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/vmalloc.h>
53 #include <linux/vmpressure.h>
54 #include <linux/mm_inline.h>
55 #include <linux/page_cgroup.h>
56 #include <linux/cpu.h>
57 #include <linux/oom.h>
58 #include <linux/lockdep.h>
59 #include <linux/file.h>
63 #include <net/tcp_memcontrol.h>
66 #include <asm/uaccess.h>
68 #include <trace/events/vmscan.h>
70 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
71 EXPORT_SYMBOL(mem_cgroup_subsys);
73 #define MEM_CGROUP_RECLAIM_RETRIES 5
74 static struct mem_cgroup *root_mem_cgroup __read_mostly;
76 #ifdef CONFIG_MEMCG_SWAP
77 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
78 int do_swap_account __read_mostly;
80 /* for remember boot option*/
81 #ifdef CONFIG_MEMCG_SWAP_ENABLED
82 static int really_do_swap_account __initdata = 1;
84 static int really_do_swap_account __initdata = 0;
88 #define do_swap_account 0
92 static const char * const mem_cgroup_stat_names[] = {
101 enum mem_cgroup_events_index {
102 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
103 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
104 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
105 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
106 MEM_CGROUP_EVENTS_NSTATS,
109 static const char * const mem_cgroup_events_names[] = {
116 static const char * const mem_cgroup_lru_names[] = {
125 * Per memcg event counter is incremented at every pagein/pageout. With THP,
126 * it will be incremated by the number of pages. This counter is used for
127 * for trigger some periodic events. This is straightforward and better
128 * than using jiffies etc. to handle periodic memcg event.
130 enum mem_cgroup_events_target {
131 MEM_CGROUP_TARGET_THRESH,
132 MEM_CGROUP_TARGET_SOFTLIMIT,
133 MEM_CGROUP_TARGET_NUMAINFO,
136 #define THRESHOLDS_EVENTS_TARGET 128
137 #define SOFTLIMIT_EVENTS_TARGET 1024
138 #define NUMAINFO_EVENTS_TARGET 1024
140 struct mem_cgroup_stat_cpu {
141 long count[MEM_CGROUP_STAT_NSTATS];
142 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
143 unsigned long nr_page_events;
144 unsigned long targets[MEM_CGROUP_NTARGETS];
147 struct mem_cgroup_reclaim_iter {
149 * last scanned hierarchy member. Valid only if last_dead_count
150 * matches memcg->dead_count of the hierarchy root group.
152 struct mem_cgroup *last_visited;
153 unsigned long last_dead_count;
155 /* scan generation, increased every round-trip */
156 unsigned int generation;
160 * per-zone information in memory controller.
162 struct mem_cgroup_per_zone {
163 struct lruvec lruvec;
164 unsigned long lru_size[NR_LRU_LISTS];
166 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
168 struct rb_node tree_node; /* RB tree node */
169 unsigned long long usage_in_excess;/* Set to the value by which */
170 /* the soft limit is exceeded*/
172 struct mem_cgroup *memcg; /* Back pointer, we cannot */
173 /* use container_of */
176 struct mem_cgroup_per_node {
177 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
181 * Cgroups above their limits are maintained in a RB-Tree, independent of
182 * their hierarchy representation
185 struct mem_cgroup_tree_per_zone {
186 struct rb_root rb_root;
190 struct mem_cgroup_tree_per_node {
191 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
194 struct mem_cgroup_tree {
195 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
198 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
200 struct mem_cgroup_threshold {
201 struct eventfd_ctx *eventfd;
206 struct mem_cgroup_threshold_ary {
207 /* An array index points to threshold just below or equal to usage. */
208 int current_threshold;
209 /* Size of entries[] */
211 /* Array of thresholds */
212 struct mem_cgroup_threshold entries[0];
215 struct mem_cgroup_thresholds {
216 /* Primary thresholds array */
217 struct mem_cgroup_threshold_ary *primary;
219 * Spare threshold array.
220 * This is needed to make mem_cgroup_unregister_event() "never fail".
221 * It must be able to store at least primary->size - 1 entries.
223 struct mem_cgroup_threshold_ary *spare;
227 struct mem_cgroup_eventfd_list {
228 struct list_head list;
229 struct eventfd_ctx *eventfd;
233 * cgroup_event represents events which userspace want to receive.
235 struct mem_cgroup_event {
237 * memcg which the event belongs to.
239 struct mem_cgroup *memcg;
241 * eventfd to signal userspace about the event.
243 struct eventfd_ctx *eventfd;
245 * Each of these stored in a list by the cgroup.
247 struct list_head list;
249 * register_event() callback will be used to add new userspace
250 * waiter for changes related to this event. Use eventfd_signal()
251 * on eventfd to send notification to userspace.
253 int (*register_event)(struct mem_cgroup *memcg,
254 struct eventfd_ctx *eventfd, const char *args);
256 * unregister_event() callback will be called when userspace closes
257 * the eventfd or on cgroup removing. This callback must be set,
258 * if you want provide notification functionality.
260 void (*unregister_event)(struct mem_cgroup *memcg,
261 struct eventfd_ctx *eventfd);
263 * All fields below needed to unregister event when
264 * userspace closes eventfd.
267 wait_queue_head_t *wqh;
269 struct work_struct remove;
272 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
273 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
276 * The memory controller data structure. The memory controller controls both
277 * page cache and RSS per cgroup. We would eventually like to provide
278 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
279 * to help the administrator determine what knobs to tune.
281 * TODO: Add a water mark for the memory controller. Reclaim will begin when
282 * we hit the water mark. May be even add a low water mark, such that
283 * no reclaim occurs from a cgroup at it's low water mark, this is
284 * a feature that will be implemented much later in the future.
287 struct cgroup_subsys_state css;
289 * the counter to account for memory usage
291 struct res_counter res;
293 /* vmpressure notifications */
294 struct vmpressure vmpressure;
297 * the counter to account for mem+swap usage.
299 struct res_counter memsw;
302 * the counter to account for kernel memory usage.
304 struct res_counter kmem;
306 * Should the accounting and control be hierarchical, per subtree?
309 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
313 atomic_t oom_wakeups;
316 /* OOM-Killer disable */
317 int oom_kill_disable;
319 /* set when res.limit == memsw.limit */
320 bool memsw_is_minimum;
322 /* protect arrays of thresholds */
323 struct mutex thresholds_lock;
325 /* thresholds for memory usage. RCU-protected */
326 struct mem_cgroup_thresholds thresholds;
328 /* thresholds for mem+swap usage. RCU-protected */
329 struct mem_cgroup_thresholds memsw_thresholds;
331 /* For oom notifier event fd */
332 struct list_head oom_notify;
335 * Should we move charges of a task when a task is moved into this
336 * mem_cgroup ? And what type of charges should we move ?
338 unsigned long move_charge_at_immigrate;
340 * set > 0 if pages under this cgroup are moving to other cgroup.
342 atomic_t moving_account;
343 /* taken only while moving_account > 0 */
344 spinlock_t move_lock;
348 struct mem_cgroup_stat_cpu __percpu *stat;
350 * used when a cpu is offlined or other synchronizations
351 * See mem_cgroup_read_stat().
353 struct mem_cgroup_stat_cpu nocpu_base;
354 spinlock_t pcp_counter_lock;
357 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
358 struct cg_proto tcp_mem;
360 #if defined(CONFIG_MEMCG_KMEM)
361 /* analogous to slab_common's slab_caches list. per-memcg */
362 struct list_head memcg_slab_caches;
363 /* Not a spinlock, we can take a lot of time walking the list */
364 struct mutex slab_caches_mutex;
365 /* Index in the kmem_cache->memcg_params->memcg_caches array */
369 int last_scanned_node;
371 nodemask_t scan_nodes;
372 atomic_t numainfo_events;
373 atomic_t numainfo_updating;
376 /* List of events which userspace want to receive */
377 struct list_head event_list;
378 spinlock_t event_list_lock;
380 struct mem_cgroup_per_node *nodeinfo[0];
381 /* WARNING: nodeinfo must be the last member here */
384 static size_t memcg_size(void)
386 return sizeof(struct mem_cgroup) +
387 nr_node_ids * sizeof(struct mem_cgroup_per_node *);
390 /* internal only representation about the status of kmem accounting. */
392 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
393 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
394 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
397 /* We account when limit is on, but only after call sites are patched */
398 #define KMEM_ACCOUNTED_MASK \
399 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
401 #ifdef CONFIG_MEMCG_KMEM
402 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
404 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
407 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
409 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
412 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
414 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
417 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
419 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
422 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
425 * Our caller must use css_get() first, because memcg_uncharge_kmem()
426 * will call css_put() if it sees the memcg is dead.
429 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
430 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
433 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
435 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
436 &memcg->kmem_account_flags);
440 /* Stuffs for move charges at task migration. */
442 * Types of charges to be moved. "move_charge_at_immitgrate" and
443 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
446 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
447 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
451 /* "mc" and its members are protected by cgroup_mutex */
452 static struct move_charge_struct {
453 spinlock_t lock; /* for from, to */
454 struct mem_cgroup *from;
455 struct mem_cgroup *to;
456 unsigned long immigrate_flags;
457 unsigned long precharge;
458 unsigned long moved_charge;
459 unsigned long moved_swap;
460 struct task_struct *moving_task; /* a task moving charges */
461 wait_queue_head_t waitq; /* a waitq for other context */
463 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
464 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
467 static bool move_anon(void)
469 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
472 static bool move_file(void)
474 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
478 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
479 * limit reclaim to prevent infinite loops, if they ever occur.
481 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
482 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
485 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
486 MEM_CGROUP_CHARGE_TYPE_ANON,
487 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
488 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
492 /* for encoding cft->private value on file */
500 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
501 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
502 #define MEMFILE_ATTR(val) ((val) & 0xffff)
503 /* Used for OOM nofiier */
504 #define OOM_CONTROL (0)
507 * Reclaim flags for mem_cgroup_hierarchical_reclaim
509 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
510 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
511 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
512 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
515 * The memcg_create_mutex will be held whenever a new cgroup is created.
516 * As a consequence, any change that needs to protect against new child cgroups
517 * appearing has to hold it as well.
519 static DEFINE_MUTEX(memcg_create_mutex);
521 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
523 return s ? container_of(s, struct mem_cgroup, css) : NULL;
526 /* Some nice accessors for the vmpressure. */
527 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
530 memcg = root_mem_cgroup;
531 return &memcg->vmpressure;
534 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
536 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
539 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
541 return (memcg == root_mem_cgroup);
545 * We restrict the id in the range of [1, 65535], so it can fit into
548 #define MEM_CGROUP_ID_MAX USHRT_MAX
550 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
553 * The ID of the root cgroup is 0, but memcg treat 0 as an
554 * invalid ID, so we return (cgroup_id + 1).
556 return memcg->css.cgroup->id + 1;
559 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
561 struct cgroup_subsys_state *css;
563 css = css_from_id(id - 1, &mem_cgroup_subsys);
564 return mem_cgroup_from_css(css);
567 /* Writing them here to avoid exposing memcg's inner layout */
568 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
570 void sock_update_memcg(struct sock *sk)
572 if (mem_cgroup_sockets_enabled) {
573 struct mem_cgroup *memcg;
574 struct cg_proto *cg_proto;
576 BUG_ON(!sk->sk_prot->proto_cgroup);
578 /* Socket cloning can throw us here with sk_cgrp already
579 * filled. It won't however, necessarily happen from
580 * process context. So the test for root memcg given
581 * the current task's memcg won't help us in this case.
583 * Respecting the original socket's memcg is a better
584 * decision in this case.
587 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
588 css_get(&sk->sk_cgrp->memcg->css);
593 memcg = mem_cgroup_from_task(current);
594 cg_proto = sk->sk_prot->proto_cgroup(memcg);
595 if (!mem_cgroup_is_root(memcg) &&
596 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
597 sk->sk_cgrp = cg_proto;
602 EXPORT_SYMBOL(sock_update_memcg);
604 void sock_release_memcg(struct sock *sk)
606 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
607 struct mem_cgroup *memcg;
608 WARN_ON(!sk->sk_cgrp->memcg);
609 memcg = sk->sk_cgrp->memcg;
610 css_put(&sk->sk_cgrp->memcg->css);
614 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
616 if (!memcg || mem_cgroup_is_root(memcg))
619 return &memcg->tcp_mem;
621 EXPORT_SYMBOL(tcp_proto_cgroup);
623 static void disarm_sock_keys(struct mem_cgroup *memcg)
625 if (!memcg_proto_activated(&memcg->tcp_mem))
627 static_key_slow_dec(&memcg_socket_limit_enabled);
630 static void disarm_sock_keys(struct mem_cgroup *memcg)
635 #ifdef CONFIG_MEMCG_KMEM
637 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
638 * The main reason for not using cgroup id for this:
639 * this works better in sparse environments, where we have a lot of memcgs,
640 * but only a few kmem-limited. Or also, if we have, for instance, 200
641 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
642 * 200 entry array for that.
644 * The current size of the caches array is stored in
645 * memcg_limited_groups_array_size. It will double each time we have to
648 static DEFINE_IDA(kmem_limited_groups);
649 int memcg_limited_groups_array_size;
652 * MIN_SIZE is different than 1, because we would like to avoid going through
653 * the alloc/free process all the time. In a small machine, 4 kmem-limited
654 * cgroups is a reasonable guess. In the future, it could be a parameter or
655 * tunable, but that is strictly not necessary.
657 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
658 * this constant directly from cgroup, but it is understandable that this is
659 * better kept as an internal representation in cgroup.c. In any case, the
660 * cgrp_id space is not getting any smaller, and we don't have to necessarily
661 * increase ours as well if it increases.
663 #define MEMCG_CACHES_MIN_SIZE 4
664 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
667 * A lot of the calls to the cache allocation functions are expected to be
668 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
669 * conditional to this static branch, we'll have to allow modules that does
670 * kmem_cache_alloc and the such to see this symbol as well
672 struct static_key memcg_kmem_enabled_key;
673 EXPORT_SYMBOL(memcg_kmem_enabled_key);
675 static void disarm_kmem_keys(struct mem_cgroup *memcg)
677 if (memcg_kmem_is_active(memcg)) {
678 static_key_slow_dec(&memcg_kmem_enabled_key);
679 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
682 * This check can't live in kmem destruction function,
683 * since the charges will outlive the cgroup
685 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
688 static void disarm_kmem_keys(struct mem_cgroup *memcg)
691 #endif /* CONFIG_MEMCG_KMEM */
693 static void disarm_static_keys(struct mem_cgroup *memcg)
695 disarm_sock_keys(memcg);
696 disarm_kmem_keys(memcg);
699 static void drain_all_stock_async(struct mem_cgroup *memcg);
701 static struct mem_cgroup_per_zone *
702 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
704 VM_BUG_ON((unsigned)nid >= nr_node_ids);
705 return &memcg->nodeinfo[nid]->zoneinfo[zid];
708 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
713 static struct mem_cgroup_per_zone *
714 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
716 int nid = page_to_nid(page);
717 int zid = page_zonenum(page);
719 return mem_cgroup_zoneinfo(memcg, nid, zid);
722 static struct mem_cgroup_tree_per_zone *
723 soft_limit_tree_node_zone(int nid, int zid)
725 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
728 static struct mem_cgroup_tree_per_zone *
729 soft_limit_tree_from_page(struct page *page)
731 int nid = page_to_nid(page);
732 int zid = page_zonenum(page);
734 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
738 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
739 struct mem_cgroup_per_zone *mz,
740 struct mem_cgroup_tree_per_zone *mctz,
741 unsigned long long new_usage_in_excess)
743 struct rb_node **p = &mctz->rb_root.rb_node;
744 struct rb_node *parent = NULL;
745 struct mem_cgroup_per_zone *mz_node;
750 mz->usage_in_excess = new_usage_in_excess;
751 if (!mz->usage_in_excess)
755 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
757 if (mz->usage_in_excess < mz_node->usage_in_excess)
760 * We can't avoid mem cgroups that are over their soft
761 * limit by the same amount
763 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
766 rb_link_node(&mz->tree_node, parent, p);
767 rb_insert_color(&mz->tree_node, &mctz->rb_root);
772 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
773 struct mem_cgroup_per_zone *mz,
774 struct mem_cgroup_tree_per_zone *mctz)
778 rb_erase(&mz->tree_node, &mctz->rb_root);
783 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
784 struct mem_cgroup_per_zone *mz,
785 struct mem_cgroup_tree_per_zone *mctz)
787 spin_lock(&mctz->lock);
788 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
789 spin_unlock(&mctz->lock);
793 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
795 unsigned long long excess;
796 struct mem_cgroup_per_zone *mz;
797 struct mem_cgroup_tree_per_zone *mctz;
798 int nid = page_to_nid(page);
799 int zid = page_zonenum(page);
800 mctz = soft_limit_tree_from_page(page);
803 * Necessary to update all ancestors when hierarchy is used.
804 * because their event counter is not touched.
806 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
807 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
808 excess = res_counter_soft_limit_excess(&memcg->res);
810 * We have to update the tree if mz is on RB-tree or
811 * mem is over its softlimit.
813 if (excess || mz->on_tree) {
814 spin_lock(&mctz->lock);
815 /* if on-tree, remove it */
817 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
819 * Insert again. mz->usage_in_excess will be updated.
820 * If excess is 0, no tree ops.
822 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
823 spin_unlock(&mctz->lock);
828 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
831 struct mem_cgroup_per_zone *mz;
832 struct mem_cgroup_tree_per_zone *mctz;
834 for_each_node(node) {
835 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
836 mz = mem_cgroup_zoneinfo(memcg, node, zone);
837 mctz = soft_limit_tree_node_zone(node, zone);
838 mem_cgroup_remove_exceeded(memcg, mz, mctz);
843 static struct mem_cgroup_per_zone *
844 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
846 struct rb_node *rightmost = NULL;
847 struct mem_cgroup_per_zone *mz;
851 rightmost = rb_last(&mctz->rb_root);
853 goto done; /* Nothing to reclaim from */
855 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
857 * Remove the node now but someone else can add it back,
858 * we will to add it back at the end of reclaim to its correct
859 * position in the tree.
861 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
862 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
863 !css_tryget(&mz->memcg->css))
869 static struct mem_cgroup_per_zone *
870 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
872 struct mem_cgroup_per_zone *mz;
874 spin_lock(&mctz->lock);
875 mz = __mem_cgroup_largest_soft_limit_node(mctz);
876 spin_unlock(&mctz->lock);
881 * Implementation Note: reading percpu statistics for memcg.
883 * Both of vmstat[] and percpu_counter has threshold and do periodic
884 * synchronization to implement "quick" read. There are trade-off between
885 * reading cost and precision of value. Then, we may have a chance to implement
886 * a periodic synchronizion of counter in memcg's counter.
888 * But this _read() function is used for user interface now. The user accounts
889 * memory usage by memory cgroup and he _always_ requires exact value because
890 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
891 * have to visit all online cpus and make sum. So, for now, unnecessary
892 * synchronization is not implemented. (just implemented for cpu hotplug)
894 * If there are kernel internal actions which can make use of some not-exact
895 * value, and reading all cpu value can be performance bottleneck in some
896 * common workload, threashold and synchonization as vmstat[] should be
899 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
900 enum mem_cgroup_stat_index idx)
906 for_each_online_cpu(cpu)
907 val += per_cpu(memcg->stat->count[idx], cpu);
908 #ifdef CONFIG_HOTPLUG_CPU
909 spin_lock(&memcg->pcp_counter_lock);
910 val += memcg->nocpu_base.count[idx];
911 spin_unlock(&memcg->pcp_counter_lock);
917 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
920 int val = (charge) ? 1 : -1;
921 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
924 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
925 enum mem_cgroup_events_index idx)
927 unsigned long val = 0;
931 for_each_online_cpu(cpu)
932 val += per_cpu(memcg->stat->events[idx], cpu);
933 #ifdef CONFIG_HOTPLUG_CPU
934 spin_lock(&memcg->pcp_counter_lock);
935 val += memcg->nocpu_base.events[idx];
936 spin_unlock(&memcg->pcp_counter_lock);
942 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
944 bool anon, int nr_pages)
949 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
950 * counted as CACHE even if it's on ANON LRU.
