1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata = 0;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct mem_cgroup_reclaim_iter {
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup *last_visited;
154 /* scan generation, increased every round-trip */
155 unsigned int generation;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone {
162 struct lruvec lruvec;
163 unsigned long lru_size[NR_LRU_LISTS];
165 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
167 struct rb_node tree_node; /* RB tree node */
168 unsigned long long usage_in_excess;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup *memcg; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node {
176 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone {
185 struct rb_root rb_root;
189 struct mem_cgroup_tree_per_node {
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
193 struct mem_cgroup_tree {
194 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
199 struct mem_cgroup_threshold {
200 struct eventfd_ctx *eventfd;
205 struct mem_cgroup_threshold_ary {
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries[0];
214 struct mem_cgroup_thresholds {
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary *primary;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary *spare;
226 struct mem_cgroup_eventfd_list {
227 struct list_head list;
228 struct eventfd_ctx *eventfd;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event {
236 * memcg which the event belongs to.
238 struct mem_cgroup *memcg;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx *eventfd;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event)(struct mem_cgroup *memcg,
253 struct eventfd_ctx *eventfd, const char *args);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event)(struct mem_cgroup *memcg,
260 struct eventfd_ctx *eventfd);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t *wqh;
268 struct work_struct remove;
271 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
272 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css;
288 * the counter to account for memory usage
290 struct res_counter res;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure;
296 * the counter to account for mem+swap usage.
298 struct res_counter memsw;
301 * the counter to account for kernel memory usage.
303 struct res_counter kmem;
305 * Should the accounting and control be hierarchical, per subtree?
308 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
312 atomic_t oom_wakeups;
315 /* OOM-Killer disable */
316 int oom_kill_disable;
318 /* set when res.limit == memsw.limit */
319 bool memsw_is_minimum;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds;
330 /* For oom notifier event fd */
331 struct list_head oom_notify;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock;
347 struct mem_cgroup_stat_cpu __percpu *stat;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base;
353 spinlock_t pcp_counter_lock;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list. per-memcg */
361 struct list_head memcg_slab_caches;
362 /* Not a spinlock, we can take a lot of time walking the list */
363 struct mutex slab_caches_mutex;
364 /* Index in the kmem_cache->memcg_params->memcg_caches array */
368 int last_scanned_node;
370 nodemask_t scan_nodes;
371 atomic_t numainfo_events;
372 atomic_t numainfo_updating;
375 /* List of events which userspace want to receive */
376 struct list_head event_list;
377 spinlock_t event_list_lock;
379 struct mem_cgroup_per_node *nodeinfo[0];
380 /* WARNING: nodeinfo must be the last member here */
383 /* internal only representation about the status of kmem accounting. */
385 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
386 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
389 #ifdef CONFIG_MEMCG_KMEM
390 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
392 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
395 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
397 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
400 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
403 * Our caller must use css_get() first, because memcg_uncharge_kmem()
404 * will call css_put() if it sees the memcg is dead.
407 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
408 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
411 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
413 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
414 &memcg->kmem_account_flags);
418 /* Stuffs for move charges at task migration. */
420 * Types of charges to be moved. "move_charge_at_immitgrate" and
421 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
424 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
425 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
429 /* "mc" and its members are protected by cgroup_mutex */
430 static struct move_charge_struct {
431 spinlock_t lock; /* for from, to */
432 struct mem_cgroup *from;
433 struct mem_cgroup *to;
434 unsigned long immigrate_flags;
435 unsigned long precharge;
436 unsigned long moved_charge;
437 unsigned long moved_swap;
438 struct task_struct *moving_task; /* a task moving charges */
439 wait_queue_head_t waitq; /* a waitq for other context */
441 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
442 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
445 static bool move_anon(void)
447 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
450 static bool move_file(void)
452 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
456 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
457 * limit reclaim to prevent infinite loops, if they ever occur.
459 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
460 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
463 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
464 MEM_CGROUP_CHARGE_TYPE_ANON,
465 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
466 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
470 /* for encoding cft->private value on file */
478 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
479 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
480 #define MEMFILE_ATTR(val) ((val) & 0xffff)
481 /* Used for OOM nofiier */
482 #define OOM_CONTROL (0)
485 * Reclaim flags for mem_cgroup_hierarchical_reclaim
487 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
488 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
489 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
490 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
493 * The memcg_create_mutex will be held whenever a new cgroup is created.
494 * As a consequence, any change that needs to protect against new child cgroups
495 * appearing has to hold it as well.
497 static DEFINE_MUTEX(memcg_create_mutex);
499 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
501 return s ? container_of(s, struct mem_cgroup, css) : NULL;
504 /* Some nice accessors for the vmpressure. */
505 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
508 memcg = root_mem_cgroup;
509 return &memcg->vmpressure;
512 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
514 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
517 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
519 return (memcg == root_mem_cgroup);
523 * We restrict the id in the range of [1, 65535], so it can fit into
526 #define MEM_CGROUP_ID_MAX USHRT_MAX
528 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
530 return memcg->css.id;
533 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
535 struct cgroup_subsys_state *css;
537 css = css_from_id(id, &memory_cgrp_subsys);
538 return mem_cgroup_from_css(css);
541 /* Writing them here to avoid exposing memcg's inner layout */
542 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
544 void sock_update_memcg(struct sock *sk)
546 if (mem_cgroup_sockets_enabled) {
547 struct mem_cgroup *memcg;
548 struct cg_proto *cg_proto;
550 BUG_ON(!sk->sk_prot->proto_cgroup);
552 /* Socket cloning can throw us here with sk_cgrp already
553 * filled. It won't however, necessarily happen from
554 * process context. So the test for root memcg given
555 * the current task's memcg won't help us in this case.
557 * Respecting the original socket's memcg is a better
558 * decision in this case.
561 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
562 css_get(&sk->sk_cgrp->memcg->css);
567 memcg = mem_cgroup_from_task(current);
568 cg_proto = sk->sk_prot->proto_cgroup(memcg);
569 if (!mem_cgroup_is_root(memcg) &&
570 memcg_proto_active(cg_proto) &&
571 css_tryget_online(&memcg->css)) {
572 sk->sk_cgrp = cg_proto;
577 EXPORT_SYMBOL(sock_update_memcg);
579 void sock_release_memcg(struct sock *sk)
581 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
582 struct mem_cgroup *memcg;
583 WARN_ON(!sk->sk_cgrp->memcg);
584 memcg = sk->sk_cgrp->memcg;
585 css_put(&sk->sk_cgrp->memcg->css);
589 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
591 if (!memcg || mem_cgroup_is_root(memcg))
594 return &memcg->tcp_mem;
596 EXPORT_SYMBOL(tcp_proto_cgroup);
598 static void disarm_sock_keys(struct mem_cgroup *memcg)
600 if (!memcg_proto_activated(&memcg->tcp_mem))
602 static_key_slow_dec(&memcg_socket_limit_enabled);
605 static void disarm_sock_keys(struct mem_cgroup *memcg)
610 #ifdef CONFIG_MEMCG_KMEM
612 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
613 * The main reason for not using cgroup id for this:
614 * this works better in sparse environments, where we have a lot of memcgs,
615 * but only a few kmem-limited. Or also, if we have, for instance, 200
616 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
617 * 200 entry array for that.
619 * The current size of the caches array is stored in
620 * memcg_limited_groups_array_size. It will double each time we have to
623 static DEFINE_IDA(kmem_limited_groups);
624 int memcg_limited_groups_array_size;
627 * MIN_SIZE is different than 1, because we would like to avoid going through
628 * the alloc/free process all the time. In a small machine, 4 kmem-limited
629 * cgroups is a reasonable guess. In the future, it could be a parameter or
630 * tunable, but that is strictly not necessary.
632 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
633 * this constant directly from cgroup, but it is understandable that this is
634 * better kept as an internal representation in cgroup.c. In any case, the
635 * cgrp_id space is not getting any smaller, and we don't have to necessarily
636 * increase ours as well if it increases.
638 #define MEMCG_CACHES_MIN_SIZE 4
639 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
642 * A lot of the calls to the cache allocation functions are expected to be
643 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
644 * conditional to this static branch, we'll have to allow modules that does
645 * kmem_cache_alloc and the such to see this symbol as well
647 struct static_key memcg_kmem_enabled_key;
648 EXPORT_SYMBOL(memcg_kmem_enabled_key);
650 static void disarm_kmem_keys(struct mem_cgroup *memcg)
652 if (memcg_kmem_is_active(memcg)) {
653 static_key_slow_dec(&memcg_kmem_enabled_key);
654 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
657 * This check can't live in kmem destruction function,
658 * since the charges will outlive the cgroup
660 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
663 static void disarm_kmem_keys(struct mem_cgroup *memcg)
666 #endif /* CONFIG_MEMCG_KMEM */
668 static void disarm_static_keys(struct mem_cgroup *memcg)
670 disarm_sock_keys(memcg);
671 disarm_kmem_keys(memcg);
674 static void drain_all_stock_async(struct mem_cgroup *memcg);
676 static struct mem_cgroup_per_zone *
677 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
679 VM_BUG_ON((unsigned)nid >= nr_node_ids);
680 return &memcg->nodeinfo[nid]->zoneinfo[zid];
683 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
688 static struct mem_cgroup_per_zone *
689 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
691 int nid = page_to_nid(page);
692 int zid = page_zonenum(page);
694 return mem_cgroup_zoneinfo(memcg, nid, zid);
697 static struct mem_cgroup_tree_per_zone *
698 soft_limit_tree_node_zone(int nid, int zid)
700 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
703 static struct mem_cgroup_tree_per_zone *
704 soft_limit_tree_from_page(struct page *page)
706 int nid = page_to_nid(page);
707 int zid = page_zonenum(page);
709 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
713 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
714 struct mem_cgroup_per_zone *mz,
715 struct mem_cgroup_tree_per_zone *mctz,
716 unsigned long long new_usage_in_excess)
718 struct rb_node **p = &mctz->rb_root.rb_node;
719 struct rb_node *parent = NULL;
720 struct mem_cgroup_per_zone *mz_node;
725 mz->usage_in_excess = new_usage_in_excess;
726 if (!mz->usage_in_excess)
730 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
732 if (mz->usage_in_excess < mz_node->usage_in_excess)
735 * We can't avoid mem cgroups that are over their soft
736 * limit by the same amount
738 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
741 rb_link_node(&mz->tree_node, parent, p);
742 rb_insert_color(&mz->tree_node, &mctz->rb_root);
747 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
748 struct mem_cgroup_per_zone *mz,
749 struct mem_cgroup_tree_per_zone *mctz)
753 rb_erase(&mz->tree_node, &mctz->rb_root);
758 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
759 struct mem_cgroup_per_zone *mz,
760 struct mem_cgroup_tree_per_zone *mctz)
762 spin_lock(&mctz->lock);
763 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
764 spin_unlock(&mctz->lock);
768 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
770 unsigned long long excess;
771 struct mem_cgroup_per_zone *mz;
772 struct mem_cgroup_tree_per_zone *mctz;
773 int nid = page_to_nid(page);
774 int zid = page_zonenum(page);
775 mctz = soft_limit_tree_from_page(page);
778 * Necessary to update all ancestors when hierarchy is used.
779 * because their event counter is not touched.
781 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
782 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
783 excess = res_counter_soft_limit_excess(&memcg->res);
785 * We have to update the tree if mz is on RB-tree or
786 * mem is over its softlimit.
788 if (excess || mz->on_tree) {
789 spin_lock(&mctz->lock);
790 /* if on-tree, remove it */
792 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
794 * Insert again. mz->usage_in_excess will be updated.
795 * If excess is 0, no tree ops.
797 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
798 spin_unlock(&mctz->lock);
803 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
806 struct mem_cgroup_per_zone *mz;
807 struct mem_cgroup_tree_per_zone *mctz;
809 for_each_node(node) {
810 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
811 mz = mem_cgroup_zoneinfo(memcg, node, zone);
812 mctz = soft_limit_tree_node_zone(node, zone);
813 mem_cgroup_remove_exceeded(memcg, mz, mctz);
818 static struct mem_cgroup_per_zone *
819 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
821 struct rb_node *rightmost = NULL;
822 struct mem_cgroup_per_zone *mz;
826 rightmost = rb_last(&mctz->rb_root);
828 goto done; /* Nothing to reclaim from */
830 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
832 * Remove the node now but someone else can add it back,
833 * we will to add it back at the end of reclaim to its correct
834 * position in the tree.
836 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
837 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
838 !css_tryget_online(&mz->memcg->css))
844 static struct mem_cgroup_per_zone *
845 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
847 struct mem_cgroup_per_zone *mz;
849 spin_lock(&mctz->lock);
850 mz = __mem_cgroup_largest_soft_limit_node(mctz);
851 spin_unlock(&mctz->lock);
856 * Implementation Note: reading percpu statistics for memcg.
858 * Both of vmstat[] and percpu_counter has threshold and do periodic
859 * synchronization to implement "quick" read. There are trade-off between
860 * reading cost and precision of value. Then, we may have a chance to implement
861 * a periodic synchronizion of counter in memcg's counter.
863 * But this _read() function is used for user interface now. The user accounts
864 * memory usage by memory cgroup and he _always_ requires exact value because
865 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
866 * have to visit all online cpus and make sum. So, for now, unnecessary
867 * synchronization is not implemented. (just implemented for cpu hotplug)
869 * If there are kernel internal actions which can make use of some not-exact
870 * value, and reading all cpu value can be performance bottleneck in some
871 * common workload, threashold and synchonization as vmstat[] should be
874 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
875 enum mem_cgroup_stat_index idx)
881 for_each_online_cpu(cpu)
882 val += per_cpu(memcg->stat->count[idx], cpu);
883 #ifdef CONFIG_HOTPLUG_CPU
884 spin_lock(&memcg->pcp_counter_lock);
885 val += memcg->nocpu_base.count[idx];
886 spin_unlock(&memcg->pcp_counter_lock);
892 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
895 int val = (charge) ? 1 : -1;
896 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
899 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
900 enum mem_cgroup_events_index idx)
902 unsigned long val = 0;
906 for_each_online_cpu(cpu)
907 val += per_cpu(memcg->stat->events[idx], cpu);
908 #ifdef CONFIG_HOTPLUG_CPU
909 spin_lock(&memcg->pcp_counter_lock);
910 val += memcg->nocpu_base.events[idx];
911 spin_unlock(&memcg->pcp_counter_lock);
917 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
919 bool anon, int nr_pages)
922 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
923 * counted as CACHE even if it's on ANON LRU.
926 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
929 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
932 if (PageTransHuge(page))
933 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
936 /* pagein of a big page is an event. So, ignore page size */
938 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
940 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
941 nr_pages = -nr_pages; /* for event */
944 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
948 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
950 struct mem_cgroup_per_zone *mz;
952 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
953 return mz->lru_size[lru];
957 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
958 unsigned int lru_mask)
960 struct mem_cgroup_per_zone *mz;
962 unsigned long ret = 0;
964 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
967 if (BIT(lru) & lru_mask)
968 ret += mz->lru_size[lru];
974 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
975 int nid, unsigned int lru_mask)
980 for (zid = 0; zid < MAX_NR_ZONES; zid++)
981 total += mem_cgroup_zone_nr_lru_pages(memcg,
987 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
988 unsigned int lru_mask)
993 for_each_node_state(nid, N_MEMORY)
994 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
998 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
999 enum mem_cgroup_events_target target)
1001 unsigned long val, next;
1003 val = __this_cpu_read(memcg->stat->nr_page_events);
1004 next = __this_cpu_read(memcg->stat->targets[target]);
1005 /* from time_after() in jiffies.h */
1006 if ((long)next - (long)val < 0) {
1008 case MEM_CGROUP_TARGET_THRESH:
1009 next = val + THRESHOLDS_EVENTS_TARGET;
1011 case MEM_CGROUP_TARGET_SOFTLIMIT:
1012 next = val + SOFTLIMIT_EVENTS_TARGET;
1014 case MEM_CGROUP_TARGET_NUMAINFO:
1015 next = val + NUMAINFO_EVENTS_TARGET;
1020 __this_cpu_write(memcg->stat->targets[target], next);
1027 * Check events in order.
1030 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1033 /* threshold event is triggered in finer grain than soft limit */
1034 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1035 MEM_CGROUP_TARGET_THRESH))) {
1037 bool do_numainfo __maybe_unused;
1039 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1040 MEM_CGROUP_TARGET_SOFTLIMIT);
1041 #if MAX_NUMNODES > 1
1042 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1043 MEM_CGROUP_TARGET_NUMAINFO);
1047 mem_cgroup_threshold(memcg);
1048 if (unlikely(do_softlimit))
1049 mem_cgroup_update_tree(memcg, page);
1050 #if MAX_NUMNODES > 1
1051 if (unlikely(do_numainfo))
1052 atomic_inc(&memcg->numainfo_events);
1058 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1061 * mm_update_next_owner() may clear mm->owner to NULL
1062 * if it races with swapoff, page migration, etc.
1063 * So this can be called with p == NULL.
1068 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1071 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1073 struct mem_cgroup *memcg = NULL;
1077 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1078 if (unlikely(!memcg))
1079 memcg = root_mem_cgroup;
1080 } while (!css_tryget_online(&memcg->css));
1086 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1087 * ref. count) or NULL if the whole root's subtree has been visited.
1089 * helper function to be used by mem_cgroup_iter
1091 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1092 struct mem_cgroup *last_visited)
1094 struct cgroup_subsys_state *prev_css, *next_css;
1096 prev_css = last_visited ? &last_visited->css : NULL;
1098 next_css = css_next_descendant_pre(prev_css, &root->css);
1101 * Even if we found a group we have to make sure it is
1102 * alive. css && !memcg means that the groups should be
1103 * skipped and we should continue the tree walk.
1104 * last_visited css is safe to use because it is
1105 * protected by css_get and the tree walk is rcu safe.
1107 * We do not take a reference on the root of the tree walk
1108 * because we might race with the root removal when it would
1109 * be the only node in the iterated hierarchy and mem_cgroup_iter
1110 * would end up in an endless loop because it expects that at
1111 * least one valid node will be returned. Root cannot disappear
1112 * because caller of the iterator should hold it already so
1113 * skipping css reference should be safe.
1116 if ((next_css == &root->css) ||
1117 ((next_css->flags & CSS_ONLINE) &&
1118 css_tryget_online(next_css)))
1119 return mem_cgroup_from_css(next_css);
1121 prev_css = next_css;
1128 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1131 * When a group in the hierarchy below root is destroyed, the
1132 * hierarchy iterator can no longer be trusted since it might
1133 * have pointed to the destroyed group. Invalidate it.
1135 atomic_inc(&root->dead_count);
1138 static struct mem_cgroup *
1139 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1140 struct mem_cgroup *root,
1143 struct mem_cgroup *position = NULL;
1145 * A cgroup destruction happens in two stages: offlining and
1146 * release. They are separated by a RCU grace period.
1148 * If the iterator is valid, we may still race with an
1149 * offlining. The RCU lock ensures the object won't be
1150 * released, tryget will fail if we lost the race.
1152 *sequence = atomic_read(&root->dead_count);
1153 if (iter->last_dead_count == *sequence) {
1155 position = iter->last_visited;
1158 * We cannot take a reference to root because we might race
1159 * with root removal and returning NULL would end up in
1160 * an endless loop on the iterator user level when root
1161 * would be returned all the time.
