1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmalloc.h>
53 #include <linux/vmpressure.h>
54 #include <linux/mm_inline.h>
55 #include <linux/page_cgroup.h>
56 #include <linux/cpu.h>
57 #include <linux/oom.h>
58 #include <linux/lockdep.h>
59 #include <linux/file.h>
63 #include <net/tcp_memcontrol.h>
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 cgroup_event {
236 * css which the event belongs to.
238 struct cgroup_subsys_state *css;
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 cgroup_subsys_state *css,
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 cgroup_subsys_state *css,
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 tcp_memcontrol 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 static size_t memcg_size(void)
385 return sizeof(struct mem_cgroup) +
386 nr_node_ids * sizeof(struct mem_cgroup_per_node);
389 /* internal only representation about the status of kmem accounting. */
391 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
392 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
393 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
396 /* We account when limit is on, but only after call sites are patched */
397 #define KMEM_ACCOUNTED_MASK \
398 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
400 #ifdef CONFIG_MEMCG_KMEM
401 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
403 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
406 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
408 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
411 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
413 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
416 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
418 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
421 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
424 * Our caller must use css_get() first, because memcg_uncharge_kmem()
425 * will call css_put() if it sees the memcg is dead.
428 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
429 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
432 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
434 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
435 &memcg->kmem_account_flags);
439 /* Stuffs for move charges at task migration. */
441 * Types of charges to be moved. "move_charge_at_immitgrate" and
442 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
445 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
446 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
450 /* "mc" and its members are protected by cgroup_mutex */
451 static struct move_charge_struct {
452 spinlock_t lock; /* for from, to */
453 struct mem_cgroup *from;
454 struct mem_cgroup *to;
455 unsigned long immigrate_flags;
456 unsigned long precharge;
457 unsigned long moved_charge;
458 unsigned long moved_swap;
459 struct task_struct *moving_task; /* a task moving charges */
460 wait_queue_head_t waitq; /* a waitq for other context */
462 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
463 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
466 static bool move_anon(void)
468 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
471 static bool move_file(void)
473 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
477 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
478 * limit reclaim to prevent infinite loops, if they ever occur.
480 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
481 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
484 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
485 MEM_CGROUP_CHARGE_TYPE_ANON,
486 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
487 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
491 /* for encoding cft->private value on file */
499 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
500 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
501 #define MEMFILE_ATTR(val) ((val) & 0xffff)
502 /* Used for OOM nofiier */
503 #define OOM_CONTROL (0)
506 * Reclaim flags for mem_cgroup_hierarchical_reclaim
508 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
509 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
510 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
511 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
514 * The memcg_create_mutex will be held whenever a new cgroup is created.
515 * As a consequence, any change that needs to protect against new child cgroups
516 * appearing has to hold it as well.
518 static DEFINE_MUTEX(memcg_create_mutex);
520 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
522 return s ? container_of(s, struct mem_cgroup, css) : NULL;
525 /* Some nice accessors for the vmpressure. */
526 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
529 memcg = root_mem_cgroup;
530 return &memcg->vmpressure;
533 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
535 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
538 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
540 return &mem_cgroup_from_css(css)->vmpressure;
543 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
545 return (memcg == root_mem_cgroup);
548 /* Writing them here to avoid exposing memcg's inner layout */
549 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
551 void sock_update_memcg(struct sock *sk)
553 if (mem_cgroup_sockets_enabled) {
554 struct mem_cgroup *memcg;
555 struct cg_proto *cg_proto;
557 BUG_ON(!sk->sk_prot->proto_cgroup);
559 /* Socket cloning can throw us here with sk_cgrp already
560 * filled. It won't however, necessarily happen from
561 * process context. So the test for root memcg given
562 * the current task's memcg won't help us in this case.
564 * Respecting the original socket's memcg is a better
565 * decision in this case.
568 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
569 css_get(&sk->sk_cgrp->memcg->css);
574 memcg = mem_cgroup_from_task(current);
575 cg_proto = sk->sk_prot->proto_cgroup(memcg);
576 if (!mem_cgroup_is_root(memcg) &&
577 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
578 sk->sk_cgrp = cg_proto;
583 EXPORT_SYMBOL(sock_update_memcg);
585 void sock_release_memcg(struct sock *sk)
587 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
588 struct mem_cgroup *memcg;
589 WARN_ON(!sk->sk_cgrp->memcg);
590 memcg = sk->sk_cgrp->memcg;
591 css_put(&sk->sk_cgrp->memcg->css);
595 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
597 if (!memcg || mem_cgroup_is_root(memcg))
600 return &memcg->tcp_mem.cg_proto;
602 EXPORT_SYMBOL(tcp_proto_cgroup);
604 static void disarm_sock_keys(struct mem_cgroup *memcg)
606 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
608 static_key_slow_dec(&memcg_socket_limit_enabled);
611 static void disarm_sock_keys(struct mem_cgroup *memcg)
616 #ifdef CONFIG_MEMCG_KMEM
618 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
619 * There are two main reasons for not using the css_id for this:
620 * 1) this works better in sparse environments, where we have a lot of memcgs,
621 * but only a few kmem-limited. Or also, if we have, for instance, 200
622 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
623 * 200 entry array for that.
625 * 2) In order not to violate the cgroup API, we would like to do all memory
626 * allocation in ->create(). At that point, we haven't yet allocated the
627 * css_id. Having a separate index prevents us from messing with the cgroup
630 * The current size of the caches array is stored in
631 * memcg_limited_groups_array_size. It will double each time we have to
634 static DEFINE_IDA(kmem_limited_groups);
635 int memcg_limited_groups_array_size;
638 * MIN_SIZE is different than 1, because we would like to avoid going through
639 * the alloc/free process all the time. In a small machine, 4 kmem-limited
640 * cgroups is a reasonable guess. In the future, it could be a parameter or
641 * tunable, but that is strictly not necessary.
643 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
644 * this constant directly from cgroup, but it is understandable that this is
645 * better kept as an internal representation in cgroup.c. In any case, the
646 * css_id space is not getting any smaller, and we don't have to necessarily
647 * increase ours as well if it increases.
649 #define MEMCG_CACHES_MIN_SIZE 4
650 #define MEMCG_CACHES_MAX_SIZE 65535
653 * A lot of the calls to the cache allocation functions are expected to be
654 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
655 * conditional to this static branch, we'll have to allow modules that does
656 * kmem_cache_alloc and the such to see this symbol as well
658 struct static_key memcg_kmem_enabled_key;
659 EXPORT_SYMBOL(memcg_kmem_enabled_key);
661 static void disarm_kmem_keys(struct mem_cgroup *memcg)
663 if (memcg_kmem_is_active(memcg)) {
664 static_key_slow_dec(&memcg_kmem_enabled_key);
665 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
668 * This check can't live in kmem destruction function,
669 * since the charges will outlive the cgroup
671 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
674 static void disarm_kmem_keys(struct mem_cgroup *memcg)
677 #endif /* CONFIG_MEMCG_KMEM */
679 static void disarm_static_keys(struct mem_cgroup *memcg)
681 disarm_sock_keys(memcg);
682 disarm_kmem_keys(memcg);
685 static void drain_all_stock_async(struct mem_cgroup *memcg);
687 static struct mem_cgroup_per_zone *
688 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
690 VM_BUG_ON((unsigned)nid >= nr_node_ids);
691 return &memcg->nodeinfo[nid]->zoneinfo[zid];
694 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
699 static struct mem_cgroup_per_zone *
700 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
702 int nid = page_to_nid(page);
703 int zid = page_zonenum(page);
705 return mem_cgroup_zoneinfo(memcg, nid, zid);
708 static struct mem_cgroup_tree_per_zone *
709 soft_limit_tree_node_zone(int nid, int zid)
711 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
714 static struct mem_cgroup_tree_per_zone *
715 soft_limit_tree_from_page(struct page *page)
717 int nid = page_to_nid(page);
718 int zid = page_zonenum(page);
720 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
724 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
725 struct mem_cgroup_per_zone *mz,
726 struct mem_cgroup_tree_per_zone *mctz,
727 unsigned long long new_usage_in_excess)
729 struct rb_node **p = &mctz->rb_root.rb_node;
730 struct rb_node *parent = NULL;
731 struct mem_cgroup_per_zone *mz_node;
736 mz->usage_in_excess = new_usage_in_excess;
737 if (!mz->usage_in_excess)
741 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
743 if (mz->usage_in_excess < mz_node->usage_in_excess)
746 * We can't avoid mem cgroups that are over their soft
747 * limit by the same amount
749 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
752 rb_link_node(&mz->tree_node, parent, p);
753 rb_insert_color(&mz->tree_node, &mctz->rb_root);
758 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
759 struct mem_cgroup_per_zone *mz,
760 struct mem_cgroup_tree_per_zone *mctz)
764 rb_erase(&mz->tree_node, &mctz->rb_root);
769 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
770 struct mem_cgroup_per_zone *mz,
771 struct mem_cgroup_tree_per_zone *mctz)
773 spin_lock(&mctz->lock);
774 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
775 spin_unlock(&mctz->lock);
779 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
781 unsigned long long excess;
782 struct mem_cgroup_per_zone *mz;
783 struct mem_cgroup_tree_per_zone *mctz;
784 int nid = page_to_nid(page);
785 int zid = page_zonenum(page);
786 mctz = soft_limit_tree_from_page(page);
789 * Necessary to update all ancestors when hierarchy is used.
790 * because their event counter is not touched.
792 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
793 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
794 excess = res_counter_soft_limit_excess(&memcg->res);
796 * We have to update the tree if mz is on RB-tree or
797 * mem is over its softlimit.
799 if (excess || mz->on_tree) {
800 spin_lock(&mctz->lock);
801 /* if on-tree, remove it */
803 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
805 * Insert again. mz->usage_in_excess will be updated.
806 * If excess is 0, no tree ops.
808 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
809 spin_unlock(&mctz->lock);
814 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
817 struct mem_cgroup_per_zone *mz;
818 struct mem_cgroup_tree_per_zone *mctz;
820 for_each_node(node) {
821 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
822 mz = mem_cgroup_zoneinfo(memcg, node, zone);
823 mctz = soft_limit_tree_node_zone(node, zone);
824 mem_cgroup_remove_exceeded(memcg, mz, mctz);
829 static struct mem_cgroup_per_zone *
830 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
832 struct rb_node *rightmost = NULL;
833 struct mem_cgroup_per_zone *mz;
837 rightmost = rb_last(&mctz->rb_root);
839 goto done; /* Nothing to reclaim from */
841 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
843 * Remove the node now but someone else can add it back,
844 * we will to add it back at the end of reclaim to its correct
845 * position in the tree.
847 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
848 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
849 !css_tryget(&mz->memcg->css))
855 static struct mem_cgroup_per_zone *
856 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
858 struct mem_cgroup_per_zone *mz;
860 spin_lock(&mctz->lock);
861 mz = __mem_cgroup_largest_soft_limit_node(mctz);
862 spin_unlock(&mctz->lock);
867 * Implementation Note: reading percpu statistics for memcg.
869 * Both of vmstat[] and percpu_counter has threshold and do periodic
870 * synchronization to implement "quick" read. There are trade-off between
871 * reading cost and precision of value. Then, we may have a chance to implement
872 * a periodic synchronizion of counter in memcg's counter.
874 * But this _read() function is used for user interface now. The user accounts
875 * memory usage by memory cgroup and he _always_ requires exact value because
876 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
877 * have to visit all online cpus and make sum. So, for now, unnecessary
878 * synchronization is not implemented. (just implemented for cpu hotplug)
880 * If there are kernel internal actions which can make use of some not-exact
881 * value, and reading all cpu value can be performance bottleneck in some
882 * common workload, threashold and synchonization as vmstat[] should be
885 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
886 enum mem_cgroup_stat_index idx)
892 for_each_online_cpu(cpu)
893 val += per_cpu(memcg->stat->count[idx], cpu);
894 #ifdef CONFIG_HOTPLUG_CPU
895 spin_lock(&memcg->pcp_counter_lock);
896 val += memcg->nocpu_base.count[idx];
897 spin_unlock(&memcg->pcp_counter_lock);
903 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
906 int val = (charge) ? 1 : -1;
907 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
910 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
911 enum mem_cgroup_events_index idx)
913 unsigned long val = 0;
917 for_each_online_cpu(cpu)
918 val += per_cpu(memcg->stat->events[idx], cpu);
919 #ifdef CONFIG_HOTPLUG_CPU
920 spin_lock(&memcg->pcp_counter_lock);
921 val += memcg->nocpu_base.events[idx];
922 spin_unlock(&memcg->pcp_counter_lock);
928 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
930 bool anon, int nr_pages)
935 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
936 * counted as CACHE even if it's on ANON LRU.
939 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
942 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
945 if (PageTransHuge(page))
946 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
949 /* pagein of a big page is an event. So, ignore page size */
951 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
953 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
954 nr_pages = -nr_pages; /* for event */
957 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
963 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
965 struct mem_cgroup_per_zone *mz;
967 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
968 return mz->lru_size[lru];
972 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
973 unsigned int lru_mask)
975 struct mem_cgroup_per_zone *mz;
977 unsigned long ret = 0;
979 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
982 if (BIT(lru) & lru_mask)
983 ret += mz->lru_size[lru];
989 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
990 int nid, unsigned int lru_mask)
995 for (zid = 0; zid < MAX_NR_ZONES; zid++)
996 total += mem_cgroup_zone_nr_lru_pages(memcg,
1002 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
1003 unsigned int lru_mask)
1008 for_each_node_state(nid, N_MEMORY)
1009 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1013 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1014 enum mem_cgroup_events_target target)
1016 unsigned long val, next;
1018 val = __this_cpu_read(memcg->stat->nr_page_events);
1019 next = __this_cpu_read(memcg->stat->targets[target]);
1020 /* from time_after() in jiffies.h */
1021 if ((long)next - (long)val < 0) {
1023 case MEM_CGROUP_TARGET_THRESH:
1024 next = val + THRESHOLDS_EVENTS_TARGET;
1026 case MEM_CGROUP_TARGET_SOFTLIMIT:
1027 next = val + SOFTLIMIT_EVENTS_TARGET;
1029 case MEM_CGROUP_TARGET_NUMAINFO:
1030 next = val + NUMAINFO_EVENTS_TARGET;
1035 __this_cpu_write(memcg->stat->targets[target], next);
1042 * Check events in order.
1045 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1048 /* threshold event is triggered in finer grain than soft limit */
1049 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1050 MEM_CGROUP_TARGET_THRESH))) {
1052 bool do_numainfo __maybe_unused;
1054 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1055 MEM_CGROUP_TARGET_SOFTLIMIT);
1056 #if MAX_NUMNODES > 1
1057 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1058 MEM_CGROUP_TARGET_NUMAINFO);
1062 mem_cgroup_threshold(memcg);
1063 if (unlikely(do_softlimit))
1064 mem_cgroup_update_tree(memcg, page);
1065 #if MAX_NUMNODES > 1
1066 if (unlikely(do_numainfo))
1067 atomic_inc(&memcg->numainfo_events);
1073 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1076 * mm_update_next_owner() may clear mm->owner to NULL
1077 * if it races with swapoff, page migration, etc.
1078 * So this can be called with p == NULL.
1083 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1086 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1088 struct mem_cgroup *memcg = NULL;
1093 * Because we have no locks, mm->owner's may be being moved to other
1094 * cgroup. We use css_tryget() here even if this looks
1095 * pessimistic (rather than adding locks here).
1099 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1100 if (unlikely(!memcg))
1102 } while (!css_tryget(&memcg->css));
1108 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1109 * ref. count) or NULL if the whole root's subtree has been visited.
1111 * helper function to be used by mem_cgroup_iter
1113 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1114 struct mem_cgroup *last_visited)
1116 struct cgroup_subsys_state *prev_css, *next_css;
1118 prev_css = last_visited ? &last_visited->css : NULL;
1120 next_css = css_next_descendant_pre(prev_css, &root->css);
1123 * Even if we found a group we have to make sure it is
1124 * alive. css && !memcg means that the groups should be
1125 * skipped and we should continue the tree walk.
1126 * last_visited css is safe to use because it is
1127 * protected by css_get and the tree walk is rcu safe.
1130 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
1132 if (css_tryget(&mem->css))
1135 prev_css = next_css;
1143 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1146 * When a group in the hierarchy below root is destroyed, the
1147 * hierarchy iterator can no longer be trusted since it might
1148 * have pointed to the destroyed group. Invalidate it.
1150 atomic_inc(&root->dead_count);
1153 static struct mem_cgroup *
1154 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1155 struct mem_cgroup *root,
1158 struct mem_cgroup *position = NULL;
1160 * A cgroup destruction happens in two stages: offlining and
1161 * release. They are separated by a RCU grace period.
1163 * If the iterator is valid, we may still race with an
1164 * offlining. The RCU lock ensures the object won't be
1165 * released, tryget will fail if we lost the race.
1167 *sequence = atomic_read(&root->dead_count);
1168 if (iter->last_dead_count == *sequence) {
1170 position = iter->last_visited;
1171 if (position && !css_tryget(&position->css))
1177 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1178 struct mem_cgroup *last_visited,
1179 struct mem_cgroup *new_position,
1183 css_put(&last_visited->css);
1185 * We store the sequence count from the time @last_visited was
1186 * loaded successfully instead of rereading it here so that we
1187 * don't lose destruction events in between. We could have
1188 * raced with the destruction of @new_position after all.
1190 iter->last_visited = new_position;
1192 iter->last_dead_count = sequence;
1196 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1197 * @root: hierarchy root
1198 * @prev: previously returned memcg, NULL on first invocation
1199 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1201 * Returns references to children of the hierarchy below @root, or
1202 * @root itself, or %NULL after a full round-trip.
1204 * Caller must pass the return value in @prev on subsequent
1205 * invocations for reference counting, or use mem_cgroup_iter_break()
1206 * to cancel a hierarchy walk before the round-trip is complete.
1208 * Reclaimers can specify a zone and a priority level in @reclaim to
1209 * divide up the memcgs in the hierarchy among all concurrent
1210 * reclaimers operating on the same zone and priority.
1212 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1213 struct mem_cgroup *prev,
1214 struct mem_cgroup_reclaim_cookie *reclaim)
1216 struct mem_cgroup *memcg = NULL;
1217 struct mem_cgroup *last_visited = NULL;
1219 if (mem_cgroup_disabled())
1223 root = root_mem_cgroup;
1225 if (prev && !reclaim)
1226 last_visited = prev;
1228 if (!root->use_hierarchy && root != root_mem_cgroup) {
1236 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1237 int uninitialized_var(seq);
1240 int nid = zone_to_nid(reclaim->zone);
1241 int zid = zone_idx(reclaim->zone);
1242 struct mem_cgroup_per_zone *mz;
1244 mz = mem_cgroup_zoneinfo(root, nid, zid);
1245 iter = &mz->reclaim_iter[reclaim->priority];
1246 if (prev && reclaim->generation != iter->generation) {
1247 iter->last_visited = NULL;
1251 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1254 memcg = __mem_cgroup_iter_next(root, last_visited);
1257 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1261 else if (!prev && memcg)
1262 reclaim->generation = iter->generation;
1271 if (prev && prev != root)
1272 css_put(&prev->css);
1278 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1279 * @root: hierarchy root
1280 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1282 void mem_cgroup_iter_break(struct mem_cgroup *root,
1283 struct mem_cgroup *prev)
1286 root = root_mem_cgroup;
1287 if (prev && prev != root)
1288 css_put(&prev->css);
1292 * Iteration constructs for visiting all cgroups (under a tree). If
1293 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1294 * be used for reference counting.
