1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct mem_cgroup_reclaim_iter {
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup *last_visited;
154 /* scan generation, increased every round-trip */
155 unsigned int generation;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone {
162 struct lruvec lruvec;
163 unsigned long lru_size[NR_LRU_LISTS];
165 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
167 struct rb_node tree_node; /* RB tree node */
168 unsigned long long usage_in_excess;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup *memcg; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node {
176 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone {
185 struct rb_root rb_root;
189 struct mem_cgroup_tree_per_node {
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
193 struct mem_cgroup_tree {
194 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
199 struct mem_cgroup_threshold {
200 struct eventfd_ctx *eventfd;
205 struct mem_cgroup_threshold_ary {
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries[0];
214 struct mem_cgroup_thresholds {
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary *primary;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary *spare;
226 struct mem_cgroup_eventfd_list {
227 struct list_head list;
228 struct eventfd_ctx *eventfd;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event {
236 * memcg which the event belongs to.
238 struct mem_cgroup *memcg;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx *eventfd;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event)(struct mem_cgroup *memcg,
253 struct eventfd_ctx *eventfd, const char *args);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event)(struct mem_cgroup *memcg,
260 struct eventfd_ctx *eventfd);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t *wqh;
268 struct work_struct remove;
271 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
272 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css;
288 * the counter to account for memory usage
290 struct res_counter res;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure;
295 /* css_online() has been completed */
299 * the counter to account for mem+swap usage.
301 struct res_counter memsw;
304 * the counter to account for kernel memory usage.
306 struct res_counter kmem;
308 * Should the accounting and control be hierarchical, per subtree?
311 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
315 atomic_t oom_wakeups;
318 /* OOM-Killer disable */
319 int oom_kill_disable;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds;
330 /* For oom notifier event fd */
331 struct list_head oom_notify;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock;
347 struct mem_cgroup_stat_cpu __percpu *stat;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base;
353 spinlock_t pcp_counter_lock;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list, but per-memcg;
361 * protected by memcg_slab_mutex */
362 struct list_head memcg_slab_caches;
363 /* Index in the kmem_cache->memcg_params->memcg_caches array */
367 int last_scanned_node;
369 nodemask_t scan_nodes;
370 atomic_t numainfo_events;
371 atomic_t numainfo_updating;
374 /* List of events which userspace want to receive */
375 struct list_head event_list;
376 spinlock_t event_list_lock;
378 struct mem_cgroup_per_node *nodeinfo[0];
379 /* WARNING: nodeinfo must be the last member here */
382 /* internal only representation about the status of kmem accounting. */
384 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
385 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
388 #ifdef CONFIG_MEMCG_KMEM
389 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
391 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
394 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
396 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
399 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
402 * Our caller must use css_get() first, because memcg_uncharge_kmem()
403 * will call css_put() if it sees the memcg is dead.
406 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
407 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
410 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
412 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
413 &memcg->kmem_account_flags);
417 /* Stuffs for move charges at task migration. */
419 * Types of charges to be moved. "move_charge_at_immitgrate" and
420 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
423 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
424 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
428 /* "mc" and its members are protected by cgroup_mutex */
429 static struct move_charge_struct {
430 spinlock_t lock; /* for from, to */
431 struct mem_cgroup *from;
432 struct mem_cgroup *to;
433 unsigned long immigrate_flags;
434 unsigned long precharge;
435 unsigned long moved_charge;
436 unsigned long moved_swap;
437 struct task_struct *moving_task; /* a task moving charges */
438 wait_queue_head_t waitq; /* a waitq for other context */
440 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
441 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
444 static bool move_anon(void)
446 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
449 static bool move_file(void)
451 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
455 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
456 * limit reclaim to prevent infinite loops, if they ever occur.
458 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
459 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
462 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
463 MEM_CGROUP_CHARGE_TYPE_ANON,
464 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
465 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
469 /* for encoding cft->private value on file */
477 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
478 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
479 #define MEMFILE_ATTR(val) ((val) & 0xffff)
480 /* Used for OOM nofiier */
481 #define OOM_CONTROL (0)
484 * The memcg_create_mutex will be held whenever a new cgroup is created.
485 * As a consequence, any change that needs to protect against new child cgroups
486 * appearing has to hold it as well.
488 static DEFINE_MUTEX(memcg_create_mutex);
490 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
492 return s ? container_of(s, struct mem_cgroup, css) : NULL;
495 /* Some nice accessors for the vmpressure. */
496 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
499 memcg = root_mem_cgroup;
500 return &memcg->vmpressure;
503 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
505 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
508 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
510 return (memcg == root_mem_cgroup);
514 * We restrict the id in the range of [1, 65535], so it can fit into
517 #define MEM_CGROUP_ID_MAX USHRT_MAX
519 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
521 return memcg->css.id;
524 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
526 struct cgroup_subsys_state *css;
528 css = css_from_id(id, &memory_cgrp_subsys);
529 return mem_cgroup_from_css(css);
532 /* Writing them here to avoid exposing memcg's inner layout */
533 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
535 void sock_update_memcg(struct sock *sk)
537 if (mem_cgroup_sockets_enabled) {
538 struct mem_cgroup *memcg;
539 struct cg_proto *cg_proto;
541 BUG_ON(!sk->sk_prot->proto_cgroup);
543 /* Socket cloning can throw us here with sk_cgrp already
544 * filled. It won't however, necessarily happen from
545 * process context. So the test for root memcg given
546 * the current task's memcg won't help us in this case.
548 * Respecting the original socket's memcg is a better
549 * decision in this case.
552 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
553 css_get(&sk->sk_cgrp->memcg->css);
558 memcg = mem_cgroup_from_task(current);
559 cg_proto = sk->sk_prot->proto_cgroup(memcg);
560 if (!mem_cgroup_is_root(memcg) &&
561 memcg_proto_active(cg_proto) &&
562 css_tryget_online(&memcg->css)) {
563 sk->sk_cgrp = cg_proto;
568 EXPORT_SYMBOL(sock_update_memcg);
570 void sock_release_memcg(struct sock *sk)
572 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
573 struct mem_cgroup *memcg;
574 WARN_ON(!sk->sk_cgrp->memcg);
575 memcg = sk->sk_cgrp->memcg;
576 css_put(&sk->sk_cgrp->memcg->css);
580 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
582 if (!memcg || mem_cgroup_is_root(memcg))
585 return &memcg->tcp_mem;
587 EXPORT_SYMBOL(tcp_proto_cgroup);
589 static void disarm_sock_keys(struct mem_cgroup *memcg)
591 if (!memcg_proto_activated(&memcg->tcp_mem))
593 static_key_slow_dec(&memcg_socket_limit_enabled);
596 static void disarm_sock_keys(struct mem_cgroup *memcg)
601 #ifdef CONFIG_MEMCG_KMEM
603 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
604 * The main reason for not using cgroup id for this:
605 * this works better in sparse environments, where we have a lot of memcgs,
606 * but only a few kmem-limited. Or also, if we have, for instance, 200
607 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
608 * 200 entry array for that.
610 * The current size of the caches array is stored in
611 * memcg_limited_groups_array_size. It will double each time we have to
614 static DEFINE_IDA(kmem_limited_groups);
615 int memcg_limited_groups_array_size;
618 * MIN_SIZE is different than 1, because we would like to avoid going through
619 * the alloc/free process all the time. In a small machine, 4 kmem-limited
620 * cgroups is a reasonable guess. In the future, it could be a parameter or
621 * tunable, but that is strictly not necessary.
623 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
624 * this constant directly from cgroup, but it is understandable that this is
625 * better kept as an internal representation in cgroup.c. In any case, the
626 * cgrp_id space is not getting any smaller, and we don't have to necessarily
627 * increase ours as well if it increases.
629 #define MEMCG_CACHES_MIN_SIZE 4
630 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
633 * A lot of the calls to the cache allocation functions are expected to be
634 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
635 * conditional to this static branch, we'll have to allow modules that does
636 * kmem_cache_alloc and the such to see this symbol as well
638 struct static_key memcg_kmem_enabled_key;
639 EXPORT_SYMBOL(memcg_kmem_enabled_key);
641 static void memcg_free_cache_id(int id);
643 static void disarm_kmem_keys(struct mem_cgroup *memcg)
645 if (memcg_kmem_is_active(memcg)) {
646 static_key_slow_dec(&memcg_kmem_enabled_key);
647 memcg_free_cache_id(memcg->kmemcg_id);
650 * This check can't live in kmem destruction function,
651 * since the charges will outlive the cgroup
653 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
656 static void disarm_kmem_keys(struct mem_cgroup *memcg)
659 #endif /* CONFIG_MEMCG_KMEM */
661 static void disarm_static_keys(struct mem_cgroup *memcg)
663 disarm_sock_keys(memcg);
664 disarm_kmem_keys(memcg);
667 static void drain_all_stock_async(struct mem_cgroup *memcg);
669 static struct mem_cgroup_per_zone *
670 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
672 int nid = zone_to_nid(zone);
673 int zid = zone_idx(zone);
675 return &memcg->nodeinfo[nid]->zoneinfo[zid];
678 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
683 static struct mem_cgroup_per_zone *
684 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
686 int nid = page_to_nid(page);
687 int zid = page_zonenum(page);
689 return &memcg->nodeinfo[nid]->zoneinfo[zid];
692 static struct mem_cgroup_tree_per_zone *
693 soft_limit_tree_node_zone(int nid, int zid)
695 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
698 static struct mem_cgroup_tree_per_zone *
699 soft_limit_tree_from_page(struct page *page)
701 int nid = page_to_nid(page);
702 int zid = page_zonenum(page);
704 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
707 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
708 struct mem_cgroup_tree_per_zone *mctz,
709 unsigned long long new_usage_in_excess)
711 struct rb_node **p = &mctz->rb_root.rb_node;
712 struct rb_node *parent = NULL;
713 struct mem_cgroup_per_zone *mz_node;
718 mz->usage_in_excess = new_usage_in_excess;
719 if (!mz->usage_in_excess)
723 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
725 if (mz->usage_in_excess < mz_node->usage_in_excess)
728 * We can't avoid mem cgroups that are over their soft
729 * limit by the same amount
731 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
734 rb_link_node(&mz->tree_node, parent, p);
735 rb_insert_color(&mz->tree_node, &mctz->rb_root);
739 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
740 struct mem_cgroup_tree_per_zone *mctz)
744 rb_erase(&mz->tree_node, &mctz->rb_root);
748 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
749 struct mem_cgroup_tree_per_zone *mctz)
753 spin_lock_irqsave(&mctz->lock, flags);
754 __mem_cgroup_remove_exceeded(mz, mctz);
755 spin_unlock_irqrestore(&mctz->lock, flags);
759 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
761 unsigned long long excess;
762 struct mem_cgroup_per_zone *mz;
763 struct mem_cgroup_tree_per_zone *mctz;
765 mctz = soft_limit_tree_from_page(page);
767 * Necessary to update all ancestors when hierarchy is used.
768 * because their event counter is not touched.
770 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
771 mz = mem_cgroup_page_zoneinfo(memcg, page);
772 excess = res_counter_soft_limit_excess(&memcg->res);
774 * We have to update the tree if mz is on RB-tree or
775 * mem is over its softlimit.
777 if (excess || mz->on_tree) {
780 spin_lock_irqsave(&mctz->lock, flags);
781 /* if on-tree, remove it */
783 __mem_cgroup_remove_exceeded(mz, mctz);
785 * Insert again. mz->usage_in_excess will be updated.
786 * If excess is 0, no tree ops.
788 __mem_cgroup_insert_exceeded(mz, mctz, excess);
789 spin_unlock_irqrestore(&mctz->lock, flags);
794 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
796 struct mem_cgroup_tree_per_zone *mctz;
797 struct mem_cgroup_per_zone *mz;
801 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
802 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
803 mctz = soft_limit_tree_node_zone(nid, zid);
804 mem_cgroup_remove_exceeded(mz, mctz);
809 static struct mem_cgroup_per_zone *
810 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
812 struct rb_node *rightmost = NULL;
813 struct mem_cgroup_per_zone *mz;
817 rightmost = rb_last(&mctz->rb_root);
819 goto done; /* Nothing to reclaim from */
821 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
823 * Remove the node now but someone else can add it back,
824 * we will to add it back at the end of reclaim to its correct
825 * position in the tree.
827 __mem_cgroup_remove_exceeded(mz, mctz);
828 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
829 !css_tryget_online(&mz->memcg->css))
835 static struct mem_cgroup_per_zone *
836 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
838 struct mem_cgroup_per_zone *mz;
840 spin_lock_irq(&mctz->lock);
841 mz = __mem_cgroup_largest_soft_limit_node(mctz);
842 spin_unlock_irq(&mctz->lock);
847 * Implementation Note: reading percpu statistics for memcg.
849 * Both of vmstat[] and percpu_counter has threshold and do periodic
850 * synchronization to implement "quick" read. There are trade-off between
851 * reading cost and precision of value. Then, we may have a chance to implement
852 * a periodic synchronizion of counter in memcg's counter.
854 * But this _read() function is used for user interface now. The user accounts
855 * memory usage by memory cgroup and he _always_ requires exact value because
856 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
857 * have to visit all online cpus and make sum. So, for now, unnecessary
858 * synchronization is not implemented. (just implemented for cpu hotplug)
860 * If there are kernel internal actions which can make use of some not-exact
861 * value, and reading all cpu value can be performance bottleneck in some
862 * common workload, threashold and synchonization as vmstat[] should be
865 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
866 enum mem_cgroup_stat_index idx)
872 for_each_online_cpu(cpu)
873 val += per_cpu(memcg->stat->count[idx], cpu);
874 #ifdef CONFIG_HOTPLUG_CPU
875 spin_lock(&memcg->pcp_counter_lock);
876 val += memcg->nocpu_base.count[idx];
877 spin_unlock(&memcg->pcp_counter_lock);
883 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
884 enum mem_cgroup_events_index idx)
886 unsigned long val = 0;
890 for_each_online_cpu(cpu)
891 val += per_cpu(memcg->stat->events[idx], cpu);
892 #ifdef CONFIG_HOTPLUG_CPU
893 spin_lock(&memcg->pcp_counter_lock);
894 val += memcg->nocpu_base.events[idx];
895 spin_unlock(&memcg->pcp_counter_lock);
901 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
906 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
907 * counted as CACHE even if it's on ANON LRU.
910 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
913 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
916 if (PageTransHuge(page))
917 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
920 /* pagein of a big page is an event. So, ignore page size */
922 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
924 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
925 nr_pages = -nr_pages; /* for event */
928 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
931 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
933 struct mem_cgroup_per_zone *mz;
935 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
936 return mz->lru_size[lru];
939 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
941 unsigned int lru_mask)
943 unsigned long nr = 0;
946 VM_BUG_ON((unsigned)nid >= nr_node_ids);
948 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
949 struct mem_cgroup_per_zone *mz;
953 if (!(BIT(lru) & lru_mask))
955 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
956 nr += mz->lru_size[lru];
962 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
963 unsigned int lru_mask)
965 unsigned long nr = 0;
968 for_each_node_state(nid, N_MEMORY)
969 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
973 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
974 enum mem_cgroup_events_target target)
976 unsigned long val, next;
978 val = __this_cpu_read(memcg->stat->nr_page_events);
979 next = __this_cpu_read(memcg->stat->targets[target]);
980 /* from time_after() in jiffies.h */
981 if ((long)next - (long)val < 0) {
983 case MEM_CGROUP_TARGET_THRESH:
984 next = val + THRESHOLDS_EVENTS_TARGET;
986 case MEM_CGROUP_TARGET_SOFTLIMIT:
987 next = val + SOFTLIMIT_EVENTS_TARGET;
989 case MEM_CGROUP_TARGET_NUMAINFO:
990 next = val + NUMAINFO_EVENTS_TARGET;
995 __this_cpu_write(memcg->stat->targets[target], next);
1002 * Check events in order.
1005 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1007 /* threshold event is triggered in finer grain than soft limit */
1008 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1009 MEM_CGROUP_TARGET_THRESH))) {
1011 bool do_numainfo __maybe_unused;
1013 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1014 MEM_CGROUP_TARGET_SOFTLIMIT);
1015 #if MAX_NUMNODES > 1
1016 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1017 MEM_CGROUP_TARGET_NUMAINFO);
1019 mem_cgroup_threshold(memcg);
1020 if (unlikely(do_softlimit))
1021 mem_cgroup_update_tree(memcg, page);
1022 #if MAX_NUMNODES > 1
1023 if (unlikely(do_numainfo))
1024 atomic_inc(&memcg->numainfo_events);
1029 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1032 * mm_update_next_owner() may clear mm->owner to NULL
1033 * if it races with swapoff, page migration, etc.
1034 * So this can be called with p == NULL.
1039 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1042 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1044 struct mem_cgroup *memcg = NULL;
1049 * Page cache insertions can happen withou an
1050 * actual mm context, e.g. during disk probing
1051 * on boot, loopback IO, acct() writes etc.
1054 memcg = root_mem_cgroup;
1056 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1057 if (unlikely(!memcg))
1058 memcg = root_mem_cgroup;
1060 } while (!css_tryget_online(&memcg->css));
1066 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1067 * ref. count) or NULL if the whole root's subtree has been visited.
1069 * helper function to be used by mem_cgroup_iter
1071 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1072 struct mem_cgroup *last_visited)
1074 struct cgroup_subsys_state *prev_css, *next_css;
1076 prev_css = last_visited ? &last_visited->css : NULL;
1078 next_css = css_next_descendant_pre(prev_css, &root->css);
1081 * Even if we found a group we have to make sure it is
1082 * alive. css && !memcg means that the groups should be
1083 * skipped and we should continue the tree walk.
1084 * last_visited css is safe to use because it is
1085 * protected by css_get and the tree walk is rcu safe.
