2718 VM_BUG_ON(mem_cgroup_is_root(memcg));
2720 /* The page allocation failed. Revert */
2722 memcg_uncharge_kmem(memcg, 1 << order);
2725 page->mem_cgroup = memcg;
2728 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2730 struct mem_cgroup *memcg = page->mem_cgroup;
2735 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2737 memcg_uncharge_kmem(memcg, 1 << order);
2738 page->mem_cgroup = NULL;
2741 struct mem_cgroup *__mem_cgroup_from_kmem(void *ptr)
2743 struct mem_cgroup *memcg = NULL;
2744 struct kmem_cache *cachep;
2747 page = virt_to_head_page(ptr);
2748 if (PageSlab(page)) {
2749 cachep = page->slab_cache;
2750 if (!is_root_cache(cachep))
2751 memcg = cachep->memcg_params.memcg;
2753 /* page allocated by alloc_kmem_pages */
2754 memcg = page->mem_cgroup;
2758 #endif /* CONFIG_MEMCG_KMEM */
2760 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2763 * Because tail pages are not marked as "used", set it. We're under
2764 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2765 * charge/uncharge will be never happen and move_account() is done under
2766 * compound_lock(), so we don't have to take care of races.
2768 void mem_cgroup_split_huge_fixup(struct page *head)
2772 if (mem_cgroup_disabled())
2775 for (i = 1; i < HPAGE_PMD_NR; i++)
2776 head[i].mem_cgroup = head->mem_cgroup;
2778 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2781 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2783 #ifdef CONFIG_MEMCG_SWAP
2784 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2787 int val = (charge) ? 1 : -1;
2788 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2792 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2793 * @entry: swap entry to be moved
2794 * @from: mem_cgroup which the entry is moved from
2795 * @to: mem_cgroup which the entry is moved to
2797 * It succeeds only when the swap_cgroup's record for this entry is the same
2798 * as the mem_cgroup's id of @from.
2800 * Returns 0 on success, -EINVAL on failure.
2802 * The caller must have charged to @to, IOW, called page_counter_charge() about
2803 * both res and memsw, and called css_get().
2805 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2806 struct mem_cgroup *from, struct mem_cgroup *to)
2808 unsigned short old_id, new_id;
2810 old_id = mem_cgroup_id(from);
2811 new_id = mem_cgroup_id(to);
2813 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2814 mem_cgroup_swap_statistics(from, false);
2815 mem_cgroup_swap_statistics(to, true);
2821 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2822 struct mem_cgroup *from, struct mem_cgroup *to)
2828 static DEFINE_MUTEX(memcg_limit_mutex);
2830 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2831 unsigned long limit)
2833 unsigned long curusage;
2834 unsigned long oldusage;
2835 bool enlarge = false;
2840 * For keeping hierarchical_reclaim simple, how long we should retry
2841 * is depends on callers. We set our retry-count to be function
2842 * of # of children which we should visit in this loop.
2844 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2845 mem_cgroup_count_children(memcg);
2847 oldusage = page_counter_read(&memcg->memory);
2850 if (signal_pending(current)) {
2855 mutex_lock(&memcg_limit_mutex);
2856 if (limit > memcg->memsw.limit) {
2857 mutex_unlock(&memcg_limit_mutex);
2861 if (limit > memcg->memory.limit)
2863 ret = page_counter_limit(&memcg->memory, limit);
2864 mutex_unlock(&memcg_limit_mutex);
2869 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2871 curusage = page_counter_read(&memcg->memory);
2872 /* Usage is reduced ? */
2873 if (curusage >= oldusage)
2876 oldusage = curusage;
2877 } while (retry_count);
2879 if (!ret && enlarge)
2880 memcg_oom_recover(memcg);
2885 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2886 unsigned long limit)
2888 unsigned long curusage;
2889 unsigned long oldusage;
2890 bool enlarge = false;
2894 /* see mem_cgroup_resize_res_limit */
2895 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2896 mem_cgroup_count_children(memcg);
2898 oldusage = page_counter_read(&memcg->memsw);
2901 if (signal_pending(current)) {
2906 mutex_lock(&memcg_limit_mutex);
2907 if (limit < memcg->memory.limit) {
2908 mutex_unlock(&memcg_limit_mutex);
2912 if (limit > memcg->memsw.limit)
2914 ret = page_counter_limit(&memcg->memsw, limit);
2915 mutex_unlock(&memcg_limit_mutex);
2920 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2922 curusage = page_counter_read(&memcg->memsw);
2923 /* Usage is reduced ? */
2924 if (curusage >= oldusage)
2927 oldusage = curusage;
2928 } while (retry_count);
2930 if (!ret && enlarge)
2931 memcg_oom_recover(memcg);
2936 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2938 unsigned long *total_scanned)
2940 unsigned long nr_reclaimed = 0;
2941 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2942 unsigned long reclaimed;
2944 struct mem_cgroup_tree_per_zone *mctz;
2945 unsigned long excess;
2946 unsigned long nr_scanned;
2951 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2953 * This loop can run a while, specially if mem_cgroup's continuously
2954 * keep exceeding their soft limit and putting the system under
2961 mz = mem_cgroup_largest_soft_limit_node(mctz);
2966 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2967 gfp_mask, &nr_scanned);
2968 nr_reclaimed += reclaimed;
2969 *total_scanned += nr_scanned;
2970 spin_lock_irq(&mctz->lock);
2971 __mem_cgroup_remove_exceeded(mz, mctz);
2974 * If we failed to reclaim anything from this memory cgroup
2975 * it is time to move on to the next cgroup
2979 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2981 excess = soft_limit_excess(mz->memcg);
2983 * One school of thought says that we should not add
2984 * back the node to the tree if reclaim returns 0.
2985 * But our reclaim could return 0, simply because due
2986 * to priority we are exposing a smaller subset of
2987 * memory to reclaim from. Consider this as a longer
2990 /* If excess == 0, no tree ops */
2991 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2992 spin_unlock_irq(&mctz->lock);
2993 css_put(&mz->memcg->css);
2996 * Could not reclaim anything and there are no more
2997 * mem cgroups to try or we seem to be looping without
2998 * reclaiming anything.
3000 if (!nr_reclaimed &&
3002 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3004 } while (!nr_reclaimed);
3006 css_put(&next_mz->memcg->css);
3007 return nr_reclaimed;
3011 * Test whether @memcg has children, dead or alive. Note that this
3012 * function doesn't care whether @memcg has use_hierarchy enabled and
3013 * returns %true if there are child csses according to the cgroup
3014 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3016 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3021 * The lock does not prevent addition or deletion of children, but
3022 * it prevents a new child from being initialized based on this
3023 * parent in css_online(), so it's enough to decide whether
3024 * hierarchically inherited attributes can still be changed or not.
3026 lockdep_assert_held(&memcg_create_mutex);
3029 ret = css_next_child(NULL, &memcg->css);
3035 * Reclaims as many pages from the given memcg as possible and moves
3036 * the rest to the parent.
3038 * Caller is responsible for holding css reference for memcg.
3040 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3042 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3044 /* we call try-to-free pages for make this cgroup empty */
3045 lru_add_drain_all();
3046 /* try to free all pages in this cgroup */
3047 while (nr_retries && page_counter_read(&memcg->memory)) {
3050 if (signal_pending(current))
3053 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3057 /* maybe some writeback is necessary */
3058 congestion_wait(BLK_RW_ASYNC, HZ/10);
3066 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3067 char *buf, size_t nbytes,
3070 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3072 if (mem_cgroup_is_root(memcg))
3074 return mem_cgroup_force_empty(memcg) ?: nbytes;
3077 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3080 return mem_cgroup_from_css(css)->use_hierarchy;
3083 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3084 struct cftype *cft, u64 val)
3087 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3088 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3090 mutex_lock(&memcg_create_mutex);
3092 if (memcg->use_hierarchy == val)
3096 * If parent's use_hierarchy is set, we can't make any modifications
3097 * in the child subtrees. If it is unset, then the change can
3098 * occur, provided the current cgroup has no children.
3100 * For the root cgroup, parent_mem is NULL, we allow value to be
3101 * set if there are no children.