1296 #define for_each_mem_cgroup_tree(iter, root) \
1297 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1299 iter = mem_cgroup_iter(root, iter, NULL))
1301 #define for_each_mem_cgroup(iter) \
1302 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1304 iter = mem_cgroup_iter(NULL, iter, NULL))
1306 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1308 struct mem_cgroup *memcg;
1311 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1312 if (unlikely(!memcg))
1317 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1320 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1328 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1331 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1332 * @zone: zone of the wanted lruvec
1333 * @memcg: memcg of the wanted lruvec
1335 * Returns the lru list vector holding pages for the given @zone and
1336 * @mem. This can be the global zone lruvec, if the memory controller
1339 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1340 struct mem_cgroup *memcg)
1342 struct mem_cgroup_per_zone *mz;
1343 struct lruvec *lruvec;
1345 if (mem_cgroup_disabled()) {
1346 lruvec = &zone->lruvec;
1350 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1351 lruvec = &mz->lruvec;
1354 * Since a node can be onlined after the mem_cgroup was created,
1355 * we have to be prepared to initialize lruvec->zone here;
1356 * and if offlined then reonlined, we need to reinitialize it.
1358 if (unlikely(lruvec->zone != zone))
1359 lruvec->zone = zone;
1364 * Following LRU functions are allowed to be used without PCG_LOCK.
1365 * Operations are called by routine of global LRU independently from memcg.
1366 * What we have to take care of here is validness of pc->mem_cgroup.
1368 * Changes to pc->mem_cgroup happens when
1371 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1372 * It is added to LRU before charge.
1373 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1374 * When moving account, the page is not on LRU. It's isolated.
1378 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1380 * @zone: zone of the page
1382 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1384 struct mem_cgroup_per_zone *mz;
1385 struct mem_cgroup *memcg;
1386 struct page_cgroup *pc;
1387 struct lruvec *lruvec;
1389 if (mem_cgroup_disabled()) {
1390 lruvec = &zone->lruvec;
1394 pc = lookup_page_cgroup(page);
1395 memcg = pc->mem_cgroup;
1398 * Surreptitiously switch any uncharged offlist page to root:
1399 * an uncharged page off lru does nothing to secure
1400 * its former mem_cgroup from sudden removal.
1402 * Our caller holds lru_lock, and PageCgroupUsed is updated
1403 * under page_cgroup lock: between them, they make all uses
1404 * of pc->mem_cgroup safe.
1406 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1407 pc->mem_cgroup = memcg = root_mem_cgroup;
1409 mz = page_cgroup_zoneinfo(memcg, page);
1410 lruvec = &mz->lruvec;
1413 * Since a node can be onlined after the mem_cgroup was created,
1414 * we have to be prepared to initialize lruvec->zone here;
1415 * and if offlined then reonlined, we need to reinitialize it.
1417 if (unlikely(lruvec->zone != zone))
1418 lruvec->zone = zone;
1423 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1424 * @lruvec: mem_cgroup per zone lru vector
1425 * @lru: index of lru list the page is sitting on
1426 * @nr_pages: positive when adding or negative when removing
1428 * This function must be called when a page is added to or removed from an
1431 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1434 struct mem_cgroup_per_zone *mz;
1435 unsigned long *lru_size;
1437 if (mem_cgroup_disabled())
1440 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1441 lru_size = mz->lru_size + lru;
1442 *lru_size += nr_pages;
1443 VM_BUG_ON((long)(*lru_size) < 0);
1447 * Checks whether given mem is same or in the root_mem_cgroup's
1450 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1451 struct mem_cgroup *memcg)
1453 if (root_memcg == memcg)
1455 if (!root_memcg->use_hierarchy || !memcg)
1457 return css_is_ancestor(&memcg->css, &root_memcg->css);
1460 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1461 struct mem_cgroup *memcg)
1466 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1471 bool task_in_mem_cgroup(struct task_struct *task,
1472 const struct mem_cgroup *memcg)
1474 struct mem_cgroup *curr = NULL;
1475 struct task_struct *p;
1478 p = find_lock_task_mm(task);
1480 curr = try_get_mem_cgroup_from_mm(p->mm);
1484 * All threads may have already detached their mm's, but the oom
1485 * killer still needs to detect if they have already been oom
1486 * killed to prevent needlessly killing additional tasks.
1489 curr = mem_cgroup_from_task(task);
1491 css_get(&curr->css);
1497 * We should check use_hierarchy of "memcg" not "curr". Because checking
1498 * use_hierarchy of "curr" here make this function true if hierarchy is
1499 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1500 * hierarchy(even if use_hierarchy is disabled in "memcg").
1502 ret = mem_cgroup_same_or_subtree(memcg, curr);
1503 css_put(&curr->css);
1507 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1509 unsigned long inactive_ratio;
1510 unsigned long inactive;
1511 unsigned long active;
1514 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1515 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1517 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1519 inactive_ratio = int_sqrt(10 * gb);
1523 return inactive * inactive_ratio < active;
1526 #define mem_cgroup_from_res_counter(counter, member) \
1527 container_of(counter, struct mem_cgroup, member)
1530 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1531 * @memcg: the memory cgroup
1533 * Returns the maximum amount of memory @mem can be charged with, in
1536 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1538 unsigned long long margin;
1540 margin = res_counter_margin(&memcg->res);
1541 if (do_swap_account)
1542 margin = min(margin, res_counter_margin(&memcg->memsw));
1543 return margin >> PAGE_SHIFT;
1546 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1549 if (!css_parent(&memcg->css))
1550 return vm_swappiness;
1552 return memcg->swappiness;
1556 * memcg->moving_account is used for checking possibility that some thread is
1557 * calling move_account(). When a thread on CPU-A starts moving pages under
1558 * a memcg, other threads should check memcg->moving_account under
1559 * rcu_read_lock(), like this:
1563 * memcg->moving_account+1 if (memcg->mocing_account)
1565 * synchronize_rcu() update something.
1570 /* for quick checking without looking up memcg */
1571 atomic_t memcg_moving __read_mostly;
1573 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1575 atomic_inc(&memcg_moving);
1576 atomic_inc(&memcg->moving_account);
1580 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1583 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1584 * We check NULL in callee rather than caller.
1587 atomic_dec(&memcg_moving);
1588 atomic_dec(&memcg->moving_account);
1593 * 2 routines for checking "mem" is under move_account() or not.
1595 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1596 * is used for avoiding races in accounting. If true,
1597 * pc->mem_cgroup may be overwritten.
1599 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1600 * under hierarchy of moving cgroups. This is for
1601 * waiting at hith-memory prressure caused by "move".
1604 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1606 VM_BUG_ON(!rcu_read_lock_held());
1607 return atomic_read(&memcg->moving_account) > 0;
1610 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1612 struct mem_cgroup *from;
1613 struct mem_cgroup *to;
1616 * Unlike task_move routines, we access mc.to, mc.from not under
1617 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1619 spin_lock(&mc.lock);
1625 ret = mem_cgroup_same_or_subtree(memcg, from)
1626 || mem_cgroup_same_or_subtree(memcg, to);
1628 spin_unlock(&mc.lock);
1632 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1634 if (mc.moving_task && current != mc.moving_task) {
1635 if (mem_cgroup_under_move(memcg)) {
1637 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1638 /* moving charge context might have finished. */
1641 finish_wait(&mc.waitq, &wait);
1649 * Take this lock when
1650 * - a code tries to modify page's memcg while it's USED.
1651 * - a code tries to modify page state accounting in a memcg.
1652 * see mem_cgroup_stolen(), too.
1654 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1655 unsigned long *flags)
1657 spin_lock_irqsave(&memcg->move_lock, *flags);
1660 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1661 unsigned long *flags)
1663 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1666 #define K(x) ((x) << (PAGE_SHIFT-10))
1668 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1669 * @memcg: The memory cgroup that went over limit
1670 * @p: Task that is going to be killed
1672 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1675 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1677 struct cgroup *task_cgrp;
1678 struct cgroup *mem_cgrp;
1680 * Need a buffer in BSS, can't rely on allocations. The code relies
1681 * on the assumption that OOM is serialized for memory controller.
1682 * If this assumption is broken, revisit this code.
1684 static char memcg_name[PATH_MAX];
1686 struct mem_cgroup *iter;
1694 mem_cgrp = memcg->css.cgroup;
1695 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1697 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1700 * Unfortunately, we are unable to convert to a useful name
1701 * But we'll still print out the usage information
1708 pr_info("Task in %s killed", memcg_name);
1711 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1719 * Continues from above, so we don't need an KERN_ level
1721 pr_cont(" as a result of limit of %s\n", memcg_name);
1724 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1725 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1726 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1727 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1728 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1729 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1730 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1731 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1732 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1733 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1734 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1735 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1737 for_each_mem_cgroup_tree(iter, memcg) {
1738 pr_info("Memory cgroup stats");
1741 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1743 pr_cont(" for %s", memcg_name);
1747 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1748 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1750 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1751 K(mem_cgroup_read_stat(iter, i)));
1754 for (i = 0; i < NR_LRU_LISTS; i++)
1755 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1756 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1763 * This function returns the number of memcg under hierarchy tree. Returns
1764 * 1(self count) if no children.
1766 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1769 struct mem_cgroup *iter;
1771 for_each_mem_cgroup_tree(iter, memcg)
1777 * Return the memory (and swap, if configured) limit for a memcg.
1779 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1783 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1786 * Do not consider swap space if we cannot swap due to swappiness
1788 if (mem_cgroup_swappiness(memcg)) {
1791 limit += total_swap_pages << PAGE_SHIFT;
1792 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1795 * If memsw is finite and limits the amount of swap space
1796 * available to this memcg, return that limit.
1798 limit = min(limit, memsw);
1804 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1807 struct mem_cgroup *iter;
1808 unsigned long chosen_points = 0;
1809 unsigned long totalpages;
1810 unsigned int points = 0;
1811 struct task_struct *chosen = NULL;
1814 * If current has a pending SIGKILL or is exiting, then automatically
1815 * select it. The goal is to allow it to allocate so that it may
1816 * quickly exit and free its memory.
1818 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1819 set_thread_flag(TIF_MEMDIE);
1823 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1824 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1825 for_each_mem_cgroup_tree(iter, memcg) {
1826 struct css_task_iter it;
1827 struct task_struct *task;
1829 css_task_iter_start(&iter->css, &it);
1830 while ((task = css_task_iter_next(&it))) {
1831 switch (oom_scan_process_thread(task, totalpages, NULL,
1833 case OOM_SCAN_SELECT:
1835 put_task_struct(chosen);
1837 chosen_points = ULONG_MAX;
1838 get_task_struct(chosen);
1840 case OOM_SCAN_CONTINUE:
1842 case OOM_SCAN_ABORT:
1843 css_task_iter_end(&it);
1844 mem_cgroup_iter_break(memcg, iter);
1846 put_task_struct(chosen);
1851 points = oom_badness(task, memcg, NULL, totalpages);
1852 if (points > chosen_points) {
1854 put_task_struct(chosen);
1856 chosen_points = points;
1857 get_task_struct(chosen);
1860 css_task_iter_end(&it);
1865 points = chosen_points * 1000 / totalpages;
1866 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1867 NULL, "Memory cgroup out of memory");
1870 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1872 unsigned long flags)
1874 unsigned long total = 0;
1875 bool noswap = false;
1878 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1880 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1883 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1885 drain_all_stock_async(memcg);
1886 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1888 * Allow limit shrinkers, which are triggered directly
1889 * by userspace, to catch signals and stop reclaim
1890 * after minimal progress, regardless of the margin.
1892 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1894 if (mem_cgroup_margin(memcg))
1897 * If nothing was reclaimed after two attempts, there
1898 * may be no reclaimable pages in this hierarchy.
1907 * test_mem_cgroup_node_reclaimable
1908 * @memcg: the target memcg
1909 * @nid: the node ID to be checked.
1910 * @noswap : specify true here if the user wants flle only information.
1912 * This function returns whether the specified memcg contains any
1913 * reclaimable pages on a node. Returns true if there are any reclaimable
1914 * pages in the node.
1916 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1917 int nid, bool noswap)
1919 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1921 if (noswap || !total_swap_pages)
1923 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1928 #if MAX_NUMNODES > 1
1931 * Always updating the nodemask is not very good - even if we have an empty
1932 * list or the wrong list here, we can start from some node and traverse all
1933 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1936 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1940 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1941 * pagein/pageout changes since the last update.
1943 if (!atomic_read(&memcg->numainfo_events))
1945 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1948 /* make a nodemask where this memcg uses memory from */
1949 memcg->scan_nodes = node_states[N_MEMORY];
1951 for_each_node_mask(nid, node_states[N_MEMORY]) {
1953 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1954 node_clear(nid, memcg->scan_nodes);
1957 atomic_set(&memcg->numainfo_events, 0);
1958 atomic_set(&memcg->numainfo_updating, 0);
1962 * Selecting a node where we start reclaim from. Because what we need is just
1963 * reducing usage counter, start from anywhere is O,K. Considering
1964 * memory reclaim from current node, there are pros. and cons.
1966 * Freeing memory from current node means freeing memory from a node which
1967 * we'll use or we've used. So, it may make LRU bad. And if several threads
1968 * hit limits, it will see a contention on a node. But freeing from remote
1969 * node means more costs for memory reclaim because of memory latency.
1971 * Now, we use round-robin. Better algorithm is welcomed.
1973 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1977 mem_cgroup_may_update_nodemask(memcg);
1978 node = memcg->last_scanned_node;
1980 node = next_node(node, memcg->scan_nodes);
1981 if (node == MAX_NUMNODES)
1982 node = first_node(memcg->scan_nodes);
1984 * We call this when we hit limit, not when pages are added to LRU.
1985 * No LRU may hold pages because all pages are UNEVICTABLE or
1986 * memcg is too small and all pages are not on LRU. In that case,
1987 * we use curret node.
1989 if (unlikely(node == MAX_NUMNODES))
1990 node = numa_node_id();
1992 memcg->last_scanned_node = node;
1997 * Check all nodes whether it contains reclaimable pages or not.
1998 * For quick scan, we make use of scan_nodes. This will allow us to skip
1999 * unused nodes. But scan_nodes is lazily updated and may not cotain
2000 * enough new information. We need to do double check.
2002 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2007 * quick check...making use of scan_node.
2008 * We can skip unused nodes.
2010 if (!nodes_empty(memcg->scan_nodes)) {
2011 for (nid = first_node(memcg->scan_nodes);
2013 nid = next_node(nid, memcg->scan_nodes)) {
2015 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2020 * Check rest of nodes.
2022 for_each_node_state(nid, N_MEMORY) {
2023 if (node_isset(nid, memcg->scan_nodes))
2025 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2032 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2037 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2039 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2043 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2046 unsigned long *total_scanned)
2048 struct mem_cgroup *victim = NULL;
2051 unsigned long excess;
2052 unsigned long nr_scanned;
2053 struct mem_cgroup_reclaim_cookie reclaim = {
2058 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2061 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2066 * If we have not been able to reclaim
2067 * anything, it might because there are
2068 * no reclaimable pages under this hierarchy
2073 * We want to do more targeted reclaim.
2074 * excess >> 2 is not to excessive so as to
2075 * reclaim too much, nor too less that we keep
2076 * coming back to reclaim from this cgroup
2078 if (total >= (excess >> 2) ||
2079 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2084 if (!mem_cgroup_reclaimable(victim, false))
2086 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2088 *total_scanned += nr_scanned;
2089 if (!res_counter_soft_limit_excess(&root_memcg->res))
2092 mem_cgroup_iter_break(root_memcg, victim);
2096 #ifdef CONFIG_LOCKDEP
2097 static struct lockdep_map memcg_oom_lock_dep_map = {
2098 .name = "memcg_oom_lock",
2102 static DEFINE_SPINLOCK(memcg_oom_lock);
2105 * Check OOM-Killer is already running under our hierarchy.
2106 * If someone is running, return false.
2108 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2110 struct mem_cgroup *iter, *failed = NULL;
2112 spin_lock(&memcg_oom_lock);
2114 for_each_mem_cgroup_tree(iter, memcg) {
2115 if (iter->oom_lock) {
2117 * this subtree of our hierarchy is already locked
2118 * so we cannot give a lock.
2121 mem_cgroup_iter_break(memcg, iter);
2124 iter->oom_lock = true;
2129 * OK, we failed to lock the whole subtree so we have
2130 * to clean up what we set up to the failing subtree
2132 for_each_mem_cgroup_tree(iter, memcg) {
2133 if (iter == failed) {
2134 mem_cgroup_iter_break(memcg, iter);
2137 iter->oom_lock = false;
2140 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2142 spin_unlock(&memcg_oom_lock);
2147 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2149 struct mem_cgroup *iter;
2151 spin_lock(&memcg_oom_lock);
2152 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2153 for_each_mem_cgroup_tree(iter, memcg)
2154 iter->oom_lock = false;
2155 spin_unlock(&memcg_oom_lock);
2158 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2160 struct mem_cgroup *iter;
2162 for_each_mem_cgroup_tree(iter, memcg)
2163 atomic_inc(&iter->under_oom);
2166 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2168 struct mem_cgroup *iter;
2171 * When a new child is created while the hierarchy is under oom,
2172 * mem_cgroup_oom_lock() may not be called. We have to use
2173 * atomic_add_unless() here.
2175 for_each_mem_cgroup_tree(iter, memcg)
2176 atomic_add_unless(&iter->under_oom, -1, 0);
2179 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2181 struct oom_wait_info {
2182 struct mem_cgroup *memcg;
2186 static int memcg_oom_wake_function(wait_queue_t *wait,
2187 unsigned mode, int sync, void *arg)
2189 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2190 struct mem_cgroup *oom_wait_memcg;
2191 struct oom_wait_info *oom_wait_info;
2193 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2194 oom_wait_memcg = oom_wait_info->memcg;
2197 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2198 * Then we can use css_is_ancestor without taking care of RCU.
2200 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2201 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2203 return autoremove_wake_function(wait, mode, sync, arg);
2206 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2208 atomic_inc(&memcg->oom_wakeups);
2209 /* for filtering, pass "memcg" as argument. */
2210 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2213 static void memcg_oom_recover(struct mem_cgroup *memcg)
2215 if (memcg && atomic_read(&memcg->under_oom))
2216 memcg_wakeup_oom(memcg);
2219 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2221 if (!current->memcg_oom.may_oom)
2224 * We are in the middle of the charge context here, so we
2225 * don't want to block when potentially sitting on a callstack
2226 * that holds all kinds of filesystem and mm locks.