1163 if (position && position != root &&
1164 !css_tryget_online(&position->css))
1170 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1171 struct mem_cgroup *last_visited,
1172 struct mem_cgroup *new_position,
1173 struct mem_cgroup *root,
1176 /* root reference counting symmetric to mem_cgroup_iter_load */
1177 if (last_visited && last_visited != root)
1178 css_put(&last_visited->css);
1180 * We store the sequence count from the time @last_visited was
1181 * loaded successfully instead of rereading it here so that we
1182 * don't lose destruction events in between. We could have
1183 * raced with the destruction of @new_position after all.
1185 iter->last_visited = new_position;
1187 iter->last_dead_count = sequence;
1191 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1192 * @root: hierarchy root
1193 * @prev: previously returned memcg, NULL on first invocation
1194 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1196 * Returns references to children of the hierarchy below @root, or
1197 * @root itself, or %NULL after a full round-trip.
1199 * Caller must pass the return value in @prev on subsequent
1200 * invocations for reference counting, or use mem_cgroup_iter_break()
1201 * to cancel a hierarchy walk before the round-trip is complete.
1203 * Reclaimers can specify a zone and a priority level in @reclaim to
1204 * divide up the memcgs in the hierarchy among all concurrent
1205 * reclaimers operating on the same zone and priority.
1207 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1208 struct mem_cgroup *prev,
1209 struct mem_cgroup_reclaim_cookie *reclaim)
1211 struct mem_cgroup *memcg = NULL;
1212 struct mem_cgroup *last_visited = NULL;
1214 if (mem_cgroup_disabled())
1218 root = root_mem_cgroup;
1220 if (prev && !reclaim)
1221 last_visited = prev;
1223 if (!root->use_hierarchy && root != root_mem_cgroup) {
1231 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1232 int uninitialized_var(seq);
1235 int nid = zone_to_nid(reclaim->zone);
1236 int zid = zone_idx(reclaim->zone);
1237 struct mem_cgroup_per_zone *mz;
1239 mz = mem_cgroup_zoneinfo(root, nid, zid);
1240 iter = &mz->reclaim_iter[reclaim->priority];
1241 if (prev && reclaim->generation != iter->generation) {
1242 iter->last_visited = NULL;
1246 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1249 memcg = __mem_cgroup_iter_next(root, last_visited);
1252 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1257 else if (!prev && memcg)
1258 reclaim->generation = iter->generation;
1267 if (prev && prev != root)
1268 css_put(&prev->css);
1274 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1275 * @root: hierarchy root
1276 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1278 void mem_cgroup_iter_break(struct mem_cgroup *root,
1279 struct mem_cgroup *prev)
1282 root = root_mem_cgroup;
1283 if (prev && prev != root)
1284 css_put(&prev->css);
1288 * Iteration constructs for visiting all cgroups (under a tree). If
1289 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1290 * be used for reference counting.
1292 #define for_each_mem_cgroup_tree(iter, root) \
1293 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1295 iter = mem_cgroup_iter(root, iter, NULL))
1297 #define for_each_mem_cgroup(iter) \
1298 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1300 iter = mem_cgroup_iter(NULL, iter, NULL))
1302 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1304 struct mem_cgroup *memcg;
1307 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1308 if (unlikely(!memcg))
1313 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1316 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1324 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1327 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1328 * @zone: zone of the wanted lruvec
1329 * @memcg: memcg of the wanted lruvec
1331 * Returns the lru list vector holding pages for the given @zone and
1332 * @mem. This can be the global zone lruvec, if the memory controller
1335 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1336 struct mem_cgroup *memcg)
1338 struct mem_cgroup_per_zone *mz;
1339 struct lruvec *lruvec;
1341 if (mem_cgroup_disabled()) {
1342 lruvec = &zone->lruvec;
1346 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1347 lruvec = &mz->lruvec;
1350 * Since a node can be onlined after the mem_cgroup was created,
1351 * we have to be prepared to initialize lruvec->zone here;
1352 * and if offlined then reonlined, we need to reinitialize it.
1354 if (unlikely(lruvec->zone != zone))
1355 lruvec->zone = zone;
1360 * Following LRU functions are allowed to be used without PCG_LOCK.
1361 * Operations are called by routine of global LRU independently from memcg.
1362 * What we have to take care of here is validness of pc->mem_cgroup.
1364 * Changes to pc->mem_cgroup happens when
1367 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1368 * It is added to LRU before charge.
1369 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1370 * When moving account, the page is not on LRU. It's isolated.
1374 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1376 * @zone: zone of the page
1378 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1380 struct mem_cgroup_per_zone *mz;
1381 struct mem_cgroup *memcg;
1382 struct page_cgroup *pc;
1383 struct lruvec *lruvec;
1385 if (mem_cgroup_disabled()) {
1386 lruvec = &zone->lruvec;
1390 pc = lookup_page_cgroup(page);
1391 memcg = pc->mem_cgroup;
1394 * Surreptitiously switch any uncharged offlist page to root:
1395 * an uncharged page off lru does nothing to secure
1396 * its former mem_cgroup from sudden removal.
1398 * Our caller holds lru_lock, and PageCgroupUsed is updated
1399 * under page_cgroup lock: between them, they make all uses
1400 * of pc->mem_cgroup safe.
1402 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1403 pc->mem_cgroup = memcg = root_mem_cgroup;
1405 mz = page_cgroup_zoneinfo(memcg, page);
1406 lruvec = &mz->lruvec;
1409 * Since a node can be onlined after the mem_cgroup was created,
1410 * we have to be prepared to initialize lruvec->zone here;
1411 * and if offlined then reonlined, we need to reinitialize it.
1413 if (unlikely(lruvec->zone != zone))
1414 lruvec->zone = zone;
1419 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1420 * @lruvec: mem_cgroup per zone lru vector
1421 * @lru: index of lru list the page is sitting on
1422 * @nr_pages: positive when adding or negative when removing
1424 * This function must be called when a page is added to or removed from an
1427 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1430 struct mem_cgroup_per_zone *mz;
1431 unsigned long *lru_size;
1433 if (mem_cgroup_disabled())
1436 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1437 lru_size = mz->lru_size + lru;
1438 *lru_size += nr_pages;
1439 VM_BUG_ON((long)(*lru_size) < 0);
1443 * Checks whether given mem is same or in the root_mem_cgroup's
1446 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1447 struct mem_cgroup *memcg)
1449 if (root_memcg == memcg)
1451 if (!root_memcg->use_hierarchy || !memcg)
1453 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1456 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1457 struct mem_cgroup *memcg)
1462 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1467 bool task_in_mem_cgroup(struct task_struct *task,
1468 const struct mem_cgroup *memcg)
1470 struct mem_cgroup *curr = NULL;
1471 struct task_struct *p;
1474 p = find_lock_task_mm(task);
1476 curr = get_mem_cgroup_from_mm(p->mm);
1480 * All threads may have already detached their mm's, but the oom
1481 * killer still needs to detect if they have already been oom
1482 * killed to prevent needlessly killing additional tasks.
1485 curr = mem_cgroup_from_task(task);
1487 css_get(&curr->css);
1491 * We should check use_hierarchy of "memcg" not "curr". Because checking
1492 * use_hierarchy of "curr" here make this function true if hierarchy is
1493 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1494 * hierarchy(even if use_hierarchy is disabled in "memcg").
1496 ret = mem_cgroup_same_or_subtree(memcg, curr);
1497 css_put(&curr->css);
1501 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1503 unsigned long inactive_ratio;
1504 unsigned long inactive;
1505 unsigned long active;
1508 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1509 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1511 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1513 inactive_ratio = int_sqrt(10 * gb);
1517 return inactive * inactive_ratio < active;
1520 #define mem_cgroup_from_res_counter(counter, member) \
1521 container_of(counter, struct mem_cgroup, member)
1524 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1525 * @memcg: the memory cgroup
1527 * Returns the maximum amount of memory @mem can be charged with, in
1530 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1532 unsigned long long margin;
1534 margin = res_counter_margin(&memcg->res);
1535 if (do_swap_account)
1536 margin = min(margin, res_counter_margin(&memcg->memsw));
1537 return margin >> PAGE_SHIFT;
1540 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1543 if (!css_parent(&memcg->css))
1544 return vm_swappiness;
1546 return memcg->swappiness;
1550 * memcg->moving_account is used for checking possibility that some thread is
1551 * calling move_account(). When a thread on CPU-A starts moving pages under
1552 * a memcg, other threads should check memcg->moving_account under
1553 * rcu_read_lock(), like this:
1557 * memcg->moving_account+1 if (memcg->mocing_account)
1559 * synchronize_rcu() update something.
1564 /* for quick checking without looking up memcg */
1565 atomic_t memcg_moving __read_mostly;
1567 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1569 atomic_inc(&memcg_moving);
1570 atomic_inc(&memcg->moving_account);
1574 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1577 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1578 * We check NULL in callee rather than caller.
1581 atomic_dec(&memcg_moving);
1582 atomic_dec(&memcg->moving_account);
1587 * 2 routines for checking "mem" is under move_account() or not.
1589 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1590 * is used for avoiding races in accounting. If true,
1591 * pc->mem_cgroup may be overwritten.
1593 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1594 * under hierarchy of moving cgroups. This is for
1595 * waiting at hith-memory prressure caused by "move".
1598 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1600 VM_BUG_ON(!rcu_read_lock_held());
1601 return atomic_read(&memcg->moving_account) > 0;
1604 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1606 struct mem_cgroup *from;
1607 struct mem_cgroup *to;
1610 * Unlike task_move routines, we access mc.to, mc.from not under
1611 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1613 spin_lock(&mc.lock);
1619 ret = mem_cgroup_same_or_subtree(memcg, from)
1620 || mem_cgroup_same_or_subtree(memcg, to);
1622 spin_unlock(&mc.lock);
1626 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1628 if (mc.moving_task && current != mc.moving_task) {
1629 if (mem_cgroup_under_move(memcg)) {
1631 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1632 /* moving charge context might have finished. */
1635 finish_wait(&mc.waitq, &wait);
1643 * Take this lock when
1644 * - a code tries to modify page's memcg while it's USED.
1645 * - a code tries to modify page state accounting in a memcg.
1646 * see mem_cgroup_stolen(), too.
1648 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1649 unsigned long *flags)
1651 spin_lock_irqsave(&memcg->move_lock, *flags);
1654 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1655 unsigned long *flags)
1657 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1660 #define K(x) ((x) << (PAGE_SHIFT-10))
1662 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1663 * @memcg: The memory cgroup that went over limit
1664 * @p: Task that is going to be killed
1666 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1669 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1671 /* oom_info_lock ensures that parallel ooms do not interleave */
1672 static DEFINE_MUTEX(oom_info_lock);
1673 struct mem_cgroup *iter;
1679 mutex_lock(&oom_info_lock);
1682 pr_info("Task in ");
1683 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1684 pr_info(" killed as a result of limit of ");
1685 pr_cont_cgroup_path(memcg->css.cgroup);
1690 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1691 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1692 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1693 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1694 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1695 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1696 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1697 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1698 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1699 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1700 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1701 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1703 for_each_mem_cgroup_tree(iter, memcg) {
1704 pr_info("Memory cgroup stats for ");
1705 pr_cont_cgroup_path(iter->css.cgroup);
1708 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1709 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1711 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1712 K(mem_cgroup_read_stat(iter, i)));
1715 for (i = 0; i < NR_LRU_LISTS; i++)
1716 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1717 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1721 mutex_unlock(&oom_info_lock);
1725 * This function returns the number of memcg under hierarchy tree. Returns
1726 * 1(self count) if no children.
1728 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1731 struct mem_cgroup *iter;
1733 for_each_mem_cgroup_tree(iter, memcg)
1739 * Return the memory (and swap, if configured) limit for a memcg.
1741 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1745 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1748 * Do not consider swap space if we cannot swap due to swappiness
1750 if (mem_cgroup_swappiness(memcg)) {
1753 limit += total_swap_pages << PAGE_SHIFT;
1754 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1757 * If memsw is finite and limits the amount of swap space
1758 * available to this memcg, return that limit.
1760 limit = min(limit, memsw);
1766 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1769 struct mem_cgroup *iter;
1770 unsigned long chosen_points = 0;
1771 unsigned long totalpages;
1772 unsigned int points = 0;
1773 struct task_struct *chosen = NULL;
1776 * If current has a pending SIGKILL or is exiting, then automatically
1777 * select it. The goal is to allow it to allocate so that it may
1778 * quickly exit and free its memory.
1780 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1781 set_thread_flag(TIF_MEMDIE);
1785 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1786 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1787 for_each_mem_cgroup_tree(iter, memcg) {
1788 struct css_task_iter it;
1789 struct task_struct *task;
1791 css_task_iter_start(&iter->css, &it);
1792 while ((task = css_task_iter_next(&it))) {
1793 switch (oom_scan_process_thread(task, totalpages, NULL,
1795 case OOM_SCAN_SELECT:
1797 put_task_struct(chosen);
1799 chosen_points = ULONG_MAX;
1800 get_task_struct(chosen);
1802 case OOM_SCAN_CONTINUE:
1804 case OOM_SCAN_ABORT:
1805 css_task_iter_end(&it);
1806 mem_cgroup_iter_break(memcg, iter);
1808 put_task_struct(chosen);
1813 points = oom_badness(task, memcg, NULL, totalpages);
1814 if (!points || points < chosen_points)
1816 /* Prefer thread group leaders for display purposes */
1817 if (points == chosen_points &&
1818 thread_group_leader(chosen))
1822 put_task_struct(chosen);
1824 chosen_points = points;
1825 get_task_struct(chosen);
1827 css_task_iter_end(&it);
1832 points = chosen_points * 1000 / totalpages;
1833 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1834 NULL, "Memory cgroup out of memory");
1837 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1839 unsigned long flags)
1841 unsigned long total = 0;
1842 bool noswap = false;
1845 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1847 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1850 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1852 drain_all_stock_async(memcg);
1853 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1855 * Allow limit shrinkers, which are triggered directly
1856 * by userspace, to catch signals and stop reclaim
1857 * after minimal progress, regardless of the margin.
1859 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1861 if (mem_cgroup_margin(memcg))
1864 * If nothing was reclaimed after two attempts, there
1865 * may be no reclaimable pages in this hierarchy.
1874 * test_mem_cgroup_node_reclaimable
1875 * @memcg: the target memcg
1876 * @nid: the node ID to be checked.
1877 * @noswap : specify true here if the user wants flle only information.
1879 * This function returns whether the specified memcg contains any
1880 * reclaimable pages on a node. Returns true if there are any reclaimable
1881 * pages in the node.
1883 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1884 int nid, bool noswap)
1886 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1888 if (noswap || !total_swap_pages)
1890 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1895 #if MAX_NUMNODES > 1
1898 * Always updating the nodemask is not very good - even if we have an empty
1899 * list or the wrong list here, we can start from some node and traverse all
1900 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1903 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1907 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1908 * pagein/pageout changes since the last update.
1910 if (!atomic_read(&memcg->numainfo_events))
1912 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1915 /* make a nodemask where this memcg uses memory from */
1916 memcg->scan_nodes = node_states[N_MEMORY];
1918 for_each_node_mask(nid, node_states[N_MEMORY]) {
1920 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1921 node_clear(nid, memcg->scan_nodes);
1924 atomic_set(&memcg->numainfo_events, 0);
1925 atomic_set(&memcg->numainfo_updating, 0);
1929 * Selecting a node where we start reclaim from. Because what we need is just
1930 * reducing usage counter, start from anywhere is O,K. Considering
1931 * memory reclaim from current node, there are pros. and cons.
1933 * Freeing memory from current node means freeing memory from a node which
1934 * we'll use or we've used. So, it may make LRU bad. And if several threads
1935 * hit limits, it will see a contention on a node. But freeing from remote
1936 * node means more costs for memory reclaim because of memory latency.
1938 * Now, we use round-robin. Better algorithm is welcomed.
1940 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1944 mem_cgroup_may_update_nodemask(memcg);
1945 node = memcg->last_scanned_node;
1947 node = next_node(node, memcg->scan_nodes);
1948 if (node == MAX_NUMNODES)
1949 node = first_node(memcg->scan_nodes);
1951 * We call this when we hit limit, not when pages are added to LRU.
1952 * No LRU may hold pages because all pages are UNEVICTABLE or
1953 * memcg is too small and all pages are not on LRU. In that case,
1954 * we use curret node.
1956 if (unlikely(node == MAX_NUMNODES))
1957 node = numa_node_id();
1959 memcg->last_scanned_node = node;
1964 * Check all nodes whether it contains reclaimable pages or not.
1965 * For quick scan, we make use of scan_nodes. This will allow us to skip
1966 * unused nodes. But scan_nodes is lazily updated and may not cotain
1967 * enough new information. We need to do double check.
1969 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1974 * quick check...making use of scan_node.
1975 * We can skip unused nodes.
1977 if (!nodes_empty(memcg->scan_nodes)) {
1978 for (nid = first_node(memcg->scan_nodes);
1980 nid = next_node(nid, memcg->scan_nodes)) {
1982 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1987 * Check rest of nodes.
1989 for_each_node_state(nid, N_MEMORY) {
1990 if (node_isset(nid, memcg->scan_nodes))
1992 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1999 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2004 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2006 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2010 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2013 unsigned long *total_scanned)
2015 struct mem_cgroup *victim = NULL;
2018 unsigned long excess;
2019 unsigned long nr_scanned;
2020 struct mem_cgroup_reclaim_cookie reclaim = {
2025 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2028 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2033 * If we have not been able to reclaim
2034 * anything, it might because there are
2035 * no reclaimable pages under this hierarchy
2040 * We want to do more targeted reclaim.
2041 * excess >> 2 is not to excessive so as to
2042 * reclaim too much, nor too less that we keep
2043 * coming back to reclaim from this cgroup
2045 if (total >= (excess >> 2) ||
2046 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2051 if (!mem_cgroup_reclaimable(victim, false))
2053 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2055 *total_scanned += nr_scanned;
2056 if (!res_counter_soft_limit_excess(&root_memcg->res))
2059 mem_cgroup_iter_break(root_memcg, victim);
2063 #ifdef CONFIG_LOCKDEP
2064 static struct lockdep_map memcg_oom_lock_dep_map = {
2065 .name = "memcg_oom_lock",
2069 static DEFINE_SPINLOCK(memcg_oom_lock);
2072 * Check OOM-Killer is already running under our hierarchy.
2073 * If someone is running, return false.
2075 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2077 struct mem_cgroup *iter, *failed = NULL;
2079 spin_lock(&memcg_oom_lock);
2081 for_each_mem_cgroup_tree(iter, memcg) {
2082 if (iter->oom_lock) {
2084 * this subtree of our hierarchy is already locked
2085 * so we cannot give a lock.
2088 mem_cgroup_iter_break(memcg, iter);
2091 iter->oom_lock = true;
2096 * OK, we failed to lock the whole subtree so we have
2097 * to clean up what we set up to the failing subtree
2099 for_each_mem_cgroup_tree(iter, memcg) {
2100 if (iter == failed) {
2101 mem_cgroup_iter_break(memcg, iter);
2104 iter->oom_lock = false;
2107 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2109 spin_unlock(&memcg_oom_lock);
2114 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2116 struct mem_cgroup *iter;
2118 spin_lock(&memcg_oom_lock);
2119 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2120 for_each_mem_cgroup_tree(iter, memcg)
2121 iter->oom_lock = false;
2122 spin_unlock(&memcg_oom_lock);
2125 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2127 struct mem_cgroup *iter;
2129 for_each_mem_cgroup_tree(iter, memcg)
2130 atomic_inc(&iter->under_oom);
2133 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2135 struct mem_cgroup *iter;
2138 * When a new child is created while the hierarchy is under oom,
2139 * mem_cgroup_oom_lock() may not be called. We have to use
2140 * atomic_add_unless() here.
