1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct reclaim_iter {
147 struct mem_cgroup *position;
148 /* scan generation, increased every round-trip */
149 unsigned int generation;
153 * per-zone information in memory controller.
155 struct mem_cgroup_per_zone {
156 struct lruvec lruvec;
157 unsigned long lru_size[NR_LRU_LISTS];
159 struct reclaim_iter iter[DEF_PRIORITY + 1];
161 struct rb_node tree_node; /* RB tree node */
162 unsigned long usage_in_excess;/* Set to the value by which */
163 /* the soft limit is exceeded*/
165 struct mem_cgroup *memcg; /* Back pointer, we cannot */
166 /* use container_of */
169 struct mem_cgroup_per_node {
170 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
174 * Cgroups above their limits are maintained in a RB-Tree, independent of
175 * their hierarchy representation
178 struct mem_cgroup_tree_per_zone {
179 struct rb_root rb_root;
183 struct mem_cgroup_tree_per_node {
184 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
187 struct mem_cgroup_tree {
188 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
191 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
193 struct mem_cgroup_threshold {
194 struct eventfd_ctx *eventfd;
195 unsigned long threshold;
199 struct mem_cgroup_threshold_ary {
200 /* An array index points to threshold just below or equal to usage. */
201 int current_threshold;
202 /* Size of entries[] */
204 /* Array of thresholds */
205 struct mem_cgroup_threshold entries[0];
208 struct mem_cgroup_thresholds {
209 /* Primary thresholds array */
210 struct mem_cgroup_threshold_ary *primary;
212 * Spare threshold array.
213 * This is needed to make mem_cgroup_unregister_event() "never fail".
214 * It must be able to store at least primary->size - 1 entries.
216 struct mem_cgroup_threshold_ary *spare;
220 struct mem_cgroup_eventfd_list {
221 struct list_head list;
222 struct eventfd_ctx *eventfd;
226 * cgroup_event represents events which userspace want to receive.
228 struct mem_cgroup_event {
230 * memcg which the event belongs to.
232 struct mem_cgroup *memcg;
234 * eventfd to signal userspace about the event.
236 struct eventfd_ctx *eventfd;
238 * Each of these stored in a list by the cgroup.
240 struct list_head list;
242 * register_event() callback will be used to add new userspace
243 * waiter for changes related to this event. Use eventfd_signal()
244 * on eventfd to send notification to userspace.
246 int (*register_event)(struct mem_cgroup *memcg,
247 struct eventfd_ctx *eventfd, const char *args);
249 * unregister_event() callback will be called when userspace closes
250 * the eventfd or on cgroup removing. This callback must be set,
251 * if you want provide notification functionality.
253 void (*unregister_event)(struct mem_cgroup *memcg,
254 struct eventfd_ctx *eventfd);
256 * All fields below needed to unregister event when
257 * userspace closes eventfd.
260 wait_queue_head_t *wqh;
262 struct work_struct remove;
265 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
266 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
269 * The memory controller data structure. The memory controller controls both
270 * page cache and RSS per cgroup. We would eventually like to provide
271 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
272 * to help the administrator determine what knobs to tune.
274 * TODO: Add a water mark for the memory controller. Reclaim will begin when
275 * we hit the water mark. May be even add a low water mark, such that
276 * no reclaim occurs from a cgroup at it's low water mark, this is
277 * a feature that will be implemented much later in the future.
280 struct cgroup_subsys_state css;
282 /* Accounted resources */
283 struct page_counter memory;
284 struct page_counter memsw;
285 struct page_counter kmem;
287 unsigned long soft_limit;
289 /* vmpressure notifications */
290 struct vmpressure vmpressure;
292 /* css_online() has been completed */
296 * Should the accounting and control be hierarchical, per subtree?
302 atomic_t oom_wakeups;
305 /* OOM-Killer disable */
306 int oom_kill_disable;
308 /* protect arrays of thresholds */
309 struct mutex thresholds_lock;
311 /* thresholds for memory usage. RCU-protected */
312 struct mem_cgroup_thresholds thresholds;
314 /* thresholds for mem+swap usage. RCU-protected */
315 struct mem_cgroup_thresholds memsw_thresholds;
317 /* For oom notifier event fd */
318 struct list_head oom_notify;
321 * Should we move charges of a task when a task is moved into this
322 * mem_cgroup ? And what type of charges should we move ?
324 unsigned long move_charge_at_immigrate;
326 * set > 0 if pages under this cgroup are moving to other cgroup.
328 atomic_t moving_account;
329 /* taken only while moving_account > 0 */
330 spinlock_t move_lock;
331 struct task_struct *move_lock_task;
332 unsigned long move_lock_flags;
336 struct mem_cgroup_stat_cpu __percpu *stat;
338 * used when a cpu is offlined or other synchronizations
339 * See mem_cgroup_read_stat().
341 struct mem_cgroup_stat_cpu nocpu_base;
342 spinlock_t pcp_counter_lock;
344 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
345 struct cg_proto tcp_mem;
347 #if defined(CONFIG_MEMCG_KMEM)
348 /* Index in the kmem_cache->memcg_params->memcg_caches array */
352 int last_scanned_node;
354 nodemask_t scan_nodes;
355 atomic_t numainfo_events;
356 atomic_t numainfo_updating;
359 /* List of events which userspace want to receive */
360 struct list_head event_list;
361 spinlock_t event_list_lock;
363 struct mem_cgroup_per_node *nodeinfo[0];
364 /* WARNING: nodeinfo must be the last member here */
367 #ifdef CONFIG_MEMCG_KMEM
368 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
370 return memcg->kmemcg_id >= 0;
374 /* Stuffs for move charges at task migration. */
376 * Types of charges to be moved. "move_charge_at_immitgrate" and
377 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
380 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
381 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
385 /* "mc" and its members are protected by cgroup_mutex */
386 static struct move_charge_struct {
387 spinlock_t lock; /* for from, to */
388 struct mem_cgroup *from;
389 struct mem_cgroup *to;
390 unsigned long immigrate_flags;
391 unsigned long precharge;
392 unsigned long moved_charge;
393 unsigned long moved_swap;
394 struct task_struct *moving_task; /* a task moving charges */
395 wait_queue_head_t waitq; /* a waitq for other context */
397 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
398 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
401 static bool move_anon(void)
403 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
406 static bool move_file(void)
408 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
412 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
413 * limit reclaim to prevent infinite loops, if they ever occur.
415 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
416 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
419 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
420 MEM_CGROUP_CHARGE_TYPE_ANON,
421 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
422 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
426 /* for encoding cft->private value on file */
434 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
435 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
436 #define MEMFILE_ATTR(val) ((val) & 0xffff)
437 /* Used for OOM nofiier */
438 #define OOM_CONTROL (0)
441 * The memcg_create_mutex will be held whenever a new cgroup is created.
442 * As a consequence, any change that needs to protect against new child cgroups
443 * appearing has to hold it as well.
445 static DEFINE_MUTEX(memcg_create_mutex);
447 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
449 return s ? container_of(s, struct mem_cgroup, css) : NULL;
452 /* Some nice accessors for the vmpressure. */
453 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
456 memcg = root_mem_cgroup;
457 return &memcg->vmpressure;
460 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
462 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
465 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
467 return (memcg == root_mem_cgroup);
471 * We restrict the id in the range of [1, 65535], so it can fit into
474 #define MEM_CGROUP_ID_MAX USHRT_MAX
476 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
478 return memcg->css.id;
481 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
483 struct cgroup_subsys_state *css;
485 css = css_from_id(id, &memory_cgrp_subsys);
486 return mem_cgroup_from_css(css);
489 /* Writing them here to avoid exposing memcg's inner layout */
490 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
492 void sock_update_memcg(struct sock *sk)
494 if (mem_cgroup_sockets_enabled) {
495 struct mem_cgroup *memcg;
496 struct cg_proto *cg_proto;
498 BUG_ON(!sk->sk_prot->proto_cgroup);
500 /* Socket cloning can throw us here with sk_cgrp already
501 * filled. It won't however, necessarily happen from
502 * process context. So the test for root memcg given
503 * the current task's memcg won't help us in this case.
505 * Respecting the original socket's memcg is a better
506 * decision in this case.
509 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
510 css_get(&sk->sk_cgrp->memcg->css);
515 memcg = mem_cgroup_from_task(current);
516 cg_proto = sk->sk_prot->proto_cgroup(memcg);
517 if (!mem_cgroup_is_root(memcg) &&
518 memcg_proto_active(cg_proto) &&
519 css_tryget_online(&memcg->css)) {
520 sk->sk_cgrp = cg_proto;
525 EXPORT_SYMBOL(sock_update_memcg);
527 void sock_release_memcg(struct sock *sk)
529 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
530 struct mem_cgroup *memcg;
531 WARN_ON(!sk->sk_cgrp->memcg);
532 memcg = sk->sk_cgrp->memcg;
533 css_put(&sk->sk_cgrp->memcg->css);
537 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
539 if (!memcg || mem_cgroup_is_root(memcg))
542 return &memcg->tcp_mem;
544 EXPORT_SYMBOL(tcp_proto_cgroup);
546 static void disarm_sock_keys(struct mem_cgroup *memcg)
548 if (!memcg_proto_activated(&memcg->tcp_mem))
550 static_key_slow_dec(&memcg_socket_limit_enabled);
553 static void disarm_sock_keys(struct mem_cgroup *memcg)
558 #ifdef CONFIG_MEMCG_KMEM
560 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
561 * The main reason for not using cgroup id for this:
562 * this works better in sparse environments, where we have a lot of memcgs,
563 * but only a few kmem-limited. Or also, if we have, for instance, 200
564 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
565 * 200 entry array for that.
567 * The current size of the caches array is stored in
568 * memcg_limited_groups_array_size. It will double each time we have to
571 static DEFINE_IDA(kmem_limited_groups);
572 int memcg_limited_groups_array_size;
575 * MIN_SIZE is different than 1, because we would like to avoid going through
576 * the alloc/free process all the time. In a small machine, 4 kmem-limited
577 * cgroups is a reasonable guess. In the future, it could be a parameter or
578 * tunable, but that is strictly not necessary.
580 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
581 * this constant directly from cgroup, but it is understandable that this is
582 * better kept as an internal representation in cgroup.c. In any case, the
583 * cgrp_id space is not getting any smaller, and we don't have to necessarily
584 * increase ours as well if it increases.
586 #define MEMCG_CACHES_MIN_SIZE 4
587 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
590 * A lot of the calls to the cache allocation functions are expected to be
591 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
592 * conditional to this static branch, we'll have to allow modules that does
593 * kmem_cache_alloc and the such to see this symbol as well
595 struct static_key memcg_kmem_enabled_key;
596 EXPORT_SYMBOL(memcg_kmem_enabled_key);
598 static void memcg_free_cache_id(int id);
600 static void disarm_kmem_keys(struct mem_cgroup *memcg)
602 if (memcg_kmem_is_active(memcg)) {
603 static_key_slow_dec(&memcg_kmem_enabled_key);
604 memcg_free_cache_id(memcg->kmemcg_id);
607 * This check can't live in kmem destruction function,
608 * since the charges will outlive the cgroup
610 WARN_ON(page_counter_read(&memcg->kmem));
613 static void disarm_kmem_keys(struct mem_cgroup *memcg)
616 #endif /* CONFIG_MEMCG_KMEM */
618 static void disarm_static_keys(struct mem_cgroup *memcg)
620 disarm_sock_keys(memcg);
621 disarm_kmem_keys(memcg);
624 static struct mem_cgroup_per_zone *
625 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
627 int nid = zone_to_nid(zone);
628 int zid = zone_idx(zone);
630 return &memcg->nodeinfo[nid]->zoneinfo[zid];
633 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
638 static struct mem_cgroup_per_zone *
639 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
641 int nid = page_to_nid(page);
642 int zid = page_zonenum(page);
644 return &memcg->nodeinfo[nid]->zoneinfo[zid];
647 static struct mem_cgroup_tree_per_zone *
648 soft_limit_tree_node_zone(int nid, int zid)
650 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
653 static struct mem_cgroup_tree_per_zone *
654 soft_limit_tree_from_page(struct page *page)
656 int nid = page_to_nid(page);
657 int zid = page_zonenum(page);
659 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
662 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
663 struct mem_cgroup_tree_per_zone *mctz,
664 unsigned long new_usage_in_excess)
666 struct rb_node **p = &mctz->rb_root.rb_node;
667 struct rb_node *parent = NULL;
668 struct mem_cgroup_per_zone *mz_node;
673 mz->usage_in_excess = new_usage_in_excess;
674 if (!mz->usage_in_excess)
678 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
680 if (mz->usage_in_excess < mz_node->usage_in_excess)
683 * We can't avoid mem cgroups that are over their soft
684 * limit by the same amount
686 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
689 rb_link_node(&mz->tree_node, parent, p);
690 rb_insert_color(&mz->tree_node, &mctz->rb_root);
694 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
695 struct mem_cgroup_tree_per_zone *mctz)
699 rb_erase(&mz->tree_node, &mctz->rb_root);
703 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
704 struct mem_cgroup_tree_per_zone *mctz)
708 spin_lock_irqsave(&mctz->lock, flags);
709 __mem_cgroup_remove_exceeded(mz, mctz);
710 spin_unlock_irqrestore(&mctz->lock, flags);
713 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
715 unsigned long nr_pages = page_counter_read(&memcg->memory);
716 unsigned long soft_limit = ACCESS_ONCE(memcg->soft_limit);
717 unsigned long excess = 0;
719 if (nr_pages > soft_limit)
720 excess = nr_pages - soft_limit;
725 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
727 unsigned long excess;
728 struct mem_cgroup_per_zone *mz;
729 struct mem_cgroup_tree_per_zone *mctz;
731 mctz = soft_limit_tree_from_page(page);
733 * Necessary to update all ancestors when hierarchy is used.
734 * because their event counter is not touched.
736 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
737 mz = mem_cgroup_page_zoneinfo(memcg, page);
738 excess = soft_limit_excess(memcg);
740 * We have to update the tree if mz is on RB-tree or
741 * mem is over its softlimit.
743 if (excess || mz->on_tree) {
746 spin_lock_irqsave(&mctz->lock, flags);
747 /* if on-tree, remove it */
749 __mem_cgroup_remove_exceeded(mz, mctz);
751 * Insert again. mz->usage_in_excess will be updated.
752 * If excess is 0, no tree ops.
754 __mem_cgroup_insert_exceeded(mz, mctz, excess);
755 spin_unlock_irqrestore(&mctz->lock, flags);
760 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
762 struct mem_cgroup_tree_per_zone *mctz;
763 struct mem_cgroup_per_zone *mz;
767 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
768 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
769 mctz = soft_limit_tree_node_zone(nid, zid);
770 mem_cgroup_remove_exceeded(mz, mctz);
775 static struct mem_cgroup_per_zone *
776 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
778 struct rb_node *rightmost = NULL;
779 struct mem_cgroup_per_zone *mz;
783 rightmost = rb_last(&mctz->rb_root);
785 goto done; /* Nothing to reclaim from */
787 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
789 * Remove the node now but someone else can add it back,
790 * we will to add it back at the end of reclaim to its correct
791 * position in the tree.