2228 * Also, the caller may handle a failed allocation gracefully
2229 * (like optional page cache readahead) and so an OOM killer
2230 * invocation might not even be necessary.
2232 * That's why we don't do anything here except remember the
2233 * OOM context and then deal with it at the end of the page
2234 * fault when the stack is unwound, the locks are released,
2235 * and when we know whether the fault was overall successful.
2237 css_get(&memcg->css);
2238 current->memcg_oom.memcg = memcg;
2239 current->memcg_oom.gfp_mask = mask;
2240 current->memcg_oom.order = order;
2244 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2245 * @handle: actually kill/wait or just clean up the OOM state
2247 * This has to be called at the end of a page fault if the memcg OOM
2248 * handler was enabled.
2250 * Memcg supports userspace OOM handling where failed allocations must
2251 * sleep on a waitqueue until the userspace task resolves the
2252 * situation. Sleeping directly in the charge context with all kinds
2253 * of locks held is not a good idea, instead we remember an OOM state
2254 * in the task and mem_cgroup_oom_synchronize() has to be called at
2255 * the end of the page fault to complete the OOM handling.
2257 * Returns %true if an ongoing memcg OOM situation was detected and
2258 * completed, %false otherwise.
2260 bool mem_cgroup_oom_synchronize(bool handle)
2262 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2263 struct oom_wait_info owait;
2266 /* OOM is global, do not handle */
2273 owait.memcg = memcg;
2274 owait.wait.flags = 0;
2275 owait.wait.func = memcg_oom_wake_function;
2276 owait.wait.private = current;
2277 INIT_LIST_HEAD(&owait.wait.task_list);
2279 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2280 mem_cgroup_mark_under_oom(memcg);
2282 locked = mem_cgroup_oom_trylock(memcg);
2285 mem_cgroup_oom_notify(memcg);
2287 if (locked && !memcg->oom_kill_disable) {
2288 mem_cgroup_unmark_under_oom(memcg);
2289 finish_wait(&memcg_oom_waitq, &owait.wait);
2290 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2291 current->memcg_oom.order);
2294 mem_cgroup_unmark_under_oom(memcg);
2295 finish_wait(&memcg_oom_waitq, &owait.wait);
2299 mem_cgroup_oom_unlock(memcg);
2301 * There is no guarantee that an OOM-lock contender
2302 * sees the wakeups triggered by the OOM kill
2303 * uncharges. Wake any sleepers explicitely.
2305 memcg_oom_recover(memcg);
2308 current->memcg_oom.memcg = NULL;
2309 css_put(&memcg->css);
2314 * Currently used to update mapped file statistics, but the routine can be
2315 * generalized to update other statistics as well.
2317 * Notes: Race condition
2319 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2320 * it tends to be costly. But considering some conditions, we doesn't need
2321 * to do so _always_.
2323 * Considering "charge", lock_page_cgroup() is not required because all
2324 * file-stat operations happen after a page is attached to radix-tree. There
2325 * are no race with "charge".
2327 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2328 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2329 * if there are race with "uncharge". Statistics itself is properly handled
2332 * Considering "move", this is an only case we see a race. To make the race
2333 * small, we check mm->moving_account and detect there are possibility of race
2334 * If there is, we take a lock.
2337 void __mem_cgroup_begin_update_page_stat(struct page *page,
2338 bool *locked, unsigned long *flags)
2340 struct mem_cgroup *memcg;
2341 struct page_cgroup *pc;
2343 pc = lookup_page_cgroup(page);
2345 memcg = pc->mem_cgroup;
2346 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2349 * If this memory cgroup is not under account moving, we don't
2350 * need to take move_lock_mem_cgroup(). Because we already hold
2351 * rcu_read_lock(), any calls to move_account will be delayed until
2352 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2354 if (!mem_cgroup_stolen(memcg))
2357 move_lock_mem_cgroup(memcg, flags);
2358 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2359 move_unlock_mem_cgroup(memcg, flags);
2365 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2367 struct page_cgroup *pc = lookup_page_cgroup(page);
2370 * It's guaranteed that pc->mem_cgroup never changes while
2371 * lock is held because a routine modifies pc->mem_cgroup
2372 * should take move_lock_mem_cgroup().
2374 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2377 void mem_cgroup_update_page_stat(struct page *page,
2378 enum mem_cgroup_stat_index idx, int val)
2380 struct mem_cgroup *memcg;
2381 struct page_cgroup *pc = lookup_page_cgroup(page);
2382 unsigned long uninitialized_var(flags);
2384 if (mem_cgroup_disabled())
2387 VM_BUG_ON(!rcu_read_lock_held());
2388 memcg = pc->mem_cgroup;
2389 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2392 this_cpu_add(memcg->stat->count[idx], val);
2396 * size of first charge trial. "32" comes from vmscan.c's magic value.
2397 * TODO: maybe necessary to use big numbers in big irons.
2399 #define CHARGE_BATCH 32U
2400 struct memcg_stock_pcp {
2401 struct mem_cgroup *cached; /* this never be root cgroup */
2402 unsigned int nr_pages;
2403 struct work_struct work;
2404 unsigned long flags;
2405 #define FLUSHING_CACHED_CHARGE 0
2407 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2408 static DEFINE_MUTEX(percpu_charge_mutex);
2411 * consume_stock: Try to consume stocked charge on this cpu.
2412 * @memcg: memcg to consume from.
2413 * @nr_pages: how many pages to charge.
2415 * The charges will only happen if @memcg matches the current cpu's memcg
2416 * stock, and at least @nr_pages are available in that stock. Failure to
2417 * service an allocation will refill the stock.
2419 * returns true if successful, false otherwise.
2421 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2423 struct memcg_stock_pcp *stock;
2426 if (nr_pages > CHARGE_BATCH)
2429 stock = &get_cpu_var(memcg_stock);
2430 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2431 stock->nr_pages -= nr_pages;
2432 else /* need to call res_counter_charge */
2434 put_cpu_var(memcg_stock);
2439 * Returns stocks cached in percpu to res_counter and reset cached information.
2441 static void drain_stock(struct memcg_stock_pcp *stock)
2443 struct mem_cgroup *old = stock->cached;
2445 if (stock->nr_pages) {
2446 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2448 res_counter_uncharge(&old->res, bytes);
2449 if (do_swap_account)
2450 res_counter_uncharge(&old->memsw, bytes);
2451 stock->nr_pages = 0;
2453 stock->cached = NULL;
2457 * This must be called under preempt disabled or must be called by
2458 * a thread which is pinned to local cpu.
2460 static void drain_local_stock(struct work_struct *dummy)
2462 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2464 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2467 static void __init memcg_stock_init(void)
2471 for_each_possible_cpu(cpu) {
2472 struct memcg_stock_pcp *stock =
2473 &per_cpu(memcg_stock, cpu);
2474 INIT_WORK(&stock->work, drain_local_stock);
2479 * Cache charges(val) which is from res_counter, to local per_cpu area.
2480 * This will be consumed by consume_stock() function, later.
2482 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2484 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2486 if (stock->cached != memcg) { /* reset if necessary */
2488 stock->cached = memcg;
2490 stock->nr_pages += nr_pages;
2491 put_cpu_var(memcg_stock);
2495 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2496 * of the hierarchy under it. sync flag says whether we should block
2497 * until the work is done.
2499 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2503 /* Notify other cpus that system-wide "drain" is running */
2506 for_each_online_cpu(cpu) {
2507 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2508 struct mem_cgroup *memcg;
2510 memcg = stock->cached;
2511 if (!memcg || !stock->nr_pages)
2513 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2515 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2517 drain_local_stock(&stock->work);
2519 schedule_work_on(cpu, &stock->work);
2527 for_each_online_cpu(cpu) {
2528 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2529 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2530 flush_work(&stock->work);
2537 * Tries to drain stocked charges in other cpus. This function is asynchronous
2538 * and just put a work per cpu for draining localy on each cpu. Caller can
2539 * expects some charges will be back to res_counter later but cannot wait for
2542 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2545 * If someone calls draining, avoid adding more kworker runs.
2547 if (!mutex_trylock(&percpu_charge_mutex))
2549 drain_all_stock(root_memcg, false);
2550 mutex_unlock(&percpu_charge_mutex);
2553 /* This is a synchronous drain interface. */
2554 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2556 /* called when force_empty is called */
2557 mutex_lock(&percpu_charge_mutex);
2558 drain_all_stock(root_memcg, true);
2559 mutex_unlock(&percpu_charge_mutex);
2563 * This function drains percpu counter value from DEAD cpu and
2564 * move it to local cpu. Note that this function can be preempted.
2566 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2570 spin_lock(&memcg->pcp_counter_lock);
2571 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2572 long x = per_cpu(memcg->stat->count[i], cpu);
2574 per_cpu(memcg->stat->count[i], cpu) = 0;
2575 memcg->nocpu_base.count[i] += x;
2577 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2578 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2580 per_cpu(memcg->stat->events[i], cpu) = 0;
2581 memcg->nocpu_base.events[i] += x;
2583 spin_unlock(&memcg->pcp_counter_lock);
2586 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2587 unsigned long action,
2590 int cpu = (unsigned long)hcpu;
2591 struct memcg_stock_pcp *stock;
2592 struct mem_cgroup *iter;
2594 if (action == CPU_ONLINE)
2597 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2600 for_each_mem_cgroup(iter)
2601 mem_cgroup_drain_pcp_counter(iter, cpu);
2603 stock = &per_cpu(memcg_stock, cpu);
2609 /* See __mem_cgroup_try_charge() for details */
2611 CHARGE_OK, /* success */
2612 CHARGE_RETRY, /* need to retry but retry is not bad */
2613 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2614 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2617 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2618 unsigned int nr_pages, unsigned int min_pages,
2621 unsigned long csize = nr_pages * PAGE_SIZE;
2622 struct mem_cgroup *mem_over_limit;
2623 struct res_counter *fail_res;
2624 unsigned long flags = 0;
2627 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2630 if (!do_swap_account)
2632 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2636 res_counter_uncharge(&memcg->res, csize);
2637 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2638 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2640 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2642 * Never reclaim on behalf of optional batching, retry with a
2643 * single page instead.
2645 if (nr_pages > min_pages)
2646 return CHARGE_RETRY;
2648 if (!(gfp_mask & __GFP_WAIT))
2649 return CHARGE_WOULDBLOCK;
2651 if (gfp_mask & __GFP_NORETRY)
2652 return CHARGE_NOMEM;
2654 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2655 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2656 return CHARGE_RETRY;
2658 * Even though the limit is exceeded at this point, reclaim
2659 * may have been able to free some pages. Retry the charge
2660 * before killing the task.
2662 * Only for regular pages, though: huge pages are rather
2663 * unlikely to succeed so close to the limit, and we fall back
2664 * to regular pages anyway in case of failure.
2666 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2667 return CHARGE_RETRY;
2670 * At task move, charge accounts can be doubly counted. So, it's
2671 * better to wait until the end of task_move if something is going on.
2673 if (mem_cgroup_wait_acct_move(mem_over_limit))
2674 return CHARGE_RETRY;
2677 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2679 return CHARGE_NOMEM;
2683 * __mem_cgroup_try_charge() does
2684 * 1. detect memcg to be charged against from passed *mm and *ptr,
2685 * 2. update res_counter
2686 * 3. call memory reclaim if necessary.
2688 * In some special case, if the task is fatal, fatal_signal_pending() or
2689 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2690 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2691 * as possible without any hazards. 2: all pages should have a valid
2692 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2693 * pointer, that is treated as a charge to root_mem_cgroup.
2695 * So __mem_cgroup_try_charge() will return
2696 * 0 ... on success, filling *ptr with a valid memcg pointer.
2697 * -ENOMEM ... charge failure because of resource limits.
2698 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2700 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2701 * the oom-killer can be invoked.
2703 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2705 unsigned int nr_pages,
2706 struct mem_cgroup **ptr,
2709 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2710 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2711 struct mem_cgroup *memcg = NULL;
2715 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2716 * in system level. So, allow to go ahead dying process in addition to
2719 if (unlikely(test_thread_flag(TIF_MEMDIE)
2720 || fatal_signal_pending(current)))
2723 if (unlikely(task_in_memcg_oom(current)))
2727 * We always charge the cgroup the mm_struct belongs to.
2728 * The mm_struct's mem_cgroup changes on task migration if the
2729 * thread group leader migrates. It's possible that mm is not
2730 * set, if so charge the root memcg (happens for pagecache usage).
2733 *ptr = root_mem_cgroup;
2735 if (*ptr) { /* css should be a valid one */
2737 if (mem_cgroup_is_root(memcg))
2739 if (consume_stock(memcg, nr_pages))
2741 css_get(&memcg->css);
2743 struct task_struct *p;
2746 p = rcu_dereference(mm->owner);
2748 * Because we don't have task_lock(), "p" can exit.
2749 * In that case, "memcg" can point to root or p can be NULL with
2750 * race with swapoff. Then, we have small risk of mis-accouning.
2751 * But such kind of mis-account by race always happens because
2752 * we don't have cgroup_mutex(). It's overkill and we allo that
2754 * (*) swapoff at el will charge against mm-struct not against
2755 * task-struct. So, mm->owner can be NULL.
2757 memcg = mem_cgroup_from_task(p);
2759 memcg = root_mem_cgroup;
2760 if (mem_cgroup_is_root(memcg)) {
2764 if (consume_stock(memcg, nr_pages)) {
2766 * It seems dagerous to access memcg without css_get().
2767 * But considering how consume_stok works, it's not
2768 * necessary. If consume_stock success, some charges
2769 * from this memcg are cached on this cpu. So, we
2770 * don't need to call css_get()/css_tryget() before
2771 * calling consume_stock().
2776 /* after here, we may be blocked. we need to get refcnt */
2777 if (!css_tryget(&memcg->css)) {
2785 bool invoke_oom = oom && !nr_oom_retries;
2787 /* If killed, bypass charge */
2788 if (fatal_signal_pending(current)) {
2789 css_put(&memcg->css);
2793 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2794 nr_pages, invoke_oom);
2798 case CHARGE_RETRY: /* not in OOM situation but retry */
2800 css_put(&memcg->css);
2803 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2804 css_put(&memcg->css);
2806 case CHARGE_NOMEM: /* OOM routine works */
2807 if (!oom || invoke_oom) {
2808 css_put(&memcg->css);
2814 } while (ret != CHARGE_OK);
2816 if (batch > nr_pages)
2817 refill_stock(memcg, batch - nr_pages);
2818 css_put(&memcg->css);
2823 if (!(gfp_mask & __GFP_NOFAIL)) {
2828 *ptr = root_mem_cgroup;
2833 * Somemtimes we have to undo a charge we got by try_charge().
2834 * This function is for that and do uncharge, put css's refcnt.
2835 * gotten by try_charge().
2837 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2838 unsigned int nr_pages)
2840 if (!mem_cgroup_is_root(memcg)) {
2841 unsigned long bytes = nr_pages * PAGE_SIZE;
2843 res_counter_uncharge(&memcg->res, bytes);
2844 if (do_swap_account)
2845 res_counter_uncharge(&memcg->memsw, bytes);
2850 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2851 * This is useful when moving usage to parent cgroup.
2853 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2854 unsigned int nr_pages)
2856 unsigned long bytes = nr_pages * PAGE_SIZE;
2858 if (mem_cgroup_is_root(memcg))
2861 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2862 if (do_swap_account)
2863 res_counter_uncharge_until(&memcg->memsw,
2864 memcg->memsw.parent, bytes);
2868 * A helper function to get mem_cgroup from ID. must be called under
2869 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2870 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2871 * called against removed memcg.)
2873 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2875 struct cgroup_subsys_state *css;
2877 /* ID 0 is unused ID */
2880 css = css_lookup(&mem_cgroup_subsys, id);
2883 return mem_cgroup_from_css(css);
2886 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2888 struct mem_cgroup *memcg = NULL;
2889 struct page_cgroup *pc;
2893 VM_BUG_ON(!PageLocked(page));
2895 pc = lookup_page_cgroup(page);
2896 lock_page_cgroup(pc);
2897 if (PageCgroupUsed(pc)) {
2898 memcg = pc->mem_cgroup;
2899 if (memcg && !css_tryget(&memcg->css))
2901 } else if (PageSwapCache(page)) {
2902 ent.val = page_private(page);
2903 id = lookup_swap_cgroup_id(ent);
2905 memcg = mem_cgroup_lookup(id);
2906 if (memcg && !css_tryget(&memcg->css))
2910 unlock_page_cgroup(pc);
2914 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2916 unsigned int nr_pages,
2917 enum charge_type ctype,
2920 struct page_cgroup *pc = lookup_page_cgroup(page);
2921 struct zone *uninitialized_var(zone);
2922 struct lruvec *lruvec;
2923 bool was_on_lru = false;
2926 lock_page_cgroup(pc);
2927 VM_BUG_ON(PageCgroupUsed(pc));
2929 * we don't need page_cgroup_lock about tail pages, becase they are not
2930 * accessed by any other context at this point.
2934 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2935 * may already be on some other mem_cgroup's LRU. Take care of it.
2938 zone = page_zone(page);
2939 spin_lock_irq(&zone->lru_lock);
2940 if (PageLRU(page)) {
2941 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2943 del_page_from_lru_list(page, lruvec, page_lru(page));
2948 pc->mem_cgroup = memcg;
2950 * We access a page_cgroup asynchronously without lock_page_cgroup().
2951 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2952 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2953 * before USED bit, we need memory barrier here.
2954 * See mem_cgroup_add_lru_list(), etc.
2957 SetPageCgroupUsed(pc);
2961 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2962 VM_BUG_ON(PageLRU(page));
2964 add_page_to_lru_list(page, lruvec, page_lru(page));
2966 spin_unlock_irq(&zone->lru_lock);
2969 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2974 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2975 unlock_page_cgroup(pc);
2978 * "charge_statistics" updated event counter. Then, check it.
2979 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2980 * if they exceeds softlimit.
2982 memcg_check_events(memcg, page);
2985 static DEFINE_MUTEX(set_limit_mutex);
2987 #ifdef CONFIG_MEMCG_KMEM
2988 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2990 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2991 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2995 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2996 * in the memcg_cache_params struct.
2998 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
3000 struct kmem_cache *cachep;
3002 VM_BUG_ON(p->is_root_cache);
3003 cachep = p->root_cache;
3004 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
3007 #ifdef CONFIG_SLABINFO
3008 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
3009 struct cftype *cft, struct seq_file *m)
3011 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3012 struct memcg_cache_params *params;
3014 if (!memcg_can_account_kmem(memcg))
3017 print_slabinfo_header(m);
3019 mutex_lock(&memcg->slab_caches_mutex);
3020 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
3021 cache_show(memcg_params_to_cache(params), m);
3022 mutex_unlock(&memcg->slab_caches_mutex);
3028 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3030 struct res_counter *fail_res;
3031 struct mem_cgroup *_memcg;
3035 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3040 * Conditions under which we can wait for the oom_killer. Those are
3041 * the same conditions tested by the core page allocator
3043 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
3046 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
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);
3098 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3103 mutex_lock(&memcg->slab_caches_mutex);
3104 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3105 mutex_unlock(&memcg->slab_caches_mutex);
3109 * helper for acessing a memcg's index. It will be used as an index in the
3110 * child cache array in kmem_cache, and also to derive its name. This function
3111 * will return -1 when this is not a kmem-limited memcg.