953 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
956 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
959 if (PageTransHuge(page))
960 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
963 /* pagein of a big page is an event. So, ignore page size */
965 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
967 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
968 nr_pages = -nr_pages; /* for event */
971 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
977 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
979 struct mem_cgroup_per_zone *mz;
981 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
982 return mz->lru_size[lru];
986 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
987 unsigned int lru_mask)
989 struct mem_cgroup_per_zone *mz;
991 unsigned long ret = 0;
993 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
996 if (BIT(lru) & lru_mask)
997 ret += mz->lru_size[lru];
1002 static unsigned long
1003 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
1004 int nid, unsigned int lru_mask)
1009 for (zid = 0; zid < MAX_NR_ZONES; zid++)
1010 total += mem_cgroup_zone_nr_lru_pages(memcg,
1011 nid, zid, lru_mask);
1016 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
1017 unsigned int lru_mask)
1022 for_each_node_state(nid, N_MEMORY)
1023 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1027 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1028 enum mem_cgroup_events_target target)
1030 unsigned long val, next;
1032 val = __this_cpu_read(memcg->stat->nr_page_events);
1033 next = __this_cpu_read(memcg->stat->targets[target]);
1034 /* from time_after() in jiffies.h */
1035 if ((long)next - (long)val < 0) {
1037 case MEM_CGROUP_TARGET_THRESH:
1038 next = val + THRESHOLDS_EVENTS_TARGET;
1040 case MEM_CGROUP_TARGET_SOFTLIMIT:
1041 next = val + SOFTLIMIT_EVENTS_TARGET;
1043 case MEM_CGROUP_TARGET_NUMAINFO:
1044 next = val + NUMAINFO_EVENTS_TARGET;
1049 __this_cpu_write(memcg->stat->targets[target], next);
1056 * Check events in order.
1059 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1062 /* threshold event is triggered in finer grain than soft limit */
1063 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1064 MEM_CGROUP_TARGET_THRESH))) {
1066 bool do_numainfo __maybe_unused;
1068 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1069 MEM_CGROUP_TARGET_SOFTLIMIT);
1070 #if MAX_NUMNODES > 1
1071 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1072 MEM_CGROUP_TARGET_NUMAINFO);
1076 mem_cgroup_threshold(memcg);
1077 if (unlikely(do_softlimit))
1078 mem_cgroup_update_tree(memcg, page);
1079 #if MAX_NUMNODES > 1
1080 if (unlikely(do_numainfo))
1081 atomic_inc(&memcg->numainfo_events);
1087 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1090 * mm_update_next_owner() may clear mm->owner to NULL
1091 * if it races with swapoff, page migration, etc.
1092 * So this can be called with p == NULL.
1097 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1100 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1102 struct mem_cgroup *memcg = NULL;
1107 * Because we have no locks, mm->owner's may be being moved to other
1108 * cgroup. We use css_tryget() here even if this looks
1109 * pessimistic (rather than adding locks here).
1113 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1114 if (unlikely(!memcg))
1116 } while (!css_tryget(&memcg->css));
1122 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1123 * ref. count) or NULL if the whole root's subtree has been visited.
1125 * helper function to be used by mem_cgroup_iter
1127 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1128 struct mem_cgroup *last_visited)
1130 struct cgroup_subsys_state *prev_css, *next_css;
1132 prev_css = last_visited ? &last_visited->css : NULL;
1134 next_css = css_next_descendant_pre(prev_css, &root->css);
1137 * Even if we found a group we have to make sure it is
1138 * alive. css && !memcg means that the groups should be
1139 * skipped and we should continue the tree walk.
1140 * last_visited css is safe to use because it is
1141 * protected by css_get and the tree walk is rcu safe.
1144 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
1146 if (css_tryget(&mem->css))
1149 prev_css = next_css;
1157 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1160 * When a group in the hierarchy below root is destroyed, the
1161 * hierarchy iterator can no longer be trusted since it might
1162 * have pointed to the destroyed group. Invalidate it.
1164 atomic_inc(&root->dead_count);
1167 static struct mem_cgroup *
1168 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1169 struct mem_cgroup *root,
1172 struct mem_cgroup *position = NULL;
1174 * A cgroup destruction happens in two stages: offlining and
1175 * release. They are separated by a RCU grace period.
1177 * If the iterator is valid, we may still race with an
1178 * offlining. The RCU lock ensures the object won't be
1179 * released, tryget will fail if we lost the race.
1181 *sequence = atomic_read(&root->dead_count);
1182 if (iter->last_dead_count == *sequence) {
1184 position = iter->last_visited;
1185 if (position && !css_tryget(&position->css))
1191 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1192 struct mem_cgroup *last_visited,
1193 struct mem_cgroup *new_position,
1197 css_put(&last_visited->css);
1199 * We store the sequence count from the time @last_visited was
1200 * loaded successfully instead of rereading it here so that we
1201 * don't lose destruction events in between. We could have
1202 * raced with the destruction of @new_position after all.
1204 iter->last_visited = new_position;
1206 iter->last_dead_count = sequence;
1210 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1211 * @root: hierarchy root
1212 * @prev: previously returned memcg, NULL on first invocation
1213 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1215 * Returns references to children of the hierarchy below @root, or
1216 * @root itself, or %NULL after a full round-trip.
1218 * Caller must pass the return value in @prev on subsequent
1219 * invocations for reference counting, or use mem_cgroup_iter_break()
1220 * to cancel a hierarchy walk before the round-trip is complete.
1222 * Reclaimers can specify a zone and a priority level in @reclaim to
1223 * divide up the memcgs in the hierarchy among all concurrent
1224 * reclaimers operating on the same zone and priority.
1226 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1227 struct mem_cgroup *prev,
1228 struct mem_cgroup_reclaim_cookie *reclaim)
1230 struct mem_cgroup *memcg = NULL;
1231 struct mem_cgroup *last_visited = NULL;
1233 if (mem_cgroup_disabled())
1237 root = root_mem_cgroup;
1239 if (prev && !reclaim)
1240 last_visited = prev;
1242 if (!root->use_hierarchy && root != root_mem_cgroup) {
1250 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1251 int uninitialized_var(seq);
1254 int nid = zone_to_nid(reclaim->zone);
1255 int zid = zone_idx(reclaim->zone);
1256 struct mem_cgroup_per_zone *mz;
1258 mz = mem_cgroup_zoneinfo(root, nid, zid);
1259 iter = &mz->reclaim_iter[reclaim->priority];
1260 if (prev && reclaim->generation != iter->generation) {
1261 iter->last_visited = NULL;
1265 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1268 memcg = __mem_cgroup_iter_next(root, last_visited);
1271 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1275 else if (!prev && memcg)
1276 reclaim->generation = iter->generation;
1285 if (prev && prev != root)
1286 css_put(&prev->css);
1292 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1293 * @root: hierarchy root
1294 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1296 void mem_cgroup_iter_break(struct mem_cgroup *root,
1297 struct mem_cgroup *prev)
1300 root = root_mem_cgroup;
1301 if (prev && prev != root)
1302 css_put(&prev->css);
1306 * Iteration constructs for visiting all cgroups (under a tree). If
1307 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1308 * be used for reference counting.
1310 #define for_each_mem_cgroup_tree(iter, root) \
1311 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1313 iter = mem_cgroup_iter(root, iter, NULL))
1315 #define for_each_mem_cgroup(iter) \
1316 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1318 iter = mem_cgroup_iter(NULL, iter, NULL))
1320 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1322 struct mem_cgroup *memcg;
1325 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1326 if (unlikely(!memcg))
1331 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1334 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1342 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1345 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1346 * @zone: zone of the wanted lruvec
1347 * @memcg: memcg of the wanted lruvec
1349 * Returns the lru list vector holding pages for the given @zone and
1350 * @mem. This can be the global zone lruvec, if the memory controller
1353 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1354 struct mem_cgroup *memcg)
1356 struct mem_cgroup_per_zone *mz;
1357 struct lruvec *lruvec;
1359 if (mem_cgroup_disabled()) {
1360 lruvec = &zone->lruvec;
1364 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1365 lruvec = &mz->lruvec;
1368 * Since a node can be onlined after the mem_cgroup was created,
1369 * we have to be prepared to initialize lruvec->zone here;
1370 * and if offlined then reonlined, we need to reinitialize it.
1372 if (unlikely(lruvec->zone != zone))
1373 lruvec->zone = zone;
1378 * Following LRU functions are allowed to be used without PCG_LOCK.
1379 * Operations are called by routine of global LRU independently from memcg.
1380 * What we have to take care of here is validness of pc->mem_cgroup.
1382 * Changes to pc->mem_cgroup happens when
1385 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1386 * It is added to LRU before charge.
1387 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1388 * When moving account, the page is not on LRU. It's isolated.
1392 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1394 * @zone: zone of the page
1396 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1398 struct mem_cgroup_per_zone *mz;
1399 struct mem_cgroup *memcg;
1400 struct page_cgroup *pc;
1401 struct lruvec *lruvec;
1403 if (mem_cgroup_disabled()) {
1404 lruvec = &zone->lruvec;
1408 pc = lookup_page_cgroup(page);
1409 memcg = pc->mem_cgroup;
1412 * Surreptitiously switch any uncharged offlist page to root:
1413 * an uncharged page off lru does nothing to secure
1414 * its former mem_cgroup from sudden removal.
1416 * Our caller holds lru_lock, and PageCgroupUsed is updated
1417 * under page_cgroup lock: between them, they make all uses
1418 * of pc->mem_cgroup safe.
1420 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1421 pc->mem_cgroup = memcg = root_mem_cgroup;
1423 mz = page_cgroup_zoneinfo(memcg, page);
1424 lruvec = &mz->lruvec;
1427 * Since a node can be onlined after the mem_cgroup was created,
1428 * we have to be prepared to initialize lruvec->zone here;
1429 * and if offlined then reonlined, we need to reinitialize it.
1431 if (unlikely(lruvec->zone != zone))
1432 lruvec->zone = zone;
1437 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1438 * @lruvec: mem_cgroup per zone lru vector
1439 * @lru: index of lru list the page is sitting on
1440 * @nr_pages: positive when adding or negative when removing
1442 * This function must be called when a page is added to or removed from an
1445 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1448 struct mem_cgroup_per_zone *mz;
1449 unsigned long *lru_size;
1451 if (mem_cgroup_disabled())
1454 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1455 lru_size = mz->lru_size + lru;
1456 *lru_size += nr_pages;
1457 VM_BUG_ON((long)(*lru_size) < 0);
1461 * Checks whether given mem is same or in the root_mem_cgroup's
1464 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1465 struct mem_cgroup *memcg)
1467 if (root_memcg == memcg)
1469 if (!root_memcg->use_hierarchy || !memcg)
1471 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1474 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1475 struct mem_cgroup *memcg)
1480 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1485 bool task_in_mem_cgroup(struct task_struct *task,
1486 const struct mem_cgroup *memcg)
1488 struct mem_cgroup *curr = NULL;
1489 struct task_struct *p;
1492 p = find_lock_task_mm(task);
1494 curr = try_get_mem_cgroup_from_mm(p->mm);
1498 * All threads may have already detached their mm's, but the oom
1499 * killer still needs to detect if they have already been oom
1500 * killed to prevent needlessly killing additional tasks.
1503 curr = mem_cgroup_from_task(task);
1505 css_get(&curr->css);
1511 * We should check use_hierarchy of "memcg" not "curr". Because checking
1512 * use_hierarchy of "curr" here make this function true if hierarchy is
1513 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1514 * hierarchy(even if use_hierarchy is disabled in "memcg").
1516 ret = mem_cgroup_same_or_subtree(memcg, curr);
1517 css_put(&curr->css);
1521 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1523 unsigned long inactive_ratio;
1524 unsigned long inactive;
1525 unsigned long active;
1528 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1529 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1531 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1533 inactive_ratio = int_sqrt(10 * gb);
1537 return inactive * inactive_ratio < active;
1540 #define mem_cgroup_from_res_counter(counter, member) \
1541 container_of(counter, struct mem_cgroup, member)
1544 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1545 * @memcg: the memory cgroup
1547 * Returns the maximum amount of memory @mem can be charged with, in
1550 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1552 unsigned long long margin;
1554 margin = res_counter_margin(&memcg->res);
1555 if (do_swap_account)
1556 margin = min(margin, res_counter_margin(&memcg->memsw));
1557 return margin >> PAGE_SHIFT;
1560 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1563 if (!css_parent(&memcg->css))
1564 return vm_swappiness;
1566 return memcg->swappiness;
1570 * memcg->moving_account is used for checking possibility that some thread is
1571 * calling move_account(). When a thread on CPU-A starts moving pages under
1572 * a memcg, other threads should check memcg->moving_account under
1573 * rcu_read_lock(), like this:
1577 * memcg->moving_account+1 if (memcg->mocing_account)
1579 * synchronize_rcu() update something.
1584 /* for quick checking without looking up memcg */
1585 atomic_t memcg_moving __read_mostly;
1587 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1589 atomic_inc(&memcg_moving);
1590 atomic_inc(&memcg->moving_account);
1594 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1597 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1598 * We check NULL in callee rather than caller.
1601 atomic_dec(&memcg_moving);
1602 atomic_dec(&memcg->moving_account);
1607 * 2 routines for checking "mem" is under move_account() or not.
1609 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1610 * is used for avoiding races in accounting. If true,
1611 * pc->mem_cgroup may be overwritten.
1613 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1614 * under hierarchy of moving cgroups. This is for
1615 * waiting at hith-memory prressure caused by "move".
1618 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1620 VM_BUG_ON(!rcu_read_lock_held());
1621 return atomic_read(&memcg->moving_account) > 0;
1624 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1626 struct mem_cgroup *from;
1627 struct mem_cgroup *to;
1630 * Unlike task_move routines, we access mc.to, mc.from not under
1631 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1633 spin_lock(&mc.lock);
1639 ret = mem_cgroup_same_or_subtree(memcg, from)
1640 || mem_cgroup_same_or_subtree(memcg, to);
1642 spin_unlock(&mc.lock);
1646 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1648 if (mc.moving_task && current != mc.moving_task) {
1649 if (mem_cgroup_under_move(memcg)) {
1651 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1652 /* moving charge context might have finished. */
1655 finish_wait(&mc.waitq, &wait);
1663 * Take this lock when
1664 * - a code tries to modify page's memcg while it's USED.
1665 * - a code tries to modify page state accounting in a memcg.
1666 * see mem_cgroup_stolen(), too.
1668 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1669 unsigned long *flags)
1671 spin_lock_irqsave(&memcg->move_lock, *flags);
1674 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1675 unsigned long *flags)
1677 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1680 #define K(x) ((x) << (PAGE_SHIFT-10))
1682 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1683 * @memcg: The memory cgroup that went over limit
1684 * @p: Task that is going to be killed
1686 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1689 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1691 struct cgroup *task_cgrp;
1692 struct cgroup *mem_cgrp;
1694 * Need a buffer in BSS, can't rely on allocations. The code relies
1695 * on the assumption that OOM is serialized for memory controller.
1696 * If this assumption is broken, revisit this code.
1698 static char memcg_name[PATH_MAX];
1700 struct mem_cgroup *iter;
1708 mem_cgrp = memcg->css.cgroup;
1709 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1711 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1714 * Unfortunately, we are unable to convert to a useful name
1715 * But we'll still print out the usage information
1722 pr_info("Task in %s killed", memcg_name);
1725 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1733 * Continues from above, so we don't need an KERN_ level
1735 pr_cont(" as a result of limit of %s\n", memcg_name);
1738 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1739 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1740 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1741 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1742 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1743 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1744 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1745 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1746 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1747 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1748 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1749 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1751 for_each_mem_cgroup_tree(iter, memcg) {
1752 pr_info("Memory cgroup stats");
1755 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1757 pr_cont(" for %s", memcg_name);
1761 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1762 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1764 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1765 K(mem_cgroup_read_stat(iter, i)));
1768 for (i = 0; i < NR_LRU_LISTS; i++)
1769 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1770 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1777 * This function returns the number of memcg under hierarchy tree. Returns
1778 * 1(self count) if no children.
1780 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1783 struct mem_cgroup *iter;
1785 for_each_mem_cgroup_tree(iter, memcg)
1791 * Return the memory (and swap, if configured) limit for a memcg.
1793 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1797 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1800 * Do not consider swap space if we cannot swap due to swappiness
1802 if (mem_cgroup_swappiness(memcg)) {
1805 limit += total_swap_pages << PAGE_SHIFT;
1806 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1809 * If memsw is finite and limits the amount of swap space
1810 * available to this memcg, return that limit.
1812 limit = min(limit, memsw);
1818 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1821 struct mem_cgroup *iter;
1822 unsigned long chosen_points = 0;
1823 unsigned long totalpages;
1824 unsigned int points = 0;
1825 struct task_struct *chosen = NULL;
1828 * If current has a pending SIGKILL or is exiting, then automatically
1829 * select it. The goal is to allow it to allocate so that it may
1830 * quickly exit and free its memory.
1832 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1833 set_thread_flag(TIF_MEMDIE);
1837 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1838 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1839 for_each_mem_cgroup_tree(iter, memcg) {
1840 struct css_task_iter it;
1841 struct task_struct *task;
1843 css_task_iter_start(&iter->css, &it);
1844 while ((task = css_task_iter_next(&it))) {
1845 switch (oom_scan_process_thread(task, totalpages, NULL,
1847 case OOM_SCAN_SELECT:
1849 put_task_struct(chosen);
1851 chosen_points = ULONG_MAX;
1852 get_task_struct(chosen);
1854 case OOM_SCAN_CONTINUE:
1856 case OOM_SCAN_ABORT:
1857 css_task_iter_end(&it);
1858 mem_cgroup_iter_break(memcg, iter);
1860 put_task_struct(chosen);
1865 points = oom_badness(task, memcg, NULL, totalpages);
1866 if (points > chosen_points) {
1868 put_task_struct(chosen);
1870 chosen_points = points;
1871 get_task_struct(chosen);
1874 css_task_iter_end(&it);
1879 points = chosen_points * 1000 / totalpages;
1880 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1881 NULL, "Memory cgroup out of memory");
1884 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1886 unsigned long flags)
1888 unsigned long total = 0;
1889 bool noswap = false;
1892 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1894 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1897 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1899 drain_all_stock_async(memcg);
1900 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1902 * Allow limit shrinkers, which are triggered directly
1903 * by userspace, to catch signals and stop reclaim
1904 * after minimal progress, regardless of the margin.
1906 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1908 if (mem_cgroup_margin(memcg))
1911 * If nothing was reclaimed after two attempts, there
1912 * may be no reclaimable pages in this hierarchy.
1921 * test_mem_cgroup_node_reclaimable
1922 * @memcg: the target memcg
1923 * @nid: the node ID to be checked.
1924 * @noswap : specify true here if the user wants flle only information.
1926 * This function returns whether the specified memcg contains any
1927 * reclaimable pages on a node. Returns true if there are any reclaimable
1928 * pages in the node.
1930 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1931 int nid, bool noswap)
1933 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1935 if (noswap || !total_swap_pages)
1937 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1942 #if MAX_NUMNODES > 1
1945 * Always updating the nodemask is not very good - even if we have an empty
1946 * list or the wrong list here, we can start from some node and traverse all
1947 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1950 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1954 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1955 * pagein/pageout changes since the last update.
1957 if (!atomic_read(&memcg->numainfo_events))
1959 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1962 /* make a nodemask where this memcg uses memory from */
1963 memcg->scan_nodes = node_states[N_MEMORY];
1965 for_each_node_mask(nid, node_states[N_MEMORY]) {
1967 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1968 node_clear(nid, memcg->scan_nodes);
1971 atomic_set(&memcg->numainfo_events, 0);
1972 atomic_set(&memcg->numainfo_updating, 0);
1976 * Selecting a node where we start reclaim from. Because what we need is just
1977 * reducing usage counter, start from anywhere is O,K. Considering
1978 * memory reclaim from current node, there are pros. and cons.
1980 * Freeing memory from current node means freeing memory from a node which
1981 * we'll use or we've used. So, it may make LRU bad. And if several threads
1982 * hit limits, it will see a contention on a node. But freeing from remote
1983 * node means more costs for memory reclaim because of memory latency.