3103 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3104 (val == 1 || val == 0)) {
3105 if (!memcg_has_children(memcg))
3106 memcg->use_hierarchy = val;
3113 mutex_unlock(&memcg_create_mutex);
3118 static unsigned long tree_stat(struct mem_cgroup *memcg,
3119 enum mem_cgroup_stat_index idx)
3121 struct mem_cgroup *iter;
3124 /* Per-cpu values can be negative, use a signed accumulator */
3125 for_each_mem_cgroup_tree(iter, memcg)
3126 val += mem_cgroup_read_stat(iter, idx);
3128 if (val < 0) /* race ? */
3133 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3137 if (mem_cgroup_is_root(memcg)) {
3138 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3139 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3141 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3144 val = page_counter_read(&memcg->memory);
3146 val = page_counter_read(&memcg->memsw);
3148 return val << PAGE_SHIFT;
3159 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3162 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3163 struct page_counter *counter;
3165 switch (MEMFILE_TYPE(cft->private)) {
3167 counter = &memcg->memory;
3170 counter = &memcg->memsw;
3173 counter = &memcg->kmem;
3179 switch (MEMFILE_ATTR(cft->private)) {
3181 if (counter == &memcg->memory)
3182 return mem_cgroup_usage(memcg, false);
3183 if (counter == &memcg->memsw)
3184 return mem_cgroup_usage(memcg, true);
3185 return (u64)page_counter_read(counter) * PAGE_SIZE;
3187 return (u64)counter->limit * PAGE_SIZE;
3189 return (u64)counter->watermark * PAGE_SIZE;
3191 return counter->failcnt;
3192 case RES_SOFT_LIMIT:
3193 return (u64)memcg->soft_limit * PAGE_SIZE;
3199 #ifdef CONFIG_MEMCG_KMEM
3200 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3201 unsigned long nr_pages)
3206 BUG_ON(memcg->kmemcg_id >= 0);
3207 BUG_ON(memcg->kmem_acct_activated);
3208 BUG_ON(memcg->kmem_acct_active);
3211 * For simplicity, we won't allow this to be disabled. It also can't
3212 * be changed if the cgroup has children already, or if tasks had
3215 * If tasks join before we set the limit, a person looking at
3216 * kmem.usage_in_bytes will have no way to determine when it took
3217 * place, which makes the value quite meaningless.
3219 * After it first became limited, changes in the value of the limit are
3220 * of course permitted.
3222 mutex_lock(&memcg_create_mutex);
3223 if (cgroup_has_tasks(memcg->css.cgroup) ||
3224 (memcg->use_hierarchy && memcg_has_children(memcg)))
3226 mutex_unlock(&memcg_create_mutex);
3230 memcg_id = memcg_alloc_cache_id();
3237 * We couldn't have accounted to this cgroup, because it hasn't got
3238 * activated yet, so this should succeed.
3240 err = page_counter_limit(&memcg->kmem, nr_pages);
3243 static_key_slow_inc(&memcg_kmem_enabled_key);
3245 * A memory cgroup is considered kmem-active as soon as it gets
3246 * kmemcg_id. Setting the id after enabling static branching will
3247 * guarantee no one starts accounting before all call sites are
3250 memcg->kmemcg_id = memcg_id;
3251 memcg->kmem_acct_activated = true;
3252 memcg->kmem_acct_active = true;
3257 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3258 unsigned long limit)
3262 mutex_lock(&memcg_limit_mutex);
3263 if (!memcg_kmem_is_active(memcg))
3264 ret = memcg_activate_kmem(memcg, limit);
3266 ret = page_counter_limit(&memcg->kmem, limit);
3267 mutex_unlock(&memcg_limit_mutex);
3271 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3274 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3279 mutex_lock(&memcg_limit_mutex);
3281 * If the parent cgroup is not kmem-active now, it cannot be activated
3282 * after this point, because it has at least one child already.
3284 if (memcg_kmem_is_active(parent))
3285 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3286 mutex_unlock(&memcg_limit_mutex);
3290 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3291 unsigned long limit)
3295 #endif /* CONFIG_MEMCG_KMEM */
3298 * The user of this function is...
3301 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3302 char *buf, size_t nbytes, loff_t off)
3304 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3305 unsigned long nr_pages;
3308 buf = strstrip(buf);
3309 ret = page_counter_memparse(buf, "-1", &nr_pages);
3313 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3315 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3319 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3321 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3324 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3327 ret = memcg_update_kmem_limit(memcg, nr_pages);
3331 case RES_SOFT_LIMIT:
3332 memcg->soft_limit = nr_pages;
3336 return ret ?: nbytes;
3339 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3340 size_t nbytes, loff_t off)
3342 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3343 struct page_counter *counter;
3345 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3347 counter = &memcg->memory;
3350 counter = &memcg->memsw;
3353 counter = &memcg->kmem;
3359 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3361 page_counter_reset_watermark(counter);
3364 counter->failcnt = 0;
3373 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3376 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3380 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3381 struct cftype *cft, u64 val)
3383 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3385 if (val & ~MOVE_MASK)
3389 * No kind of locking is needed in here, because ->can_attach() will
3390 * check this value once in the beginning of the process, and then carry
3391 * on with stale data. This means that changes to this value will only
3392 * affect task migrations starting after the change.
3394 memcg->move_charge_at_immigrate = val;
3398 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3399 struct cftype *cft, u64 val)
3406 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3410 unsigned int lru_mask;
3413 static const struct numa_stat stats[] = {
3414 { "total", LRU_ALL },
3415 { "file", LRU_ALL_FILE },
3416 { "anon", LRU_ALL_ANON },
3417 { "unevictable", BIT(LRU_UNEVICTABLE) },
3419 const struct numa_stat *stat;
3422 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3424 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3425 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3426 seq_printf(m, "%s=%lu", stat->name, nr);
3427 for_each_node_state(nid, N_MEMORY) {
3428 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3430 seq_printf(m, " N%d=%lu", nid, nr);
3435 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3436 struct mem_cgroup *iter;
3439 for_each_mem_cgroup_tree(iter, memcg)
3440 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3441 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3442 for_each_node_state(nid, N_MEMORY) {
3444 for_each_mem_cgroup_tree(iter, memcg)
3445 nr += mem_cgroup_node_nr_lru_pages(
3446 iter, nid, stat->lru_mask);
3447 seq_printf(m, " N%d=%lu", nid, nr);
3454 #endif /* CONFIG_NUMA */
3456 static int memcg_stat_show(struct seq_file *m, void *v)
3458 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3459 unsigned long memory, memsw;
3460 struct mem_cgroup *mi;
3463 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3464 MEM_CGROUP_STAT_NSTATS);
3465 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3466 MEM_CGROUP_EVENTS_NSTATS);
3467 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3469 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3470 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3472 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3473 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3476 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3477 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3478 mem_cgroup_read_events(memcg, i));
3480 for (i = 0; i < NR_LRU_LISTS; i++)
3481 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3482 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3484 /* Hierarchical information */
3485 memory = memsw = PAGE_COUNTER_MAX;
3486 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3487 memory = min(memory, mi->memory.limit);
3488 memsw = min(memsw, mi->memsw.limit);
3490 seq_printf(m, "hierarchical_memory_limit %llu\n",
3491 (u64)memory * PAGE_SIZE);
3492 if (do_swap_account)
3493 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3494 (u64)memsw * PAGE_SIZE);
3496 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3499 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3501 for_each_mem_cgroup_tree(mi, memcg)
3502 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3503 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3506 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3507 unsigned long long val = 0;
3509 for_each_mem_cgroup_tree(mi, memcg)
3510 val += mem_cgroup_read_events(mi, i);
3511 seq_printf(m, "total_%s %llu\n",
3512 mem_cgroup_events_names[i], val);
3515 for (i = 0; i < NR_LRU_LISTS; i++) {
3516 unsigned long long val = 0;
3518 for_each_mem_cgroup_tree(mi, memcg)
3519 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3520 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3523 #ifdef CONFIG_DEBUG_VM
3526 struct mem_cgroup_per_zone *mz;
3527 struct zone_reclaim_stat *rstat;
3528 unsigned long recent_rotated[2] = {0, 0};
3529 unsigned long recent_scanned[2] = {0, 0};
3531 for_each_online_node(nid)
3532 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3533 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3534 rstat = &mz->lruvec.reclaim_stat;
3536 recent_rotated[0] += rstat->recent_rotated[0];
3537 recent_rotated[1] += rstat->recent_rotated[1];
3538 recent_scanned[0] += rstat->recent_scanned[0];
3539 recent_scanned[1] += rstat->recent_scanned[1];
3541 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3542 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3543 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3544 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3551 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3554 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3556 return mem_cgroup_swappiness(memcg);
3559 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3560 struct cftype *cft, u64 val)
3562 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3568 memcg->swappiness = val;
3570 vm_swappiness = val;
3575 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3577 struct mem_cgroup_threshold_ary *t;
3578 unsigned long usage;
3583 t = rcu_dereference(memcg->thresholds.primary);
3585 t = rcu_dereference(memcg->memsw_thresholds.primary);
3590 usage = mem_cgroup_usage(memcg, swap);
3593 * current_threshold points to threshold just below or equal to usage.
3594 * If it's not true, a threshold was crossed after last
3595 * call of __mem_cgroup_threshold().
