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/page_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/swap_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 reclaim_iter {
147 struct mem_cgroup *position;
148 /* scan generation, increased every round-trip */
149 unsigned int generation;
153 * per-zone information in memory controller.
155 struct mem_cgroup_per_zone {
156 struct lruvec lruvec;
157 unsigned long lru_size[NR_LRU_LISTS];
159 struct reclaim_iter iter[DEF_PRIORITY + 1];
161 struct rb_node tree_node; /* RB tree node */
162 unsigned long usage_in_excess;/* Set to the value by which */
163 /* the soft limit is exceeded*/
165 struct mem_cgroup *memcg; /* Back pointer, we cannot */
166 /* use container_of */
169 struct mem_cgroup_per_node {
170 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
174 * Cgroups above their limits are maintained in a RB-Tree, independent of
175 * their hierarchy representation
178 struct mem_cgroup_tree_per_zone {
179 struct rb_root rb_root;
183 struct mem_cgroup_tree_per_node {
184 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
187 struct mem_cgroup_tree {
188 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
191 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
193 struct mem_cgroup_threshold {
194 struct eventfd_ctx *eventfd;
195 unsigned long threshold;
199 struct mem_cgroup_threshold_ary {
200 /* An array index points to threshold just below or equal to usage. */
201 int current_threshold;
202 /* Size of entries[] */
204 /* Array of thresholds */
205 struct mem_cgroup_threshold entries[0];
208 struct mem_cgroup_thresholds {
209 /* Primary thresholds array */
210 struct mem_cgroup_threshold_ary *primary;
212 * Spare threshold array.
213 * This is needed to make mem_cgroup_unregister_event() "never fail".
214 * It must be able to store at least primary->size - 1 entries.
216 struct mem_cgroup_threshold_ary *spare;
220 struct mem_cgroup_eventfd_list {
221 struct list_head list;
222 struct eventfd_ctx *eventfd;
226 * cgroup_event represents events which userspace want to receive.
228 struct mem_cgroup_event {
230 * memcg which the event belongs to.
232 struct mem_cgroup *memcg;
234 * eventfd to signal userspace about the event.
236 struct eventfd_ctx *eventfd;
238 * Each of these stored in a list by the cgroup.
240 struct list_head list;
242 * register_event() callback will be used to add new userspace
243 * waiter for changes related to this event. Use eventfd_signal()
244 * on eventfd to send notification to userspace.
246 int (*register_event)(struct mem_cgroup *memcg,
247 struct eventfd_ctx *eventfd, const char *args);
249 * unregister_event() callback will be called when userspace closes
250 * the eventfd or on cgroup removing. This callback must be set,
251 * if you want provide notification functionality.
253 void (*unregister_event)(struct mem_cgroup *memcg,
254 struct eventfd_ctx *eventfd);
256 * All fields below needed to unregister event when
257 * userspace closes eventfd.
260 wait_queue_head_t *wqh;
262 struct work_struct remove;
265 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
266 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
269 * The memory controller data structure. The memory controller controls both
270 * page cache and RSS per cgroup. We would eventually like to provide
271 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
272 * to help the administrator determine what knobs to tune.
274 * TODO: Add a water mark for the memory controller. Reclaim will begin when
275 * we hit the water mark. May be even add a low water mark, such that
276 * no reclaim occurs from a cgroup at it's low water mark, this is
277 * a feature that will be implemented much later in the future.
280 struct cgroup_subsys_state css;
282 /* Accounted resources */
283 struct page_counter memory;
284 struct page_counter memsw;
285 struct page_counter kmem;
287 unsigned long soft_limit;
289 /* vmpressure notifications */
290 struct vmpressure vmpressure;
292 /* css_online() has been completed */
296 * Should the accounting and control be hierarchical, per subtree?
302 atomic_t oom_wakeups;
305 /* OOM-Killer disable */
306 int oom_kill_disable;
308 /* protect arrays of thresholds */
309 struct mutex thresholds_lock;
311 /* thresholds for memory usage. RCU-protected */
312 struct mem_cgroup_thresholds thresholds;
314 /* thresholds for mem+swap usage. RCU-protected */
315 struct mem_cgroup_thresholds memsw_thresholds;
317 /* For oom notifier event fd */
318 struct list_head oom_notify;
321 * Should we move charges of a task when a task is moved into this
322 * mem_cgroup ? And what type of charges should we move ?
324 unsigned long move_charge_at_immigrate;
326 * set > 0 if pages under this cgroup are moving to other cgroup.
328 atomic_t moving_account;
329 /* taken only while moving_account > 0 */
330 spinlock_t move_lock;
334 struct mem_cgroup_stat_cpu __percpu *stat;
336 * used when a cpu is offlined or other synchronizations
337 * See mem_cgroup_read_stat().
339 struct mem_cgroup_stat_cpu nocpu_base;
340 spinlock_t pcp_counter_lock;
342 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
343 struct cg_proto tcp_mem;
345 #if defined(CONFIG_MEMCG_KMEM)
346 /* Index in the kmem_cache->memcg_params->memcg_caches array */
350 int last_scanned_node;
352 nodemask_t scan_nodes;
353 atomic_t numainfo_events;
354 atomic_t numainfo_updating;
357 /* List of events which userspace want to receive */
358 struct list_head event_list;
359 spinlock_t event_list_lock;
361 struct mem_cgroup_per_node *nodeinfo[0];
362 /* WARNING: nodeinfo must be the last member here */
365 #ifdef CONFIG_MEMCG_KMEM
366 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
368 return memcg->kmemcg_id >= 0;
372 /* Stuffs for move charges at task migration. */
374 * Types of charges to be moved. "move_charge_at_immitgrate" and
375 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
378 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
379 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
383 /* "mc" and its members are protected by cgroup_mutex */
384 static struct move_charge_struct {
385 spinlock_t lock; /* for from, to */
386 struct mem_cgroup *from;
387 struct mem_cgroup *to;
388 unsigned long immigrate_flags;
389 unsigned long precharge;
390 unsigned long moved_charge;
391 unsigned long moved_swap;
392 struct task_struct *moving_task; /* a task moving charges */
393 wait_queue_head_t waitq; /* a waitq for other context */
395 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
396 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
399 static bool move_anon(void)
401 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
404 static bool move_file(void)
406 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
410 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
411 * limit reclaim to prevent infinite loops, if they ever occur.
413 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
414 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
417 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
418 MEM_CGROUP_CHARGE_TYPE_ANON,
419 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
420 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
424 /* for encoding cft->private value on file */
432 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
433 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
434 #define MEMFILE_ATTR(val) ((val) & 0xffff)
435 /* Used for OOM nofiier */
436 #define OOM_CONTROL (0)
439 * The memcg_create_mutex will be held whenever a new cgroup is created.
440 * As a consequence, any change that needs to protect against new child cgroups
441 * appearing has to hold it as well.
443 static DEFINE_MUTEX(memcg_create_mutex);
445 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
447 return s ? container_of(s, struct mem_cgroup, css) : NULL;
450 /* Some nice accessors for the vmpressure. */
451 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
454 memcg = root_mem_cgroup;
455 return &memcg->vmpressure;
458 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
460 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
463 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
465 return (memcg == root_mem_cgroup);
469 * We restrict the id in the range of [1, 65535], so it can fit into
472 #define MEM_CGROUP_ID_MAX USHRT_MAX
474 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
476 return memcg->css.id;
479 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
481 struct cgroup_subsys_state *css;
483 css = css_from_id(id, &memory_cgrp_subsys);
484 return mem_cgroup_from_css(css);
487 /* Writing them here to avoid exposing memcg's inner layout */
488 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
490 void sock_update_memcg(struct sock *sk)
492 if (mem_cgroup_sockets_enabled) {
493 struct mem_cgroup *memcg;
494 struct cg_proto *cg_proto;
496 BUG_ON(!sk->sk_prot->proto_cgroup);
498 /* Socket cloning can throw us here with sk_cgrp already
499 * filled. It won't however, necessarily happen from
500 * process context. So the test for root memcg given
501 * the current task's memcg won't help us in this case.
503 * Respecting the original socket's memcg is a better
504 * decision in this case.
507 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
508 css_get(&sk->sk_cgrp->memcg->css);
513 memcg = mem_cgroup_from_task(current);
514 cg_proto = sk->sk_prot->proto_cgroup(memcg);
515 if (!mem_cgroup_is_root(memcg) &&
516 memcg_proto_active(cg_proto) &&
517 css_tryget_online(&memcg->css)) {
518 sk->sk_cgrp = cg_proto;
523 EXPORT_SYMBOL(sock_update_memcg);
525 void sock_release_memcg(struct sock *sk)
527 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
528 struct mem_cgroup *memcg;
529 WARN_ON(!sk->sk_cgrp->memcg);
530 memcg = sk->sk_cgrp->memcg;
531 css_put(&sk->sk_cgrp->memcg->css);
535 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
537 if (!memcg || mem_cgroup_is_root(memcg))
540 return &memcg->tcp_mem;
542 EXPORT_SYMBOL(tcp_proto_cgroup);
544 static void disarm_sock_keys(struct mem_cgroup *memcg)
546 if (!memcg_proto_activated(&memcg->tcp_mem))
548 static_key_slow_dec(&memcg_socket_limit_enabled);
551 static void disarm_sock_keys(struct mem_cgroup *memcg)
556 #ifdef CONFIG_MEMCG_KMEM
558 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
559 * The main reason for not using cgroup id for this:
560 * this works better in sparse environments, where we have a lot of memcgs,
561 * but only a few kmem-limited. Or also, if we have, for instance, 200
562 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
563 * 200 entry array for that.
565 * The current size of the caches array is stored in
566 * memcg_limited_groups_array_size. It will double each time we have to
569 static DEFINE_IDA(kmem_limited_groups);
570 int memcg_limited_groups_array_size;
573 * MIN_SIZE is different than 1, because we would like to avoid going through
574 * the alloc/free process all the time. In a small machine, 4 kmem-limited
575 * cgroups is a reasonable guess. In the future, it could be a parameter or
576 * tunable, but that is strictly not necessary.
578 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
579 * this constant directly from cgroup, but it is understandable that this is
580 * better kept as an internal representation in cgroup.c. In any case, the
581 * cgrp_id space is not getting any smaller, and we don't have to necessarily
582 * increase ours as well if it increases.
584 #define MEMCG_CACHES_MIN_SIZE 4
585 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
588 * A lot of the calls to the cache allocation functions are expected to be
589 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
590 * conditional to this static branch, we'll have to allow modules that does
591 * kmem_cache_alloc and the such to see this symbol as well
593 struct static_key memcg_kmem_enabled_key;
594 EXPORT_SYMBOL(memcg_kmem_enabled_key);
596 static void memcg_free_cache_id(int id);
598 static void disarm_kmem_keys(struct mem_cgroup *memcg)
600 if (memcg_kmem_is_active(memcg)) {
601 static_key_slow_dec(&memcg_kmem_enabled_key);
602 memcg_free_cache_id(memcg->kmemcg_id);
605 * This check can't live in kmem destruction function,
606 * since the charges will outlive the cgroup
608 WARN_ON(page_counter_read(&memcg->kmem));
611 static void disarm_kmem_keys(struct mem_cgroup *memcg)
614 #endif /* CONFIG_MEMCG_KMEM */
616 static void disarm_static_keys(struct mem_cgroup *memcg)
618 disarm_sock_keys(memcg);
619 disarm_kmem_keys(memcg);
622 static struct mem_cgroup_per_zone *
623 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
625 int nid = zone_to_nid(zone);
626 int zid = zone_idx(zone);
628 return &memcg->nodeinfo[nid]->zoneinfo[zid];
631 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
636 static struct mem_cgroup_per_zone *
637 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
639 int nid = page_to_nid(page);
640 int zid = page_zonenum(page);
642 return &memcg->nodeinfo[nid]->zoneinfo[zid];
645 static struct mem_cgroup_tree_per_zone *
646 soft_limit_tree_node_zone(int nid, int zid)
648 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
651 static struct mem_cgroup_tree_per_zone *
652 soft_limit_tree_from_page(struct page *page)
654 int nid = page_to_nid(page);
655 int zid = page_zonenum(page);
657 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
660 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
661 struct mem_cgroup_tree_per_zone *mctz,
662 unsigned long new_usage_in_excess)
664 struct rb_node **p = &mctz->rb_root.rb_node;
665 struct rb_node *parent = NULL;
666 struct mem_cgroup_per_zone *mz_node;
671 mz->usage_in_excess = new_usage_in_excess;
672 if (!mz->usage_in_excess)
676 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
678 if (mz->usage_in_excess < mz_node->usage_in_excess)
681 * We can't avoid mem cgroups that are over their soft
682 * limit by the same amount
684 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
687 rb_link_node(&mz->tree_node, parent, p);
688 rb_insert_color(&mz->tree_node, &mctz->rb_root);
692 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
693 struct mem_cgroup_tree_per_zone *mctz)
697 rb_erase(&mz->tree_node, &mctz->rb_root);
701 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
702 struct mem_cgroup_tree_per_zone *mctz)
706 spin_lock_irqsave(&mctz->lock, flags);
707 __mem_cgroup_remove_exceeded(mz, mctz);
708 spin_unlock_irqrestore(&mctz->lock, flags);
711 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
713 unsigned long nr_pages = page_counter_read(&memcg->memory);
714 unsigned long soft_limit = ACCESS_ONCE(memcg->soft_limit);
715 unsigned long excess = 0;
717 if (nr_pages > soft_limit)
718 excess = nr_pages - soft_limit;
723 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
725 unsigned long excess;
726 struct mem_cgroup_per_zone *mz;
727 struct mem_cgroup_tree_per_zone *mctz;
729 mctz = soft_limit_tree_from_page(page);
731 * Necessary to update all ancestors when hierarchy is used.
732 * because their event counter is not touched.
734 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
735 mz = mem_cgroup_page_zoneinfo(memcg, page);
736 excess = soft_limit_excess(memcg);
738 * We have to update the tree if mz is on RB-tree or
739 * mem is over its softlimit.
741 if (excess || mz->on_tree) {
744 spin_lock_irqsave(&mctz->lock, flags);
745 /* if on-tree, remove it */
747 __mem_cgroup_remove_exceeded(mz, mctz);
749 * Insert again. mz->usage_in_excess will be updated.
750 * If excess is 0, no tree ops.
752 __mem_cgroup_insert_exceeded(mz, mctz, excess);
753 spin_unlock_irqrestore(&mctz->lock, flags);
758 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
760 struct mem_cgroup_tree_per_zone *mctz;
761 struct mem_cgroup_per_zone *mz;
765 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
766 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
767 mctz = soft_limit_tree_node_zone(nid, zid);
768 mem_cgroup_remove_exceeded(mz, mctz);
773 static struct mem_cgroup_per_zone *
774 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
776 struct rb_node *rightmost = NULL;
777 struct mem_cgroup_per_zone *mz;
781 rightmost = rb_last(&mctz->rb_root);
783 goto done; /* Nothing to reclaim from */
785 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
787 * Remove the node now but someone else can add it back,
788 * we will to add it back at the end of reclaim to its correct
789 * position in the tree.