1985 * Now, we use round-robin. Better algorithm is welcomed.
1987 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1991 mem_cgroup_may_update_nodemask(memcg);
1992 node = memcg->last_scanned_node;
1994 node = next_node(node, memcg->scan_nodes);
1995 if (node == MAX_NUMNODES)
1996 node = first_node(memcg->scan_nodes);
1998 * We call this when we hit limit, not when pages are added to LRU.
1999 * No LRU may hold pages because all pages are UNEVICTABLE or
2000 * memcg is too small and all pages are not on LRU. In that case,
2001 * we use curret node.
2003 if (unlikely(node == MAX_NUMNODES))
2004 node = numa_node_id();
2006 memcg->last_scanned_node = node;
2011 * Check all nodes whether it contains reclaimable pages or not.
2012 * For quick scan, we make use of scan_nodes. This will allow us to skip
2013 * unused nodes. But scan_nodes is lazily updated and may not cotain
2014 * enough new information. We need to do double check.
2016 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2021 * quick check...making use of scan_node.
2022 * We can skip unused nodes.
2024 if (!nodes_empty(memcg->scan_nodes)) {
2025 for (nid = first_node(memcg->scan_nodes);
2027 nid = next_node(nid, memcg->scan_nodes)) {
2029 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2034 * Check rest of nodes.
2036 for_each_node_state(nid, N_MEMORY) {
2037 if (node_isset(nid, memcg->scan_nodes))
2039 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2046 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2051 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2053 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2057 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2060 unsigned long *total_scanned)
2062 struct mem_cgroup *victim = NULL;
2065 unsigned long excess;
2066 unsigned long nr_scanned;
2067 struct mem_cgroup_reclaim_cookie reclaim = {
2072 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2075 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2080 * If we have not been able to reclaim
2081 * anything, it might because there are
2082 * no reclaimable pages under this hierarchy
2087 * We want to do more targeted reclaim.
2088 * excess >> 2 is not to excessive so as to
2089 * reclaim too much, nor too less that we keep
2090 * coming back to reclaim from this cgroup
2092 if (total >= (excess >> 2) ||
2093 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2098 if (!mem_cgroup_reclaimable(victim, false))
2100 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2102 *total_scanned += nr_scanned;
2103 if (!res_counter_soft_limit_excess(&root_memcg->res))
2106 mem_cgroup_iter_break(root_memcg, victim);
2110 #ifdef CONFIG_LOCKDEP
2111 static struct lockdep_map memcg_oom_lock_dep_map = {
2112 .name = "memcg_oom_lock",
2116 static DEFINE_SPINLOCK(memcg_oom_lock);
2119 * Check OOM-Killer is already running under our hierarchy.
2120 * If someone is running, return false.
2122 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2124 struct mem_cgroup *iter, *failed = NULL;
2126 spin_lock(&memcg_oom_lock);
2128 for_each_mem_cgroup_tree(iter, memcg) {
2129 if (iter->oom_lock) {
2131 * this subtree of our hierarchy is already locked
2132 * so we cannot give a lock.
2135 mem_cgroup_iter_break(memcg, iter);
2138 iter->oom_lock = true;
2143 * OK, we failed to lock the whole subtree so we have
2144 * to clean up what we set up to the failing subtree
2146 for_each_mem_cgroup_tree(iter, memcg) {
2147 if (iter == failed) {
2148 mem_cgroup_iter_break(memcg, iter);
2151 iter->oom_lock = false;
2154 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2156 spin_unlock(&memcg_oom_lock);
2161 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2163 struct mem_cgroup *iter;
2165 spin_lock(&memcg_oom_lock);
2166 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2167 for_each_mem_cgroup_tree(iter, memcg)
2168 iter->oom_lock = false;
2169 spin_unlock(&memcg_oom_lock);
2172 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2174 struct mem_cgroup *iter;
2176 for_each_mem_cgroup_tree(iter, memcg)
2177 atomic_inc(&iter->under_oom);
2180 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2182 struct mem_cgroup *iter;
2185 * When a new child is created while the hierarchy is under oom,
2186 * mem_cgroup_oom_lock() may not be called. We have to use
2187 * atomic_add_unless() here.
2189 for_each_mem_cgroup_tree(iter, memcg)
2190 atomic_add_unless(&iter->under_oom, -1, 0);
2193 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2195 struct oom_wait_info {
2196 struct mem_cgroup *memcg;
2200 static int memcg_oom_wake_function(wait_queue_t *wait,
2201 unsigned mode, int sync, void *arg)
2203 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2204 struct mem_cgroup *oom_wait_memcg;
2205 struct oom_wait_info *oom_wait_info;
2207 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2208 oom_wait_memcg = oom_wait_info->memcg;
2211 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2212 * Then we can use css_is_ancestor without taking care of RCU.
2214 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2215 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2217 return autoremove_wake_function(wait, mode, sync, arg);
2220 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2222 atomic_inc(&memcg->oom_wakeups);
2223 /* for filtering, pass "memcg" as argument. */
2224 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2227 static void memcg_oom_recover(struct mem_cgroup *memcg)
2229 if (memcg && atomic_read(&memcg->under_oom))
2230 memcg_wakeup_oom(memcg);
2233 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2235 if (!current->memcg_oom.may_oom)
2238 * We are in the middle of the charge context here, so we
2239 * don't want to block when potentially sitting on a callstack
2240 * that holds all kinds of filesystem and mm locks.
2242 * Also, the caller may handle a failed allocation gracefully
2243 * (like optional page cache readahead) and so an OOM killer
2244 * invocation might not even be necessary.
2246 * That's why we don't do anything here except remember the
2247 * OOM context and then deal with it at the end of the page
2248 * fault when the stack is unwound, the locks are released,
2249 * and when we know whether the fault was overall successful.
2251 css_get(&memcg->css);
2252 current->memcg_oom.memcg = memcg;
2253 current->memcg_oom.gfp_mask = mask;
2254 current->memcg_oom.order = order;
2258 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2259 * @handle: actually kill/wait or just clean up the OOM state
2261 * This has to be called at the end of a page fault if the memcg OOM
2262 * handler was enabled.
2264 * Memcg supports userspace OOM handling where failed allocations must
2265 * sleep on a waitqueue until the userspace task resolves the
2266 * situation. Sleeping directly in the charge context with all kinds
2267 * of locks held is not a good idea, instead we remember an OOM state
2268 * in the task and mem_cgroup_oom_synchronize() has to be called at
2269 * the end of the page fault to complete the OOM handling.
2271 * Returns %true if an ongoing memcg OOM situation was detected and
2272 * completed, %false otherwise.
2274 bool mem_cgroup_oom_synchronize(bool handle)
2276 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2277 struct oom_wait_info owait;
2280 /* OOM is global, do not handle */
2287 owait.memcg = memcg;
2288 owait.wait.flags = 0;
2289 owait.wait.func = memcg_oom_wake_function;
2290 owait.wait.private = current;
2291 INIT_LIST_HEAD(&owait.wait.task_list);
2293 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2294 mem_cgroup_mark_under_oom(memcg);
2296 locked = mem_cgroup_oom_trylock(memcg);
2299 mem_cgroup_oom_notify(memcg);
2301 if (locked && !memcg->oom_kill_disable) {
2302 mem_cgroup_unmark_under_oom(memcg);
2303 finish_wait(&memcg_oom_waitq, &owait.wait);
2304 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2305 current->memcg_oom.order);
2308 mem_cgroup_unmark_under_oom(memcg);
2309 finish_wait(&memcg_oom_waitq, &owait.wait);
2313 mem_cgroup_oom_unlock(memcg);
2315 * There is no guarantee that an OOM-lock contender
2316 * sees the wakeups triggered by the OOM kill
2317 * uncharges. Wake any sleepers explicitely.
2319 memcg_oom_recover(memcg);
2322 current->memcg_oom.memcg = NULL;
2323 css_put(&memcg->css);
2328 * Currently used to update mapped file statistics, but the routine can be
2329 * generalized to update other statistics as well.
2331 * Notes: Race condition
2333 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2334 * it tends to be costly. But considering some conditions, we doesn't need
2335 * to do so _always_.
2337 * Considering "charge", lock_page_cgroup() is not required because all
2338 * file-stat operations happen after a page is attached to radix-tree. There
2339 * are no race with "charge".
2341 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2342 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2343 * if there are race with "uncharge". Statistics itself is properly handled
2346 * Considering "move", this is an only case we see a race. To make the race
2347 * small, we check mm->moving_account and detect there are possibility of race
2348 * If there is, we take a lock.
2351 void __mem_cgroup_begin_update_page_stat(struct page *page,
2352 bool *locked, unsigned long *flags)
2354 struct mem_cgroup *memcg;
2355 struct page_cgroup *pc;
2357 pc = lookup_page_cgroup(page);
2359 memcg = pc->mem_cgroup;
2360 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2363 * If this memory cgroup is not under account moving, we don't
2364 * need to take move_lock_mem_cgroup(). Because we already hold
2365 * rcu_read_lock(), any calls to move_account will be delayed until
2366 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2368 if (!mem_cgroup_stolen(memcg))
2371 move_lock_mem_cgroup(memcg, flags);
2372 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2373 move_unlock_mem_cgroup(memcg, flags);
2379 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2381 struct page_cgroup *pc = lookup_page_cgroup(page);
2384 * It's guaranteed that pc->mem_cgroup never changes while
2385 * lock is held because a routine modifies pc->mem_cgroup
2386 * should take move_lock_mem_cgroup().
2388 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2391 void mem_cgroup_update_page_stat(struct page *page,
2392 enum mem_cgroup_stat_index idx, int val)
2394 struct mem_cgroup *memcg;
2395 struct page_cgroup *pc = lookup_page_cgroup(page);
2396 unsigned long uninitialized_var(flags);
2398 if (mem_cgroup_disabled())
2401 VM_BUG_ON(!rcu_read_lock_held());
2402 memcg = pc->mem_cgroup;
2403 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2406 this_cpu_add(memcg->stat->count[idx], val);
2410 * size of first charge trial. "32" comes from vmscan.c's magic value.
2411 * TODO: maybe necessary to use big numbers in big irons.
2413 #define CHARGE_BATCH 32U
2414 struct memcg_stock_pcp {
2415 struct mem_cgroup *cached; /* this never be root cgroup */
2416 unsigned int nr_pages;
2417 struct work_struct work;
2418 unsigned long flags;
2419 #define FLUSHING_CACHED_CHARGE 0
2421 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2422 static DEFINE_MUTEX(percpu_charge_mutex);
2425 * consume_stock: Try to consume stocked charge on this cpu.
2426 * @memcg: memcg to consume from.
2427 * @nr_pages: how many pages to charge.
2429 * The charges will only happen if @memcg matches the current cpu's memcg
2430 * stock, and at least @nr_pages are available in that stock. Failure to
2431 * service an allocation will refill the stock.
2433 * returns true if successful, false otherwise.
2435 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2437 struct memcg_stock_pcp *stock;
2440 if (nr_pages > CHARGE_BATCH)
2443 stock = &get_cpu_var(memcg_stock);
2444 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2445 stock->nr_pages -= nr_pages;
2446 else /* need to call res_counter_charge */
2448 put_cpu_var(memcg_stock);
2453 * Returns stocks cached in percpu to res_counter and reset cached information.
2455 static void drain_stock(struct memcg_stock_pcp *stock)
2457 struct mem_cgroup *old = stock->cached;
2459 if (stock->nr_pages) {
2460 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2462 res_counter_uncharge(&old->res, bytes);
2463 if (do_swap_account)
2464 res_counter_uncharge(&old->memsw, bytes);
2465 stock->nr_pages = 0;
2467 stock->cached = NULL;
2471 * This must be called under preempt disabled or must be called by
2472 * a thread which is pinned to local cpu.
2474 static void drain_local_stock(struct work_struct *dummy)
2476 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2478 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2481 static void __init memcg_stock_init(void)
2485 for_each_possible_cpu(cpu) {
2486 struct memcg_stock_pcp *stock =
2487 &per_cpu(memcg_stock, cpu);
2488 INIT_WORK(&stock->work, drain_local_stock);
2493 * Cache charges(val) which is from res_counter, to local per_cpu area.
2494 * This will be consumed by consume_stock() function, later.
2496 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2498 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2500 if (stock->cached != memcg) { /* reset if necessary */
2502 stock->cached = memcg;
2504 stock->nr_pages += nr_pages;
2505 put_cpu_var(memcg_stock);
2509 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2510 * of the hierarchy under it. sync flag says whether we should block
2511 * until the work is done.
2513 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2517 /* Notify other cpus that system-wide "drain" is running */
2520 for_each_online_cpu(cpu) {
2521 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2522 struct mem_cgroup *memcg;
2524 memcg = stock->cached;
2525 if (!memcg || !stock->nr_pages)
2527 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2529 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2531 drain_local_stock(&stock->work);
2533 schedule_work_on(cpu, &stock->work);
2541 for_each_online_cpu(cpu) {
2542 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2543 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2544 flush_work(&stock->work);
2551 * Tries to drain stocked charges in other cpus. This function is asynchronous
2552 * and just put a work per cpu for draining localy on each cpu. Caller can
2553 * expects some charges will be back to res_counter later but cannot wait for
2556 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2559 * If someone calls draining, avoid adding more kworker runs.
2561 if (!mutex_trylock(&percpu_charge_mutex))
2563 drain_all_stock(root_memcg, false);
2564 mutex_unlock(&percpu_charge_mutex);
2567 /* This is a synchronous drain interface. */
2568 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2570 /* called when force_empty is called */
2571 mutex_lock(&percpu_charge_mutex);
2572 drain_all_stock(root_memcg, true);
2573 mutex_unlock(&percpu_charge_mutex);
2577 * This function drains percpu counter value from DEAD cpu and
2578 * move it to local cpu. Note that this function can be preempted.
2580 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2584 spin_lock(&memcg->pcp_counter_lock);
2585 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2586 long x = per_cpu(memcg->stat->count[i], cpu);
2588 per_cpu(memcg->stat->count[i], cpu) = 0;
2589 memcg->nocpu_base.count[i] += x;
2591 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2592 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2594 per_cpu(memcg->stat->events[i], cpu) = 0;
2595 memcg->nocpu_base.events[i] += x;
2597 spin_unlock(&memcg->pcp_counter_lock);
2600 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2601 unsigned long action,
2604 int cpu = (unsigned long)hcpu;
2605 struct memcg_stock_pcp *stock;
2606 struct mem_cgroup *iter;
2608 if (action == CPU_ONLINE)
2611 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2614 for_each_mem_cgroup(iter)
2615 mem_cgroup_drain_pcp_counter(iter, cpu);
2617 stock = &per_cpu(memcg_stock, cpu);
2623 /* See __mem_cgroup_try_charge() for details */
2625 CHARGE_OK, /* success */
2626 CHARGE_RETRY, /* need to retry but retry is not bad */
2627 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2628 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2631 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2632 unsigned int nr_pages, unsigned int min_pages,
2635 unsigned long csize = nr_pages * PAGE_SIZE;
2636 struct mem_cgroup *mem_over_limit;
2637 struct res_counter *fail_res;
2638 unsigned long flags = 0;
2641 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2644 if (!do_swap_account)
2646 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2650 res_counter_uncharge(&memcg->res, csize);
2651 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2652 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2654 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2656 * Never reclaim on behalf of optional batching, retry with a
2657 * single page instead.
2659 if (nr_pages > min_pages)
2660 return CHARGE_RETRY;
2662 if (!(gfp_mask & __GFP_WAIT))
2663 return CHARGE_WOULDBLOCK;
2665 if (gfp_mask & __GFP_NORETRY)
2666 return CHARGE_NOMEM;
2668 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2669 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2670 return CHARGE_RETRY;
2672 * Even though the limit is exceeded at this point, reclaim
2673 * may have been able to free some pages. Retry the charge
2674 * before killing the task.
2676 * Only for regular pages, though: huge pages are rather
2677 * unlikely to succeed so close to the limit, and we fall back
2678 * to regular pages anyway in case of failure.
2680 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2681 return CHARGE_RETRY;
2684 * At task move, charge accounts can be doubly counted. So, it's
2685 * better to wait until the end of task_move if something is going on.
2687 if (mem_cgroup_wait_acct_move(mem_over_limit))
2688 return CHARGE_RETRY;
2691 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2693 return CHARGE_NOMEM;
2697 * __mem_cgroup_try_charge() does
2698 * 1. detect memcg to be charged against from passed *mm and *ptr,
2699 * 2. update res_counter
2700 * 3. call memory reclaim if necessary.
2702 * In some special case, if the task is fatal, fatal_signal_pending() or
2703 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2704 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2705 * as possible without any hazards. 2: all pages should have a valid
2706 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2707 * pointer, that is treated as a charge to root_mem_cgroup.
2709 * So __mem_cgroup_try_charge() will return
2710 * 0 ... on success, filling *ptr with a valid memcg pointer.
2711 * -ENOMEM ... charge failure because of resource limits.
2712 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2714 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2715 * the oom-killer can be invoked.
2717 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2719 unsigned int nr_pages,
2720 struct mem_cgroup **ptr,
2723 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2724 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2725 struct mem_cgroup *memcg = NULL;
2729 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2730 * in system level. So, allow to go ahead dying process in addition to
2733 if (unlikely(test_thread_flag(TIF_MEMDIE)
2734 || fatal_signal_pending(current)))
2737 if (unlikely(task_in_memcg_oom(current)))
2740 if (gfp_mask & __GFP_NOFAIL)
2744 * We always charge the cgroup the mm_struct belongs to.
2745 * The mm_struct's mem_cgroup changes on task migration if the
2746 * thread group leader migrates. It's possible that mm is not
2747 * set, if so charge the root memcg (happens for pagecache usage).
2750 *ptr = root_mem_cgroup;
2752 if (*ptr) { /* css should be a valid one */
2754 if (mem_cgroup_is_root(memcg))
2756 if (consume_stock(memcg, nr_pages))
2758 css_get(&memcg->css);
2760 struct task_struct *p;
2763 p = rcu_dereference(mm->owner);
2765 * Because we don't have task_lock(), "p" can exit.
2766 * In that case, "memcg" can point to root or p can be NULL with
2767 * race with swapoff. Then, we have small risk of mis-accouning.
2768 * But such kind of mis-account by race always happens because
2769 * we don't have cgroup_mutex(). It's overkill and we allo that
2771 * (*) swapoff at el will charge against mm-struct not against
2772 * task-struct. So, mm->owner can be NULL.
2774 memcg = mem_cgroup_from_task(p);
2776 memcg = root_mem_cgroup;
2777 if (mem_cgroup_is_root(memcg)) {
2781 if (consume_stock(memcg, nr_pages)) {
2783 * It seems dagerous to access memcg without css_get().
2784 * But considering how consume_stok works, it's not
2785 * necessary. If consume_stock success, some charges
2786 * from this memcg are cached on this cpu. So, we
2787 * don't need to call css_get()/css_tryget() before
2788 * calling consume_stock().
2793 /* after here, we may be blocked. we need to get refcnt */
2794 if (!css_tryget(&memcg->css)) {
2802 bool invoke_oom = oom && !nr_oom_retries;
2804 /* If killed, bypass charge */
2805 if (fatal_signal_pending(current)) {
2806 css_put(&memcg->css);
2810 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2811 nr_pages, invoke_oom);
2815 case CHARGE_RETRY: /* not in OOM situation but retry */
2817 css_put(&memcg->css);
2820 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2821 css_put(&memcg->css);
2823 case CHARGE_NOMEM: /* OOM routine works */
2824 if (!oom || invoke_oom) {
2825 css_put(&memcg->css);
2831 } while (ret != CHARGE_OK);
2833 if (batch > nr_pages)
2834 refill_stock(memcg, batch - nr_pages);
2835 css_put(&memcg->css);
2840 if (!(gfp_mask & __GFP_NOFAIL)) {
2845 *ptr = root_mem_cgroup;
2850 * Somemtimes we have to undo a charge we got by try_charge().
2851 * This function is for that and do uncharge, put css's refcnt.
2852 * gotten by try_charge().