1087 * We do not take a reference on the root of the tree walk
1088 * because we might race with the root removal when it would
1089 * be the only node in the iterated hierarchy and mem_cgroup_iter
1090 * would end up in an endless loop because it expects that at
1091 * least one valid node will be returned. Root cannot disappear
1092 * because caller of the iterator should hold it already so
1093 * skipping css reference should be safe.
1096 struct mem_cgroup *memcg = mem_cgroup_from_css(next_css);
1098 if (next_css == &root->css)
1101 if (css_tryget_online(next_css)) {
1103 * Make sure the memcg is initialized:
1104 * mem_cgroup_css_online() orders the the
1105 * initialization against setting the flag.
1107 if (smp_load_acquire(&memcg->initialized))
1112 prev_css = next_css;
1119 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1122 * When a group in the hierarchy below root is destroyed, the
1123 * hierarchy iterator can no longer be trusted since it might
1124 * have pointed to the destroyed group. Invalidate it.
1126 atomic_inc(&root->dead_count);
1129 static struct mem_cgroup *
1130 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1131 struct mem_cgroup *root,
1134 struct mem_cgroup *position = NULL;
1136 * A cgroup destruction happens in two stages: offlining and
1137 * release. They are separated by a RCU grace period.
1139 * If the iterator is valid, we may still race with an
1140 * offlining. The RCU lock ensures the object won't be
1141 * released, tryget will fail if we lost the race.
1143 *sequence = atomic_read(&root->dead_count);
1144 if (iter->last_dead_count == *sequence) {
1146 position = iter->last_visited;
1149 * We cannot take a reference to root because we might race
1150 * with root removal and returning NULL would end up in
1151 * an endless loop on the iterator user level when root
1152 * would be returned all the time.
1154 if (position && position != root &&
1155 !css_tryget_online(&position->css))
1161 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1162 struct mem_cgroup *last_visited,
1163 struct mem_cgroup *new_position,
1164 struct mem_cgroup *root,
1167 /* root reference counting symmetric to mem_cgroup_iter_load */
1168 if (last_visited && last_visited != root)
1169 css_put(&last_visited->css);
1171 * We store the sequence count from the time @last_visited was
1172 * loaded successfully instead of rereading it here so that we
1173 * don't lose destruction events in between. We could have
1174 * raced with the destruction of @new_position after all.
1176 iter->last_visited = new_position;
1178 iter->last_dead_count = sequence;
1182 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1183 * @root: hierarchy root
1184 * @prev: previously returned memcg, NULL on first invocation
1185 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1187 * Returns references to children of the hierarchy below @root, or
1188 * @root itself, or %NULL after a full round-trip.
1190 * Caller must pass the return value in @prev on subsequent
1191 * invocations for reference counting, or use mem_cgroup_iter_break()
1192 * to cancel a hierarchy walk before the round-trip is complete.
1194 * Reclaimers can specify a zone and a priority level in @reclaim to
1195 * divide up the memcgs in the hierarchy among all concurrent
1196 * reclaimers operating on the same zone and priority.
1198 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1199 struct mem_cgroup *prev,
1200 struct mem_cgroup_reclaim_cookie *reclaim)
1202 struct mem_cgroup *memcg = NULL;
1203 struct mem_cgroup *last_visited = NULL;
1205 if (mem_cgroup_disabled())
1209 root = root_mem_cgroup;
1211 if (prev && !reclaim)
1212 last_visited = prev;
1214 if (!root->use_hierarchy && root != root_mem_cgroup) {
1222 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1223 int uninitialized_var(seq);
1226 struct mem_cgroup_per_zone *mz;
1228 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1229 iter = &mz->reclaim_iter[reclaim->priority];
1230 if (prev && reclaim->generation != iter->generation) {
1231 iter->last_visited = NULL;
1235 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1238 memcg = __mem_cgroup_iter_next(root, last_visited);
1241 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1246 else if (!prev && memcg)
1247 reclaim->generation = iter->generation;
1256 if (prev && prev != root)
1257 css_put(&prev->css);
1263 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1264 * @root: hierarchy root
1265 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1267 void mem_cgroup_iter_break(struct mem_cgroup *root,
1268 struct mem_cgroup *prev)
1271 root = root_mem_cgroup;
1272 if (prev && prev != root)
1273 css_put(&prev->css);
1277 * Iteration constructs for visiting all cgroups (under a tree). If
1278 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1279 * be used for reference counting.
1281 #define for_each_mem_cgroup_tree(iter, root) \
1282 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1284 iter = mem_cgroup_iter(root, iter, NULL))
1286 #define for_each_mem_cgroup(iter) \
1287 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1289 iter = mem_cgroup_iter(NULL, iter, NULL))
1291 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1293 struct mem_cgroup *memcg;
1296 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1297 if (unlikely(!memcg))
1302 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1305 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1313 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1316 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1317 * @zone: zone of the wanted lruvec
1318 * @memcg: memcg of the wanted lruvec
1320 * Returns the lru list vector holding pages for the given @zone and
1321 * @mem. This can be the global zone lruvec, if the memory controller
1324 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1325 struct mem_cgroup *memcg)
1327 struct mem_cgroup_per_zone *mz;
1328 struct lruvec *lruvec;
1330 if (mem_cgroup_disabled()) {
1331 lruvec = &zone->lruvec;
1335 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1336 lruvec = &mz->lruvec;
1339 * Since a node can be onlined after the mem_cgroup was created,
1340 * we have to be prepared to initialize lruvec->zone here;
1341 * and if offlined then reonlined, we need to reinitialize it.
1343 if (unlikely(lruvec->zone != zone))
1344 lruvec->zone = zone;
1349 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1351 * @zone: zone of the page
1353 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1355 struct mem_cgroup_per_zone *mz;
1356 struct mem_cgroup *memcg;
1357 struct page_cgroup *pc;
1358 struct lruvec *lruvec;
1360 if (mem_cgroup_disabled()) {
1361 lruvec = &zone->lruvec;
1365 pc = lookup_page_cgroup(page);
1366 memcg = pc->mem_cgroup;
1369 * Surreptitiously switch any uncharged offlist page to root:
1370 * an uncharged page off lru does nothing to secure
1371 * its former mem_cgroup from sudden removal.
1373 * Our caller holds lru_lock, and PageCgroupUsed is updated
1374 * under page_cgroup lock: between them, they make all uses
1375 * of pc->mem_cgroup safe.
1377 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1378 pc->mem_cgroup = memcg = root_mem_cgroup;
1380 mz = mem_cgroup_page_zoneinfo(memcg, page);
1381 lruvec = &mz->lruvec;
1384 * Since a node can be onlined after the mem_cgroup was created,
1385 * we have to be prepared to initialize lruvec->zone here;
1386 * and if offlined then reonlined, we need to reinitialize it.
1388 if (unlikely(lruvec->zone != zone))
1389 lruvec->zone = zone;
1394 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1395 * @lruvec: mem_cgroup per zone lru vector
1396 * @lru: index of lru list the page is sitting on
1397 * @nr_pages: positive when adding or negative when removing
1399 * This function must be called when a page is added to or removed from an
1402 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1405 struct mem_cgroup_per_zone *mz;
1406 unsigned long *lru_size;
1408 if (mem_cgroup_disabled())
1411 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1412 lru_size = mz->lru_size + lru;
1413 *lru_size += nr_pages;
1414 VM_BUG_ON((long)(*lru_size) < 0);
1418 * Checks whether given mem is same or in the root_mem_cgroup's
1421 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1422 struct mem_cgroup *memcg)
1424 if (root_memcg == memcg)
1426 if (!root_memcg->use_hierarchy || !memcg)
1428 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1431 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1432 struct mem_cgroup *memcg)
1437 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1442 bool task_in_mem_cgroup(struct task_struct *task,
1443 const struct mem_cgroup *memcg)
1445 struct mem_cgroup *curr = NULL;
1446 struct task_struct *p;
1449 p = find_lock_task_mm(task);
1451 curr = get_mem_cgroup_from_mm(p->mm);
1455 * All threads may have already detached their mm's, but the oom
1456 * killer still needs to detect if they have already been oom
1457 * killed to prevent needlessly killing additional tasks.
1460 curr = mem_cgroup_from_task(task);
1462 css_get(&curr->css);
1466 * We should check use_hierarchy of "memcg" not "curr". Because checking
1467 * use_hierarchy of "curr" here make this function true if hierarchy is
1468 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1469 * hierarchy(even if use_hierarchy is disabled in "memcg").
1471 ret = mem_cgroup_same_or_subtree(memcg, curr);
1472 css_put(&curr->css);
1476 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1478 unsigned long inactive_ratio;
1479 unsigned long inactive;
1480 unsigned long active;
1483 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1484 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1486 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1488 inactive_ratio = int_sqrt(10 * gb);
1492 return inactive * inactive_ratio < active;
1495 #define mem_cgroup_from_res_counter(counter, member) \
1496 container_of(counter, struct mem_cgroup, member)
1499 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1500 * @memcg: the memory cgroup
1502 * Returns the maximum amount of memory @mem can be charged with, in
1505 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1507 unsigned long long margin;
1509 margin = res_counter_margin(&memcg->res);
1510 if (do_swap_account)
1511 margin = min(margin, res_counter_margin(&memcg->memsw));
1512 return margin >> PAGE_SHIFT;
1515 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1518 if (mem_cgroup_disabled() || !memcg->css.parent)
1519 return vm_swappiness;
1521 return memcg->swappiness;
1525 * memcg->moving_account is used for checking possibility that some thread is
1526 * calling move_account(). When a thread on CPU-A starts moving pages under
1527 * a memcg, other threads should check memcg->moving_account under
1528 * rcu_read_lock(), like this:
1532 * memcg->moving_account+1 if (memcg->mocing_account)
1534 * synchronize_rcu() update something.
1539 /* for quick checking without looking up memcg */
1540 atomic_t memcg_moving __read_mostly;
1542 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1544 atomic_inc(&memcg_moving);
1545 atomic_inc(&memcg->moving_account);
1549 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1552 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1553 * We check NULL in callee rather than caller.
1556 atomic_dec(&memcg_moving);
1557 atomic_dec(&memcg->moving_account);
1562 * A routine for checking "mem" is under move_account() or not.
1564 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1565 * moving cgroups. This is for waiting at high-memory pressure
1568 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1570 struct mem_cgroup *from;
1571 struct mem_cgroup *to;
1574 * Unlike task_move routines, we access mc.to, mc.from not under
1575 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1577 spin_lock(&mc.lock);
1583 ret = mem_cgroup_same_or_subtree(memcg, from)
1584 || mem_cgroup_same_or_subtree(memcg, to);
1586 spin_unlock(&mc.lock);
1590 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1592 if (mc.moving_task && current != mc.moving_task) {
1593 if (mem_cgroup_under_move(memcg)) {
1595 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1596 /* moving charge context might have finished. */
1599 finish_wait(&mc.waitq, &wait);
1607 * Take this lock when
1608 * - a code tries to modify page's memcg while it's USED.
1609 * - a code tries to modify page state accounting in a memcg.
1611 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1612 unsigned long *flags)
1614 spin_lock_irqsave(&memcg->move_lock, *flags);
1617 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1618 unsigned long *flags)
1620 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1623 #define K(x) ((x) << (PAGE_SHIFT-10))
1625 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1626 * @memcg: The memory cgroup that went over limit
1627 * @p: Task that is going to be killed
1629 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1632 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1634 /* oom_info_lock ensures that parallel ooms do not interleave */
1635 static DEFINE_MUTEX(oom_info_lock);
1636 struct mem_cgroup *iter;
1642 mutex_lock(&oom_info_lock);
1645 pr_info("Task in ");
1646 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1647 pr_info(" killed as a result of limit of ");
1648 pr_cont_cgroup_path(memcg->css.cgroup);
1653 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1654 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1655 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1656 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1657 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1658 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1659 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1660 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1661 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1662 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1663 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1664 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1666 for_each_mem_cgroup_tree(iter, memcg) {
1667 pr_info("Memory cgroup stats for ");
1668 pr_cont_cgroup_path(iter->css.cgroup);
1671 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1672 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1674 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1675 K(mem_cgroup_read_stat(iter, i)));
1678 for (i = 0; i < NR_LRU_LISTS; i++)
1679 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1680 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1684 mutex_unlock(&oom_info_lock);
1688 * This function returns the number of memcg under hierarchy tree. Returns
1689 * 1(self count) if no children.
1691 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1694 struct mem_cgroup *iter;
1696 for_each_mem_cgroup_tree(iter, memcg)
1702 * Return the memory (and swap, if configured) limit for a memcg.
1704 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1708 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1711 * Do not consider swap space if we cannot swap due to swappiness
1713 if (mem_cgroup_swappiness(memcg)) {
1716 limit += total_swap_pages << PAGE_SHIFT;
1717 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1720 * If memsw is finite and limits the amount of swap space
1721 * available to this memcg, return that limit.
1723 limit = min(limit, memsw);
1729 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1732 struct mem_cgroup *iter;
1733 unsigned long chosen_points = 0;
1734 unsigned long totalpages;
1735 unsigned int points = 0;
1736 struct task_struct *chosen = NULL;
1739 * If current has a pending SIGKILL or is exiting, then automatically
1740 * select it. The goal is to allow it to allocate so that it may
1741 * quickly exit and free its memory.
1743 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1744 set_thread_flag(TIF_MEMDIE);
1748 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1749 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1750 for_each_mem_cgroup_tree(iter, memcg) {
1751 struct css_task_iter it;
1752 struct task_struct *task;
1754 css_task_iter_start(&iter->css, &it);
1755 while ((task = css_task_iter_next(&it))) {
1756 switch (oom_scan_process_thread(task, totalpages, NULL,
1758 case OOM_SCAN_SELECT:
1760 put_task_struct(chosen);
1762 chosen_points = ULONG_MAX;
1763 get_task_struct(chosen);
1765 case OOM_SCAN_CONTINUE:
1767 case OOM_SCAN_ABORT:
1768 css_task_iter_end(&it);
1769 mem_cgroup_iter_break(memcg, iter);
1771 put_task_struct(chosen);
1776 points = oom_badness(task, memcg, NULL, totalpages);
1777 if (!points || points < chosen_points)
1779 /* Prefer thread group leaders for display purposes */
1780 if (points == chosen_points &&
1781 thread_group_leader(chosen))
1785 put_task_struct(chosen);
1787 chosen_points = points;
1788 get_task_struct(chosen);
1790 css_task_iter_end(&it);
1795 points = chosen_points * 1000 / totalpages;
1796 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1797 NULL, "Memory cgroup out of memory");
1801 * test_mem_cgroup_node_reclaimable
1802 * @memcg: the target memcg
1803 * @nid: the node ID to be checked.
1804 * @noswap : specify true here if the user wants flle only information.
1806 * This function returns whether the specified memcg contains any
1807 * reclaimable pages on a node. Returns true if there are any reclaimable
1808 * pages in the node.
1810 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1811 int nid, bool noswap)
1813 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1815 if (noswap || !total_swap_pages)
1817 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1822 #if MAX_NUMNODES > 1
1825 * Always updating the nodemask is not very good - even if we have an empty
1826 * list or the wrong list here, we can start from some node and traverse all
1827 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1830 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1834 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1835 * pagein/pageout changes since the last update.
1837 if (!atomic_read(&memcg->numainfo_events))
1839 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1842 /* make a nodemask where this memcg uses memory from */
1843 memcg->scan_nodes = node_states[N_MEMORY];
1845 for_each_node_mask(nid, node_states[N_MEMORY]) {
1847 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1848 node_clear(nid, memcg->scan_nodes);
1851 atomic_set(&memcg->numainfo_events, 0);
1852 atomic_set(&memcg->numainfo_updating, 0);
1856 * Selecting a node where we start reclaim from. Because what we need is just
1857 * reducing usage counter, start from anywhere is O,K. Considering
1858 * memory reclaim from current node, there are pros. and cons.
1860 * Freeing memory from current node means freeing memory from a node which
1861 * we'll use or we've used. So, it may make LRU bad. And if several threads
1862 * hit limits, it will see a contention on a node. But freeing from remote
1863 * node means more costs for memory reclaim because of memory latency.
1865 * Now, we use round-robin. Better algorithm is welcomed.
1867 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1871 mem_cgroup_may_update_nodemask(memcg);
1872 node = memcg->last_scanned_node;
1874 node = next_node(node, memcg->scan_nodes);
1875 if (node == MAX_NUMNODES)
1876 node = first_node(memcg->scan_nodes);
1878 * We call this when we hit limit, not when pages are added to LRU.
1879 * No LRU may hold pages because all pages are UNEVICTABLE or
1880 * memcg is too small and all pages are not on LRU. In that case,
1881 * we use curret node.
1883 if (unlikely(node == MAX_NUMNODES))
1884 node = numa_node_id();
1886 memcg->last_scanned_node = node;
1891 * Check all nodes whether it contains reclaimable pages or not.
1892 * For quick scan, we make use of scan_nodes. This will allow us to skip
1893 * unused nodes. But scan_nodes is lazily updated and may not cotain
1894 * enough new information. We need to do double check.
1896 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1901 * quick check...making use of scan_node.
1902 * We can skip unused nodes.
1904 if (!nodes_empty(memcg->scan_nodes)) {
1905 for (nid = first_node(memcg->scan_nodes);
1907 nid = next_node(nid, memcg->scan_nodes)) {
1909 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1914 * Check rest of nodes.
1916 for_each_node_state(nid, N_MEMORY) {
1917 if (node_isset(nid, memcg->scan_nodes))
1919 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1926 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1931 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1933 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1937 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1940 unsigned long *total_scanned)
1942 struct mem_cgroup *victim = NULL;
1945 unsigned long excess;
1946 unsigned long nr_scanned;
1947 struct mem_cgroup_reclaim_cookie reclaim = {
1952 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1955 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1960 * If we have not been able to reclaim
1961 * anything, it might because there are
1962 * no reclaimable pages under this hierarchy
1967 * We want to do more targeted reclaim.