791 __mem_cgroup_remove_exceeded(mz, mctz);
792 if (!soft_limit_excess(mz->memcg) ||
793 !css_tryget_online(&mz->memcg->css))
799 static struct mem_cgroup_per_zone *
800 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
802 struct mem_cgroup_per_zone *mz;
804 spin_lock_irq(&mctz->lock);
805 mz = __mem_cgroup_largest_soft_limit_node(mctz);
806 spin_unlock_irq(&mctz->lock);
811 * Implementation Note: reading percpu statistics for memcg.
813 * Both of vmstat[] and percpu_counter has threshold and do periodic
814 * synchronization to implement "quick" read. There are trade-off between
815 * reading cost and precision of value. Then, we may have a chance to implement
816 * a periodic synchronizion of counter in memcg's counter.
818 * But this _read() function is used for user interface now. The user accounts
819 * memory usage by memory cgroup and he _always_ requires exact value because
820 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
821 * have to visit all online cpus and make sum. So, for now, unnecessary
822 * synchronization is not implemented. (just implemented for cpu hotplug)
824 * If there are kernel internal actions which can make use of some not-exact
825 * value, and reading all cpu value can be performance bottleneck in some
826 * common workload, threashold and synchonization as vmstat[] should be
829 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
830 enum mem_cgroup_stat_index idx)
836 for_each_online_cpu(cpu)
837 val += per_cpu(memcg->stat->count[idx], cpu);
838 #ifdef CONFIG_HOTPLUG_CPU
839 spin_lock(&memcg->pcp_counter_lock);
840 val += memcg->nocpu_base.count[idx];
841 spin_unlock(&memcg->pcp_counter_lock);
847 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
848 enum mem_cgroup_events_index idx)
850 unsigned long val = 0;
854 for_each_online_cpu(cpu)
855 val += per_cpu(memcg->stat->events[idx], cpu);
856 #ifdef CONFIG_HOTPLUG_CPU
857 spin_lock(&memcg->pcp_counter_lock);
858 val += memcg->nocpu_base.events[idx];
859 spin_unlock(&memcg->pcp_counter_lock);
865 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
870 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
871 * counted as CACHE even if it's on ANON LRU.
874 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
877 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
880 if (PageTransHuge(page))
881 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
884 /* pagein of a big page is an event. So, ignore page size */
886 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
888 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
889 nr_pages = -nr_pages; /* for event */
892 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
895 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
897 struct mem_cgroup_per_zone *mz;
899 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
900 return mz->lru_size[lru];
903 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
905 unsigned int lru_mask)
907 unsigned long nr = 0;
910 VM_BUG_ON((unsigned)nid >= nr_node_ids);
912 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
913 struct mem_cgroup_per_zone *mz;
917 if (!(BIT(lru) & lru_mask))
919 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
920 nr += mz->lru_size[lru];
926 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
927 unsigned int lru_mask)
929 unsigned long nr = 0;
932 for_each_node_state(nid, N_MEMORY)
933 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
937 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
938 enum mem_cgroup_events_target target)
940 unsigned long val, next;
942 val = __this_cpu_read(memcg->stat->nr_page_events);
943 next = __this_cpu_read(memcg->stat->targets[target]);
944 /* from time_after() in jiffies.h */
945 if ((long)next - (long)val < 0) {
947 case MEM_CGROUP_TARGET_THRESH:
948 next = val + THRESHOLDS_EVENTS_TARGET;
950 case MEM_CGROUP_TARGET_SOFTLIMIT:
951 next = val + SOFTLIMIT_EVENTS_TARGET;
953 case MEM_CGROUP_TARGET_NUMAINFO:
954 next = val + NUMAINFO_EVENTS_TARGET;
959 __this_cpu_write(memcg->stat->targets[target], next);
966 * Check events in order.
969 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
971 /* threshold event is triggered in finer grain than soft limit */
972 if (unlikely(mem_cgroup_event_ratelimit(memcg,
973 MEM_CGROUP_TARGET_THRESH))) {
975 bool do_numainfo __maybe_unused;
977 do_softlimit = mem_cgroup_event_ratelimit(memcg,
978 MEM_CGROUP_TARGET_SOFTLIMIT);
980 do_numainfo = mem_cgroup_event_ratelimit(memcg,
981 MEM_CGROUP_TARGET_NUMAINFO);
983 mem_cgroup_threshold(memcg);
984 if (unlikely(do_softlimit))
985 mem_cgroup_update_tree(memcg, page);
987 if (unlikely(do_numainfo))
988 atomic_inc(&memcg->numainfo_events);
993 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
996 * mm_update_next_owner() may clear mm->owner to NULL
997 * if it races with swapoff, page migration, etc.
998 * So this can be called with p == NULL.
1003 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1006 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1008 struct mem_cgroup *memcg = NULL;
1013 * Page cache insertions can happen withou an
1014 * actual mm context, e.g. during disk probing
1015 * on boot, loopback IO, acct() writes etc.
1018 memcg = root_mem_cgroup;
1020 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1021 if (unlikely(!memcg))
1022 memcg = root_mem_cgroup;
1024 } while (!css_tryget_online(&memcg->css));
1030 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1031 * @root: hierarchy root
1032 * @prev: previously returned memcg, NULL on first invocation
1033 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1035 * Returns references to children of the hierarchy below @root, or
1036 * @root itself, or %NULL after a full round-trip.
1038 * Caller must pass the return value in @prev on subsequent
1039 * invocations for reference counting, or use mem_cgroup_iter_break()
1040 * to cancel a hierarchy walk before the round-trip is complete.
1042 * Reclaimers can specify a zone and a priority level in @reclaim to
1043 * divide up the memcgs in the hierarchy among all concurrent
1044 * reclaimers operating on the same zone and priority.
1046 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1047 struct mem_cgroup *prev,
1048 struct mem_cgroup_reclaim_cookie *reclaim)
1050 struct reclaim_iter *uninitialized_var(iter);
1051 struct cgroup_subsys_state *css = NULL;
1052 struct mem_cgroup *memcg = NULL;
1053 struct mem_cgroup *pos = NULL;
1055 if (mem_cgroup_disabled())
1059 root = root_mem_cgroup;
1061 if (prev && !reclaim)
1064 if (!root->use_hierarchy && root != root_mem_cgroup) {
1073 struct mem_cgroup_per_zone *mz;
1075 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1076 iter = &mz->iter[reclaim->priority];
1078 if (prev && reclaim->generation != iter->generation)
1082 pos = ACCESS_ONCE(iter->position);
1084 * A racing update may change the position and
1085 * put the last reference, hence css_tryget(),
1086 * or retry to see the updated position.
1088 } while (pos && !css_tryget(&pos->css));
1095 css = css_next_descendant_pre(css, &root->css);
1098 * Reclaimers share the hierarchy walk, and a
1099 * new one might jump in right at the end of
1100 * the hierarchy - make sure they see at least
1101 * one group and restart from the beginning.
1109 * Verify the css and acquire a reference. The root
1110 * is provided by the caller, so we know it's alive
1111 * and kicking, and don't take an extra reference.
1113 memcg = mem_cgroup_from_css(css);
1115 if (css == &root->css)
1118 if (css_tryget(css)) {
1120 * Make sure the memcg is initialized:
1121 * mem_cgroup_css_online() orders the the
1122 * initialization against setting the flag.
1124 if (smp_load_acquire(&memcg->initialized))
1134 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1136 css_get(&memcg->css);
1142 * pairs with css_tryget when dereferencing iter->position
1151 reclaim->generation = iter->generation;
1157 if (prev && prev != root)
1158 css_put(&prev->css);
1164 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1165 * @root: hierarchy root
1166 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1168 void mem_cgroup_iter_break(struct mem_cgroup *root,
1169 struct mem_cgroup *prev)
1172 root = root_mem_cgroup;
1173 if (prev && prev != root)
1174 css_put(&prev->css);
1178 * Iteration constructs for visiting all cgroups (under a tree). If
1179 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1180 * be used for reference counting.
1182 #define for_each_mem_cgroup_tree(iter, root) \
1183 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1185 iter = mem_cgroup_iter(root, iter, NULL))
1187 #define for_each_mem_cgroup(iter) \
1188 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1190 iter = mem_cgroup_iter(NULL, iter, NULL))
1192 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1194 struct mem_cgroup *memcg;
1197 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1198 if (unlikely(!memcg))
1203 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1206 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1214 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1217 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1218 * @zone: zone of the wanted lruvec
1219 * @memcg: memcg of the wanted lruvec
1221 * Returns the lru list vector holding pages for the given @zone and
1222 * @mem. This can be the global zone lruvec, if the memory controller
1225 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1226 struct mem_cgroup *memcg)
1228 struct mem_cgroup_per_zone *mz;
1229 struct lruvec *lruvec;
1231 if (mem_cgroup_disabled()) {
1232 lruvec = &zone->lruvec;
1236 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1237 lruvec = &mz->lruvec;
1240 * Since a node can be onlined after the mem_cgroup was created,
1241 * we have to be prepared to initialize lruvec->zone here;
1242 * and if offlined then reonlined, we need to reinitialize it.
1244 if (unlikely(lruvec->zone != zone))
1245 lruvec->zone = zone;
1250 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1252 * @zone: zone of the page
1254 * This function is only safe when following the LRU page isolation
1255 * and putback protocol: the LRU lock must be held, and the page must
1256 * either be PageLRU() or the caller must have isolated/allocated it.
1258 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1260 struct mem_cgroup_per_zone *mz;
1261 struct mem_cgroup *memcg;
1262 struct lruvec *lruvec;
1264 if (mem_cgroup_disabled()) {
1265 lruvec = &zone->lruvec;
1269 memcg = page->mem_cgroup;
1271 * Swapcache readahead pages are added to the LRU - and
1272 * possibly migrated - before they are charged.
1275 memcg = root_mem_cgroup;
1277 mz = mem_cgroup_page_zoneinfo(memcg, page);
1278 lruvec = &mz->lruvec;
1281 * Since a node can be onlined after the mem_cgroup was created,
1282 * we have to be prepared to initialize lruvec->zone here;
1283 * and if offlined then reonlined, we need to reinitialize it.
1285 if (unlikely(lruvec->zone != zone))
1286 lruvec->zone = zone;
1291 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1292 * @lruvec: mem_cgroup per zone lru vector
1293 * @lru: index of lru list the page is sitting on
1294 * @nr_pages: positive when adding or negative when removing
1296 * This function must be called when a page is added to or removed from an
1299 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1302 struct mem_cgroup_per_zone *mz;
1303 unsigned long *lru_size;
1305 if (mem_cgroup_disabled())
1308 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1309 lru_size = mz->lru_size + lru;
1310 *lru_size += nr_pages;
1311 VM_BUG_ON((long)(*lru_size) < 0);
1314 bool mem_cgroup_is_descendant(struct mem_cgroup *memcg, struct mem_cgroup *root)
1318 if (!root->use_hierarchy)
1320 return cgroup_is_descendant(memcg->css.cgroup, root->css.cgroup);
1323 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1325 struct mem_cgroup *task_memcg;
1326 struct task_struct *p;
1329 p = find_lock_task_mm(task);
1331 task_memcg = get_mem_cgroup_from_mm(p->mm);
1335 * All threads may have already detached their mm's, but the oom
1336 * killer still needs to detect if they have already been oom
1337 * killed to prevent needlessly killing additional tasks.
1340 task_memcg = mem_cgroup_from_task(task);
1341 css_get(&task_memcg->css);
1344 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1345 css_put(&task_memcg->css);
1349 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1351 unsigned long inactive_ratio;
1352 unsigned long inactive;
1353 unsigned long active;
1356 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1357 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1359 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1361 inactive_ratio = int_sqrt(10 * gb);
1365 return inactive * inactive_ratio < active;
1368 #define mem_cgroup_from_counter(counter, member) \
1369 container_of(counter, struct mem_cgroup, member)
1372 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1373 * @memcg: the memory cgroup
1375 * Returns the maximum amount of memory @mem can be charged with, in
1378 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1380 unsigned long margin = 0;
1381 unsigned long count;
1382 unsigned long limit;
1384 count = page_counter_read(&memcg->memory);
1385 limit = ACCESS_ONCE(memcg->memory.limit);
1387 margin = limit - count;
1389 if (do_swap_account) {
1390 count = page_counter_read(&memcg->memsw);
1391 limit = ACCESS_ONCE(memcg->memsw.limit);
1393 margin = min(margin, limit - count);
1399 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1402 if (mem_cgroup_disabled() || !memcg->css.parent)
1403 return vm_swappiness;
1405 return memcg->swappiness;
1409 * A routine for checking "mem" is under move_account() or not.
1411 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1412 * moving cgroups. This is for waiting at high-memory pressure
1415 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1417 struct mem_cgroup *from;
1418 struct mem_cgroup *to;
1421 * Unlike task_move routines, we access mc.to, mc.from not under
1422 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1424 spin_lock(&mc.lock);
1430 ret = mem_cgroup_is_descendant(from, memcg) ||
1431 mem_cgroup_is_descendant(to, memcg);
1433 spin_unlock(&mc.lock);
1437 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1439 if (mc.moving_task && current != mc.moving_task) {
1440 if (mem_cgroup_under_move(memcg)) {
1442 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1443 /* moving charge context might have finished. */
1446 finish_wait(&mc.waitq, &wait);
1453 #define K(x) ((x) << (PAGE_SHIFT-10))
1455 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1456 * @memcg: The memory cgroup that went over limit
1457 * @p: Task that is going to be killed
1459 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1462 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1464 /* oom_info_lock ensures that parallel ooms do not interleave */
1465 static DEFINE_MUTEX(oom_info_lock);
1466 struct mem_cgroup *iter;
1472 mutex_lock(&oom_info_lock);
1475 pr_info("Task in ");
1476 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1477 pr_cont(" killed as a result of limit of ");
1478 pr_cont_cgroup_path(memcg->css.cgroup);
1483 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1484 K((u64)page_counter_read(&memcg->memory)),
1485 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1486 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1487 K((u64)page_counter_read(&memcg->memsw)),
1488 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1489 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1490 K((u64)page_counter_read(&memcg->kmem)),
1491 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1493 for_each_mem_cgroup_tree(iter, memcg) {
1494 pr_info("Memory cgroup stats for ");
1495 pr_cont_cgroup_path(iter->css.cgroup);
1498 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1499 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1501 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1502 K(mem_cgroup_read_stat(iter, i)));
1505 for (i = 0; i < NR_LRU_LISTS; i++)
1506 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1507 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1511 mutex_unlock(&oom_info_lock);
1515 * This function returns the number of memcg under hierarchy tree. Returns
1516 * 1(self count) if no children.
1518 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1521 struct mem_cgroup *iter;
1523 for_each_mem_cgroup_tree(iter, memcg)
1529 * Return the memory (and swap, if configured) limit for a memcg.
1531 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1533 unsigned long limit;
1535 limit = memcg->memory.limit;
1536 if (mem_cgroup_swappiness(memcg)) {
1537 unsigned long memsw_limit;
1539 memsw_limit = memcg->memsw.limit;
1540 limit = min(limit + total_swap_pages, memsw_limit);
1545 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1548 struct mem_cgroup *iter;
1549 unsigned long chosen_points = 0;
1550 unsigned long totalpages;
1551 unsigned int points = 0;
1552 struct task_struct *chosen = NULL;
1555 * If current has a pending SIGKILL or is exiting, then automatically
1556 * select it. The goal is to allow it to allocate so that it may
1557 * quickly exit and free its memory.
