1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/page_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/swap_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
57 #include <linux/lockdep.h>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct reclaim_iter {
147 struct mem_cgroup *position;
148 /* scan generation, increased every round-trip */
149 unsigned int generation;
153 * per-zone information in memory controller.
155 struct mem_cgroup_per_zone {
156 struct lruvec lruvec;
157 unsigned long lru_size[NR_LRU_LISTS];
159 struct reclaim_iter iter[DEF_PRIORITY + 1];
161 struct rb_node tree_node; /* RB tree node */
162 unsigned long usage_in_excess;/* Set to the value by which */
163 /* the soft limit is exceeded*/
165 struct mem_cgroup *memcg; /* Back pointer, we cannot */
166 /* use container_of */
169 struct mem_cgroup_per_node {
170 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
174 * Cgroups above their limits are maintained in a RB-Tree, independent of
175 * their hierarchy representation
178 struct mem_cgroup_tree_per_zone {
179 struct rb_root rb_root;
183 struct mem_cgroup_tree_per_node {
184 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
187 struct mem_cgroup_tree {
188 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
191 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
193 struct mem_cgroup_threshold {
194 struct eventfd_ctx *eventfd;
195 unsigned long threshold;
199 struct mem_cgroup_threshold_ary {
200 /* An array index points to threshold just below or equal to usage. */
201 int current_threshold;
202 /* Size of entries[] */
204 /* Array of thresholds */
205 struct mem_cgroup_threshold entries[0];
208 struct mem_cgroup_thresholds {
209 /* Primary thresholds array */
210 struct mem_cgroup_threshold_ary *primary;
212 * Spare threshold array.
213 * This is needed to make mem_cgroup_unregister_event() "never fail".
214 * It must be able to store at least primary->size - 1 entries.
216 struct mem_cgroup_threshold_ary *spare;
220 struct mem_cgroup_eventfd_list {
221 struct list_head list;
222 struct eventfd_ctx *eventfd;
226 * cgroup_event represents events which userspace want to receive.
228 struct mem_cgroup_event {
230 * memcg which the event belongs to.
232 struct mem_cgroup *memcg;
234 * eventfd to signal userspace about the event.
236 struct eventfd_ctx *eventfd;
238 * Each of these stored in a list by the cgroup.
240 struct list_head list;
242 * register_event() callback will be used to add new userspace
243 * waiter for changes related to this event. Use eventfd_signal()
244 * on eventfd to send notification to userspace.
246 int (*register_event)(struct mem_cgroup *memcg,
247 struct eventfd_ctx *eventfd, const char *args);
249 * unregister_event() callback will be called when userspace closes
250 * the eventfd or on cgroup removing. This callback must be set,
251 * if you want provide notification functionality.
253 void (*unregister_event)(struct mem_cgroup *memcg,
254 struct eventfd_ctx *eventfd);
256 * All fields below needed to unregister event when
257 * userspace closes eventfd.
260 wait_queue_head_t *wqh;
262 struct work_struct remove;
265 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
266 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
269 * The memory controller data structure. The memory controller controls both
270 * page cache and RSS per cgroup. We would eventually like to provide
271 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
272 * to help the administrator determine what knobs to tune.
274 * TODO: Add a water mark for the memory controller. Reclaim will begin when
275 * we hit the water mark. May be even add a low water mark, such that
276 * no reclaim occurs from a cgroup at it's low water mark, this is
277 * a feature that will be implemented much later in the future.
280 struct cgroup_subsys_state css;
282 /* Accounted resources */
283 struct page_counter memory;
284 struct page_counter memsw;
285 struct page_counter kmem;
287 unsigned long soft_limit;
289 /* vmpressure notifications */
290 struct vmpressure vmpressure;
292 /* css_online() has been completed */
296 * Should the accounting and control be hierarchical, per subtree?
302 atomic_t oom_wakeups;
305 /* OOM-Killer disable */
306 int oom_kill_disable;
308 /* protect arrays of thresholds */
309 struct mutex thresholds_lock;
311 /* thresholds for memory usage. RCU-protected */
312 struct mem_cgroup_thresholds thresholds;
314 /* thresholds for mem+swap usage. RCU-protected */
315 struct mem_cgroup_thresholds memsw_thresholds;
317 /* For oom notifier event fd */
318 struct list_head oom_notify;
321 * Should we move charges of a task when a task is moved into this
322 * mem_cgroup ? And what type of charges should we move ?
324 unsigned long move_charge_at_immigrate;
326 * set > 0 if pages under this cgroup are moving to other cgroup.
328 atomic_t moving_account;
329 /* taken only while moving_account > 0 */
330 spinlock_t move_lock;
334 struct mem_cgroup_stat_cpu __percpu *stat;
336 * used when a cpu is offlined or other synchronizations
337 * See mem_cgroup_read_stat().
339 struct mem_cgroup_stat_cpu nocpu_base;
340 spinlock_t pcp_counter_lock;
342 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
343 struct cg_proto tcp_mem;
345 #if defined(CONFIG_MEMCG_KMEM)
346 /* analogous to slab_common's slab_caches list, but per-memcg;
347 * protected by memcg_slab_mutex */
348 struct list_head memcg_slab_caches;
349 /* Index in the kmem_cache->memcg_params->memcg_caches array */
353 int last_scanned_node;
355 nodemask_t scan_nodes;
356 atomic_t numainfo_events;
357 atomic_t numainfo_updating;
360 /* List of events which userspace want to receive */
361 struct list_head event_list;
362 spinlock_t event_list_lock;
364 struct mem_cgroup_per_node *nodeinfo[0];
365 /* WARNING: nodeinfo must be the last member here */
368 #ifdef CONFIG_MEMCG_KMEM
369 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
371 return memcg->kmemcg_id >= 0;
375 /* Stuffs for move charges at task migration. */
377 * Types of charges to be moved. "move_charge_at_immitgrate" and
378 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
381 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
382 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
386 /* "mc" and its members are protected by cgroup_mutex */
387 static struct move_charge_struct {
388 spinlock_t lock; /* for from, to */
389 struct mem_cgroup *from;
390 struct mem_cgroup *to;
391 unsigned long immigrate_flags;
392 unsigned long precharge;
393 unsigned long moved_charge;
394 unsigned long moved_swap;
395 struct task_struct *moving_task; /* a task moving charges */
396 wait_queue_head_t waitq; /* a waitq for other context */
398 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
399 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
402 static bool move_anon(void)
404 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
407 static bool move_file(void)
409 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
413 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
414 * limit reclaim to prevent infinite loops, if they ever occur.
416 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
417 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
420 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
421 MEM_CGROUP_CHARGE_TYPE_ANON,
422 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
423 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
427 /* for encoding cft->private value on file */
435 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
436 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
437 #define MEMFILE_ATTR(val) ((val) & 0xffff)
438 /* Used for OOM nofiier */
439 #define OOM_CONTROL (0)
442 * The memcg_create_mutex will be held whenever a new cgroup is created.
443 * As a consequence, any change that needs to protect against new child cgroups
444 * appearing has to hold it as well.
446 static DEFINE_MUTEX(memcg_create_mutex);
448 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
450 return s ? container_of(s, struct mem_cgroup, css) : NULL;
453 /* Some nice accessors for the vmpressure. */
454 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
457 memcg = root_mem_cgroup;
458 return &memcg->vmpressure;
461 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
463 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
466 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
468 return (memcg == root_mem_cgroup);
472 * We restrict the id in the range of [1, 65535], so it can fit into
475 #define MEM_CGROUP_ID_MAX USHRT_MAX
477 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
479 return memcg->css.id;
482 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
484 struct cgroup_subsys_state *css;
486 css = css_from_id(id, &memory_cgrp_subsys);
487 return mem_cgroup_from_css(css);
490 /* Writing them here to avoid exposing memcg's inner layout */
491 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
493 void sock_update_memcg(struct sock *sk)
495 if (mem_cgroup_sockets_enabled) {
496 struct mem_cgroup *memcg;
497 struct cg_proto *cg_proto;
499 BUG_ON(!sk->sk_prot->proto_cgroup);
501 /* Socket cloning can throw us here with sk_cgrp already
502 * filled. It won't however, necessarily happen from
503 * process context. So the test for root memcg given
504 * the current task's memcg won't help us in this case.