1968 * excess >> 2 is not to excessive so as to
1969 * reclaim too much, nor too less that we keep
1970 * coming back to reclaim from this cgroup
1972 if (total >= (excess >> 2) ||
1973 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1978 if (!mem_cgroup_reclaimable(victim, false))
1980 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1982 *total_scanned += nr_scanned;
1983 if (!res_counter_soft_limit_excess(&root_memcg->res))
1986 mem_cgroup_iter_break(root_memcg, victim);
1990 #ifdef CONFIG_LOCKDEP
1991 static struct lockdep_map memcg_oom_lock_dep_map = {
1992 .name = "memcg_oom_lock",
1996 static DEFINE_SPINLOCK(memcg_oom_lock);
1999 * Check OOM-Killer is already running under our hierarchy.
2000 * If someone is running, return false.
2002 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2004 struct mem_cgroup *iter, *failed = NULL;
2006 spin_lock(&memcg_oom_lock);
2008 for_each_mem_cgroup_tree(iter, memcg) {
2009 if (iter->oom_lock) {
2011 * this subtree of our hierarchy is already locked
2012 * so we cannot give a lock.
2015 mem_cgroup_iter_break(memcg, iter);
2018 iter->oom_lock = true;
2023 * OK, we failed to lock the whole subtree so we have
2024 * to clean up what we set up to the failing subtree
2026 for_each_mem_cgroup_tree(iter, memcg) {
2027 if (iter == failed) {
2028 mem_cgroup_iter_break(memcg, iter);
2031 iter->oom_lock = false;
2034 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2036 spin_unlock(&memcg_oom_lock);
2041 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2043 struct mem_cgroup *iter;
2045 spin_lock(&memcg_oom_lock);
2046 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2047 for_each_mem_cgroup_tree(iter, memcg)
2048 iter->oom_lock = false;
2049 spin_unlock(&memcg_oom_lock);
2052 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2054 struct mem_cgroup *iter;
2056 for_each_mem_cgroup_tree(iter, memcg)
2057 atomic_inc(&iter->under_oom);
2060 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2062 struct mem_cgroup *iter;
2065 * When a new child is created while the hierarchy is under oom,
2066 * mem_cgroup_oom_lock() may not be called. We have to use
2067 * atomic_add_unless() here.
2069 for_each_mem_cgroup_tree(iter, memcg)
2070 atomic_add_unless(&iter->under_oom, -1, 0);
2073 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2075 struct oom_wait_info {
2076 struct mem_cgroup *memcg;
2080 static int memcg_oom_wake_function(wait_queue_t *wait,
2081 unsigned mode, int sync, void *arg)
2083 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2084 struct mem_cgroup *oom_wait_memcg;
2085 struct oom_wait_info *oom_wait_info;
2087 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2088 oom_wait_memcg = oom_wait_info->memcg;
2091 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2092 * Then we can use css_is_ancestor without taking care of RCU.
2094 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2095 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2097 return autoremove_wake_function(wait, mode, sync, arg);
2100 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2102 atomic_inc(&memcg->oom_wakeups);
2103 /* for filtering, pass "memcg" as argument. */
2104 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2107 static void memcg_oom_recover(struct mem_cgroup *memcg)
2109 if (memcg && atomic_read(&memcg->under_oom))
2110 memcg_wakeup_oom(memcg);
2113 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2115 if (!current->memcg_oom.may_oom)
2118 * We are in the middle of the charge context here, so we
2119 * don't want to block when potentially sitting on a callstack
2120 * that holds all kinds of filesystem and mm locks.
2122 * Also, the caller may handle a failed allocation gracefully
2123 * (like optional page cache readahead) and so an OOM killer
2124 * invocation might not even be necessary.
2126 * That's why we don't do anything here except remember the
2127 * OOM context and then deal with it at the end of the page
2128 * fault when the stack is unwound, the locks are released,
2129 * and when we know whether the fault was overall successful.
2131 css_get(&memcg->css);
2132 current->memcg_oom.memcg = memcg;
2133 current->memcg_oom.gfp_mask = mask;
2134 current->memcg_oom.order = order;
2138 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2139 * @handle: actually kill/wait or just clean up the OOM state
2141 * This has to be called at the end of a page fault if the memcg OOM
2142 * handler was enabled.
2144 * Memcg supports userspace OOM handling where failed allocations must
2145 * sleep on a waitqueue until the userspace task resolves the
2146 * situation. Sleeping directly in the charge context with all kinds
2147 * of locks held is not a good idea, instead we remember an OOM state
2148 * in the task and mem_cgroup_oom_synchronize() has to be called at
2149 * the end of the page fault to complete the OOM handling.
2151 * Returns %true if an ongoing memcg OOM situation was detected and
2152 * completed, %false otherwise.
2154 bool mem_cgroup_oom_synchronize(bool handle)
2156 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2157 struct oom_wait_info owait;
2160 /* OOM is global, do not handle */
2167 owait.memcg = memcg;
2168 owait.wait.flags = 0;
2169 owait.wait.func = memcg_oom_wake_function;
2170 owait.wait.private = current;
2171 INIT_LIST_HEAD(&owait.wait.task_list);
2173 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2174 mem_cgroup_mark_under_oom(memcg);
2176 locked = mem_cgroup_oom_trylock(memcg);
2179 mem_cgroup_oom_notify(memcg);
2181 if (locked && !memcg->oom_kill_disable) {
2182 mem_cgroup_unmark_under_oom(memcg);
2183 finish_wait(&memcg_oom_waitq, &owait.wait);
2184 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2185 current->memcg_oom.order);
2188 mem_cgroup_unmark_under_oom(memcg);
2189 finish_wait(&memcg_oom_waitq, &owait.wait);
2193 mem_cgroup_oom_unlock(memcg);
2195 * There is no guarantee that an OOM-lock contender
2196 * sees the wakeups triggered by the OOM kill
2197 * uncharges. Wake any sleepers explicitely.
2199 memcg_oom_recover(memcg);
2202 current->memcg_oom.memcg = NULL;
2203 css_put(&memcg->css);
2208 * Used to update mapped file or writeback or other statistics.
2210 * Notes: Race condition
2212 * Charging occurs during page instantiation, while the page is
2213 * unmapped and locked in page migration, or while the page table is
2214 * locked in THP migration. No race is possible.
2216 * Uncharge happens to pages with zero references, no race possible.
2218 * Charge moving between groups is protected by checking mm->moving
2219 * account and taking the move_lock in the slowpath.
2222 void __mem_cgroup_begin_update_page_stat(struct page *page,
2223 bool *locked, unsigned long *flags)
2225 struct mem_cgroup *memcg;
2226 struct page_cgroup *pc;
2228 pc = lookup_page_cgroup(page);
2230 memcg = pc->mem_cgroup;
2231 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2234 * If this memory cgroup is not under account moving, we don't
2235 * need to take move_lock_mem_cgroup(). Because we already hold
2236 * rcu_read_lock(), any calls to move_account will be delayed until
2237 * rcu_read_unlock().
2239 VM_BUG_ON(!rcu_read_lock_held());
2240 if (atomic_read(&memcg->moving_account) <= 0)
2243 move_lock_mem_cgroup(memcg, flags);
2244 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2245 move_unlock_mem_cgroup(memcg, flags);
2251 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2253 struct page_cgroup *pc = lookup_page_cgroup(page);
2256 * It's guaranteed that pc->mem_cgroup never changes while
2257 * lock is held because a routine modifies pc->mem_cgroup
2258 * should take move_lock_mem_cgroup().
2260 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2263 void mem_cgroup_update_page_stat(struct page *page,
2264 enum mem_cgroup_stat_index idx, int val)
2266 struct mem_cgroup *memcg;
2267 struct page_cgroup *pc = lookup_page_cgroup(page);
2268 unsigned long uninitialized_var(flags);
2270 if (mem_cgroup_disabled())
2273 VM_BUG_ON(!rcu_read_lock_held());
2274 memcg = pc->mem_cgroup;
2275 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2278 this_cpu_add(memcg->stat->count[idx], val);
2282 * size of first charge trial. "32" comes from vmscan.c's magic value.
2283 * TODO: maybe necessary to use big numbers in big irons.
2285 #define CHARGE_BATCH 32U
2286 struct memcg_stock_pcp {
2287 struct mem_cgroup *cached; /* this never be root cgroup */
2288 unsigned int nr_pages;
2289 struct work_struct work;
2290 unsigned long flags;
2291 #define FLUSHING_CACHED_CHARGE 0
2293 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2294 static DEFINE_MUTEX(percpu_charge_mutex);
2297 * consume_stock: Try to consume stocked charge on this cpu.
2298 * @memcg: memcg to consume from.
2299 * @nr_pages: how many pages to charge.
2301 * The charges will only happen if @memcg matches the current cpu's memcg
2302 * stock, and at least @nr_pages are available in that stock. Failure to
2303 * service an allocation will refill the stock.
2305 * returns true if successful, false otherwise.
2307 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2309 struct memcg_stock_pcp *stock;
2312 if (nr_pages > CHARGE_BATCH)
2315 stock = &get_cpu_var(memcg_stock);
2316 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2317 stock->nr_pages -= nr_pages;
2318 else /* need to call res_counter_charge */
2320 put_cpu_var(memcg_stock);
2325 * Returns stocks cached in percpu to res_counter and reset cached information.
2327 static void drain_stock(struct memcg_stock_pcp *stock)
2329 struct mem_cgroup *old = stock->cached;
2331 if (stock->nr_pages) {
2332 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2334 res_counter_uncharge(&old->res, bytes);
2335 if (do_swap_account)
2336 res_counter_uncharge(&old->memsw, bytes);
2337 stock->nr_pages = 0;
2339 stock->cached = NULL;
2343 * This must be called under preempt disabled or must be called by
2344 * a thread which is pinned to local cpu.
2346 static void drain_local_stock(struct work_struct *dummy)
2348 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2350 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2353 static void __init memcg_stock_init(void)
2357 for_each_possible_cpu(cpu) {
2358 struct memcg_stock_pcp *stock =
2359 &per_cpu(memcg_stock, cpu);
2360 INIT_WORK(&stock->work, drain_local_stock);
2365 * Cache charges(val) which is from res_counter, to local per_cpu area.
2366 * This will be consumed by consume_stock() function, later.
2368 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2370 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2372 if (stock->cached != memcg) { /* reset if necessary */
2374 stock->cached = memcg;
2376 stock->nr_pages += nr_pages;
2377 put_cpu_var(memcg_stock);
2381 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2382 * of the hierarchy under it. sync flag says whether we should block
2383 * until the work is done.
2385 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2389 /* Notify other cpus that system-wide "drain" is running */
2392 for_each_online_cpu(cpu) {
2393 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2394 struct mem_cgroup *memcg;
2396 memcg = stock->cached;
2397 if (!memcg || !stock->nr_pages)
2399 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2401 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2403 drain_local_stock(&stock->work);
2405 schedule_work_on(cpu, &stock->work);
2413 for_each_online_cpu(cpu) {
2414 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2415 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2416 flush_work(&stock->work);
2423 * Tries to drain stocked charges in other cpus. This function is asynchronous
2424 * and just put a work per cpu for draining localy on each cpu. Caller can
2425 * expects some charges will be back to res_counter later but cannot wait for
2428 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2431 * If someone calls draining, avoid adding more kworker runs.
2433 if (!mutex_trylock(&percpu_charge_mutex))
2435 drain_all_stock(root_memcg, false);
2436 mutex_unlock(&percpu_charge_mutex);
2439 /* This is a synchronous drain interface. */
2440 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2442 /* called when force_empty is called */
2443 mutex_lock(&percpu_charge_mutex);
2444 drain_all_stock(root_memcg, true);
2445 mutex_unlock(&percpu_charge_mutex);
2449 * This function drains percpu counter value from DEAD cpu and
2450 * move it to local cpu. Note that this function can be preempted.
2452 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2456 spin_lock(&memcg->pcp_counter_lock);
2457 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2458 long x = per_cpu(memcg->stat->count[i], cpu);
2460 per_cpu(memcg->stat->count[i], cpu) = 0;
2461 memcg->nocpu_base.count[i] += x;
2463 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2464 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2466 per_cpu(memcg->stat->events[i], cpu) = 0;
2467 memcg->nocpu_base.events[i] += x;
2469 spin_unlock(&memcg->pcp_counter_lock);
2472 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2473 unsigned long action,
2476 int cpu = (unsigned long)hcpu;
2477 struct memcg_stock_pcp *stock;
2478 struct mem_cgroup *iter;
2480 if (action == CPU_ONLINE)
2483 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2486 for_each_mem_cgroup(iter)
2487 mem_cgroup_drain_pcp_counter(iter, cpu);
2489 stock = &per_cpu(memcg_stock, cpu);
2494 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2495 unsigned int nr_pages)
2497 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2498 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2499 struct mem_cgroup *mem_over_limit;
2500 struct res_counter *fail_res;
2501 unsigned long nr_reclaimed;
2502 unsigned long long size;
2503 bool may_swap = true;
2504 bool drained = false;
2507 if (mem_cgroup_is_root(memcg))
2510 if (consume_stock(memcg, nr_pages))
2513 size = batch * PAGE_SIZE;
2514 if (!do_swap_account ||
2515 !res_counter_charge(&memcg->memsw, size, &fail_res)) {
2516 if (!res_counter_charge(&memcg->res, size, &fail_res))
2518 if (do_swap_account)
2519 res_counter_uncharge(&memcg->memsw, size);
2520 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2522 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2526 if (batch > nr_pages) {
2532 * Unlike in global OOM situations, memcg is not in a physical
2533 * memory shortage. Allow dying and OOM-killed tasks to
2534 * bypass the last charges so that they can exit quickly and
2535 * free their memory.
2537 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2538 fatal_signal_pending(current) ||
2539 current->flags & PF_EXITING))
2542 if (unlikely(task_in_memcg_oom(current)))
2545 if (!(gfp_mask & __GFP_WAIT))
2548 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2549 gfp_mask, may_swap);
2551 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2555 drain_all_stock_async(mem_over_limit);
2560 if (gfp_mask & __GFP_NORETRY)
2563 * Even though the limit is exceeded at this point, reclaim
2564 * may have been able to free some pages. Retry the charge
2565 * before killing the task.
2567 * Only for regular pages, though: huge pages are rather
2568 * unlikely to succeed so close to the limit, and we fall back
2569 * to regular pages anyway in case of failure.
2571 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2574 * At task move, charge accounts can be doubly counted. So, it's
2575 * better to wait until the end of task_move if something is going on.
2577 if (mem_cgroup_wait_acct_move(mem_over_limit))
2583 if (gfp_mask & __GFP_NOFAIL)
2586 if (fatal_signal_pending(current))
2589 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2591 if (!(gfp_mask & __GFP_NOFAIL))
2597 if (batch > nr_pages)
2598 refill_stock(memcg, batch - nr_pages);
2603 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2605 unsigned long bytes = nr_pages * PAGE_SIZE;
2607 if (mem_cgroup_is_root(memcg))
2610 res_counter_uncharge(&memcg->res, bytes);
2611 if (do_swap_account)
2612 res_counter_uncharge(&memcg->memsw, bytes);
2616 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2617 * This is useful when moving usage to parent cgroup.
2619 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2620 unsigned int nr_pages)
2622 unsigned long bytes = nr_pages * PAGE_SIZE;
2624 if (mem_cgroup_is_root(memcg))
2627 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2628 if (do_swap_account)
2629 res_counter_uncharge_until(&memcg->memsw,
2630 memcg->memsw.parent, bytes);
2634 * A helper function to get mem_cgroup from ID. must be called under
2635 * rcu_read_lock(). The caller is responsible for calling
2636 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2637 * refcnt from swap can be called against removed memcg.)
2639 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2641 /* ID 0 is unused ID */
2644 return mem_cgroup_from_id(id);
2648 * try_get_mem_cgroup_from_page - look up page's memcg association
2651 * Look up, get a css reference, and return the memcg that owns @page.
2653 * The page must be locked to prevent racing with swap-in and page
2654 * cache charges. If coming from an unlocked page table, the caller
2655 * must ensure the page is on the LRU or this can race with charging.
2657 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2659 struct mem_cgroup *memcg = NULL;
2660 struct page_cgroup *pc;
2664 VM_BUG_ON_PAGE(!PageLocked(page), page);
2666 pc = lookup_page_cgroup(page);
2667 if (PageCgroupUsed(pc)) {
2668 memcg = pc->mem_cgroup;
2669 if (memcg && !css_tryget_online(&memcg->css))
2671 } else if (PageSwapCache(page)) {
2672 ent.val = page_private(page);
2673 id = lookup_swap_cgroup_id(ent);
2675 memcg = mem_cgroup_lookup(id);
2676 if (memcg && !css_tryget_online(&memcg->css))
2683 static void lock_page_lru(struct page *page, int *isolated)
2685 struct zone *zone = page_zone(page);
2687 spin_lock_irq(&zone->lru_lock);
2688 if (PageLRU(page)) {
2689 struct lruvec *lruvec;
2691 lruvec = mem_cgroup_page_lruvec(page, zone);
2693 del_page_from_lru_list(page, lruvec, page_lru(page));
2699 static void unlock_page_lru(struct page *page, int isolated)
2701 struct zone *zone = page_zone(page);
2704 struct lruvec *lruvec;
2706 lruvec = mem_cgroup_page_lruvec(page, zone);
2707 VM_BUG_ON_PAGE(PageLRU(page), page);
2709 add_page_to_lru_list(page, lruvec, page_lru(page));
2711 spin_unlock_irq(&zone->lru_lock);
2714 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2717 struct page_cgroup *pc = lookup_page_cgroup(page);
2720 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2722 * we don't need page_cgroup_lock about tail pages, becase they are not
2723 * accessed by any other context at this point.
2727 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2728 * may already be on some other mem_cgroup's LRU. Take care of it.