2142 for_each_mem_cgroup_tree(iter, memcg)
2143 atomic_add_unless(&iter->under_oom, -1, 0);
2146 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2148 struct oom_wait_info {
2149 struct mem_cgroup *memcg;
2153 static int memcg_oom_wake_function(wait_queue_t *wait,
2154 unsigned mode, int sync, void *arg)
2156 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2157 struct mem_cgroup *oom_wait_memcg;
2158 struct oom_wait_info *oom_wait_info;
2160 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2161 oom_wait_memcg = oom_wait_info->memcg;
2164 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2165 * Then we can use css_is_ancestor without taking care of RCU.
2167 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2168 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2170 return autoremove_wake_function(wait, mode, sync, arg);
2173 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2175 atomic_inc(&memcg->oom_wakeups);
2176 /* for filtering, pass "memcg" as argument. */
2177 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2180 static void memcg_oom_recover(struct mem_cgroup *memcg)
2182 if (memcg && atomic_read(&memcg->under_oom))
2183 memcg_wakeup_oom(memcg);
2186 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2188 if (!current->memcg_oom.may_oom)
2191 * We are in the middle of the charge context here, so we
2192 * don't want to block when potentially sitting on a callstack
2193 * that holds all kinds of filesystem and mm locks.
2195 * Also, the caller may handle a failed allocation gracefully
2196 * (like optional page cache readahead) and so an OOM killer
2197 * invocation might not even be necessary.
2199 * That's why we don't do anything here except remember the
2200 * OOM context and then deal with it at the end of the page
2201 * fault when the stack is unwound, the locks are released,
2202 * and when we know whether the fault was overall successful.
2204 css_get(&memcg->css);
2205 current->memcg_oom.memcg = memcg;
2206 current->memcg_oom.gfp_mask = mask;
2207 current->memcg_oom.order = order;
2211 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2212 * @handle: actually kill/wait or just clean up the OOM state
2214 * This has to be called at the end of a page fault if the memcg OOM
2215 * handler was enabled.
2217 * Memcg supports userspace OOM handling where failed allocations must
2218 * sleep on a waitqueue until the userspace task resolves the
2219 * situation. Sleeping directly in the charge context with all kinds
2220 * of locks held is not a good idea, instead we remember an OOM state
2221 * in the task and mem_cgroup_oom_synchronize() has to be called at
2222 * the end of the page fault to complete the OOM handling.
2224 * Returns %true if an ongoing memcg OOM situation was detected and
2225 * completed, %false otherwise.
2227 bool mem_cgroup_oom_synchronize(bool handle)
2229 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2230 struct oom_wait_info owait;
2233 /* OOM is global, do not handle */
2240 owait.memcg = memcg;
2241 owait.wait.flags = 0;
2242 owait.wait.func = memcg_oom_wake_function;
2243 owait.wait.private = current;
2244 INIT_LIST_HEAD(&owait.wait.task_list);
2246 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2247 mem_cgroup_mark_under_oom(memcg);
2249 locked = mem_cgroup_oom_trylock(memcg);
2252 mem_cgroup_oom_notify(memcg);
2254 if (locked && !memcg->oom_kill_disable) {
2255 mem_cgroup_unmark_under_oom(memcg);
2256 finish_wait(&memcg_oom_waitq, &owait.wait);
2257 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2258 current->memcg_oom.order);
2261 mem_cgroup_unmark_under_oom(memcg);
2262 finish_wait(&memcg_oom_waitq, &owait.wait);
2266 mem_cgroup_oom_unlock(memcg);
2268 * There is no guarantee that an OOM-lock contender
2269 * sees the wakeups triggered by the OOM kill
2270 * uncharges. Wake any sleepers explicitely.
2272 memcg_oom_recover(memcg);
2275 current->memcg_oom.memcg = NULL;
2276 css_put(&memcg->css);
2281 * Currently used to update mapped file statistics, but the routine can be
2282 * generalized to update other statistics as well.
2284 * Notes: Race condition
2286 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2287 * it tends to be costly. But considering some conditions, we doesn't need
2288 * to do so _always_.
2290 * Considering "charge", lock_page_cgroup() is not required because all
2291 * file-stat operations happen after a page is attached to radix-tree. There
2292 * are no race with "charge".
2294 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2295 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2296 * if there are race with "uncharge". Statistics itself is properly handled
2299 * Considering "move", this is an only case we see a race. To make the race
2300 * small, we check mm->moving_account and detect there are possibility of race
2301 * If there is, we take a lock.
2304 void __mem_cgroup_begin_update_page_stat(struct page *page,
2305 bool *locked, unsigned long *flags)
2307 struct mem_cgroup *memcg;
2308 struct page_cgroup *pc;
2310 pc = lookup_page_cgroup(page);
2312 memcg = pc->mem_cgroup;
2313 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2316 * If this memory cgroup is not under account moving, we don't
2317 * need to take move_lock_mem_cgroup(). Because we already hold
2318 * rcu_read_lock(), any calls to move_account will be delayed until
2319 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2321 if (!mem_cgroup_stolen(memcg))
2324 move_lock_mem_cgroup(memcg, flags);
2325 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2326 move_unlock_mem_cgroup(memcg, flags);
2332 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2334 struct page_cgroup *pc = lookup_page_cgroup(page);
2337 * It's guaranteed that pc->mem_cgroup never changes while
2338 * lock is held because a routine modifies pc->mem_cgroup
2339 * should take move_lock_mem_cgroup().
2341 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2344 void mem_cgroup_update_page_stat(struct page *page,
2345 enum mem_cgroup_stat_index idx, int val)
2347 struct mem_cgroup *memcg;
2348 struct page_cgroup *pc = lookup_page_cgroup(page);
2349 unsigned long uninitialized_var(flags);
2351 if (mem_cgroup_disabled())
2354 VM_BUG_ON(!rcu_read_lock_held());
2355 memcg = pc->mem_cgroup;
2356 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2359 this_cpu_add(memcg->stat->count[idx], val);
2363 * size of first charge trial. "32" comes from vmscan.c's magic value.
2364 * TODO: maybe necessary to use big numbers in big irons.
2366 #define CHARGE_BATCH 32U
2367 struct memcg_stock_pcp {
2368 struct mem_cgroup *cached; /* this never be root cgroup */
2369 unsigned int nr_pages;
2370 struct work_struct work;
2371 unsigned long flags;
2372 #define FLUSHING_CACHED_CHARGE 0
2374 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2375 static DEFINE_MUTEX(percpu_charge_mutex);
2378 * consume_stock: Try to consume stocked charge on this cpu.
2379 * @memcg: memcg to consume from.
2380 * @nr_pages: how many pages to charge.
2382 * The charges will only happen if @memcg matches the current cpu's memcg
2383 * stock, and at least @nr_pages are available in that stock. Failure to
2384 * service an allocation will refill the stock.
2386 * returns true if successful, false otherwise.
2388 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2390 struct memcg_stock_pcp *stock;
2393 if (nr_pages > CHARGE_BATCH)
2396 stock = &get_cpu_var(memcg_stock);
2397 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2398 stock->nr_pages -= nr_pages;
2399 else /* need to call res_counter_charge */
2401 put_cpu_var(memcg_stock);
2406 * Returns stocks cached in percpu to res_counter and reset cached information.
2408 static void drain_stock(struct memcg_stock_pcp *stock)
2410 struct mem_cgroup *old = stock->cached;
2412 if (stock->nr_pages) {
2413 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2415 res_counter_uncharge(&old->res, bytes);
2416 if (do_swap_account)
2417 res_counter_uncharge(&old->memsw, bytes);
2418 stock->nr_pages = 0;
2420 stock->cached = NULL;
2424 * This must be called under preempt disabled or must be called by
2425 * a thread which is pinned to local cpu.
2427 static void drain_local_stock(struct work_struct *dummy)
2429 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2431 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2434 static void __init memcg_stock_init(void)
2438 for_each_possible_cpu(cpu) {
2439 struct memcg_stock_pcp *stock =
2440 &per_cpu(memcg_stock, cpu);
2441 INIT_WORK(&stock->work, drain_local_stock);
2446 * Cache charges(val) which is from res_counter, to local per_cpu area.
2447 * This will be consumed by consume_stock() function, later.
2449 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2451 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2453 if (stock->cached != memcg) { /* reset if necessary */
2455 stock->cached = memcg;
2457 stock->nr_pages += nr_pages;
2458 put_cpu_var(memcg_stock);
2462 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2463 * of the hierarchy under it. sync flag says whether we should block
2464 * until the work is done.
2466 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2470 /* Notify other cpus that system-wide "drain" is running */
2473 for_each_online_cpu(cpu) {
2474 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2475 struct mem_cgroup *memcg;
2477 memcg = stock->cached;
2478 if (!memcg || !stock->nr_pages)
2480 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2482 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2484 drain_local_stock(&stock->work);
2486 schedule_work_on(cpu, &stock->work);
2494 for_each_online_cpu(cpu) {
2495 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2496 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2497 flush_work(&stock->work);
2504 * Tries to drain stocked charges in other cpus. This function is asynchronous
2505 * and just put a work per cpu for draining localy on each cpu. Caller can
2506 * expects some charges will be back to res_counter later but cannot wait for
2509 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2512 * If someone calls draining, avoid adding more kworker runs.
2514 if (!mutex_trylock(&percpu_charge_mutex))
2516 drain_all_stock(root_memcg, false);
2517 mutex_unlock(&percpu_charge_mutex);
2520 /* This is a synchronous drain interface. */
2521 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2523 /* called when force_empty is called */
2524 mutex_lock(&percpu_charge_mutex);
2525 drain_all_stock(root_memcg, true);
2526 mutex_unlock(&percpu_charge_mutex);
2530 * This function drains percpu counter value from DEAD cpu and
2531 * move it to local cpu. Note that this function can be preempted.
2533 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2537 spin_lock(&memcg->pcp_counter_lock);
2538 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2539 long x = per_cpu(memcg->stat->count[i], cpu);
2541 per_cpu(memcg->stat->count[i], cpu) = 0;
2542 memcg->nocpu_base.count[i] += x;
2544 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2545 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2547 per_cpu(memcg->stat->events[i], cpu) = 0;
2548 memcg->nocpu_base.events[i] += x;
2550 spin_unlock(&memcg->pcp_counter_lock);
2553 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2554 unsigned long action,
2557 int cpu = (unsigned long)hcpu;
2558 struct memcg_stock_pcp *stock;
2559 struct mem_cgroup *iter;
2561 if (action == CPU_ONLINE)
2564 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2567 for_each_mem_cgroup(iter)
2568 mem_cgroup_drain_pcp_counter(iter, cpu);
2570 stock = &per_cpu(memcg_stock, cpu);
2576 /* See mem_cgroup_try_charge() for details */
2578 CHARGE_OK, /* success */
2579 CHARGE_RETRY, /* need to retry but retry is not bad */
2580 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2581 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2584 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2585 unsigned int nr_pages, unsigned int min_pages,
2588 unsigned long csize = nr_pages * PAGE_SIZE;
2589 struct mem_cgroup *mem_over_limit;
2590 struct res_counter *fail_res;
2591 unsigned long flags = 0;
2594 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2597 if (!do_swap_account)
2599 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2603 res_counter_uncharge(&memcg->res, csize);
2604 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2605 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2607 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2609 * Never reclaim on behalf of optional batching, retry with a
2610 * single page instead.
2612 if (nr_pages > min_pages)
2613 return CHARGE_RETRY;
2615 if (!(gfp_mask & __GFP_WAIT))
2616 return CHARGE_WOULDBLOCK;
2618 if (gfp_mask & __GFP_NORETRY)
2619 return CHARGE_NOMEM;
2621 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2622 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2623 return CHARGE_RETRY;
2625 * Even though the limit is exceeded at this point, reclaim
2626 * may have been able to free some pages. Retry the charge
2627 * before killing the task.
2629 * Only for regular pages, though: huge pages are rather
2630 * unlikely to succeed so close to the limit, and we fall back
2631 * to regular pages anyway in case of failure.
2633 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2634 return CHARGE_RETRY;
2637 * At task move, charge accounts can be doubly counted. So, it's
2638 * better to wait until the end of task_move if something is going on.
2640 if (mem_cgroup_wait_acct_move(mem_over_limit))
2641 return CHARGE_RETRY;
2644 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2646 return CHARGE_NOMEM;
2650 * mem_cgroup_try_charge - try charging a memcg
2651 * @memcg: memcg to charge
2652 * @nr_pages: number of pages to charge
2653 * @oom: trigger OOM if reclaim fails
2655 * Returns 0 if @memcg was charged successfully, -EINTR if the charge
2656 * was bypassed to root_mem_cgroup, and -ENOMEM if the charge failed.
2658 static int mem_cgroup_try_charge(struct mem_cgroup *memcg,
2660 unsigned int nr_pages,
2663 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2664 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2667 if (mem_cgroup_is_root(memcg))
2670 * Unlike in global OOM situations, memcg is not in a physical
2671 * memory shortage. Allow dying and OOM-killed tasks to
2672 * bypass the last charges so that they can exit quickly and
2673 * free their memory.
2675 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2676 fatal_signal_pending(current)))
2679 if (unlikely(task_in_memcg_oom(current)))
2682 if (gfp_mask & __GFP_NOFAIL)
2685 if (consume_stock(memcg, nr_pages))
2689 bool invoke_oom = oom && !nr_oom_retries;
2691 /* If killed, bypass charge */
2692 if (fatal_signal_pending(current))
2695 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2696 nr_pages, invoke_oom);
2700 case CHARGE_RETRY: /* not in OOM situation but retry */
2703 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2705 case CHARGE_NOMEM: /* OOM routine works */
2706 if (!oom || invoke_oom)
2711 } while (ret != CHARGE_OK);
2713 if (batch > nr_pages)
2714 refill_stock(memcg, batch - nr_pages);
2718 if (!(gfp_mask & __GFP_NOFAIL))
2725 * mem_cgroup_try_charge_mm - try charging a mm
2726 * @mm: mm_struct to charge
2727 * @nr_pages: number of pages to charge
2728 * @oom: trigger OOM if reclaim fails
2730 * Returns the charged mem_cgroup associated with the given mm_struct or
2731 * NULL the charge failed.
2733 static struct mem_cgroup *mem_cgroup_try_charge_mm(struct mm_struct *mm,
2735 unsigned int nr_pages,
2739 struct mem_cgroup *memcg;
2742 memcg = get_mem_cgroup_from_mm(mm);
2743 ret = mem_cgroup_try_charge(memcg, gfp_mask, nr_pages, oom);
2744 css_put(&memcg->css);
2746 memcg = root_mem_cgroup;
2754 * Somemtimes we have to undo a charge we got by try_charge().
2755 * This function is for that and do uncharge, put css's refcnt.
2756 * gotten by try_charge().
2758 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2759 unsigned int nr_pages)
2761 if (!mem_cgroup_is_root(memcg)) {
2762 unsigned long bytes = nr_pages * PAGE_SIZE;
2764 res_counter_uncharge(&memcg->res, bytes);
2765 if (do_swap_account)
2766 res_counter_uncharge(&memcg->memsw, bytes);
2771 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2772 * This is useful when moving usage to parent cgroup.
2774 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2775 unsigned int nr_pages)
2777 unsigned long bytes = nr_pages * PAGE_SIZE;
2779 if (mem_cgroup_is_root(memcg))
2782 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2783 if (do_swap_account)
2784 res_counter_uncharge_until(&memcg->memsw,
2785 memcg->memsw.parent, bytes);
2789 * A helper function to get mem_cgroup from ID. must be called under
2790 * rcu_read_lock(). The caller is responsible for calling
2791 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2792 * refcnt from swap can be called against removed memcg.)
2794 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2796 /* ID 0 is unused ID */
2799 return mem_cgroup_from_id(id);
2802 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2804 struct mem_cgroup *memcg = NULL;
2805 struct page_cgroup *pc;
2809 VM_BUG_ON_PAGE(!PageLocked(page), page);
2811 pc = lookup_page_cgroup(page);
2812 lock_page_cgroup(pc);
2813 if (PageCgroupUsed(pc)) {
2814 memcg = pc->mem_cgroup;
2815 if (memcg && !css_tryget_online(&memcg->css))
2817 } else if (PageSwapCache(page)) {
2818 ent.val = page_private(page);
2819 id = lookup_swap_cgroup_id(ent);
2821 memcg = mem_cgroup_lookup(id);
2822 if (memcg && !css_tryget_online(&memcg->css))
2826 unlock_page_cgroup(pc);
2830 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2832 unsigned int nr_pages,
2833 enum charge_type ctype,
2836 struct page_cgroup *pc = lookup_page_cgroup(page);
2837 struct zone *uninitialized_var(zone);
2838 struct lruvec *lruvec;
2839 bool was_on_lru = false;
2842 lock_page_cgroup(pc);
2843 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2845 * we don't need page_cgroup_lock about tail pages, becase they are not
2846 * accessed by any other context at this point.
2850 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2851 * may already be on some other mem_cgroup's LRU. Take care of it.
2854 zone = page_zone(page);
2855 spin_lock_irq(&zone->lru_lock);
2856 if (PageLRU(page)) {
2857 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2859 del_page_from_lru_list(page, lruvec, page_lru(page));
2864 pc->mem_cgroup = memcg;
2866 * We access a page_cgroup asynchronously without lock_page_cgroup().
2867 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2868 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2869 * before USED bit, we need memory barrier here.
2870 * See mem_cgroup_add_lru_list(), etc.
2873 SetPageCgroupUsed(pc);
2877 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2878 VM_BUG_ON_PAGE(PageLRU(page), page);
2880 add_page_to_lru_list(page, lruvec, page_lru(page));
2882 spin_unlock_irq(&zone->lru_lock);
2885 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2890 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2891 unlock_page_cgroup(pc);
2894 * "charge_statistics" updated event counter. Then, check it.
2895 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2896 * if they exceeds softlimit.
2898 memcg_check_events(memcg, page);
2901 static DEFINE_MUTEX(set_limit_mutex);
2903 #ifdef CONFIG_MEMCG_KMEM
2904 static DEFINE_MUTEX(activate_kmem_mutex);
2906 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2908 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2909 memcg_kmem_is_active(memcg);
2913 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2914 * in the memcg_cache_params struct.