2854 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2855 unsigned int nr_pages)
2857 if (!mem_cgroup_is_root(memcg)) {
2858 unsigned long bytes = nr_pages * PAGE_SIZE;
2860 res_counter_uncharge(&memcg->res, bytes);
2861 if (do_swap_account)
2862 res_counter_uncharge(&memcg->memsw, bytes);
2867 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2868 * This is useful when moving usage to parent cgroup.
2870 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2871 unsigned int nr_pages)
2873 unsigned long bytes = nr_pages * PAGE_SIZE;
2875 if (mem_cgroup_is_root(memcg))
2878 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2879 if (do_swap_account)
2880 res_counter_uncharge_until(&memcg->memsw,
2881 memcg->memsw.parent, bytes);
2885 * A helper function to get mem_cgroup from ID. must be called under
2886 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2887 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2888 * called against removed memcg.)
2890 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2892 /* ID 0 is unused ID */
2895 return mem_cgroup_from_id(id);
2898 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2900 struct mem_cgroup *memcg = NULL;
2901 struct page_cgroup *pc;
2905 VM_BUG_ON(!PageLocked(page));
2907 pc = lookup_page_cgroup(page);
2908 lock_page_cgroup(pc);
2909 if (PageCgroupUsed(pc)) {
2910 memcg = pc->mem_cgroup;
2911 if (memcg && !css_tryget(&memcg->css))
2913 } else if (PageSwapCache(page)) {
2914 ent.val = page_private(page);
2915 id = lookup_swap_cgroup_id(ent);
2917 memcg = mem_cgroup_lookup(id);
2918 if (memcg && !css_tryget(&memcg->css))
2922 unlock_page_cgroup(pc);
2926 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2928 unsigned int nr_pages,
2929 enum charge_type ctype,
2932 struct page_cgroup *pc = lookup_page_cgroup(page);
2933 struct zone *uninitialized_var(zone);
2934 struct lruvec *lruvec;
2935 bool was_on_lru = false;
2938 lock_page_cgroup(pc);
2939 VM_BUG_ON(PageCgroupUsed(pc));
2941 * we don't need page_cgroup_lock about tail pages, becase they are not
2942 * accessed by any other context at this point.
2946 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2947 * may already be on some other mem_cgroup's LRU. Take care of it.
2950 zone = page_zone(page);
2951 spin_lock_irq(&zone->lru_lock);
2952 if (PageLRU(page)) {
2953 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2955 del_page_from_lru_list(page, lruvec, page_lru(page));
2960 pc->mem_cgroup = memcg;
2962 * We access a page_cgroup asynchronously without lock_page_cgroup().
2963 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2964 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2965 * before USED bit, we need memory barrier here.
2966 * See mem_cgroup_add_lru_list(), etc.
2969 SetPageCgroupUsed(pc);
2973 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2974 VM_BUG_ON(PageLRU(page));
2976 add_page_to_lru_list(page, lruvec, page_lru(page));
2978 spin_unlock_irq(&zone->lru_lock);
2981 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2986 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2987 unlock_page_cgroup(pc);
2990 * "charge_statistics" updated event counter. Then, check it.
2991 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2992 * if they exceeds softlimit.
2994 memcg_check_events(memcg, page);
2997 static DEFINE_MUTEX(set_limit_mutex);
2999 #ifdef CONFIG_MEMCG_KMEM
3000 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
3002 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
3003 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
3007 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
3008 * in the memcg_cache_params struct.
3010 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
3012 struct kmem_cache *cachep;
3014 VM_BUG_ON(p->is_root_cache);
3015 cachep = p->root_cache;
3016 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
3019 #ifdef CONFIG_SLABINFO
3020 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
3022 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3023 struct memcg_cache_params *params;
3025 if (!memcg_can_account_kmem(memcg))
3028 print_slabinfo_header(m);
3030 mutex_lock(&memcg->slab_caches_mutex);
3031 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
3032 cache_show(memcg_params_to_cache(params), m);
3033 mutex_unlock(&memcg->slab_caches_mutex);
3039 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3041 struct res_counter *fail_res;
3042 struct mem_cgroup *_memcg;
3045 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3050 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3051 &_memcg, oom_gfp_allowed(gfp));
3053 if (ret == -EINTR) {
3055 * __mem_cgroup_try_charge() chosed to bypass to root due to
3056 * OOM kill or fatal signal. Since our only options are to
3057 * either fail the allocation or charge it to this cgroup, do
3058 * it as a temporary condition. But we can't fail. From a
3059 * kmem/slab perspective, the cache has already been selected,
3060 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3063 * This condition will only trigger if the task entered
3064 * memcg_charge_kmem in a sane state, but was OOM-killed during
3065 * __mem_cgroup_try_charge() above. Tasks that were already
3066 * dying when the allocation triggers should have been already
3067 * directed to the root cgroup in memcontrol.h
3069 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3070 if (do_swap_account)
3071 res_counter_charge_nofail(&memcg->memsw, size,
3075 res_counter_uncharge(&memcg->kmem, size);
3080 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3082 res_counter_uncharge(&memcg->res, size);
3083 if (do_swap_account)
3084 res_counter_uncharge(&memcg->memsw, size);
3087 if (res_counter_uncharge(&memcg->kmem, size))
3091 * Releases a reference taken in kmem_cgroup_css_offline in case
3092 * this last uncharge is racing with the offlining code or it is
3093 * outliving the memcg existence.
3095 * The memory barrier imposed by test&clear is paired with the
3096 * explicit one in memcg_kmem_mark_dead().
3098 if (memcg_kmem_test_and_clear_dead(memcg))
3099 css_put(&memcg->css);
3102 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3107 mutex_lock(&memcg->slab_caches_mutex);
3108 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3109 mutex_unlock(&memcg->slab_caches_mutex);
3113 * helper for acessing a memcg's index. It will be used as an index in the
3114 * child cache array in kmem_cache, and also to derive its name. This function
3115 * will return -1 when this is not a kmem-limited memcg.
3117 int memcg_cache_id(struct mem_cgroup *memcg)
3119 return memcg ? memcg->kmemcg_id : -1;
3123 * This ends up being protected by the set_limit mutex, during normal
3124 * operation, because that is its main call site.
3126 * But when we create a new cache, we can call this as well if its parent
3127 * is kmem-limited. That will have to hold set_limit_mutex as well.
3129 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3133 num = ida_simple_get(&kmem_limited_groups,
3134 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3138 * After this point, kmem_accounted (that we test atomically in
3139 * the beginning of this conditional), is no longer 0. This
3140 * guarantees only one process will set the following boolean
3141 * to true. We don't need test_and_set because we're protected
3142 * by the set_limit_mutex anyway.
3144 memcg_kmem_set_activated(memcg);
3146 ret = memcg_update_all_caches(num+1);
3148 ida_simple_remove(&kmem_limited_groups, num);
3149 memcg_kmem_clear_activated(memcg);
3153 memcg->kmemcg_id = num;
3154 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3155 mutex_init(&memcg->slab_caches_mutex);
3159 static size_t memcg_caches_array_size(int num_groups)
3162 if (num_groups <= 0)
3165 size = 2 * num_groups;
3166 if (size < MEMCG_CACHES_MIN_SIZE)
3167 size = MEMCG_CACHES_MIN_SIZE;
3168 else if (size > MEMCG_CACHES_MAX_SIZE)
3169 size = MEMCG_CACHES_MAX_SIZE;
3175 * We should update the current array size iff all caches updates succeed. This
3176 * can only be done from the slab side. The slab mutex needs to be held when
3179 void memcg_update_array_size(int num)
3181 if (num > memcg_limited_groups_array_size)
3182 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3185 static void kmem_cache_destroy_work_func(struct work_struct *w);
3187 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3189 struct memcg_cache_params *cur_params = s->memcg_params;
3191 VM_BUG_ON(!is_root_cache(s));
3193 if (num_groups > memcg_limited_groups_array_size) {
3195 ssize_t size = memcg_caches_array_size(num_groups);
3197 size *= sizeof(void *);
3198 size += offsetof(struct memcg_cache_params, memcg_caches);
3200 s->memcg_params = kzalloc(size, GFP_KERNEL);
3201 if (!s->memcg_params) {
3202 s->memcg_params = cur_params;
3206 s->memcg_params->is_root_cache = true;
3209 * There is the chance it will be bigger than
3210 * memcg_limited_groups_array_size, if we failed an allocation
3211 * in a cache, in which case all caches updated before it, will
3212 * have a bigger array.
3214 * But if that is the case, the data after
3215 * memcg_limited_groups_array_size is certainly unused
3217 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3218 if (!cur_params->memcg_caches[i])
3220 s->memcg_params->memcg_caches[i] =
3221 cur_params->memcg_caches[i];
3225 * Ideally, we would wait until all caches succeed, and only
3226 * then free the old one. But this is not worth the extra
3227 * pointer per-cache we'd have to have for this.
3229 * It is not a big deal if some caches are left with a size
3230 * bigger than the others. And all updates will reset this
3238 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3239 struct kmem_cache *root_cache)
3243 if (!memcg_kmem_enabled())
3247 size = offsetof(struct memcg_cache_params, memcg_caches);
3248 size += memcg_limited_groups_array_size * sizeof(void *);
3250 size = sizeof(struct memcg_cache_params);
3252 s->memcg_params = kzalloc(size, GFP_KERNEL);
3253 if (!s->memcg_params)
3257 s->memcg_params->memcg = memcg;
3258 s->memcg_params->root_cache = root_cache;
3259 INIT_WORK(&s->memcg_params->destroy,
3260 kmem_cache_destroy_work_func);
3262 s->memcg_params->is_root_cache = true;
3267 void memcg_release_cache(struct kmem_cache *s)
3269 struct kmem_cache *root;
3270 struct mem_cgroup *memcg;
3274 * This happens, for instance, when a root cache goes away before we
3277 if (!s->memcg_params)
3280 if (s->memcg_params->is_root_cache)
3283 memcg = s->memcg_params->memcg;
3284 id = memcg_cache_id(memcg);
3286 root = s->memcg_params->root_cache;
3287 root->memcg_params->memcg_caches[id] = NULL;
3289 mutex_lock(&memcg->slab_caches_mutex);
3290 list_del(&s->memcg_params->list);
3291 mutex_unlock(&memcg->slab_caches_mutex);
3293 css_put(&memcg->css);
3295 kfree(s->memcg_params);
3299 * During the creation a new cache, we need to disable our accounting mechanism
3300 * altogether. This is true even if we are not creating, but rather just
3301 * enqueing new caches to be created.
3303 * This is because that process will trigger allocations; some visible, like
3304 * explicit kmallocs to auxiliary data structures, name strings and internal
3305 * cache structures; some well concealed, like INIT_WORK() that can allocate
3306 * objects during debug.
3308 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3309 * to it. This may not be a bounded recursion: since the first cache creation
3310 * failed to complete (waiting on the allocation), we'll just try to create the
3311 * cache again, failing at the same point.
3313 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3314 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3315 * inside the following two functions.
3317 static inline void memcg_stop_kmem_account(void)
3319 VM_BUG_ON(!current->mm);
3320 current->memcg_kmem_skip_account++;
3323 static inline void memcg_resume_kmem_account(void)
3325 VM_BUG_ON(!current->mm);
3326 current->memcg_kmem_skip_account--;
3329 static void kmem_cache_destroy_work_func(struct work_struct *w)
3331 struct kmem_cache *cachep;
3332 struct memcg_cache_params *p;
3334 p = container_of(w, struct memcg_cache_params, destroy);
3336 cachep = memcg_params_to_cache(p);
3339 * If we get down to 0 after shrink, we could delete right away.
3340 * However, memcg_release_pages() already puts us back in the workqueue
3341 * in that case. If we proceed deleting, we'll get a dangling
3342 * reference, and removing the object from the workqueue in that case
3343 * is unnecessary complication. We are not a fast path.
3345 * Note that this case is fundamentally different from racing with
3346 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3347 * kmem_cache_shrink, not only we would be reinserting a dead cache
3348 * into the queue, but doing so from inside the worker racing to
3351 * So if we aren't down to zero, we'll just schedule a worker and try
3354 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3355 kmem_cache_shrink(cachep);
3356 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3359 kmem_cache_destroy(cachep);
3362 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3364 if (!cachep->memcg_params->dead)
3368 * There are many ways in which we can get here.
3370 * We can get to a memory-pressure situation while the delayed work is
3371 * still pending to run. The vmscan shrinkers can then release all
3372 * cache memory and get us to destruction. If this is the case, we'll
3373 * be executed twice, which is a bug (the second time will execute over
3374 * bogus data). In this case, cancelling the work should be fine.
3376 * But we can also get here from the worker itself, if
3377 * kmem_cache_shrink is enough to shake all the remaining objects and
3378 * get the page count to 0. In this case, we'll deadlock if we try to
3379 * cancel the work (the worker runs with an internal lock held, which
3380 * is the same lock we would hold for cancel_work_sync().)
3382 * Since we can't possibly know who got us here, just refrain from
3383 * running if there is already work pending
3385 if (work_pending(&cachep->memcg_params->destroy))
3388 * We have to defer the actual destroying to a workqueue, because
3389 * we might currently be in a context that cannot sleep.
3391 schedule_work(&cachep->memcg_params->destroy);
3395 * This lock protects updaters, not readers. We want readers to be as fast as
3396 * they can, and they will either see NULL or a valid cache value. Our model
3397 * allow them to see NULL, in which case the root memcg will be selected.
3399 * We need this lock because multiple allocations to the same cache from a non
3400 * will span more than one worker. Only one of them can create the cache.
3402 static DEFINE_MUTEX(memcg_cache_mutex);
3405 * Called with memcg_cache_mutex held
3407 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3408 struct kmem_cache *s)
3410 struct kmem_cache *new;
3411 static char *tmp_name = NULL;
3413 lockdep_assert_held(&memcg_cache_mutex);
3416 * kmem_cache_create_memcg duplicates the given name and
3417 * cgroup_name for this name requires RCU context.
3418 * This static temporary buffer is used to prevent from
3419 * pointless shortliving allocation.
3422 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3428 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3429 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3432 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3433 (s->flags & ~SLAB_PANIC), s->ctor, s);
3436 new->allocflags |= __GFP_KMEMCG;
3441 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3442 struct kmem_cache *cachep)
3444 struct kmem_cache *new_cachep;
3447 BUG_ON(!memcg_can_account_kmem(memcg));
3449 idx = memcg_cache_id(memcg);
3451 mutex_lock(&memcg_cache_mutex);
3452 new_cachep = cache_from_memcg_idx(cachep, idx);
3454 css_put(&memcg->css);
3458 new_cachep = kmem_cache_dup(memcg, cachep);
3459 if (new_cachep == NULL) {
3460 new_cachep = cachep;
3461 css_put(&memcg->css);
3465 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3467 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3469 * the readers won't lock, make sure everybody sees the updated value,
3470 * so they won't put stuff in the queue again for no reason
3474 mutex_unlock(&memcg_cache_mutex);
3478 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3480 struct kmem_cache *c;
3483 if (!s->memcg_params)
3485 if (!s->memcg_params->is_root_cache)
3489 * If the cache is being destroyed, we trust that there is no one else
3490 * requesting objects from it. Even if there are, the sanity checks in
3491 * kmem_cache_destroy should caught this ill-case.
3493 * Still, we don't want anyone else freeing memcg_caches under our
3494 * noses, which can happen if a new memcg comes to life. As usual,
3495 * we'll take the set_limit_mutex to protect ourselves against this.
3497 mutex_lock(&set_limit_mutex);
3498 for_each_memcg_cache_index(i) {
3499 c = cache_from_memcg_idx(s, i);
3504 * We will now manually delete the caches, so to avoid races
3505 * we need to cancel all pending destruction workers and
3506 * proceed with destruction ourselves.
3508 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3509 * and that could spawn the workers again: it is likely that
3510 * the cache still have active pages until this very moment.
3511 * This would lead us back to mem_cgroup_destroy_cache.
3513 * But that will not execute at all if the "dead" flag is not
3514 * set, so flip it down to guarantee we are in control.
3516 c->memcg_params->dead = false;
3517 cancel_work_sync(&c->memcg_params->destroy);
3518 kmem_cache_destroy(c);
3520 mutex_unlock(&set_limit_mutex);
3523 struct create_work {
3524 struct mem_cgroup *memcg;
3525 struct kmem_cache *cachep;
3526 struct work_struct work;
3529 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3531 struct kmem_cache *cachep;
3532 struct memcg_cache_params *params;
3534 if (!memcg_kmem_is_active(memcg))
3537 mutex_lock(&memcg->slab_caches_mutex);
3538 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3539 cachep = memcg_params_to_cache(params);
3540 cachep->memcg_params->dead = true;
3541 schedule_work(&cachep->memcg_params->destroy);
3543 mutex_unlock(&memcg->slab_caches_mutex);
3546 static void memcg_create_cache_work_func(struct work_struct *w)
3548 struct create_work *cw;
3550 cw = container_of(w, struct create_work, work);
3551 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3556 * Enqueue the creation of a per-memcg kmem_cache.
3558 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3559 struct kmem_cache *cachep)
3561 struct create_work *cw;
3563 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3565 css_put(&memcg->css);
3570 cw->cachep = cachep;
3572 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3573 schedule_work(&cw->work);
3576 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3577 struct kmem_cache *cachep)
3580 * We need to stop accounting when we kmalloc, because if the
3581 * corresponding kmalloc cache is not yet created, the first allocation
3582 * in __memcg_create_cache_enqueue will recurse.
3584 * However, it is better to enclose the whole function. Depending on
3585 * the debugging options enabled, INIT_WORK(), for instance, can
3586 * trigger an allocation. This too, will make us recurse. Because at
3587 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3588 * the safest choice is to do it like this, wrapping the whole function.
3590 memcg_stop_kmem_account();
3591 __memcg_create_cache_enqueue(memcg, cachep);
3592 memcg_resume_kmem_account();
3595 * Return the kmem_cache we're supposed to use for a slab allocation.
3596 * We try to use the current memcg's version of the cache.
3598 * If the cache does not exist yet, if we are the first user of it,
3599 * we either create it immediately, if possible, or create it asynchronously
3601 * In the latter case, we will let the current allocation go through with
3602 * the original cache.
3604 * Can't be called in interrupt context or from kernel threads.
3605 * This function needs to be called with rcu_read_lock() held.
3607 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3610 struct mem_cgroup *memcg;
3613 VM_BUG_ON(!cachep->memcg_params);
3614 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3616 if (!current->mm || current->memcg_kmem_skip_account)
3620 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3622 if (!memcg_can_account_kmem(memcg))
3625 idx = memcg_cache_id(memcg);
3628 * barrier to mare sure we're always seeing the up to date value. The
3629 * code updating memcg_caches will issue a write barrier to match this.
3631 read_barrier_depends();
3632 if (likely(cache_from_memcg_idx(cachep, idx))) {
3633 cachep = cache_from_memcg_idx(cachep, idx);
3637 /* The corresponding put will be done in the workqueue. */
3638 if (!css_tryget(&memcg->css))
3643 * If we are in a safe context (can wait, and not in interrupt
3644 * context), we could be be predictable and return right away.
3645 * This would guarantee that the allocation being performed
3646 * already belongs in the new cache.
3648 * However, there are some clashes that can arrive from locking.
3649 * For instance, because we acquire the slab_mutex while doing
3650 * kmem_cache_dup, this means no further allocation could happen
3651 * with the slab_mutex held.
3653 * Also, because cache creation issue get_online_cpus(), this
3654 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3655 * that ends up reversed during cpu hotplug. (cpuset allocates
3656 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3657 * better to defer everything.
3659 memcg_create_cache_enqueue(memcg, cachep);
3665 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3668 * We need to verify if the allocation against current->mm->owner's memcg is
3669 * possible for the given order. But the page is not allocated yet, so we'll
3670 * need a further commit step to do the final arrangements.
3672 * It is possible for the task to switch cgroups in this mean time, so at
3673 * commit time, we can't rely on task conversion any longer. We'll then use
3674 * the handle argument to return to the caller which cgroup we should commit
3675 * against. We could also return the memcg directly and avoid the pointer
3676 * passing, but a boolean return value gives better semantics considering
3677 * the compiled-out case as well.
3679 * Returning true means the allocation is possible.
