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 mem_cgroup_subsys __read_mostly;
70 EXPORT_SYMBOL(mem_cgroup_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;
152 unsigned long last_dead_count;
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 = 0, /* accounted by this cgroup itself */
386 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
387 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
390 /* We account when limit is on, but only after call sites are patched */
391 #define KMEM_ACCOUNTED_MASK \
392 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
394 #ifdef CONFIG_MEMCG_KMEM
395 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
397 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
400 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
402 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
405 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
407 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
410 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
412 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
415 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
418 * Our caller must use css_get() first, because memcg_uncharge_kmem()
419 * will call css_put() if it sees the memcg is dead.
422 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
423 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
426 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
428 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
429 &memcg->kmem_account_flags);
433 /* Stuffs for move charges at task migration. */
435 * Types of charges to be moved. "move_charge_at_immitgrate" and
436 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
439 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
440 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
444 /* "mc" and its members are protected by cgroup_mutex */
445 static struct move_charge_struct {
446 spinlock_t lock; /* for from, to */
447 struct mem_cgroup *from;
448 struct mem_cgroup *to;
449 unsigned long immigrate_flags;
450 unsigned long precharge;
451 unsigned long moved_charge;
452 unsigned long moved_swap;
453 struct task_struct *moving_task; /* a task moving charges */
454 wait_queue_head_t waitq; /* a waitq for other context */
456 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
457 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
460 static bool move_anon(void)
462 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
465 static bool move_file(void)
467 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
471 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
472 * limit reclaim to prevent infinite loops, if they ever occur.
474 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
475 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
478 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
479 MEM_CGROUP_CHARGE_TYPE_ANON,
480 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
481 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
485 /* for encoding cft->private value on file */
493 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
494 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
495 #define MEMFILE_ATTR(val) ((val) & 0xffff)
496 /* Used for OOM nofiier */
497 #define OOM_CONTROL (0)
500 * Reclaim flags for mem_cgroup_hierarchical_reclaim
502 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
503 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
504 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
505 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
508 * The memcg_create_mutex will be held whenever a new cgroup is created.
509 * As a consequence, any change that needs to protect against new child cgroups
510 * appearing has to hold it as well.
512 static DEFINE_MUTEX(memcg_create_mutex);
514 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
516 return s ? container_of(s, struct mem_cgroup, css) : NULL;
519 /* Some nice accessors for the vmpressure. */
520 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
523 memcg = root_mem_cgroup;
524 return &memcg->vmpressure;
527 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
529 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
532 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
534 return (memcg == root_mem_cgroup);
538 * We restrict the id in the range of [1, 65535], so it can fit into
541 #define MEM_CGROUP_ID_MAX USHRT_MAX
543 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
546 * The ID of the root cgroup is 0, but memcg treat 0 as an
547 * invalid ID, so we return (cgroup_id + 1).
549 return memcg->css.cgroup->id + 1;
552 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
554 struct cgroup_subsys_state *css;
556 css = css_from_id(id - 1, &mem_cgroup_subsys);
557 return mem_cgroup_from_css(css);
560 /* Writing them here to avoid exposing memcg's inner layout */
561 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
563 void sock_update_memcg(struct sock *sk)
565 if (mem_cgroup_sockets_enabled) {
566 struct mem_cgroup *memcg;
567 struct cg_proto *cg_proto;
569 BUG_ON(!sk->sk_prot->proto_cgroup);
571 /* Socket cloning can throw us here with sk_cgrp already
572 * filled. It won't however, necessarily happen from
573 * process context. So the test for root memcg given
574 * the current task's memcg won't help us in this case.
576 * Respecting the original socket's memcg is a better
577 * decision in this case.
580 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
581 css_get(&sk->sk_cgrp->memcg->css);
586 memcg = mem_cgroup_from_task(current);
587 cg_proto = sk->sk_prot->proto_cgroup(memcg);
588 if (!mem_cgroup_is_root(memcg) &&
589 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
590 sk->sk_cgrp = cg_proto;
595 EXPORT_SYMBOL(sock_update_memcg);
597 void sock_release_memcg(struct sock *sk)
599 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
600 struct mem_cgroup *memcg;
601 WARN_ON(!sk->sk_cgrp->memcg);
602 memcg = sk->sk_cgrp->memcg;
603 css_put(&sk->sk_cgrp->memcg->css);
607 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
609 if (!memcg || mem_cgroup_is_root(memcg))
612 return &memcg->tcp_mem;
614 EXPORT_SYMBOL(tcp_proto_cgroup);
616 static void disarm_sock_keys(struct mem_cgroup *memcg)
618 if (!memcg_proto_activated(&memcg->tcp_mem))
620 static_key_slow_dec(&memcg_socket_limit_enabled);
623 static void disarm_sock_keys(struct mem_cgroup *memcg)
628 #ifdef CONFIG_MEMCG_KMEM
630 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
631 * The main reason for not using cgroup id for this:
632 * this works better in sparse environments, where we have a lot of memcgs,
633 * but only a few kmem-limited. Or also, if we have, for instance, 200
634 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
635 * 200 entry array for that.
637 * The current size of the caches array is stored in
638 * memcg_limited_groups_array_size. It will double each time we have to
641 static DEFINE_IDA(kmem_limited_groups);
642 int memcg_limited_groups_array_size;
645 * MIN_SIZE is different than 1, because we would like to avoid going through
646 * the alloc/free process all the time. In a small machine, 4 kmem-limited
647 * cgroups is a reasonable guess. In the future, it could be a parameter or
648 * tunable, but that is strictly not necessary.
650 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
651 * this constant directly from cgroup, but it is understandable that this is
652 * better kept as an internal representation in cgroup.c. In any case, the
653 * cgrp_id space is not getting any smaller, and we don't have to necessarily
654 * increase ours as well if it increases.
656 #define MEMCG_CACHES_MIN_SIZE 4
657 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
660 * A lot of the calls to the cache allocation functions are expected to be
661 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
662 * conditional to this static branch, we'll have to allow modules that does
663 * kmem_cache_alloc and the such to see this symbol as well
665 struct static_key memcg_kmem_enabled_key;
666 EXPORT_SYMBOL(memcg_kmem_enabled_key);
668 static void disarm_kmem_keys(struct mem_cgroup *memcg)
670 if (memcg_kmem_is_active(memcg)) {
671 static_key_slow_dec(&memcg_kmem_enabled_key);
672 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
675 * This check can't live in kmem destruction function,
676 * since the charges will outlive the cgroup
678 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
681 static void disarm_kmem_keys(struct mem_cgroup *memcg)
684 #endif /* CONFIG_MEMCG_KMEM */
686 static void disarm_static_keys(struct mem_cgroup *memcg)
688 disarm_sock_keys(memcg);
689 disarm_kmem_keys(memcg);
692 static void drain_all_stock_async(struct mem_cgroup *memcg);
694 static struct mem_cgroup_per_zone *
695 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
697 VM_BUG_ON((unsigned)nid >= nr_node_ids);
698 return &memcg->nodeinfo[nid]->zoneinfo[zid];
701 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
706 static struct mem_cgroup_per_zone *
707 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
709 int nid = page_to_nid(page);
710 int zid = page_zonenum(page);
712 return mem_cgroup_zoneinfo(memcg, nid, zid);
715 static struct mem_cgroup_tree_per_zone *
716 soft_limit_tree_node_zone(int nid, int zid)
718 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
721 static struct mem_cgroup_tree_per_zone *
722 soft_limit_tree_from_page(struct page *page)
724 int nid = page_to_nid(page);
725 int zid = page_zonenum(page);
727 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
731 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
732 struct mem_cgroup_per_zone *mz,
733 struct mem_cgroup_tree_per_zone *mctz,
734 unsigned long long new_usage_in_excess)
736 struct rb_node **p = &mctz->rb_root.rb_node;
737 struct rb_node *parent = NULL;
738 struct mem_cgroup_per_zone *mz_node;
743 mz->usage_in_excess = new_usage_in_excess;
744 if (!mz->usage_in_excess)
748 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
750 if (mz->usage_in_excess < mz_node->usage_in_excess)
753 * We can't avoid mem cgroups that are over their soft
754 * limit by the same amount
756 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
759 rb_link_node(&mz->tree_node, parent, p);
760 rb_insert_color(&mz->tree_node, &mctz->rb_root);
765 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
766 struct mem_cgroup_per_zone *mz,
767 struct mem_cgroup_tree_per_zone *mctz)
771 rb_erase(&mz->tree_node, &mctz->rb_root);
776 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
777 struct mem_cgroup_per_zone *mz,
778 struct mem_cgroup_tree_per_zone *mctz)
780 spin_lock(&mctz->lock);
781 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
782 spin_unlock(&mctz->lock);
786 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
788 unsigned long long excess;
789 struct mem_cgroup_per_zone *mz;
790 struct mem_cgroup_tree_per_zone *mctz;
791 int nid = page_to_nid(page);
792 int zid = page_zonenum(page);
793 mctz = soft_limit_tree_from_page(page);
796 * Necessary to update all ancestors when hierarchy is used.
797 * because their event counter is not touched.
799 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
800 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
801 excess = res_counter_soft_limit_excess(&memcg->res);
803 * We have to update the tree if mz is on RB-tree or
804 * mem is over its softlimit.
806 if (excess || mz->on_tree) {
807 spin_lock(&mctz->lock);
808 /* if on-tree, remove it */
810 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
812 * Insert again. mz->usage_in_excess will be updated.
813 * If excess is 0, no tree ops.
815 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
816 spin_unlock(&mctz->lock);
821 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
824 struct mem_cgroup_per_zone *mz;
825 struct mem_cgroup_tree_per_zone *mctz;
827 for_each_node(node) {
828 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
829 mz = mem_cgroup_zoneinfo(memcg, node, zone);
830 mctz = soft_limit_tree_node_zone(node, zone);
831 mem_cgroup_remove_exceeded(memcg, mz, mctz);
836 static struct mem_cgroup_per_zone *
837 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
839 struct rb_node *rightmost = NULL;
840 struct mem_cgroup_per_zone *mz;
844 rightmost = rb_last(&mctz->rb_root);
846 goto done; /* Nothing to reclaim from */
848 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
850 * Remove the node now but someone else can add it back,
851 * we will to add it back at the end of reclaim to its correct
852 * position in the tree.
854 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
855 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
856 !css_tryget(&mz->memcg->css))
862 static struct mem_cgroup_per_zone *
863 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
865 struct mem_cgroup_per_zone *mz;
867 spin_lock(&mctz->lock);
868 mz = __mem_cgroup_largest_soft_limit_node(mctz);
869 spin_unlock(&mctz->lock);
874 * Implementation Note: reading percpu statistics for memcg.
876 * Both of vmstat[] and percpu_counter has threshold and do periodic
877 * synchronization to implement "quick" read. There are trade-off between
878 * reading cost and precision of value. Then, we may have a chance to implement
879 * a periodic synchronizion of counter in memcg's counter.
881 * But this _read() function is used for user interface now. The user accounts
882 * memory usage by memory cgroup and he _always_ requires exact value because
883 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
884 * have to visit all online cpus and make sum. So, for now, unnecessary
885 * synchronization is not implemented. (just implemented for cpu hotplug)
887 * If there are kernel internal actions which can make use of some not-exact
888 * value, and reading all cpu value can be performance bottleneck in some
889 * common workload, threashold and synchonization as vmstat[] should be
892 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
893 enum mem_cgroup_stat_index idx)
899 for_each_online_cpu(cpu)
900 val += per_cpu(memcg->stat->count[idx], cpu);
901 #ifdef CONFIG_HOTPLUG_CPU
902 spin_lock(&memcg->pcp_counter_lock);
903 val += memcg->nocpu_base.count[idx];
904 spin_unlock(&memcg->pcp_counter_lock);
910 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
913 int val = (charge) ? 1 : -1;
914 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
917 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
918 enum mem_cgroup_events_index idx)
920 unsigned long val = 0;
924 for_each_online_cpu(cpu)
925 val += per_cpu(memcg->stat->events[idx], cpu);
926 #ifdef CONFIG_HOTPLUG_CPU
927 spin_lock(&memcg->pcp_counter_lock);
928 val += memcg->nocpu_base.events[idx];
929 spin_unlock(&memcg->pcp_counter_lock);
935 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
937 bool anon, int nr_pages)
942 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
943 * counted as CACHE even if it's on ANON LRU.
946 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
949 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
952 if (PageTransHuge(page))
953 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
956 /* pagein of a big page is an event. So, ignore page size */
958 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
960 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
961 nr_pages = -nr_pages; /* for event */
964 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
970 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
972 struct mem_cgroup_per_zone *mz;
974 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
975 return mz->lru_size[lru];
979 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
980 unsigned int lru_mask)
982 struct mem_cgroup_per_zone *mz;
984 unsigned long ret = 0;
986 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
989 if (BIT(lru) & lru_mask)
990 ret += mz->lru_size[lru];
996 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
997 int nid, unsigned int lru_mask)
1002 for (zid = 0; zid < MAX_NR_ZONES; zid++)
1003 total += mem_cgroup_zone_nr_lru_pages(memcg,
1004 nid, zid, lru_mask);
1009 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
1010 unsigned int lru_mask)
1015 for_each_node_state(nid, N_MEMORY)
1016 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1020 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1021 enum mem_cgroup_events_target target)
1023 unsigned long val, next;
1025 val = __this_cpu_read(memcg->stat->nr_page_events);
1026 next = __this_cpu_read(memcg->stat->targets[target]);
1027 /* from time_after() in jiffies.h */
1028 if ((long)next - (long)val < 0) {
1030 case MEM_CGROUP_TARGET_THRESH:
1031 next = val + THRESHOLDS_EVENTS_TARGET;
1033 case MEM_CGROUP_TARGET_SOFTLIMIT:
1034 next = val + SOFTLIMIT_EVENTS_TARGET;
1036 case MEM_CGROUP_TARGET_NUMAINFO:
1037 next = val + NUMAINFO_EVENTS_TARGET;
1042 __this_cpu_write(memcg->stat->targets[target], next);
1049 * Check events in order.
1052 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1055 /* threshold event is triggered in finer grain than soft limit */
1056 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1057 MEM_CGROUP_TARGET_THRESH))) {
1059 bool do_numainfo __maybe_unused;
1061 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1062 MEM_CGROUP_TARGET_SOFTLIMIT);
1063 #if MAX_NUMNODES > 1
1064 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1065 MEM_CGROUP_TARGET_NUMAINFO);
1069 mem_cgroup_threshold(memcg);
1070 if (unlikely(do_softlimit))
1071 mem_cgroup_update_tree(memcg, page);
1072 #if MAX_NUMNODES > 1
1073 if (unlikely(do_numainfo))
1074 atomic_inc(&memcg->numainfo_events);
1080 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1083 * mm_update_next_owner() may clear mm->owner to NULL
1084 * if it races with swapoff, page migration, etc.
1085 * So this can be called with p == NULL.
1090 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1093 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1095 struct mem_cgroup *memcg = NULL;
1100 * Because we have no locks, mm->owner's may be being moved to other
1101 * cgroup. We use css_tryget() here even if this looks
1102 * pessimistic (rather than adding locks here).
1106 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1107 if (unlikely(!memcg))
1109 } while (!css_tryget(&memcg->css));
1115 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1116 * ref. count) or NULL if the whole root's subtree has been visited.
1118 * helper function to be used by mem_cgroup_iter
1120 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1121 struct mem_cgroup *last_visited)
1123 struct cgroup_subsys_state *prev_css, *next_css;
1125 prev_css = last_visited ? &last_visited->css : NULL;
1127 next_css = css_next_descendant_pre(prev_css, &root->css);
1130 * Even if we found a group we have to make sure it is
1131 * alive. css && !memcg means that the groups should be
1132 * skipped and we should continue the tree walk.
1133 * last_visited css is safe to use because it is
1134 * protected by css_get and the tree walk is rcu safe.
1137 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
1139 if (css_tryget(&mem->css))
1142 prev_css = next_css;
1150 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1153 * When a group in the hierarchy below root is destroyed, the
1154 * hierarchy iterator can no longer be trusted since it might
1155 * have pointed to the destroyed group. Invalidate it.
1157 atomic_inc(&root->dead_count);
1160 static struct mem_cgroup *
1161 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1162 struct mem_cgroup *root,
1165 struct mem_cgroup *position = NULL;
1167 * A cgroup destruction happens in two stages: offlining and
1168 * release. They are separated by a RCU grace period.
1170 * If the iterator is valid, we may still race with an
1171 * offlining. The RCU lock ensures the object won't be
1172 * released, tryget will fail if we lost the race.
1174 *sequence = atomic_read(&root->dead_count);
1175 if (iter->last_dead_count == *sequence) {
1177 position = iter->last_visited;
1178 if (position && !css_tryget(&position->css))
1184 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1185 struct mem_cgroup *last_visited,
1186 struct mem_cgroup *new_position,
1190 css_put(&last_visited->css);
1192 * We store the sequence count from the time @last_visited was
1193 * loaded successfully instead of rereading it here so that we
1194 * don't lose destruction events in between. We could have
1195 * raced with the destruction of @new_position after all.
1197 iter->last_visited = new_position;
1199 iter->last_dead_count = sequence;
1203 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1204 * @root: hierarchy root
1205 * @prev: previously returned memcg, NULL on first invocation
1206 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1208 * Returns references to children of the hierarchy below @root, or
1209 * @root itself, or %NULL after a full round-trip.
1211 * Caller must pass the return value in @prev on subsequent
1212 * invocations for reference counting, or use mem_cgroup_iter_break()
1213 * to cancel a hierarchy walk before the round-trip is complete.
1215 * Reclaimers can specify a zone and a priority level in @reclaim to
1216 * divide up the memcgs in the hierarchy among all concurrent
1217 * reclaimers operating on the same zone and priority.
1219 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1220 struct mem_cgroup *prev,
1221 struct mem_cgroup_reclaim_cookie *reclaim)
1223 struct mem_cgroup *memcg = NULL;
1224 struct mem_cgroup *last_visited = NULL;
1226 if (mem_cgroup_disabled())
1230 root = root_mem_cgroup;
1232 if (prev && !reclaim)
1233 last_visited = prev;
1235 if (!root->use_hierarchy && root != root_mem_cgroup) {
1243 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1244 int uninitialized_var(seq);
1247 int nid = zone_to_nid(reclaim->zone);
1248 int zid = zone_idx(reclaim->zone);
1249 struct mem_cgroup_per_zone *mz;
1251 mz = mem_cgroup_zoneinfo(root, nid, zid);
1252 iter = &mz->reclaim_iter[reclaim->priority];
1253 if (prev && reclaim->generation != iter->generation) {
1254 iter->last_visited = NULL;
1258 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1261 memcg = __mem_cgroup_iter_next(root, last_visited);
1264 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1268 else if (!prev && memcg)
1269 reclaim->generation = iter->generation;
1278 if (prev && prev != root)
1279 css_put(&prev->css);
1285 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1286 * @root: hierarchy root
1287 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1289 void mem_cgroup_iter_break(struct mem_cgroup *root,
1290 struct mem_cgroup *prev)
1293 root = root_mem_cgroup;
1294 if (prev && prev != root)
1295 css_put(&prev->css);
1299 * Iteration constructs for visiting all cgroups (under a tree). If
1300 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1301 * be used for reference counting.