2916 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2918 struct kmem_cache *cachep;
2920 VM_BUG_ON(p->is_root_cache);
2921 cachep = p->root_cache;
2922 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2925 #ifdef CONFIG_SLABINFO
2926 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2928 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2929 struct memcg_cache_params *params;
2931 if (!memcg_can_account_kmem(memcg))
2934 print_slabinfo_header(m);
2936 mutex_lock(&memcg->slab_caches_mutex);
2937 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2938 cache_show(memcg_params_to_cache(params), m);
2939 mutex_unlock(&memcg->slab_caches_mutex);
2945 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2947 struct res_counter *fail_res;
2950 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2954 ret = mem_cgroup_try_charge(memcg, gfp, size >> PAGE_SHIFT,
2955 oom_gfp_allowed(gfp));
2956 if (ret == -EINTR) {
2958 * mem_cgroup_try_charge() chosed to bypass to root due to
2959 * OOM kill or fatal signal. Since our only options are to
2960 * either fail the allocation or charge it to this cgroup, do
2961 * it as a temporary condition. But we can't fail. From a
2962 * kmem/slab perspective, the cache has already been selected,
2963 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2966 * This condition will only trigger if the task entered
2967 * memcg_charge_kmem in a sane state, but was OOM-killed during
2968 * mem_cgroup_try_charge() above. Tasks that were already
2969 * dying when the allocation triggers should have been already
2970 * directed to the root cgroup in memcontrol.h
2972 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2973 if (do_swap_account)
2974 res_counter_charge_nofail(&memcg->memsw, size,
2978 res_counter_uncharge(&memcg->kmem, size);
2983 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2985 res_counter_uncharge(&memcg->res, size);
2986 if (do_swap_account)
2987 res_counter_uncharge(&memcg->memsw, size);
2990 if (res_counter_uncharge(&memcg->kmem, size))
2994 * Releases a reference taken in kmem_cgroup_css_offline in case
2995 * this last uncharge is racing with the offlining code or it is
2996 * outliving the memcg existence.
2998 * The memory barrier imposed by test&clear is paired with the
2999 * explicit one in memcg_kmem_mark_dead().
3001 if (memcg_kmem_test_and_clear_dead(memcg))
3002 css_put(&memcg->css);
3006 * helper for acessing a memcg's index. It will be used as an index in the
3007 * child cache array in kmem_cache, and also to derive its name. This function
3008 * will return -1 when this is not a kmem-limited memcg.
3010 int memcg_cache_id(struct mem_cgroup *memcg)
3012 return memcg ? memcg->kmemcg_id : -1;
3015 static size_t memcg_caches_array_size(int num_groups)
3018 if (num_groups <= 0)
3021 size = 2 * num_groups;
3022 if (size < MEMCG_CACHES_MIN_SIZE)
3023 size = MEMCG_CACHES_MIN_SIZE;
3024 else if (size > MEMCG_CACHES_MAX_SIZE)
3025 size = MEMCG_CACHES_MAX_SIZE;
3031 * We should update the current array size iff all caches updates succeed. This
3032 * can only be done from the slab side. The slab mutex needs to be held when
3035 void memcg_update_array_size(int num)
3037 if (num > memcg_limited_groups_array_size)
3038 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3041 static void kmem_cache_destroy_work_func(struct work_struct *w);
3043 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3045 struct memcg_cache_params *cur_params = s->memcg_params;
3047 VM_BUG_ON(!is_root_cache(s));
3049 if (num_groups > memcg_limited_groups_array_size) {
3051 struct memcg_cache_params *new_params;
3052 ssize_t size = memcg_caches_array_size(num_groups);
3054 size *= sizeof(void *);
3055 size += offsetof(struct memcg_cache_params, memcg_caches);
3057 new_params = kzalloc(size, GFP_KERNEL);
3061 new_params->is_root_cache = true;
3064 * There is the chance it will be bigger than
3065 * memcg_limited_groups_array_size, if we failed an allocation
3066 * in a cache, in which case all caches updated before it, will
3067 * have a bigger array.
3069 * But if that is the case, the data after
3070 * memcg_limited_groups_array_size is certainly unused
3072 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3073 if (!cur_params->memcg_caches[i])
3075 new_params->memcg_caches[i] =
3076 cur_params->memcg_caches[i];
3080 * Ideally, we would wait until all caches succeed, and only
3081 * then free the old one. But this is not worth the extra
3082 * pointer per-cache we'd have to have for this.
3084 * It is not a big deal if some caches are left with a size
3085 * bigger than the others. And all updates will reset this
3088 rcu_assign_pointer(s->memcg_params, new_params);
3090 kfree_rcu(cur_params, rcu_head);
3095 char *memcg_create_cache_name(struct mem_cgroup *memcg,
3096 struct kmem_cache *root_cache)
3098 static char *buf = NULL;
3101 * We need a mutex here to protect the shared buffer. Since this is
3102 * expected to be called only on cache creation, we can employ the
3103 * slab_mutex for that purpose.
3105 lockdep_assert_held(&slab_mutex);
3108 buf = kmalloc(NAME_MAX + 1, GFP_KERNEL);
3113 cgroup_name(memcg->css.cgroup, buf, NAME_MAX + 1);
3114 return kasprintf(GFP_KERNEL, "%s(%d:%s)", root_cache->name,
3115 memcg_cache_id(memcg), buf);
3118 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3119 struct kmem_cache *root_cache)
3123 if (!memcg_kmem_enabled())
3127 size = offsetof(struct memcg_cache_params, memcg_caches);
3128 size += memcg_limited_groups_array_size * sizeof(void *);
3130 size = sizeof(struct memcg_cache_params);
3132 s->memcg_params = kzalloc(size, GFP_KERNEL);
3133 if (!s->memcg_params)
3137 s->memcg_params->memcg = memcg;
3138 s->memcg_params->root_cache = root_cache;
3139 INIT_WORK(&s->memcg_params->destroy,
3140 kmem_cache_destroy_work_func);
3141 css_get(&memcg->css);
3143 s->memcg_params->is_root_cache = true;
3148 void memcg_free_cache_params(struct kmem_cache *s)
3150 if (!s->memcg_params)
3152 if (!s->memcg_params->is_root_cache)
3153 css_put(&s->memcg_params->memcg->css);
3154 kfree(s->memcg_params);
3157 void memcg_register_cache(struct kmem_cache *s)
3159 struct kmem_cache *root;
3160 struct mem_cgroup *memcg;
3163 if (is_root_cache(s))
3167 * Holding the slab_mutex assures nobody will touch the memcg_caches
3168 * array while we are modifying it.
3170 lockdep_assert_held(&slab_mutex);
3172 root = s->memcg_params->root_cache;
3173 memcg = s->memcg_params->memcg;
3174 id = memcg_cache_id(memcg);
3177 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3178 * barrier here to ensure nobody will see the kmem_cache partially
3184 * Initialize the pointer to this cache in its parent's memcg_params
3185 * before adding it to the memcg_slab_caches list, otherwise we can
3186 * fail to convert memcg_params_to_cache() while traversing the list.
3188 VM_BUG_ON(root->memcg_params->memcg_caches[id]);
3189 root->memcg_params->memcg_caches[id] = s;
3191 mutex_lock(&memcg->slab_caches_mutex);
3192 list_add(&s->memcg_params->list, &memcg->memcg_slab_caches);
3193 mutex_unlock(&memcg->slab_caches_mutex);
3196 void memcg_unregister_cache(struct kmem_cache *s)
3198 struct kmem_cache *root;
3199 struct mem_cgroup *memcg;
3202 if (is_root_cache(s))
3206 * Holding the slab_mutex assures nobody will touch the memcg_caches
3207 * array while we are modifying it.
3209 lockdep_assert_held(&slab_mutex);
3211 root = s->memcg_params->root_cache;
3212 memcg = s->memcg_params->memcg;
3213 id = memcg_cache_id(memcg);
3215 mutex_lock(&memcg->slab_caches_mutex);
3216 list_del(&s->memcg_params->list);
3217 mutex_unlock(&memcg->slab_caches_mutex);
3220 * Clear the pointer to this cache in its parent's memcg_params only
3221 * after removing it from the memcg_slab_caches list, otherwise we can
3222 * fail to convert memcg_params_to_cache() while traversing the list.
3224 VM_BUG_ON(root->memcg_params->memcg_caches[id] != s);
3225 root->memcg_params->memcg_caches[id] = NULL;
3229 * During the creation a new cache, we need to disable our accounting mechanism
3230 * altogether. This is true even if we are not creating, but rather just
3231 * enqueing new caches to be created.
3233 * This is because that process will trigger allocations; some visible, like
3234 * explicit kmallocs to auxiliary data structures, name strings and internal
3235 * cache structures; some well concealed, like INIT_WORK() that can allocate
3236 * objects during debug.
3238 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3239 * to it. This may not be a bounded recursion: since the first cache creation
3240 * failed to complete (waiting on the allocation), we'll just try to create the
3241 * cache again, failing at the same point.
3243 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3244 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3245 * inside the following two functions.
3247 static inline void memcg_stop_kmem_account(void)
3249 VM_BUG_ON(!current->mm);
3250 current->memcg_kmem_skip_account++;
3253 static inline void memcg_resume_kmem_account(void)
3255 VM_BUG_ON(!current->mm);
3256 current->memcg_kmem_skip_account--;
3259 static void kmem_cache_destroy_work_func(struct work_struct *w)
3261 struct kmem_cache *cachep;
3262 struct memcg_cache_params *p;
3264 p = container_of(w, struct memcg_cache_params, destroy);
3266 cachep = memcg_params_to_cache(p);
3269 * If we get down to 0 after shrink, we could delete right away.
3270 * However, memcg_release_pages() already puts us back in the workqueue
3271 * in that case. If we proceed deleting, we'll get a dangling
3272 * reference, and removing the object from the workqueue in that case
3273 * is unnecessary complication. We are not a fast path.
3275 * Note that this case is fundamentally different from racing with
3276 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3277 * kmem_cache_shrink, not only we would be reinserting a dead cache
3278 * into the queue, but doing so from inside the worker racing to
3281 * So if we aren't down to zero, we'll just schedule a worker and try
3284 if (atomic_read(&cachep->memcg_params->nr_pages) != 0)
3285 kmem_cache_shrink(cachep);
3287 kmem_cache_destroy(cachep);
3290 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3292 if (!cachep->memcg_params->dead)
3296 * There are many ways in which we can get here.
3298 * We can get to a memory-pressure situation while the delayed work is
3299 * still pending to run. The vmscan shrinkers can then release all
3300 * cache memory and get us to destruction. If this is the case, we'll
3301 * be executed twice, which is a bug (the second time will execute over
3302 * bogus data). In this case, cancelling the work should be fine.
3304 * But we can also get here from the worker itself, if
3305 * kmem_cache_shrink is enough to shake all the remaining objects and
3306 * get the page count to 0. In this case, we'll deadlock if we try to
3307 * cancel the work (the worker runs with an internal lock held, which
3308 * is the same lock we would hold for cancel_work_sync().)
3310 * Since we can't possibly know who got us here, just refrain from
3311 * running if there is already work pending
3313 if (work_pending(&cachep->memcg_params->destroy))
3316 * We have to defer the actual destroying to a workqueue, because
3317 * we might currently be in a context that cannot sleep.
3319 schedule_work(&cachep->memcg_params->destroy);
3322 int __kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3324 struct kmem_cache *c;
3328 * If the cache is being destroyed, we trust that there is no one else
3329 * requesting objects from it. Even if there are, the sanity checks in
3330 * kmem_cache_destroy should caught this ill-case.
3332 * Still, we don't want anyone else freeing memcg_caches under our
3333 * noses, which can happen if a new memcg comes to life. As usual,
3334 * we'll take the activate_kmem_mutex to protect ourselves against
3337 mutex_lock(&activate_kmem_mutex);
3338 for_each_memcg_cache_index(i) {
3339 c = cache_from_memcg_idx(s, i);
3344 * We will now manually delete the caches, so to avoid races
3345 * we need to cancel all pending destruction workers and
3346 * proceed with destruction ourselves.
3348 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3349 * and that could spawn the workers again: it is likely that
3350 * the cache still have active pages until this very moment.
3351 * This would lead us back to mem_cgroup_destroy_cache.
3353 * But that will not execute at all if the "dead" flag is not
3354 * set, so flip it down to guarantee we are in control.
3356 c->memcg_params->dead = false;
3357 cancel_work_sync(&c->memcg_params->destroy);
3358 kmem_cache_destroy(c);
3360 if (cache_from_memcg_idx(s, i))
3363 mutex_unlock(&activate_kmem_mutex);
3367 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3369 struct kmem_cache *cachep;
3370 struct memcg_cache_params *params;
3372 if (!memcg_kmem_is_active(memcg))
3375 mutex_lock(&memcg->slab_caches_mutex);
3376 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3377 cachep = memcg_params_to_cache(params);
3378 cachep->memcg_params->dead = true;
3379 schedule_work(&cachep->memcg_params->destroy);
3381 mutex_unlock(&memcg->slab_caches_mutex);
3384 struct create_work {
3385 struct mem_cgroup *memcg;
3386 struct kmem_cache *cachep;
3387 struct work_struct work;
3390 static void memcg_create_cache_work_func(struct work_struct *w)
3392 struct create_work *cw = container_of(w, struct create_work, work);
3393 struct mem_cgroup *memcg = cw->memcg;
3394 struct kmem_cache *cachep = cw->cachep;
3396 kmem_cache_create_memcg(memcg, cachep);
3397 css_put(&memcg->css);
3402 * Enqueue the creation of a per-memcg kmem_cache.
3404 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3405 struct kmem_cache *cachep)
3407 struct create_work *cw;
3409 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3411 css_put(&memcg->css);
3416 cw->cachep = cachep;
3418 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3419 schedule_work(&cw->work);
3422 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3423 struct kmem_cache *cachep)
3426 * We need to stop accounting when we kmalloc, because if the
3427 * corresponding kmalloc cache is not yet created, the first allocation
3428 * in __memcg_create_cache_enqueue will recurse.
3430 * However, it is better to enclose the whole function. Depending on
3431 * the debugging options enabled, INIT_WORK(), for instance, can
3432 * trigger an allocation. This too, will make us recurse. Because at
3433 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3434 * the safest choice is to do it like this, wrapping the whole function.
3436 memcg_stop_kmem_account();
3437 __memcg_create_cache_enqueue(memcg, cachep);
3438 memcg_resume_kmem_account();
3441 * Return the kmem_cache we're supposed to use for a slab allocation.
3442 * We try to use the current memcg's version of the cache.
3444 * If the cache does not exist yet, if we are the first user of it,
3445 * we either create it immediately, if possible, or create it asynchronously
3447 * In the latter case, we will let the current allocation go through with
3448 * the original cache.
3450 * Can't be called in interrupt context or from kernel threads.
3451 * This function needs to be called with rcu_read_lock() held.
3453 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3456 struct mem_cgroup *memcg;
3457 struct kmem_cache *memcg_cachep;
3459 VM_BUG_ON(!cachep->memcg_params);
3460 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3462 if (!current->mm || current->memcg_kmem_skip_account)
3466 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3468 if (!memcg_can_account_kmem(memcg))
3471 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3472 if (likely(memcg_cachep)) {
3473 cachep = memcg_cachep;
3477 /* The corresponding put will be done in the workqueue. */
3478 if (!css_tryget_online(&memcg->css))
3483 * If we are in a safe context (can wait, and not in interrupt
3484 * context), we could be be predictable and return right away.
3485 * This would guarantee that the allocation being performed
3486 * already belongs in the new cache.
3488 * However, there are some clashes that can arrive from locking.
3489 * For instance, because we acquire the slab_mutex while doing
3490 * kmem_cache_dup, this means no further allocation could happen
3491 * with the slab_mutex held.
3493 * Also, because cache creation issue get_online_cpus(), this
3494 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3495 * that ends up reversed during cpu hotplug. (cpuset allocates
3496 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3497 * better to defer everything.
3499 memcg_create_cache_enqueue(memcg, cachep);
3505 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3508 * We need to verify if the allocation against current->mm->owner's memcg is
3509 * possible for the given order. But the page is not allocated yet, so we'll
3510 * need a further commit step to do the final arrangements.
3512 * It is possible for the task to switch cgroups in this mean time, so at
3513 * commit time, we can't rely on task conversion any longer. We'll then use
3514 * the handle argument to return to the caller which cgroup we should commit
3515 * against. We could also return the memcg directly and avoid the pointer
3516 * passing, but a boolean return value gives better semantics considering
3517 * the compiled-out case as well.
3519 * Returning true means the allocation is possible.
3522 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3524 struct mem_cgroup *memcg;
3530 * Disabling accounting is only relevant for some specific memcg
3531 * internal allocations. Therefore we would initially not have such
3532 * check here, since direct calls to the page allocator that are marked
3533 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3534 * concerned with cache allocations, and by having this test at
3535 * memcg_kmem_get_cache, we are already able to relay the allocation to
3536 * the root cache and bypass the memcg cache altogether.
3538 * There is one exception, though: the SLUB allocator does not create
3539 * large order caches, but rather service large kmallocs directly from
3540 * the page allocator. Therefore, the following sequence when backed by
3541 * the SLUB allocator:
3543 * memcg_stop_kmem_account();
3544 * kmalloc(<large_number>)
3545 * memcg_resume_kmem_account();
3547 * would effectively ignore the fact that we should skip accounting,
3548 * since it will drive us directly to this function without passing
3549 * through the cache selector memcg_kmem_get_cache. Such large
3550 * allocations are extremely rare but can happen, for instance, for the
3551 * cache arrays. We bring this test here.
3553 if (!current->mm || current->memcg_kmem_skip_account)
3556 memcg = get_mem_cgroup_from_mm(current->mm);
3558 if (!memcg_can_account_kmem(memcg)) {
3559 css_put(&memcg->css);
3563 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3567 css_put(&memcg->css);
3571 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3574 struct page_cgroup *pc;
3576 VM_BUG_ON(mem_cgroup_is_root(memcg));
3578 /* The page allocation failed. Revert */
3580 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3584 pc = lookup_page_cgroup(page);
3585 lock_page_cgroup(pc);
3586 pc->mem_cgroup = memcg;
3587 SetPageCgroupUsed(pc);
3588 unlock_page_cgroup(pc);
3591 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3593 struct mem_cgroup *memcg = NULL;
3594 struct page_cgroup *pc;
3597 pc = lookup_page_cgroup(page);
3599 * Fast unlocked return. Theoretically might have changed, have to
3600 * check again after locking.
3602 if (!PageCgroupUsed(pc))
3605 lock_page_cgroup(pc);
3606 if (PageCgroupUsed(pc)) {
3607 memcg = pc->mem_cgroup;
3608 ClearPageCgroupUsed(pc);
3610 unlock_page_cgroup(pc);
3613 * We trust that only if there is a memcg associated with the page, it
3614 * is a valid allocation
3619 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3620 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3623 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3626 #endif /* CONFIG_MEMCG_KMEM */
3628 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3630 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3632 * Because tail pages are not marked as "used", set it. We're under
3633 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3634 * charge/uncharge will be never happen and move_account() is done under
3635 * compound_lock(), so we don't have to take care of races.
3637 void mem_cgroup_split_huge_fixup(struct page *head)
3639 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3640 struct page_cgroup *pc;
3641 struct mem_cgroup *memcg;
3644 if (mem_cgroup_disabled())
3647 memcg = head_pc->mem_cgroup;
3648 for (i = 1; i < HPAGE_PMD_NR; i++) {
3650 pc->mem_cgroup = memcg;
3651 smp_wmb();/* see __commit_charge() */
3652 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3654 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3657 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3660 * mem_cgroup_move_account - move account of the page
3662 * @nr_pages: number of regular pages (>1 for huge pages)
3663 * @pc: page_cgroup of the page.
3664 * @from: mem_cgroup which the page is moved from.
3665 * @to: mem_cgroup which the page is moved to. @from != @to.
3667 * The caller must confirm following.
3668 * - page is not on LRU (isolate_page() is useful.)