3682 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3684 struct mem_cgroup *memcg;
3690 * Disabling accounting is only relevant for some specific memcg
3691 * internal allocations. Therefore we would initially not have such
3692 * check here, since direct calls to the page allocator that are marked
3693 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3694 * concerned with cache allocations, and by having this test at
3695 * memcg_kmem_get_cache, we are already able to relay the allocation to
3696 * the root cache and bypass the memcg cache altogether.
3698 * There is one exception, though: the SLUB allocator does not create
3699 * large order caches, but rather service large kmallocs directly from
3700 * the page allocator. Therefore, the following sequence when backed by
3701 * the SLUB allocator:
3703 * memcg_stop_kmem_account();
3704 * kmalloc(<large_number>)
3705 * memcg_resume_kmem_account();
3707 * would effectively ignore the fact that we should skip accounting,
3708 * since it will drive us directly to this function without passing
3709 * through the cache selector memcg_kmem_get_cache. Such large
3710 * allocations are extremely rare but can happen, for instance, for the
3711 * cache arrays. We bring this test here.
3713 if (!current->mm || current->memcg_kmem_skip_account)
3716 memcg = try_get_mem_cgroup_from_mm(current->mm);
3719 * very rare case described in mem_cgroup_from_task. Unfortunately there
3720 * isn't much we can do without complicating this too much, and it would
3721 * be gfp-dependent anyway. Just let it go
3723 if (unlikely(!memcg))
3726 if (!memcg_can_account_kmem(memcg)) {
3727 css_put(&memcg->css);
3731 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3735 css_put(&memcg->css);
3739 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3742 struct page_cgroup *pc;
3744 VM_BUG_ON(mem_cgroup_is_root(memcg));
3746 /* The page allocation failed. Revert */
3748 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3752 pc = lookup_page_cgroup(page);
3753 lock_page_cgroup(pc);
3754 pc->mem_cgroup = memcg;
3755 SetPageCgroupUsed(pc);
3756 unlock_page_cgroup(pc);
3759 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3761 struct mem_cgroup *memcg = NULL;
3762 struct page_cgroup *pc;
3765 pc = lookup_page_cgroup(page);
3767 * Fast unlocked return. Theoretically might have changed, have to
3768 * check again after locking.
3770 if (!PageCgroupUsed(pc))
3773 lock_page_cgroup(pc);
3774 if (PageCgroupUsed(pc)) {
3775 memcg = pc->mem_cgroup;
3776 ClearPageCgroupUsed(pc);
3778 unlock_page_cgroup(pc);
3781 * We trust that only if there is a memcg associated with the page, it
3782 * is a valid allocation
3787 VM_BUG_ON(mem_cgroup_is_root(memcg));
3788 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3791 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3794 #endif /* CONFIG_MEMCG_KMEM */
3796 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3798 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3800 * Because tail pages are not marked as "used", set it. We're under
3801 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3802 * charge/uncharge will be never happen and move_account() is done under
3803 * compound_lock(), so we don't have to take care of races.
3805 void mem_cgroup_split_huge_fixup(struct page *head)
3807 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3808 struct page_cgroup *pc;
3809 struct mem_cgroup *memcg;
3812 if (mem_cgroup_disabled())
3815 memcg = head_pc->mem_cgroup;
3816 for (i = 1; i < HPAGE_PMD_NR; i++) {
3818 pc->mem_cgroup = memcg;
3819 smp_wmb();/* see __commit_charge() */
3820 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3822 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3825 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3828 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3829 struct mem_cgroup *to,
3830 unsigned int nr_pages,
3831 enum mem_cgroup_stat_index idx)
3833 /* Update stat data for mem_cgroup */
3835 __this_cpu_sub(from->stat->count[idx], nr_pages);
3836 __this_cpu_add(to->stat->count[idx], nr_pages);
3841 * mem_cgroup_move_account - move account of the page
3843 * @nr_pages: number of regular pages (>1 for huge pages)
3844 * @pc: page_cgroup of the page.
3845 * @from: mem_cgroup which the page is moved from.
3846 * @to: mem_cgroup which the page is moved to. @from != @to.
3848 * The caller must confirm following.
3849 * - page is not on LRU (isolate_page() is useful.)
3850 * - compound_lock is held when nr_pages > 1
3852 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3855 static int mem_cgroup_move_account(struct page *page,
3856 unsigned int nr_pages,
3857 struct page_cgroup *pc,
3858 struct mem_cgroup *from,
3859 struct mem_cgroup *to)
3861 unsigned long flags;
3863 bool anon = PageAnon(page);
3865 VM_BUG_ON(from == to);
3866 VM_BUG_ON(PageLRU(page));
3868 * The page is isolated from LRU. So, collapse function
3869 * will not handle this page. But page splitting can happen.
3870 * Do this check under compound_page_lock(). The caller should
3874 if (nr_pages > 1 && !PageTransHuge(page))
3877 lock_page_cgroup(pc);
3880 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3883 move_lock_mem_cgroup(from, &flags);
3885 if (!anon && page_mapped(page))
3886 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3887 MEM_CGROUP_STAT_FILE_MAPPED);
3889 if (PageWriteback(page))
3890 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3891 MEM_CGROUP_STAT_WRITEBACK);
3893 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3895 /* caller should have done css_get */
3896 pc->mem_cgroup = to;
3897 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3898 move_unlock_mem_cgroup(from, &flags);
3901 unlock_page_cgroup(pc);
3905 memcg_check_events(to, page);
3906 memcg_check_events(from, page);
3912 * mem_cgroup_move_parent - moves page to the parent group
3913 * @page: the page to move
3914 * @pc: page_cgroup of the page
3915 * @child: page's cgroup
3917 * move charges to its parent or the root cgroup if the group has no
3918 * parent (aka use_hierarchy==0).
3919 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3920 * mem_cgroup_move_account fails) the failure is always temporary and
3921 * it signals a race with a page removal/uncharge or migration. In the
3922 * first case the page is on the way out and it will vanish from the LRU
3923 * on the next attempt and the call should be retried later.
3924 * Isolation from the LRU fails only if page has been isolated from
3925 * the LRU since we looked at it and that usually means either global
3926 * reclaim or migration going on. The page will either get back to the
3928 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3929 * (!PageCgroupUsed) or moved to a different group. The page will
3930 * disappear in the next attempt.
3932 static int mem_cgroup_move_parent(struct page *page,
3933 struct page_cgroup *pc,
3934 struct mem_cgroup *child)
3936 struct mem_cgroup *parent;
3937 unsigned int nr_pages;
3938 unsigned long uninitialized_var(flags);
3941 VM_BUG_ON(mem_cgroup_is_root(child));
3944 if (!get_page_unless_zero(page))
3946 if (isolate_lru_page(page))
3949 nr_pages = hpage_nr_pages(page);
3951 parent = parent_mem_cgroup(child);
3953 * If no parent, move charges to root cgroup.
3956 parent = root_mem_cgroup;
3959 VM_BUG_ON(!PageTransHuge(page));
3960 flags = compound_lock_irqsave(page);
3963 ret = mem_cgroup_move_account(page, nr_pages,
3966 __mem_cgroup_cancel_local_charge(child, nr_pages);
3969 compound_unlock_irqrestore(page, flags);
3970 putback_lru_page(page);
3978 * Charge the memory controller for page usage.
3980 * 0 if the charge was successful
3981 * < 0 if the cgroup is over its limit
3983 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3984 gfp_t gfp_mask, enum charge_type ctype)
3986 struct mem_cgroup *memcg = NULL;
3987 unsigned int nr_pages = 1;
3991 if (PageTransHuge(page)) {
3992 nr_pages <<= compound_order(page);
3993 VM_BUG_ON(!PageTransHuge(page));
3995 * Never OOM-kill a process for a huge page. The
3996 * fault handler will fall back to regular pages.
4001 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
4004 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
4008 int mem_cgroup_newpage_charge(struct page *page,
4009 struct mm_struct *mm, gfp_t gfp_mask)
4011 if (mem_cgroup_disabled())
4013 VM_BUG_ON(page_mapped(page));
4014 VM_BUG_ON(page->mapping && !PageAnon(page));
4016 return mem_cgroup_charge_common(page, mm, gfp_mask,
4017 MEM_CGROUP_CHARGE_TYPE_ANON);
4021 * While swap-in, try_charge -> commit or cancel, the page is locked.
4022 * And when try_charge() successfully returns, one refcnt to memcg without
4023 * struct page_cgroup is acquired. This refcnt will be consumed by
4024 * "commit()" or removed by "cancel()"
4026 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
4029 struct mem_cgroup **memcgp)
4031 struct mem_cgroup *memcg;
4032 struct page_cgroup *pc;
4035 pc = lookup_page_cgroup(page);
4037 * Every swap fault against a single page tries to charge the
4038 * page, bail as early as possible. shmem_unuse() encounters
4039 * already charged pages, too. The USED bit is protected by
4040 * the page lock, which serializes swap cache removal, which
4041 * in turn serializes uncharging.
4043 if (PageCgroupUsed(pc))
4045 if (!do_swap_account)
4047 memcg = try_get_mem_cgroup_from_page(page);
4051 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4052 css_put(&memcg->css);
4057 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4063 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4064 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4067 if (mem_cgroup_disabled())
4070 * A racing thread's fault, or swapoff, may have already
4071 * updated the pte, and even removed page from swap cache: in
4072 * those cases unuse_pte()'s pte_same() test will fail; but
4073 * there's also a KSM case which does need to charge the page.
4075 if (!PageSwapCache(page)) {
4078 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4083 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4086 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4088 if (mem_cgroup_disabled())
4092 __mem_cgroup_cancel_charge(memcg, 1);
4096 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4097 enum charge_type ctype)
4099 if (mem_cgroup_disabled())
4104 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4106 * Now swap is on-memory. This means this page may be
4107 * counted both as mem and swap....double count.
4108 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4109 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4110 * may call delete_from_swap_cache() before reach here.
4112 if (do_swap_account && PageSwapCache(page)) {
4113 swp_entry_t ent = {.val = page_private(page)};
4114 mem_cgroup_uncharge_swap(ent);
4118 void mem_cgroup_commit_charge_swapin(struct page *page,
4119 struct mem_cgroup *memcg)
4121 __mem_cgroup_commit_charge_swapin(page, memcg,
4122 MEM_CGROUP_CHARGE_TYPE_ANON);
4125 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4128 struct mem_cgroup *memcg = NULL;
4129 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4132 if (mem_cgroup_disabled())
4134 if (PageCompound(page))
4137 if (!PageSwapCache(page))
4138 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4139 else { /* page is swapcache/shmem */
4140 ret = __mem_cgroup_try_charge_swapin(mm, page,
4143 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4148 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4149 unsigned int nr_pages,
4150 const enum charge_type ctype)
4152 struct memcg_batch_info *batch = NULL;
4153 bool uncharge_memsw = true;
4155 /* If swapout, usage of swap doesn't decrease */
4156 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4157 uncharge_memsw = false;
4159 batch = ¤t->memcg_batch;
4161 * In usual, we do css_get() when we remember memcg pointer.
4162 * But in this case, we keep res->usage until end of a series of
4163 * uncharges. Then, it's ok to ignore memcg's refcnt.
4166 batch->memcg = memcg;
4168 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4169 * In those cases, all pages freed continuously can be expected to be in
4170 * the same cgroup and we have chance to coalesce uncharges.
4171 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4172 * because we want to do uncharge as soon as possible.
4175 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4176 goto direct_uncharge;
4179 goto direct_uncharge;
4182 * In typical case, batch->memcg == mem. This means we can
4183 * merge a series of uncharges to an uncharge of res_counter.
4184 * If not, we uncharge res_counter ony by one.
4186 if (batch->memcg != memcg)
4187 goto direct_uncharge;
4188 /* remember freed charge and uncharge it later */
4191 batch->memsw_nr_pages++;
4194 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4196 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4197 if (unlikely(batch->memcg != memcg))
4198 memcg_oom_recover(memcg);
4202 * uncharge if !page_mapped(page)
4204 static struct mem_cgroup *
4205 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4208 struct mem_cgroup *memcg = NULL;
4209 unsigned int nr_pages = 1;
4210 struct page_cgroup *pc;
4213 if (mem_cgroup_disabled())
4216 if (PageTransHuge(page)) {
4217 nr_pages <<= compound_order(page);
4218 VM_BUG_ON(!PageTransHuge(page));
4221 * Check if our page_cgroup is valid
4223 pc = lookup_page_cgroup(page);
4224 if (unlikely(!PageCgroupUsed(pc)))
4227 lock_page_cgroup(pc);
4229 memcg = pc->mem_cgroup;
4231 if (!PageCgroupUsed(pc))
4234 anon = PageAnon(page);
4237 case MEM_CGROUP_CHARGE_TYPE_ANON:
4239 * Generally PageAnon tells if it's the anon statistics to be
4240 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4241 * used before page reached the stage of being marked PageAnon.
4245 case MEM_CGROUP_CHARGE_TYPE_DROP:
4246 /* See mem_cgroup_prepare_migration() */
4247 if (page_mapped(page))
4250 * Pages under migration may not be uncharged. But
4251 * end_migration() /must/ be the one uncharging the
4252 * unused post-migration page and so it has to call
4253 * here with the migration bit still set. See the
4254 * res_counter handling below.
4256 if (!end_migration && PageCgroupMigration(pc))
4259 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4260 if (!PageAnon(page)) { /* Shared memory */
4261 if (page->mapping && !page_is_file_cache(page))
4263 } else if (page_mapped(page)) /* Anon */
4270 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4272 ClearPageCgroupUsed(pc);
4274 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4275 * freed from LRU. This is safe because uncharged page is expected not
4276 * to be reused (freed soon). Exception is SwapCache, it's handled by
4277 * special functions.
4280 unlock_page_cgroup(pc);
4282 * even after unlock, we have memcg->res.usage here and this memcg
4283 * will never be freed, so it's safe to call css_get().
4285 memcg_check_events(memcg, page);
4286 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4287 mem_cgroup_swap_statistics(memcg, true);
4288 css_get(&memcg->css);
4291 * Migration does not charge the res_counter for the
4292 * replacement page, so leave it alone when phasing out the
4293 * page that is unused after the migration.
4295 if (!end_migration && !mem_cgroup_is_root(memcg))
4296 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4301 unlock_page_cgroup(pc);
4305 void mem_cgroup_uncharge_page(struct page *page)
4308 if (page_mapped(page))
4310 VM_BUG_ON(page->mapping && !PageAnon(page));
4312 * If the page is in swap cache, uncharge should be deferred
4313 * to the swap path, which also properly accounts swap usage
4314 * and handles memcg lifetime.
4316 * Note that this check is not stable and reclaim may add the
4317 * page to swap cache at any time after this. However, if the
4318 * page is not in swap cache by the time page->mapcount hits
4319 * 0, there won't be any page table references to the swap
4320 * slot, and reclaim will free it and not actually write the
4323 if (PageSwapCache(page))
4325 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4328 void mem_cgroup_uncharge_cache_page(struct page *page)
4330 VM_BUG_ON(page_mapped(page));
4331 VM_BUG_ON(page->mapping);
4332 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4336 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4337 * In that cases, pages are freed continuously and we can expect pages
4338 * are in the same memcg. All these calls itself limits the number of
4339 * pages freed at once, then uncharge_start/end() is called properly.
4340 * This may be called prural(2) times in a context,
4343 void mem_cgroup_uncharge_start(void)
4345 current->memcg_batch.do_batch++;
4346 /* We can do nest. */
4347 if (current->memcg_batch.do_batch == 1) {
4348 current->memcg_batch.memcg = NULL;
4349 current->memcg_batch.nr_pages = 0;
4350 current->memcg_batch.memsw_nr_pages = 0;
4354 void mem_cgroup_uncharge_end(void)
4356 struct memcg_batch_info *batch = ¤t->memcg_batch;
4358 if (!batch->do_batch)
4362 if (batch->do_batch) /* If stacked, do nothing. */
4368 * This "batch->memcg" is valid without any css_get/put etc...
4369 * bacause we hide charges behind us.
4371 if (batch->nr_pages)
4372 res_counter_uncharge(&batch->memcg->res,
4373 batch->nr_pages * PAGE_SIZE);
4374 if (batch->memsw_nr_pages)
4375 res_counter_uncharge(&batch->memcg->memsw,
4376 batch->memsw_nr_pages * PAGE_SIZE);
4377 memcg_oom_recover(batch->memcg);
4378 /* forget this pointer (for sanity check) */
4379 batch->memcg = NULL;
4384 * called after __delete_from_swap_cache() and drop "page" account.
4385 * memcg information is recorded to swap_cgroup of "ent"
4388 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4390 struct mem_cgroup *memcg;
4391 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4393 if (!swapout) /* this was a swap cache but the swap is unused ! */
4394 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4396 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4399 * record memcg information, if swapout && memcg != NULL,
4400 * css_get() was called in uncharge().
4402 if (do_swap_account && swapout && memcg)
4403 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4407 #ifdef CONFIG_MEMCG_SWAP
4409 * called from swap_entry_free(). remove record in swap_cgroup and
4410 * uncharge "memsw" account.
4412 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4414 struct mem_cgroup *memcg;
4417 if (!do_swap_account)
4420 id = swap_cgroup_record(ent, 0);
4422 memcg = mem_cgroup_lookup(id);
4425 * We uncharge this because swap is freed.
4426 * This memcg can be obsolete one. We avoid calling css_tryget
4428 if (!mem_cgroup_is_root(memcg))
4429 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4430 mem_cgroup_swap_statistics(memcg, false);
4431 css_put(&memcg->css);
4437 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4438 * @entry: swap entry to be moved
4439 * @from: mem_cgroup which the entry is moved from
4440 * @to: mem_cgroup which the entry is moved to
4442 * It succeeds only when the swap_cgroup's record for this entry is the same
4443 * as the mem_cgroup's id of @from.
4445 * Returns 0 on success, -EINVAL on failure.
4447 * The caller must have charged to @to, IOW, called res_counter_charge() about
4448 * both res and memsw, and called css_get().
4450 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4451 struct mem_cgroup *from, struct mem_cgroup *to)
4453 unsigned short old_id, new_id;
4455 old_id = mem_cgroup_id(from);
4456 new_id = mem_cgroup_id(to);
4458 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4459 mem_cgroup_swap_statistics(from, false);
4460 mem_cgroup_swap_statistics(to, true);
4462 * This function is only called from task migration context now.
4463 * It postpones res_counter and refcount handling till the end
4464 * of task migration(mem_cgroup_clear_mc()) for performance
4465 * improvement. But we cannot postpone css_get(to) because if
4466 * the process that has been moved to @to does swap-in, the
4467 * refcount of @to might be decreased to 0.
4469 * We are in attach() phase, so the cgroup is guaranteed to be
4470 * alive, so we can just call css_get().
4478 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4479 struct mem_cgroup *from, struct mem_cgroup *to)
4486 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4489 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4490 struct mem_cgroup **memcgp)
4492 struct mem_cgroup *memcg = NULL;
4493 unsigned int nr_pages = 1;
4494 struct page_cgroup *pc;
4495 enum charge_type ctype;
4499 if (mem_cgroup_disabled())
4502 if (PageTransHuge(page))
4503 nr_pages <<= compound_order(page);
4505 pc = lookup_page_cgroup(page);
4506 lock_page_cgroup(pc);
4507 if (PageCgroupUsed(pc)) {
4508 memcg = pc->mem_cgroup;
4509 css_get(&memcg->css);
4511 * At migrating an anonymous page, its mapcount goes down
4512 * to 0 and uncharge() will be called. But, even if it's fully
4513 * unmapped, migration may fail and this page has to be
4514 * charged again. We set MIGRATION flag here and delay uncharge
4515 * until end_migration() is called
4517 * Corner Case Thinking
4519 * When the old page was mapped as Anon and it's unmap-and-freed
4520 * while migration was ongoing.
4521 * If unmap finds the old page, uncharge() of it will be delayed
4522 * until end_migration(). If unmap finds a new page, it's
4523 * uncharged when it make mapcount to be 1->0. If unmap code
4524 * finds swap_migration_entry, the new page will not be mapped
4525 * and end_migration() will find it(mapcount==0).
4528 * When the old page was mapped but migraion fails, the kernel
4529 * remaps it. A charge for it is kept by MIGRATION flag even
4530 * if mapcount goes down to 0. We can do remap successfully
4531 * without charging it again.