3113 int memcg_cache_id(struct mem_cgroup *memcg)
3115 return memcg ? memcg->kmemcg_id : -1;
3119 * This ends up being protected by the set_limit mutex, during normal
3120 * operation, because that is its main call site.
3122 * But when we create a new cache, we can call this as well if its parent
3123 * is kmem-limited. That will have to hold set_limit_mutex as well.
3125 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3129 num = ida_simple_get(&kmem_limited_groups,
3130 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3134 * After this point, kmem_accounted (that we test atomically in
3135 * the beginning of this conditional), is no longer 0. This
3136 * guarantees only one process will set the following boolean
3137 * to true. We don't need test_and_set because we're protected
3138 * by the set_limit_mutex anyway.
3140 memcg_kmem_set_activated(memcg);
3142 ret = memcg_update_all_caches(num+1);
3144 ida_simple_remove(&kmem_limited_groups, num);
3145 memcg_kmem_clear_activated(memcg);
3149 memcg->kmemcg_id = num;
3150 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3151 mutex_init(&memcg->slab_caches_mutex);
3155 static size_t memcg_caches_array_size(int num_groups)
3158 if (num_groups <= 0)
3161 size = 2 * num_groups;
3162 if (size < MEMCG_CACHES_MIN_SIZE)
3163 size = MEMCG_CACHES_MIN_SIZE;
3164 else if (size > MEMCG_CACHES_MAX_SIZE)
3165 size = MEMCG_CACHES_MAX_SIZE;
3171 * We should update the current array size iff all caches updates succeed. This
3172 * can only be done from the slab side. The slab mutex needs to be held when
3175 void memcg_update_array_size(int num)
3177 if (num > memcg_limited_groups_array_size)
3178 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3181 static void kmem_cache_destroy_work_func(struct work_struct *w);
3183 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3185 struct memcg_cache_params *cur_params = s->memcg_params;
3187 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3189 if (num_groups > memcg_limited_groups_array_size) {
3191 ssize_t size = memcg_caches_array_size(num_groups);
3193 size *= sizeof(void *);
3194 size += offsetof(struct memcg_cache_params, memcg_caches);
3196 s->memcg_params = kzalloc(size, GFP_KERNEL);
3197 if (!s->memcg_params) {
3198 s->memcg_params = cur_params;
3202 s->memcg_params->is_root_cache = true;
3205 * There is the chance it will be bigger than
3206 * memcg_limited_groups_array_size, if we failed an allocation
3207 * in a cache, in which case all caches updated before it, will
3208 * have a bigger array.
3210 * But if that is the case, the data after
3211 * memcg_limited_groups_array_size is certainly unused
3213 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3214 if (!cur_params->memcg_caches[i])
3216 s->memcg_params->memcg_caches[i] =
3217 cur_params->memcg_caches[i];
3221 * Ideally, we would wait until all caches succeed, and only
3222 * then free the old one. But this is not worth the extra
3223 * pointer per-cache we'd have to have for this.
3225 * It is not a big deal if some caches are left with a size
3226 * bigger than the others. And all updates will reset this
3234 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3235 struct kmem_cache *root_cache)
3239 if (!memcg_kmem_enabled())
3243 size = offsetof(struct memcg_cache_params, memcg_caches);
3244 size += memcg_limited_groups_array_size * sizeof(void *);
3246 size = sizeof(struct memcg_cache_params);
3248 s->memcg_params = kzalloc(size, GFP_KERNEL);
3249 if (!s->memcg_params)
3253 s->memcg_params->memcg = memcg;
3254 s->memcg_params->root_cache = root_cache;
3255 INIT_WORK(&s->memcg_params->destroy,
3256 kmem_cache_destroy_work_func);
3258 s->memcg_params->is_root_cache = true;
3263 void memcg_release_cache(struct kmem_cache *s)
3265 struct kmem_cache *root;
3266 struct mem_cgroup *memcg;
3270 * This happens, for instance, when a root cache goes away before we
3273 if (!s->memcg_params)
3276 if (s->memcg_params->is_root_cache)
3279 memcg = s->memcg_params->memcg;
3280 id = memcg_cache_id(memcg);
3282 root = s->memcg_params->root_cache;
3283 root->memcg_params->memcg_caches[id] = NULL;
3285 mutex_lock(&memcg->slab_caches_mutex);
3286 list_del(&s->memcg_params->list);
3287 mutex_unlock(&memcg->slab_caches_mutex);
3289 css_put(&memcg->css);
3291 kfree(s->memcg_params);
3295 * During the creation a new cache, we need to disable our accounting mechanism
3296 * altogether. This is true even if we are not creating, but rather just
3297 * enqueing new caches to be created.
3299 * This is because that process will trigger allocations; some visible, like
3300 * explicit kmallocs to auxiliary data structures, name strings and internal
3301 * cache structures; some well concealed, like INIT_WORK() that can allocate
3302 * objects during debug.
3304 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3305 * to it. This may not be a bounded recursion: since the first cache creation
3306 * failed to complete (waiting on the allocation), we'll just try to create the
3307 * cache again, failing at the same point.
3309 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3310 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3311 * inside the following two functions.
3313 static inline void memcg_stop_kmem_account(void)
3315 VM_BUG_ON(!current->mm);
3316 current->memcg_kmem_skip_account++;
3319 static inline void memcg_resume_kmem_account(void)
3321 VM_BUG_ON(!current->mm);
3322 current->memcg_kmem_skip_account--;
3325 static void kmem_cache_destroy_work_func(struct work_struct *w)
3327 struct kmem_cache *cachep;
3328 struct memcg_cache_params *p;
3330 p = container_of(w, struct memcg_cache_params, destroy);
3332 cachep = memcg_params_to_cache(p);
3335 * If we get down to 0 after shrink, we could delete right away.
3336 * However, memcg_release_pages() already puts us back in the workqueue
3337 * in that case. If we proceed deleting, we'll get a dangling
3338 * reference, and removing the object from the workqueue in that case
3339 * is unnecessary complication. We are not a fast path.
3341 * Note that this case is fundamentally different from racing with
3342 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3343 * kmem_cache_shrink, not only we would be reinserting a dead cache
3344 * into the queue, but doing so from inside the worker racing to
3347 * So if we aren't down to zero, we'll just schedule a worker and try
3350 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3351 kmem_cache_shrink(cachep);
3352 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3355 kmem_cache_destroy(cachep);
3358 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3360 if (!cachep->memcg_params->dead)
3364 * There are many ways in which we can get here.
3366 * We can get to a memory-pressure situation while the delayed work is
3367 * still pending to run. The vmscan shrinkers can then release all
3368 * cache memory and get us to destruction. If this is the case, we'll
3369 * be executed twice, which is a bug (the second time will execute over
3370 * bogus data). In this case, cancelling the work should be fine.
3372 * But we can also get here from the worker itself, if
3373 * kmem_cache_shrink is enough to shake all the remaining objects and
3374 * get the page count to 0. In this case, we'll deadlock if we try to
3375 * cancel the work (the worker runs with an internal lock held, which
3376 * is the same lock we would hold for cancel_work_sync().)
3378 * Since we can't possibly know who got us here, just refrain from
3379 * running if there is already work pending
3381 if (work_pending(&cachep->memcg_params->destroy))
3384 * We have to defer the actual destroying to a workqueue, because
3385 * we might currently be in a context that cannot sleep.
3387 schedule_work(&cachep->memcg_params->destroy);
3391 * This lock protects updaters, not readers. We want readers to be as fast as
3392 * they can, and they will either see NULL or a valid cache value. Our model
3393 * allow them to see NULL, in which case the root memcg will be selected.
3395 * We need this lock because multiple allocations to the same cache from a non
3396 * will span more than one worker. Only one of them can create the cache.
3398 static DEFINE_MUTEX(memcg_cache_mutex);
3401 * Called with memcg_cache_mutex held
3403 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3404 struct kmem_cache *s)
3406 struct kmem_cache *new;
3407 static char *tmp_name = NULL;
3409 lockdep_assert_held(&memcg_cache_mutex);
3412 * kmem_cache_create_memcg duplicates the given name and
3413 * cgroup_name for this name requires RCU context.
3414 * This static temporary buffer is used to prevent from
3415 * pointless shortliving allocation.
3418 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3424 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3425 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3428 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3429 (s->flags & ~SLAB_PANIC), s->ctor, s);
3432 new->allocflags |= __GFP_KMEMCG;
3437 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3438 struct kmem_cache *cachep)
3440 struct kmem_cache *new_cachep;
3443 BUG_ON(!memcg_can_account_kmem(memcg));
3445 idx = memcg_cache_id(memcg);
3447 mutex_lock(&memcg_cache_mutex);
3448 new_cachep = cachep->memcg_params->memcg_caches[idx];
3450 css_put(&memcg->css);
3454 new_cachep = kmem_cache_dup(memcg, cachep);
3455 if (new_cachep == NULL) {
3456 new_cachep = cachep;
3457 css_put(&memcg->css);
3461 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3463 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3465 * the readers won't lock, make sure everybody sees the updated value,
3466 * so they won't put stuff in the queue again for no reason
3470 mutex_unlock(&memcg_cache_mutex);
3474 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3476 struct kmem_cache *c;
3479 if (!s->memcg_params)
3481 if (!s->memcg_params->is_root_cache)
3485 * If the cache is being destroyed, we trust that there is no one else
3486 * requesting objects from it. Even if there are, the sanity checks in
3487 * kmem_cache_destroy should caught this ill-case.
3489 * Still, we don't want anyone else freeing memcg_caches under our
3490 * noses, which can happen if a new memcg comes to life. As usual,
3491 * we'll take the set_limit_mutex to protect ourselves against this.
3493 mutex_lock(&set_limit_mutex);
3494 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3495 c = s->memcg_params->memcg_caches[i];
3500 * We will now manually delete the caches, so to avoid races
3501 * we need to cancel all pending destruction workers and
3502 * proceed with destruction ourselves.
3504 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3505 * and that could spawn the workers again: it is likely that
3506 * the cache still have active pages until this very moment.
3507 * This would lead us back to mem_cgroup_destroy_cache.
3509 * But that will not execute at all if the "dead" flag is not
3510 * set, so flip it down to guarantee we are in control.
3512 c->memcg_params->dead = false;
3513 cancel_work_sync(&c->memcg_params->destroy);
3514 kmem_cache_destroy(c);
3516 mutex_unlock(&set_limit_mutex);
3519 struct create_work {
3520 struct mem_cgroup *memcg;
3521 struct kmem_cache *cachep;
3522 struct work_struct work;
3525 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3527 struct kmem_cache *cachep;
3528 struct memcg_cache_params *params;
3530 if (!memcg_kmem_is_active(memcg))
3533 mutex_lock(&memcg->slab_caches_mutex);
3534 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3535 cachep = memcg_params_to_cache(params);
3536 cachep->memcg_params->dead = true;
3537 schedule_work(&cachep->memcg_params->destroy);
3539 mutex_unlock(&memcg->slab_caches_mutex);
3542 static void memcg_create_cache_work_func(struct work_struct *w)
3544 struct create_work *cw;
3546 cw = container_of(w, struct create_work, work);
3547 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3552 * Enqueue the creation of a per-memcg kmem_cache.
3554 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3555 struct kmem_cache *cachep)
3557 struct create_work *cw;
3559 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3561 css_put(&memcg->css);
3566 cw->cachep = cachep;
3568 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3569 schedule_work(&cw->work);
3572 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3573 struct kmem_cache *cachep)
3576 * We need to stop accounting when we kmalloc, because if the
3577 * corresponding kmalloc cache is not yet created, the first allocation
3578 * in __memcg_create_cache_enqueue will recurse.
3580 * However, it is better to enclose the whole function. Depending on
3581 * the debugging options enabled, INIT_WORK(), for instance, can
3582 * trigger an allocation. This too, will make us recurse. Because at
3583 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3584 * the safest choice is to do it like this, wrapping the whole function.
3586 memcg_stop_kmem_account();
3587 __memcg_create_cache_enqueue(memcg, cachep);
3588 memcg_resume_kmem_account();
3591 * Return the kmem_cache we're supposed to use for a slab allocation.
3592 * We try to use the current memcg's version of the cache.
3594 * If the cache does not exist yet, if we are the first user of it,
3595 * we either create it immediately, if possible, or create it asynchronously
3597 * In the latter case, we will let the current allocation go through with
3598 * the original cache.
3600 * Can't be called in interrupt context or from kernel threads.
3601 * This function needs to be called with rcu_read_lock() held.
3603 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3606 struct mem_cgroup *memcg;
3609 VM_BUG_ON(!cachep->memcg_params);
3610 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3612 if (!current->mm || current->memcg_kmem_skip_account)
3616 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3618 if (!memcg_can_account_kmem(memcg))
3621 idx = memcg_cache_id(memcg);
3624 * barrier to mare sure we're always seeing the up to date value. The
3625 * code updating memcg_caches will issue a write barrier to match this.
3627 read_barrier_depends();
3628 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3629 cachep = cachep->memcg_params->memcg_caches[idx];
3633 /* The corresponding put will be done in the workqueue. */
3634 if (!css_tryget(&memcg->css))
3639 * If we are in a safe context (can wait, and not in interrupt
3640 * context), we could be be predictable and return right away.
3641 * This would guarantee that the allocation being performed
3642 * already belongs in the new cache.
3644 * However, there are some clashes that can arrive from locking.
3645 * For instance, because we acquire the slab_mutex while doing
3646 * kmem_cache_dup, this means no further allocation could happen
3647 * with the slab_mutex held.
3649 * Also, because cache creation issue get_online_cpus(), this
3650 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3651 * that ends up reversed during cpu hotplug. (cpuset allocates
3652 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3653 * better to defer everything.
3655 memcg_create_cache_enqueue(memcg, cachep);
3661 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3664 * We need to verify if the allocation against current->mm->owner's memcg is
3665 * possible for the given order. But the page is not allocated yet, so we'll
3666 * need a further commit step to do the final arrangements.
3668 * It is possible for the task to switch cgroups in this mean time, so at
3669 * commit time, we can't rely on task conversion any longer. We'll then use
3670 * the handle argument to return to the caller which cgroup we should commit
3671 * against. We could also return the memcg directly and avoid the pointer
3672 * passing, but a boolean return value gives better semantics considering
3673 * the compiled-out case as well.
3675 * Returning true means the allocation is possible.
3678 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3680 struct mem_cgroup *memcg;
3686 * Disabling accounting is only relevant for some specific memcg
3687 * internal allocations. Therefore we would initially not have such
3688 * check here, since direct calls to the page allocator that are marked
3689 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3690 * concerned with cache allocations, and by having this test at
3691 * memcg_kmem_get_cache, we are already able to relay the allocation to
3692 * the root cache and bypass the memcg cache altogether.
3694 * There is one exception, though: the SLUB allocator does not create
3695 * large order caches, but rather service large kmallocs directly from
3696 * the page allocator. Therefore, the following sequence when backed by
3697 * the SLUB allocator:
3699 * memcg_stop_kmem_account();
3700 * kmalloc(<large_number>)
3701 * memcg_resume_kmem_account();
3703 * would effectively ignore the fact that we should skip accounting,
3704 * since it will drive us directly to this function without passing
3705 * through the cache selector memcg_kmem_get_cache. Such large
3706 * allocations are extremely rare but can happen, for instance, for the
3707 * cache arrays. We bring this test here.
3709 if (!current->mm || current->memcg_kmem_skip_account)
3712 memcg = try_get_mem_cgroup_from_mm(current->mm);
3715 * very rare case described in mem_cgroup_from_task. Unfortunately there
3716 * isn't much we can do without complicating this too much, and it would
3717 * be gfp-dependent anyway. Just let it go
3719 if (unlikely(!memcg))
3722 if (!memcg_can_account_kmem(memcg)) {
3723 css_put(&memcg->css);
3727 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3731 css_put(&memcg->css);
3735 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3738 struct page_cgroup *pc;
3740 VM_BUG_ON(mem_cgroup_is_root(memcg));
3742 /* The page allocation failed. Revert */
3744 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3748 pc = lookup_page_cgroup(page);
3749 lock_page_cgroup(pc);
3750 pc->mem_cgroup = memcg;
3751 SetPageCgroupUsed(pc);
3752 unlock_page_cgroup(pc);
3755 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3757 struct mem_cgroup *memcg = NULL;
3758 struct page_cgroup *pc;
3761 pc = lookup_page_cgroup(page);
3763 * Fast unlocked return. Theoretically might have changed, have to
3764 * check again after locking.
3766 if (!PageCgroupUsed(pc))
3769 lock_page_cgroup(pc);
3770 if (PageCgroupUsed(pc)) {
3771 memcg = pc->mem_cgroup;
3772 ClearPageCgroupUsed(pc);
3774 unlock_page_cgroup(pc);
3777 * We trust that only if there is a memcg associated with the page, it
3778 * is a valid allocation
3783 VM_BUG_ON(mem_cgroup_is_root(memcg));
3784 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3787 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3790 #endif /* CONFIG_MEMCG_KMEM */
3792 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3794 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3796 * Because tail pages are not marked as "used", set it. We're under
3797 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3798 * charge/uncharge will be never happen and move_account() is done under
3799 * compound_lock(), so we don't have to take care of races.
3801 void mem_cgroup_split_huge_fixup(struct page *head)
3803 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3804 struct page_cgroup *pc;
3805 struct mem_cgroup *memcg;
3808 if (mem_cgroup_disabled())
3811 memcg = head_pc->mem_cgroup;
3812 for (i = 1; i < HPAGE_PMD_NR; i++) {
3814 pc->mem_cgroup = memcg;
3815 smp_wmb();/* see __commit_charge() */
3816 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3818 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3821 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3824 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3825 struct mem_cgroup *to,
3826 unsigned int nr_pages,
3827 enum mem_cgroup_stat_index idx)
3829 /* Update stat data for mem_cgroup */
3831 __this_cpu_sub(from->stat->count[idx], nr_pages);
3832 __this_cpu_add(to->stat->count[idx], nr_pages);
3837 * mem_cgroup_move_account - move account of the page
3839 * @nr_pages: number of regular pages (>1 for huge pages)
3840 * @pc: page_cgroup of the page.
3841 * @from: mem_cgroup which the page is moved from.
3842 * @to: mem_cgroup which the page is moved to. @from != @to.
3844 * The caller must confirm following.
3845 * - page is not on LRU (isolate_page() is useful.)