1559 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1560 set_thread_flag(TIF_MEMDIE);
1564 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1565 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1566 for_each_mem_cgroup_tree(iter, memcg) {
1567 struct css_task_iter it;
1568 struct task_struct *task;
1570 css_task_iter_start(&iter->css, &it);
1571 while ((task = css_task_iter_next(&it))) {
1572 switch (oom_scan_process_thread(task, totalpages, NULL,
1574 case OOM_SCAN_SELECT:
1576 put_task_struct(chosen);
1578 chosen_points = ULONG_MAX;
1579 get_task_struct(chosen);
1581 case OOM_SCAN_CONTINUE:
1583 case OOM_SCAN_ABORT:
1584 css_task_iter_end(&it);
1585 mem_cgroup_iter_break(memcg, iter);
1587 put_task_struct(chosen);
1592 points = oom_badness(task, memcg, NULL, totalpages);
1593 if (!points || points < chosen_points)
1595 /* Prefer thread group leaders for display purposes */
1596 if (points == chosen_points &&
1597 thread_group_leader(chosen))
1601 put_task_struct(chosen);
1603 chosen_points = points;
1604 get_task_struct(chosen);
1606 css_task_iter_end(&it);
1611 points = chosen_points * 1000 / totalpages;
1612 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1613 NULL, "Memory cgroup out of memory");
1616 #if MAX_NUMNODES > 1
1619 * test_mem_cgroup_node_reclaimable
1620 * @memcg: the target memcg
1621 * @nid: the node ID to be checked.
1622 * @noswap : specify true here if the user wants flle only information.
1624 * This function returns whether the specified memcg contains any
1625 * reclaimable pages on a node. Returns true if there are any reclaimable
1626 * pages in the node.
1628 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1629 int nid, bool noswap)
1631 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1633 if (noswap || !total_swap_pages)
1635 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1642 * Always updating the nodemask is not very good - even if we have an empty
1643 * list or the wrong list here, we can start from some node and traverse all
1644 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1647 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1651 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1652 * pagein/pageout changes since the last update.
1654 if (!atomic_read(&memcg->numainfo_events))
1656 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1659 /* make a nodemask where this memcg uses memory from */
1660 memcg->scan_nodes = node_states[N_MEMORY];
1662 for_each_node_mask(nid, node_states[N_MEMORY]) {
1664 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1665 node_clear(nid, memcg->scan_nodes);
1668 atomic_set(&memcg->numainfo_events, 0);
1669 atomic_set(&memcg->numainfo_updating, 0);
1673 * Selecting a node where we start reclaim from. Because what we need is just
1674 * reducing usage counter, start from anywhere is O,K. Considering
1675 * memory reclaim from current node, there are pros. and cons.
1677 * Freeing memory from current node means freeing memory from a node which
1678 * we'll use or we've used. So, it may make LRU bad. And if several threads
1679 * hit limits, it will see a contention on a node. But freeing from remote
1680 * node means more costs for memory reclaim because of memory latency.
1682 * Now, we use round-robin. Better algorithm is welcomed.
1684 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1688 mem_cgroup_may_update_nodemask(memcg);
1689 node = memcg->last_scanned_node;
1691 node = next_node(node, memcg->scan_nodes);
1692 if (node == MAX_NUMNODES)
1693 node = first_node(memcg->scan_nodes);
1695 * We call this when we hit limit, not when pages are added to LRU.
1696 * No LRU may hold pages because all pages are UNEVICTABLE or
1697 * memcg is too small and all pages are not on LRU. In that case,
1698 * we use curret node.
1700 if (unlikely(node == MAX_NUMNODES))
1701 node = numa_node_id();
1703 memcg->last_scanned_node = node;
1707 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1713 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1716 unsigned long *total_scanned)
1718 struct mem_cgroup *victim = NULL;
1721 unsigned long excess;
1722 unsigned long nr_scanned;
1723 struct mem_cgroup_reclaim_cookie reclaim = {
1728 excess = soft_limit_excess(root_memcg);
1731 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1736 * If we have not been able to reclaim
1737 * anything, it might because there are
1738 * no reclaimable pages under this hierarchy
1743 * We want to do more targeted reclaim.
1744 * excess >> 2 is not to excessive so as to
1745 * reclaim too much, nor too less that we keep
1746 * coming back to reclaim from this cgroup
1748 if (total >= (excess >> 2) ||
1749 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1754 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1756 *total_scanned += nr_scanned;
1757 if (!soft_limit_excess(root_memcg))
1760 mem_cgroup_iter_break(root_memcg, victim);
1764 #ifdef CONFIG_LOCKDEP
1765 static struct lockdep_map memcg_oom_lock_dep_map = {
1766 .name = "memcg_oom_lock",
1770 static DEFINE_SPINLOCK(memcg_oom_lock);
1773 * Check OOM-Killer is already running under our hierarchy.
1774 * If someone is running, return false.
1776 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1778 struct mem_cgroup *iter, *failed = NULL;
1780 spin_lock(&memcg_oom_lock);
1782 for_each_mem_cgroup_tree(iter, memcg) {
1783 if (iter->oom_lock) {
1785 * this subtree of our hierarchy is already locked
1786 * so we cannot give a lock.
1789 mem_cgroup_iter_break(memcg, iter);
1792 iter->oom_lock = true;
1797 * OK, we failed to lock the whole subtree so we have
1798 * to clean up what we set up to the failing subtree
1800 for_each_mem_cgroup_tree(iter, memcg) {
1801 if (iter == failed) {
1802 mem_cgroup_iter_break(memcg, iter);
1805 iter->oom_lock = false;
1808 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1810 spin_unlock(&memcg_oom_lock);
1815 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1817 struct mem_cgroup *iter;
1819 spin_lock(&memcg_oom_lock);
1820 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1821 for_each_mem_cgroup_tree(iter, memcg)
1822 iter->oom_lock = false;
1823 spin_unlock(&memcg_oom_lock);
1826 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1828 struct mem_cgroup *iter;
1830 for_each_mem_cgroup_tree(iter, memcg)
1831 atomic_inc(&iter->under_oom);
1834 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1836 struct mem_cgroup *iter;
1839 * When a new child is created while the hierarchy is under oom,
1840 * mem_cgroup_oom_lock() may not be called. We have to use
1841 * atomic_add_unless() here.
1843 for_each_mem_cgroup_tree(iter, memcg)
1844 atomic_add_unless(&iter->under_oom, -1, 0);
1847 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1849 struct oom_wait_info {
1850 struct mem_cgroup *memcg;
1854 static int memcg_oom_wake_function(wait_queue_t *wait,
1855 unsigned mode, int sync, void *arg)
1857 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1858 struct mem_cgroup *oom_wait_memcg;
1859 struct oom_wait_info *oom_wait_info;
1861 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1862 oom_wait_memcg = oom_wait_info->memcg;
1864 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1865 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1867 return autoremove_wake_function(wait, mode, sync, arg);
1870 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1872 atomic_inc(&memcg->oom_wakeups);
1873 /* for filtering, pass "memcg" as argument. */
1874 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1877 static void memcg_oom_recover(struct mem_cgroup *memcg)
1879 if (memcg && atomic_read(&memcg->under_oom))
1880 memcg_wakeup_oom(memcg);
1883 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1885 if (!current->memcg_oom.may_oom)
1888 * We are in the middle of the charge context here, so we
1889 * don't want to block when potentially sitting on a callstack
1890 * that holds all kinds of filesystem and mm locks.
1892 * Also, the caller may handle a failed allocation gracefully
1893 * (like optional page cache readahead) and so an OOM killer
1894 * invocation might not even be necessary.
1896 * That's why we don't do anything here except remember the
1897 * OOM context and then deal with it at the end of the page
1898 * fault when the stack is unwound, the locks are released,
1899 * and when we know whether the fault was overall successful.
1901 css_get(&memcg->css);
1902 current->memcg_oom.memcg = memcg;
1903 current->memcg_oom.gfp_mask = mask;
1904 current->memcg_oom.order = order;
1908 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1909 * @handle: actually kill/wait or just clean up the OOM state
1911 * This has to be called at the end of a page fault if the memcg OOM
1912 * handler was enabled.
1914 * Memcg supports userspace OOM handling where failed allocations must
1915 * sleep on a waitqueue until the userspace task resolves the
1916 * situation. Sleeping directly in the charge context with all kinds
1917 * of locks held is not a good idea, instead we remember an OOM state
1918 * in the task and mem_cgroup_oom_synchronize() has to be called at
1919 * the end of the page fault to complete the OOM handling.
1921 * Returns %true if an ongoing memcg OOM situation was detected and
1922 * completed, %false otherwise.
1924 bool mem_cgroup_oom_synchronize(bool handle)
1926 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1927 struct oom_wait_info owait;
1930 /* OOM is global, do not handle */
1937 owait.memcg = memcg;
1938 owait.wait.flags = 0;
1939 owait.wait.func = memcg_oom_wake_function;
1940 owait.wait.private = current;
1941 INIT_LIST_HEAD(&owait.wait.task_list);
1943 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1944 mem_cgroup_mark_under_oom(memcg);
1946 locked = mem_cgroup_oom_trylock(memcg);
1949 mem_cgroup_oom_notify(memcg);
1951 if (locked && !memcg->oom_kill_disable) {
1952 mem_cgroup_unmark_under_oom(memcg);
1953 finish_wait(&memcg_oom_waitq, &owait.wait);
1954 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1955 current->memcg_oom.order);
1958 mem_cgroup_unmark_under_oom(memcg);
1959 finish_wait(&memcg_oom_waitq, &owait.wait);
1963 mem_cgroup_oom_unlock(memcg);
1965 * There is no guarantee that an OOM-lock contender
1966 * sees the wakeups triggered by the OOM kill
1967 * uncharges. Wake any sleepers explicitely.
1969 memcg_oom_recover(memcg);
1972 current->memcg_oom.memcg = NULL;
1973 css_put(&memcg->css);
1978 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1979 * @page: page that is going to change accounted state
1980 * @locked: &memcg->move_lock slowpath was taken
1981 * @flags: IRQ-state flags for &memcg->move_lock
1983 * This function must mark the beginning of an accounted page state
1984 * change to prevent double accounting when the page is concurrently
1985 * being moved to another memcg:
1987 * memcg = mem_cgroup_begin_page_stat(page, &locked, &flags);
1988 * if (TestClearPageState(page))
1989 * mem_cgroup_update_page_stat(memcg, state, -1);
1990 * mem_cgroup_end_page_stat(memcg, locked, flags);
1992 * The RCU lock is held throughout the transaction. The fast path can
1993 * get away without acquiring the memcg->move_lock (@locked is false)
1994 * because page moving starts with an RCU grace period.
1996 * The RCU lock also protects the memcg from being freed when the page
1997 * state that is going to change is the only thing preventing the page
1998 * from being uncharged. E.g. end-writeback clearing PageWriteback(),
1999 * which allows migration to go ahead and uncharge the page before the
2000 * account transaction might be complete.
2002 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page,
2004 unsigned long *flags)
2006 struct mem_cgroup *memcg;
2010 if (mem_cgroup_disabled())
2013 memcg = page->mem_cgroup;
2014 if (unlikely(!memcg))
2018 if (atomic_read(&memcg->moving_account) <= 0)
2021 spin_lock_irqsave(&memcg->move_lock, *flags);
2022 if (memcg != page->mem_cgroup) {
2023 spin_unlock_irqrestore(&memcg->move_lock, *flags);
2032 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2033 * @memcg: the memcg that was accounted against
2034 * @locked: value received from mem_cgroup_begin_page_stat()
2035 * @flags: value received from mem_cgroup_begin_page_stat()
2037 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg, bool *locked,
2038 unsigned long *flags)
2040 if (memcg && *locked)
2041 spin_unlock_irqrestore(&memcg->move_lock, *flags);
2047 * mem_cgroup_update_page_stat - update page state statistics
2048 * @memcg: memcg to account against
2049 * @idx: page state item to account
2050 * @val: number of pages (positive or negative)
2052 * See mem_cgroup_begin_page_stat() for locking requirements.
2054 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2055 enum mem_cgroup_stat_index idx, int val)
2057 VM_BUG_ON(!rcu_read_lock_held());
2060 this_cpu_add(memcg->stat->count[idx], val);
2064 * size of first charge trial. "32" comes from vmscan.c's magic value.
2065 * TODO: maybe necessary to use big numbers in big irons.
2067 #define CHARGE_BATCH 32U
2068 struct memcg_stock_pcp {
2069 struct mem_cgroup *cached; /* this never be root cgroup */
2070 unsigned int nr_pages;
2071 struct work_struct work;
2072 unsigned long flags;
2073 #define FLUSHING_CACHED_CHARGE 0
2075 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2076 static DEFINE_MUTEX(percpu_charge_mutex);
2079 * consume_stock: Try to consume stocked charge on this cpu.
2080 * @memcg: memcg to consume from.
2081 * @nr_pages: how many pages to charge.
2083 * The charges will only happen if @memcg matches the current cpu's memcg
2084 * stock, and at least @nr_pages are available in that stock. Failure to
2085 * service an allocation will refill the stock.
2087 * returns true if successful, false otherwise.
2089 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2091 struct memcg_stock_pcp *stock;
2094 if (nr_pages > CHARGE_BATCH)
2097 stock = &get_cpu_var(memcg_stock);
2098 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2099 stock->nr_pages -= nr_pages;
2102 put_cpu_var(memcg_stock);
2107 * Returns stocks cached in percpu and reset cached information.
2109 static void drain_stock(struct memcg_stock_pcp *stock)
2111 struct mem_cgroup *old = stock->cached;
2113 if (stock->nr_pages) {
2114 page_counter_uncharge(&old->memory, stock->nr_pages);
2115 if (do_swap_account)
2116 page_counter_uncharge(&old->memsw, stock->nr_pages);
2117 css_put_many(&old->css, stock->nr_pages);
2118 stock->nr_pages = 0;
2120 stock->cached = NULL;
2124 * This must be called under preempt disabled or must be called by
2125 * a thread which is pinned to local cpu.
2127 static void drain_local_stock(struct work_struct *dummy)
2129 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2131 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2134 static void __init memcg_stock_init(void)
2138 for_each_possible_cpu(cpu) {
2139 struct memcg_stock_pcp *stock =
2140 &per_cpu(memcg_stock, cpu);
2141 INIT_WORK(&stock->work, drain_local_stock);
2146 * Cache charges(val) to local per_cpu area.
2147 * This will be consumed by consume_stock() function, later.
2149 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2151 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2153 if (stock->cached != memcg) { /* reset if necessary */
2155 stock->cached = memcg;
2157 stock->nr_pages += nr_pages;
2158 put_cpu_var(memcg_stock);
2162 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2163 * of the hierarchy under it.