4534 * The "old" page is under lock_page() until the end of
4535 * migration, so, the old page itself will not be swapped-out.
4536 * If the new page is swapped out before end_migraton, our
4537 * hook to usual swap-out path will catch the event.
4540 SetPageCgroupMigration(pc);
4542 unlock_page_cgroup(pc);
4544 * If the page is not charged at this point,
4552 * We charge new page before it's used/mapped. So, even if unlock_page()
4553 * is called before end_migration, we can catch all events on this new
4554 * page. In the case new page is migrated but not remapped, new page's
4555 * mapcount will be finally 0 and we call uncharge in end_migration().
4558 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4560 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4562 * The page is committed to the memcg, but it's not actually
4563 * charged to the res_counter since we plan on replacing the
4564 * old one and only one page is going to be left afterwards.
4566 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4569 /* remove redundant charge if migration failed*/
4570 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4571 struct page *oldpage, struct page *newpage, bool migration_ok)
4573 struct page *used, *unused;
4574 struct page_cgroup *pc;
4580 if (!migration_ok) {
4587 anon = PageAnon(used);
4588 __mem_cgroup_uncharge_common(unused,
4589 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4590 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4592 css_put(&memcg->css);
4594 * We disallowed uncharge of pages under migration because mapcount
4595 * of the page goes down to zero, temporarly.
4596 * Clear the flag and check the page should be charged.
4598 pc = lookup_page_cgroup(oldpage);
4599 lock_page_cgroup(pc);
4600 ClearPageCgroupMigration(pc);
4601 unlock_page_cgroup(pc);
4604 * If a page is a file cache, radix-tree replacement is very atomic
4605 * and we can skip this check. When it was an Anon page, its mapcount
4606 * goes down to 0. But because we added MIGRATION flage, it's not
4607 * uncharged yet. There are several case but page->mapcount check
4608 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4609 * check. (see prepare_charge() also)
4612 mem_cgroup_uncharge_page(used);
4616 * At replace page cache, newpage is not under any memcg but it's on
4617 * LRU. So, this function doesn't touch res_counter but handles LRU
4618 * in correct way. Both pages are locked so we cannot race with uncharge.
4620 void mem_cgroup_replace_page_cache(struct page *oldpage,
4621 struct page *newpage)
4623 struct mem_cgroup *memcg = NULL;
4624 struct page_cgroup *pc;
4625 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4627 if (mem_cgroup_disabled())
4630 pc = lookup_page_cgroup(oldpage);
4631 /* fix accounting on old pages */
4632 lock_page_cgroup(pc);
4633 if (PageCgroupUsed(pc)) {
4634 memcg = pc->mem_cgroup;
4635 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4636 ClearPageCgroupUsed(pc);
4638 unlock_page_cgroup(pc);
4641 * When called from shmem_replace_page(), in some cases the
4642 * oldpage has already been charged, and in some cases not.
4647 * Even if newpage->mapping was NULL before starting replacement,
4648 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4649 * LRU while we overwrite pc->mem_cgroup.
4651 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4654 #ifdef CONFIG_DEBUG_VM
4655 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4657 struct page_cgroup *pc;
4659 pc = lookup_page_cgroup(page);
4661 * Can be NULL while feeding pages into the page allocator for
4662 * the first time, i.e. during boot or memory hotplug;
4663 * or when mem_cgroup_disabled().
4665 if (likely(pc) && PageCgroupUsed(pc))
4670 bool mem_cgroup_bad_page_check(struct page *page)
4672 if (mem_cgroup_disabled())
4675 return lookup_page_cgroup_used(page) != NULL;
4678 void mem_cgroup_print_bad_page(struct page *page)
4680 struct page_cgroup *pc;
4682 pc = lookup_page_cgroup_used(page);
4684 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4685 pc, pc->flags, pc->mem_cgroup);
4690 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4691 unsigned long long val)
4694 u64 memswlimit, memlimit;
4696 int children = mem_cgroup_count_children(memcg);
4697 u64 curusage, oldusage;
4701 * For keeping hierarchical_reclaim simple, how long we should retry
4702 * is depends on callers. We set our retry-count to be function
4703 * of # of children which we should visit in this loop.
4705 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4707 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4710 while (retry_count) {
4711 if (signal_pending(current)) {
4716 * Rather than hide all in some function, I do this in
4717 * open coded manner. You see what this really does.
4718 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4720 mutex_lock(&set_limit_mutex);
4721 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4722 if (memswlimit < val) {
4724 mutex_unlock(&set_limit_mutex);
4728 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4732 ret = res_counter_set_limit(&memcg->res, val);
4734 if (memswlimit == val)
4735 memcg->memsw_is_minimum = true;
4737 memcg->memsw_is_minimum = false;
4739 mutex_unlock(&set_limit_mutex);
4744 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4745 MEM_CGROUP_RECLAIM_SHRINK);
4746 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4747 /* Usage is reduced ? */
4748 if (curusage >= oldusage)
4751 oldusage = curusage;
4753 if (!ret && enlarge)
4754 memcg_oom_recover(memcg);
4759 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4760 unsigned long long val)
4763 u64 memlimit, memswlimit, oldusage, curusage;
4764 int children = mem_cgroup_count_children(memcg);
4768 /* see mem_cgroup_resize_res_limit */
4769 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4770 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4771 while (retry_count) {
4772 if (signal_pending(current)) {
4777 * Rather than hide all in some function, I do this in
4778 * open coded manner. You see what this really does.
4779 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4781 mutex_lock(&set_limit_mutex);
4782 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4783 if (memlimit > val) {
4785 mutex_unlock(&set_limit_mutex);
4788 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4789 if (memswlimit < val)
4791 ret = res_counter_set_limit(&memcg->memsw, val);
4793 if (memlimit == val)
4794 memcg->memsw_is_minimum = true;
4796 memcg->memsw_is_minimum = false;
4798 mutex_unlock(&set_limit_mutex);
4803 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4804 MEM_CGROUP_RECLAIM_NOSWAP |
4805 MEM_CGROUP_RECLAIM_SHRINK);
4806 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4807 /* Usage is reduced ? */
4808 if (curusage >= oldusage)
4811 oldusage = curusage;
4813 if (!ret && enlarge)
4814 memcg_oom_recover(memcg);
4818 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4820 unsigned long *total_scanned)
4822 unsigned long nr_reclaimed = 0;
4823 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4824 unsigned long reclaimed;
4826 struct mem_cgroup_tree_per_zone *mctz;
4827 unsigned long long excess;
4828 unsigned long nr_scanned;
4833 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4835 * This loop can run a while, specially if mem_cgroup's continuously
4836 * keep exceeding their soft limit and putting the system under
4843 mz = mem_cgroup_largest_soft_limit_node(mctz);
4848 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4849 gfp_mask, &nr_scanned);
4850 nr_reclaimed += reclaimed;
4851 *total_scanned += nr_scanned;
4852 spin_lock(&mctz->lock);
4855 * If we failed to reclaim anything from this memory cgroup
4856 * it is time to move on to the next cgroup
4862 * Loop until we find yet another one.
4864 * By the time we get the soft_limit lock
4865 * again, someone might have aded the
4866 * group back on the RB tree. Iterate to
4867 * make sure we get a different mem.
4868 * mem_cgroup_largest_soft_limit_node returns
4869 * NULL if no other cgroup is present on
4873 __mem_cgroup_largest_soft_limit_node(mctz);
4875 css_put(&next_mz->memcg->css);
4876 else /* next_mz == NULL or other memcg */
4880 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4881 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4883 * One school of thought says that we should not add
4884 * back the node to the tree if reclaim returns 0.
4885 * But our reclaim could return 0, simply because due
4886 * to priority we are exposing a smaller subset of
4887 * memory to reclaim from. Consider this as a longer
4890 /* If excess == 0, no tree ops */
4891 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4892 spin_unlock(&mctz->lock);
4893 css_put(&mz->memcg->css);
4896 * Could not reclaim anything and there are no more
4897 * mem cgroups to try or we seem to be looping without
4898 * reclaiming anything.
4900 if (!nr_reclaimed &&
4902 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4904 } while (!nr_reclaimed);
4906 css_put(&next_mz->memcg->css);
4907 return nr_reclaimed;
4911 * mem_cgroup_force_empty_list - clears LRU of a group
4912 * @memcg: group to clear
4915 * @lru: lru to to clear
4917 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4918 * reclaim the pages page themselves - pages are moved to the parent (or root)
4921 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4922 int node, int zid, enum lru_list lru)
4924 struct lruvec *lruvec;
4925 unsigned long flags;
4926 struct list_head *list;
4930 zone = &NODE_DATA(node)->node_zones[zid];
4931 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4932 list = &lruvec->lists[lru];
4936 struct page_cgroup *pc;
4939 spin_lock_irqsave(&zone->lru_lock, flags);
4940 if (list_empty(list)) {
4941 spin_unlock_irqrestore(&zone->lru_lock, flags);
4944 page = list_entry(list->prev, struct page, lru);
4946 list_move(&page->lru, list);
4948 spin_unlock_irqrestore(&zone->lru_lock, flags);
4951 spin_unlock_irqrestore(&zone->lru_lock, flags);
4953 pc = lookup_page_cgroup(page);
4955 if (mem_cgroup_move_parent(page, pc, memcg)) {
4956 /* found lock contention or "pc" is obsolete. */
4961 } while (!list_empty(list));
4965 * make mem_cgroup's charge to be 0 if there is no task by moving
4966 * all the charges and pages to the parent.
4967 * This enables deleting this mem_cgroup.
4969 * Caller is responsible for holding css reference on the memcg.
4971 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4977 /* This is for making all *used* pages to be on LRU. */
4978 lru_add_drain_all();
4979 drain_all_stock_sync(memcg);
4980 mem_cgroup_start_move(memcg);
4981 for_each_node_state(node, N_MEMORY) {
4982 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4985 mem_cgroup_force_empty_list(memcg,
4990 mem_cgroup_end_move(memcg);
4991 memcg_oom_recover(memcg);
4995 * Kernel memory may not necessarily be trackable to a specific
4996 * process. So they are not migrated, and therefore we can't
4997 * expect their value to drop to 0 here.
4998 * Having res filled up with kmem only is enough.
5000 * This is a safety check because mem_cgroup_force_empty_list
5001 * could have raced with mem_cgroup_replace_page_cache callers
5002 * so the lru seemed empty but the page could have been added
5003 * right after the check. RES_USAGE should be safe as we always
5004 * charge before adding to the LRU.
5006 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
5007 res_counter_read_u64(&memcg->kmem, RES_USAGE);
5008 } while (usage > 0);
5011 static inline bool memcg_has_children(struct mem_cgroup *memcg)
5013 lockdep_assert_held(&memcg_create_mutex);
5015 * The lock does not prevent addition or deletion to the list
5016 * of children, but it prevents a new child from being
5017 * initialized based on this parent in css_online(), so it's
5018 * enough to decide whether hierarchically inherited
5019 * attributes can still be changed or not.
5021 return memcg->use_hierarchy &&
5022 !list_empty(&memcg->css.cgroup->children);
5026 * Reclaims as many pages from the given memcg as possible and moves
5027 * the rest to the parent.
5029 * Caller is responsible for holding css reference for memcg.
5031 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
5033 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
5034 struct cgroup *cgrp = memcg->css.cgroup;
5036 /* returns EBUSY if there is a task or if we come here twice. */
5037 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
5040 /* we call try-to-free pages for make this cgroup empty */
5041 lru_add_drain_all();
5042 /* try to free all pages in this cgroup */
5043 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
5046 if (signal_pending(current))
5049 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5053 /* maybe some writeback is necessary */
5054 congestion_wait(BLK_RW_ASYNC, HZ/10);
5059 mem_cgroup_reparent_charges(memcg);
5064 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
5067 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5069 if (mem_cgroup_is_root(memcg))
5071 return mem_cgroup_force_empty(memcg);
5074 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5077 return mem_cgroup_from_css(css)->use_hierarchy;
5080 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5081 struct cftype *cft, u64 val)
5084 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5085 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5087 mutex_lock(&memcg_create_mutex);
5089 if (memcg->use_hierarchy == val)
5093 * If parent's use_hierarchy is set, we can't make any modifications
5094 * in the child subtrees. If it is unset, then the change can
5095 * occur, provided the current cgroup has no children.
5097 * For the root cgroup, parent_mem is NULL, we allow value to be
5098 * set if there are no children.
5100 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5101 (val == 1 || val == 0)) {
5102 if (list_empty(&memcg->css.cgroup->children))
5103 memcg->use_hierarchy = val;
5110 mutex_unlock(&memcg_create_mutex);
5116 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5117 enum mem_cgroup_stat_index idx)
5119 struct mem_cgroup *iter;
5122 /* Per-cpu values can be negative, use a signed accumulator */
5123 for_each_mem_cgroup_tree(iter, memcg)
5124 val += mem_cgroup_read_stat(iter, idx);
5126 if (val < 0) /* race ? */
5131 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5135 if (!mem_cgroup_is_root(memcg)) {
5137 return res_counter_read_u64(&memcg->res, RES_USAGE);
5139 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5143 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5144 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5146 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5147 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5150 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5152 return val << PAGE_SHIFT;
5155 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
5158 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5163 type = MEMFILE_TYPE(cft->private);
5164 name = MEMFILE_ATTR(cft->private);
5168 if (name == RES_USAGE)
5169 val = mem_cgroup_usage(memcg, false);
5171 val = res_counter_read_u64(&memcg->res, name);
5174 if (name == RES_USAGE)
5175 val = mem_cgroup_usage(memcg, true);
5177 val = res_counter_read_u64(&memcg->memsw, name);
5180 val = res_counter_read_u64(&memcg->kmem, name);
5189 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
5192 #ifdef CONFIG_MEMCG_KMEM
5193 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5195 * For simplicity, we won't allow this to be disabled. It also can't
5196 * be changed if the cgroup has children already, or if tasks had
5199 * If tasks join before we set the limit, a person looking at
5200 * kmem.usage_in_bytes will have no way to determine when it took
5201 * place, which makes the value quite meaningless.
5203 * After it first became limited, changes in the value of the limit are
5204 * of course permitted.
5206 mutex_lock(&memcg_create_mutex);
5207 mutex_lock(&set_limit_mutex);
5208 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
5209 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
5213 ret = res_counter_set_limit(&memcg->kmem, val);
5216 ret = memcg_update_cache_sizes(memcg);
5218 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
5221 static_key_slow_inc(&memcg_kmem_enabled_key);
5223 * setting the active bit after the inc will guarantee no one
5224 * starts accounting before all call sites are patched
5226 memcg_kmem_set_active(memcg);
5228 ret = res_counter_set_limit(&memcg->kmem, val);
5230 mutex_unlock(&set_limit_mutex);
5231 mutex_unlock(&memcg_create_mutex);
5236 #ifdef CONFIG_MEMCG_KMEM
5237 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5240 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5244 memcg->kmem_account_flags = parent->kmem_account_flags;
5246 * When that happen, we need to disable the static branch only on those
5247 * memcgs that enabled it. To achieve this, we would be forced to
5248 * complicate the code by keeping track of which memcgs were the ones
5249 * that actually enabled limits, and which ones got it from its
5252 * It is a lot simpler just to do static_key_slow_inc() on every child
5253 * that is accounted.
5255 if (!memcg_kmem_is_active(memcg))
5259 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5260 * memcg is active already. If the later initialization fails then the
5261 * cgroup core triggers the cleanup so we do not have to do it here.
5263 static_key_slow_inc(&memcg_kmem_enabled_key);
5265 mutex_lock(&set_limit_mutex);
5266 memcg_stop_kmem_account();
5267 ret = memcg_update_cache_sizes(memcg);
5268 memcg_resume_kmem_account();
5269 mutex_unlock(&set_limit_mutex);
5273 #endif /* CONFIG_MEMCG_KMEM */
5276 * The user of this function is...
5279 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5282 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5285 unsigned long long val;
5288 type = MEMFILE_TYPE(cft->private);
5289 name = MEMFILE_ATTR(cft->private);
5293 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5297 /* This function does all necessary parse...reuse it */
5298 ret = res_counter_memparse_write_strategy(buffer, &val);
5302 ret = mem_cgroup_resize_limit(memcg, val);
5303 else if (type == _MEMSWAP)
5304 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5305 else if (type == _KMEM)
5306 ret = memcg_update_kmem_limit(css, val);
5310 case RES_SOFT_LIMIT:
5311 ret = res_counter_memparse_write_strategy(buffer, &val);
5315 * For memsw, soft limits are hard to implement in terms
5316 * of semantics, for now, we support soft limits for
5317 * control without swap
5320 ret = res_counter_set_soft_limit(&memcg->res, val);
5325 ret = -EINVAL; /* should be BUG() ? */
5331 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5332 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5334 unsigned long long min_limit, min_memsw_limit, tmp;
5336 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5337 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5338 if (!memcg->use_hierarchy)
5341 while (css_parent(&memcg->css)) {
5342 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5343 if (!memcg->use_hierarchy)
5345 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5346 min_limit = min(min_limit, tmp);
5347 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5348 min_memsw_limit = min(min_memsw_limit, tmp);
5351 *mem_limit = min_limit;
5352 *memsw_limit = min_memsw_limit;
5355 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5357 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5361 type = MEMFILE_TYPE(event);
5362 name = MEMFILE_ATTR(event);
5367 res_counter_reset_max(&memcg->res);
5368 else if (type == _MEMSWAP)
5369 res_counter_reset_max(&memcg->memsw);
5370 else if (type == _KMEM)
5371 res_counter_reset_max(&memcg->kmem);
5377 res_counter_reset_failcnt(&memcg->res);
5378 else if (type == _MEMSWAP)
5379 res_counter_reset_failcnt(&memcg->memsw);
5380 else if (type == _KMEM)
5381 res_counter_reset_failcnt(&memcg->kmem);
5390 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5393 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5397 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5398 struct cftype *cft, u64 val)
5400 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5402 if (val >= (1 << NR_MOVE_TYPE))
5406 * No kind of locking is needed in here, because ->can_attach() will
5407 * check this value once in the beginning of the process, and then carry
5408 * on with stale data. This means that changes to this value will only
5409 * affect task migrations starting after the change.
5411 memcg->move_charge_at_immigrate = val;
5415 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5416 struct cftype *cft, u64 val)
5423 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5427 unsigned int lru_mask;
5430 static const struct numa_stat stats[] = {
5431 { "total", LRU_ALL },
5432 { "file", LRU_ALL_FILE },
5433 { "anon", LRU_ALL_ANON },
5434 { "unevictable", BIT(LRU_UNEVICTABLE) },
5436 const struct numa_stat *stat;
5439 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5441 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5442 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5443 seq_printf(m, "%s=%lu", stat->name, nr);
5444 for_each_node_state(nid, N_MEMORY) {
5445 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5447 seq_printf(m, " N%d=%lu", nid, nr);
5452 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5453 struct mem_cgroup *iter;
5456 for_each_mem_cgroup_tree(iter, memcg)
5457 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5458 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5459 for_each_node_state(nid, N_MEMORY) {
5461 for_each_mem_cgroup_tree(iter, memcg)
5462 nr += mem_cgroup_node_nr_lru_pages(
5463 iter, nid, stat->lru_mask);
5464 seq_printf(m, " N%d=%lu", nid, nr);
5471 #endif /* CONFIG_NUMA */
5473 static inline void mem_cgroup_lru_names_not_uptodate(void)
5475 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5478 static int memcg_stat_show(struct seq_file *m, void *v)
5480 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5481 struct mem_cgroup *mi;
5484 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5485 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5487 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5488 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5491 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5492 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5493 mem_cgroup_read_events(memcg, i));
5495 for (i = 0; i < NR_LRU_LISTS; i++)
5496 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5497 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5499 /* Hierarchical information */
5501 unsigned long long limit, memsw_limit;
5502 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5503 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5504 if (do_swap_account)
5505 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5509 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5512 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5514 for_each_mem_cgroup_tree(mi, memcg)
5515 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5516 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5519 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5520 unsigned long long val = 0;
5522 for_each_mem_cgroup_tree(mi, memcg)
5523 val += mem_cgroup_read_events(mi, i);
5524 seq_printf(m, "total_%s %llu\n",
5525 mem_cgroup_events_names[i], val);
5528 for (i = 0; i < NR_LRU_LISTS; i++) {
5529 unsigned long long val = 0;
5531 for_each_mem_cgroup_tree(mi, memcg)
5532 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5533 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5536 #ifdef CONFIG_DEBUG_VM
5539 struct mem_cgroup_per_zone *mz;
5540 struct zone_reclaim_stat *rstat;
5541 unsigned long recent_rotated[2] = {0, 0};
5542 unsigned long recent_scanned[2] = {0, 0};
5544 for_each_online_node(nid)
5545 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5546 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5547 rstat = &mz->lruvec.reclaim_stat;
5549 recent_rotated[0] += rstat->recent_rotated[0];
5550 recent_rotated[1] += rstat->recent_rotated[1];
5551 recent_scanned[0] += rstat->recent_scanned[0];
5552 recent_scanned[1] += rstat->recent_scanned[1];
5554 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5555 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5556 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5557 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5564 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5567 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5569 return mem_cgroup_swappiness(memcg);
5572 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5573 struct cftype *cft, u64 val)
5575 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5576 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5578 if (val > 100 || !parent)
5581 mutex_lock(&memcg_create_mutex);
5583 /* If under hierarchy, only empty-root can set this value */
5584 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5585 mutex_unlock(&memcg_create_mutex);
5589 memcg->swappiness = val;
5591 mutex_unlock(&memcg_create_mutex);
5596 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5598 struct mem_cgroup_threshold_ary *t;
5604 t = rcu_dereference(memcg->thresholds.primary);
5606 t = rcu_dereference(memcg->memsw_thresholds.primary);
5611 usage = mem_cgroup_usage(memcg, swap);
5614 * current_threshold points to threshold just below or equal to usage.
