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
18 * Charge lifetime sanitation
19 * Lockless page tracking & accounting
20 * Unified hierarchy configuration model
21 * Copyright (C) 2015 Red Hat, Inc., Johannes Weiner
23 * This program is free software; you can redistribute it and/or modify
24 * it under the terms of the GNU General Public License as published by
25 * the Free Software Foundation; either version 2 of the License, or
26 * (at your option) any later version.
28 * This program is distributed in the hope that it will be useful,
29 * but WITHOUT ANY WARRANTY; without even the implied warranty of
30 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
31 * GNU General Public License for more details.
34 #include <linux/page_counter.h>
35 #include <linux/memcontrol.h>
36 #include <linux/cgroup.h>
38 #include <linux/hugetlb.h>
39 #include <linux/pagemap.h>
40 #include <linux/smp.h>
41 #include <linux/page-flags.h>
42 #include <linux/backing-dev.h>
43 #include <linux/bit_spinlock.h>
44 #include <linux/rcupdate.h>
45 #include <linux/limits.h>
46 #include <linux/export.h>
47 #include <linux/mutex.h>
48 #include <linux/rbtree.h>
49 #include <linux/slab.h>
50 #include <linux/swap.h>
51 #include <linux/swapops.h>
52 #include <linux/spinlock.h>
53 #include <linux/eventfd.h>
54 #include <linux/poll.h>
55 #include <linux/sort.h>
57 #include <linux/seq_file.h>
58 #include <linux/vmpressure.h>
59 #include <linux/mm_inline.h>
60 #include <linux/swap_cgroup.h>
61 #include <linux/cpu.h>
62 #include <linux/oom.h>
63 #include <linux/lockdep.h>
64 #include <linux/file.h>
68 #include <net/tcp_memcontrol.h>
71 #include <asm/uaccess.h>
73 #include <trace/events/vmscan.h>
75 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
76 EXPORT_SYMBOL(memory_cgrp_subsys);
78 #define MEM_CGROUP_RECLAIM_RETRIES 5
79 static struct mem_cgroup *root_mem_cgroup __read_mostly;
80 struct cgroup_subsys_state *mem_cgroup_root_css __read_mostly;
82 /* Whether the swap controller is active */
83 #ifdef CONFIG_MEMCG_SWAP
84 int do_swap_account __read_mostly;
86 #define do_swap_account 0
89 static const char * const mem_cgroup_stat_names[] = {
99 static const char * const mem_cgroup_events_names[] = {
106 static const char * const mem_cgroup_lru_names[] = {
115 * Per memcg event counter is incremented at every pagein/pageout. With THP,
116 * it will be incremated by the number of pages. This counter is used for
117 * for trigger some periodic events. This is straightforward and better
118 * than using jiffies etc. to handle periodic memcg event.
120 enum mem_cgroup_events_target {
121 MEM_CGROUP_TARGET_THRESH,
122 MEM_CGROUP_TARGET_SOFTLIMIT,
123 MEM_CGROUP_TARGET_NUMAINFO,
126 #define THRESHOLDS_EVENTS_TARGET 128
127 #define SOFTLIMIT_EVENTS_TARGET 1024
128 #define NUMAINFO_EVENTS_TARGET 1024
130 struct mem_cgroup_stat_cpu {
131 long count[MEM_CGROUP_STAT_NSTATS];
132 unsigned long events[MEMCG_NR_EVENTS];
133 unsigned long nr_page_events;
134 unsigned long targets[MEM_CGROUP_NTARGETS];
137 struct reclaim_iter {
138 struct mem_cgroup *position;
139 /* scan generation, increased every round-trip */
140 unsigned int generation;
144 * per-zone information in memory controller.
146 struct mem_cgroup_per_zone {
147 struct lruvec lruvec;
148 unsigned long lru_size[NR_LRU_LISTS];
150 struct reclaim_iter iter[DEF_PRIORITY + 1];
152 struct rb_node tree_node; /* RB tree node */
153 unsigned long usage_in_excess;/* Set to the value by which */
154 /* the soft limit is exceeded*/
156 struct mem_cgroup *memcg; /* Back pointer, we cannot */
157 /* use container_of */
160 struct mem_cgroup_per_node {
161 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
165 * Cgroups above their limits are maintained in a RB-Tree, independent of
166 * their hierarchy representation
169 struct mem_cgroup_tree_per_zone {
170 struct rb_root rb_root;
174 struct mem_cgroup_tree_per_node {
175 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
178 struct mem_cgroup_tree {
179 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
182 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
184 struct mem_cgroup_threshold {
185 struct eventfd_ctx *eventfd;
186 unsigned long threshold;
190 struct mem_cgroup_threshold_ary {
191 /* An array index points to threshold just below or equal to usage. */
192 int current_threshold;
193 /* Size of entries[] */
195 /* Array of thresholds */
196 struct mem_cgroup_threshold entries[0];
199 struct mem_cgroup_thresholds {
200 /* Primary thresholds array */
201 struct mem_cgroup_threshold_ary *primary;
203 * Spare threshold array.
204 * This is needed to make mem_cgroup_unregister_event() "never fail".
205 * It must be able to store at least primary->size - 1 entries.
207 struct mem_cgroup_threshold_ary *spare;
211 struct mem_cgroup_eventfd_list {
212 struct list_head list;
213 struct eventfd_ctx *eventfd;
217 * cgroup_event represents events which userspace want to receive.
219 struct mem_cgroup_event {
221 * memcg which the event belongs to.
223 struct mem_cgroup *memcg;
225 * eventfd to signal userspace about the event.
227 struct eventfd_ctx *eventfd;
229 * Each of these stored in a list by the cgroup.
231 struct list_head list;
233 * register_event() callback will be used to add new userspace
234 * waiter for changes related to this event. Use eventfd_signal()
235 * on eventfd to send notification to userspace.
237 int (*register_event)(struct mem_cgroup *memcg,
238 struct eventfd_ctx *eventfd, const char *args);
240 * unregister_event() callback will be called when userspace closes
241 * the eventfd or on cgroup removing. This callback must be set,
242 * if you want provide notification functionality.
244 void (*unregister_event)(struct mem_cgroup *memcg,
245 struct eventfd_ctx *eventfd);
247 * All fields below needed to unregister event when
248 * userspace closes eventfd.
251 wait_queue_head_t *wqh;
253 struct work_struct remove;
256 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
257 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
260 * The memory controller data structure. The memory controller controls both
261 * page cache and RSS per cgroup. We would eventually like to provide
262 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
263 * to help the administrator determine what knobs to tune.
266 struct cgroup_subsys_state css;
268 /* Accounted resources */
269 struct page_counter memory;
270 struct page_counter memsw;
271 struct page_counter kmem;
273 /* Normal memory consumption range */
277 unsigned long soft_limit;
279 /* vmpressure notifications */
280 struct vmpressure vmpressure;
282 /* css_online() has been completed */
286 * Should the accounting and control be hierarchical, per subtree?
292 atomic_t oom_wakeups;
295 /* OOM-Killer disable */
296 int oom_kill_disable;
298 /* protect arrays of thresholds */
299 struct mutex thresholds_lock;
301 /* thresholds for memory usage. RCU-protected */
302 struct mem_cgroup_thresholds thresholds;
304 /* thresholds for mem+swap usage. RCU-protected */
305 struct mem_cgroup_thresholds memsw_thresholds;
307 /* For oom notifier event fd */
308 struct list_head oom_notify;
311 * Should we move charges of a task when a task is moved into this
312 * mem_cgroup ? And what type of charges should we move ?
314 unsigned long move_charge_at_immigrate;
316 * set > 0 if pages under this cgroup are moving to other cgroup.
318 atomic_t moving_account;
319 /* taken only while moving_account > 0 */
320 spinlock_t move_lock;
321 struct task_struct *move_lock_task;
322 unsigned long move_lock_flags;
326 struct mem_cgroup_stat_cpu __percpu *stat;
327 spinlock_t pcp_counter_lock;
329 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
330 struct cg_proto tcp_mem;
332 #if defined(CONFIG_MEMCG_KMEM)
333 /* Index in the kmem_cache->memcg_params.memcg_caches array */
335 bool kmem_acct_activated;
336 bool kmem_acct_active;
339 int last_scanned_node;
341 nodemask_t scan_nodes;
342 atomic_t numainfo_events;
343 atomic_t numainfo_updating;
346 #ifdef CONFIG_CGROUP_WRITEBACK
347 struct list_head cgwb_list;
348 struct wb_domain cgwb_domain;
351 /* List of events which userspace want to receive */
352 struct list_head event_list;
353 spinlock_t event_list_lock;
355 struct mem_cgroup_per_node *nodeinfo[0];
356 /* WARNING: nodeinfo must be the last member here */
359 #ifdef CONFIG_MEMCG_KMEM
360 bool memcg_kmem_is_active(struct mem_cgroup *memcg)
362 return memcg->kmem_acct_active;
366 /* Stuffs for move charges at task migration. */
368 * Types of charges to be moved.
370 #define MOVE_ANON 0x1U
371 #define MOVE_FILE 0x2U
372 #define MOVE_MASK (MOVE_ANON | MOVE_FILE)
374 /* "mc" and its members are protected by cgroup_mutex */
375 static struct move_charge_struct {
376 spinlock_t lock; /* for from, to */
377 struct mem_cgroup *from;
378 struct mem_cgroup *to;
380 unsigned long precharge;
381 unsigned long moved_charge;
382 unsigned long moved_swap;
383 struct task_struct *moving_task; /* a task moving charges */
384 wait_queue_head_t waitq; /* a waitq for other context */
386 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
387 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
391 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
392 * limit reclaim to prevent infinite loops, if they ever occur.
394 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
395 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
398 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
399 MEM_CGROUP_CHARGE_TYPE_ANON,
400 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
401 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
405 /* for encoding cft->private value on file */
413 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
414 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
415 #define MEMFILE_ATTR(val) ((val) & 0xffff)
416 /* Used for OOM nofiier */
417 #define OOM_CONTROL (0)
420 * The memcg_create_mutex will be held whenever a new cgroup is created.
421 * As a consequence, any change that needs to protect against new child cgroups
422 * appearing has to hold it as well.
424 static DEFINE_MUTEX(memcg_create_mutex);
426 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
428 return s ? container_of(s, struct mem_cgroup, css) : NULL;
431 /* Some nice accessors for the vmpressure. */
432 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
435 memcg = root_mem_cgroup;
436 return &memcg->vmpressure;
439 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
441 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
444 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
446 return (memcg == root_mem_cgroup);
450 * We restrict the id in the range of [1, 65535], so it can fit into
453 #define MEM_CGROUP_ID_MAX USHRT_MAX
455 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
457 return memcg->css.id;
461 * A helper function to get mem_cgroup from ID. must be called under
462 * rcu_read_lock(). The caller is responsible for calling
463 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
464 * refcnt from swap can be called against removed memcg.)
466 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
468 struct cgroup_subsys_state *css;
470 css = css_from_id(id, &memory_cgrp_subsys);
471 return mem_cgroup_from_css(css);
474 /* Writing them here to avoid exposing memcg's inner layout */
475 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
477 void sock_update_memcg(struct sock *sk)
479 if (mem_cgroup_sockets_enabled) {
480 struct mem_cgroup *memcg;
481 struct cg_proto *cg_proto;
483 BUG_ON(!sk->sk_prot->proto_cgroup);
485 /* Socket cloning can throw us here with sk_cgrp already
486 * filled. It won't however, necessarily happen from
487 * process context. So the test for root memcg given
488 * the current task's memcg won't help us in this case.
490 * Respecting the original socket's memcg is a better
491 * decision in this case.
494 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
495 css_get(&sk->sk_cgrp->memcg->css);
500 memcg = mem_cgroup_from_task(current);
501 cg_proto = sk->sk_prot->proto_cgroup(memcg);
502 if (!mem_cgroup_is_root(memcg) &&
503 memcg_proto_active(cg_proto) &&
504 css_tryget_online(&memcg->css)) {
505 sk->sk_cgrp = cg_proto;
510 EXPORT_SYMBOL(sock_update_memcg);
512 void sock_release_memcg(struct sock *sk)
514 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
515 struct mem_cgroup *memcg;
516 WARN_ON(!sk->sk_cgrp->memcg);
517 memcg = sk->sk_cgrp->memcg;
518 css_put(&sk->sk_cgrp->memcg->css);
522 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
524 if (!memcg || mem_cgroup_is_root(memcg))
527 return &memcg->tcp_mem;
529 EXPORT_SYMBOL(tcp_proto_cgroup);
533 #ifdef CONFIG_MEMCG_KMEM
535 * This will be the memcg's index in each cache's ->memcg_params.memcg_caches.
536 * The main reason for not using cgroup id for this:
537 * this works better in sparse environments, where we have a lot of memcgs,
538 * but only a few kmem-limited. Or also, if we have, for instance, 200
539 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
540 * 200 entry array for that.
542 * The current size of the caches array is stored in memcg_nr_cache_ids. It
543 * will double each time we have to increase it.
545 static DEFINE_IDA(memcg_cache_ida);
546 int memcg_nr_cache_ids;
548 /* Protects memcg_nr_cache_ids */
549 static DECLARE_RWSEM(memcg_cache_ids_sem);
551 void memcg_get_cache_ids(void)
553 down_read(&memcg_cache_ids_sem);
556 void memcg_put_cache_ids(void)
558 up_read(&memcg_cache_ids_sem);
562 * MIN_SIZE is different than 1, because we would like to avoid going through
563 * the alloc/free process all the time. In a small machine, 4 kmem-limited
564 * cgroups is a reasonable guess. In the future, it could be a parameter or
565 * tunable, but that is strictly not necessary.
567 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
568 * this constant directly from cgroup, but it is understandable that this is
569 * better kept as an internal representation in cgroup.c. In any case, the
570 * cgrp_id space is not getting any smaller, and we don't have to necessarily
571 * increase ours as well if it increases.
573 #define MEMCG_CACHES_MIN_SIZE 4
574 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
577 * A lot of the calls to the cache allocation functions are expected to be
578 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
579 * conditional to this static branch, we'll have to allow modules that does
580 * kmem_cache_alloc and the such to see this symbol as well
582 struct static_key memcg_kmem_enabled_key;
583 EXPORT_SYMBOL(memcg_kmem_enabled_key);
585 #endif /* CONFIG_MEMCG_KMEM */
587 static struct mem_cgroup_per_zone *
588 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
590 int nid = zone_to_nid(zone);
591 int zid = zone_idx(zone);
593 return &memcg->nodeinfo[nid]->zoneinfo[zid];
596 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
602 * mem_cgroup_css_from_page - css of the memcg associated with a page
603 * @page: page of interest
605 * If memcg is bound to the default hierarchy, css of the memcg associated
606 * with @page is returned. The returned css remains associated with @page
607 * until it is released.