3597 i = t->current_threshold;
3600 * Iterate backward over array of thresholds starting from
3601 * current_threshold and check if a threshold is crossed.
3602 * If none of thresholds below usage is crossed, we read
3603 * only one element of the array here.
3605 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3606 eventfd_signal(t->entries[i].eventfd, 1);
3608 /* i = current_threshold + 1 */
3612 * Iterate forward over array of thresholds starting from
3613 * current_threshold+1 and check if a threshold is crossed.
3614 * If none of thresholds above usage is crossed, we read
3615 * only one element of the array here.
3617 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3618 eventfd_signal(t->entries[i].eventfd, 1);
3620 /* Update current_threshold */
3621 t->current_threshold = i - 1;
3626 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3629 __mem_cgroup_threshold(memcg, false);
3630 if (do_swap_account)
3631 __mem_cgroup_threshold(memcg, true);
3633 memcg = parent_mem_cgroup(memcg);
3637 static int compare_thresholds(const void *a, const void *b)
3639 const struct mem_cgroup_threshold *_a = a;
3640 const struct mem_cgroup_threshold *_b = b;
3642 if (_a->threshold > _b->threshold)
3645 if (_a->threshold < _b->threshold)
3651 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3653 struct mem_cgroup_eventfd_list *ev;
3655 spin_lock(&memcg_oom_lock);
3657 list_for_each_entry(ev, &memcg->oom_notify, list)
3658 eventfd_signal(ev->eventfd, 1);
3660 spin_unlock(&memcg_oom_lock);
3664 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3666 struct mem_cgroup *iter;
3668 for_each_mem_cgroup_tree(iter, memcg)
3669 mem_cgroup_oom_notify_cb(iter);
3672 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3673 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3675 struct mem_cgroup_thresholds *thresholds;
3676 struct mem_cgroup_threshold_ary *new;
3677 unsigned long threshold;
3678 unsigned long usage;
3681 ret = page_counter_memparse(args, "-1", &threshold);
3685 mutex_lock(&memcg->thresholds_lock);
3688 thresholds = &memcg->thresholds;
3689 usage = mem_cgroup_usage(memcg, false);
3690 } else if (type == _MEMSWAP) {
3691 thresholds = &memcg->memsw_thresholds;
3692 usage = mem_cgroup_usage(memcg, true);
3696 /* Check if a threshold crossed before adding a new one */
3697 if (thresholds->primary)
3698 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3700 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3702 /* Allocate memory for new array of thresholds */
3703 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3711 /* Copy thresholds (if any) to new array */
3712 if (thresholds->primary) {
3713 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3714 sizeof(struct mem_cgroup_threshold));
3717 /* Add new threshold */
3718 new->entries[size - 1].eventfd = eventfd;
3719 new->entries[size - 1].threshold = threshold;
3721 /* Sort thresholds. Registering of new threshold isn't time-critical */
3722 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3723 compare_thresholds, NULL);
3725 /* Find current threshold */
3726 new->current_threshold = -1;
3727 for (i = 0; i < size; i++) {
3728 if (new->entries[i].threshold <= usage) {
3730 * new->current_threshold will not be used until
3731 * rcu_assign_pointer(), so it's safe to increment
3734 ++new->current_threshold;
3739 /* Free old spare buffer and save old primary buffer as spare */
3740 kfree(thresholds->spare);
3741 thresholds->spare = thresholds->primary;
3743 rcu_assign_pointer(thresholds->primary, new);
3745 /* To be sure that nobody uses thresholds */
3749 mutex_unlock(&memcg->thresholds_lock);
3754 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3755 struct eventfd_ctx *eventfd, const char *args)
3757 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3760 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3761 struct eventfd_ctx *eventfd, const char *args)
3763 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3766 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3767 struct eventfd_ctx *eventfd, enum res_type type)
3769 struct mem_cgroup_thresholds *thresholds;
3770 struct mem_cgroup_threshold_ary *new;
3771 unsigned long usage;
3774 mutex_lock(&memcg->thresholds_lock);
3777 thresholds = &memcg->thresholds;
3778 usage = mem_cgroup_usage(memcg, false);
3779 } else if (type == _MEMSWAP) {
3780 thresholds = &memcg->memsw_thresholds;
3781 usage = mem_cgroup_usage(memcg, true);
3785 if (!thresholds->primary)
3788 /* Check if a threshold crossed before removing */
3789 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3791 /* Calculate new number of threshold */
3793 for (i = 0; i < thresholds->primary->size; i++) {
3794 if (thresholds->primary->entries[i].eventfd != eventfd)
3798 new = thresholds->spare;
3800 /* Set thresholds array to NULL if we don't have thresholds */
3809 /* Copy thresholds and find current threshold */
3810 new->current_threshold = -1;
3811 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3812 if (thresholds->primary->entries[i].eventfd == eventfd)
3815 new->entries[j] = thresholds->primary->entries[i];
3816 if (new->entries[j].threshold <= usage) {
3818 * new->current_threshold will not be used
3819 * until rcu_assign_pointer(), so it's safe to increment
3822 ++new->current_threshold;
3828 /* Swap primary and spare array */
3829 thresholds->spare = thresholds->primary;
3830 /* If all events are unregistered, free the spare array */
3832 kfree(thresholds->spare);
3833 thresholds->spare = NULL;
3836 rcu_assign_pointer(thresholds->primary, new);
3838 /* To be sure that nobody uses thresholds */
3841 mutex_unlock(&memcg->thresholds_lock);
3844 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3845 struct eventfd_ctx *eventfd)
3847 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3850 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3851 struct eventfd_ctx *eventfd)
3853 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3856 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3857 struct eventfd_ctx *eventfd, const char *args)
3859 struct mem_cgroup_eventfd_list *event;
3861 event = kmalloc(sizeof(*event), GFP_KERNEL);
3865 spin_lock(&memcg_oom_lock);
3867 event->eventfd = eventfd;
3868 list_add(&event->list, &memcg->oom_notify);
3870 /* already in OOM ? */
3871 if (atomic_read(&memcg->under_oom))
3872 eventfd_signal(eventfd, 1);
3873 spin_unlock(&memcg_oom_lock);
3878 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3879 struct eventfd_ctx *eventfd)
3881 struct mem_cgroup_eventfd_list *ev, *tmp;
3883 spin_lock(&memcg_oom_lock);
3885 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3886 if (ev->eventfd == eventfd) {
3887 list_del(&ev->list);
3892 spin_unlock(&memcg_oom_lock);
3895 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3897 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3899 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3900 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
3904 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3905 struct cftype *cft, u64 val)
3907 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3909 /* cannot set to root cgroup and only 0 and 1 are allowed */
3910 if (!css->parent || !((val == 0) || (val == 1)))
3913 memcg->oom_kill_disable = val;
3915 memcg_oom_recover(memcg);
3920 #ifdef CONFIG_MEMCG_KMEM
3921 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3925 ret = memcg_propagate_kmem(memcg);
3929 return mem_cgroup_sockets_init(memcg, ss);
3932 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3934 struct cgroup_subsys_state *css;
3935 struct mem_cgroup *parent, *child;
3938 if (!memcg->kmem_acct_active)
3942 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3943 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3944 * guarantees no cache will be created for this cgroup after we are
3945 * done (see memcg_create_kmem_cache()).
3947 memcg->kmem_acct_active = false;
3949 memcg_deactivate_kmem_caches(memcg);
3951 kmemcg_id = memcg->kmemcg_id;
3952 BUG_ON(kmemcg_id < 0);
3954 parent = parent_mem_cgroup(memcg);
3956 parent = root_mem_cgroup;
3959 * Change kmemcg_id of this cgroup and all its descendants to the
3960 * parent's id, and then move all entries from this cgroup's list_lrus
3961 * to ones of the parent. After we have finished, all list_lrus
3962 * corresponding to this cgroup are guaranteed to remain empty. The
3963 * ordering is imposed by list_lru_node->lock taken by
3964 * memcg_drain_all_list_lrus().
3966 css_for_each_descendant_pre(css, &memcg->css) {
3967 child = mem_cgroup_from_css(css);
3968 BUG_ON(child->kmemcg_id != kmemcg_id);
3969 child->kmemcg_id = parent->kmemcg_id;
3970 if (!memcg->use_hierarchy)
3973 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3975 memcg_free_cache_id(kmemcg_id);
3978 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3980 if (memcg->kmem_acct_activated) {
3981 memcg_destroy_kmem_caches(memcg);
3982 static_key_slow_dec(&memcg_kmem_enabled_key);
3983 WARN_ON(page_counter_read(&memcg->kmem));
3985 mem_cgroup_sockets_destroy(memcg);
3988 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3993 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3997 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4003 * DO NOT USE IN NEW FILES.