3846 * - compound_lock is held when nr_pages > 1
3848 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3851 static int mem_cgroup_move_account(struct page *page,
3852 unsigned int nr_pages,
3853 struct page_cgroup *pc,
3854 struct mem_cgroup *from,
3855 struct mem_cgroup *to)
3857 unsigned long flags;
3859 bool anon = PageAnon(page);
3861 VM_BUG_ON(from == to);
3862 VM_BUG_ON(PageLRU(page));
3864 * The page is isolated from LRU. So, collapse function
3865 * will not handle this page. But page splitting can happen.
3866 * Do this check under compound_page_lock(). The caller should
3870 if (nr_pages > 1 && !PageTransHuge(page))
3873 lock_page_cgroup(pc);
3876 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3879 move_lock_mem_cgroup(from, &flags);
3881 if (!anon && page_mapped(page))
3882 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3883 MEM_CGROUP_STAT_FILE_MAPPED);
3885 if (PageWriteback(page))
3886 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3887 MEM_CGROUP_STAT_WRITEBACK);
3889 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3891 /* caller should have done css_get */
3892 pc->mem_cgroup = to;
3893 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3894 move_unlock_mem_cgroup(from, &flags);
3897 unlock_page_cgroup(pc);
3901 memcg_check_events(to, page);
3902 memcg_check_events(from, page);
3908 * mem_cgroup_move_parent - moves page to the parent group
3909 * @page: the page to move
3910 * @pc: page_cgroup of the page
3911 * @child: page's cgroup
3913 * move charges to its parent or the root cgroup if the group has no
3914 * parent (aka use_hierarchy==0).
3915 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3916 * mem_cgroup_move_account fails) the failure is always temporary and
3917 * it signals a race with a page removal/uncharge or migration. In the
3918 * first case the page is on the way out and it will vanish from the LRU
3919 * on the next attempt and the call should be retried later.
3920 * Isolation from the LRU fails only if page has been isolated from
3921 * the LRU since we looked at it and that usually means either global
3922 * reclaim or migration going on. The page will either get back to the
3924 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3925 * (!PageCgroupUsed) or moved to a different group. The page will
3926 * disappear in the next attempt.
3928 static int mem_cgroup_move_parent(struct page *page,
3929 struct page_cgroup *pc,
3930 struct mem_cgroup *child)
3932 struct mem_cgroup *parent;
3933 unsigned int nr_pages;
3934 unsigned long uninitialized_var(flags);
3937 VM_BUG_ON(mem_cgroup_is_root(child));
3940 if (!get_page_unless_zero(page))
3942 if (isolate_lru_page(page))
3945 nr_pages = hpage_nr_pages(page);
3947 parent = parent_mem_cgroup(child);
3949 * If no parent, move charges to root cgroup.
3952 parent = root_mem_cgroup;
3955 VM_BUG_ON(!PageTransHuge(page));
3956 flags = compound_lock_irqsave(page);
3959 ret = mem_cgroup_move_account(page, nr_pages,
3962 __mem_cgroup_cancel_local_charge(child, nr_pages);
3965 compound_unlock_irqrestore(page, flags);
3966 putback_lru_page(page);
3974 * Charge the memory controller for page usage.
3976 * 0 if the charge was successful
3977 * < 0 if the cgroup is over its limit
3979 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3980 gfp_t gfp_mask, enum charge_type ctype)
3982 struct mem_cgroup *memcg = NULL;
3983 unsigned int nr_pages = 1;
3987 if (PageTransHuge(page)) {
3988 nr_pages <<= compound_order(page);
3989 VM_BUG_ON(!PageTransHuge(page));
3991 * Never OOM-kill a process for a huge page. The
3992 * fault handler will fall back to regular pages.
3997 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
4000 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
4004 int mem_cgroup_newpage_charge(struct page *page,
4005 struct mm_struct *mm, gfp_t gfp_mask)
4007 if (mem_cgroup_disabled())
4009 VM_BUG_ON(page_mapped(page));
4010 VM_BUG_ON(page->mapping && !PageAnon(page));
4012 return mem_cgroup_charge_common(page, mm, gfp_mask,
4013 MEM_CGROUP_CHARGE_TYPE_ANON);
4017 * While swap-in, try_charge -> commit or cancel, the page is locked.
4018 * And when try_charge() successfully returns, one refcnt to memcg without
4019 * struct page_cgroup is acquired. This refcnt will be consumed by
4020 * "commit()" or removed by "cancel()"
4022 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
4025 struct mem_cgroup **memcgp)
4027 struct mem_cgroup *memcg;
4028 struct page_cgroup *pc;
4031 pc = lookup_page_cgroup(page);
4033 * Every swap fault against a single page tries to charge the
4034 * page, bail as early as possible. shmem_unuse() encounters
4035 * already charged pages, too. The USED bit is protected by
4036 * the page lock, which serializes swap cache removal, which
4037 * in turn serializes uncharging.
4039 if (PageCgroupUsed(pc))
4041 if (!do_swap_account)
4043 memcg = try_get_mem_cgroup_from_page(page);
4047 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4048 css_put(&memcg->css);
4053 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4059 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4060 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4063 if (mem_cgroup_disabled())
4066 * A racing thread's fault, or swapoff, may have already
4067 * updated the pte, and even removed page from swap cache: in
4068 * those cases unuse_pte()'s pte_same() test will fail; but
4069 * there's also a KSM case which does need to charge the page.
4071 if (!PageSwapCache(page)) {
4074 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4079 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4082 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4084 if (mem_cgroup_disabled())
4088 __mem_cgroup_cancel_charge(memcg, 1);
4092 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4093 enum charge_type ctype)
4095 if (mem_cgroup_disabled())
4100 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4102 * Now swap is on-memory. This means this page may be
4103 * counted both as mem and swap....double count.
4104 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4105 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4106 * may call delete_from_swap_cache() before reach here.
4108 if (do_swap_account && PageSwapCache(page)) {
4109 swp_entry_t ent = {.val = page_private(page)};
4110 mem_cgroup_uncharge_swap(ent);
4114 void mem_cgroup_commit_charge_swapin(struct page *page,
4115 struct mem_cgroup *memcg)
4117 __mem_cgroup_commit_charge_swapin(page, memcg,
4118 MEM_CGROUP_CHARGE_TYPE_ANON);
4121 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4124 struct mem_cgroup *memcg = NULL;
4125 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4128 if (mem_cgroup_disabled())
4130 if (PageCompound(page))
4133 if (!PageSwapCache(page))
4134 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4135 else { /* page is swapcache/shmem */
4136 ret = __mem_cgroup_try_charge_swapin(mm, page,
4139 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4144 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4145 unsigned int nr_pages,
4146 const enum charge_type ctype)
4148 struct memcg_batch_info *batch = NULL;
4149 bool uncharge_memsw = true;
4151 /* If swapout, usage of swap doesn't decrease */
4152 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4153 uncharge_memsw = false;
4155 batch = ¤t->memcg_batch;
4157 * In usual, we do css_get() when we remember memcg pointer.
4158 * But in this case, we keep res->usage until end of a series of
4159 * uncharges. Then, it's ok to ignore memcg's refcnt.
4162 batch->memcg = memcg;
4164 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4165 * In those cases, all pages freed continuously can be expected to be in
4166 * the same cgroup and we have chance to coalesce uncharges.
4167 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4168 * because we want to do uncharge as soon as possible.
4171 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4172 goto direct_uncharge;
4175 goto direct_uncharge;
4178 * In typical case, batch->memcg == mem. This means we can
4179 * merge a series of uncharges to an uncharge of res_counter.
4180 * If not, we uncharge res_counter ony by one.
4182 if (batch->memcg != memcg)
4183 goto direct_uncharge;
4184 /* remember freed charge and uncharge it later */
4187 batch->memsw_nr_pages++;
4190 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4192 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4193 if (unlikely(batch->memcg != memcg))
4194 memcg_oom_recover(memcg);
4198 * uncharge if !page_mapped(page)
4200 static struct mem_cgroup *
4201 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4204 struct mem_cgroup *memcg = NULL;
4205 unsigned int nr_pages = 1;
4206 struct page_cgroup *pc;
4209 if (mem_cgroup_disabled())
4212 if (PageTransHuge(page)) {
4213 nr_pages <<= compound_order(page);
4214 VM_BUG_ON(!PageTransHuge(page));
4217 * Check if our page_cgroup is valid
4219 pc = lookup_page_cgroup(page);
4220 if (unlikely(!PageCgroupUsed(pc)))
4223 lock_page_cgroup(pc);
4225 memcg = pc->mem_cgroup;
4227 if (!PageCgroupUsed(pc))
4230 anon = PageAnon(page);
4233 case MEM_CGROUP_CHARGE_TYPE_ANON:
4235 * Generally PageAnon tells if it's the anon statistics to be
4236 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4237 * used before page reached the stage of being marked PageAnon.
4241 case MEM_CGROUP_CHARGE_TYPE_DROP:
4242 /* See mem_cgroup_prepare_migration() */
4243 if (page_mapped(page))
4246 * Pages under migration may not be uncharged. But
4247 * end_migration() /must/ be the one uncharging the
4248 * unused post-migration page and so it has to call
4249 * here with the migration bit still set. See the
4250 * res_counter handling below.
4252 if (!end_migration && PageCgroupMigration(pc))
4255 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4256 if (!PageAnon(page)) { /* Shared memory */
4257 if (page->mapping && !page_is_file_cache(page))
4259 } else if (page_mapped(page)) /* Anon */
4266 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4268 ClearPageCgroupUsed(pc);
4270 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4271 * freed from LRU. This is safe because uncharged page is expected not
4272 * to be reused (freed soon). Exception is SwapCache, it's handled by
4273 * special functions.
4276 unlock_page_cgroup(pc);
4278 * even after unlock, we have memcg->res.usage here and this memcg
4279 * will never be freed, so it's safe to call css_get().
4281 memcg_check_events(memcg, page);
4282 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4283 mem_cgroup_swap_statistics(memcg, true);
4284 css_get(&memcg->css);
4287 * Migration does not charge the res_counter for the
4288 * replacement page, so leave it alone when phasing out the
4289 * page that is unused after the migration.
4291 if (!end_migration && !mem_cgroup_is_root(memcg))
4292 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4297 unlock_page_cgroup(pc);
4301 void mem_cgroup_uncharge_page(struct page *page)
4304 if (page_mapped(page))
4306 VM_BUG_ON(page->mapping && !PageAnon(page));
4308 * If the page is in swap cache, uncharge should be deferred
4309 * to the swap path, which also properly accounts swap usage
4310 * and handles memcg lifetime.
4312 * Note that this check is not stable and reclaim may add the
4313 * page to swap cache at any time after this. However, if the
4314 * page is not in swap cache by the time page->mapcount hits
4315 * 0, there won't be any page table references to the swap
4316 * slot, and reclaim will free it and not actually write the
4319 if (PageSwapCache(page))
4321 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4324 void mem_cgroup_uncharge_cache_page(struct page *page)
4326 VM_BUG_ON(page_mapped(page));
4327 VM_BUG_ON(page->mapping);
4328 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4332 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4333 * In that cases, pages are freed continuously and we can expect pages
4334 * are in the same memcg. All these calls itself limits the number of
4335 * pages freed at once, then uncharge_start/end() is called properly.
4336 * This may be called prural(2) times in a context,
4339 void mem_cgroup_uncharge_start(void)
4341 current->memcg_batch.do_batch++;
4342 /* We can do nest. */
4343 if (current->memcg_batch.do_batch == 1) {
4344 current->memcg_batch.memcg = NULL;
4345 current->memcg_batch.nr_pages = 0;
4346 current->memcg_batch.memsw_nr_pages = 0;
4350 void mem_cgroup_uncharge_end(void)
4352 struct memcg_batch_info *batch = ¤t->memcg_batch;
4354 if (!batch->do_batch)
4358 if (batch->do_batch) /* If stacked, do nothing. */
4364 * This "batch->memcg" is valid without any css_get/put etc...
4365 * bacause we hide charges behind us.
4367 if (batch->nr_pages)
4368 res_counter_uncharge(&batch->memcg->res,
4369 batch->nr_pages * PAGE_SIZE);
4370 if (batch->memsw_nr_pages)
4371 res_counter_uncharge(&batch->memcg->memsw,
4372 batch->memsw_nr_pages * PAGE_SIZE);
4373 memcg_oom_recover(batch->memcg);
4374 /* forget this pointer (for sanity check) */
4375 batch->memcg = NULL;
4380 * called after __delete_from_swap_cache() and drop "page" account.
4381 * memcg information is recorded to swap_cgroup of "ent"
4384 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4386 struct mem_cgroup *memcg;
4387 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4389 if (!swapout) /* this was a swap cache but the swap is unused ! */
4390 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4392 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4395 * record memcg information, if swapout && memcg != NULL,
4396 * css_get() was called in uncharge().
4398 if (do_swap_account && swapout && memcg)
4399 swap_cgroup_record(ent, css_id(&memcg->css));
4403 #ifdef CONFIG_MEMCG_SWAP
4405 * called from swap_entry_free(). remove record in swap_cgroup and
4406 * uncharge "memsw" account.
4408 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4410 struct mem_cgroup *memcg;
4413 if (!do_swap_account)
4416 id = swap_cgroup_record(ent, 0);
4418 memcg = mem_cgroup_lookup(id);
4421 * We uncharge this because swap is freed.
4422 * This memcg can be obsolete one. We avoid calling css_tryget
4424 if (!mem_cgroup_is_root(memcg))
4425 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4426 mem_cgroup_swap_statistics(memcg, false);
4427 css_put(&memcg->css);
4433 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4434 * @entry: swap entry to be moved
4435 * @from: mem_cgroup which the entry is moved from
4436 * @to: mem_cgroup which the entry is moved to
4438 * It succeeds only when the swap_cgroup's record for this entry is the same
4439 * as the mem_cgroup's id of @from.
4441 * Returns 0 on success, -EINVAL on failure.
4443 * The caller must have charged to @to, IOW, called res_counter_charge() about
4444 * both res and memsw, and called css_get().
4446 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4447 struct mem_cgroup *from, struct mem_cgroup *to)
4449 unsigned short old_id, new_id;
4451 old_id = css_id(&from->css);
4452 new_id = css_id(&to->css);
4454 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4455 mem_cgroup_swap_statistics(from, false);
4456 mem_cgroup_swap_statistics(to, true);
4458 * This function is only called from task migration context now.
4459 * It postpones res_counter and refcount handling till the end
4460 * of task migration(mem_cgroup_clear_mc()) for performance
4461 * improvement. But we cannot postpone css_get(to) because if
4462 * the process that has been moved to @to does swap-in, the
4463 * refcount of @to might be decreased to 0.
4465 * We are in attach() phase, so the cgroup is guaranteed to be
4466 * alive, so we can just call css_get().
4474 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4475 struct mem_cgroup *from, struct mem_cgroup *to)
4482 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4485 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4486 struct mem_cgroup **memcgp)
4488 struct mem_cgroup *memcg = NULL;
4489 unsigned int nr_pages = 1;
4490 struct page_cgroup *pc;
4491 enum charge_type ctype;
4495 if (mem_cgroup_disabled())
4498 if (PageTransHuge(page))
4499 nr_pages <<= compound_order(page);
4501 pc = lookup_page_cgroup(page);
4502 lock_page_cgroup(pc);
4503 if (PageCgroupUsed(pc)) {
4504 memcg = pc->mem_cgroup;
4505 css_get(&memcg->css);
4507 * At migrating an anonymous page, its mapcount goes down
4508 * to 0 and uncharge() will be called. But, even if it's fully
4509 * unmapped, migration may fail and this page has to be
4510 * charged again. We set MIGRATION flag here and delay uncharge
4511 * until end_migration() is called
4513 * Corner Case Thinking
4515 * When the old page was mapped as Anon and it's unmap-and-freed
4516 * while migration was ongoing.
4517 * If unmap finds the old page, uncharge() of it will be delayed
4518 * until end_migration(). If unmap finds a new page, it's
4519 * uncharged when it make mapcount to be 1->0. If unmap code
4520 * finds swap_migration_entry, the new page will not be mapped
4521 * and end_migration() will find it(mapcount==0).
4524 * When the old page was mapped but migraion fails, the kernel
4525 * remaps it. A charge for it is kept by MIGRATION flag even
4526 * if mapcount goes down to 0. We can do remap successfully
4527 * without charging it again.
4530 * The "old" page is under lock_page() until the end of
4531 * migration, so, the old page itself will not be swapped-out.
4532 * If the new page is swapped out before end_migraton, our
4533 * hook to usual swap-out path will catch the event.
4536 SetPageCgroupMigration(pc);
4538 unlock_page_cgroup(pc);
4540 * If the page is not charged at this point,
4548 * We charge new page before it's used/mapped. So, even if unlock_page()
4549 * is called before end_migration, we can catch all events on this new
4550 * page. In the case new page is migrated but not remapped, new page's
4551 * mapcount will be finally 0 and we call uncharge in end_migration().
4554 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4556 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4558 * The page is committed to the memcg, but it's not actually
4559 * charged to the res_counter since we plan on replacing the
4560 * old one and only one page is going to be left afterwards.
4562 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4565 /* remove redundant charge if migration failed*/
4566 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4567 struct page *oldpage, struct page *newpage, bool migration_ok)
4569 struct page *used, *unused;
4570 struct page_cgroup *pc;
4576 if (!migration_ok) {
4583 anon = PageAnon(used);
4584 __mem_cgroup_uncharge_common(unused,
4585 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4586 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4588 css_put(&memcg->css);
4590 * We disallowed uncharge of pages under migration because mapcount
4591 * of the page goes down to zero, temporarly.
4592 * Clear the flag and check the page should be charged.
4594 pc = lookup_page_cgroup(oldpage);
4595 lock_page_cgroup(pc);
4596 ClearPageCgroupMigration(pc);
4597 unlock_page_cgroup(pc);
4600 * If a page is a file cache, radix-tree replacement is very atomic
4601 * and we can skip this check. When it was an Anon page, its mapcount
4602 * goes down to 0. But because we added MIGRATION flage, it's not
4603 * uncharged yet. There are several case but page->mapcount check
4604 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4605 * check. (see prepare_charge() also)
4608 mem_cgroup_uncharge_page(used);
4612 * At replace page cache, newpage is not under any memcg but it's on
4613 * LRU. So, this function doesn't touch res_counter but handles LRU
4614 * in correct way. Both pages are locked so we cannot race with uncharge.
4616 void mem_cgroup_replace_page_cache(struct page *oldpage,
4617 struct page *newpage)
4619 struct mem_cgroup *memcg = NULL;
4620 struct page_cgroup *pc;
4621 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4623 if (mem_cgroup_disabled())
4626 pc = lookup_page_cgroup(oldpage);
4627 /* fix accounting on old pages */
4628 lock_page_cgroup(pc);
4629 if (PageCgroupUsed(pc)) {
4630 memcg = pc->mem_cgroup;
4631 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4632 ClearPageCgroupUsed(pc);
4634 unlock_page_cgroup(pc);
4637 * When called from shmem_replace_page(), in some cases the
4638 * oldpage has already been charged, and in some cases not.
