1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct mem_cgroup_reclaim_iter {
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup *last_visited;
154 /* scan generation, increased every round-trip */
155 unsigned int generation;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone {
162 struct lruvec lruvec;
163 unsigned long lru_size[NR_LRU_LISTS];
165 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
167 struct rb_node tree_node; /* RB tree node */
168 unsigned long long usage_in_excess;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup *memcg; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node {
176 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone {
185 struct rb_root rb_root;
189 struct mem_cgroup_tree_per_node {
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
193 struct mem_cgroup_tree {
194 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
199 struct mem_cgroup_threshold {
200 struct eventfd_ctx *eventfd;
205 struct mem_cgroup_threshold_ary {
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries[0];
214 struct mem_cgroup_thresholds {
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary *primary;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary *spare;
226 struct mem_cgroup_eventfd_list {
227 struct list_head list;
228 struct eventfd_ctx *eventfd;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event {
236 * memcg which the event belongs to.
238 struct mem_cgroup *memcg;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx *eventfd;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event)(struct mem_cgroup *memcg,
253 struct eventfd_ctx *eventfd, const char *args);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event)(struct mem_cgroup *memcg,
260 struct eventfd_ctx *eventfd);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t *wqh;
268 struct work_struct remove;
271 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
272 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css;
288 * the counter to account for memory usage
290 struct res_counter res;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure;
296 * the counter to account for mem+swap usage.
298 struct res_counter memsw;
301 * the counter to account for kernel memory usage.
303 struct res_counter kmem;
305 * Should the accounting and control be hierarchical, per subtree?
308 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
312 atomic_t oom_wakeups;
315 /* OOM-Killer disable */
316 int oom_kill_disable;
318 /* set when res.limit == memsw.limit */
319 bool memsw_is_minimum;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds;
330 /* For oom notifier event fd */
331 struct list_head oom_notify;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock;
347 struct mem_cgroup_stat_cpu __percpu *stat;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base;
353 spinlock_t pcp_counter_lock;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list, but per-memcg;
361 * protected by memcg_slab_mutex */
362 struct list_head memcg_slab_caches;
363 /* Index in the kmem_cache->memcg_params->memcg_caches array */
367 int last_scanned_node;
369 nodemask_t scan_nodes;
370 atomic_t numainfo_events;
371 atomic_t numainfo_updating;
374 /* List of events which userspace want to receive */
375 struct list_head event_list;
376 spinlock_t event_list_lock;
378 struct mem_cgroup_per_node *nodeinfo[0];
379 /* WARNING: nodeinfo must be the last member here */
382 /* internal only representation about the status of kmem accounting. */
384 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
385 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
388 #ifdef CONFIG_MEMCG_KMEM
389 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
391 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
394 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
396 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
399 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
402 * Our caller must use css_get() first, because memcg_uncharge_kmem()
403 * will call css_put() if it sees the memcg is dead.
406 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
407 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
410 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
412 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
413 &memcg->kmem_account_flags);
417 /* Stuffs for move charges at task migration. */
419 * Types of charges to be moved. "move_charge_at_immitgrate" and
420 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
423 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
424 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
428 /* "mc" and its members are protected by cgroup_mutex */
429 static struct move_charge_struct {
430 spinlock_t lock; /* for from, to */
431 struct mem_cgroup *from;
432 struct mem_cgroup *to;
433 unsigned long immigrate_flags;
434 unsigned long precharge;
435 unsigned long moved_charge;
436 unsigned long moved_swap;
437 struct task_struct *moving_task; /* a task moving charges */
438 wait_queue_head_t waitq; /* a waitq for other context */
440 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
441 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
444 static bool move_anon(void)
446 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
449 static bool move_file(void)
451 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
455 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
456 * limit reclaim to prevent infinite loops, if they ever occur.
458 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
459 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
462 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
463 MEM_CGROUP_CHARGE_TYPE_ANON,
464 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
465 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
469 /* for encoding cft->private value on file */
477 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
478 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
479 #define MEMFILE_ATTR(val) ((val) & 0xffff)
480 /* Used for OOM nofiier */
481 #define OOM_CONTROL (0)
484 * Reclaim flags for mem_cgroup_hierarchical_reclaim
486 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
487 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
488 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
489 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
492 * The memcg_create_mutex will be held whenever a new cgroup is created.
493 * As a consequence, any change that needs to protect against new child cgroups
494 * appearing has to hold it as well.
496 static DEFINE_MUTEX(memcg_create_mutex);
498 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
500 return s ? container_of(s, struct mem_cgroup, css) : NULL;
503 /* Some nice accessors for the vmpressure. */
504 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
507 memcg = root_mem_cgroup;
508 return &memcg->vmpressure;
511 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
513 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
516 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
518 return (memcg == root_mem_cgroup);
522 * We restrict the id in the range of [1, 65535], so it can fit into
525 #define MEM_CGROUP_ID_MAX USHRT_MAX
527 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
529 return memcg->css.id;
532 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
534 struct cgroup_subsys_state *css;
536 css = css_from_id(id, &memory_cgrp_subsys);
537 return mem_cgroup_from_css(css);
540 /* Writing them here to avoid exposing memcg's inner layout */
541 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
543 void sock_update_memcg(struct sock *sk)
545 if (mem_cgroup_sockets_enabled) {
546 struct mem_cgroup *memcg;
547 struct cg_proto *cg_proto;
549 BUG_ON(!sk->sk_prot->proto_cgroup);
551 /* Socket cloning can throw us here with sk_cgrp already
552 * filled. It won't however, necessarily happen from
553 * process context. So the test for root memcg given
554 * the current task's memcg won't help us in this case.
556 * Respecting the original socket's memcg is a better
557 * decision in this case.
560 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
561 css_get(&sk->sk_cgrp->memcg->css);
566 memcg = mem_cgroup_from_task(current);
567 cg_proto = sk->sk_prot->proto_cgroup(memcg);
568 if (!mem_cgroup_is_root(memcg) &&
569 memcg_proto_active(cg_proto) &&
570 css_tryget_online(&memcg->css)) {
571 sk->sk_cgrp = cg_proto;
576 EXPORT_SYMBOL(sock_update_memcg);
578 void sock_release_memcg(struct sock *sk)
580 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
581 struct mem_cgroup *memcg;
582 WARN_ON(!sk->sk_cgrp->memcg);
583 memcg = sk->sk_cgrp->memcg;
584 css_put(&sk->sk_cgrp->memcg->css);
588 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
590 if (!memcg || mem_cgroup_is_root(memcg))
593 return &memcg->tcp_mem;
595 EXPORT_SYMBOL(tcp_proto_cgroup);
597 static void disarm_sock_keys(struct mem_cgroup *memcg)
599 if (!memcg_proto_activated(&memcg->tcp_mem))
601 static_key_slow_dec(&memcg_socket_limit_enabled);
604 static void disarm_sock_keys(struct mem_cgroup *memcg)
609 #ifdef CONFIG_MEMCG_KMEM
611 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
612 * The main reason for not using cgroup id for this:
613 * this works better in sparse environments, where we have a lot of memcgs,
614 * but only a few kmem-limited. Or also, if we have, for instance, 200
615 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
616 * 200 entry array for that.
618 * The current size of the caches array is stored in
619 * memcg_limited_groups_array_size. It will double each time we have to
622 static DEFINE_IDA(kmem_limited_groups);
623 int memcg_limited_groups_array_size;
626 * MIN_SIZE is different than 1, because we would like to avoid going through
627 * the alloc/free process all the time. In a small machine, 4 kmem-limited
628 * cgroups is a reasonable guess. In the future, it could be a parameter or
629 * tunable, but that is strictly not necessary.
631 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
632 * this constant directly from cgroup, but it is understandable that this is
633 * better kept as an internal representation in cgroup.c. In any case, the
634 * cgrp_id space is not getting any smaller, and we don't have to necessarily
635 * increase ours as well if it increases.
637 #define MEMCG_CACHES_MIN_SIZE 4
638 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
641 * A lot of the calls to the cache allocation functions are expected to be
642 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
643 * conditional to this static branch, we'll have to allow modules that does
644 * kmem_cache_alloc and the such to see this symbol as well
646 struct static_key memcg_kmem_enabled_key;
647 EXPORT_SYMBOL(memcg_kmem_enabled_key);
649 static void disarm_kmem_keys(struct mem_cgroup *memcg)
651 if (memcg_kmem_is_active(memcg)) {
652 static_key_slow_dec(&memcg_kmem_enabled_key);
653 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
656 * This check can't live in kmem destruction function,
657 * since the charges will outlive the cgroup
659 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
662 static void disarm_kmem_keys(struct mem_cgroup *memcg)
665 #endif /* CONFIG_MEMCG_KMEM */
667 static void disarm_static_keys(struct mem_cgroup *memcg)
669 disarm_sock_keys(memcg);
670 disarm_kmem_keys(memcg);
673 static void drain_all_stock_async(struct mem_cgroup *memcg);
675 static struct mem_cgroup_per_zone *
676 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
678 VM_BUG_ON((unsigned)nid >= nr_node_ids);
679 return &memcg->nodeinfo[nid]->zoneinfo[zid];
682 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
687 static struct mem_cgroup_per_zone *
688 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
690 int nid = page_to_nid(page);
691 int zid = page_zonenum(page);
693 return mem_cgroup_zoneinfo(memcg, nid, zid);
696 static struct mem_cgroup_tree_per_zone *
697 soft_limit_tree_node_zone(int nid, int zid)
699 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
702 static struct mem_cgroup_tree_per_zone *
703 soft_limit_tree_from_page(struct page *page)
705 int nid = page_to_nid(page);
706 int zid = page_zonenum(page);
708 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
712 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
713 struct mem_cgroup_per_zone *mz,
714 struct mem_cgroup_tree_per_zone *mctz,
715 unsigned long long new_usage_in_excess)
717 struct rb_node **p = &mctz->rb_root.rb_node;
718 struct rb_node *parent = NULL;
719 struct mem_cgroup_per_zone *mz_node;
724 mz->usage_in_excess = new_usage_in_excess;
725 if (!mz->usage_in_excess)
729 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
731 if (mz->usage_in_excess < mz_node->usage_in_excess)
734 * We can't avoid mem cgroups that are over their soft
735 * limit by the same amount
737 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
740 rb_link_node(&mz->tree_node, parent, p);
741 rb_insert_color(&mz->tree_node, &mctz->rb_root);
746 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
747 struct mem_cgroup_per_zone *mz,
748 struct mem_cgroup_tree_per_zone *mctz)
752 rb_erase(&mz->tree_node, &mctz->rb_root);
757 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
758 struct mem_cgroup_per_zone *mz,
759 struct mem_cgroup_tree_per_zone *mctz)
761 spin_lock(&mctz->lock);
762 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
763 spin_unlock(&mctz->lock);
767 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
769 unsigned long long excess;
770 struct mem_cgroup_per_zone *mz;
771 struct mem_cgroup_tree_per_zone *mctz;
772 int nid = page_to_nid(page);
773 int zid = page_zonenum(page);
774 mctz = soft_limit_tree_from_page(page);
777 * Necessary to update all ancestors when hierarchy is used.
778 * because their event counter is not touched.
780 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
781 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
782 excess = res_counter_soft_limit_excess(&memcg->res);
784 * We have to update the tree if mz is on RB-tree or
785 * mem is over its softlimit.
787 if (excess || mz->on_tree) {
788 spin_lock(&mctz->lock);
789 /* if on-tree, remove it */
791 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
793 * Insert again. mz->usage_in_excess will be updated.
794 * If excess is 0, no tree ops.
796 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
797 spin_unlock(&mctz->lock);
802 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
805 struct mem_cgroup_per_zone *mz;
806 struct mem_cgroup_tree_per_zone *mctz;
808 for_each_node(node) {
809 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
810 mz = mem_cgroup_zoneinfo(memcg, node, zone);
811 mctz = soft_limit_tree_node_zone(node, zone);
812 mem_cgroup_remove_exceeded(memcg, mz, mctz);
817 static struct mem_cgroup_per_zone *
818 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
820 struct rb_node *rightmost = NULL;
821 struct mem_cgroup_per_zone *mz;
825 rightmost = rb_last(&mctz->rb_root);
827 goto done; /* Nothing to reclaim from */
829 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
831 * Remove the node now but someone else can add it back,
832 * we will to add it back at the end of reclaim to its correct
833 * position in the tree.
835 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
836 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
837 !css_tryget_online(&mz->memcg->css))
843 static struct mem_cgroup_per_zone *
844 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
846 struct mem_cgroup_per_zone *mz;
848 spin_lock(&mctz->lock);
849 mz = __mem_cgroup_largest_soft_limit_node(mctz);
850 spin_unlock(&mctz->lock);
855 * Implementation Note: reading percpu statistics for memcg.
857 * Both of vmstat[] and percpu_counter has threshold and do periodic
858 * synchronization to implement "quick" read. There are trade-off between
859 * reading cost and precision of value. Then, we may have a chance to implement
860 * a periodic synchronizion of counter in memcg's counter.
862 * But this _read() function is used for user interface now. The user accounts
863 * memory usage by memory cgroup and he _always_ requires exact value because
864 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
865 * have to visit all online cpus and make sum. So, for now, unnecessary
866 * synchronization is not implemented. (just implemented for cpu hotplug)
868 * If there are kernel internal actions which can make use of some not-exact
869 * value, and reading all cpu value can be performance bottleneck in some
870 * common workload, threashold and synchonization as vmstat[] should be
873 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
874 enum mem_cgroup_stat_index idx)
880 for_each_online_cpu(cpu)
881 val += per_cpu(memcg->stat->count[idx], cpu);
882 #ifdef CONFIG_HOTPLUG_CPU
883 spin_lock(&memcg->pcp_counter_lock);
884 val += memcg->nocpu_base.count[idx];
885 spin_unlock(&memcg->pcp_counter_lock);
891 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
894 int val = (charge) ? 1 : -1;
895 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
898 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
899 enum mem_cgroup_events_index idx)
901 unsigned long val = 0;
905 for_each_online_cpu(cpu)
906 val += per_cpu(memcg->stat->events[idx], cpu);
907 #ifdef CONFIG_HOTPLUG_CPU
908 spin_lock(&memcg->pcp_counter_lock);
909 val += memcg->nocpu_base.events[idx];
910 spin_unlock(&memcg->pcp_counter_lock);
916 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
918 bool anon, int nr_pages)
921 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
922 * counted as CACHE even if it's on ANON LRU.
925 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
928 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
931 if (PageTransHuge(page))
932 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
935 /* pagein of a big page is an event. So, ignore page size */
937 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
939 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
940 nr_pages = -nr_pages; /* for event */
943 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
947 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
949 struct mem_cgroup_per_zone *mz;
951 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
952 return mz->lru_size[lru];
956 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
957 unsigned int lru_mask)
959 struct mem_cgroup_per_zone *mz;
961 unsigned long ret = 0;
963 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
966 if (BIT(lru) & lru_mask)
967 ret += mz->lru_size[lru];
973 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
974 int nid, unsigned int lru_mask)
979 for (zid = 0; zid < MAX_NR_ZONES; zid++)
980 total += mem_cgroup_zone_nr_lru_pages(memcg,
986 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
987 unsigned int lru_mask)
992 for_each_node_state(nid, N_MEMORY)
993 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
997 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
998 enum mem_cgroup_events_target target)
1000 unsigned long val, next;
1002 val = __this_cpu_read(memcg->stat->nr_page_events);
1003 next = __this_cpu_read(memcg->stat->targets[target]);
1004 /* from time_after() in jiffies.h */
1005 if ((long)next - (long)val < 0) {
1007 case MEM_CGROUP_TARGET_THRESH:
1008 next = val + THRESHOLDS_EVENTS_TARGET;
1010 case MEM_CGROUP_TARGET_SOFTLIMIT:
1011 next = val + SOFTLIMIT_EVENTS_TARGET;
1013 case MEM_CGROUP_TARGET_NUMAINFO:
1014 next = val + NUMAINFO_EVENTS_TARGET;
1019 __this_cpu_write(memcg->stat->targets[target], next);
1026 * Check events in order.
1029 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1032 /* threshold event is triggered in finer grain than soft limit */
1033 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1034 MEM_CGROUP_TARGET_THRESH))) {
1036 bool do_numainfo __maybe_unused;
1038 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1039 MEM_CGROUP_TARGET_SOFTLIMIT);
1040 #if MAX_NUMNODES > 1
1041 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1042 MEM_CGROUP_TARGET_NUMAINFO);
1046 mem_cgroup_threshold(memcg);
1047 if (unlikely(do_softlimit))
1048 mem_cgroup_update_tree(memcg, page);
1049 #if MAX_NUMNODES > 1
1050 if (unlikely(do_numainfo))
1051 atomic_inc(&memcg->numainfo_events);
1057 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1060 * mm_update_next_owner() may clear mm->owner to NULL
1061 * if it races with swapoff, page migration, etc.
1062 * So this can be called with p == NULL.
1067 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1070 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1072 struct mem_cgroup *memcg = NULL;
1076 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1077 if (unlikely(!memcg))
1078 memcg = root_mem_cgroup;
1079 } while (!css_tryget_online(&memcg->css));
1085 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1086 * ref. count) or NULL if the whole root's subtree has been visited.
1088 * helper function to be used by mem_cgroup_iter
1090 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1091 struct mem_cgroup *last_visited)
1093 struct cgroup_subsys_state *prev_css, *next_css;
1095 prev_css = last_visited ? &last_visited->css : NULL;
1097 next_css = css_next_descendant_pre(prev_css, &root->css);
1100 * Even if we found a group we have to make sure it is
1101 * alive. css && !memcg means that the groups should be
1102 * skipped and we should continue the tree walk.
1103 * last_visited css is safe to use because it is
1104 * protected by css_get and the tree walk is rcu safe.
1106 * We do not take a reference on the root of the tree walk
1107 * because we might race with the root removal when it would
1108 * be the only node in the iterated hierarchy and mem_cgroup_iter
1109 * would end up in an endless loop because it expects that at
1110 * least one valid node will be returned. Root cannot disappear
1111 * because caller of the iterator should hold it already so
1112 * skipping css reference should be safe.
1115 if ((next_css == &root->css) ||
1116 ((next_css->flags & CSS_ONLINE) &&
1117 css_tryget_online(next_css)))
1118 return mem_cgroup_from_css(next_css);
1120 prev_css = next_css;
1127 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1130 * When a group in the hierarchy below root is destroyed, the
1131 * hierarchy iterator can no longer be trusted since it might
1132 * have pointed to the destroyed group. Invalidate it.
1134 atomic_inc(&root->dead_count);
1137 static struct mem_cgroup *
1138 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1139 struct mem_cgroup *root,
1142 struct mem_cgroup *position = NULL;
1144 * A cgroup destruction happens in two stages: offlining and
1145 * release. They are separated by a RCU grace period.