506 * Respecting the original socket's memcg is a better
507 * decision in this case.
510 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
511 css_get(&sk->sk_cgrp->memcg->css);
516 memcg = mem_cgroup_from_task(current);
517 cg_proto = sk->sk_prot->proto_cgroup(memcg);
518 if (!mem_cgroup_is_root(memcg) &&
519 memcg_proto_active(cg_proto) &&
520 css_tryget_online(&memcg->css)) {
521 sk->sk_cgrp = cg_proto;
526 EXPORT_SYMBOL(sock_update_memcg);
528 void sock_release_memcg(struct sock *sk)
530 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
531 struct mem_cgroup *memcg;
532 WARN_ON(!sk->sk_cgrp->memcg);
533 memcg = sk->sk_cgrp->memcg;
534 css_put(&sk->sk_cgrp->memcg->css);
538 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
540 if (!memcg || mem_cgroup_is_root(memcg))
543 return &memcg->tcp_mem;
545 EXPORT_SYMBOL(tcp_proto_cgroup);
547 static void disarm_sock_keys(struct mem_cgroup *memcg)
549 if (!memcg_proto_activated(&memcg->tcp_mem))
551 static_key_slow_dec(&memcg_socket_limit_enabled);
554 static void disarm_sock_keys(struct mem_cgroup *memcg)
559 #ifdef CONFIG_MEMCG_KMEM
561 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
562 * The main reason for not using cgroup id for this:
563 * this works better in sparse environments, where we have a lot of memcgs,
564 * but only a few kmem-limited. Or also, if we have, for instance, 200
565 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
566 * 200 entry array for that.
568 * The current size of the caches array is stored in
569 * memcg_limited_groups_array_size. It will double each time we have to
572 static DEFINE_IDA(kmem_limited_groups);
573 int memcg_limited_groups_array_size;
576 * MIN_SIZE is different than 1, because we would like to avoid going through
577 * the alloc/free process all the time. In a small machine, 4 kmem-limited
578 * cgroups is a reasonable guess. In the future, it could be a parameter or
579 * tunable, but that is strictly not necessary.
581 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
582 * this constant directly from cgroup, but it is understandable that this is
583 * better kept as an internal representation in cgroup.c. In any case, the
584 * cgrp_id space is not getting any smaller, and we don't have to necessarily
585 * increase ours as well if it increases.
587 #define MEMCG_CACHES_MIN_SIZE 4
588 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
591 * A lot of the calls to the cache allocation functions are expected to be
592 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
593 * conditional to this static branch, we'll have to allow modules that does
594 * kmem_cache_alloc and the such to see this symbol as well
596 struct static_key memcg_kmem_enabled_key;
597 EXPORT_SYMBOL(memcg_kmem_enabled_key);
599 static void memcg_free_cache_id(int id);
601 static void disarm_kmem_keys(struct mem_cgroup *memcg)
603 if (memcg_kmem_is_active(memcg)) {
604 static_key_slow_dec(&memcg_kmem_enabled_key);
605 memcg_free_cache_id(memcg->kmemcg_id);
608 * This check can't live in kmem destruction function,
609 * since the charges will outlive the cgroup
611 WARN_ON(page_counter_read(&memcg->kmem));
614 static void disarm_kmem_keys(struct mem_cgroup *memcg)
617 #endif /* CONFIG_MEMCG_KMEM */
619 static void disarm_static_keys(struct mem_cgroup *memcg)
621 disarm_sock_keys(memcg);
622 disarm_kmem_keys(memcg);
625 static struct mem_cgroup_per_zone *
626 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
628 int nid = zone_to_nid(zone);
629 int zid = zone_idx(zone);
631 return &memcg->nodeinfo[nid]->zoneinfo[zid];
634 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
639 static struct mem_cgroup_per_zone *
640 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
642 int nid = page_to_nid(page);
643 int zid = page_zonenum(page);
645 return &memcg->nodeinfo[nid]->zoneinfo[zid];
648 static struct mem_cgroup_tree_per_zone *
649 soft_limit_tree_node_zone(int nid, int zid)
651 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
654 static struct mem_cgroup_tree_per_zone *
655 soft_limit_tree_from_page(struct page *page)
657 int nid = page_to_nid(page);
658 int zid = page_zonenum(page);
660 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
663 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
664 struct mem_cgroup_tree_per_zone *mctz,
665 unsigned long new_usage_in_excess)
667 struct rb_node **p = &mctz->rb_root.rb_node;
668 struct rb_node *parent = NULL;
669 struct mem_cgroup_per_zone *mz_node;
674 mz->usage_in_excess = new_usage_in_excess;
675 if (!mz->usage_in_excess)
679 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
681 if (mz->usage_in_excess < mz_node->usage_in_excess)
684 * We can't avoid mem cgroups that are over their soft
685 * limit by the same amount
687 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
690 rb_link_node(&mz->tree_node, parent, p);
691 rb_insert_color(&mz->tree_node, &mctz->rb_root);
695 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
696 struct mem_cgroup_tree_per_zone *mctz)
700 rb_erase(&mz->tree_node, &mctz->rb_root);
704 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
705 struct mem_cgroup_tree_per_zone *mctz)
709 spin_lock_irqsave(&mctz->lock, flags);
710 __mem_cgroup_remove_exceeded(mz, mctz);
711 spin_unlock_irqrestore(&mctz->lock, flags);
714 static unsigned long soft_limit_excess(struct mem_cgroup *memcg)
716 unsigned long nr_pages = page_counter_read(&memcg->memory);
717 unsigned long soft_limit = ACCESS_ONCE(memcg->soft_limit);
718 unsigned long excess = 0;
720 if (nr_pages > soft_limit)
721 excess = nr_pages - soft_limit;
726 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
728 unsigned long excess;
729 struct mem_cgroup_per_zone *mz;
730 struct mem_cgroup_tree_per_zone *mctz;
732 mctz = soft_limit_tree_from_page(page);
734 * Necessary to update all ancestors when hierarchy is used.
735 * because their event counter is not touched.
737 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
738 mz = mem_cgroup_page_zoneinfo(memcg, page);
739 excess = soft_limit_excess(memcg);
741 * We have to update the tree if mz is on RB-tree or
742 * mem is over its softlimit.
744 if (excess || mz->on_tree) {
747 spin_lock_irqsave(&mctz->lock, flags);
748 /* if on-tree, remove it */
750 __mem_cgroup_remove_exceeded(mz, mctz);
752 * Insert again. mz->usage_in_excess will be updated.
753 * If excess is 0, no tree ops.
755 __mem_cgroup_insert_exceeded(mz, mctz, excess);
756 spin_unlock_irqrestore(&mctz->lock, flags);
761 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
763 struct mem_cgroup_tree_per_zone *mctz;
764 struct mem_cgroup_per_zone *mz;
768 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
769 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
770 mctz = soft_limit_tree_node_zone(nid, zid);
771 mem_cgroup_remove_exceeded(mz, mctz);
776 static struct mem_cgroup_per_zone *
777 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
779 struct rb_node *rightmost = NULL;
780 struct mem_cgroup_per_zone *mz;
784 rightmost = rb_last(&mctz->rb_root);
786 goto done; /* Nothing to reclaim from */
788 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
790 * Remove the node now but someone else can add it back,
791 * we will to add it back at the end of reclaim to its correct
792 * position in the tree.
794 __mem_cgroup_remove_exceeded(mz, mctz);
795 if (!soft_limit_excess(mz->memcg) ||
796 !css_tryget_online(&mz->memcg->css))
802 static struct mem_cgroup_per_zone *
803 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
805 struct mem_cgroup_per_zone *mz;
807 spin_lock_irq(&mctz->lock);
808 mz = __mem_cgroup_largest_soft_limit_node(mctz);
809 spin_unlock_irq(&mctz->lock);
814 * Implementation Note: reading percpu statistics for memcg.