4005 * "cgroup.event_control" implementation.
4007 * This is way over-engineered. It tries to support fully configurable
4008 * events for each user. Such level of flexibility is completely
4009 * unnecessary especially in the light of the planned unified hierarchy.
4011 * Please deprecate this and replace with something simpler if at all
4016 * Unregister event and free resources.
4018 * Gets called from workqueue.
4020 static void memcg_event_remove(struct work_struct *work)
4022 struct mem_cgroup_event *event =
4023 container_of(work, struct mem_cgroup_event, remove);
4024 struct mem_cgroup *memcg = event->memcg;
4026 remove_wait_queue(event->wqh, &event->wait);
4028 event->unregister_event(memcg, event->eventfd);
4030 /* Notify userspace the event is going away. */
4031 eventfd_signal(event->eventfd, 1);
4033 eventfd_ctx_put(event->eventfd);
4035 css_put(&memcg->css);
4039 * Gets called on POLLHUP on eventfd when user closes it.
4041 * Called with wqh->lock held and interrupts disabled.
4043 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4044 int sync, void *key)
4046 struct mem_cgroup_event *event =
4047 container_of(wait, struct mem_cgroup_event, wait);
4048 struct mem_cgroup *memcg = event->memcg;
4049 unsigned long flags = (unsigned long)key;
4051 if (flags & POLLHUP) {
4053 * If the event has been detached at cgroup removal, we
4054 * can simply return knowing the other side will cleanup
4057 * We can't race against event freeing since the other
4058 * side will require wqh->lock via remove_wait_queue(),
4061 spin_lock(&memcg->event_list_lock);
4062 if (!list_empty(&event->list)) {
4063 list_del_init(&event->list);
4065 * We are in atomic context, but cgroup_event_remove()
4066 * may sleep, so we have to call it in workqueue.
4068 schedule_work(&event->remove);
4070 spin_unlock(&memcg->event_list_lock);
4076 static void memcg_event_ptable_queue_proc(struct file *file,
4077 wait_queue_head_t *wqh, poll_table *pt)
4079 struct mem_cgroup_event *event =
4080 container_of(pt, struct mem_cgroup_event, pt);
4083 add_wait_queue(wqh, &event->wait);
4087 * DO NOT USE IN NEW FILES.
4089 * Parse input and register new cgroup event handler.
4091 * Input must be in format '<event_fd> <control_fd> <args>'.
4092 * Interpretation of args is defined by control file implementation.
4094 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4095 char *buf, size_t nbytes, loff_t off)
4097 struct cgroup_subsys_state *css = of_css(of);
4098 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4099 struct mem_cgroup_event *event;
4100 struct cgroup_subsys_state *cfile_css;
4101 unsigned int efd, cfd;
4108 buf = strstrip(buf);
4110 efd = simple_strtoul(buf, &endp, 10);
4115 cfd = simple_strtoul(buf, &endp, 10);
4116 if ((*endp != ' ') && (*endp != '\0'))
4120 event = kzalloc(sizeof(*event), GFP_KERNEL);
4124 event->memcg = memcg;
4125 INIT_LIST_HEAD(&event->list);
4126 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4127 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4128 INIT_WORK(&event->remove, memcg_event_remove);
4136 event->eventfd = eventfd_ctx_fileget(efile.file);
4137 if (IS_ERR(event->eventfd)) {
4138 ret = PTR_ERR(event->eventfd);
4145 goto out_put_eventfd;
4148 /* the process need read permission on control file */
4149 /* AV: shouldn't we check that it's been opened for read instead? */
4150 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4155 * Determine the event callbacks and set them in @event. This used
4156 * to be done via struct cftype but cgroup core no longer knows
4157 * about these events. The following is crude but the whole thing
4158 * is for compatibility anyway.
4160 * DO NOT ADD NEW FILES.
4162 name = cfile.file->f_path.dentry->d_name.name;
4164 if (!strcmp(name, "memory.usage_in_bytes")) {
4165 event->register_event = mem_cgroup_usage_register_event;
4166 event->unregister_event = mem_cgroup_usage_unregister_event;
4167 } else if (!strcmp(name, "memory.oom_control")) {
4168 event->register_event = mem_cgroup_oom_register_event;
4169 event->unregister_event = mem_cgroup_oom_unregister_event;
4170 } else if (!strcmp(name, "memory.pressure_level")) {
4171 event->register_event = vmpressure_register_event;
4172 event->unregister_event = vmpressure_unregister_event;
4173 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4174 event->register_event = memsw_cgroup_usage_register_event;
4175 event->unregister_event = memsw_cgroup_usage_unregister_event;
4182 * Verify @cfile should belong to @css. Also, remaining events are
4183 * automatically removed on cgroup destruction but the removal is
4184 * asynchronous, so take an extra ref on @css.
4186 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4187 &memory_cgrp_subsys);
4189 if (IS_ERR(cfile_css))
4191 if (cfile_css != css) {
4196 ret = event->register_event(memcg, event->eventfd, buf);
4200 efile.file->f_op->poll(efile.file, &event->pt);
4202 spin_lock(&memcg->event_list_lock);
4203 list_add(&event->list, &memcg->event_list);
4204 spin_unlock(&memcg->event_list_lock);
4216 eventfd_ctx_put(event->eventfd);
4225 static struct cftype mem_cgroup_legacy_files[] = {
4227 .name = "usage_in_bytes",
4228 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4229 .read_u64 = mem_cgroup_read_u64,
4232 .name = "max_usage_in_bytes",
4233 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4234 .write = mem_cgroup_reset,
4235 .read_u64 = mem_cgroup_read_u64,
4238 .name = "limit_in_bytes",
4239 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4240 .write = mem_cgroup_write,
4241 .read_u64 = mem_cgroup_read_u64,
4244 .name = "soft_limit_in_bytes",
4245 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4246 .write = mem_cgroup_write,
4247 .read_u64 = mem_cgroup_read_u64,
4251 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4252 .write = mem_cgroup_reset,
4253 .read_u64 = mem_cgroup_read_u64,
4257 .seq_show = memcg_stat_show,
4260 .name = "force_empty",
4261 .write = mem_cgroup_force_empty_write,
4264 .name = "use_hierarchy",
4265 .write_u64 = mem_cgroup_hierarchy_write,
4266 .read_u64 = mem_cgroup_hierarchy_read,
4269 .name = "cgroup.event_control", /* XXX: for compat */
4270 .write = memcg_write_event_control,
4271 .flags = CFTYPE_NO_PREFIX,
4275 .name = "swappiness",
4276 .read_u64 = mem_cgroup_swappiness_read,
4277 .write_u64 = mem_cgroup_swappiness_write,
4280 .name = "move_charge_at_immigrate",
4281 .read_u64 = mem_cgroup_move_charge_read,
4282 .write_u64 = mem_cgroup_move_charge_write,
4285 .name = "oom_control",
4286 .seq_show = mem_cgroup_oom_control_read,
4287 .write_u64 = mem_cgroup_oom_control_write,
4288 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4291 .name = "pressure_level",
4295 .name = "numa_stat",
4296 .seq_show = memcg_numa_stat_show,
4299 #ifdef CONFIG_MEMCG_KMEM
4301 .name = "kmem.limit_in_bytes",
4302 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4303 .write = mem_cgroup_write,
4304 .read_u64 = mem_cgroup_read_u64,
4307 .name = "kmem.usage_in_bytes",
4308 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4309 .read_u64 = mem_cgroup_read_u64,
4312 .name = "kmem.failcnt",
4313 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4314 .write = mem_cgroup_reset,
4315 .read_u64 = mem_cgroup_read_u64,
4318 .name = "kmem.max_usage_in_bytes",
4319 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4320 .write = mem_cgroup_reset,
4321 .read_u64 = mem_cgroup_read_u64,
4323 #ifdef CONFIG_SLABINFO
4325 .name = "kmem.slabinfo",
4326 .seq_start = slab_start,
4327 .seq_next = slab_next,
4328 .seq_stop = slab_stop,
4329 .seq_show = memcg_slab_show,
4333 { }, /* terminate */
4336 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4338 struct mem_cgroup_per_node *pn;
4339 struct mem_cgroup_per_zone *mz;
4340 int zone, tmp = node;
4342 * This routine is called against possible nodes.
4343 * But it's BUG to call kmalloc() against offline node.