793 __mem_cgroup_remove_exceeded(mz, mctz);
794 if (!soft_limit_excess(mz->memcg) ||
795 !css_tryget_online(&mz->memcg->css))
801 static struct mem_cgroup_per_zone *
802 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
804 struct mem_cgroup_per_zone *mz;
806 spin_lock_irq(&mctz->lock);
807 mz = __mem_cgroup_largest_soft_limit_node(mctz);
808 spin_unlock_irq(&mctz->lock);
813 * Implementation Note: reading percpu statistics for memcg.
815 * Both of vmstat[] and percpu_counter has threshold and do periodic
816 * synchronization to implement "quick" read. There are trade-off between
817 * reading cost and precision of value. Then, we may have a chance to implement
818 * a periodic synchronizion of counter in memcg's counter.
820 * But this _read() function is used for user interface now. The user accounts
821 * memory usage by memory cgroup and he _always_ requires exact value because
822 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
823 * have to visit all online cpus and make sum. So, for now, unnecessary
824 * synchronization is not implemented. (just implemented for cpu hotplug)
826 * If there are kernel internal actions which can make use of some not-exact
827 * value, and reading all cpu value can be performance bottleneck in some
828 * common workload, threashold and synchonization as vmstat[] should be
831 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
832 enum mem_cgroup_stat_index idx)
838 for_each_online_cpu(cpu)
839 val += per_cpu(memcg->stat->count[idx], cpu);
840 #ifdef CONFIG_HOTPLUG_CPU
841 spin_lock(&memcg->pcp_counter_lock);
842 val += memcg->nocpu_base.count[idx];
843 spin_unlock(&memcg->pcp_counter_lock);
849 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
850 enum mem_cgroup_events_index idx)
852 unsigned long val = 0;
856 for_each_online_cpu(cpu)
857 val += per_cpu(memcg->stat->events[idx], cpu);
858 #ifdef CONFIG_HOTPLUG_CPU
859 spin_lock(&memcg->pcp_counter_lock);
860 val += memcg->nocpu_base.events[idx];
861 spin_unlock(&memcg->pcp_counter_lock);
867 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
872 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
873 * counted as CACHE even if it's on ANON LRU.
876 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
879 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
882 if (PageTransHuge(page))
883 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
886 /* pagein of a big page is an event. So, ignore page size */
888 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
890 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
891 nr_pages = -nr_pages; /* for event */
894 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
897 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
899 struct mem_cgroup_per_zone *mz;
901 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
902 return mz->lru_size[lru];
905 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
907 unsigned int lru_mask)
909 unsigned long nr = 0;
912 VM_BUG_ON((unsigned)nid >= nr_node_ids);
914 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
915 struct mem_cgroup_per_zone *mz;
919 if (!(BIT(lru) & lru_mask))
921 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
922 nr += mz->lru_size[lru];
928 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
929 unsigned int lru_mask)
931 unsigned long nr = 0;
934 for_each_node_state(nid, N_MEMORY)
935 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
939 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
940 enum mem_cgroup_events_target target)
942 unsigned long val, next;
944 val = __this_cpu_read(memcg->stat->nr_page_events);
945 next = __this_cpu_read(memcg->stat->targets[target]);
946 /* from time_after() in jiffies.h */
947 if ((long)next - (long)val < 0) {
949 case MEM_CGROUP_TARGET_THRESH:
950 next = val + THRESHOLDS_EVENTS_TARGET;
952 case MEM_CGROUP_TARGET_SOFTLIMIT:
953 next = val + SOFTLIMIT_EVENTS_TARGET;
955 case MEM_CGROUP_TARGET_NUMAINFO:
956 next = val + NUMAINFO_EVENTS_TARGET;
961 __this_cpu_write(memcg->stat->targets[target], next);
968 * Check events in order.
971 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
973 /* threshold event is triggered in finer grain than soft limit */
974 if (unlikely(mem_cgroup_event_ratelimit(memcg,
975 MEM_CGROUP_TARGET_THRESH))) {
977 bool do_numainfo __maybe_unused;
979 do_softlimit = mem_cgroup_event_ratelimit(memcg,
980 MEM_CGROUP_TARGET_SOFTLIMIT);
982 do_numainfo = mem_cgroup_event_ratelimit(memcg,
983 MEM_CGROUP_TARGET_NUMAINFO);
985 mem_cgroup_threshold(memcg);
986 if (unlikely(do_softlimit))
987 mem_cgroup_update_tree(memcg, page);
989 if (unlikely(do_numainfo))
990 atomic_inc(&memcg->numainfo_events);
995 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
998 * mm_update_next_owner() may clear mm->owner to NULL
999 * if it races with swapoff, page migration, etc.
1000 * So this can be called with p == NULL.
1005 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1008 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1010 struct mem_cgroup *memcg = NULL;
1015 * Page cache insertions can happen withou an
1016 * actual mm context, e.g. during disk probing
1017 * on boot, loopback IO, acct() writes etc.
1020 memcg = root_mem_cgroup;
1022 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1023 if (unlikely(!memcg))
1024 memcg = root_mem_cgroup;
1026 } while (!css_tryget_online(&memcg->css));
1032 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1033 * @root: hierarchy root
1034 * @prev: previously returned memcg, NULL on first invocation
1035 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1037 * Returns references to children of the hierarchy below @root, or
1038 * @root itself, or %NULL after a full round-trip.
1040 * Caller must pass the return value in @prev on subsequent
1041 * invocations for reference counting, or use mem_cgroup_iter_break()
1042 * to cancel a hierarchy walk before the round-trip is complete.
1044 * Reclaimers can specify a zone and a priority level in @reclaim to
1045 * divide up the memcgs in the hierarchy among all concurrent
1046 * reclaimers operating on the same zone and priority.
1048 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1049 struct mem_cgroup *prev,
1050 struct mem_cgroup_reclaim_cookie *reclaim)
1052 struct reclaim_iter *uninitialized_var(iter);
1053 struct cgroup_subsys_state *css = NULL;
1054 struct mem_cgroup *memcg = NULL;
1055 struct mem_cgroup *pos = NULL;
1057 if (mem_cgroup_disabled())
1061 root = root_mem_cgroup;
1063 if (prev && !reclaim)
1066 if (!root->use_hierarchy && root != root_mem_cgroup) {
1075 struct mem_cgroup_per_zone *mz;
1077 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1078 iter = &mz->iter[reclaim->priority];
1080 if (prev && reclaim->generation != iter->generation)
1084 pos = ACCESS_ONCE(iter->position);
1086 * A racing update may change the position and
1087 * put the last reference, hence css_tryget(),
1088 * or retry to see the updated position.
1090 } while (pos && !css_tryget(&pos->css));
1097 css = css_next_descendant_pre(css, &root->css);
1100 * Reclaimers share the hierarchy walk, and a
1101 * new one might jump in right at the end of
1102 * the hierarchy - make sure they see at least
1103 * one group and restart from the beginning.
1111 * Verify the css and acquire a reference. The root
1112 * is provided by the caller, so we know it's alive
1113 * and kicking, and don't take an extra reference.
1115 memcg = mem_cgroup_from_css(css);
1117 if (css == &root->css)
1120 if (css_tryget(css)) {
1122 * Make sure the memcg is initialized:
1123 * mem_cgroup_css_online() orders the the
1124 * initialization against setting the flag.
1126 if (smp_load_acquire(&memcg->initialized))
1136 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1138 css_get(&memcg->css);
1144 * pairs with css_tryget when dereferencing iter->position
1153 reclaim->generation = iter->generation;
1159 if (prev && prev != root)
1160 css_put(&prev->css);
1166 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1167 * @root: hierarchy root
1168 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1170 void mem_cgroup_iter_break(struct mem_cgroup *root,
1171 struct mem_cgroup *prev)
1174 root = root_mem_cgroup;
1175 if (prev && prev != root)
1176 css_put(&prev->css);
1180 * Iteration constructs for visiting all cgroups (under a tree). If
1181 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1182 * be used for reference counting.
1184 #define for_each_mem_cgroup_tree(iter, root) \
1185 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1187 iter = mem_cgroup_iter(root, iter, NULL))
1189 #define for_each_mem_cgroup(iter) \
1190 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1192 iter = mem_cgroup_iter(NULL, iter, NULL))
1194 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1196 struct mem_cgroup *memcg;
1199 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1200 if (unlikely(!memcg))
1205 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1208 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1216 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1219 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1220 * @zone: zone of the wanted lruvec
1221 * @memcg: memcg of the wanted lruvec
1223 * Returns the lru list vector holding pages for the given @zone and
1224 * @mem. This can be the global zone lruvec, if the memory controller
1227 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1228 struct mem_cgroup *memcg)
1230 struct mem_cgroup_per_zone *mz;
1231 struct lruvec *lruvec;
1233 if (mem_cgroup_disabled()) {
1234 lruvec = &zone->lruvec;
1238 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1239 lruvec = &mz->lruvec;
1242 * Since a node can be onlined after the mem_cgroup was created,
1243 * we have to be prepared to initialize lruvec->zone here;
1244 * and if offlined then reonlined, we need to reinitialize it.
1246 if (unlikely(lruvec->zone != zone))
1247 lruvec->zone = zone;
1252 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1254 * @zone: zone of the page
1256 * This function is only safe when following the LRU page isolation
1257 * and putback protocol: the LRU lock must be held, and the page must
1258 * either be PageLRU() or the caller must have isolated/allocated it.
1260 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1262 struct mem_cgroup_per_zone *mz;
1263 struct mem_cgroup *memcg;
1264 struct lruvec *lruvec;
1266 if (mem_cgroup_disabled()) {
1267 lruvec = &zone->lruvec;
1271 memcg = page->mem_cgroup;
1273 * Swapcache readahead pages are added to the LRU - and
1274 * possibly migrated - before they are charged.
1277 memcg = root_mem_cgroup;
1279 mz = mem_cgroup_page_zoneinfo(memcg, page);
1280 lruvec = &mz->lruvec;
1283 * Since a node can be onlined after the mem_cgroup was created,
1284 * we have to be prepared to initialize lruvec->zone here;
1285 * and if offlined then reonlined, we need to reinitialize it.
1287 if (unlikely(lruvec->zone != zone))
1288 lruvec->zone = zone;
1293 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1294 * @lruvec: mem_cgroup per zone lru vector
1295 * @lru: index of lru list the page is sitting on
1296 * @nr_pages: positive when adding or negative when removing
1298 * This function must be called when a page is added to or removed from an
1301 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1304 struct mem_cgroup_per_zone *mz;
1305 unsigned long *lru_size;
1307 if (mem_cgroup_disabled())
1310 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1311 lru_size = mz->lru_size + lru;
1312 *lru_size += nr_pages;
1313 VM_BUG_ON((long)(*lru_size) < 0);
1316 bool mem_cgroup_is_descendant(struct mem_cgroup *memcg, struct mem_cgroup *root)
1320 if (!root->use_hierarchy)
1322 return cgroup_is_descendant(memcg->css.cgroup, root->css.cgroup);
1325 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1327 struct mem_cgroup *task_memcg;
1328 struct task_struct *p;
1331 p = find_lock_task_mm(task);
1333 task_memcg = get_mem_cgroup_from_mm(p->mm);
1337 * All threads may have already detached their mm's, but the oom
1338 * killer still needs to detect if they have already been oom
1339 * killed to prevent needlessly killing additional tasks.
1342 task_memcg = mem_cgroup_from_task(task);
1343 css_get(&task_memcg->css);
1346 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1347 css_put(&task_memcg->css);
1351 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1353 unsigned long inactive_ratio;
1354 unsigned long inactive;
1355 unsigned long active;
1358 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1359 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1361 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1363 inactive_ratio = int_sqrt(10 * gb);
1367 return inactive * inactive_ratio < active;
1370 #define mem_cgroup_from_counter(counter, member) \
1371 container_of(counter, struct mem_cgroup, member)
1374 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1375 * @memcg: the memory cgroup
1377 * Returns the maximum amount of memory @mem can be charged with, in
1380 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1382 unsigned long margin = 0;
1383 unsigned long count;
1384 unsigned long limit;
1386 count = page_counter_read(&memcg->memory);
1387 limit = ACCESS_ONCE(memcg->memory.limit);
1389 margin = limit - count;
1391 if (do_swap_account) {
1392 count = page_counter_read(&memcg->memsw);
1393 limit = ACCESS_ONCE(memcg->memsw.limit);
1395 margin = min(margin, limit - count);
1401 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1404 if (mem_cgroup_disabled() || !memcg->css.parent)
1405 return vm_swappiness;
1407 return memcg->swappiness;
1411 * A routine for checking "mem" is under move_account() or not.
1413 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1414 * moving cgroups. This is for waiting at high-memory pressure
1417 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1419 struct mem_cgroup *from;
1420 struct mem_cgroup *to;
1423 * Unlike task_move routines, we access mc.to, mc.from not under
1424 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1426 spin_lock(&mc.lock);
1432 ret = mem_cgroup_is_descendant(from, memcg) ||
1433 mem_cgroup_is_descendant(to, memcg);
1435 spin_unlock(&mc.lock);
1439 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1441 if (mc.moving_task && current != mc.moving_task) {
1442 if (mem_cgroup_under_move(memcg)) {
1444 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1445 /* moving charge context might have finished. */
1448 finish_wait(&mc.waitq, &wait);
1455 #define K(x) ((x) << (PAGE_SHIFT-10))
1457 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1458 * @memcg: The memory cgroup that went over limit
1459 * @p: Task that is going to be killed
1461 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1464 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1466 /* oom_info_lock ensures that parallel ooms do not interleave */
1467 static DEFINE_MUTEX(oom_info_lock);
1468 struct mem_cgroup *iter;
1474 mutex_lock(&oom_info_lock);
1477 pr_info("Task in ");
1478 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1479 pr_cont(" killed as a result of limit of ");
1480 pr_cont_cgroup_path(memcg->css.cgroup);
1485 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1486 K((u64)page_counter_read(&memcg->memory)),
1487 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1488 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1489 K((u64)page_counter_read(&memcg->memsw)),
1490 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1491 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1492 K((u64)page_counter_read(&memcg->kmem)),
1493 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1495 for_each_mem_cgroup_tree(iter, memcg) {
1496 pr_info("Memory cgroup stats for ");
1497 pr_cont_cgroup_path(iter->css.cgroup);
1500 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1501 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1503 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1504 K(mem_cgroup_read_stat(iter, i)));
1507 for (i = 0; i < NR_LRU_LISTS; i++)
1508 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1509 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1513 mutex_unlock(&oom_info_lock);
1517 * This function returns the number of memcg under hierarchy tree. Returns
1518 * 1(self count) if no children.
1520 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1523 struct mem_cgroup *iter;
1525 for_each_mem_cgroup_tree(iter, memcg)
1531 * Return the memory (and swap, if configured) limit for a memcg.