4643 * Even if newpage->mapping was NULL before starting replacement,
4644 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4645 * LRU while we overwrite pc->mem_cgroup.
4647 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4650 #ifdef CONFIG_DEBUG_VM
4651 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4653 struct page_cgroup *pc;
4655 pc = lookup_page_cgroup(page);
4657 * Can be NULL while feeding pages into the page allocator for
4658 * the first time, i.e. during boot or memory hotplug;
4659 * or when mem_cgroup_disabled().
4661 if (likely(pc) && PageCgroupUsed(pc))
4666 bool mem_cgroup_bad_page_check(struct page *page)
4668 if (mem_cgroup_disabled())
4671 return lookup_page_cgroup_used(page) != NULL;
4674 void mem_cgroup_print_bad_page(struct page *page)
4676 struct page_cgroup *pc;
4678 pc = lookup_page_cgroup_used(page);
4680 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4681 pc, pc->flags, pc->mem_cgroup);
4686 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4687 unsigned long long val)
4690 u64 memswlimit, memlimit;
4692 int children = mem_cgroup_count_children(memcg);
4693 u64 curusage, oldusage;
4697 * For keeping hierarchical_reclaim simple, how long we should retry
4698 * is depends on callers. We set our retry-count to be function
4699 * of # of children which we should visit in this loop.
4701 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4703 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4706 while (retry_count) {
4707 if (signal_pending(current)) {
4712 * Rather than hide all in some function, I do this in
4713 * open coded manner. You see what this really does.
4714 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4716 mutex_lock(&set_limit_mutex);
4717 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4718 if (memswlimit < val) {
4720 mutex_unlock(&set_limit_mutex);
4724 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4728 ret = res_counter_set_limit(&memcg->res, val);
4730 if (memswlimit == val)
4731 memcg->memsw_is_minimum = true;
4733 memcg->memsw_is_minimum = false;
4735 mutex_unlock(&set_limit_mutex);
4740 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4741 MEM_CGROUP_RECLAIM_SHRINK);
4742 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4743 /* Usage is reduced ? */
4744 if (curusage >= oldusage)
4747 oldusage = curusage;
4749 if (!ret && enlarge)
4750 memcg_oom_recover(memcg);
4755 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4756 unsigned long long val)
4759 u64 memlimit, memswlimit, oldusage, curusage;
4760 int children = mem_cgroup_count_children(memcg);
4764 /* see mem_cgroup_resize_res_limit */
4765 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4766 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4767 while (retry_count) {
4768 if (signal_pending(current)) {
4773 * Rather than hide all in some function, I do this in
4774 * open coded manner. You see what this really does.
4775 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4777 mutex_lock(&set_limit_mutex);
4778 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4779 if (memlimit > val) {
4781 mutex_unlock(&set_limit_mutex);
4784 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4785 if (memswlimit < val)
4787 ret = res_counter_set_limit(&memcg->memsw, val);
4789 if (memlimit == val)
4790 memcg->memsw_is_minimum = true;
4792 memcg->memsw_is_minimum = false;
4794 mutex_unlock(&set_limit_mutex);
4799 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4800 MEM_CGROUP_RECLAIM_NOSWAP |
4801 MEM_CGROUP_RECLAIM_SHRINK);
4802 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4803 /* Usage is reduced ? */
4804 if (curusage >= oldusage)
4807 oldusage = curusage;
4809 if (!ret && enlarge)
4810 memcg_oom_recover(memcg);
4814 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4816 unsigned long *total_scanned)
4818 unsigned long nr_reclaimed = 0;
4819 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4820 unsigned long reclaimed;
4822 struct mem_cgroup_tree_per_zone *mctz;
4823 unsigned long long excess;
4824 unsigned long nr_scanned;
4829 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4831 * This loop can run a while, specially if mem_cgroup's continuously
4832 * keep exceeding their soft limit and putting the system under
4839 mz = mem_cgroup_largest_soft_limit_node(mctz);
4844 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4845 gfp_mask, &nr_scanned);
4846 nr_reclaimed += reclaimed;
4847 *total_scanned += nr_scanned;
4848 spin_lock(&mctz->lock);
4851 * If we failed to reclaim anything from this memory cgroup
4852 * it is time to move on to the next cgroup
4858 * Loop until we find yet another one.
4860 * By the time we get the soft_limit lock
4861 * again, someone might have aded the
4862 * group back on the RB tree. Iterate to
4863 * make sure we get a different mem.
4864 * mem_cgroup_largest_soft_limit_node returns
4865 * NULL if no other cgroup is present on
4869 __mem_cgroup_largest_soft_limit_node(mctz);
4871 css_put(&next_mz->memcg->css);
4872 else /* next_mz == NULL or other memcg */
4876 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4877 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4879 * One school of thought says that we should not add
4880 * back the node to the tree if reclaim returns 0.
4881 * But our reclaim could return 0, simply because due
4882 * to priority we are exposing a smaller subset of
4883 * memory to reclaim from. Consider this as a longer
4886 /* If excess == 0, no tree ops */
4887 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4888 spin_unlock(&mctz->lock);
4889 css_put(&mz->memcg->css);
4892 * Could not reclaim anything and there are no more
4893 * mem cgroups to try or we seem to be looping without
4894 * reclaiming anything.
4896 if (!nr_reclaimed &&
4898 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4900 } while (!nr_reclaimed);
4902 css_put(&next_mz->memcg->css);
4903 return nr_reclaimed;
4907 * mem_cgroup_force_empty_list - clears LRU of a group
4908 * @memcg: group to clear
4911 * @lru: lru to to clear
4913 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4914 * reclaim the pages page themselves - pages are moved to the parent (or root)
4917 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4918 int node, int zid, enum lru_list lru)
4920 struct lruvec *lruvec;
4921 unsigned long flags;
4922 struct list_head *list;
4926 zone = &NODE_DATA(node)->node_zones[zid];
4927 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4928 list = &lruvec->lists[lru];
4932 struct page_cgroup *pc;
4935 spin_lock_irqsave(&zone->lru_lock, flags);
4936 if (list_empty(list)) {
4937 spin_unlock_irqrestore(&zone->lru_lock, flags);
4940 page = list_entry(list->prev, struct page, lru);
4942 list_move(&page->lru, list);
4944 spin_unlock_irqrestore(&zone->lru_lock, flags);
4947 spin_unlock_irqrestore(&zone->lru_lock, flags);
4949 pc = lookup_page_cgroup(page);
4951 if (mem_cgroup_move_parent(page, pc, memcg)) {
4952 /* found lock contention or "pc" is obsolete. */
4957 } while (!list_empty(list));
4961 * make mem_cgroup's charge to be 0 if there is no task by moving
4962 * all the charges and pages to the parent.
4963 * This enables deleting this mem_cgroup.
4965 * Caller is responsible for holding css reference on the memcg.
4967 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4973 /* This is for making all *used* pages to be on LRU. */
4974 lru_add_drain_all();
4975 drain_all_stock_sync(memcg);
4976 mem_cgroup_start_move(memcg);
4977 for_each_node_state(node, N_MEMORY) {
4978 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4981 mem_cgroup_force_empty_list(memcg,
4986 mem_cgroup_end_move(memcg);
4987 memcg_oom_recover(memcg);
4991 * Kernel memory may not necessarily be trackable to a specific
4992 * process. So they are not migrated, and therefore we can't
4993 * expect their value to drop to 0 here.
4994 * Having res filled up with kmem only is enough.
4996 * This is a safety check because mem_cgroup_force_empty_list
4997 * could have raced with mem_cgroup_replace_page_cache callers
4998 * so the lru seemed empty but the page could have been added
4999 * right after the check. RES_USAGE should be safe as we always
5000 * charge before adding to the LRU.
5002 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
5003 res_counter_read_u64(&memcg->kmem, RES_USAGE);
5004 } while (usage > 0);
5007 static inline bool memcg_has_children(struct mem_cgroup *memcg)
5009 lockdep_assert_held(&memcg_create_mutex);
5011 * The lock does not prevent addition or deletion to the list
5012 * of children, but it prevents a new child from being
5013 * initialized based on this parent in css_online(), so it's
5014 * enough to decide whether hierarchically inherited
5015 * attributes can still be changed or not.
5017 return memcg->use_hierarchy &&
5018 !list_empty(&memcg->css.cgroup->children);
5022 * Reclaims as many pages from the given memcg as possible and moves
5023 * the rest to the parent.
5025 * Caller is responsible for holding css reference for memcg.
5027 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
5029 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
5030 struct cgroup *cgrp = memcg->css.cgroup;
5032 /* returns EBUSY if there is a task or if we come here twice. */
5033 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
5036 /* we call try-to-free pages for make this cgroup empty */
5037 lru_add_drain_all();
5038 /* try to free all pages in this cgroup */
5039 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
5042 if (signal_pending(current))
5045 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5049 /* maybe some writeback is necessary */
5050 congestion_wait(BLK_RW_ASYNC, HZ/10);
5055 mem_cgroup_reparent_charges(memcg);
5060 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
5063 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5065 if (mem_cgroup_is_root(memcg))
5067 return mem_cgroup_force_empty(memcg);
5070 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5073 return mem_cgroup_from_css(css)->use_hierarchy;
5076 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5077 struct cftype *cft, u64 val)
5080 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5081 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5083 mutex_lock(&memcg_create_mutex);
5085 if (memcg->use_hierarchy == val)
5089 * If parent's use_hierarchy is set, we can't make any modifications
5090 * in the child subtrees. If it is unset, then the change can
5091 * occur, provided the current cgroup has no children.
5093 * For the root cgroup, parent_mem is NULL, we allow value to be
5094 * set if there are no children.
5096 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5097 (val == 1 || val == 0)) {
5098 if (list_empty(&memcg->css.cgroup->children))
5099 memcg->use_hierarchy = val;
5106 mutex_unlock(&memcg_create_mutex);
5112 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5113 enum mem_cgroup_stat_index idx)
5115 struct mem_cgroup *iter;
5118 /* Per-cpu values can be negative, use a signed accumulator */
5119 for_each_mem_cgroup_tree(iter, memcg)
5120 val += mem_cgroup_read_stat(iter, idx);
5122 if (val < 0) /* race ? */
5127 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5131 if (!mem_cgroup_is_root(memcg)) {
5133 return res_counter_read_u64(&memcg->res, RES_USAGE);
5135 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5139 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5140 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5142 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5143 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5146 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5148 return val << PAGE_SHIFT;
5151 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
5152 struct cftype *cft, struct file *file,
5153 char __user *buf, size_t nbytes, loff_t *ppos)
5155 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5161 type = MEMFILE_TYPE(cft->private);
5162 name = MEMFILE_ATTR(cft->private);
5166 if (name == RES_USAGE)
5167 val = mem_cgroup_usage(memcg, false);
5169 val = res_counter_read_u64(&memcg->res, name);
5172 if (name == RES_USAGE)
5173 val = mem_cgroup_usage(memcg, true);
5175 val = res_counter_read_u64(&memcg->memsw, name);
5178 val = res_counter_read_u64(&memcg->kmem, name);
5184 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5185 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5188 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
5191 #ifdef CONFIG_MEMCG_KMEM
5192 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5194 * For simplicity, we won't allow this to be disabled. It also can't
5195 * be changed if the cgroup has children already, or if tasks had
5198 * If tasks join before we set the limit, a person looking at
5199 * kmem.usage_in_bytes will have no way to determine when it took
5200 * place, which makes the value quite meaningless.
5202 * After it first became limited, changes in the value of the limit are
5203 * of course permitted.
5205 mutex_lock(&memcg_create_mutex);
5206 mutex_lock(&set_limit_mutex);
5207 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
5208 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
5212 ret = res_counter_set_limit(&memcg->kmem, val);
5215 ret = memcg_update_cache_sizes(memcg);
5217 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
5220 static_key_slow_inc(&memcg_kmem_enabled_key);
5222 * setting the active bit after the inc will guarantee no one
5223 * starts accounting before all call sites are patched
5225 memcg_kmem_set_active(memcg);
5227 ret = res_counter_set_limit(&memcg->kmem, val);
5229 mutex_unlock(&set_limit_mutex);
5230 mutex_unlock(&memcg_create_mutex);
5235 #ifdef CONFIG_MEMCG_KMEM
5236 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5239 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5243 memcg->kmem_account_flags = parent->kmem_account_flags;
5245 * When that happen, we need to disable the static branch only on those
5246 * memcgs that enabled it. To achieve this, we would be forced to
5247 * complicate the code by keeping track of which memcgs were the ones
5248 * that actually enabled limits, and which ones got it from its
5251 * It is a lot simpler just to do static_key_slow_inc() on every child
5252 * that is accounted.
5254 if (!memcg_kmem_is_active(memcg))
5258 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5259 * memcg is active already. If the later initialization fails then the
5260 * cgroup core triggers the cleanup so we do not have to do it here.
5262 static_key_slow_inc(&memcg_kmem_enabled_key);
5264 mutex_lock(&set_limit_mutex);
5265 memcg_stop_kmem_account();
5266 ret = memcg_update_cache_sizes(memcg);
5267 memcg_resume_kmem_account();
5268 mutex_unlock(&set_limit_mutex);
5272 #endif /* CONFIG_MEMCG_KMEM */
5275 * The user of this function is...
5278 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5281 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5284 unsigned long long val;
5287 type = MEMFILE_TYPE(cft->private);
5288 name = MEMFILE_ATTR(cft->private);
5292 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5296 /* This function does all necessary parse...reuse it */
5297 ret = res_counter_memparse_write_strategy(buffer, &val);
5301 ret = mem_cgroup_resize_limit(memcg, val);
5302 else if (type == _MEMSWAP)
5303 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5304 else if (type == _KMEM)
5305 ret = memcg_update_kmem_limit(css, val);
5309 case RES_SOFT_LIMIT:
5310 ret = res_counter_memparse_write_strategy(buffer, &val);
5314 * For memsw, soft limits are hard to implement in terms
5315 * of semantics, for now, we support soft limits for
5316 * control without swap
5319 ret = res_counter_set_soft_limit(&memcg->res, val);
5324 ret = -EINVAL; /* should be BUG() ? */
5330 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5331 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5333 unsigned long long min_limit, min_memsw_limit, tmp;
5335 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5336 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5337 if (!memcg->use_hierarchy)
5340 while (css_parent(&memcg->css)) {
5341 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5342 if (!memcg->use_hierarchy)
5344 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5345 min_limit = min(min_limit, tmp);
5346 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5347 min_memsw_limit = min(min_memsw_limit, tmp);
5350 *mem_limit = min_limit;
5351 *memsw_limit = min_memsw_limit;
5354 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5356 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5360 type = MEMFILE_TYPE(event);
5361 name = MEMFILE_ATTR(event);
5366 res_counter_reset_max(&memcg->res);
5367 else if (type == _MEMSWAP)
5368 res_counter_reset_max(&memcg->memsw);
5369 else if (type == _KMEM)
5370 res_counter_reset_max(&memcg->kmem);
5376 res_counter_reset_failcnt(&memcg->res);
5377 else if (type == _MEMSWAP)
5378 res_counter_reset_failcnt(&memcg->memsw);
5379 else if (type == _KMEM)
5380 res_counter_reset_failcnt(&memcg->kmem);
5389 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5392 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5396 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5397 struct cftype *cft, u64 val)
5399 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5401 if (val >= (1 << NR_MOVE_TYPE))
5405 * No kind of locking is needed in here, because ->can_attach() will
5406 * check this value once in the beginning of the process, and then carry
5407 * on with stale data. This means that changes to this value will only
5408 * affect task migrations starting after the change.
5410 memcg->move_charge_at_immigrate = val;
5414 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5415 struct cftype *cft, u64 val)
5422 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5423 struct cftype *cft, struct seq_file *m)
5426 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5427 unsigned long node_nr;
5428 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5430 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5431 seq_printf(m, "total=%lu", total_nr);
5432 for_each_node_state(nid, N_MEMORY) {
5433 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5434 seq_printf(m, " N%d=%lu", nid, node_nr);
5438 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5439 seq_printf(m, "file=%lu", file_nr);
5440 for_each_node_state(nid, N_MEMORY) {
5441 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5443 seq_printf(m, " N%d=%lu", nid, node_nr);
5447 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5448 seq_printf(m, "anon=%lu", anon_nr);
5449 for_each_node_state(nid, N_MEMORY) {
5450 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5452 seq_printf(m, " N%d=%lu", nid, node_nr);
5456 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5457 seq_printf(m, "unevictable=%lu", unevictable_nr);
5458 for_each_node_state(nid, N_MEMORY) {
5459 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5460 BIT(LRU_UNEVICTABLE));
5461 seq_printf(m, " N%d=%lu", nid, node_nr);
5466 #endif /* CONFIG_NUMA */
5468 static inline void mem_cgroup_lru_names_not_uptodate(void)
5470 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5473 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5476 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5477 struct mem_cgroup *mi;
5480 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5481 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5483 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5484 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5487 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5488 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5489 mem_cgroup_read_events(memcg, i));
5491 for (i = 0; i < NR_LRU_LISTS; i++)
5492 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5493 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5495 /* Hierarchical information */
5497 unsigned long long limit, memsw_limit;
5498 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5499 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5500 if (do_swap_account)
5501 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5505 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5508 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5510 for_each_mem_cgroup_tree(mi, memcg)
5511 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5512 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5515 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5516 unsigned long long val = 0;
5518 for_each_mem_cgroup_tree(mi, memcg)
5519 val += mem_cgroup_read_events(mi, i);
5520 seq_printf(m, "total_%s %llu\n",
5521 mem_cgroup_events_names[i], val);
5524 for (i = 0; i < NR_LRU_LISTS; i++) {
5525 unsigned long long val = 0;
5527 for_each_mem_cgroup_tree(mi, memcg)
5528 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5529 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5532 #ifdef CONFIG_DEBUG_VM
5535 struct mem_cgroup_per_zone *mz;
5536 struct zone_reclaim_stat *rstat;
5537 unsigned long recent_rotated[2] = {0, 0};
5538 unsigned long recent_scanned[2] = {0, 0};
5540 for_each_online_node(nid)
5541 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5542 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5543 rstat = &mz->lruvec.reclaim_stat;
5545 recent_rotated[0] += rstat->recent_rotated[0];
5546 recent_rotated[1] += rstat->recent_rotated[1];
5547 recent_scanned[0] += rstat->recent_scanned[0];
5548 recent_scanned[1] += rstat->recent_scanned[1];
5550 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5551 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5552 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5553 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5560 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5563 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5565 return mem_cgroup_swappiness(memcg);
5568 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5569 struct cftype *cft, u64 val)
5571 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5572 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5574 if (val > 100 || !parent)
5577 mutex_lock(&memcg_create_mutex);
5579 /* If under hierarchy, only empty-root can set this value */
5580 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5581 mutex_unlock(&memcg_create_mutex);
5585 memcg->swappiness = val;
5587 mutex_unlock(&memcg_create_mutex);
5592 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5594 struct mem_cgroup_threshold_ary *t;
5600 t = rcu_dereference(memcg->thresholds.primary);
5602 t = rcu_dereference(memcg->memsw_thresholds.primary);
5607 usage = mem_cgroup_usage(memcg, swap);
5610 * current_threshold points to threshold just below or equal to usage.