1147 * If the iterator is valid, we may still race with an
1148 * offlining. The RCU lock ensures the object won't be
1149 * released, tryget will fail if we lost the race.
1151 *sequence = atomic_read(&root->dead_count);
1152 if (iter->last_dead_count == *sequence) {
1154 position = iter->last_visited;
1157 * We cannot take a reference to root because we might race
1158 * with root removal and returning NULL would end up in
1159 * an endless loop on the iterator user level when root
1160 * would be returned all the time.
1162 if (position && position != root &&
1163 !css_tryget_online(&position->css))
1169 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1170 struct mem_cgroup *last_visited,
1171 struct mem_cgroup *new_position,
1172 struct mem_cgroup *root,
1175 /* root reference counting symmetric to mem_cgroup_iter_load */
1176 if (last_visited && last_visited != root)
1177 css_put(&last_visited->css);
1179 * We store the sequence count from the time @last_visited was
1180 * loaded successfully instead of rereading it here so that we
1181 * don't lose destruction events in between. We could have
1182 * raced with the destruction of @new_position after all.
1184 iter->last_visited = new_position;
1186 iter->last_dead_count = sequence;
1190 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1191 * @root: hierarchy root
1192 * @prev: previously returned memcg, NULL on first invocation
1193 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1195 * Returns references to children of the hierarchy below @root, or
1196 * @root itself, or %NULL after a full round-trip.
1198 * Caller must pass the return value in @prev on subsequent
1199 * invocations for reference counting, or use mem_cgroup_iter_break()
1200 * to cancel a hierarchy walk before the round-trip is complete.
1202 * Reclaimers can specify a zone and a priority level in @reclaim to
1203 * divide up the memcgs in the hierarchy among all concurrent
1204 * reclaimers operating on the same zone and priority.
1206 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1207 struct mem_cgroup *prev,
1208 struct mem_cgroup_reclaim_cookie *reclaim)
1210 struct mem_cgroup *memcg = NULL;
1211 struct mem_cgroup *last_visited = NULL;
1213 if (mem_cgroup_disabled())
1217 root = root_mem_cgroup;
1219 if (prev && !reclaim)
1220 last_visited = prev;
1222 if (!root->use_hierarchy && root != root_mem_cgroup) {
1230 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1231 int uninitialized_var(seq);
1234 int nid = zone_to_nid(reclaim->zone);
1235 int zid = zone_idx(reclaim->zone);
1236 struct mem_cgroup_per_zone *mz;
1238 mz = mem_cgroup_zoneinfo(root, nid, zid);
1239 iter = &mz->reclaim_iter[reclaim->priority];
1240 if (prev && reclaim->generation != iter->generation) {
1241 iter->last_visited = NULL;
1245 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1248 memcg = __mem_cgroup_iter_next(root, last_visited);
1251 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1256 else if (!prev && memcg)
1257 reclaim->generation = iter->generation;
1266 if (prev && prev != root)
1267 css_put(&prev->css);
1273 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1274 * @root: hierarchy root
1275 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1277 void mem_cgroup_iter_break(struct mem_cgroup *root,
1278 struct mem_cgroup *prev)
1281 root = root_mem_cgroup;
1282 if (prev && prev != root)
1283 css_put(&prev->css);
1287 * Iteration constructs for visiting all cgroups (under a tree). If
1288 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1289 * be used for reference counting.
1291 #define for_each_mem_cgroup_tree(iter, root) \
1292 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1294 iter = mem_cgroup_iter(root, iter, NULL))
1296 #define for_each_mem_cgroup(iter) \
1297 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1299 iter = mem_cgroup_iter(NULL, iter, NULL))
1301 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1303 struct mem_cgroup *memcg;
1306 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1307 if (unlikely(!memcg))
1312 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1315 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1323 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1326 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1327 * @zone: zone of the wanted lruvec
1328 * @memcg: memcg of the wanted lruvec
1330 * Returns the lru list vector holding pages for the given @zone and
1331 * @mem. This can be the global zone lruvec, if the memory controller
1334 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1335 struct mem_cgroup *memcg)
1337 struct mem_cgroup_per_zone *mz;
1338 struct lruvec *lruvec;
1340 if (mem_cgroup_disabled()) {
1341 lruvec = &zone->lruvec;
1345 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1346 lruvec = &mz->lruvec;
1349 * Since a node can be onlined after the mem_cgroup was created,
1350 * we have to be prepared to initialize lruvec->zone here;
1351 * and if offlined then reonlined, we need to reinitialize it.
1353 if (unlikely(lruvec->zone != zone))
1354 lruvec->zone = zone;
1359 * Following LRU functions are allowed to be used without PCG_LOCK.
1360 * Operations are called by routine of global LRU independently from memcg.
1361 * What we have to take care of here is validness of pc->mem_cgroup.
1363 * Changes to pc->mem_cgroup happens when
1366 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1367 * It is added to LRU before charge.
1368 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1369 * When moving account, the page is not on LRU. It's isolated.
1373 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1375 * @zone: zone of the page
1377 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1379 struct mem_cgroup_per_zone *mz;
1380 struct mem_cgroup *memcg;
1381 struct page_cgroup *pc;
1382 struct lruvec *lruvec;
1384 if (mem_cgroup_disabled()) {
1385 lruvec = &zone->lruvec;
1389 pc = lookup_page_cgroup(page);
1390 memcg = pc->mem_cgroup;
1393 * Surreptitiously switch any uncharged offlist page to root:
1394 * an uncharged page off lru does nothing to secure
1395 * its former mem_cgroup from sudden removal.
1397 * Our caller holds lru_lock, and PageCgroupUsed is updated
1398 * under page_cgroup lock: between them, they make all uses
1399 * of pc->mem_cgroup safe.
1401 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1402 pc->mem_cgroup = memcg = root_mem_cgroup;
1404 mz = page_cgroup_zoneinfo(memcg, page);
1405 lruvec = &mz->lruvec;
1408 * Since a node can be onlined after the mem_cgroup was created,
1409 * we have to be prepared to initialize lruvec->zone here;
1410 * and if offlined then reonlined, we need to reinitialize it.
1412 if (unlikely(lruvec->zone != zone))
1413 lruvec->zone = zone;
1418 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1419 * @lruvec: mem_cgroup per zone lru vector
1420 * @lru: index of lru list the page is sitting on
1421 * @nr_pages: positive when adding or negative when removing
1423 * This function must be called when a page is added to or removed from an
1426 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1429 struct mem_cgroup_per_zone *mz;
1430 unsigned long *lru_size;
1432 if (mem_cgroup_disabled())
1435 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1436 lru_size = mz->lru_size + lru;
1437 *lru_size += nr_pages;
1438 VM_BUG_ON((long)(*lru_size) < 0);
1442 * Checks whether given mem is same or in the root_mem_cgroup's
1445 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1446 struct mem_cgroup *memcg)
1448 if (root_memcg == memcg)
1450 if (!root_memcg->use_hierarchy || !memcg)
1452 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1455 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1456 struct mem_cgroup *memcg)
1461 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1466 bool task_in_mem_cgroup(struct task_struct *task,
1467 const struct mem_cgroup *memcg)
1469 struct mem_cgroup *curr = NULL;
1470 struct task_struct *p;
1473 p = find_lock_task_mm(task);
1475 curr = get_mem_cgroup_from_mm(p->mm);
1479 * All threads may have already detached their mm's, but the oom
1480 * killer still needs to detect if they have already been oom
1481 * killed to prevent needlessly killing additional tasks.
1484 curr = mem_cgroup_from_task(task);
1486 css_get(&curr->css);
1490 * We should check use_hierarchy of "memcg" not "curr". Because checking
1491 * use_hierarchy of "curr" here make this function true if hierarchy is
1492 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1493 * hierarchy(even if use_hierarchy is disabled in "memcg").
1495 ret = mem_cgroup_same_or_subtree(memcg, curr);
1496 css_put(&curr->css);
1500 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1502 unsigned long inactive_ratio;
1503 unsigned long inactive;
1504 unsigned long active;
1507 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1508 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1510 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1512 inactive_ratio = int_sqrt(10 * gb);
1516 return inactive * inactive_ratio < active;
1519 #define mem_cgroup_from_res_counter(counter, member) \
1520 container_of(counter, struct mem_cgroup, member)
1523 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1524 * @memcg: the memory cgroup
1526 * Returns the maximum amount of memory @mem can be charged with, in
1529 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1531 unsigned long long margin;
1533 margin = res_counter_margin(&memcg->res);
1534 if (do_swap_account)
1535 margin = min(margin, res_counter_margin(&memcg->memsw));
1536 return margin >> PAGE_SHIFT;
1539 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1542 if (!memcg->css.parent)
1543 return vm_swappiness;
1545 return memcg->swappiness;
1549 * memcg->moving_account is used for checking possibility that some thread is
1550 * calling move_account(). When a thread on CPU-A starts moving pages under
1551 * a memcg, other threads should check memcg->moving_account under
1552 * rcu_read_lock(), like this:
1556 * memcg->moving_account+1 if (memcg->mocing_account)
1558 * synchronize_rcu() update something.
1563 /* for quick checking without looking up memcg */
1564 atomic_t memcg_moving __read_mostly;
1566 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1568 atomic_inc(&memcg_moving);
1569 atomic_inc(&memcg->moving_account);
1573 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1576 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1577 * We check NULL in callee rather than caller.
1580 atomic_dec(&memcg_moving);
1581 atomic_dec(&memcg->moving_account);
1586 * A routine for checking "mem" is under move_account() or not.
1588 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1589 * moving cgroups. This is for waiting at high-memory pressure
1592 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1594 struct mem_cgroup *from;
1595 struct mem_cgroup *to;
1598 * Unlike task_move routines, we access mc.to, mc.from not under
1599 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1601 spin_lock(&mc.lock);
1607 ret = mem_cgroup_same_or_subtree(memcg, from)
1608 || mem_cgroup_same_or_subtree(memcg, to);
1610 spin_unlock(&mc.lock);
1614 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1616 if (mc.moving_task && current != mc.moving_task) {
1617 if (mem_cgroup_under_move(memcg)) {
1619 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1620 /* moving charge context might have finished. */
1623 finish_wait(&mc.waitq, &wait);
1631 * Take this lock when
1632 * - a code tries to modify page's memcg while it's USED.
1633 * - a code tries to modify page state accounting in a memcg.
1635 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1636 unsigned long *flags)
1638 spin_lock_irqsave(&memcg->move_lock, *flags);
1641 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1642 unsigned long *flags)
1644 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1647 #define K(x) ((x) << (PAGE_SHIFT-10))
1649 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1650 * @memcg: The memory cgroup that went over limit
1651 * @p: Task that is going to be killed
1653 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1656 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1658 /* oom_info_lock ensures that parallel ooms do not interleave */
1659 static DEFINE_MUTEX(oom_info_lock);
1660 struct mem_cgroup *iter;
1666 mutex_lock(&oom_info_lock);
1669 pr_info("Task in ");
1670 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1671 pr_info(" killed as a result of limit of ");
1672 pr_cont_cgroup_path(memcg->css.cgroup);
1677 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1678 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1679 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1680 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1681 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1682 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1683 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1684 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1685 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1686 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1687 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1688 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1690 for_each_mem_cgroup_tree(iter, memcg) {
1691 pr_info("Memory cgroup stats for ");
1692 pr_cont_cgroup_path(iter->css.cgroup);
1695 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1696 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1698 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1699 K(mem_cgroup_read_stat(iter, i)));
1702 for (i = 0; i < NR_LRU_LISTS; i++)
1703 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1704 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1708 mutex_unlock(&oom_info_lock);
1712 * This function returns the number of memcg under hierarchy tree. Returns
1713 * 1(self count) if no children.
1715 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1718 struct mem_cgroup *iter;
1720 for_each_mem_cgroup_tree(iter, memcg)
1726 * Return the memory (and swap, if configured) limit for a memcg.
1728 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1732 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1735 * Do not consider swap space if we cannot swap due to swappiness
1737 if (mem_cgroup_swappiness(memcg)) {
1740 limit += total_swap_pages << PAGE_SHIFT;
1741 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1744 * If memsw is finite and limits the amount of swap space
1745 * available to this memcg, return that limit.
1747 limit = min(limit, memsw);
1753 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1756 struct mem_cgroup *iter;
1757 unsigned long chosen_points = 0;
1758 unsigned long totalpages;
1759 unsigned int points = 0;
1760 struct task_struct *chosen = NULL;
1763 * If current has a pending SIGKILL or is exiting, then automatically
1764 * select it. The goal is to allow it to allocate so that it may
1765 * quickly exit and free its memory.
1767 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1768 set_thread_flag(TIF_MEMDIE);
1772 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1773 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1774 for_each_mem_cgroup_tree(iter, memcg) {
1775 struct css_task_iter it;
1776 struct task_struct *task;
1778 css_task_iter_start(&iter->css, &it);
1779 while ((task = css_task_iter_next(&it))) {
1780 switch (oom_scan_process_thread(task, totalpages, NULL,
1782 case OOM_SCAN_SELECT:
1784 put_task_struct(chosen);
1786 chosen_points = ULONG_MAX;
1787 get_task_struct(chosen);
1789 case OOM_SCAN_CONTINUE:
1791 case OOM_SCAN_ABORT:
1792 css_task_iter_end(&it);
1793 mem_cgroup_iter_break(memcg, iter);
1795 put_task_struct(chosen);
1800 points = oom_badness(task, memcg, NULL, totalpages);
1801 if (!points || points < chosen_points)
1803 /* Prefer thread group leaders for display purposes */
1804 if (points == chosen_points &&
1805 thread_group_leader(chosen))
1809 put_task_struct(chosen);
1811 chosen_points = points;
1812 get_task_struct(chosen);
1814 css_task_iter_end(&it);
1819 points = chosen_points * 1000 / totalpages;
1820 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1821 NULL, "Memory cgroup out of memory");
1824 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1826 unsigned long flags)
1828 unsigned long total = 0;
1829 bool noswap = false;
1832 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1834 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1837 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1839 drain_all_stock_async(memcg);
1840 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1842 * Allow limit shrinkers, which are triggered directly
1843 * by userspace, to catch signals and stop reclaim
1844 * after minimal progress, regardless of the margin.
1846 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1848 if (mem_cgroup_margin(memcg))
1851 * If nothing was reclaimed after two attempts, there
1852 * may be no reclaimable pages in this hierarchy.
1861 * test_mem_cgroup_node_reclaimable
1862 * @memcg: the target memcg
1863 * @nid: the node ID to be checked.
1864 * @noswap : specify true here if the user wants flle only information.
1866 * This function returns whether the specified memcg contains any
1867 * reclaimable pages on a node. Returns true if there are any reclaimable
1868 * pages in the node.
1870 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1871 int nid, bool noswap)
1873 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1875 if (noswap || !total_swap_pages)
1877 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1882 #if MAX_NUMNODES > 1
1885 * Always updating the nodemask is not very good - even if we have an empty
1886 * list or the wrong list here, we can start from some node and traverse all
1887 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1890 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1894 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1895 * pagein/pageout changes since the last update.
1897 if (!atomic_read(&memcg->numainfo_events))
1899 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1902 /* make a nodemask where this memcg uses memory from */
1903 memcg->scan_nodes = node_states[N_MEMORY];
1905 for_each_node_mask(nid, node_states[N_MEMORY]) {
1907 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1908 node_clear(nid, memcg->scan_nodes);
1911 atomic_set(&memcg->numainfo_events, 0);
1912 atomic_set(&memcg->numainfo_updating, 0);
1916 * Selecting a node where we start reclaim from. Because what we need is just
1917 * reducing usage counter, start from anywhere is O,K. Considering
1918 * memory reclaim from current node, there are pros. and cons.
1920 * Freeing memory from current node means freeing memory from a node which
1921 * we'll use or we've used. So, it may make LRU bad. And if several threads
1922 * hit limits, it will see a contention on a node. But freeing from remote
1923 * node means more costs for memory reclaim because of memory latency.
1925 * Now, we use round-robin. Better algorithm is welcomed.
1927 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1931 mem_cgroup_may_update_nodemask(memcg);
1932 node = memcg->last_scanned_node;
1934 node = next_node(node, memcg->scan_nodes);
1935 if (node == MAX_NUMNODES)
1936 node = first_node(memcg->scan_nodes);
1938 * We call this when we hit limit, not when pages are added to LRU.
1939 * No LRU may hold pages because all pages are UNEVICTABLE or
1940 * memcg is too small and all pages are not on LRU. In that case,
1941 * we use curret node.
1943 if (unlikely(node == MAX_NUMNODES))
1944 node = numa_node_id();
1946 memcg->last_scanned_node = node;
1951 * Check all nodes whether it contains reclaimable pages or not.
1952 * For quick scan, we make use of scan_nodes. This will allow us to skip
1953 * unused nodes. But scan_nodes is lazily updated and may not cotain
1954 * enough new information. We need to do double check.
1956 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1961 * quick check...making use of scan_node.
1962 * We can skip unused nodes.
1964 if (!nodes_empty(memcg->scan_nodes)) {
1965 for (nid = first_node(memcg->scan_nodes);
1967 nid = next_node(nid, memcg->scan_nodes)) {
1969 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1974 * Check rest of nodes.
1976 for_each_node_state(nid, N_MEMORY) {
1977 if (node_isset(nid, memcg->scan_nodes))
1979 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1986 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1991 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1993 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1997 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2000 unsigned long *total_scanned)
2002 struct mem_cgroup *victim = NULL;
2005 unsigned long excess;
2006 unsigned long nr_scanned;
2007 struct mem_cgroup_reclaim_cookie reclaim = {
2012 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2015 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2020 * If we have not been able to reclaim
2021 * anything, it might because there are
2022 * no reclaimable pages under this hierarchy
2027 * We want to do more targeted reclaim.
2028 * excess >> 2 is not to excessive so as to
2029 * reclaim too much, nor too less that we keep
2030 * coming back to reclaim from this cgroup
2032 if (total >= (excess >> 2) ||
2033 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2038 if (!mem_cgroup_reclaimable(victim, false))
2040 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2042 *total_scanned += nr_scanned;
2043 if (!res_counter_soft_limit_excess(&root_memcg->res))
2046 mem_cgroup_iter_break(root_memcg, victim);
2050 #ifdef CONFIG_LOCKDEP
2051 static struct lockdep_map memcg_oom_lock_dep_map = {
2052 .name = "memcg_oom_lock",
2056 static DEFINE_SPINLOCK(memcg_oom_lock);
2059 * Check OOM-Killer is already running under our hierarchy.
2060 * If someone is running, return false.
2062 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2064 struct mem_cgroup *iter, *failed = NULL;
2066 spin_lock(&memcg_oom_lock);
2068 for_each_mem_cgroup_tree(iter, memcg) {
2069 if (iter->oom_lock) {
2071 * this subtree of our hierarchy is already locked
2072 * so we cannot give a lock.