816 * Both of vmstat[] and percpu_counter has threshold and do periodic
817 * synchronization to implement "quick" read. There are trade-off between
818 * reading cost and precision of value. Then, we may have a chance to implement
819 * a periodic synchronizion of counter in memcg's counter.
821 * But this _read() function is used for user interface now. The user accounts
822 * memory usage by memory cgroup and he _always_ requires exact value because
823 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
824 * have to visit all online cpus and make sum. So, for now, unnecessary
825 * synchronization is not implemented. (just implemented for cpu hotplug)
827 * If there are kernel internal actions which can make use of some not-exact
828 * value, and reading all cpu value can be performance bottleneck in some
829 * common workload, threashold and synchonization as vmstat[] should be
832 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
833 enum mem_cgroup_stat_index idx)
839 for_each_online_cpu(cpu)
840 val += per_cpu(memcg->stat->count[idx], cpu);
841 #ifdef CONFIG_HOTPLUG_CPU
842 spin_lock(&memcg->pcp_counter_lock);
843 val += memcg->nocpu_base.count[idx];
844 spin_unlock(&memcg->pcp_counter_lock);
850 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
851 enum mem_cgroup_events_index idx)
853 unsigned long val = 0;
857 for_each_online_cpu(cpu)
858 val += per_cpu(memcg->stat->events[idx], cpu);
859 #ifdef CONFIG_HOTPLUG_CPU
860 spin_lock(&memcg->pcp_counter_lock);
861 val += memcg->nocpu_base.events[idx];
862 spin_unlock(&memcg->pcp_counter_lock);
868 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
873 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
874 * counted as CACHE even if it's on ANON LRU.
877 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
880 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
883 if (PageTransHuge(page))
884 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
887 /* pagein of a big page is an event. So, ignore page size */
889 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
891 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
892 nr_pages = -nr_pages; /* for event */
895 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
898 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
900 struct mem_cgroup_per_zone *mz;
902 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
903 return mz->lru_size[lru];
906 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
908 unsigned int lru_mask)
910 unsigned long nr = 0;
913 VM_BUG_ON((unsigned)nid >= nr_node_ids);
915 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
916 struct mem_cgroup_per_zone *mz;
920 if (!(BIT(lru) & lru_mask))
922 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
923 nr += mz->lru_size[lru];
929 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
930 unsigned int lru_mask)
932 unsigned long nr = 0;
935 for_each_node_state(nid, N_MEMORY)
936 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
940 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
941 enum mem_cgroup_events_target target)
943 unsigned long val, next;
945 val = __this_cpu_read(memcg->stat->nr_page_events);
946 next = __this_cpu_read(memcg->stat->targets[target]);
947 /* from time_after() in jiffies.h */
948 if ((long)next - (long)val < 0) {
950 case MEM_CGROUP_TARGET_THRESH:
951 next = val + THRESHOLDS_EVENTS_TARGET;
953 case MEM_CGROUP_TARGET_SOFTLIMIT:
954 next = val + SOFTLIMIT_EVENTS_TARGET;
956 case MEM_CGROUP_TARGET_NUMAINFO:
957 next = val + NUMAINFO_EVENTS_TARGET;
962 __this_cpu_write(memcg->stat->targets[target], next);
969 * Check events in order.
972 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
974 /* threshold event is triggered in finer grain than soft limit */
975 if (unlikely(mem_cgroup_event_ratelimit(memcg,
976 MEM_CGROUP_TARGET_THRESH))) {
978 bool do_numainfo __maybe_unused;
980 do_softlimit = mem_cgroup_event_ratelimit(memcg,
981 MEM_CGROUP_TARGET_SOFTLIMIT);
983 do_numainfo = mem_cgroup_event_ratelimit(memcg,
984 MEM_CGROUP_TARGET_NUMAINFO);
986 mem_cgroup_threshold(memcg);
987 if (unlikely(do_softlimit))
988 mem_cgroup_update_tree(memcg, page);
990 if (unlikely(do_numainfo))
991 atomic_inc(&memcg->numainfo_events);
996 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
999 * mm_update_next_owner() may clear mm->owner to NULL
1000 * if it races with swapoff, page migration, etc.
1001 * So this can be called with p == NULL.
1006 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1009 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1011 struct mem_cgroup *memcg = NULL;
1016 * Page cache insertions can happen withou an
1017 * actual mm context, e.g. during disk probing
1018 * on boot, loopback IO, acct() writes etc.
1021 memcg = root_mem_cgroup;
1023 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1024 if (unlikely(!memcg))
1025 memcg = root_mem_cgroup;
1027 } while (!css_tryget_online(&memcg->css));
1033 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1034 * @root: hierarchy root
1035 * @prev: previously returned memcg, NULL on first invocation
1036 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1038 * Returns references to children of the hierarchy below @root, or
1039 * @root itself, or %NULL after a full round-trip.
1041 * Caller must pass the return value in @prev on subsequent
1042 * invocations for reference counting, or use mem_cgroup_iter_break()
1043 * to cancel a hierarchy walk before the round-trip is complete.
1045 * Reclaimers can specify a zone and a priority level in @reclaim to
1046 * divide up the memcgs in the hierarchy among all concurrent
1047 * reclaimers operating on the same zone and priority.
1049 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1050 struct mem_cgroup *prev,
1051 struct mem_cgroup_reclaim_cookie *reclaim)
1053 struct reclaim_iter *uninitialized_var(iter);
1054 struct cgroup_subsys_state *css = NULL;
1055 struct mem_cgroup *memcg = NULL;
1056 struct mem_cgroup *pos = NULL;
1058 if (mem_cgroup_disabled())
1062 root = root_mem_cgroup;
1064 if (prev && !reclaim)
1067 if (!root->use_hierarchy && root != root_mem_cgroup) {
1076 struct mem_cgroup_per_zone *mz;
1078 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1079 iter = &mz->iter[reclaim->priority];
1081 if (prev && reclaim->generation != iter->generation)
1085 pos = ACCESS_ONCE(iter->position);
1087 * A racing update may change the position and
1088 * put the last reference, hence css_tryget(),
1089 * or retry to see the updated position.
1091 } while (pos && !css_tryget(&pos->css));
1098 css = css_next_descendant_pre(css, &root->css);
1101 * Reclaimers share the hierarchy walk, and a
1102 * new one might jump in right at the end of
1103 * the hierarchy - make sure they see at least
1104 * one group and restart from the beginning.
1112 * Verify the css and acquire a reference. The root
1113 * is provided by the caller, so we know it's alive
1114 * and kicking, and don't take an extra reference.
1116 memcg = mem_cgroup_from_css(css);
1118 if (css == &root->css)
1121 if (css_tryget(css)) {
1123 * Make sure the memcg is initialized:
1124 * mem_cgroup_css_online() orders the the
1125 * initialization against setting the flag.
1127 if (smp_load_acquire(&memcg->initialized))
1137 if (cmpxchg(&iter->position, pos, memcg) == pos) {
1139 css_get(&memcg->css);
1145 * pairs with css_tryget when dereferencing iter->position
1154 reclaim->generation = iter->generation;
1160 if (prev && prev != root)
1161 css_put(&prev->css);
1167 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1168 * @root: hierarchy root
1169 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1171 void mem_cgroup_iter_break(struct mem_cgroup *root,
1172 struct mem_cgroup *prev)
1175 root = root_mem_cgroup;
1176 if (prev && prev != root)
1177 css_put(&prev->css);
1181 * Iteration constructs for visiting all cgroups (under a tree). If
1182 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1183 * be used for reference counting.