1533 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1535 unsigned long limit;
1537 limit = memcg->memory.limit;
1538 if (mem_cgroup_swappiness(memcg)) {
1539 unsigned long memsw_limit;
1541 memsw_limit = memcg->memsw.limit;
1542 limit = min(limit + total_swap_pages, memsw_limit);
1547 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1550 struct mem_cgroup *iter;
1551 unsigned long chosen_points = 0;
1552 unsigned long totalpages;
1553 unsigned int points = 0;
1554 struct task_struct *chosen = NULL;
1557 * If current has a pending SIGKILL or is exiting, then automatically
1558 * select it. The goal is to allow it to allocate so that it may
1559 * quickly exit and free its memory.
1561 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1562 set_thread_flag(TIF_MEMDIE);
1566 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1567 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1568 for_each_mem_cgroup_tree(iter, memcg) {
1569 struct css_task_iter it;
1570 struct task_struct *task;
1572 css_task_iter_start(&iter->css, &it);
1573 while ((task = css_task_iter_next(&it))) {
1574 switch (oom_scan_process_thread(task, totalpages, NULL,
1576 case OOM_SCAN_SELECT:
1578 put_task_struct(chosen);
1580 chosen_points = ULONG_MAX;
1581 get_task_struct(chosen);
1583 case OOM_SCAN_CONTINUE:
1585 case OOM_SCAN_ABORT:
1586 css_task_iter_end(&it);
1587 mem_cgroup_iter_break(memcg, iter);
1589 put_task_struct(chosen);
1594 points = oom_badness(task, memcg, NULL, totalpages);
1595 if (!points || points < chosen_points)
1597 /* Prefer thread group leaders for display purposes */
1598 if (points == chosen_points &&
1599 thread_group_leader(chosen))
1603 put_task_struct(chosen);
1605 chosen_points = points;
1606 get_task_struct(chosen);
1608 css_task_iter_end(&it);
1613 points = chosen_points * 1000 / totalpages;
1614 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1615 NULL, "Memory cgroup out of memory");
1618 #if MAX_NUMNODES > 1
1621 * test_mem_cgroup_node_reclaimable
1622 * @memcg: the target memcg
1623 * @nid: the node ID to be checked.
1624 * @noswap : specify true here if the user wants flle only information.
1626 * This function returns whether the specified memcg contains any
1627 * reclaimable pages on a node. Returns true if there are any reclaimable
1628 * pages in the node.
1630 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1631 int nid, bool noswap)
1633 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1635 if (noswap || !total_swap_pages)
1637 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1644 * Always updating the nodemask is not very good - even if we have an empty
1645 * list or the wrong list here, we can start from some node and traverse all
1646 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1649 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1653 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1654 * pagein/pageout changes since the last update.
1656 if (!atomic_read(&memcg->numainfo_events))
1658 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1661 /* make a nodemask where this memcg uses memory from */
1662 memcg->scan_nodes = node_states[N_MEMORY];
1664 for_each_node_mask(nid, node_states[N_MEMORY]) {
1666 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1667 node_clear(nid, memcg->scan_nodes);
1670 atomic_set(&memcg->numainfo_events, 0);
1671 atomic_set(&memcg->numainfo_updating, 0);
1675 * Selecting a node where we start reclaim from. Because what we need is just
1676 * reducing usage counter, start from anywhere is O,K. Considering
1677 * memory reclaim from current node, there are pros. and cons.
1679 * Freeing memory from current node means freeing memory from a node which
1680 * we'll use or we've used. So, it may make LRU bad. And if several threads
1681 * hit limits, it will see a contention on a node. But freeing from remote
1682 * node means more costs for memory reclaim because of memory latency.
1684 * Now, we use round-robin. Better algorithm is welcomed.
1686 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1690 mem_cgroup_may_update_nodemask(memcg);
1691 node = memcg->last_scanned_node;
1693 node = next_node(node, memcg->scan_nodes);
1694 if (node == MAX_NUMNODES)
1695 node = first_node(memcg->scan_nodes);
1697 * We call this when we hit limit, not when pages are added to LRU.
1698 * No LRU may hold pages because all pages are UNEVICTABLE or
1699 * memcg is too small and all pages are not on LRU. In that case,
1700 * we use curret node.
1702 if (unlikely(node == MAX_NUMNODES))
1703 node = numa_node_id();
1705 memcg->last_scanned_node = node;
1709 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1715 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1718 unsigned long *total_scanned)
1720 struct mem_cgroup *victim = NULL;
1723 unsigned long excess;
1724 unsigned long nr_scanned;
1725 struct mem_cgroup_reclaim_cookie reclaim = {
1730 excess = soft_limit_excess(root_memcg);
1733 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1738 * If we have not been able to reclaim
1739 * anything, it might because there are
1740 * no reclaimable pages under this hierarchy
1745 * We want to do more targeted reclaim.
1746 * excess >> 2 is not to excessive so as to
1747 * reclaim too much, nor too less that we keep
1748 * coming back to reclaim from this cgroup
1750 if (total >= (excess >> 2) ||
1751 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1756 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1758 *total_scanned += nr_scanned;
1759 if (!soft_limit_excess(root_memcg))
1762 mem_cgroup_iter_break(root_memcg, victim);
1766 #ifdef CONFIG_LOCKDEP
1767 static struct lockdep_map memcg_oom_lock_dep_map = {
1768 .name = "memcg_oom_lock",
1772 static DEFINE_SPINLOCK(memcg_oom_lock);
1775 * Check OOM-Killer is already running under our hierarchy.
1776 * If someone is running, return false.
1778 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1780 struct mem_cgroup *iter, *failed = NULL;
1782 spin_lock(&memcg_oom_lock);
1784 for_each_mem_cgroup_tree(iter, memcg) {
1785 if (iter->oom_lock) {
1787 * this subtree of our hierarchy is already locked
1788 * so we cannot give a lock.
1791 mem_cgroup_iter_break(memcg, iter);
1794 iter->oom_lock = true;
1799 * OK, we failed to lock the whole subtree so we have
1800 * to clean up what we set up to the failing subtree
1802 for_each_mem_cgroup_tree(iter, memcg) {
1803 if (iter == failed) {
1804 mem_cgroup_iter_break(memcg, iter);
1807 iter->oom_lock = false;
1810 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1812 spin_unlock(&memcg_oom_lock);
1817 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1819 struct mem_cgroup *iter;
1821 spin_lock(&memcg_oom_lock);
1822 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1823 for_each_mem_cgroup_tree(iter, memcg)
1824 iter->oom_lock = false;
1825 spin_unlock(&memcg_oom_lock);
1828 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1830 struct mem_cgroup *iter;
1832 for_each_mem_cgroup_tree(iter, memcg)
1833 atomic_inc(&iter->under_oom);
1836 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1838 struct mem_cgroup *iter;
1841 * When a new child is created while the hierarchy is under oom,
1842 * mem_cgroup_oom_lock() may not be called. We have to use
1843 * atomic_add_unless() here.
1845 for_each_mem_cgroup_tree(iter, memcg)
1846 atomic_add_unless(&iter->under_oom, -1, 0);
1849 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1851 struct oom_wait_info {
1852 struct mem_cgroup *memcg;
1856 static int memcg_oom_wake_function(wait_queue_t *wait,
1857 unsigned mode, int sync, void *arg)
1859 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1860 struct mem_cgroup *oom_wait_memcg;
1861 struct oom_wait_info *oom_wait_info;
1863 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1864 oom_wait_memcg = oom_wait_info->memcg;
1866 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1867 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1869 return autoremove_wake_function(wait, mode, sync, arg);
1872 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1874 atomic_inc(&memcg->oom_wakeups);
1875 /* for filtering, pass "memcg" as argument. */
1876 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1879 static void memcg_oom_recover(struct mem_cgroup *memcg)
1881 if (memcg && atomic_read(&memcg->under_oom))
1882 memcg_wakeup_oom(memcg);
1885 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1887 if (!current->memcg_oom.may_oom)
1890 * We are in the middle of the charge context here, so we
1891 * don't want to block when potentially sitting on a callstack
1892 * that holds all kinds of filesystem and mm locks.
1894 * Also, the caller may handle a failed allocation gracefully
1895 * (like optional page cache readahead) and so an OOM killer
1896 * invocation might not even be necessary.
1898 * That's why we don't do anything here except remember the
1899 * OOM context and then deal with it at the end of the page
1900 * fault when the stack is unwound, the locks are released,
1901 * and when we know whether the fault was overall successful.
1903 css_get(&memcg->css);
1904 current->memcg_oom.memcg = memcg;
1905 current->memcg_oom.gfp_mask = mask;
1906 current->memcg_oom.order = order;
1910 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1911 * @handle: actually kill/wait or just clean up the OOM state
1913 * This has to be called at the end of a page fault if the memcg OOM
1914 * handler was enabled.
1916 * Memcg supports userspace OOM handling where failed allocations must
1917 * sleep on a waitqueue until the userspace task resolves the
1918 * situation. Sleeping directly in the charge context with all kinds
1919 * of locks held is not a good idea, instead we remember an OOM state
1920 * in the task and mem_cgroup_oom_synchronize() has to be called at
1921 * the end of the page fault to complete the OOM handling.
1923 * Returns %true if an ongoing memcg OOM situation was detected and
1924 * completed, %false otherwise.
1926 bool mem_cgroup_oom_synchronize(bool handle)
1928 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1929 struct oom_wait_info owait;
1932 /* OOM is global, do not handle */
1939 owait.memcg = memcg;
1940 owait.wait.flags = 0;
1941 owait.wait.func = memcg_oom_wake_function;
1942 owait.wait.private = current;
1943 INIT_LIST_HEAD(&owait.wait.task_list);
1945 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1946 mem_cgroup_mark_under_oom(memcg);
1948 locked = mem_cgroup_oom_trylock(memcg);
1951 mem_cgroup_oom_notify(memcg);
1953 if (locked && !memcg->oom_kill_disable) {
1954 mem_cgroup_unmark_under_oom(memcg);
1955 finish_wait(&memcg_oom_waitq, &owait.wait);
1956 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1957 current->memcg_oom.order);
1960 mem_cgroup_unmark_under_oom(memcg);
1961 finish_wait(&memcg_oom_waitq, &owait.wait);
1965 mem_cgroup_oom_unlock(memcg);
1967 * There is no guarantee that an OOM-lock contender
1968 * sees the wakeups triggered by the OOM kill
1969 * uncharges. Wake any sleepers explicitely.
1971 memcg_oom_recover(memcg);
1974 current->memcg_oom.memcg = NULL;
1975 css_put(&memcg->css);
1980 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1981 * @page: page that is going to change accounted state
1983 * This function must mark the beginning of an accounted page state
1984 * change to prevent double accounting when the page is concurrently
1985 * being moved to another memcg:
1987 * memcg = mem_cgroup_begin_page_stat(page);
1988 * if (TestClearPageState(page))
1989 * mem_cgroup_update_page_stat(memcg, state, -1);
1990 * mem_cgroup_end_page_stat(memcg);
1992 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1994 struct mem_cgroup *memcg;
1995 unsigned long flags;
1998 * The RCU lock is held throughout the transaction. The fast
1999 * path can get away without acquiring the memcg->move_lock
2000 * because page moving starts with an RCU grace period.
2002 * The RCU lock also protects the memcg from being freed when
2003 * the page state that is going to change is the only thing
2004 * preventing the page from being uncharged.
2005 * E.g. end-writeback clearing PageWriteback(), which allows
2006 * migration to go ahead and uncharge the page before the
2007 * account transaction might be complete.
2011 if (mem_cgroup_disabled())
2014 memcg = page->mem_cgroup;
2015 if (unlikely(!memcg))
2018 if (atomic_read(&memcg->moving_account) <= 0)
2021 spin_lock_irqsave(&memcg->move_lock, flags);
2022 if (memcg != page->mem_cgroup) {
2023 spin_unlock_irqrestore(&memcg->move_lock, flags);
2028 * When charge migration first begins, we can have locked and
2029 * unlocked page stat updates happening concurrently. Track
2030 * the task who has the lock for mem_cgroup_end_page_stat().
2032 memcg->move_lock_task = current;
2033 memcg->move_lock_flags = flags;
2039 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2040 * @memcg: the memcg that was accounted against
2042 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
2044 if (memcg && memcg->move_lock_task == current) {
2045 unsigned long flags = memcg->move_lock_flags;
2047 memcg->move_lock_task = NULL;
2048 memcg->move_lock_flags = 0;
2050 spin_unlock_irqrestore(&memcg->move_lock, flags);
2057 * mem_cgroup_update_page_stat - update page state statistics
2058 * @memcg: memcg to account against
2059 * @idx: page state item to account
2060 * @val: number of pages (positive or negative)
2062 * See mem_cgroup_begin_page_stat() for locking requirements.
2064 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2065 enum mem_cgroup_stat_index idx, int val)
2067 VM_BUG_ON(!rcu_read_lock_held());
2070 this_cpu_add(memcg->stat->count[idx], val);
2074 * size of first charge trial. "32" comes from vmscan.c's magic value.
2075 * TODO: maybe necessary to use big numbers in big irons.
2077 #define CHARGE_BATCH 32U
2078 struct memcg_stock_pcp {
2079 struct mem_cgroup *cached; /* this never be root cgroup */
2080 unsigned int nr_pages;
2081 struct work_struct work;
2082 unsigned long flags;
2083 #define FLUSHING_CACHED_CHARGE 0
2085 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2086 static DEFINE_MUTEX(percpu_charge_mutex);
2089 * consume_stock: Try to consume stocked charge on this cpu.
2090 * @memcg: memcg to consume from.
2091 * @nr_pages: how many pages to charge.
2093 * The charges will only happen if @memcg matches the current cpu's memcg
2094 * stock, and at least @nr_pages are available in that stock. Failure to
2095 * service an allocation will refill the stock.
2097 * returns true if successful, false otherwise.
2099 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2101 struct memcg_stock_pcp *stock;
2104 if (nr_pages > CHARGE_BATCH)
2107 stock = &get_cpu_var(memcg_stock);
2108 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2109 stock->nr_pages -= nr_pages;
2112 put_cpu_var(memcg_stock);
2117 * Returns stocks cached in percpu and reset cached information.
2119 static void drain_stock(struct memcg_stock_pcp *stock)
2121 struct mem_cgroup *old = stock->cached;
2123 if (stock->nr_pages) {
2124 page_counter_uncharge(&old->memory, stock->nr_pages);
2125 if (do_swap_account)
2126 page_counter_uncharge(&old->memsw, stock->nr_pages);
2127 css_put_many(&old->css, stock->nr_pages);
2128 stock->nr_pages = 0;
2130 stock->cached = NULL;
2134 * This must be called under preempt disabled or must be called by
2135 * a thread which is pinned to local cpu.
2137 static void drain_local_stock(struct work_struct *dummy)
2139 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2141 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2144 static void __init memcg_stock_init(void)
2148 for_each_possible_cpu(cpu) {
2149 struct memcg_stock_pcp *stock =
2150 &per_cpu(memcg_stock, cpu);
2151 INIT_WORK(&stock->work, drain_local_stock);
2156 * Cache charges(val) to local per_cpu area.
2157 * This will be consumed by consume_stock() function, later.
2159 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2161 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2163 if (stock->cached != memcg) { /* reset if necessary */
2165 stock->cached = memcg;
2167 stock->nr_pages += nr_pages;
2168 put_cpu_var(memcg_stock);
2172 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2173 * of the hierarchy under it.