2075 mem_cgroup_iter_break(memcg, iter);
2078 iter->oom_lock = true;
2083 * OK, we failed to lock the whole subtree so we have
2084 * to clean up what we set up to the failing subtree
2086 for_each_mem_cgroup_tree(iter, memcg) {
2087 if (iter == failed) {
2088 mem_cgroup_iter_break(memcg, iter);
2091 iter->oom_lock = false;
2094 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2096 spin_unlock(&memcg_oom_lock);
2101 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2103 struct mem_cgroup *iter;
2105 spin_lock(&memcg_oom_lock);
2106 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2107 for_each_mem_cgroup_tree(iter, memcg)
2108 iter->oom_lock = false;
2109 spin_unlock(&memcg_oom_lock);
2112 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2114 struct mem_cgroup *iter;
2116 for_each_mem_cgroup_tree(iter, memcg)
2117 atomic_inc(&iter->under_oom);
2120 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2122 struct mem_cgroup *iter;
2125 * When a new child is created while the hierarchy is under oom,
2126 * mem_cgroup_oom_lock() may not be called. We have to use
2127 * atomic_add_unless() here.
2129 for_each_mem_cgroup_tree(iter, memcg)
2130 atomic_add_unless(&iter->under_oom, -1, 0);
2133 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2135 struct oom_wait_info {
2136 struct mem_cgroup *memcg;
2140 static int memcg_oom_wake_function(wait_queue_t *wait,
2141 unsigned mode, int sync, void *arg)
2143 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2144 struct mem_cgroup *oom_wait_memcg;
2145 struct oom_wait_info *oom_wait_info;
2147 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2148 oom_wait_memcg = oom_wait_info->memcg;
2151 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2152 * Then we can use css_is_ancestor without taking care of RCU.
2154 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2155 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2157 return autoremove_wake_function(wait, mode, sync, arg);
2160 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2162 atomic_inc(&memcg->oom_wakeups);
2163 /* for filtering, pass "memcg" as argument. */
2164 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2167 static void memcg_oom_recover(struct mem_cgroup *memcg)
2169 if (memcg && atomic_read(&memcg->under_oom))
2170 memcg_wakeup_oom(memcg);
2173 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2175 if (!current->memcg_oom.may_oom)
2178 * We are in the middle of the charge context here, so we
2179 * don't want to block when potentially sitting on a callstack
2180 * that holds all kinds of filesystem and mm locks.
2182 * Also, the caller may handle a failed allocation gracefully
2183 * (like optional page cache readahead) and so an OOM killer
2184 * invocation might not even be necessary.
2186 * That's why we don't do anything here except remember the
2187 * OOM context and then deal with it at the end of the page
2188 * fault when the stack is unwound, the locks are released,
2189 * and when we know whether the fault was overall successful.
2191 css_get(&memcg->css);
2192 current->memcg_oom.memcg = memcg;
2193 current->memcg_oom.gfp_mask = mask;
2194 current->memcg_oom.order = order;
2198 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2199 * @handle: actually kill/wait or just clean up the OOM state
2201 * This has to be called at the end of a page fault if the memcg OOM
2202 * handler was enabled.
2204 * Memcg supports userspace OOM handling where failed allocations must
2205 * sleep on a waitqueue until the userspace task resolves the
2206 * situation. Sleeping directly in the charge context with all kinds
2207 * of locks held is not a good idea, instead we remember an OOM state
2208 * in the task and mem_cgroup_oom_synchronize() has to be called at
2209 * the end of the page fault to complete the OOM handling.
2211 * Returns %true if an ongoing memcg OOM situation was detected and
2212 * completed, %false otherwise.
2214 bool mem_cgroup_oom_synchronize(bool handle)
2216 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2217 struct oom_wait_info owait;
2220 /* OOM is global, do not handle */
2227 owait.memcg = memcg;
2228 owait.wait.flags = 0;
2229 owait.wait.func = memcg_oom_wake_function;
2230 owait.wait.private = current;
2231 INIT_LIST_HEAD(&owait.wait.task_list);
2233 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2234 mem_cgroup_mark_under_oom(memcg);
2236 locked = mem_cgroup_oom_trylock(memcg);
2239 mem_cgroup_oom_notify(memcg);
2241 if (locked && !memcg->oom_kill_disable) {
2242 mem_cgroup_unmark_under_oom(memcg);
2243 finish_wait(&memcg_oom_waitq, &owait.wait);
2244 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2245 current->memcg_oom.order);
2248 mem_cgroup_unmark_under_oom(memcg);
2249 finish_wait(&memcg_oom_waitq, &owait.wait);
2253 mem_cgroup_oom_unlock(memcg);
2255 * There is no guarantee that an OOM-lock contender
2256 * sees the wakeups triggered by the OOM kill
2257 * uncharges. Wake any sleepers explicitely.
2259 memcg_oom_recover(memcg);
2262 current->memcg_oom.memcg = NULL;
2263 css_put(&memcg->css);
2268 * Used to update mapped file or writeback or other statistics.
2270 * Notes: Race condition
2272 * We usually use lock_page_cgroup() for accessing page_cgroup member but
2273 * it tends to be costly. But considering some conditions, we doesn't need
2274 * to do so _always_.
2276 * Considering "charge", lock_page_cgroup() is not required because all
2277 * file-stat operations happen after a page is attached to radix-tree. There
2278 * are no race with "charge".
2280 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2281 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2282 * if there are race with "uncharge". Statistics itself is properly handled
2285 * Considering "move", this is an only case we see a race. To make the race
2286 * small, we check memcg->moving_account and detect there are possibility
2287 * of race or not. If there is, we take a lock.
2290 void __mem_cgroup_begin_update_page_stat(struct page *page,
2291 bool *locked, unsigned long *flags)
2293 struct mem_cgroup *memcg;
2294 struct page_cgroup *pc;
2296 pc = lookup_page_cgroup(page);
2298 memcg = pc->mem_cgroup;
2299 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2302 * If this memory cgroup is not under account moving, we don't
2303 * need to take move_lock_mem_cgroup(). Because we already hold
2304 * rcu_read_lock(), any calls to move_account will be delayed until
2305 * rcu_read_unlock().
2307 VM_BUG_ON(!rcu_read_lock_held());
2308 if (atomic_read(&memcg->moving_account) <= 0)
2311 move_lock_mem_cgroup(memcg, flags);
2312 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2313 move_unlock_mem_cgroup(memcg, flags);
2319 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2321 struct page_cgroup *pc = lookup_page_cgroup(page);
2324 * It's guaranteed that pc->mem_cgroup never changes while
2325 * lock is held because a routine modifies pc->mem_cgroup
2326 * should take move_lock_mem_cgroup().
2328 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2331 void mem_cgroup_update_page_stat(struct page *page,
2332 enum mem_cgroup_stat_index idx, int val)
2334 struct mem_cgroup *memcg;
2335 struct page_cgroup *pc = lookup_page_cgroup(page);
2336 unsigned long uninitialized_var(flags);
2338 if (mem_cgroup_disabled())
2341 VM_BUG_ON(!rcu_read_lock_held());
2342 memcg = pc->mem_cgroup;
2343 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2346 this_cpu_add(memcg->stat->count[idx], val);
2350 * size of first charge trial. "32" comes from vmscan.c's magic value.
2351 * TODO: maybe necessary to use big numbers in big irons.
2353 #define CHARGE_BATCH 32U
2354 struct memcg_stock_pcp {
2355 struct mem_cgroup *cached; /* this never be root cgroup */
2356 unsigned int nr_pages;
2357 struct work_struct work;
2358 unsigned long flags;
2359 #define FLUSHING_CACHED_CHARGE 0
2361 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2362 static DEFINE_MUTEX(percpu_charge_mutex);
2365 * consume_stock: Try to consume stocked charge on this cpu.
2366 * @memcg: memcg to consume from.
2367 * @nr_pages: how many pages to charge.
2369 * The charges will only happen if @memcg matches the current cpu's memcg
2370 * stock, and at least @nr_pages are available in that stock. Failure to
2371 * service an allocation will refill the stock.
2373 * returns true if successful, false otherwise.
2375 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2377 struct memcg_stock_pcp *stock;
2380 if (nr_pages > CHARGE_BATCH)
2383 stock = &get_cpu_var(memcg_stock);
2384 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2385 stock->nr_pages -= nr_pages;
2386 else /* need to call res_counter_charge */
2388 put_cpu_var(memcg_stock);
2393 * Returns stocks cached in percpu to res_counter and reset cached information.
2395 static void drain_stock(struct memcg_stock_pcp *stock)
2397 struct mem_cgroup *old = stock->cached;
2399 if (stock->nr_pages) {
2400 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2402 res_counter_uncharge(&old->res, bytes);
2403 if (do_swap_account)
2404 res_counter_uncharge(&old->memsw, bytes);
2405 stock->nr_pages = 0;
2407 stock->cached = NULL;
2411 * This must be called under preempt disabled or must be called by
2412 * a thread which is pinned to local cpu.
2414 static void drain_local_stock(struct work_struct *dummy)
2416 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2418 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2421 static void __init memcg_stock_init(void)
2425 for_each_possible_cpu(cpu) {
2426 struct memcg_stock_pcp *stock =
2427 &per_cpu(memcg_stock, cpu);
2428 INIT_WORK(&stock->work, drain_local_stock);
2433 * Cache charges(val) which is from res_counter, to local per_cpu area.
2434 * This will be consumed by consume_stock() function, later.
2436 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2438 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2440 if (stock->cached != memcg) { /* reset if necessary */
2442 stock->cached = memcg;
2444 stock->nr_pages += nr_pages;
2445 put_cpu_var(memcg_stock);
2449 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2450 * of the hierarchy under it. sync flag says whether we should block
2451 * until the work is done.
2453 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2457 /* Notify other cpus that system-wide "drain" is running */
2460 for_each_online_cpu(cpu) {
2461 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2462 struct mem_cgroup *memcg;
2464 memcg = stock->cached;
2465 if (!memcg || !stock->nr_pages)
2467 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2469 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2471 drain_local_stock(&stock->work);
2473 schedule_work_on(cpu, &stock->work);
2481 for_each_online_cpu(cpu) {
2482 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2483 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2484 flush_work(&stock->work);
2491 * Tries to drain stocked charges in other cpus. This function is asynchronous
2492 * and just put a work per cpu for draining localy on each cpu. Caller can
2493 * expects some charges will be back to res_counter later but cannot wait for
2496 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2499 * If someone calls draining, avoid adding more kworker runs.
2501 if (!mutex_trylock(&percpu_charge_mutex))
2503 drain_all_stock(root_memcg, false);
2504 mutex_unlock(&percpu_charge_mutex);
2507 /* This is a synchronous drain interface. */
2508 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2510 /* called when force_empty is called */
2511 mutex_lock(&percpu_charge_mutex);
2512 drain_all_stock(root_memcg, true);
2513 mutex_unlock(&percpu_charge_mutex);
2517 * This function drains percpu counter value from DEAD cpu and
2518 * move it to local cpu. Note that this function can be preempted.
2520 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2524 spin_lock(&memcg->pcp_counter_lock);
2525 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2526 long x = per_cpu(memcg->stat->count[i], cpu);
2528 per_cpu(memcg->stat->count[i], cpu) = 0;
2529 memcg->nocpu_base.count[i] += x;
2531 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2532 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2534 per_cpu(memcg->stat->events[i], cpu) = 0;
2535 memcg->nocpu_base.events[i] += x;
2537 spin_unlock(&memcg->pcp_counter_lock);
2540 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2541 unsigned long action,
2544 int cpu = (unsigned long)hcpu;
2545 struct memcg_stock_pcp *stock;
2546 struct mem_cgroup *iter;
2548 if (action == CPU_ONLINE)
2551 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2554 for_each_mem_cgroup(iter)
2555 mem_cgroup_drain_pcp_counter(iter, cpu);
2557 stock = &per_cpu(memcg_stock, cpu);
2563 /* See mem_cgroup_try_charge() for details */
2565 CHARGE_OK, /* success */
2566 CHARGE_RETRY, /* need to retry but retry is not bad */
2567 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2568 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2571 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2572 unsigned int nr_pages, unsigned int min_pages,
2575 unsigned long csize = nr_pages * PAGE_SIZE;
2576 struct mem_cgroup *mem_over_limit;
2577 struct res_counter *fail_res;
2578 unsigned long flags = 0;
2581 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2584 if (!do_swap_account)
2586 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2590 res_counter_uncharge(&memcg->res, csize);
2591 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2592 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2594 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2596 * Never reclaim on behalf of optional batching, retry with a
2597 * single page instead.
2599 if (nr_pages > min_pages)
2600 return CHARGE_RETRY;
2602 if (!(gfp_mask & __GFP_WAIT))
2603 return CHARGE_WOULDBLOCK;
2605 if (gfp_mask & __GFP_NORETRY)
2606 return CHARGE_NOMEM;
2608 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2609 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2610 return CHARGE_RETRY;
2612 * Even though the limit is exceeded at this point, reclaim
2613 * may have been able to free some pages. Retry the charge
2614 * before killing the task.
2616 * Only for regular pages, though: huge pages are rather
2617 * unlikely to succeed so close to the limit, and we fall back
2618 * to regular pages anyway in case of failure.
2620 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2621 return CHARGE_RETRY;
2624 * At task move, charge accounts can be doubly counted. So, it's
2625 * better to wait until the end of task_move if something is going on.
2627 if (mem_cgroup_wait_acct_move(mem_over_limit))
2628 return CHARGE_RETRY;
2631 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2633 return CHARGE_NOMEM;
2637 * mem_cgroup_try_charge - try charging a memcg
2638 * @memcg: memcg to charge
2639 * @nr_pages: number of pages to charge
2640 * @oom: trigger OOM if reclaim fails
2642 * Returns 0 if @memcg was charged successfully, -EINTR if the charge
2643 * was bypassed to root_mem_cgroup, and -ENOMEM if the charge failed.
2645 static int mem_cgroup_try_charge(struct mem_cgroup *memcg,
2647 unsigned int nr_pages,
2650 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2651 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2654 if (mem_cgroup_is_root(memcg))
2657 * Unlike in global OOM situations, memcg is not in a physical
2658 * memory shortage. Allow dying and OOM-killed tasks to
2659 * bypass the last charges so that they can exit quickly and
2660 * free their memory.
2662 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2663 fatal_signal_pending(current) ||
2664 current->flags & PF_EXITING))
2667 if (unlikely(task_in_memcg_oom(current)))
2670 if (gfp_mask & __GFP_NOFAIL)
2673 if (consume_stock(memcg, nr_pages))
2677 bool invoke_oom = oom && !nr_oom_retries;
2679 /* If killed, bypass charge */
2680 if (fatal_signal_pending(current))
2683 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2684 nr_pages, invoke_oom);
2688 case CHARGE_RETRY: /* not in OOM situation but retry */
2691 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2693 case CHARGE_NOMEM: /* OOM routine works */
2694 if (!oom || invoke_oom)
2699 } while (ret != CHARGE_OK);
2701 if (batch > nr_pages)
2702 refill_stock(memcg, batch - nr_pages);
2706 if (!(gfp_mask & __GFP_NOFAIL))
2713 * mem_cgroup_try_charge_mm - try charging a mm
2714 * @mm: mm_struct to charge
2715 * @nr_pages: number of pages to charge
2716 * @oom: trigger OOM if reclaim fails
2718 * Returns the charged mem_cgroup associated with the given mm_struct or
2719 * NULL the charge failed.
2721 static struct mem_cgroup *mem_cgroup_try_charge_mm(struct mm_struct *mm,
2723 unsigned int nr_pages,
2727 struct mem_cgroup *memcg;
2730 memcg = get_mem_cgroup_from_mm(mm);
2731 ret = mem_cgroup_try_charge(memcg, gfp_mask, nr_pages, oom);
2732 css_put(&memcg->css);
2734 memcg = root_mem_cgroup;
2742 * Somemtimes we have to undo a charge we got by try_charge().
2743 * This function is for that and do uncharge, put css's refcnt.
2744 * gotten by try_charge().
2746 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2747 unsigned int nr_pages)
2749 if (!mem_cgroup_is_root(memcg)) {
2750 unsigned long bytes = nr_pages * PAGE_SIZE;
2752 res_counter_uncharge(&memcg->res, bytes);
2753 if (do_swap_account)
2754 res_counter_uncharge(&memcg->memsw, bytes);
2759 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2760 * This is useful when moving usage to parent cgroup.
2762 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2763 unsigned int nr_pages)
2765 unsigned long bytes = nr_pages * PAGE_SIZE;
2767 if (mem_cgroup_is_root(memcg))
2770 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2771 if (do_swap_account)
2772 res_counter_uncharge_until(&memcg->memsw,
2773 memcg->memsw.parent, bytes);
2777 * A helper function to get mem_cgroup from ID. must be called under
2778 * rcu_read_lock(). The caller is responsible for calling
2779 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2780 * refcnt from swap can be called against removed memcg.)
2782 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2784 /* ID 0 is unused ID */
2787 return mem_cgroup_from_id(id);
2790 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2792 struct mem_cgroup *memcg = NULL;
2793 struct page_cgroup *pc;
2797 VM_BUG_ON_PAGE(!PageLocked(page), page);
2799 pc = lookup_page_cgroup(page);
2800 lock_page_cgroup(pc);
2801 if (PageCgroupUsed(pc)) {
2802 memcg = pc->mem_cgroup;
2803 if (memcg && !css_tryget_online(&memcg->css))
2805 } else if (PageSwapCache(page)) {
2806 ent.val = page_private(page);
2807 id = lookup_swap_cgroup_id(ent);
2809 memcg = mem_cgroup_lookup(id);
2810 if (memcg && !css_tryget_online(&memcg->css))
2814 unlock_page_cgroup(pc);
2818 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2820 unsigned int nr_pages,
2821 enum charge_type ctype,
2824 struct page_cgroup *pc = lookup_page_cgroup(page);
2825 struct zone *uninitialized_var(zone);
2826 struct lruvec *lruvec;
2827 bool was_on_lru = false;
2830 lock_page_cgroup(pc);
2831 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2833 * we don't need page_cgroup_lock about tail pages, becase they are not
2834 * accessed by any other context at this point.
2838 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2839 * may already be on some other mem_cgroup's LRU. Take care of it.
2842 zone = page_zone(page);
2843 spin_lock_irq(&zone->lru_lock);
2844 if (PageLRU(page)) {
2845 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2847 del_page_from_lru_list(page, lruvec, page_lru(page));
2852 pc->mem_cgroup = memcg;
2854 * We access a page_cgroup asynchronously without lock_page_cgroup().
2855 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2856 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2857 * before USED bit, we need memory barrier here.
2858 * See mem_cgroup_add_lru_list(), etc.
2861 SetPageCgroupUsed(pc);
2865 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2866 VM_BUG_ON_PAGE(PageLRU(page), page);
2868 add_page_to_lru_list(page, lruvec, page_lru(page));
2870 spin_unlock_irq(&zone->lru_lock);
2873 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2878 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2879 unlock_page_cgroup(pc);
2882 * "charge_statistics" updated event counter. Then, check it.