1185 #define for_each_mem_cgroup_tree(iter, root) \
1186 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1188 iter = mem_cgroup_iter(root, iter, NULL))
1190 #define for_each_mem_cgroup(iter) \
1191 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1193 iter = mem_cgroup_iter(NULL, iter, NULL))
1195 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1197 struct mem_cgroup *memcg;
1200 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1201 if (unlikely(!memcg))
1206 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1209 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1217 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1220 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1221 * @zone: zone of the wanted lruvec
1222 * @memcg: memcg of the wanted lruvec
1224 * Returns the lru list vector holding pages for the given @zone and
1225 * @mem. This can be the global zone lruvec, if the memory controller
1228 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1229 struct mem_cgroup *memcg)
1231 struct mem_cgroup_per_zone *mz;
1232 struct lruvec *lruvec;
1234 if (mem_cgroup_disabled()) {
1235 lruvec = &zone->lruvec;
1239 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1240 lruvec = &mz->lruvec;
1243 * Since a node can be onlined after the mem_cgroup was created,
1244 * we have to be prepared to initialize lruvec->zone here;
1245 * and if offlined then reonlined, we need to reinitialize it.
1247 if (unlikely(lruvec->zone != zone))
1248 lruvec->zone = zone;
1253 * mem_cgroup_page_lruvec - return lruvec for isolating/putting an LRU page
1255 * @zone: zone of the page
1257 * This function is only safe when following the LRU page isolation
1258 * and putback protocol: the LRU lock must be held, and the page must
1259 * either be PageLRU() or the caller must have isolated/allocated it.
1261 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1263 struct mem_cgroup_per_zone *mz;
1264 struct mem_cgroup *memcg;
1265 struct lruvec *lruvec;
1267 if (mem_cgroup_disabled()) {
1268 lruvec = &zone->lruvec;
1272 memcg = page->mem_cgroup;
1274 * Swapcache readahead pages are added to the LRU - and
1275 * possibly migrated - before they are charged.
1278 memcg = root_mem_cgroup;
1280 mz = mem_cgroup_page_zoneinfo(memcg, page);
1281 lruvec = &mz->lruvec;
1284 * Since a node can be onlined after the mem_cgroup was created,
1285 * we have to be prepared to initialize lruvec->zone here;
1286 * and if offlined then reonlined, we need to reinitialize it.
1288 if (unlikely(lruvec->zone != zone))
1289 lruvec->zone = zone;
1294 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1295 * @lruvec: mem_cgroup per zone lru vector
1296 * @lru: index of lru list the page is sitting on
1297 * @nr_pages: positive when adding or negative when removing
1299 * This function must be called when a page is added to or removed from an
1302 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1305 struct mem_cgroup_per_zone *mz;
1306 unsigned long *lru_size;
1308 if (mem_cgroup_disabled())
1311 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1312 lru_size = mz->lru_size + lru;
1313 *lru_size += nr_pages;
1314 VM_BUG_ON((long)(*lru_size) < 0);
1317 bool mem_cgroup_is_descendant(struct mem_cgroup *memcg, struct mem_cgroup *root)
1321 if (!root->use_hierarchy)
1323 return cgroup_is_descendant(memcg->css.cgroup, root->css.cgroup);
1326 bool task_in_mem_cgroup(struct task_struct *task, struct mem_cgroup *memcg)
1328 struct mem_cgroup *task_memcg;
1329 struct task_struct *p;
1332 p = find_lock_task_mm(task);
1334 task_memcg = get_mem_cgroup_from_mm(p->mm);
1338 * All threads may have already detached their mm's, but the oom
1339 * killer still needs to detect if they have already been oom
1340 * killed to prevent needlessly killing additional tasks.
1343 task_memcg = mem_cgroup_from_task(task);
1344 css_get(&task_memcg->css);
1347 ret = mem_cgroup_is_descendant(task_memcg, memcg);
1348 css_put(&task_memcg->css);
1352 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1354 unsigned long inactive_ratio;
1355 unsigned long inactive;
1356 unsigned long active;
1359 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1360 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1362 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1364 inactive_ratio = int_sqrt(10 * gb);
1368 return inactive * inactive_ratio < active;
1371 #define mem_cgroup_from_counter(counter, member) \
1372 container_of(counter, struct mem_cgroup, member)
1375 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1376 * @memcg: the memory cgroup
1378 * Returns the maximum amount of memory @mem can be charged with, in
1381 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1383 unsigned long margin = 0;
1384 unsigned long count;
1385 unsigned long limit;
1387 count = page_counter_read(&memcg->memory);
1388 limit = ACCESS_ONCE(memcg->memory.limit);
1390 margin = limit - count;
1392 if (do_swap_account) {
1393 count = page_counter_read(&memcg->memsw);
1394 limit = ACCESS_ONCE(memcg->memsw.limit);
1396 margin = min(margin, limit - count);
1402 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1405 if (mem_cgroup_disabled() || !memcg->css.parent)
1406 return vm_swappiness;
1408 return memcg->swappiness;
1412 * A routine for checking "mem" is under move_account() or not.
1414 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1415 * moving cgroups. This is for waiting at high-memory pressure
1418 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1420 struct mem_cgroup *from;
1421 struct mem_cgroup *to;
1424 * Unlike task_move routines, we access mc.to, mc.from not under
1425 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1427 spin_lock(&mc.lock);
1433 ret = mem_cgroup_is_descendant(from, memcg) ||
1434 mem_cgroup_is_descendant(to, memcg);
1436 spin_unlock(&mc.lock);
1440 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1442 if (mc.moving_task && current != mc.moving_task) {
1443 if (mem_cgroup_under_move(memcg)) {
1445 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1446 /* moving charge context might have finished. */
1449 finish_wait(&mc.waitq, &wait);
1456 #define K(x) ((x) << (PAGE_SHIFT-10))
1458 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1459 * @memcg: The memory cgroup that went over limit
1460 * @p: Task that is going to be killed
1462 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1465 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1467 /* oom_info_lock ensures that parallel ooms do not interleave */
1468 static DEFINE_MUTEX(oom_info_lock);
1469 struct mem_cgroup *iter;
1475 mutex_lock(&oom_info_lock);
1478 pr_info("Task in ");
1479 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1480 pr_cont(" killed as a result of limit of ");
1481 pr_cont_cgroup_path(memcg->css.cgroup);
1486 pr_info("memory: usage %llukB, limit %llukB, failcnt %lu\n",
1487 K((u64)page_counter_read(&memcg->memory)),
1488 K((u64)memcg->memory.limit), memcg->memory.failcnt);
1489 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %lu\n",
1490 K((u64)page_counter_read(&memcg->memsw)),
1491 K((u64)memcg->memsw.limit), memcg->memsw.failcnt);
1492 pr_info("kmem: usage %llukB, limit %llukB, failcnt %lu\n",
1493 K((u64)page_counter_read(&memcg->kmem)),
1494 K((u64)memcg->kmem.limit), memcg->kmem.failcnt);
1496 for_each_mem_cgroup_tree(iter, memcg) {
1497 pr_info("Memory cgroup stats for ");
1498 pr_cont_cgroup_path(iter->css.cgroup);
1501 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1502 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1504 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1505 K(mem_cgroup_read_stat(iter, i)));
1508 for (i = 0; i < NR_LRU_LISTS; i++)
1509 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1510 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1514 mutex_unlock(&oom_info_lock);
1518 * This function returns the number of memcg under hierarchy tree. Returns
1519 * 1(self count) if no children.
1521 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1524 struct mem_cgroup *iter;
1526 for_each_mem_cgroup_tree(iter, memcg)
1532 * Return the memory (and swap, if configured) limit for a memcg.
1534 static unsigned long mem_cgroup_get_limit(struct mem_cgroup *memcg)
1536 unsigned long limit;
1538 limit = memcg->memory.limit;
1539 if (mem_cgroup_swappiness(memcg)) {
1540 unsigned long memsw_limit;
1542 memsw_limit = memcg->memsw.limit;
1543 limit = min(limit + total_swap_pages, memsw_limit);
1548 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1551 struct mem_cgroup *iter;
1552 unsigned long chosen_points = 0;
1553 unsigned long totalpages;
1554 unsigned int points = 0;
1555 struct task_struct *chosen = NULL;
1558 * If current has a pending SIGKILL or is exiting, then automatically
1559 * select it. The goal is to allow it to allocate so that it may
1560 * quickly exit and free its memory.