1303 #define for_each_mem_cgroup_tree(iter, root) \
1304 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1306 iter = mem_cgroup_iter(root, iter, NULL))
1308 #define for_each_mem_cgroup(iter) \
1309 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1311 iter = mem_cgroup_iter(NULL, iter, NULL))
1313 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1315 struct mem_cgroup *memcg;
1318 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1319 if (unlikely(!memcg))
1324 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1327 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1335 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1338 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1339 * @zone: zone of the wanted lruvec
1340 * @memcg: memcg of the wanted lruvec
1342 * Returns the lru list vector holding pages for the given @zone and
1343 * @mem. This can be the global zone lruvec, if the memory controller
1346 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1347 struct mem_cgroup *memcg)
1349 struct mem_cgroup_per_zone *mz;
1350 struct lruvec *lruvec;
1352 if (mem_cgroup_disabled()) {
1353 lruvec = &zone->lruvec;
1357 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1358 lruvec = &mz->lruvec;
1361 * Since a node can be onlined after the mem_cgroup was created,
1362 * we have to be prepared to initialize lruvec->zone here;
1363 * and if offlined then reonlined, we need to reinitialize it.
1365 if (unlikely(lruvec->zone != zone))
1366 lruvec->zone = zone;
1371 * Following LRU functions are allowed to be used without PCG_LOCK.
1372 * Operations are called by routine of global LRU independently from memcg.
1373 * What we have to take care of here is validness of pc->mem_cgroup.
1375 * Changes to pc->mem_cgroup happens when
1378 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1379 * It is added to LRU before charge.
1380 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1381 * When moving account, the page is not on LRU. It's isolated.
1385 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1387 * @zone: zone of the page
1389 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1391 struct mem_cgroup_per_zone *mz;
1392 struct mem_cgroup *memcg;
1393 struct page_cgroup *pc;
1394 struct lruvec *lruvec;
1396 if (mem_cgroup_disabled()) {
1397 lruvec = &zone->lruvec;
1401 pc = lookup_page_cgroup(page);
1402 memcg = pc->mem_cgroup;
1405 * Surreptitiously switch any uncharged offlist page to root:
1406 * an uncharged page off lru does nothing to secure
1407 * its former mem_cgroup from sudden removal.
1409 * Our caller holds lru_lock, and PageCgroupUsed is updated
1410 * under page_cgroup lock: between them, they make all uses
1411 * of pc->mem_cgroup safe.
1413 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1414 pc->mem_cgroup = memcg = root_mem_cgroup;
1416 mz = page_cgroup_zoneinfo(memcg, page);
1417 lruvec = &mz->lruvec;
1420 * Since a node can be onlined after the mem_cgroup was created,
1421 * we have to be prepared to initialize lruvec->zone here;
1422 * and if offlined then reonlined, we need to reinitialize it.
1424 if (unlikely(lruvec->zone != zone))
1425 lruvec->zone = zone;
1430 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1431 * @lruvec: mem_cgroup per zone lru vector
1432 * @lru: index of lru list the page is sitting on
1433 * @nr_pages: positive when adding or negative when removing
1435 * This function must be called when a page is added to or removed from an
1438 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1441 struct mem_cgroup_per_zone *mz;
1442 unsigned long *lru_size;
1444 if (mem_cgroup_disabled())
1447 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1448 lru_size = mz->lru_size + lru;
1449 *lru_size += nr_pages;
1450 VM_BUG_ON((long)(*lru_size) < 0);
1454 * Checks whether given mem is same or in the root_mem_cgroup's
1457 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1458 struct mem_cgroup *memcg)
1460 if (root_memcg == memcg)
1462 if (!root_memcg->use_hierarchy || !memcg)
1464 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1467 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1468 struct mem_cgroup *memcg)
1473 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1478 bool task_in_mem_cgroup(struct task_struct *task,
1479 const struct mem_cgroup *memcg)
1481 struct mem_cgroup *curr = NULL;
1482 struct task_struct *p;
1485 p = find_lock_task_mm(task);
1487 curr = try_get_mem_cgroup_from_mm(p->mm);
1491 * All threads may have already detached their mm's, but the oom
1492 * killer still needs to detect if they have already been oom
1493 * killed to prevent needlessly killing additional tasks.
1496 curr = mem_cgroup_from_task(task);
1498 css_get(&curr->css);
1504 * We should check use_hierarchy of "memcg" not "curr". Because checking
1505 * use_hierarchy of "curr" here make this function true if hierarchy is
1506 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1507 * hierarchy(even if use_hierarchy is disabled in "memcg").
1509 ret = mem_cgroup_same_or_subtree(memcg, curr);
1510 css_put(&curr->css);
1514 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1516 unsigned long inactive_ratio;
1517 unsigned long inactive;
1518 unsigned long active;
1521 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1522 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1524 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1526 inactive_ratio = int_sqrt(10 * gb);
1530 return inactive * inactive_ratio < active;
1533 #define mem_cgroup_from_res_counter(counter, member) \
1534 container_of(counter, struct mem_cgroup, member)
1537 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1538 * @memcg: the memory cgroup
1540 * Returns the maximum amount of memory @mem can be charged with, in
1543 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1545 unsigned long long margin;
1547 margin = res_counter_margin(&memcg->res);
1548 if (do_swap_account)
1549 margin = min(margin, res_counter_margin(&memcg->memsw));
1550 return margin >> PAGE_SHIFT;
1553 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1556 if (!css_parent(&memcg->css))
1557 return vm_swappiness;
1559 return memcg->swappiness;
1563 * memcg->moving_account is used for checking possibility that some thread is
1564 * calling move_account(). When a thread on CPU-A starts moving pages under
1565 * a memcg, other threads should check memcg->moving_account under
1566 * rcu_read_lock(), like this:
1570 * memcg->moving_account+1 if (memcg->mocing_account)
1572 * synchronize_rcu() update something.
1577 /* for quick checking without looking up memcg */
1578 atomic_t memcg_moving __read_mostly;
1580 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1582 atomic_inc(&memcg_moving);
1583 atomic_inc(&memcg->moving_account);
1587 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1590 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1591 * We check NULL in callee rather than caller.
1594 atomic_dec(&memcg_moving);
1595 atomic_dec(&memcg->moving_account);
1600 * 2 routines for checking "mem" is under move_account() or not.
1602 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1603 * is used for avoiding races in accounting. If true,
1604 * pc->mem_cgroup may be overwritten.
1606 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1607 * under hierarchy of moving cgroups. This is for
1608 * waiting at hith-memory prressure caused by "move".
1611 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1613 VM_BUG_ON(!rcu_read_lock_held());
1614 return atomic_read(&memcg->moving_account) > 0;
1617 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1619 struct mem_cgroup *from;
1620 struct mem_cgroup *to;
1623 * Unlike task_move routines, we access mc.to, mc.from not under
1624 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1626 spin_lock(&mc.lock);
1632 ret = mem_cgroup_same_or_subtree(memcg, from)
1633 || mem_cgroup_same_or_subtree(memcg, to);
1635 spin_unlock(&mc.lock);
1639 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1641 if (mc.moving_task && current != mc.moving_task) {
1642 if (mem_cgroup_under_move(memcg)) {
1644 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1645 /* moving charge context might have finished. */
1648 finish_wait(&mc.waitq, &wait);
1656 * Take this lock when
1657 * - a code tries to modify page's memcg while it's USED.
1658 * - a code tries to modify page state accounting in a memcg.
1659 * see mem_cgroup_stolen(), too.
1661 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1662 unsigned long *flags)
1664 spin_lock_irqsave(&memcg->move_lock, *flags);
1667 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1668 unsigned long *flags)
1670 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1673 #define K(x) ((x) << (PAGE_SHIFT-10))
1675 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1676 * @memcg: The memory cgroup that went over limit
1677 * @p: Task that is going to be killed
1679 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1682 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1685 * protects memcg_name and makes sure that parallel ooms do not
1688 static DEFINE_SPINLOCK(oom_info_lock);
1689 struct cgroup *task_cgrp;
1690 struct cgroup *mem_cgrp;
1691 static char memcg_name[PATH_MAX];
1693 struct mem_cgroup *iter;
1699 spin_lock(&oom_info_lock);
1702 mem_cgrp = memcg->css.cgroup;
1703 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1705 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1708 * Unfortunately, we are unable to convert to a useful name
1709 * But we'll still print out the usage information
1716 pr_info("Task in %s killed", memcg_name);
1719 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1727 * Continues from above, so we don't need an KERN_ level
1729 pr_cont(" as a result of limit of %s\n", memcg_name);
1732 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1733 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1734 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1735 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1736 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1737 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1738 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1739 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1740 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1741 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1742 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1743 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1745 for_each_mem_cgroup_tree(iter, memcg) {
1746 pr_info("Memory cgroup stats");
1749 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1751 pr_cont(" for %s", memcg_name);
1755 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1756 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1758 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1759 K(mem_cgroup_read_stat(iter, i)));
1762 for (i = 0; i < NR_LRU_LISTS; i++)
1763 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1764 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1768 spin_unlock(&oom_info_lock);
1772 * This function returns the number of memcg under hierarchy tree. Returns
1773 * 1(self count) if no children.
1775 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1778 struct mem_cgroup *iter;
1780 for_each_mem_cgroup_tree(iter, memcg)
1786 * Return the memory (and swap, if configured) limit for a memcg.
1788 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1792 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1795 * Do not consider swap space if we cannot swap due to swappiness
1797 if (mem_cgroup_swappiness(memcg)) {
1800 limit += total_swap_pages << PAGE_SHIFT;
1801 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1804 * If memsw is finite and limits the amount of swap space
1805 * available to this memcg, return that limit.
1807 limit = min(limit, memsw);
1813 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1816 struct mem_cgroup *iter;
1817 unsigned long chosen_points = 0;
1818 unsigned long totalpages;
1819 unsigned int points = 0;
1820 struct task_struct *chosen = NULL;
1823 * If current has a pending SIGKILL or is exiting, then automatically
1824 * select it. The goal is to allow it to allocate so that it may
1825 * quickly exit and free its memory.
1827 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1828 set_thread_flag(TIF_MEMDIE);
1832 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1833 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1834 for_each_mem_cgroup_tree(iter, memcg) {
1835 struct css_task_iter it;
1836 struct task_struct *task;
1838 css_task_iter_start(&iter->css, &it);
1839 while ((task = css_task_iter_next(&it))) {
1840 switch (oom_scan_process_thread(task, totalpages, NULL,
1842 case OOM_SCAN_SELECT:
1844 put_task_struct(chosen);
1846 chosen_points = ULONG_MAX;
1847 get_task_struct(chosen);
1849 case OOM_SCAN_CONTINUE:
1851 case OOM_SCAN_ABORT:
1852 css_task_iter_end(&it);
1853 mem_cgroup_iter_break(memcg, iter);
1855 put_task_struct(chosen);
1860 points = oom_badness(task, memcg, NULL, totalpages);
1861 if (points > chosen_points) {
1863 put_task_struct(chosen);
1865 chosen_points = points;
1866 get_task_struct(chosen);
1869 css_task_iter_end(&it);
1874 points = chosen_points * 1000 / totalpages;
1875 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1876 NULL, "Memory cgroup out of memory");
1879 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1881 unsigned long flags)
1883 unsigned long total = 0;
1884 bool noswap = false;
1887 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1889 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1892 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1894 drain_all_stock_async(memcg);
1895 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1897 * Allow limit shrinkers, which are triggered directly
1898 * by userspace, to catch signals and stop reclaim
1899 * after minimal progress, regardless of the margin.
1901 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1903 if (mem_cgroup_margin(memcg))
1906 * If nothing was reclaimed after two attempts, there
1907 * may be no reclaimable pages in this hierarchy.
1916 * test_mem_cgroup_node_reclaimable
1917 * @memcg: the target memcg
1918 * @nid: the node ID to be checked.
1919 * @noswap : specify true here if the user wants flle only information.
1921 * This function returns whether the specified memcg contains any
1922 * reclaimable pages on a node. Returns true if there are any reclaimable
1923 * pages in the node.
1925 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1926 int nid, bool noswap)
1928 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1930 if (noswap || !total_swap_pages)
1932 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1937 #if MAX_NUMNODES > 1
1940 * Always updating the nodemask is not very good - even if we have an empty
1941 * list or the wrong list here, we can start from some node and traverse all
1942 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1945 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1949 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1950 * pagein/pageout changes since the last update.
1952 if (!atomic_read(&memcg->numainfo_events))
1954 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1957 /* make a nodemask where this memcg uses memory from */
1958 memcg->scan_nodes = node_states[N_MEMORY];
1960 for_each_node_mask(nid, node_states[N_MEMORY]) {
1962 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1963 node_clear(nid, memcg->scan_nodes);
1966 atomic_set(&memcg->numainfo_events, 0);
1967 atomic_set(&memcg->numainfo_updating, 0);
1971 * Selecting a node where we start reclaim from. Because what we need is just
1972 * reducing usage counter, start from anywhere is O,K. Considering
1973 * memory reclaim from current node, there are pros. and cons.
1975 * Freeing memory from current node means freeing memory from a node which
1976 * we'll use or we've used. So, it may make LRU bad. And if several threads
1977 * hit limits, it will see a contention on a node. But freeing from remote
1978 * node means more costs for memory reclaim because of memory latency.
1980 * Now, we use round-robin. Better algorithm is welcomed.
1982 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1986 mem_cgroup_may_update_nodemask(memcg);
1987 node = memcg->last_scanned_node;
1989 node = next_node(node, memcg->scan_nodes);
1990 if (node == MAX_NUMNODES)
1991 node = first_node(memcg->scan_nodes);
1993 * We call this when we hit limit, not when pages are added to LRU.
1994 * No LRU may hold pages because all pages are UNEVICTABLE or
1995 * memcg is too small and all pages are not on LRU. In that case,
1996 * we use curret node.
1998 if (unlikely(node == MAX_NUMNODES))
1999 node = numa_node_id();
2001 memcg->last_scanned_node = node;
2006 * Check all nodes whether it contains reclaimable pages or not.
2007 * For quick scan, we make use of scan_nodes. This will allow us to skip
2008 * unused nodes. But scan_nodes is lazily updated and may not cotain
2009 * enough new information. We need to do double check.
2011 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2016 * quick check...making use of scan_node.
2017 * We can skip unused nodes.
2019 if (!nodes_empty(memcg->scan_nodes)) {
2020 for (nid = first_node(memcg->scan_nodes);
2022 nid = next_node(nid, memcg->scan_nodes)) {
2024 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2029 * Check rest of nodes.
2031 for_each_node_state(nid, N_MEMORY) {
2032 if (node_isset(nid, memcg->scan_nodes))
2034 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2041 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2046 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2048 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2052 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2055 unsigned long *total_scanned)
2057 struct mem_cgroup *victim = NULL;
2060 unsigned long excess;
2061 unsigned long nr_scanned;
2062 struct mem_cgroup_reclaim_cookie reclaim = {
2067 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2070 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2075 * If we have not been able to reclaim
2076 * anything, it might because there are
2077 * no reclaimable pages under this hierarchy
2082 * We want to do more targeted reclaim.
2083 * excess >> 2 is not to excessive so as to
2084 * reclaim too much, nor too less that we keep
2085 * coming back to reclaim from this cgroup
2087 if (total >= (excess >> 2) ||
2088 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2093 if (!mem_cgroup_reclaimable(victim, false))
2095 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2097 *total_scanned += nr_scanned;
2098 if (!res_counter_soft_limit_excess(&root_memcg->res))
2101 mem_cgroup_iter_break(root_memcg, victim);
2105 #ifdef CONFIG_LOCKDEP
2106 static struct lockdep_map memcg_oom_lock_dep_map = {
2107 .name = "memcg_oom_lock",
2111 static DEFINE_SPINLOCK(memcg_oom_lock);
2114 * Check OOM-Killer is already running under our hierarchy.
2115 * If someone is running, return false.
2117 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2119 struct mem_cgroup *iter, *failed = NULL;
2121 spin_lock(&memcg_oom_lock);
2123 for_each_mem_cgroup_tree(iter, memcg) {
2124 if (iter->oom_lock) {
2126 * this subtree of our hierarchy is already locked
2127 * so we cannot give a lock.
2130 mem_cgroup_iter_break(memcg, iter);
2133 iter->oom_lock = true;
2138 * OK, we failed to lock the whole subtree so we have
2139 * to clean up what we set up to the failing subtree
2141 for_each_mem_cgroup_tree(iter, memcg) {
2142 if (iter == failed) {
2143 mem_cgroup_iter_break(memcg, iter);
2146 iter->oom_lock = false;
2149 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2151 spin_unlock(&memcg_oom_lock);
2156 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2158 struct mem_cgroup *iter;
2160 spin_lock(&memcg_oom_lock);
2161 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2162 for_each_mem_cgroup_tree(iter, memcg)
2163 iter->oom_lock = false;
2164 spin_unlock(&memcg_oom_lock);
2167 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2169 struct mem_cgroup *iter;
2171 for_each_mem_cgroup_tree(iter, memcg)
2172 atomic_inc(&iter->under_oom);
2175 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2177 struct mem_cgroup *iter;
2180 * When a new child is created while the hierarchy is under oom,
2181 * mem_cgroup_oom_lock() may not be called. We have to use
2182 * atomic_add_unless() here.
2184 for_each_mem_cgroup_tree(iter, memcg)
2185 atomic_add_unless(&iter->under_oom, -1, 0);
2188 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2190 struct oom_wait_info {
2191 struct mem_cgroup *memcg;
2195 static int memcg_oom_wake_function(wait_queue_t *wait,
2196 unsigned mode, int sync, void *arg)
2198 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2199 struct mem_cgroup *oom_wait_memcg;
2200 struct oom_wait_info *oom_wait_info;
2202 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2203 oom_wait_memcg = oom_wait_info->memcg;
2206 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2207 * Then we can use css_is_ancestor without taking care of RCU.
2209 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2210 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2212 return autoremove_wake_function(wait, mode, sync, arg);
2215 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2217 atomic_inc(&memcg->oom_wakeups);
2218 /* for filtering, pass "memcg" as argument. */
2219 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2222 static void memcg_oom_recover(struct mem_cgroup *memcg)
2224 if (memcg && atomic_read(&memcg->under_oom))
2225 memcg_wakeup_oom(memcg);
2228 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2230 if (!current->memcg_oom.may_oom)
2233 * We are in the middle of the charge context here, so we
2234 * don't want to block when potentially sitting on a callstack
2235 * that holds all kinds of filesystem and mm locks.
2237 * Also, the caller may handle a failed allocation gracefully
2238 * (like optional page cache readahead) and so an OOM killer
2239 * invocation might not even be necessary.
2241 * That's why we don't do anything here except remember the
2242 * OOM context and then deal with it at the end of the page
2243 * fault when the stack is unwound, the locks are released,
2244 * and when we know whether the fault was overall successful.
2246 css_get(&memcg->css);
2247 current->memcg_oom.memcg = memcg;
2248 current->memcg_oom.gfp_mask = mask;
2249 current->memcg_oom.order = order;
2253 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2254 * @handle: actually kill/wait or just clean up the OOM state
2256 * This has to be called at the end of a page fault if the memcg OOM
2257 * handler was enabled.