2165 static void drain_all_stock(struct mem_cgroup *root_memcg)
2169 /* If someone's already draining, avoid adding running more workers. */
2170 if (!mutex_trylock(&percpu_charge_mutex))
2172 /* Notify other cpus that system-wide "drain" is running */
2175 for_each_online_cpu(cpu) {
2176 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2177 struct mem_cgroup *memcg;
2179 memcg = stock->cached;
2180 if (!memcg || !stock->nr_pages)
2182 if (!mem_cgroup_is_descendant(memcg, root_memcg))
2184 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2186 drain_local_stock(&stock->work);
2188 schedule_work_on(cpu, &stock->work);
2193 mutex_unlock(&percpu_charge_mutex);
2197 * This function drains percpu counter value from DEAD cpu and
2198 * move it to local cpu. Note that this function can be preempted.
2200 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2204 spin_lock(&memcg->pcp_counter_lock);
2205 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2206 long x = per_cpu(memcg->stat->count[i], cpu);
2208 per_cpu(memcg->stat->count[i], cpu) = 0;
2209 memcg->nocpu_base.count[i] += x;
2211 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2212 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2214 per_cpu(memcg->stat->events[i], cpu) = 0;
2215 memcg->nocpu_base.events[i] += x;
2217 spin_unlock(&memcg->pcp_counter_lock);
2220 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2221 unsigned long action,
2224 int cpu = (unsigned long)hcpu;
2225 struct memcg_stock_pcp *stock;
2226 struct mem_cgroup *iter;
2228 if (action == CPU_ONLINE)
2231 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2234 for_each_mem_cgroup(iter)
2235 mem_cgroup_drain_pcp_counter(iter, cpu);
2237 stock = &per_cpu(memcg_stock, cpu);
2242 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2243 unsigned int nr_pages)
2245 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2246 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2247 struct mem_cgroup *mem_over_limit;
2248 struct page_counter *counter;
2249 unsigned long nr_reclaimed;
2250 bool may_swap = true;
2251 bool drained = false;
2254 if (mem_cgroup_is_root(memcg))
2257 if (consume_stock(memcg, nr_pages))
2260 if (!do_swap_account ||
2261 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2262 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2264 if (do_swap_account)
2265 page_counter_uncharge(&memcg->memsw, batch);
2266 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2268 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2272 if (batch > nr_pages) {
2278 * Unlike in global OOM situations, memcg is not in a physical
2279 * memory shortage. Allow dying and OOM-killed tasks to
2280 * bypass the last charges so that they can exit quickly and
2281 * free their memory.
2283 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2284 fatal_signal_pending(current) ||
2285 current->flags & PF_EXITING))
2288 if (unlikely(task_in_memcg_oom(current)))
2291 if (!(gfp_mask & __GFP_WAIT))
2294 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2295 gfp_mask, may_swap);
2297 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2301 drain_all_stock(mem_over_limit);
2306 if (gfp_mask & __GFP_NORETRY)
2309 * Even though the limit is exceeded at this point, reclaim
2310 * may have been able to free some pages. Retry the charge
2311 * before killing the task.
2313 * Only for regular pages, though: huge pages are rather
2314 * unlikely to succeed so close to the limit, and we fall back
2315 * to regular pages anyway in case of failure.
2317 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2320 * At task move, charge accounts can be doubly counted. So, it's
2321 * better to wait until the end of task_move if something is going on.
2323 if (mem_cgroup_wait_acct_move(mem_over_limit))
2329 if (gfp_mask & __GFP_NOFAIL)
2332 if (fatal_signal_pending(current))
2335 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2337 if (!(gfp_mask & __GFP_NOFAIL))
2343 css_get_many(&memcg->css, batch);
2344 if (batch > nr_pages)
2345 refill_stock(memcg, batch - nr_pages);
2350 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2352 if (mem_cgroup_is_root(memcg))
2355 page_counter_uncharge(&memcg->memory, nr_pages);
2356 if (do_swap_account)
2357 page_counter_uncharge(&memcg->memsw, nr_pages);
2359 css_put_many(&memcg->css, nr_pages);
2363 * A helper function to get mem_cgroup from ID. must be called under
2364 * rcu_read_lock(). The caller is responsible for calling
2365 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2366 * refcnt from swap can be called against removed memcg.)
2368 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2370 /* ID 0 is unused ID */
2373 return mem_cgroup_from_id(id);
2377 * try_get_mem_cgroup_from_page - look up page's memcg association
2380 * Look up, get a css reference, and return the memcg that owns @page.
2382 * The page must be locked to prevent racing with swap-in and page
2383 * cache charges. If coming from an unlocked page table, the caller
2384 * must ensure the page is on the LRU or this can race with charging.
2386 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2388 struct mem_cgroup *memcg;
2392 VM_BUG_ON_PAGE(!PageLocked(page), page);
2394 memcg = page->mem_cgroup;
2396 if (!css_tryget_online(&memcg->css))
2398 } else if (PageSwapCache(page)) {
2399 ent.val = page_private(page);
2400 id = lookup_swap_cgroup_id(ent);
2402 memcg = mem_cgroup_lookup(id);
2403 if (memcg && !css_tryget_online(&memcg->css))
2410 static void lock_page_lru(struct page *page, int *isolated)
2412 struct zone *zone = page_zone(page);
2414 spin_lock_irq(&zone->lru_lock);
2415 if (PageLRU(page)) {
2416 struct lruvec *lruvec;
2418 lruvec = mem_cgroup_page_lruvec(page, zone);
2420 del_page_from_lru_list(page, lruvec, page_lru(page));
2426 static void unlock_page_lru(struct page *page, int isolated)
2428 struct zone *zone = page_zone(page);
2431 struct lruvec *lruvec;
2433 lruvec = mem_cgroup_page_lruvec(page, zone);
2434 VM_BUG_ON_PAGE(PageLRU(page), page);
2436 add_page_to_lru_list(page, lruvec, page_lru(page));
2438 spin_unlock_irq(&zone->lru_lock);
2441 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2446 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2449 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2450 * may already be on some other mem_cgroup's LRU. Take care of it.
2453 lock_page_lru(page, &isolated);
2456 * Nobody should be changing or seriously looking at
2457 * page->mem_cgroup at this point:
2459 * - the page is uncharged
2461 * - the page is off-LRU
2463 * - an anonymous fault has exclusive page access, except for
2464 * a locked page table
2466 * - a page cache insertion, a swapin fault, or a migration
2467 * have the page locked
2469 page->mem_cgroup = memcg;
2472 unlock_page_lru(page, isolated);
2475 #ifdef CONFIG_MEMCG_KMEM
2476 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2477 unsigned long nr_pages)
2479 struct page_counter *counter;
2482 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2486 ret = try_charge(memcg, gfp, nr_pages);
2487 if (ret == -EINTR) {
2489 * try_charge() chose to bypass to root due to OOM kill or
2490 * fatal signal. Since our only options are to either fail
2491 * the allocation or charge it to this cgroup, do it as a
2492 * temporary condition. But we can't fail. From a kmem/slab
2493 * perspective, the cache has already been selected, by
2494 * mem_cgroup_kmem_get_cache(), so it is too late to change
2497 * This condition will only trigger if the task entered
2498 * memcg_charge_kmem in a sane state, but was OOM-killed
2499 * during try_charge() above. Tasks that were already dying
2500 * when the allocation triggers should have been already
2501 * directed to the root cgroup in memcontrol.h
2503 page_counter_charge(&memcg->memory, nr_pages);
2504 if (do_swap_account)
2505 page_counter_charge(&memcg->memsw, nr_pages);
2506 css_get_many(&memcg->css, nr_pages);
2509 page_counter_uncharge(&memcg->kmem, nr_pages);
2514 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2516 page_counter_uncharge(&memcg->memory, nr_pages);
2517 if (do_swap_account)
2518 page_counter_uncharge(&memcg->memsw, nr_pages);
2520 page_counter_uncharge(&memcg->kmem, nr_pages);
2522 css_put_many(&memcg->css, nr_pages);
2526 * helper for acessing a memcg's index. It will be used as an index in the
2527 * child cache array in kmem_cache, and also to derive its name. This function
2528 * will return -1 when this is not a kmem-limited memcg.
2530 int memcg_cache_id(struct mem_cgroup *memcg)
2532 return memcg ? memcg->kmemcg_id : -1;
2535 static int memcg_alloc_cache_id(void)
2540 id = ida_simple_get(&kmem_limited_groups,
2541 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2545 if (id < memcg_limited_groups_array_size)
2549 * There's no space for the new id in memcg_caches arrays,
2550 * so we have to grow them.
2553 size = 2 * (id + 1);
2554 if (size < MEMCG_CACHES_MIN_SIZE)
2555 size = MEMCG_CACHES_MIN_SIZE;
2556 else if (size > MEMCG_CACHES_MAX_SIZE)
2557 size = MEMCG_CACHES_MAX_SIZE;
2559 err = memcg_update_all_caches(size);
2561 ida_simple_remove(&kmem_limited_groups, id);
2567 static void memcg_free_cache_id(int id)
2569 ida_simple_remove(&kmem_limited_groups, id);
2573 * We should update the current array size iff all caches updates succeed. This
2574 * can only be done from the slab side. The slab mutex needs to be held when
2577 void memcg_update_array_size(int num)
2579 memcg_limited_groups_array_size = num;
2582 struct memcg_kmem_cache_create_work {
2583 struct mem_cgroup *memcg;
2584 struct kmem_cache *cachep;
2585 struct work_struct work;
2588 static void memcg_kmem_cache_create_func(struct work_struct *w)
2590 struct memcg_kmem_cache_create_work *cw =
2591 container_of(w, struct memcg_kmem_cache_create_work, work);
2592 struct mem_cgroup *memcg = cw->memcg;
2593 struct kmem_cache *cachep = cw->cachep;
2595 memcg_create_kmem_cache(memcg, cachep);
2597 css_put(&memcg->css);
2602 * Enqueue the creation of a per-memcg kmem_cache.
2604 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2605 struct kmem_cache *cachep)
2607 struct memcg_kmem_cache_create_work *cw;
2609 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2613 css_get(&memcg->css);
2616 cw->cachep = cachep;
2617 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2619 schedule_work(&cw->work);
2622 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2623 struct kmem_cache *cachep)
2626 * We need to stop accounting when we kmalloc, because if the
2627 * corresponding kmalloc cache is not yet created, the first allocation
2628 * in __memcg_schedule_kmem_cache_create will recurse.
2630 * However, it is better to enclose the whole function. Depending on
2631 * the debugging options enabled, INIT_WORK(), for instance, can
2632 * trigger an allocation. This too, will make us recurse. Because at
2633 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2634 * the safest choice is to do it like this, wrapping the whole function.
2636 current->memcg_kmem_skip_account = 1;
2637 __memcg_schedule_kmem_cache_create(memcg, cachep);
2638 current->memcg_kmem_skip_account = 0;
2642 * Return the kmem_cache we're supposed to use for a slab allocation.
2643 * We try to use the current memcg's version of the cache.
2645 * If the cache does not exist yet, if we are the first user of it,
2646 * we either create it immediately, if possible, or create it asynchronously
2648 * In the latter case, we will let the current allocation go through with
2649 * the original cache.
2651 * Can't be called in interrupt context or from kernel threads.
2652 * This function needs to be called with rcu_read_lock() held.
2654 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2656 struct mem_cgroup *memcg;
2657 struct kmem_cache *memcg_cachep;
2659 VM_BUG_ON(!cachep->memcg_params);
2660 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
2662 if (current->memcg_kmem_skip_account)
2665 memcg = get_mem_cgroup_from_mm(current->mm);
2666 if (!memcg_kmem_is_active(memcg))
2669 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
2670 if (likely(memcg_cachep))
2671 return memcg_cachep;
2674 * If we are in a safe context (can wait, and not in interrupt
2675 * context), we could be be predictable and return right away.
2676 * This would guarantee that the allocation being performed
2677 * already belongs in the new cache.
2679 * However, there are some clashes that can arrive from locking.
2680 * For instance, because we acquire the slab_mutex while doing
2681 * memcg_create_kmem_cache, this means no further allocation
2682 * could happen with the slab_mutex held. So it's better to
2685 memcg_schedule_kmem_cache_create(memcg, cachep);
2687 css_put(&memcg->css);
2691 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2693 if (!is_root_cache(cachep))
2694 css_put(&cachep->memcg_params->memcg->css);
2698 * We need to verify if the allocation against current->mm->owner's memcg is
2699 * possible for the given order. But the page is not allocated yet, so we'll
2700 * need a further commit step to do the final arrangements.
2702 * It is possible for the task to switch cgroups in this mean time, so at
2703 * commit time, we can't rely on task conversion any longer. We'll then use
2704 * the handle argument to return to the caller which cgroup we should commit
2705 * against. We could also return the memcg directly and avoid the pointer
2706 * passing, but a boolean return value gives better semantics considering
2707 * the compiled-out case as well.
2709 * Returning true means the allocation is possible.
2712 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2714 struct mem_cgroup *memcg;
2719 memcg = get_mem_cgroup_from_mm(current->mm);
2721 if (!memcg_kmem_is_active(memcg)) {
2722 css_put(&memcg->css);
2726 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2730 css_put(&memcg->css);
2734 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2737 VM_BUG_ON(mem_cgroup_is_root(memcg));
2739 /* The page allocation failed. Revert */
2741 memcg_uncharge_kmem(memcg, 1 << order);
2744 page->mem_cgroup = memcg;
2747 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2749 struct mem_cgroup *memcg = page->mem_cgroup;
2754 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2756 memcg_uncharge_kmem(memcg, 1 << order);
2757 page->mem_cgroup = NULL;
2759 #endif /* CONFIG_MEMCG_KMEM */
2761 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2764 * Because tail pages are not marked as "used", set it. We're under
2765 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2766 * charge/uncharge will be never happen and move_account() is done under
2767 * compound_lock(), so we don't have to take care of races.
2769 void mem_cgroup_split_huge_fixup(struct page *head)
2773 if (mem_cgroup_disabled())
2776 for (i = 1; i < HPAGE_PMD_NR; i++)
2777 head[i].mem_cgroup = head->mem_cgroup;
2779 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2782 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2785 * mem_cgroup_move_account - move account of the page
2787 * @nr_pages: number of regular pages (>1 for huge pages)
2788 * @from: mem_cgroup which the page is moved from.
2789 * @to: mem_cgroup which the page is moved to. @from != @to.
2791 * The caller must confirm following.
2792 * - page is not on LRU (isolate_page() is useful.)
2793 * - compound_lock is held when nr_pages > 1
2795 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2798 static int mem_cgroup_move_account(struct page *page,
2799 unsigned int nr_pages,
2800 struct mem_cgroup *from,
2801 struct mem_cgroup *to)
2803 unsigned long flags;
2806 VM_BUG_ON(from == to);
2807 VM_BUG_ON_PAGE(PageLRU(page), page);
2809 * The page is isolated from LRU. So, collapse function
2810 * will not handle this page. But page splitting can happen.
2811 * Do this check under compound_page_lock(). The caller should
2815 if (nr_pages > 1 && !PageTransHuge(page))
2819 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
2820 * of its source page while we change it: page migration takes
2821 * both pages off the LRU, but page cache replacement doesn't.
2823 if (!trylock_page(page))
2827 if (page->mem_cgroup != from)
2830 spin_lock_irqsave(&from->move_lock, flags);
2832 if (!PageAnon(page) && page_mapped(page)) {
2833 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2835 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2839 if (PageWriteback(page)) {
2840 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2842 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2847 * It is safe to change page->mem_cgroup here because the page
2848 * is referenced, charged, and isolated - we can't race with
2849 * uncharging, charging, migration, or LRU putback.