4345 * TODO: this routine can waste much memory for nodes which will
4346 * never be onlined. It's better to use memory hotplug callback
4349 if (!node_state(node, N_NORMAL_MEMORY))
4351 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4355 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4356 mz = &pn->zoneinfo[zone];
4357 lruvec_init(&mz->lruvec);
4358 mz->usage_in_excess = 0;
4359 mz->on_tree = false;
4362 memcg->nodeinfo[node] = pn;
4366 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4368 kfree(memcg->nodeinfo[node]);
4371 static struct mem_cgroup *mem_cgroup_alloc(void)
4373 struct mem_cgroup *memcg;
4376 size = sizeof(struct mem_cgroup);
4377 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4379 memcg = kzalloc(size, GFP_KERNEL);
4383 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4386 spin_lock_init(&memcg->pcp_counter_lock);
4395 * At destroying mem_cgroup, references from swap_cgroup can remain.
4396 * (scanning all at force_empty is too costly...)
4398 * Instead of clearing all references at force_empty, we remember
4399 * the number of reference from swap_cgroup and free mem_cgroup when
4400 * it goes down to 0.
4402 * Removal of cgroup itself succeeds regardless of refs from swap.
4405 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4409 mem_cgroup_remove_from_trees(memcg);
4412 free_mem_cgroup_per_zone_info(memcg, node);
4414 free_percpu(memcg->stat);
4419 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4421 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4423 if (!memcg->memory.parent)
4425 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4427 EXPORT_SYMBOL(parent_mem_cgroup);
4429 static struct cgroup_subsys_state * __ref
4430 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4432 struct mem_cgroup *memcg;
4433 long error = -ENOMEM;
4436 memcg = mem_cgroup_alloc();
4438 return ERR_PTR(error);
4441 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4445 if (parent_css == NULL) {
4446 root_mem_cgroup = memcg;
4447 page_counter_init(&memcg->memory, NULL);
4448 memcg->high = PAGE_COUNTER_MAX;
4449 memcg->soft_limit = PAGE_COUNTER_MAX;
4450 page_counter_init(&memcg->memsw, NULL);
4451 page_counter_init(&memcg->kmem, NULL);
4454 memcg->last_scanned_node = MAX_NUMNODES;
4455 INIT_LIST_HEAD(&memcg->oom_notify);
4456 memcg->move_charge_at_immigrate = 0;
4457 mutex_init(&memcg->thresholds_lock);
4458 spin_lock_init(&memcg->move_lock);
4459 vmpressure_init(&memcg->vmpressure);
4460 INIT_LIST_HEAD(&memcg->event_list);
4461 spin_lock_init(&memcg->event_list_lock);
4462 #ifdef CONFIG_MEMCG_KMEM
4463 memcg->kmemcg_id = -1;
4469 __mem_cgroup_free(memcg);
4470 return ERR_PTR(error);
4474 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4476 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4477 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4480 if (css->id > MEM_CGROUP_ID_MAX)
4486 mutex_lock(&memcg_create_mutex);
4488 memcg->use_hierarchy = parent->use_hierarchy;
4489 memcg->oom_kill_disable = parent->oom_kill_disable;
4490 memcg->swappiness = mem_cgroup_swappiness(parent);
4492 if (parent->use_hierarchy) {
4493 page_counter_init(&memcg->memory, &parent->memory);
4494 memcg->high = PAGE_COUNTER_MAX;
4495 memcg->soft_limit = PAGE_COUNTER_MAX;
4496 page_counter_init(&memcg->memsw, &parent->memsw);
4497 page_counter_init(&memcg->kmem, &parent->kmem);
4500 * No need to take a reference to the parent because cgroup
4501 * core guarantees its existence.
4504 page_counter_init(&memcg->memory, NULL);
4505 memcg->high = PAGE_COUNTER_MAX;
4506 memcg->soft_limit = PAGE_COUNTER_MAX;
4507 page_counter_init(&memcg->memsw, NULL);
4508 page_counter_init(&memcg->kmem, NULL);
4510 * Deeper hierachy with use_hierarchy == false doesn't make
4511 * much sense so let cgroup subsystem know about this
4512 * unfortunate state in our controller.
4514 if (parent != root_mem_cgroup)
4515 memory_cgrp_subsys.broken_hierarchy = true;
4517 mutex_unlock(&memcg_create_mutex);
4519 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4524 * Make sure the memcg is initialized: mem_cgroup_iter()
4525 * orders reading memcg->initialized against its callers
4526 * reading the memcg members.
4528 smp_store_release(&memcg->initialized, 1);
4533 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4535 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4536 struct mem_cgroup_event *event, *tmp;
4539 * Unregister events and notify userspace.
4540 * Notify userspace about cgroup removing only after rmdir of cgroup
4541 * directory to avoid race between userspace and kernelspace.
4543 spin_lock(&memcg->event_list_lock);
4544 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4545 list_del_init(&event->list);
4546 schedule_work(&event->remove);
4548 spin_unlock(&memcg->event_list_lock);
4550 vmpressure_cleanup(&memcg->vmpressure);
4552 memcg_deactivate_kmem(memcg);
4555 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4557 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4559 memcg_destroy_kmem(memcg);
4560 __mem_cgroup_free(memcg);
4564 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4565 * @css: the target css
4567 * Reset the states of the mem_cgroup associated with @css. This is
4568 * invoked when the userland requests disabling on the default hierarchy
4569 * but the memcg is pinned through dependency. The memcg should stop
4570 * applying policies and should revert to the vanilla state as it may be
4571 * made visible again.
4573 * The current implementation only resets the essential configurations.
4574 * This needs to be expanded to cover all the visible parts.
4576 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4578 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4580 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4581 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4582 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4584 memcg->high = PAGE_COUNTER_MAX;
4585 memcg->soft_limit = PAGE_COUNTER_MAX;
4589 /* Handlers for move charge at task migration. */
4590 static int mem_cgroup_do_precharge(unsigned long count)
4594 /* Try a single bulk charge without reclaim first */
4595 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4597 mc.precharge += count;
4600 if (ret == -EINTR) {
4601 cancel_charge(root_mem_cgroup, count);
4605 /* Try charges one by one with reclaim */
4607 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4609 * In case of failure, any residual charges against
4610 * mc.to will be dropped by mem_cgroup_clear_mc()
4611 * later on. However, cancel any charges that are
4612 * bypassed to root right away or they'll be lost.
4615 cancel_charge(root_mem_cgroup, 1);
4625 * get_mctgt_type - get target type of moving charge
4626 * @vma: the vma the pte to be checked belongs
4627 * @addr: the address corresponding to the pte to be checked
4628 * @ptent: the pte to be checked
4629 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4632 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4633 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4634 * move charge. if @target is not NULL, the page is stored in target->page
4635 * with extra refcnt got(Callers should handle it).
4636 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4637 * target for charge migration. if @target is not NULL, the entry is stored
4640 * Called with pte lock held.
4647 enum mc_target_type {
4653 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4654 unsigned long addr, pte_t ptent)
4656 struct page *page = vm_normal_page(vma, addr, ptent);
4658 if (!page || !page_mapped(page))
4660 if (PageAnon(page)) {
4661 if (!(mc.flags & MOVE_ANON))
4664 if (!(mc.flags & MOVE_FILE))
4667 if (!get_page_unless_zero(page))
4674 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4675 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4677 struct page *page = NULL;
4678 swp_entry_t ent = pte_to_swp_entry(ptent);
4680 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4683 * Because lookup_swap_cache() updates some statistics counter,
4684 * we call find_get_page() with swapper_space directly.
4686 page = find_get_page(swap_address_space(ent), ent.val);
4687 if (do_swap_account)
4688 entry->val = ent.val;
4693 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4694 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4700 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4701 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4703 struct page *page = NULL;
4704 struct address_space *mapping;
4707 if (!vma->vm_file) /* anonymous vma */
4709 if (!(mc.flags & MOVE_FILE))
4712 mapping = vma->vm_file->f_mapping;
4713 pgoff = linear_page_index(vma, addr);
4715 /* page is moved even if it's not RSS of this task(page-faulted). */
4717 /* shmem/tmpfs may report page out on swap: account for that too. */
4718 if (shmem_mapping(mapping)) {
4719 page = find_get_entry(mapping, pgoff);
4720 if (radix_tree_exceptional_entry(page)) {
4721 swp_entry_t swp = radix_to_swp_entry(page);
4722 if (do_swap_account)
4724 page = find_get_page(swap_address_space(swp), swp.val);
4727 page = find_get_page(mapping, pgoff);
4729 page = find_get_page(mapping, pgoff);
4735 * mem_cgroup_move_account - move account of the page
4737 * @nr_pages: number of regular pages (>1 for huge pages)
4738 * @from: mem_cgroup which the page is moved from.
4739 * @to: mem_cgroup which the page is moved to. @from != @to.
4741 * The caller must confirm following.
4742 * - page is not on LRU (isolate_page() is useful.)