3669 * - compound_lock is held when nr_pages > 1
3671 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3674 static int mem_cgroup_move_account(struct page *page,
3675 unsigned int nr_pages,
3676 struct page_cgroup *pc,
3677 struct mem_cgroup *from,
3678 struct mem_cgroup *to)
3680 unsigned long flags;
3682 bool anon = PageAnon(page);
3684 VM_BUG_ON(from == to);
3685 VM_BUG_ON_PAGE(PageLRU(page), page);
3687 * The page is isolated from LRU. So, collapse function
3688 * will not handle this page. But page splitting can happen.
3689 * Do this check under compound_page_lock(). The caller should
3693 if (nr_pages > 1 && !PageTransHuge(page))
3696 lock_page_cgroup(pc);
3699 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3702 move_lock_mem_cgroup(from, &flags);
3704 if (!anon && page_mapped(page)) {
3705 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3707 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3711 if (PageWriteback(page)) {
3712 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3714 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3718 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3720 /* caller should have done css_get */
3721 pc->mem_cgroup = to;
3722 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3723 move_unlock_mem_cgroup(from, &flags);
3726 unlock_page_cgroup(pc);
3730 memcg_check_events(to, page);
3731 memcg_check_events(from, page);
3737 * mem_cgroup_move_parent - moves page to the parent group
3738 * @page: the page to move
3739 * @pc: page_cgroup of the page
3740 * @child: page's cgroup
3742 * move charges to its parent or the root cgroup if the group has no
3743 * parent (aka use_hierarchy==0).
3744 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3745 * mem_cgroup_move_account fails) the failure is always temporary and
3746 * it signals a race with a page removal/uncharge or migration. In the
3747 * first case the page is on the way out and it will vanish from the LRU
3748 * on the next attempt and the call should be retried later.
3749 * Isolation from the LRU fails only if page has been isolated from
3750 * the LRU since we looked at it and that usually means either global
3751 * reclaim or migration going on. The page will either get back to the
3753 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3754 * (!PageCgroupUsed) or moved to a different group. The page will
3755 * disappear in the next attempt.
3757 static int mem_cgroup_move_parent(struct page *page,
3758 struct page_cgroup *pc,
3759 struct mem_cgroup *child)
3761 struct mem_cgroup *parent;
3762 unsigned int nr_pages;
3763 unsigned long uninitialized_var(flags);
3766 VM_BUG_ON(mem_cgroup_is_root(child));
3769 if (!get_page_unless_zero(page))
3771 if (isolate_lru_page(page))
3774 nr_pages = hpage_nr_pages(page);
3776 parent = parent_mem_cgroup(child);
3778 * If no parent, move charges to root cgroup.
3781 parent = root_mem_cgroup;
3784 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3785 flags = compound_lock_irqsave(page);
3788 ret = mem_cgroup_move_account(page, nr_pages,
3791 __mem_cgroup_cancel_local_charge(child, nr_pages);
3794 compound_unlock_irqrestore(page, flags);
3795 putback_lru_page(page);
3802 int mem_cgroup_charge_anon(struct page *page,
3803 struct mm_struct *mm, gfp_t gfp_mask)
3805 unsigned int nr_pages = 1;
3806 struct mem_cgroup *memcg;
3809 if (mem_cgroup_disabled())
3812 VM_BUG_ON_PAGE(page_mapped(page), page);
3813 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3816 if (PageTransHuge(page)) {
3817 nr_pages <<= compound_order(page);
3818 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3820 * Never OOM-kill a process for a huge page. The
3821 * fault handler will fall back to regular pages.
3826 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, nr_pages, oom);
3829 __mem_cgroup_commit_charge(memcg, page, nr_pages,
3830 MEM_CGROUP_CHARGE_TYPE_ANON, false);
3835 * While swap-in, try_charge -> commit or cancel, the page is locked.
3836 * And when try_charge() successfully returns, one refcnt to memcg without
3837 * struct page_cgroup is acquired. This refcnt will be consumed by
3838 * "commit()" or removed by "cancel()"
3840 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3843 struct mem_cgroup **memcgp)
3845 struct mem_cgroup *memcg = NULL;
3846 struct page_cgroup *pc;
3849 pc = lookup_page_cgroup(page);
3851 * Every swap fault against a single page tries to charge the
3852 * page, bail as early as possible. shmem_unuse() encounters
3853 * already charged pages, too. The USED bit is protected by
3854 * the page lock, which serializes swap cache removal, which
3855 * in turn serializes uncharging.
3857 if (PageCgroupUsed(pc))
3859 if (do_swap_account)
3860 memcg = try_get_mem_cgroup_from_page(page);
3862 memcg = get_mem_cgroup_from_mm(mm);
3863 ret = mem_cgroup_try_charge(memcg, mask, 1, true);
3864 css_put(&memcg->css);
3866 memcg = root_mem_cgroup;
3874 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3875 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3877 if (mem_cgroup_disabled()) {
3882 * A racing thread's fault, or swapoff, may have already
3883 * updated the pte, and even removed page from swap cache: in
3884 * those cases unuse_pte()'s pte_same() test will fail; but
3885 * there's also a KSM case which does need to charge the page.
3887 if (!PageSwapCache(page)) {
3888 struct mem_cgroup *memcg;
3890 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3896 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3899 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3901 if (mem_cgroup_disabled())
3905 __mem_cgroup_cancel_charge(memcg, 1);
3909 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3910 enum charge_type ctype)
3912 if (mem_cgroup_disabled())
3917 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3919 * Now swap is on-memory. This means this page may be
3920 * counted both as mem and swap....double count.
3921 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3922 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3923 * may call delete_from_swap_cache() before reach here.
3925 if (do_swap_account && PageSwapCache(page)) {
3926 swp_entry_t ent = {.val = page_private(page)};
3927 mem_cgroup_uncharge_swap(ent);
3931 void mem_cgroup_commit_charge_swapin(struct page *page,
3932 struct mem_cgroup *memcg)
3934 __mem_cgroup_commit_charge_swapin(page, memcg,
3935 MEM_CGROUP_CHARGE_TYPE_ANON);
3938 int mem_cgroup_charge_file(struct page *page, struct mm_struct *mm,
3941 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3942 struct mem_cgroup *memcg;
3945 if (mem_cgroup_disabled())
3947 if (PageCompound(page))
3950 if (PageSwapCache(page)) { /* shmem */
3951 ret = __mem_cgroup_try_charge_swapin(mm, page,
3955 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3960 * Page cache insertions can happen without an actual mm
3961 * context, e.g. during disk probing on boot.
3964 memcg = root_mem_cgroup;
3966 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3970 __mem_cgroup_commit_charge(memcg, page, 1, type, false);
3974 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3975 unsigned int nr_pages,
3976 const enum charge_type ctype)
3978 struct memcg_batch_info *batch = NULL;
3979 bool uncharge_memsw = true;
3981 /* If swapout, usage of swap doesn't decrease */
3982 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3983 uncharge_memsw = false;
3985 batch = ¤t->memcg_batch;
3987 * In usual, we do css_get() when we remember memcg pointer.
3988 * But in this case, we keep res->usage until end of a series of
3989 * uncharges. Then, it's ok to ignore memcg's refcnt.
3992 batch->memcg = memcg;
3994 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3995 * In those cases, all pages freed continuously can be expected to be in
3996 * the same cgroup and we have chance to coalesce uncharges.
3997 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3998 * because we want to do uncharge as soon as possible.
4001 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4002 goto direct_uncharge;
4005 goto direct_uncharge;
4008 * In typical case, batch->memcg == mem. This means we can
4009 * merge a series of uncharges to an uncharge of res_counter.
4010 * If not, we uncharge res_counter ony by one.
4012 if (batch->memcg != memcg)
4013 goto direct_uncharge;
4014 /* remember freed charge and uncharge it later */
4017 batch->memsw_nr_pages++;
4020 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4022 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4023 if (unlikely(batch->memcg != memcg))
4024 memcg_oom_recover(memcg);
4028 * uncharge if !page_mapped(page)
4030 static struct mem_cgroup *
4031 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4034 struct mem_cgroup *memcg = NULL;
4035 unsigned int nr_pages = 1;
4036 struct page_cgroup *pc;
4039 if (mem_cgroup_disabled())
4042 if (PageTransHuge(page)) {
4043 nr_pages <<= compound_order(page);
4044 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
4047 * Check if our page_cgroup is valid
4049 pc = lookup_page_cgroup(page);
4050 if (unlikely(!PageCgroupUsed(pc)))
4053 lock_page_cgroup(pc);
4055 memcg = pc->mem_cgroup;
4057 if (!PageCgroupUsed(pc))
4060 anon = PageAnon(page);
4063 case MEM_CGROUP_CHARGE_TYPE_ANON:
4065 * Generally PageAnon tells if it's the anon statistics to be
4066 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4067 * used before page reached the stage of being marked PageAnon.
4071 case MEM_CGROUP_CHARGE_TYPE_DROP:
4072 /* See mem_cgroup_prepare_migration() */
4073 if (page_mapped(page))
4076 * Pages under migration may not be uncharged. But
4077 * end_migration() /must/ be the one uncharging the
4078 * unused post-migration page and so it has to call
4079 * here with the migration bit still set. See the
4080 * res_counter handling below.
4082 if (!end_migration && PageCgroupMigration(pc))
4085 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4086 if (!PageAnon(page)) { /* Shared memory */
4087 if (page->mapping && !page_is_file_cache(page))
4089 } else if (page_mapped(page)) /* Anon */
4096 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4098 ClearPageCgroupUsed(pc);
4100 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4101 * freed from LRU. This is safe because uncharged page is expected not
4102 * to be reused (freed soon). Exception is SwapCache, it's handled by
4103 * special functions.
4106 unlock_page_cgroup(pc);
4108 * even after unlock, we have memcg->res.usage here and this memcg
4109 * will never be freed, so it's safe to call css_get().
4111 memcg_check_events(memcg, page);
4112 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4113 mem_cgroup_swap_statistics(memcg, true);
4114 css_get(&memcg->css);
4117 * Migration does not charge the res_counter for the
4118 * replacement page, so leave it alone when phasing out the
4119 * page that is unused after the migration.
4121 if (!end_migration && !mem_cgroup_is_root(memcg))
4122 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4127 unlock_page_cgroup(pc);
4131 void mem_cgroup_uncharge_page(struct page *page)
4134 if (page_mapped(page))
4136 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4138 * If the page is in swap cache, uncharge should be deferred
4139 * to the swap path, which also properly accounts swap usage
4140 * and handles memcg lifetime.
4142 * Note that this check is not stable and reclaim may add the
4143 * page to swap cache at any time after this. However, if the
4144 * page is not in swap cache by the time page->mapcount hits
4145 * 0, there won't be any page table references to the swap
4146 * slot, and reclaim will free it and not actually write the
4149 if (PageSwapCache(page))
4151 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4154 void mem_cgroup_uncharge_cache_page(struct page *page)
4156 VM_BUG_ON_PAGE(page_mapped(page), page);
4157 VM_BUG_ON_PAGE(page->mapping, page);
4158 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4162 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4163 * In that cases, pages are freed continuously and we can expect pages
4164 * are in the same memcg. All these calls itself limits the number of
4165 * pages freed at once, then uncharge_start/end() is called properly.
4166 * This may be called prural(2) times in a context,
4169 void mem_cgroup_uncharge_start(void)
4171 current->memcg_batch.do_batch++;
4172 /* We can do nest. */
4173 if (current->memcg_batch.do_batch == 1) {
4174 current->memcg_batch.memcg = NULL;
4175 current->memcg_batch.nr_pages = 0;
4176 current->memcg_batch.memsw_nr_pages = 0;
4180 void mem_cgroup_uncharge_end(void)
4182 struct memcg_batch_info *batch = ¤t->memcg_batch;
4184 if (!batch->do_batch)
4188 if (batch->do_batch) /* If stacked, do nothing. */
4194 * This "batch->memcg" is valid without any css_get/put etc...
4195 * bacause we hide charges behind us.
4197 if (batch->nr_pages)
4198 res_counter_uncharge(&batch->memcg->res,
4199 batch->nr_pages * PAGE_SIZE);
4200 if (batch->memsw_nr_pages)
4201 res_counter_uncharge(&batch->memcg->memsw,
4202 batch->memsw_nr_pages * PAGE_SIZE);
4203 memcg_oom_recover(batch->memcg);
4204 /* forget this pointer (for sanity check) */
4205 batch->memcg = NULL;
4210 * called after __delete_from_swap_cache() and drop "page" account.
4211 * memcg information is recorded to swap_cgroup of "ent"
4214 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4216 struct mem_cgroup *memcg;
4217 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4219 if (!swapout) /* this was a swap cache but the swap is unused ! */
4220 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4222 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4225 * record memcg information, if swapout && memcg != NULL,
4226 * css_get() was called in uncharge().
4228 if (do_swap_account && swapout && memcg)
4229 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4233 #ifdef CONFIG_MEMCG_SWAP
4235 * called from swap_entry_free(). remove record in swap_cgroup and
4236 * uncharge "memsw" account.
4238 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4240 struct mem_cgroup *memcg;
4243 if (!do_swap_account)
4246 id = swap_cgroup_record(ent, 0);
4248 memcg = mem_cgroup_lookup(id);
4251 * We uncharge this because swap is freed. This memcg can
4252 * be obsolete one. We avoid calling css_tryget_online().
4254 if (!mem_cgroup_is_root(memcg))
4255 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4256 mem_cgroup_swap_statistics(memcg, false);
4257 css_put(&memcg->css);
4263 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4264 * @entry: swap entry to be moved
4265 * @from: mem_cgroup which the entry is moved from
4266 * @to: mem_cgroup which the entry is moved to
4268 * It succeeds only when the swap_cgroup's record for this entry is the same
4269 * as the mem_cgroup's id of @from.
4271 * Returns 0 on success, -EINVAL on failure.
4273 * The caller must have charged to @to, IOW, called res_counter_charge() about
4274 * both res and memsw, and called css_get().
4276 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4277 struct mem_cgroup *from, struct mem_cgroup *to)
4279 unsigned short old_id, new_id;
4281 old_id = mem_cgroup_id(from);
4282 new_id = mem_cgroup_id(to);
4284 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4285 mem_cgroup_swap_statistics(from, false);
4286 mem_cgroup_swap_statistics(to, true);
4288 * This function is only called from task migration context now.
4289 * It postpones res_counter and refcount handling till the end
4290 * of task migration(mem_cgroup_clear_mc()) for performance
4291 * improvement. But we cannot postpone css_get(to) because if
4292 * the process that has been moved to @to does swap-in, the
4293 * refcount of @to might be decreased to 0.
4295 * We are in attach() phase, so the cgroup is guaranteed to be
4296 * alive, so we can just call css_get().
4304 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4305 struct mem_cgroup *from, struct mem_cgroup *to)
4312 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4315 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4316 struct mem_cgroup **memcgp)
4318 struct mem_cgroup *memcg = NULL;
4319 unsigned int nr_pages = 1;
4320 struct page_cgroup *pc;
4321 enum charge_type ctype;
4325 if (mem_cgroup_disabled())
4328 if (PageTransHuge(page))
4329 nr_pages <<= compound_order(page);
4331 pc = lookup_page_cgroup(page);
4332 lock_page_cgroup(pc);
4333 if (PageCgroupUsed(pc)) {
4334 memcg = pc->mem_cgroup;
4335 css_get(&memcg->css);
4337 * At migrating an anonymous page, its mapcount goes down
4338 * to 0 and uncharge() will be called. But, even if it's fully
4339 * unmapped, migration may fail and this page has to be
4340 * charged again. We set MIGRATION flag here and delay uncharge
4341 * until end_migration() is called
4343 * Corner Case Thinking
4345 * When the old page was mapped as Anon and it's unmap-and-freed
4346 * while migration was ongoing.
4347 * If unmap finds the old page, uncharge() of it will be delayed
4348 * until end_migration(). If unmap finds a new page, it's
4349 * uncharged when it make mapcount to be 1->0. If unmap code
4350 * finds swap_migration_entry, the new page will not be mapped
4351 * and end_migration() will find it(mapcount==0).
4354 * When the old page was mapped but migraion fails, the kernel
4355 * remaps it. A charge for it is kept by MIGRATION flag even
4356 * if mapcount goes down to 0. We can do remap successfully
4357 * without charging it again.
4360 * The "old" page is under lock_page() until the end of
4361 * migration, so, the old page itself will not be swapped-out.
4362 * If the new page is swapped out before end_migraton, our
4363 * hook to usual swap-out path will catch the event.
4366 SetPageCgroupMigration(pc);
4368 unlock_page_cgroup(pc);
4370 * If the page is not charged at this point,
4378 * We charge new page before it's used/mapped. So, even if unlock_page()
4379 * is called before end_migration, we can catch all events on this new
4380 * page. In the case new page is migrated but not remapped, new page's
4381 * mapcount will be finally 0 and we call uncharge in end_migration().
4384 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4386 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4388 * The page is committed to the memcg, but it's not actually
4389 * charged to the res_counter since we plan on replacing the
4390 * old one and only one page is going to be left afterwards.
4392 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4395 /* remove redundant charge if migration failed*/
4396 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4397 struct page *oldpage, struct page *newpage, bool migration_ok)
4399 struct page *used, *unused;
4400 struct page_cgroup *pc;
4406 if (!migration_ok) {
4413 anon = PageAnon(used);
4414 __mem_cgroup_uncharge_common(unused,
4415 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4416 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4418 css_put(&memcg->css);
4420 * We disallowed uncharge of pages under migration because mapcount
4421 * of the page goes down to zero, temporarly.
4422 * Clear the flag and check the page should be charged.
4424 pc = lookup_page_cgroup(oldpage);
4425 lock_page_cgroup(pc);
4426 ClearPageCgroupMigration(pc);
4427 unlock_page_cgroup(pc);
4430 * If a page is a file cache, radix-tree replacement is very atomic
4431 * and we can skip this check. When it was an Anon page, its mapcount
4432 * goes down to 0. But because we added MIGRATION flage, it's not
4433 * uncharged yet. There are several case but page->mapcount check
4434 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4435 * check. (see prepare_charge() also)
4438 mem_cgroup_uncharge_page(used);
4442 * At replace page cache, newpage is not under any memcg but it's on
4443 * LRU. So, this function doesn't touch res_counter but handles LRU
4444 * in correct way. Both pages are locked so we cannot race with uncharge.
4446 void mem_cgroup_replace_page_cache(struct page *oldpage,
4447 struct page *newpage)
4449 struct mem_cgroup *memcg = NULL;
4450 struct page_cgroup *pc;
4451 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4453 if (mem_cgroup_disabled())
4456 pc = lookup_page_cgroup(oldpage);
4457 /* fix accounting on old pages */
4458 lock_page_cgroup(pc);
4459 if (PageCgroupUsed(pc)) {
4460 memcg = pc->mem_cgroup;
4461 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4462 ClearPageCgroupUsed(pc);
4464 unlock_page_cgroup(pc);
4467 * When called from shmem_replace_page(), in some cases the
4468 * oldpage has already been charged, and in some cases not.
4473 * Even if newpage->mapping was NULL before starting replacement,
4474 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4475 * LRU while we overwrite pc->mem_cgroup.
4477 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4480 #ifdef CONFIG_DEBUG_VM
4481 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4483 struct page_cgroup *pc;
4485 pc = lookup_page_cgroup(page);
4487 * Can be NULL while feeding pages into the page allocator for
4488 * the first time, i.e. during boot or memory hotplug;
4489 * or when mem_cgroup_disabled().