609 * If memcg is bound to a traditional hierarchy, the css of root_mem_cgroup
612 * XXX: The above description of behavior on the default hierarchy isn't
613 * strictly true yet as replace_page_cache_page() can modify the
614 * association before @page is released even on the default hierarchy;
615 * however, the current and planned usages don't mix the the two functions
616 * and replace_page_cache_page() will soon be updated to make the invariant
619 struct cgroup_subsys_state *mem_cgroup_css_from_page(struct page *page)
621 struct mem_cgroup *memcg;
625 memcg = page->mem_cgroup;
627 if (!memcg || !cgroup_on_dfl(memcg->css.cgroup))
628 memcg = root_mem_cgroup;
634 static struct mem_cgroup_per_zone *
635 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
637 int nid = page_to_nid(page);
638 int zid = page_zonenum(page);
640 return &memcg->nodeinfo[nid]->zoneinfo[zid];
643 static struct mem_cgroup_tree_per_zone *
644 soft_limit_tree_node_zone(int nid, int zid)
646 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
649 static struct mem_cgroup_tree_per_zone *
650 soft_limit_tree_from_page(struct page *page)
652 int nid = page_to_nid(page);
653 int zid = page_zonenum(page);
655 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
658 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
659 struct mem_cgroup_tree_per_zone *mctz,
660 unsigned long new_usage_in_excess)
662 struct rb_node **p = &mctz->rb_root.rb_node;
663 struct rb_node *parent = NULL;
664 struct mem_cgroup_per_zone *mz_node;
669 mz->usage_in_excess = new_usage_in_excess;
670 if (!mz->usage_in_excess)
674 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
676 if (mz->usage_in_excess < mz_node->usage_in_excess)
679 * We can't avoid mem cgroups that are over their soft
680 * limit by the same amount
682 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
685 rb_link_node(&mz->tree_node, parent, p);
686 rb_insert_color(&mz->tree_node, &mctz->rb_root);
690 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
691 struct mem_cgroup_tree_per_zone *mctz)
695 rb_erase(&mz->tree_node, &mctz->rb_root);
699 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
700 struct mem_cgroup_tree_per_zone *mctz)
704 spin_lock_irqsave(&mctz->lock, flags);
705 __mem_cgroup_remove_exceeded(mz, mctz);
706 spin_unlock_irqrestore(&mctz->lock, flags);
709 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
711 unsigned long nr_pages = page_counter_read(&memcg->memory);
712 unsigned long soft_limit = READ_ONCE(memcg->soft_limit);
713 unsigned long excess = 0;
715 if (nr_pages > soft_limit)
716 excess = nr_pages - soft_limit;
721 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
723 unsigned long excess;
724 struct mem_cgroup_per_zone *mz;
725 struct mem_cgroup_tree_per_zone *mctz;
727 mctz = soft_limit_tree_from_page(page);
729 * Necessary to update all ancestors when hierarchy is used.
730 * because their event counter is not touched.
732 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
733 mz = mem_cgroup_page_zoneinfo(memcg, page);
734 excess = soft_limit_excess(memcg);
736 * We have to update the tree if mz is on RB-tree or
737 * mem is over its softlimit.
739 if (excess || mz->on_tree) {
742 spin_lock_irqsave(&mctz->lock, flags);
743 /* if on-tree, remove it */
745 __mem_cgroup_remove_exceeded(mz, mctz);
747 * Insert again. mz->usage_in_excess will be updated.
748 * If excess is 0, no tree ops.
750 __mem_cgroup_insert_exceeded(mz, mctz, excess);
751 spin_unlock_irqrestore(&mctz->lock, flags);
756 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
758 struct mem_cgroup_tree_per_zone *mctz;
759 struct mem_cgroup_per_zone *mz;
763 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
764 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
765 mctz = soft_limit_tree_node_zone(nid, zid);
766 mem_cgroup_remove_exceeded(mz, mctz);
771 static struct mem_cgroup_per_zone *
772 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
774 struct rb_node *rightmost = NULL;
775 struct mem_cgroup_per_zone *mz;
779 rightmost = rb_last(&mctz->rb_root);
781 goto done; /* Nothing to reclaim from */
783 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
785 * Remove the node now but someone else can add it back,
786 * we will to add it back at the end of reclaim to its correct
787 * position in the tree.
789 __mem_cgroup_remove_exceeded(mz, mctz);
790 if (!soft_limit_excess(mz->memcg) ||
791 !css_tryget_online(&mz->memcg->css))
797 static struct mem_cgroup_per_zone *
798 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
800 struct mem_cgroup_per_zone *mz;
802 spin_lock_irq(&mctz->lock);
803 mz = __mem_cgroup_largest_soft_limit_node(mctz);
804 spin_unlock_irq(&mctz->lock);
809 * Implementation Note: reading percpu statistics for memcg.
811 * Both of vmstat[] and percpu_counter has threshold and do periodic
812 * synchronization to implement "quick" read. There are trade-off between
813 * reading cost and precision of value. Then, we may have a chance to implement
814 * a periodic synchronizion of counter in memcg's counter.
816 * But this _read() function is used for user interface now. The user accounts
817 * memory usage by memory cgroup and he _always_ requires exact value because
818 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
819 * have to visit all online cpus and make sum. So, for now, unnecessary
820 * synchronization is not implemented. (just implemented for cpu hotplug)
822 * If there are kernel internal actions which can make use of some not-exact
823 * value, and reading all cpu value can be performance bottleneck in some
824 * common workload, threashold and synchonization as vmstat[] should be
827 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
828 enum mem_cgroup_stat_index idx)
833 for_each_possible_cpu(cpu)
834 val += per_cpu(memcg->stat->count[idx], cpu);
838 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
839 enum mem_cgroup_events_index idx)
841 unsigned long val = 0;
844 for_each_possible_cpu(cpu)
845 val += per_cpu(memcg->stat->events[idx], cpu);
849 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
854 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
855 * counted as CACHE even if it's on ANON LRU.
858 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
861 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
864 if (PageTransHuge(page))
865 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
868 /* pagein of a big page is an event. So, ignore page size */
870 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
872 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
873 nr_pages = -nr_pages; /* for event */
876 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
879 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
881 struct mem_cgroup_per_zone *mz;
883 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
884 return mz->lru_size[lru];
887 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
889 unsigned int lru_mask)
891 unsigned long nr = 0;
894 VM_BUG_ON((unsigned)nid >= nr_node_ids);
896 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
897 struct mem_cgroup_per_zone *mz;
901 if (!(BIT(lru) & lru_mask))
903 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
904 nr += mz->lru_size[lru];
910 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
911 unsigned int lru_mask)
913 unsigned long nr = 0;
916 for_each_node_state(nid, N_MEMORY)
917 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
921 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
922 enum mem_cgroup_events_target target)
924 unsigned long val, next;
926 val = __this_cpu_read(memcg->stat->nr_page_events);
927 next = __this_cpu_read(memcg->stat->targets[target]);
928 /* from time_after() in jiffies.h */
929 if ((long)next - (long)val < 0) {
931 case MEM_CGROUP_TARGET_THRESH:
932 next = val + THRESHOLDS_EVENTS_TARGET;
934 case MEM_CGROUP_TARGET_SOFTLIMIT:
935 next = val + SOFTLIMIT_EVENTS_TARGET;
937 case MEM_CGROUP_TARGET_NUMAINFO:
938 next = val + NUMAINFO_EVENTS_TARGET;
943 __this_cpu_write(memcg->stat->targets[target], next);
950 * Check events in order.
953 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
955 /* threshold event is triggered in finer grain than soft limit */
956 if (unlikely(mem_cgroup_event_ratelimit(memcg,
957 MEM_CGROUP_TARGET_THRESH))) {
959 bool do_numainfo __maybe_unused;
961 do_softlimit = mem_cgroup_event_ratelimit(memcg,
962 MEM_CGROUP_TARGET_SOFTLIMIT);
964 do_numainfo = mem_cgroup_event_ratelimit(memcg,
965 MEM_CGROUP_TARGET_NUMAINFO);
967 mem_cgroup_threshold(memcg);
968 if (unlikely(do_softlimit))
969 mem_cgroup_update_tree(memcg, page);
971 if (unlikely(do_numainfo))
972 atomic_inc(&memcg->numainfo_events);
977 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
980 * mm_update_next_owner() may clear mm->owner to NULL
981 * if it races with swapoff, page migration, etc.
982 * So this can be called with p == NULL.
987 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
990 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
992 struct mem_cgroup *memcg = NULL;
997 * Page cache insertions can happen withou an
998 * actual mm context, e.g. during disk probing
999 * on boot, loopback IO, acct() writes etc.
1002 memcg = root_mem_cgroup;
1004 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1005 if (unlikely(!memcg))
1006 memcg = root_mem_cgroup;
1008 } while (!css_tryget_online(&memcg->css));
1014 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1015 * @root: hierarchy root
1016 * @prev: previously returned memcg, NULL on first invocation
1017 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1019 * Returns references to children of the hierarchy below @root, or
1020 * @root itself, or %NULL after a full round-trip.
1022 * Caller must pass the return value in @prev on subsequent
1023 * invocations for reference counting, or use mem_cgroup_iter_break()
1024 * to cancel a hierarchy walk before the round-trip is complete.
1026 * Reclaimers can specify a zone and a priority level in @reclaim to
1027 * divide up the memcgs in the hierarchy among all concurrent
1028 * reclaimers operating on the same zone and priority.
1030 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1031 struct mem_cgroup *prev,
1032 struct mem_cgroup_reclaim_cookie *reclaim)
1034 struct reclaim_iter *uninitialized_var(iter);
1035 struct cgroup_subsys_state *css = NULL;
1036 struct mem_cgroup *memcg = NULL;
1037 struct mem_cgroup *pos = NULL;
1039 if (mem_cgroup_disabled())
1043 root = root_mem_cgroup;
1045 if (prev && !reclaim)
1048 if (!root->use_hierarchy && root != root_mem_cgroup) {
1057 struct mem_cgroup_per_zone *mz;
1059 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1060 iter = &mz->iter[reclaim->priority];
1062 if (prev && reclaim->generation != iter->generation)
1066 pos = READ_ONCE(iter->position);
1068 * A racing update may change the position and
1069 * put the last reference, hence css_tryget(),
1070 * or retry to see the updated position.
1072 } while (pos && !css_tryget(&pos->css));
1079 css = css_next_descendant_pre(css, &root->css);
1082 * Reclaimers share the hierarchy walk, and a
1083 * new one might jump in right at the end of
1084 * the hierarchy - make sure they see at least
1085 * one group and restart from the beginning.
1093 * Verify the css and acquire a reference. The root
1094 * is provided by the caller, so we know it's alive
1095 * and kicking, and don't take an extra reference.
1097 memcg = mem_cgroup_from_css(css);
1099 if (css == &root->css)
1102 if (css_tryget(css)) {
1104 * Make sure the memcg is initialized:
1105 * mem_cgroup_css_online() orders the the
1106 * initialization against setting the flag.
1108 if (smp_load_acquire(&memcg->initialized))
1118 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1120 css_get(&memcg->css);
1126 * pairs with css_tryget when dereferencing iter->position
1135 reclaim->generation = iter->generation;
1141 if (prev && prev != root)
1142 css_put(&prev->css);
1148 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1149 * @root: hierarchy root
1150 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1152 void mem_cgroup_iter_break(struct mem_cgroup *root,
1153 struct mem_cgroup *prev)
1156 root = root_mem_cgroup;
1157 if (prev && prev != root)
1158 css_put(&prev->css);
1162 * Iteration constructs for visiting all cgroups (under a tree). If
1163 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1164 * be used for reference counting.
1166 #define for_each_mem_cgroup_tree(iter, root) \
1167 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1169 iter = mem_cgroup_iter(root, iter, NULL))
1171 #define for_each_mem_cgroup(iter) \
1172 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1174 iter = mem_cgroup_iter(NULL, iter, NULL))
1176 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1178 struct mem_cgroup *memcg;
1181 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1182 if (unlikely(!memcg))
1187 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1190 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1198 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1201 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1202 * @zone: zone of the wanted lruvec
1203 * @memcg: memcg of the wanted lruvec
1205 * Returns the lru list vector holding pages for the given @zone and
1206 * @mem. This can be the global zone lruvec, if the memory controller
1209 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1210 struct mem_cgroup *memcg)
1212 struct mem_cgroup_per_zone *mz;
1213 struct lruvec *lruvec;
1215 if (mem_cgroup_disabled()) {
1216 lruvec = &zone->lruvec;
1220 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1221 lruvec = &mz->lruvec;
1224 * Since a node can be onlined after the mem_cgroup was created,
1225 * we have to be prepared to initialize lruvec->zone here;
1226 * and if offlined then reonlined, we need to reinitialize it.
1228 if (unlikely(lruvec->zone != zone))
1229 lruvec->zone = zone;
1234 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1236 * @zone: zone of the page
1238 * This function is only safe when following the LRU page isolation
1239 * and putback protocol: the LRU lock must be held, and the page must
1240 * either be PageLRU() or the caller must have isolated/allocated it.
1242 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1244 struct mem_cgroup_per_zone *mz;
1245 struct mem_cgroup *memcg;
1246 struct lruvec *lruvec;
1248 if (mem_cgroup_disabled()) {
1249 lruvec = &zone->lruvec;
1253 memcg = page->mem_cgroup;
1255 * Swapcache readahead pages are added to the LRU - and
1256 * possibly migrated - before they are charged.
1259 memcg = root_mem_cgroup;
1261 mz = mem_cgroup_page_zoneinfo(memcg, page);
1262 lruvec = &mz->lruvec;
1265 * Since a node can be onlined after the mem_cgroup was created,
1266 * we have to be prepared to initialize lruvec->zone here;
1267 * and if offlined then reonlined, we need to reinitialize it.
1269 if (unlikely(lruvec->zone != zone))
1270 lruvec->zone = zone;
1275 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1276 * @lruvec: mem_cgroup per zone lru vector
1277 * @lru: index of lru list the page is sitting on
1278 * @nr_pages: positive when adding or negative when removing
1280 * This function must be called when a page is added to or removed from an
1283 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1286 struct mem_cgroup_per_zone *mz;
1287 unsigned long *lru_size;
1289 if (mem_cgroup_disabled())
1292 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1293 lru_size = mz->lru_size + lru;
1294 *lru_size += nr_pages;
1295 VM_BUG_ON((long)(*lru_size) < 0);
1298 bool mem_cgroup_is_descendant(struct mem_cgroup *memcg, struct mem_cgroup *root)
1302 if (!root->use_hierarchy)
1304 return cgroup_is_descendant(memcg->css.cgroup, root->css.cgroup);
1307 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1309 struct mem_cgroup *task_memcg;
1310 struct task_struct *p;
1313 p = find_lock_task_mm(task);
1315 task_memcg = get_mem_cgroup_from_mm(p->mm);
1319 * All threads may have already detached their mm's, but the oom
1320 * killer still needs to detect if they have already been oom
1321 * killed to prevent needlessly killing additional tasks.
1324 task_memcg = mem_cgroup_from_task(task);
1325 css_get(&task_memcg->css);
1328 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1329 css_put(&task_memcg->css);
1333 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1335 unsigned long inactive_ratio;
1336 unsigned long inactive;
1337 unsigned long active;
1340 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1341 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1343 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1345 inactive_ratio = int_sqrt(10 * gb);
1349 return inactive * inactive_ratio < active;
1352 bool mem_cgroup_lruvec_online(struct lruvec *lruvec)
1354 struct mem_cgroup_per_zone *mz;
1355 struct mem_cgroup *memcg;
1357 if (mem_cgroup_disabled())
1360 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1363 return !!(memcg->css.flags & CSS_ONLINE);
1366 #define mem_cgroup_from_counter(counter, member) \
1367 container_of(counter, struct mem_cgroup, member)
1370 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1371 * @memcg: the memory cgroup
1373 * Returns the maximum amount of memory @mem can be charged with, in
1376 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1378 unsigned long margin = 0;
1379 unsigned long count;
1380 unsigned long limit;
1382 count = page_counter_read(&memcg->memory);
1383 limit = READ_ONCE(memcg->memory.limit);
1385 margin = limit - count;
1387 if (do_swap_account) {
1388 count = page_counter_read(&memcg->memsw);
1389 limit = READ_ONCE(memcg->memsw.limit);
1391 margin = min(margin, limit - count);
1397 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1400 if (mem_cgroup_disabled() || !memcg->css.parent)
1401 return vm_swappiness;
1403 return memcg->swappiness;
1407 * A routine for checking "mem" is under move_account() or not.