5611 * If it's not true, a threshold was crossed after last
5612 * call of __mem_cgroup_threshold().
5614 i = t->current_threshold;
5617 * Iterate backward over array of thresholds starting from
5618 * current_threshold and check if a threshold is crossed.
5619 * If none of thresholds below usage is crossed, we read
5620 * only one element of the array here.
5622 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5623 eventfd_signal(t->entries[i].eventfd, 1);
5625 /* i = current_threshold + 1 */
5629 * Iterate forward over array of thresholds starting from
5630 * current_threshold+1 and check if a threshold is crossed.
5631 * If none of thresholds above usage is crossed, we read
5632 * only one element of the array here.
5634 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5635 eventfd_signal(t->entries[i].eventfd, 1);
5637 /* Update current_threshold */
5638 t->current_threshold = i - 1;
5643 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5646 __mem_cgroup_threshold(memcg, false);
5647 if (do_swap_account)
5648 __mem_cgroup_threshold(memcg, true);
5650 memcg = parent_mem_cgroup(memcg);
5654 static int compare_thresholds(const void *a, const void *b)
5656 const struct mem_cgroup_threshold *_a = a;
5657 const struct mem_cgroup_threshold *_b = b;
5659 if (_a->threshold > _b->threshold)
5662 if (_a->threshold < _b->threshold)
5668 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5670 struct mem_cgroup_eventfd_list *ev;
5672 list_for_each_entry(ev, &memcg->oom_notify, list)
5673 eventfd_signal(ev->eventfd, 1);
5677 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5679 struct mem_cgroup *iter;
5681 for_each_mem_cgroup_tree(iter, memcg)
5682 mem_cgroup_oom_notify_cb(iter);
5685 static int __mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5686 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5688 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5689 struct mem_cgroup_thresholds *thresholds;
5690 struct mem_cgroup_threshold_ary *new;
5691 u64 threshold, usage;
5694 ret = res_counter_memparse_write_strategy(args, &threshold);
5698 mutex_lock(&memcg->thresholds_lock);
5701 thresholds = &memcg->thresholds;
5702 else if (type == _MEMSWAP)
5703 thresholds = &memcg->memsw_thresholds;
5707 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5709 /* Check if a threshold crossed before adding a new one */
5710 if (thresholds->primary)
5711 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5713 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5715 /* Allocate memory for new array of thresholds */
5716 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5724 /* Copy thresholds (if any) to new array */
5725 if (thresholds->primary) {
5726 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5727 sizeof(struct mem_cgroup_threshold));
5730 /* Add new threshold */
5731 new->entries[size - 1].eventfd = eventfd;
5732 new->entries[size - 1].threshold = threshold;
5734 /* Sort thresholds. Registering of new threshold isn't time-critical */
5735 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5736 compare_thresholds, NULL);
5738 /* Find current threshold */
5739 new->current_threshold = -1;
5740 for (i = 0; i < size; i++) {
5741 if (new->entries[i].threshold <= usage) {
5743 * new->current_threshold will not be used until
5744 * rcu_assign_pointer(), so it's safe to increment
5747 ++new->current_threshold;
5752 /* Free old spare buffer and save old primary buffer as spare */
5753 kfree(thresholds->spare);
5754 thresholds->spare = thresholds->primary;
5756 rcu_assign_pointer(thresholds->primary, new);
5758 /* To be sure that nobody uses thresholds */
5762 mutex_unlock(&memcg->thresholds_lock);
5767 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5768 struct eventfd_ctx *eventfd, const char *args)
5770 return __mem_cgroup_usage_register_event(css, eventfd, args, _MEM);
5773 static int memsw_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5774 struct eventfd_ctx *eventfd, const char *args)
5776 return __mem_cgroup_usage_register_event(css, eventfd, args, _MEMSWAP);
5779 static void __mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5780 struct eventfd_ctx *eventfd, enum res_type type)
5782 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5783 struct mem_cgroup_thresholds *thresholds;
5784 struct mem_cgroup_threshold_ary *new;
5788 mutex_lock(&memcg->thresholds_lock);
5790 thresholds = &memcg->thresholds;
5791 else if (type == _MEMSWAP)
5792 thresholds = &memcg->memsw_thresholds;
5796 if (!thresholds->primary)
5799 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5801 /* Check if a threshold crossed before removing */
5802 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5804 /* Calculate new number of threshold */
5806 for (i = 0; i < thresholds->primary->size; i++) {
5807 if (thresholds->primary->entries[i].eventfd != eventfd)
5811 new = thresholds->spare;
5813 /* Set thresholds array to NULL if we don't have thresholds */
5822 /* Copy thresholds and find current threshold */
5823 new->current_threshold = -1;
5824 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5825 if (thresholds->primary->entries[i].eventfd == eventfd)
5828 new->entries[j] = thresholds->primary->entries[i];
5829 if (new->entries[j].threshold <= usage) {
5831 * new->current_threshold will not be used
5832 * until rcu_assign_pointer(), so it's safe to increment
5835 ++new->current_threshold;
5841 /* Swap primary and spare array */
5842 thresholds->spare = thresholds->primary;
5843 /* If all events are unregistered, free the spare array */
5845 kfree(thresholds->spare);
5846 thresholds->spare = NULL;
5849 rcu_assign_pointer(thresholds->primary, new);
5851 /* To be sure that nobody uses thresholds */
5854 mutex_unlock(&memcg->thresholds_lock);
5857 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5858 struct eventfd_ctx *eventfd)
5860 return __mem_cgroup_usage_unregister_event(css, eventfd, _MEM);
5863 static void memsw_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5864 struct eventfd_ctx *eventfd)
5866 return __mem_cgroup_usage_unregister_event(css, eventfd, _MEMSWAP);
5869 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5870 struct eventfd_ctx *eventfd, const char *args)
5872 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5873 struct mem_cgroup_eventfd_list *event;
5875 event = kmalloc(sizeof(*event), GFP_KERNEL);
5879 spin_lock(&memcg_oom_lock);
5881 event->eventfd = eventfd;
5882 list_add(&event->list, &memcg->oom_notify);
5884 /* already in OOM ? */
5885 if (atomic_read(&memcg->under_oom))
5886 eventfd_signal(eventfd, 1);
5887 spin_unlock(&memcg_oom_lock);
5892 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5893 struct eventfd_ctx *eventfd)
5895 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5896 struct mem_cgroup_eventfd_list *ev, *tmp;
5898 spin_lock(&memcg_oom_lock);
5900 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5901 if (ev->eventfd == eventfd) {
5902 list_del(&ev->list);
5907 spin_unlock(&memcg_oom_lock);
5910 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5911 struct cftype *cft, struct cgroup_map_cb *cb)
5913 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5915 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5917 if (atomic_read(&memcg->under_oom))
5918 cb->fill(cb, "under_oom", 1);
5920 cb->fill(cb, "under_oom", 0);
5924 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5925 struct cftype *cft, u64 val)
5927 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5928 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5930 /* cannot set to root cgroup and only 0 and 1 are allowed */
5931 if (!parent || !((val == 0) || (val == 1)))
5934 mutex_lock(&memcg_create_mutex);
5935 /* oom-kill-disable is a flag for subhierarchy. */
5936 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5937 mutex_unlock(&memcg_create_mutex);
5940 memcg->oom_kill_disable = val;
5942 memcg_oom_recover(memcg);
5943 mutex_unlock(&memcg_create_mutex);
5947 #ifdef CONFIG_MEMCG_KMEM
5948 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5952 memcg->kmemcg_id = -1;
5953 ret = memcg_propagate_kmem(memcg);
5957 return mem_cgroup_sockets_init(memcg, ss);
5960 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5962 mem_cgroup_sockets_destroy(memcg);
5965 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5967 if (!memcg_kmem_is_active(memcg))
5971 * kmem charges can outlive the cgroup. In the case of slab
5972 * pages, for instance, a page contain objects from various
5973 * processes. As we prevent from taking a reference for every
5974 * such allocation we have to be careful when doing uncharge
5975 * (see memcg_uncharge_kmem) and here during offlining.
5977 * The idea is that that only the _last_ uncharge which sees
5978 * the dead memcg will drop the last reference. An additional
5979 * reference is taken here before the group is marked dead
5980 * which is then paired with css_put during uncharge resp. here.
5982 * Although this might sound strange as this path is called from
5983 * css_offline() when the referencemight have dropped down to 0
5984 * and shouldn't be incremented anymore (css_tryget would fail)
5985 * we do not have other options because of the kmem allocations
5988 css_get(&memcg->css);
5990 memcg_kmem_mark_dead(memcg);
5992 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5995 if (memcg_kmem_test_and_clear_dead(memcg))
5996 css_put(&memcg->css);
5999 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
6004 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
6008 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
6014 * Unregister event and free resources.
6016 * Gets called from workqueue.
6018 static void cgroup_event_remove(struct work_struct *work)
6020 struct cgroup_event *event = container_of(work, struct cgroup_event,
6022 struct cgroup_subsys_state *css = event->css;
6024 remove_wait_queue(event->wqh, &event->wait);
6026 event->unregister_event(css, event->eventfd);
6028 /* Notify userspace the event is going away. */
6029 eventfd_signal(event->eventfd, 1);
6031 eventfd_ctx_put(event->eventfd);
6037 * Gets called on POLLHUP on eventfd when user closes it.
6039 * Called with wqh->lock held and interrupts disabled.
6041 static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
6042 int sync, void *key)
6044 struct cgroup_event *event = container_of(wait,
6045 struct cgroup_event, wait);
6046 struct mem_cgroup *memcg = mem_cgroup_from_css(event->css);
6047 unsigned long flags = (unsigned long)key;
6049 if (flags & POLLHUP) {
6051 * If the event has been detached at cgroup removal, we
6052 * can simply return knowing the other side will cleanup
6055 * We can't race against event freeing since the other
6056 * side will require wqh->lock via remove_wait_queue(),
6059 spin_lock(&memcg->event_list_lock);
6060 if (!list_empty(&event->list)) {
6061 list_del_init(&event->list);
6063 * We are in atomic context, but cgroup_event_remove()
6064 * may sleep, so we have to call it in workqueue.
6066 schedule_work(&event->remove);
6068 spin_unlock(&memcg->event_list_lock);
6074 static void cgroup_event_ptable_queue_proc(struct file *file,
6075 wait_queue_head_t *wqh, poll_table *pt)
6077 struct cgroup_event *event = container_of(pt,
6078 struct cgroup_event, pt);
6081 add_wait_queue(wqh, &event->wait);
6085 * Parse input and register new cgroup event handler.
6087 * Input must be in format '<event_fd> <control_fd> <args>'.
6088 * Interpretation of args is defined by control file implementation.
6090 static int cgroup_write_event_control(struct cgroup_subsys_state *css,
6091 struct cftype *cft, const char *buffer)
6093 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6094 struct cgroup_event *event;
6095 struct cgroup_subsys_state *cfile_css;
6096 unsigned int efd, cfd;
6103 efd = simple_strtoul(buffer, &endp, 10);
6108 cfd = simple_strtoul(buffer, &endp, 10);
6109 if ((*endp != ' ') && (*endp != '\0'))
6113 event = kzalloc(sizeof(*event), GFP_KERNEL);
6118 INIT_LIST_HEAD(&event->list);
6119 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
6120 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
6121 INIT_WORK(&event->remove, cgroup_event_remove);
6129 event->eventfd = eventfd_ctx_fileget(efile.file);
6130 if (IS_ERR(event->eventfd)) {
6131 ret = PTR_ERR(event->eventfd);
6138 goto out_put_eventfd;
6141 /* the process need read permission on control file */
6142 /* AV: shouldn't we check that it's been opened for read instead? */
6143 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6148 * Determine the event callbacks and set them in @event. This used
6149 * to be done via struct cftype but cgroup core no longer knows
6150 * about these events. The following is crude but the whole thing
6151 * is for compatibility anyway.
6153 name = cfile.file->f_dentry->d_name.name;
6155 if (!strcmp(name, "memory.usage_in_bytes")) {
6156 event->register_event = mem_cgroup_usage_register_event;
6157 event->unregister_event = mem_cgroup_usage_unregister_event;
6158 } else if (!strcmp(name, "memory.oom_control")) {
6159 event->register_event = mem_cgroup_oom_register_event;
6160 event->unregister_event = mem_cgroup_oom_unregister_event;
6161 } else if (!strcmp(name, "memory.pressure_level")) {
6162 event->register_event = vmpressure_register_event;
6163 event->unregister_event = vmpressure_unregister_event;
6164 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
6165 event->register_event = memsw_cgroup_usage_register_event;
6166 event->unregister_event = memsw_cgroup_usage_unregister_event;
6173 * Verify @cfile should belong to @css. Also, remaining events are
6174 * automatically removed on cgroup destruction but the removal is
6175 * asynchronous, so take an extra ref on @css.
6180 cfile_css = css_from_dir(cfile.file->f_dentry->d_parent,
6181 &mem_cgroup_subsys);
6182 if (cfile_css == css && css_tryget(css))
6189 ret = event->register_event(css, event->eventfd, buffer);
6193 efile.file->f_op->poll(efile.file, &event->pt);
6195 spin_lock(&memcg->event_list_lock);
6196 list_add(&event->list, &memcg->event_list);
6197 spin_unlock(&memcg->event_list_lock);
6209 eventfd_ctx_put(event->eventfd);
6218 static struct cftype mem_cgroup_files[] = {
6220 .name = "usage_in_bytes",
6221 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6222 .read = mem_cgroup_read,
6225 .name = "max_usage_in_bytes",
6226 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6227 .trigger = mem_cgroup_reset,
6228 .read = mem_cgroup_read,
6231 .name = "limit_in_bytes",
6232 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6233 .write_string = mem_cgroup_write,
6234 .read = mem_cgroup_read,
6237 .name = "soft_limit_in_bytes",
6238 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6239 .write_string = mem_cgroup_write,
6240 .read = mem_cgroup_read,
6244 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6245 .trigger = mem_cgroup_reset,
6246 .read = mem_cgroup_read,
6250 .read_seq_string = memcg_stat_show,
6253 .name = "force_empty",
6254 .trigger = mem_cgroup_force_empty_write,
6257 .name = "use_hierarchy",
6258 .flags = CFTYPE_INSANE,
6259 .write_u64 = mem_cgroup_hierarchy_write,
6260 .read_u64 = mem_cgroup_hierarchy_read,
6263 .name = "cgroup.event_control",
6264 .write_string = cgroup_write_event_control,
6265 .flags = CFTYPE_NO_PREFIX,
6269 .name = "swappiness",
6270 .read_u64 = mem_cgroup_swappiness_read,
6271 .write_u64 = mem_cgroup_swappiness_write,
6274 .name = "move_charge_at_immigrate",
6275 .read_u64 = mem_cgroup_move_charge_read,
6276 .write_u64 = mem_cgroup_move_charge_write,
6279 .name = "oom_control",
6280 .read_map = mem_cgroup_oom_control_read,
6281 .write_u64 = mem_cgroup_oom_control_write,
6282 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6285 .name = "pressure_level",
6289 .name = "numa_stat",
6290 .read_seq_string = memcg_numa_stat_show,
6293 #ifdef CONFIG_MEMCG_KMEM
6295 .name = "kmem.limit_in_bytes",
6296 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6297 .write_string = mem_cgroup_write,
6298 .read = mem_cgroup_read,
6301 .name = "kmem.usage_in_bytes",
6302 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6303 .read = mem_cgroup_read,
6306 .name = "kmem.failcnt",
6307 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6308 .trigger = mem_cgroup_reset,
6309 .read = mem_cgroup_read,
6312 .name = "kmem.max_usage_in_bytes",
6313 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6314 .trigger = mem_cgroup_reset,
6315 .read = mem_cgroup_read,
6317 #ifdef CONFIG_SLABINFO
6319 .name = "kmem.slabinfo",
6320 .read_seq_string = mem_cgroup_slabinfo_read,
6324 { }, /* terminate */
6327 #ifdef CONFIG_MEMCG_SWAP
6328 static struct cftype memsw_cgroup_files[] = {
6330 .name = "memsw.usage_in_bytes",
6331 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6332 .read = mem_cgroup_read,
6335 .name = "memsw.max_usage_in_bytes",
6336 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6337 .trigger = mem_cgroup_reset,
6338 .read = mem_cgroup_read,
6341 .name = "memsw.limit_in_bytes",
6342 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6343 .write_string = mem_cgroup_write,
6344 .read = mem_cgroup_read,
6347 .name = "memsw.failcnt",
6348 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6349 .trigger = mem_cgroup_reset,
6350 .read = mem_cgroup_read,
6352 { }, /* terminate */
6355 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6357 struct mem_cgroup_per_node *pn;
6358 struct mem_cgroup_per_zone *mz;
6359 int zone, tmp = node;
6361 * This routine is called against possible nodes.
6362 * But it's BUG to call kmalloc() against offline node.
6364 * TODO: this routine can waste much memory for nodes which will
6365 * never be onlined. It's better to use memory hotplug callback
6368 if (!node_state(node, N_NORMAL_MEMORY))
6370 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6374 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6375 mz = &pn->zoneinfo[zone];
6376 lruvec_init(&mz->lruvec);
6377 mz->usage_in_excess = 0;
6378 mz->on_tree = false;
6381 memcg->nodeinfo[node] = pn;
6385 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6387 kfree(memcg->nodeinfo[node]);
6390 static struct mem_cgroup *mem_cgroup_alloc(void)
6392 struct mem_cgroup *memcg;
6393 size_t size = memcg_size();
6395 /* Can be very big if nr_node_ids is very big */
6396 if (size < PAGE_SIZE)
6397 memcg = kzalloc(size, GFP_KERNEL);
6399 memcg = vzalloc(size);
6404 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6407 spin_lock_init(&memcg->pcp_counter_lock);
6411 if (size < PAGE_SIZE)
6419 * At destroying mem_cgroup, references from swap_cgroup can remain.
6420 * (scanning all at force_empty is too costly...)
6422 * Instead of clearing all references at force_empty, we remember
6423 * the number of reference from swap_cgroup and free mem_cgroup when
6424 * it goes down to 0.
6426 * Removal of cgroup itself succeeds regardless of refs from swap.
6429 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6432 size_t size = memcg_size();
6434 mem_cgroup_remove_from_trees(memcg);
6435 free_css_id(&mem_cgroup_subsys, &memcg->css);
6438 free_mem_cgroup_per_zone_info(memcg, node);
6440 free_percpu(memcg->stat);
6443 * We need to make sure that (at least for now), the jump label
6444 * destruction code runs outside of the cgroup lock. This is because
6445 * get_online_cpus(), which is called from the static_branch update,
6446 * can't be called inside the cgroup_lock. cpusets are the ones
6447 * enforcing this dependency, so if they ever change, we might as well.