2731 lock_page_lru(page, &isolated);
2734 * Nobody should be changing or seriously looking at
2735 * pc->mem_cgroup and pc->flags at this point:
2737 * - the page is uncharged
2739 * - the page is off-LRU
2741 * - an anonymous fault has exclusive page access, except for
2742 * a locked page table
2744 * - a page cache insertion, a swapin fault, or a migration
2745 * have the page locked
2747 pc->mem_cgroup = memcg;
2748 pc->flags = PCG_USED | PCG_MEM | (do_swap_account ? PCG_MEMSW : 0);
2751 unlock_page_lru(page, isolated);
2754 static DEFINE_MUTEX(set_limit_mutex);
2756 #ifdef CONFIG_MEMCG_KMEM
2758 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2759 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2761 static DEFINE_MUTEX(memcg_slab_mutex);
2763 static DEFINE_MUTEX(activate_kmem_mutex);
2765 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2767 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2768 memcg_kmem_is_active(memcg);
2772 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2773 * in the memcg_cache_params struct.
2775 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2777 struct kmem_cache *cachep;
2779 VM_BUG_ON(p->is_root_cache);
2780 cachep = p->root_cache;
2781 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2784 #ifdef CONFIG_SLABINFO
2785 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2787 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2788 struct memcg_cache_params *params;
2790 if (!memcg_can_account_kmem(memcg))
2793 print_slabinfo_header(m);
2795 mutex_lock(&memcg_slab_mutex);
2796 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2797 cache_show(memcg_params_to_cache(params), m);
2798 mutex_unlock(&memcg_slab_mutex);
2804 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2806 struct res_counter *fail_res;
2809 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2813 ret = try_charge(memcg, gfp, size >> PAGE_SHIFT);
2814 if (ret == -EINTR) {
2816 * try_charge() chose to bypass to root due to OOM kill or
2817 * fatal signal. Since our only options are to either fail
2818 * the allocation or charge it to this cgroup, do it as a
2819 * temporary condition. But we can't fail. From a kmem/slab
2820 * perspective, the cache has already been selected, by
2821 * mem_cgroup_kmem_get_cache(), so it is too late to change
2824 * This condition will only trigger if the task entered
2825 * memcg_charge_kmem in a sane state, but was OOM-killed
2826 * during try_charge() above. Tasks that were already dying
2827 * when the allocation triggers should have been already
2828 * directed to the root cgroup in memcontrol.h
2830 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2831 if (do_swap_account)
2832 res_counter_charge_nofail(&memcg->memsw, size,
2836 res_counter_uncharge(&memcg->kmem, size);
2841 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2843 res_counter_uncharge(&memcg->res, size);
2844 if (do_swap_account)
2845 res_counter_uncharge(&memcg->memsw, size);
2848 if (res_counter_uncharge(&memcg->kmem, size))
2852 * Releases a reference taken in kmem_cgroup_css_offline in case
2853 * this last uncharge is racing with the offlining code or it is
2854 * outliving the memcg existence.
2856 * The memory barrier imposed by test&clear is paired with the
2857 * explicit one in memcg_kmem_mark_dead().
2859 if (memcg_kmem_test_and_clear_dead(memcg))
2860 css_put(&memcg->css);
2864 * helper for acessing a memcg's index. It will be used as an index in the
2865 * child cache array in kmem_cache, and also to derive its name. This function
2866 * will return -1 when this is not a kmem-limited memcg.
2868 int memcg_cache_id(struct mem_cgroup *memcg)
2870 return memcg ? memcg->kmemcg_id : -1;
2873 static int memcg_alloc_cache_id(void)
2878 id = ida_simple_get(&kmem_limited_groups,
2879 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2883 if (id < memcg_limited_groups_array_size)
2887 * There's no space for the new id in memcg_caches arrays,
2888 * so we have to grow them.
2891 size = 2 * (id + 1);
2892 if (size < MEMCG_CACHES_MIN_SIZE)
2893 size = MEMCG_CACHES_MIN_SIZE;
2894 else if (size > MEMCG_CACHES_MAX_SIZE)
2895 size = MEMCG_CACHES_MAX_SIZE;
2897 mutex_lock(&memcg_slab_mutex);
2898 err = memcg_update_all_caches(size);
2899 mutex_unlock(&memcg_slab_mutex);
2902 ida_simple_remove(&kmem_limited_groups, id);
2908 static void memcg_free_cache_id(int id)
2910 ida_simple_remove(&kmem_limited_groups, id);
2914 * We should update the current array size iff all caches updates succeed. This
2915 * can only be done from the slab side. The slab mutex needs to be held when
2918 void memcg_update_array_size(int num)
2920 memcg_limited_groups_array_size = num;
2923 static void memcg_register_cache(struct mem_cgroup *memcg,
2924 struct kmem_cache *root_cache)
2926 static char memcg_name_buf[NAME_MAX + 1]; /* protected by
2928 struct kmem_cache *cachep;
2931 lockdep_assert_held(&memcg_slab_mutex);
2933 id = memcg_cache_id(memcg);
2936 * Since per-memcg caches are created asynchronously on first
2937 * allocation (see memcg_kmem_get_cache()), several threads can try to
2938 * create the same cache, but only one of them may succeed.
2940 if (cache_from_memcg_idx(root_cache, id))
2943 cgroup_name(memcg->css.cgroup, memcg_name_buf, NAME_MAX + 1);
2944 cachep = memcg_create_kmem_cache(memcg, root_cache, memcg_name_buf);
2946 * If we could not create a memcg cache, do not complain, because
2947 * that's not critical at all as we can always proceed with the root
2953 css_get(&memcg->css);
2954 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2957 * Since readers won't lock (see cache_from_memcg_idx()), we need a
2958 * barrier here to ensure nobody will see the kmem_cache partially
2963 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
2964 root_cache->memcg_params->memcg_caches[id] = cachep;
2967 static void memcg_unregister_cache(struct kmem_cache *cachep)
2969 struct kmem_cache *root_cache;
2970 struct mem_cgroup *memcg;
2973 lockdep_assert_held(&memcg_slab_mutex);
2975 BUG_ON(is_root_cache(cachep));
2977 root_cache = cachep->memcg_params->root_cache;
2978 memcg = cachep->memcg_params->memcg;
2979 id = memcg_cache_id(memcg);
2981 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
2982 root_cache->memcg_params->memcg_caches[id] = NULL;
2984 list_del(&cachep->memcg_params->list);
2986 kmem_cache_destroy(cachep);
2988 /* drop the reference taken in memcg_register_cache */
2989 css_put(&memcg->css);
2993 * During the creation a new cache, we need to disable our accounting mechanism
2994 * altogether. This is true even if we are not creating, but rather just
2995 * enqueing new caches to be created.
2997 * This is because that process will trigger allocations; some visible, like
2998 * explicit kmallocs to auxiliary data structures, name strings and internal
2999 * cache structures; some well concealed, like INIT_WORK() that can allocate
3000 * objects during debug.
3002 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3003 * to it. This may not be a bounded recursion: since the first cache creation
3004 * failed to complete (waiting on the allocation), we'll just try to create the
3005 * cache again, failing at the same point.
3007 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3008 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3009 * inside the following two functions.
3011 static inline void memcg_stop_kmem_account(void)
3013 VM_BUG_ON(!current->mm);
3014 current->memcg_kmem_skip_account++;
3017 static inline void memcg_resume_kmem_account(void)
3019 VM_BUG_ON(!current->mm);
3020 current->memcg_kmem_skip_account--;
3023 int __memcg_cleanup_cache_params(struct kmem_cache *s)
3025 struct kmem_cache *c;
3028 mutex_lock(&memcg_slab_mutex);
3029 for_each_memcg_cache_index(i) {
3030 c = cache_from_memcg_idx(s, i);
3034 memcg_unregister_cache(c);
3036 if (cache_from_memcg_idx(s, i))
3039 mutex_unlock(&memcg_slab_mutex);
3043 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3045 struct kmem_cache *cachep;
3046 struct memcg_cache_params *params, *tmp;
3048 if (!memcg_kmem_is_active(memcg))
3051 mutex_lock(&memcg_slab_mutex);
3052 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
3053 cachep = memcg_params_to_cache(params);
3054 kmem_cache_shrink(cachep);
3055 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3056 memcg_unregister_cache(cachep);
3058 mutex_unlock(&memcg_slab_mutex);
3061 struct memcg_register_cache_work {
3062 struct mem_cgroup *memcg;
3063 struct kmem_cache *cachep;
3064 struct work_struct work;
3067 static void memcg_register_cache_func(struct work_struct *w)
3069 struct memcg_register_cache_work *cw =
3070 container_of(w, struct memcg_register_cache_work, work);
3071 struct mem_cgroup *memcg = cw->memcg;
3072 struct kmem_cache *cachep = cw->cachep;
3074 mutex_lock(&memcg_slab_mutex);
3075 memcg_register_cache(memcg, cachep);
3076 mutex_unlock(&memcg_slab_mutex);
3078 css_put(&memcg->css);
3083 * Enqueue the creation of a per-memcg kmem_cache.
3085 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
3086 struct kmem_cache *cachep)
3088 struct memcg_register_cache_work *cw;
3090 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
3092 css_put(&memcg->css);
3097 cw->cachep = cachep;
3099 INIT_WORK(&cw->work, memcg_register_cache_func);
3100 schedule_work(&cw->work);
3103 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
3104 struct kmem_cache *cachep)
3107 * We need to stop accounting when we kmalloc, because if the
3108 * corresponding kmalloc cache is not yet created, the first allocation
3109 * in __memcg_schedule_register_cache will recurse.
3111 * However, it is better to enclose the whole function. Depending on
3112 * the debugging options enabled, INIT_WORK(), for instance, can
3113 * trigger an allocation. This too, will make us recurse. Because at
3114 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3115 * the safest choice is to do it like this, wrapping the whole function.
3117 memcg_stop_kmem_account();
3118 __memcg_schedule_register_cache(memcg, cachep);
3119 memcg_resume_kmem_account();
3122 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
3126 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp,
3127 PAGE_SIZE << order);
3129 atomic_add(1 << order, &cachep->memcg_params->nr_pages);
3133 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
3135 memcg_uncharge_kmem(cachep->memcg_params->memcg, PAGE_SIZE << order);
3136 atomic_sub(1 << order, &cachep->memcg_params->nr_pages);
3140 * Return the kmem_cache we're supposed to use for a slab allocation.
3141 * We try to use the current memcg's version of the cache.
3143 * If the cache does not exist yet, if we are the first user of it,
3144 * we either create it immediately, if possible, or create it asynchronously
3146 * In the latter case, we will let the current allocation go through with
3147 * the original cache.
3149 * Can't be called in interrupt context or from kernel threads.
3150 * This function needs to be called with rcu_read_lock() held.
3152 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3155 struct mem_cgroup *memcg;
3156 struct kmem_cache *memcg_cachep;
3158 VM_BUG_ON(!cachep->memcg_params);
3159 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3161 if (!current->mm || current->memcg_kmem_skip_account)
3165 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3167 if (!memcg_can_account_kmem(memcg))
3170 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3171 if (likely(memcg_cachep)) {
3172 cachep = memcg_cachep;
3176 /* The corresponding put will be done in the workqueue. */
3177 if (!css_tryget_online(&memcg->css))
3182 * If we are in a safe context (can wait, and not in interrupt
3183 * context), we could be be predictable and return right away.
3184 * This would guarantee that the allocation being performed
3185 * already belongs in the new cache.
3187 * However, there are some clashes that can arrive from locking.
3188 * For instance, because we acquire the slab_mutex while doing
3189 * memcg_create_kmem_cache, this means no further allocation
3190 * could happen with the slab_mutex held. So it's better to
3193 memcg_schedule_register_cache(memcg, cachep);
3201 * We need to verify if the allocation against current->mm->owner's memcg is
3202 * possible for the given order. But the page is not allocated yet, so we'll
3203 * need a further commit step to do the final arrangements.
3205 * It is possible for the task to switch cgroups in this mean time, so at
3206 * commit time, we can't rely on task conversion any longer. We'll then use
3207 * the handle argument to return to the caller which cgroup we should commit
3208 * against. We could also return the memcg directly and avoid the pointer
3209 * passing, but a boolean return value gives better semantics considering
3210 * the compiled-out case as well.
3212 * Returning true means the allocation is possible.
3215 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3217 struct mem_cgroup *memcg;
3223 * Disabling accounting is only relevant for some specific memcg
3224 * internal allocations. Therefore we would initially not have such
3225 * check here, since direct calls to the page allocator that are
3226 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3227 * outside memcg core. We are mostly concerned with cache allocations,
3228 * and by having this test at memcg_kmem_get_cache, we are already able
3229 * to relay the allocation to the root cache and bypass the memcg cache
3232 * There is one exception, though: the SLUB allocator does not create
3233 * large order caches, but rather service large kmallocs directly from
3234 * the page allocator. Therefore, the following sequence when backed by
3235 * the SLUB allocator:
3237 * memcg_stop_kmem_account();
3238 * kmalloc(<large_number>)
3239 * memcg_resume_kmem_account();
3241 * would effectively ignore the fact that we should skip accounting,
3242 * since it will drive us directly to this function without passing
3243 * through the cache selector memcg_kmem_get_cache. Such large
3244 * allocations are extremely rare but can happen, for instance, for the
3245 * cache arrays. We bring this test here.
3247 if (!current->mm || current->memcg_kmem_skip_account)
3250 memcg = get_mem_cgroup_from_mm(current->mm);
3252 if (!memcg_can_account_kmem(memcg)) {
3253 css_put(&memcg->css);
3257 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3261 css_put(&memcg->css);
3265 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3268 struct page_cgroup *pc;
3270 VM_BUG_ON(mem_cgroup_is_root(memcg));
3272 /* The page allocation failed. Revert */
3274 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3278 * The page is freshly allocated and not visible to any
3279 * outside callers yet. Set up pc non-atomically.
3281 pc = lookup_page_cgroup(page);
3282 pc->mem_cgroup = memcg;
3283 pc->flags = PCG_USED;
3286 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3288 struct mem_cgroup *memcg = NULL;
3289 struct page_cgroup *pc;
3292 pc = lookup_page_cgroup(page);
3293 if (!PageCgroupUsed(pc))
3296 memcg = pc->mem_cgroup;
3300 * We trust that only if there is a memcg associated with the page, it
3301 * is a valid allocation
3306 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3307 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3310 static inline void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3313 #endif /* CONFIG_MEMCG_KMEM */
3315 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3318 * Because tail pages are not marked as "used", set it. We're under
3319 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3320 * charge/uncharge will be never happen and move_account() is done under
3321 * compound_lock(), so we don't have to take care of races.
3323 void mem_cgroup_split_huge_fixup(struct page *head)
3325 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3326 struct page_cgroup *pc;
3327 struct mem_cgroup *memcg;
3330 if (mem_cgroup_disabled())
3333 memcg = head_pc->mem_cgroup;
3334 for (i = 1; i < HPAGE_PMD_NR; i++) {
3336 pc->mem_cgroup = memcg;
3337 pc->flags = head_pc->flags;
3339 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3342 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3345 * mem_cgroup_move_account - move account of the page
3347 * @nr_pages: number of regular pages (>1 for huge pages)
3348 * @pc: page_cgroup of the page.
3349 * @from: mem_cgroup which the page is moved from.
3350 * @to: mem_cgroup which the page is moved to. @from != @to.
3352 * The caller must confirm following.
3353 * - page is not on LRU (isolate_page() is useful.)
3354 * - compound_lock is held when nr_pages > 1
3356 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3359 static int mem_cgroup_move_account(struct page *page,
3360 unsigned int nr_pages,
3361 struct page_cgroup *pc,
3362 struct mem_cgroup *from,
3363 struct mem_cgroup *to)
3365 unsigned long flags;
3368 VM_BUG_ON(from == to);
3369 VM_BUG_ON_PAGE(PageLRU(page), page);
3371 * The page is isolated from LRU. So, collapse function
3372 * will not handle this page. But page splitting can happen.
3373 * Do this check under compound_page_lock(). The caller should
3377 if (nr_pages > 1 && !PageTransHuge(page))
3381 * Prevent mem_cgroup_migrate() from looking at pc->mem_cgroup
3382 * of its source page while we change it: page migration takes
3383 * both pages off the LRU, but page cache replacement doesn't.
3385 if (!trylock_page(page))
3389 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3392 move_lock_mem_cgroup(from, &flags);
3394 if (!PageAnon(page) && page_mapped(page)) {
3395 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3397 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3401 if (PageWriteback(page)) {
3402 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3404 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3409 * It is safe to change pc->mem_cgroup here because the page
3410 * is referenced, charged, and isolated - we can't race with
3411 * uncharging, charging, migration, or LRU putback.
3414 /* caller should have done css_get */
3415 pc->mem_cgroup = to;
3416 move_unlock_mem_cgroup(from, &flags);
3419 local_irq_disable();
3420 mem_cgroup_charge_statistics(to, page, nr_pages);
3421 memcg_check_events(to, page);
3422 mem_cgroup_charge_statistics(from, page, -nr_pages);
3423 memcg_check_events(from, page);
3432 * mem_cgroup_move_parent - moves page to the parent group
3433 * @page: the page to move
3434 * @pc: page_cgroup of the page
3435 * @child: page's cgroup
3437 * move charges to its parent or the root cgroup if the group has no
3438 * parent (aka use_hierarchy==0).
3439 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3440 * mem_cgroup_move_account fails) the failure is always temporary and
3441 * it signals a race with a page removal/uncharge or migration. In the
3442 * first case the page is on the way out and it will vanish from the LRU
3443 * on the next attempt and the call should be retried later.
3444 * Isolation from the LRU fails only if page has been isolated from
3445 * the LRU since we looked at it and that usually means either global
3446 * reclaim or migration going on. The page will either get back to the
3448 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3449 * (!PageCgroupUsed) or moved to a different group. The page will
3450 * disappear in the next attempt.