1409 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1410 * moving cgroups. This is for waiting at high-memory pressure
1413 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1415 struct mem_cgroup *from;
1416 struct mem_cgroup *to;
1419 * Unlike task_move routines, we access mc.to, mc.from not under
1420 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1422 spin_lock(&mc.lock);
1428 ret = mem_cgroup_is_descendant(from, memcg) ||
1429 mem_cgroup_is_descendant(to, memcg);
1431 spin_unlock(&mc.lock);
1435 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1437 if (mc.moving_task && current != mc.moving_task) {
1438 if (mem_cgroup_under_move(memcg)) {
1440 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1441 /* moving charge context might have finished. */
1444 finish_wait(&mc.waitq, &wait);
1451 #define K(x) ((x) << (PAGE_SHIFT-10))
1453 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1454 * @memcg: The memory cgroup that went over limit
1455 * @p: Task that is going to be killed
1457 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1460 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1462 /* oom_info_lock ensures that parallel ooms do not interleave */
1463 static DEFINE_MUTEX(oom_info_lock);
1464 struct mem_cgroup *iter;
1467 mutex_lock(&oom_info_lock);
1471 pr_info("Task in ");
1472 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1473 pr_cont(" killed as a result of limit of ");
1475 pr_info("Memory limit reached of cgroup ");
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 mark_tsk_oom_victim(current);
1564 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL, memcg);
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 */
1934 if (!handle || oom_killer_disabled)
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
1981 * This function must mark the beginning of an accounted page state
1982 * change to prevent double accounting when the page is concurrently
1983 * being moved to another memcg:
1985 * memcg = mem_cgroup_begin_page_stat(page);
1986 * if (TestClearPageState(page))
1987 * mem_cgroup_update_page_stat(memcg, state, -1);
1988 * mem_cgroup_end_page_stat(memcg);
1990 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page)
1992 struct mem_cgroup *memcg;
1993 unsigned long flags;
1996 * The RCU lock is held throughout the transaction. The fast
1997 * path can get away without acquiring the memcg->move_lock
1998 * because page moving starts with an RCU grace period.
2000 * The RCU lock also protects the memcg from being freed when
2001 * the page state that is going to change is the only thing
2002 * preventing the page from being uncharged.
2003 * E.g. end-writeback clearing PageWriteback(), which allows
2004 * migration to go ahead and uncharge the page before the
2005 * account transaction might be complete.
2009 if (mem_cgroup_disabled())
2012 memcg = page->mem_cgroup;
2013 if (unlikely(!memcg))
2016 if (atomic_read(&memcg->moving_account) <= 0)
2019 spin_lock_irqsave(&memcg->move_lock, flags);
2020 if (memcg != page->mem_cgroup) {
2021 spin_unlock_irqrestore(&memcg->move_lock, flags);
2026 * When charge migration first begins, we can have locked and
2027 * unlocked page stat updates happening concurrently. Track
2028 * the task who has the lock for mem_cgroup_end_page_stat().
2030 memcg->move_lock_task = current;
2031 memcg->move_lock_flags = flags;
2035 EXPORT_SYMBOL(mem_cgroup_begin_page_stat);
2038 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2039 * @memcg: the memcg that was accounted against
2041 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg)
2043 if (memcg && memcg->move_lock_task == current) {
2044 unsigned long flags = memcg->move_lock_flags;
2046 memcg->move_lock_task = NULL;
2047 memcg->move_lock_flags = 0;
2049 spin_unlock_irqrestore(&memcg->move_lock, flags);
2054 EXPORT_SYMBOL(mem_cgroup_end_page_stat);
2057 * mem_cgroup_update_page_stat - update page state statistics
2058 * @memcg: memcg to account against
2059 * @idx: page state item to account
2060 * @val: number of pages (positive or negative)
2062 * See mem_cgroup_begin_page_stat() for locking requirements.
2064 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2065 enum mem_cgroup_stat_index idx, int val)
2067 VM_BUG_ON(!rcu_read_lock_held());
2070 this_cpu_add(memcg->stat->count[idx], val);
2074 * size of first charge trial. "32" comes from vmscan.c's magic value.
2075 * TODO: maybe necessary to use big numbers in big irons.
2077 #define CHARGE_BATCH 32U
2078 struct memcg_stock_pcp {
2079 struct mem_cgroup *cached; /* this never be root cgroup */
2080 unsigned int nr_pages;
2081 struct work_struct work;
2082 unsigned long flags;
2083 #define FLUSHING_CACHED_CHARGE 0
2085 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2086 static DEFINE_MUTEX(percpu_charge_mutex);
2089 * consume_stock: Try to consume stocked charge on this cpu.
2090 * @memcg: memcg to consume from.
2091 * @nr_pages: how many pages to charge.
2093 * The charges will only happen if @memcg matches the current cpu's memcg
2094 * stock, and at least @nr_pages are available in that stock. Failure to
2095 * service an allocation will refill the stock.
2097 * returns true if successful, false otherwise.
2099 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2101 struct memcg_stock_pcp *stock;
2104 if (nr_pages > CHARGE_BATCH)
2107 stock = &get_cpu_var(memcg_stock);
2108 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2109 stock->nr_pages -= nr_pages;
2112 put_cpu_var(memcg_stock);
2117 * Returns stocks cached in percpu and reset cached information.
2119 static void drain_stock(struct memcg_stock_pcp *stock)
2121 struct mem_cgroup *old = stock->cached;
2123 if (stock->nr_pages) {
2124 page_counter_uncharge(&old->memory, stock->nr_pages);
2125 if (do_swap_account)
2126 page_counter_uncharge(&old->memsw, stock->nr_pages);
2127 css_put_many(&old->css, stock->nr_pages);
2128 stock->nr_pages = 0;
2130 stock->cached = NULL;
2134 * This must be called under preempt disabled or must be called by
2135 * a thread which is pinned to local cpu.
2137 static void drain_local_stock(struct work_struct *dummy)
2139 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2141 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2145 * Cache charges(val) to local per_cpu area.
2146 * This will be consumed by consume_stock() function, later.
2148 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2150 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2152 if (stock->cached != memcg) { /* reset if necessary */
2154 stock->cached = memcg;
2156 stock->nr_pages += nr_pages;
2157 put_cpu_var(memcg_stock);
2161 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2162 * of the hierarchy under it.
2164 static void drain_all_stock(struct mem_cgroup *root_memcg)
2168 /* If someone's already draining, avoid adding running more workers. */
2169 if (!mutex_trylock(&percpu_charge_mutex))
2171 /* Notify other cpus that system-wide "drain" is running */
2174 for_each_online_cpu(cpu) {
2175 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2176 struct mem_cgroup *memcg;
2178 memcg = stock->cached;
2179 if (!memcg || !stock->nr_pages)
2181 if (!mem_cgroup_is_descendant(memcg, root_memcg))
2183 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2185 drain_local_stock(&stock->work);
2187 schedule_work_on(cpu, &stock->work);
2192 mutex_unlock(&percpu_charge_mutex);
2195 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2196 unsigned long action,
2199 int cpu = (unsigned long)hcpu;
2200 struct memcg_stock_pcp *stock;
2202 if (action == CPU_ONLINE)
2205 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2208 stock = &per_cpu(memcg_stock, cpu);
2213 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2214 unsigned int nr_pages)
2216 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2217 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2218 struct mem_cgroup *mem_over_limit;
2219 struct page_counter *counter;
2220 unsigned long nr_reclaimed;
2221 bool may_swap = true;
2222 bool drained = false;
2225 if (mem_cgroup_is_root(memcg))
2228 if (consume_stock(memcg, nr_pages))
2231 if (!do_swap_account ||
2232 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2233 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2235 if (do_swap_account)
2236 page_counter_uncharge(&memcg->memsw, batch);
2237 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2239 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2243 if (batch > nr_pages) {
2249 * Unlike in global OOM situations, memcg is not in a physical
2250 * memory shortage. Allow dying and OOM-killed tasks to
2251 * bypass the last charges so that they can exit quickly and
2252 * free their memory.
2254 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2255 fatal_signal_pending(current) ||
2256 current->flags & PF_EXITING))
2259 if (unlikely(task_in_memcg_oom(current)))
2262 if (!(gfp_mask & __GFP_WAIT))
2265 mem_cgroup_events(mem_over_limit, MEMCG_MAX, 1);
2267 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2268 gfp_mask, may_swap);
2270 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2274 drain_all_stock(mem_over_limit);
2279 if (gfp_mask & __GFP_NORETRY)
2282 * Even though the limit is exceeded at this point, reclaim
2283 * may have been able to free some pages. Retry the charge
2284 * before killing the task.
2286 * Only for regular pages, though: huge pages are rather
2287 * unlikely to succeed so close to the limit, and we fall back
2288 * to regular pages anyway in case of failure.
2290 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2293 * At task move, charge accounts can be doubly counted. So, it's
2294 * better to wait until the end of task_move if something is going on.
2296 if (mem_cgroup_wait_acct_move(mem_over_limit))
2302 if (gfp_mask & __GFP_NOFAIL)
2305 if (fatal_signal_pending(current))
2308 mem_cgroup_events(mem_over_limit, MEMCG_OOM, 1);
2310 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2312 if (!(gfp_mask & __GFP_NOFAIL))
2318 css_get_many(&memcg->css, batch);
2319 if (batch > nr_pages)
2320 refill_stock(memcg, batch - nr_pages);
2322 * If the hierarchy is above the normal consumption range,
2323 * make the charging task trim their excess contribution.
2326 if (page_counter_read(&memcg->memory) <= memcg->high)
2328 mem_cgroup_events(memcg, MEMCG_HIGH, 1);
2329 try_to_free_mem_cgroup_pages(memcg, nr_pages, gfp_mask, true);
2330 } while ((memcg = parent_mem_cgroup(memcg)));
2335 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2337 if (mem_cgroup_is_root(memcg))
2340 page_counter_uncharge(&memcg->memory, nr_pages);
2341 if (do_swap_account)
2342 page_counter_uncharge(&memcg->memsw, nr_pages);
2344 css_put_many(&memcg->css, nr_pages);
2348 * try_get_mem_cgroup_from_page - look up page's memcg association
2351 * Look up, get a css reference, and return the memcg that owns @page.
2353 * The page must be locked to prevent racing with swap-in and page
2354 * cache charges. If coming from an unlocked page table, the caller
2355 * must ensure the page is on the LRU or this can race with charging.
2357 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2359 struct mem_cgroup *memcg;
2363 VM_BUG_ON_PAGE(!PageLocked(page), page);
2365 memcg = page->mem_cgroup;
2367 if (!css_tryget_online(&memcg->css))
2369 } else if (PageSwapCache(page)) {
2370 ent.val = page_private(page);
2371 id = lookup_swap_cgroup_id(ent);
2373 memcg = mem_cgroup_from_id(id);
2374 if (memcg && !css_tryget_online(&memcg->css))
2381 static void lock_page_lru(struct page *page, int *isolated)
2383 struct zone *zone = page_zone(page);
2385 spin_lock_irq(&zone->lru_lock);
2386 if (PageLRU(page)) {
2387 struct lruvec *lruvec;
2389 lruvec = mem_cgroup_page_lruvec(page, zone);
2391 del_page_from_lru_list(page, lruvec, page_lru(page));
2397 static void unlock_page_lru(struct page *page, int isolated)
2399 struct zone *zone = page_zone(page);
2402 struct lruvec *lruvec;
2404 lruvec = mem_cgroup_page_lruvec(page, zone);
2405 VM_BUG_ON_PAGE(PageLRU(page), page);
2407 add_page_to_lru_list(page, lruvec, page_lru(page));
2409 spin_unlock_irq(&zone->lru_lock);
2412 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2417 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2420 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2421 * may already be on some other mem_cgroup's LRU. Take care of it.
2424 lock_page_lru(page, &isolated);
2427 * Nobody should be changing or seriously looking at
2428 * page->mem_cgroup at this point:
2430 * - the page is uncharged
2432 * - the page is off-LRU
2434 * - an anonymous fault has exclusive page access, except for
2435 * a locked page table
2437 * - a page cache insertion, a swapin fault, or a migration
2438 * have the page locked
2440 page->mem_cgroup = memcg;
2443 unlock_page_lru(page, isolated);
2446 #ifdef CONFIG_MEMCG_KMEM
2447 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2448 unsigned long nr_pages)
2450 struct page_counter *counter;
2453 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2457 ret = try_charge(memcg, gfp, nr_pages);
2458 if (ret == -EINTR) {
2460 * try_charge() chose to bypass to root due to OOM kill or
2461 * fatal signal. Since our only options are to either fail
2462 * the allocation or charge it to this cgroup, do it as a
2463 * temporary condition. But we can't fail. From a kmem/slab
2464 * perspective, the cache has already been selected, by
2465 * mem_cgroup_kmem_get_cache(), so it is too late to change
2468 * This condition will only trigger if the task entered
2469 * memcg_charge_kmem in a sane state, but was OOM-killed
2470 * during try_charge() above. Tasks that were already dying
2471 * when the allocation triggers should have been already
2472 * directed to the root cgroup in memcontrol.h
2474 page_counter_charge(&memcg->memory, nr_pages);
2475 if (do_swap_account)
2476 page_counter_charge(&memcg->memsw, nr_pages);
2477 css_get_many(&memcg->css, nr_pages);
2480 page_counter_uncharge(&memcg->kmem, nr_pages);
2485 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2487 page_counter_uncharge(&memcg->memory, nr_pages);
2488 if (do_swap_account)
2489 page_counter_uncharge(&memcg->memsw, nr_pages);
2491 page_counter_uncharge(&memcg->kmem, nr_pages);
2493 css_put_many(&memcg->css, nr_pages);
2497 * helper for acessing a memcg's index. It will be used as an index in the
2498 * child cache array in kmem_cache, and also to derive its name. This function
2499 * will return -1 when this is not a kmem-limited memcg.
2501 int memcg_cache_id(struct mem_cgroup *memcg)
2503 return memcg ? memcg->kmemcg_id : -1;
2506 static int memcg_alloc_cache_id(void)
2511 id = ida_simple_get(&memcg_cache_ida,
2512 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2516 if (id < memcg_nr_cache_ids)
2520 * There's no space for the new id in memcg_caches arrays,
2521 * so we have to grow them.
2523 down_write(&memcg_cache_ids_sem);
2525 size = 2 * (id + 1);
2526 if (size < MEMCG_CACHES_MIN_SIZE)
2527 size = MEMCG_CACHES_MIN_SIZE;
2528 else if (size > MEMCG_CACHES_MAX_SIZE)
2529 size = MEMCG_CACHES_MAX_SIZE;
2531 err = memcg_update_all_caches(size);
2533 err = memcg_update_all_list_lrus(size);
2535 memcg_nr_cache_ids = size;
2537 up_write(&memcg_cache_ids_sem);
2540 ida_simple_remove(&memcg_cache_ida, id);
2546 static void memcg_free_cache_id(int id)
2548 ida_simple_remove(&memcg_cache_ida, id);
2551 struct memcg_kmem_cache_create_work {
2552 struct mem_cgroup *memcg;
2553 struct kmem_cache *cachep;
2554 struct work_struct work;
2557 static void memcg_kmem_cache_create_func(struct work_struct *w)
2559 struct memcg_kmem_cache_create_work *cw =
2560 container_of(w, struct memcg_kmem_cache_create_work, work);
2561 struct mem_cgroup *memcg = cw->memcg;
2562 struct kmem_cache *cachep = cw->cachep;
2564 memcg_create_kmem_cache(memcg, cachep);
2566 css_put(&memcg->css);
2571 * Enqueue the creation of a per-memcg kmem_cache.