2259 * Memcg supports userspace OOM handling where failed allocations must
2260 * sleep on a waitqueue until the userspace task resolves the
2261 * situation. Sleeping directly in the charge context with all kinds
2262 * of locks held is not a good idea, instead we remember an OOM state
2263 * in the task and mem_cgroup_oom_synchronize() has to be called at
2264 * the end of the page fault to complete the OOM handling.
2266 * Returns %true if an ongoing memcg OOM situation was detected and
2267 * completed, %false otherwise.
2269 bool mem_cgroup_oom_synchronize(bool handle)
2271 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2272 struct oom_wait_info owait;
2275 /* OOM is global, do not handle */
2282 owait.memcg = memcg;
2283 owait.wait.flags = 0;
2284 owait.wait.func = memcg_oom_wake_function;
2285 owait.wait.private = current;
2286 INIT_LIST_HEAD(&owait.wait.task_list);
2288 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2289 mem_cgroup_mark_under_oom(memcg);
2291 locked = mem_cgroup_oom_trylock(memcg);
2294 mem_cgroup_oom_notify(memcg);
2296 if (locked && !memcg->oom_kill_disable) {
2297 mem_cgroup_unmark_under_oom(memcg);
2298 finish_wait(&memcg_oom_waitq, &owait.wait);
2299 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2300 current->memcg_oom.order);
2303 mem_cgroup_unmark_under_oom(memcg);
2304 finish_wait(&memcg_oom_waitq, &owait.wait);
2308 mem_cgroup_oom_unlock(memcg);
2310 * There is no guarantee that an OOM-lock contender
2311 * sees the wakeups triggered by the OOM kill
2312 * uncharges. Wake any sleepers explicitely.
2314 memcg_oom_recover(memcg);
2317 current->memcg_oom.memcg = NULL;
2318 css_put(&memcg->css);
2323 * Currently used to update mapped file statistics, but the routine can be
2324 * generalized to update other statistics as well.
2326 * Notes: Race condition
2328 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2329 * it tends to be costly. But considering some conditions, we doesn't need
2330 * to do so _always_.
2332 * Considering "charge", lock_page_cgroup() is not required because all
2333 * file-stat operations happen after a page is attached to radix-tree. There
2334 * are no race with "charge".
2336 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2337 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2338 * if there are race with "uncharge". Statistics itself is properly handled
2341 * Considering "move", this is an only case we see a race. To make the race
2342 * small, we check mm->moving_account and detect there are possibility of race
2343 * If there is, we take a lock.
2346 void __mem_cgroup_begin_update_page_stat(struct page *page,
2347 bool *locked, unsigned long *flags)
2349 struct mem_cgroup *memcg;
2350 struct page_cgroup *pc;
2352 pc = lookup_page_cgroup(page);
2354 memcg = pc->mem_cgroup;
2355 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2358 * If this memory cgroup is not under account moving, we don't
2359 * need to take move_lock_mem_cgroup(). Because we already hold
2360 * rcu_read_lock(), any calls to move_account will be delayed until
2361 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2363 if (!mem_cgroup_stolen(memcg))
2366 move_lock_mem_cgroup(memcg, flags);
2367 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2368 move_unlock_mem_cgroup(memcg, flags);
2374 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2376 struct page_cgroup *pc = lookup_page_cgroup(page);
2379 * It's guaranteed that pc->mem_cgroup never changes while
2380 * lock is held because a routine modifies pc->mem_cgroup
2381 * should take move_lock_mem_cgroup().
2383 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2386 void mem_cgroup_update_page_stat(struct page *page,
2387 enum mem_cgroup_stat_index idx, int val)
2389 struct mem_cgroup *memcg;
2390 struct page_cgroup *pc = lookup_page_cgroup(page);
2391 unsigned long uninitialized_var(flags);
2393 if (mem_cgroup_disabled())
2396 VM_BUG_ON(!rcu_read_lock_held());
2397 memcg = pc->mem_cgroup;
2398 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2401 this_cpu_add(memcg->stat->count[idx], val);
2405 * size of first charge trial. "32" comes from vmscan.c's magic value.
2406 * TODO: maybe necessary to use big numbers in big irons.
2408 #define CHARGE_BATCH 32U
2409 struct memcg_stock_pcp {
2410 struct mem_cgroup *cached; /* this never be root cgroup */
2411 unsigned int nr_pages;
2412 struct work_struct work;
2413 unsigned long flags;
2414 #define FLUSHING_CACHED_CHARGE 0
2416 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2417 static DEFINE_MUTEX(percpu_charge_mutex);
2420 * consume_stock: Try to consume stocked charge on this cpu.
2421 * @memcg: memcg to consume from.
2422 * @nr_pages: how many pages to charge.
2424 * The charges will only happen if @memcg matches the current cpu's memcg
2425 * stock, and at least @nr_pages are available in that stock. Failure to
2426 * service an allocation will refill the stock.
2428 * returns true if successful, false otherwise.
2430 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2432 struct memcg_stock_pcp *stock;
2435 if (nr_pages > CHARGE_BATCH)
2438 stock = &get_cpu_var(memcg_stock);
2439 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2440 stock->nr_pages -= nr_pages;
2441 else /* need to call res_counter_charge */
2443 put_cpu_var(memcg_stock);
2448 * Returns stocks cached in percpu to res_counter and reset cached information.
2450 static void drain_stock(struct memcg_stock_pcp *stock)
2452 struct mem_cgroup *old = stock->cached;
2454 if (stock->nr_pages) {
2455 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2457 res_counter_uncharge(&old->res, bytes);
2458 if (do_swap_account)
2459 res_counter_uncharge(&old->memsw, bytes);
2460 stock->nr_pages = 0;
2462 stock->cached = NULL;
2466 * This must be called under preempt disabled or must be called by
2467 * a thread which is pinned to local cpu.
2469 static void drain_local_stock(struct work_struct *dummy)
2471 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2473 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2476 static void __init memcg_stock_init(void)
2480 for_each_possible_cpu(cpu) {
2481 struct memcg_stock_pcp *stock =
2482 &per_cpu(memcg_stock, cpu);
2483 INIT_WORK(&stock->work, drain_local_stock);
2488 * Cache charges(val) which is from res_counter, to local per_cpu area.
2489 * This will be consumed by consume_stock() function, later.
2491 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2493 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2495 if (stock->cached != memcg) { /* reset if necessary */
2497 stock->cached = memcg;
2499 stock->nr_pages += nr_pages;
2500 put_cpu_var(memcg_stock);
2504 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2505 * of the hierarchy under it. sync flag says whether we should block
2506 * until the work is done.
2508 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2512 /* Notify other cpus that system-wide "drain" is running */
2515 for_each_online_cpu(cpu) {
2516 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2517 struct mem_cgroup *memcg;
2519 memcg = stock->cached;
2520 if (!memcg || !stock->nr_pages)
2522 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2524 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2526 drain_local_stock(&stock->work);
2528 schedule_work_on(cpu, &stock->work);
2536 for_each_online_cpu(cpu) {
2537 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2538 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2539 flush_work(&stock->work);
2546 * Tries to drain stocked charges in other cpus. This function is asynchronous
2547 * and just put a work per cpu for draining localy on each cpu. Caller can
2548 * expects some charges will be back to res_counter later but cannot wait for
2551 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2554 * If someone calls draining, avoid adding more kworker runs.
2556 if (!mutex_trylock(&percpu_charge_mutex))
2558 drain_all_stock(root_memcg, false);
2559 mutex_unlock(&percpu_charge_mutex);
2562 /* This is a synchronous drain interface. */
2563 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2565 /* called when force_empty is called */
2566 mutex_lock(&percpu_charge_mutex);
2567 drain_all_stock(root_memcg, true);
2568 mutex_unlock(&percpu_charge_mutex);
2572 * This function drains percpu counter value from DEAD cpu and
2573 * move it to local cpu. Note that this function can be preempted.
2575 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2579 spin_lock(&memcg->pcp_counter_lock);
2580 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2581 long x = per_cpu(memcg->stat->count[i], cpu);
2583 per_cpu(memcg->stat->count[i], cpu) = 0;
2584 memcg->nocpu_base.count[i] += x;
2586 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2587 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2589 per_cpu(memcg->stat->events[i], cpu) = 0;
2590 memcg->nocpu_base.events[i] += x;
2592 spin_unlock(&memcg->pcp_counter_lock);
2595 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2596 unsigned long action,
2599 int cpu = (unsigned long)hcpu;
2600 struct memcg_stock_pcp *stock;
2601 struct mem_cgroup *iter;
2603 if (action == CPU_ONLINE)
2606 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2609 for_each_mem_cgroup(iter)
2610 mem_cgroup_drain_pcp_counter(iter, cpu);
2612 stock = &per_cpu(memcg_stock, cpu);
2618 /* See __mem_cgroup_try_charge() for details */
2620 CHARGE_OK, /* success */
2621 CHARGE_RETRY, /* need to retry but retry is not bad */
2622 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2623 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2626 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2627 unsigned int nr_pages, unsigned int min_pages,
2630 unsigned long csize = nr_pages * PAGE_SIZE;
2631 struct mem_cgroup *mem_over_limit;
2632 struct res_counter *fail_res;
2633 unsigned long flags = 0;
2636 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2639 if (!do_swap_account)
2641 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2645 res_counter_uncharge(&memcg->res, csize);
2646 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2647 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2649 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2651 * Never reclaim on behalf of optional batching, retry with a
2652 * single page instead.
2654 if (nr_pages > min_pages)
2655 return CHARGE_RETRY;
2657 if (!(gfp_mask & __GFP_WAIT))
2658 return CHARGE_WOULDBLOCK;
2660 if (gfp_mask & __GFP_NORETRY)
2661 return CHARGE_NOMEM;
2663 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2664 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2665 return CHARGE_RETRY;
2667 * Even though the limit is exceeded at this point, reclaim
2668 * may have been able to free some pages. Retry the charge
2669 * before killing the task.
2671 * Only for regular pages, though: huge pages are rather
2672 * unlikely to succeed so close to the limit, and we fall back
2673 * to regular pages anyway in case of failure.
2675 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2676 return CHARGE_RETRY;
2679 * At task move, charge accounts can be doubly counted. So, it's
2680 * better to wait until the end of task_move if something is going on.
2682 if (mem_cgroup_wait_acct_move(mem_over_limit))
2683 return CHARGE_RETRY;
2686 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2688 return CHARGE_NOMEM;
2692 * __mem_cgroup_try_charge() does
2693 * 1. detect memcg to be charged against from passed *mm and *ptr,
2694 * 2. update res_counter
2695 * 3. call memory reclaim if necessary.
2697 * In some special case, if the task is fatal, fatal_signal_pending() or
2698 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2699 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2700 * as possible without any hazards. 2: all pages should have a valid
2701 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2702 * pointer, that is treated as a charge to root_mem_cgroup.
2704 * So __mem_cgroup_try_charge() will return
2705 * 0 ... on success, filling *ptr with a valid memcg pointer.
2706 * -ENOMEM ... charge failure because of resource limits.
2707 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2709 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2710 * the oom-killer can be invoked.
2712 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2714 unsigned int nr_pages,
2715 struct mem_cgroup **ptr,
2718 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2719 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2720 struct mem_cgroup *memcg = NULL;
2724 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2725 * in system level. So, allow to go ahead dying process in addition to
2728 if (unlikely(test_thread_flag(TIF_MEMDIE)
2729 || fatal_signal_pending(current)))
2732 if (unlikely(task_in_memcg_oom(current)))
2735 if (gfp_mask & __GFP_NOFAIL)
2739 * We always charge the cgroup the mm_struct belongs to.
2740 * The mm_struct's mem_cgroup changes on task migration if the
2741 * thread group leader migrates. It's possible that mm is not
2742 * set, if so charge the root memcg (happens for pagecache usage).
2745 *ptr = root_mem_cgroup;
2747 if (*ptr) { /* css should be a valid one */
2749 if (mem_cgroup_is_root(memcg))
2751 if (consume_stock(memcg, nr_pages))
2753 css_get(&memcg->css);
2755 struct task_struct *p;
2758 p = rcu_dereference(mm->owner);
2760 * Because we don't have task_lock(), "p" can exit.
2761 * In that case, "memcg" can point to root or p can be NULL with
2762 * race with swapoff. Then, we have small risk of mis-accouning.
2763 * But such kind of mis-account by race always happens because
2764 * we don't have cgroup_mutex(). It's overkill and we allo that
2766 * (*) swapoff at el will charge against mm-struct not against
2767 * task-struct. So, mm->owner can be NULL.
2769 memcg = mem_cgroup_from_task(p);
2771 memcg = root_mem_cgroup;
2772 if (mem_cgroup_is_root(memcg)) {
2776 if (consume_stock(memcg, nr_pages)) {
2778 * It seems dagerous to access memcg without css_get().
2779 * But considering how consume_stok works, it's not
2780 * necessary. If consume_stock success, some charges
2781 * from this memcg are cached on this cpu. So, we
2782 * don't need to call css_get()/css_tryget() before
2783 * calling consume_stock().
2788 /* after here, we may be blocked. we need to get refcnt */
2789 if (!css_tryget(&memcg->css)) {
2797 bool invoke_oom = oom && !nr_oom_retries;
2799 /* If killed, bypass charge */
2800 if (fatal_signal_pending(current)) {
2801 css_put(&memcg->css);
2805 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2806 nr_pages, invoke_oom);
2810 case CHARGE_RETRY: /* not in OOM situation but retry */
2812 css_put(&memcg->css);
2815 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2816 css_put(&memcg->css);
2818 case CHARGE_NOMEM: /* OOM routine works */
2819 if (!oom || invoke_oom) {
2820 css_put(&memcg->css);
2826 } while (ret != CHARGE_OK);
2828 if (batch > nr_pages)
2829 refill_stock(memcg, batch - nr_pages);
2830 css_put(&memcg->css);
2835 if (!(gfp_mask & __GFP_NOFAIL)) {
2840 *ptr = root_mem_cgroup;
2845 * Somemtimes we have to undo a charge we got by try_charge().
2846 * This function is for that and do uncharge, put css's refcnt.
2847 * gotten by try_charge().
2849 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2850 unsigned int nr_pages)
2852 if (!mem_cgroup_is_root(memcg)) {
2853 unsigned long bytes = nr_pages * PAGE_SIZE;
2855 res_counter_uncharge(&memcg->res, bytes);
2856 if (do_swap_account)
2857 res_counter_uncharge(&memcg->memsw, bytes);
2862 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2863 * This is useful when moving usage to parent cgroup.
2865 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2866 unsigned int nr_pages)
2868 unsigned long bytes = nr_pages * PAGE_SIZE;
2870 if (mem_cgroup_is_root(memcg))
2873 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2874 if (do_swap_account)
2875 res_counter_uncharge_until(&memcg->memsw,
2876 memcg->memsw.parent, bytes);
2880 * A helper function to get mem_cgroup from ID. must be called under
2881 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2882 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2883 * called against removed memcg.)
2885 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2887 /* ID 0 is unused ID */
2890 return mem_cgroup_from_id(id);
2893 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2895 struct mem_cgroup *memcg = NULL;
2896 struct page_cgroup *pc;
2900 VM_BUG_ON_PAGE(!PageLocked(page), page);
2902 pc = lookup_page_cgroup(page);
2903 lock_page_cgroup(pc);
2904 if (PageCgroupUsed(pc)) {
2905 memcg = pc->mem_cgroup;
2906 if (memcg && !css_tryget(&memcg->css))
2908 } else if (PageSwapCache(page)) {
2909 ent.val = page_private(page);
2910 id = lookup_swap_cgroup_id(ent);
2912 memcg = mem_cgroup_lookup(id);
2913 if (memcg && !css_tryget(&memcg->css))
2917 unlock_page_cgroup(pc);
2921 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2923 unsigned int nr_pages,
2924 enum charge_type ctype,
2927 struct page_cgroup *pc = lookup_page_cgroup(page);
2928 struct zone *uninitialized_var(zone);
2929 struct lruvec *lruvec;
2930 bool was_on_lru = false;
2933 lock_page_cgroup(pc);
2934 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2936 * we don't need page_cgroup_lock about tail pages, becase they are not
2937 * accessed by any other context at this point.
2941 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2942 * may already be on some other mem_cgroup's LRU. Take care of it.
2945 zone = page_zone(page);
2946 spin_lock_irq(&zone->lru_lock);
2947 if (PageLRU(page)) {
2948 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2950 del_page_from_lru_list(page, lruvec, page_lru(page));
2955 pc->mem_cgroup = memcg;
2957 * We access a page_cgroup asynchronously without lock_page_cgroup().
2958 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2959 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2960 * before USED bit, we need memory barrier here.
2961 * See mem_cgroup_add_lru_list(), etc.
2964 SetPageCgroupUsed(pc);
2968 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2969 VM_BUG_ON_PAGE(PageLRU(page), page);
2971 add_page_to_lru_list(page, lruvec, page_lru(page));
2973 spin_unlock_irq(&zone->lru_lock);
2976 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2981 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2982 unlock_page_cgroup(pc);
2985 * "charge_statistics" updated event counter. Then, check it.
2986 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2987 * if they exceeds softlimit.
2989 memcg_check_events(memcg, page);
2992 static DEFINE_MUTEX(set_limit_mutex);
2994 #ifdef CONFIG_MEMCG_KMEM
2995 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2997 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2998 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK) ==
2999 KMEM_ACCOUNTED_MASK;
3003 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
3004 * in the memcg_cache_params struct.
3006 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
3008 struct kmem_cache *cachep;
3010 VM_BUG_ON(p->is_root_cache);
3011 cachep = p->root_cache;
3012 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
3015 #ifdef CONFIG_SLABINFO
3016 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
3018 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3019 struct memcg_cache_params *params;
3021 if (!memcg_can_account_kmem(memcg))
3024 print_slabinfo_header(m);
3026 mutex_lock(&memcg->slab_caches_mutex);
3027 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
3028 cache_show(memcg_params_to_cache(params), m);
3029 mutex_unlock(&memcg->slab_caches_mutex);
3035 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3037 struct res_counter *fail_res;
3038 struct mem_cgroup *_memcg;
3041 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3046 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3047 &_memcg, oom_gfp_allowed(gfp));
3049 if (ret == -EINTR) {
3051 * __mem_cgroup_try_charge() chosed to bypass to root due to
3052 * OOM kill or fatal signal. Since our only options are to
3053 * either fail the allocation or charge it to this cgroup, do
3054 * it as a temporary condition. But we can't fail. From a
3055 * kmem/slab perspective, the cache has already been selected,
3056 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3059 * This condition will only trigger if the task entered
3060 * memcg_charge_kmem in a sane state, but was OOM-killed during
3061 * __mem_cgroup_try_charge() above. Tasks that were already
3062 * dying when the allocation triggers should have been already
3063 * directed to the root cgroup in memcontrol.h
3065 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3066 if (do_swap_account)
3067 res_counter_charge_nofail(&memcg->memsw, size,
3071 res_counter_uncharge(&memcg->kmem, size);
3076 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3078 res_counter_uncharge(&memcg->res, size);
3079 if (do_swap_account)
3080 res_counter_uncharge(&memcg->memsw, size);
3083 if (res_counter_uncharge(&memcg->kmem, size))
3087 * Releases a reference taken in kmem_cgroup_css_offline in case
3088 * this last uncharge is racing with the offlining code or it is
3089 * outliving the memcg existence.