3452 static int mem_cgroup_move_parent(struct page *page,
3453 struct page_cgroup *pc,
3454 struct mem_cgroup *child)
3456 struct mem_cgroup *parent;
3457 unsigned int nr_pages;
3458 unsigned long uninitialized_var(flags);
3461 VM_BUG_ON(mem_cgroup_is_root(child));
3464 if (!get_page_unless_zero(page))
3466 if (isolate_lru_page(page))
3469 nr_pages = hpage_nr_pages(page);
3471 parent = parent_mem_cgroup(child);
3473 * If no parent, move charges to root cgroup.
3476 parent = root_mem_cgroup;
3479 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3480 flags = compound_lock_irqsave(page);
3483 ret = mem_cgroup_move_account(page, nr_pages,
3486 __mem_cgroup_cancel_local_charge(child, nr_pages);
3489 compound_unlock_irqrestore(page, flags);
3490 putback_lru_page(page);
3497 #ifdef CONFIG_MEMCG_SWAP
3498 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
3501 int val = (charge) ? 1 : -1;
3502 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
3506 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3507 * @entry: swap entry to be moved
3508 * @from: mem_cgroup which the entry is moved from
3509 * @to: mem_cgroup which the entry is moved to
3511 * It succeeds only when the swap_cgroup's record for this entry is the same
3512 * as the mem_cgroup's id of @from.
3514 * Returns 0 on success, -EINVAL on failure.
3516 * The caller must have charged to @to, IOW, called res_counter_charge() about
3517 * both res and memsw, and called css_get().
3519 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3520 struct mem_cgroup *from, struct mem_cgroup *to)
3522 unsigned short old_id, new_id;
3524 old_id = mem_cgroup_id(from);
3525 new_id = mem_cgroup_id(to);
3527 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3528 mem_cgroup_swap_statistics(from, false);
3529 mem_cgroup_swap_statistics(to, true);
3531 * This function is only called from task migration context now.
3532 * It postpones res_counter and refcount handling till the end
3533 * of task migration(mem_cgroup_clear_mc()) for performance
3534 * improvement. But we cannot postpone css_get(to) because if
3535 * the process that has been moved to @to does swap-in, the
3536 * refcount of @to might be decreased to 0.
3538 * We are in attach() phase, so the cgroup is guaranteed to be
3539 * alive, so we can just call css_get().
3547 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3548 struct mem_cgroup *from, struct mem_cgroup *to)
3554 #ifdef CONFIG_DEBUG_VM
3555 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3557 struct page_cgroup *pc;
3559 pc = lookup_page_cgroup(page);
3561 * Can be NULL while feeding pages into the page allocator for
3562 * the first time, i.e. during boot or memory hotplug;
3563 * or when mem_cgroup_disabled().
3565 if (likely(pc) && PageCgroupUsed(pc))
3570 bool mem_cgroup_bad_page_check(struct page *page)
3572 if (mem_cgroup_disabled())
3575 return lookup_page_cgroup_used(page) != NULL;
3578 void mem_cgroup_print_bad_page(struct page *page)
3580 struct page_cgroup *pc;
3582 pc = lookup_page_cgroup_used(page);
3584 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3585 pc, pc->flags, pc->mem_cgroup);
3590 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3591 unsigned long long val)
3595 int children = mem_cgroup_count_children(memcg);
3596 u64 curusage, oldusage;
3600 * For keeping hierarchical_reclaim simple, how long we should retry
3601 * is depends on callers. We set our retry-count to be function
3602 * of # of children which we should visit in this loop.
3604 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3606 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3609 while (retry_count) {
3610 if (signal_pending(current)) {
3615 * Rather than hide all in some function, I do this in
3616 * open coded manner. You see what this really does.
3617 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3619 mutex_lock(&set_limit_mutex);
3620 if (res_counter_read_u64(&memcg->memsw, RES_LIMIT) < val) {
3622 mutex_unlock(&set_limit_mutex);
3626 if (res_counter_read_u64(&memcg->res, RES_LIMIT) < val)
3629 ret = res_counter_set_limit(&memcg->res, val);
3630 mutex_unlock(&set_limit_mutex);
3635 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
3637 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3638 /* Usage is reduced ? */
3639 if (curusage >= oldusage)
3642 oldusage = curusage;
3644 if (!ret && enlarge)
3645 memcg_oom_recover(memcg);
3650 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3651 unsigned long long val)
3654 u64 oldusage, curusage;
3655 int children = mem_cgroup_count_children(memcg);
3659 /* see mem_cgroup_resize_res_limit */
3660 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3661 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3662 while (retry_count) {
3663 if (signal_pending(current)) {
3668 * Rather than hide all in some function, I do this in
3669 * open coded manner. You see what this really does.
3670 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3672 mutex_lock(&set_limit_mutex);
3673 if (res_counter_read_u64(&memcg->res, RES_LIMIT) > val) {
3675 mutex_unlock(&set_limit_mutex);
3678 if (res_counter_read_u64(&memcg->memsw, RES_LIMIT) < val)
3680 ret = res_counter_set_limit(&memcg->memsw, val);
3681 mutex_unlock(&set_limit_mutex);
3686 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3688 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3689 /* Usage is reduced ? */
3690 if (curusage >= oldusage)
3693 oldusage = curusage;
3695 if (!ret && enlarge)
3696 memcg_oom_recover(memcg);
3700 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3702 unsigned long *total_scanned)
3704 unsigned long nr_reclaimed = 0;
3705 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3706 unsigned long reclaimed;
3708 struct mem_cgroup_tree_per_zone *mctz;
3709 unsigned long long excess;
3710 unsigned long nr_scanned;
3715 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3717 * This loop can run a while, specially if mem_cgroup's continuously
3718 * keep exceeding their soft limit and putting the system under
3725 mz = mem_cgroup_largest_soft_limit_node(mctz);
3730 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3731 gfp_mask, &nr_scanned);
3732 nr_reclaimed += reclaimed;
3733 *total_scanned += nr_scanned;
3734 spin_lock_irq(&mctz->lock);
3737 * If we failed to reclaim anything from this memory cgroup
3738 * it is time to move on to the next cgroup
3744 * Loop until we find yet another one.
3746 * By the time we get the soft_limit lock
3747 * again, someone might have aded the
3748 * group back on the RB tree. Iterate to
3749 * make sure we get a different mem.
3750 * mem_cgroup_largest_soft_limit_node returns
3751 * NULL if no other cgroup is present on
3755 __mem_cgroup_largest_soft_limit_node(mctz);
3757 css_put(&next_mz->memcg->css);
3758 else /* next_mz == NULL or other memcg */
3762 __mem_cgroup_remove_exceeded(mz, mctz);
3763 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3765 * One school of thought says that we should not add
3766 * back the node to the tree if reclaim returns 0.
3767 * But our reclaim could return 0, simply because due
3768 * to priority we are exposing a smaller subset of
3769 * memory to reclaim from. Consider this as a longer
3772 /* If excess == 0, no tree ops */
3773 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3774 spin_unlock_irq(&mctz->lock);
3775 css_put(&mz->memcg->css);
3778 * Could not reclaim anything and there are no more
3779 * mem cgroups to try or we seem to be looping without
3780 * reclaiming anything.
3782 if (!nr_reclaimed &&
3784 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3786 } while (!nr_reclaimed);
3788 css_put(&next_mz->memcg->css);
3789 return nr_reclaimed;
3793 * mem_cgroup_force_empty_list - clears LRU of a group
3794 * @memcg: group to clear
3797 * @lru: lru to to clear
3799 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3800 * reclaim the pages page themselves - pages are moved to the parent (or root)
3803 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3804 int node, int zid, enum lru_list lru)
3806 struct lruvec *lruvec;
3807 unsigned long flags;
3808 struct list_head *list;
3812 zone = &NODE_DATA(node)->node_zones[zid];
3813 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
3814 list = &lruvec->lists[lru];
3818 struct page_cgroup *pc;
3821 spin_lock_irqsave(&zone->lru_lock, flags);
3822 if (list_empty(list)) {
3823 spin_unlock_irqrestore(&zone->lru_lock, flags);
3826 page = list_entry(list->prev, struct page, lru);
3828 list_move(&page->lru, list);
3830 spin_unlock_irqrestore(&zone->lru_lock, flags);
3833 spin_unlock_irqrestore(&zone->lru_lock, flags);
3835 pc = lookup_page_cgroup(page);
3837 if (mem_cgroup_move_parent(page, pc, memcg)) {
3838 /* found lock contention or "pc" is obsolete. */
3843 } while (!list_empty(list));
3847 * make mem_cgroup's charge to be 0 if there is no task by moving
3848 * all the charges and pages to the parent.
3849 * This enables deleting this mem_cgroup.
3851 * Caller is responsible for holding css reference on the memcg.
3853 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
3859 /* This is for making all *used* pages to be on LRU. */
3860 lru_add_drain_all();
3861 drain_all_stock_sync(memcg);
3862 mem_cgroup_start_move(memcg);
3863 for_each_node_state(node, N_MEMORY) {
3864 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3867 mem_cgroup_force_empty_list(memcg,
3872 mem_cgroup_end_move(memcg);
3873 memcg_oom_recover(memcg);
3877 * Kernel memory may not necessarily be trackable to a specific
3878 * process. So they are not migrated, and therefore we can't
3879 * expect their value to drop to 0 here.
3880 * Having res filled up with kmem only is enough.
3882 * This is a safety check because mem_cgroup_force_empty_list
3883 * could have raced with mem_cgroup_replace_page_cache callers
3884 * so the lru seemed empty but the page could have been added
3885 * right after the check. RES_USAGE should be safe as we always
3886 * charge before adding to the LRU.
3888 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
3889 res_counter_read_u64(&memcg->kmem, RES_USAGE);
3890 } while (usage > 0);
3894 * Test whether @memcg has children, dead or alive. Note that this
3895 * function doesn't care whether @memcg has use_hierarchy enabled and
3896 * returns %true if there are child csses according to the cgroup
3897 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3899 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3904 * The lock does not prevent addition or deletion of children, but
3905 * it prevents a new child from being initialized based on this
3906 * parent in css_online(), so it's enough to decide whether
3907 * hierarchically inherited attributes can still be changed or not.
3909 lockdep_assert_held(&memcg_create_mutex);
3912 ret = css_next_child(NULL, &memcg->css);
3918 * Reclaims as many pages from the given memcg as possible and moves
3919 * the rest to the parent.
3921 * Caller is responsible for holding css reference for memcg.
3923 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3925 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3927 /* we call try-to-free pages for make this cgroup empty */
3928 lru_add_drain_all();
3929 /* try to free all pages in this cgroup */
3930 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
3933 if (signal_pending(current))
3936 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3940 /* maybe some writeback is necessary */
3941 congestion_wait(BLK_RW_ASYNC, HZ/10);
3949 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3950 char *buf, size_t nbytes,
3953 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3955 if (mem_cgroup_is_root(memcg))
3957 return mem_cgroup_force_empty(memcg) ?: nbytes;
3960 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3963 return mem_cgroup_from_css(css)->use_hierarchy;
3966 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3967 struct cftype *cft, u64 val)
3970 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3971 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3973 mutex_lock(&memcg_create_mutex);
3975 if (memcg->use_hierarchy == val)
3979 * If parent's use_hierarchy is set, we can't make any modifications
3980 * in the child subtrees. If it is unset, then the change can
3981 * occur, provided the current cgroup has no children.
3983 * For the root cgroup, parent_mem is NULL, we allow value to be
3984 * set if there are no children.
3986 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3987 (val == 1 || val == 0)) {
3988 if (!memcg_has_children(memcg))
3989 memcg->use_hierarchy = val;
3996 mutex_unlock(&memcg_create_mutex);
4001 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4002 enum mem_cgroup_stat_index idx)
4004 struct mem_cgroup *iter;
4007 /* Per-cpu values can be negative, use a signed accumulator */
4008 for_each_mem_cgroup_tree(iter, memcg)
4009 val += mem_cgroup_read_stat(iter, idx);
4011 if (val < 0) /* race ? */
4016 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4020 if (!mem_cgroup_is_root(memcg)) {
4022 return res_counter_read_u64(&memcg->res, RES_USAGE);
4024 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4028 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4029 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4031 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4032 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4035 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4037 return val << PAGE_SHIFT;
4041 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
4044 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4045 enum res_type type = MEMFILE_TYPE(cft->private);
4046 int name = MEMFILE_ATTR(cft->private);
4050 if (name == RES_USAGE)
4051 return mem_cgroup_usage(memcg, false);
4052 return res_counter_read_u64(&memcg->res, name);
4054 if (name == RES_USAGE)
4055 return mem_cgroup_usage(memcg, true);
4056 return res_counter_read_u64(&memcg->memsw, name);
4058 return res_counter_read_u64(&memcg->kmem, name);
4065 #ifdef CONFIG_MEMCG_KMEM
4066 /* should be called with activate_kmem_mutex held */
4067 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
4068 unsigned long long limit)
4073 if (memcg_kmem_is_active(memcg))
4077 * We are going to allocate memory for data shared by all memory
4078 * cgroups so let's stop accounting here.
4080 memcg_stop_kmem_account();
4083 * For simplicity, we won't allow this to be disabled. It also can't
4084 * be changed if the cgroup has children already, or if tasks had
4087 * If tasks join before we set the limit, a person looking at
4088 * kmem.usage_in_bytes will have no way to determine when it took
4089 * place, which makes the value quite meaningless.
4091 * After it first became limited, changes in the value of the limit are
4092 * of course permitted.
4094 mutex_lock(&memcg_create_mutex);
4095 if (cgroup_has_tasks(memcg->css.cgroup) ||
4096 (memcg->use_hierarchy && memcg_has_children(memcg)))
4098 mutex_unlock(&memcg_create_mutex);
4102 memcg_id = memcg_alloc_cache_id();
4108 memcg->kmemcg_id = memcg_id;
4109 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
4112 * We couldn't have accounted to this cgroup, because it hasn't got the
4113 * active bit set yet, so this should succeed.
4115 err = res_counter_set_limit(&memcg->kmem, limit);
4118 static_key_slow_inc(&memcg_kmem_enabled_key);
4120 * Setting the active bit after enabling static branching will
4121 * guarantee no one starts accounting before all call sites are
4124 memcg_kmem_set_active(memcg);
4126 memcg_resume_kmem_account();
4130 static int memcg_activate_kmem(struct mem_cgroup *memcg,
4131 unsigned long long limit)
4135 mutex_lock(&activate_kmem_mutex);
4136 ret = __memcg_activate_kmem(memcg, limit);
4137 mutex_unlock(&activate_kmem_mutex);
4141 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4142 unsigned long long val)
4146 if (!memcg_kmem_is_active(memcg))
4147 ret = memcg_activate_kmem(memcg, val);
4149 ret = res_counter_set_limit(&memcg->kmem, val);
4153 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4156 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4161 mutex_lock(&activate_kmem_mutex);
4163 * If the parent cgroup is not kmem-active now, it cannot be activated
4164 * after this point, because it has at least one child already.
4166 if (memcg_kmem_is_active(parent))
4167 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
4168 mutex_unlock(&activate_kmem_mutex);
4172 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4173 unsigned long long val)
4177 #endif /* CONFIG_MEMCG_KMEM */
4180 * The user of this function is...
4183 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
4184 char *buf, size_t nbytes, loff_t off)
4186 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4189 unsigned long long val;
4192 buf = strstrip(buf);
4193 type = MEMFILE_TYPE(of_cft(of)->private);
4194 name = MEMFILE_ATTR(of_cft(of)->private);
4198 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4202 /* This function does all necessary parse...reuse it */
4203 ret = res_counter_memparse_write_strategy(buf, &val);
4207 ret = mem_cgroup_resize_limit(memcg, val);
4208 else if (type == _MEMSWAP)
4209 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4210 else if (type == _KMEM)
4211 ret = memcg_update_kmem_limit(memcg, val);
4215 case RES_SOFT_LIMIT:
4216 ret = res_counter_memparse_write_strategy(buf, &val);
4220 * For memsw, soft limits are hard to implement in terms
4221 * of semantics, for now, we support soft limits for
4222 * control without swap
4225 ret = res_counter_set_soft_limit(&memcg->res, val);
4230 ret = -EINVAL; /* should be BUG() ? */
4233 return ret ?: nbytes;
4236 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4237 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4239 unsigned long long min_limit, min_memsw_limit, tmp;
4241 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4242 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4243 if (!memcg->use_hierarchy)
4246 while (memcg->css.parent) {
4247 memcg = mem_cgroup_from_css(memcg->css.parent);
4248 if (!memcg->use_hierarchy)
4250 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4251 min_limit = min(min_limit, tmp);
4252 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4253 min_memsw_limit = min(min_memsw_limit, tmp);
4256 *mem_limit = min_limit;
4257 *memsw_limit = min_memsw_limit;
4260 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
4261 size_t nbytes, loff_t off)
4263 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4267 type = MEMFILE_TYPE(of_cft(of)->private);
4268 name = MEMFILE_ATTR(of_cft(of)->private);
4273 res_counter_reset_max(&memcg->res);
4274 else if (type == _MEMSWAP)
4275 res_counter_reset_max(&memcg->memsw);
4276 else if (type == _KMEM)
4277 res_counter_reset_max(&memcg->kmem);
4283 res_counter_reset_failcnt(&memcg->res);
4284 else if (type == _MEMSWAP)
4285 res_counter_reset_failcnt(&memcg->memsw);
4286 else if (type == _KMEM)
4287 res_counter_reset_failcnt(&memcg->kmem);
4296 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
4299 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
4303 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4304 struct cftype *cft, u64 val)
4306 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4308 if (val >= (1 << NR_MOVE_TYPE))
4312 * No kind of locking is needed in here, because ->can_attach() will
4313 * check this value once in the beginning of the process, and then carry
4314 * on with stale data. This means that changes to this value will only
4315 * affect task migrations starting after the change.