2883 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2884 * if they exceeds softlimit.
2886 memcg_check_events(memcg, page);
2889 static DEFINE_MUTEX(set_limit_mutex);
2891 #ifdef CONFIG_MEMCG_KMEM
2893 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2894 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2896 static DEFINE_MUTEX(memcg_slab_mutex);
2898 static DEFINE_MUTEX(activate_kmem_mutex);
2900 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2902 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2903 memcg_kmem_is_active(memcg);
2907 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2908 * in the memcg_cache_params struct.
2910 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2912 struct kmem_cache *cachep;
2914 VM_BUG_ON(p->is_root_cache);
2915 cachep = p->root_cache;
2916 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2919 #ifdef CONFIG_SLABINFO
2920 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2922 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2923 struct memcg_cache_params *params;
2925 if (!memcg_can_account_kmem(memcg))
2928 print_slabinfo_header(m);
2930 mutex_lock(&memcg_slab_mutex);
2931 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2932 cache_show(memcg_params_to_cache(params), m);
2933 mutex_unlock(&memcg_slab_mutex);
2939 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2941 struct res_counter *fail_res;
2944 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2948 ret = mem_cgroup_try_charge(memcg, gfp, size >> PAGE_SHIFT,
2949 oom_gfp_allowed(gfp));
2950 if (ret == -EINTR) {
2952 * mem_cgroup_try_charge() chosed to bypass to root due to
2953 * OOM kill or fatal signal. Since our only options are to
2954 * either fail the allocation or charge it to this cgroup, do
2955 * it as a temporary condition. But we can't fail. From a
2956 * kmem/slab perspective, the cache has already been selected,
2957 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2960 * This condition will only trigger if the task entered
2961 * memcg_charge_kmem in a sane state, but was OOM-killed during
2962 * mem_cgroup_try_charge() above. Tasks that were already
2963 * dying when the allocation triggers should have been already
2964 * directed to the root cgroup in memcontrol.h
2966 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2967 if (do_swap_account)
2968 res_counter_charge_nofail(&memcg->memsw, size,
2972 res_counter_uncharge(&memcg->kmem, size);
2977 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2979 res_counter_uncharge(&memcg->res, size);
2980 if (do_swap_account)
2981 res_counter_uncharge(&memcg->memsw, size);
2984 if (res_counter_uncharge(&memcg->kmem, size))
2988 * Releases a reference taken in kmem_cgroup_css_offline in case
2989 * this last uncharge is racing with the offlining code or it is
2990 * outliving the memcg existence.
2992 * The memory barrier imposed by test&clear is paired with the
2993 * explicit one in memcg_kmem_mark_dead().
2995 if (memcg_kmem_test_and_clear_dead(memcg))
2996 css_put(&memcg->css);
3000 * helper for acessing a memcg's index. It will be used as an index in the
3001 * child cache array in kmem_cache, and also to derive its name. This function
3002 * will return -1 when this is not a kmem-limited memcg.
3004 int memcg_cache_id(struct mem_cgroup *memcg)
3006 return memcg ? memcg->kmemcg_id : -1;
3009 static size_t memcg_caches_array_size(int num_groups)
3012 if (num_groups <= 0)
3015 size = 2 * num_groups;
3016 if (size < MEMCG_CACHES_MIN_SIZE)
3017 size = MEMCG_CACHES_MIN_SIZE;
3018 else if (size > MEMCG_CACHES_MAX_SIZE)
3019 size = MEMCG_CACHES_MAX_SIZE;
3025 * We should update the current array size iff all caches updates succeed. This
3026 * can only be done from the slab side. The slab mutex needs to be held when
3029 void memcg_update_array_size(int num)
3031 if (num > memcg_limited_groups_array_size)
3032 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3035 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3037 struct memcg_cache_params *cur_params = s->memcg_params;
3039 VM_BUG_ON(!is_root_cache(s));
3041 if (num_groups > memcg_limited_groups_array_size) {
3043 struct memcg_cache_params *new_params;
3044 ssize_t size = memcg_caches_array_size(num_groups);
3046 size *= sizeof(void *);
3047 size += offsetof(struct memcg_cache_params, memcg_caches);
3049 new_params = kzalloc(size, GFP_KERNEL);
3053 new_params->is_root_cache = true;
3056 * There is the chance it will be bigger than
3057 * memcg_limited_groups_array_size, if we failed an allocation
3058 * in a cache, in which case all caches updated before it, will
3059 * have a bigger array.
3061 * But if that is the case, the data after
3062 * memcg_limited_groups_array_size is certainly unused
3064 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3065 if (!cur_params->memcg_caches[i])
3067 new_params->memcg_caches[i] =
3068 cur_params->memcg_caches[i];
3072 * Ideally, we would wait until all caches succeed, and only
3073 * then free the old one. But this is not worth the extra
3074 * pointer per-cache we'd have to have for this.
3076 * It is not a big deal if some caches are left with a size
3077 * bigger than the others. And all updates will reset this
3080 rcu_assign_pointer(s->memcg_params, new_params);
3082 kfree_rcu(cur_params, rcu_head);
3087 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3088 struct kmem_cache *root_cache)
3092 if (!memcg_kmem_enabled())
3096 size = offsetof(struct memcg_cache_params, memcg_caches);
3097 size += memcg_limited_groups_array_size * sizeof(void *);
3099 size = sizeof(struct memcg_cache_params);
3101 s->memcg_params = kzalloc(size, GFP_KERNEL);
3102 if (!s->memcg_params)
3106 s->memcg_params->memcg = memcg;
3107 s->memcg_params->root_cache = root_cache;
3108 css_get(&memcg->css);
3110 s->memcg_params->is_root_cache = true;
3115 void memcg_free_cache_params(struct kmem_cache *s)
3117 if (!s->memcg_params)
3119 if (!s->memcg_params->is_root_cache)
3120 css_put(&s->memcg_params->memcg->css);
3121 kfree(s->memcg_params);
3124 static void memcg_register_cache(struct mem_cgroup *memcg,
3125 struct kmem_cache *root_cache)
3127 static char memcg_name_buf[NAME_MAX + 1]; /* protected by
3129 struct kmem_cache *cachep;
3132 lockdep_assert_held(&memcg_slab_mutex);
3134 id = memcg_cache_id(memcg);
3137 * Since per-memcg caches are created asynchronously on first
3138 * allocation (see memcg_kmem_get_cache()), several threads can try to
3139 * create the same cache, but only one of them may succeed.
3141 if (cache_from_memcg_idx(root_cache, id))
3144 cgroup_name(memcg->css.cgroup, memcg_name_buf, NAME_MAX + 1);
3145 cachep = memcg_create_kmem_cache(memcg, root_cache, memcg_name_buf);
3147 * If we could not create a memcg cache, do not complain, because
3148 * that's not critical at all as we can always proceed with the root
3154 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3157 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3158 * barrier here to ensure nobody will see the kmem_cache partially
3163 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
3164 root_cache->memcg_params->memcg_caches[id] = cachep;
3167 static void memcg_unregister_cache(struct kmem_cache *cachep)
3169 struct kmem_cache *root_cache;
3170 struct mem_cgroup *memcg;
3173 lockdep_assert_held(&memcg_slab_mutex);
3175 BUG_ON(is_root_cache(cachep));
3177 root_cache = cachep->memcg_params->root_cache;
3178 memcg = cachep->memcg_params->memcg;
3179 id = memcg_cache_id(memcg);
3181 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
3182 root_cache->memcg_params->memcg_caches[id] = NULL;
3184 list_del(&cachep->memcg_params->list);
3186 kmem_cache_destroy(cachep);
3190 * During the creation a new cache, we need to disable our accounting mechanism
3191 * altogether. This is true even if we are not creating, but rather just
3192 * enqueing new caches to be created.
3194 * This is because that process will trigger allocations; some visible, like
3195 * explicit kmallocs to auxiliary data structures, name strings and internal
3196 * cache structures; some well concealed, like INIT_WORK() that can allocate
3197 * objects during debug.
3199 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3200 * to it. This may not be a bounded recursion: since the first cache creation
3201 * failed to complete (waiting on the allocation), we'll just try to create the
3202 * cache again, failing at the same point.
3204 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3205 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3206 * inside the following two functions.
3208 static inline void memcg_stop_kmem_account(void)
3210 VM_BUG_ON(!current->mm);
3211 current->memcg_kmem_skip_account++;
3214 static inline void memcg_resume_kmem_account(void)
3216 VM_BUG_ON(!current->mm);
3217 current->memcg_kmem_skip_account--;
3220 int __memcg_cleanup_cache_params(struct kmem_cache *s)
3222 struct kmem_cache *c;
3225 mutex_lock(&memcg_slab_mutex);
3226 for_each_memcg_cache_index(i) {
3227 c = cache_from_memcg_idx(s, i);
3231 memcg_unregister_cache(c);
3233 if (cache_from_memcg_idx(s, i))
3236 mutex_unlock(&memcg_slab_mutex);
3240 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3242 struct kmem_cache *cachep;
3243 struct memcg_cache_params *params, *tmp;
3245 if (!memcg_kmem_is_active(memcg))
3248 mutex_lock(&memcg_slab_mutex);
3249 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
3250 cachep = memcg_params_to_cache(params);
3251 kmem_cache_shrink(cachep);
3252 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3253 memcg_unregister_cache(cachep);
3255 mutex_unlock(&memcg_slab_mutex);
3258 struct memcg_register_cache_work {
3259 struct mem_cgroup *memcg;
3260 struct kmem_cache *cachep;
3261 struct work_struct work;
3264 static void memcg_register_cache_func(struct work_struct *w)
3266 struct memcg_register_cache_work *cw =
3267 container_of(w, struct memcg_register_cache_work, work);
3268 struct mem_cgroup *memcg = cw->memcg;
3269 struct kmem_cache *cachep = cw->cachep;
3271 mutex_lock(&memcg_slab_mutex);
3272 memcg_register_cache(memcg, cachep);
3273 mutex_unlock(&memcg_slab_mutex);
3275 css_put(&memcg->css);
3280 * Enqueue the creation of a per-memcg kmem_cache.
3282 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
3283 struct kmem_cache *cachep)
3285 struct memcg_register_cache_work *cw;
3287 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
3289 css_put(&memcg->css);
3294 cw->cachep = cachep;
3296 INIT_WORK(&cw->work, memcg_register_cache_func);
3297 schedule_work(&cw->work);
3300 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
3301 struct kmem_cache *cachep)
3304 * We need to stop accounting when we kmalloc, because if the
3305 * corresponding kmalloc cache is not yet created, the first allocation
3306 * in __memcg_schedule_register_cache will recurse.
3308 * However, it is better to enclose the whole function. Depending on
3309 * the debugging options enabled, INIT_WORK(), for instance, can
3310 * trigger an allocation. This too, will make us recurse. Because at
3311 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3312 * the safest choice is to do it like this, wrapping the whole function.
3314 memcg_stop_kmem_account();
3315 __memcg_schedule_register_cache(memcg, cachep);
3316 memcg_resume_kmem_account();
3319 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
3323 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp,
3324 PAGE_SIZE << order);
3326 atomic_add(1 << order, &cachep->memcg_params->nr_pages);
3330 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
3332 memcg_uncharge_kmem(cachep->memcg_params->memcg, PAGE_SIZE << order);
3333 atomic_sub(1 << order, &cachep->memcg_params->nr_pages);
3337 * Return the kmem_cache we're supposed to use for a slab allocation.
3338 * We try to use the current memcg's version of the cache.
3340 * If the cache does not exist yet, if we are the first user of it,
3341 * we either create it immediately, if possible, or create it asynchronously
3343 * In the latter case, we will let the current allocation go through with
3344 * the original cache.
3346 * Can't be called in interrupt context or from kernel threads.
3347 * This function needs to be called with rcu_read_lock() held.
3349 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3352 struct mem_cgroup *memcg;
3353 struct kmem_cache *memcg_cachep;
3355 VM_BUG_ON(!cachep->memcg_params);
3356 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3358 if (!current->mm || current->memcg_kmem_skip_account)
3362 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3364 if (!memcg_can_account_kmem(memcg))
3367 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3368 if (likely(memcg_cachep)) {
3369 cachep = memcg_cachep;
3373 /* The corresponding put will be done in the workqueue. */
3374 if (!css_tryget_online(&memcg->css))
3379 * If we are in a safe context (can wait, and not in interrupt
3380 * context), we could be be predictable and return right away.
3381 * This would guarantee that the allocation being performed
3382 * already belongs in the new cache.
3384 * However, there are some clashes that can arrive from locking.
3385 * For instance, because we acquire the slab_mutex while doing
3386 * memcg_create_kmem_cache, this means no further allocation
3387 * could happen with the slab_mutex held. So it's better to
3390 memcg_schedule_register_cache(memcg, cachep);
3398 * We need to verify if the allocation against current->mm->owner's memcg is
3399 * possible for the given order. But the page is not allocated yet, so we'll
3400 * need a further commit step to do the final arrangements.
3402 * It is possible for the task to switch cgroups in this mean time, so at
3403 * commit time, we can't rely on task conversion any longer. We'll then use
3404 * the handle argument to return to the caller which cgroup we should commit
3405 * against. We could also return the memcg directly and avoid the pointer
3406 * passing, but a boolean return value gives better semantics considering
3407 * the compiled-out case as well.
3409 * Returning true means the allocation is possible.
3412 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3414 struct mem_cgroup *memcg;
3420 * Disabling accounting is only relevant for some specific memcg
3421 * internal allocations. Therefore we would initially not have such
3422 * check here, since direct calls to the page allocator that are
3423 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3424 * outside memcg core. We are mostly concerned with cache allocations,
3425 * and by having this test at memcg_kmem_get_cache, we are already able
3426 * to relay the allocation to the root cache and bypass the memcg cache
3429 * There is one exception, though: the SLUB allocator does not create
3430 * large order caches, but rather service large kmallocs directly from
3431 * the page allocator. Therefore, the following sequence when backed by
3432 * the SLUB allocator:
3434 * memcg_stop_kmem_account();
3435 * kmalloc(<large_number>)
3436 * memcg_resume_kmem_account();
3438 * would effectively ignore the fact that we should skip accounting,
3439 * since it will drive us directly to this function without passing
3440 * through the cache selector memcg_kmem_get_cache. Such large
3441 * allocations are extremely rare but can happen, for instance, for the
3442 * cache arrays. We bring this test here.
3444 if (!current->mm || current->memcg_kmem_skip_account)
3447 memcg = get_mem_cgroup_from_mm(current->mm);
3449 if (!memcg_can_account_kmem(memcg)) {
3450 css_put(&memcg->css);
3454 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3458 css_put(&memcg->css);
3462 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3465 struct page_cgroup *pc;
3467 VM_BUG_ON(mem_cgroup_is_root(memcg));
3469 /* The page allocation failed. Revert */
3471 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3475 pc = lookup_page_cgroup(page);
3476 lock_page_cgroup(pc);
3477 pc->mem_cgroup = memcg;
3478 SetPageCgroupUsed(pc);
3479 unlock_page_cgroup(pc);
3482 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3484 struct mem_cgroup *memcg = NULL;
3485 struct page_cgroup *pc;
3488 pc = lookup_page_cgroup(page);
3490 * Fast unlocked return. Theoretically might have changed, have to
3491 * check again after locking.
3493 if (!PageCgroupUsed(pc))
3496 lock_page_cgroup(pc);
3497 if (PageCgroupUsed(pc)) {
3498 memcg = pc->mem_cgroup;
3499 ClearPageCgroupUsed(pc);
3501 unlock_page_cgroup(pc);
3504 * We trust that only if there is a memcg associated with the page, it
3505 * is a valid allocation
3510 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3511 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3514 static inline void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3517 #endif /* CONFIG_MEMCG_KMEM */
3519 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3521 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3523 * Because tail pages are not marked as "used", set it. We're under
3524 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3525 * charge/uncharge will be never happen and move_account() is done under
3526 * compound_lock(), so we don't have to take care of races.
3528 void mem_cgroup_split_huge_fixup(struct page *head)
3530 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3531 struct page_cgroup *pc;
3532 struct mem_cgroup *memcg;
3535 if (mem_cgroup_disabled())
3538 memcg = head_pc->mem_cgroup;
3539 for (i = 1; i < HPAGE_PMD_NR; i++) {
3541 pc->mem_cgroup = memcg;
3542 smp_wmb();/* see __commit_charge() */
3543 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3545 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3548 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3551 * mem_cgroup_move_account - move account of the page
3553 * @nr_pages: number of regular pages (>1 for huge pages)
3554 * @pc: page_cgroup of the page.
3555 * @from: mem_cgroup which the page is moved from.
3556 * @to: mem_cgroup which the page is moved to. @from != @to.
3558 * The caller must confirm following.
3559 * - page is not on LRU (isolate_page() is useful.)
3560 * - compound_lock is held when nr_pages > 1
3562 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3565 static int mem_cgroup_move_account(struct page *page,
3566 unsigned int nr_pages,
3567 struct page_cgroup *pc,
3568 struct mem_cgroup *from,
3569 struct mem_cgroup *to)
3571 unsigned long flags;
3573 bool anon = PageAnon(page);
3575 VM_BUG_ON(from == to);
3576 VM_BUG_ON_PAGE(PageLRU(page), page);
3578 * The page is isolated from LRU. So, collapse function
3579 * will not handle this page. But page splitting can happen.
3580 * Do this check under compound_page_lock(). The caller should
3584 if (nr_pages > 1 && !PageTransHuge(page))
3587 lock_page_cgroup(pc);
3590 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3593 move_lock_mem_cgroup(from, &flags);
3595 if (!anon && page_mapped(page)) {
3596 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3598 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3602 if (PageWriteback(page)) {
3603 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3605 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3609 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3611 /* caller should have done css_get */
3612 pc->mem_cgroup = to;
3613 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3614 move_unlock_mem_cgroup(from, &flags);
3617 unlock_page_cgroup(pc);
3621 memcg_check_events(to, page);
3622 memcg_check_events(from, page);
3628 * mem_cgroup_move_parent - moves page to the parent group
3629 * @page: the page to move
3630 * @pc: page_cgroup of the page
3631 * @child: page's cgroup
3633 * move charges to its parent or the root cgroup if the group has no
3634 * parent (aka use_hierarchy==0).
3635 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3636 * mem_cgroup_move_account fails) the failure is always temporary and
3637 * it signals a race with a page removal/uncharge or migration. In the
3638 * first case the page is on the way out and it will vanish from the LRU
3639 * on the next attempt and the call should be retried later.
3640 * Isolation from the LRU fails only if page has been isolated from
3641 * the LRU since we looked at it and that usually means either global
3642 * reclaim or migration going on. The page will either get back to the
3644 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3645 * (!PageCgroupUsed) or moved to a different group. The page will
3646 * disappear in the next attempt.