2175 static void drain_all_stock(struct mem_cgroup *root_memcg)
2179 /* If someone's already draining, avoid adding running more workers. */
2180 if (!mutex_trylock(&percpu_charge_mutex))
2182 /* Notify other cpus that system-wide "drain" is running */
2185 for_each_online_cpu(cpu) {
2186 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2187 struct mem_cgroup *memcg;
2189 memcg = stock->cached;
2190 if (!memcg || !stock->nr_pages)
2192 if (!mem_cgroup_is_descendant(memcg, root_memcg))
2194 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2196 drain_local_stock(&stock->work);
2198 schedule_work_on(cpu, &stock->work);
2203 mutex_unlock(&percpu_charge_mutex);
2207 * This function drains percpu counter value from DEAD cpu and
2208 * move it to local cpu. Note that this function can be preempted.
2210 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2214 spin_lock(&memcg->pcp_counter_lock);
2215 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2216 long x = per_cpu(memcg->stat->count[i], cpu);
2218 per_cpu(memcg->stat->count[i], cpu) = 0;
2219 memcg->nocpu_base.count[i] += x;
2221 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2222 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2224 per_cpu(memcg->stat->events[i], cpu) = 0;
2225 memcg->nocpu_base.events[i] += x;
2227 spin_unlock(&memcg->pcp_counter_lock);
2230 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2231 unsigned long action,
2234 int cpu = (unsigned long)hcpu;
2235 struct memcg_stock_pcp *stock;
2236 struct mem_cgroup *iter;
2238 if (action == CPU_ONLINE)
2241 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2244 for_each_mem_cgroup(iter)
2245 mem_cgroup_drain_pcp_counter(iter, cpu);
2247 stock = &per_cpu(memcg_stock, cpu);
2252 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2253 unsigned int nr_pages)
2255 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2256 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2257 struct mem_cgroup *mem_over_limit;
2258 struct page_counter *counter;
2259 unsigned long nr_reclaimed;
2260 bool may_swap = true;
2261 bool drained = false;
2264 if (mem_cgroup_is_root(memcg))
2267 if (consume_stock(memcg, nr_pages))
2270 if (!do_swap_account ||
2271 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2272 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2274 if (do_swap_account)
2275 page_counter_uncharge(&memcg->memsw, batch);
2276 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2278 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2282 if (batch > nr_pages) {
2288 * Unlike in global OOM situations, memcg is not in a physical
2289 * memory shortage. Allow dying and OOM-killed tasks to
2290 * bypass the last charges so that they can exit quickly and
2291 * free their memory.
2293 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2294 fatal_signal_pending(current) ||
2295 current->flags & PF_EXITING))
2298 if (unlikely(task_in_memcg_oom(current)))
2301 if (!(gfp_mask & __GFP_WAIT))
2304 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2305 gfp_mask, may_swap);
2307 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2311 drain_all_stock(mem_over_limit);
2316 if (gfp_mask & __GFP_NORETRY)
2319 * Even though the limit is exceeded at this point, reclaim
2320 * may have been able to free some pages. Retry the charge
2321 * before killing the task.
2323 * Only for regular pages, though: huge pages are rather
2324 * unlikely to succeed so close to the limit, and we fall back
2325 * to regular pages anyway in case of failure.
2327 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2330 * At task move, charge accounts can be doubly counted. So, it's
2331 * better to wait until the end of task_move if something is going on.
2333 if (mem_cgroup_wait_acct_move(mem_over_limit))
2339 if (gfp_mask & __GFP_NOFAIL)
2342 if (fatal_signal_pending(current))
2345 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2347 if (!(gfp_mask & __GFP_NOFAIL))
2353 css_get_many(&memcg->css, batch);
2354 if (batch > nr_pages)
2355 refill_stock(memcg, batch - nr_pages);
2360 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2362 if (mem_cgroup_is_root(memcg))
2365 page_counter_uncharge(&memcg->memory, nr_pages);
2366 if (do_swap_account)
2367 page_counter_uncharge(&memcg->memsw, nr_pages);
2369 css_put_many(&memcg->css, nr_pages);
2373 * A helper function to get mem_cgroup from ID. must be called under
2374 * rcu_read_lock(). The caller is responsible for calling
2375 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2376 * refcnt from swap can be called against removed memcg.)
2378 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2380 /* ID 0 is unused ID */
2383 return mem_cgroup_from_id(id);
2387 * try_get_mem_cgroup_from_page - look up page's memcg association
2390 * Look up, get a css reference, and return the memcg that owns @page.
2392 * The page must be locked to prevent racing with swap-in and page
2393 * cache charges. If coming from an unlocked page table, the caller
2394 * must ensure the page is on the LRU or this can race with charging.
2396 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2398 struct mem_cgroup *memcg;
2402 VM_BUG_ON_PAGE(!PageLocked(page), page);
2404 memcg = page->mem_cgroup;
2406 if (!css_tryget_online(&memcg->css))
2408 } else if (PageSwapCache(page)) {
2409 ent.val = page_private(page);
2410 id = lookup_swap_cgroup_id(ent);
2412 memcg = mem_cgroup_lookup(id);
2413 if (memcg && !css_tryget_online(&memcg->css))
2420 static void lock_page_lru(struct page *page, int *isolated)
2422 struct zone *zone = page_zone(page);
2424 spin_lock_irq(&zone->lru_lock);
2425 if (PageLRU(page)) {
2426 struct lruvec *lruvec;
2428 lruvec = mem_cgroup_page_lruvec(page, zone);
2430 del_page_from_lru_list(page, lruvec, page_lru(page));
2436 static void unlock_page_lru(struct page *page, int isolated)
2438 struct zone *zone = page_zone(page);
2441 struct lruvec *lruvec;
2443 lruvec = mem_cgroup_page_lruvec(page, zone);
2444 VM_BUG_ON_PAGE(PageLRU(page), page);
2446 add_page_to_lru_list(page, lruvec, page_lru(page));
2448 spin_unlock_irq(&zone->lru_lock);
2451 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2456 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2459 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2460 * may already be on some other mem_cgroup's LRU. Take care of it.
2463 lock_page_lru(page, &isolated);
2466 * Nobody should be changing or seriously looking at
2467 * page->mem_cgroup at this point:
2469 * - the page is uncharged
2471 * - the page is off-LRU
2473 * - an anonymous fault has exclusive page access, except for
2474 * a locked page table
2476 * - a page cache insertion, a swapin fault, or a migration
2477 * have the page locked
2479 page->mem_cgroup = memcg;
2482 unlock_page_lru(page, isolated);
2485 #ifdef CONFIG_MEMCG_KMEM
2486 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2487 unsigned long nr_pages)
2489 struct page_counter *counter;
2492 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2496 ret = try_charge(memcg, gfp, nr_pages);
2497 if (ret == -EINTR) {
2499 * try_charge() chose to bypass to root due to OOM kill or
2500 * fatal signal. Since our only options are to either fail
2501 * the allocation or charge it to this cgroup, do it as a
2502 * temporary condition. But we can't fail. From a kmem/slab
2503 * perspective, the cache has already been selected, by
2504 * mem_cgroup_kmem_get_cache(), so it is too late to change
2507 * This condition will only trigger if the task entered
2508 * memcg_charge_kmem in a sane state, but was OOM-killed
2509 * during try_charge() above. Tasks that were already dying
2510 * when the allocation triggers should have been already
2511 * directed to the root cgroup in memcontrol.h
2513 page_counter_charge(&memcg->memory, nr_pages);
2514 if (do_swap_account)
2515 page_counter_charge(&memcg->memsw, nr_pages);
2516 css_get_many(&memcg->css, nr_pages);
2519 page_counter_uncharge(&memcg->kmem, nr_pages);
2524 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2526 page_counter_uncharge(&memcg->memory, nr_pages);
2527 if (do_swap_account)
2528 page_counter_uncharge(&memcg->memsw, nr_pages);
2530 page_counter_uncharge(&memcg->kmem, nr_pages);
2532 css_put_many(&memcg->css, nr_pages);
2536 * helper for acessing a memcg's index. It will be used as an index in the
2537 * child cache array in kmem_cache, and also to derive its name. This function
2538 * will return -1 when this is not a kmem-limited memcg.
2540 int memcg_cache_id(struct mem_cgroup *memcg)
2542 return memcg ? memcg->kmemcg_id : -1;
2545 static int memcg_alloc_cache_id(void)
2550 id = ida_simple_get(&kmem_limited_groups,
2551 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2555 if (id < memcg_limited_groups_array_size)
2559 * There's no space for the new id in memcg_caches arrays,
2560 * so we have to grow them.
2563 size = 2 * (id + 1);
2564 if (size < MEMCG_CACHES_MIN_SIZE)
2565 size = MEMCG_CACHES_MIN_SIZE;
2566 else if (size > MEMCG_CACHES_MAX_SIZE)
2567 size = MEMCG_CACHES_MAX_SIZE;
2569 err = memcg_update_all_caches(size);
2571 ida_simple_remove(&kmem_limited_groups, id);
2577 static void memcg_free_cache_id(int id)
2579 ida_simple_remove(&kmem_limited_groups, id);
2583 * We should update the current array size iff all caches updates succeed. This
2584 * can only be done from the slab side. The slab mutex needs to be held when
2587 void memcg_update_array_size(int num)
2589 memcg_limited_groups_array_size = num;
2592 struct memcg_kmem_cache_create_work {
2593 struct mem_cgroup *memcg;
2594 struct kmem_cache *cachep;
2595 struct work_struct work;
2598 static void memcg_kmem_cache_create_func(struct work_struct *w)
2600 struct memcg_kmem_cache_create_work *cw =
2601 container_of(w, struct memcg_kmem_cache_create_work, work);
2602 struct mem_cgroup *memcg = cw->memcg;
2603 struct kmem_cache *cachep = cw->cachep;
2605 memcg_create_kmem_cache(memcg, cachep);
2607 css_put(&memcg->css);
2612 * Enqueue the creation of a per-memcg kmem_cache.
2614 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2615 struct kmem_cache *cachep)
2617 struct memcg_kmem_cache_create_work *cw;
2619 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2623 css_get(&memcg->css);
2626 cw->cachep = cachep;
2627 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2629 schedule_work(&cw->work);
2632 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2633 struct kmem_cache *cachep)
2636 * We need to stop accounting when we kmalloc, because if the
2637 * corresponding kmalloc cache is not yet created, the first allocation
2638 * in __memcg_schedule_kmem_cache_create will recurse.
2640 * However, it is better to enclose the whole function. Depending on
2641 * the debugging options enabled, INIT_WORK(), for instance, can
2642 * trigger an allocation. This too, will make us recurse. Because at
2643 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2644 * the safest choice is to do it like this, wrapping the whole function.
2646 current->memcg_kmem_skip_account = 1;
2647 __memcg_schedule_kmem_cache_create(memcg, cachep);
2648 current->memcg_kmem_skip_account = 0;
2652 * Return the kmem_cache we're supposed to use for a slab allocation.
2653 * We try to use the current memcg's version of the cache.
2655 * If the cache does not exist yet, if we are the first user of it,
2656 * we either create it immediately, if possible, or create it asynchronously
2658 * In the latter case, we will let the current allocation go through with
2659 * the original cache.
2661 * Can't be called in interrupt context or from kernel threads.
2662 * This function needs to be called with rcu_read_lock() held.
2664 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2666 struct mem_cgroup *memcg;
2667 struct kmem_cache *memcg_cachep;
2669 VM_BUG_ON(!cachep->memcg_params);
2670 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
2672 if (current->memcg_kmem_skip_account)
2675 memcg = get_mem_cgroup_from_mm(current->mm);
2676 if (!memcg_kmem_is_active(memcg))
2679 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
2680 if (likely(memcg_cachep))
2681 return memcg_cachep;
2684 * If we are in a safe context (can wait, and not in interrupt
2685 * context), we could be be predictable and return right away.
2686 * This would guarantee that the allocation being performed
2687 * already belongs in the new cache.
2689 * However, there are some clashes that can arrive from locking.
2690 * For instance, because we acquire the slab_mutex while doing
2691 * memcg_create_kmem_cache, this means no further allocation
2692 * could happen with the slab_mutex held. So it's better to
2695 memcg_schedule_kmem_cache_create(memcg, cachep);
2697 css_put(&memcg->css);
2701 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2703 if (!is_root_cache(cachep))
2704 css_put(&cachep->memcg_params->memcg->css);
2708 * We need to verify if the allocation against current->mm->owner's memcg is
2709 * possible for the given order. But the page is not allocated yet, so we'll
2710 * need a further commit step to do the final arrangements.
2712 * It is possible for the task to switch cgroups in this mean time, so at
2713 * commit time, we can't rely on task conversion any longer. We'll then use
2714 * the handle argument to return to the caller which cgroup we should commit
2715 * against. We could also return the memcg directly and avoid the pointer
2716 * passing, but a boolean return value gives better semantics considering
2717 * the compiled-out case as well.
2719 * Returning true means the allocation is possible.
2722 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2724 struct mem_cgroup *memcg;
2729 memcg = get_mem_cgroup_from_mm(current->mm);
2731 if (!memcg_kmem_is_active(memcg)) {
2732 css_put(&memcg->css);
2736 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2740 css_put(&memcg->css);
2744 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2747 VM_BUG_ON(mem_cgroup_is_root(memcg));
2749 /* The page allocation failed. Revert */
2751 memcg_uncharge_kmem(memcg, 1 << order);
2754 page->mem_cgroup = memcg;
2757 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2759 struct mem_cgroup *memcg = page->mem_cgroup;
2764 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2766 memcg_uncharge_kmem(memcg, 1 << order);
2767 page->mem_cgroup = NULL;
2769 #endif /* CONFIG_MEMCG_KMEM */
2771 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2774 * Because tail pages are not marked as "used", set it. We're under
2775 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2776 * charge/uncharge will be never happen and move_account() is done under
2777 * compound_lock(), so we don't have to take care of races.
2779 void mem_cgroup_split_huge_fixup(struct page *head)
2783 if (mem_cgroup_disabled())
2786 for (i = 1; i < HPAGE_PMD_NR; i++)
2787 head[i].mem_cgroup = head->mem_cgroup;
2789 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2792 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2795 * mem_cgroup_move_account - move account of the page
2797 * @nr_pages: number of regular pages (>1 for huge pages)
2798 * @from: mem_cgroup which the page is moved from.
2799 * @to: mem_cgroup which the page is moved to. @from != @to.
2801 * The caller must confirm following.
2802 * - page is not on LRU (isolate_page() is useful.)
2803 * - compound_lock is held when nr_pages > 1
2805 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2808 static int mem_cgroup_move_account(struct page *page,
2809 unsigned int nr_pages,
2810 struct mem_cgroup *from,
2811 struct mem_cgroup *to)
2813 unsigned long flags;
2816 VM_BUG_ON(from == to);
2817 VM_BUG_ON_PAGE(PageLRU(page), page);
2819 * The page is isolated from LRU. So, collapse function
2820 * will not handle this page. But page splitting can happen.
2821 * Do this check under compound_page_lock(). The caller should
2825 if (nr_pages > 1 && !PageTransHuge(page))
2829 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
2830 * of its source page while we change it: page migration takes
2831 * both pages off the LRU, but page cache replacement doesn't.