4743 * - compound_lock is held when nr_pages > 1
4745 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4748 static int mem_cgroup_move_account(struct page *page,
4749 unsigned int nr_pages,
4750 struct mem_cgroup *from,
4751 struct mem_cgroup *to)
4753 unsigned long flags;
4756 VM_BUG_ON(from == to);
4757 VM_BUG_ON_PAGE(PageLRU(page), page);
4759 * The page is isolated from LRU. So, collapse function
4760 * will not handle this page. But page splitting can happen.
4761 * Do this check under compound_page_lock(). The caller should
4765 if (nr_pages > 1 && !PageTransHuge(page))
4769 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
4770 * of its source page while we change it: page migration takes
4771 * both pages off the LRU, but page cache replacement doesn't.
4773 if (!trylock_page(page))
4777 if (page->mem_cgroup != from)
4780 spin_lock_irqsave(&from->move_lock, flags);
4782 if (!PageAnon(page) && page_mapped(page)) {
4783 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4785 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4789 if (PageWriteback(page)) {
4790 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4792 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4797 * It is safe to change page->mem_cgroup here because the page
4798 * is referenced, charged, and isolated - we can't race with
4799 * uncharging, charging, migration, or LRU putback.
4802 /* caller should have done css_get */
4803 page->mem_cgroup = to;
4804 spin_unlock_irqrestore(&from->move_lock, flags);
4808 local_irq_disable();
4809 mem_cgroup_charge_statistics(to, page, nr_pages);
4810 memcg_check_events(to, page);
4811 mem_cgroup_charge_statistics(from, page, -nr_pages);
4812 memcg_check_events(from, page);
4820 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4821 unsigned long addr, pte_t ptent, union mc_target *target)
4823 struct page *page = NULL;
4824 enum mc_target_type ret = MC_TARGET_NONE;
4825 swp_entry_t ent = { .val = 0 };
4827 if (pte_present(ptent))
4828 page = mc_handle_present_pte(vma, addr, ptent);
4829 else if (is_swap_pte(ptent))
4830 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4831 else if (pte_none(ptent))
4832 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4834 if (!page && !ent.val)
4838 * Do only loose check w/o serialization.
4839 * mem_cgroup_move_account() checks the page is valid or
4840 * not under LRU exclusion.
4842 if (page->mem_cgroup == mc.from) {
4843 ret = MC_TARGET_PAGE;
4845 target->page = page;
4847 if (!ret || !target)
4850 /* There is a swap entry and a page doesn't exist or isn't charged */
4851 if (ent.val && !ret &&
4852 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4853 ret = MC_TARGET_SWAP;
4860 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4862 * We don't consider swapping or file mapped pages because THP does not
4863 * support them for now.
4864 * Caller should make sure that pmd_trans_huge(pmd) is true.
4866 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4867 unsigned long addr, pmd_t pmd, union mc_target *target)
4869 struct page *page = NULL;
4870 enum mc_target_type ret = MC_TARGET_NONE;
4872 page = pmd_page(pmd);
4873 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4874 if (!(mc.flags & MOVE_ANON))
4876 if (page->mem_cgroup == mc.from) {
4877 ret = MC_TARGET_PAGE;
4880 target->page = page;
4886 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4887 unsigned long addr, pmd_t pmd, union mc_target *target)
4889 return MC_TARGET_NONE;
4893 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4894 unsigned long addr, unsigned long end,
4895 struct mm_walk *walk)
4897 struct vm_area_struct *vma = walk->vma;
4901 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4902 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4903 mc.precharge += HPAGE_PMD_NR;
4908 if (pmd_trans_unstable(pmd))
4910 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4911 for (; addr != end; pte++, addr += PAGE_SIZE)
4912 if (get_mctgt_type(vma, addr, *pte, NULL))
4913 mc.precharge++; /* increment precharge temporarily */
4914 pte_unmap_unlock(pte - 1, ptl);
4920 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4922 unsigned long precharge;
4924 struct mm_walk mem_cgroup_count_precharge_walk = {
4925 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4928 down_read(&mm->mmap_sem);
4929 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
4930 up_read(&mm->mmap_sem);
4932 precharge = mc.precharge;
4938 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4940 unsigned long precharge = mem_cgroup_count_precharge(mm);
4942 VM_BUG_ON(mc.moving_task);
4943 mc.moving_task = current;
4944 return mem_cgroup_do_precharge(precharge);
4947 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4948 static void __mem_cgroup_clear_mc(void)
4950 struct mem_cgroup *from = mc.from;
4951 struct mem_cgroup *to = mc.to;
4953 /* we must uncharge all the leftover precharges from mc.to */
4955 cancel_charge(mc.to, mc.precharge);
4959 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4960 * we must uncharge here.
4962 if (mc.moved_charge) {
4963 cancel_charge(mc.from, mc.moved_charge);
4964 mc.moved_charge = 0;
4966 /* we must fixup refcnts and charges */
4967 if (mc.moved_swap) {
4968 /* uncharge swap account from the old cgroup */
4969 if (!mem_cgroup_is_root(mc.from))
4970 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4973 * we charged both to->memory and to->memsw, so we
4974 * should uncharge to->memory.
4976 if (!mem_cgroup_is_root(mc.to))
4977 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4979 css_put_many(&mc.from->css, mc.moved_swap);
4981 /* we've already done css_get(mc.to) */
4984 memcg_oom_recover(from);
4985 memcg_oom_recover(to);
4986 wake_up_all(&mc.waitq);
4989 static void mem_cgroup_clear_mc(void)
4992 * we must clear moving_task before waking up waiters at the end of
4995 mc.moving_task = NULL;
4996 __mem_cgroup_clear_mc();
4997 spin_lock(&mc.lock);
5000 spin_unlock(&mc.lock);
5003 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5004 struct cgroup_taskset *tset)
5006 struct task_struct *p = cgroup_taskset_first(tset);
5008 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5009 unsigned long move_flags;
5012 * We are now commited to this value whatever it is. Changes in this
5013 * tunable will only affect upcoming migrations, not the current one.
5014 * So we need to save it, and keep it going.
5016 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5018 struct mm_struct *mm;
5019 struct mem_cgroup *from = mem_cgroup_from_task(p);
5021 VM_BUG_ON(from == memcg);
5023 mm = get_task_mm(p);
5026 /* We move charges only when we move a owner of the mm */
5027 if (mm->owner == p) {
5030 VM_BUG_ON(mc.precharge);
5031 VM_BUG_ON(mc.moved_charge);
5032 VM_BUG_ON(mc.moved_swap);
5034 spin_lock(&mc.lock);
5037 mc.flags = move_flags;
5038 spin_unlock(&mc.lock);
5039 /* We set mc.moving_task later */
5041 ret = mem_cgroup_precharge_mc(mm);
5043 mem_cgroup_clear_mc();
5050 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5051 struct cgroup_taskset *tset)
5054 mem_cgroup_clear_mc();
5057 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5058 unsigned long addr, unsigned long end,
5059 struct mm_walk *walk)
5062 struct vm_area_struct *vma = walk->vma;
5065 enum mc_target_type target_type;
5066 union mc_target target;
5070 * We don't take compound_lock() here but no race with splitting thp
5072 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5073 * under splitting, which means there's no concurrent thp split,
5074 * - if another thread runs into split_huge_page() just after we
5075 * entered this if-block, the thread must wait for page table lock
5076 * to be unlocked in __split_huge_page_splitting(), where the main
5077 * part of thp split is not executed yet.
5079 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5080 if (mc.precharge < HPAGE_PMD_NR) {
5084 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5085 if (target_type == MC_TARGET_PAGE) {
5087 if (!isolate_lru_page(page)) {
5088 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5090 mc.precharge -= HPAGE_PMD_NR;
5091 mc.moved_charge += HPAGE_PMD_NR;
5093 putback_lru_page(page);
5101 if (pmd_trans_unstable(pmd))
5104 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5105 for (; addr != end; addr += PAGE_SIZE) {
5106 pte_t ptent = *(pte++);
5112 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5113 case MC_TARGET_PAGE:
5115 if (isolate_lru_page(page))
5117 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5119 /* we uncharge from mc.from later. */
5122 putback_lru_page(page);
5123 put: /* get_mctgt_type() gets the page */
5126 case MC_TARGET_SWAP:
5128 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5130 /* we fixup refcnts and charges later. */
5138 pte_unmap_unlock(pte - 1, ptl);
5143 * We have consumed all precharges we got in can_attach().
5144 * We try charge one by one, but don't do any additional
5145 * charges to mc.to if we have failed in charge once in attach()
5148 ret = mem_cgroup_do_precharge(1);
5156 static void mem_cgroup_move_charge(struct mm_struct *mm)
5158 struct mm_walk mem_cgroup_move_charge_walk = {
5159 .pmd_entry = mem_cgroup_move_charge_pte_range,
5163 lru_add_drain_all();
5165 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5166 * move_lock while we're moving its pages to another memcg.