2573 static void __memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2574 struct kmem_cache *cachep)
2576 struct memcg_kmem_cache_create_work *cw;
2578 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2582 css_get(&memcg->css);
2585 cw->cachep = cachep;
2586 INIT_WORK(&cw->work, memcg_kmem_cache_create_func);
2588 schedule_work(&cw->work);
2591 static void memcg_schedule_kmem_cache_create(struct mem_cgroup *memcg,
2592 struct kmem_cache *cachep)
2595 * We need to stop accounting when we kmalloc, because if the
2596 * corresponding kmalloc cache is not yet created, the first allocation
2597 * in __memcg_schedule_kmem_cache_create will recurse.
2599 * However, it is better to enclose the whole function. Depending on
2600 * the debugging options enabled, INIT_WORK(), for instance, can
2601 * trigger an allocation. This too, will make us recurse. Because at
2602 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2603 * the safest choice is to do it like this, wrapping the whole function.
2605 current->memcg_kmem_skip_account = 1;
2606 __memcg_schedule_kmem_cache_create(memcg, cachep);
2607 current->memcg_kmem_skip_account = 0;
2611 * Return the kmem_cache we're supposed to use for a slab allocation.
2612 * We try to use the current memcg's version of the cache.
2614 * If the cache does not exist yet, if we are the first user of it,
2615 * we either create it immediately, if possible, or create it asynchronously
2617 * In the latter case, we will let the current allocation go through with
2618 * the original cache.
2620 * Can't be called in interrupt context or from kernel threads.
2621 * This function needs to be called with rcu_read_lock() held.
2623 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2625 struct mem_cgroup *memcg;
2626 struct kmem_cache *memcg_cachep;
2629 VM_BUG_ON(!is_root_cache(cachep));
2631 if (current->memcg_kmem_skip_account)
2634 memcg = get_mem_cgroup_from_mm(current->mm);
2635 kmemcg_id = READ_ONCE(memcg->kmemcg_id);
2639 memcg_cachep = cache_from_memcg_idx(cachep, kmemcg_id);
2640 if (likely(memcg_cachep))
2641 return memcg_cachep;
2644 * If we are in a safe context (can wait, and not in interrupt
2645 * context), we could be be predictable and return right away.
2646 * This would guarantee that the allocation being performed
2647 * already belongs in the new cache.
2649 * However, there are some clashes that can arrive from locking.
2650 * For instance, because we acquire the slab_mutex while doing
2651 * memcg_create_kmem_cache, this means no further allocation
2652 * could happen with the slab_mutex held. So it's better to
2655 memcg_schedule_kmem_cache_create(memcg, cachep);
2657 css_put(&memcg->css);
2661 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2663 if (!is_root_cache(cachep))
2664 css_put(&cachep->memcg_params.memcg->css);
2668 * We need to verify if the allocation against current->mm->owner's memcg is
2669 * possible for the given order. But the page is not allocated yet, so we'll
2670 * need a further commit step to do the final arrangements.
2672 * It is possible for the task to switch cgroups in this mean time, so at
2673 * commit time, we can't rely on task conversion any longer. We'll then use
2674 * the handle argument to return to the caller which cgroup we should commit
2675 * against. We could also return the memcg directly and avoid the pointer
2676 * passing, but a boolean return value gives better semantics considering
2677 * the compiled-out case as well.
2679 * Returning true means the allocation is possible.
2682 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2684 struct mem_cgroup *memcg;
2689 memcg = get_mem_cgroup_from_mm(current->mm);
2691 if (!memcg_kmem_is_active(memcg)) {
2692 css_put(&memcg->css);
2696 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2700 css_put(&memcg->css);
2704 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2707 VM_BUG_ON(mem_cgroup_is_root(memcg));
2709 /* The page allocation failed. Revert */
2711 memcg_uncharge_kmem(memcg, 1 << order);
2714 page->mem_cgroup = memcg;
2717 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2719 struct mem_cgroup *memcg = page->mem_cgroup;
2724 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2726 memcg_uncharge_kmem(memcg, 1 << order);
2727 page->mem_cgroup = NULL;
2730 struct mem_cgroup *__mem_cgroup_from_kmem(void *ptr)
2732 struct mem_cgroup *memcg = NULL;
2733 struct kmem_cache *cachep;
2736 page = virt_to_head_page(ptr);
2737 if (PageSlab(page)) {
2738 cachep = page->slab_cache;
2739 if (!is_root_cache(cachep))
2740 memcg = cachep->memcg_params.memcg;
2742 /* page allocated by alloc_kmem_pages */
2743 memcg = page->mem_cgroup;
2747 #endif /* CONFIG_MEMCG_KMEM */
2749 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2752 * Because tail pages are not marked as "used", set it. We're under
2753 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2754 * charge/uncharge will be never happen and move_account() is done under
2755 * compound_lock(), so we don't have to take care of races.
2757 void mem_cgroup_split_huge_fixup(struct page *head)
2761 if (mem_cgroup_disabled())
2764 for (i = 1; i < HPAGE_PMD_NR; i++)
2765 head[i].mem_cgroup = head->mem_cgroup;
2767 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2770 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2772 #ifdef CONFIG_MEMCG_SWAP
2773 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
2776 int val = (charge) ? 1 : -1;
2777 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
2781 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2782 * @entry: swap entry to be moved
2783 * @from: mem_cgroup which the entry is moved from
2784 * @to: mem_cgroup which the entry is moved to
2786 * It succeeds only when the swap_cgroup's record for this entry is the same
2787 * as the mem_cgroup's id of @from.
2789 * Returns 0 on success, -EINVAL on failure.
2791 * The caller must have charged to @to, IOW, called page_counter_charge() about
2792 * both res and memsw, and called css_get().
2794 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2795 struct mem_cgroup *from, struct mem_cgroup *to)
2797 unsigned short old_id, new_id;
2799 old_id = mem_cgroup_id(from);
2800 new_id = mem_cgroup_id(to);
2802 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2803 mem_cgroup_swap_statistics(from, false);
2804 mem_cgroup_swap_statistics(to, true);
2810 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2811 struct mem_cgroup *from, struct mem_cgroup *to)
2817 static DEFINE_MUTEX(memcg_limit_mutex);
2819 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2820 unsigned long limit)
2822 unsigned long curusage;
2823 unsigned long oldusage;
2824 bool enlarge = false;
2829 * For keeping hierarchical_reclaim simple, how long we should retry
2830 * is depends on callers. We set our retry-count to be function
2831 * of # of children which we should visit in this loop.
2833 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2834 mem_cgroup_count_children(memcg);
2836 oldusage = page_counter_read(&memcg->memory);
2839 if (signal_pending(current)) {
2844 mutex_lock(&memcg_limit_mutex);
2845 if (limit > memcg->memsw.limit) {
2846 mutex_unlock(&memcg_limit_mutex);
2850 if (limit > memcg->memory.limit)
2852 ret = page_counter_limit(&memcg->memory, limit);
2853 mutex_unlock(&memcg_limit_mutex);
2858 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
2860 curusage = page_counter_read(&memcg->memory);
2861 /* Usage is reduced ? */
2862 if (curusage >= oldusage)
2865 oldusage = curusage;
2866 } while (retry_count);
2868 if (!ret && enlarge)
2869 memcg_oom_recover(memcg);
2874 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2875 unsigned long limit)
2877 unsigned long curusage;
2878 unsigned long oldusage;
2879 bool enlarge = false;
2883 /* see mem_cgroup_resize_res_limit */
2884 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
2885 mem_cgroup_count_children(memcg);
2887 oldusage = page_counter_read(&memcg->memsw);
2890 if (signal_pending(current)) {
2895 mutex_lock(&memcg_limit_mutex);
2896 if (limit < memcg->memory.limit) {
2897 mutex_unlock(&memcg_limit_mutex);
2901 if (limit > memcg->memsw.limit)
2903 ret = page_counter_limit(&memcg->memsw, limit);
2904 mutex_unlock(&memcg_limit_mutex);
2909 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
2911 curusage = page_counter_read(&memcg->memsw);
2912 /* Usage is reduced ? */
2913 if (curusage >= oldusage)
2916 oldusage = curusage;
2917 } while (retry_count);
2919 if (!ret && enlarge)
2920 memcg_oom_recover(memcg);
2925 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2927 unsigned long *total_scanned)
2929 unsigned long nr_reclaimed = 0;
2930 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2931 unsigned long reclaimed;
2933 struct mem_cgroup_tree_per_zone *mctz;
2934 unsigned long excess;
2935 unsigned long nr_scanned;
2940 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
2942 * This loop can run a while, specially if mem_cgroup's continuously
2943 * keep exceeding their soft limit and putting the system under
2950 mz = mem_cgroup_largest_soft_limit_node(mctz);
2955 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
2956 gfp_mask, &nr_scanned);
2957 nr_reclaimed += reclaimed;
2958 *total_scanned += nr_scanned;
2959 spin_lock_irq(&mctz->lock);
2960 __mem_cgroup_remove_exceeded(mz, mctz);
2963 * If we failed to reclaim anything from this memory cgroup
2964 * it is time to move on to the next cgroup
2968 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
2970 excess = soft_limit_excess(mz->memcg);
2972 * One school of thought says that we should not add
2973 * back the node to the tree if reclaim returns 0.
2974 * But our reclaim could return 0, simply because due
2975 * to priority we are exposing a smaller subset of
2976 * memory to reclaim from. Consider this as a longer
2979 /* If excess == 0, no tree ops */
2980 __mem_cgroup_insert_exceeded(mz, mctz, excess);
2981 spin_unlock_irq(&mctz->lock);
2982 css_put(&mz->memcg->css);
2985 * Could not reclaim anything and there are no more
2986 * mem cgroups to try or we seem to be looping without
2987 * reclaiming anything.
2989 if (!nr_reclaimed &&
2991 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2993 } while (!nr_reclaimed);
2995 css_put(&next_mz->memcg->css);
2996 return nr_reclaimed;
3000 * Test whether @memcg has children, dead or alive. Note that this
3001 * function doesn't care whether @memcg has use_hierarchy enabled and
3002 * returns %true if there are child csses according to the cgroup
3003 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3005 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3010 * The lock does not prevent addition or deletion of children, but
3011 * it prevents a new child from being initialized based on this
3012 * parent in css_online(), so it's enough to decide whether
3013 * hierarchically inherited attributes can still be changed or not.
3015 lockdep_assert_held(&memcg_create_mutex);
3018 ret = css_next_child(NULL, &memcg->css);
3024 * Reclaims as many pages from the given memcg as possible and moves
3025 * the rest to the parent.
3027 * Caller is responsible for holding css reference for memcg.
3029 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3031 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3033 /* we call try-to-free pages for make this cgroup empty */
3034 lru_add_drain_all();
3035 /* try to free all pages in this cgroup */
3036 while (nr_retries && page_counter_read(&memcg->memory)) {
3039 if (signal_pending(current))
3042 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3046 /* maybe some writeback is necessary */
3047 congestion_wait(BLK_RW_ASYNC, HZ/10);
3055 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3056 char *buf, size_t nbytes,
3059 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3061 if (mem_cgroup_is_root(memcg))
3063 return mem_cgroup_force_empty(memcg) ?: nbytes;
3066 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3069 return mem_cgroup_from_css(css)->use_hierarchy;
3072 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3073 struct cftype *cft, u64 val)
3076 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3077 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3079 mutex_lock(&memcg_create_mutex);
3081 if (memcg->use_hierarchy == val)
3085 * If parent's use_hierarchy is set, we can't make any modifications
3086 * in the child subtrees. If it is unset, then the change can
3087 * occur, provided the current cgroup has no children.
3089 * For the root cgroup, parent_mem is NULL, we allow value to be
3090 * set if there are no children.
3092 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3093 (val == 1 || val == 0)) {
3094 if (!memcg_has_children(memcg))
3095 memcg->use_hierarchy = val;
3102 mutex_unlock(&memcg_create_mutex);
3107 static unsigned long tree_stat(struct mem_cgroup *memcg,
3108 enum mem_cgroup_stat_index idx)
3110 struct mem_cgroup *iter;
3113 /* Per-cpu values can be negative, use a signed accumulator */
3114 for_each_mem_cgroup_tree(iter, memcg)
3115 val += mem_cgroup_read_stat(iter, idx);
3117 if (val < 0) /* race ? */
3122 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3126 if (mem_cgroup_is_root(memcg)) {
3127 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3128 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3130 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3133 val = page_counter_read(&memcg->memory);
3135 val = page_counter_read(&memcg->memsw);
3137 return val << PAGE_SHIFT;
3148 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3151 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3152 struct page_counter *counter;
3154 switch (MEMFILE_TYPE(cft->private)) {
3156 counter = &memcg->memory;
3159 counter = &memcg->memsw;
3162 counter = &memcg->kmem;
3168 switch (MEMFILE_ATTR(cft->private)) {
3170 if (counter == &memcg->memory)
3171 return mem_cgroup_usage(memcg, false);
3172 if (counter == &memcg->memsw)
3173 return mem_cgroup_usage(memcg, true);
3174 return (u64)page_counter_read(counter) * PAGE_SIZE;
3176 return (u64)counter->limit * PAGE_SIZE;
3178 return (u64)counter->watermark * PAGE_SIZE;
3180 return counter->failcnt;
3181 case RES_SOFT_LIMIT:
3182 return (u64)memcg->soft_limit * PAGE_SIZE;
3188 #ifdef CONFIG_MEMCG_KMEM
3189 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3190 unsigned long nr_pages)
3195 BUG_ON(memcg->kmemcg_id >= 0);
3196 BUG_ON(memcg->kmem_acct_activated);
3197 BUG_ON(memcg->kmem_acct_active);
3200 * For simplicity, we won't allow this to be disabled. It also can't
3201 * be changed if the cgroup has children already, or if tasks had
3204 * If tasks join before we set the limit, a person looking at
3205 * kmem.usage_in_bytes will have no way to determine when it took
3206 * place, which makes the value quite meaningless.
3208 * After it first became limited, changes in the value of the limit are
3209 * of course permitted.
3211 mutex_lock(&memcg_create_mutex);
3212 if (cgroup_has_tasks(memcg->css.cgroup) ||
3213 (memcg->use_hierarchy && memcg_has_children(memcg)))
3215 mutex_unlock(&memcg_create_mutex);
3219 memcg_id = memcg_alloc_cache_id();
3226 * We couldn't have accounted to this cgroup, because it hasn't got
3227 * activated yet, so this should succeed.
3229 err = page_counter_limit(&memcg->kmem, nr_pages);
3232 static_key_slow_inc(&memcg_kmem_enabled_key);
3234 * A memory cgroup is considered kmem-active as soon as it gets
3235 * kmemcg_id. Setting the id after enabling static branching will
3236 * guarantee no one starts accounting before all call sites are
3239 memcg->kmemcg_id = memcg_id;
3240 memcg->kmem_acct_activated = true;
3241 memcg->kmem_acct_active = true;
3246 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3247 unsigned long limit)
3251 mutex_lock(&memcg_limit_mutex);
3252 if (!memcg_kmem_is_active(memcg))
3253 ret = memcg_activate_kmem(memcg, limit);
3255 ret = page_counter_limit(&memcg->kmem, limit);
3256 mutex_unlock(&memcg_limit_mutex);
3260 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3263 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3268 mutex_lock(&memcg_limit_mutex);
3270 * If the parent cgroup is not kmem-active now, it cannot be activated
3271 * after this point, because it has at least one child already.