3091 * The memory barrier imposed by test&clear is paired with the
3092 * explicit one in memcg_kmem_mark_dead().
3094 if (memcg_kmem_test_and_clear_dead(memcg))
3095 css_put(&memcg->css);
3099 * helper for acessing a memcg's index. It will be used as an index in the
3100 * child cache array in kmem_cache, and also to derive its name. This function
3101 * will return -1 when this is not a kmem-limited memcg.
3103 int memcg_cache_id(struct mem_cgroup *memcg)
3105 return memcg ? memcg->kmemcg_id : -1;
3109 * This ends up being protected by the set_limit mutex, during normal
3110 * operation, because that is its main call site.
3112 * But when we create a new cache, we can call this as well if its parent
3113 * is kmem-limited. That will have to hold set_limit_mutex as well.
3115 static int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3119 num = ida_simple_get(&kmem_limited_groups,
3120 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3124 * After this point, kmem_accounted (that we test atomically in
3125 * the beginning of this conditional), is no longer 0. This
3126 * guarantees only one process will set the following boolean
3127 * to true. We don't need test_and_set because we're protected
3128 * by the set_limit_mutex anyway.
3130 memcg_kmem_set_activated(memcg);
3132 ret = memcg_update_all_caches(num+1);
3134 ida_simple_remove(&kmem_limited_groups, num);
3135 memcg_kmem_clear_activated(memcg);
3139 memcg->kmemcg_id = num;
3140 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3141 mutex_init(&memcg->slab_caches_mutex);
3145 static size_t memcg_caches_array_size(int num_groups)
3148 if (num_groups <= 0)
3151 size = 2 * num_groups;
3152 if (size < MEMCG_CACHES_MIN_SIZE)
3153 size = MEMCG_CACHES_MIN_SIZE;
3154 else if (size > MEMCG_CACHES_MAX_SIZE)
3155 size = MEMCG_CACHES_MAX_SIZE;
3161 * We should update the current array size iff all caches updates succeed. This
3162 * can only be done from the slab side. The slab mutex needs to be held when
3165 void memcg_update_array_size(int num)
3167 if (num > memcg_limited_groups_array_size)
3168 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3171 static void kmem_cache_destroy_work_func(struct work_struct *w);
3173 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3175 struct memcg_cache_params *cur_params = s->memcg_params;
3177 VM_BUG_ON(!is_root_cache(s));
3179 if (num_groups > memcg_limited_groups_array_size) {
3181 ssize_t size = memcg_caches_array_size(num_groups);
3183 size *= sizeof(void *);
3184 size += offsetof(struct memcg_cache_params, memcg_caches);
3186 s->memcg_params = kzalloc(size, GFP_KERNEL);
3187 if (!s->memcg_params) {
3188 s->memcg_params = cur_params;
3192 s->memcg_params->is_root_cache = true;
3195 * There is the chance it will be bigger than
3196 * memcg_limited_groups_array_size, if we failed an allocation
3197 * in a cache, in which case all caches updated before it, will
3198 * have a bigger array.
3200 * But if that is the case, the data after
3201 * memcg_limited_groups_array_size is certainly unused
3203 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3204 if (!cur_params->memcg_caches[i])
3206 s->memcg_params->memcg_caches[i] =
3207 cur_params->memcg_caches[i];
3211 * Ideally, we would wait until all caches succeed, and only
3212 * then free the old one. But this is not worth the extra
3213 * pointer per-cache we'd have to have for this.
3215 * It is not a big deal if some caches are left with a size
3216 * bigger than the others. And all updates will reset this
3224 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3225 struct kmem_cache *root_cache)
3229 if (!memcg_kmem_enabled())
3233 size = offsetof(struct memcg_cache_params, memcg_caches);
3234 size += memcg_limited_groups_array_size * sizeof(void *);
3236 size = sizeof(struct memcg_cache_params);
3238 s->memcg_params = kzalloc(size, GFP_KERNEL);
3239 if (!s->memcg_params)
3243 s->memcg_params->memcg = memcg;
3244 s->memcg_params->root_cache = root_cache;
3245 INIT_WORK(&s->memcg_params->destroy,
3246 kmem_cache_destroy_work_func);
3248 s->memcg_params->is_root_cache = true;
3253 void memcg_free_cache_params(struct kmem_cache *s)
3255 kfree(s->memcg_params);
3258 void memcg_register_cache(struct kmem_cache *s)
3260 struct kmem_cache *root;
3261 struct mem_cgroup *memcg;
3264 if (is_root_cache(s))
3268 * Holding the slab_mutex assures nobody will touch the memcg_caches
3269 * array while we are modifying it.
3271 lockdep_assert_held(&slab_mutex);
3273 root = s->memcg_params->root_cache;
3274 memcg = s->memcg_params->memcg;
3275 id = memcg_cache_id(memcg);
3277 css_get(&memcg->css);
3281 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3282 * barrier here to ensure nobody will see the kmem_cache partially
3288 * Initialize the pointer to this cache in its parent's memcg_params
3289 * before adding it to the memcg_slab_caches list, otherwise we can
3290 * fail to convert memcg_params_to_cache() while traversing the list.
3292 VM_BUG_ON(root->memcg_params->memcg_caches[id]);
3293 root->memcg_params->memcg_caches[id] = s;
3295 mutex_lock(&memcg->slab_caches_mutex);
3296 list_add(&s->memcg_params->list, &memcg->memcg_slab_caches);
3297 mutex_unlock(&memcg->slab_caches_mutex);
3300 void memcg_unregister_cache(struct kmem_cache *s)
3302 struct kmem_cache *root;
3303 struct mem_cgroup *memcg;
3306 if (is_root_cache(s))
3310 * Holding the slab_mutex assures nobody will touch the memcg_caches
3311 * array while we are modifying it.
3313 lockdep_assert_held(&slab_mutex);
3315 root = s->memcg_params->root_cache;
3316 memcg = s->memcg_params->memcg;
3317 id = memcg_cache_id(memcg);
3319 mutex_lock(&memcg->slab_caches_mutex);
3320 list_del(&s->memcg_params->list);
3321 mutex_unlock(&memcg->slab_caches_mutex);
3324 * Clear the pointer to this cache in its parent's memcg_params only
3325 * after removing it from the memcg_slab_caches list, otherwise we can
3326 * fail to convert memcg_params_to_cache() while traversing the list.
3328 VM_BUG_ON(!root->memcg_params->memcg_caches[id]);
3329 root->memcg_params->memcg_caches[id] = NULL;
3331 css_put(&memcg->css);
3335 * During the creation a new cache, we need to disable our accounting mechanism
3336 * altogether. This is true even if we are not creating, but rather just
3337 * enqueing new caches to be created.
3339 * This is because that process will trigger allocations; some visible, like
3340 * explicit kmallocs to auxiliary data structures, name strings and internal
3341 * cache structures; some well concealed, like INIT_WORK() that can allocate
3342 * objects during debug.
3344 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3345 * to it. This may not be a bounded recursion: since the first cache creation
3346 * failed to complete (waiting on the allocation), we'll just try to create the
3347 * cache again, failing at the same point.
3349 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3350 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3351 * inside the following two functions.
3353 static inline void memcg_stop_kmem_account(void)
3355 VM_BUG_ON(!current->mm);
3356 current->memcg_kmem_skip_account++;
3359 static inline void memcg_resume_kmem_account(void)
3361 VM_BUG_ON(!current->mm);
3362 current->memcg_kmem_skip_account--;
3365 static void kmem_cache_destroy_work_func(struct work_struct *w)
3367 struct kmem_cache *cachep;
3368 struct memcg_cache_params *p;
3370 p = container_of(w, struct memcg_cache_params, destroy);
3372 cachep = memcg_params_to_cache(p);
3375 * If we get down to 0 after shrink, we could delete right away.
3376 * However, memcg_release_pages() already puts us back in the workqueue
3377 * in that case. If we proceed deleting, we'll get a dangling
3378 * reference, and removing the object from the workqueue in that case
3379 * is unnecessary complication. We are not a fast path.
3381 * Note that this case is fundamentally different from racing with
3382 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3383 * kmem_cache_shrink, not only we would be reinserting a dead cache
3384 * into the queue, but doing so from inside the worker racing to
3387 * So if we aren't down to zero, we'll just schedule a worker and try
3390 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3391 kmem_cache_shrink(cachep);
3392 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3395 kmem_cache_destroy(cachep);
3398 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3400 if (!cachep->memcg_params->dead)
3404 * There are many ways in which we can get here.
3406 * We can get to a memory-pressure situation while the delayed work is
3407 * still pending to run. The vmscan shrinkers can then release all
3408 * cache memory and get us to destruction. If this is the case, we'll
3409 * be executed twice, which is a bug (the second time will execute over
3410 * bogus data). In this case, cancelling the work should be fine.
3412 * But we can also get here from the worker itself, if
3413 * kmem_cache_shrink is enough to shake all the remaining objects and
3414 * get the page count to 0. In this case, we'll deadlock if we try to
3415 * cancel the work (the worker runs with an internal lock held, which
3416 * is the same lock we would hold for cancel_work_sync().)
3418 * Since we can't possibly know who got us here, just refrain from
3419 * running if there is already work pending
3421 if (work_pending(&cachep->memcg_params->destroy))
3424 * We have to defer the actual destroying to a workqueue, because
3425 * we might currently be in a context that cannot sleep.
3427 schedule_work(&cachep->memcg_params->destroy);
3430 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3431 struct kmem_cache *s)
3433 struct kmem_cache *new;
3434 static char *tmp_name = NULL;
3435 static DEFINE_MUTEX(mutex); /* protects tmp_name */
3437 BUG_ON(!memcg_can_account_kmem(memcg));
3441 * kmem_cache_create_memcg duplicates the given name and
3442 * cgroup_name for this name requires RCU context.
3443 * This static temporary buffer is used to prevent from
3444 * pointless shortliving allocation.
3447 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3453 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3454 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3457 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3458 (s->flags & ~SLAB_PANIC), s->ctor, s);
3461 new->allocflags |= __GFP_KMEMCG;
3465 mutex_unlock(&mutex);
3469 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3471 struct kmem_cache *c;
3474 if (!s->memcg_params)
3476 if (!s->memcg_params->is_root_cache)
3480 * If the cache is being destroyed, we trust that there is no one else
3481 * requesting objects from it. Even if there are, the sanity checks in
3482 * kmem_cache_destroy should caught this ill-case.
3484 * Still, we don't want anyone else freeing memcg_caches under our
3485 * noses, which can happen if a new memcg comes to life. As usual,
3486 * we'll take the set_limit_mutex to protect ourselves against this.
3488 mutex_lock(&set_limit_mutex);
3489 for_each_memcg_cache_index(i) {
3490 c = cache_from_memcg_idx(s, i);
3495 * We will now manually delete the caches, so to avoid races
3496 * we need to cancel all pending destruction workers and
3497 * proceed with destruction ourselves.
3499 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3500 * and that could spawn the workers again: it is likely that
3501 * the cache still have active pages until this very moment.
3502 * This would lead us back to mem_cgroup_destroy_cache.
3504 * But that will not execute at all if the "dead" flag is not
3505 * set, so flip it down to guarantee we are in control.
3507 c->memcg_params->dead = false;
3508 cancel_work_sync(&c->memcg_params->destroy);
3509 kmem_cache_destroy(c);
3511 mutex_unlock(&set_limit_mutex);
3514 struct create_work {
3515 struct mem_cgroup *memcg;
3516 struct kmem_cache *cachep;
3517 struct work_struct work;
3520 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3522 struct kmem_cache *cachep;
3523 struct memcg_cache_params *params;
3525 if (!memcg_kmem_is_active(memcg))
3528 mutex_lock(&memcg->slab_caches_mutex);
3529 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3530 cachep = memcg_params_to_cache(params);
3531 cachep->memcg_params->dead = true;
3532 schedule_work(&cachep->memcg_params->destroy);
3534 mutex_unlock(&memcg->slab_caches_mutex);
3537 static void memcg_create_cache_work_func(struct work_struct *w)
3539 struct create_work *cw;
3541 cw = container_of(w, struct create_work, work);
3542 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3543 css_put(&cw->memcg->css);
3548 * Enqueue the creation of a per-memcg kmem_cache.
3550 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3551 struct kmem_cache *cachep)
3553 struct create_work *cw;
3555 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3557 css_put(&memcg->css);
3562 cw->cachep = cachep;
3564 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3565 schedule_work(&cw->work);
3568 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3569 struct kmem_cache *cachep)
3572 * We need to stop accounting when we kmalloc, because if the
3573 * corresponding kmalloc cache is not yet created, the first allocation
3574 * in __memcg_create_cache_enqueue will recurse.
3576 * However, it is better to enclose the whole function. Depending on
3577 * the debugging options enabled, INIT_WORK(), for instance, can
3578 * trigger an allocation. This too, will make us recurse. Because at
3579 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3580 * the safest choice is to do it like this, wrapping the whole function.
3582 memcg_stop_kmem_account();
3583 __memcg_create_cache_enqueue(memcg, cachep);
3584 memcg_resume_kmem_account();
3587 * Return the kmem_cache we're supposed to use for a slab allocation.
3588 * We try to use the current memcg's version of the cache.
3590 * If the cache does not exist yet, if we are the first user of it,
3591 * we either create it immediately, if possible, or create it asynchronously
3593 * In the latter case, we will let the current allocation go through with
3594 * the original cache.
3596 * Can't be called in interrupt context or from kernel threads.
3597 * This function needs to be called with rcu_read_lock() held.
3599 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3602 struct mem_cgroup *memcg;
3603 struct kmem_cache *memcg_cachep;
3605 VM_BUG_ON(!cachep->memcg_params);
3606 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3608 if (!current->mm || current->memcg_kmem_skip_account)
3612 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3614 if (!memcg_can_account_kmem(memcg))
3617 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3618 if (likely(memcg_cachep)) {
3619 cachep = memcg_cachep;
3623 /* The corresponding put will be done in the workqueue. */
3624 if (!css_tryget(&memcg->css))
3629 * If we are in a safe context (can wait, and not in interrupt
3630 * context), we could be be predictable and return right away.
3631 * This would guarantee that the allocation being performed
3632 * already belongs in the new cache.
3634 * However, there are some clashes that can arrive from locking.
3635 * For instance, because we acquire the slab_mutex while doing
3636 * kmem_cache_dup, this means no further allocation could happen
3637 * with the slab_mutex held.
3639 * Also, because cache creation issue get_online_cpus(), this
3640 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3641 * that ends up reversed during cpu hotplug. (cpuset allocates
3642 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3643 * better to defer everything.
3645 memcg_create_cache_enqueue(memcg, cachep);
3651 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3654 * We need to verify if the allocation against current->mm->owner's memcg is
3655 * possible for the given order. But the page is not allocated yet, so we'll
3656 * need a further commit step to do the final arrangements.
3658 * It is possible for the task to switch cgroups in this mean time, so at
3659 * commit time, we can't rely on task conversion any longer. We'll then use
3660 * the handle argument to return to the caller which cgroup we should commit
3661 * against. We could also return the memcg directly and avoid the pointer
3662 * passing, but a boolean return value gives better semantics considering
3663 * the compiled-out case as well.
3665 * Returning true means the allocation is possible.
3668 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3670 struct mem_cgroup *memcg;
3676 * Disabling accounting is only relevant for some specific memcg
3677 * internal allocations. Therefore we would initially not have such
3678 * check here, since direct calls to the page allocator that are marked
3679 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3680 * concerned with cache allocations, and by having this test at
3681 * memcg_kmem_get_cache, we are already able to relay the allocation to
3682 * the root cache and bypass the memcg cache altogether.
3684 * There is one exception, though: the SLUB allocator does not create
3685 * large order caches, but rather service large kmallocs directly from
3686 * the page allocator. Therefore, the following sequence when backed by
3687 * the SLUB allocator:
3689 * memcg_stop_kmem_account();
3690 * kmalloc(<large_number>)
3691 * memcg_resume_kmem_account();
3693 * would effectively ignore the fact that we should skip accounting,
3694 * since it will drive us directly to this function without passing
3695 * through the cache selector memcg_kmem_get_cache. Such large
3696 * allocations are extremely rare but can happen, for instance, for the
3697 * cache arrays. We bring this test here.
3699 if (!current->mm || current->memcg_kmem_skip_account)
3702 memcg = try_get_mem_cgroup_from_mm(current->mm);
3705 * very rare case described in mem_cgroup_from_task. Unfortunately there
3706 * isn't much we can do without complicating this too much, and it would
3707 * be gfp-dependent anyway. Just let it go
3709 if (unlikely(!memcg))
3712 if (!memcg_can_account_kmem(memcg)) {
3713 css_put(&memcg->css);
3717 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3721 css_put(&memcg->css);
3725 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3728 struct page_cgroup *pc;
3730 VM_BUG_ON(mem_cgroup_is_root(memcg));
3732 /* The page allocation failed. Revert */
3734 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3738 pc = lookup_page_cgroup(page);
3739 lock_page_cgroup(pc);
3740 pc->mem_cgroup = memcg;
3741 SetPageCgroupUsed(pc);
3742 unlock_page_cgroup(pc);
3745 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3747 struct mem_cgroup *memcg = NULL;
3748 struct page_cgroup *pc;
3751 pc = lookup_page_cgroup(page);
3753 * Fast unlocked return. Theoretically might have changed, have to
3754 * check again after locking.
3756 if (!PageCgroupUsed(pc))
3759 lock_page_cgroup(pc);
3760 if (PageCgroupUsed(pc)) {
3761 memcg = pc->mem_cgroup;
3762 ClearPageCgroupUsed(pc);
3764 unlock_page_cgroup(pc);
3767 * We trust that only if there is a memcg associated with the page, it
3768 * is a valid allocation
3773 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3774 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3777 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3780 #endif /* CONFIG_MEMCG_KMEM */
3782 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3784 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3786 * Because tail pages are not marked as "used", set it. We're under
3787 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3788 * charge/uncharge will be never happen and move_account() is done under
3789 * compound_lock(), so we don't have to take care of races.
3791 void mem_cgroup_split_huge_fixup(struct page *head)
3793 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3794 struct page_cgroup *pc;
3795 struct mem_cgroup *memcg;
3798 if (mem_cgroup_disabled())
3801 memcg = head_pc->mem_cgroup;
3802 for (i = 1; i < HPAGE_PMD_NR; i++) {
3804 pc->mem_cgroup = memcg;
3805 smp_wmb();/* see __commit_charge() */
3806 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3808 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3811 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3814 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3815 struct mem_cgroup *to,
3816 unsigned int nr_pages,
3817 enum mem_cgroup_stat_index idx)
3819 /* Update stat data for mem_cgroup */
3821 __this_cpu_sub(from->stat->count[idx], nr_pages);
3822 __this_cpu_add(to->stat->count[idx], nr_pages);
3827 * mem_cgroup_move_account - move account of the page
3829 * @nr_pages: number of regular pages (>1 for huge pages)
3830 * @pc: page_cgroup of the page.