1562 if (fatal_signal_pending(current) || task_will_free_mem(current)) {
1563 set_thread_flag(TIF_MEMDIE);
1567 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1568 totalpages = mem_cgroup_get_limit(memcg) ? : 1;
1569 for_each_mem_cgroup_tree(iter, memcg) {
1570 struct css_task_iter it;
1571 struct task_struct *task;
1573 css_task_iter_start(&iter->css, &it);
1574 while ((task = css_task_iter_next(&it))) {
1575 switch (oom_scan_process_thread(task, totalpages, NULL,
1577 case OOM_SCAN_SELECT:
1579 put_task_struct(chosen);
1581 chosen_points = ULONG_MAX;
1582 get_task_struct(chosen);
1584 case OOM_SCAN_CONTINUE:
1586 case OOM_SCAN_ABORT:
1587 css_task_iter_end(&it);
1588 mem_cgroup_iter_break(memcg, iter);
1590 put_task_struct(chosen);
1595 points = oom_badness(task, memcg, NULL, totalpages);
1596 if (!points || points < chosen_points)
1598 /* Prefer thread group leaders for display purposes */
1599 if (points == chosen_points &&
1600 thread_group_leader(chosen))
1604 put_task_struct(chosen);
1606 chosen_points = points;
1607 get_task_struct(chosen);
1609 css_task_iter_end(&it);
1614 points = chosen_points * 1000 / totalpages;
1615 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1616 NULL, "Memory cgroup out of memory");
1619 #if MAX_NUMNODES > 1
1622 * test_mem_cgroup_node_reclaimable
1623 * @memcg: the target memcg
1624 * @nid: the node ID to be checked.
1625 * @noswap : specify true here if the user wants flle only information.
1627 * This function returns whether the specified memcg contains any
1628 * reclaimable pages on a node. Returns true if there are any reclaimable
1629 * pages in the node.
1631 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1632 int nid, bool noswap)
1634 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1636 if (noswap || !total_swap_pages)
1638 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1645 * Always updating the nodemask is not very good - even if we have an empty
1646 * list or the wrong list here, we can start from some node and traverse all
1647 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1650 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1654 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1655 * pagein/pageout changes since the last update.
1657 if (!atomic_read(&memcg->numainfo_events))
1659 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1662 /* make a nodemask where this memcg uses memory from */
1663 memcg->scan_nodes = node_states[N_MEMORY];
1665 for_each_node_mask(nid, node_states[N_MEMORY]) {
1667 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1668 node_clear(nid, memcg->scan_nodes);
1671 atomic_set(&memcg->numainfo_events, 0);
1672 atomic_set(&memcg->numainfo_updating, 0);
1676 * Selecting a node where we start reclaim from. Because what we need is just
1677 * reducing usage counter, start from anywhere is O,K. Considering
1678 * memory reclaim from current node, there are pros. and cons.
1680 * Freeing memory from current node means freeing memory from a node which
1681 * we'll use or we've used. So, it may make LRU bad. And if several threads
1682 * hit limits, it will see a contention on a node. But freeing from remote
1683 * node means more costs for memory reclaim because of memory latency.
1685 * Now, we use round-robin. Better algorithm is welcomed.
1687 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1691 mem_cgroup_may_update_nodemask(memcg);
1692 node = memcg->last_scanned_node;
1694 node = next_node(node, memcg->scan_nodes);
1695 if (node == MAX_NUMNODES)
1696 node = first_node(memcg->scan_nodes);
1698 * We call this when we hit limit, not when pages are added to LRU.
1699 * No LRU may hold pages because all pages are UNEVICTABLE or
1700 * memcg is too small and all pages are not on LRU. In that case,
1701 * we use curret node.
1703 if (unlikely(node == MAX_NUMNODES))
1704 node = numa_node_id();
1706 memcg->last_scanned_node = node;
1710 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1716 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1719 unsigned long *total_scanned)
1721 struct mem_cgroup *victim = NULL;
1724 unsigned long excess;
1725 unsigned long nr_scanned;
1726 struct mem_cgroup_reclaim_cookie reclaim = {
1731 excess = soft_limit_excess(root_memcg);
1734 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1739 * If we have not been able to reclaim
1740 * anything, it might because there are
1741 * no reclaimable pages under this hierarchy
1746 * We want to do more targeted reclaim.
1747 * excess >> 2 is not to excessive so as to
1748 * reclaim too much, nor too less that we keep
1749 * coming back to reclaim from this cgroup
1751 if (total >= (excess >> 2) ||
1752 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1757 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1759 *total_scanned += nr_scanned;
1760 if (!soft_limit_excess(root_memcg))
1763 mem_cgroup_iter_break(root_memcg, victim);
1767 #ifdef CONFIG_LOCKDEP
1768 static struct lockdep_map memcg_oom_lock_dep_map = {
1769 .name = "memcg_oom_lock",
1773 static DEFINE_SPINLOCK(memcg_oom_lock);
1776 * Check OOM-Killer is already running under our hierarchy.
1777 * If someone is running, return false.
1779 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1781 struct mem_cgroup *iter, *failed = NULL;
1783 spin_lock(&memcg_oom_lock);
1785 for_each_mem_cgroup_tree(iter, memcg) {
1786 if (iter->oom_lock) {
1788 * this subtree of our hierarchy is already locked
1789 * so we cannot give a lock.
1792 mem_cgroup_iter_break(memcg, iter);
1795 iter->oom_lock = true;
1800 * OK, we failed to lock the whole subtree so we have
1801 * to clean up what we set up to the failing subtree
1803 for_each_mem_cgroup_tree(iter, memcg) {
1804 if (iter == failed) {
1805 mem_cgroup_iter_break(memcg, iter);
1808 iter->oom_lock = false;
1811 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1813 spin_unlock(&memcg_oom_lock);
1818 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1820 struct mem_cgroup *iter;
1822 spin_lock(&memcg_oom_lock);
1823 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1824 for_each_mem_cgroup_tree(iter, memcg)
1825 iter->oom_lock = false;
1826 spin_unlock(&memcg_oom_lock);
1829 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1831 struct mem_cgroup *iter;
1833 for_each_mem_cgroup_tree(iter, memcg)
1834 atomic_inc(&iter->under_oom);
1837 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1839 struct mem_cgroup *iter;
1842 * When a new child is created while the hierarchy is under oom,
1843 * mem_cgroup_oom_lock() may not be called. We have to use
1844 * atomic_add_unless() here.
1846 for_each_mem_cgroup_tree(iter, memcg)
1847 atomic_add_unless(&iter->under_oom, -1, 0);
1850 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1852 struct oom_wait_info {
1853 struct mem_cgroup *memcg;
1857 static int memcg_oom_wake_function(wait_queue_t *wait,
1858 unsigned mode, int sync, void *arg)
1860 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1861 struct mem_cgroup *oom_wait_memcg;
1862 struct oom_wait_info *oom_wait_info;
1864 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1865 oom_wait_memcg = oom_wait_info->memcg;
1867 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1868 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1870 return autoremove_wake_function(wait, mode, sync, arg);
1873 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1875 atomic_inc(&memcg->oom_wakeups);
1876 /* for filtering, pass "memcg" as argument. */
1877 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1880 static void memcg_oom_recover(struct mem_cgroup *memcg)
1882 if (memcg && atomic_read(&memcg->under_oom))
1883 memcg_wakeup_oom(memcg);
1886 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1888 if (!current->memcg_oom.may_oom)
1891 * We are in the middle of the charge context here, so we
1892 * don't want to block when potentially sitting on a callstack
1893 * that holds all kinds of filesystem and mm locks.
1895 * Also, the caller may handle a failed allocation gracefully
1896 * (like optional page cache readahead) and so an OOM killer
1897 * invocation might not even be necessary.
1899 * That's why we don't do anything here except remember the
1900 * OOM context and then deal with it at the end of the page
1901 * fault when the stack is unwound, the locks are released,
1902 * and when we know whether the fault was overall successful.
1904 css_get(&memcg->css);
1905 current->memcg_oom.memcg = memcg;
1906 current->memcg_oom.gfp_mask = mask;
1907 current->memcg_oom.order = order;
1911 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1912 * @handle: actually kill/wait or just clean up the OOM state
1914 * This has to be called at the end of a page fault if the memcg OOM
1915 * handler was enabled.