2852 /* caller should have done css_get */
2853 page->mem_cgroup = to;
2854 spin_unlock_irqrestore(&from->move_lock, flags);
2858 local_irq_disable();
2859 mem_cgroup_charge_statistics(to, page, nr_pages);
2860 memcg_check_events(to, page);
2861 mem_cgroup_charge_statistics(from, page, -nr_pages);
2862 memcg_check_events(from, page);
2870 #ifdef CONFIG_MEMCG_SWAP
2871 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2874 int val = (charge) ? 1 : -1;
2875 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2879 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2880 * @entry: swap entry to be moved
2881 * @from: mem_cgroup which the entry is moved from
2882 * @to: mem_cgroup which the entry is moved to
2884 * It succeeds only when the swap_cgroup's record for this entry is the same
2885 * as the mem_cgroup's id of @from.
2887 * Returns 0 on success, -EINVAL on failure.
2889 * The caller must have charged to @to, IOW, called page_counter_charge() about
2890 * both res and memsw, and called css_get().
2892 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2893 struct mem_cgroup *from, struct mem_cgroup *to)
2895 unsigned short old_id, new_id;
2897 old_id = mem_cgroup_id(from);
2898 new_id = mem_cgroup_id(to);
2900 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2901 mem_cgroup_swap_statistics(from, false);
2902 mem_cgroup_swap_statistics(to, true);
2908 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2909 struct mem_cgroup *from, struct mem_cgroup *to)
2915 static DEFINE_MUTEX(memcg_limit_mutex);
2917 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2918 unsigned long limit)
2920 unsigned long curusage;
2921 unsigned long oldusage;
2922 bool enlarge = false;
2927 * For keeping hierarchical_reclaim simple, how long we should retry
2928 * is depends on callers. We set our retry-count to be function
2929 * of # of children which we should visit in this loop.
2931 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2932 mem_cgroup_count_children(memcg);
2934 oldusage = page_counter_read(&memcg->memory);
2937 if (signal_pending(current)) {
2942 mutex_lock(&memcg_limit_mutex);
2943 if (limit > memcg->memsw.limit) {
2944 mutex_unlock(&memcg_limit_mutex);
2948 if (limit > memcg->memory.limit)
2950 ret = page_counter_limit(&memcg->memory, limit);
2951 mutex_unlock(&memcg_limit_mutex);
2956 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2958 curusage = page_counter_read(&memcg->memory);
2959 /* Usage is reduced ? */
2960 if (curusage >= oldusage)
2963 oldusage = curusage;
2964 } while (retry_count);
2966 if (!ret && enlarge)
2967 memcg_oom_recover(memcg);
2972 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2973 unsigned long limit)
2975 unsigned long curusage;
2976 unsigned long oldusage;
2977 bool enlarge = false;
2981 /* see mem_cgroup_resize_res_limit */
2982 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2983 mem_cgroup_count_children(memcg);
2985 oldusage = page_counter_read(&memcg->memsw);
2988 if (signal_pending(current)) {
2993 mutex_lock(&memcg_limit_mutex);
2994 if (limit < memcg->memory.limit) {
2995 mutex_unlock(&memcg_limit_mutex);
2999 if (limit > memcg->memsw.limit)
3001 ret = page_counter_limit(&memcg->memsw, limit);
3002 mutex_unlock(&memcg_limit_mutex);
3007 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3009 curusage = page_counter_read(&memcg->memsw);
3010 /* Usage is reduced ? */
3011 if (curusage >= oldusage)
3014 oldusage = curusage;
3015 } while (retry_count);
3017 if (!ret && enlarge)
3018 memcg_oom_recover(memcg);
3023 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3025 unsigned long *total_scanned)
3027 unsigned long nr_reclaimed = 0;
3028 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3029 unsigned long reclaimed;
3031 struct mem_cgroup_tree_per_zone *mctz;
3032 unsigned long excess;
3033 unsigned long nr_scanned;
3038 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3040 * This loop can run a while, specially if mem_cgroup's continuously
3041 * keep exceeding their soft limit and putting the system under
3048 mz = mem_cgroup_largest_soft_limit_node(mctz);
3053 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3054 gfp_mask, &nr_scanned);
3055 nr_reclaimed += reclaimed;
3056 *total_scanned += nr_scanned;
3057 spin_lock_irq(&mctz->lock);
3058 __mem_cgroup_remove_exceeded(mz, mctz);
3061 * If we failed to reclaim anything from this memory cgroup
3062 * it is time to move on to the next cgroup
3066 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3068 excess = soft_limit_excess(mz->memcg);
3070 * One school of thought says that we should not add
3071 * back the node to the tree if reclaim returns 0.
3072 * But our reclaim could return 0, simply because due
3073 * to priority we are exposing a smaller subset of
3074 * memory to reclaim from. Consider this as a longer
3077 /* If excess == 0, no tree ops */
3078 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3079 spin_unlock_irq(&mctz->lock);
3080 css_put(&mz->memcg->css);
3083 * Could not reclaim anything and there are no more
3084 * mem cgroups to try or we seem to be looping without
3085 * reclaiming anything.
3087 if (!nr_reclaimed &&
3089 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3091 } while (!nr_reclaimed);
3093 css_put(&next_mz->memcg->css);
3094 return nr_reclaimed;
3098 * Test whether @memcg has children, dead or alive. Note that this
3099 * function doesn't care whether @memcg has use_hierarchy enabled and
3100 * returns %true if there are child csses according to the cgroup
3101 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3103 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3108 * The lock does not prevent addition or deletion of children, but
3109 * it prevents a new child from being initialized based on this
3110 * parent in css_online(), so it's enough to decide whether
3111 * hierarchically inherited attributes can still be changed or not.
3113 lockdep_assert_held(&memcg_create_mutex);
3116 ret = css_next_child(NULL, &memcg->css);
3122 * Reclaims as many pages from the given memcg as possible and moves
3123 * the rest to the parent.
3125 * Caller is responsible for holding css reference for memcg.
3127 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3129 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3131 /* we call try-to-free pages for make this cgroup empty */
3132 lru_add_drain_all();
3133 /* try to free all pages in this cgroup */
3134 while (nr_retries && page_counter_read(&memcg->memory)) {
3137 if (signal_pending(current))
3140 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3144 /* maybe some writeback is necessary */
3145 congestion_wait(BLK_RW_ASYNC, HZ/10);
3153 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3154 char *buf, size_t nbytes,
3157 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3159 if (mem_cgroup_is_root(memcg))
3161 return mem_cgroup_force_empty(memcg) ?: nbytes;
3164 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3167 return mem_cgroup_from_css(css)->use_hierarchy;
3170 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3171 struct cftype *cft, u64 val)
3174 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3175 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3177 mutex_lock(&memcg_create_mutex);
3179 if (memcg->use_hierarchy == val)
3183 * If parent's use_hierarchy is set, we can't make any modifications
3184 * in the child subtrees. If it is unset, then the change can
3185 * occur, provided the current cgroup has no children.
3187 * For the root cgroup, parent_mem is NULL, we allow value to be
3188 * set if there are no children.
3190 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3191 (val == 1 || val == 0)) {
3192 if (!memcg_has_children(memcg))
3193 memcg->use_hierarchy = val;
3200 mutex_unlock(&memcg_create_mutex);
3205 static unsigned long tree_stat(struct mem_cgroup *memcg,
3206 enum mem_cgroup_stat_index idx)
3208 struct mem_cgroup *iter;
3211 /* Per-cpu values can be negative, use a signed accumulator */
3212 for_each_mem_cgroup_tree(iter, memcg)
3213 val += mem_cgroup_read_stat(iter, idx);
3215 if (val < 0) /* race ? */
3220 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3224 if (mem_cgroup_is_root(memcg)) {
3225 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3226 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3228 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3231 val = page_counter_read(&memcg->memory);
3233 val = page_counter_read(&memcg->memsw);
3235 return val << PAGE_SHIFT;
3246 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3249 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3250 struct page_counter *counter;
3252 switch (MEMFILE_TYPE(cft->private)) {
3254 counter = &memcg->memory;
3257 counter = &memcg->memsw;
3260 counter = &memcg->kmem;
3266 switch (MEMFILE_ATTR(cft->private)) {
3268 if (counter == &memcg->memory)
3269 return mem_cgroup_usage(memcg, false);
3270 if (counter == &memcg->memsw)
3271 return mem_cgroup_usage(memcg, true);
3272 return (u64)page_counter_read(counter) * PAGE_SIZE;
3274 return (u64)counter->limit * PAGE_SIZE;
3276 return (u64)counter->watermark * PAGE_SIZE;
3278 return counter->failcnt;
3279 case RES_SOFT_LIMIT:
3280 return (u64)memcg->soft_limit * PAGE_SIZE;
3286 #ifdef CONFIG_MEMCG_KMEM
3287 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3288 unsigned long nr_pages)
3293 if (memcg_kmem_is_active(memcg))
3297 * For simplicity, we won't allow this to be disabled. It also can't
3298 * be changed if the cgroup has children already, or if tasks had
3301 * If tasks join before we set the limit, a person looking at
3302 * kmem.usage_in_bytes will have no way to determine when it took
3303 * place, which makes the value quite meaningless.
3305 * After it first became limited, changes in the value of the limit are
3306 * of course permitted.
3308 mutex_lock(&memcg_create_mutex);
3309 if (cgroup_has_tasks(memcg->css.cgroup) ||
3310 (memcg->use_hierarchy && memcg_has_children(memcg)))
3312 mutex_unlock(&memcg_create_mutex);
3316 memcg_id = memcg_alloc_cache_id();
3323 * We couldn't have accounted to this cgroup, because it hasn't got
3324 * activated yet, so this should succeed.
3326 err = page_counter_limit(&memcg->kmem, nr_pages);
3329 static_key_slow_inc(&memcg_kmem_enabled_key);
3331 * A memory cgroup is considered kmem-active as soon as it gets
3332 * kmemcg_id. Setting the id after enabling static branching will
3333 * guarantee no one starts accounting before all call sites are
3336 memcg->kmemcg_id = memcg_id;
3341 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3342 unsigned long limit)
3346 mutex_lock(&memcg_limit_mutex);
3347 if (!memcg_kmem_is_active(memcg))
3348 ret = memcg_activate_kmem(memcg, limit);
3350 ret = page_counter_limit(&memcg->kmem, limit);
3351 mutex_unlock(&memcg_limit_mutex);
3355 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3358 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3363 mutex_lock(&memcg_limit_mutex);
3365 * If the parent cgroup is not kmem-active now, it cannot be activated
3366 * after this point, because it has at least one child already.
3368 if (memcg_kmem_is_active(parent))
3369 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3370 mutex_unlock(&memcg_limit_mutex);
3374 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3375 unsigned long limit)
3379 #endif /* CONFIG_MEMCG_KMEM */
3382 * The user of this function is...
3385 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3386 char *buf, size_t nbytes, loff_t off)
3388 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3389 unsigned long nr_pages;
3392 buf = strstrip(buf);
3393 ret = page_counter_memparse(buf, &nr_pages);
3397 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3399 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3403 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3405 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3408 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3411 ret = memcg_update_kmem_limit(memcg, nr_pages);
3415 case RES_SOFT_LIMIT:
3416 memcg->soft_limit = nr_pages;
3420 return ret ?: nbytes;
3423 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3424 size_t nbytes, loff_t off)
3426 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3427 struct page_counter *counter;
3429 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3431 counter = &memcg->memory;
3434 counter = &memcg->memsw;
3437 counter = &memcg->kmem;
3443 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3445 page_counter_reset_watermark(counter);
3448 counter->failcnt = 0;
3457 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3460 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3464 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3465 struct cftype *cft, u64 val)
3467 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3469 if (val >= (1 << NR_MOVE_TYPE))
3473 * No kind of locking is needed in here, because ->can_attach() will
3474 * check this value once in the beginning of the process, and then carry
3475 * on with stale data. This means that changes to this value will only
3476 * affect task migrations starting after the change.
3478 memcg->move_charge_at_immigrate = val;
3482 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3483 struct cftype *cft, u64 val)
3490 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3494 unsigned int lru_mask;
3497 static const struct numa_stat stats[] = {
3498 { "total", LRU_ALL },
3499 { "file", LRU_ALL_FILE },
3500 { "anon", LRU_ALL_ANON },
3501 { "unevictable", BIT(LRU_UNEVICTABLE) },
3503 const struct numa_stat *stat;
3506 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3508 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3509 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3510 seq_printf(m, "%s=%lu", stat->name, nr);
3511 for_each_node_state(nid, N_MEMORY) {
3512 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3514 seq_printf(m, " N%d=%lu", nid, nr);
3519 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3520 struct mem_cgroup *iter;
3523 for_each_mem_cgroup_tree(iter, memcg)
3524 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3525 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3526 for_each_node_state(nid, N_MEMORY) {
3528 for_each_mem_cgroup_tree(iter, memcg)
3529 nr += mem_cgroup_node_nr_lru_pages(
3530 iter, nid, stat->lru_mask);
3531 seq_printf(m, " N%d=%lu", nid, nr);
3538 #endif /* CONFIG_NUMA */
3540 static int memcg_stat_show(struct seq_file *m, void *v)
3542 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3543 unsigned long memory, memsw;
3544 struct mem_cgroup *mi;
3547 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3549 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3550 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3552 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3553 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3556 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3557 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3558 mem_cgroup_read_events(memcg, i));
3560 for (i = 0; i < NR_LRU_LISTS; i++)
3561 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3562 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3564 /* Hierarchical information */
3565 memory = memsw = PAGE_COUNTER_MAX;
3566 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3567 memory = min(memory, mi->memory.limit);
3568 memsw = min(memsw, mi->memsw.limit);
3570 seq_printf(m, "hierarchical_memory_limit %llu\n",
3571 (u64)memory * PAGE_SIZE);
3572 if (do_swap_account)
3573 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3574 (u64)memsw * PAGE_SIZE);
3576 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3579 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3581 for_each_mem_cgroup_tree(mi, memcg)
3582 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3583 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3586 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3587 unsigned long long val = 0;
3589 for_each_mem_cgroup_tree(mi, memcg)
3590 val += mem_cgroup_read_events(mi, i);
3591 seq_printf(m, "total_%s %llu\n",
3592 mem_cgroup_events_names[i], val);
3595 for (i = 0; i < NR_LRU_LISTS; i++) {
3596 unsigned long long val = 0;
3598 for_each_mem_cgroup_tree(mi, memcg)
3599 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3600 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3603 #ifdef CONFIG_DEBUG_VM
3606 struct mem_cgroup_per_zone *mz;
3607 struct zone_reclaim_stat *rstat;
3608 unsigned long recent_rotated[2] = {0, 0};
3609 unsigned long recent_scanned[2] = {0, 0};
3611 for_each_online_node(nid)
3612 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3613 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3614 rstat = &mz->lruvec.reclaim_stat;
3616 recent_rotated[0] += rstat->recent_rotated[0];
3617 recent_rotated[1] += rstat->recent_rotated[1];
3618 recent_scanned[0] += rstat->recent_scanned[0];
3619 recent_scanned[1] += rstat->recent_scanned[1];
3621 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3622 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3623 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3624 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3631 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3634 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3636 return mem_cgroup_swappiness(memcg);
3639 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3640 struct cftype *cft, u64 val)
3642 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3648 memcg->swappiness = val;
3650 vm_swappiness = val;
3655 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3657 struct mem_cgroup_threshold_ary *t;
3658 unsigned long usage;
3663 t = rcu_dereference(memcg->thresholds.primary);
3665 t = rcu_dereference(memcg->memsw_thresholds.primary);
3670 usage = mem_cgroup_usage(memcg, swap);
3673 * current_threshold points to threshold just below or equal to usage.