4491 if (likely(pc) && PageCgroupUsed(pc))
4496 bool mem_cgroup_bad_page_check(struct page *page)
4498 if (mem_cgroup_disabled())
4501 return lookup_page_cgroup_used(page) != NULL;
4504 void mem_cgroup_print_bad_page(struct page *page)
4506 struct page_cgroup *pc;
4508 pc = lookup_page_cgroup_used(page);
4510 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4511 pc, pc->flags, pc->mem_cgroup);
4516 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4517 unsigned long long val)
4520 u64 memswlimit, memlimit;
4522 int children = mem_cgroup_count_children(memcg);
4523 u64 curusage, oldusage;
4527 * For keeping hierarchical_reclaim simple, how long we should retry
4528 * is depends on callers. We set our retry-count to be function
4529 * of # of children which we should visit in this loop.
4531 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4533 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4536 while (retry_count) {
4537 if (signal_pending(current)) {
4542 * Rather than hide all in some function, I do this in
4543 * open coded manner. You see what this really does.
4544 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4546 mutex_lock(&set_limit_mutex);
4547 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4548 if (memswlimit < val) {
4550 mutex_unlock(&set_limit_mutex);
4554 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4558 ret = res_counter_set_limit(&memcg->res, val);
4560 if (memswlimit == val)
4561 memcg->memsw_is_minimum = true;
4563 memcg->memsw_is_minimum = false;
4565 mutex_unlock(&set_limit_mutex);
4570 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4571 MEM_CGROUP_RECLAIM_SHRINK);
4572 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4573 /* Usage is reduced ? */
4574 if (curusage >= oldusage)
4577 oldusage = curusage;
4579 if (!ret && enlarge)
4580 memcg_oom_recover(memcg);
4585 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4586 unsigned long long val)
4589 u64 memlimit, memswlimit, oldusage, curusage;
4590 int children = mem_cgroup_count_children(memcg);
4594 /* see mem_cgroup_resize_res_limit */
4595 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4596 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4597 while (retry_count) {
4598 if (signal_pending(current)) {
4603 * Rather than hide all in some function, I do this in
4604 * open coded manner. You see what this really does.
4605 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4607 mutex_lock(&set_limit_mutex);
4608 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4609 if (memlimit > val) {
4611 mutex_unlock(&set_limit_mutex);
4614 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4615 if (memswlimit < val)
4617 ret = res_counter_set_limit(&memcg->memsw, val);
4619 if (memlimit == val)
4620 memcg->memsw_is_minimum = true;
4622 memcg->memsw_is_minimum = false;
4624 mutex_unlock(&set_limit_mutex);
4629 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4630 MEM_CGROUP_RECLAIM_NOSWAP |
4631 MEM_CGROUP_RECLAIM_SHRINK);
4632 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4633 /* Usage is reduced ? */
4634 if (curusage >= oldusage)
4637 oldusage = curusage;
4639 if (!ret && enlarge)
4640 memcg_oom_recover(memcg);
4644 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4646 unsigned long *total_scanned)
4648 unsigned long nr_reclaimed = 0;
4649 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4650 unsigned long reclaimed;
4652 struct mem_cgroup_tree_per_zone *mctz;
4653 unsigned long long excess;
4654 unsigned long nr_scanned;
4659 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4661 * This loop can run a while, specially if mem_cgroup's continuously
4662 * keep exceeding their soft limit and putting the system under
4669 mz = mem_cgroup_largest_soft_limit_node(mctz);
4674 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4675 gfp_mask, &nr_scanned);
4676 nr_reclaimed += reclaimed;
4677 *total_scanned += nr_scanned;
4678 spin_lock(&mctz->lock);
4681 * If we failed to reclaim anything from this memory cgroup
4682 * it is time to move on to the next cgroup
4688 * Loop until we find yet another one.
4690 * By the time we get the soft_limit lock
4691 * again, someone might have aded the
4692 * group back on the RB tree. Iterate to
4693 * make sure we get a different mem.
4694 * mem_cgroup_largest_soft_limit_node returns
4695 * NULL if no other cgroup is present on
4699 __mem_cgroup_largest_soft_limit_node(mctz);
4701 css_put(&next_mz->memcg->css);
4702 else /* next_mz == NULL or other memcg */
4706 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4707 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4709 * One school of thought says that we should not add
4710 * back the node to the tree if reclaim returns 0.
4711 * But our reclaim could return 0, simply because due
4712 * to priority we are exposing a smaller subset of
4713 * memory to reclaim from. Consider this as a longer
4716 /* If excess == 0, no tree ops */
4717 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4718 spin_unlock(&mctz->lock);
4719 css_put(&mz->memcg->css);
4722 * Could not reclaim anything and there are no more
4723 * mem cgroups to try or we seem to be looping without
4724 * reclaiming anything.
4726 if (!nr_reclaimed &&
4728 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4730 } while (!nr_reclaimed);
4732 css_put(&next_mz->memcg->css);
4733 return nr_reclaimed;
4737 * mem_cgroup_force_empty_list - clears LRU of a group
4738 * @memcg: group to clear
4741 * @lru: lru to to clear
4743 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4744 * reclaim the pages page themselves - pages are moved to the parent (or root)
4747 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4748 int node, int zid, enum lru_list lru)
4750 struct lruvec *lruvec;
4751 unsigned long flags;
4752 struct list_head *list;
4756 zone = &NODE_DATA(node)->node_zones[zid];
4757 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4758 list = &lruvec->lists[lru];
4762 struct page_cgroup *pc;
4765 spin_lock_irqsave(&zone->lru_lock, flags);
4766 if (list_empty(list)) {
4767 spin_unlock_irqrestore(&zone->lru_lock, flags);
4770 page = list_entry(list->prev, struct page, lru);
4772 list_move(&page->lru, list);
4774 spin_unlock_irqrestore(&zone->lru_lock, flags);
4777 spin_unlock_irqrestore(&zone->lru_lock, flags);
4779 pc = lookup_page_cgroup(page);
4781 if (mem_cgroup_move_parent(page, pc, memcg)) {
4782 /* found lock contention or "pc" is obsolete. */
4787 } while (!list_empty(list));
4791 * make mem_cgroup's charge to be 0 if there is no task by moving
4792 * all the charges and pages to the parent.
4793 * This enables deleting this mem_cgroup.
4795 * Caller is responsible for holding css reference on the memcg.
4797 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4803 /* This is for making all *used* pages to be on LRU. */
4804 lru_add_drain_all();
4805 drain_all_stock_sync(memcg);
4806 mem_cgroup_start_move(memcg);
4807 for_each_node_state(node, N_MEMORY) {
4808 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4811 mem_cgroup_force_empty_list(memcg,
4816 mem_cgroup_end_move(memcg);
4817 memcg_oom_recover(memcg);
4821 * Kernel memory may not necessarily be trackable to a specific
4822 * process. So they are not migrated, and therefore we can't
4823 * expect their value to drop to 0 here.
4824 * Having res filled up with kmem only is enough.
4826 * This is a safety check because mem_cgroup_force_empty_list
4827 * could have raced with mem_cgroup_replace_page_cache callers
4828 * so the lru seemed empty but the page could have been added
4829 * right after the check. RES_USAGE should be safe as we always
4830 * charge before adding to the LRU.
4832 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4833 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4834 } while (usage > 0);
4837 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4839 lockdep_assert_held(&memcg_create_mutex);
4841 * The lock does not prevent addition or deletion to the list
4842 * of children, but it prevents a new child from being
4843 * initialized based on this parent in css_online(), so it's
4844 * enough to decide whether hierarchically inherited
4845 * attributes can still be changed or not.
4847 return memcg->use_hierarchy &&
4848 !list_empty(&memcg->css.cgroup->children);
4852 * Reclaims as many pages from the given memcg as possible and moves
4853 * the rest to the parent.
4855 * Caller is responsible for holding css reference for memcg.
4857 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4859 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4860 struct cgroup *cgrp = memcg->css.cgroup;
4862 /* returns EBUSY if there is a task or if we come here twice. */
4863 if (cgroup_has_tasks(cgrp) || !list_empty(&cgrp->children))
4866 /* we call try-to-free pages for make this cgroup empty */
4867 lru_add_drain_all();
4868 /* try to free all pages in this cgroup */
4869 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4872 if (signal_pending(current))
4875 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4879 /* maybe some writeback is necessary */
4880 congestion_wait(BLK_RW_ASYNC, HZ/10);
4885 mem_cgroup_reparent_charges(memcg);
4890 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4893 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4895 if (mem_cgroup_is_root(memcg))
4897 return mem_cgroup_force_empty(memcg);
4900 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4903 return mem_cgroup_from_css(css)->use_hierarchy;
4906 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4907 struct cftype *cft, u64 val)
4910 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4911 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4913 mutex_lock(&memcg_create_mutex);
4915 if (memcg->use_hierarchy == val)
4919 * If parent's use_hierarchy is set, we can't make any modifications
4920 * in the child subtrees. If it is unset, then the change can
4921 * occur, provided the current cgroup has no children.
4923 * For the root cgroup, parent_mem is NULL, we allow value to be
4924 * set if there are no children.
4926 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4927 (val == 1 || val == 0)) {
4928 if (list_empty(&memcg->css.cgroup->children))
4929 memcg->use_hierarchy = val;
4936 mutex_unlock(&memcg_create_mutex);
4942 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4943 enum mem_cgroup_stat_index idx)
4945 struct mem_cgroup *iter;
4948 /* Per-cpu values can be negative, use a signed accumulator */
4949 for_each_mem_cgroup_tree(iter, memcg)
4950 val += mem_cgroup_read_stat(iter, idx);
4952 if (val < 0) /* race ? */
4957 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4961 if (!mem_cgroup_is_root(memcg)) {
4963 return res_counter_read_u64(&memcg->res, RES_USAGE);
4965 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4969 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4970 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4972 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4973 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4976 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4978 return val << PAGE_SHIFT;
4981 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
4984 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4989 type = MEMFILE_TYPE(cft->private);
4990 name = MEMFILE_ATTR(cft->private);
4994 if (name == RES_USAGE)
4995 val = mem_cgroup_usage(memcg, false);
4997 val = res_counter_read_u64(&memcg->res, name);
5000 if (name == RES_USAGE)
5001 val = mem_cgroup_usage(memcg, true);
5003 val = res_counter_read_u64(&memcg->memsw, name);
5006 val = res_counter_read_u64(&memcg->kmem, name);
5015 #ifdef CONFIG_MEMCG_KMEM
5016 /* should be called with activate_kmem_mutex held */
5017 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
5018 unsigned long long limit)
5023 if (memcg_kmem_is_active(memcg))
5027 * We are going to allocate memory for data shared by all memory
5028 * cgroups so let's stop accounting here.
5030 memcg_stop_kmem_account();
5033 * For simplicity, we won't allow this to be disabled. It also can't
5034 * be changed if the cgroup has children already, or if tasks had
5037 * If tasks join before we set the limit, a person looking at
5038 * kmem.usage_in_bytes will have no way to determine when it took
5039 * place, which makes the value quite meaningless.
5041 * After it first became limited, changes in the value of the limit are
5042 * of course permitted.
5044 mutex_lock(&memcg_create_mutex);
5045 if (cgroup_has_tasks(memcg->css.cgroup) || memcg_has_children(memcg))
5047 mutex_unlock(&memcg_create_mutex);
5051 memcg_id = ida_simple_get(&kmem_limited_groups,
5052 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
5059 * Make sure we have enough space for this cgroup in each root cache's
5062 err = memcg_update_all_caches(memcg_id + 1);
5066 memcg->kmemcg_id = memcg_id;
5067 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
5068 mutex_init(&memcg->slab_caches_mutex);
5071 * We couldn't have accounted to this cgroup, because it hasn't got the
5072 * active bit set yet, so this should succeed.
5074 err = res_counter_set_limit(&memcg->kmem, limit);
5077 static_key_slow_inc(&memcg_kmem_enabled_key);
5079 * Setting the active bit after enabling static branching will
5080 * guarantee no one starts accounting before all call sites are
5083 memcg_kmem_set_active(memcg);
5085 memcg_resume_kmem_account();
5089 ida_simple_remove(&kmem_limited_groups, memcg_id);
5093 static int memcg_activate_kmem(struct mem_cgroup *memcg,
5094 unsigned long long limit)
5098 mutex_lock(&activate_kmem_mutex);
5099 ret = __memcg_activate_kmem(memcg, limit);
5100 mutex_unlock(&activate_kmem_mutex);
5104 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5105 unsigned long long val)
5109 if (!memcg_kmem_is_active(memcg))
5110 ret = memcg_activate_kmem(memcg, val);
5112 ret = res_counter_set_limit(&memcg->kmem, val);
5116 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5119 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5124 mutex_lock(&activate_kmem_mutex);
5126 * If the parent cgroup is not kmem-active now, it cannot be activated
5127 * after this point, because it has at least one child already.
5129 if (memcg_kmem_is_active(parent))
5130 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
5131 mutex_unlock(&activate_kmem_mutex);
5135 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5136 unsigned long long val)
5140 #endif /* CONFIG_MEMCG_KMEM */
5143 * The user of this function is...
5146 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
5147 char *buf, size_t nbytes, loff_t off)
5149 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5152 unsigned long long val;
5155 buf = strstrip(buf);
5156 type = MEMFILE_TYPE(of_cft(of)->private);
5157 name = MEMFILE_ATTR(of_cft(of)->private);
5161 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5165 /* This function does all necessary parse...reuse it */
5166 ret = res_counter_memparse_write_strategy(buf, &val);
5170 ret = mem_cgroup_resize_limit(memcg, val);
5171 else if (type == _MEMSWAP)
5172 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5173 else if (type == _KMEM)
5174 ret = memcg_update_kmem_limit(memcg, val);
5178 case RES_SOFT_LIMIT:
5179 ret = res_counter_memparse_write_strategy(buf, &val);
5183 * For memsw, soft limits are hard to implement in terms
5184 * of semantics, for now, we support soft limits for
5185 * control without swap
5188 ret = res_counter_set_soft_limit(&memcg->res, val);
5193 ret = -EINVAL; /* should be BUG() ? */
5196 return ret ?: nbytes;
5199 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5200 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5202 unsigned long long min_limit, min_memsw_limit, tmp;
5204 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5205 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5206 if (!memcg->use_hierarchy)
5209 while (css_parent(&memcg->css)) {
5210 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5211 if (!memcg->use_hierarchy)
5213 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5214 min_limit = min(min_limit, tmp);
5215 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5216 min_memsw_limit = min(min_memsw_limit, tmp);
5219 *mem_limit = min_limit;
5220 *memsw_limit = min_memsw_limit;
5223 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5225 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5229 type = MEMFILE_TYPE(event);
5230 name = MEMFILE_ATTR(event);
5235 res_counter_reset_max(&memcg->res);
5236 else if (type == _MEMSWAP)
5237 res_counter_reset_max(&memcg->memsw);
5238 else if (type == _KMEM)
5239 res_counter_reset_max(&memcg->kmem);
5245 res_counter_reset_failcnt(&memcg->res);
5246 else if (type == _MEMSWAP)
5247 res_counter_reset_failcnt(&memcg->memsw);
5248 else if (type == _KMEM)
5249 res_counter_reset_failcnt(&memcg->kmem);
5258 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5261 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5265 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5266 struct cftype *cft, u64 val)
5268 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5270 if (val >= (1 << NR_MOVE_TYPE))
5274 * No kind of locking is needed in here, because ->can_attach() will
5275 * check this value once in the beginning of the process, and then carry
5276 * on with stale data. This means that changes to this value will only
5277 * affect task migrations starting after the change.
5279 memcg->move_charge_at_immigrate = val;
5283 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5284 struct cftype *cft, u64 val)
5291 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5295 unsigned int lru_mask;
5298 static const struct numa_stat stats[] = {
5299 { "total", LRU_ALL },
5300 { "file", LRU_ALL_FILE },
5301 { "anon", LRU_ALL_ANON },
5302 { "unevictable", BIT(LRU_UNEVICTABLE) },
5304 const struct numa_stat *stat;
5307 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5309 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5310 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5311 seq_printf(m, "%s=%lu", stat->name, nr);
5312 for_each_node_state(nid, N_MEMORY) {
5313 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5315 seq_printf(m, " N%d=%lu", nid, nr);
5320 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5321 struct mem_cgroup *iter;
5324 for_each_mem_cgroup_tree(iter, memcg)
5325 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5326 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5327 for_each_node_state(nid, N_MEMORY) {
5329 for_each_mem_cgroup_tree(iter, memcg)
5330 nr += mem_cgroup_node_nr_lru_pages(
5331 iter, nid, stat->lru_mask);
5332 seq_printf(m, " N%d=%lu", nid, nr);
5339 #endif /* CONFIG_NUMA */
5341 static inline void mem_cgroup_lru_names_not_uptodate(void)
5343 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5346 static int memcg_stat_show(struct seq_file *m, void *v)
5348 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5349 struct mem_cgroup *mi;
5352 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5353 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5355 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5356 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5359 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5360 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5361 mem_cgroup_read_events(memcg, i));
5363 for (i = 0; i < NR_LRU_LISTS; i++)
5364 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5365 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5367 /* Hierarchical information */
5369 unsigned long long limit, memsw_limit;
5370 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5371 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5372 if (do_swap_account)
5373 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5377 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5380 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5382 for_each_mem_cgroup_tree(mi, memcg)
5383 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5384 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5387 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5388 unsigned long long val = 0;
5390 for_each_mem_cgroup_tree(mi, memcg)
5391 val += mem_cgroup_read_events(mi, i);
5392 seq_printf(m, "total_%s %llu\n",
5393 mem_cgroup_events_names[i], val);
5396 for (i = 0; i < NR_LRU_LISTS; i++) {
5397 unsigned long long val = 0;
5399 for_each_mem_cgroup_tree(mi, memcg)
5400 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5401 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5404 #ifdef CONFIG_DEBUG_VM
5407 struct mem_cgroup_per_zone *mz;
5408 struct zone_reclaim_stat *rstat;
5409 unsigned long recent_rotated[2] = {0, 0};
5410 unsigned long recent_scanned[2] = {0, 0};
5412 for_each_online_node(nid)
5413 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5414 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5415 rstat = &mz->lruvec.reclaim_stat;
5417 recent_rotated[0] += rstat->recent_rotated[0];
5418 recent_rotated[1] += rstat->recent_rotated[1];
5419 recent_scanned[0] += rstat->recent_scanned[0];
5420 recent_scanned[1] += rstat->recent_scanned[1];
5422 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5423 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5424 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5425 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5432 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5435 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5437 return mem_cgroup_swappiness(memcg);
5440 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5441 struct cftype *cft, u64 val)
5443 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5444 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5446 if (val > 100 || !parent)
5449 mutex_lock(&memcg_create_mutex);
5451 /* If under hierarchy, only empty-root can set this value */
5452 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5453 mutex_unlock(&memcg_create_mutex);
5457 memcg->swappiness = val;
5459 mutex_unlock(&memcg_create_mutex);
5464 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5466 struct mem_cgroup_threshold_ary *t;
5472 t = rcu_dereference(memcg->thresholds.primary);
5474 t = rcu_dereference(memcg->memsw_thresholds.primary);
5479 usage = mem_cgroup_usage(memcg, swap);
5482 * current_threshold points to threshold just below or equal to usage.