1917 * Memcg supports userspace OOM handling where failed allocations must
1918 * sleep on a waitqueue until the userspace task resolves the
1919 * situation. Sleeping directly in the charge context with all kinds
1920 * of locks held is not a good idea, instead we remember an OOM state
1921 * in the task and mem_cgroup_oom_synchronize() has to be called at
1922 * the end of the page fault to complete the OOM handling.
1924 * Returns %true if an ongoing memcg OOM situation was detected and
1925 * completed, %false otherwise.
1927 bool mem_cgroup_oom_synchronize(bool handle)
1929 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1930 struct oom_wait_info owait;
1933 /* OOM is global, do not handle */
1940 owait.memcg = memcg;
1941 owait.wait.flags = 0;
1942 owait.wait.func = memcg_oom_wake_function;
1943 owait.wait.private = current;
1944 INIT_LIST_HEAD(&owait.wait.task_list);
1946 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1947 mem_cgroup_mark_under_oom(memcg);
1949 locked = mem_cgroup_oom_trylock(memcg);
1952 mem_cgroup_oom_notify(memcg);
1954 if (locked && !memcg->oom_kill_disable) {
1955 mem_cgroup_unmark_under_oom(memcg);
1956 finish_wait(&memcg_oom_waitq, &owait.wait);
1957 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1958 current->memcg_oom.order);
1961 mem_cgroup_unmark_under_oom(memcg);
1962 finish_wait(&memcg_oom_waitq, &owait.wait);
1966 mem_cgroup_oom_unlock(memcg);
1968 * There is no guarantee that an OOM-lock contender
1969 * sees the wakeups triggered by the OOM kill
1970 * uncharges. Wake any sleepers explicitely.
1972 memcg_oom_recover(memcg);
1975 current->memcg_oom.memcg = NULL;
1976 css_put(&memcg->css);
1981 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1982 * @page: page that is going to change accounted state
1983 * @locked: &memcg->move_lock slowpath was taken
1984 * @flags: IRQ-state flags for &memcg->move_lock
1986 * This function must mark the beginning of an accounted page state
1987 * change to prevent double accounting when the page is concurrently
1988 * being moved to another memcg:
1990 * memcg = mem_cgroup_begin_page_stat(page, &locked, &flags);
1991 * if (TestClearPageState(page))
1992 * mem_cgroup_update_page_stat(memcg, state, -1);
1993 * mem_cgroup_end_page_stat(memcg, locked, flags);
1995 * The RCU lock is held throughout the transaction. The fast path can
1996 * get away without acquiring the memcg->move_lock (@locked is false)
1997 * because page moving starts with an RCU grace period.
1999 * The RCU lock also protects the memcg from being freed when the page
2000 * state that is going to change is the only thing preventing the page
2001 * from being uncharged. E.g. end-writeback clearing PageWriteback(),
2002 * which allows migration to go ahead and uncharge the page before the
2003 * account transaction might be complete.
2005 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page,
2007 unsigned long *flags)
2009 struct mem_cgroup *memcg;
2013 if (mem_cgroup_disabled())
2016 memcg = page->mem_cgroup;
2017 if (unlikely(!memcg))
2021 if (atomic_read(&memcg->moving_account) <= 0)
2024 spin_lock_irqsave(&memcg->move_lock, *flags);
2025 if (memcg != page->mem_cgroup) {
2026 spin_unlock_irqrestore(&memcg->move_lock, *flags);
2035 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2036 * @memcg: the memcg that was accounted against
2037 * @locked: value received from mem_cgroup_begin_page_stat()
2038 * @flags: value received from mem_cgroup_begin_page_stat()
2040 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg, bool *locked,
2041 unsigned long *flags)
2043 if (memcg && *locked)
2044 spin_unlock_irqrestore(&memcg->move_lock, *flags);
2050 * mem_cgroup_update_page_stat - update page state statistics
2051 * @memcg: memcg to account against
2052 * @idx: page state item to account
2053 * @val: number of pages (positive or negative)
2055 * See mem_cgroup_begin_page_stat() for locking requirements.
2057 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2058 enum mem_cgroup_stat_index idx, int val)
2060 VM_BUG_ON(!rcu_read_lock_held());
2063 this_cpu_add(memcg->stat->count[idx], val);
2067 * size of first charge trial. "32" comes from vmscan.c's magic value.
2068 * TODO: maybe necessary to use big numbers in big irons.
2070 #define CHARGE_BATCH 32U
2071 struct memcg_stock_pcp {
2072 struct mem_cgroup *cached; /* this never be root cgroup */
2073 unsigned int nr_pages;
2074 struct work_struct work;
2075 unsigned long flags;
2076 #define FLUSHING_CACHED_CHARGE 0
2078 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2079 static DEFINE_MUTEX(percpu_charge_mutex);
2082 * consume_stock: Try to consume stocked charge on this cpu.
2083 * @memcg: memcg to consume from.
2084 * @nr_pages: how many pages to charge.
2086 * The charges will only happen if @memcg matches the current cpu's memcg
2087 * stock, and at least @nr_pages are available in that stock. Failure to
2088 * service an allocation will refill the stock.
2090 * returns true if successful, false otherwise.
2092 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2094 struct memcg_stock_pcp *stock;
2097 if (nr_pages > CHARGE_BATCH)
2100 stock = &get_cpu_var(memcg_stock);
2101 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2102 stock->nr_pages -= nr_pages;
2105 put_cpu_var(memcg_stock);
2110 * Returns stocks cached in percpu and reset cached information.
2112 static void drain_stock(struct memcg_stock_pcp *stock)
2114 struct mem_cgroup *old = stock->cached;
2116 if (stock->nr_pages) {
2117 page_counter_uncharge(&old->memory, stock->nr_pages);
2118 if (do_swap_account)
2119 page_counter_uncharge(&old->memsw, stock->nr_pages);
2120 css_put_many(&old->css, stock->nr_pages);
2121 stock->nr_pages = 0;
2123 stock->cached = NULL;
2127 * This must be called under preempt disabled or must be called by
2128 * a thread which is pinned to local cpu.
2130 static void drain_local_stock(struct work_struct *dummy)
2132 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2134 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2137 static void __init memcg_stock_init(void)
2141 for_each_possible_cpu(cpu) {
2142 struct memcg_stock_pcp *stock =
2143 &per_cpu(memcg_stock, cpu);
2144 INIT_WORK(&stock->work, drain_local_stock);
2149 * Cache charges(val) to local per_cpu area.
2150 * This will be consumed by consume_stock() function, later.
2152 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2154 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2156 if (stock->cached != memcg) { /* reset if necessary */
2158 stock->cached = memcg;
2160 stock->nr_pages += nr_pages;
2161 put_cpu_var(memcg_stock);
2165 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2166 * of the hierarchy under it.
2168 static void drain_all_stock(struct mem_cgroup *root_memcg)
2172 /* If someone's already draining, avoid adding running more workers. */
2173 if (!mutex_trylock(&percpu_charge_mutex))
2175 /* Notify other cpus that system-wide "drain" is running */
2178 for_each_online_cpu(cpu) {
2179 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2180 struct mem_cgroup *memcg;
2182 memcg = stock->cached;
2183 if (!memcg || !stock->nr_pages)
2185 if (!mem_cgroup_is_descendant(memcg, root_memcg))
2187 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2189 drain_local_stock(&stock->work);
2191 schedule_work_on(cpu, &stock->work);
2196 mutex_unlock(&percpu_charge_mutex);
2200 * This function drains percpu counter value from DEAD cpu and
2201 * move it to local cpu. Note that this function can be preempted.