5167 * Then wait for already started RCU-only updates to finish.
5169 atomic_inc(&mc.from->moving_account);
5172 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5174 * Someone who are holding the mmap_sem might be waiting in
5175 * waitq. So we cancel all extra charges, wake up all waiters,
5176 * and retry. Because we cancel precharges, we might not be able
5177 * to move enough charges, but moving charge is a best-effort
5178 * feature anyway, so it wouldn't be a big problem.
5180 __mem_cgroup_clear_mc();
5185 * When we have consumed all precharges and failed in doing
5186 * additional charge, the page walk just aborts.
5188 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5189 up_read(&mm->mmap_sem);
5190 atomic_dec(&mc.from->moving_account);
5193 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5194 struct cgroup_taskset *tset)
5196 struct task_struct *p = cgroup_taskset_first(tset);
5197 struct mm_struct *mm = get_task_mm(p);
5201 mem_cgroup_move_charge(mm);
5205 mem_cgroup_clear_mc();
5207 #else /* !CONFIG_MMU */
5208 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5209 struct cgroup_taskset *tset)
5213 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5214 struct cgroup_taskset *tset)
5217 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5218 struct cgroup_taskset *tset)
5224 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5225 * to verify whether we're attached to the default hierarchy on each mount
5228 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5231 * use_hierarchy is forced on the default hierarchy. cgroup core
5232 * guarantees that @root doesn't have any children, so turning it
5233 * on for the root memcg is enough.
5235 if (cgroup_on_dfl(root_css->cgroup))
5236 root_mem_cgroup->use_hierarchy = true;
5238 root_mem_cgroup->use_hierarchy = false;
5241 static u64 memory_current_read(struct cgroup_subsys_state *css,
5244 return mem_cgroup_usage(mem_cgroup_from_css(css), false);
5247 static int memory_low_show(struct seq_file *m, void *v)
5249 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5250 unsigned long low = READ_ONCE(memcg->low);
5252 if (low == PAGE_COUNTER_MAX)
5253 seq_puts(m, "max\n");
5255 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5260 static ssize_t memory_low_write(struct kernfs_open_file *of,
5261 char *buf, size_t nbytes, loff_t off)
5263 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5267 buf = strstrip(buf);
5268 err = page_counter_memparse(buf, "max", &low);
5277 static int memory_high_show(struct seq_file *m, void *v)
5279 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5280 unsigned long high = READ_ONCE(memcg->high);
5282 if (high == PAGE_COUNTER_MAX)
5283 seq_puts(m, "max\n");
5285 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5290 static ssize_t memory_high_write(struct kernfs_open_file *of,
5291 char *buf, size_t nbytes, loff_t off)
5293 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5297 buf = strstrip(buf);
5298 err = page_counter_memparse(buf, "max", &high);
5307 static int memory_max_show(struct seq_file *m, void *v)
5309 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5310 unsigned long max = READ_ONCE(memcg->memory.limit);
5312 if (max == PAGE_COUNTER_MAX)
5313 seq_puts(m, "max\n");
5315 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5320 static ssize_t memory_max_write(struct kernfs_open_file *of,
5321 char *buf, size_t nbytes, loff_t off)
5323 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5327 buf = strstrip(buf);
5328 err = page_counter_memparse(buf, "max", &max);
5332 err = mem_cgroup_resize_limit(memcg, max);
5339 static int memory_events_show(struct seq_file *m, void *v)
5341 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5343 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5344 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5345 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5346 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5351 static struct cftype memory_files[] = {
5354 .read_u64 = memory_current_read,
5358 .flags = CFTYPE_NOT_ON_ROOT,
5359 .seq_show = memory_low_show,
5360 .write = memory_low_write,
5364 .flags = CFTYPE_NOT_ON_ROOT,
5365 .seq_show = memory_high_show,
5366 .write = memory_high_write,
5370 .flags = CFTYPE_NOT_ON_ROOT,
5371 .seq_show = memory_max_show,
5372 .write = memory_max_write,
5376 .flags = CFTYPE_NOT_ON_ROOT,
5377 .seq_show = memory_events_show,
5382 struct cgroup_subsys memory_cgrp_subsys = {
5383 .css_alloc = mem_cgroup_css_alloc,
5384 .css_online = mem_cgroup_css_online,
5385 .css_offline = mem_cgroup_css_offline,
5386 .css_free = mem_cgroup_css_free,
5387 .css_reset = mem_cgroup_css_reset,
5388 .can_attach = mem_cgroup_can_attach,
5389 .cancel_attach = mem_cgroup_cancel_attach,
5390 .attach = mem_cgroup_move_task,
5391 .bind = mem_cgroup_bind,
5392 .dfl_cftypes = memory_files,
5393 .legacy_cftypes = mem_cgroup_legacy_files,
5398 * mem_cgroup_events - count memory events against a cgroup
5399 * @memcg: the memory cgroup
5400 * @idx: the event index
5401 * @nr: the number of events to account for
5403 void mem_cgroup_events(struct mem_cgroup *memcg,
5404 enum mem_cgroup_events_index idx,
5407 this_cpu_add(memcg->stat->events[idx], nr);
5411 * mem_cgroup_low - check if memory consumption is below the normal range
5412 * @root: the highest ancestor to consider
5413 * @memcg: the memory cgroup to check
5415 * Returns %true if memory consumption of @memcg, and that of all
5416 * configurable ancestors up to @root, is below the normal range.
5418 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5420 if (mem_cgroup_disabled())
5424 * The toplevel group doesn't have a configurable range, so
5425 * it's never low when looked at directly, and it is not
5426 * considered an ancestor when assessing the hierarchy.
5429 if (memcg == root_mem_cgroup)
5432 if (page_counter_read(&memcg->memory) >= memcg->low)
5435 while (memcg != root) {
5436 memcg = parent_mem_cgroup(memcg);
5438 if (memcg == root_mem_cgroup)
5441 if (page_counter_read(&memcg->memory) >= memcg->low)
5448 * mem_cgroup_try_charge - try charging a page
5449 * @page: page to charge
5450 * @mm: mm context of the victim
5451 * @gfp_mask: reclaim mode
5452 * @memcgp: charged memcg return
5454 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5455 * pages according to @gfp_mask if necessary.
5457 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5458 * Otherwise, an error code is returned.
5460 * After page->mapping has been set up, the caller must finalize the
5461 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5462 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5464 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5465 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5467 struct mem_cgroup *memcg = NULL;
5468 unsigned int nr_pages = 1;
5471 if (mem_cgroup_disabled())
5474 if (PageSwapCache(page)) {
5476 * Every swap fault against a single page tries to charge the
5477 * page, bail as early as possible. shmem_unuse() encounters
5478 * already charged pages, too. The USED bit is protected by
5479 * the page lock, which serializes swap cache removal, which
5480 * in turn serializes uncharging.
5482 if (page->mem_cgroup)
5486 if (PageTransHuge(page)) {
5487 nr_pages <<= compound_order(page);
5488 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5491 if (do_swap_account && PageSwapCache(page))
5492 memcg = try_get_mem_cgroup_from_page(page);
5494 memcg = get_mem_cgroup_from_mm(mm);
5496 ret = try_charge(memcg, gfp_mask, nr_pages);
5498 css_put(&memcg->css);
5500 if (ret == -EINTR) {
5501 memcg = root_mem_cgroup;
5510 * mem_cgroup_commit_charge - commit a page charge
5511 * @page: page to charge
5512 * @memcg: memcg to charge the page to
5513 * @lrucare: page might be on LRU already
5515 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5516 * after page->mapping has been set up. This must happen atomically
5517 * as part of the page instantiation, i.e. under the page table lock
5518 * for anonymous pages, under the page lock for page and swap cache.
5520 * In addition, the page must not be on the LRU during the commit, to
5521 * prevent racing with task migration. If it might be, use @lrucare.
5523 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5525 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5528 unsigned int nr_pages = 1;
5530 VM_BUG_ON_PAGE(!page->mapping, page);
5531 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5533 if (mem_cgroup_disabled())
5536 * Swap faults will attempt to charge the same page multiple
5537 * times. But reuse_swap_page() might have removed the page
5538 * from swapcache already, so we can't check PageSwapCache().
5543 commit_charge(page, memcg, lrucare);
5545 if (PageTransHuge(page)) {
5546 nr_pages <<= compound_order(page);
5547 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5550 local_irq_disable();
5551 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5552 memcg_check_events(memcg, page);
5555 if (do_swap_account && PageSwapCache(page)) {
5556 swp_entry_t entry = { .val = page_private(page) };
5558 * The swap entry might not get freed for a long time,
5559 * let's not wait for it. The page already received a
5560 * memory+swap charge, drop the swap entry duplicate.