2833 if (!trylock_page(page))
2837 if (page->mem_cgroup != from)
2840 spin_lock_irqsave(&from->move_lock, flags);
2842 if (!PageAnon(page) && page_mapped(page)) {
2843 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2845 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2849 if (PageWriteback(page)) {
2850 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2852 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2857 * It is safe to change page->mem_cgroup here because the page
2858 * is referenced, charged, and isolated - we can't race with
2859 * uncharging, charging, migration, or LRU putback.
2862 /* caller should have done css_get */
2863 page->mem_cgroup = to;
2864 spin_unlock_irqrestore(&from->move_lock, flags);
2868 local_irq_disable();
2869 mem_cgroup_charge_statistics(to, page, nr_pages);
2870 memcg_check_events(to, page);
2871 mem_cgroup_charge_statistics(from, page, -nr_pages);
2872 memcg_check_events(from, page);
2880 #ifdef CONFIG_MEMCG_SWAP
2881 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2884 int val = (charge) ? 1 : -1;
2885 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2889 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2890 * @entry: swap entry to be moved
2891 * @from: mem_cgroup which the entry is moved from
2892 * @to: mem_cgroup which the entry is moved to
2894 * It succeeds only when the swap_cgroup's record for this entry is the same
2895 * as the mem_cgroup's id of @from.
2897 * Returns 0 on success, -EINVAL on failure.
2899 * The caller must have charged to @to, IOW, called page_counter_charge() about
2900 * both res and memsw, and called css_get().
2902 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2903 struct mem_cgroup *from, struct mem_cgroup *to)
2905 unsigned short old_id, new_id;
2907 old_id = mem_cgroup_id(from);
2908 new_id = mem_cgroup_id(to);
2910 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2911 mem_cgroup_swap_statistics(from, false);
2912 mem_cgroup_swap_statistics(to, true);
2918 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2919 struct mem_cgroup *from, struct mem_cgroup *to)
2925 static DEFINE_MUTEX(memcg_limit_mutex);
2927 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2928 unsigned long limit)
2930 unsigned long curusage;
2931 unsigned long oldusage;
2932 bool enlarge = false;
2937 * For keeping hierarchical_reclaim simple, how long we should retry
2938 * is depends on callers. We set our retry-count to be function
2939 * of # of children which we should visit in this loop.
2941 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2942 mem_cgroup_count_children(memcg);
2944 oldusage = page_counter_read(&memcg->memory);
2947 if (signal_pending(current)) {
2952 mutex_lock(&memcg_limit_mutex);
2953 if (limit > memcg->memsw.limit) {
2954 mutex_unlock(&memcg_limit_mutex);
2958 if (limit > memcg->memory.limit)
2960 ret = page_counter_limit(&memcg->memory, limit);
2961 mutex_unlock(&memcg_limit_mutex);
2966 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2968 curusage = page_counter_read(&memcg->memory);
2969 /* Usage is reduced ? */
2970 if (curusage >= oldusage)
2973 oldusage = curusage;
2974 } while (retry_count);
2976 if (!ret && enlarge)
2977 memcg_oom_recover(memcg);
2982 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2983 unsigned long limit)
2985 unsigned long curusage;
2986 unsigned long oldusage;
2987 bool enlarge = false;
2991 /* see mem_cgroup_resize_res_limit */
2992 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2993 mem_cgroup_count_children(memcg);
2995 oldusage = page_counter_read(&memcg->memsw);
2998 if (signal_pending(current)) {
3003 mutex_lock(&memcg_limit_mutex);
3004 if (limit < memcg->memory.limit) {
3005 mutex_unlock(&memcg_limit_mutex);
3009 if (limit > memcg->memsw.limit)
3011 ret = page_counter_limit(&memcg->memsw, limit);
3012 mutex_unlock(&memcg_limit_mutex);
3017 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3019 curusage = page_counter_read(&memcg->memsw);
3020 /* Usage is reduced ? */
3021 if (curusage >= oldusage)
3024 oldusage = curusage;
3025 } while (retry_count);
3027 if (!ret && enlarge)
3028 memcg_oom_recover(memcg);
3033 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3035 unsigned long *total_scanned)
3037 unsigned long nr_reclaimed = 0;
3038 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3039 unsigned long reclaimed;
3041 struct mem_cgroup_tree_per_zone *mctz;
3042 unsigned long excess;
3043 unsigned long nr_scanned;
3048 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3050 * This loop can run a while, specially if mem_cgroup's continuously
3051 * keep exceeding their soft limit and putting the system under
3058 mz = mem_cgroup_largest_soft_limit_node(mctz);
3063 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3064 gfp_mask, &nr_scanned);
3065 nr_reclaimed += reclaimed;
3066 *total_scanned += nr_scanned;
3067 spin_lock_irq(&mctz->lock);
3068 __mem_cgroup_remove_exceeded(mz, mctz);
3071 * If we failed to reclaim anything from this memory cgroup
3072 * it is time to move on to the next cgroup
3076 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3078 excess = soft_limit_excess(mz->memcg);
3080 * One school of thought says that we should not add
3081 * back the node to the tree if reclaim returns 0.
3082 * But our reclaim could return 0, simply because due
3083 * to priority we are exposing a smaller subset of
3084 * memory to reclaim from. Consider this as a longer
3087 /* If excess == 0, no tree ops */
3088 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3089 spin_unlock_irq(&mctz->lock);
3090 css_put(&mz->memcg->css);
3093 * Could not reclaim anything and there are no more
3094 * mem cgroups to try or we seem to be looping without
3095 * reclaiming anything.
3097 if (!nr_reclaimed &&
3099 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3101 } while (!nr_reclaimed);
3103 css_put(&next_mz->memcg->css);
3104 return nr_reclaimed;
3108 * Test whether @memcg has children, dead or alive. Note that this
3109 * function doesn't care whether @memcg has use_hierarchy enabled and
3110 * returns %true if there are child csses according to the cgroup
3111 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3113 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3118 * The lock does not prevent addition or deletion of children, but
3119 * it prevents a new child from being initialized based on this
3120 * parent in css_online(), so it's enough to decide whether
3121 * hierarchically inherited attributes can still be changed or not.
3123 lockdep_assert_held(&memcg_create_mutex);
3126 ret = css_next_child(NULL, &memcg->css);
3132 * Reclaims as many pages from the given memcg as possible and moves
3133 * the rest to the parent.
3135 * Caller is responsible for holding css reference for memcg.
3137 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3139 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3141 /* we call try-to-free pages for make this cgroup empty */
3142 lru_add_drain_all();
3143 /* try to free all pages in this cgroup */
3144 while (nr_retries && page_counter_read(&memcg->memory)) {
3147 if (signal_pending(current))
3150 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3154 /* maybe some writeback is necessary */
3155 congestion_wait(BLK_RW_ASYNC, HZ/10);
3163 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3164 char *buf, size_t nbytes,
3167 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3169 if (mem_cgroup_is_root(memcg))
3171 return mem_cgroup_force_empty(memcg) ?: nbytes;
3174 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3177 return mem_cgroup_from_css(css)->use_hierarchy;
3180 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3181 struct cftype *cft, u64 val)
3184 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3185 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3187 mutex_lock(&memcg_create_mutex);
3189 if (memcg->use_hierarchy == val)
3193 * If parent's use_hierarchy is set, we can't make any modifications
3194 * in the child subtrees. If it is unset, then the change can
3195 * occur, provided the current cgroup has no children.
3197 * For the root cgroup, parent_mem is NULL, we allow value to be
3198 * set if there are no children.
3200 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3201 (val == 1 || val == 0)) {
3202 if (!memcg_has_children(memcg))
3203 memcg->use_hierarchy = val;
3210 mutex_unlock(&memcg_create_mutex);
3215 static unsigned long tree_stat(struct mem_cgroup *memcg,
3216 enum mem_cgroup_stat_index idx)
3218 struct mem_cgroup *iter;
3221 /* Per-cpu values can be negative, use a signed accumulator */
3222 for_each_mem_cgroup_tree(iter, memcg)
3223 val += mem_cgroup_read_stat(iter, idx);
3225 if (val < 0) /* race ? */
3230 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3234 if (mem_cgroup_is_root(memcg)) {
3235 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3236 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3238 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3241 val = page_counter_read(&memcg->memory);
3243 val = page_counter_read(&memcg->memsw);
3245 return val << PAGE_SHIFT;
3256 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3259 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3260 struct page_counter *counter;
3262 switch (MEMFILE_TYPE(cft->private)) {
3264 counter = &memcg->memory;
3267 counter = &memcg->memsw;
3270 counter = &memcg->kmem;
3276 switch (MEMFILE_ATTR(cft->private)) {
3278 if (counter == &memcg->memory)
3279 return mem_cgroup_usage(memcg, false);
3280 if (counter == &memcg->memsw)
3281 return mem_cgroup_usage(memcg, true);
3282 return (u64)page_counter_read(counter) * PAGE_SIZE;
3284 return (u64)counter->limit * PAGE_SIZE;
3286 return (u64)counter->watermark * PAGE_SIZE;
3288 return counter->failcnt;
3289 case RES_SOFT_LIMIT:
3290 return (u64)memcg->soft_limit * PAGE_SIZE;
3296 #ifdef CONFIG_MEMCG_KMEM
3297 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3298 unsigned long nr_pages)
3303 if (memcg_kmem_is_active(memcg))
3307 * For simplicity, we won't allow this to be disabled. It also can't
3308 * be changed if the cgroup has children already, or if tasks had
3311 * If tasks join before we set the limit, a person looking at
3312 * kmem.usage_in_bytes will have no way to determine when it took
3313 * place, which makes the value quite meaningless.
3315 * After it first became limited, changes in the value of the limit are
3316 * of course permitted.
3318 mutex_lock(&memcg_create_mutex);
3319 if (cgroup_has_tasks(memcg->css.cgroup) ||
3320 (memcg->use_hierarchy && memcg_has_children(memcg)))
3322 mutex_unlock(&memcg_create_mutex);
3326 memcg_id = memcg_alloc_cache_id();
3333 * We couldn't have accounted to this cgroup, because it hasn't got
3334 * activated yet, so this should succeed.
3336 err = page_counter_limit(&memcg->kmem, nr_pages);
3339 static_key_slow_inc(&memcg_kmem_enabled_key);
3341 * A memory cgroup is considered kmem-active as soon as it gets
3342 * kmemcg_id. Setting the id after enabling static branching will
3343 * guarantee no one starts accounting before all call sites are
3346 memcg->kmemcg_id = memcg_id;
3351 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3352 unsigned long limit)
3356 mutex_lock(&memcg_limit_mutex);
3357 if (!memcg_kmem_is_active(memcg))
3358 ret = memcg_activate_kmem(memcg, limit);
3360 ret = page_counter_limit(&memcg->kmem, limit);
3361 mutex_unlock(&memcg_limit_mutex);
3365 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3368 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3373 mutex_lock(&memcg_limit_mutex);
3375 * If the parent cgroup is not kmem-active now, it cannot be activated
3376 * after this point, because it has at least one child already.
3378 if (memcg_kmem_is_active(parent))
3379 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3380 mutex_unlock(&memcg_limit_mutex);
3384 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3385 unsigned long limit)
3389 #endif /* CONFIG_MEMCG_KMEM */
3392 * The user of this function is...
3395 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3396 char *buf, size_t nbytes, loff_t off)
3398 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3399 unsigned long nr_pages;
3402 buf = strstrip(buf);
3403 ret = page_counter_memparse(buf, &nr_pages);
3407 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3409 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3413 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3415 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3418 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3421 ret = memcg_update_kmem_limit(memcg, nr_pages);
3425 case RES_SOFT_LIMIT:
3426 memcg->soft_limit = nr_pages;
3430 return ret ?: nbytes;
3433 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3434 size_t nbytes, loff_t off)
3436 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3437 struct page_counter *counter;
3439 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3441 counter = &memcg->memory;
3444 counter = &memcg->memsw;
3447 counter = &memcg->kmem;
3453 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3455 page_counter_reset_watermark(counter);
3458 counter->failcnt = 0;
3467 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3470 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3474 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3475 struct cftype *cft, u64 val)
3477 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3479 if (val >= (1 << NR_MOVE_TYPE))
3483 * No kind of locking is needed in here, because ->can_attach() will
3484 * check this value once in the beginning of the process, and then carry
3485 * on with stale data. This means that changes to this value will only
3486 * affect task migrations starting after the change.
3488 memcg->move_charge_at_immigrate = val;
3492 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3493 struct cftype *cft, u64 val)
3500 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3504 unsigned int lru_mask;
3507 static const struct numa_stat stats[] = {
3508 { "total", LRU_ALL },
3509 { "file", LRU_ALL_FILE },
3510 { "anon", LRU_ALL_ANON },
3511 { "unevictable", BIT(LRU_UNEVICTABLE) },
3513 const struct numa_stat *stat;
3516 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3518 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3519 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3520 seq_printf(m, "%s=%lu", stat->name, nr);
3521 for_each_node_state(nid, N_MEMORY) {
3522 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3524 seq_printf(m, " N%d=%lu", nid, nr);
3529 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3530 struct mem_cgroup *iter;
3533 for_each_mem_cgroup_tree(iter, memcg)
3534 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3535 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3536 for_each_node_state(nid, N_MEMORY) {
3538 for_each_mem_cgroup_tree(iter, memcg)
3539 nr += mem_cgroup_node_nr_lru_pages(
3540 iter, nid, stat->lru_mask);
3541 seq_printf(m, " N%d=%lu", nid, nr);
3548 #endif /* CONFIG_NUMA */
3550 static int memcg_stat_show(struct seq_file *m, void *v)
3552 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3553 unsigned long memory, memsw;
3554 struct mem_cgroup *mi;
3557 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3559 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3560 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3562 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3563 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3566 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3567 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3568 mem_cgroup_read_events(memcg, i));
3570 for (i = 0; i < NR_LRU_LISTS; i++)
3571 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3572 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3574 /* Hierarchical information */
3575 memory = memsw = PAGE_COUNTER_MAX;
3576 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3577 memory = min(memory, mi->memory.limit);
3578 memsw = min(memsw, mi->memsw.limit);
3580 seq_printf(m, "hierarchical_memory_limit %llu\n",
3581 (u64)memory * PAGE_SIZE);
3582 if (do_swap_account)
3583 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3584 (u64)memsw * PAGE_SIZE);
3586 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3589 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3591 for_each_mem_cgroup_tree(mi, memcg)
3592 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3593 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3596 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3597 unsigned long long val = 0;
3599 for_each_mem_cgroup_tree(mi, memcg)
3600 val += mem_cgroup_read_events(mi, i);
3601 seq_printf(m, "total_%s %llu\n",
3602 mem_cgroup_events_names[i], val);
3605 for (i = 0; i < NR_LRU_LISTS; i++) {
3606 unsigned long long val = 0;
3608 for_each_mem_cgroup_tree(mi, memcg)
3609 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3610 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3613 #ifdef CONFIG_DEBUG_VM
3616 struct mem_cgroup_per_zone *mz;
3617 struct zone_reclaim_stat *rstat;
3618 unsigned long recent_rotated[2] = {0, 0};
3619 unsigned long recent_scanned[2] = {0, 0};
3621 for_each_online_node(nid)
3622 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3623 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3624 rstat = &mz->lruvec.reclaim_stat;
3626 recent_rotated[0] += rstat->recent_rotated[0];
3627 recent_rotated[1] += rstat->recent_rotated[1];
3628 recent_scanned[0] += rstat->recent_scanned[0];
3629 recent_scanned[1] += rstat->recent_scanned[1];
3631 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3632 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3633 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3634 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3641 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3644 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3646 return mem_cgroup_swappiness(memcg);
3649 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3650 struct cftype *cft, u64 val)
3652 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3658 memcg->swappiness = val;
3660 vm_swappiness = val;
3665 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3667 struct mem_cgroup_threshold_ary *t;
3668 unsigned long usage;
3673 t = rcu_dereference(memcg->thresholds.primary);
3675 t = rcu_dereference(memcg->memsw_thresholds.primary);
3680 usage = mem_cgroup_usage(memcg, swap);
3683 * current_threshold points to threshold just below or equal to usage.