6449 * schedule_work() will guarantee this happens. Be careful if you need
6450 * to move this code around, and make sure it is outside
6453 disarm_static_keys(memcg);
6454 if (size < PAGE_SIZE)
6461 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6463 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6465 if (!memcg->res.parent)
6467 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6469 EXPORT_SYMBOL(parent_mem_cgroup);
6471 static void __init mem_cgroup_soft_limit_tree_init(void)
6473 struct mem_cgroup_tree_per_node *rtpn;
6474 struct mem_cgroup_tree_per_zone *rtpz;
6475 int tmp, node, zone;
6477 for_each_node(node) {
6479 if (!node_state(node, N_NORMAL_MEMORY))
6481 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6484 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6486 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6487 rtpz = &rtpn->rb_tree_per_zone[zone];
6488 rtpz->rb_root = RB_ROOT;
6489 spin_lock_init(&rtpz->lock);
6494 static struct cgroup_subsys_state * __ref
6495 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6497 struct mem_cgroup *memcg;
6498 long error = -ENOMEM;
6501 memcg = mem_cgroup_alloc();
6503 return ERR_PTR(error);
6506 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6510 if (parent_css == NULL) {
6511 root_mem_cgroup = memcg;
6512 res_counter_init(&memcg->res, NULL);
6513 res_counter_init(&memcg->memsw, NULL);
6514 res_counter_init(&memcg->kmem, NULL);
6517 memcg->last_scanned_node = MAX_NUMNODES;
6518 INIT_LIST_HEAD(&memcg->oom_notify);
6519 memcg->move_charge_at_immigrate = 0;
6520 mutex_init(&memcg->thresholds_lock);
6521 spin_lock_init(&memcg->move_lock);
6522 vmpressure_init(&memcg->vmpressure);
6523 INIT_LIST_HEAD(&memcg->event_list);
6524 spin_lock_init(&memcg->event_list_lock);
6529 __mem_cgroup_free(memcg);
6530 return ERR_PTR(error);
6534 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6536 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6537 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6543 mutex_lock(&memcg_create_mutex);
6545 memcg->use_hierarchy = parent->use_hierarchy;
6546 memcg->oom_kill_disable = parent->oom_kill_disable;
6547 memcg->swappiness = mem_cgroup_swappiness(parent);
6549 if (parent->use_hierarchy) {
6550 res_counter_init(&memcg->res, &parent->res);
6551 res_counter_init(&memcg->memsw, &parent->memsw);
6552 res_counter_init(&memcg->kmem, &parent->kmem);
6555 * No need to take a reference to the parent because cgroup
6556 * core guarantees its existence.
6559 res_counter_init(&memcg->res, NULL);
6560 res_counter_init(&memcg->memsw, NULL);
6561 res_counter_init(&memcg->kmem, NULL);
6563 * Deeper hierachy with use_hierarchy == false doesn't make
6564 * much sense so let cgroup subsystem know about this
6565 * unfortunate state in our controller.
6567 if (parent != root_mem_cgroup)
6568 mem_cgroup_subsys.broken_hierarchy = true;
6571 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6572 mutex_unlock(&memcg_create_mutex);
6577 * Announce all parents that a group from their hierarchy is gone.
6579 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6581 struct mem_cgroup *parent = memcg;
6583 while ((parent = parent_mem_cgroup(parent)))
6584 mem_cgroup_iter_invalidate(parent);
6587 * if the root memcg is not hierarchical we have to check it
6590 if (!root_mem_cgroup->use_hierarchy)
6591 mem_cgroup_iter_invalidate(root_mem_cgroup);
6594 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6596 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6597 struct cgroup_event *event, *tmp;
6600 * Unregister events and notify userspace.
6601 * Notify userspace about cgroup removing only after rmdir of cgroup
6602 * directory to avoid race between userspace and kernelspace.
6604 spin_lock(&memcg->event_list_lock);
6605 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6606 list_del_init(&event->list);
6607 schedule_work(&event->remove);
6609 spin_unlock(&memcg->event_list_lock);
6611 kmem_cgroup_css_offline(memcg);
6613 mem_cgroup_invalidate_reclaim_iterators(memcg);
6614 mem_cgroup_reparent_charges(memcg);
6615 mem_cgroup_destroy_all_caches(memcg);
6616 vmpressure_cleanup(&memcg->vmpressure);
6619 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6621 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6623 memcg_destroy_kmem(memcg);
6624 __mem_cgroup_free(memcg);
6628 /* Handlers for move charge at task migration. */
6629 #define PRECHARGE_COUNT_AT_ONCE 256
6630 static int mem_cgroup_do_precharge(unsigned long count)
6633 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6634 struct mem_cgroup *memcg = mc.to;
6636 if (mem_cgroup_is_root(memcg)) {
6637 mc.precharge += count;
6638 /* we don't need css_get for root */
6641 /* try to charge at once */
6643 struct res_counter *dummy;
6645 * "memcg" cannot be under rmdir() because we've already checked
6646 * by cgroup_lock_live_cgroup() that it is not removed and we
6647 * are still under the same cgroup_mutex. So we can postpone
6650 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6652 if (do_swap_account && res_counter_charge(&memcg->memsw,
6653 PAGE_SIZE * count, &dummy)) {
6654 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6657 mc.precharge += count;
6661 /* fall back to one by one charge */
6663 if (signal_pending(current)) {
6667 if (!batch_count--) {
6668 batch_count = PRECHARGE_COUNT_AT_ONCE;
6671 ret = __mem_cgroup_try_charge(NULL,
6672 GFP_KERNEL, 1, &memcg, false);
6674 /* mem_cgroup_clear_mc() will do uncharge later */
6682 * get_mctgt_type - get target type of moving charge
6683 * @vma: the vma the pte to be checked belongs
6684 * @addr: the address corresponding to the pte to be checked
6685 * @ptent: the pte to be checked
6686 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6689 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6690 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6691 * move charge. if @target is not NULL, the page is stored in target->page
6692 * with extra refcnt got(Callers should handle it).
6693 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6694 * target for charge migration. if @target is not NULL, the entry is stored
6697 * Called with pte lock held.
6704 enum mc_target_type {
6710 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6711 unsigned long addr, pte_t ptent)
6713 struct page *page = vm_normal_page(vma, addr, ptent);
6715 if (!page || !page_mapped(page))
6717 if (PageAnon(page)) {
6718 /* we don't move shared anon */
6721 } else if (!move_file())
6722 /* we ignore mapcount for file pages */
6724 if (!get_page_unless_zero(page))
6731 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6732 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6734 struct page *page = NULL;
6735 swp_entry_t ent = pte_to_swp_entry(ptent);
6737 if (!move_anon() || non_swap_entry(ent))
6740 * Because lookup_swap_cache() updates some statistics counter,
6741 * we call find_get_page() with swapper_space directly.
6743 page = find_get_page(swap_address_space(ent), ent.val);
6744 if (do_swap_account)
6745 entry->val = ent.val;
6750 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6751 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6757 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6758 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6760 struct page *page = NULL;
6761 struct address_space *mapping;
6764 if (!vma->vm_file) /* anonymous vma */
6769 mapping = vma->vm_file->f_mapping;
6770 if (pte_none(ptent))
6771 pgoff = linear_page_index(vma, addr);
6772 else /* pte_file(ptent) is true */
6773 pgoff = pte_to_pgoff(ptent);
6775 /* page is moved even if it's not RSS of this task(page-faulted). */
6776 page = find_get_page(mapping, pgoff);
6779 /* shmem/tmpfs may report page out on swap: account for that too. */
6780 if (radix_tree_exceptional_entry(page)) {
6781 swp_entry_t swap = radix_to_swp_entry(page);
6782 if (do_swap_account)
6784 page = find_get_page(swap_address_space(swap), swap.val);
6790 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6791 unsigned long addr, pte_t ptent, union mc_target *target)
6793 struct page *page = NULL;
6794 struct page_cgroup *pc;
6795 enum mc_target_type ret = MC_TARGET_NONE;
6796 swp_entry_t ent = { .val = 0 };
6798 if (pte_present(ptent))
6799 page = mc_handle_present_pte(vma, addr, ptent);
6800 else if (is_swap_pte(ptent))
6801 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6802 else if (pte_none(ptent) || pte_file(ptent))
6803 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6805 if (!page && !ent.val)
6808 pc = lookup_page_cgroup(page);
6810 * Do only loose check w/o page_cgroup lock.
6811 * mem_cgroup_move_account() checks the pc is valid or not under
6814 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6815 ret = MC_TARGET_PAGE;
6817 target->page = page;
6819 if (!ret || !target)
6822 /* There is a swap entry and a page doesn't exist or isn't charged */
6823 if (ent.val && !ret &&
6824 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6825 ret = MC_TARGET_SWAP;
6832 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6834 * We don't consider swapping or file mapped pages because THP does not
6835 * support them for now.
6836 * Caller should make sure that pmd_trans_huge(pmd) is true.
6838 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6839 unsigned long addr, pmd_t pmd, union mc_target *target)
6841 struct page *page = NULL;
6842 struct page_cgroup *pc;
6843 enum mc_target_type ret = MC_TARGET_NONE;
6845 page = pmd_page(pmd);
6846 VM_BUG_ON(!page || !PageHead(page));
6849 pc = lookup_page_cgroup(page);
6850 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6851 ret = MC_TARGET_PAGE;
6854 target->page = page;
6860 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6861 unsigned long addr, pmd_t pmd, union mc_target *target)
6863 return MC_TARGET_NONE;
6867 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6868 unsigned long addr, unsigned long end,
6869 struct mm_walk *walk)
6871 struct vm_area_struct *vma = walk->private;
6875 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6876 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6877 mc.precharge += HPAGE_PMD_NR;
6878 spin_unlock(&vma->vm_mm->page_table_lock);
6882 if (pmd_trans_unstable(pmd))
6884 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6885 for (; addr != end; pte++, addr += PAGE_SIZE)
6886 if (get_mctgt_type(vma, addr, *pte, NULL))
6887 mc.precharge++; /* increment precharge temporarily */
6888 pte_unmap_unlock(pte - 1, ptl);
6894 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6896 unsigned long precharge;
6897 struct vm_area_struct *vma;
6899 down_read(&mm->mmap_sem);
6900 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6901 struct mm_walk mem_cgroup_count_precharge_walk = {
6902 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6906 if (is_vm_hugetlb_page(vma))
6908 walk_page_range(vma->vm_start, vma->vm_end,
6909 &mem_cgroup_count_precharge_walk);
6911 up_read(&mm->mmap_sem);
6913 precharge = mc.precharge;
6919 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6921 unsigned long precharge = mem_cgroup_count_precharge(mm);
6923 VM_BUG_ON(mc.moving_task);
6924 mc.moving_task = current;
6925 return mem_cgroup_do_precharge(precharge);
6928 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6929 static void __mem_cgroup_clear_mc(void)
6931 struct mem_cgroup *from = mc.from;
6932 struct mem_cgroup *to = mc.to;
6935 /* we must uncharge all the leftover precharges from mc.to */
6937 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6941 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6942 * we must uncharge here.
6944 if (mc.moved_charge) {
6945 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6946 mc.moved_charge = 0;
6948 /* we must fixup refcnts and charges */
6949 if (mc.moved_swap) {
6950 /* uncharge swap account from the old cgroup */
6951 if (!mem_cgroup_is_root(mc.from))
6952 res_counter_uncharge(&mc.from->memsw,
6953 PAGE_SIZE * mc.moved_swap);
6955 for (i = 0; i < mc.moved_swap; i++)
6956 css_put(&mc.from->css);
6958 if (!mem_cgroup_is_root(mc.to)) {
6960 * we charged both to->res and to->memsw, so we should
6963 res_counter_uncharge(&mc.to->res,
6964 PAGE_SIZE * mc.moved_swap);
6966 /* we've already done css_get(mc.to) */
6969 memcg_oom_recover(from);
6970 memcg_oom_recover(to);
6971 wake_up_all(&mc.waitq);
6974 static void mem_cgroup_clear_mc(void)
6976 struct mem_cgroup *from = mc.from;
6979 * we must clear moving_task before waking up waiters at the end of
6982 mc.moving_task = NULL;
6983 __mem_cgroup_clear_mc();
6984 spin_lock(&mc.lock);
6987 spin_unlock(&mc.lock);
6988 mem_cgroup_end_move(from);
6991 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6992 struct cgroup_taskset *tset)
6994 struct task_struct *p = cgroup_taskset_first(tset);
6996 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6997 unsigned long move_charge_at_immigrate;
7000 * We are now commited to this value whatever it is. Changes in this
7001 * tunable will only affect upcoming migrations, not the current one.
7002 * So we need to save it, and keep it going.
7004 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
7005 if (move_charge_at_immigrate) {
7006 struct mm_struct *mm;
7007 struct mem_cgroup *from = mem_cgroup_from_task(p);
7009 VM_BUG_ON(from == memcg);
7011 mm = get_task_mm(p);
7014 /* We move charges only when we move a owner of the mm */
7015 if (mm->owner == p) {
7018 VM_BUG_ON(mc.precharge);
7019 VM_BUG_ON(mc.moved_charge);
7020 VM_BUG_ON(mc.moved_swap);
7021 mem_cgroup_start_move(from);
7022 spin_lock(&mc.lock);
7025 mc.immigrate_flags = move_charge_at_immigrate;
7026 spin_unlock(&mc.lock);
7027 /* We set mc.moving_task later */
7029 ret = mem_cgroup_precharge_mc(mm);
7031 mem_cgroup_clear_mc();
7038 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7039 struct cgroup_taskset *tset)
7041 mem_cgroup_clear_mc();
7044 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
7045 unsigned long addr, unsigned long end,
7046 struct mm_walk *walk)
7049 struct vm_area_struct *vma = walk->private;
7052 enum mc_target_type target_type;
7053 union mc_target target;
7055 struct page_cgroup *pc;
7058 * We don't take compound_lock() here but no race with splitting thp
7060 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7061 * under splitting, which means there's no concurrent thp split,
7062 * - if another thread runs into split_huge_page() just after we
7063 * entered this if-block, the thread must wait for page table lock
7064 * to be unlocked in __split_huge_page_splitting(), where the main
7065 * part of thp split is not executed yet.
7067 if (pmd_trans_huge_lock(pmd, vma) == 1) {
7068 if (mc.precharge < HPAGE_PMD_NR) {
7069 spin_unlock(&vma->vm_mm->page_table_lock);
7072 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
7073 if (target_type == MC_TARGET_PAGE) {
7075 if (!isolate_lru_page(page)) {
7076 pc = lookup_page_cgroup(page);
7077 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
7078 pc, mc.from, mc.to)) {
7079 mc.precharge -= HPAGE_PMD_NR;
7080 mc.moved_charge += HPAGE_PMD_NR;
7082 putback_lru_page(page);
7086 spin_unlock(&vma->vm_mm->page_table_lock);
7090 if (pmd_trans_unstable(pmd))
7093 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7094 for (; addr != end; addr += PAGE_SIZE) {
7095 pte_t ptent = *(pte++);
7101 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7102 case MC_TARGET_PAGE:
7104 if (isolate_lru_page(page))
7106 pc = lookup_page_cgroup(page);
7107 if (!mem_cgroup_move_account(page, 1, pc,
7110 /* we uncharge from mc.from later. */
7113 putback_lru_page(page);
7114 put: /* get_mctgt_type() gets the page */
7117 case MC_TARGET_SWAP:
7119 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7121 /* we fixup refcnts and charges later. */
7129 pte_unmap_unlock(pte - 1, ptl);
7134 * We have consumed all precharges we got in can_attach().
7135 * We try charge one by one, but don't do any additional
7136 * charges to mc.to if we have failed in charge once in attach()
7139 ret = mem_cgroup_do_precharge(1);
7147 static void mem_cgroup_move_charge(struct mm_struct *mm)
7149 struct vm_area_struct *vma;
7151 lru_add_drain_all();
7153 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7155 * Someone who are holding the mmap_sem might be waiting in
7156 * waitq. So we cancel all extra charges, wake up all waiters,
7157 * and retry. Because we cancel precharges, we might not be able
7158 * to move enough charges, but moving charge is a best-effort
7159 * feature anyway, so it wouldn't be a big problem.
7161 __mem_cgroup_clear_mc();
7165 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7167 struct mm_walk mem_cgroup_move_charge_walk = {
7168 .pmd_entry = mem_cgroup_move_charge_pte_range,
7172 if (is_vm_hugetlb_page(vma))
7174 ret = walk_page_range(vma->vm_start, vma->vm_end,
7175 &mem_cgroup_move_charge_walk);
7178 * means we have consumed all precharges and failed in
7179 * doing additional charge. Just abandon here.
7183 up_read(&mm->mmap_sem);
7186 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7187 struct cgroup_taskset *tset)
7189 struct task_struct *p = cgroup_taskset_first(tset);
7190 struct mm_struct *mm = get_task_mm(p);
7194 mem_cgroup_move_charge(mm);
7198 mem_cgroup_clear_mc();
7200 #else /* !CONFIG_MMU */
7201 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7202 struct cgroup_taskset *tset)
7206 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7207 struct cgroup_taskset *tset)
7210 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7211 struct cgroup_taskset *tset)
7217 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7218 * to verify sane_behavior flag on each mount attempt.
7220 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7223 * use_hierarchy is forced with sane_behavior. cgroup core
7224 * guarantees that @root doesn't have any children, so turning it
7225 * on for the root memcg is enough.
7227 if (cgroup_sane_behavior(root_css->cgroup))
7228 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7231 struct cgroup_subsys mem_cgroup_subsys = {
7233 .subsys_id = mem_cgroup_subsys_id,
7234 .css_alloc = mem_cgroup_css_alloc,
7235 .css_online = mem_cgroup_css_online,
7236 .css_offline = mem_cgroup_css_offline,
7237 .css_free = mem_cgroup_css_free,
7238 .can_attach = mem_cgroup_can_attach,
7239 .cancel_attach = mem_cgroup_cancel_attach,
7240 .attach = mem_cgroup_move_task,
7241 .bind = mem_cgroup_bind,
7242 .base_cftypes = mem_cgroup_files,
7247 #ifdef CONFIG_MEMCG_SWAP
7248 static int __init enable_swap_account(char *s)
7250 if (!strcmp(s, "1"))
7251 really_do_swap_account = 1;
7252 else if (!strcmp(s, "0"))
7253 really_do_swap_account = 0;
7256 __setup("swapaccount=", enable_swap_account);
7258 static void __init memsw_file_init(void)
7260 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7263 static void __init enable_swap_cgroup(void)
7265 if (!mem_cgroup_disabled() && really_do_swap_account) {
7266 do_swap_account = 1;
7272 static void __init enable_swap_cgroup(void)
7278 * subsys_initcall() for memory controller.
7280 * Some parts like hotcpu_notifier() have to be initialized from this context
7281 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7282 * everything that doesn't depend on a specific mem_cgroup structure should
7283 * be initialized from here.
7285 static int __init mem_cgroup_init(void)
7287 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7288 enable_swap_cgroup();
7289 mem_cgroup_soft_limit_tree_init();
7293 subsys_initcall(mem_cgroup_init);