2203 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2207 spin_lock(&memcg->pcp_counter_lock);
2208 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2209 long x = per_cpu(memcg->stat->count[i], cpu);
2211 per_cpu(memcg->stat->count[i], cpu) = 0;
2212 memcg->nocpu_base.count[i] += x;
2214 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2215 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2217 per_cpu(memcg->stat->events[i], cpu) = 0;
2218 memcg->nocpu_base.events[i] += x;
2220 spin_unlock(&memcg->pcp_counter_lock);
2223 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2224 unsigned long action,
2227 int cpu = (unsigned long)hcpu;
2228 struct memcg_stock_pcp *stock;
2229 struct mem_cgroup *iter;
2231 if (action == CPU_ONLINE)
2234 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2237 for_each_mem_cgroup(iter)
2238 mem_cgroup_drain_pcp_counter(iter, cpu);
2240 stock = &per_cpu(memcg_stock, cpu);
2245 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2246 unsigned int nr_pages)
2248 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2249 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2250 struct mem_cgroup *mem_over_limit;
2251 struct page_counter *counter;
2252 unsigned long nr_reclaimed;
2253 bool may_swap = true;
2254 bool drained = false;
2257 if (mem_cgroup_is_root(memcg))
2260 if (consume_stock(memcg, nr_pages))
2263 if (!do_swap_account ||
2264 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2265 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2267 if (do_swap_account)
2268 page_counter_uncharge(&memcg->memsw, batch);
2269 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2271 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2275 if (batch > nr_pages) {
2281 * Unlike in global OOM situations, memcg is not in a physical
2282 * memory shortage. Allow dying and OOM-killed tasks to
2283 * bypass the last charges so that they can exit quickly and
2284 * free their memory.
2286 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2287 fatal_signal_pending(current) ||
2288 current->flags & PF_EXITING))
2291 if (unlikely(task_in_memcg_oom(current)))
2294 if (!(gfp_mask & __GFP_WAIT))
2297 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2298 gfp_mask, may_swap);
2300 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2304 drain_all_stock(mem_over_limit);
2309 if (gfp_mask & __GFP_NORETRY)
2312 * Even though the limit is exceeded at this point, reclaim
2313 * may have been able to free some pages. Retry the charge
2314 * before killing the task.
2316 * Only for regular pages, though: huge pages are rather
2317 * unlikely to succeed so close to the limit, and we fall back
2318 * to regular pages anyway in case of failure.
2320 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2323 * At task move, charge accounts can be doubly counted. So, it's
2324 * better to wait until the end of task_move if something is going on.
2326 if (mem_cgroup_wait_acct_move(mem_over_limit))
2332 if (gfp_mask & __GFP_NOFAIL)
2335 if (fatal_signal_pending(current))
2338 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2340 if (!(gfp_mask & __GFP_NOFAIL))
2346 css_get_many(&memcg->css, batch);
2347 if (batch > nr_pages)
2348 refill_stock(memcg, batch - nr_pages);
2353 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2355 if (mem_cgroup_is_root(memcg))
2358 page_counter_uncharge(&memcg->memory, nr_pages);
2359 if (do_swap_account)
2360 page_counter_uncharge(&memcg->memsw, nr_pages);
2362 css_put_many(&memcg->css, nr_pages);
2366 * A helper function to get mem_cgroup from ID. must be called under
2367 * rcu_read_lock(). The caller is responsible for calling
2368 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2369 * refcnt from swap can be called against removed memcg.)
2371 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2373 /* ID 0 is unused ID */
2376 return mem_cgroup_from_id(id);
2380 * try_get_mem_cgroup_from_page - look up page's memcg association
2383 * Look up, get a css reference, and return the memcg that owns @page.
2385 * The page must be locked to prevent racing with swap-in and page
2386 * cache charges. If coming from an unlocked page table, the caller
2387 * must ensure the page is on the LRU or this can race with charging.
2389 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2391 struct mem_cgroup *memcg;
2395 VM_BUG_ON_PAGE(!PageLocked(page), page);
2397 memcg = page->mem_cgroup;
2399 if (!css_tryget_online(&memcg->css))
2401 } else if (PageSwapCache(page)) {
2402 ent.val = page_private(page);
2403 id = lookup_swap_cgroup_id(ent);
2405 memcg = mem_cgroup_lookup(id);
2406 if (memcg && !css_tryget_online(&memcg->css))
2413 static void lock_page_lru(struct page *page, int *isolated)
2415 struct zone *zone = page_zone(page);
2417 spin_lock_irq(&zone->lru_lock);
2418 if (PageLRU(page)) {
2419 struct lruvec *lruvec;
2421 lruvec = mem_cgroup_page_lruvec(page, zone);
2423 del_page_from_lru_list(page, lruvec, page_lru(page));
2429 static void unlock_page_lru(struct page *page, int isolated)
2431 struct zone *zone = page_zone(page);
2434 struct lruvec *lruvec;
2436 lruvec = mem_cgroup_page_lruvec(page, zone);
2437 VM_BUG_ON_PAGE(PageLRU(page), page);
2439 add_page_to_lru_list(page, lruvec, page_lru(page));
2441 spin_unlock_irq(&zone->lru_lock);
2444 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2449 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2452 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2453 * may already be on some other mem_cgroup's LRU. Take care of it.
2456 lock_page_lru(page, &isolated);
2459 * Nobody should be changing or seriously looking at
2460 * page->mem_cgroup at this point:
2462 * - the page is uncharged
2464 * - the page is off-LRU
2466 * - an anonymous fault has exclusive page access, except for
2467 * a locked page table
2469 * - a page cache insertion, a swapin fault, or a migration
2470 * have the page locked
2472 page->mem_cgroup = memcg;
2475 unlock_page_lru(page, isolated);
2478 #ifdef CONFIG_MEMCG_KMEM
2480 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2481 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2483 static DEFINE_MUTEX(memcg_slab_mutex);
2486 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2487 * in the memcg_cache_params struct.
2489 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2491 struct kmem_cache *cachep;
2493 VM_BUG_ON(p->is_root_cache);
2494 cachep = p->root_cache;
2495 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2498 int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2499 unsigned long nr_pages)
2501 struct page_counter *counter;
2504 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2508 ret = try_charge(memcg, gfp, nr_pages);
2509 if (ret == -EINTR) {
2511 * try_charge() chose to bypass to root due to OOM kill or
2512 * fatal signal. Since our only options are to either fail
2513 * the allocation or charge it to this cgroup, do it as a
2514 * temporary condition. But we can't fail. From a kmem/slab
2515 * perspective, the cache has already been selected, by
2516 * mem_cgroup_kmem_get_cache(), so it is too late to change
2519 * This condition will only trigger if the task entered
2520 * memcg_charge_kmem in a sane state, but was OOM-killed
2521 * during try_charge() above. Tasks that were already dying
2522 * when the allocation triggers should have been already
2523 * directed to the root cgroup in memcontrol.h
2525 page_counter_charge(&memcg->memory, nr_pages);
2526 if (do_swap_account)
2527 page_counter_charge(&memcg->memsw, nr_pages);
2528 css_get_many(&memcg->css, nr_pages);
2531 page_counter_uncharge(&memcg->kmem, nr_pages);
2536 void memcg_uncharge_kmem(struct mem_cgroup *memcg, unsigned long nr_pages)
2538 page_counter_uncharge(&memcg->memory, nr_pages);
2539 if (do_swap_account)
2540 page_counter_uncharge(&memcg->memsw, nr_pages);
2542 page_counter_uncharge(&memcg->kmem, nr_pages);
2544 css_put_many(&memcg->css, nr_pages);
2548 * helper for acessing a memcg's index. It will be used as an index in the
2549 * child cache array in kmem_cache, and also to derive its name. This function
2550 * will return -1 when this is not a kmem-limited memcg.
2552 int memcg_cache_id(struct mem_cgroup *memcg)
2554 return memcg ? memcg->kmemcg_id : -1;
2557 static int memcg_alloc_cache_id(void)
2562 id = ida_simple_get(&kmem_limited_groups,
2563 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2567 if (id < memcg_limited_groups_array_size)
2571 * There's no space for the new id in memcg_caches arrays,
2572 * so we have to grow them.