3684 * If it's not true, a threshold was crossed after last
3685 * call of __mem_cgroup_threshold().
3687 i = t->current_threshold;
3690 * Iterate backward over array of thresholds starting from
3691 * current_threshold and check if a threshold is crossed.
3692 * If none of thresholds below usage is crossed, we read
3693 * only one element of the array here.
3695 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3696 eventfd_signal(t->entries[i].eventfd, 1);
3698 /* i = current_threshold + 1 */
3702 * Iterate forward over array of thresholds starting from
3703 * current_threshold+1 and check if a threshold is crossed.
3704 * If none of thresholds above usage is crossed, we read
3705 * only one element of the array here.
3707 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3708 eventfd_signal(t->entries[i].eventfd, 1);
3710 /* Update current_threshold */
3711 t->current_threshold = i - 1;
3716 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3719 __mem_cgroup_threshold(memcg, false);
3720 if (do_swap_account)
3721 __mem_cgroup_threshold(memcg, true);
3723 memcg = parent_mem_cgroup(memcg);
3727 static int compare_thresholds(const void *a, const void *b)
3729 const struct mem_cgroup_threshold *_a = a;
3730 const struct mem_cgroup_threshold *_b = b;
3732 if (_a->threshold > _b->threshold)
3735 if (_a->threshold < _b->threshold)
3741 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3743 struct mem_cgroup_eventfd_list *ev;
3745 spin_lock(&memcg_oom_lock);
3747 list_for_each_entry(ev, &memcg->oom_notify, list)
3748 eventfd_signal(ev->eventfd, 1);
3750 spin_unlock(&memcg_oom_lock);
3754 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3756 struct mem_cgroup *iter;
3758 for_each_mem_cgroup_tree(iter, memcg)
3759 mem_cgroup_oom_notify_cb(iter);
3762 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3763 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3765 struct mem_cgroup_thresholds *thresholds;
3766 struct mem_cgroup_threshold_ary *new;
3767 unsigned long threshold;
3768 unsigned long usage;
3771 ret = page_counter_memparse(args, &threshold);
3775 mutex_lock(&memcg->thresholds_lock);
3778 thresholds = &memcg->thresholds;
3779 usage = mem_cgroup_usage(memcg, false);
3780 } else if (type == _MEMSWAP) {
3781 thresholds = &memcg->memsw_thresholds;
3782 usage = mem_cgroup_usage(memcg, true);
3786 /* Check if a threshold crossed before adding a new one */
3787 if (thresholds->primary)
3788 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3790 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3792 /* Allocate memory for new array of thresholds */
3793 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3801 /* Copy thresholds (if any) to new array */
3802 if (thresholds->primary) {
3803 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3804 sizeof(struct mem_cgroup_threshold));
3807 /* Add new threshold */
3808 new->entries[size - 1].eventfd = eventfd;
3809 new->entries[size - 1].threshold = threshold;
3811 /* Sort thresholds. Registering of new threshold isn't time-critical */
3812 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3813 compare_thresholds, NULL);
3815 /* Find current threshold */
3816 new->current_threshold = -1;
3817 for (i = 0; i < size; i++) {
3818 if (new->entries[i].threshold <= usage) {
3820 * new->current_threshold will not be used until
3821 * rcu_assign_pointer(), so it's safe to increment
3824 ++new->current_threshold;
3829 /* Free old spare buffer and save old primary buffer as spare */
3830 kfree(thresholds->spare);
3831 thresholds->spare = thresholds->primary;
3833 rcu_assign_pointer(thresholds->primary, new);
3835 /* To be sure that nobody uses thresholds */
3839 mutex_unlock(&memcg->thresholds_lock);
3844 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3845 struct eventfd_ctx *eventfd, const char *args)
3847 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3850 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3851 struct eventfd_ctx *eventfd, const char *args)
3853 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3856 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3857 struct eventfd_ctx *eventfd, enum res_type type)
3859 struct mem_cgroup_thresholds *thresholds;
3860 struct mem_cgroup_threshold_ary *new;
3861 unsigned long usage;
3864 mutex_lock(&memcg->thresholds_lock);
3867 thresholds = &memcg->thresholds;
3868 usage = mem_cgroup_usage(memcg, false);
3869 } else if (type == _MEMSWAP) {
3870 thresholds = &memcg->memsw_thresholds;
3871 usage = mem_cgroup_usage(memcg, true);
3875 if (!thresholds->primary)
3878 /* Check if a threshold crossed before removing */
3879 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3881 /* Calculate new number of threshold */
3883 for (i = 0; i < thresholds->primary->size; i++) {
3884 if (thresholds->primary->entries[i].eventfd != eventfd)
3888 new = thresholds->spare;
3890 /* Set thresholds array to NULL if we don't have thresholds */
3899 /* Copy thresholds and find current threshold */
3900 new->current_threshold = -1;
3901 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3902 if (thresholds->primary->entries[i].eventfd == eventfd)
3905 new->entries[j] = thresholds->primary->entries[i];
3906 if (new->entries[j].threshold <= usage) {
3908 * new->current_threshold will not be used
3909 * until rcu_assign_pointer(), so it's safe to increment
3912 ++new->current_threshold;
3918 /* Swap primary and spare array */
3919 thresholds->spare = thresholds->primary;
3920 /* If all events are unregistered, free the spare array */
3922 kfree(thresholds->spare);
3923 thresholds->spare = NULL;
3926 rcu_assign_pointer(thresholds->primary, new);
3928 /* To be sure that nobody uses thresholds */
3931 mutex_unlock(&memcg->thresholds_lock);
3934 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3935 struct eventfd_ctx *eventfd)
3937 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3940 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3941 struct eventfd_ctx *eventfd)
3943 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3946 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3947 struct eventfd_ctx *eventfd, const char *args)
3949 struct mem_cgroup_eventfd_list *event;
3951 event = kmalloc(sizeof(*event), GFP_KERNEL);
3955 spin_lock(&memcg_oom_lock);
3957 event->eventfd = eventfd;
3958 list_add(&event->list, &memcg->oom_notify);
3960 /* already in OOM ? */
3961 if (atomic_read(&memcg->under_oom))
3962 eventfd_signal(eventfd, 1);
3963 spin_unlock(&memcg_oom_lock);
3968 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3969 struct eventfd_ctx *eventfd)
3971 struct mem_cgroup_eventfd_list *ev, *tmp;
3973 spin_lock(&memcg_oom_lock);
3975 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3976 if (ev->eventfd == eventfd) {
3977 list_del(&ev->list);
3982 spin_unlock(&memcg_oom_lock);
3985 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3987 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3989 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3990 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
3994 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3995 struct cftype *cft, u64 val)
3997 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3999 /* cannot set to root cgroup and only 0 and 1 are allowed */
4000 if (!css->parent || !((val == 0) || (val == 1)))
4003 memcg->oom_kill_disable = val;
4005 memcg_oom_recover(memcg);
4010 #ifdef CONFIG_MEMCG_KMEM
4011 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4015 ret = memcg_propagate_kmem(memcg);
4019 return mem_cgroup_sockets_init(memcg, ss);
4022 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4024 memcg_destroy_kmem_caches(memcg);
4025 mem_cgroup_sockets_destroy(memcg);
4028 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4033 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4039 * DO NOT USE IN NEW FILES.
4041 * "cgroup.event_control" implementation.
4043 * This is way over-engineered. It tries to support fully configurable
4044 * events for each user. Such level of flexibility is completely
4045 * unnecessary especially in the light of the planned unified hierarchy.
4047 * Please deprecate this and replace with something simpler if at all
4052 * Unregister event and free resources.
4054 * Gets called from workqueue.
4056 static void memcg_event_remove(struct work_struct *work)
4058 struct mem_cgroup_event *event =
4059 container_of(work, struct mem_cgroup_event, remove);
4060 struct mem_cgroup *memcg = event->memcg;
4062 remove_wait_queue(event->wqh, &event->wait);
4064 event->unregister_event(memcg, event->eventfd);
4066 /* Notify userspace the event is going away. */
4067 eventfd_signal(event->eventfd, 1);
4069 eventfd_ctx_put(event->eventfd);
4071 css_put(&memcg->css);
4075 * Gets called on POLLHUP on eventfd when user closes it.
4077 * Called with wqh->lock held and interrupts disabled.
4079 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4080 int sync, void *key)
4082 struct mem_cgroup_event *event =
4083 container_of(wait, struct mem_cgroup_event, wait);
4084 struct mem_cgroup *memcg = event->memcg;
4085 unsigned long flags = (unsigned long)key;
4087 if (flags & POLLHUP) {
4089 * If the event has been detached at cgroup removal, we
4090 * can simply return knowing the other side will cleanup
4093 * We can't race against event freeing since the other
4094 * side will require wqh->lock via remove_wait_queue(),
4097 spin_lock(&memcg->event_list_lock);
4098 if (!list_empty(&event->list)) {
4099 list_del_init(&event->list);
4101 * We are in atomic context, but cgroup_event_remove()
4102 * may sleep, so we have to call it in workqueue.
4104 schedule_work(&event->remove);
4106 spin_unlock(&memcg->event_list_lock);
4112 static void memcg_event_ptable_queue_proc(struct file *file,
4113 wait_queue_head_t *wqh, poll_table *pt)
4115 struct mem_cgroup_event *event =
4116 container_of(pt, struct mem_cgroup_event, pt);
4119 add_wait_queue(wqh, &event->wait);
4123 * DO NOT USE IN NEW FILES.
4125 * Parse input and register new cgroup event handler.
4127 * Input must be in format '<event_fd> <control_fd> <args>'.
4128 * Interpretation of args is defined by control file implementation.
4130 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4131 char *buf, size_t nbytes, loff_t off)
4133 struct cgroup_subsys_state *css = of_css(of);
4134 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4135 struct mem_cgroup_event *event;
4136 struct cgroup_subsys_state *cfile_css;
4137 unsigned int efd, cfd;
4144 buf = strstrip(buf);
4146 efd = simple_strtoul(buf, &endp, 10);
4151 cfd = simple_strtoul(buf, &endp, 10);
4152 if ((*endp != ' ') && (*endp != '\0'))
4156 event = kzalloc(sizeof(*event), GFP_KERNEL);
4160 event->memcg = memcg;
4161 INIT_LIST_HEAD(&event->list);
4162 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4163 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4164 INIT_WORK(&event->remove, memcg_event_remove);
4172 event->eventfd = eventfd_ctx_fileget(efile.file);
4173 if (IS_ERR(event->eventfd)) {
4174 ret = PTR_ERR(event->eventfd);
4181 goto out_put_eventfd;
4184 /* the process need read permission on control file */
4185 /* AV: shouldn't we check that it's been opened for read instead? */
4186 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4191 * Determine the event callbacks and set them in @event. This used
4192 * to be done via struct cftype but cgroup core no longer knows
4193 * about these events. The following is crude but the whole thing
4194 * is for compatibility anyway.
4196 * DO NOT ADD NEW FILES.
4198 name = cfile.file->f_path.dentry->d_name.name;
4200 if (!strcmp(name, "memory.usage_in_bytes")) {
4201 event->register_event = mem_cgroup_usage_register_event;
4202 event->unregister_event = mem_cgroup_usage_unregister_event;
4203 } else if (!strcmp(name, "memory.oom_control")) {
4204 event->register_event = mem_cgroup_oom_register_event;
4205 event->unregister_event = mem_cgroup_oom_unregister_event;
4206 } else if (!strcmp(name, "memory.pressure_level")) {
4207 event->register_event = vmpressure_register_event;
4208 event->unregister_event = vmpressure_unregister_event;
4209 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4210 event->register_event = memsw_cgroup_usage_register_event;
4211 event->unregister_event = memsw_cgroup_usage_unregister_event;
4218 * Verify @cfile should belong to @css. Also, remaining events are
4219 * automatically removed on cgroup destruction but the removal is
4220 * asynchronous, so take an extra ref on @css.
4222 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4223 &memory_cgrp_subsys);
4225 if (IS_ERR(cfile_css))
4227 if (cfile_css != css) {
4232 ret = event->register_event(memcg, event->eventfd, buf);
4236 efile.file->f_op->poll(efile.file, &event->pt);
4238 spin_lock(&memcg->event_list_lock);
4239 list_add(&event->list, &memcg->event_list);
4240 spin_unlock(&memcg->event_list_lock);
4252 eventfd_ctx_put(event->eventfd);
4261 static struct cftype mem_cgroup_files[] = {
4263 .name = "usage_in_bytes",
4264 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4265 .read_u64 = mem_cgroup_read_u64,
4268 .name = "max_usage_in_bytes",
4269 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4270 .write = mem_cgroup_reset,
4271 .read_u64 = mem_cgroup_read_u64,
4274 .name = "limit_in_bytes",
4275 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4276 .write = mem_cgroup_write,
4277 .read_u64 = mem_cgroup_read_u64,
4280 .name = "soft_limit_in_bytes",
4281 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4282 .write = mem_cgroup_write,
4283 .read_u64 = mem_cgroup_read_u64,
4287 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4288 .write = mem_cgroup_reset,
4289 .read_u64 = mem_cgroup_read_u64,
4293 .seq_show = memcg_stat_show,
4296 .name = "force_empty",
4297 .write = mem_cgroup_force_empty_write,
4300 .name = "use_hierarchy",
4301 .write_u64 = mem_cgroup_hierarchy_write,
4302 .read_u64 = mem_cgroup_hierarchy_read,
4305 .name = "cgroup.event_control", /* XXX: for compat */
4306 .write = memcg_write_event_control,
4307 .flags = CFTYPE_NO_PREFIX,
4311 .name = "swappiness",
4312 .read_u64 = mem_cgroup_swappiness_read,
4313 .write_u64 = mem_cgroup_swappiness_write,
4316 .name = "move_charge_at_immigrate",
4317 .read_u64 = mem_cgroup_move_charge_read,
4318 .write_u64 = mem_cgroup_move_charge_write,
4321 .name = "oom_control",
4322 .seq_show = mem_cgroup_oom_control_read,
4323 .write_u64 = mem_cgroup_oom_control_write,
4324 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4327 .name = "pressure_level",
4331 .name = "numa_stat",
4332 .seq_show = memcg_numa_stat_show,
4335 #ifdef CONFIG_MEMCG_KMEM
4337 .name = "kmem.limit_in_bytes",
4338 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4339 .write = mem_cgroup_write,
4340 .read_u64 = mem_cgroup_read_u64,
4343 .name = "kmem.usage_in_bytes",
4344 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4345 .read_u64 = mem_cgroup_read_u64,
4348 .name = "kmem.failcnt",
4349 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4350 .write = mem_cgroup_reset,
4351 .read_u64 = mem_cgroup_read_u64,
4354 .name = "kmem.max_usage_in_bytes",
4355 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4356 .write = mem_cgroup_reset,
4357 .read_u64 = mem_cgroup_read_u64,
4359 #ifdef CONFIG_SLABINFO
4361 .name = "kmem.slabinfo",
4362 .seq_start = slab_start,
4363 .seq_next = slab_next,
4364 .seq_stop = slab_stop,
4365 .seq_show = memcg_slab_show,
4369 { }, /* terminate */
4372 #ifdef CONFIG_MEMCG_SWAP
4373 static struct cftype memsw_cgroup_files[] = {
4375 .name = "memsw.usage_in_bytes",
4376 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4377 .read_u64 = mem_cgroup_read_u64,
4380 .name = "memsw.max_usage_in_bytes",
4381 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4382 .write = mem_cgroup_reset,
4383 .read_u64 = mem_cgroup_read_u64,
4386 .name = "memsw.limit_in_bytes",
4387 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4388 .write = mem_cgroup_write,
4389 .read_u64 = mem_cgroup_read_u64,
4392 .name = "memsw.failcnt",
4393 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4394 .write = mem_cgroup_reset,
4395 .read_u64 = mem_cgroup_read_u64,
4397 { }, /* terminate */
4400 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4402 struct mem_cgroup_per_node *pn;
4403 struct mem_cgroup_per_zone *mz;
4404 int zone, tmp = node;
4406 * This routine is called against possible nodes.