3273 if (memcg_kmem_is_active(parent))
3274 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3275 mutex_unlock(&memcg_limit_mutex);
3279 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3280 unsigned long limit)
3284 #endif /* CONFIG_MEMCG_KMEM */
3287 * The user of this function is...
3290 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3291 char *buf, size_t nbytes, loff_t off)
3293 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3294 unsigned long nr_pages;
3297 buf = strstrip(buf);
3298 ret = page_counter_memparse(buf, "-1", &nr_pages);
3302 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3304 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3308 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3310 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3313 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3316 ret = memcg_update_kmem_limit(memcg, nr_pages);
3320 case RES_SOFT_LIMIT:
3321 memcg->soft_limit = nr_pages;
3325 return ret ?: nbytes;
3328 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3329 size_t nbytes, loff_t off)
3331 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3332 struct page_counter *counter;
3334 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3336 counter = &memcg->memory;
3339 counter = &memcg->memsw;
3342 counter = &memcg->kmem;
3348 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3350 page_counter_reset_watermark(counter);
3353 counter->failcnt = 0;
3362 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3365 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3369 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3370 struct cftype *cft, u64 val)
3372 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3374 if (val & ~MOVE_MASK)
3378 * No kind of locking is needed in here, because ->can_attach() will
3379 * check this value once in the beginning of the process, and then carry
3380 * on with stale data. This means that changes to this value will only
3381 * affect task migrations starting after the change.
3383 memcg->move_charge_at_immigrate = val;
3387 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3388 struct cftype *cft, u64 val)
3395 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3399 unsigned int lru_mask;
3402 static const struct numa_stat stats[] = {
3403 { "total", LRU_ALL },
3404 { "file", LRU_ALL_FILE },
3405 { "anon", LRU_ALL_ANON },
3406 { "unevictable", BIT(LRU_UNEVICTABLE) },
3408 const struct numa_stat *stat;
3411 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3413 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3414 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3415 seq_printf(m, "%s=%lu", stat->name, nr);
3416 for_each_node_state(nid, N_MEMORY) {
3417 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3419 seq_printf(m, " N%d=%lu", nid, nr);
3424 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3425 struct mem_cgroup *iter;
3428 for_each_mem_cgroup_tree(iter, memcg)
3429 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3430 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3431 for_each_node_state(nid, N_MEMORY) {
3433 for_each_mem_cgroup_tree(iter, memcg)
3434 nr += mem_cgroup_node_nr_lru_pages(
3435 iter, nid, stat->lru_mask);
3436 seq_printf(m, " N%d=%lu", nid, nr);
3443 #endif /* CONFIG_NUMA */
3445 static int memcg_stat_show(struct seq_file *m, void *v)
3447 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3448 unsigned long memory, memsw;
3449 struct mem_cgroup *mi;
3452 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_stat_names) !=
3453 MEM_CGROUP_STAT_NSTATS);
3454 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_events_names) !=
3455 MEM_CGROUP_EVENTS_NSTATS);
3456 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3458 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3459 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3461 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3462 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3465 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3466 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3467 mem_cgroup_read_events(memcg, i));
3469 for (i = 0; i < NR_LRU_LISTS; i++)
3470 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3471 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3473 /* Hierarchical information */
3474 memory = memsw = PAGE_COUNTER_MAX;
3475 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3476 memory = min(memory, mi->memory.limit);
3477 memsw = min(memsw, mi->memsw.limit);
3479 seq_printf(m, "hierarchical_memory_limit %llu\n",
3480 (u64)memory * PAGE_SIZE);
3481 if (do_swap_account)
3482 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3483 (u64)memsw * PAGE_SIZE);
3485 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3488 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3490 for_each_mem_cgroup_tree(mi, memcg)
3491 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3492 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3495 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3496 unsigned long long val = 0;
3498 for_each_mem_cgroup_tree(mi, memcg)
3499 val += mem_cgroup_read_events(mi, i);
3500 seq_printf(m, "total_%s %llu\n",
3501 mem_cgroup_events_names[i], val);
3504 for (i = 0; i < NR_LRU_LISTS; i++) {
3505 unsigned long long val = 0;
3507 for_each_mem_cgroup_tree(mi, memcg)
3508 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3509 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3512 #ifdef CONFIG_DEBUG_VM
3515 struct mem_cgroup_per_zone *mz;
3516 struct zone_reclaim_stat *rstat;
3517 unsigned long recent_rotated[2] = {0, 0};
3518 unsigned long recent_scanned[2] = {0, 0};
3520 for_each_online_node(nid)
3521 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3522 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3523 rstat = &mz->lruvec.reclaim_stat;
3525 recent_rotated[0] += rstat->recent_rotated[0];
3526 recent_rotated[1] += rstat->recent_rotated[1];
3527 recent_scanned[0] += rstat->recent_scanned[0];
3528 recent_scanned[1] += rstat->recent_scanned[1];
3530 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3531 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3532 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3533 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3540 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3543 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3545 return mem_cgroup_swappiness(memcg);
3548 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3549 struct cftype *cft, u64 val)
3551 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3557 memcg->swappiness = val;
3559 vm_swappiness = val;
3564 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3566 struct mem_cgroup_threshold_ary *t;
3567 unsigned long usage;
3572 t = rcu_dereference(memcg->thresholds.primary);
3574 t = rcu_dereference(memcg->memsw_thresholds.primary);
3579 usage = mem_cgroup_usage(memcg, swap);
3582 * current_threshold points to threshold just below or equal to usage.
3583 * If it's not true, a threshold was crossed after last
3584 * call of __mem_cgroup_threshold().
3586 i = t->current_threshold;
3589 * Iterate backward over array of thresholds starting from
3590 * current_threshold and check if a threshold is crossed.
3591 * If none of thresholds below usage is crossed, we read
3592 * only one element of the array here.
3594 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3595 eventfd_signal(t->entries[i].eventfd, 1);
3597 /* i = current_threshold + 1 */
3601 * Iterate forward over array of thresholds starting from
3602 * current_threshold+1 and check if a threshold is crossed.
3603 * If none of thresholds above usage is crossed, we read
3604 * only one element of the array here.
3606 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3607 eventfd_signal(t->entries[i].eventfd, 1);
3609 /* Update current_threshold */
3610 t->current_threshold = i - 1;
3615 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3618 __mem_cgroup_threshold(memcg, false);
3619 if (do_swap_account)
3620 __mem_cgroup_threshold(memcg, true);
3622 memcg = parent_mem_cgroup(memcg);
3626 static int compare_thresholds(const void *a, const void *b)
3628 const struct mem_cgroup_threshold *_a = a;
3629 const struct mem_cgroup_threshold *_b = b;
3631 if (_a->threshold > _b->threshold)
3634 if (_a->threshold < _b->threshold)
3640 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3642 struct mem_cgroup_eventfd_list *ev;
3644 spin_lock(&memcg_oom_lock);
3646 list_for_each_entry(ev, &memcg->oom_notify, list)
3647 eventfd_signal(ev->eventfd, 1);
3649 spin_unlock(&memcg_oom_lock);
3653 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3655 struct mem_cgroup *iter;
3657 for_each_mem_cgroup_tree(iter, memcg)
3658 mem_cgroup_oom_notify_cb(iter);
3661 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3662 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3664 struct mem_cgroup_thresholds *thresholds;
3665 struct mem_cgroup_threshold_ary *new;
3666 unsigned long threshold;
3667 unsigned long usage;
3670 ret = page_counter_memparse(args, "-1", &threshold);
3674 mutex_lock(&memcg->thresholds_lock);
3677 thresholds = &memcg->thresholds;
3678 usage = mem_cgroup_usage(memcg, false);
3679 } else if (type == _MEMSWAP) {
3680 thresholds = &memcg->memsw_thresholds;
3681 usage = mem_cgroup_usage(memcg, true);
3685 /* Check if a threshold crossed before adding a new one */
3686 if (thresholds->primary)
3687 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3689 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3691 /* Allocate memory for new array of thresholds */
3692 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3700 /* Copy thresholds (if any) to new array */
3701 if (thresholds->primary) {
3702 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3703 sizeof(struct mem_cgroup_threshold));
3706 /* Add new threshold */
3707 new->entries[size - 1].eventfd = eventfd;
3708 new->entries[size - 1].threshold = threshold;
3710 /* Sort thresholds. Registering of new threshold isn't time-critical */
3711 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3712 compare_thresholds, NULL);
3714 /* Find current threshold */
3715 new->current_threshold = -1;
3716 for (i = 0; i < size; i++) {
3717 if (new->entries[i].threshold <= usage) {
3719 * new->current_threshold will not be used until
3720 * rcu_assign_pointer(), so it's safe to increment
3723 ++new->current_threshold;
3728 /* Free old spare buffer and save old primary buffer as spare */
3729 kfree(thresholds->spare);
3730 thresholds->spare = thresholds->primary;
3732 rcu_assign_pointer(thresholds->primary, new);
3734 /* To be sure that nobody uses thresholds */
3738 mutex_unlock(&memcg->thresholds_lock);
3743 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3744 struct eventfd_ctx *eventfd, const char *args)
3746 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3749 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3750 struct eventfd_ctx *eventfd, const char *args)
3752 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3755 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3756 struct eventfd_ctx *eventfd, enum res_type type)
3758 struct mem_cgroup_thresholds *thresholds;
3759 struct mem_cgroup_threshold_ary *new;
3760 unsigned long usage;
3763 mutex_lock(&memcg->thresholds_lock);
3766 thresholds = &memcg->thresholds;
3767 usage = mem_cgroup_usage(memcg, false);
3768 } else if (type == _MEMSWAP) {
3769 thresholds = &memcg->memsw_thresholds;
3770 usage = mem_cgroup_usage(memcg, true);
3774 if (!thresholds->primary)
3777 /* Check if a threshold crossed before removing */
3778 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3780 /* Calculate new number of threshold */
3782 for (i = 0; i < thresholds->primary->size; i++) {
3783 if (thresholds->primary->entries[i].eventfd != eventfd)
3787 new = thresholds->spare;
3789 /* Set thresholds array to NULL if we don't have thresholds */
3798 /* Copy thresholds and find current threshold */
3799 new->current_threshold = -1;
3800 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3801 if (thresholds->primary->entries[i].eventfd == eventfd)
3804 new->entries[j] = thresholds->primary->entries[i];
3805 if (new->entries[j].threshold <= usage) {
3807 * new->current_threshold will not be used
3808 * until rcu_assign_pointer(), so it's safe to increment
3811 ++new->current_threshold;
3817 /* Swap primary and spare array */
3818 thresholds->spare = thresholds->primary;
3819 /* If all events are unregistered, free the spare array */
3821 kfree(thresholds->spare);
3822 thresholds->spare = NULL;
3825 rcu_assign_pointer(thresholds->primary, new);
3827 /* To be sure that nobody uses thresholds */
3830 mutex_unlock(&memcg->thresholds_lock);
3833 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3834 struct eventfd_ctx *eventfd)
3836 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
3839 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3840 struct eventfd_ctx *eventfd)
3842 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
3845 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
3846 struct eventfd_ctx *eventfd, const char *args)
3848 struct mem_cgroup_eventfd_list *event;
3850 event = kmalloc(sizeof(*event), GFP_KERNEL);
3854 spin_lock(&memcg_oom_lock);
3856 event->eventfd = eventfd;
3857 list_add(&event->list, &memcg->oom_notify);
3859 /* already in OOM ? */
3860 if (atomic_read(&memcg->under_oom))
3861 eventfd_signal(eventfd, 1);
3862 spin_unlock(&memcg_oom_lock);
3867 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
3868 struct eventfd_ctx *eventfd)
3870 struct mem_cgroup_eventfd_list *ev, *tmp;
3872 spin_lock(&memcg_oom_lock);
3874 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
3875 if (ev->eventfd == eventfd) {
3876 list_del(&ev->list);
3881 spin_unlock(&memcg_oom_lock);
3884 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
3886 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
3888 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
3889 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
3893 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
3894 struct cftype *cft, u64 val)
3896 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3898 /* cannot set to root cgroup and only 0 and 1 are allowed */
3899 if (!css->parent || !((val == 0) || (val == 1)))
3902 memcg->oom_kill_disable = val;
3904 memcg_oom_recover(memcg);
3909 #ifdef CONFIG_MEMCG_KMEM
3910 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3914 ret = memcg_propagate_kmem(memcg);
3918 return mem_cgroup_sockets_init(memcg, ss);
3921 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3923 struct cgroup_subsys_state *css;
3924 struct mem_cgroup *parent, *child;
3927 if (!memcg->kmem_acct_active)
3931 * Clear the 'active' flag before clearing memcg_caches arrays entries.
3932 * Since we take the slab_mutex in memcg_deactivate_kmem_caches(), it
3933 * guarantees no cache will be created for this cgroup after we are
3934 * done (see memcg_create_kmem_cache()).
3936 memcg->kmem_acct_active = false;
3938 memcg_deactivate_kmem_caches(memcg);
3940 kmemcg_id = memcg->kmemcg_id;
3941 BUG_ON(kmemcg_id < 0);
3943 parent = parent_mem_cgroup(memcg);
3945 parent = root_mem_cgroup;
3948 * Change kmemcg_id of this cgroup and all its descendants to the
3949 * parent's id, and then move all entries from this cgroup's list_lrus
3950 * to ones of the parent. After we have finished, all list_lrus
3951 * corresponding to this cgroup are guaranteed to remain empty. The
3952 * ordering is imposed by list_lru_node->lock taken by
3953 * memcg_drain_all_list_lrus().
3955 css_for_each_descendant_pre(css, &memcg->css) {
3956 child = mem_cgroup_from_css(css);
3957 BUG_ON(child->kmemcg_id != kmemcg_id);
3958 child->kmemcg_id = parent->kmemcg_id;
3959 if (!memcg->use_hierarchy)
3962 memcg_drain_all_list_lrus(kmemcg_id, parent->kmemcg_id);
3964 memcg_free_cache_id(kmemcg_id);
3967 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3969 if (memcg->kmem_acct_activated) {
3970 memcg_destroy_kmem_caches(memcg);
3971 static_key_slow_dec(&memcg_kmem_enabled_key);
3972 WARN_ON(page_counter_read(&memcg->kmem));
3974 mem_cgroup_sockets_destroy(memcg);
3977 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
3982 static void memcg_deactivate_kmem(struct mem_cgroup *memcg)
3986 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
3991 #ifdef CONFIG_CGROUP_WRITEBACK
3993 struct list_head *mem_cgroup_cgwb_list(struct mem_cgroup *memcg)
3995 return &memcg->cgwb_list;
3998 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4000 return wb_domain_init(&memcg->cgwb_domain, gfp);
4003 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4005 wb_domain_exit(&memcg->cgwb_domain);
4008 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4010 wb_domain_size_changed(&memcg->cgwb_domain);
4013 struct wb_domain *mem_cgroup_wb_domain(struct bdi_writeback *wb)
4015 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4017 if (!memcg->css.parent)
4020 return &memcg->cgwb_domain;
4024 * mem_cgroup_wb_stats - retrieve writeback related stats from its memcg
4025 * @wb: bdi_writeback in question
4026 * @pavail: out parameter for number of available pages
4027 * @pdirty: out parameter for number of dirty pages
4028 * @pwriteback: out parameter for number of pages under writeback
4030 * Determine the numbers of available, dirty, and writeback pages in @wb's
4031 * memcg. Dirty and writeback are self-explanatory. Available is a bit
4034 * A memcg's headroom is "min(max, high) - used". The available memory is
4035 * calculated as the lowest headroom of itself and the ancestors plus the
4036 * number of pages already being used for file pages. Note that this
4037 * doesn't consider the actual amount of available memory in the system.