3648 static int mem_cgroup_move_parent(struct page *page,
3649 struct page_cgroup *pc,
3650 struct mem_cgroup *child)
3652 struct mem_cgroup *parent;
3653 unsigned int nr_pages;
3654 unsigned long uninitialized_var(flags);
3657 VM_BUG_ON(mem_cgroup_is_root(child));
3660 if (!get_page_unless_zero(page))
3662 if (isolate_lru_page(page))
3665 nr_pages = hpage_nr_pages(page);
3667 parent = parent_mem_cgroup(child);
3669 * If no parent, move charges to root cgroup.
3672 parent = root_mem_cgroup;
3675 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3676 flags = compound_lock_irqsave(page);
3679 ret = mem_cgroup_move_account(page, nr_pages,
3682 __mem_cgroup_cancel_local_charge(child, nr_pages);
3685 compound_unlock_irqrestore(page, flags);
3686 putback_lru_page(page);
3693 int mem_cgroup_charge_anon(struct page *page,
3694 struct mm_struct *mm, gfp_t gfp_mask)
3696 unsigned int nr_pages = 1;
3697 struct mem_cgroup *memcg;
3700 if (mem_cgroup_disabled())
3703 VM_BUG_ON_PAGE(page_mapped(page), page);
3704 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3707 if (PageTransHuge(page)) {
3708 nr_pages <<= compound_order(page);
3709 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3711 * Never OOM-kill a process for a huge page. The
3712 * fault handler will fall back to regular pages.
3717 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, nr_pages, oom);
3720 __mem_cgroup_commit_charge(memcg, page, nr_pages,
3721 MEM_CGROUP_CHARGE_TYPE_ANON, false);
3726 * While swap-in, try_charge -> commit or cancel, the page is locked.
3727 * And when try_charge() successfully returns, one refcnt to memcg without
3728 * struct page_cgroup is acquired. This refcnt will be consumed by
3729 * "commit()" or removed by "cancel()"
3731 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3734 struct mem_cgroup **memcgp)
3736 struct mem_cgroup *memcg = NULL;
3737 struct page_cgroup *pc;
3740 pc = lookup_page_cgroup(page);
3742 * Every swap fault against a single page tries to charge the
3743 * page, bail as early as possible. shmem_unuse() encounters
3744 * already charged pages, too. The USED bit is protected by
3745 * the page lock, which serializes swap cache removal, which
3746 * in turn serializes uncharging.
3748 if (PageCgroupUsed(pc))
3750 if (do_swap_account)
3751 memcg = try_get_mem_cgroup_from_page(page);
3753 memcg = get_mem_cgroup_from_mm(mm);
3754 ret = mem_cgroup_try_charge(memcg, mask, 1, true);
3755 css_put(&memcg->css);
3757 memcg = root_mem_cgroup;
3765 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3766 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3768 if (mem_cgroup_disabled()) {
3773 * A racing thread's fault, or swapoff, may have already
3774 * updated the pte, and even removed page from swap cache: in
3775 * those cases unuse_pte()'s pte_same() test will fail; but
3776 * there's also a KSM case which does need to charge the page.
3778 if (!PageSwapCache(page)) {
3779 struct mem_cgroup *memcg;
3781 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3787 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3790 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3792 if (mem_cgroup_disabled())
3796 __mem_cgroup_cancel_charge(memcg, 1);
3800 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3801 enum charge_type ctype)
3803 if (mem_cgroup_disabled())
3808 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3810 * Now swap is on-memory. This means this page may be
3811 * counted both as mem and swap....double count.
3812 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3813 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3814 * may call delete_from_swap_cache() before reach here.
3816 if (do_swap_account && PageSwapCache(page)) {
3817 swp_entry_t ent = {.val = page_private(page)};
3818 mem_cgroup_uncharge_swap(ent);
3822 void mem_cgroup_commit_charge_swapin(struct page *page,
3823 struct mem_cgroup *memcg)
3825 __mem_cgroup_commit_charge_swapin(page, memcg,
3826 MEM_CGROUP_CHARGE_TYPE_ANON);
3829 int mem_cgroup_charge_file(struct page *page, struct mm_struct *mm,
3832 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3833 struct mem_cgroup *memcg;
3836 if (mem_cgroup_disabled())
3838 if (PageCompound(page))
3841 if (PageSwapCache(page)) { /* shmem */
3842 ret = __mem_cgroup_try_charge_swapin(mm, page,
3846 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3851 * Page cache insertions can happen without an actual mm
3852 * context, e.g. during disk probing on boot.
3855 memcg = root_mem_cgroup;
3857 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3861 __mem_cgroup_commit_charge(memcg, page, 1, type, false);
3865 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3866 unsigned int nr_pages,
3867 const enum charge_type ctype)
3869 struct memcg_batch_info *batch = NULL;
3870 bool uncharge_memsw = true;
3872 /* If swapout, usage of swap doesn't decrease */
3873 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3874 uncharge_memsw = false;
3876 batch = ¤t->memcg_batch;
3878 * In usual, we do css_get() when we remember memcg pointer.
3879 * But in this case, we keep res->usage until end of a series of
3880 * uncharges. Then, it's ok to ignore memcg's refcnt.
3883 batch->memcg = memcg;
3885 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3886 * In those cases, all pages freed continuously can be expected to be in
3887 * the same cgroup and we have chance to coalesce uncharges.
3888 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3889 * because we want to do uncharge as soon as possible.
3892 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3893 goto direct_uncharge;
3896 goto direct_uncharge;
3899 * In typical case, batch->memcg == mem. This means we can
3900 * merge a series of uncharges to an uncharge of res_counter.
3901 * If not, we uncharge res_counter ony by one.
3903 if (batch->memcg != memcg)
3904 goto direct_uncharge;
3905 /* remember freed charge and uncharge it later */
3908 batch->memsw_nr_pages++;
3911 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3913 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3914 if (unlikely(batch->memcg != memcg))
3915 memcg_oom_recover(memcg);
3919 * uncharge if !page_mapped(page)
3921 static struct mem_cgroup *
3922 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3925 struct mem_cgroup *memcg = NULL;
3926 unsigned int nr_pages = 1;
3927 struct page_cgroup *pc;
3930 if (mem_cgroup_disabled())
3933 if (PageTransHuge(page)) {
3934 nr_pages <<= compound_order(page);
3935 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3938 * Check if our page_cgroup is valid
3940 pc = lookup_page_cgroup(page);
3941 if (unlikely(!PageCgroupUsed(pc)))
3944 lock_page_cgroup(pc);
3946 memcg = pc->mem_cgroup;
3948 if (!PageCgroupUsed(pc))
3951 anon = PageAnon(page);
3954 case MEM_CGROUP_CHARGE_TYPE_ANON:
3956 * Generally PageAnon tells if it's the anon statistics to be
3957 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3958 * used before page reached the stage of being marked PageAnon.
3962 case MEM_CGROUP_CHARGE_TYPE_DROP:
3963 /* See mem_cgroup_prepare_migration() */
3964 if (page_mapped(page))
3967 * Pages under migration may not be uncharged. But
3968 * end_migration() /must/ be the one uncharging the
3969 * unused post-migration page and so it has to call
3970 * here with the migration bit still set. See the
3971 * res_counter handling below.
3973 if (!end_migration && PageCgroupMigration(pc))
3976 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3977 if (!PageAnon(page)) { /* Shared memory */
3978 if (page->mapping && !page_is_file_cache(page))
3980 } else if (page_mapped(page)) /* Anon */
3987 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
3989 ClearPageCgroupUsed(pc);
3991 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3992 * freed from LRU. This is safe because uncharged page is expected not
3993 * to be reused (freed soon). Exception is SwapCache, it's handled by
3994 * special functions.
3997 unlock_page_cgroup(pc);
3999 * even after unlock, we have memcg->res.usage here and this memcg
4000 * will never be freed, so it's safe to call css_get().
4002 memcg_check_events(memcg, page);
4003 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4004 mem_cgroup_swap_statistics(memcg, true);
4005 css_get(&memcg->css);
4008 * Migration does not charge the res_counter for the
4009 * replacement page, so leave it alone when phasing out the
4010 * page that is unused after the migration.
4012 if (!end_migration && !mem_cgroup_is_root(memcg))
4013 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4018 unlock_page_cgroup(pc);
4022 void mem_cgroup_uncharge_page(struct page *page)
4025 if (page_mapped(page))
4027 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
4029 * If the page is in swap cache, uncharge should be deferred
4030 * to the swap path, which also properly accounts swap usage
4031 * and handles memcg lifetime.
4033 * Note that this check is not stable and reclaim may add the
4034 * page to swap cache at any time after this. However, if the
4035 * page is not in swap cache by the time page->mapcount hits
4036 * 0, there won't be any page table references to the swap
4037 * slot, and reclaim will free it and not actually write the
4040 if (PageSwapCache(page))
4042 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4045 void mem_cgroup_uncharge_cache_page(struct page *page)
4047 VM_BUG_ON_PAGE(page_mapped(page), page);
4048 VM_BUG_ON_PAGE(page->mapping, page);
4049 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4053 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4054 * In that cases, pages are freed continuously and we can expect pages
4055 * are in the same memcg. All these calls itself limits the number of
4056 * pages freed at once, then uncharge_start/end() is called properly.
4057 * This may be called prural(2) times in a context,
4060 void mem_cgroup_uncharge_start(void)
4062 current->memcg_batch.do_batch++;
4063 /* We can do nest. */
4064 if (current->memcg_batch.do_batch == 1) {
4065 current->memcg_batch.memcg = NULL;
4066 current->memcg_batch.nr_pages = 0;
4067 current->memcg_batch.memsw_nr_pages = 0;
4071 void mem_cgroup_uncharge_end(void)
4073 struct memcg_batch_info *batch = ¤t->memcg_batch;
4075 if (!batch->do_batch)
4079 if (batch->do_batch) /* If stacked, do nothing. */
4085 * This "batch->memcg" is valid without any css_get/put etc...
4086 * bacause we hide charges behind us.
4088 if (batch->nr_pages)
4089 res_counter_uncharge(&batch->memcg->res,
4090 batch->nr_pages * PAGE_SIZE);
4091 if (batch->memsw_nr_pages)
4092 res_counter_uncharge(&batch->memcg->memsw,
4093 batch->memsw_nr_pages * PAGE_SIZE);
4094 memcg_oom_recover(batch->memcg);
4095 /* forget this pointer (for sanity check) */
4096 batch->memcg = NULL;
4101 * called after __delete_from_swap_cache() and drop "page" account.
4102 * memcg information is recorded to swap_cgroup of "ent"
4105 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4107 struct mem_cgroup *memcg;
4108 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4110 if (!swapout) /* this was a swap cache but the swap is unused ! */
4111 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4113 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4116 * record memcg information, if swapout && memcg != NULL,
4117 * css_get() was called in uncharge().
4119 if (do_swap_account && swapout && memcg)
4120 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4124 #ifdef CONFIG_MEMCG_SWAP
4126 * called from swap_entry_free(). remove record in swap_cgroup and
4127 * uncharge "memsw" account.
4129 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4131 struct mem_cgroup *memcg;
4134 if (!do_swap_account)
4137 id = swap_cgroup_record(ent, 0);
4139 memcg = mem_cgroup_lookup(id);
4142 * We uncharge this because swap is freed. This memcg can
4143 * be obsolete one. We avoid calling css_tryget_online().
4145 if (!mem_cgroup_is_root(memcg))
4146 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4147 mem_cgroup_swap_statistics(memcg, false);
4148 css_put(&memcg->css);
4154 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4155 * @entry: swap entry to be moved
4156 * @from: mem_cgroup which the entry is moved from
4157 * @to: mem_cgroup which the entry is moved to
4159 * It succeeds only when the swap_cgroup's record for this entry is the same
4160 * as the mem_cgroup's id of @from.
4162 * Returns 0 on success, -EINVAL on failure.
4164 * The caller must have charged to @to, IOW, called res_counter_charge() about
4165 * both res and memsw, and called css_get().
4167 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4168 struct mem_cgroup *from, struct mem_cgroup *to)
4170 unsigned short old_id, new_id;
4172 old_id = mem_cgroup_id(from);
4173 new_id = mem_cgroup_id(to);
4175 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4176 mem_cgroup_swap_statistics(from, false);
4177 mem_cgroup_swap_statistics(to, true);
4179 * This function is only called from task migration context now.
4180 * It postpones res_counter and refcount handling till the end
4181 * of task migration(mem_cgroup_clear_mc()) for performance
4182 * improvement. But we cannot postpone css_get(to) because if
4183 * the process that has been moved to @to does swap-in, the
4184 * refcount of @to might be decreased to 0.
4186 * We are in attach() phase, so the cgroup is guaranteed to be
4187 * alive, so we can just call css_get().
4195 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4196 struct mem_cgroup *from, struct mem_cgroup *to)
4203 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4206 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4207 struct mem_cgroup **memcgp)
4209 struct mem_cgroup *memcg = NULL;
4210 unsigned int nr_pages = 1;
4211 struct page_cgroup *pc;
4212 enum charge_type ctype;
4216 if (mem_cgroup_disabled())
4219 if (PageTransHuge(page))
4220 nr_pages <<= compound_order(page);
4222 pc = lookup_page_cgroup(page);
4223 lock_page_cgroup(pc);
4224 if (PageCgroupUsed(pc)) {
4225 memcg = pc->mem_cgroup;
4226 css_get(&memcg->css);
4228 * At migrating an anonymous page, its mapcount goes down
4229 * to 0 and uncharge() will be called. But, even if it's fully
4230 * unmapped, migration may fail and this page has to be
4231 * charged again. We set MIGRATION flag here and delay uncharge
4232 * until end_migration() is called
4234 * Corner Case Thinking
4236 * When the old page was mapped as Anon and it's unmap-and-freed
4237 * while migration was ongoing.
4238 * If unmap finds the old page, uncharge() of it will be delayed
4239 * until end_migration(). If unmap finds a new page, it's
4240 * uncharged when it make mapcount to be 1->0. If unmap code
4241 * finds swap_migration_entry, the new page will not be mapped
4242 * and end_migration() will find it(mapcount==0).
4245 * When the old page was mapped but migraion fails, the kernel
4246 * remaps it. A charge for it is kept by MIGRATION flag even
4247 * if mapcount goes down to 0. We can do remap successfully
4248 * without charging it again.
4251 * The "old" page is under lock_page() until the end of
4252 * migration, so, the old page itself will not be swapped-out.
4253 * If the new page is swapped out before end_migraton, our
4254 * hook to usual swap-out path will catch the event.
4257 SetPageCgroupMigration(pc);
4259 unlock_page_cgroup(pc);
4261 * If the page is not charged at this point,
4269 * We charge new page before it's used/mapped. So, even if unlock_page()
4270 * is called before end_migration, we can catch all events on this new
4271 * page. In the case new page is migrated but not remapped, new page's
4272 * mapcount will be finally 0 and we call uncharge in end_migration().
4275 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4277 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4279 * The page is committed to the memcg, but it's not actually
4280 * charged to the res_counter since we plan on replacing the
4281 * old one and only one page is going to be left afterwards.
4283 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4286 /* remove redundant charge if migration failed*/
4287 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4288 struct page *oldpage, struct page *newpage, bool migration_ok)
4290 struct page *used, *unused;
4291 struct page_cgroup *pc;
4297 if (!migration_ok) {
4304 anon = PageAnon(used);
4305 __mem_cgroup_uncharge_common(unused,
4306 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4307 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4309 css_put(&memcg->css);
4311 * We disallowed uncharge of pages under migration because mapcount
4312 * of the page goes down to zero, temporarly.
4313 * Clear the flag and check the page should be charged.
4315 pc = lookup_page_cgroup(oldpage);
4316 lock_page_cgroup(pc);
4317 ClearPageCgroupMigration(pc);
4318 unlock_page_cgroup(pc);
4321 * If a page is a file cache, radix-tree replacement is very atomic
4322 * and we can skip this check. When it was an Anon page, its mapcount
4323 * goes down to 0. But because we added MIGRATION flage, it's not
4324 * uncharged yet. There are several case but page->mapcount check
4325 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4326 * check. (see prepare_charge() also)
4329 mem_cgroup_uncharge_page(used);
4333 * At replace page cache, newpage is not under any memcg but it's on
4334 * LRU. So, this function doesn't touch res_counter but handles LRU
4335 * in correct way. Both pages are locked so we cannot race with uncharge.
4337 void mem_cgroup_replace_page_cache(struct page *oldpage,
4338 struct page *newpage)
4340 struct mem_cgroup *memcg = NULL;
4341 struct page_cgroup *pc;
4342 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4344 if (mem_cgroup_disabled())
4347 pc = lookup_page_cgroup(oldpage);
4348 /* fix accounting on old pages */
4349 lock_page_cgroup(pc);
4350 if (PageCgroupUsed(pc)) {
4351 memcg = pc->mem_cgroup;
4352 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4353 ClearPageCgroupUsed(pc);
4355 unlock_page_cgroup(pc);
4358 * When called from shmem_replace_page(), in some cases the
4359 * oldpage has already been charged, and in some cases not.
4364 * Even if newpage->mapping was NULL before starting replacement,
4365 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4366 * LRU while we overwrite pc->mem_cgroup.
4368 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4371 #ifdef CONFIG_DEBUG_VM
4372 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4374 struct page_cgroup *pc;
4376 pc = lookup_page_cgroup(page);
4378 * Can be NULL while feeding pages into the page allocator for
4379 * the first time, i.e. during boot or memory hotplug;
4380 * or when mem_cgroup_disabled().
4382 if (likely(pc) && PageCgroupUsed(pc))
4387 bool mem_cgroup_bad_page_check(struct page *page)
4389 if (mem_cgroup_disabled())
4392 return lookup_page_cgroup_used(page) != NULL;
4395 void mem_cgroup_print_bad_page(struct page *page)
4397 struct page_cgroup *pc;
4399 pc = lookup_page_cgroup_used(page);
4401 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4402 pc, pc->flags, pc->mem_cgroup);
4407 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4408 unsigned long long val)
4411 u64 memswlimit, memlimit;
4413 int children = mem_cgroup_count_children(memcg);
4414 u64 curusage, oldusage;
4418 * For keeping hierarchical_reclaim simple, how long we should retry
4419 * is depends on callers. We set our retry-count to be function
4420 * of # of children which we should visit in this loop.
4422 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4424 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4427 while (retry_count) {
4428 if (signal_pending(current)) {
4433 * Rather than hide all in some function, I do this in
4434 * open coded manner. You see what this really does.