4407 * But it's BUG to call kmalloc() against offline node.
4409 * TODO: this routine can waste much memory for nodes which will
4410 * never be onlined. It's better to use memory hotplug callback
4413 if (!node_state(node, N_NORMAL_MEMORY))
4415 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4419 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4420 mz = &pn->zoneinfo[zone];
4421 lruvec_init(&mz->lruvec);
4422 mz->usage_in_excess = 0;
4423 mz->on_tree = false;
4426 memcg->nodeinfo[node] = pn;
4430 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4432 kfree(memcg->nodeinfo[node]);
4435 static struct mem_cgroup *mem_cgroup_alloc(void)
4437 struct mem_cgroup *memcg;
4440 size = sizeof(struct mem_cgroup);
4441 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4443 memcg = kzalloc(size, GFP_KERNEL);
4447 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4450 spin_lock_init(&memcg->pcp_counter_lock);
4459 * At destroying mem_cgroup, references from swap_cgroup can remain.
4460 * (scanning all at force_empty is too costly...)
4462 * Instead of clearing all references at force_empty, we remember
4463 * the number of reference from swap_cgroup and free mem_cgroup when
4464 * it goes down to 0.
4466 * Removal of cgroup itself succeeds regardless of refs from swap.
4469 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4473 mem_cgroup_remove_from_trees(memcg);
4476 free_mem_cgroup_per_zone_info(memcg, node);
4478 free_percpu(memcg->stat);
4480 disarm_static_keys(memcg);
4485 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4487 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4489 if (!memcg->memory.parent)
4491 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4493 EXPORT_SYMBOL(parent_mem_cgroup);
4495 static void __init mem_cgroup_soft_limit_tree_init(void)
4497 struct mem_cgroup_tree_per_node *rtpn;
4498 struct mem_cgroup_tree_per_zone *rtpz;
4499 int tmp, node, zone;
4501 for_each_node(node) {
4503 if (!node_state(node, N_NORMAL_MEMORY))
4505 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4508 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4510 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4511 rtpz = &rtpn->rb_tree_per_zone[zone];
4512 rtpz->rb_root = RB_ROOT;
4513 spin_lock_init(&rtpz->lock);
4518 static struct cgroup_subsys_state * __ref
4519 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4521 struct mem_cgroup *memcg;
4522 long error = -ENOMEM;
4525 memcg = mem_cgroup_alloc();
4527 return ERR_PTR(error);
4530 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4534 if (parent_css == NULL) {
4535 root_mem_cgroup = memcg;
4536 page_counter_init(&memcg->memory, NULL);
4537 memcg->soft_limit = PAGE_COUNTER_MAX;
4538 page_counter_init(&memcg->memsw, NULL);
4539 page_counter_init(&memcg->kmem, NULL);
4542 memcg->last_scanned_node = MAX_NUMNODES;
4543 INIT_LIST_HEAD(&memcg->oom_notify);
4544 memcg->move_charge_at_immigrate = 0;
4545 mutex_init(&memcg->thresholds_lock);
4546 spin_lock_init(&memcg->move_lock);
4547 vmpressure_init(&memcg->vmpressure);
4548 INIT_LIST_HEAD(&memcg->event_list);
4549 spin_lock_init(&memcg->event_list_lock);
4550 #ifdef CONFIG_MEMCG_KMEM
4551 memcg->kmemcg_id = -1;
4557 __mem_cgroup_free(memcg);
4558 return ERR_PTR(error);
4562 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4564 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4565 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4568 if (css->id > MEM_CGROUP_ID_MAX)
4574 mutex_lock(&memcg_create_mutex);
4576 memcg->use_hierarchy = parent->use_hierarchy;
4577 memcg->oom_kill_disable = parent->oom_kill_disable;
4578 memcg->swappiness = mem_cgroup_swappiness(parent);
4580 if (parent->use_hierarchy) {
4581 page_counter_init(&memcg->memory, &parent->memory);
4582 memcg->soft_limit = PAGE_COUNTER_MAX;
4583 page_counter_init(&memcg->memsw, &parent->memsw);
4584 page_counter_init(&memcg->kmem, &parent->kmem);
4587 * No need to take a reference to the parent because cgroup
4588 * core guarantees its existence.
4591 page_counter_init(&memcg->memory, NULL);
4592 memcg->soft_limit = PAGE_COUNTER_MAX;
4593 page_counter_init(&memcg->memsw, NULL);
4594 page_counter_init(&memcg->kmem, NULL);
4596 * Deeper hierachy with use_hierarchy == false doesn't make
4597 * much sense so let cgroup subsystem know about this
4598 * unfortunate state in our controller.
4600 if (parent != root_mem_cgroup)
4601 memory_cgrp_subsys.broken_hierarchy = true;
4603 mutex_unlock(&memcg_create_mutex);
4605 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4610 * Make sure the memcg is initialized: mem_cgroup_iter()
4611 * orders reading memcg->initialized against its callers
4612 * reading the memcg members.
4614 smp_store_release(&memcg->initialized, 1);
4619 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4621 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4622 struct mem_cgroup_event *event, *tmp;
4625 * Unregister events and notify userspace.
4626 * Notify userspace about cgroup removing only after rmdir of cgroup
4627 * directory to avoid race between userspace and kernelspace.
4629 spin_lock(&memcg->event_list_lock);
4630 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4631 list_del_init(&event->list);
4632 schedule_work(&event->remove);
4634 spin_unlock(&memcg->event_list_lock);
4636 vmpressure_cleanup(&memcg->vmpressure);
4639 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4641 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4643 memcg_destroy_kmem(memcg);
4644 __mem_cgroup_free(memcg);
4648 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4649 * @css: the target css
4651 * Reset the states of the mem_cgroup associated with @css. This is
4652 * invoked when the userland requests disabling on the default hierarchy
4653 * but the memcg is pinned through dependency. The memcg should stop
4654 * applying policies and should revert to the vanilla state as it may be
4655 * made visible again.
4657 * The current implementation only resets the essential configurations.
4658 * This needs to be expanded to cover all the visible parts.
4660 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4662 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4664 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4665 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4666 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4667 memcg->soft_limit = PAGE_COUNTER_MAX;
4671 /* Handlers for move charge at task migration. */
4672 static int mem_cgroup_do_precharge(unsigned long count)
4676 /* Try a single bulk charge without reclaim first */
4677 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4679 mc.precharge += count;
4682 if (ret == -EINTR) {
4683 cancel_charge(root_mem_cgroup, count);
4687 /* Try charges one by one with reclaim */
4689 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4691 * In case of failure, any residual charges against
4692 * mc.to will be dropped by mem_cgroup_clear_mc()
4693 * later on. However, cancel any charges that are
4694 * bypassed to root right away or they'll be lost.
4697 cancel_charge(root_mem_cgroup, 1);
4707 * get_mctgt_type - get target type of moving charge
4708 * @vma: the vma the pte to be checked belongs
4709 * @addr: the address corresponding to the pte to be checked
4710 * @ptent: the pte to be checked
4711 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4714 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4715 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4716 * move charge. if @target is not NULL, the page is stored in target->page
4717 * with extra refcnt got(Callers should handle it).
4718 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4719 * target for charge migration. if @target is not NULL, the entry is stored
4722 * Called with pte lock held.
4729 enum mc_target_type {
4735 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4736 unsigned long addr, pte_t ptent)
4738 struct page *page = vm_normal_page(vma, addr, ptent);
4740 if (!page || !page_mapped(page))
4742 if (PageAnon(page)) {
4743 /* we don't move shared anon */
4746 } else if (!move_file())
4747 /* we ignore mapcount for file pages */
4749 if (!get_page_unless_zero(page))
4756 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4757 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4759 struct page *page = NULL;
4760 swp_entry_t ent = pte_to_swp_entry(ptent);
4762 if (!move_anon() || non_swap_entry(ent))
4765 * Because lookup_swap_cache() updates some statistics counter,
4766 * we call find_get_page() with swapper_space directly.
4768 page = find_get_page(swap_address_space(ent), ent.val);
4769 if (do_swap_account)
4770 entry->val = ent.val;
4775 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4776 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4782 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4783 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4785 struct page *page = NULL;
4786 struct address_space *mapping;
4789 if (!vma->vm_file) /* anonymous vma */
4794 mapping = vma->vm_file->f_mapping;
4795 pgoff = linear_page_index(vma, addr);
4797 /* page is moved even if it's not RSS of this task(page-faulted). */
4799 /* shmem/tmpfs may report page out on swap: account for that too. */
4800 if (shmem_mapping(mapping)) {
4801 page = find_get_entry(mapping, pgoff);
4802 if (radix_tree_exceptional_entry(page)) {
4803 swp_entry_t swp = radix_to_swp_entry(page);
4804 if (do_swap_account)
4806 page = find_get_page(swap_address_space(swp), swp.val);
4809 page = find_get_page(mapping, pgoff);
4811 page = find_get_page(mapping, pgoff);
4816 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4817 unsigned long addr, pte_t ptent, union mc_target *target)
4819 struct page *page = NULL;
4820 enum mc_target_type ret = MC_TARGET_NONE;
4821 swp_entry_t ent = { .val = 0 };
4823 if (pte_present(ptent))
4824 page = mc_handle_present_pte(vma, addr, ptent);
4825 else if (is_swap_pte(ptent))
4826 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4827 else if (pte_none(ptent))
4828 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4830 if (!page && !ent.val)
4834 * Do only loose check w/o serialization.
4835 * mem_cgroup_move_account() checks the page is valid or
4836 * not under LRU exclusion.
4838 if (page->mem_cgroup == mc.from) {
4839 ret = MC_TARGET_PAGE;
4841 target->page = page;
4843 if (!ret || !target)
4846 /* There is a swap entry and a page doesn't exist or isn't charged */
4847 if (ent.val && !ret &&
4848 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4849 ret = MC_TARGET_SWAP;
4856 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4858 * We don't consider swapping or file mapped pages because THP does not
4859 * support them for now.
4860 * Caller should make sure that pmd_trans_huge(pmd) is true.
4862 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4863 unsigned long addr, pmd_t pmd, union mc_target *target)
4865 struct page *page = NULL;
4866 enum mc_target_type ret = MC_TARGET_NONE;
4868 page = pmd_page(pmd);
4869 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4872 if (page->mem_cgroup == mc.from) {
4873 ret = MC_TARGET_PAGE;
4876 target->page = page;
4882 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4883 unsigned long addr, pmd_t pmd, union mc_target *target)
4885 return MC_TARGET_NONE;
4889 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4890 unsigned long addr, unsigned long end,
4891 struct mm_walk *walk)
4893 struct vm_area_struct *vma = walk->private;
4897 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4898 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4899 mc.precharge += HPAGE_PMD_NR;
4904 if (pmd_trans_unstable(pmd))
4906 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4907 for (; addr != end; pte++, addr += PAGE_SIZE)
4908 if (get_mctgt_type(vma, addr, *pte, NULL))
4909 mc.precharge++; /* increment precharge temporarily */
4910 pte_unmap_unlock(pte - 1, ptl);
4916 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4918 unsigned long precharge;
4919 struct vm_area_struct *vma;
4921 down_read(&mm->mmap_sem);
4922 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4923 struct mm_walk mem_cgroup_count_precharge_walk = {
4924 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4928 if (is_vm_hugetlb_page(vma))
4930 walk_page_range(vma->vm_start, vma->vm_end,
4931 &mem_cgroup_count_precharge_walk);
4933 up_read(&mm->mmap_sem);
4935 precharge = mc.precharge;
4941 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4943 unsigned long precharge = mem_cgroup_count_precharge(mm);
4945 VM_BUG_ON(mc.moving_task);
4946 mc.moving_task = current;
4947 return mem_cgroup_do_precharge(precharge);
4950 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4951 static void __mem_cgroup_clear_mc(void)
4953 struct mem_cgroup *from = mc.from;
4954 struct mem_cgroup *to = mc.to;
4956 /* we must uncharge all the leftover precharges from mc.to */
4958 cancel_charge(mc.to, mc.precharge);
4962 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4963 * we must uncharge here.
4965 if (mc.moved_charge) {
4966 cancel_charge(mc.from, mc.moved_charge);
4967 mc.moved_charge = 0;
4969 /* we must fixup refcnts and charges */
4970 if (mc.moved_swap) {
4971 /* uncharge swap account from the old cgroup */
4972 if (!mem_cgroup_is_root(mc.from))
4973 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4976 * we charged both to->memory and to->memsw, so we
4977 * should uncharge to->memory.
4979 if (!mem_cgroup_is_root(mc.to))
4980 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4982 css_put_many(&mc.from->css, mc.moved_swap);
4984 /* we've already done css_get(mc.to) */
4987 memcg_oom_recover(from);
4988 memcg_oom_recover(to);
4989 wake_up_all(&mc.waitq);
4992 static void mem_cgroup_clear_mc(void)
4995 * we must clear moving_task before waking up waiters at the end of
4998 mc.moving_task = NULL;
4999 __mem_cgroup_clear_mc();
5000 spin_lock(&mc.lock);
5003 spin_unlock(&mc.lock);
5006 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5007 struct cgroup_taskset *tset)
5009 struct task_struct *p = cgroup_taskset_first(tset);
5011 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5012 unsigned long move_charge_at_immigrate;
5015 * We are now commited to this value whatever it is. Changes in this
5016 * tunable will only affect upcoming migrations, not the current one.