4038 * The caller should further cap *@pavail accordingly.
4040 void mem_cgroup_wb_stats(struct bdi_writeback *wb, unsigned long *pavail,
4041 unsigned long *pdirty, unsigned long *pwriteback)
4043 struct mem_cgroup *memcg = mem_cgroup_from_css(wb->memcg_css);
4044 struct mem_cgroup *parent;
4045 unsigned long head_room = PAGE_COUNTER_MAX;
4046 unsigned long file_pages;
4048 *pdirty = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_DIRTY);
4050 /* this should eventually include NR_UNSTABLE_NFS */
4051 *pwriteback = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_WRITEBACK);
4053 file_pages = mem_cgroup_nr_lru_pages(memcg, (1 << LRU_INACTIVE_FILE) |
4054 (1 << LRU_ACTIVE_FILE));
4055 while ((parent = parent_mem_cgroup(memcg))) {
4056 unsigned long ceiling = min(memcg->memory.limit, memcg->high);
4057 unsigned long used = page_counter_read(&memcg->memory);
4059 head_room = min(head_room, ceiling - min(ceiling, used));
4063 *pavail = file_pages + head_room;
4066 #else /* CONFIG_CGROUP_WRITEBACK */
4068 static int memcg_wb_domain_init(struct mem_cgroup *memcg, gfp_t gfp)
4073 static void memcg_wb_domain_exit(struct mem_cgroup *memcg)
4077 static void memcg_wb_domain_size_changed(struct mem_cgroup *memcg)
4081 #endif /* CONFIG_CGROUP_WRITEBACK */
4084 * DO NOT USE IN NEW FILES.
4086 * "cgroup.event_control" implementation.
4088 * This is way over-engineered. It tries to support fully configurable
4089 * events for each user. Such level of flexibility is completely
4090 * unnecessary especially in the light of the planned unified hierarchy.
4092 * Please deprecate this and replace with something simpler if at all
4097 * Unregister event and free resources.
4099 * Gets called from workqueue.
4101 static void memcg_event_remove(struct work_struct *work)
4103 struct mem_cgroup_event *event =
4104 container_of(work, struct mem_cgroup_event, remove);
4105 struct mem_cgroup *memcg = event->memcg;
4107 remove_wait_queue(event->wqh, &event->wait);
4109 event->unregister_event(memcg, event->eventfd);
4111 /* Notify userspace the event is going away. */
4112 eventfd_signal(event->eventfd, 1);
4114 eventfd_ctx_put(event->eventfd);
4116 css_put(&memcg->css);
4120 * Gets called on POLLHUP on eventfd when user closes it.
4122 * Called with wqh->lock held and interrupts disabled.
4124 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4125 int sync, void *key)
4127 struct mem_cgroup_event *event =
4128 container_of(wait, struct mem_cgroup_event, wait);
4129 struct mem_cgroup *memcg = event->memcg;
4130 unsigned long flags = (unsigned long)key;
4132 if (flags & POLLHUP) {
4134 * If the event has been detached at cgroup removal, we
4135 * can simply return knowing the other side will cleanup
4138 * We can't race against event freeing since the other
4139 * side will require wqh->lock via remove_wait_queue(),
4142 spin_lock(&memcg->event_list_lock);
4143 if (!list_empty(&event->list)) {
4144 list_del_init(&event->list);
4146 * We are in atomic context, but cgroup_event_remove()
4147 * may sleep, so we have to call it in workqueue.
4149 schedule_work(&event->remove);
4151 spin_unlock(&memcg->event_list_lock);
4157 static void memcg_event_ptable_queue_proc(struct file *file,
4158 wait_queue_head_t *wqh, poll_table *pt)
4160 struct mem_cgroup_event *event =
4161 container_of(pt, struct mem_cgroup_event, pt);
4164 add_wait_queue(wqh, &event->wait);
4168 * DO NOT USE IN NEW FILES.
4170 * Parse input and register new cgroup event handler.
4172 * Input must be in format '<event_fd> <control_fd> <args>'.
4173 * Interpretation of args is defined by control file implementation.
4175 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4176 char *buf, size_t nbytes, loff_t off)
4178 struct cgroup_subsys_state *css = of_css(of);
4179 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4180 struct mem_cgroup_event *event;
4181 struct cgroup_subsys_state *cfile_css;
4182 unsigned int efd, cfd;
4189 buf = strstrip(buf);
4191 efd = simple_strtoul(buf, &endp, 10);
4196 cfd = simple_strtoul(buf, &endp, 10);
4197 if ((*endp != ' ') && (*endp != '\0'))
4201 event = kzalloc(sizeof(*event), GFP_KERNEL);
4205 event->memcg = memcg;
4206 INIT_LIST_HEAD(&event->list);
4207 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4208 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4209 INIT_WORK(&event->remove, memcg_event_remove);
4217 event->eventfd = eventfd_ctx_fileget(efile.file);
4218 if (IS_ERR(event->eventfd)) {
4219 ret = PTR_ERR(event->eventfd);
4226 goto out_put_eventfd;
4229 /* the process need read permission on control file */
4230 /* AV: shouldn't we check that it's been opened for read instead? */
4231 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4236 * Determine the event callbacks and set them in @event. This used
4237 * to be done via struct cftype but cgroup core no longer knows
4238 * about these events. The following is crude but the whole thing
4239 * is for compatibility anyway.
4241 * DO NOT ADD NEW FILES.
4243 name = cfile.file->f_path.dentry->d_name.name;
4245 if (!strcmp(name, "memory.usage_in_bytes")) {
4246 event->register_event = mem_cgroup_usage_register_event;
4247 event->unregister_event = mem_cgroup_usage_unregister_event;
4248 } else if (!strcmp(name, "memory.oom_control")) {
4249 event->register_event = mem_cgroup_oom_register_event;
4250 event->unregister_event = mem_cgroup_oom_unregister_event;
4251 } else if (!strcmp(name, "memory.pressure_level")) {
4252 event->register_event = vmpressure_register_event;
4253 event->unregister_event = vmpressure_unregister_event;
4254 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4255 event->register_event = memsw_cgroup_usage_register_event;
4256 event->unregister_event = memsw_cgroup_usage_unregister_event;
4263 * Verify @cfile should belong to @css. Also, remaining events are
4264 * automatically removed on cgroup destruction but the removal is
4265 * asynchronous, so take an extra ref on @css.
4267 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4268 &memory_cgrp_subsys);
4270 if (IS_ERR(cfile_css))
4272 if (cfile_css != css) {
4277 ret = event->register_event(memcg, event->eventfd, buf);
4281 efile.file->f_op->poll(efile.file, &event->pt);
4283 spin_lock(&memcg->event_list_lock);
4284 list_add(&event->list, &memcg->event_list);
4285 spin_unlock(&memcg->event_list_lock);
4297 eventfd_ctx_put(event->eventfd);
4306 static struct cftype mem_cgroup_legacy_files[] = {
4308 .name = "usage_in_bytes",
4309 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4310 .read_u64 = mem_cgroup_read_u64,
4313 .name = "max_usage_in_bytes",
4314 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4315 .write = mem_cgroup_reset,
4316 .read_u64 = mem_cgroup_read_u64,
4319 .name = "limit_in_bytes",
4320 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4321 .write = mem_cgroup_write,
4322 .read_u64 = mem_cgroup_read_u64,
4325 .name = "soft_limit_in_bytes",
4326 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4327 .write = mem_cgroup_write,
4328 .read_u64 = mem_cgroup_read_u64,
4332 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4333 .write = mem_cgroup_reset,
4334 .read_u64 = mem_cgroup_read_u64,
4338 .seq_show = memcg_stat_show,
4341 .name = "force_empty",
4342 .write = mem_cgroup_force_empty_write,
4345 .name = "use_hierarchy",
4346 .write_u64 = mem_cgroup_hierarchy_write,
4347 .read_u64 = mem_cgroup_hierarchy_read,
4350 .name = "cgroup.event_control", /* XXX: for compat */
4351 .write = memcg_write_event_control,
4352 .flags = CFTYPE_NO_PREFIX,
4356 .name = "swappiness",
4357 .read_u64 = mem_cgroup_swappiness_read,
4358 .write_u64 = mem_cgroup_swappiness_write,
4361 .name = "move_charge_at_immigrate",
4362 .read_u64 = mem_cgroup_move_charge_read,
4363 .write_u64 = mem_cgroup_move_charge_write,
4366 .name = "oom_control",
4367 .seq_show = mem_cgroup_oom_control_read,
4368 .write_u64 = mem_cgroup_oom_control_write,
4369 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4372 .name = "pressure_level",
4376 .name = "numa_stat",
4377 .seq_show = memcg_numa_stat_show,
4380 #ifdef CONFIG_MEMCG_KMEM
4382 .name = "kmem.limit_in_bytes",
4383 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4384 .write = mem_cgroup_write,
4385 .read_u64 = mem_cgroup_read_u64,
4388 .name = "kmem.usage_in_bytes",
4389 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4390 .read_u64 = mem_cgroup_read_u64,
4393 .name = "kmem.failcnt",
4394 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4395 .write = mem_cgroup_reset,
4396 .read_u64 = mem_cgroup_read_u64,
4399 .name = "kmem.max_usage_in_bytes",
4400 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4401 .write = mem_cgroup_reset,
4402 .read_u64 = mem_cgroup_read_u64,
4404 #ifdef CONFIG_SLABINFO
4406 .name = "kmem.slabinfo",
4407 .seq_start = slab_start,
4408 .seq_next = slab_next,
4409 .seq_stop = slab_stop,
4410 .seq_show = memcg_slab_show,
4414 { }, /* terminate */
4417 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4419 struct mem_cgroup_per_node *pn;
4420 struct mem_cgroup_per_zone *mz;
4421 int zone, tmp = node;
4423 * This routine is called against possible nodes.
4424 * But it's BUG to call kmalloc() against offline node.
4426 * TODO: this routine can waste much memory for nodes which will
4427 * never be onlined. It's better to use memory hotplug callback
4430 if (!node_state(node, N_NORMAL_MEMORY))
4432 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4436 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4437 mz = &pn->zoneinfo[zone];
4438 lruvec_init(&mz->lruvec);
4439 mz->usage_in_excess = 0;
4440 mz->on_tree = false;
4443 memcg->nodeinfo[node] = pn;
4447 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4449 kfree(memcg->nodeinfo[node]);
4452 static struct mem_cgroup *mem_cgroup_alloc(void)
4454 struct mem_cgroup *memcg;
4457 size = sizeof(struct mem_cgroup);
4458 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4460 memcg = kzalloc(size, GFP_KERNEL);
4464 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4468 if (memcg_wb_domain_init(memcg, GFP_KERNEL))
4471 spin_lock_init(&memcg->pcp_counter_lock);
4475 free_percpu(memcg->stat);
4482 * At destroying mem_cgroup, references from swap_cgroup can remain.
4483 * (scanning all at force_empty is too costly...)
4485 * Instead of clearing all references at force_empty, we remember
4486 * the number of reference from swap_cgroup and free mem_cgroup when
4487 * it goes down to 0.
4489 * Removal of cgroup itself succeeds regardless of refs from swap.
4492 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4496 mem_cgroup_remove_from_trees(memcg);
4499 free_mem_cgroup_per_zone_info(memcg, node);
4501 free_percpu(memcg->stat);
4502 memcg_wb_domain_exit(memcg);
4507 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4509 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4511 if (!memcg->memory.parent)
4513 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4515 EXPORT_SYMBOL(parent_mem_cgroup);
4517 static struct cgroup_subsys_state * __ref
4518 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4520 struct mem_cgroup *memcg;
4521 long error = -ENOMEM;
4524 memcg = mem_cgroup_alloc();
4526 return ERR_PTR(error);
4529 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4533 if (parent_css == NULL) {
4534 root_mem_cgroup = memcg;
4535 mem_cgroup_root_css = &memcg->css;
4536 page_counter_init(&memcg->memory, NULL);
4537 memcg->high = PAGE_COUNTER_MAX;
4538 memcg->soft_limit = PAGE_COUNTER_MAX;
4539 page_counter_init(&memcg->memsw, NULL);
4540 page_counter_init(&memcg->kmem, NULL);
4543 memcg->last_scanned_node = MAX_NUMNODES;
4544 INIT_LIST_HEAD(&memcg->oom_notify);
4545 memcg->move_charge_at_immigrate = 0;
4546 mutex_init(&memcg->thresholds_lock);
4547 spin_lock_init(&memcg->move_lock);
4548 vmpressure_init(&memcg->vmpressure);
4549 INIT_LIST_HEAD(&memcg->event_list);
4550 spin_lock_init(&memcg->event_list_lock);
4551 #ifdef CONFIG_MEMCG_KMEM
4552 memcg->kmemcg_id = -1;
4554 #ifdef CONFIG_CGROUP_WRITEBACK
4555 INIT_LIST_HEAD(&memcg->cgwb_list);
4560 __mem_cgroup_free(memcg);
4561 return ERR_PTR(error);
4565 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4567 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4568 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4571 if (css->id > MEM_CGROUP_ID_MAX)
4577 mutex_lock(&memcg_create_mutex);
4579 memcg->use_hierarchy = parent->use_hierarchy;
4580 memcg->oom_kill_disable = parent->oom_kill_disable;
4581 memcg->swappiness = mem_cgroup_swappiness(parent);
4583 if (parent->use_hierarchy) {
4584 page_counter_init(&memcg->memory, &parent->memory);
4585 memcg->high = PAGE_COUNTER_MAX;
4586 memcg->soft_limit = PAGE_COUNTER_MAX;
4587 page_counter_init(&memcg->memsw, &parent->memsw);
4588 page_counter_init(&memcg->kmem, &parent->kmem);
4591 * No need to take a reference to the parent because cgroup
4592 * core guarantees its existence.
4595 page_counter_init(&memcg->memory, NULL);
4596 memcg->high = PAGE_COUNTER_MAX;
4597 memcg->soft_limit = PAGE_COUNTER_MAX;
4598 page_counter_init(&memcg->memsw, NULL);
4599 page_counter_init(&memcg->kmem, NULL);
4601 * Deeper hierachy with use_hierarchy == false doesn't make
4602 * much sense so let cgroup subsystem know about this
4603 * unfortunate state in our controller.
4605 if (parent != root_mem_cgroup)
4606 memory_cgrp_subsys.broken_hierarchy = true;
4608 mutex_unlock(&memcg_create_mutex);
4610 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4615 * Make sure the memcg is initialized: mem_cgroup_iter()
4616 * orders reading memcg->initialized against its callers
4617 * reading the memcg members.
4619 smp_store_release(&memcg->initialized, 1);
4624 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4626 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4627 struct mem_cgroup_event *event, *tmp;
4630 * Unregister events and notify userspace.
4631 * Notify userspace about cgroup removing only after rmdir of cgroup
4632 * directory to avoid race between userspace and kernelspace.
4634 spin_lock(&memcg->event_list_lock);
4635 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4636 list_del_init(&event->list);
4637 schedule_work(&event->remove);
4639 spin_unlock(&memcg->event_list_lock);
4641 vmpressure_cleanup(&memcg->vmpressure);
4643 memcg_deactivate_kmem(memcg);
4645 wb_memcg_offline(memcg);
4648 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4650 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4652 memcg_destroy_kmem(memcg);
4653 __mem_cgroup_free(memcg);
4657 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4658 * @css: the target css
4660 * Reset the states of the mem_cgroup associated with @css. This is
4661 * invoked when the userland requests disabling on the default hierarchy
4662 * but the memcg is pinned through dependency. The memcg should stop
4663 * applying policies and should revert to the vanilla state as it may be
4664 * made visible again.