4435 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4437 mutex_lock(&set_limit_mutex);
4438 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4439 if (memswlimit < val) {
4441 mutex_unlock(&set_limit_mutex);
4445 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4449 ret = res_counter_set_limit(&memcg->res, val);
4451 if (memswlimit == val)
4452 memcg->memsw_is_minimum = true;
4454 memcg->memsw_is_minimum = false;
4456 mutex_unlock(&set_limit_mutex);
4461 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4462 MEM_CGROUP_RECLAIM_SHRINK);
4463 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4464 /* Usage is reduced ? */
4465 if (curusage >= oldusage)
4468 oldusage = curusage;
4470 if (!ret && enlarge)
4471 memcg_oom_recover(memcg);
4476 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4477 unsigned long long val)
4480 u64 memlimit, memswlimit, oldusage, curusage;
4481 int children = mem_cgroup_count_children(memcg);
4485 /* see mem_cgroup_resize_res_limit */
4486 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4487 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4488 while (retry_count) {
4489 if (signal_pending(current)) {
4494 * Rather than hide all in some function, I do this in
4495 * open coded manner. You see what this really does.
4496 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4498 mutex_lock(&set_limit_mutex);
4499 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4500 if (memlimit > val) {
4502 mutex_unlock(&set_limit_mutex);
4505 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4506 if (memswlimit < val)
4508 ret = res_counter_set_limit(&memcg->memsw, val);
4510 if (memlimit == val)
4511 memcg->memsw_is_minimum = true;
4513 memcg->memsw_is_minimum = false;
4515 mutex_unlock(&set_limit_mutex);
4520 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4521 MEM_CGROUP_RECLAIM_NOSWAP |
4522 MEM_CGROUP_RECLAIM_SHRINK);
4523 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4524 /* Usage is reduced ? */
4525 if (curusage >= oldusage)
4528 oldusage = curusage;
4530 if (!ret && enlarge)
4531 memcg_oom_recover(memcg);
4535 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4537 unsigned long *total_scanned)
4539 unsigned long nr_reclaimed = 0;
4540 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4541 unsigned long reclaimed;
4543 struct mem_cgroup_tree_per_zone *mctz;
4544 unsigned long long excess;
4545 unsigned long nr_scanned;
4550 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4552 * This loop can run a while, specially if mem_cgroup's continuously
4553 * keep exceeding their soft limit and putting the system under
4560 mz = mem_cgroup_largest_soft_limit_node(mctz);
4565 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4566 gfp_mask, &nr_scanned);
4567 nr_reclaimed += reclaimed;
4568 *total_scanned += nr_scanned;
4569 spin_lock(&mctz->lock);
4572 * If we failed to reclaim anything from this memory cgroup
4573 * it is time to move on to the next cgroup
4579 * Loop until we find yet another one.
4581 * By the time we get the soft_limit lock
4582 * again, someone might have aded the
4583 * group back on the RB tree. Iterate to
4584 * make sure we get a different mem.
4585 * mem_cgroup_largest_soft_limit_node returns
4586 * NULL if no other cgroup is present on
4590 __mem_cgroup_largest_soft_limit_node(mctz);
4592 css_put(&next_mz->memcg->css);
4593 else /* next_mz == NULL or other memcg */
4597 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4598 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4600 * One school of thought says that we should not add
4601 * back the node to the tree if reclaim returns 0.
4602 * But our reclaim could return 0, simply because due
4603 * to priority we are exposing a smaller subset of
4604 * memory to reclaim from. Consider this as a longer
4607 /* If excess == 0, no tree ops */
4608 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4609 spin_unlock(&mctz->lock);
4610 css_put(&mz->memcg->css);
4613 * Could not reclaim anything and there are no more
4614 * mem cgroups to try or we seem to be looping without
4615 * reclaiming anything.
4617 if (!nr_reclaimed &&
4619 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4621 } while (!nr_reclaimed);
4623 css_put(&next_mz->memcg->css);
4624 return nr_reclaimed;
4628 * mem_cgroup_force_empty_list - clears LRU of a group
4629 * @memcg: group to clear
4632 * @lru: lru to to clear
4634 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4635 * reclaim the pages page themselves - pages are moved to the parent (or root)
4638 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4639 int node, int zid, enum lru_list lru)
4641 struct lruvec *lruvec;
4642 unsigned long flags;
4643 struct list_head *list;
4647 zone = &NODE_DATA(node)->node_zones[zid];
4648 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4649 list = &lruvec->lists[lru];
4653 struct page_cgroup *pc;
4656 spin_lock_irqsave(&zone->lru_lock, flags);
4657 if (list_empty(list)) {
4658 spin_unlock_irqrestore(&zone->lru_lock, flags);
4661 page = list_entry(list->prev, struct page, lru);
4663 list_move(&page->lru, list);
4665 spin_unlock_irqrestore(&zone->lru_lock, flags);
4668 spin_unlock_irqrestore(&zone->lru_lock, flags);
4670 pc = lookup_page_cgroup(page);
4672 if (mem_cgroup_move_parent(page, pc, memcg)) {
4673 /* found lock contention or "pc" is obsolete. */
4678 } while (!list_empty(list));
4682 * make mem_cgroup's charge to be 0 if there is no task by moving
4683 * all the charges and pages to the parent.
4684 * This enables deleting this mem_cgroup.
4686 * Caller is responsible for holding css reference on the memcg.
4688 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4694 /* This is for making all *used* pages to be on LRU. */
4695 lru_add_drain_all();
4696 drain_all_stock_sync(memcg);
4697 mem_cgroup_start_move(memcg);
4698 for_each_node_state(node, N_MEMORY) {
4699 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4702 mem_cgroup_force_empty_list(memcg,
4707 mem_cgroup_end_move(memcg);
4708 memcg_oom_recover(memcg);
4712 * Kernel memory may not necessarily be trackable to a specific
4713 * process. So they are not migrated, and therefore we can't
4714 * expect their value to drop to 0 here.
4715 * Having res filled up with kmem only is enough.
4717 * This is a safety check because mem_cgroup_force_empty_list
4718 * could have raced with mem_cgroup_replace_page_cache callers
4719 * so the lru seemed empty but the page could have been added
4720 * right after the check. RES_USAGE should be safe as we always
4721 * charge before adding to the LRU.
4723 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4724 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4725 } while (usage > 0);
4729 * Test whether @memcg has children, dead or alive. Note that this
4730 * function doesn't care whether @memcg has use_hierarchy enabled and
4731 * returns %true if there are child csses according to the cgroup
4732 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
4734 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4739 * The lock does not prevent addition or deletion of children, but
4740 * it prevents a new child from being initialized based on this
4741 * parent in css_online(), so it's enough to decide whether
4742 * hierarchically inherited attributes can still be changed or not.
4744 lockdep_assert_held(&memcg_create_mutex);
4747 ret = css_next_child(NULL, &memcg->css);
4753 * Reclaims as many pages from the given memcg as possible and moves
4754 * the rest to the parent.
4756 * Caller is responsible for holding css reference for memcg.
4758 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4760 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4762 /* we call try-to-free pages for make this cgroup empty */
4763 lru_add_drain_all();
4764 /* try to free all pages in this cgroup */
4765 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4768 if (signal_pending(current))
4771 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4775 /* maybe some writeback is necessary */
4776 congestion_wait(BLK_RW_ASYNC, HZ/10);
4784 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
4785 char *buf, size_t nbytes,
4788 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4790 if (mem_cgroup_is_root(memcg))
4792 return mem_cgroup_force_empty(memcg) ?: nbytes;
4795 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4798 return mem_cgroup_from_css(css)->use_hierarchy;
4801 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4802 struct cftype *cft, u64 val)
4805 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4806 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
4808 mutex_lock(&memcg_create_mutex);
4810 if (memcg->use_hierarchy == val)
4814 * If parent's use_hierarchy is set, we can't make any modifications
4815 * in the child subtrees. If it is unset, then the change can
4816 * occur, provided the current cgroup has no children.
4818 * For the root cgroup, parent_mem is NULL, we allow value to be
4819 * set if there are no children.
4821 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4822 (val == 1 || val == 0)) {
4823 if (!memcg_has_children(memcg))
4824 memcg->use_hierarchy = val;
4831 mutex_unlock(&memcg_create_mutex);
4837 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4838 enum mem_cgroup_stat_index idx)
4840 struct mem_cgroup *iter;
4843 /* Per-cpu values can be negative, use a signed accumulator */
4844 for_each_mem_cgroup_tree(iter, memcg)
4845 val += mem_cgroup_read_stat(iter, idx);
4847 if (val < 0) /* race ? */
4852 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4856 if (!mem_cgroup_is_root(memcg)) {
4858 return res_counter_read_u64(&memcg->res, RES_USAGE);
4860 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4864 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4865 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4867 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4868 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4871 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4873 return val << PAGE_SHIFT;
4876 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
4879 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4884 type = MEMFILE_TYPE(cft->private);
4885 name = MEMFILE_ATTR(cft->private);
4889 if (name == RES_USAGE)
4890 val = mem_cgroup_usage(memcg, false);
4892 val = res_counter_read_u64(&memcg->res, name);
4895 if (name == RES_USAGE)
4896 val = mem_cgroup_usage(memcg, true);
4898 val = res_counter_read_u64(&memcg->memsw, name);
4901 val = res_counter_read_u64(&memcg->kmem, name);
4910 #ifdef CONFIG_MEMCG_KMEM
4911 /* should be called with activate_kmem_mutex held */
4912 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
4913 unsigned long long limit)
4918 if (memcg_kmem_is_active(memcg))
4922 * We are going to allocate memory for data shared by all memory
4923 * cgroups so let's stop accounting here.
4925 memcg_stop_kmem_account();
4928 * For simplicity, we won't allow this to be disabled. It also can't
4929 * be changed if the cgroup has children already, or if tasks had
4932 * If tasks join before we set the limit, a person looking at
4933 * kmem.usage_in_bytes will have no way to determine when it took
4934 * place, which makes the value quite meaningless.
4936 * After it first became limited, changes in the value of the limit are
4937 * of course permitted.
4939 mutex_lock(&memcg_create_mutex);
4940 if (cgroup_has_tasks(memcg->css.cgroup) ||
4941 (memcg->use_hierarchy && memcg_has_children(memcg)))
4943 mutex_unlock(&memcg_create_mutex);
4947 memcg_id = ida_simple_get(&kmem_limited_groups,
4948 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
4955 * Make sure we have enough space for this cgroup in each root cache's
4958 mutex_lock(&memcg_slab_mutex);
4959 err = memcg_update_all_caches(memcg_id + 1);
4960 mutex_unlock(&memcg_slab_mutex);
4964 memcg->kmemcg_id = memcg_id;
4965 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
4968 * We couldn't have accounted to this cgroup, because it hasn't got the
4969 * active bit set yet, so this should succeed.
4971 err = res_counter_set_limit(&memcg->kmem, limit);
4974 static_key_slow_inc(&memcg_kmem_enabled_key);
4976 * Setting the active bit after enabling static branching will
4977 * guarantee no one starts accounting before all call sites are
4980 memcg_kmem_set_active(memcg);
4982 memcg_resume_kmem_account();
4986 ida_simple_remove(&kmem_limited_groups, memcg_id);
4990 static int memcg_activate_kmem(struct mem_cgroup *memcg,
4991 unsigned long long limit)
4995 mutex_lock(&activate_kmem_mutex);
4996 ret = __memcg_activate_kmem(memcg, limit);
4997 mutex_unlock(&activate_kmem_mutex);
5001 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5002 unsigned long long val)
5006 if (!memcg_kmem_is_active(memcg))
5007 ret = memcg_activate_kmem(memcg, val);
5009 ret = res_counter_set_limit(&memcg->kmem, val);
5013 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5016 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5021 mutex_lock(&activate_kmem_mutex);
5023 * If the parent cgroup is not kmem-active now, it cannot be activated
5024 * after this point, because it has at least one child already.
5026 if (memcg_kmem_is_active(parent))
5027 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
5028 mutex_unlock(&activate_kmem_mutex);
5032 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
5033 unsigned long long val)
5037 #endif /* CONFIG_MEMCG_KMEM */
5040 * The user of this function is...
5043 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
5044 char *buf, size_t nbytes, loff_t off)
5046 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5049 unsigned long long val;
5052 buf = strstrip(buf);
5053 type = MEMFILE_TYPE(of_cft(of)->private);
5054 name = MEMFILE_ATTR(of_cft(of)->private);
5058 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5062 /* This function does all necessary parse...reuse it */
5063 ret = res_counter_memparse_write_strategy(buf, &val);
5067 ret = mem_cgroup_resize_limit(memcg, val);
5068 else if (type == _MEMSWAP)
5069 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5070 else if (type == _KMEM)
5071 ret = memcg_update_kmem_limit(memcg, val);
5075 case RES_SOFT_LIMIT:
5076 ret = res_counter_memparse_write_strategy(buf, &val);
5080 * For memsw, soft limits are hard to implement in terms
5081 * of semantics, for now, we support soft limits for
5082 * control without swap
5085 ret = res_counter_set_soft_limit(&memcg->res, val);
5090 ret = -EINVAL; /* should be BUG() ? */
5093 return ret ?: nbytes;
5096 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5097 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5099 unsigned long long min_limit, min_memsw_limit, tmp;
5101 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5102 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5103 if (!memcg->use_hierarchy)
5106 while (memcg->css.parent) {
5107 memcg = mem_cgroup_from_css(memcg->css.parent);
5108 if (!memcg->use_hierarchy)
5110 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5111 min_limit = min(min_limit, tmp);
5112 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5113 min_memsw_limit = min(min_memsw_limit, tmp);
5116 *mem_limit = min_limit;
5117 *memsw_limit = min_memsw_limit;
5120 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
5121 size_t nbytes, loff_t off)
5123 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5127 type = MEMFILE_TYPE(of_cft(of)->private);
5128 name = MEMFILE_ATTR(of_cft(of)->private);
5133 res_counter_reset_max(&memcg->res);
5134 else if (type == _MEMSWAP)
5135 res_counter_reset_max(&memcg->memsw);
5136 else if (type == _KMEM)
5137 res_counter_reset_max(&memcg->kmem);
5143 res_counter_reset_failcnt(&memcg->res);
5144 else if (type == _MEMSWAP)
5145 res_counter_reset_failcnt(&memcg->memsw);
5146 else if (type == _KMEM)
5147 res_counter_reset_failcnt(&memcg->kmem);
5156 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5159 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5163 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5164 struct cftype *cft, u64 val)
5166 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5168 if (val >= (1 << NR_MOVE_TYPE))
5172 * No kind of locking is needed in here, because ->can_attach() will
5173 * check this value once in the beginning of the process, and then carry
5174 * on with stale data. This means that changes to this value will only
5175 * affect task migrations starting after the change.
5177 memcg->move_charge_at_immigrate = val;
5181 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5182 struct cftype *cft, u64 val)
5189 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5193 unsigned int lru_mask;
5196 static const struct numa_stat stats[] = {
5197 { "total", LRU_ALL },
5198 { "file", LRU_ALL_FILE },
5199 { "anon", LRU_ALL_ANON },
5200 { "unevictable", BIT(LRU_UNEVICTABLE) },
5202 const struct numa_stat *stat;
5205 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5207 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5208 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5209 seq_printf(m, "%s=%lu", stat->name, nr);
5210 for_each_node_state(nid, N_MEMORY) {
5211 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5213 seq_printf(m, " N%d=%lu", nid, nr);
5218 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5219 struct mem_cgroup *iter;
5222 for_each_mem_cgroup_tree(iter, memcg)
5223 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5224 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5225 for_each_node_state(nid, N_MEMORY) {
5227 for_each_mem_cgroup_tree(iter, memcg)
5228 nr += mem_cgroup_node_nr_lru_pages(
5229 iter, nid, stat->lru_mask);
5230 seq_printf(m, " N%d=%lu", nid, nr);
5237 #endif /* CONFIG_NUMA */
5239 static inline void mem_cgroup_lru_names_not_uptodate(void)
5241 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5244 static int memcg_stat_show(struct seq_file *m, void *v)
5246 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5247 struct mem_cgroup *mi;
5250 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5251 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5253 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5254 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5257 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5258 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5259 mem_cgroup_read_events(memcg, i));
5261 for (i = 0; i < NR_LRU_LISTS; i++)
5262 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5263 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5265 /* Hierarchical information */
5267 unsigned long long limit, memsw_limit;
5268 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5269 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5270 if (do_swap_account)
5271 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5275 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5278 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5280 for_each_mem_cgroup_tree(mi, memcg)
5281 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5282 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5285 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5286 unsigned long long val = 0;
5288 for_each_mem_cgroup_tree(mi, memcg)
5289 val += mem_cgroup_read_events(mi, i);
5290 seq_printf(m, "total_%s %llu\n",
5291 mem_cgroup_events_names[i], val);
5294 for (i = 0; i < NR_LRU_LISTS; i++) {
5295 unsigned long long val = 0;
5297 for_each_mem_cgroup_tree(mi, memcg)
5298 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5299 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5302 #ifdef CONFIG_DEBUG_VM
5305 struct mem_cgroup_per_zone *mz;
5306 struct zone_reclaim_stat *rstat;
5307 unsigned long recent_rotated[2] = {0, 0};
5308 unsigned long recent_scanned[2] = {0, 0};
5310 for_each_online_node(nid)
5311 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5312 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5313 rstat = &mz->lruvec.reclaim_stat;
5315 recent_rotated[0] += rstat->recent_rotated[0];
5316 recent_rotated[1] += rstat->recent_rotated[1];
5317 recent_scanned[0] += rstat->recent_scanned[0];
5318 recent_scanned[1] += rstat->recent_scanned[1];
5320 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5321 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5322 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5323 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5330 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5333 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5335 return mem_cgroup_swappiness(memcg);
5338 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5339 struct cftype *cft, u64 val)
5341 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5347 memcg->swappiness = val;
5349 vm_swappiness = val;
5354 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5356 struct mem_cgroup_threshold_ary *t;
5362 t = rcu_dereference(memcg->thresholds.primary);
5364 t = rcu_dereference(memcg->memsw_thresholds.primary);
5369 usage = mem_cgroup_usage(memcg, swap);
5372 * current_threshold points to threshold just below or equal to usage.
5373 * If it's not true, a threshold was crossed after last
5374 * call of __mem_cgroup_threshold().
5376 i = t->current_threshold;
5379 * Iterate backward over array of thresholds starting from
5380 * current_threshold and check if a threshold is crossed.
5381 * If none of thresholds below usage is crossed, we read
5382 * only one element of the array here.
5384 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5385 eventfd_signal(t->entries[i].eventfd, 1);
5387 /* i = current_threshold + 1 */
5391 * Iterate forward over array of thresholds starting from
5392 * current_threshold+1 and check if a threshold is crossed.
5393 * If none of thresholds above usage is crossed, we read
5394 * only one element of the array here.