3831 * @from: mem_cgroup which the page is moved from.
3832 * @to: mem_cgroup which the page is moved to. @from != @to.
3834 * The caller must confirm following.
3835 * - page is not on LRU (isolate_page() is useful.)
3836 * - compound_lock is held when nr_pages > 1
3838 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3841 static int mem_cgroup_move_account(struct page *page,
3842 unsigned int nr_pages,
3843 struct page_cgroup *pc,
3844 struct mem_cgroup *from,
3845 struct mem_cgroup *to)
3847 unsigned long flags;
3849 bool anon = PageAnon(page);
3851 VM_BUG_ON(from == to);
3852 VM_BUG_ON_PAGE(PageLRU(page), page);
3854 * The page is isolated from LRU. So, collapse function
3855 * will not handle this page. But page splitting can happen.
3856 * Do this check under compound_page_lock(). The caller should
3860 if (nr_pages > 1 && !PageTransHuge(page))
3863 lock_page_cgroup(pc);
3866 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3869 move_lock_mem_cgroup(from, &flags);
3871 if (!anon && page_mapped(page))
3872 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3873 MEM_CGROUP_STAT_FILE_MAPPED);
3875 if (PageWriteback(page))
3876 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3877 MEM_CGROUP_STAT_WRITEBACK);
3879 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3881 /* caller should have done css_get */
3882 pc->mem_cgroup = to;
3883 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3884 move_unlock_mem_cgroup(from, &flags);
3887 unlock_page_cgroup(pc);
3891 memcg_check_events(to, page);
3892 memcg_check_events(from, page);
3898 * mem_cgroup_move_parent - moves page to the parent group
3899 * @page: the page to move
3900 * @pc: page_cgroup of the page
3901 * @child: page's cgroup
3903 * move charges to its parent or the root cgroup if the group has no
3904 * parent (aka use_hierarchy==0).
3905 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3906 * mem_cgroup_move_account fails) the failure is always temporary and
3907 * it signals a race with a page removal/uncharge or migration. In the
3908 * first case the page is on the way out and it will vanish from the LRU
3909 * on the next attempt and the call should be retried later.
3910 * Isolation from the LRU fails only if page has been isolated from
3911 * the LRU since we looked at it and that usually means either global
3912 * reclaim or migration going on. The page will either get back to the
3914 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3915 * (!PageCgroupUsed) or moved to a different group. The page will
3916 * disappear in the next attempt.
3918 static int mem_cgroup_move_parent(struct page *page,
3919 struct page_cgroup *pc,
3920 struct mem_cgroup *child)
3922 struct mem_cgroup *parent;
3923 unsigned int nr_pages;
3924 unsigned long uninitialized_var(flags);
3927 VM_BUG_ON(mem_cgroup_is_root(child));
3930 if (!get_page_unless_zero(page))
3932 if (isolate_lru_page(page))
3935 nr_pages = hpage_nr_pages(page);
3937 parent = parent_mem_cgroup(child);
3939 * If no parent, move charges to root cgroup.
3942 parent = root_mem_cgroup;
3945 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3946 flags = compound_lock_irqsave(page);
3949 ret = mem_cgroup_move_account(page, nr_pages,
3952 __mem_cgroup_cancel_local_charge(child, nr_pages);
3955 compound_unlock_irqrestore(page, flags);
3956 putback_lru_page(page);
3964 * Charge the memory controller for page usage.
3966 * 0 if the charge was successful
3967 * < 0 if the cgroup is over its limit
3969 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3970 gfp_t gfp_mask, enum charge_type ctype)
3972 struct mem_cgroup *memcg = NULL;
3973 unsigned int nr_pages = 1;
3977 if (PageTransHuge(page)) {
3978 nr_pages <<= compound_order(page);
3979 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3981 * Never OOM-kill a process for a huge page. The
3982 * fault handler will fall back to regular pages.
3987 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3990 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3994 int mem_cgroup_newpage_charge(struct page *page,
3995 struct mm_struct *mm, gfp_t gfp_mask)
3997 if (mem_cgroup_disabled())
3999 VM_BUG_ON_PAGE(page_mapped(page), page);
4000 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4002 return mem_cgroup_charge_common(page, mm, gfp_mask,
4003 MEM_CGROUP_CHARGE_TYPE_ANON);
4007 * While swap-in, try_charge -> commit or cancel, the page is locked.
4008 * And when try_charge() successfully returns, one refcnt to memcg without
4009 * struct page_cgroup is acquired. This refcnt will be consumed by
4010 * "commit()" or removed by "cancel()"
4012 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
4015 struct mem_cgroup **memcgp)
4017 struct mem_cgroup *memcg;
4018 struct page_cgroup *pc;
4021 pc = lookup_page_cgroup(page);
4023 * Every swap fault against a single page tries to charge the
4024 * page, bail as early as possible. shmem_unuse() encounters
4025 * already charged pages, too. The USED bit is protected by
4026 * the page lock, which serializes swap cache removal, which
4027 * in turn serializes uncharging.
4029 if (PageCgroupUsed(pc))
4031 if (!do_swap_account)
4033 memcg = try_get_mem_cgroup_from_page(page);
4037 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4038 css_put(&memcg->css);
4043 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4049 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4050 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4053 if (mem_cgroup_disabled())
4056 * A racing thread's fault, or swapoff, may have already
4057 * updated the pte, and even removed page from swap cache: in
4058 * those cases unuse_pte()'s pte_same() test will fail; but
4059 * there's also a KSM case which does need to charge the page.
4061 if (!PageSwapCache(page)) {
4064 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4069 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4072 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4074 if (mem_cgroup_disabled())
4078 __mem_cgroup_cancel_charge(memcg, 1);
4082 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4083 enum charge_type ctype)
4085 if (mem_cgroup_disabled())
4090 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4092 * Now swap is on-memory. This means this page may be
4093 * counted both as mem and swap....double count.
4094 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4095 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4096 * may call delete_from_swap_cache() before reach here.
4098 if (do_swap_account && PageSwapCache(page)) {
4099 swp_entry_t ent = {.val = page_private(page)};
4100 mem_cgroup_uncharge_swap(ent);
4104 void mem_cgroup_commit_charge_swapin(struct page *page,
4105 struct mem_cgroup *memcg)
4107 __mem_cgroup_commit_charge_swapin(page, memcg,
4108 MEM_CGROUP_CHARGE_TYPE_ANON);
4111 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4114 struct mem_cgroup *memcg = NULL;
4115 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4118 if (mem_cgroup_disabled())
4120 if (PageCompound(page))
4123 if (!PageSwapCache(page))
4124 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4125 else { /* page is swapcache/shmem */
4126 ret = __mem_cgroup_try_charge_swapin(mm, page,
4129 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4134 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4135 unsigned int nr_pages,
4136 const enum charge_type ctype)
4138 struct memcg_batch_info *batch = NULL;
4139 bool uncharge_memsw = true;
4141 /* If swapout, usage of swap doesn't decrease */
4142 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4143 uncharge_memsw = false;
4145 batch = ¤t->memcg_batch;
4147 * In usual, we do css_get() when we remember memcg pointer.
4148 * But in this case, we keep res->usage until end of a series of
4149 * uncharges. Then, it's ok to ignore memcg's refcnt.
4152 batch->memcg = memcg;
4154 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4155 * In those cases, all pages freed continuously can be expected to be in
4156 * the same cgroup and we have chance to coalesce uncharges.
4157 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4158 * because we want to do uncharge as soon as possible.
4161 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4162 goto direct_uncharge;
4165 goto direct_uncharge;
4168 * In typical case, batch->memcg == mem. This means we can
4169 * merge a series of uncharges to an uncharge of res_counter.
4170 * If not, we uncharge res_counter ony by one.
4172 if (batch->memcg != memcg)
4173 goto direct_uncharge;
4174 /* remember freed charge and uncharge it later */
4177 batch->memsw_nr_pages++;
4180 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4182 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4183 if (unlikely(batch->memcg != memcg))
4184 memcg_oom_recover(memcg);
4188 * uncharge if !page_mapped(page)
4190 static struct mem_cgroup *
4191 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4194 struct mem_cgroup *memcg = NULL;
4195 unsigned int nr_pages = 1;
4196 struct page_cgroup *pc;
4199 if (mem_cgroup_disabled())
4202 if (PageTransHuge(page)) {
4203 nr_pages <<= compound_order(page);
4204 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
4207 * Check if our page_cgroup is valid
4209 pc = lookup_page_cgroup(page);
4210 if (unlikely(!PageCgroupUsed(pc)))
4213 lock_page_cgroup(pc);
4215 memcg = pc->mem_cgroup;
4217 if (!PageCgroupUsed(pc))
4220 anon = PageAnon(page);
4223 case MEM_CGROUP_CHARGE_TYPE_ANON:
4225 * Generally PageAnon tells if it's the anon statistics to be
4226 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4227 * used before page reached the stage of being marked PageAnon.
4231 case MEM_CGROUP_CHARGE_TYPE_DROP:
4232 /* See mem_cgroup_prepare_migration() */
4233 if (page_mapped(page))
4236 * Pages under migration may not be uncharged. But
4237 * end_migration() /must/ be the one uncharging the
4238 * unused post-migration page and so it has to call
4239 * here with the migration bit still set. See the
4240 * res_counter handling below.
4242 if (!end_migration && PageCgroupMigration(pc))
4245 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4246 if (!PageAnon(page)) { /* Shared memory */
4247 if (page->mapping && !page_is_file_cache(page))
4249 } else if (page_mapped(page)) /* Anon */
4256 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4258 ClearPageCgroupUsed(pc);
4260 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4261 * freed from LRU. This is safe because uncharged page is expected not
4262 * to be reused (freed soon). Exception is SwapCache, it's handled by
4263 * special functions.
4266 unlock_page_cgroup(pc);
4268 * even after unlock, we have memcg->res.usage here and this memcg
4269 * will never be freed, so it's safe to call css_get().
4271 memcg_check_events(memcg, page);
4272 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4273 mem_cgroup_swap_statistics(memcg, true);
4274 css_get(&memcg->css);
4277 * Migration does not charge the res_counter for the
4278 * replacement page, so leave it alone when phasing out the
4279 * page that is unused after the migration.
4281 if (!end_migration && !mem_cgroup_is_root(memcg))
4282 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4287 unlock_page_cgroup(pc);
4291 void mem_cgroup_uncharge_page(struct page *page)
4294 if (page_mapped(page))
4296 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4298 * If the page is in swap cache, uncharge should be deferred
4299 * to the swap path, which also properly accounts swap usage
4300 * and handles memcg lifetime.
4302 * Note that this check is not stable and reclaim may add the
4303 * page to swap cache at any time after this. However, if the
4304 * page is not in swap cache by the time page->mapcount hits
4305 * 0, there won't be any page table references to the swap
4306 * slot, and reclaim will free it and not actually write the
4309 if (PageSwapCache(page))
4311 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4314 void mem_cgroup_uncharge_cache_page(struct page *page)
4316 VM_BUG_ON_PAGE(page_mapped(page), page);
4317 VM_BUG_ON_PAGE(page->mapping, page);
4318 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4322 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4323 * In that cases, pages are freed continuously and we can expect pages
4324 * are in the same memcg. All these calls itself limits the number of
4325 * pages freed at once, then uncharge_start/end() is called properly.
4326 * This may be called prural(2) times in a context,
4329 void mem_cgroup_uncharge_start(void)
4331 current->memcg_batch.do_batch++;
4332 /* We can do nest. */
4333 if (current->memcg_batch.do_batch == 1) {
4334 current->memcg_batch.memcg = NULL;
4335 current->memcg_batch.nr_pages = 0;
4336 current->memcg_batch.memsw_nr_pages = 0;
4340 void mem_cgroup_uncharge_end(void)
4342 struct memcg_batch_info *batch = ¤t->memcg_batch;
4344 if (!batch->do_batch)
4348 if (batch->do_batch) /* If stacked, do nothing. */
4354 * This "batch->memcg" is valid without any css_get/put etc...
4355 * bacause we hide charges behind us.
4357 if (batch->nr_pages)
4358 res_counter_uncharge(&batch->memcg->res,
4359 batch->nr_pages * PAGE_SIZE);
4360 if (batch->memsw_nr_pages)
4361 res_counter_uncharge(&batch->memcg->memsw,
4362 batch->memsw_nr_pages * PAGE_SIZE);
4363 memcg_oom_recover(batch->memcg);
4364 /* forget this pointer (for sanity check) */
4365 batch->memcg = NULL;
4370 * called after __delete_from_swap_cache() and drop "page" account.
4371 * memcg information is recorded to swap_cgroup of "ent"
4374 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4376 struct mem_cgroup *memcg;
4377 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4379 if (!swapout) /* this was a swap cache but the swap is unused ! */
4380 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4382 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4385 * record memcg information, if swapout && memcg != NULL,
4386 * css_get() was called in uncharge().
4388 if (do_swap_account && swapout && memcg)
4389 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4393 #ifdef CONFIG_MEMCG_SWAP
4395 * called from swap_entry_free(). remove record in swap_cgroup and
4396 * uncharge "memsw" account.
4398 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4400 struct mem_cgroup *memcg;
4403 if (!do_swap_account)
4406 id = swap_cgroup_record(ent, 0);
4408 memcg = mem_cgroup_lookup(id);
4411 * We uncharge this because swap is freed.
4412 * This memcg can be obsolete one. We avoid calling css_tryget
4414 if (!mem_cgroup_is_root(memcg))
4415 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4416 mem_cgroup_swap_statistics(memcg, false);
4417 css_put(&memcg->css);
4423 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4424 * @entry: swap entry to be moved
4425 * @from: mem_cgroup which the entry is moved from
4426 * @to: mem_cgroup which the entry is moved to
4428 * It succeeds only when the swap_cgroup's record for this entry is the same
4429 * as the mem_cgroup's id of @from.
4431 * Returns 0 on success, -EINVAL on failure.
4433 * The caller must have charged to @to, IOW, called res_counter_charge() about
4434 * both res and memsw, and called css_get().
4436 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4437 struct mem_cgroup *from, struct mem_cgroup *to)
4439 unsigned short old_id, new_id;
4441 old_id = mem_cgroup_id(from);
4442 new_id = mem_cgroup_id(to);
4444 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4445 mem_cgroup_swap_statistics(from, false);
4446 mem_cgroup_swap_statistics(to, true);
4448 * This function is only called from task migration context now.
4449 * It postpones res_counter and refcount handling till the end
4450 * of task migration(mem_cgroup_clear_mc()) for performance
4451 * improvement. But we cannot postpone css_get(to) because if
4452 * the process that has been moved to @to does swap-in, the
4453 * refcount of @to might be decreased to 0.
4455 * We are in attach() phase, so the cgroup is guaranteed to be
4456 * alive, so we can just call css_get().
4464 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4465 struct mem_cgroup *from, struct mem_cgroup *to)
4472 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4475 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4476 struct mem_cgroup **memcgp)
4478 struct mem_cgroup *memcg = NULL;
4479 unsigned int nr_pages = 1;
4480 struct page_cgroup *pc;
4481 enum charge_type ctype;
4485 if (mem_cgroup_disabled())
4488 if (PageTransHuge(page))
4489 nr_pages <<= compound_order(page);
4491 pc = lookup_page_cgroup(page);
4492 lock_page_cgroup(pc);
4493 if (PageCgroupUsed(pc)) {
4494 memcg = pc->mem_cgroup;
4495 css_get(&memcg->css);
4497 * At migrating an anonymous page, its mapcount goes down
4498 * to 0 and uncharge() will be called. But, even if it's fully
4499 * unmapped, migration may fail and this page has to be
4500 * charged again. We set MIGRATION flag here and delay uncharge
4501 * until end_migration() is called
4503 * Corner Case Thinking
4505 * When the old page was mapped as Anon and it's unmap-and-freed
4506 * while migration was ongoing.
4507 * If unmap finds the old page, uncharge() of it will be delayed
4508 * until end_migration(). If unmap finds a new page, it's
4509 * uncharged when it make mapcount to be 1->0. If unmap code
4510 * finds swap_migration_entry, the new page will not be mapped
4511 * and end_migration() will find it(mapcount==0).
4514 * When the old page was mapped but migraion fails, the kernel
4515 * remaps it. A charge for it is kept by MIGRATION flag even
4516 * if mapcount goes down to 0. We can do remap successfully
4517 * without charging it again.
4520 * The "old" page is under lock_page() until the end of
4521 * migration, so, the old page itself will not be swapped-out.
4522 * If the new page is swapped out before end_migraton, our
4523 * hook to usual swap-out path will catch the event.
4526 SetPageCgroupMigration(pc);
4528 unlock_page_cgroup(pc);
4530 * If the page is not charged at this point,
4538 * We charge new page before it's used/mapped. So, even if unlock_page()
4539 * is called before end_migration, we can catch all events on this new
4540 * page. In the case new page is migrated but not remapped, new page's
4541 * mapcount will be finally 0 and we call uncharge in end_migration().
4544 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4546 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4548 * The page is committed to the memcg, but it's not actually
4549 * charged to the res_counter since we plan on replacing the
4550 * old one and only one page is going to be left afterwards.
4552 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4555 /* remove redundant charge if migration failed*/
4556 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4557 struct page *oldpage, struct page *newpage, bool migration_ok)
4559 struct page *used, *unused;
4560 struct page_cgroup *pc;
4566 if (!migration_ok) {
4573 anon = PageAnon(used);
4574 __mem_cgroup_uncharge_common(unused,
4575 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4576 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4578 css_put(&memcg->css);
4580 * We disallowed uncharge of pages under migration because mapcount
4581 * of the page goes down to zero, temporarly.
4582 * Clear the flag and check the page should be charged.
4584 pc = lookup_page_cgroup(oldpage);
4585 lock_page_cgroup(pc);
4586 ClearPageCgroupMigration(pc);
4587 unlock_page_cgroup(pc);
4590 * If a page is a file cache, radix-tree replacement is very atomic
4591 * and we can skip this check. When it was an Anon page, its mapcount
4592 * goes down to 0. But because we added MIGRATION flage, it's not
4593 * uncharged yet. There are several case but page->mapcount check
4594 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4595 * check. (see prepare_charge() also)
4598 mem_cgroup_uncharge_page(used);
4602 * At replace page cache, newpage is not under any memcg but it's on
4603 * LRU. So, this function doesn't touch res_counter but handles LRU
4604 * in correct way. Both pages are locked so we cannot race with uncharge.
4606 void mem_cgroup_replace_page_cache(struct page *oldpage,
4607 struct page *newpage)
4609 struct mem_cgroup *memcg = NULL;
4610 struct page_cgroup *pc;
4611 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4613 if (mem_cgroup_disabled())
4616 pc = lookup_page_cgroup(oldpage);
4617 /* fix accounting on old pages */
4618 lock_page_cgroup(pc);
4619 if (PageCgroupUsed(pc)) {
4620 memcg = pc->mem_cgroup;
4621 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4622 ClearPageCgroupUsed(pc);
4624 unlock_page_cgroup(pc);
4627 * When called from shmem_replace_page(), in some cases the
4628 * oldpage has already been charged, and in some cases not.