4666 * The current implementation only resets the essential configurations.
4667 * This needs to be expanded to cover all the visible parts.
4669 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4671 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4673 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4674 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4675 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4677 memcg->high = PAGE_COUNTER_MAX;
4678 memcg->soft_limit = PAGE_COUNTER_MAX;
4679 memcg_wb_domain_size_changed(memcg);
4683 /* Handlers for move charge at task migration. */
4684 static int mem_cgroup_do_precharge(unsigned long count)
4688 /* Try a single bulk charge without reclaim first */
4689 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4691 mc.precharge += count;
4694 if (ret == -EINTR) {
4695 cancel_charge(root_mem_cgroup, count);
4699 /* Try charges one by one with reclaim */
4701 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4703 * In case of failure, any residual charges against
4704 * mc.to will be dropped by mem_cgroup_clear_mc()
4705 * later on. However, cancel any charges that are
4706 * bypassed to root right away or they'll be lost.
4709 cancel_charge(root_mem_cgroup, 1);
4719 * get_mctgt_type - get target type of moving charge
4720 * @vma: the vma the pte to be checked belongs
4721 * @addr: the address corresponding to the pte to be checked
4722 * @ptent: the pte to be checked
4723 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4726 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4727 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4728 * move charge. if @target is not NULL, the page is stored in target->page
4729 * with extra refcnt got(Callers should handle it).
4730 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4731 * target for charge migration. if @target is not NULL, the entry is stored
4734 * Called with pte lock held.
4741 enum mc_target_type {
4747 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4748 unsigned long addr, pte_t ptent)
4750 struct page *page = vm_normal_page(vma, addr, ptent);
4752 if (!page || !page_mapped(page))
4754 if (PageAnon(page)) {
4755 if (!(mc.flags & MOVE_ANON))
4758 if (!(mc.flags & MOVE_FILE))
4761 if (!get_page_unless_zero(page))
4768 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4769 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4771 struct page *page = NULL;
4772 swp_entry_t ent = pte_to_swp_entry(ptent);
4774 if (!(mc.flags & MOVE_ANON) || non_swap_entry(ent))
4777 * Because lookup_swap_cache() updates some statistics counter,
4778 * we call find_get_page() with swapper_space directly.
4780 page = find_get_page(swap_address_space(ent), ent.val);
4781 if (do_swap_account)
4782 entry->val = ent.val;
4787 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4788 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4794 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4795 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4797 struct page *page = NULL;
4798 struct address_space *mapping;
4801 if (!vma->vm_file) /* anonymous vma */
4803 if (!(mc.flags & MOVE_FILE))
4806 mapping = vma->vm_file->f_mapping;
4807 pgoff = linear_page_index(vma, addr);
4809 /* page is moved even if it's not RSS of this task(page-faulted). */
4811 /* shmem/tmpfs may report page out on swap: account for that too. */
4812 if (shmem_mapping(mapping)) {
4813 page = find_get_entry(mapping, pgoff);
4814 if (radix_tree_exceptional_entry(page)) {
4815 swp_entry_t swp = radix_to_swp_entry(page);
4816 if (do_swap_account)
4818 page = find_get_page(swap_address_space(swp), swp.val);
4821 page = find_get_page(mapping, pgoff);
4823 page = find_get_page(mapping, pgoff);
4829 * mem_cgroup_move_account - move account of the page
4831 * @nr_pages: number of regular pages (>1 for huge pages)
4832 * @from: mem_cgroup which the page is moved from.
4833 * @to: mem_cgroup which the page is moved to. @from != @to.
4835 * The caller must confirm following.
4836 * - page is not on LRU (isolate_page() is useful.)
4837 * - compound_lock is held when nr_pages > 1
4839 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
4842 static int mem_cgroup_move_account(struct page *page,
4843 unsigned int nr_pages,
4844 struct mem_cgroup *from,
4845 struct mem_cgroup *to)
4847 unsigned long flags;
4851 VM_BUG_ON(from == to);
4852 VM_BUG_ON_PAGE(PageLRU(page), page);
4854 * The page is isolated from LRU. So, collapse function
4855 * will not handle this page. But page splitting can happen.
4856 * Do this check under compound_page_lock(). The caller should
4860 if (nr_pages > 1 && !PageTransHuge(page))
4864 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
4865 * of its source page while we change it: page migration takes
4866 * both pages off the LRU, but page cache replacement doesn't.
4868 if (!trylock_page(page))
4872 if (page->mem_cgroup != from)
4875 anon = PageAnon(page);
4877 spin_lock_irqsave(&from->move_lock, flags);
4879 if (!anon && page_mapped(page)) {
4880 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4882 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
4887 * move_lock grabbed above and caller set from->moving_account, so
4888 * mem_cgroup_update_page_stat() will serialize updates to PageDirty.
4889 * So mapping should be stable for dirty pages.
4891 if (!anon && PageDirty(page)) {
4892 struct address_space *mapping = page_mapping(page);
4894 if (mapping_cap_account_dirty(mapping)) {
4895 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_DIRTY],
4897 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_DIRTY],
4902 if (PageWriteback(page)) {
4903 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4905 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
4910 * It is safe to change page->mem_cgroup here because the page
4911 * is referenced, charged, and isolated - we can't race with
4912 * uncharging, charging, migration, or LRU putback.
4915 /* caller should have done css_get */
4916 page->mem_cgroup = to;
4917 spin_unlock_irqrestore(&from->move_lock, flags);
4921 local_irq_disable();
4922 mem_cgroup_charge_statistics(to, page, nr_pages);
4923 memcg_check_events(to, page);
4924 mem_cgroup_charge_statistics(from, page, -nr_pages);
4925 memcg_check_events(from, page);
4933 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4934 unsigned long addr, pte_t ptent, union mc_target *target)
4936 struct page *page = NULL;
4937 enum mc_target_type ret = MC_TARGET_NONE;
4938 swp_entry_t ent = { .val = 0 };
4940 if (pte_present(ptent))
4941 page = mc_handle_present_pte(vma, addr, ptent);
4942 else if (is_swap_pte(ptent))
4943 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4944 else if (pte_none(ptent))
4945 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4947 if (!page && !ent.val)
4951 * Do only loose check w/o serialization.
4952 * mem_cgroup_move_account() checks the page is valid or
4953 * not under LRU exclusion.
4955 if (page->mem_cgroup == mc.from) {
4956 ret = MC_TARGET_PAGE;
4958 target->page = page;
4960 if (!ret || !target)
4963 /* There is a swap entry and a page doesn't exist or isn't charged */
4964 if (ent.val && !ret &&
4965 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4966 ret = MC_TARGET_SWAP;
4973 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4975 * We don't consider swapping or file mapped pages because THP does not
4976 * support them for now.
4977 * Caller should make sure that pmd_trans_huge(pmd) is true.
4979 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4980 unsigned long addr, pmd_t pmd, union mc_target *target)
4982 struct page *page = NULL;
4983 enum mc_target_type ret = MC_TARGET_NONE;
4985 page = pmd_page(pmd);
4986 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4987 if (!(mc.flags & MOVE_ANON))
4989 if (page->mem_cgroup == mc.from) {
4990 ret = MC_TARGET_PAGE;
4993 target->page = page;
4999 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5000 unsigned long addr, pmd_t pmd, union mc_target *target)
5002 return MC_TARGET_NONE;
5006 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5007 unsigned long addr, unsigned long end,
5008 struct mm_walk *walk)
5010 struct vm_area_struct *vma = walk->vma;
5014 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5015 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5016 mc.precharge += HPAGE_PMD_NR;
5021 if (pmd_trans_unstable(pmd))
5023 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5024 for (; addr != end; pte++, addr += PAGE_SIZE)
5025 if (get_mctgt_type(vma, addr, *pte, NULL))
5026 mc.precharge++; /* increment precharge temporarily */
5027 pte_unmap_unlock(pte - 1, ptl);
5033 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5035 unsigned long precharge;
5037 struct mm_walk mem_cgroup_count_precharge_walk = {
5038 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5041 down_read(&mm->mmap_sem);
5042 walk_page_range(0, ~0UL, &mem_cgroup_count_precharge_walk);
5043 up_read(&mm->mmap_sem);
5045 precharge = mc.precharge;
5051 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5053 unsigned long precharge = mem_cgroup_count_precharge(mm);
5055 VM_BUG_ON(mc.moving_task);
5056 mc.moving_task = current;
5057 return mem_cgroup_do_precharge(precharge);
5060 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5061 static void __mem_cgroup_clear_mc(void)
5063 struct mem_cgroup *from = mc.from;
5064 struct mem_cgroup *to = mc.to;
5066 /* we must uncharge all the leftover precharges from mc.to */
5068 cancel_charge(mc.to, mc.precharge);
5072 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5073 * we must uncharge here.
5075 if (mc.moved_charge) {
5076 cancel_charge(mc.from, mc.moved_charge);
5077 mc.moved_charge = 0;
5079 /* we must fixup refcnts and charges */
5080 if (mc.moved_swap) {
5081 /* uncharge swap account from the old cgroup */
5082 if (!mem_cgroup_is_root(mc.from))
5083 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5086 * we charged both to->memory and to->memsw, so we
5087 * should uncharge to->memory.
5089 if (!mem_cgroup_is_root(mc.to))
5090 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5092 css_put_many(&mc.from->css, mc.moved_swap);
5094 /* we've already done css_get(mc.to) */
5097 memcg_oom_recover(from);
5098 memcg_oom_recover(to);
5099 wake_up_all(&mc.waitq);
5102 static void mem_cgroup_clear_mc(void)
5105 * we must clear moving_task before waking up waiters at the end of
5108 mc.moving_task = NULL;
5109 __mem_cgroup_clear_mc();
5110 spin_lock(&mc.lock);
5113 spin_unlock(&mc.lock);
5116 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5117 struct cgroup_taskset *tset)
5119 struct task_struct *p = cgroup_taskset_first(tset);
5121 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5122 unsigned long move_flags;
5125 * We are now commited to this value whatever it is. Changes in this
5126 * tunable will only affect upcoming migrations, not the current one.
5127 * So we need to save it, and keep it going.
5129 move_flags = READ_ONCE(memcg->move_charge_at_immigrate);
5131 struct mm_struct *mm;
5132 struct mem_cgroup *from = mem_cgroup_from_task(p);
5134 VM_BUG_ON(from == memcg);
5136 mm = get_task_mm(p);
5139 /* We move charges only when we move a owner of the mm */
5140 if (mm->owner == p) {
5143 VM_BUG_ON(mc.precharge);
5144 VM_BUG_ON(mc.moved_charge);
5145 VM_BUG_ON(mc.moved_swap);
5147 spin_lock(&mc.lock);
5150 mc.flags = move_flags;
5151 spin_unlock(&mc.lock);
5152 /* We set mc.moving_task later */
5154 ret = mem_cgroup_precharge_mc(mm);
5156 mem_cgroup_clear_mc();
5163 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5164 struct cgroup_taskset *tset)
5167 mem_cgroup_clear_mc();
5170 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5171 unsigned long addr, unsigned long end,
5172 struct mm_walk *walk)
5175 struct vm_area_struct *vma = walk->vma;
5178 enum mc_target_type target_type;
5179 union mc_target target;
5183 * We don't take compound_lock() here but no race with splitting thp
5185 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5186 * under splitting, which means there's no concurrent thp split,
5187 * - if another thread runs into split_huge_page() just after we
5188 * entered this if-block, the thread must wait for page table lock
5189 * to be unlocked in __split_huge_page_splitting(), where the main
5190 * part of thp split is not executed yet.
5192 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5193 if (mc.precharge < HPAGE_PMD_NR) {
5197 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5198 if (target_type == MC_TARGET_PAGE) {
5200 if (!isolate_lru_page(page)) {
5201 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5203 mc.precharge -= HPAGE_PMD_NR;
5204 mc.moved_charge += HPAGE_PMD_NR;
5206 putback_lru_page(page);
5214 if (pmd_trans_unstable(pmd))
5217 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5218 for (; addr != end; addr += PAGE_SIZE) {
5219 pte_t ptent = *(pte++);
5225 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5226 case MC_TARGET_PAGE:
5228 if (isolate_lru_page(page))
5230 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5232 /* we uncharge from mc.from later. */
5235 putback_lru_page(page);
5236 put: /* get_mctgt_type() gets the page */
5239 case MC_TARGET_SWAP:
5241 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5243 /* we fixup refcnts and charges later. */
5251 pte_unmap_unlock(pte - 1, ptl);
5256 * We have consumed all precharges we got in can_attach().
5257 * We try charge one by one, but don't do any additional
5258 * charges to mc.to if we have failed in charge once in attach()
5261 ret = mem_cgroup_do_precharge(1);
5269 static void mem_cgroup_move_charge(struct mm_struct *mm)
5271 struct mm_walk mem_cgroup_move_charge_walk = {
5272 .pmd_entry = mem_cgroup_move_charge_pte_range,
5276 lru_add_drain_all();
5278 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5279 * move_lock while we're moving its pages to another memcg.
5280 * Then wait for already started RCU-only updates to finish.
5282 atomic_inc(&mc.from->moving_account);
5285 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5287 * Someone who are holding the mmap_sem might be waiting in
5288 * waitq. So we cancel all extra charges, wake up all waiters,
5289 * and retry. Because we cancel precharges, we might not be able
5290 * to move enough charges, but moving charge is a best-effort
5291 * feature anyway, so it wouldn't be a big problem.
5293 __mem_cgroup_clear_mc();
5298 * When we have consumed all precharges and failed in doing
5299 * additional charge, the page walk just aborts.
5301 walk_page_range(0, ~0UL, &mem_cgroup_move_charge_walk);
5302 up_read(&mm->mmap_sem);
5303 atomic_dec(&mc.from->moving_account);
5306 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5307 struct cgroup_taskset *tset)
5309 struct task_struct *p = cgroup_taskset_first(tset);
5310 struct mm_struct *mm = get_task_mm(p);
5314 mem_cgroup_move_charge(mm);
5318 mem_cgroup_clear_mc();
5320 #else /* !CONFIG_MMU */
5321 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5322 struct cgroup_taskset *tset)
5326 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5327 struct cgroup_taskset *tset)
5330 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5331 struct cgroup_taskset *tset)
5337 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5338 * to verify whether we're attached to the default hierarchy on each mount
5341 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5344 * use_hierarchy is forced on the default hierarchy. cgroup core
5345 * guarantees that @root doesn't have any children, so turning it
5346 * on for the root memcg is enough.