5396 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5397 eventfd_signal(t->entries[i].eventfd, 1);
5399 /* Update current_threshold */
5400 t->current_threshold = i - 1;
5405 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5408 __mem_cgroup_threshold(memcg, false);
5409 if (do_swap_account)
5410 __mem_cgroup_threshold(memcg, true);
5412 memcg = parent_mem_cgroup(memcg);
5416 static int compare_thresholds(const void *a, const void *b)
5418 const struct mem_cgroup_threshold *_a = a;
5419 const struct mem_cgroup_threshold *_b = b;
5421 if (_a->threshold > _b->threshold)
5424 if (_a->threshold < _b->threshold)
5430 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5432 struct mem_cgroup_eventfd_list *ev;
5434 list_for_each_entry(ev, &memcg->oom_notify, list)
5435 eventfd_signal(ev->eventfd, 1);
5439 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5441 struct mem_cgroup *iter;
5443 for_each_mem_cgroup_tree(iter, memcg)
5444 mem_cgroup_oom_notify_cb(iter);
5447 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5448 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5450 struct mem_cgroup_thresholds *thresholds;
5451 struct mem_cgroup_threshold_ary *new;
5452 u64 threshold, usage;
5455 ret = res_counter_memparse_write_strategy(args, &threshold);
5459 mutex_lock(&memcg->thresholds_lock);
5462 thresholds = &memcg->thresholds;
5463 else if (type == _MEMSWAP)
5464 thresholds = &memcg->memsw_thresholds;
5468 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5470 /* Check if a threshold crossed before adding a new one */
5471 if (thresholds->primary)
5472 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5474 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5476 /* Allocate memory for new array of thresholds */
5477 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5485 /* Copy thresholds (if any) to new array */
5486 if (thresholds->primary) {
5487 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5488 sizeof(struct mem_cgroup_threshold));
5491 /* Add new threshold */
5492 new->entries[size - 1].eventfd = eventfd;
5493 new->entries[size - 1].threshold = threshold;
5495 /* Sort thresholds. Registering of new threshold isn't time-critical */
5496 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5497 compare_thresholds, NULL);
5499 /* Find current threshold */
5500 new->current_threshold = -1;
5501 for (i = 0; i < size; i++) {
5502 if (new->entries[i].threshold <= usage) {
5504 * new->current_threshold will not be used until
5505 * rcu_assign_pointer(), so it's safe to increment
5508 ++new->current_threshold;
5513 /* Free old spare buffer and save old primary buffer as spare */
5514 kfree(thresholds->spare);
5515 thresholds->spare = thresholds->primary;
5517 rcu_assign_pointer(thresholds->primary, new);
5519 /* To be sure that nobody uses thresholds */
5523 mutex_unlock(&memcg->thresholds_lock);
5528 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5529 struct eventfd_ctx *eventfd, const char *args)
5531 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5534 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5535 struct eventfd_ctx *eventfd, const char *args)
5537 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5540 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5541 struct eventfd_ctx *eventfd, enum res_type type)
5543 struct mem_cgroup_thresholds *thresholds;
5544 struct mem_cgroup_threshold_ary *new;
5548 mutex_lock(&memcg->thresholds_lock);
5550 thresholds = &memcg->thresholds;
5551 else if (type == _MEMSWAP)
5552 thresholds = &memcg->memsw_thresholds;
5556 if (!thresholds->primary)
5559 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5561 /* Check if a threshold crossed before removing */
5562 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5564 /* Calculate new number of threshold */
5566 for (i = 0; i < thresholds->primary->size; i++) {
5567 if (thresholds->primary->entries[i].eventfd != eventfd)
5571 new = thresholds->spare;
5573 /* Set thresholds array to NULL if we don't have thresholds */
5582 /* Copy thresholds and find current threshold */
5583 new->current_threshold = -1;
5584 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5585 if (thresholds->primary->entries[i].eventfd == eventfd)
5588 new->entries[j] = thresholds->primary->entries[i];
5589 if (new->entries[j].threshold <= usage) {
5591 * new->current_threshold will not be used
5592 * until rcu_assign_pointer(), so it's safe to increment
5595 ++new->current_threshold;
5601 /* Swap primary and spare array */
5602 thresholds->spare = thresholds->primary;
5603 /* If all events are unregistered, free the spare array */
5605 kfree(thresholds->spare);
5606 thresholds->spare = NULL;
5609 rcu_assign_pointer(thresholds->primary, new);
5611 /* To be sure that nobody uses thresholds */
5614 mutex_unlock(&memcg->thresholds_lock);
5617 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5618 struct eventfd_ctx *eventfd)
5620 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5623 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5624 struct eventfd_ctx *eventfd)
5626 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5629 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5630 struct eventfd_ctx *eventfd, const char *args)
5632 struct mem_cgroup_eventfd_list *event;
5634 event = kmalloc(sizeof(*event), GFP_KERNEL);
5638 spin_lock(&memcg_oom_lock);
5640 event->eventfd = eventfd;
5641 list_add(&event->list, &memcg->oom_notify);
5643 /* already in OOM ? */
5644 if (atomic_read(&memcg->under_oom))
5645 eventfd_signal(eventfd, 1);
5646 spin_unlock(&memcg_oom_lock);
5651 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5652 struct eventfd_ctx *eventfd)
5654 struct mem_cgroup_eventfd_list *ev, *tmp;
5656 spin_lock(&memcg_oom_lock);
5658 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5659 if (ev->eventfd == eventfd) {
5660 list_del(&ev->list);
5665 spin_unlock(&memcg_oom_lock);
5668 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5670 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5672 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5673 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5677 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5678 struct cftype *cft, u64 val)
5680 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5682 /* cannot set to root cgroup and only 0 and 1 are allowed */
5683 if (!css->parent || !((val == 0) || (val == 1)))
5686 memcg->oom_kill_disable = val;
5688 memcg_oom_recover(memcg);
5693 #ifdef CONFIG_MEMCG_KMEM
5694 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5698 memcg->kmemcg_id = -1;
5699 ret = memcg_propagate_kmem(memcg);
5703 return mem_cgroup_sockets_init(memcg, ss);
5706 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5708 mem_cgroup_sockets_destroy(memcg);
5711 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5713 if (!memcg_kmem_is_active(memcg))
5717 * kmem charges can outlive the cgroup. In the case of slab
5718 * pages, for instance, a page contain objects from various
5719 * processes. As we prevent from taking a reference for every
5720 * such allocation we have to be careful when doing uncharge
5721 * (see memcg_uncharge_kmem) and here during offlining.
5723 * The idea is that that only the _last_ uncharge which sees
5724 * the dead memcg will drop the last reference. An additional
5725 * reference is taken here before the group is marked dead
5726 * which is then paired with css_put during uncharge resp. here.
5728 * Although this might sound strange as this path is called from
5729 * css_offline() when the referencemight have dropped down to 0 and
5730 * shouldn't be incremented anymore (css_tryget_online() would
5731 * fail) we do not have other options because of the kmem
5732 * allocations lifetime.
5734 css_get(&memcg->css);
5736 memcg_kmem_mark_dead(memcg);
5738 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5741 if (memcg_kmem_test_and_clear_dead(memcg))
5742 css_put(&memcg->css);
5745 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5750 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5754 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5760 * DO NOT USE IN NEW FILES.
5762 * "cgroup.event_control" implementation.
5764 * This is way over-engineered. It tries to support fully configurable
5765 * events for each user. Such level of flexibility is completely
5766 * unnecessary especially in the light of the planned unified hierarchy.
5768 * Please deprecate this and replace with something simpler if at all
5773 * Unregister event and free resources.
5775 * Gets called from workqueue.
5777 static void memcg_event_remove(struct work_struct *work)
5779 struct mem_cgroup_event *event =
5780 container_of(work, struct mem_cgroup_event, remove);
5781 struct mem_cgroup *memcg = event->memcg;
5783 remove_wait_queue(event->wqh, &event->wait);
5785 event->unregister_event(memcg, event->eventfd);
5787 /* Notify userspace the event is going away. */
5788 eventfd_signal(event->eventfd, 1);
5790 eventfd_ctx_put(event->eventfd);
5792 css_put(&memcg->css);
5796 * Gets called on POLLHUP on eventfd when user closes it.
5798 * Called with wqh->lock held and interrupts disabled.
5800 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
5801 int sync, void *key)
5803 struct mem_cgroup_event *event =
5804 container_of(wait, struct mem_cgroup_event, wait);
5805 struct mem_cgroup *memcg = event->memcg;
5806 unsigned long flags = (unsigned long)key;
5808 if (flags & POLLHUP) {
5810 * If the event has been detached at cgroup removal, we
5811 * can simply return knowing the other side will cleanup
5814 * We can't race against event freeing since the other
5815 * side will require wqh->lock via remove_wait_queue(),
5818 spin_lock(&memcg->event_list_lock);
5819 if (!list_empty(&event->list)) {
5820 list_del_init(&event->list);
5822 * We are in atomic context, but cgroup_event_remove()
5823 * may sleep, so we have to call it in workqueue.
5825 schedule_work(&event->remove);
5827 spin_unlock(&memcg->event_list_lock);
5833 static void memcg_event_ptable_queue_proc(struct file *file,
5834 wait_queue_head_t *wqh, poll_table *pt)
5836 struct mem_cgroup_event *event =
5837 container_of(pt, struct mem_cgroup_event, pt);
5840 add_wait_queue(wqh, &event->wait);
5844 * DO NOT USE IN NEW FILES.
5846 * Parse input and register new cgroup event handler.
5848 * Input must be in format '<event_fd> <control_fd> <args>'.
5849 * Interpretation of args is defined by control file implementation.
5851 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
5852 char *buf, size_t nbytes, loff_t off)
5854 struct cgroup_subsys_state *css = of_css(of);
5855 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5856 struct mem_cgroup_event *event;
5857 struct cgroup_subsys_state *cfile_css;
5858 unsigned int efd, cfd;
5865 buf = strstrip(buf);
5867 efd = simple_strtoul(buf, &endp, 10);
5872 cfd = simple_strtoul(buf, &endp, 10);
5873 if ((*endp != ' ') && (*endp != '\0'))
5877 event = kzalloc(sizeof(*event), GFP_KERNEL);
5881 event->memcg = memcg;
5882 INIT_LIST_HEAD(&event->list);
5883 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
5884 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
5885 INIT_WORK(&event->remove, memcg_event_remove);
5893 event->eventfd = eventfd_ctx_fileget(efile.file);
5894 if (IS_ERR(event->eventfd)) {
5895 ret = PTR_ERR(event->eventfd);
5902 goto out_put_eventfd;
5905 /* the process need read permission on control file */
5906 /* AV: shouldn't we check that it's been opened for read instead? */
5907 ret = inode_permission(file_inode(cfile.file), MAY_READ);
5912 * Determine the event callbacks and set them in @event. This used
5913 * to be done via struct cftype but cgroup core no longer knows
5914 * about these events. The following is crude but the whole thing
5915 * is for compatibility anyway.
5917 * DO NOT ADD NEW FILES.
5919 name = cfile.file->f_dentry->d_name.name;
5921 if (!strcmp(name, "memory.usage_in_bytes")) {
5922 event->register_event = mem_cgroup_usage_register_event;
5923 event->unregister_event = mem_cgroup_usage_unregister_event;
5924 } else if (!strcmp(name, "memory.oom_control")) {
5925 event->register_event = mem_cgroup_oom_register_event;
5926 event->unregister_event = mem_cgroup_oom_unregister_event;
5927 } else if (!strcmp(name, "memory.pressure_level")) {
5928 event->register_event = vmpressure_register_event;
5929 event->unregister_event = vmpressure_unregister_event;
5930 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
5931 event->register_event = memsw_cgroup_usage_register_event;
5932 event->unregister_event = memsw_cgroup_usage_unregister_event;
5939 * Verify @cfile should belong to @css. Also, remaining events are
5940 * automatically removed on cgroup destruction but the removal is
5941 * asynchronous, so take an extra ref on @css.
5943 cfile_css = css_tryget_online_from_dir(cfile.file->f_dentry->d_parent,
5944 &memory_cgrp_subsys);
5946 if (IS_ERR(cfile_css))
5948 if (cfile_css != css) {
5953 ret = event->register_event(memcg, event->eventfd, buf);
5957 efile.file->f_op->poll(efile.file, &event->pt);
5959 spin_lock(&memcg->event_list_lock);
5960 list_add(&event->list, &memcg->event_list);
5961 spin_unlock(&memcg->event_list_lock);
5973 eventfd_ctx_put(event->eventfd);
5982 static struct cftype mem_cgroup_files[] = {
5984 .name = "usage_in_bytes",
5985 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5986 .read_u64 = mem_cgroup_read_u64,
5989 .name = "max_usage_in_bytes",
5990 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5991 .write = mem_cgroup_reset,
5992 .read_u64 = mem_cgroup_read_u64,
5995 .name = "limit_in_bytes",
5996 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5997 .write = mem_cgroup_write,
5998 .read_u64 = mem_cgroup_read_u64,
6001 .name = "soft_limit_in_bytes",
6002 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6003 .write = mem_cgroup_write,
6004 .read_u64 = mem_cgroup_read_u64,
6008 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6009 .write = mem_cgroup_reset,
6010 .read_u64 = mem_cgroup_read_u64,
6014 .seq_show = memcg_stat_show,
6017 .name = "force_empty",
6018 .write = mem_cgroup_force_empty_write,
6021 .name = "use_hierarchy",
6022 .flags = CFTYPE_INSANE,
6023 .write_u64 = mem_cgroup_hierarchy_write,
6024 .read_u64 = mem_cgroup_hierarchy_read,
6027 .name = "cgroup.event_control", /* XXX: for compat */
6028 .write = memcg_write_event_control,
6029 .flags = CFTYPE_NO_PREFIX,
6033 .name = "swappiness",
6034 .read_u64 = mem_cgroup_swappiness_read,
6035 .write_u64 = mem_cgroup_swappiness_write,
6038 .name = "move_charge_at_immigrate",
6039 .read_u64 = mem_cgroup_move_charge_read,
6040 .write_u64 = mem_cgroup_move_charge_write,
6043 .name = "oom_control",
6044 .seq_show = mem_cgroup_oom_control_read,
6045 .write_u64 = mem_cgroup_oom_control_write,
6046 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6049 .name = "pressure_level",
6053 .name = "numa_stat",
6054 .seq_show = memcg_numa_stat_show,
6057 #ifdef CONFIG_MEMCG_KMEM
6059 .name = "kmem.limit_in_bytes",
6060 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6061 .write = mem_cgroup_write,
6062 .read_u64 = mem_cgroup_read_u64,
6065 .name = "kmem.usage_in_bytes",
6066 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6067 .read_u64 = mem_cgroup_read_u64,
6070 .name = "kmem.failcnt",
6071 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6072 .write = mem_cgroup_reset,
6073 .read_u64 = mem_cgroup_read_u64,
6076 .name = "kmem.max_usage_in_bytes",
6077 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6078 .write = mem_cgroup_reset,
6079 .read_u64 = mem_cgroup_read_u64,
6081 #ifdef CONFIG_SLABINFO
6083 .name = "kmem.slabinfo",
6084 .seq_show = mem_cgroup_slabinfo_read,
6088 { }, /* terminate */
6091 #ifdef CONFIG_MEMCG_SWAP
6092 static struct cftype memsw_cgroup_files[] = {
6094 .name = "memsw.usage_in_bytes",
6095 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6096 .read_u64 = mem_cgroup_read_u64,
6099 .name = "memsw.max_usage_in_bytes",
6100 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6101 .write = mem_cgroup_reset,
6102 .read_u64 = mem_cgroup_read_u64,
6105 .name = "memsw.limit_in_bytes",
6106 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6107 .write = mem_cgroup_write,
6108 .read_u64 = mem_cgroup_read_u64,
6111 .name = "memsw.failcnt",
6112 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6113 .write = mem_cgroup_reset,
6114 .read_u64 = mem_cgroup_read_u64,
6116 { }, /* terminate */
6119 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6121 struct mem_cgroup_per_node *pn;
6122 struct mem_cgroup_per_zone *mz;
6123 int zone, tmp = node;
6125 * This routine is called against possible nodes.
6126 * But it's BUG to call kmalloc() against offline node.
6128 * TODO: this routine can waste much memory for nodes which will
6129 * never be onlined. It's better to use memory hotplug callback
6132 if (!node_state(node, N_NORMAL_MEMORY))
6134 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6138 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6139 mz = &pn->zoneinfo[zone];
6140 lruvec_init(&mz->lruvec);
6141 mz->usage_in_excess = 0;
6142 mz->on_tree = false;
6145 memcg->nodeinfo[node] = pn;
6149 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6151 kfree(memcg->nodeinfo[node]);
6154 static struct mem_cgroup *mem_cgroup_alloc(void)
6156 struct mem_cgroup *memcg;
6159 size = sizeof(struct mem_cgroup);
6160 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6162 memcg = kzalloc(size, GFP_KERNEL);
6166 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6169 spin_lock_init(&memcg->pcp_counter_lock);
6178 * At destroying mem_cgroup, references from swap_cgroup can remain.
6179 * (scanning all at force_empty is too costly...)
6181 * Instead of clearing all references at force_empty, we remember
6182 * the number of reference from swap_cgroup and free mem_cgroup when
6183 * it goes down to 0.
6185 * Removal of cgroup itself succeeds regardless of refs from swap.
6188 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6192 mem_cgroup_remove_from_trees(memcg);
6195 free_mem_cgroup_per_zone_info(memcg, node);
6197 free_percpu(memcg->stat);
6200 * We need to make sure that (at least for now), the jump label
6201 * destruction code runs outside of the cgroup lock. This is because
6202 * get_online_cpus(), which is called from the static_branch update,
6203 * can't be called inside the cgroup_lock. cpusets are the ones
6204 * enforcing this dependency, so if they ever change, we might as well.
6206 * schedule_work() will guarantee this happens. Be careful if you need
6207 * to move this code around, and make sure it is outside
6210 disarm_static_keys(memcg);
6215 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6217 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6219 if (!memcg->res.parent)
6221 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6223 EXPORT_SYMBOL(parent_mem_cgroup);
6225 static void __init mem_cgroup_soft_limit_tree_init(void)
6227 struct mem_cgroup_tree_per_node *rtpn;
6228 struct mem_cgroup_tree_per_zone *rtpz;
6229 int tmp, node, zone;
6231 for_each_node(node) {
6233 if (!node_state(node, N_NORMAL_MEMORY))
6235 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6238 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6240 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6241 rtpz = &rtpn->rb_tree_per_zone[zone];
6242 rtpz->rb_root = RB_ROOT;
6243 spin_lock_init(&rtpz->lock);
6248 static struct cgroup_subsys_state * __ref
6249 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6251 struct mem_cgroup *memcg;
6252 long error = -ENOMEM;
6255 memcg = mem_cgroup_alloc();
6257 return ERR_PTR(error);
6260 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6264 if (parent_css == NULL) {
6265 root_mem_cgroup = memcg;
6266 res_counter_init(&memcg->res, NULL);
6267 res_counter_init(&memcg->memsw, NULL);
6268 res_counter_init(&memcg->kmem, NULL);
6271 memcg->last_scanned_node = MAX_NUMNODES;
6272 INIT_LIST_HEAD(&memcg->oom_notify);
6273 memcg->move_charge_at_immigrate = 0;
6274 mutex_init(&memcg->thresholds_lock);
6275 spin_lock_init(&memcg->move_lock);
6276 vmpressure_init(&memcg->vmpressure);
6277 INIT_LIST_HEAD(&memcg->event_list);
6278 spin_lock_init(&memcg->event_list_lock);
6283 __mem_cgroup_free(memcg);
6284 return ERR_PTR(error);
6288 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6290 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6291 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
6293 if (css->id > MEM_CGROUP_ID_MAX)
6299 mutex_lock(&memcg_create_mutex);
6301 memcg->use_hierarchy = parent->use_hierarchy;
6302 memcg->oom_kill_disable = parent->oom_kill_disable;
6303 memcg->swappiness = mem_cgroup_swappiness(parent);
6305 if (parent->use_hierarchy) {
6306 res_counter_init(&memcg->res, &parent->res);
6307 res_counter_init(&memcg->memsw, &parent->memsw);
6308 res_counter_init(&memcg->kmem, &parent->kmem);
6311 * No need to take a reference to the parent because cgroup
6312 * core guarantees its existence.