4633 * Even if newpage->mapping was NULL before starting replacement,
4634 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4635 * LRU while we overwrite pc->mem_cgroup.
4637 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4640 #ifdef CONFIG_DEBUG_VM
4641 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4643 struct page_cgroup *pc;
4645 pc = lookup_page_cgroup(page);
4647 * Can be NULL while feeding pages into the page allocator for
4648 * the first time, i.e. during boot or memory hotplug;
4649 * or when mem_cgroup_disabled().
4651 if (likely(pc) && PageCgroupUsed(pc))
4656 bool mem_cgroup_bad_page_check(struct page *page)
4658 if (mem_cgroup_disabled())
4661 return lookup_page_cgroup_used(page) != NULL;
4664 void mem_cgroup_print_bad_page(struct page *page)
4666 struct page_cgroup *pc;
4668 pc = lookup_page_cgroup_used(page);
4670 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4671 pc, pc->flags, pc->mem_cgroup);
4676 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4677 unsigned long long val)
4680 u64 memswlimit, memlimit;
4682 int children = mem_cgroup_count_children(memcg);
4683 u64 curusage, oldusage;
4687 * For keeping hierarchical_reclaim simple, how long we should retry
4688 * is depends on callers. We set our retry-count to be function
4689 * of # of children which we should visit in this loop.
4691 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4693 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4696 while (retry_count) {
4697 if (signal_pending(current)) {
4702 * Rather than hide all in some function, I do this in
4703 * open coded manner. You see what this really does.
4704 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4706 mutex_lock(&set_limit_mutex);
4707 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4708 if (memswlimit < val) {
4710 mutex_unlock(&set_limit_mutex);
4714 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4718 ret = res_counter_set_limit(&memcg->res, val);
4720 if (memswlimit == val)
4721 memcg->memsw_is_minimum = true;
4723 memcg->memsw_is_minimum = false;
4725 mutex_unlock(&set_limit_mutex);
4730 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4731 MEM_CGROUP_RECLAIM_SHRINK);
4732 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4733 /* Usage is reduced ? */
4734 if (curusage >= oldusage)
4737 oldusage = curusage;
4739 if (!ret && enlarge)
4740 memcg_oom_recover(memcg);
4745 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4746 unsigned long long val)
4749 u64 memlimit, memswlimit, oldusage, curusage;
4750 int children = mem_cgroup_count_children(memcg);
4754 /* see mem_cgroup_resize_res_limit */
4755 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4756 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4757 while (retry_count) {
4758 if (signal_pending(current)) {
4763 * Rather than hide all in some function, I do this in
4764 * open coded manner. You see what this really does.
4765 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4767 mutex_lock(&set_limit_mutex);
4768 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4769 if (memlimit > val) {
4771 mutex_unlock(&set_limit_mutex);
4774 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4775 if (memswlimit < val)
4777 ret = res_counter_set_limit(&memcg->memsw, val);
4779 if (memlimit == val)
4780 memcg->memsw_is_minimum = true;
4782 memcg->memsw_is_minimum = false;
4784 mutex_unlock(&set_limit_mutex);
4789 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4790 MEM_CGROUP_RECLAIM_NOSWAP |
4791 MEM_CGROUP_RECLAIM_SHRINK);
4792 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4793 /* Usage is reduced ? */
4794 if (curusage >= oldusage)
4797 oldusage = curusage;
4799 if (!ret && enlarge)
4800 memcg_oom_recover(memcg);
4804 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4806 unsigned long *total_scanned)
4808 unsigned long nr_reclaimed = 0;
4809 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4810 unsigned long reclaimed;
4812 struct mem_cgroup_tree_per_zone *mctz;
4813 unsigned long long excess;
4814 unsigned long nr_scanned;
4819 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4821 * This loop can run a while, specially if mem_cgroup's continuously
4822 * keep exceeding their soft limit and putting the system under
4829 mz = mem_cgroup_largest_soft_limit_node(mctz);
4834 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4835 gfp_mask, &nr_scanned);
4836 nr_reclaimed += reclaimed;
4837 *total_scanned += nr_scanned;
4838 spin_lock(&mctz->lock);
4841 * If we failed to reclaim anything from this memory cgroup
4842 * it is time to move on to the next cgroup
4848 * Loop until we find yet another one.
4850 * By the time we get the soft_limit lock
4851 * again, someone might have aded the
4852 * group back on the RB tree. Iterate to
4853 * make sure we get a different mem.
4854 * mem_cgroup_largest_soft_limit_node returns
4855 * NULL if no other cgroup is present on
4859 __mem_cgroup_largest_soft_limit_node(mctz);
4861 css_put(&next_mz->memcg->css);
4862 else /* next_mz == NULL or other memcg */
4866 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4867 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4869 * One school of thought says that we should not add
4870 * back the node to the tree if reclaim returns 0.
4871 * But our reclaim could return 0, simply because due
4872 * to priority we are exposing a smaller subset of
4873 * memory to reclaim from. Consider this as a longer
4876 /* If excess == 0, no tree ops */
4877 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4878 spin_unlock(&mctz->lock);
4879 css_put(&mz->memcg->css);
4882 * Could not reclaim anything and there are no more
4883 * mem cgroups to try or we seem to be looping without
4884 * reclaiming anything.
4886 if (!nr_reclaimed &&
4888 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4890 } while (!nr_reclaimed);
4892 css_put(&next_mz->memcg->css);
4893 return nr_reclaimed;
4897 * mem_cgroup_force_empty_list - clears LRU of a group
4898 * @memcg: group to clear
4901 * @lru: lru to to clear
4903 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4904 * reclaim the pages page themselves - pages are moved to the parent (or root)
4907 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4908 int node, int zid, enum lru_list lru)
4910 struct lruvec *lruvec;
4911 unsigned long flags;
4912 struct list_head *list;
4916 zone = &NODE_DATA(node)->node_zones[zid];
4917 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4918 list = &lruvec->lists[lru];
4922 struct page_cgroup *pc;
4925 spin_lock_irqsave(&zone->lru_lock, flags);
4926 if (list_empty(list)) {
4927 spin_unlock_irqrestore(&zone->lru_lock, flags);
4930 page = list_entry(list->prev, struct page, lru);
4932 list_move(&page->lru, list);
4934 spin_unlock_irqrestore(&zone->lru_lock, flags);
4937 spin_unlock_irqrestore(&zone->lru_lock, flags);
4939 pc = lookup_page_cgroup(page);
4941 if (mem_cgroup_move_parent(page, pc, memcg)) {
4942 /* found lock contention or "pc" is obsolete. */
4947 } while (!list_empty(list));
4951 * make mem_cgroup's charge to be 0 if there is no task by moving
4952 * all the charges and pages to the parent.
4953 * This enables deleting this mem_cgroup.
4955 * Caller is responsible for holding css reference on the memcg.
4957 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4963 /* This is for making all *used* pages to be on LRU. */
4964 lru_add_drain_all();
4965 drain_all_stock_sync(memcg);
4966 mem_cgroup_start_move(memcg);
4967 for_each_node_state(node, N_MEMORY) {
4968 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4971 mem_cgroup_force_empty_list(memcg,
4976 mem_cgroup_end_move(memcg);
4977 memcg_oom_recover(memcg);
4981 * Kernel memory may not necessarily be trackable to a specific
4982 * process. So they are not migrated, and therefore we can't
4983 * expect their value to drop to 0 here.
4984 * Having res filled up with kmem only is enough.
4986 * This is a safety check because mem_cgroup_force_empty_list
4987 * could have raced with mem_cgroup_replace_page_cache callers
4988 * so the lru seemed empty but the page could have been added
4989 * right after the check. RES_USAGE should be safe as we always
4990 * charge before adding to the LRU.
4992 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4993 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4994 } while (usage > 0);
4997 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4999 lockdep_assert_held(&memcg_create_mutex);
5001 * The lock does not prevent addition or deletion to the list
5002 * of children, but it prevents a new child from being
5003 * initialized based on this parent in css_online(), so it's
5004 * enough to decide whether hierarchically inherited
5005 * attributes can still be changed or not.
5007 return memcg->use_hierarchy &&
5008 !list_empty(&memcg->css.cgroup->children);
5012 * Reclaims as many pages from the given memcg as possible and moves
5013 * the rest to the parent.
5015 * Caller is responsible for holding css reference for memcg.
5017 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
5019 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
5020 struct cgroup *cgrp = memcg->css.cgroup;
5022 /* returns EBUSY if there is a task or if we come here twice. */
5023 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
5026 /* we call try-to-free pages for make this cgroup empty */
5027 lru_add_drain_all();
5028 /* try to free all pages in this cgroup */
5029 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
5032 if (signal_pending(current))
5035 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5039 /* maybe some writeback is necessary */
5040 congestion_wait(BLK_RW_ASYNC, HZ/10);
5045 mem_cgroup_reparent_charges(memcg);
5050 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
5053 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5055 if (mem_cgroup_is_root(memcg))
5057 return mem_cgroup_force_empty(memcg);
5060 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5063 return mem_cgroup_from_css(css)->use_hierarchy;
5066 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5067 struct cftype *cft, u64 val)
5070 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5071 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5073 mutex_lock(&memcg_create_mutex);
5075 if (memcg->use_hierarchy == val)
5079 * If parent's use_hierarchy is set, we can't make any modifications
5080 * in the child subtrees. If it is unset, then the change can
5081 * occur, provided the current cgroup has no children.
5083 * For the root cgroup, parent_mem is NULL, we allow value to be
5084 * set if there are no children.
5086 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5087 (val == 1 || val == 0)) {
5088 if (list_empty(&memcg->css.cgroup->children))
5089 memcg->use_hierarchy = val;
5096 mutex_unlock(&memcg_create_mutex);
5102 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5103 enum mem_cgroup_stat_index idx)
5105 struct mem_cgroup *iter;
5108 /* Per-cpu values can be negative, use a signed accumulator */
5109 for_each_mem_cgroup_tree(iter, memcg)
5110 val += mem_cgroup_read_stat(iter, idx);
5112 if (val < 0) /* race ? */
5117 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5121 if (!mem_cgroup_is_root(memcg)) {
5123 return res_counter_read_u64(&memcg->res, RES_USAGE);
5125 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5129 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5130 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5132 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5133 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5136 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5138 return val << PAGE_SHIFT;
5141 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
5144 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5149 type = MEMFILE_TYPE(cft->private);
5150 name = MEMFILE_ATTR(cft->private);
5154 if (name == RES_USAGE)
5155 val = mem_cgroup_usage(memcg, false);
5157 val = res_counter_read_u64(&memcg->res, name);
5160 if (name == RES_USAGE)
5161 val = mem_cgroup_usage(memcg, true);
5163 val = res_counter_read_u64(&memcg->memsw, name);
5166 val = res_counter_read_u64(&memcg->kmem, name);
5175 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
5178 #ifdef CONFIG_MEMCG_KMEM
5179 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5181 * For simplicity, we won't allow this to be disabled. It also can't
5182 * be changed if the cgroup has children already, or if tasks had
5185 * If tasks join before we set the limit, a person looking at
5186 * kmem.usage_in_bytes will have no way to determine when it took
5187 * place, which makes the value quite meaningless.
5189 * After it first became limited, changes in the value of the limit are
5190 * of course permitted.
5192 mutex_lock(&memcg_create_mutex);
5193 mutex_lock(&set_limit_mutex);
5194 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
5195 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
5199 ret = res_counter_set_limit(&memcg->kmem, val);
5202 ret = memcg_update_cache_sizes(memcg);
5204 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
5207 static_key_slow_inc(&memcg_kmem_enabled_key);
5209 * setting the active bit after the inc will guarantee no one
5210 * starts accounting before all call sites are patched
5212 memcg_kmem_set_active(memcg);
5214 ret = res_counter_set_limit(&memcg->kmem, val);
5216 mutex_unlock(&set_limit_mutex);
5217 mutex_unlock(&memcg_create_mutex);
5222 #ifdef CONFIG_MEMCG_KMEM
5223 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5226 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5230 memcg->kmem_account_flags = parent->kmem_account_flags;
5232 * When that happen, we need to disable the static branch only on those
5233 * memcgs that enabled it. To achieve this, we would be forced to
5234 * complicate the code by keeping track of which memcgs were the ones
5235 * that actually enabled limits, and which ones got it from its
5238 * It is a lot simpler just to do static_key_slow_inc() on every child
5239 * that is accounted.
5241 if (!memcg_kmem_is_active(memcg))
5245 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5246 * memcg is active already. If the later initialization fails then the
5247 * cgroup core triggers the cleanup so we do not have to do it here.
5249 static_key_slow_inc(&memcg_kmem_enabled_key);
5251 mutex_lock(&set_limit_mutex);
5252 memcg_stop_kmem_account();
5253 ret = memcg_update_cache_sizes(memcg);
5254 memcg_resume_kmem_account();
5255 mutex_unlock(&set_limit_mutex);
5259 #endif /* CONFIG_MEMCG_KMEM */
5262 * The user of this function is...
5265 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5268 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5271 unsigned long long val;
5274 type = MEMFILE_TYPE(cft->private);
5275 name = MEMFILE_ATTR(cft->private);
5279 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5283 /* This function does all necessary parse...reuse it */
5284 ret = res_counter_memparse_write_strategy(buffer, &val);
5288 ret = mem_cgroup_resize_limit(memcg, val);
5289 else if (type == _MEMSWAP)
5290 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5291 else if (type == _KMEM)
5292 ret = memcg_update_kmem_limit(css, val);
5296 case RES_SOFT_LIMIT:
5297 ret = res_counter_memparse_write_strategy(buffer, &val);
5301 * For memsw, soft limits are hard to implement in terms
5302 * of semantics, for now, we support soft limits for
5303 * control without swap
5306 ret = res_counter_set_soft_limit(&memcg->res, val);
5311 ret = -EINVAL; /* should be BUG() ? */
5317 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5318 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5320 unsigned long long min_limit, min_memsw_limit, tmp;
5322 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5323 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5324 if (!memcg->use_hierarchy)
5327 while (css_parent(&memcg->css)) {
5328 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5329 if (!memcg->use_hierarchy)
5331 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5332 min_limit = min(min_limit, tmp);
5333 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5334 min_memsw_limit = min(min_memsw_limit, tmp);
5337 *mem_limit = min_limit;
5338 *memsw_limit = min_memsw_limit;
5341 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5343 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5347 type = MEMFILE_TYPE(event);
5348 name = MEMFILE_ATTR(event);
5353 res_counter_reset_max(&memcg->res);
5354 else if (type == _MEMSWAP)
5355 res_counter_reset_max(&memcg->memsw);
5356 else if (type == _KMEM)
5357 res_counter_reset_max(&memcg->kmem);
5363 res_counter_reset_failcnt(&memcg->res);
5364 else if (type == _MEMSWAP)
5365 res_counter_reset_failcnt(&memcg->memsw);
5366 else if (type == _KMEM)
5367 res_counter_reset_failcnt(&memcg->kmem);
5376 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5379 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5383 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5384 struct cftype *cft, u64 val)
5386 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5388 if (val >= (1 << NR_MOVE_TYPE))
5392 * No kind of locking is needed in here, because ->can_attach() will
5393 * check this value once in the beginning of the process, and then carry
5394 * on with stale data. This means that changes to this value will only
5395 * affect task migrations starting after the change.
5397 memcg->move_charge_at_immigrate = val;
5401 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5402 struct cftype *cft, u64 val)
5409 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5413 unsigned int lru_mask;
5416 static const struct numa_stat stats[] = {
5417 { "total", LRU_ALL },
5418 { "file", LRU_ALL_FILE },
5419 { "anon", LRU_ALL_ANON },
5420 { "unevictable", BIT(LRU_UNEVICTABLE) },
5422 const struct numa_stat *stat;
5425 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5427 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5428 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5429 seq_printf(m, "%s=%lu", stat->name, nr);
5430 for_each_node_state(nid, N_MEMORY) {
5431 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5433 seq_printf(m, " N%d=%lu", nid, nr);
5438 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5439 struct mem_cgroup *iter;
5442 for_each_mem_cgroup_tree(iter, memcg)
5443 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5444 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5445 for_each_node_state(nid, N_MEMORY) {
5447 for_each_mem_cgroup_tree(iter, memcg)
5448 nr += mem_cgroup_node_nr_lru_pages(
5449 iter, nid, stat->lru_mask);
5450 seq_printf(m, " N%d=%lu", nid, nr);
5457 #endif /* CONFIG_NUMA */
5459 static inline void mem_cgroup_lru_names_not_uptodate(void)
5461 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5464 static int memcg_stat_show(struct seq_file *m, void *v)
5466 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5467 struct mem_cgroup *mi;
5470 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5471 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5473 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5474 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5477 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5478 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5479 mem_cgroup_read_events(memcg, i));
5481 for (i = 0; i < NR_LRU_LISTS; i++)
5482 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5483 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5485 /* Hierarchical information */
5487 unsigned long long limit, memsw_limit;
5488 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5489 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5490 if (do_swap_account)
5491 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5495 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5498 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5500 for_each_mem_cgroup_tree(mi, memcg)
5501 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5502 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5505 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5506 unsigned long long val = 0;
5508 for_each_mem_cgroup_tree(mi, memcg)
5509 val += mem_cgroup_read_events(mi, i);
5510 seq_printf(m, "total_%s %llu\n",
5511 mem_cgroup_events_names[i], val);
5514 for (i = 0; i < NR_LRU_LISTS; i++) {
5515 unsigned long long val = 0;
5517 for_each_mem_cgroup_tree(mi, memcg)
5518 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5519 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5522 #ifdef CONFIG_DEBUG_VM
5525 struct mem_cgroup_per_zone *mz;
5526 struct zone_reclaim_stat *rstat;
5527 unsigned long recent_rotated[2] = {0, 0};
5528 unsigned long recent_scanned[2] = {0, 0};
5530 for_each_online_node(nid)
5531 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5532 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5533 rstat = &mz->lruvec.reclaim_stat;
5535 recent_rotated[0] += rstat->recent_rotated[0];
5536 recent_rotated[1] += rstat->recent_rotated[1];
5537 recent_scanned[0] += rstat->recent_scanned[0];
5538 recent_scanned[1] += rstat->recent_scanned[1];
5540 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5541 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5542 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5543 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5550 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5553 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5555 return mem_cgroup_swappiness(memcg);
5558 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5559 struct cftype *cft, u64 val)
5561 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5562 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5564 if (val > 100 || !parent)
5567 mutex_lock(&memcg_create_mutex);
5569 /* If under hierarchy, only empty-root can set this value */
5570 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5571 mutex_unlock(&memcg_create_mutex);
5575 memcg->swappiness = val;
5577 mutex_unlock(&memcg_create_mutex);
5582 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5584 struct mem_cgroup_threshold_ary *t;
5590 t = rcu_dereference(memcg->thresholds.primary);
5592 t = rcu_dereference(memcg->memsw_thresholds.primary);
5597 usage = mem_cgroup_usage(memcg, swap);
5600 * current_threshold points to threshold just below or equal to usage.
5601 * If it's not true, a threshold was crossed after last
5602 * call of __mem_cgroup_threshold().