5348 if (cgroup_on_dfl(root_css->cgroup))
5349 root_mem_cgroup->use_hierarchy = true;
5351 root_mem_cgroup->use_hierarchy = false;
5354 static u64 memory_current_read(struct cgroup_subsys_state *css,
5357 return mem_cgroup_usage(mem_cgroup_from_css(css), false);
5360 static int memory_low_show(struct seq_file *m, void *v)
5362 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5363 unsigned long low = READ_ONCE(memcg->low);
5365 if (low == PAGE_COUNTER_MAX)
5366 seq_puts(m, "max\n");
5368 seq_printf(m, "%llu\n", (u64)low * PAGE_SIZE);
5373 static ssize_t memory_low_write(struct kernfs_open_file *of,
5374 char *buf, size_t nbytes, loff_t off)
5376 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5380 buf = strstrip(buf);
5381 err = page_counter_memparse(buf, "max", &low);
5390 static int memory_high_show(struct seq_file *m, void *v)
5392 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5393 unsigned long high = READ_ONCE(memcg->high);
5395 if (high == PAGE_COUNTER_MAX)
5396 seq_puts(m, "max\n");
5398 seq_printf(m, "%llu\n", (u64)high * PAGE_SIZE);
5403 static ssize_t memory_high_write(struct kernfs_open_file *of,
5404 char *buf, size_t nbytes, loff_t off)
5406 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5410 buf = strstrip(buf);
5411 err = page_counter_memparse(buf, "max", &high);
5417 memcg_wb_domain_size_changed(memcg);
5421 static int memory_max_show(struct seq_file *m, void *v)
5423 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5424 unsigned long max = READ_ONCE(memcg->memory.limit);
5426 if (max == PAGE_COUNTER_MAX)
5427 seq_puts(m, "max\n");
5429 seq_printf(m, "%llu\n", (u64)max * PAGE_SIZE);
5434 static ssize_t memory_max_write(struct kernfs_open_file *of,
5435 char *buf, size_t nbytes, loff_t off)
5437 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5441 buf = strstrip(buf);
5442 err = page_counter_memparse(buf, "max", &max);
5446 err = mem_cgroup_resize_limit(memcg, max);
5450 memcg_wb_domain_size_changed(memcg);
5454 static int memory_events_show(struct seq_file *m, void *v)
5456 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5458 seq_printf(m, "low %lu\n", mem_cgroup_read_events(memcg, MEMCG_LOW));
5459 seq_printf(m, "high %lu\n", mem_cgroup_read_events(memcg, MEMCG_HIGH));
5460 seq_printf(m, "max %lu\n", mem_cgroup_read_events(memcg, MEMCG_MAX));
5461 seq_printf(m, "oom %lu\n", mem_cgroup_read_events(memcg, MEMCG_OOM));
5466 static struct cftype memory_files[] = {
5469 .read_u64 = memory_current_read,
5473 .flags = CFTYPE_NOT_ON_ROOT,
5474 .seq_show = memory_low_show,
5475 .write = memory_low_write,
5479 .flags = CFTYPE_NOT_ON_ROOT,
5480 .seq_show = memory_high_show,
5481 .write = memory_high_write,
5485 .flags = CFTYPE_NOT_ON_ROOT,
5486 .seq_show = memory_max_show,
5487 .write = memory_max_write,
5491 .flags = CFTYPE_NOT_ON_ROOT,
5492 .seq_show = memory_events_show,
5497 struct cgroup_subsys memory_cgrp_subsys = {
5498 .css_alloc = mem_cgroup_css_alloc,
5499 .css_online = mem_cgroup_css_online,
5500 .css_offline = mem_cgroup_css_offline,
5501 .css_free = mem_cgroup_css_free,
5502 .css_reset = mem_cgroup_css_reset,
5503 .can_attach = mem_cgroup_can_attach,
5504 .cancel_attach = mem_cgroup_cancel_attach,
5505 .attach = mem_cgroup_move_task,
5506 .bind = mem_cgroup_bind,
5507 .dfl_cftypes = memory_files,
5508 .legacy_cftypes = mem_cgroup_legacy_files,
5513 * mem_cgroup_events - count memory events against a cgroup
5514 * @memcg: the memory cgroup
5515 * @idx: the event index
5516 * @nr: the number of events to account for
5518 void mem_cgroup_events(struct mem_cgroup *memcg,
5519 enum mem_cgroup_events_index idx,
5522 this_cpu_add(memcg->stat->events[idx], nr);
5526 * mem_cgroup_low - check if memory consumption is below the normal range
5527 * @root: the highest ancestor to consider
5528 * @memcg: the memory cgroup to check
5530 * Returns %true if memory consumption of @memcg, and that of all
5531 * configurable ancestors up to @root, is below the normal range.
5533 bool mem_cgroup_low(struct mem_cgroup *root, struct mem_cgroup *memcg)
5535 if (mem_cgroup_disabled())
5539 * The toplevel group doesn't have a configurable range, so
5540 * it's never low when looked at directly, and it is not
5541 * considered an ancestor when assessing the hierarchy.
5544 if (memcg == root_mem_cgroup)
5547 if (page_counter_read(&memcg->memory) >= memcg->low)
5550 while (memcg != root) {
5551 memcg = parent_mem_cgroup(memcg);
5553 if (memcg == root_mem_cgroup)
5556 if (page_counter_read(&memcg->memory) >= memcg->low)
5563 * mem_cgroup_try_charge - try charging a page
5564 * @page: page to charge
5565 * @mm: mm context of the victim
5566 * @gfp_mask: reclaim mode
5567 * @memcgp: charged memcg return
5569 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5570 * pages according to @gfp_mask if necessary.
5572 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5573 * Otherwise, an error code is returned.
5575 * After page->mapping has been set up, the caller must finalize the
5576 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5577 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5579 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5580 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5582 struct mem_cgroup *memcg = NULL;
5583 unsigned int nr_pages = 1;
5586 if (mem_cgroup_disabled())
5589 if (PageSwapCache(page)) {
5591 * Every swap fault against a single page tries to charge the
5592 * page, bail as early as possible. shmem_unuse() encounters
5593 * already charged pages, too. The USED bit is protected by
5594 * the page lock, which serializes swap cache removal, which
5595 * in turn serializes uncharging.
5597 if (page->mem_cgroup)
5601 if (PageTransHuge(page)) {
5602 nr_pages <<= compound_order(page);
5603 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5606 if (do_swap_account && PageSwapCache(page))
5607 memcg = try_get_mem_cgroup_from_page(page);
5609 memcg = get_mem_cgroup_from_mm(mm);
5611 ret = try_charge(memcg, gfp_mask, nr_pages);
5613 css_put(&memcg->css);
5615 if (ret == -EINTR) {
5616 memcg = root_mem_cgroup;
5625 * mem_cgroup_commit_charge - commit a page charge
5626 * @page: page to charge
5627 * @memcg: memcg to charge the page to
5628 * @lrucare: page might be on LRU already
5630 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5631 * after page->mapping has been set up. This must happen atomically
5632 * as part of the page instantiation, i.e. under the page table lock
5633 * for anonymous pages, under the page lock for page and swap cache.
5635 * In addition, the page must not be on the LRU during the commit, to
5636 * prevent racing with task migration. If it might be, use @lrucare.
5638 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5640 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5643 unsigned int nr_pages = 1;
5645 VM_BUG_ON_PAGE(!page->mapping, page);
5646 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5648 if (mem_cgroup_disabled())
5651 * Swap faults will attempt to charge the same page multiple
5652 * times. But reuse_swap_page() might have removed the page
5653 * from swapcache already, so we can't check PageSwapCache().
5658 commit_charge(page, memcg, lrucare);
5660 if (PageTransHuge(page)) {
5661 nr_pages <<= compound_order(page);
5662 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5665 local_irq_disable();
5666 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5667 memcg_check_events(memcg, page);
5670 if (do_swap_account && PageSwapCache(page)) {
5671 swp_entry_t entry = { .val = page_private(page) };
5673 * The swap entry might not get freed for a long time,
5674 * let's not wait for it. The page already received a
5675 * memory+swap charge, drop the swap entry duplicate.
5677 mem_cgroup_uncharge_swap(entry);
5682 * mem_cgroup_cancel_charge - cancel a page charge
5683 * @page: page to charge
5684 * @memcg: memcg to charge the page to
5686 * Cancel a charge transaction started by mem_cgroup_try_charge().
5688 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5690 unsigned int nr_pages = 1;
5692 if (mem_cgroup_disabled())
5695 * Swap faults will attempt to charge the same page multiple
5696 * times. But reuse_swap_page() might have removed the page
5697 * from swapcache already, so we can't check PageSwapCache().
5702 if (PageTransHuge(page)) {
5703 nr_pages <<= compound_order(page);
5704 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5707 cancel_charge(memcg, nr_pages);
5710 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5711 unsigned long nr_anon, unsigned long nr_file,
5712 unsigned long nr_huge, struct page *dummy_page)
5714 unsigned long nr_pages = nr_anon + nr_file;
5715 unsigned long flags;
5717 if (!mem_cgroup_is_root(memcg)) {
5718 page_counter_uncharge(&memcg->memory, nr_pages);
5719 if (do_swap_account)
5720 page_counter_uncharge(&memcg->memsw, nr_pages);
5721 memcg_oom_recover(memcg);
5724 local_irq_save(flags);
5725 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5726 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5727 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5728 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5729 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5730 memcg_check_events(memcg, dummy_page);
5731 local_irq_restore(flags);
5733 if (!mem_cgroup_is_root(memcg))
5734 css_put_many(&memcg->css, nr_pages);
5737 static void uncharge_list(struct list_head *page_list)
5739 struct mem_cgroup *memcg = NULL;
5740 unsigned long nr_anon = 0;
5741 unsigned long nr_file = 0;
5742 unsigned long nr_huge = 0;
5743 unsigned long pgpgout = 0;
5744 struct list_head *next;
5747 next = page_list->next;
5749 unsigned int nr_pages = 1;
5751 page = list_entry(next, struct page, lru);
5752 next = page->lru.next;
5754 VM_BUG_ON_PAGE(PageLRU(page), page);
5755 VM_BUG_ON_PAGE(page_count(page), page);
5757 if (!page->mem_cgroup)
5761 * Nobody should be changing or seriously looking at
5762 * page->mem_cgroup at this point, we have fully
5763 * exclusive access to the page.
5766 if (memcg != page->mem_cgroup) {
5768 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5770 pgpgout = nr_anon = nr_file = nr_huge = 0;
5772 memcg = page->mem_cgroup;
5775 if (PageTransHuge(page)) {
5776 nr_pages <<= compound_order(page);
5777 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5778 nr_huge += nr_pages;
5782 nr_anon += nr_pages;
5784 nr_file += nr_pages;
5786 page->mem_cgroup = NULL;
5789 } while (next != page_list);
5792 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5797 * mem_cgroup_uncharge - uncharge a page
5798 * @page: page to uncharge
5800 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5801 * mem_cgroup_commit_charge().
5803 void mem_cgroup_uncharge(struct page *page)
5805 if (mem_cgroup_disabled())
5808 /* Don't touch page->lru of any random page, pre-check: */
5809 if (!page->mem_cgroup)
5812 INIT_LIST_HEAD(&page->lru);
5813 uncharge_list(&page->lru);
5817 * mem_cgroup_uncharge_list - uncharge a list of page
5818 * @page_list: list of pages to uncharge
5820 * Uncharge a list of pages previously charged with
5821 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5823 void mem_cgroup_uncharge_list(struct list_head *page_list)
5825 if (mem_cgroup_disabled())
5828 if (!list_empty(page_list))
5829 uncharge_list(page_list);
5833 * mem_cgroup_migrate - migrate a charge to another page
5834 * @oldpage: currently charged page
5835 * @newpage: page to transfer the charge to
5836 * @lrucare: either or both pages might be on the LRU already
5838 * Migrate the charge from @oldpage to @newpage.
5840 * Both pages must be locked, @newpage->mapping must be set up.
5842 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5845 struct mem_cgroup *memcg;
5848 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5849 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5850 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5851 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5852 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5853 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5856 if (mem_cgroup_disabled())
5859 /* Page cache replacement: new page already charged? */
5860 if (newpage->mem_cgroup)
5864 * Swapcache readahead pages can get migrated before being
5865 * charged, and migration from compaction can happen to an
5866 * uncharged page when the PFN walker finds a page that
5867 * reclaim just put back on the LRU but has not released yet.
5869 memcg = oldpage->mem_cgroup;
5874 lock_page_lru(oldpage, &isolated);
5876 oldpage->mem_cgroup = NULL;
5879 unlock_page_lru(oldpage, isolated);
5881 commit_charge(newpage, memcg, lrucare);
5885 * subsys_initcall() for memory controller.
5887 * Some parts like hotcpu_notifier() have to be initialized from this context
5888 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5889 * everything that doesn't depend on a specific mem_cgroup structure should
5890 * be initialized from here.
5892 static int __init mem_cgroup_init(void)
5896 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5898 for_each_possible_cpu(cpu)
5899 INIT_WORK(&per_cpu_ptr(&memcg_stock, cpu)->work,
5902 for_each_node(node) {
5903 struct mem_cgroup_tree_per_node *rtpn;
5906 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL,
5907 node_online(node) ? node : NUMA_NO_NODE);
5909 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5910 struct mem_cgroup_tree_per_zone *rtpz;
5912 rtpz = &rtpn->rb_tree_per_zone[zone];
5913 rtpz->rb_root = RB_ROOT;
5914 spin_lock_init(&rtpz->lock);
5916 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5921 subsys_initcall(mem_cgroup_init);
5923 #ifdef CONFIG_MEMCG_SWAP
5925 * mem_cgroup_swapout - transfer a memsw charge to swap
5926 * @page: page whose memsw charge to transfer
5927 * @entry: swap entry to move the charge to
5929 * Transfer the memsw charge of @page to @entry.
5931 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5933 struct mem_cgroup *memcg;
5934 unsigned short oldid;
5936 VM_BUG_ON_PAGE(PageLRU(page), page);
5937 VM_BUG_ON_PAGE(page_count(page), page);
5939 if (!do_swap_account)
5942 memcg = page->mem_cgroup;
5944 /* Readahead page, never charged */
5948 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5949 VM_BUG_ON_PAGE(oldid, page);
5950 mem_cgroup_swap_statistics(memcg, true);
5952 page->mem_cgroup = NULL;
5954 if (!mem_cgroup_is_root(memcg))
5955 page_counter_uncharge(&memcg->memory, 1);
5957 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5958 VM_BUG_ON(!irqs_disabled());
5960 mem_cgroup_charge_statistics(memcg, page, -1);
5961 memcg_check_events(memcg, page);
5965 * mem_cgroup_uncharge_swap - uncharge a swap entry
5966 * @entry: swap entry to uncharge
5968 * Drop the memsw charge associated with @entry.
5970 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5972 struct mem_cgroup *memcg;
5975 if (!do_swap_account)
5978 id = swap_cgroup_record(entry, 0);
5980 memcg = mem_cgroup_from_id(id);
5982 if (!mem_cgroup_is_root(memcg))
5983 page_counter_uncharge(&memcg->memsw, 1);
5984 mem_cgroup_swap_statistics(memcg, false);
5985 css_put(&memcg->css);
5990 /* for remember boot option*/
5991 #ifdef CONFIG_MEMCG_SWAP_ENABLED
5992 static int really_do_swap_account __initdata = 1;
5994 static int really_do_swap_account __initdata;
5997 static int __init enable_swap_account(char *s)
5999 if (!strcmp(s, "1"))
6000 really_do_swap_account = 1;
6001 else if (!strcmp(s, "0"))
6002 really_do_swap_account = 0;
6005 __setup("swapaccount=", enable_swap_account);
6007 static struct cftype memsw_cgroup_files[] = {
6009 .name = "memsw.usage_in_bytes",
6010 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6011 .read_u64 = mem_cgroup_read_u64,
6014 .name = "memsw.max_usage_in_bytes",
6015 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6016 .write = mem_cgroup_reset,
6017 .read_u64 = mem_cgroup_read_u64,
6020 .name = "memsw.limit_in_bytes",
6021 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6022 .write = mem_cgroup_write,
6023 .read_u64 = mem_cgroup_read_u64,
6026 .name = "memsw.failcnt",
6027 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6028 .write = mem_cgroup_reset,
6029 .read_u64 = mem_cgroup_read_u64,
6031 { }, /* terminate */
6034 static int __init mem_cgroup_swap_init(void)
6036 if (!mem_cgroup_disabled() && really_do_swap_account) {
6037 do_swap_account = 1;
6038 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
6039 memsw_cgroup_files));
6043 subsys_initcall(mem_cgroup_swap_init);
6045 #endif /* CONFIG_MEMCG_SWAP */