3674 * If it's not true, a threshold was crossed after last
3675 * call of __mem_cgroup_threshold().
3677 i = t->current_threshold;
3680 * Iterate backward over array of thresholds starting from
3681 * current_threshold and check if a threshold is crossed.
3682 * If none of thresholds below usage is crossed, we read
3683 * only one element of the array here.
3685 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3686 eventfd_signal(t->entries[i].eventfd, 1);
3688 /* i = current_threshold + 1 */
3692 * Iterate forward over array of thresholds starting from
3693 * current_threshold+1 and check if a threshold is crossed.
3694 * If none of thresholds above usage is crossed, we read
3695 * only one element of the array here.
3697 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3698 eventfd_signal(t->entries[i].eventfd, 1);
3700 /* Update current_threshold */
3701 t->current_threshold = i - 1;
3706 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3709 __mem_cgroup_threshold(memcg, false);
3710 if (do_swap_account)
3711 __mem_cgroup_threshold(memcg, true);
3713 memcg = parent_mem_cgroup(memcg);
3717 static int compare_thresholds(const void *a, const void *b)
3719 const struct mem_cgroup_threshold *_a = a;
3720 const struct mem_cgroup_threshold *_b = b;
3722 if (_a->threshold > _b->threshold)
3725 if (_a->threshold < _b->threshold)
3731 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3733 struct mem_cgroup_eventfd_list *ev;
3735 spin_lock(&memcg_oom_lock);
3737 list_for_each_entry(ev, &memcg->oom_notify, list)
3738 eventfd_signal(ev->eventfd, 1);
3740 spin_unlock(&memcg_oom_lock);
3744 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3746 struct mem_cgroup *iter;
3748 for_each_mem_cgroup_tree(iter, memcg)
3749 mem_cgroup_oom_notify_cb(iter);
3752 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3753 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3755 struct mem_cgroup_thresholds *thresholds;
3756 struct mem_cgroup_threshold_ary *new;
3757 unsigned long threshold;
3758 unsigned long usage;
3761 ret = page_counter_memparse(args, &threshold);
3765 mutex_lock(&memcg->thresholds_lock);
3768 thresholds = &memcg->thresholds;
3769 usage = mem_cgroup_usage(memcg, false);
3770 } else if (type == _MEMSWAP) {
3771 thresholds = &memcg->memsw_thresholds;
3772 usage = mem_cgroup_usage(memcg, true);
3776 /* Check if a threshold crossed before adding a new one */
3777 if (thresholds->primary)
3778 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3780 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3782 /* Allocate memory for new array of thresholds */
3783 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3791 /* Copy thresholds (if any) to new array */
3792 if (thresholds->primary) {
3793 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3794 sizeof(struct mem_cgroup_threshold));
3797 /* Add new threshold */
3798 new->entries[size - 1].eventfd = eventfd;
3799 new->entries[size - 1].threshold = threshold;
3801 /* Sort thresholds. Registering of new threshold isn't time-critical */
3802 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3803 compare_thresholds, NULL);
3805 /* Find current threshold */
3806 new->current_threshold = -1;
3807 for (i = 0; i < size; i++) {
3808 if (new->entries[i].threshold <= usage) {
3810 * new->current_threshold will not be used until
3811 * rcu_assign_pointer(), so it's safe to increment
3814 ++new->current_threshold;
3819 /* Free old spare buffer and save old primary buffer as spare */
3820 kfree(thresholds->spare);
3821 thresholds->spare = thresholds->primary;
3823 rcu_assign_pointer(thresholds->primary, new);
3825 /* To be sure that nobody uses thresholds */
3829 mutex_unlock(&memcg->thresholds_lock);
3834 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3835 struct eventfd_ctx *eventfd, const char *args)
3837 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3840 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3841 struct eventfd_ctx *eventfd, const char *args)
3843 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3846 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3847 struct eventfd_ctx *eventfd, enum res_type type)
3849 struct mem_cgroup_thresholds *thresholds;
3850 struct mem_cgroup_threshold_ary *new;
3851 unsigned long usage;
3854 mutex_lock(&memcg->thresholds_lock);
3857 thresholds = &memcg->thresholds;
3858 usage = mem_cgroup_usage(memcg, false);
3859 } else if (type == _MEMSWAP) {
3860 thresholds = &memcg->memsw_thresholds;
3861 usage = mem_cgroup_usage(memcg, true);
3865 if (!thresholds->primary)
3868 /* Check if a threshold crossed before removing */
3869 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3871 /* Calculate new number of threshold */
3873 for (i = 0; i < thresholds->primary->size; i++) {
3874 if (thresholds->primary->entries[i].eventfd != eventfd)
3878 new = thresholds->spare;
3880 /* Set thresholds array to NULL if we don't have thresholds */
3889 /* Copy thresholds and find current threshold */
3890 new->current_threshold = -1;
3891 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3892 if (thresholds->primary->entries[i].eventfd == eventfd)
3895 new->entries[j] = thresholds->primary->entries[i];
3896 if (new->entries[j].threshold <= usage) {
3898 * new->current_threshold will not be used
3899 * until rcu_assign_pointer(), so it's safe to increment
3902 ++new->current_threshold;
3908 /* Swap primary and spare array */
3909 thresholds->spare = thresholds->primary;
3910 /* If all events are unregistered, free the spare array */
3912 kfree(thresholds->spare);
3913 thresholds->spare = NULL;
3916 rcu_assign_pointer(thresholds->primary, new);
3918 /* To be sure that nobody uses thresholds */
3921 mutex_unlock(&memcg->thresholds_lock);
3924 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3925 struct eventfd_ctx *eventfd)
3927 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3930 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3931 struct eventfd_ctx *eventfd)
3933 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3936 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3937 struct eventfd_ctx *eventfd, const char *args)
3939 struct mem_cgroup_eventfd_list *event;
3941 event = kmalloc(sizeof(*event), GFP_KERNEL);
3945 spin_lock(&memcg_oom_lock);
3947 event->eventfd = eventfd;
3948 list_add(&event->list, &memcg->oom_notify);
3950 /* already in OOM ? */
3951 if (atomic_read(&memcg->under_oom))
3952 eventfd_signal(eventfd, 1);
3953 spin_unlock(&memcg_oom_lock);
3958 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3959 struct eventfd_ctx *eventfd)
3961 struct mem_cgroup_eventfd_list *ev, *tmp;
3963 spin_lock(&memcg_oom_lock);
3965 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3966 if (ev->eventfd == eventfd) {
3967 list_del(&ev->list);
3972 spin_unlock(&memcg_oom_lock);
3975 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3977 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3979 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3980 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
3984 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3985 struct cftype *cft, u64 val)
3987 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3989 /* cannot set to root cgroup and only 0 and 1 are allowed */
3990 if (!css->parent || !((val == 0) || (val == 1)))
3993 memcg->oom_kill_disable = val;
3995 memcg_oom_recover(memcg);
4000 #ifdef CONFIG_MEMCG_KMEM
4001 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4005 ret = memcg_propagate_kmem(memcg);
4009 return mem_cgroup_sockets_init(memcg, ss);
4012 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4014 memcg_destroy_kmem_caches(memcg);
4015 mem_cgroup_sockets_destroy(memcg);
4018 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4023 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4029 * DO NOT USE IN NEW FILES.
4031 * "cgroup.event_control" implementation.
4033 * This is way over-engineered. It tries to support fully configurable
4034 * events for each user. Such level of flexibility is completely
4035 * unnecessary especially in the light of the planned unified hierarchy.
4037 * Please deprecate this and replace with something simpler if at all
4042 * Unregister event and free resources.
4044 * Gets called from workqueue.
4046 static void memcg_event_remove(struct work_struct *work)
4048 struct mem_cgroup_event *event =
4049 container_of(work, struct mem_cgroup_event, remove);
4050 struct mem_cgroup *memcg = event->memcg;
4052 remove_wait_queue(event->wqh, &event->wait);
4054 event->unregister_event(memcg, event->eventfd);
4056 /* Notify userspace the event is going away. */
4057 eventfd_signal(event->eventfd, 1);
4059 eventfd_ctx_put(event->eventfd);
4061 css_put(&memcg->css);
4065 * Gets called on POLLHUP on eventfd when user closes it.
4067 * Called with wqh->lock held and interrupts disabled.
4069 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4070 int sync, void *key)
4072 struct mem_cgroup_event *event =
4073 container_of(wait, struct mem_cgroup_event, wait);
4074 struct mem_cgroup *memcg = event->memcg;
4075 unsigned long flags = (unsigned long)key;
4077 if (flags & POLLHUP) {
4079 * If the event has been detached at cgroup removal, we
4080 * can simply return knowing the other side will cleanup
4083 * We can't race against event freeing since the other
4084 * side will require wqh->lock via remove_wait_queue(),
4087 spin_lock(&memcg->event_list_lock);
4088 if (!list_empty(&event->list)) {
4089 list_del_init(&event->list);
4091 * We are in atomic context, but cgroup_event_remove()
4092 * may sleep, so we have to call it in workqueue.
4094 schedule_work(&event->remove);
4096 spin_unlock(&memcg->event_list_lock);
4102 static void memcg_event_ptable_queue_proc(struct file *file,
4103 wait_queue_head_t *wqh, poll_table *pt)
4105 struct mem_cgroup_event *event =
4106 container_of(pt, struct mem_cgroup_event, pt);
4109 add_wait_queue(wqh, &event->wait);
4113 * DO NOT USE IN NEW FILES.
4115 * Parse input and register new cgroup event handler.
4117 * Input must be in format '<event_fd> <control_fd> <args>'.
4118 * Interpretation of args is defined by control file implementation.
4120 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4121 char *buf, size_t nbytes, loff_t off)
4123 struct cgroup_subsys_state *css = of_css(of);
4124 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4125 struct mem_cgroup_event *event;
4126 struct cgroup_subsys_state *cfile_css;
4127 unsigned int efd, cfd;
4134 buf = strstrip(buf);
4136 efd = simple_strtoul(buf, &endp, 10);
4141 cfd = simple_strtoul(buf, &endp, 10);
4142 if ((*endp != ' ') && (*endp != '\0'))
4146 event = kzalloc(sizeof(*event), GFP_KERNEL);
4150 event->memcg = memcg;
4151 INIT_LIST_HEAD(&event->list);
4152 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4153 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4154 INIT_WORK(&event->remove, memcg_event_remove);
4162 event->eventfd = eventfd_ctx_fileget(efile.file);
4163 if (IS_ERR(event->eventfd)) {
4164 ret = PTR_ERR(event->eventfd);
4171 goto out_put_eventfd;
4174 /* the process need read permission on control file */
4175 /* AV: shouldn't we check that it's been opened for read instead? */
4176 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4181 * Determine the event callbacks and set them in @event. This used
4182 * to be done via struct cftype but cgroup core no longer knows
4183 * about these events. The following is crude but the whole thing
4184 * is for compatibility anyway.
4186 * DO NOT ADD NEW FILES.
4188 name = cfile.file->f_path.dentry->d_name.name;
4190 if (!strcmp(name, "memory.usage_in_bytes")) {
4191 event->register_event = mem_cgroup_usage_register_event;
4192 event->unregister_event = mem_cgroup_usage_unregister_event;
4193 } else if (!strcmp(name, "memory.oom_control")) {
4194 event->register_event = mem_cgroup_oom_register_event;
4195 event->unregister_event = mem_cgroup_oom_unregister_event;
4196 } else if (!strcmp(name, "memory.pressure_level")) {
4197 event->register_event = vmpressure_register_event;
4198 event->unregister_event = vmpressure_unregister_event;
4199 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4200 event->register_event = memsw_cgroup_usage_register_event;
4201 event->unregister_event = memsw_cgroup_usage_unregister_event;
4208 * Verify @cfile should belong to @css. Also, remaining events are
4209 * automatically removed on cgroup destruction but the removal is
4210 * asynchronous, so take an extra ref on @css.
4212 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4213 &memory_cgrp_subsys);
4215 if (IS_ERR(cfile_css))
4217 if (cfile_css != css) {
4222 ret = event->register_event(memcg, event->eventfd, buf);
4226 efile.file->f_op->poll(efile.file, &event->pt);
4228 spin_lock(&memcg->event_list_lock);
4229 list_add(&event->list, &memcg->event_list);
4230 spin_unlock(&memcg->event_list_lock);
4242 eventfd_ctx_put(event->eventfd);
4251 static struct cftype mem_cgroup_files[] = {
4253 .name = "usage_in_bytes",
4254 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4255 .read_u64 = mem_cgroup_read_u64,
4258 .name = "max_usage_in_bytes",
4259 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4260 .write = mem_cgroup_reset,
4261 .read_u64 = mem_cgroup_read_u64,
4264 .name = "limit_in_bytes",
4265 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4266 .write = mem_cgroup_write,
4267 .read_u64 = mem_cgroup_read_u64,
4270 .name = "soft_limit_in_bytes",
4271 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4272 .write = mem_cgroup_write,
4273 .read_u64 = mem_cgroup_read_u64,
4277 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4278 .write = mem_cgroup_reset,
4279 .read_u64 = mem_cgroup_read_u64,
4283 .seq_show = memcg_stat_show,
4286 .name = "force_empty",
4287 .write = mem_cgroup_force_empty_write,
4290 .name = "use_hierarchy",
4291 .write_u64 = mem_cgroup_hierarchy_write,
4292 .read_u64 = mem_cgroup_hierarchy_read,
4295 .name = "cgroup.event_control", /* XXX: for compat */
4296 .write = memcg_write_event_control,
4297 .flags = CFTYPE_NO_PREFIX,
4301 .name = "swappiness",
4302 .read_u64 = mem_cgroup_swappiness_read,
4303 .write_u64 = mem_cgroup_swappiness_write,
4306 .name = "move_charge_at_immigrate",
4307 .read_u64 = mem_cgroup_move_charge_read,
4308 .write_u64 = mem_cgroup_move_charge_write,
4311 .name = "oom_control",
4312 .seq_show = mem_cgroup_oom_control_read,
4313 .write_u64 = mem_cgroup_oom_control_write,
4314 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4317 .name = "pressure_level",
4321 .name = "numa_stat",
4322 .seq_show = memcg_numa_stat_show,
4325 #ifdef CONFIG_MEMCG_KMEM
4327 .name = "kmem.limit_in_bytes",
4328 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4329 .write = mem_cgroup_write,
4330 .read_u64 = mem_cgroup_read_u64,
4333 .name = "kmem.usage_in_bytes",
4334 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4335 .read_u64 = mem_cgroup_read_u64,
4338 .name = "kmem.failcnt",
4339 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4340 .write = mem_cgroup_reset,
4341 .read_u64 = mem_cgroup_read_u64,
4344 .name = "kmem.max_usage_in_bytes",
4345 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4346 .write = mem_cgroup_reset,
4347 .read_u64 = mem_cgroup_read_u64,
4349 #ifdef CONFIG_SLABINFO
4351 .name = "kmem.slabinfo",
4352 .seq_start = slab_start,
4353 .seq_next = slab_next,
4354 .seq_stop = slab_stop,
4355 .seq_show = memcg_slab_show,
4359 { }, /* terminate */
4362 #ifdef CONFIG_MEMCG_SWAP
4363 static struct cftype memsw_cgroup_files[] = {
4365 .name = "memsw.usage_in_bytes",
4366 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4367 .read_u64 = mem_cgroup_read_u64,
4370 .name = "memsw.max_usage_in_bytes",
4371 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4372 .write = mem_cgroup_reset,
4373 .read_u64 = mem_cgroup_read_u64,
4376 .name = "memsw.limit_in_bytes",
4377 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4378 .write = mem_cgroup_write,
4379 .read_u64 = mem_cgroup_read_u64,
4382 .name = "memsw.failcnt",
4383 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4384 .write = mem_cgroup_reset,
4385 .read_u64 = mem_cgroup_read_u64,
4387 { }, /* terminate */
4390 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4392 struct mem_cgroup_per_node *pn;
4393 struct mem_cgroup_per_zone *mz;
4394 int zone, tmp = node;
4396 * This routine is called against possible nodes.