5017 * So we need to save it, and keep it going.
5019 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
5020 if (move_charge_at_immigrate) {
5021 struct mm_struct *mm;
5022 struct mem_cgroup *from = mem_cgroup_from_task(p);
5024 VM_BUG_ON(from == memcg);
5026 mm = get_task_mm(p);
5029 /* We move charges only when we move a owner of the mm */
5030 if (mm->owner == p) {
5033 VM_BUG_ON(mc.precharge);
5034 VM_BUG_ON(mc.moved_charge);
5035 VM_BUG_ON(mc.moved_swap);
5037 spin_lock(&mc.lock);
5040 mc.immigrate_flags = move_charge_at_immigrate;
5041 spin_unlock(&mc.lock);
5042 /* We set mc.moving_task later */
5044 ret = mem_cgroup_precharge_mc(mm);
5046 mem_cgroup_clear_mc();
5053 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5054 struct cgroup_taskset *tset)
5057 mem_cgroup_clear_mc();
5060 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5061 unsigned long addr, unsigned long end,
5062 struct mm_walk *walk)
5065 struct vm_area_struct *vma = walk->private;
5068 enum mc_target_type target_type;
5069 union mc_target target;
5073 * We don't take compound_lock() here but no race with splitting thp
5075 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5076 * under splitting, which means there's no concurrent thp split,
5077 * - if another thread runs into split_huge_page() just after we
5078 * entered this if-block, the thread must wait for page table lock
5079 * to be unlocked in __split_huge_page_splitting(), where the main
5080 * part of thp split is not executed yet.
5082 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5083 if (mc.precharge < HPAGE_PMD_NR) {
5087 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5088 if (target_type == MC_TARGET_PAGE) {
5090 if (!isolate_lru_page(page)) {
5091 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5093 mc.precharge -= HPAGE_PMD_NR;
5094 mc.moved_charge += HPAGE_PMD_NR;
5096 putback_lru_page(page);
5104 if (pmd_trans_unstable(pmd))
5107 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5108 for (; addr != end; addr += PAGE_SIZE) {
5109 pte_t ptent = *(pte++);
5115 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5116 case MC_TARGET_PAGE:
5118 if (isolate_lru_page(page))
5120 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5122 /* we uncharge from mc.from later. */
5125 putback_lru_page(page);
5126 put: /* get_mctgt_type() gets the page */
5129 case MC_TARGET_SWAP:
5131 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5133 /* we fixup refcnts and charges later. */
5141 pte_unmap_unlock(pte - 1, ptl);
5146 * We have consumed all precharges we got in can_attach().
5147 * We try charge one by one, but don't do any additional
5148 * charges to mc.to if we have failed in charge once in attach()
5151 ret = mem_cgroup_do_precharge(1);
5159 static void mem_cgroup_move_charge(struct mm_struct *mm)
5161 struct vm_area_struct *vma;
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();
5184 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5186 struct mm_walk mem_cgroup_move_charge_walk = {
5187 .pmd_entry = mem_cgroup_move_charge_pte_range,
5191 if (is_vm_hugetlb_page(vma))
5193 ret = walk_page_range(vma->vm_start, vma->vm_end,
5194 &mem_cgroup_move_charge_walk);
5197 * means we have consumed all precharges and failed in
5198 * doing additional charge. Just abandon here.
5202 up_read(&mm->mmap_sem);
5203 atomic_dec(&mc.from->moving_account);
5206 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5207 struct cgroup_taskset *tset)
5209 struct task_struct *p = cgroup_taskset_first(tset);
5210 struct mm_struct *mm = get_task_mm(p);
5214 mem_cgroup_move_charge(mm);
5218 mem_cgroup_clear_mc();
5220 #else /* !CONFIG_MMU */
5221 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5222 struct cgroup_taskset *tset)
5226 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5227 struct cgroup_taskset *tset)
5230 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5231 struct cgroup_taskset *tset)
5237 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5238 * to verify whether we're attached to the default hierarchy on each mount
5241 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5244 * use_hierarchy is forced on the default hierarchy. cgroup core
5245 * guarantees that @root doesn't have any children, so turning it
5246 * on for the root memcg is enough.
5248 if (cgroup_on_dfl(root_css->cgroup))
5249 mem_cgroup_from_css(root_css)->use_hierarchy = true;
5252 struct cgroup_subsys memory_cgrp_subsys = {
5253 .css_alloc = mem_cgroup_css_alloc,
5254 .css_online = mem_cgroup_css_online,
5255 .css_offline = mem_cgroup_css_offline,
5256 .css_free = mem_cgroup_css_free,
5257 .css_reset = mem_cgroup_css_reset,
5258 .can_attach = mem_cgroup_can_attach,
5259 .cancel_attach = mem_cgroup_cancel_attach,
5260 .attach = mem_cgroup_move_task,
5261 .bind = mem_cgroup_bind,
5262 .legacy_cftypes = mem_cgroup_files,
5266 #ifdef CONFIG_MEMCG_SWAP
5267 static int __init enable_swap_account(char *s)
5269 if (!strcmp(s, "1"))
5270 really_do_swap_account = 1;
5271 else if (!strcmp(s, "0"))
5272 really_do_swap_account = 0;
5275 __setup("swapaccount=", enable_swap_account);
5277 static void __init memsw_file_init(void)
5279 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5280 memsw_cgroup_files));
5283 static void __init enable_swap_cgroup(void)
5285 if (!mem_cgroup_disabled() && really_do_swap_account) {
5286 do_swap_account = 1;
5292 static void __init enable_swap_cgroup(void)
5297 #ifdef CONFIG_MEMCG_SWAP
5299 * mem_cgroup_swapout - transfer a memsw charge to swap
5300 * @page: page whose memsw charge to transfer
5301 * @entry: swap entry to move the charge to
5303 * Transfer the memsw charge of @page to @entry.
5305 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5307 struct mem_cgroup *memcg;
5308 unsigned short oldid;
5310 VM_BUG_ON_PAGE(PageLRU(page), page);
5311 VM_BUG_ON_PAGE(page_count(page), page);
5313 if (!do_swap_account)
5316 memcg = page->mem_cgroup;
5318 /* Readahead page, never charged */
5322 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5323 VM_BUG_ON_PAGE(oldid, page);
5324 mem_cgroup_swap_statistics(memcg, true);
5326 page->mem_cgroup = NULL;
5328 if (!mem_cgroup_is_root(memcg))
5329 page_counter_uncharge(&memcg->memory, 1);
5331 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5332 VM_BUG_ON(!irqs_disabled());
5334 mem_cgroup_charge_statistics(memcg, page, -1);
5335 memcg_check_events(memcg, page);
5339 * mem_cgroup_uncharge_swap - uncharge a swap entry
5340 * @entry: swap entry to uncharge
5342 * Drop the memsw charge associated with @entry.
5344 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5346 struct mem_cgroup *memcg;
5349 if (!do_swap_account)
5352 id = swap_cgroup_record(entry, 0);
5354 memcg = mem_cgroup_lookup(id);
5356 if (!mem_cgroup_is_root(memcg))
5357 page_counter_uncharge(&memcg->memsw, 1);
5358 mem_cgroup_swap_statistics(memcg, false);
5359 css_put(&memcg->css);
5366 * mem_cgroup_try_charge - try charging a page
5367 * @page: page to charge
5368 * @mm: mm context of the victim
5369 * @gfp_mask: reclaim mode
5370 * @memcgp: charged memcg return
5372 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5373 * pages according to @gfp_mask if necessary.
5375 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5376 * Otherwise, an error code is returned.
5378 * After page->mapping has been set up, the caller must finalize the
5379 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5380 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5382 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5383 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5385 struct mem_cgroup *memcg = NULL;
5386 unsigned int nr_pages = 1;
5389 if (mem_cgroup_disabled())
5392 if (PageSwapCache(page)) {
5394 * Every swap fault against a single page tries to charge the
5395 * page, bail as early as possible. shmem_unuse() encounters
5396 * already charged pages, too. The USED bit is protected by
5397 * the page lock, which serializes swap cache removal, which
5398 * in turn serializes uncharging.
5400 if (page->mem_cgroup)
5404 if (PageTransHuge(page)) {
5405 nr_pages <<= compound_order(page);
5406 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5409 if (do_swap_account && PageSwapCache(page))
5410 memcg = try_get_mem_cgroup_from_page(page);
5412 memcg = get_mem_cgroup_from_mm(mm);
5414 ret = try_charge(memcg, gfp_mask, nr_pages);
5416 css_put(&memcg->css);
5418 if (ret == -EINTR) {
5419 memcg = root_mem_cgroup;
5428 * mem_cgroup_commit_charge - commit a page charge
5429 * @page: page to charge
5430 * @memcg: memcg to charge the page to
5431 * @lrucare: page might be on LRU already
5433 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5434 * after page->mapping has been set up. This must happen atomically
5435 * as part of the page instantiation, i.e. under the page table lock
5436 * for anonymous pages, under the page lock for page and swap cache.
5438 * In addition, the page must not be on the LRU during the commit, to
5439 * prevent racing with task migration. If it might be, use @lrucare.
5441 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5443 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5446 unsigned int nr_pages = 1;
5448 VM_BUG_ON_PAGE(!page->mapping, page);
5449 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5451 if (mem_cgroup_disabled())
5454 * Swap faults will attempt to charge the same page multiple
5455 * times. But reuse_swap_page() might have removed the page
5456 * from swapcache already, so we can't check PageSwapCache().
5461 commit_charge(page, memcg, lrucare);
5463 if (PageTransHuge(page)) {
5464 nr_pages <<= compound_order(page);
5465 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5468 local_irq_disable();
5469 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5470 memcg_check_events(memcg, page);
5473 if (do_swap_account && PageSwapCache(page)) {
5474 swp_entry_t entry = { .val = page_private(page) };
5476 * The swap entry might not get freed for a long time,
5477 * let's not wait for it. The page already received a
5478 * memory+swap charge, drop the swap entry duplicate.
5480 mem_cgroup_uncharge_swap(entry);
5485 * mem_cgroup_cancel_charge - cancel a page charge
5486 * @page: page to charge
5487 * @memcg: memcg to charge the page to
5489 * Cancel a charge transaction started by mem_cgroup_try_charge().
5491 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5493 unsigned int nr_pages = 1;
5495 if (mem_cgroup_disabled())
5498 * Swap faults will attempt to charge the same page multiple
5499 * times. But reuse_swap_page() might have removed the page
5500 * from swapcache already, so we can't check PageSwapCache().
5505 if (PageTransHuge(page)) {
5506 nr_pages <<= compound_order(page);
5507 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5510 cancel_charge(memcg, nr_pages);
5513 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5514 unsigned long nr_anon, unsigned long nr_file,
5515 unsigned long nr_huge, struct page *dummy_page)
5517 unsigned long nr_pages = nr_anon + nr_file;
5518 unsigned long flags;
5520 if (!mem_cgroup_is_root(memcg)) {
5521 page_counter_uncharge(&memcg->memory, nr_pages);
5522 if (do_swap_account)
5523 page_counter_uncharge(&memcg->memsw, nr_pages);
5524 memcg_oom_recover(memcg);
5527 local_irq_save(flags);
5528 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5529 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5530 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5531 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5532 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5533 memcg_check_events(memcg, dummy_page);
5534 local_irq_restore(flags);
5536 if (!mem_cgroup_is_root(memcg))
5537 css_put_many(&memcg->css, nr_pages);
5540 static void uncharge_list(struct list_head *page_list)
5542 struct mem_cgroup *memcg = NULL;
5543 unsigned long nr_anon = 0;
5544 unsigned long nr_file = 0;
5545 unsigned long nr_huge = 0;
5546 unsigned long pgpgout = 0;
5547 struct list_head *next;
5550 next = page_list->next;
5552 unsigned int nr_pages = 1;
5554 page = list_entry(next, struct page, lru);
5555 next = page->lru.next;
5557 VM_BUG_ON_PAGE(PageLRU(page), page);
5558 VM_BUG_ON_PAGE(page_count(page), page);
5560 if (!page->mem_cgroup)
5564 * Nobody should be changing or seriously looking at
5565 * page->mem_cgroup at this point, we have fully
5566 * exclusive access to the page.
5569 if (memcg != page->mem_cgroup) {
5571 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5573 pgpgout = nr_anon = nr_file = nr_huge = 0;
5575 memcg = page->mem_cgroup;
5578 if (PageTransHuge(page)) {
5579 nr_pages <<= compound_order(page);
5580 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5581 nr_huge += nr_pages;
5585 nr_anon += nr_pages;
5587 nr_file += nr_pages;
5589 page->mem_cgroup = NULL;
5592 } while (next != page_list);
5595 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5600 * mem_cgroup_uncharge - uncharge a page
5601 * @page: page to uncharge
5603 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5604 * mem_cgroup_commit_charge().
5606 void mem_cgroup_uncharge(struct page *page)
5608 if (mem_cgroup_disabled())
5611 /* Don't touch page->lru of any random page, pre-check: */
5612 if (!page->mem_cgroup)
5615 INIT_LIST_HEAD(&page->lru);
5616 uncharge_list(&page->lru);
5620 * mem_cgroup_uncharge_list - uncharge a list of page
5621 * @page_list: list of pages to uncharge
5623 * Uncharge a list of pages previously charged with
5624 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5626 void mem_cgroup_uncharge_list(struct list_head *page_list)
5628 if (mem_cgroup_disabled())
5631 if (!list_empty(page_list))
5632 uncharge_list(page_list);
5636 * mem_cgroup_migrate - migrate a charge to another page
5637 * @oldpage: currently charged page
5638 * @newpage: page to transfer the charge to
5639 * @lrucare: either or both pages might be on the LRU already
5641 * Migrate the charge from @oldpage to @newpage.
5643 * Both pages must be locked, @newpage->mapping must be set up.
5645 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5648 struct mem_cgroup *memcg;
5651 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5652 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5653 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5654 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5655 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5656 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5659 if (mem_cgroup_disabled())
5662 /* Page cache replacement: new page already charged? */
5663 if (newpage->mem_cgroup)
5667 * Swapcache readahead pages can get migrated before being
5668 * charged, and migration from compaction can happen to an
5669 * uncharged page when the PFN walker finds a page that
5670 * reclaim just put back on the LRU but has not released yet.
5672 memcg = oldpage->mem_cgroup;
5677 lock_page_lru(oldpage, &isolated);
5679 oldpage->mem_cgroup = NULL;
5682 unlock_page_lru(oldpage, isolated);
5684 commit_charge(newpage, memcg, lrucare);
5688 * subsys_initcall() for memory controller.
5690 * Some parts like hotcpu_notifier() have to be initialized from this context
5691 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5692 * everything that doesn't depend on a specific mem_cgroup structure should
5693 * be initialized from here.
5695 static int __init mem_cgroup_init(void)
5697 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5698 enable_swap_cgroup();
5699 mem_cgroup_soft_limit_tree_init();
5703 subsys_initcall(mem_cgroup_init);