5562 mem_cgroup_uncharge_swap(entry);
5567 * mem_cgroup_cancel_charge - cancel a page charge
5568 * @page: page to charge
5569 * @memcg: memcg to charge the page to
5571 * Cancel a charge transaction started by mem_cgroup_try_charge().
5573 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5575 unsigned int nr_pages = 1;
5577 if (mem_cgroup_disabled())
5580 * Swap faults will attempt to charge the same page multiple
5581 * times. But reuse_swap_page() might have removed the page
5582 * from swapcache already, so we can't check PageSwapCache().
5587 if (PageTransHuge(page)) {
5588 nr_pages <<= compound_order(page);
5589 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5592 cancel_charge(memcg, nr_pages);
5595 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5596 unsigned long nr_anon, unsigned long nr_file,
5597 unsigned long nr_huge, struct page *dummy_page)
5599 unsigned long nr_pages = nr_anon + nr_file;
5600 unsigned long flags;
5602 if (!mem_cgroup_is_root(memcg)) {
5603 page_counter_uncharge(&memcg->memory, nr_pages);
5604 if (do_swap_account)
5605 page_counter_uncharge(&memcg->memsw, nr_pages);
5606 memcg_oom_recover(memcg);
5609 local_irq_save(flags);
5610 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5611 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5612 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5613 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5614 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5615 memcg_check_events(memcg, dummy_page);
5616 local_irq_restore(flags);
5618 if (!mem_cgroup_is_root(memcg))
5619 css_put_many(&memcg->css, nr_pages);
5622 static void uncharge_list(struct list_head *page_list)
5624 struct mem_cgroup *memcg = NULL;
5625 unsigned long nr_anon = 0;
5626 unsigned long nr_file = 0;
5627 unsigned long nr_huge = 0;
5628 unsigned long pgpgout = 0;
5629 struct list_head *next;
5632 next = page_list->next;
5634 unsigned int nr_pages = 1;
5636 page = list_entry(next, struct page, lru);
5637 next = page->lru.next;
5639 VM_BUG_ON_PAGE(PageLRU(page), page);
5640 VM_BUG_ON_PAGE(page_count(page), page);
5642 if (!page->mem_cgroup)
5646 * Nobody should be changing or seriously looking at
5647 * page->mem_cgroup at this point, we have fully
5648 * exclusive access to the page.
5651 if (memcg != page->mem_cgroup) {
5653 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5655 pgpgout = nr_anon = nr_file = nr_huge = 0;
5657 memcg = page->mem_cgroup;
5660 if (PageTransHuge(page)) {
5661 nr_pages <<= compound_order(page);
5662 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5663 nr_huge += nr_pages;
5667 nr_anon += nr_pages;
5669 nr_file += nr_pages;
5671 page->mem_cgroup = NULL;
5674 } while (next != page_list);
5677 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5682 * mem_cgroup_uncharge - uncharge a page
5683 * @page: page to uncharge
5685 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5686 * mem_cgroup_commit_charge().
5688 void mem_cgroup_uncharge(struct page *page)
5690 if (mem_cgroup_disabled())
5693 /* Don't touch page->lru of any random page, pre-check: */
5694 if (!page->mem_cgroup)
5697 INIT_LIST_HEAD(&page->lru);
5698 uncharge_list(&page->lru);
5702 * mem_cgroup_uncharge_list - uncharge a list of page
5703 * @page_list: list of pages to uncharge
5705 * Uncharge a list of pages previously charged with
5706 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5708 void mem_cgroup_uncharge_list(struct list_head *page_list)
5710 if (mem_cgroup_disabled())
5713 if (!list_empty(page_list))
5714 uncharge_list(page_list);
5718 * mem_cgroup_migrate - migrate a charge to another page
5719 * @oldpage: currently charged page
5720 * @newpage: page to transfer the charge to
5721 * @lrucare: either or both pages might be on the LRU already
5723 * Migrate the charge from @oldpage to @newpage.
5725 * Both pages must be locked, @newpage->mapping must be set up.
5727 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5730 struct mem_cgroup *memcg;
5733 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5734 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5735 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5736 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5737 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5738 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5741 if (mem_cgroup_disabled())
5744 /* Page cache replacement: new page already charged? */
5745 if (newpage->mem_cgroup)
5749 * Swapcache readahead pages can get migrated before being
5750 * charged, and migration from compaction can happen to an
5751 * uncharged page when the PFN walker finds a page that
5752 * reclaim just put back on the LRU but has not released yet.
5754 memcg = oldpage->mem_cgroup;
5759 lock_page_lru(oldpage, &isolated);
5761 oldpage->mem_cgroup = NULL;
5764 unlock_page_lru(oldpage, isolated);
5766 commit_charge(newpage, memcg, lrucare);
5770 * subsys_initcall() for memory controller.
5772 * Some parts like hotcpu_notifier() have to be initialized from this context
5773 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5774 * everything that doesn't depend on a specific mem_cgroup structure should
5775 * be initialized from here.
5777 static int __init mem_cgroup_init(void)
5781 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5783 for_each_possible_cpu(cpu)
5784 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5787 for_each_node(node) {
5788 struct mem_cgroup_tree_per_node *rtpn;
5791 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5792 node_online(node) ? node : NUMA_NO_NODE);
5794 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5795 struct mem_cgroup_tree_per_zone *rtpz;
5797 rtpz = &rtpn->rb_tree_per_zone[zone];
5798 rtpz->rb_root = RB_ROOT;
5799 spin_lock_init(&rtpz->lock);
5801 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5806 subsys_initcall(mem_cgroup_init);
5808 #ifdef CONFIG_MEMCG_SWAP
5810 * mem_cgroup_swapout - transfer a memsw charge to swap
5811 * @page: page whose memsw charge to transfer
5812 * @entry: swap entry to move the charge to
5814 * Transfer the memsw charge of @page to @entry.
5816 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5818 struct mem_cgroup *memcg;
5819 unsigned short oldid;
5821 VM_BUG_ON_PAGE(PageLRU(page), page);
5822 VM_BUG_ON_PAGE(page_count(page), page);
5824 if (!do_swap_account)
5827 memcg = page->mem_cgroup;
5829 /* Readahead page, never charged */
5833 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5834 VM_BUG_ON_PAGE(oldid, page);
5835 mem_cgroup_swap_statistics(memcg, true);
5837 page->mem_cgroup = NULL;
5839 if (!mem_cgroup_is_root(memcg))
5840 page_counter_uncharge(&memcg->memory, 1);
5842 /* Caller disabled preemption with mapping->tree_lock */
5843 mem_cgroup_charge_statistics(memcg, page, -1);
5844 memcg_check_events(memcg, page);
5848 * mem_cgroup_uncharge_swap - uncharge a swap entry
5849 * @entry: swap entry to uncharge
5851 * Drop the memsw charge associated with @entry.
5853 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5855 struct mem_cgroup *memcg;
5858 if (!do_swap_account)
5861 id = swap_cgroup_record(entry, 0);
5863 memcg = mem_cgroup_from_id(id);
5865 if (!mem_cgroup_is_root(memcg))
5866 page_counter_uncharge(&memcg->memsw, 1);
5867 mem_cgroup_swap_statistics(memcg, false);
5868 css_put(&memcg->css);
5873 /* for remember boot option*/
5874 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5875 static int really_do_swap_account __initdata = 1;
5877 static int really_do_swap_account __initdata;
5880 static int __init enable_swap_account(char *s)
5882 if (!strcmp(s, "1"))
5883 really_do_swap_account = 1;
5884 else if (!strcmp(s, "0"))
5885 really_do_swap_account = 0;
5888 __setup("swapaccount=", enable_swap_account);
5890 static struct cftype memsw_cgroup_files[] = {
5892 .name = "memsw.usage_in_bytes",
5893 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5894 .read_u64 = mem_cgroup_read_u64,
5897 .name = "memsw.max_usage_in_bytes",
5898 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5899 .write = mem_cgroup_reset,
5900 .read_u64 = mem_cgroup_read_u64,
5903 .name = "memsw.limit_in_bytes",
5904 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5905 .write = mem_cgroup_write,
5906 .read_u64 = mem_cgroup_read_u64,
5909 .name = "memsw.failcnt",
5910 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5911 .write = mem_cgroup_reset,
5912 .read_u64 = mem_cgroup_read_u64,
5914 { }, /* terminate */
5917 static int __init mem_cgroup_swap_init(void)
5919 if (!mem_cgroup_disabled() && really_do_swap_account) {
5920 do_swap_account = 1;
5921 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5922 memsw_cgroup_files));
5926 subsys_initcall(mem_cgroup_swap_init);
5928 #endif /* CONFIG_MEMCG_SWAP */