4397 * But it's BUG to call kmalloc() against offline node.
4399 * TODO: this routine can waste much memory for nodes which will
4400 * never be onlined. It's better to use memory hotplug callback
4403 if (!node_state(node, N_NORMAL_MEMORY))
4405 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4409 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4410 mz = &pn->zoneinfo[zone];
4411 lruvec_init(&mz->lruvec);
4412 mz->usage_in_excess = 0;
4413 mz->on_tree = false;
4416 memcg->nodeinfo[node] = pn;
4420 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4422 kfree(memcg->nodeinfo[node]);
4425 static struct mem_cgroup *mem_cgroup_alloc(void)
4427 struct mem_cgroup *memcg;
4430 size = sizeof(struct mem_cgroup);
4431 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4433 memcg = kzalloc(size, GFP_KERNEL);
4437 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4440 spin_lock_init(&memcg->pcp_counter_lock);
4449 * At destroying mem_cgroup, references from swap_cgroup can remain.
4450 * (scanning all at force_empty is too costly...)
4452 * Instead of clearing all references at force_empty, we remember
4453 * the number of reference from swap_cgroup and free mem_cgroup when
4454 * it goes down to 0.
4456 * Removal of cgroup itself succeeds regardless of refs from swap.
4459 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4463 mem_cgroup_remove_from_trees(memcg);
4466 free_mem_cgroup_per_zone_info(memcg, node);
4468 free_percpu(memcg->stat);
4470 disarm_static_keys(memcg);
4475 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4477 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4479 if (!memcg->memory.parent)
4481 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4483 EXPORT_SYMBOL(parent_mem_cgroup);
4485 static void __init mem_cgroup_soft_limit_tree_init(void)
4487 struct mem_cgroup_tree_per_node *rtpn;
4488 struct mem_cgroup_tree_per_zone *rtpz;
4489 int tmp, node, zone;
4491 for_each_node(node) {
4493 if (!node_state(node, N_NORMAL_MEMORY))
4495 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4498 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4500 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4501 rtpz = &rtpn->rb_tree_per_zone[zone];
4502 rtpz->rb_root = RB_ROOT;
4503 spin_lock_init(&rtpz->lock);
4508 static struct cgroup_subsys_state * __ref
4509 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4511 struct mem_cgroup *memcg;
4512 long error = -ENOMEM;
4515 memcg = mem_cgroup_alloc();
4517 return ERR_PTR(error);
4520 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4524 if (parent_css == NULL) {
4525 root_mem_cgroup = memcg;
4526 page_counter_init(&memcg->memory, NULL);
4527 memcg->soft_limit = PAGE_COUNTER_MAX;
4528 page_counter_init(&memcg->memsw, NULL);
4529 page_counter_init(&memcg->kmem, NULL);
4532 memcg->last_scanned_node = MAX_NUMNODES;
4533 INIT_LIST_HEAD(&memcg->oom_notify);
4534 memcg->move_charge_at_immigrate = 0;
4535 mutex_init(&memcg->thresholds_lock);
4536 spin_lock_init(&memcg->move_lock);
4537 vmpressure_init(&memcg->vmpressure);
4538 INIT_LIST_HEAD(&memcg->event_list);
4539 spin_lock_init(&memcg->event_list_lock);
4540 #ifdef CONFIG_MEMCG_KMEM
4541 memcg->kmemcg_id = -1;
4547 __mem_cgroup_free(memcg);
4548 return ERR_PTR(error);
4552 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4554 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4555 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4558 if (css->id > MEM_CGROUP_ID_MAX)
4564 mutex_lock(&memcg_create_mutex);
4566 memcg->use_hierarchy = parent->use_hierarchy;
4567 memcg->oom_kill_disable = parent->oom_kill_disable;
4568 memcg->swappiness = mem_cgroup_swappiness(parent);
4570 if (parent->use_hierarchy) {
4571 page_counter_init(&memcg->memory, &parent->memory);
4572 memcg->soft_limit = PAGE_COUNTER_MAX;
4573 page_counter_init(&memcg->memsw, &parent->memsw);
4574 page_counter_init(&memcg->kmem, &parent->kmem);
4577 * No need to take a reference to the parent because cgroup
4578 * core guarantees its existence.
4581 page_counter_init(&memcg->memory, NULL);
4582 memcg->soft_limit = PAGE_COUNTER_MAX;
4583 page_counter_init(&memcg->memsw, NULL);
4584 page_counter_init(&memcg->kmem, NULL);
4586 * Deeper hierachy with use_hierarchy == false doesn't make
4587 * much sense so let cgroup subsystem know about this
4588 * unfortunate state in our controller.
4590 if (parent != root_mem_cgroup)
4591 memory_cgrp_subsys.broken_hierarchy = true;
4593 mutex_unlock(&memcg_create_mutex);
4595 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4600 * Make sure the memcg is initialized: mem_cgroup_iter()
4601 * orders reading memcg->initialized against its callers
4602 * reading the memcg members.
4604 smp_store_release(&memcg->initialized, 1);
4609 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4611 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4612 struct mem_cgroup_event *event, *tmp;
4615 * Unregister events and notify userspace.
4616 * Notify userspace about cgroup removing only after rmdir of cgroup
4617 * directory to avoid race between userspace and kernelspace.
4619 spin_lock(&memcg->event_list_lock);
4620 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4621 list_del_init(&event->list);
4622 schedule_work(&event->remove);
4624 spin_unlock(&memcg->event_list_lock);
4626 vmpressure_cleanup(&memcg->vmpressure);
4629 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4631 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4633 memcg_destroy_kmem(memcg);
4634 __mem_cgroup_free(memcg);
4638 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4639 * @css: the target css
4641 * Reset the states of the mem_cgroup associated with @css. This is
4642 * invoked when the userland requests disabling on the default hierarchy
4643 * but the memcg is pinned through dependency. The memcg should stop
4644 * applying policies and should revert to the vanilla state as it may be
4645 * made visible again.
4647 * The current implementation only resets the essential configurations.
4648 * This needs to be expanded to cover all the visible parts.
4650 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4652 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4654 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4655 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4656 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4657 memcg->soft_limit = PAGE_COUNTER_MAX;
4661 /* Handlers for move charge at task migration. */
4662 static int mem_cgroup_do_precharge(unsigned long count)
4666 /* Try a single bulk charge without reclaim first */
4667 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4669 mc.precharge += count;
4672 if (ret == -EINTR) {
4673 cancel_charge(root_mem_cgroup, count);
4677 /* Try charges one by one with reclaim */
4679 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4681 * In case of failure, any residual charges against
4682 * mc.to will be dropped by mem_cgroup_clear_mc()
4683 * later on. However, cancel any charges that are
4684 * bypassed to root right away or they'll be lost.
4687 cancel_charge(root_mem_cgroup, 1);
4697 * get_mctgt_type - get target type of moving charge
4698 * @vma: the vma the pte to be checked belongs
4699 * @addr: the address corresponding to the pte to be checked
4700 * @ptent: the pte to be checked
4701 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4704 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4705 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4706 * move charge. if @target is not NULL, the page is stored in target->page
4707 * with extra refcnt got(Callers should handle it).
4708 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4709 * target for charge migration. if @target is not NULL, the entry is stored
4712 * Called with pte lock held.
4719 enum mc_target_type {
4725 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4726 unsigned long addr, pte_t ptent)
4728 struct page *page = vm_normal_page(vma, addr, ptent);
4730 if (!page || !page_mapped(page))
4732 if (PageAnon(page)) {
4733 /* we don't move shared anon */
4736 } else if (!move_file())
4737 /* we ignore mapcount for file pages */
4739 if (!get_page_unless_zero(page))
4746 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4747 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4749 struct page *page = NULL;
4750 swp_entry_t ent = pte_to_swp_entry(ptent);
4752 if (!move_anon() || non_swap_entry(ent))
4755 * Because lookup_swap_cache() updates some statistics counter,
4756 * we call find_get_page() with swapper_space directly.
4758 page = find_get_page(swap_address_space(ent), ent.val);
4759 if (do_swap_account)
4760 entry->val = ent.val;
4765 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4766 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4772 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4773 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4775 struct page *page = NULL;
4776 struct address_space *mapping;
4779 if (!vma->vm_file) /* anonymous vma */
4784 mapping = vma->vm_file->f_mapping;
4785 pgoff = linear_page_index(vma, addr);
4787 /* page is moved even if it's not RSS of this task(page-faulted). */
4789 /* shmem/tmpfs may report page out on swap: account for that too. */
4790 if (shmem_mapping(mapping)) {
4791 page = find_get_entry(mapping, pgoff);
4792 if (radix_tree_exceptional_entry(page)) {
4793 swp_entry_t swp = radix_to_swp_entry(page);
4794 if (do_swap_account)
4796 page = find_get_page(swap_address_space(swp), swp.val);
4799 page = find_get_page(mapping, pgoff);
4801 page = find_get_page(mapping, pgoff);
4806 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4807 unsigned long addr, pte_t ptent, union mc_target *target)
4809 struct page *page = NULL;
4810 enum mc_target_type ret = MC_TARGET_NONE;
4811 swp_entry_t ent = { .val = 0 };
4813 if (pte_present(ptent))
4814 page = mc_handle_present_pte(vma, addr, ptent);
4815 else if (is_swap_pte(ptent))
4816 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4817 else if (pte_none(ptent))
4818 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4820 if (!page && !ent.val)
4824 * Do only loose check w/o serialization.
4825 * mem_cgroup_move_account() checks the page is valid or
4826 * not under LRU exclusion.
4828 if (page->mem_cgroup == mc.from) {
4829 ret = MC_TARGET_PAGE;
4831 target->page = page;
4833 if (!ret || !target)
4836 /* There is a swap entry and a page doesn't exist or isn't charged */
4837 if (ent.val && !ret &&
4838 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4839 ret = MC_TARGET_SWAP;
4846 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4848 * We don't consider swapping or file mapped pages because THP does not
4849 * support them for now.
4850 * Caller should make sure that pmd_trans_huge(pmd) is true.
4852 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4853 unsigned long addr, pmd_t pmd, union mc_target *target)
4855 struct page *page = NULL;
4856 enum mc_target_type ret = MC_TARGET_NONE;
4858 page = pmd_page(pmd);
4859 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4862 if (page->mem_cgroup == mc.from) {
4863 ret = MC_TARGET_PAGE;
4866 target->page = page;
4872 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4873 unsigned long addr, pmd_t pmd, union mc_target *target)
4875 return MC_TARGET_NONE;
4879 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4880 unsigned long addr, unsigned long end,
4881 struct mm_walk *walk)
4883 struct vm_area_struct *vma = walk->private;
4887 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
4888 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
4889 mc.precharge += HPAGE_PMD_NR;
4894 if (pmd_trans_unstable(pmd))
4896 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4897 for (; addr != end; pte++, addr += PAGE_SIZE)
4898 if (get_mctgt_type(vma, addr, *pte, NULL))
4899 mc.precharge++; /* increment precharge temporarily */
4900 pte_unmap_unlock(pte - 1, ptl);
4906 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4908 unsigned long precharge;
4909 struct vm_area_struct *vma;
4911 down_read(&mm->mmap_sem);
4912 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4913 struct mm_walk mem_cgroup_count_precharge_walk = {
4914 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4918 if (is_vm_hugetlb_page(vma))
4920 walk_page_range(vma->vm_start, vma->vm_end,
4921 &mem_cgroup_count_precharge_walk);
4923 up_read(&mm->mmap_sem);
4925 precharge = mc.precharge;
4931 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4933 unsigned long precharge = mem_cgroup_count_precharge(mm);
4935 VM_BUG_ON(mc.moving_task);
4936 mc.moving_task = current;
4937 return mem_cgroup_do_precharge(precharge);
4940 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4941 static void __mem_cgroup_clear_mc(void)
4943 struct mem_cgroup *from = mc.from;
4944 struct mem_cgroup *to = mc.to;
4946 /* we must uncharge all the leftover precharges from mc.to */
4948 cancel_charge(mc.to, mc.precharge);
4952 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4953 * we must uncharge here.
4955 if (mc.moved_charge) {
4956 cancel_charge(mc.from, mc.moved_charge);
4957 mc.moved_charge = 0;
4959 /* we must fixup refcnts and charges */
4960 if (mc.moved_swap) {
4961 /* uncharge swap account from the old cgroup */
4962 if (!mem_cgroup_is_root(mc.from))
4963 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
4966 * we charged both to->memory and to->memsw, so we
4967 * should uncharge to->memory.
4969 if (!mem_cgroup_is_root(mc.to))
4970 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
4972 css_put_many(&mc.from->css, mc.moved_swap);
4974 /* we've already done css_get(mc.to) */
4977 memcg_oom_recover(from);
4978 memcg_oom_recover(to);
4979 wake_up_all(&mc.waitq);
4982 static void mem_cgroup_clear_mc(void)
4985 * we must clear moving_task before waking up waiters at the end of
4988 mc.moving_task = NULL;
4989 __mem_cgroup_clear_mc();
4990 spin_lock(&mc.lock);
4993 spin_unlock(&mc.lock);
4996 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
4997 struct cgroup_taskset *tset)
4999 struct task_struct *p = cgroup_taskset_first(tset);
5001 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5002 unsigned long move_charge_at_immigrate;
5005 * We are now commited to this value whatever it is. Changes in this
5006 * tunable will only affect upcoming migrations, not the current one.
5007 * So we need to save it, and keep it going.