4317 memcg->move_charge_at_immigrate = val;
4321 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
4322 struct cftype *cft, u64 val)
4329 static int memcg_numa_stat_show(struct seq_file *m, void *v)
4333 unsigned int lru_mask;
4336 static const struct numa_stat stats[] = {
4337 { "total", LRU_ALL },
4338 { "file", LRU_ALL_FILE },
4339 { "anon", LRU_ALL_ANON },
4340 { "unevictable", BIT(LRU_UNEVICTABLE) },
4342 const struct numa_stat *stat;
4345 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4347 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4348 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
4349 seq_printf(m, "%s=%lu", stat->name, nr);
4350 for_each_node_state(nid, N_MEMORY) {
4351 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4353 seq_printf(m, " N%d=%lu", nid, nr);
4358 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
4359 struct mem_cgroup *iter;
4362 for_each_mem_cgroup_tree(iter, memcg)
4363 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
4364 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
4365 for_each_node_state(nid, N_MEMORY) {
4367 for_each_mem_cgroup_tree(iter, memcg)
4368 nr += mem_cgroup_node_nr_lru_pages(
4369 iter, nid, stat->lru_mask);
4370 seq_printf(m, " N%d=%lu", nid, nr);
4377 #endif /* CONFIG_NUMA */
4379 static inline void mem_cgroup_lru_names_not_uptodate(void)
4381 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4384 static int memcg_stat_show(struct seq_file *m, void *v)
4386 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
4387 struct mem_cgroup *mi;
4390 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4391 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4393 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4394 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4397 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4398 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4399 mem_cgroup_read_events(memcg, i));
4401 for (i = 0; i < NR_LRU_LISTS; i++)
4402 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4403 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4405 /* Hierarchical information */
4407 unsigned long long limit, memsw_limit;
4408 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4409 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
4410 if (do_swap_account)
4411 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4415 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4418 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4420 for_each_mem_cgroup_tree(mi, memcg)
4421 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4422 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4425 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4426 unsigned long long val = 0;
4428 for_each_mem_cgroup_tree(mi, memcg)
4429 val += mem_cgroup_read_events(mi, i);
4430 seq_printf(m, "total_%s %llu\n",
4431 mem_cgroup_events_names[i], val);
4434 for (i = 0; i < NR_LRU_LISTS; i++) {
4435 unsigned long long val = 0;
4437 for_each_mem_cgroup_tree(mi, memcg)
4438 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4439 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4442 #ifdef CONFIG_DEBUG_VM
4445 struct mem_cgroup_per_zone *mz;
4446 struct zone_reclaim_stat *rstat;
4447 unsigned long recent_rotated[2] = {0, 0};
4448 unsigned long recent_scanned[2] = {0, 0};
4450 for_each_online_node(nid)
4451 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4452 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
4453 rstat = &mz->lruvec.reclaim_stat;
4455 recent_rotated[0] += rstat->recent_rotated[0];
4456 recent_rotated[1] += rstat->recent_rotated[1];
4457 recent_scanned[0] += rstat->recent_scanned[0];
4458 recent_scanned[1] += rstat->recent_scanned[1];
4460 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4461 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4462 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4463 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4470 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
4473 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4475 return mem_cgroup_swappiness(memcg);
4478 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
4479 struct cftype *cft, u64 val)
4481 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4487 memcg->swappiness = val;
4489 vm_swappiness = val;
4494 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4496 struct mem_cgroup_threshold_ary *t;
4502 t = rcu_dereference(memcg->thresholds.primary);
4504 t = rcu_dereference(memcg->memsw_thresholds.primary);
4509 usage = mem_cgroup_usage(memcg, swap);
4512 * current_threshold points to threshold just below or equal to usage.
4513 * If it's not true, a threshold was crossed after last
4514 * call of __mem_cgroup_threshold().
4516 i = t->current_threshold;
4519 * Iterate backward over array of thresholds starting from
4520 * current_threshold and check if a threshold is crossed.
4521 * If none of thresholds below usage is crossed, we read
4522 * only one element of the array here.
4524 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4525 eventfd_signal(t->entries[i].eventfd, 1);
4527 /* i = current_threshold + 1 */
4531 * Iterate forward over array of thresholds starting from
4532 * current_threshold+1 and check if a threshold is crossed.
4533 * If none of thresholds above usage is crossed, we read
4534 * only one element of the array here.
4536 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4537 eventfd_signal(t->entries[i].eventfd, 1);
4539 /* Update current_threshold */
4540 t->current_threshold = i - 1;
4545 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4548 __mem_cgroup_threshold(memcg, false);
4549 if (do_swap_account)
4550 __mem_cgroup_threshold(memcg, true);
4552 memcg = parent_mem_cgroup(memcg);
4556 static int compare_thresholds(const void *a, const void *b)
4558 const struct mem_cgroup_threshold *_a = a;
4559 const struct mem_cgroup_threshold *_b = b;
4561 if (_a->threshold > _b->threshold)
4564 if (_a->threshold < _b->threshold)
4570 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4572 struct mem_cgroup_eventfd_list *ev;
4574 spin_lock(&memcg_oom_lock);
4576 list_for_each_entry(ev, &memcg->oom_notify, list)
4577 eventfd_signal(ev->eventfd, 1);
4579 spin_unlock(&memcg_oom_lock);
4583 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4585 struct mem_cgroup *iter;
4587 for_each_mem_cgroup_tree(iter, memcg)
4588 mem_cgroup_oom_notify_cb(iter);
4591 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4592 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
4594 struct mem_cgroup_thresholds *thresholds;
4595 struct mem_cgroup_threshold_ary *new;
4596 u64 threshold, usage;
4599 ret = res_counter_memparse_write_strategy(args, &threshold);
4603 mutex_lock(&memcg->thresholds_lock);
4606 thresholds = &memcg->thresholds;
4607 usage = mem_cgroup_usage(memcg, false);
4608 } else if (type == _MEMSWAP) {
4609 thresholds = &memcg->memsw_thresholds;
4610 usage = mem_cgroup_usage(memcg, true);
4614 /* Check if a threshold crossed before adding a new one */
4615 if (thresholds->primary)
4616 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4618 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4620 /* Allocate memory for new array of thresholds */
4621 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4629 /* Copy thresholds (if any) to new array */
4630 if (thresholds->primary) {
4631 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4632 sizeof(struct mem_cgroup_threshold));
4635 /* Add new threshold */
4636 new->entries[size - 1].eventfd = eventfd;
4637 new->entries[size - 1].threshold = threshold;
4639 /* Sort thresholds. Registering of new threshold isn't time-critical */
4640 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4641 compare_thresholds, NULL);
4643 /* Find current threshold */
4644 new->current_threshold = -1;
4645 for (i = 0; i < size; i++) {
4646 if (new->entries[i].threshold <= usage) {
4648 * new->current_threshold will not be used until
4649 * rcu_assign_pointer(), so it's safe to increment
4652 ++new->current_threshold;
4657 /* Free old spare buffer and save old primary buffer as spare */
4658 kfree(thresholds->spare);
4659 thresholds->spare = thresholds->primary;
4661 rcu_assign_pointer(thresholds->primary, new);
4663 /* To be sure that nobody uses thresholds */
4667 mutex_unlock(&memcg->thresholds_lock);
4672 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4673 struct eventfd_ctx *eventfd, const char *args)
4675 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4678 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4679 struct eventfd_ctx *eventfd, const char *args)
4681 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4684 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4685 struct eventfd_ctx *eventfd, enum res_type type)
4687 struct mem_cgroup_thresholds *thresholds;
4688 struct mem_cgroup_threshold_ary *new;
4692 mutex_lock(&memcg->thresholds_lock);
4695 thresholds = &memcg->thresholds;
4696 usage = mem_cgroup_usage(memcg, false);
4697 } else if (type == _MEMSWAP) {
4698 thresholds = &memcg->memsw_thresholds;
4699 usage = mem_cgroup_usage(memcg, true);
4703 if (!thresholds->primary)
4706 /* Check if a threshold crossed before removing */
4707 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4709 /* Calculate new number of threshold */
4711 for (i = 0; i < thresholds->primary->size; i++) {
4712 if (thresholds->primary->entries[i].eventfd != eventfd)
4716 new = thresholds->spare;
4718 /* Set thresholds array to NULL if we don't have thresholds */
4727 /* Copy thresholds and find current threshold */
4728 new->current_threshold = -1;
4729 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4730 if (thresholds->primary->entries[i].eventfd == eventfd)
4733 new->entries[j] = thresholds->primary->entries[i];
4734 if (new->entries[j].threshold <= usage) {
4736 * new->current_threshold will not be used
4737 * until rcu_assign_pointer(), so it's safe to increment
4740 ++new->current_threshold;
4746 /* Swap primary and spare array */
4747 thresholds->spare = thresholds->primary;
4748 /* If all events are unregistered, free the spare array */
4750 kfree(thresholds->spare);
4751 thresholds->spare = NULL;
4754 rcu_assign_pointer(thresholds->primary, new);
4756 /* To be sure that nobody uses thresholds */
4759 mutex_unlock(&memcg->thresholds_lock);
4762 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4763 struct eventfd_ctx *eventfd)
4765 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4768 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4769 struct eventfd_ctx *eventfd)
4771 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4774 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4775 struct eventfd_ctx *eventfd, const char *args)
4777 struct mem_cgroup_eventfd_list *event;
4779 event = kmalloc(sizeof(*event), GFP_KERNEL);
4783 spin_lock(&memcg_oom_lock);
4785 event->eventfd = eventfd;
4786 list_add(&event->list, &memcg->oom_notify);
4788 /* already in OOM ? */
4789 if (atomic_read(&memcg->under_oom))
4790 eventfd_signal(eventfd, 1);
4791 spin_unlock(&memcg_oom_lock);
4796 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4797 struct eventfd_ctx *eventfd)
4799 struct mem_cgroup_eventfd_list *ev, *tmp;
4801 spin_lock(&memcg_oom_lock);
4803 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4804 if (ev->eventfd == eventfd) {
4805 list_del(&ev->list);
4810 spin_unlock(&memcg_oom_lock);
4813 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4815 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
4817 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4818 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
4822 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4823 struct cftype *cft, u64 val)
4825 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4827 /* cannot set to root cgroup and only 0 and 1 are allowed */
4828 if (!css->parent || !((val == 0) || (val == 1)))
4831 memcg->oom_kill_disable = val;
4833 memcg_oom_recover(memcg);
4838 #ifdef CONFIG_MEMCG_KMEM
4839 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4843 memcg->kmemcg_id = -1;
4844 ret = memcg_propagate_kmem(memcg);
4848 return mem_cgroup_sockets_init(memcg, ss);
4851 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4853 mem_cgroup_sockets_destroy(memcg);
4856 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
4858 if (!memcg_kmem_is_active(memcg))
4862 * kmem charges can outlive the cgroup. In the case of slab
4863 * pages, for instance, a page contain objects from various
4864 * processes. As we prevent from taking a reference for every
4865 * such allocation we have to be careful when doing uncharge
4866 * (see memcg_uncharge_kmem) and here during offlining.
4868 * The idea is that that only the _last_ uncharge which sees
4869 * the dead memcg will drop the last reference. An additional
4870 * reference is taken here before the group is marked dead
4871 * which is then paired with css_put during uncharge resp. here.
4873 * Although this might sound strange as this path is called from
4874 * css_offline() when the referencemight have dropped down to 0 and
4875 * shouldn't be incremented anymore (css_tryget_online() would
4876 * fail) we do not have other options because of the kmem
4877 * allocations lifetime.
4879 css_get(&memcg->css);
4881 memcg_kmem_mark_dead(memcg);
4883 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
4886 if (memcg_kmem_test_and_clear_dead(memcg))
4887 css_put(&memcg->css);
4890 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4895 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4899 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
4905 * DO NOT USE IN NEW FILES.
4907 * "cgroup.event_control" implementation.
4909 * This is way over-engineered. It tries to support fully configurable
4910 * events for each user. Such level of flexibility is completely
4911 * unnecessary especially in the light of the planned unified hierarchy.
4913 * Please deprecate this and replace with something simpler if at all
4918 * Unregister event and free resources.
4920 * Gets called from workqueue.
4922 static void memcg_event_remove(struct work_struct *work)
4924 struct mem_cgroup_event *event =
4925 container_of(work, struct mem_cgroup_event, remove);
4926 struct mem_cgroup *memcg = event->memcg;
4928 remove_wait_queue(event->wqh, &event->wait);
4930 event->unregister_event(memcg, event->eventfd);
4932 /* Notify userspace the event is going away. */
4933 eventfd_signal(event->eventfd, 1);
4935 eventfd_ctx_put(event->eventfd);
4937 css_put(&memcg->css);
4941 * Gets called on POLLHUP on eventfd when user closes it.
4943 * Called with wqh->lock held and interrupts disabled.
4945 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4946 int sync, void *key)
4948 struct mem_cgroup_event *event =
4949 container_of(wait, struct mem_cgroup_event, wait);
4950 struct mem_cgroup *memcg = event->memcg;
4951 unsigned long flags = (unsigned long)key;
4953 if (flags & POLLHUP) {
4955 * If the event has been detached at cgroup removal, we
4956 * can simply return knowing the other side will cleanup
4959 * We can't race against event freeing since the other
4960 * side will require wqh->lock via remove_wait_queue(),
4963 spin_lock(&memcg->event_list_lock);
4964 if (!list_empty(&event->list)) {
4965 list_del_init(&event->list);
4967 * We are in atomic context, but cgroup_event_remove()
4968 * may sleep, so we have to call it in workqueue.
4970 schedule_work(&event->remove);
4972 spin_unlock(&memcg->event_list_lock);
4978 static void memcg_event_ptable_queue_proc(struct file *file,
4979 wait_queue_head_t *wqh, poll_table *pt)
4981 struct mem_cgroup_event *event =
4982 container_of(pt, struct mem_cgroup_event, pt);
4985 add_wait_queue(wqh, &event->wait);
4989 * DO NOT USE IN NEW FILES.
4991 * Parse input and register new cgroup event handler.
4993 * Input must be in format '<event_fd> <control_fd> <args>'.
4994 * Interpretation of args is defined by control file implementation.
4996 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4997 char *buf, size_t nbytes, loff_t off)
4999 struct cgroup_subsys_state *css = of_css(of);
5000 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5001 struct mem_cgroup_event *event;
5002 struct cgroup_subsys_state *cfile_css;
5003 unsigned int efd, cfd;
5010 buf = strstrip(buf);
5012 efd = simple_strtoul(buf, &endp, 10);
5017 cfd = simple_strtoul(buf, &endp, 10);
5018 if ((*endp != ' ') && (*endp != '\0'))
5022 event = kzalloc(sizeof(*event), GFP_KERNEL);
5026 event->memcg = memcg;
5027 INIT_LIST_HEAD(&event->list);
5028 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
5029 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
5030 INIT_WORK(&event->remove, memcg_event_remove);
5038 event->eventfd = eventfd_ctx_fileget(efile.file);
5039 if (IS_ERR(event->eventfd)) {
5040 ret = PTR_ERR(event->eventfd);
5047 goto out_put_eventfd;
5050 /* the process need read permission on control file */
5051 /* AV: shouldn't we check that it's been opened for read instead? */
5052 ret = inode_permission(file_inode(cfile.file), MAY_READ);
5057 * Determine the event callbacks and set them in @event. This used
5058 * to be done via struct cftype but cgroup core no longer knows
5059 * about these events. The following is crude but the whole thing
5060 * is for compatibility anyway.
5062 * DO NOT ADD NEW FILES.
5064 name = cfile.file->f_dentry->d_name.name;
5066 if (!strcmp(name, "memory.usage_in_bytes")) {
5067 event->register_event = mem_cgroup_usage_register_event;
5068 event->unregister_event = mem_cgroup_usage_unregister_event;
5069 } else if (!strcmp(name, "memory.oom_control")) {
5070 event->register_event = mem_cgroup_oom_register_event;
5071 event->unregister_event = mem_cgroup_oom_unregister_event;
5072 } else if (!strcmp(name, "memory.pressure_level")) {
5073 event->register_event = vmpressure_register_event;
5074 event->unregister_event = vmpressure_unregister_event;
5075 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
5076 event->register_event = memsw_cgroup_usage_register_event;
5077 event->unregister_event = memsw_cgroup_usage_unregister_event;
5084 * Verify @cfile should belong to @css. Also, remaining events are
5085 * automatically removed on cgroup destruction but the removal is
5086 * asynchronous, so take an extra ref on @css.