5604 i = t->current_threshold;
5607 * Iterate backward over array of thresholds starting from
5608 * current_threshold and check if a threshold is crossed.
5609 * If none of thresholds below usage is crossed, we read
5610 * only one element of the array here.
5612 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5613 eventfd_signal(t->entries[i].eventfd, 1);
5615 /* i = current_threshold + 1 */
5619 * Iterate forward over array of thresholds starting from
5620 * current_threshold+1 and check if a threshold is crossed.
5621 * If none of thresholds above usage is crossed, we read
5622 * only one element of the array here.
5624 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5625 eventfd_signal(t->entries[i].eventfd, 1);
5627 /* Update current_threshold */
5628 t->current_threshold = i - 1;
5633 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5636 __mem_cgroup_threshold(memcg, false);
5637 if (do_swap_account)
5638 __mem_cgroup_threshold(memcg, true);
5640 memcg = parent_mem_cgroup(memcg);
5644 static int compare_thresholds(const void *a, const void *b)
5646 const struct mem_cgroup_threshold *_a = a;
5647 const struct mem_cgroup_threshold *_b = b;
5649 if (_a->threshold > _b->threshold)
5652 if (_a->threshold < _b->threshold)
5658 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5660 struct mem_cgroup_eventfd_list *ev;
5662 list_for_each_entry(ev, &memcg->oom_notify, list)
5663 eventfd_signal(ev->eventfd, 1);
5667 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5669 struct mem_cgroup *iter;
5671 for_each_mem_cgroup_tree(iter, memcg)
5672 mem_cgroup_oom_notify_cb(iter);
5675 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5676 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5678 struct mem_cgroup_thresholds *thresholds;
5679 struct mem_cgroup_threshold_ary *new;
5680 u64 threshold, usage;
5683 ret = res_counter_memparse_write_strategy(args, &threshold);
5687 mutex_lock(&memcg->thresholds_lock);
5690 thresholds = &memcg->thresholds;
5691 else if (type == _MEMSWAP)
5692 thresholds = &memcg->memsw_thresholds;
5696 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5698 /* Check if a threshold crossed before adding a new one */
5699 if (thresholds->primary)
5700 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5702 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5704 /* Allocate memory for new array of thresholds */
5705 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5713 /* Copy thresholds (if any) to new array */
5714 if (thresholds->primary) {
5715 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5716 sizeof(struct mem_cgroup_threshold));
5719 /* Add new threshold */
5720 new->entries[size - 1].eventfd = eventfd;
5721 new->entries[size - 1].threshold = threshold;
5723 /* Sort thresholds. Registering of new threshold isn't time-critical */
5724 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5725 compare_thresholds, NULL);
5727 /* Find current threshold */
5728 new->current_threshold = -1;
5729 for (i = 0; i < size; i++) {
5730 if (new->entries[i].threshold <= usage) {
5732 * new->current_threshold will not be used until
5733 * rcu_assign_pointer(), so it's safe to increment
5736 ++new->current_threshold;
5741 /* Free old spare buffer and save old primary buffer as spare */
5742 kfree(thresholds->spare);
5743 thresholds->spare = thresholds->primary;
5745 rcu_assign_pointer(thresholds->primary, new);
5747 /* To be sure that nobody uses thresholds */
5751 mutex_unlock(&memcg->thresholds_lock);
5756 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5757 struct eventfd_ctx *eventfd, const char *args)
5759 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5762 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5763 struct eventfd_ctx *eventfd, const char *args)
5765 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5768 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5769 struct eventfd_ctx *eventfd, enum res_type type)
5771 struct mem_cgroup_thresholds *thresholds;
5772 struct mem_cgroup_threshold_ary *new;
5776 mutex_lock(&memcg->thresholds_lock);
5778 thresholds = &memcg->thresholds;
5779 else if (type == _MEMSWAP)
5780 thresholds = &memcg->memsw_thresholds;
5784 if (!thresholds->primary)
5787 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5789 /* Check if a threshold crossed before removing */
5790 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5792 /* Calculate new number of threshold */
5794 for (i = 0; i < thresholds->primary->size; i++) {
5795 if (thresholds->primary->entries[i].eventfd != eventfd)
5799 new = thresholds->spare;
5801 /* Set thresholds array to NULL if we don't have thresholds */
5810 /* Copy thresholds and find current threshold */
5811 new->current_threshold = -1;
5812 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5813 if (thresholds->primary->entries[i].eventfd == eventfd)
5816 new->entries[j] = thresholds->primary->entries[i];
5817 if (new->entries[j].threshold <= usage) {
5819 * new->current_threshold will not be used
5820 * until rcu_assign_pointer(), so it's safe to increment
5823 ++new->current_threshold;
5829 /* Swap primary and spare array */
5830 thresholds->spare = thresholds->primary;
5831 /* If all events are unregistered, free the spare array */
5833 kfree(thresholds->spare);
5834 thresholds->spare = NULL;
5837 rcu_assign_pointer(thresholds->primary, new);
5839 /* To be sure that nobody uses thresholds */
5842 mutex_unlock(&memcg->thresholds_lock);
5845 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5846 struct eventfd_ctx *eventfd)
5848 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5851 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5852 struct eventfd_ctx *eventfd)
5854 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5857 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5858 struct eventfd_ctx *eventfd, const char *args)
5860 struct mem_cgroup_eventfd_list *event;
5862 event = kmalloc(sizeof(*event), GFP_KERNEL);
5866 spin_lock(&memcg_oom_lock);
5868 event->eventfd = eventfd;
5869 list_add(&event->list, &memcg->oom_notify);
5871 /* already in OOM ? */
5872 if (atomic_read(&memcg->under_oom))
5873 eventfd_signal(eventfd, 1);
5874 spin_unlock(&memcg_oom_lock);
5879 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5880 struct eventfd_ctx *eventfd)
5882 struct mem_cgroup_eventfd_list *ev, *tmp;
5884 spin_lock(&memcg_oom_lock);
5886 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5887 if (ev->eventfd == eventfd) {
5888 list_del(&ev->list);
5893 spin_unlock(&memcg_oom_lock);
5896 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5898 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5900 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5901 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5905 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5906 struct cftype *cft, u64 val)
5908 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5909 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5911 /* cannot set to root cgroup and only 0 and 1 are allowed */
5912 if (!parent || !((val == 0) || (val == 1)))
5915 mutex_lock(&memcg_create_mutex);
5916 /* oom-kill-disable is a flag for subhierarchy. */
5917 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5918 mutex_unlock(&memcg_create_mutex);
5921 memcg->oom_kill_disable = val;
5923 memcg_oom_recover(memcg);
5924 mutex_unlock(&memcg_create_mutex);
5928 #ifdef CONFIG_MEMCG_KMEM
5929 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5933 memcg->kmemcg_id = -1;
5934 ret = memcg_propagate_kmem(memcg);
5938 return mem_cgroup_sockets_init(memcg, ss);
5941 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5943 mem_cgroup_sockets_destroy(memcg);
5946 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5948 if (!memcg_kmem_is_active(memcg))
5952 * kmem charges can outlive the cgroup. In the case of slab
5953 * pages, for instance, a page contain objects from various
5954 * processes. As we prevent from taking a reference for every
5955 * such allocation we have to be careful when doing uncharge
5956 * (see memcg_uncharge_kmem) and here during offlining.
5958 * The idea is that that only the _last_ uncharge which sees
5959 * the dead memcg will drop the last reference. An additional
5960 * reference is taken here before the group is marked dead
5961 * which is then paired with css_put during uncharge resp. here.
5963 * Although this might sound strange as this path is called from
5964 * css_offline() when the referencemight have dropped down to 0
5965 * and shouldn't be incremented anymore (css_tryget would fail)
5966 * we do not have other options because of the kmem allocations
5969 css_get(&memcg->css);
5971 memcg_kmem_mark_dead(memcg);
5973 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5976 if (memcg_kmem_test_and_clear_dead(memcg))
5977 css_put(&memcg->css);
5980 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5985 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5989 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5995 * DO NOT USE IN NEW FILES.
5997 * "cgroup.event_control" implementation.
5999 * This is way over-engineered. It tries to support fully configurable
6000 * events for each user. Such level of flexibility is completely
6001 * unnecessary especially in the light of the planned unified hierarchy.
6003 * Please deprecate this and replace with something simpler if at all
6008 * Unregister event and free resources.
6010 * Gets called from workqueue.
6012 static void memcg_event_remove(struct work_struct *work)
6014 struct mem_cgroup_event *event =
6015 container_of(work, struct mem_cgroup_event, remove);
6016 struct mem_cgroup *memcg = event->memcg;
6018 remove_wait_queue(event->wqh, &event->wait);
6020 event->unregister_event(memcg, event->eventfd);
6022 /* Notify userspace the event is going away. */
6023 eventfd_signal(event->eventfd, 1);
6025 eventfd_ctx_put(event->eventfd);
6027 css_put(&memcg->css);
6031 * Gets called on POLLHUP on eventfd when user closes it.
6033 * Called with wqh->lock held and interrupts disabled.
6035 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
6036 int sync, void *key)
6038 struct mem_cgroup_event *event =
6039 container_of(wait, struct mem_cgroup_event, wait);
6040 struct mem_cgroup *memcg = event->memcg;
6041 unsigned long flags = (unsigned long)key;
6043 if (flags & POLLHUP) {
6045 * If the event has been detached at cgroup removal, we
6046 * can simply return knowing the other side will cleanup
6049 * We can't race against event freeing since the other
6050 * side will require wqh->lock via remove_wait_queue(),
6053 spin_lock(&memcg->event_list_lock);
6054 if (!list_empty(&event->list)) {
6055 list_del_init(&event->list);
6057 * We are in atomic context, but cgroup_event_remove()
6058 * may sleep, so we have to call it in workqueue.
6060 schedule_work(&event->remove);
6062 spin_unlock(&memcg->event_list_lock);
6068 static void memcg_event_ptable_queue_proc(struct file *file,
6069 wait_queue_head_t *wqh, poll_table *pt)
6071 struct mem_cgroup_event *event =
6072 container_of(pt, struct mem_cgroup_event, pt);
6075 add_wait_queue(wqh, &event->wait);
6079 * DO NOT USE IN NEW FILES.
6081 * Parse input and register new cgroup event handler.
6083 * Input must be in format '<event_fd> <control_fd> <args>'.
6084 * Interpretation of args is defined by control file implementation.
6086 static int memcg_write_event_control(struct cgroup_subsys_state *css,
6087 struct cftype *cft, const char *buffer)
6089 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6090 struct mem_cgroup_event *event;
6091 struct cgroup_subsys_state *cfile_css;
6092 unsigned int efd, cfd;
6099 efd = simple_strtoul(buffer, &endp, 10);
6104 cfd = simple_strtoul(buffer, &endp, 10);
6105 if ((*endp != ' ') && (*endp != '\0'))
6109 event = kzalloc(sizeof(*event), GFP_KERNEL);
6113 event->memcg = memcg;
6114 INIT_LIST_HEAD(&event->list);
6115 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
6116 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
6117 INIT_WORK(&event->remove, memcg_event_remove);
6125 event->eventfd = eventfd_ctx_fileget(efile.file);
6126 if (IS_ERR(event->eventfd)) {
6127 ret = PTR_ERR(event->eventfd);
6134 goto out_put_eventfd;
6137 /* the process need read permission on control file */
6138 /* AV: shouldn't we check that it's been opened for read instead? */
6139 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6144 * Determine the event callbacks and set them in @event. This used
6145 * to be done via struct cftype but cgroup core no longer knows
6146 * about these events. The following is crude but the whole thing
6147 * is for compatibility anyway.
6149 * DO NOT ADD NEW FILES.
6151 name = cfile.file->f_dentry->d_name.name;
6153 if (!strcmp(name, "memory.usage_in_bytes")) {
6154 event->register_event = mem_cgroup_usage_register_event;
6155 event->unregister_event = mem_cgroup_usage_unregister_event;
6156 } else if (!strcmp(name, "memory.oom_control")) {
6157 event->register_event = mem_cgroup_oom_register_event;
6158 event->unregister_event = mem_cgroup_oom_unregister_event;
6159 } else if (!strcmp(name, "memory.pressure_level")) {
6160 event->register_event = vmpressure_register_event;
6161 event->unregister_event = vmpressure_unregister_event;
6162 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
6163 event->register_event = memsw_cgroup_usage_register_event;
6164 event->unregister_event = memsw_cgroup_usage_unregister_event;
6171 * Verify @cfile should belong to @css. Also, remaining events are
6172 * automatically removed on cgroup destruction but the removal is
6173 * asynchronous, so take an extra ref on @css.
6178 cfile_css = css_from_dir(cfile.file->f_dentry->d_parent,
6179 &mem_cgroup_subsys);
6180 if (cfile_css == css && css_tryget(css))
6187 ret = event->register_event(memcg, event->eventfd, buffer);
6191 efile.file->f_op->poll(efile.file, &event->pt);
6193 spin_lock(&memcg->event_list_lock);
6194 list_add(&event->list, &memcg->event_list);
6195 spin_unlock(&memcg->event_list_lock);
6207 eventfd_ctx_put(event->eventfd);
6216 static struct cftype mem_cgroup_files[] = {
6218 .name = "usage_in_bytes",
6219 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6220 .read_u64 = mem_cgroup_read_u64,
6223 .name = "max_usage_in_bytes",
6224 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6225 .trigger = mem_cgroup_reset,
6226 .read_u64 = mem_cgroup_read_u64,
6229 .name = "limit_in_bytes",
6230 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6231 .write_string = mem_cgroup_write,
6232 .read_u64 = mem_cgroup_read_u64,
6235 .name = "soft_limit_in_bytes",
6236 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6237 .write_string = mem_cgroup_write,
6238 .read_u64 = mem_cgroup_read_u64,
6242 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6243 .trigger = mem_cgroup_reset,
6244 .read_u64 = mem_cgroup_read_u64,
6248 .seq_show = memcg_stat_show,
6251 .name = "force_empty",
6252 .trigger = mem_cgroup_force_empty_write,
6255 .name = "use_hierarchy",
6256 .flags = CFTYPE_INSANE,
6257 .write_u64 = mem_cgroup_hierarchy_write,
6258 .read_u64 = mem_cgroup_hierarchy_read,
6261 .name = "cgroup.event_control", /* XXX: for compat */
6262 .write_string = memcg_write_event_control,
6263 .flags = CFTYPE_NO_PREFIX,
6267 .name = "swappiness",
6268 .read_u64 = mem_cgroup_swappiness_read,
6269 .write_u64 = mem_cgroup_swappiness_write,
6272 .name = "move_charge_at_immigrate",
6273 .read_u64 = mem_cgroup_move_charge_read,
6274 .write_u64 = mem_cgroup_move_charge_write,
6277 .name = "oom_control",
6278 .seq_show = mem_cgroup_oom_control_read,
6279 .write_u64 = mem_cgroup_oom_control_write,
6280 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6283 .name = "pressure_level",
6287 .name = "numa_stat",
6288 .seq_show = memcg_numa_stat_show,
6291 #ifdef CONFIG_MEMCG_KMEM
6293 .name = "kmem.limit_in_bytes",
6294 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6295 .write_string = mem_cgroup_write,
6296 .read_u64 = mem_cgroup_read_u64,
6299 .name = "kmem.usage_in_bytes",
6300 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6301 .read_u64 = mem_cgroup_read_u64,
6304 .name = "kmem.failcnt",
6305 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6306 .trigger = mem_cgroup_reset,
6307 .read_u64 = mem_cgroup_read_u64,
6310 .name = "kmem.max_usage_in_bytes",
6311 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6312 .trigger = mem_cgroup_reset,
6313 .read_u64 = mem_cgroup_read_u64,
6315 #ifdef CONFIG_SLABINFO
6317 .name = "kmem.slabinfo",
6318 .seq_show = mem_cgroup_slabinfo_read,
6322 { }, /* terminate */
6325 #ifdef CONFIG_MEMCG_SWAP
6326 static struct cftype memsw_cgroup_files[] = {
6328 .name = "memsw.usage_in_bytes",
6329 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6330 .read_u64 = mem_cgroup_read_u64,
6333 .name = "memsw.max_usage_in_bytes",
6334 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6335 .trigger = mem_cgroup_reset,
6336 .read_u64 = mem_cgroup_read_u64,
6339 .name = "memsw.limit_in_bytes",
6340 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6341 .write_string = mem_cgroup_write,
6342 .read_u64 = mem_cgroup_read_u64,
6345 .name = "memsw.failcnt",
6346 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6347 .trigger = mem_cgroup_reset,
6348 .read_u64 = mem_cgroup_read_u64,
6350 { }, /* terminate */
6353 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6355 struct mem_cgroup_per_node *pn;
6356 struct mem_cgroup_per_zone *mz;
6357 int zone, tmp = node;
6359 * This routine is called against possible nodes.
6360 * But it's BUG to call kmalloc() against offline node.
6362 * TODO: this routine can waste much memory for nodes which will
6363 * never be onlined. It's better to use memory hotplug callback
6366 if (!node_state(node, N_NORMAL_MEMORY))
6368 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6372 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6373 mz = &pn->zoneinfo[zone];
6374 lruvec_init(&mz->lruvec);
6375 mz->usage_in_excess = 0;
6376 mz->on_tree = false;
6379 memcg->nodeinfo[node] = pn;
6383 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6385 kfree(memcg->nodeinfo[node]);
6388 static struct mem_cgroup *mem_cgroup_alloc(void)
6390 struct mem_cgroup *memcg;
6393 size = sizeof(struct mem_cgroup);
6394 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6396 memcg = kzalloc(size, GFP_KERNEL);
6400 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6403 spin_lock_init(&memcg->pcp_counter_lock);
6412 * At destroying mem_cgroup, references from swap_cgroup can remain.
6413 * (scanning all at force_empty is too costly...)
6415 * Instead of clearing all references at force_empty, we remember
6416 * the number of reference from swap_cgroup and free mem_cgroup when
6417 * it goes down to 0.
6419 * Removal of cgroup itself succeeds regardless of refs from swap.
6422 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6426 mem_cgroup_remove_from_trees(memcg);
6429 free_mem_cgroup_per_zone_info(memcg, node);
6431 free_percpu(memcg->stat);
6434 * We need to make sure that (at least for now), the jump label
6435 * destruction code runs outside of the cgroup lock. This is because
6436 * get_online_cpus(), which is called from the static_branch update,
6437 * can't be called inside the cgroup_lock. cpusets are the ones
6438 * enforcing this dependency, so if they ever change, we might as well.