5483 * If it's not true, a threshold was crossed after last
5484 * call of __mem_cgroup_threshold().
5486 i = t->current_threshold;
5489 * Iterate backward over array of thresholds starting from
5490 * current_threshold and check if a threshold is crossed.
5491 * If none of thresholds below usage is crossed, we read
5492 * only one element of the array here.
5494 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5495 eventfd_signal(t->entries[i].eventfd, 1);
5497 /* i = current_threshold + 1 */
5501 * Iterate forward over array of thresholds starting from
5502 * current_threshold+1 and check if a threshold is crossed.
5503 * If none of thresholds above usage is crossed, we read
5504 * only one element of the array here.
5506 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5507 eventfd_signal(t->entries[i].eventfd, 1);
5509 /* Update current_threshold */
5510 t->current_threshold = i - 1;
5515 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5518 __mem_cgroup_threshold(memcg, false);
5519 if (do_swap_account)
5520 __mem_cgroup_threshold(memcg, true);
5522 memcg = parent_mem_cgroup(memcg);
5526 static int compare_thresholds(const void *a, const void *b)
5528 const struct mem_cgroup_threshold *_a = a;
5529 const struct mem_cgroup_threshold *_b = b;
5531 if (_a->threshold > _b->threshold)
5534 if (_a->threshold < _b->threshold)
5540 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5542 struct mem_cgroup_eventfd_list *ev;
5544 list_for_each_entry(ev, &memcg->oom_notify, list)
5545 eventfd_signal(ev->eventfd, 1);
5549 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5551 struct mem_cgroup *iter;
5553 for_each_mem_cgroup_tree(iter, memcg)
5554 mem_cgroup_oom_notify_cb(iter);
5557 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5558 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5560 struct mem_cgroup_thresholds *thresholds;
5561 struct mem_cgroup_threshold_ary *new;
5562 u64 threshold, usage;
5565 ret = res_counter_memparse_write_strategy(args, &threshold);
5569 mutex_lock(&memcg->thresholds_lock);
5572 thresholds = &memcg->thresholds;
5573 else if (type == _MEMSWAP)
5574 thresholds = &memcg->memsw_thresholds;
5578 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5580 /* Check if a threshold crossed before adding a new one */
5581 if (thresholds->primary)
5582 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5584 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5586 /* Allocate memory for new array of thresholds */
5587 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5595 /* Copy thresholds (if any) to new array */
5596 if (thresholds->primary) {
5597 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5598 sizeof(struct mem_cgroup_threshold));
5601 /* Add new threshold */
5602 new->entries[size - 1].eventfd = eventfd;
5603 new->entries[size - 1].threshold = threshold;
5605 /* Sort thresholds. Registering of new threshold isn't time-critical */
5606 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5607 compare_thresholds, NULL);
5609 /* Find current threshold */
5610 new->current_threshold = -1;
5611 for (i = 0; i < size; i++) {
5612 if (new->entries[i].threshold <= usage) {
5614 * new->current_threshold will not be used until
5615 * rcu_assign_pointer(), so it's safe to increment
5618 ++new->current_threshold;
5623 /* Free old spare buffer and save old primary buffer as spare */
5624 kfree(thresholds->spare);
5625 thresholds->spare = thresholds->primary;
5627 rcu_assign_pointer(thresholds->primary, new);
5629 /* To be sure that nobody uses thresholds */
5633 mutex_unlock(&memcg->thresholds_lock);
5638 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5639 struct eventfd_ctx *eventfd, const char *args)
5641 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5644 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5645 struct eventfd_ctx *eventfd, const char *args)
5647 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5650 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5651 struct eventfd_ctx *eventfd, enum res_type type)
5653 struct mem_cgroup_thresholds *thresholds;
5654 struct mem_cgroup_threshold_ary *new;
5658 mutex_lock(&memcg->thresholds_lock);
5660 thresholds = &memcg->thresholds;
5661 else if (type == _MEMSWAP)
5662 thresholds = &memcg->memsw_thresholds;
5666 if (!thresholds->primary)
5669 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5671 /* Check if a threshold crossed before removing */
5672 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5674 /* Calculate new number of threshold */
5676 for (i = 0; i < thresholds->primary->size; i++) {
5677 if (thresholds->primary->entries[i].eventfd != eventfd)
5681 new = thresholds->spare;
5683 /* Set thresholds array to NULL if we don't have thresholds */
5692 /* Copy thresholds and find current threshold */
5693 new->current_threshold = -1;
5694 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5695 if (thresholds->primary->entries[i].eventfd == eventfd)
5698 new->entries[j] = thresholds->primary->entries[i];
5699 if (new->entries[j].threshold <= usage) {
5701 * new->current_threshold will not be used
5702 * until rcu_assign_pointer(), so it's safe to increment
5705 ++new->current_threshold;
5711 /* Swap primary and spare array */
5712 thresholds->spare = thresholds->primary;
5713 /* If all events are unregistered, free the spare array */
5715 kfree(thresholds->spare);
5716 thresholds->spare = NULL;
5719 rcu_assign_pointer(thresholds->primary, new);
5721 /* To be sure that nobody uses thresholds */
5724 mutex_unlock(&memcg->thresholds_lock);
5727 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5728 struct eventfd_ctx *eventfd)
5730 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5733 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5734 struct eventfd_ctx *eventfd)
5736 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5739 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5740 struct eventfd_ctx *eventfd, const char *args)
5742 struct mem_cgroup_eventfd_list *event;
5744 event = kmalloc(sizeof(*event), GFP_KERNEL);
5748 spin_lock(&memcg_oom_lock);
5750 event->eventfd = eventfd;
5751 list_add(&event->list, &memcg->oom_notify);
5753 /* already in OOM ? */
5754 if (atomic_read(&memcg->under_oom))
5755 eventfd_signal(eventfd, 1);
5756 spin_unlock(&memcg_oom_lock);
5761 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5762 struct eventfd_ctx *eventfd)
5764 struct mem_cgroup_eventfd_list *ev, *tmp;
5766 spin_lock(&memcg_oom_lock);
5768 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5769 if (ev->eventfd == eventfd) {
5770 list_del(&ev->list);
5775 spin_unlock(&memcg_oom_lock);
5778 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5780 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5782 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5783 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5787 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5788 struct cftype *cft, u64 val)
5790 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5791 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5793 /* cannot set to root cgroup and only 0 and 1 are allowed */
5794 if (!parent || !((val == 0) || (val == 1)))
5797 mutex_lock(&memcg_create_mutex);
5798 /* oom-kill-disable is a flag for subhierarchy. */
5799 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5800 mutex_unlock(&memcg_create_mutex);
5803 memcg->oom_kill_disable = val;
5805 memcg_oom_recover(memcg);
5806 mutex_unlock(&memcg_create_mutex);
5810 #ifdef CONFIG_MEMCG_KMEM
5811 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5815 memcg->kmemcg_id = -1;
5816 ret = memcg_propagate_kmem(memcg);
5820 return mem_cgroup_sockets_init(memcg, ss);
5823 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5825 mem_cgroup_sockets_destroy(memcg);
5828 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5830 if (!memcg_kmem_is_active(memcg))
5834 * kmem charges can outlive the cgroup. In the case of slab
5835 * pages, for instance, a page contain objects from various
5836 * processes. As we prevent from taking a reference for every
5837 * such allocation we have to be careful when doing uncharge
5838 * (see memcg_uncharge_kmem) and here during offlining.
5840 * The idea is that that only the _last_ uncharge which sees
5841 * the dead memcg will drop the last reference. An additional
5842 * reference is taken here before the group is marked dead
5843 * which is then paired with css_put during uncharge resp. here.
5845 * Although this might sound strange as this path is called from
5846 * css_offline() when the referencemight have dropped down to 0 and
5847 * shouldn't be incremented anymore (css_tryget_online() would
5848 * fail) we do not have other options because of the kmem
5849 * allocations lifetime.
5851 css_get(&memcg->css);
5853 memcg_kmem_mark_dead(memcg);
5855 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5858 if (memcg_kmem_test_and_clear_dead(memcg))
5859 css_put(&memcg->css);
5862 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5867 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5871 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5877 * DO NOT USE IN NEW FILES.
5879 * "cgroup.event_control" implementation.
5881 * This is way over-engineered. It tries to support fully configurable
5882 * events for each user. Such level of flexibility is completely
5883 * unnecessary especially in the light of the planned unified hierarchy.
5885 * Please deprecate this and replace with something simpler if at all
5890 * Unregister event and free resources.
5892 * Gets called from workqueue.
5894 static void memcg_event_remove(struct work_struct *work)
5896 struct mem_cgroup_event *event =
5897 container_of(work, struct mem_cgroup_event, remove);
5898 struct mem_cgroup *memcg = event->memcg;
5900 remove_wait_queue(event->wqh, &event->wait);
5902 event->unregister_event(memcg, event->eventfd);
5904 /* Notify userspace the event is going away. */
5905 eventfd_signal(event->eventfd, 1);
5907 eventfd_ctx_put(event->eventfd);
5909 css_put(&memcg->css);
5913 * Gets called on POLLHUP on eventfd when user closes it.
5915 * Called with wqh->lock held and interrupts disabled.
5917 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
5918 int sync, void *key)
5920 struct mem_cgroup_event *event =
5921 container_of(wait, struct mem_cgroup_event, wait);
5922 struct mem_cgroup *memcg = event->memcg;
5923 unsigned long flags = (unsigned long)key;
5925 if (flags & POLLHUP) {
5927 * If the event has been detached at cgroup removal, we
5928 * can simply return knowing the other side will cleanup
5931 * We can't race against event freeing since the other
5932 * side will require wqh->lock via remove_wait_queue(),
5935 spin_lock(&memcg->event_list_lock);
5936 if (!list_empty(&event->list)) {
5937 list_del_init(&event->list);
5939 * We are in atomic context, but cgroup_event_remove()
5940 * may sleep, so we have to call it in workqueue.
5942 schedule_work(&event->remove);
5944 spin_unlock(&memcg->event_list_lock);
5950 static void memcg_event_ptable_queue_proc(struct file *file,
5951 wait_queue_head_t *wqh, poll_table *pt)
5953 struct mem_cgroup_event *event =
5954 container_of(pt, struct mem_cgroup_event, pt);
5957 add_wait_queue(wqh, &event->wait);
5961 * DO NOT USE IN NEW FILES.
5963 * Parse input and register new cgroup event handler.
5965 * Input must be in format '<event_fd> <control_fd> <args>'.
5966 * Interpretation of args is defined by control file implementation.
5968 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
5969 char *buf, size_t nbytes, loff_t off)
5971 struct cgroup_subsys_state *css = of_css(of);
5972 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5973 struct mem_cgroup_event *event;
5974 struct cgroup_subsys_state *cfile_css;
5975 unsigned int efd, cfd;
5982 buf = strstrip(buf);
5984 efd = simple_strtoul(buf, &endp, 10);
5989 cfd = simple_strtoul(buf, &endp, 10);
5990 if ((*endp != ' ') && (*endp != '\0'))
5994 event = kzalloc(sizeof(*event), GFP_KERNEL);
5998 event->memcg = memcg;
5999 INIT_LIST_HEAD(&event->list);
6000 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
6001 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
6002 INIT_WORK(&event->remove, memcg_event_remove);
6010 event->eventfd = eventfd_ctx_fileget(efile.file);
6011 if (IS_ERR(event->eventfd)) {
6012 ret = PTR_ERR(event->eventfd);
6019 goto out_put_eventfd;
6022 /* the process need read permission on control file */
6023 /* AV: shouldn't we check that it's been opened for read instead? */
6024 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6029 * Determine the event callbacks and set them in @event. This used
6030 * to be done via struct cftype but cgroup core no longer knows
6031 * about these events. The following is crude but the whole thing
6032 * is for compatibility anyway.
6034 * DO NOT ADD NEW FILES.
6036 name = cfile.file->f_dentry->d_name.name;
6038 if (!strcmp(name, "memory.usage_in_bytes")) {
6039 event->register_event = mem_cgroup_usage_register_event;
6040 event->unregister_event = mem_cgroup_usage_unregister_event;
6041 } else if (!strcmp(name, "memory.oom_control")) {
6042 event->register_event = mem_cgroup_oom_register_event;
6043 event->unregister_event = mem_cgroup_oom_unregister_event;
6044 } else if (!strcmp(name, "memory.pressure_level")) {
6045 event->register_event = vmpressure_register_event;
6046 event->unregister_event = vmpressure_unregister_event;
6047 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
6048 event->register_event = memsw_cgroup_usage_register_event;
6049 event->unregister_event = memsw_cgroup_usage_unregister_event;
6056 * Verify @cfile should belong to @css. Also, remaining events are
6057 * automatically removed on cgroup destruction but the removal is
6058 * asynchronous, so take an extra ref on @css.
6060 cfile_css = css_tryget_online_from_dir(cfile.file->f_dentry->d_parent,
6061 &memory_cgrp_subsys);
6063 if (IS_ERR(cfile_css))
6065 if (cfile_css != css) {
6070 ret = event->register_event(memcg, event->eventfd, buf);
6074 efile.file->f_op->poll(efile.file, &event->pt);
6076 spin_lock(&memcg->event_list_lock);
6077 list_add(&event->list, &memcg->event_list);
6078 spin_unlock(&memcg->event_list_lock);
6090 eventfd_ctx_put(event->eventfd);
6099 static struct cftype mem_cgroup_files[] = {
6101 .name = "usage_in_bytes",
6102 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6103 .read_u64 = mem_cgroup_read_u64,
6106 .name = "max_usage_in_bytes",
6107 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6108 .trigger = mem_cgroup_reset,
6109 .read_u64 = mem_cgroup_read_u64,
6112 .name = "limit_in_bytes",
6113 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6114 .write = mem_cgroup_write,
6115 .read_u64 = mem_cgroup_read_u64,
6118 .name = "soft_limit_in_bytes",
6119 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6120 .write = mem_cgroup_write,
6121 .read_u64 = mem_cgroup_read_u64,
6125 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6126 .trigger = mem_cgroup_reset,
6127 .read_u64 = mem_cgroup_read_u64,
6131 .seq_show = memcg_stat_show,
6134 .name = "force_empty",
6135 .trigger = mem_cgroup_force_empty_write,
6138 .name = "use_hierarchy",
6139 .flags = CFTYPE_INSANE,
6140 .write_u64 = mem_cgroup_hierarchy_write,
6141 .read_u64 = mem_cgroup_hierarchy_read,
6144 .name = "cgroup.event_control", /* XXX: for compat */
6145 .write = memcg_write_event_control,
6146 .flags = CFTYPE_NO_PREFIX,
6150 .name = "swappiness",
6151 .read_u64 = mem_cgroup_swappiness_read,
6152 .write_u64 = mem_cgroup_swappiness_write,
6155 .name = "move_charge_at_immigrate",
6156 .read_u64 = mem_cgroup_move_charge_read,
6157 .write_u64 = mem_cgroup_move_charge_write,
6160 .name = "oom_control",
6161 .seq_show = mem_cgroup_oom_control_read,
6162 .write_u64 = mem_cgroup_oom_control_write,
6163 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6166 .name = "pressure_level",
6170 .name = "numa_stat",
6171 .seq_show = memcg_numa_stat_show,
6174 #ifdef CONFIG_MEMCG_KMEM
6176 .name = "kmem.limit_in_bytes",
6177 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6178 .write = mem_cgroup_write,
6179 .read_u64 = mem_cgroup_read_u64,
6182 .name = "kmem.usage_in_bytes",
6183 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6184 .read_u64 = mem_cgroup_read_u64,
6187 .name = "kmem.failcnt",
6188 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6189 .trigger = mem_cgroup_reset,
6190 .read_u64 = mem_cgroup_read_u64,
6193 .name = "kmem.max_usage_in_bytes",
6194 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6195 .trigger = mem_cgroup_reset,
6196 .read_u64 = mem_cgroup_read_u64,
6198 #ifdef CONFIG_SLABINFO
6200 .name = "kmem.slabinfo",
6201 .seq_show = mem_cgroup_slabinfo_read,
6205 { }, /* terminate */
6208 #ifdef CONFIG_MEMCG_SWAP
6209 static struct cftype memsw_cgroup_files[] = {
6211 .name = "memsw.usage_in_bytes",
6212 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6213 .read_u64 = mem_cgroup_read_u64,
6216 .name = "memsw.max_usage_in_bytes",
6217 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6218 .trigger = mem_cgroup_reset,
6219 .read_u64 = mem_cgroup_read_u64,
6222 .name = "memsw.limit_in_bytes",
6223 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6224 .write = mem_cgroup_write,
6225 .read_u64 = mem_cgroup_read_u64,
6228 .name = "memsw.failcnt",
6229 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6230 .trigger = mem_cgroup_reset,
6231 .read_u64 = mem_cgroup_read_u64,
6233 { }, /* terminate */
6236 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6238 struct mem_cgroup_per_node *pn;
6239 struct mem_cgroup_per_zone *mz;
6240 int zone, tmp = node;
6242 * This routine is called against possible nodes.
6243 * But it's BUG to call kmalloc() against offline node.
6245 * TODO: this routine can waste much memory for nodes which will
6246 * never be onlined. It's better to use memory hotplug callback
6249 if (!node_state(node, N_NORMAL_MEMORY))
6251 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6255 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6256 mz = &pn->zoneinfo[zone];
6257 lruvec_init(&mz->lruvec);
6258 mz->usage_in_excess = 0;
6259 mz->on_tree = false;
6262 memcg->nodeinfo[node] = pn;
6266 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6268 kfree(memcg->nodeinfo[node]);
6271 static struct mem_cgroup *mem_cgroup_alloc(void)
6273 struct mem_cgroup *memcg;
6276 size = sizeof(struct mem_cgroup);
6277 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6279 memcg = kzalloc(size, GFP_KERNEL);
6283 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6286 spin_lock_init(&memcg->pcp_counter_lock);
6295 * At destroying mem_cgroup, references from swap_cgroup can remain.
6296 * (scanning all at force_empty is too costly...)
6298 * Instead of clearing all references at force_empty, we remember
6299 * the number of reference from swap_cgroup and free mem_cgroup when
6300 * it goes down to 0.
6302 * Removal of cgroup itself succeeds regardless of refs from swap.
6305 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6309 mem_cgroup_remove_from_trees(memcg);
6312 free_mem_cgroup_per_zone_info(memcg, node);
6314 free_percpu(memcg->stat);
6317 * We need to make sure that (at least for now), the jump label
6318 * destruction code runs outside of the cgroup lock. This is because
6319 * get_online_cpus(), which is called from the static_branch update,
6320 * can't be called inside the cgroup_lock. cpusets are the ones
6321 * enforcing this dependency, so if they ever change, we might as well.