5088 cfile_css = css_tryget_online_from_dir(cfile.file->f_dentry->d_parent,
5089 &memory_cgrp_subsys);
5091 if (IS_ERR(cfile_css))
5093 if (cfile_css != css) {
5098 ret = event->register_event(memcg, event->eventfd, buf);
5102 efile.file->f_op->poll(efile.file, &event->pt);
5104 spin_lock(&memcg->event_list_lock);
5105 list_add(&event->list, &memcg->event_list);
5106 spin_unlock(&memcg->event_list_lock);
5118 eventfd_ctx_put(event->eventfd);
5127 static struct cftype mem_cgroup_files[] = {
5129 .name = "usage_in_bytes",
5130 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5131 .read_u64 = mem_cgroup_read_u64,
5134 .name = "max_usage_in_bytes",
5135 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5136 .write = mem_cgroup_reset,
5137 .read_u64 = mem_cgroup_read_u64,
5140 .name = "limit_in_bytes",
5141 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5142 .write = mem_cgroup_write,
5143 .read_u64 = mem_cgroup_read_u64,
5146 .name = "soft_limit_in_bytes",
5147 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5148 .write = mem_cgroup_write,
5149 .read_u64 = mem_cgroup_read_u64,
5153 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5154 .write = mem_cgroup_reset,
5155 .read_u64 = mem_cgroup_read_u64,
5159 .seq_show = memcg_stat_show,
5162 .name = "force_empty",
5163 .write = mem_cgroup_force_empty_write,
5166 .name = "use_hierarchy",
5167 .write_u64 = mem_cgroup_hierarchy_write,
5168 .read_u64 = mem_cgroup_hierarchy_read,
5171 .name = "cgroup.event_control", /* XXX: for compat */
5172 .write = memcg_write_event_control,
5173 .flags = CFTYPE_NO_PREFIX,
5177 .name = "swappiness",
5178 .read_u64 = mem_cgroup_swappiness_read,
5179 .write_u64 = mem_cgroup_swappiness_write,
5182 .name = "move_charge_at_immigrate",
5183 .read_u64 = mem_cgroup_move_charge_read,
5184 .write_u64 = mem_cgroup_move_charge_write,
5187 .name = "oom_control",
5188 .seq_show = mem_cgroup_oom_control_read,
5189 .write_u64 = mem_cgroup_oom_control_write,
5190 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5193 .name = "pressure_level",
5197 .name = "numa_stat",
5198 .seq_show = memcg_numa_stat_show,
5201 #ifdef CONFIG_MEMCG_KMEM
5203 .name = "kmem.limit_in_bytes",
5204 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5205 .write = mem_cgroup_write,
5206 .read_u64 = mem_cgroup_read_u64,
5209 .name = "kmem.usage_in_bytes",
5210 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5211 .read_u64 = mem_cgroup_read_u64,
5214 .name = "kmem.failcnt",
5215 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5216 .write = mem_cgroup_reset,
5217 .read_u64 = mem_cgroup_read_u64,
5220 .name = "kmem.max_usage_in_bytes",
5221 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5222 .write = mem_cgroup_reset,
5223 .read_u64 = mem_cgroup_read_u64,
5225 #ifdef CONFIG_SLABINFO
5227 .name = "kmem.slabinfo",
5228 .seq_show = mem_cgroup_slabinfo_read,
5232 { }, /* terminate */
5235 #ifdef CONFIG_MEMCG_SWAP
5236 static struct cftype memsw_cgroup_files[] = {
5238 .name = "memsw.usage_in_bytes",
5239 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5240 .read_u64 = mem_cgroup_read_u64,
5243 .name = "memsw.max_usage_in_bytes",
5244 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5245 .write = mem_cgroup_reset,
5246 .read_u64 = mem_cgroup_read_u64,
5249 .name = "memsw.limit_in_bytes",
5250 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5251 .write = mem_cgroup_write,
5252 .read_u64 = mem_cgroup_read_u64,
5255 .name = "memsw.failcnt",
5256 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5257 .write = mem_cgroup_reset,
5258 .read_u64 = mem_cgroup_read_u64,
5260 { }, /* terminate */
5263 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5265 struct mem_cgroup_per_node *pn;
5266 struct mem_cgroup_per_zone *mz;
5267 int zone, tmp = node;
5269 * This routine is called against possible nodes.
5270 * But it's BUG to call kmalloc() against offline node.
5272 * TODO: this routine can waste much memory for nodes which will
5273 * never be onlined. It's better to use memory hotplug callback
5276 if (!node_state(node, N_NORMAL_MEMORY))
5278 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5282 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5283 mz = &pn->zoneinfo[zone];
5284 lruvec_init(&mz->lruvec);
5285 mz->usage_in_excess = 0;
5286 mz->on_tree = false;
5289 memcg->nodeinfo[node] = pn;
5293 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5295 kfree(memcg->nodeinfo[node]);
5298 static struct mem_cgroup *mem_cgroup_alloc(void)
5300 struct mem_cgroup *memcg;
5303 size = sizeof(struct mem_cgroup);
5304 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
5306 memcg = kzalloc(size, GFP_KERNEL);
5310 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5313 spin_lock_init(&memcg->pcp_counter_lock);
5322 * At destroying mem_cgroup, references from swap_cgroup can remain.
5323 * (scanning all at force_empty is too costly...)
5325 * Instead of clearing all references at force_empty, we remember
5326 * the number of reference from swap_cgroup and free mem_cgroup when
5327 * it goes down to 0.
5329 * Removal of cgroup itself succeeds regardless of refs from swap.
5332 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5336 mem_cgroup_remove_from_trees(memcg);
5339 free_mem_cgroup_per_zone_info(memcg, node);
5341 free_percpu(memcg->stat);
5344 * We need to make sure that (at least for now), the jump label
5345 * destruction code runs outside of the cgroup lock. This is because
5346 * get_online_cpus(), which is called from the static_branch update,
5347 * can't be called inside the cgroup_lock. cpusets are the ones
5348 * enforcing this dependency, so if they ever change, we might as well.
5350 * schedule_work() will guarantee this happens. Be careful if you need
5351 * to move this code around, and make sure it is outside
5354 disarm_static_keys(memcg);
5359 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5361 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5363 if (!memcg->res.parent)
5365 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5367 EXPORT_SYMBOL(parent_mem_cgroup);
5369 static void __init mem_cgroup_soft_limit_tree_init(void)
5371 struct mem_cgroup_tree_per_node *rtpn;
5372 struct mem_cgroup_tree_per_zone *rtpz;
5373 int tmp, node, zone;
5375 for_each_node(node) {
5377 if (!node_state(node, N_NORMAL_MEMORY))
5379 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5382 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5384 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5385 rtpz = &rtpn->rb_tree_per_zone[zone];
5386 rtpz->rb_root = RB_ROOT;
5387 spin_lock_init(&rtpz->lock);
5392 static struct cgroup_subsys_state * __ref
5393 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5395 struct mem_cgroup *memcg;
5396 long error = -ENOMEM;
5399 memcg = mem_cgroup_alloc();
5401 return ERR_PTR(error);
5404 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5408 if (parent_css == NULL) {
5409 root_mem_cgroup = memcg;
5410 res_counter_init(&memcg->res, NULL);
5411 res_counter_init(&memcg->memsw, NULL);
5412 res_counter_init(&memcg->kmem, NULL);
5415 memcg->last_scanned_node = MAX_NUMNODES;
5416 INIT_LIST_HEAD(&memcg->oom_notify);
5417 memcg->move_charge_at_immigrate = 0;
5418 mutex_init(&memcg->thresholds_lock);
5419 spin_lock_init(&memcg->move_lock);
5420 vmpressure_init(&memcg->vmpressure);
5421 INIT_LIST_HEAD(&memcg->event_list);
5422 spin_lock_init(&memcg->event_list_lock);
5427 __mem_cgroup_free(memcg);
5428 return ERR_PTR(error);
5432 mem_cgroup_css_online(struct cgroup_subsys_state *css)
5434 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5435 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
5438 if (css->id > MEM_CGROUP_ID_MAX)
5444 mutex_lock(&memcg_create_mutex);
5446 memcg->use_hierarchy = parent->use_hierarchy;
5447 memcg->oom_kill_disable = parent->oom_kill_disable;
5448 memcg->swappiness = mem_cgroup_swappiness(parent);
5450 if (parent->use_hierarchy) {
5451 res_counter_init(&memcg->res, &parent->res);
5452 res_counter_init(&memcg->memsw, &parent->memsw);
5453 res_counter_init(&memcg->kmem, &parent->kmem);
5456 * No need to take a reference to the parent because cgroup
5457 * core guarantees its existence.
5460 res_counter_init(&memcg->res, NULL);
5461 res_counter_init(&memcg->memsw, NULL);
5462 res_counter_init(&memcg->kmem, NULL);
5464 * Deeper hierachy with use_hierarchy == false doesn't make
5465 * much sense so let cgroup subsystem know about this
5466 * unfortunate state in our controller.
5468 if (parent != root_mem_cgroup)
5469 memory_cgrp_subsys.broken_hierarchy = true;
5471 mutex_unlock(&memcg_create_mutex);
5473 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
5478 * Make sure the memcg is initialized: mem_cgroup_iter()
5479 * orders reading memcg->initialized against its callers
5480 * reading the memcg members.
5482 smp_store_release(&memcg->initialized, 1);
5488 * Announce all parents that a group from their hierarchy is gone.
5490 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
5492 struct mem_cgroup *parent = memcg;
5494 while ((parent = parent_mem_cgroup(parent)))
5495 mem_cgroup_iter_invalidate(parent);
5498 * if the root memcg is not hierarchical we have to check it
5501 if (!root_mem_cgroup->use_hierarchy)
5502 mem_cgroup_iter_invalidate(root_mem_cgroup);
5505 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5507 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5508 struct mem_cgroup_event *event, *tmp;
5509 struct cgroup_subsys_state *iter;
5512 * Unregister events and notify userspace.
5513 * Notify userspace about cgroup removing only after rmdir of cgroup
5514 * directory to avoid race between userspace and kernelspace.
5516 spin_lock(&memcg->event_list_lock);
5517 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
5518 list_del_init(&event->list);
5519 schedule_work(&event->remove);
5521 spin_unlock(&memcg->event_list_lock);
5523 kmem_cgroup_css_offline(memcg);
5525 mem_cgroup_invalidate_reclaim_iterators(memcg);
5528 * This requires that offlining is serialized. Right now that is
5529 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
5531 css_for_each_descendant_post(iter, css)
5532 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
5534 memcg_unregister_all_caches(memcg);
5535 vmpressure_cleanup(&memcg->vmpressure);
5538 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5540 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5542 * XXX: css_offline() would be where we should reparent all
5543 * memory to prepare the cgroup for destruction. However,
5544 * memcg does not do css_tryget_online() and res_counter charging
5545 * under the same RCU lock region, which means that charging
5546 * could race with offlining. Offlining only happens to
5547 * cgroups with no tasks in them but charges can show up
5548 * without any tasks from the swapin path when the target
5549 * memcg is looked up from the swapout record and not from the
5550 * current task as it usually is. A race like this can leak
5551 * charges and put pages with stale cgroup pointers into
5555 * lookup_swap_cgroup_id()
5557 * mem_cgroup_lookup()
5558 * css_tryget_online()
5560 * disable css_tryget_online()
5563 * reparent_charges()
5564 * res_counter_charge()
5567 * pc->mem_cgroup = dead memcg
5570 * The bulk of the charges are still moved in offline_css() to
5571 * avoid pinning a lot of pages in case a long-term reference
5572 * like a swapout record is deferring the css_free() to long
5573 * after offlining. But this makes sure we catch any charges
5574 * made after offlining:
5576 mem_cgroup_reparent_charges(memcg);
5578 memcg_destroy_kmem(memcg);
5579 __mem_cgroup_free(memcg);
5583 * mem_cgroup_css_reset - reset the states of a mem_cgroup
5584 * @css: the target css
5586 * Reset the states of the mem_cgroup associated with @css. This is
5587 * invoked when the userland requests disabling on the default hierarchy
5588 * but the memcg is pinned through dependency. The memcg should stop
5589 * applying policies and should revert to the vanilla state as it may be
5590 * made visible again.
5592 * The current implementation only resets the essential configurations.
5593 * This needs to be expanded to cover all the visible parts.
5595 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
5597 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5599 mem_cgroup_resize_limit(memcg, ULLONG_MAX);
5600 mem_cgroup_resize_memsw_limit(memcg, ULLONG_MAX);
5601 memcg_update_kmem_limit(memcg, ULLONG_MAX);
5602 res_counter_set_soft_limit(&memcg->res, ULLONG_MAX);
5606 /* Handlers for move charge at task migration. */
5607 static int mem_cgroup_do_precharge(unsigned long count)
5611 /* Try a single bulk charge without reclaim first */
5612 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
5614 mc.precharge += count;
5617 if (ret == -EINTR) {
5618 cancel_charge(root_mem_cgroup, count);
5622 /* Try charges one by one with reclaim */
5624 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
5626 * In case of failure, any residual charges against
5627 * mc.to will be dropped by mem_cgroup_clear_mc()
5628 * later on. However, cancel any charges that are
5629 * bypassed to root right away or they'll be lost.
5632 cancel_charge(root_mem_cgroup, 1);
5642 * get_mctgt_type - get target type of moving charge
5643 * @vma: the vma the pte to be checked belongs
5644 * @addr: the address corresponding to the pte to be checked
5645 * @ptent: the pte to be checked
5646 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5649 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5650 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5651 * move charge. if @target is not NULL, the page is stored in target->page
5652 * with extra refcnt got(Callers should handle it).
5653 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5654 * target for charge migration. if @target is not NULL, the entry is stored
5657 * Called with pte lock held.
5664 enum mc_target_type {
5670 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5671 unsigned long addr, pte_t ptent)
5673 struct page *page = vm_normal_page(vma, addr, ptent);
5675 if (!page || !page_mapped(page))
5677 if (PageAnon(page)) {
5678 /* we don't move shared anon */
5681 } else if (!move_file())
5682 /* we ignore mapcount for file pages */
5684 if (!get_page_unless_zero(page))
5691 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5692 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5694 struct page *page = NULL;
5695 swp_entry_t ent = pte_to_swp_entry(ptent);
5697 if (!move_anon() || non_swap_entry(ent))
5700 * Because lookup_swap_cache() updates some statistics counter,
5701 * we call find_get_page() with swapper_space directly.
5703 page = find_get_page(swap_address_space(ent), ent.val);
5704 if (do_swap_account)
5705 entry->val = ent.val;
5710 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5711 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5717 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5718 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5720 struct page *page = NULL;
5721 struct address_space *mapping;
5724 if (!vma->vm_file) /* anonymous vma */
5729 mapping = vma->vm_file->f_mapping;
5730 if (pte_none(ptent))
5731 pgoff = linear_page_index(vma, addr);
5732 else /* pte_file(ptent) is true */
5733 pgoff = pte_to_pgoff(ptent);
5735 /* page is moved even if it's not RSS of this task(page-faulted). */
5737 /* shmem/tmpfs may report page out on swap: account for that too. */
5738 if (shmem_mapping(mapping)) {
5739 page = find_get_entry(mapping, pgoff);
5740 if (radix_tree_exceptional_entry(page)) {
5741 swp_entry_t swp = radix_to_swp_entry(page);
5742 if (do_swap_account)
5744 page = find_get_page(swap_address_space(swp), swp.val);
5747 page = find_get_page(mapping, pgoff);
5749 page = find_get_page(mapping, pgoff);
5754 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5755 unsigned long addr, pte_t ptent, union mc_target *target)
5757 struct page *page = NULL;
5758 struct page_cgroup *pc;
5759 enum mc_target_type ret = MC_TARGET_NONE;
5760 swp_entry_t ent = { .val = 0 };
5762 if (pte_present(ptent))
5763 page = mc_handle_present_pte(vma, addr, ptent);
5764 else if (is_swap_pte(ptent))
5765 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5766 else if (pte_none(ptent) || pte_file(ptent))
5767 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5769 if (!page && !ent.val)
5772 pc = lookup_page_cgroup(page);
5774 * Do only loose check w/o serialization.
5775 * mem_cgroup_move_account() checks the pc is valid or
5776 * not under LRU exclusion.
5778 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5779 ret = MC_TARGET_PAGE;
5781 target->page = page;
5783 if (!ret || !target)
5786 /* There is a swap entry and a page doesn't exist or isn't charged */
5787 if (ent.val && !ret &&
5788 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5789 ret = MC_TARGET_SWAP;
5796 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5798 * We don't consider swapping or file mapped pages because THP does not
5799 * support them for now.
5800 * Caller should make sure that pmd_trans_huge(pmd) is true.
5802 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5803 unsigned long addr, pmd_t pmd, union mc_target *target)
5805 struct page *page = NULL;
5806 struct page_cgroup *pc;
5807 enum mc_target_type ret = MC_TARGET_NONE;
5809 page = pmd_page(pmd);
5810 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5813 pc = lookup_page_cgroup(page);
5814 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5815 ret = MC_TARGET_PAGE;
5818 target->page = page;
5824 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5825 unsigned long addr, pmd_t pmd, union mc_target *target)
5827 return MC_TARGET_NONE;
5831 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5832 unsigned long addr, unsigned long end,
5833 struct mm_walk *walk)
5835 struct vm_area_struct *vma = walk->private;
5839 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5840 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5841 mc.precharge += HPAGE_PMD_NR;
5846 if (pmd_trans_unstable(pmd))
5848 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5849 for (; addr != end; pte++, addr += PAGE_SIZE)
5850 if (get_mctgt_type(vma, addr, *pte, NULL))
5851 mc.precharge++; /* increment precharge temporarily */
5852 pte_unmap_unlock(pte - 1, ptl);
5858 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5860 unsigned long precharge;
5861 struct vm_area_struct *vma;
5863 down_read(&mm->mmap_sem);
5864 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5865 struct mm_walk mem_cgroup_count_precharge_walk = {
5866 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5870 if (is_vm_hugetlb_page(vma))
5872 walk_page_range(vma->vm_start, vma->vm_end,
5873 &mem_cgroup_count_precharge_walk);
5875 up_read(&mm->mmap_sem);
5877 precharge = mc.precharge;
5883 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5885 unsigned long precharge = mem_cgroup_count_precharge(mm);
5887 VM_BUG_ON(mc.moving_task);
5888 mc.moving_task = current;
5889 return mem_cgroup_do_precharge(precharge);
5892 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5893 static void __mem_cgroup_clear_mc(void)
5895 struct mem_cgroup *from = mc.from;
5896 struct mem_cgroup *to = mc.to;
5899 /* we must uncharge all the leftover precharges from mc.to */
5901 cancel_charge(mc.to, mc.precharge);
5905 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5906 * we must uncharge here.