6440 * schedule_work() will guarantee this happens. Be careful if you need
6441 * to move this code around, and make sure it is outside
6444 disarm_static_keys(memcg);
6449 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6451 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6453 if (!memcg->res.parent)
6455 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6457 EXPORT_SYMBOL(parent_mem_cgroup);
6459 static void __init mem_cgroup_soft_limit_tree_init(void)
6461 struct mem_cgroup_tree_per_node *rtpn;
6462 struct mem_cgroup_tree_per_zone *rtpz;
6463 int tmp, node, zone;
6465 for_each_node(node) {
6467 if (!node_state(node, N_NORMAL_MEMORY))
6469 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6472 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6474 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6475 rtpz = &rtpn->rb_tree_per_zone[zone];
6476 rtpz->rb_root = RB_ROOT;
6477 spin_lock_init(&rtpz->lock);
6482 static struct cgroup_subsys_state * __ref
6483 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6485 struct mem_cgroup *memcg;
6486 long error = -ENOMEM;
6489 memcg = mem_cgroup_alloc();
6491 return ERR_PTR(error);
6494 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6498 if (parent_css == NULL) {
6499 root_mem_cgroup = memcg;
6500 res_counter_init(&memcg->res, NULL);
6501 res_counter_init(&memcg->memsw, NULL);
6502 res_counter_init(&memcg->kmem, NULL);
6505 memcg->last_scanned_node = MAX_NUMNODES;
6506 INIT_LIST_HEAD(&memcg->oom_notify);
6507 memcg->move_charge_at_immigrate = 0;
6508 mutex_init(&memcg->thresholds_lock);
6509 spin_lock_init(&memcg->move_lock);
6510 vmpressure_init(&memcg->vmpressure);
6511 INIT_LIST_HEAD(&memcg->event_list);
6512 spin_lock_init(&memcg->event_list_lock);
6517 __mem_cgroup_free(memcg);
6518 return ERR_PTR(error);
6522 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6524 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6525 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6528 if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6534 mutex_lock(&memcg_create_mutex);
6536 memcg->use_hierarchy = parent->use_hierarchy;
6537 memcg->oom_kill_disable = parent->oom_kill_disable;
6538 memcg->swappiness = mem_cgroup_swappiness(parent);
6540 if (parent->use_hierarchy) {
6541 res_counter_init(&memcg->res, &parent->res);
6542 res_counter_init(&memcg->memsw, &parent->memsw);
6543 res_counter_init(&memcg->kmem, &parent->kmem);
6546 * No need to take a reference to the parent because cgroup
6547 * core guarantees its existence.
6550 res_counter_init(&memcg->res, NULL);
6551 res_counter_init(&memcg->memsw, NULL);
6552 res_counter_init(&memcg->kmem, NULL);
6554 * Deeper hierachy with use_hierarchy == false doesn't make
6555 * much sense so let cgroup subsystem know about this
6556 * unfortunate state in our controller.
6558 if (parent != root_mem_cgroup)
6559 mem_cgroup_subsys.broken_hierarchy = true;
6562 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6563 mutex_unlock(&memcg_create_mutex);
6568 * Announce all parents that a group from their hierarchy is gone.
6570 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6572 struct mem_cgroup *parent = memcg;
6574 while ((parent = parent_mem_cgroup(parent)))
6575 mem_cgroup_iter_invalidate(parent);
6578 * if the root memcg is not hierarchical we have to check it
6581 if (!root_mem_cgroup->use_hierarchy)
6582 mem_cgroup_iter_invalidate(root_mem_cgroup);
6585 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6587 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6588 struct mem_cgroup_event *event, *tmp;
6591 * Unregister events and notify userspace.
6592 * Notify userspace about cgroup removing only after rmdir of cgroup
6593 * directory to avoid race between userspace and kernelspace.
6595 spin_lock(&memcg->event_list_lock);
6596 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6597 list_del_init(&event->list);
6598 schedule_work(&event->remove);
6600 spin_unlock(&memcg->event_list_lock);
6602 kmem_cgroup_css_offline(memcg);
6604 mem_cgroup_invalidate_reclaim_iterators(memcg);
6605 mem_cgroup_reparent_charges(memcg);
6606 mem_cgroup_destroy_all_caches(memcg);
6607 vmpressure_cleanup(&memcg->vmpressure);
6610 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6612 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6614 * XXX: css_offline() would be where we should reparent all
6615 * memory to prepare the cgroup for destruction. However,
6616 * memcg does not do css_tryget() and res_counter charging
6617 * under the same RCU lock region, which means that charging
6618 * could race with offlining. Offlining only happens to
6619 * cgroups with no tasks in them but charges can show up
6620 * without any tasks from the swapin path when the target
6621 * memcg is looked up from the swapout record and not from the
6622 * current task as it usually is. A race like this can leak
6623 * charges and put pages with stale cgroup pointers into
6627 * lookup_swap_cgroup_id()
6629 * mem_cgroup_lookup()
6632 * disable css_tryget()
6635 * reparent_charges()
6636 * res_counter_charge()
6639 * pc->mem_cgroup = dead memcg
6642 * The bulk of the charges are still moved in offline_css() to
6643 * avoid pinning a lot of pages in case a long-term reference
6644 * like a swapout record is deferring the css_free() to long
6645 * after offlining. But this makes sure we catch any charges
6646 * made after offlining:
6648 mem_cgroup_reparent_charges(memcg);
6650 memcg_destroy_kmem(memcg);
6651 __mem_cgroup_free(memcg);
6655 /* Handlers for move charge at task migration. */
6656 #define PRECHARGE_COUNT_AT_ONCE 256
6657 static int mem_cgroup_do_precharge(unsigned long count)
6660 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6661 struct mem_cgroup *memcg = mc.to;
6663 if (mem_cgroup_is_root(memcg)) {
6664 mc.precharge += count;
6665 /* we don't need css_get for root */
6668 /* try to charge at once */
6670 struct res_counter *dummy;
6672 * "memcg" cannot be under rmdir() because we've already checked
6673 * by cgroup_lock_live_cgroup() that it is not removed and we
6674 * are still under the same cgroup_mutex. So we can postpone
6677 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6679 if (do_swap_account && res_counter_charge(&memcg->memsw,
6680 PAGE_SIZE * count, &dummy)) {
6681 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6684 mc.precharge += count;
6688 /* fall back to one by one charge */
6690 if (signal_pending(current)) {
6694 if (!batch_count--) {
6695 batch_count = PRECHARGE_COUNT_AT_ONCE;
6698 ret = __mem_cgroup_try_charge(NULL,
6699 GFP_KERNEL, 1, &memcg, false);
6701 /* mem_cgroup_clear_mc() will do uncharge later */
6709 * get_mctgt_type - get target type of moving charge
6710 * @vma: the vma the pte to be checked belongs
6711 * @addr: the address corresponding to the pte to be checked
6712 * @ptent: the pte to be checked
6713 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6716 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6717 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6718 * move charge. if @target is not NULL, the page is stored in target->page
6719 * with extra refcnt got(Callers should handle it).
6720 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6721 * target for charge migration. if @target is not NULL, the entry is stored
6724 * Called with pte lock held.
6731 enum mc_target_type {
6737 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6738 unsigned long addr, pte_t ptent)
6740 struct page *page = vm_normal_page(vma, addr, ptent);
6742 if (!page || !page_mapped(page))
6744 if (PageAnon(page)) {
6745 /* we don't move shared anon */
6748 } else if (!move_file())
6749 /* we ignore mapcount for file pages */
6751 if (!get_page_unless_zero(page))
6758 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6759 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6761 struct page *page = NULL;
6762 swp_entry_t ent = pte_to_swp_entry(ptent);
6764 if (!move_anon() || non_swap_entry(ent))
6767 * Because lookup_swap_cache() updates some statistics counter,
6768 * we call find_get_page() with swapper_space directly.
6770 page = find_get_page(swap_address_space(ent), ent.val);
6771 if (do_swap_account)
6772 entry->val = ent.val;
6777 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6778 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6784 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6785 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6787 struct page *page = NULL;
6788 struct address_space *mapping;
6791 if (!vma->vm_file) /* anonymous vma */
6796 mapping = vma->vm_file->f_mapping;
6797 if (pte_none(ptent))
6798 pgoff = linear_page_index(vma, addr);
6799 else /* pte_file(ptent) is true */
6800 pgoff = pte_to_pgoff(ptent);
6802 /* page is moved even if it's not RSS of this task(page-faulted). */
6803 page = find_get_page(mapping, pgoff);
6806 /* shmem/tmpfs may report page out on swap: account for that too. */
6807 if (radix_tree_exceptional_entry(page)) {
6808 swp_entry_t swap = radix_to_swp_entry(page);
6809 if (do_swap_account)
6811 page = find_get_page(swap_address_space(swap), swap.val);
6817 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6818 unsigned long addr, pte_t ptent, union mc_target *target)
6820 struct page *page = NULL;
6821 struct page_cgroup *pc;
6822 enum mc_target_type ret = MC_TARGET_NONE;
6823 swp_entry_t ent = { .val = 0 };
6825 if (pte_present(ptent))
6826 page = mc_handle_present_pte(vma, addr, ptent);
6827 else if (is_swap_pte(ptent))
6828 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6829 else if (pte_none(ptent) || pte_file(ptent))
6830 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6832 if (!page && !ent.val)
6835 pc = lookup_page_cgroup(page);
6837 * Do only loose check w/o page_cgroup lock.
6838 * mem_cgroup_move_account() checks the pc is valid or not under
6841 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6842 ret = MC_TARGET_PAGE;
6844 target->page = page;
6846 if (!ret || !target)
6849 /* There is a swap entry and a page doesn't exist or isn't charged */
6850 if (ent.val && !ret &&
6851 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6852 ret = MC_TARGET_SWAP;
6859 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6861 * We don't consider swapping or file mapped pages because THP does not
6862 * support them for now.
6863 * Caller should make sure that pmd_trans_huge(pmd) is true.
6865 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6866 unsigned long addr, pmd_t pmd, union mc_target *target)
6868 struct page *page = NULL;
6869 struct page_cgroup *pc;
6870 enum mc_target_type ret = MC_TARGET_NONE;
6872 page = pmd_page(pmd);
6873 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6876 pc = lookup_page_cgroup(page);
6877 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6878 ret = MC_TARGET_PAGE;
6881 target->page = page;
6887 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6888 unsigned long addr, pmd_t pmd, union mc_target *target)
6890 return MC_TARGET_NONE;
6894 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6895 unsigned long addr, unsigned long end,
6896 struct mm_walk *walk)
6898 struct vm_area_struct *vma = walk->private;
6902 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6903 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6904 mc.precharge += HPAGE_PMD_NR;
6909 if (pmd_trans_unstable(pmd))
6911 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6912 for (; addr != end; pte++, addr += PAGE_SIZE)
6913 if (get_mctgt_type(vma, addr, *pte, NULL))
6914 mc.precharge++; /* increment precharge temporarily */
6915 pte_unmap_unlock(pte - 1, ptl);
6921 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6923 unsigned long precharge;
6924 struct vm_area_struct *vma;
6926 down_read(&mm->mmap_sem);
6927 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6928 struct mm_walk mem_cgroup_count_precharge_walk = {
6929 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6933 if (is_vm_hugetlb_page(vma))
6935 walk_page_range(vma->vm_start, vma->vm_end,
6936 &mem_cgroup_count_precharge_walk);
6938 up_read(&mm->mmap_sem);
6940 precharge = mc.precharge;
6946 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6948 unsigned long precharge = mem_cgroup_count_precharge(mm);
6950 VM_BUG_ON(mc.moving_task);
6951 mc.moving_task = current;
6952 return mem_cgroup_do_precharge(precharge);
6955 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6956 static void __mem_cgroup_clear_mc(void)
6958 struct mem_cgroup *from = mc.from;
6959 struct mem_cgroup *to = mc.to;
6962 /* we must uncharge all the leftover precharges from mc.to */
6964 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6968 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6969 * we must uncharge here.
6971 if (mc.moved_charge) {
6972 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6973 mc.moved_charge = 0;
6975 /* we must fixup refcnts and charges */
6976 if (mc.moved_swap) {
6977 /* uncharge swap account from the old cgroup */
6978 if (!mem_cgroup_is_root(mc.from))
6979 res_counter_uncharge(&mc.from->memsw,
6980 PAGE_SIZE * mc.moved_swap);
6982 for (i = 0; i < mc.moved_swap; i++)
6983 css_put(&mc.from->css);
6985 if (!mem_cgroup_is_root(mc.to)) {
6987 * we charged both to->res and to->memsw, so we should
6990 res_counter_uncharge(&mc.to->res,
6991 PAGE_SIZE * mc.moved_swap);
6993 /* we've already done css_get(mc.to) */
6996 memcg_oom_recover(from);
6997 memcg_oom_recover(to);
6998 wake_up_all(&mc.waitq);
7001 static void mem_cgroup_clear_mc(void)
7003 struct mem_cgroup *from = mc.from;
7006 * we must clear moving_task before waking up waiters at the end of
7009 mc.moving_task = NULL;
7010 __mem_cgroup_clear_mc();
7011 spin_lock(&mc.lock);
7014 spin_unlock(&mc.lock);
7015 mem_cgroup_end_move(from);
7018 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7019 struct cgroup_taskset *tset)
7021 struct task_struct *p = cgroup_taskset_first(tset);
7023 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
7024 unsigned long move_charge_at_immigrate;
7027 * We are now commited to this value whatever it is. Changes in this
7028 * tunable will only affect upcoming migrations, not the current one.
7029 * So we need to save it, and keep it going.
7031 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
7032 if (move_charge_at_immigrate) {
7033 struct mm_struct *mm;
7034 struct mem_cgroup *from = mem_cgroup_from_task(p);
7036 VM_BUG_ON(from == memcg);
7038 mm = get_task_mm(p);
7041 /* We move charges only when we move a owner of the mm */
7042 if (mm->owner == p) {
7045 VM_BUG_ON(mc.precharge);
7046 VM_BUG_ON(mc.moved_charge);
7047 VM_BUG_ON(mc.moved_swap);
7048 mem_cgroup_start_move(from);
7049 spin_lock(&mc.lock);
7052 mc.immigrate_flags = move_charge_at_immigrate;
7053 spin_unlock(&mc.lock);
7054 /* We set mc.moving_task later */
7056 ret = mem_cgroup_precharge_mc(mm);
7058 mem_cgroup_clear_mc();
7065 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7066 struct cgroup_taskset *tset)
7068 mem_cgroup_clear_mc();
7071 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
7072 unsigned long addr, unsigned long end,
7073 struct mm_walk *walk)
7076 struct vm_area_struct *vma = walk->private;
7079 enum mc_target_type target_type;
7080 union mc_target target;
7082 struct page_cgroup *pc;
7085 * We don't take compound_lock() here but no race with splitting thp
7087 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7088 * under splitting, which means there's no concurrent thp split,
7089 * - if another thread runs into split_huge_page() just after we
7090 * entered this if-block, the thread must wait for page table lock
7091 * to be unlocked in __split_huge_page_splitting(), where the main
7092 * part of thp split is not executed yet.
7094 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
7095 if (mc.precharge < HPAGE_PMD_NR) {
7099 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
7100 if (target_type == MC_TARGET_PAGE) {
7102 if (!isolate_lru_page(page)) {
7103 pc = lookup_page_cgroup(page);
7104 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
7105 pc, mc.from, mc.to)) {
7106 mc.precharge -= HPAGE_PMD_NR;
7107 mc.moved_charge += HPAGE_PMD_NR;
7109 putback_lru_page(page);
7117 if (pmd_trans_unstable(pmd))
7120 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7121 for (; addr != end; addr += PAGE_SIZE) {
7122 pte_t ptent = *(pte++);
7128 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7129 case MC_TARGET_PAGE:
7131 if (isolate_lru_page(page))
7133 pc = lookup_page_cgroup(page);
7134 if (!mem_cgroup_move_account(page, 1, pc,
7137 /* we uncharge from mc.from later. */
7140 putback_lru_page(page);
7141 put: /* get_mctgt_type() gets the page */
7144 case MC_TARGET_SWAP:
7146 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7148 /* we fixup refcnts and charges later. */
7156 pte_unmap_unlock(pte - 1, ptl);
7161 * We have consumed all precharges we got in can_attach().
7162 * We try charge one by one, but don't do any additional
7163 * charges to mc.to if we have failed in charge once in attach()
7166 ret = mem_cgroup_do_precharge(1);
7174 static void mem_cgroup_move_charge(struct mm_struct *mm)
7176 struct vm_area_struct *vma;
7178 lru_add_drain_all();
7180 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7182 * Someone who are holding the mmap_sem might be waiting in
7183 * waitq. So we cancel all extra charges, wake up all waiters,
7184 * and retry. Because we cancel precharges, we might not be able
7185 * to move enough charges, but moving charge is a best-effort
7186 * feature anyway, so it wouldn't be a big problem.
7188 __mem_cgroup_clear_mc();
7192 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7194 struct mm_walk mem_cgroup_move_charge_walk = {
7195 .pmd_entry = mem_cgroup_move_charge_pte_range,
7199 if (is_vm_hugetlb_page(vma))
7201 ret = walk_page_range(vma->vm_start, vma->vm_end,
7202 &mem_cgroup_move_charge_walk);
7205 * means we have consumed all precharges and failed in
7206 * doing additional charge. Just abandon here.
7210 up_read(&mm->mmap_sem);
7213 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7214 struct cgroup_taskset *tset)
7216 struct task_struct *p = cgroup_taskset_first(tset);
7217 struct mm_struct *mm = get_task_mm(p);
7221 mem_cgroup_move_charge(mm);
7225 mem_cgroup_clear_mc();
7227 #else /* !CONFIG_MMU */
7228 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7229 struct cgroup_taskset *tset)
7233 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7234 struct cgroup_taskset *tset)
7237 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7238 struct cgroup_taskset *tset)
7244 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7245 * to verify sane_behavior flag on each mount attempt.
7247 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7250 * use_hierarchy is forced with sane_behavior. cgroup core
7251 * guarantees that @root doesn't have any children, so turning it
7252 * on for the root memcg is enough.
7254 if (cgroup_sane_behavior(root_css->cgroup))
7255 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7258 struct cgroup_subsys mem_cgroup_subsys = {
7260 .subsys_id = mem_cgroup_subsys_id,
7261 .css_alloc = mem_cgroup_css_alloc,
7262 .css_online = mem_cgroup_css_online,
7263 .css_offline = mem_cgroup_css_offline,
7264 .css_free = mem_cgroup_css_free,
7265 .can_attach = mem_cgroup_can_attach,
7266 .cancel_attach = mem_cgroup_cancel_attach,
7267 .attach = mem_cgroup_move_task,
7268 .bind = mem_cgroup_bind,
7269 .base_cftypes = mem_cgroup_files,
7273 #ifdef CONFIG_MEMCG_SWAP
7274 static int __init enable_swap_account(char *s)
7276 if (!strcmp(s, "1"))
7277 really_do_swap_account = 1;
7278 else if (!strcmp(s, "0"))
7279 really_do_swap_account = 0;
7282 __setup("swapaccount=", enable_swap_account);
7284 static void __init memsw_file_init(void)
7286 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7289 static void __init enable_swap_cgroup(void)
7291 if (!mem_cgroup_disabled() && really_do_swap_account) {
7292 do_swap_account = 1;
7298 static void __init enable_swap_cgroup(void)
7304 * subsys_initcall() for memory controller.
7306 * Some parts like hotcpu_notifier() have to be initialized from this context
7307 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7308 * everything that doesn't depend on a specific mem_cgroup structure should
7309 * be initialized from here.
7311 static int __init mem_cgroup_init(void)
7313 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7314 enable_swap_cgroup();
7315 mem_cgroup_soft_limit_tree_init();
7319 subsys_initcall(mem_cgroup_init);