6323 * schedule_work() will guarantee this happens. Be careful if you need
6324 * to move this code around, and make sure it is outside
6327 disarm_static_keys(memcg);
6332 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6334 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6336 if (!memcg->res.parent)
6338 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6340 EXPORT_SYMBOL(parent_mem_cgroup);
6342 static void __init mem_cgroup_soft_limit_tree_init(void)
6344 struct mem_cgroup_tree_per_node *rtpn;
6345 struct mem_cgroup_tree_per_zone *rtpz;
6346 int tmp, node, zone;
6348 for_each_node(node) {
6350 if (!node_state(node, N_NORMAL_MEMORY))
6352 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6355 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6357 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6358 rtpz = &rtpn->rb_tree_per_zone[zone];
6359 rtpz->rb_root = RB_ROOT;
6360 spin_lock_init(&rtpz->lock);
6365 static struct cgroup_subsys_state * __ref
6366 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6368 struct mem_cgroup *memcg;
6369 long error = -ENOMEM;
6372 memcg = mem_cgroup_alloc();
6374 return ERR_PTR(error);
6377 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6381 if (parent_css == NULL) {
6382 root_mem_cgroup = memcg;
6383 res_counter_init(&memcg->res, NULL);
6384 res_counter_init(&memcg->memsw, NULL);
6385 res_counter_init(&memcg->kmem, NULL);
6388 memcg->last_scanned_node = MAX_NUMNODES;
6389 INIT_LIST_HEAD(&memcg->oom_notify);
6390 memcg->move_charge_at_immigrate = 0;
6391 mutex_init(&memcg->thresholds_lock);
6392 spin_lock_init(&memcg->move_lock);
6393 vmpressure_init(&memcg->vmpressure);
6394 INIT_LIST_HEAD(&memcg->event_list);
6395 spin_lock_init(&memcg->event_list_lock);
6400 __mem_cgroup_free(memcg);
6401 return ERR_PTR(error);
6405 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6407 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6408 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6410 if (css->id > MEM_CGROUP_ID_MAX)
6416 mutex_lock(&memcg_create_mutex);
6418 memcg->use_hierarchy = parent->use_hierarchy;
6419 memcg->oom_kill_disable = parent->oom_kill_disable;
6420 memcg->swappiness = mem_cgroup_swappiness(parent);
6422 if (parent->use_hierarchy) {
6423 res_counter_init(&memcg->res, &parent->res);
6424 res_counter_init(&memcg->memsw, &parent->memsw);
6425 res_counter_init(&memcg->kmem, &parent->kmem);
6428 * No need to take a reference to the parent because cgroup
6429 * core guarantees its existence.
6432 res_counter_init(&memcg->res, NULL);
6433 res_counter_init(&memcg->memsw, NULL);
6434 res_counter_init(&memcg->kmem, NULL);
6436 * Deeper hierachy with use_hierarchy == false doesn't make
6437 * much sense so let cgroup subsystem know about this
6438 * unfortunate state in our controller.
6440 if (parent != root_mem_cgroup)
6441 memory_cgrp_subsys.broken_hierarchy = true;
6443 mutex_unlock(&memcg_create_mutex);
6445 return memcg_init_kmem(memcg, &memory_cgrp_subsys);
6449 * Announce all parents that a group from their hierarchy is gone.
6451 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6453 struct mem_cgroup *parent = memcg;
6455 while ((parent = parent_mem_cgroup(parent)))
6456 mem_cgroup_iter_invalidate(parent);
6459 * if the root memcg is not hierarchical we have to check it
6462 if (!root_mem_cgroup->use_hierarchy)
6463 mem_cgroup_iter_invalidate(root_mem_cgroup);
6466 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6468 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6469 struct mem_cgroup_event *event, *tmp;
6470 struct cgroup_subsys_state *iter;
6473 * Unregister events and notify userspace.
6474 * Notify userspace about cgroup removing only after rmdir of cgroup
6475 * directory to avoid race between userspace and kernelspace.
6477 spin_lock(&memcg->event_list_lock);
6478 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6479 list_del_init(&event->list);
6480 schedule_work(&event->remove);
6482 spin_unlock(&memcg->event_list_lock);
6484 kmem_cgroup_css_offline(memcg);
6486 mem_cgroup_invalidate_reclaim_iterators(memcg);
6489 * This requires that offlining is serialized. Right now that is
6490 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6492 css_for_each_descendant_post(iter, css)
6493 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
6495 mem_cgroup_destroy_all_caches(memcg);
6496 vmpressure_cleanup(&memcg->vmpressure);
6499 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6501 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6503 * XXX: css_offline() would be where we should reparent all
6504 * memory to prepare the cgroup for destruction. However,
6505 * memcg does not do css_tryget_online() and res_counter charging
6506 * under the same RCU lock region, which means that charging
6507 * could race with offlining. Offlining only happens to
6508 * cgroups with no tasks in them but charges can show up
6509 * without any tasks from the swapin path when the target
6510 * memcg is looked up from the swapout record and not from the
6511 * current task as it usually is. A race like this can leak
6512 * charges and put pages with stale cgroup pointers into
6516 * lookup_swap_cgroup_id()
6518 * mem_cgroup_lookup()
6519 * css_tryget_online()
6521 * disable css_tryget_online()
6524 * reparent_charges()
6525 * res_counter_charge()
6528 * pc->mem_cgroup = dead memcg
6531 * The bulk of the charges are still moved in offline_css() to
6532 * avoid pinning a lot of pages in case a long-term reference
6533 * like a swapout record is deferring the css_free() to long
6534 * after offlining. But this makes sure we catch any charges
6535 * made after offlining:
6537 mem_cgroup_reparent_charges(memcg);
6539 memcg_destroy_kmem(memcg);
6540 __mem_cgroup_free(memcg);
6544 /* Handlers for move charge at task migration. */
6545 #define PRECHARGE_COUNT_AT_ONCE 256
6546 static int mem_cgroup_do_precharge(unsigned long count)
6549 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6550 struct mem_cgroup *memcg = mc.to;
6552 if (mem_cgroup_is_root(memcg)) {
6553 mc.precharge += count;
6554 /* we don't need css_get for root */
6557 /* try to charge at once */
6559 struct res_counter *dummy;
6561 * "memcg" cannot be under rmdir() because we've already checked
6562 * by cgroup_lock_live_cgroup() that it is not removed and we
6563 * are still under the same cgroup_mutex. So we can postpone
6566 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6568 if (do_swap_account && res_counter_charge(&memcg->memsw,
6569 PAGE_SIZE * count, &dummy)) {
6570 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6573 mc.precharge += count;
6577 /* fall back to one by one charge */
6579 if (signal_pending(current)) {
6583 if (!batch_count--) {
6584 batch_count = PRECHARGE_COUNT_AT_ONCE;
6587 ret = mem_cgroup_try_charge(memcg, GFP_KERNEL, 1, false);
6589 /* mem_cgroup_clear_mc() will do uncharge later */
6597 * get_mctgt_type - get target type of moving charge
6598 * @vma: the vma the pte to be checked belongs
6599 * @addr: the address corresponding to the pte to be checked
6600 * @ptent: the pte to be checked
6601 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6604 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6605 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6606 * move charge. if @target is not NULL, the page is stored in target->page
6607 * with extra refcnt got(Callers should handle it).
6608 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6609 * target for charge migration. if @target is not NULL, the entry is stored
6612 * Called with pte lock held.
6619 enum mc_target_type {
6625 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6626 unsigned long addr, pte_t ptent)
6628 struct page *page = vm_normal_page(vma, addr, ptent);
6630 if (!page || !page_mapped(page))
6632 if (PageAnon(page)) {
6633 /* we don't move shared anon */
6636 } else if (!move_file())
6637 /* we ignore mapcount for file pages */
6639 if (!get_page_unless_zero(page))
6646 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6647 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6649 struct page *page = NULL;
6650 swp_entry_t ent = pte_to_swp_entry(ptent);
6652 if (!move_anon() || non_swap_entry(ent))
6655 * Because lookup_swap_cache() updates some statistics counter,
6656 * we call find_get_page() with swapper_space directly.
6658 page = find_get_page(swap_address_space(ent), ent.val);
6659 if (do_swap_account)
6660 entry->val = ent.val;
6665 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6666 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6672 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6673 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6675 struct page *page = NULL;
6676 struct address_space *mapping;
6679 if (!vma->vm_file) /* anonymous vma */
6684 mapping = vma->vm_file->f_mapping;
6685 if (pte_none(ptent))
6686 pgoff = linear_page_index(vma, addr);
6687 else /* pte_file(ptent) is true */
6688 pgoff = pte_to_pgoff(ptent);
6690 /* page is moved even if it's not RSS of this task(page-faulted). */
6691 page = find_get_page(mapping, pgoff);
6694 /* shmem/tmpfs may report page out on swap: account for that too. */
6695 if (radix_tree_exceptional_entry(page)) {
6696 swp_entry_t swap = radix_to_swp_entry(page);
6697 if (do_swap_account)
6699 page = find_get_page(swap_address_space(swap), swap.val);
6705 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6706 unsigned long addr, pte_t ptent, union mc_target *target)
6708 struct page *page = NULL;
6709 struct page_cgroup *pc;
6710 enum mc_target_type ret = MC_TARGET_NONE;
6711 swp_entry_t ent = { .val = 0 };
6713 if (pte_present(ptent))
6714 page = mc_handle_present_pte(vma, addr, ptent);
6715 else if (is_swap_pte(ptent))
6716 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6717 else if (pte_none(ptent) || pte_file(ptent))
6718 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6720 if (!page && !ent.val)
6723 pc = lookup_page_cgroup(page);
6725 * Do only loose check w/o page_cgroup lock.
6726 * mem_cgroup_move_account() checks the pc is valid or not under
6729 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6730 ret = MC_TARGET_PAGE;
6732 target->page = page;
6734 if (!ret || !target)
6737 /* There is a swap entry and a page doesn't exist or isn't charged */
6738 if (ent.val && !ret &&
6739 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6740 ret = MC_TARGET_SWAP;
6747 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6749 * We don't consider swapping or file mapped pages because THP does not
6750 * support them for now.
6751 * Caller should make sure that pmd_trans_huge(pmd) is true.
6753 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6754 unsigned long addr, pmd_t pmd, union mc_target *target)
6756 struct page *page = NULL;
6757 struct page_cgroup *pc;
6758 enum mc_target_type ret = MC_TARGET_NONE;
6760 page = pmd_page(pmd);
6761 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6764 pc = lookup_page_cgroup(page);
6765 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6766 ret = MC_TARGET_PAGE;
6769 target->page = page;
6775 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6776 unsigned long addr, pmd_t pmd, union mc_target *target)
6778 return MC_TARGET_NONE;
6782 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6783 unsigned long addr, unsigned long end,
6784 struct mm_walk *walk)
6786 struct vm_area_struct *vma = walk->private;
6790 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6791 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6792 mc.precharge += HPAGE_PMD_NR;
6797 if (pmd_trans_unstable(pmd))
6799 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6800 for (; addr != end; pte++, addr += PAGE_SIZE)
6801 if (get_mctgt_type(vma, addr, *pte, NULL))
6802 mc.precharge++; /* increment precharge temporarily */
6803 pte_unmap_unlock(pte - 1, ptl);
6809 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6811 unsigned long precharge;
6812 struct vm_area_struct *vma;
6814 down_read(&mm->mmap_sem);
6815 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6816 struct mm_walk mem_cgroup_count_precharge_walk = {
6817 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6821 if (is_vm_hugetlb_page(vma))
6823 walk_page_range(vma->vm_start, vma->vm_end,
6824 &mem_cgroup_count_precharge_walk);
6826 up_read(&mm->mmap_sem);
6828 precharge = mc.precharge;
6834 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6836 unsigned long precharge = mem_cgroup_count_precharge(mm);
6838 VM_BUG_ON(mc.moving_task);
6839 mc.moving_task = current;
6840 return mem_cgroup_do_precharge(precharge);
6843 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6844 static void __mem_cgroup_clear_mc(void)
6846 struct mem_cgroup *from = mc.from;
6847 struct mem_cgroup *to = mc.to;
6850 /* we must uncharge all the leftover precharges from mc.to */
6852 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6856 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6857 * we must uncharge here.
6859 if (mc.moved_charge) {
6860 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6861 mc.moved_charge = 0;
6863 /* we must fixup refcnts and charges */
6864 if (mc.moved_swap) {
6865 /* uncharge swap account from the old cgroup */
6866 if (!mem_cgroup_is_root(mc.from))
6867 res_counter_uncharge(&mc.from->memsw,
6868 PAGE_SIZE * mc.moved_swap);
6870 for (i = 0; i < mc.moved_swap; i++)
6871 css_put(&mc.from->css);
6873 if (!mem_cgroup_is_root(mc.to)) {
6875 * we charged both to->res and to->memsw, so we should
6878 res_counter_uncharge(&mc.to->res,
6879 PAGE_SIZE * mc.moved_swap);
6881 /* we've already done css_get(mc.to) */
6884 memcg_oom_recover(from);
6885 memcg_oom_recover(to);
6886 wake_up_all(&mc.waitq);
6889 static void mem_cgroup_clear_mc(void)
6891 struct mem_cgroup *from = mc.from;
6894 * we must clear moving_task before waking up waiters at the end of
6897 mc.moving_task = NULL;
6898 __mem_cgroup_clear_mc();
6899 spin_lock(&mc.lock);
6902 spin_unlock(&mc.lock);
6903 mem_cgroup_end_move(from);
6906 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6907 struct cgroup_taskset *tset)
6909 struct task_struct *p = cgroup_taskset_first(tset);
6911 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6912 unsigned long move_charge_at_immigrate;
6915 * We are now commited to this value whatever it is. Changes in this
6916 * tunable will only affect upcoming migrations, not the current one.
6917 * So we need to save it, and keep it going.
6919 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6920 if (move_charge_at_immigrate) {
6921 struct mm_struct *mm;
6922 struct mem_cgroup *from = mem_cgroup_from_task(p);
6924 VM_BUG_ON(from == memcg);
6926 mm = get_task_mm(p);
6929 /* We move charges only when we move a owner of the mm */
6930 if (mm->owner == p) {
6933 VM_BUG_ON(mc.precharge);
6934 VM_BUG_ON(mc.moved_charge);
6935 VM_BUG_ON(mc.moved_swap);
6936 mem_cgroup_start_move(from);
6937 spin_lock(&mc.lock);
6940 mc.immigrate_flags = move_charge_at_immigrate;
6941 spin_unlock(&mc.lock);
6942 /* We set mc.moving_task later */
6944 ret = mem_cgroup_precharge_mc(mm);
6946 mem_cgroup_clear_mc();
6953 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6954 struct cgroup_taskset *tset)
6956 mem_cgroup_clear_mc();
6959 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6960 unsigned long addr, unsigned long end,
6961 struct mm_walk *walk)
6964 struct vm_area_struct *vma = walk->private;
6967 enum mc_target_type target_type;
6968 union mc_target target;
6970 struct page_cgroup *pc;
6973 * We don't take compound_lock() here but no race with splitting thp
6975 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6976 * under splitting, which means there's no concurrent thp split,
6977 * - if another thread runs into split_huge_page() just after we
6978 * entered this if-block, the thread must wait for page table lock
6979 * to be unlocked in __split_huge_page_splitting(), where the main
6980 * part of thp split is not executed yet.
6982 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6983 if (mc.precharge < HPAGE_PMD_NR) {
6987 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6988 if (target_type == MC_TARGET_PAGE) {
6990 if (!isolate_lru_page(page)) {
6991 pc = lookup_page_cgroup(page);
6992 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6993 pc, mc.from, mc.to)) {
6994 mc.precharge -= HPAGE_PMD_NR;
6995 mc.moved_charge += HPAGE_PMD_NR;
6997 putback_lru_page(page);
7005 if (pmd_trans_unstable(pmd))
7008 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7009 for (; addr != end; addr += PAGE_SIZE) {
7010 pte_t ptent = *(pte++);
7016 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7017 case MC_TARGET_PAGE:
7019 if (isolate_lru_page(page))
7021 pc = lookup_page_cgroup(page);
7022 if (!mem_cgroup_move_account(page, 1, pc,
7025 /* we uncharge from mc.from later. */
7028 putback_lru_page(page);
7029 put: /* get_mctgt_type() gets the page */
7032 case MC_TARGET_SWAP:
7034 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7036 /* we fixup refcnts and charges later. */
7044 pte_unmap_unlock(pte - 1, ptl);
7049 * We have consumed all precharges we got in can_attach().
7050 * We try charge one by one, but don't do any additional
7051 * charges to mc.to if we have failed in charge once in attach()
7054 ret = mem_cgroup_do_precharge(1);
7062 static void mem_cgroup_move_charge(struct mm_struct *mm)
7064 struct vm_area_struct *vma;
7066 lru_add_drain_all();
7068 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7070 * Someone who are holding the mmap_sem might be waiting in
7071 * waitq. So we cancel all extra charges, wake up all waiters,
7072 * and retry. Because we cancel precharges, we might not be able
7073 * to move enough charges, but moving charge is a best-effort
7074 * feature anyway, so it wouldn't be a big problem.
7076 __mem_cgroup_clear_mc();
7080 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7082 struct mm_walk mem_cgroup_move_charge_walk = {
7083 .pmd_entry = mem_cgroup_move_charge_pte_range,
7087 if (is_vm_hugetlb_page(vma))
7089 ret = walk_page_range(vma->vm_start, vma->vm_end,
7090 &mem_cgroup_move_charge_walk);
7093 * means we have consumed all precharges and failed in
7094 * doing additional charge. Just abandon here.
7098 up_read(&mm->mmap_sem);
7101 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7102 struct cgroup_taskset *tset)
7104 struct task_struct *p = cgroup_taskset_first(tset);
7105 struct mm_struct *mm = get_task_mm(p);
7109 mem_cgroup_move_charge(mm);
7113 mem_cgroup_clear_mc();
7115 #else /* !CONFIG_MMU */
7116 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7117 struct cgroup_taskset *tset)
7121 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7122 struct cgroup_taskset *tset)
7125 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7126 struct cgroup_taskset *tset)
7132 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7133 * to verify sane_behavior flag on each mount attempt.
7135 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7138 * use_hierarchy is forced with sane_behavior. cgroup core
7139 * guarantees that @root doesn't have any children, so turning it
7140 * on for the root memcg is enough.
7142 if (cgroup_sane_behavior(root_css->cgroup))
7143 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7146 struct cgroup_subsys memory_cgrp_subsys = {
7147 .css_alloc = mem_cgroup_css_alloc,
7148 .css_online = mem_cgroup_css_online,
7149 .css_offline = mem_cgroup_css_offline,
7150 .css_free = mem_cgroup_css_free,
7151 .can_attach = mem_cgroup_can_attach,
7152 .cancel_attach = mem_cgroup_cancel_attach,
7153 .attach = mem_cgroup_move_task,
7154 .bind = mem_cgroup_bind,
7155 .base_cftypes = mem_cgroup_files,
7159 #ifdef CONFIG_MEMCG_SWAP
7160 static int __init enable_swap_account(char *s)
7162 if (!strcmp(s, "1"))
7163 really_do_swap_account = 1;
7164 else if (!strcmp(s, "0"))
7165 really_do_swap_account = 0;
7168 __setup("swapaccount=", enable_swap_account);
7170 static void __init memsw_file_init(void)
7172 WARN_ON(cgroup_add_cftypes(&memory_cgrp_subsys, memsw_cgroup_files));
7175 static void __init enable_swap_cgroup(void)
7177 if (!mem_cgroup_disabled() && really_do_swap_account) {
7178 do_swap_account = 1;
7184 static void __init enable_swap_cgroup(void)
7190 * subsys_initcall() for memory controller.
7192 * Some parts like hotcpu_notifier() have to be initialized from this context
7193 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7194 * everything that doesn't depend on a specific mem_cgroup structure should
7195 * be initialized from here.
7197 static int __init mem_cgroup_init(void)
7199 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7200 enable_swap_cgroup();
7201 mem_cgroup_soft_limit_tree_init();
7205 subsys_initcall(mem_cgroup_init);