5009 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
5010 if (move_charge_at_immigrate) {
5011 struct mm_struct *mm;
5012 struct mem_cgroup *from = mem_cgroup_from_task(p);
5014 VM_BUG_ON(from == memcg);
5016 mm = get_task_mm(p);
5019 /* We move charges only when we move a owner of the mm */
5020 if (mm->owner == p) {
5023 VM_BUG_ON(mc.precharge);
5024 VM_BUG_ON(mc.moved_charge);
5025 VM_BUG_ON(mc.moved_swap);
5027 spin_lock(&mc.lock);
5030 mc.immigrate_flags = move_charge_at_immigrate;
5031 spin_unlock(&mc.lock);
5032 /* We set mc.moving_task later */
5034 ret = mem_cgroup_precharge_mc(mm);
5036 mem_cgroup_clear_mc();
5043 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5044 struct cgroup_taskset *tset)
5047 mem_cgroup_clear_mc();
5050 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5051 unsigned long addr, unsigned long end,
5052 struct mm_walk *walk)
5055 struct vm_area_struct *vma = walk->private;
5058 enum mc_target_type target_type;
5059 union mc_target target;
5063 * We don't take compound_lock() here but no race with splitting thp
5065 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5066 * under splitting, which means there's no concurrent thp split,
5067 * - if another thread runs into split_huge_page() just after we
5068 * entered this if-block, the thread must wait for page table lock
5069 * to be unlocked in __split_huge_page_splitting(), where the main
5070 * part of thp split is not executed yet.
5072 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5073 if (mc.precharge < HPAGE_PMD_NR) {
5077 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5078 if (target_type == MC_TARGET_PAGE) {
5080 if (!isolate_lru_page(page)) {
5081 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5083 mc.precharge -= HPAGE_PMD_NR;
5084 mc.moved_charge += HPAGE_PMD_NR;
5086 putback_lru_page(page);
5094 if (pmd_trans_unstable(pmd))
5097 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5098 for (; addr != end; addr += PAGE_SIZE) {
5099 pte_t ptent = *(pte++);
5105 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5106 case MC_TARGET_PAGE:
5108 if (isolate_lru_page(page))
5110 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5112 /* we uncharge from mc.from later. */
5115 putback_lru_page(page);
5116 put: /* get_mctgt_type() gets the page */
5119 case MC_TARGET_SWAP:
5121 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5123 /* we fixup refcnts and charges later. */
5131 pte_unmap_unlock(pte - 1, ptl);
5136 * We have consumed all precharges we got in can_attach().
5137 * We try charge one by one, but don't do any additional
5138 * charges to mc.to if we have failed in charge once in attach()
5141 ret = mem_cgroup_do_precharge(1);
5149 static void mem_cgroup_move_charge(struct mm_struct *mm)
5151 struct vm_area_struct *vma;
5153 lru_add_drain_all();
5155 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5156 * move_lock while we're moving its pages to another memcg.
5157 * Then wait for already started RCU-only updates to finish.
5159 atomic_inc(&mc.from->moving_account);
5162 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5164 * Someone who are holding the mmap_sem might be waiting in
5165 * waitq. So we cancel all extra charges, wake up all waiters,
5166 * and retry. Because we cancel precharges, we might not be able
5167 * to move enough charges, but moving charge is a best-effort
5168 * feature anyway, so it wouldn't be a big problem.
5170 __mem_cgroup_clear_mc();
5174 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5176 struct mm_walk mem_cgroup_move_charge_walk = {
5177 .pmd_entry = mem_cgroup_move_charge_pte_range,
5181 if (is_vm_hugetlb_page(vma))
5183 ret = walk_page_range(vma->vm_start, vma->vm_end,
5184 &mem_cgroup_move_charge_walk);
5187 * means we have consumed all precharges and failed in
5188 * doing additional charge. Just abandon here.
5192 up_read(&mm->mmap_sem);
5193 atomic_dec(&mc.from->moving_account);
5196 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5197 struct cgroup_taskset *tset)
5199 struct task_struct *p = cgroup_taskset_first(tset);
5200 struct mm_struct *mm = get_task_mm(p);
5204 mem_cgroup_move_charge(mm);
5208 mem_cgroup_clear_mc();
5210 #else /* !CONFIG_MMU */
5211 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5212 struct cgroup_taskset *tset)
5216 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5217 struct cgroup_taskset *tset)
5220 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5221 struct cgroup_taskset *tset)
5227 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5228 * to verify whether we're attached to the default hierarchy on each mount
5231 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5234 * use_hierarchy is forced on the default hierarchy. cgroup core
5235 * guarantees that @root doesn't have any children, so turning it
5236 * on for the root memcg is enough.
5238 if (cgroup_on_dfl(root_css->cgroup))
5239 mem_cgroup_from_css(root_css)->use_hierarchy = true;
5242 struct cgroup_subsys memory_cgrp_subsys = {
5243 .css_alloc = mem_cgroup_css_alloc,
5244 .css_online = mem_cgroup_css_online,
5245 .css_offline = mem_cgroup_css_offline,
5246 .css_free = mem_cgroup_css_free,
5247 .css_reset = mem_cgroup_css_reset,
5248 .can_attach = mem_cgroup_can_attach,
5249 .cancel_attach = mem_cgroup_cancel_attach,
5250 .attach = mem_cgroup_move_task,
5251 .bind = mem_cgroup_bind,
5252 .legacy_cftypes = mem_cgroup_files,
5256 #ifdef CONFIG_MEMCG_SWAP
5257 static int __init enable_swap_account(char *s)
5259 if (!strcmp(s, "1"))
5260 really_do_swap_account = 1;
5261 else if (!strcmp(s, "0"))
5262 really_do_swap_account = 0;
5265 __setup("swapaccount=", enable_swap_account);
5267 static void __init memsw_file_init(void)
5269 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5270 memsw_cgroup_files));
5273 static void __init enable_swap_cgroup(void)
5275 if (!mem_cgroup_disabled() && really_do_swap_account) {
5276 do_swap_account = 1;
5282 static void __init enable_swap_cgroup(void)
5287 #ifdef CONFIG_MEMCG_SWAP
5289 * mem_cgroup_swapout - transfer a memsw charge to swap
5290 * @page: page whose memsw charge to transfer
5291 * @entry: swap entry to move the charge to
5293 * Transfer the memsw charge of @page to @entry.
5295 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5297 struct mem_cgroup *memcg;
5298 unsigned short oldid;
5300 VM_BUG_ON_PAGE(PageLRU(page), page);
5301 VM_BUG_ON_PAGE(page_count(page), page);
5303 if (!do_swap_account)
5306 memcg = page->mem_cgroup;
5308 /* Readahead page, never charged */
5312 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5313 VM_BUG_ON_PAGE(oldid, page);
5314 mem_cgroup_swap_statistics(memcg, true);
5316 page->mem_cgroup = NULL;
5318 if (!mem_cgroup_is_root(memcg))
5319 page_counter_uncharge(&memcg->memory, 1);
5321 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5322 VM_BUG_ON(!irqs_disabled());
5324 mem_cgroup_charge_statistics(memcg, page, -1);
5325 memcg_check_events(memcg, page);
5329 * mem_cgroup_uncharge_swap - uncharge a swap entry
5330 * @entry: swap entry to uncharge
5332 * Drop the memsw charge associated with @entry.
5334 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5336 struct mem_cgroup *memcg;
5339 if (!do_swap_account)
5342 id = swap_cgroup_record(entry, 0);
5344 memcg = mem_cgroup_lookup(id);
5346 if (!mem_cgroup_is_root(memcg))
5347 page_counter_uncharge(&memcg->memsw, 1);
5348 mem_cgroup_swap_statistics(memcg, false);
5349 css_put(&memcg->css);
5356 * mem_cgroup_try_charge - try charging a page
5357 * @page: page to charge
5358 * @mm: mm context of the victim
5359 * @gfp_mask: reclaim mode
5360 * @memcgp: charged memcg return
5362 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5363 * pages according to @gfp_mask if necessary.
5365 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5366 * Otherwise, an error code is returned.
5368 * After page->mapping has been set up, the caller must finalize the
5369 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5370 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5372 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5373 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5375 struct mem_cgroup *memcg = NULL;
5376 unsigned int nr_pages = 1;
5379 if (mem_cgroup_disabled())
5382 if (PageSwapCache(page)) {
5384 * Every swap fault against a single page tries to charge the
5385 * page, bail as early as possible. shmem_unuse() encounters
5386 * already charged pages, too. The USED bit is protected by
5387 * the page lock, which serializes swap cache removal, which
5388 * in turn serializes uncharging.
5390 if (page->mem_cgroup)
5394 if (PageTransHuge(page)) {
5395 nr_pages <<= compound_order(page);
5396 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5399 if (do_swap_account && PageSwapCache(page))
5400 memcg = try_get_mem_cgroup_from_page(page);
5402 memcg = get_mem_cgroup_from_mm(mm);
5404 ret = try_charge(memcg, gfp_mask, nr_pages);
5406 css_put(&memcg->css);
5408 if (ret == -EINTR) {
5409 memcg = root_mem_cgroup;
5418 * mem_cgroup_commit_charge - commit a page charge
5419 * @page: page to charge
5420 * @memcg: memcg to charge the page to
5421 * @lrucare: page might be on LRU already
5423 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5424 * after page->mapping has been set up. This must happen atomically
5425 * as part of the page instantiation, i.e. under the page table lock
5426 * for anonymous pages, under the page lock for page and swap cache.
5428 * In addition, the page must not be on the LRU during the commit, to
5429 * prevent racing with task migration. If it might be, use @lrucare.
5431 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5433 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5436 unsigned int nr_pages = 1;
5438 VM_BUG_ON_PAGE(!page->mapping, page);
5439 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5441 if (mem_cgroup_disabled())
5444 * Swap faults will attempt to charge the same page multiple
5445 * times. But reuse_swap_page() might have removed the page
5446 * from swapcache already, so we can't check PageSwapCache().
5451 commit_charge(page, memcg, lrucare);
5453 if (PageTransHuge(page)) {
5454 nr_pages <<= compound_order(page);
5455 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5458 local_irq_disable();
5459 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5460 memcg_check_events(memcg, page);
5463 if (do_swap_account && PageSwapCache(page)) {
5464 swp_entry_t entry = { .val = page_private(page) };
5466 * The swap entry might not get freed for a long time,
5467 * let's not wait for it. The page already received a
5468 * memory+swap charge, drop the swap entry duplicate.
5470 mem_cgroup_uncharge_swap(entry);
5475 * mem_cgroup_cancel_charge - cancel a page charge
5476 * @page: page to charge
5477 * @memcg: memcg to charge the page to
5479 * Cancel a charge transaction started by mem_cgroup_try_charge().
5481 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5483 unsigned int nr_pages = 1;
5485 if (mem_cgroup_disabled())
5488 * Swap faults will attempt to charge the same page multiple
5489 * times. But reuse_swap_page() might have removed the page
5490 * from swapcache already, so we can't check PageSwapCache().
5495 if (PageTransHuge(page)) {
5496 nr_pages <<= compound_order(page);
5497 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5500 cancel_charge(memcg, nr_pages);
5503 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5504 unsigned long nr_anon, unsigned long nr_file,
5505 unsigned long nr_huge, struct page *dummy_page)
5507 unsigned long nr_pages = nr_anon + nr_file;
5508 unsigned long flags;
5510 if (!mem_cgroup_is_root(memcg)) {
5511 page_counter_uncharge(&memcg->memory, nr_pages);
5512 if (do_swap_account)
5513 page_counter_uncharge(&memcg->memsw, nr_pages);
5514 memcg_oom_recover(memcg);
5517 local_irq_save(flags);
5518 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5519 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5520 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5521 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5522 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5523 memcg_check_events(memcg, dummy_page);
5524 local_irq_restore(flags);
5526 if (!mem_cgroup_is_root(memcg))
5527 css_put_many(&memcg->css, nr_pages);
5530 static void uncharge_list(struct list_head *page_list)
5532 struct mem_cgroup *memcg = NULL;
5533 unsigned long nr_anon = 0;
5534 unsigned long nr_file = 0;
5535 unsigned long nr_huge = 0;
5536 unsigned long pgpgout = 0;
5537 struct list_head *next;
5540 next = page_list->next;
5542 unsigned int nr_pages = 1;
5544 page = list_entry(next, struct page, lru);
5545 next = page->lru.next;
5547 VM_BUG_ON_PAGE(PageLRU(page), page);
5548 VM_BUG_ON_PAGE(page_count(page), page);
5550 if (!page->mem_cgroup)
5554 * Nobody should be changing or seriously looking at
5555 * page->mem_cgroup at this point, we have fully
5556 * exclusive access to the page.
5559 if (memcg != page->mem_cgroup) {
5561 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5563 pgpgout = nr_anon = nr_file = nr_huge = 0;
5565 memcg = page->mem_cgroup;
5568 if (PageTransHuge(page)) {
5569 nr_pages <<= compound_order(page);
5570 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5571 nr_huge += nr_pages;
5575 nr_anon += nr_pages;
5577 nr_file += nr_pages;
5579 page->mem_cgroup = NULL;
5582 } while (next != page_list);
5585 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5590 * mem_cgroup_uncharge - uncharge a page
5591 * @page: page to uncharge
5593 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5594 * mem_cgroup_commit_charge().
5596 void mem_cgroup_uncharge(struct page *page)
5598 if (mem_cgroup_disabled())
5601 /* Don't touch page->lru of any random page, pre-check: */
5602 if (!page->mem_cgroup)
5605 INIT_LIST_HEAD(&page->lru);
5606 uncharge_list(&page->lru);
5610 * mem_cgroup_uncharge_list - uncharge a list of page
5611 * @page_list: list of pages to uncharge
5613 * Uncharge a list of pages previously charged with
5614 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5616 void mem_cgroup_uncharge_list(struct list_head *page_list)
5618 if (mem_cgroup_disabled())
5621 if (!list_empty(page_list))
5622 uncharge_list(page_list);
5626 * mem_cgroup_migrate - migrate a charge to another page
5627 * @oldpage: currently charged page
5628 * @newpage: page to transfer the charge to
5629 * @lrucare: either or both pages might be on the LRU already
5631 * Migrate the charge from @oldpage to @newpage.
5633 * Both pages must be locked, @newpage->mapping must be set up.
5635 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5638 struct mem_cgroup *memcg;
5641 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5642 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5643 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5644 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5645 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5646 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5649 if (mem_cgroup_disabled())
5652 /* Page cache replacement: new page already charged? */
5653 if (newpage->mem_cgroup)
5657 * Swapcache readahead pages can get migrated before being
5658 * charged, and migration from compaction can happen to an
5659 * uncharged page when the PFN walker finds a page that
5660 * reclaim just put back on the LRU but has not released yet.
5662 memcg = oldpage->mem_cgroup;
5667 lock_page_lru(oldpage, &isolated);
5669 oldpage->mem_cgroup = NULL;
5672 unlock_page_lru(oldpage, isolated);
5674 commit_charge(newpage, memcg, lrucare);
5678 * subsys_initcall() for memory controller.
5680 * Some parts like hotcpu_notifier() have to be initialized from this context
5681 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5682 * everything that doesn't depend on a specific mem_cgroup structure should
5683 * be initialized from here.
5685 static int __init mem_cgroup_init(void)
5687 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5688 enable_swap_cgroup();
5689 mem_cgroup_soft_limit_tree_init();
5693 subsys_initcall(mem_cgroup_init);