5615 * If it's not true, a threshold was crossed after last
5616 * call of __mem_cgroup_threshold().
5618 i = t->current_threshold;
5621 * Iterate backward over array of thresholds starting from
5622 * current_threshold and check if a threshold is crossed.
5623 * If none of thresholds below usage is crossed, we read
5624 * only one element of the array here.
5626 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5627 eventfd_signal(t->entries[i].eventfd, 1);
5629 /* i = current_threshold + 1 */
5633 * Iterate forward over array of thresholds starting from
5634 * current_threshold+1 and check if a threshold is crossed.
5635 * If none of thresholds above usage is crossed, we read
5636 * only one element of the array here.
5638 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5639 eventfd_signal(t->entries[i].eventfd, 1);
5641 /* Update current_threshold */
5642 t->current_threshold = i - 1;
5647 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5650 __mem_cgroup_threshold(memcg, false);
5651 if (do_swap_account)
5652 __mem_cgroup_threshold(memcg, true);
5654 memcg = parent_mem_cgroup(memcg);
5658 static int compare_thresholds(const void *a, const void *b)
5660 const struct mem_cgroup_threshold *_a = a;
5661 const struct mem_cgroup_threshold *_b = b;
5663 if (_a->threshold > _b->threshold)
5666 if (_a->threshold < _b->threshold)
5672 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5674 struct mem_cgroup_eventfd_list *ev;
5676 list_for_each_entry(ev, &memcg->oom_notify, list)
5677 eventfd_signal(ev->eventfd, 1);
5681 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5683 struct mem_cgroup *iter;
5685 for_each_mem_cgroup_tree(iter, memcg)
5686 mem_cgroup_oom_notify_cb(iter);
5689 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5690 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5692 struct mem_cgroup_thresholds *thresholds;
5693 struct mem_cgroup_threshold_ary *new;
5694 u64 threshold, usage;
5697 ret = res_counter_memparse_write_strategy(args, &threshold);
5701 mutex_lock(&memcg->thresholds_lock);
5704 thresholds = &memcg->thresholds;
5705 else if (type == _MEMSWAP)
5706 thresholds = &memcg->memsw_thresholds;
5710 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5712 /* Check if a threshold crossed before adding a new one */
5713 if (thresholds->primary)
5714 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5716 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5718 /* Allocate memory for new array of thresholds */
5719 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5727 /* Copy thresholds (if any) to new array */
5728 if (thresholds->primary) {
5729 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5730 sizeof(struct mem_cgroup_threshold));
5733 /* Add new threshold */
5734 new->entries[size - 1].eventfd = eventfd;
5735 new->entries[size - 1].threshold = threshold;
5737 /* Sort thresholds. Registering of new threshold isn't time-critical */
5738 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5739 compare_thresholds, NULL);
5741 /* Find current threshold */
5742 new->current_threshold = -1;
5743 for (i = 0; i < size; i++) {
5744 if (new->entries[i].threshold <= usage) {
5746 * new->current_threshold will not be used until
5747 * rcu_assign_pointer(), so it's safe to increment
5750 ++new->current_threshold;
5755 /* Free old spare buffer and save old primary buffer as spare */
5756 kfree(thresholds->spare);
5757 thresholds->spare = thresholds->primary;
5759 rcu_assign_pointer(thresholds->primary, new);
5761 /* To be sure that nobody uses thresholds */
5765 mutex_unlock(&memcg->thresholds_lock);
5770 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5771 struct eventfd_ctx *eventfd, const char *args)
5773 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5776 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5777 struct eventfd_ctx *eventfd, const char *args)
5779 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5782 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5783 struct eventfd_ctx *eventfd, enum res_type type)
5785 struct mem_cgroup_thresholds *thresholds;
5786 struct mem_cgroup_threshold_ary *new;
5790 mutex_lock(&memcg->thresholds_lock);
5792 thresholds = &memcg->thresholds;
5793 else if (type == _MEMSWAP)
5794 thresholds = &memcg->memsw_thresholds;
5798 if (!thresholds->primary)
5801 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5803 /* Check if a threshold crossed before removing */
5804 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5806 /* Calculate new number of threshold */
5808 for (i = 0; i < thresholds->primary->size; i++) {
5809 if (thresholds->primary->entries[i].eventfd != eventfd)
5813 new = thresholds->spare;
5815 /* Set thresholds array to NULL if we don't have thresholds */
5824 /* Copy thresholds and find current threshold */
5825 new->current_threshold = -1;
5826 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5827 if (thresholds->primary->entries[i].eventfd == eventfd)
5830 new->entries[j] = thresholds->primary->entries[i];
5831 if (new->entries[j].threshold <= usage) {
5833 * new->current_threshold will not be used
5834 * until rcu_assign_pointer(), so it's safe to increment
5837 ++new->current_threshold;
5843 /* Swap primary and spare array */
5844 thresholds->spare = thresholds->primary;
5845 /* If all events are unregistered, free the spare array */
5847 kfree(thresholds->spare);
5848 thresholds->spare = NULL;
5851 rcu_assign_pointer(thresholds->primary, new);
5853 /* To be sure that nobody uses thresholds */
5856 mutex_unlock(&memcg->thresholds_lock);
5859 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5860 struct eventfd_ctx *eventfd)
5862 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5865 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5866 struct eventfd_ctx *eventfd)
5868 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5871 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5872 struct eventfd_ctx *eventfd, const char *args)
5874 struct mem_cgroup_eventfd_list *event;
5876 event = kmalloc(sizeof(*event), GFP_KERNEL);
5880 spin_lock(&memcg_oom_lock);
5882 event->eventfd = eventfd;
5883 list_add(&event->list, &memcg->oom_notify);
5885 /* already in OOM ? */
5886 if (atomic_read(&memcg->under_oom))
5887 eventfd_signal(eventfd, 1);
5888 spin_unlock(&memcg_oom_lock);
5893 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5894 struct eventfd_ctx *eventfd)
5896 struct mem_cgroup_eventfd_list *ev, *tmp;
5898 spin_lock(&memcg_oom_lock);
5900 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5901 if (ev->eventfd == eventfd) {
5902 list_del(&ev->list);
5907 spin_unlock(&memcg_oom_lock);
5910 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5912 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5914 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5915 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5919 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5920 struct cftype *cft, u64 val)
5922 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5923 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5925 /* cannot set to root cgroup and only 0 and 1 are allowed */
5926 if (!parent || !((val == 0) || (val == 1)))
5929 mutex_lock(&memcg_create_mutex);
5930 /* oom-kill-disable is a flag for subhierarchy. */
5931 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5932 mutex_unlock(&memcg_create_mutex);
5935 memcg->oom_kill_disable = val;
5937 memcg_oom_recover(memcg);
5938 mutex_unlock(&memcg_create_mutex);
5942 #ifdef CONFIG_MEMCG_KMEM
5943 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5947 memcg->kmemcg_id = -1;
5948 ret = memcg_propagate_kmem(memcg);
5952 return mem_cgroup_sockets_init(memcg, ss);
5955 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5957 mem_cgroup_sockets_destroy(memcg);
5960 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5962 if (!memcg_kmem_is_active(memcg))
5966 * kmem charges can outlive the cgroup. In the case of slab
5967 * pages, for instance, a page contain objects from various
5968 * processes. As we prevent from taking a reference for every
5969 * such allocation we have to be careful when doing uncharge
5970 * (see memcg_uncharge_kmem) and here during offlining.
5972 * The idea is that that only the _last_ uncharge which sees
5973 * the dead memcg will drop the last reference. An additional
5974 * reference is taken here before the group is marked dead
5975 * which is then paired with css_put during uncharge resp. here.
5977 * Although this might sound strange as this path is called from
5978 * css_offline() when the referencemight have dropped down to 0
5979 * and shouldn't be incremented anymore (css_tryget would fail)
5980 * we do not have other options because of the kmem allocations
5983 css_get(&memcg->css);
5985 memcg_kmem_mark_dead(memcg);
5987 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5990 if (memcg_kmem_test_and_clear_dead(memcg))
5991 css_put(&memcg->css);
5994 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5999 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
6003 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
6009 * DO NOT USE IN NEW FILES.
6011 * "cgroup.event_control" implementation.
6013 * This is way over-engineered. It tries to support fully configurable
6014 * events for each user. Such level of flexibility is completely
6015 * unnecessary especially in the light of the planned unified hierarchy.
6017 * Please deprecate this and replace with something simpler if at all
6022 * Unregister event and free resources.
6024 * Gets called from workqueue.
6026 static void memcg_event_remove(struct work_struct *work)
6028 struct mem_cgroup_event *event =
6029 container_of(work, struct mem_cgroup_event, remove);
6030 struct mem_cgroup *memcg = event->memcg;
6032 remove_wait_queue(event->wqh, &event->wait);
6034 event->unregister_event(memcg, event->eventfd);
6036 /* Notify userspace the event is going away. */
6037 eventfd_signal(event->eventfd, 1);
6039 eventfd_ctx_put(event->eventfd);
6041 css_put(&memcg->css);
6045 * Gets called on POLLHUP on eventfd when user closes it.
6047 * Called with wqh->lock held and interrupts disabled.
6049 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
6050 int sync, void *key)
6052 struct mem_cgroup_event *event =
6053 container_of(wait, struct mem_cgroup_event, wait);
6054 struct mem_cgroup *memcg = event->memcg;
6055 unsigned long flags = (unsigned long)key;
6057 if (flags & POLLHUP) {
6059 * If the event has been detached at cgroup removal, we
6060 * can simply return knowing the other side will cleanup
6063 * We can't race against event freeing since the other
6064 * side will require wqh->lock via remove_wait_queue(),
6067 spin_lock(&memcg->event_list_lock);
6068 if (!list_empty(&event->list)) {
6069 list_del_init(&event->list);
6071 * We are in atomic context, but cgroup_event_remove()
6072 * may sleep, so we have to call it in workqueue.
6074 schedule_work(&event->remove);
6076 spin_unlock(&memcg->event_list_lock);
6082 static void memcg_event_ptable_queue_proc(struct file *file,
6083 wait_queue_head_t *wqh, poll_table *pt)
6085 struct mem_cgroup_event *event =
6086 container_of(pt, struct mem_cgroup_event, pt);
6089 add_wait_queue(wqh, &event->wait);
6093 * DO NOT USE IN NEW FILES.
6095 * Parse input and register new cgroup event handler.
6097 * Input must be in format '<event_fd> <control_fd> <args>'.
6098 * Interpretation of args is defined by control file implementation.
6100 static int memcg_write_event_control(struct cgroup_subsys_state *css,
6101 struct cftype *cft, const char *buffer)
6103 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6104 struct mem_cgroup_event *event;
6105 struct cgroup_subsys_state *cfile_css;
6106 unsigned int efd, cfd;
6113 efd = simple_strtoul(buffer, &endp, 10);
6118 cfd = simple_strtoul(buffer, &endp, 10);
6119 if ((*endp != ' ') && (*endp != '\0'))
6123 event = kzalloc(sizeof(*event), GFP_KERNEL);
6127 event->memcg = memcg;
6128 INIT_LIST_HEAD(&event->list);
6129 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
6130 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
6131 INIT_WORK(&event->remove, memcg_event_remove);
6139 event->eventfd = eventfd_ctx_fileget(efile.file);
6140 if (IS_ERR(event->eventfd)) {
6141 ret = PTR_ERR(event->eventfd);
6148 goto out_put_eventfd;
6151 /* the process need read permission on control file */
6152 /* AV: shouldn't we check that it's been opened for read instead? */
6153 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6158 * Determine the event callbacks and set them in @event. This used
6159 * to be done via struct cftype but cgroup core no longer knows
6160 * about these events. The following is crude but the whole thing
6161 * is for compatibility anyway.
6163 * DO NOT ADD NEW FILES.
6165 name = cfile.file->f_dentry->d_name.name;
6167 if (!strcmp(name, "memory.usage_in_bytes")) {
6168 event->register_event = mem_cgroup_usage_register_event;
6169 event->unregister_event = mem_cgroup_usage_unregister_event;
6170 } else if (!strcmp(name, "memory.oom_control")) {
6171 event->register_event = mem_cgroup_oom_register_event;
6172 event->unregister_event = mem_cgroup_oom_unregister_event;
6173 } else if (!strcmp(name, "memory.pressure_level")) {
6174 event->register_event = vmpressure_register_event;
6175 event->unregister_event = vmpressure_unregister_event;
6176 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
6177 event->register_event = memsw_cgroup_usage_register_event;
6178 event->unregister_event = memsw_cgroup_usage_unregister_event;
6185 * Verify @cfile should belong to @css. Also, remaining events are
6186 * automatically removed on cgroup destruction but the removal is
6187 * asynchronous, so take an extra ref on @css.
6192 cfile_css = css_from_dir(cfile.file->f_dentry->d_parent,
6193 &mem_cgroup_subsys);
6194 if (cfile_css == css && css_tryget(css))
6201 ret = event->register_event(memcg, event->eventfd, buffer);
6205 efile.file->f_op->poll(efile.file, &event->pt);
6207 spin_lock(&memcg->event_list_lock);
6208 list_add(&event->list, &memcg->event_list);
6209 spin_unlock(&memcg->event_list_lock);
6221 eventfd_ctx_put(event->eventfd);
6230 static struct cftype mem_cgroup_files[] = {
6232 .name = "usage_in_bytes",
6233 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6234 .read_u64 = mem_cgroup_read_u64,
6237 .name = "max_usage_in_bytes",
6238 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6239 .trigger = mem_cgroup_reset,
6240 .read_u64 = mem_cgroup_read_u64,
6243 .name = "limit_in_bytes",
6244 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6245 .write_string = mem_cgroup_write,
6246 .read_u64 = mem_cgroup_read_u64,
6249 .name = "soft_limit_in_bytes",
6250 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6251 .write_string = mem_cgroup_write,
6252 .read_u64 = mem_cgroup_read_u64,
6256 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6257 .trigger = mem_cgroup_reset,
6258 .read_u64 = mem_cgroup_read_u64,
6262 .seq_show = memcg_stat_show,
6265 .name = "force_empty",
6266 .trigger = mem_cgroup_force_empty_write,
6269 .name = "use_hierarchy",
6270 .flags = CFTYPE_INSANE,
6271 .write_u64 = mem_cgroup_hierarchy_write,
6272 .read_u64 = mem_cgroup_hierarchy_read,
6275 .name = "cgroup.event_control", /* XXX: for compat */
6276 .write_string = memcg_write_event_control,
6277 .flags = CFTYPE_NO_PREFIX,
6281 .name = "swappiness",
6282 .read_u64 = mem_cgroup_swappiness_read,
6283 .write_u64 = mem_cgroup_swappiness_write,
6286 .name = "move_charge_at_immigrate",
6287 .read_u64 = mem_cgroup_move_charge_read,
6288 .write_u64 = mem_cgroup_move_charge_write,
6291 .name = "oom_control",
6292 .seq_show = mem_cgroup_oom_control_read,
6293 .write_u64 = mem_cgroup_oom_control_write,
6294 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6297 .name = "pressure_level",
6301 .name = "numa_stat",
6302 .seq_show = memcg_numa_stat_show,
6305 #ifdef CONFIG_MEMCG_KMEM
6307 .name = "kmem.limit_in_bytes",
6308 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6309 .write_string = mem_cgroup_write,
6310 .read_u64 = mem_cgroup_read_u64,
6313 .name = "kmem.usage_in_bytes",
6314 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6315 .read_u64 = mem_cgroup_read_u64,
6318 .name = "kmem.failcnt",
6319 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6320 .trigger = mem_cgroup_reset,
6321 .read_u64 = mem_cgroup_read_u64,
6324 .name = "kmem.max_usage_in_bytes",
6325 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6326 .trigger = mem_cgroup_reset,
6327 .read_u64 = mem_cgroup_read_u64,
6329 #ifdef CONFIG_SLABINFO
6331 .name = "kmem.slabinfo",
6332 .seq_show = mem_cgroup_slabinfo_read,
6336 { }, /* terminate */
6339 #ifdef CONFIG_MEMCG_SWAP
6340 static struct cftype memsw_cgroup_files[] = {
6342 .name = "memsw.usage_in_bytes",
6343 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6344 .read_u64 = mem_cgroup_read_u64,
6347 .name = "memsw.max_usage_in_bytes",
6348 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6349 .trigger = mem_cgroup_reset,
6350 .read_u64 = mem_cgroup_read_u64,
6353 .name = "memsw.limit_in_bytes",
6354 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6355 .write_string = mem_cgroup_write,
6356 .read_u64 = mem_cgroup_read_u64,
6359 .name = "memsw.failcnt",
6360 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6361 .trigger = mem_cgroup_reset,
6362 .read_u64 = mem_cgroup_read_u64,
6364 { }, /* terminate */
6367 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6369 struct mem_cgroup_per_node *pn;
6370 struct mem_cgroup_per_zone *mz;
6371 int zone, tmp = node;
6373 * This routine is called against possible nodes.
6374 * But it's BUG to call kmalloc() against offline node.
6376 * TODO: this routine can waste much memory for nodes which will
6377 * never be onlined. It's better to use memory hotplug callback
6380 if (!node_state(node, N_NORMAL_MEMORY))
6382 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6386 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6387 mz = &pn->zoneinfo[zone];
6388 lruvec_init(&mz->lruvec);
6389 mz->usage_in_excess = 0;
6390 mz->on_tree = false;
6393 memcg->nodeinfo[node] = pn;
6397 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6399 kfree(memcg->nodeinfo[node]);
6402 static struct mem_cgroup *mem_cgroup_alloc(void)
6404 struct mem_cgroup *memcg;
6405 size_t size = memcg_size();
6407 /* Can be very big if nr_node_ids is very big */
6408 if (size < PAGE_SIZE)
6409 memcg = kzalloc(size, GFP_KERNEL);
6411 memcg = vzalloc(size);
6416 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6419 spin_lock_init(&memcg->pcp_counter_lock);
6423 if (size < PAGE_SIZE)
6431 * At destroying mem_cgroup, references from swap_cgroup can remain.
6432 * (scanning all at force_empty is too costly...)
6434 * Instead of clearing all references at force_empty, we remember
6435 * the number of reference from swap_cgroup and free mem_cgroup when
6436 * it goes down to 0.
6438 * Removal of cgroup itself succeeds regardless of refs from swap.