6315 res_counter_init(&memcg->res, NULL);
6316 res_counter_init(&memcg->memsw, NULL);
6317 res_counter_init(&memcg->kmem, NULL);
6319 * Deeper hierachy with use_hierarchy == false doesn't make
6320 * much sense so let cgroup subsystem know about this
6321 * unfortunate state in our controller.
6323 if (parent != root_mem_cgroup)
6324 memory_cgrp_subsys.broken_hierarchy = true;
6326 mutex_unlock(&memcg_create_mutex);
6328 return memcg_init_kmem(memcg, &memory_cgrp_subsys);
6332 * Announce all parents that a group from their hierarchy is gone.
6334 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6336 struct mem_cgroup *parent = memcg;
6338 while ((parent = parent_mem_cgroup(parent)))
6339 mem_cgroup_iter_invalidate(parent);
6342 * if the root memcg is not hierarchical we have to check it
6345 if (!root_mem_cgroup->use_hierarchy)
6346 mem_cgroup_iter_invalidate(root_mem_cgroup);
6349 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6351 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6352 struct mem_cgroup_event *event, *tmp;
6353 struct cgroup_subsys_state *iter;
6356 * Unregister events and notify userspace.
6357 * Notify userspace about cgroup removing only after rmdir of cgroup
6358 * directory to avoid race between userspace and kernelspace.
6360 spin_lock(&memcg->event_list_lock);
6361 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6362 list_del_init(&event->list);
6363 schedule_work(&event->remove);
6365 spin_unlock(&memcg->event_list_lock);
6367 kmem_cgroup_css_offline(memcg);
6369 mem_cgroup_invalidate_reclaim_iterators(memcg);
6372 * This requires that offlining is serialized. Right now that is
6373 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6375 css_for_each_descendant_post(iter, css)
6376 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
6378 memcg_unregister_all_caches(memcg);
6379 vmpressure_cleanup(&memcg->vmpressure);
6382 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6384 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6386 * XXX: css_offline() would be where we should reparent all
6387 * memory to prepare the cgroup for destruction. However,
6388 * memcg does not do css_tryget_online() and res_counter charging
6389 * under the same RCU lock region, which means that charging
6390 * could race with offlining. Offlining only happens to
6391 * cgroups with no tasks in them but charges can show up
6392 * without any tasks from the swapin path when the target
6393 * memcg is looked up from the swapout record and not from the
6394 * current task as it usually is. A race like this can leak
6395 * charges and put pages with stale cgroup pointers into
6399 * lookup_swap_cgroup_id()
6401 * mem_cgroup_lookup()
6402 * css_tryget_online()
6404 * disable css_tryget_online()
6407 * reparent_charges()
6408 * res_counter_charge()
6411 * pc->mem_cgroup = dead memcg
6414 * The bulk of the charges are still moved in offline_css() to
6415 * avoid pinning a lot of pages in case a long-term reference
6416 * like a swapout record is deferring the css_free() to long
6417 * after offlining. But this makes sure we catch any charges
6418 * made after offlining:
6420 mem_cgroup_reparent_charges(memcg);
6422 memcg_destroy_kmem(memcg);
6423 __mem_cgroup_free(memcg);
6427 /* Handlers for move charge at task migration. */
6428 #define PRECHARGE_COUNT_AT_ONCE 256
6429 static int mem_cgroup_do_precharge(unsigned long count)
6432 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6433 struct mem_cgroup *memcg = mc.to;
6435 if (mem_cgroup_is_root(memcg)) {
6436 mc.precharge += count;
6437 /* we don't need css_get for root */
6440 /* try to charge at once */
6442 struct res_counter *dummy;
6444 * "memcg" cannot be under rmdir() because we've already checked
6445 * by cgroup_lock_live_cgroup() that it is not removed and we
6446 * are still under the same cgroup_mutex. So we can postpone
6449 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6451 if (do_swap_account && res_counter_charge(&memcg->memsw,
6452 PAGE_SIZE * count, &dummy)) {
6453 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6456 mc.precharge += count;
6460 /* fall back to one by one charge */
6462 if (signal_pending(current)) {
6466 if (!batch_count--) {
6467 batch_count = PRECHARGE_COUNT_AT_ONCE;
6470 ret = mem_cgroup_try_charge(memcg, GFP_KERNEL, 1, false);
6472 /* mem_cgroup_clear_mc() will do uncharge later */
6480 * get_mctgt_type - get target type of moving charge
6481 * @vma: the vma the pte to be checked belongs
6482 * @addr: the address corresponding to the pte to be checked
6483 * @ptent: the pte to be checked
6484 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6487 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6488 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6489 * move charge. if @target is not NULL, the page is stored in target->page
6490 * with extra refcnt got(Callers should handle it).
6491 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6492 * target for charge migration. if @target is not NULL, the entry is stored
6495 * Called with pte lock held.
6502 enum mc_target_type {
6508 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6509 unsigned long addr, pte_t ptent)
6511 struct page *page = vm_normal_page(vma, addr, ptent);
6513 if (!page || !page_mapped(page))
6515 if (PageAnon(page)) {
6516 /* we don't move shared anon */
6519 } else if (!move_file())
6520 /* we ignore mapcount for file pages */
6522 if (!get_page_unless_zero(page))
6529 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6530 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6532 struct page *page = NULL;
6533 swp_entry_t ent = pte_to_swp_entry(ptent);
6535 if (!move_anon() || non_swap_entry(ent))
6538 * Because lookup_swap_cache() updates some statistics counter,
6539 * we call find_get_page() with swapper_space directly.
6541 page = find_get_page(swap_address_space(ent), ent.val);
6542 if (do_swap_account)
6543 entry->val = ent.val;
6548 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6549 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6555 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6556 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6558 struct page *page = NULL;
6559 struct address_space *mapping;
6562 if (!vma->vm_file) /* anonymous vma */
6567 mapping = vma->vm_file->f_mapping;
6568 if (pte_none(ptent))
6569 pgoff = linear_page_index(vma, addr);
6570 else /* pte_file(ptent) is true */
6571 pgoff = pte_to_pgoff(ptent);
6573 /* page is moved even if it's not RSS of this task(page-faulted). */
6575 /* shmem/tmpfs may report page out on swap: account for that too. */
6576 if (shmem_mapping(mapping)) {
6577 page = find_get_entry(mapping, pgoff);
6578 if (radix_tree_exceptional_entry(page)) {
6579 swp_entry_t swp = radix_to_swp_entry(page);
6580 if (do_swap_account)
6582 page = find_get_page(swap_address_space(swp), swp.val);
6585 page = find_get_page(mapping, pgoff);
6587 page = find_get_page(mapping, pgoff);
6592 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6593 unsigned long addr, pte_t ptent, union mc_target *target)
6595 struct page *page = NULL;
6596 struct page_cgroup *pc;
6597 enum mc_target_type ret = MC_TARGET_NONE;
6598 swp_entry_t ent = { .val = 0 };
6600 if (pte_present(ptent))
6601 page = mc_handle_present_pte(vma, addr, ptent);
6602 else if (is_swap_pte(ptent))
6603 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6604 else if (pte_none(ptent) || pte_file(ptent))
6605 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6607 if (!page && !ent.val)
6610 pc = lookup_page_cgroup(page);
6612 * Do only loose check w/o page_cgroup lock.
6613 * mem_cgroup_move_account() checks the pc is valid or not under
6616 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6617 ret = MC_TARGET_PAGE;
6619 target->page = page;
6621 if (!ret || !target)
6624 /* There is a swap entry and a page doesn't exist or isn't charged */
6625 if (ent.val && !ret &&
6626 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6627 ret = MC_TARGET_SWAP;
6634 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6636 * We don't consider swapping or file mapped pages because THP does not
6637 * support them for now.
6638 * Caller should make sure that pmd_trans_huge(pmd) is true.
6640 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6641 unsigned long addr, pmd_t pmd, union mc_target *target)
6643 struct page *page = NULL;
6644 struct page_cgroup *pc;
6645 enum mc_target_type ret = MC_TARGET_NONE;
6647 page = pmd_page(pmd);
6648 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6651 pc = lookup_page_cgroup(page);
6652 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6653 ret = MC_TARGET_PAGE;
6656 target->page = page;
6662 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6663 unsigned long addr, pmd_t pmd, union mc_target *target)
6665 return MC_TARGET_NONE;
6669 static int mem_cgroup_count_precharge_pte(pte_t *pte,
6670 unsigned long addr, unsigned long end,
6671 struct mm_walk *walk)
6673 if (get_mctgt_type(walk->vma, addr, *pte, NULL))
6674 mc.precharge++; /* increment precharge temporarily */
6678 static int mem_cgroup_count_precharge_pmd(pmd_t *pmd,
6679 unsigned long addr, unsigned long end,
6680 struct mm_walk *walk)
6682 struct vm_area_struct *vma = walk->vma;
6685 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6686 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6687 mc.precharge += HPAGE_PMD_NR;
6689 /* don't call mem_cgroup_count_precharge_pte() */
6695 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6697 unsigned long precharge;
6698 struct vm_area_struct *vma;
6700 struct mm_walk mem_cgroup_count_precharge_walk = {
6701 .pmd_entry = mem_cgroup_count_precharge_pmd,
6702 .pte_entry = mem_cgroup_count_precharge_pte,
6705 down_read(&mm->mmap_sem);
6706 for (vma = mm->mmap; vma; vma = vma->vm_next)
6707 walk_page_vma(vma, &mem_cgroup_count_precharge_walk);
6708 up_read(&mm->mmap_sem);
6710 precharge = mc.precharge;
6716 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6718 unsigned long precharge = mem_cgroup_count_precharge(mm);
6720 VM_BUG_ON(mc.moving_task);
6721 mc.moving_task = current;
6722 return mem_cgroup_do_precharge(precharge);
6725 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6726 static void __mem_cgroup_clear_mc(void)
6728 struct mem_cgroup *from = mc.from;
6729 struct mem_cgroup *to = mc.to;
6732 /* we must uncharge all the leftover precharges from mc.to */
6734 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6738 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6739 * we must uncharge here.
6741 if (mc.moved_charge) {
6742 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6743 mc.moved_charge = 0;
6745 /* we must fixup refcnts and charges */
6746 if (mc.moved_swap) {
6747 /* uncharge swap account from the old cgroup */
6748 if (!mem_cgroup_is_root(mc.from))
6749 res_counter_uncharge(&mc.from->memsw,
6750 PAGE_SIZE * mc.moved_swap);
6752 for (i = 0; i < mc.moved_swap; i++)
6753 css_put(&mc.from->css);
6755 if (!mem_cgroup_is_root(mc.to)) {
6757 * we charged both to->res and to->memsw, so we should
6760 res_counter_uncharge(&mc.to->res,
6761 PAGE_SIZE * mc.moved_swap);
6763 /* we've already done css_get(mc.to) */
6766 memcg_oom_recover(from);
6767 memcg_oom_recover(to);
6768 wake_up_all(&mc.waitq);
6771 static void mem_cgroup_clear_mc(void)
6773 struct mem_cgroup *from = mc.from;
6776 * we must clear moving_task before waking up waiters at the end of
6779 mc.moving_task = NULL;
6780 __mem_cgroup_clear_mc();
6781 spin_lock(&mc.lock);
6784 spin_unlock(&mc.lock);
6785 mem_cgroup_end_move(from);
6788 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6789 struct cgroup_taskset *tset)
6791 struct task_struct *p = cgroup_taskset_first(tset);
6793 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6794 unsigned long move_charge_at_immigrate;
6797 * We are now commited to this value whatever it is. Changes in this
6798 * tunable will only affect upcoming migrations, not the current one.
6799 * So we need to save it, and keep it going.
6801 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6802 if (move_charge_at_immigrate) {
6803 struct mm_struct *mm;
6804 struct mem_cgroup *from = mem_cgroup_from_task(p);
6806 VM_BUG_ON(from == memcg);
6808 mm = get_task_mm(p);
6811 /* We move charges only when we move a owner of the mm */
6812 if (mm->owner == p) {
6815 VM_BUG_ON(mc.precharge);
6816 VM_BUG_ON(mc.moved_charge);
6817 VM_BUG_ON(mc.moved_swap);
6818 mem_cgroup_start_move(from);
6819 spin_lock(&mc.lock);
6822 mc.immigrate_flags = move_charge_at_immigrate;
6823 spin_unlock(&mc.lock);
6824 /* We set mc.moving_task later */
6826 ret = mem_cgroup_precharge_mc(mm);
6828 mem_cgroup_clear_mc();
6835 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6836 struct cgroup_taskset *tset)
6838 mem_cgroup_clear_mc();
6841 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6842 unsigned long addr, unsigned long end,
6843 struct mm_walk *walk)
6846 struct vm_area_struct *vma = walk->vma;
6849 enum mc_target_type target_type;
6850 union mc_target target;
6852 struct page_cgroup *pc;
6855 * We don't take compound_lock() here but no race with splitting thp
6857 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6858 * under splitting, which means there's no concurrent thp split,
6859 * - if another thread runs into split_huge_page() just after we
6860 * entered this if-block, the thread must wait for page table lock
6861 * to be unlocked in __split_huge_page_splitting(), where the main
6862 * part of thp split is not executed yet.
6864 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6865 if (mc.precharge < HPAGE_PMD_NR) {
6869 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6870 if (target_type == MC_TARGET_PAGE) {
6872 if (!isolate_lru_page(page)) {
6873 pc = lookup_page_cgroup(page);
6874 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6875 pc, mc.from, mc.to)) {
6876 mc.precharge -= HPAGE_PMD_NR;
6877 mc.moved_charge += HPAGE_PMD_NR;
6879 putback_lru_page(page);
6887 if (pmd_trans_unstable(pmd))
6890 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6891 for (; addr != end; addr += PAGE_SIZE) {
6892 pte_t ptent = *(pte++);
6898 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6899 case MC_TARGET_PAGE:
6901 if (isolate_lru_page(page))
6903 pc = lookup_page_cgroup(page);
6904 if (!mem_cgroup_move_account(page, 1, pc,
6907 /* we uncharge from mc.from later. */
6910 putback_lru_page(page);
6911 put: /* get_mctgt_type() gets the page */
6914 case MC_TARGET_SWAP:
6916 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6918 /* we fixup refcnts and charges later. */
6926 pte_unmap_unlock(pte - 1, ptl);
6931 * We have consumed all precharges we got in can_attach().
6932 * We try charge one by one, but don't do any additional
6933 * charges to mc.to if we have failed in charge once in attach()
6936 ret = mem_cgroup_do_precharge(1);
6944 static void mem_cgroup_move_charge(struct mm_struct *mm)
6946 struct vm_area_struct *vma;
6947 struct mm_walk mem_cgroup_move_charge_walk = {
6948 .pmd_entry = mem_cgroup_move_charge_pte_range,
6952 lru_add_drain_all();
6954 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6956 * Someone who are holding the mmap_sem might be waiting in
6957 * waitq. So we cancel all extra charges, wake up all waiters,
6958 * and retry. Because we cancel precharges, we might not be able
6959 * to move enough charges, but moving charge is a best-effort
6960 * feature anyway, so it wouldn't be a big problem.
6962 __mem_cgroup_clear_mc();
6966 for (vma = mm->mmap; vma; vma = vma->vm_next)
6967 walk_page_vma(vma, &mem_cgroup_move_charge_walk);
6968 up_read(&mm->mmap_sem);
6971 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6972 struct cgroup_taskset *tset)
6974 struct task_struct *p = cgroup_taskset_first(tset);
6975 struct mm_struct *mm = get_task_mm(p);
6979 mem_cgroup_move_charge(mm);
6983 mem_cgroup_clear_mc();
6985 #else /* !CONFIG_MMU */
6986 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6987 struct cgroup_taskset *tset)
6991 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6992 struct cgroup_taskset *tset)
6995 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6996 struct cgroup_taskset *tset)
7002 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7003 * to verify sane_behavior flag on each mount attempt.
7005 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7008 * use_hierarchy is forced with sane_behavior. cgroup core
7009 * guarantees that @root doesn't have any children, so turning it
7010 * on for the root memcg is enough.
7012 if (cgroup_sane_behavior(root_css->cgroup))
7013 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7016 struct cgroup_subsys memory_cgrp_subsys = {
7017 .css_alloc = mem_cgroup_css_alloc,
7018 .css_online = mem_cgroup_css_online,
7019 .css_offline = mem_cgroup_css_offline,
7020 .css_free = mem_cgroup_css_free,
7021 .can_attach = mem_cgroup_can_attach,
7022 .cancel_attach = mem_cgroup_cancel_attach,
7023 .attach = mem_cgroup_move_task,
7024 .bind = mem_cgroup_bind,
7025 .base_cftypes = mem_cgroup_files,
7029 #ifdef CONFIG_MEMCG_SWAP
7030 static int __init enable_swap_account(char *s)
7032 if (!strcmp(s, "1"))
7033 really_do_swap_account = 1;
7034 else if (!strcmp(s, "0"))
7035 really_do_swap_account = 0;
7038 __setup("swapaccount=", enable_swap_account);
7040 static void __init memsw_file_init(void)
7042 WARN_ON(cgroup_add_cftypes(&memory_cgrp_subsys, memsw_cgroup_files));
7045 static void __init enable_swap_cgroup(void)
7047 if (!mem_cgroup_disabled() && really_do_swap_account) {
7048 do_swap_account = 1;
7054 static void __init enable_swap_cgroup(void)
7060 * subsys_initcall() for memory controller.
7062 * Some parts like hotcpu_notifier() have to be initialized from this context
7063 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7064 * everything that doesn't depend on a specific mem_cgroup structure should
7065 * be initialized from here.
7067 static int __init mem_cgroup_init(void)
7069 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7070 enable_swap_cgroup();
7071 mem_cgroup_soft_limit_tree_init();
7075 subsys_initcall(mem_cgroup_init);