5908 if (mc.moved_charge) {
5909 cancel_charge(mc.from, mc.moved_charge);
5910 mc.moved_charge = 0;
5912 /* we must fixup refcnts and charges */
5913 if (mc.moved_swap) {
5914 /* uncharge swap account from the old cgroup */
5915 if (!mem_cgroup_is_root(mc.from))
5916 res_counter_uncharge(&mc.from->memsw,
5917 PAGE_SIZE * mc.moved_swap);
5919 for (i = 0; i < mc.moved_swap; i++)
5920 css_put(&mc.from->css);
5923 * we charged both to->res and to->memsw, so we should
5926 if (!mem_cgroup_is_root(mc.to))
5927 res_counter_uncharge(&mc.to->res,
5928 PAGE_SIZE * mc.moved_swap);
5929 /* we've already done css_get(mc.to) */
5932 memcg_oom_recover(from);
5933 memcg_oom_recover(to);
5934 wake_up_all(&mc.waitq);
5937 static void mem_cgroup_clear_mc(void)
5939 struct mem_cgroup *from = mc.from;
5942 * we must clear moving_task before waking up waiters at the end of
5945 mc.moving_task = NULL;
5946 __mem_cgroup_clear_mc();
5947 spin_lock(&mc.lock);
5950 spin_unlock(&mc.lock);
5951 mem_cgroup_end_move(from);
5954 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5955 struct cgroup_taskset *tset)
5957 struct task_struct *p = cgroup_taskset_first(tset);
5959 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5960 unsigned long move_charge_at_immigrate;
5963 * We are now commited to this value whatever it is. Changes in this
5964 * tunable will only affect upcoming migrations, not the current one.
5965 * So we need to save it, and keep it going.
5967 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
5968 if (move_charge_at_immigrate) {
5969 struct mm_struct *mm;
5970 struct mem_cgroup *from = mem_cgroup_from_task(p);
5972 VM_BUG_ON(from == memcg);
5974 mm = get_task_mm(p);
5977 /* We move charges only when we move a owner of the mm */
5978 if (mm->owner == p) {
5981 VM_BUG_ON(mc.precharge);
5982 VM_BUG_ON(mc.moved_charge);
5983 VM_BUG_ON(mc.moved_swap);
5984 mem_cgroup_start_move(from);
5985 spin_lock(&mc.lock);
5988 mc.immigrate_flags = move_charge_at_immigrate;
5989 spin_unlock(&mc.lock);
5990 /* We set mc.moving_task later */
5992 ret = mem_cgroup_precharge_mc(mm);
5994 mem_cgroup_clear_mc();
6001 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6002 struct cgroup_taskset *tset)
6004 mem_cgroup_clear_mc();
6007 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6008 unsigned long addr, unsigned long end,
6009 struct mm_walk *walk)
6012 struct vm_area_struct *vma = walk->private;
6015 enum mc_target_type target_type;
6016 union mc_target target;
6018 struct page_cgroup *pc;
6021 * We don't take compound_lock() here but no race with splitting thp
6023 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6024 * under splitting, which means there's no concurrent thp split,
6025 * - if another thread runs into split_huge_page() just after we
6026 * entered this if-block, the thread must wait for page table lock
6027 * to be unlocked in __split_huge_page_splitting(), where the main
6028 * part of thp split is not executed yet.
6030 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6031 if (mc.precharge < HPAGE_PMD_NR) {
6035 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6036 if (target_type == MC_TARGET_PAGE) {
6038 if (!isolate_lru_page(page)) {
6039 pc = lookup_page_cgroup(page);
6040 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6041 pc, mc.from, mc.to)) {
6042 mc.precharge -= HPAGE_PMD_NR;
6043 mc.moved_charge += HPAGE_PMD_NR;
6045 putback_lru_page(page);
6053 if (pmd_trans_unstable(pmd))
6056 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6057 for (; addr != end; addr += PAGE_SIZE) {
6058 pte_t ptent = *(pte++);
6064 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6065 case MC_TARGET_PAGE:
6067 if (isolate_lru_page(page))
6069 pc = lookup_page_cgroup(page);
6070 if (!mem_cgroup_move_account(page, 1, pc,
6073 /* we uncharge from mc.from later. */
6076 putback_lru_page(page);
6077 put: /* get_mctgt_type() gets the page */
6080 case MC_TARGET_SWAP:
6082 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6084 /* we fixup refcnts and charges later. */
6092 pte_unmap_unlock(pte - 1, ptl);
6097 * We have consumed all precharges we got in can_attach().
6098 * We try charge one by one, but don't do any additional
6099 * charges to mc.to if we have failed in charge once in attach()
6102 ret = mem_cgroup_do_precharge(1);
6110 static void mem_cgroup_move_charge(struct mm_struct *mm)
6112 struct vm_area_struct *vma;
6114 lru_add_drain_all();
6116 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6118 * Someone who are holding the mmap_sem might be waiting in
6119 * waitq. So we cancel all extra charges, wake up all waiters,
6120 * and retry. Because we cancel precharges, we might not be able
6121 * to move enough charges, but moving charge is a best-effort
6122 * feature anyway, so it wouldn't be a big problem.
6124 __mem_cgroup_clear_mc();
6128 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6130 struct mm_walk mem_cgroup_move_charge_walk = {
6131 .pmd_entry = mem_cgroup_move_charge_pte_range,
6135 if (is_vm_hugetlb_page(vma))
6137 ret = walk_page_range(vma->vm_start, vma->vm_end,
6138 &mem_cgroup_move_charge_walk);
6141 * means we have consumed all precharges and failed in
6142 * doing additional charge. Just abandon here.
6146 up_read(&mm->mmap_sem);
6149 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6150 struct cgroup_taskset *tset)
6152 struct task_struct *p = cgroup_taskset_first(tset);
6153 struct mm_struct *mm = get_task_mm(p);
6157 mem_cgroup_move_charge(mm);
6161 mem_cgroup_clear_mc();
6163 #else /* !CONFIG_MMU */
6164 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6165 struct cgroup_taskset *tset)
6169 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6170 struct cgroup_taskset *tset)
6173 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6174 struct cgroup_taskset *tset)
6180 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6181 * to verify whether we're attached to the default hierarchy on each mount
6184 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6187 * use_hierarchy is forced on the default hierarchy. cgroup core
6188 * guarantees that @root doesn't have any children, so turning it
6189 * on for the root memcg is enough.
6191 if (cgroup_on_dfl(root_css->cgroup))
6192 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6195 struct cgroup_subsys memory_cgrp_subsys = {
6196 .css_alloc = mem_cgroup_css_alloc,
6197 .css_online = mem_cgroup_css_online,
6198 .css_offline = mem_cgroup_css_offline,
6199 .css_free = mem_cgroup_css_free,
6200 .css_reset = mem_cgroup_css_reset,
6201 .can_attach = mem_cgroup_can_attach,
6202 .cancel_attach = mem_cgroup_cancel_attach,
6203 .attach = mem_cgroup_move_task,
6204 .bind = mem_cgroup_bind,
6205 .legacy_cftypes = mem_cgroup_files,
6209 #ifdef CONFIG_MEMCG_SWAP
6210 static int __init enable_swap_account(char *s)
6212 if (!strcmp(s, "1"))
6213 really_do_swap_account = 1;
6214 else if (!strcmp(s, "0"))
6215 really_do_swap_account = 0;
6218 __setup("swapaccount=", enable_swap_account);
6220 static void __init memsw_file_init(void)
6222 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6223 memsw_cgroup_files));
6226 static void __init enable_swap_cgroup(void)
6228 if (!mem_cgroup_disabled() && really_do_swap_account) {
6229 do_swap_account = 1;
6235 static void __init enable_swap_cgroup(void)
6240 #ifdef CONFIG_MEMCG_SWAP
6242 * mem_cgroup_swapout - transfer a memsw charge to swap
6243 * @page: page whose memsw charge to transfer
6244 * @entry: swap entry to move the charge to
6246 * Transfer the memsw charge of @page to @entry.
6248 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
6250 struct page_cgroup *pc;
6251 unsigned short oldid;
6253 VM_BUG_ON_PAGE(PageLRU(page), page);
6254 VM_BUG_ON_PAGE(page_count(page), page);
6256 if (!do_swap_account)
6259 pc = lookup_page_cgroup(page);
6261 /* Readahead page, never charged */
6262 if (!PageCgroupUsed(pc))
6265 VM_BUG_ON_PAGE(!(pc->flags & PCG_MEMSW), page);
6267 oldid = swap_cgroup_record(entry, mem_cgroup_id(pc->mem_cgroup));
6268 VM_BUG_ON_PAGE(oldid, page);
6270 pc->flags &= ~PCG_MEMSW;
6271 css_get(&pc->mem_cgroup->css);
6272 mem_cgroup_swap_statistics(pc->mem_cgroup, true);
6276 * mem_cgroup_uncharge_swap - uncharge a swap entry
6277 * @entry: swap entry to uncharge
6279 * Drop the memsw charge associated with @entry.
6281 void mem_cgroup_uncharge_swap(swp_entry_t entry)
6283 struct mem_cgroup *memcg;
6286 if (!do_swap_account)
6289 id = swap_cgroup_record(entry, 0);
6291 memcg = mem_cgroup_lookup(id);
6293 if (!mem_cgroup_is_root(memcg))
6294 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
6295 mem_cgroup_swap_statistics(memcg, false);
6296 css_put(&memcg->css);
6303 * mem_cgroup_try_charge - try charging a page
6304 * @page: page to charge
6305 * @mm: mm context of the victim
6306 * @gfp_mask: reclaim mode
6307 * @memcgp: charged memcg return
6309 * Try to charge @page to the memcg that @mm belongs to, reclaiming
6310 * pages according to @gfp_mask if necessary.
6312 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
6313 * Otherwise, an error code is returned.
6315 * After page->mapping has been set up, the caller must finalize the
6316 * charge with mem_cgroup_commit_charge(). Or abort the transaction
6317 * with mem_cgroup_cancel_charge() in case page instantiation fails.
6319 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
6320 gfp_t gfp_mask, struct mem_cgroup **memcgp)
6322 struct mem_cgroup *memcg = NULL;
6323 unsigned int nr_pages = 1;
6326 if (mem_cgroup_disabled())
6329 if (PageSwapCache(page)) {
6330 struct page_cgroup *pc = lookup_page_cgroup(page);
6332 * Every swap fault against a single page tries to charge the
6333 * page, bail as early as possible. shmem_unuse() encounters
6334 * already charged pages, too. The USED bit is protected by
6335 * the page lock, which serializes swap cache removal, which
6336 * in turn serializes uncharging.
6338 if (PageCgroupUsed(pc))
6342 if (PageTransHuge(page)) {
6343 nr_pages <<= compound_order(page);
6344 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6347 if (do_swap_account && PageSwapCache(page))
6348 memcg = try_get_mem_cgroup_from_page(page);
6350 memcg = get_mem_cgroup_from_mm(mm);
6352 ret = try_charge(memcg, gfp_mask, nr_pages);
6354 css_put(&memcg->css);
6356 if (ret == -EINTR) {
6357 memcg = root_mem_cgroup;
6366 * mem_cgroup_commit_charge - commit a page charge
6367 * @page: page to charge
6368 * @memcg: memcg to charge the page to
6369 * @lrucare: page might be on LRU already
6371 * Finalize a charge transaction started by mem_cgroup_try_charge(),
6372 * after page->mapping has been set up. This must happen atomically
6373 * as part of the page instantiation, i.e. under the page table lock
6374 * for anonymous pages, under the page lock for page and swap cache.
6376 * In addition, the page must not be on the LRU during the commit, to
6377 * prevent racing with task migration. If it might be, use @lrucare.
6379 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
6381 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
6384 unsigned int nr_pages = 1;
6386 VM_BUG_ON_PAGE(!page->mapping, page);
6387 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
6389 if (mem_cgroup_disabled())
6392 * Swap faults will attempt to charge the same page multiple
6393 * times. But reuse_swap_page() might have removed the page
6394 * from swapcache already, so we can't check PageSwapCache().
6399 commit_charge(page, memcg, lrucare);
6401 if (PageTransHuge(page)) {
6402 nr_pages <<= compound_order(page);
6403 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6406 local_irq_disable();
6407 mem_cgroup_charge_statistics(memcg, page, nr_pages);
6408 memcg_check_events(memcg, page);
6411 if (do_swap_account && PageSwapCache(page)) {
6412 swp_entry_t entry = { .val = page_private(page) };
6414 * The swap entry might not get freed for a long time,
6415 * let's not wait for it. The page already received a
6416 * memory+swap charge, drop the swap entry duplicate.
6418 mem_cgroup_uncharge_swap(entry);
6423 * mem_cgroup_cancel_charge - cancel a page charge
6424 * @page: page to charge
6425 * @memcg: memcg to charge the page to
6427 * Cancel a charge transaction started by mem_cgroup_try_charge().
6429 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
6431 unsigned int nr_pages = 1;
6433 if (mem_cgroup_disabled())
6436 * Swap faults will attempt to charge the same page multiple
6437 * times. But reuse_swap_page() might have removed the page
6438 * from swapcache already, so we can't check PageSwapCache().
6443 if (PageTransHuge(page)) {
6444 nr_pages <<= compound_order(page);
6445 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6448 cancel_charge(memcg, nr_pages);
6451 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
6452 unsigned long nr_mem, unsigned long nr_memsw,
6453 unsigned long nr_anon, unsigned long nr_file,
6454 unsigned long nr_huge, struct page *dummy_page)
6456 unsigned long flags;
6458 if (!mem_cgroup_is_root(memcg)) {
6460 res_counter_uncharge(&memcg->res,
6461 nr_mem * PAGE_SIZE);
6463 res_counter_uncharge(&memcg->memsw,
6464 nr_memsw * PAGE_SIZE);
6465 memcg_oom_recover(memcg);
6468 local_irq_save(flags);
6469 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
6470 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
6471 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
6472 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
6473 __this_cpu_add(memcg->stat->nr_page_events, nr_anon + nr_file);
6474 memcg_check_events(memcg, dummy_page);
6475 local_irq_restore(flags);
6478 static void uncharge_list(struct list_head *page_list)
6480 struct mem_cgroup *memcg = NULL;
6481 unsigned long nr_memsw = 0;
6482 unsigned long nr_anon = 0;
6483 unsigned long nr_file = 0;
6484 unsigned long nr_huge = 0;
6485 unsigned long pgpgout = 0;
6486 unsigned long nr_mem = 0;
6487 struct list_head *next;
6490 next = page_list->next;
6492 unsigned int nr_pages = 1;
6493 struct page_cgroup *pc;
6495 page = list_entry(next, struct page, lru);
6496 next = page->lru.next;
6498 VM_BUG_ON_PAGE(PageLRU(page), page);
6499 VM_BUG_ON_PAGE(page_count(page), page);
6501 pc = lookup_page_cgroup(page);
6502 if (!PageCgroupUsed(pc))
6506 * Nobody should be changing or seriously looking at
6507 * pc->mem_cgroup and pc->flags at this point, we have
6508 * fully exclusive access to the page.
6511 if (memcg != pc->mem_cgroup) {
6513 uncharge_batch(memcg, pgpgout, nr_mem, nr_memsw,
6514 nr_anon, nr_file, nr_huge, page);
6515 pgpgout = nr_mem = nr_memsw = 0;
6516 nr_anon = nr_file = nr_huge = 0;
6518 memcg = pc->mem_cgroup;
6521 if (PageTransHuge(page)) {
6522 nr_pages <<= compound_order(page);
6523 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
6524 nr_huge += nr_pages;
6528 nr_anon += nr_pages;
6530 nr_file += nr_pages;
6532 if (pc->flags & PCG_MEM)
6534 if (pc->flags & PCG_MEMSW)
6535 nr_memsw += nr_pages;
6539 } while (next != page_list);
6542 uncharge_batch(memcg, pgpgout, nr_mem, nr_memsw,
6543 nr_anon, nr_file, nr_huge, page);
6547 * mem_cgroup_uncharge - uncharge a page
6548 * @page: page to uncharge
6550 * Uncharge a page previously charged with mem_cgroup_try_charge() and
6551 * mem_cgroup_commit_charge().
6553 void mem_cgroup_uncharge(struct page *page)
6555 struct page_cgroup *pc;
6557 if (mem_cgroup_disabled())
6560 /* Don't touch page->lru of any random page, pre-check: */
6561 pc = lookup_page_cgroup(page);
6562 if (!PageCgroupUsed(pc))
6565 INIT_LIST_HEAD(&page->lru);
6566 uncharge_list(&page->lru);
6570 * mem_cgroup_uncharge_list - uncharge a list of page
6571 * @page_list: list of pages to uncharge
6573 * Uncharge a list of pages previously charged with
6574 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
6576 void mem_cgroup_uncharge_list(struct list_head *page_list)
6578 if (mem_cgroup_disabled())
6581 if (!list_empty(page_list))
6582 uncharge_list(page_list);
6586 * mem_cgroup_migrate - migrate a charge to another page
6587 * @oldpage: currently charged page
6588 * @newpage: page to transfer the charge to
6589 * @lrucare: both pages might be on the LRU already
6591 * Migrate the charge from @oldpage to @newpage.
6593 * Both pages must be locked, @newpage->mapping must be set up.
6595 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
6598 struct page_cgroup *pc;
6601 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
6602 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
6603 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
6604 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
6605 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
6606 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
6609 if (mem_cgroup_disabled())
6612 /* Page cache replacement: new page already charged? */
6613 pc = lookup_page_cgroup(newpage);
6614 if (PageCgroupUsed(pc))
6617 /* Re-entrant migration: old page already uncharged? */
6618 pc = lookup_page_cgroup(oldpage);
6619 if (!PageCgroupUsed(pc))
6622 VM_BUG_ON_PAGE(!(pc->flags & PCG_MEM), oldpage);
6623 VM_BUG_ON_PAGE(do_swap_account && !(pc->flags & PCG_MEMSW), oldpage);
6626 lock_page_lru(oldpage, &isolated);
6631 unlock_page_lru(oldpage, isolated);
6633 commit_charge(newpage, pc->mem_cgroup, lrucare);
6637 * subsys_initcall() for memory controller.
6639 * Some parts like hotcpu_notifier() have to be initialized from this context
6640 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6641 * everything that doesn't depend on a specific mem_cgroup structure should
6642 * be initialized from here.
6644 static int __init mem_cgroup_init(void)
6646 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6647 enable_swap_cgroup();
6648 mem_cgroup_soft_limit_tree_init();
6652 subsys_initcall(mem_cgroup_init);