6441 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6444 size_t size = memcg_size();
6446 mem_cgroup_remove_from_trees(memcg);
6449 free_mem_cgroup_per_zone_info(memcg, node);
6451 free_percpu(memcg->stat);
6454 * We need to make sure that (at least for now), the jump label
6455 * destruction code runs outside of the cgroup lock. This is because
6456 * get_online_cpus(), which is called from the static_branch update,
6457 * can't be called inside the cgroup_lock. cpusets are the ones
6458 * enforcing this dependency, so if they ever change, we might as well.
6460 * schedule_work() will guarantee this happens. Be careful if you need
6461 * to move this code around, and make sure it is outside
6464 disarm_static_keys(memcg);
6465 if (size < PAGE_SIZE)
6472 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6474 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6476 if (!memcg->res.parent)
6478 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6480 EXPORT_SYMBOL(parent_mem_cgroup);
6482 static void __init mem_cgroup_soft_limit_tree_init(void)
6484 struct mem_cgroup_tree_per_node *rtpn;
6485 struct mem_cgroup_tree_per_zone *rtpz;
6486 int tmp, node, zone;
6488 for_each_node(node) {
6490 if (!node_state(node, N_NORMAL_MEMORY))
6492 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6495 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6497 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6498 rtpz = &rtpn->rb_tree_per_zone[zone];
6499 rtpz->rb_root = RB_ROOT;
6500 spin_lock_init(&rtpz->lock);
6505 static struct cgroup_subsys_state * __ref
6506 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6508 struct mem_cgroup *memcg;
6509 long error = -ENOMEM;
6512 memcg = mem_cgroup_alloc();
6514 return ERR_PTR(error);
6517 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6521 if (parent_css == NULL) {
6522 root_mem_cgroup = memcg;
6523 res_counter_init(&memcg->res, NULL);
6524 res_counter_init(&memcg->memsw, NULL);
6525 res_counter_init(&memcg->kmem, NULL);
6528 memcg->last_scanned_node = MAX_NUMNODES;
6529 INIT_LIST_HEAD(&memcg->oom_notify);
6530 memcg->move_charge_at_immigrate = 0;
6531 mutex_init(&memcg->thresholds_lock);
6532 spin_lock_init(&memcg->move_lock);
6533 vmpressure_init(&memcg->vmpressure);
6534 INIT_LIST_HEAD(&memcg->event_list);
6535 spin_lock_init(&memcg->event_list_lock);
6540 __mem_cgroup_free(memcg);
6541 return ERR_PTR(error);
6545 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6547 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6548 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6551 if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6557 mutex_lock(&memcg_create_mutex);
6559 memcg->use_hierarchy = parent->use_hierarchy;
6560 memcg->oom_kill_disable = parent->oom_kill_disable;
6561 memcg->swappiness = mem_cgroup_swappiness(parent);
6563 if (parent->use_hierarchy) {
6564 res_counter_init(&memcg->res, &parent->res);
6565 res_counter_init(&memcg->memsw, &parent->memsw);
6566 res_counter_init(&memcg->kmem, &parent->kmem);
6569 * No need to take a reference to the parent because cgroup
6570 * core guarantees its existence.
6573 res_counter_init(&memcg->res, NULL);
6574 res_counter_init(&memcg->memsw, NULL);
6575 res_counter_init(&memcg->kmem, NULL);
6577 * Deeper hierachy with use_hierarchy == false doesn't make
6578 * much sense so let cgroup subsystem know about this
6579 * unfortunate state in our controller.
6581 if (parent != root_mem_cgroup)
6582 mem_cgroup_subsys.broken_hierarchy = true;
6585 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6586 mutex_unlock(&memcg_create_mutex);
6591 * Announce all parents that a group from their hierarchy is gone.
6593 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6595 struct mem_cgroup *parent = memcg;
6597 while ((parent = parent_mem_cgroup(parent)))
6598 mem_cgroup_iter_invalidate(parent);
6601 * if the root memcg is not hierarchical we have to check it
6604 if (!root_mem_cgroup->use_hierarchy)
6605 mem_cgroup_iter_invalidate(root_mem_cgroup);
6608 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6610 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6611 struct mem_cgroup_event *event, *tmp;
6614 * Unregister events and notify userspace.
6615 * Notify userspace about cgroup removing only after rmdir of cgroup
6616 * directory to avoid race between userspace and kernelspace.
6618 spin_lock(&memcg->event_list_lock);
6619 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6620 list_del_init(&event->list);
6621 schedule_work(&event->remove);
6623 spin_unlock(&memcg->event_list_lock);
6625 kmem_cgroup_css_offline(memcg);
6627 mem_cgroup_invalidate_reclaim_iterators(memcg);
6628 mem_cgroup_reparent_charges(memcg);
6629 mem_cgroup_destroy_all_caches(memcg);
6630 vmpressure_cleanup(&memcg->vmpressure);
6633 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6635 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6637 * XXX: css_offline() would be where we should reparent all
6638 * memory to prepare the cgroup for destruction. However,
6639 * memcg does not do css_tryget() and res_counter charging
6640 * under the same RCU lock region, which means that charging
6641 * could race with offlining. Offlining only happens to
6642 * cgroups with no tasks in them but charges can show up
6643 * without any tasks from the swapin path when the target
6644 * memcg is looked up from the swapout record and not from the
6645 * current task as it usually is. A race like this can leak
6646 * charges and put pages with stale cgroup pointers into
6650 * lookup_swap_cgroup_id()
6652 * mem_cgroup_lookup()
6655 * disable css_tryget()
6658 * reparent_charges()
6659 * res_counter_charge()
6662 * pc->mem_cgroup = dead memcg
6665 * The bulk of the charges are still moved in offline_css() to
6666 * avoid pinning a lot of pages in case a long-term reference
6667 * like a swapout record is deferring the css_free() to long
6668 * after offlining. But this makes sure we catch any charges
6669 * made after offlining:
6671 mem_cgroup_reparent_charges(memcg);
6673 memcg_destroy_kmem(memcg);
6674 __mem_cgroup_free(memcg);
6678 /* Handlers for move charge at task migration. */
6679 #define PRECHARGE_COUNT_AT_ONCE 256
6680 static int mem_cgroup_do_precharge(unsigned long count)
6683 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6684 struct mem_cgroup *memcg = mc.to;
6686 if (mem_cgroup_is_root(memcg)) {
6687 mc.precharge += count;
6688 /* we don't need css_get for root */
6691 /* try to charge at once */
6693 struct res_counter *dummy;
6695 * "memcg" cannot be under rmdir() because we've already checked
6696 * by cgroup_lock_live_cgroup() that it is not removed and we
6697 * are still under the same cgroup_mutex. So we can postpone
6700 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6702 if (do_swap_account && res_counter_charge(&memcg->memsw,
6703 PAGE_SIZE * count, &dummy)) {
6704 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6707 mc.precharge += count;
6711 /* fall back to one by one charge */
6713 if (signal_pending(current)) {
6717 if (!batch_count--) {
6718 batch_count = PRECHARGE_COUNT_AT_ONCE;
6721 ret = __mem_cgroup_try_charge(NULL,
6722 GFP_KERNEL, 1, &memcg, false);
6724 /* mem_cgroup_clear_mc() will do uncharge later */
6732 * get_mctgt_type - get target type of moving charge
6733 * @vma: the vma the pte to be checked belongs
6734 * @addr: the address corresponding to the pte to be checked
6735 * @ptent: the pte to be checked
6736 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6739 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6740 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6741 * move charge. if @target is not NULL, the page is stored in target->page
6742 * with extra refcnt got(Callers should handle it).
6743 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6744 * target for charge migration. if @target is not NULL, the entry is stored
6747 * Called with pte lock held.
6754 enum mc_target_type {
6760 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6761 unsigned long addr, pte_t ptent)
6763 struct page *page = vm_normal_page(vma, addr, ptent);
6765 if (!page || !page_mapped(page))
6767 if (PageAnon(page)) {
6768 /* we don't move shared anon */
6771 } else if (!move_file())
6772 /* we ignore mapcount for file pages */
6774 if (!get_page_unless_zero(page))
6781 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6782 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6784 struct page *page = NULL;
6785 swp_entry_t ent = pte_to_swp_entry(ptent);
6787 if (!move_anon() || non_swap_entry(ent))
6790 * Because lookup_swap_cache() updates some statistics counter,
6791 * we call find_get_page() with swapper_space directly.
6793 page = find_get_page(swap_address_space(ent), ent.val);
6794 if (do_swap_account)
6795 entry->val = ent.val;
6800 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6801 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6807 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6808 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6810 struct page *page = NULL;
6811 struct address_space *mapping;
6814 if (!vma->vm_file) /* anonymous vma */
6819 mapping = vma->vm_file->f_mapping;
6820 if (pte_none(ptent))
6821 pgoff = linear_page_index(vma, addr);
6822 else /* pte_file(ptent) is true */
6823 pgoff = pte_to_pgoff(ptent);
6825 /* page is moved even if it's not RSS of this task(page-faulted). */
6826 page = find_get_page(mapping, pgoff);
6829 /* shmem/tmpfs may report page out on swap: account for that too. */
6830 if (radix_tree_exceptional_entry(page)) {
6831 swp_entry_t swap = radix_to_swp_entry(page);
6832 if (do_swap_account)
6834 page = find_get_page(swap_address_space(swap), swap.val);
6840 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6841 unsigned long addr, pte_t ptent, union mc_target *target)
6843 struct page *page = NULL;
6844 struct page_cgroup *pc;
6845 enum mc_target_type ret = MC_TARGET_NONE;
6846 swp_entry_t ent = { .val = 0 };
6848 if (pte_present(ptent))
6849 page = mc_handle_present_pte(vma, addr, ptent);
6850 else if (is_swap_pte(ptent))
6851 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6852 else if (pte_none(ptent) || pte_file(ptent))
6853 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6855 if (!page && !ent.val)
6858 pc = lookup_page_cgroup(page);
6860 * Do only loose check w/o page_cgroup lock.
6861 * mem_cgroup_move_account() checks the pc is valid or not under
6864 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6865 ret = MC_TARGET_PAGE;
6867 target->page = page;
6869 if (!ret || !target)
6872 /* There is a swap entry and a page doesn't exist or isn't charged */
6873 if (ent.val && !ret &&
6874 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6875 ret = MC_TARGET_SWAP;
6882 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6884 * We don't consider swapping or file mapped pages because THP does not
6885 * support them for now.
6886 * Caller should make sure that pmd_trans_huge(pmd) is true.
6888 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6889 unsigned long addr, pmd_t pmd, union mc_target *target)
6891 struct page *page = NULL;
6892 struct page_cgroup *pc;
6893 enum mc_target_type ret = MC_TARGET_NONE;
6895 page = pmd_page(pmd);
6896 VM_BUG_ON(!page || !PageHead(page));
6899 pc = lookup_page_cgroup(page);
6900 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6901 ret = MC_TARGET_PAGE;
6904 target->page = page;
6910 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6911 unsigned long addr, pmd_t pmd, union mc_target *target)
6913 return MC_TARGET_NONE;
6917 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6918 unsigned long addr, unsigned long end,
6919 struct mm_walk *walk)
6921 struct vm_area_struct *vma = walk->private;
6925 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6926 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6927 mc.precharge += HPAGE_PMD_NR;
6932 if (pmd_trans_unstable(pmd))
6934 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6935 for (; addr != end; pte++, addr += PAGE_SIZE)
6936 if (get_mctgt_type(vma, addr, *pte, NULL))
6937 mc.precharge++; /* increment precharge temporarily */
6938 pte_unmap_unlock(pte - 1, ptl);
6944 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6946 unsigned long precharge;
6947 struct vm_area_struct *vma;
6949 down_read(&mm->mmap_sem);
6950 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6951 struct mm_walk mem_cgroup_count_precharge_walk = {
6952 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6956 if (is_vm_hugetlb_page(vma))
6958 walk_page_range(vma->vm_start, vma->vm_end,
6959 &mem_cgroup_count_precharge_walk);
6961 up_read(&mm->mmap_sem);
6963 precharge = mc.precharge;
6969 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6971 unsigned long precharge = mem_cgroup_count_precharge(mm);
6973 VM_BUG_ON(mc.moving_task);
6974 mc.moving_task = current;
6975 return mem_cgroup_do_precharge(precharge);
6978 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6979 static void __mem_cgroup_clear_mc(void)
6981 struct mem_cgroup *from = mc.from;
6982 struct mem_cgroup *to = mc.to;
6985 /* we must uncharge all the leftover precharges from mc.to */
6987 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6991 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6992 * we must uncharge here.
6994 if (mc.moved_charge) {
6995 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6996 mc.moved_charge = 0;
6998 /* we must fixup refcnts and charges */
6999 if (mc.moved_swap) {
7000 /* uncharge swap account from the old cgroup */
7001 if (!mem_cgroup_is_root(mc.from))
7002 res_counter_uncharge(&mc.from->memsw,
7003 PAGE_SIZE * mc.moved_swap);
7005 for (i = 0; i < mc.moved_swap; i++)
7006 css_put(&mc.from->css);
7008 if (!mem_cgroup_is_root(mc.to)) {
7010 * we charged both to->res and to->memsw, so we should
7013 res_counter_uncharge(&mc.to->res,
7014 PAGE_SIZE * mc.moved_swap);
7016 /* we've already done css_get(mc.to) */
7019 memcg_oom_recover(from);
7020 memcg_oom_recover(to);
7021 wake_up_all(&mc.waitq);
7024 static void mem_cgroup_clear_mc(void)
7026 struct mem_cgroup *from = mc.from;
7029 * we must clear moving_task before waking up waiters at the end of
7032 mc.moving_task = NULL;
7033 __mem_cgroup_clear_mc();
7034 spin_lock(&mc.lock);
7037 spin_unlock(&mc.lock);
7038 mem_cgroup_end_move(from);
7041 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7042 struct cgroup_taskset *tset)
7044 struct task_struct *p = cgroup_taskset_first(tset);
7046 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7047 unsigned long move_charge_at_immigrate;
7050 * We are now commited to this value whatever it is. Changes in this
7051 * tunable will only affect upcoming migrations, not the current one.
7052 * So we need to save it, and keep it going.
7054 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
7055 if (move_charge_at_immigrate) {
7056 struct mm_struct *mm;
7057 struct mem_cgroup *from = mem_cgroup_from_task(p);
7059 VM_BUG_ON(from == memcg);
7061 mm = get_task_mm(p);
7064 /* We move charges only when we move a owner of the mm */
7065 if (mm->owner == p) {
7068 VM_BUG_ON(mc.precharge);
7069 VM_BUG_ON(mc.moved_charge);
7070 VM_BUG_ON(mc.moved_swap);
7071 mem_cgroup_start_move(from);
7072 spin_lock(&mc.lock);
7075 mc.immigrate_flags = move_charge_at_immigrate;
7076 spin_unlock(&mc.lock);
7077 /* We set mc.moving_task later */
7079 ret = mem_cgroup_precharge_mc(mm);
7081 mem_cgroup_clear_mc();
7088 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7089 struct cgroup_taskset *tset)
7091 mem_cgroup_clear_mc();
7094 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
7095 unsigned long addr, unsigned long end,
7096 struct mm_walk *walk)
7099 struct vm_area_struct *vma = walk->private;
7102 enum mc_target_type target_type;
7103 union mc_target target;
7105 struct page_cgroup *pc;
7108 * We don't take compound_lock() here but no race with splitting thp
7110 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7111 * under splitting, which means there's no concurrent thp split,
7112 * - if another thread runs into split_huge_page() just after we
7113 * entered this if-block, the thread must wait for page table lock
7114 * to be unlocked in __split_huge_page_splitting(), where the main
7115 * part of thp split is not executed yet.
7117 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
7118 if (mc.precharge < HPAGE_PMD_NR) {
7122 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
7123 if (target_type == MC_TARGET_PAGE) {
7125 if (!isolate_lru_page(page)) {
7126 pc = lookup_page_cgroup(page);
7127 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
7128 pc, mc.from, mc.to)) {
7129 mc.precharge -= HPAGE_PMD_NR;
7130 mc.moved_charge += HPAGE_PMD_NR;
7132 putback_lru_page(page);
7140 if (pmd_trans_unstable(pmd))
7143 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7144 for (; addr != end; addr += PAGE_SIZE) {
7145 pte_t ptent = *(pte++);
7151 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7152 case MC_TARGET_PAGE:
7154 if (isolate_lru_page(page))
7156 pc = lookup_page_cgroup(page);
7157 if (!mem_cgroup_move_account(page, 1, pc,
7160 /* we uncharge from mc.from later. */
7163 putback_lru_page(page);
7164 put: /* get_mctgt_type() gets the page */
7167 case MC_TARGET_SWAP:
7169 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7171 /* we fixup refcnts and charges later. */
7179 pte_unmap_unlock(pte - 1, ptl);
7184 * We have consumed all precharges we got in can_attach().
7185 * We try charge one by one, but don't do any additional
7186 * charges to mc.to if we have failed in charge once in attach()
7189 ret = mem_cgroup_do_precharge(1);
7197 static void mem_cgroup_move_charge(struct mm_struct *mm)
7199 struct vm_area_struct *vma;
7201 lru_add_drain_all();
7203 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7205 * Someone who are holding the mmap_sem might be waiting in
7206 * waitq. So we cancel all extra charges, wake up all waiters,
7207 * and retry. Because we cancel precharges, we might not be able
7208 * to move enough charges, but moving charge is a best-effort
7209 * feature anyway, so it wouldn't be a big problem.
7211 __mem_cgroup_clear_mc();
7215 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7217 struct mm_walk mem_cgroup_move_charge_walk = {
7218 .pmd_entry = mem_cgroup_move_charge_pte_range,
7222 if (is_vm_hugetlb_page(vma))
7224 ret = walk_page_range(vma->vm_start, vma->vm_end,
7225 &mem_cgroup_move_charge_walk);
7228 * means we have consumed all precharges and failed in
7229 * doing additional charge. Just abandon here.
7233 up_read(&mm->mmap_sem);
7236 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7237 struct cgroup_taskset *tset)
7239 struct task_struct *p = cgroup_taskset_first(tset);
7240 struct mm_struct *mm = get_task_mm(p);
7244 mem_cgroup_move_charge(mm);
7248 mem_cgroup_clear_mc();
7250 #else /* !CONFIG_MMU */
7251 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7252 struct cgroup_taskset *tset)
7256 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7257 struct cgroup_taskset *tset)
7260 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7261 struct cgroup_taskset *tset)
7267 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7268 * to verify sane_behavior flag on each mount attempt.
7270 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7273 * use_hierarchy is forced with sane_behavior. cgroup core
7274 * guarantees that @root doesn't have any children, so turning it
7275 * on for the root memcg is enough.
7277 if (cgroup_sane_behavior(root_css->cgroup))
7278 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7281 struct cgroup_subsys mem_cgroup_subsys = {
7283 .subsys_id = mem_cgroup_subsys_id,
7284 .css_alloc = mem_cgroup_css_alloc,
7285 .css_online = mem_cgroup_css_online,
7286 .css_offline = mem_cgroup_css_offline,
7287 .css_free = mem_cgroup_css_free,
7288 .can_attach = mem_cgroup_can_attach,
7289 .cancel_attach = mem_cgroup_cancel_attach,
7290 .attach = mem_cgroup_move_task,
7291 .bind = mem_cgroup_bind,
7292 .base_cftypes = mem_cgroup_files,
7296 #ifdef CONFIG_MEMCG_SWAP
7297 static int __init enable_swap_account(char *s)
7299 if (!strcmp(s, "1"))
7300 really_do_swap_account = 1;
7301 else if (!strcmp(s, "0"))
7302 really_do_swap_account = 0;
7305 __setup("swapaccount=", enable_swap_account);
7307 static void __init memsw_file_init(void)
7309 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7312 static void __init enable_swap_cgroup(void)
7314 if (!mem_cgroup_disabled() && really_do_swap_account) {
7315 do_swap_account = 1;
7321 static void __init enable_swap_cgroup(void)
7327 * subsys_initcall() for memory controller.
7329 * Some parts like hotcpu_notifier() have to be initialized from this context
7330 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7331 * everything that doesn't depend on a specific mem_cgroup structure should
7332 * be initialized from here.
7334 static int __init mem_cgroup_init(void)
7336 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7337 enable_swap_cgroup();
7338 mem_cgroup_soft_limit_tree_init();
7342 subsys_initcall(mem_cgroup_init);