2575 size = 2 * (id + 1);
2576 if (size < MEMCG_CACHES_MIN_SIZE)
2577 size = MEMCG_CACHES_MIN_SIZE;
2578 else if (size > MEMCG_CACHES_MAX_SIZE)
2579 size = MEMCG_CACHES_MAX_SIZE;
2581 mutex_lock(&memcg_slab_mutex);
2582 err = memcg_update_all_caches(size);
2583 mutex_unlock(&memcg_slab_mutex);
2586 ida_simple_remove(&kmem_limited_groups, id);
2592 static void memcg_free_cache_id(int id)
2594 ida_simple_remove(&kmem_limited_groups, id);
2598 * We should update the current array size iff all caches updates succeed. This
2599 * can only be done from the slab side. The slab mutex needs to be held when
2602 void memcg_update_array_size(int num)
2604 memcg_limited_groups_array_size = num;
2607 static void memcg_register_cache(struct mem_cgroup *memcg,
2608 struct kmem_cache *root_cache)
2610 static char memcg_name_buf[NAME_MAX + 1]; /* protected by
2612 struct kmem_cache *cachep;
2615 lockdep_assert_held(&memcg_slab_mutex);
2617 id = memcg_cache_id(memcg);
2620 * Since per-memcg caches are created asynchronously on first
2621 * allocation (see memcg_kmem_get_cache()), several threads can try to
2622 * create the same cache, but only one of them may succeed.
2624 if (cache_from_memcg_idx(root_cache, id))
2627 cgroup_name(memcg->css.cgroup, memcg_name_buf, NAME_MAX + 1);
2628 cachep = memcg_create_kmem_cache(memcg, root_cache, memcg_name_buf);
2630 * If we could not create a memcg cache, do not complain, because
2631 * that's not critical at all as we can always proceed with the root
2637 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2640 * Since readers won't lock (see cache_from_memcg_idx()), we need a
2641 * barrier here to ensure nobody will see the kmem_cache partially
2646 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
2647 root_cache->memcg_params->memcg_caches[id] = cachep;
2650 static void memcg_unregister_cache(struct kmem_cache *cachep)
2652 struct kmem_cache *root_cache;
2653 struct mem_cgroup *memcg;
2656 lockdep_assert_held(&memcg_slab_mutex);
2658 BUG_ON(is_root_cache(cachep));
2660 root_cache = cachep->memcg_params->root_cache;
2661 memcg = cachep->memcg_params->memcg;
2662 id = memcg_cache_id(memcg);
2664 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
2665 root_cache->memcg_params->memcg_caches[id] = NULL;
2667 list_del(&cachep->memcg_params->list);
2669 kmem_cache_destroy(cachep);
2672 int __memcg_cleanup_cache_params(struct kmem_cache *s)
2674 struct kmem_cache *c;
2677 mutex_lock(&memcg_slab_mutex);
2678 for_each_memcg_cache_index(i) {
2679 c = cache_from_memcg_idx(s, i);
2683 memcg_unregister_cache(c);
2685 if (cache_from_memcg_idx(s, i))
2688 mutex_unlock(&memcg_slab_mutex);
2692 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
2694 struct kmem_cache *cachep;
2695 struct memcg_cache_params *params, *tmp;
2697 if (!memcg_kmem_is_active(memcg))
2700 mutex_lock(&memcg_slab_mutex);
2701 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
2702 cachep = memcg_params_to_cache(params);
2703 memcg_unregister_cache(cachep);
2705 mutex_unlock(&memcg_slab_mutex);
2708 struct memcg_register_cache_work {
2709 struct mem_cgroup *memcg;
2710 struct kmem_cache *cachep;
2711 struct work_struct work;
2714 static void memcg_register_cache_func(struct work_struct *w)
2716 struct memcg_register_cache_work *cw =
2717 container_of(w, struct memcg_register_cache_work, work);
2718 struct mem_cgroup *memcg = cw->memcg;
2719 struct kmem_cache *cachep = cw->cachep;
2721 mutex_lock(&memcg_slab_mutex);
2722 memcg_register_cache(memcg, cachep);
2723 mutex_unlock(&memcg_slab_mutex);
2725 css_put(&memcg->css);
2730 * Enqueue the creation of a per-memcg kmem_cache.
2732 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
2733 struct kmem_cache *cachep)
2735 struct memcg_register_cache_work *cw;
2737 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2741 css_get(&memcg->css);
2744 cw->cachep = cachep;
2746 INIT_WORK(&cw->work, memcg_register_cache_func);
2747 schedule_work(&cw->work);
2750 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
2751 struct kmem_cache *cachep)
2754 * We need to stop accounting when we kmalloc, because if the
2755 * corresponding kmalloc cache is not yet created, the first allocation
2756 * in __memcg_schedule_register_cache will recurse.
2758 * However, it is better to enclose the whole function. Depending on
2759 * the debugging options enabled, INIT_WORK(), for instance, can
2760 * trigger an allocation. This too, will make us recurse. Because at
2761 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2762 * the safest choice is to do it like this, wrapping the whole function.
2764 current->memcg_kmem_skip_account = 1;
2765 __memcg_schedule_register_cache(memcg, cachep);
2766 current->memcg_kmem_skip_account = 0;
2770 * Return the kmem_cache we're supposed to use for a slab allocation.
2771 * We try to use the current memcg's version of the cache.
2773 * If the cache does not exist yet, if we are the first user of it,
2774 * we either create it immediately, if possible, or create it asynchronously
2776 * In the latter case, we will let the current allocation go through with
2777 * the original cache.
2779 * Can't be called in interrupt context or from kernel threads.
2780 * This function needs to be called with rcu_read_lock() held.
2782 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep)
2784 struct mem_cgroup *memcg;
2785 struct kmem_cache *memcg_cachep;
2787 VM_BUG_ON(!cachep->memcg_params);
2788 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
2790 if (current->memcg_kmem_skip_account)
2793 memcg = get_mem_cgroup_from_mm(current->mm);
2794 if (!memcg_kmem_is_active(memcg))
2797 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
2798 if (likely(memcg_cachep))
2799 return memcg_cachep;
2802 * If we are in a safe context (can wait, and not in interrupt
2803 * context), we could be be predictable and return right away.
2804 * This would guarantee that the allocation being performed
2805 * already belongs in the new cache.
2807 * However, there are some clashes that can arrive from locking.
2808 * For instance, because we acquire the slab_mutex while doing
2809 * memcg_create_kmem_cache, this means no further allocation
2810 * could happen with the slab_mutex held. So it's better to
2813 memcg_schedule_register_cache(memcg, cachep);
2815 css_put(&memcg->css);
2819 void __memcg_kmem_put_cache(struct kmem_cache *cachep)
2821 if (!is_root_cache(cachep))
2822 css_put(&cachep->memcg_params->memcg->css);
2826 * We need to verify if the allocation against current->mm->owner's memcg is
2827 * possible for the given order. But the page is not allocated yet, so we'll
2828 * need a further commit step to do the final arrangements.
2830 * It is possible for the task to switch cgroups in this mean time, so at
2831 * commit time, we can't rely on task conversion any longer. We'll then use
2832 * the handle argument to return to the caller which cgroup we should commit
2833 * against. We could also return the memcg directly and avoid the pointer
2834 * passing, but a boolean return value gives better semantics considering
2835 * the compiled-out case as well.
2837 * Returning true means the allocation is possible.
2840 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2842 struct mem_cgroup *memcg;
2847 memcg = get_mem_cgroup_from_mm(current->mm);
2849 if (!memcg_kmem_is_active(memcg)) {
2850 css_put(&memcg->css);
2854 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2858 css_put(&memcg->css);
2862 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2865 VM_BUG_ON(mem_cgroup_is_root(memcg));
2867 /* The page allocation failed. Revert */
2869 memcg_uncharge_kmem(memcg, 1 << order);
2872 page->mem_cgroup = memcg;
2875 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2877 struct mem_cgroup *memcg = page->mem_cgroup;
2882 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2884 memcg_uncharge_kmem(memcg, 1 << order);
2885 page->mem_cgroup = NULL;
2887 #endif /* CONFIG_MEMCG_KMEM */
2889 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2892 * Because tail pages are not marked as "used", set it. We're under
2893 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2894 * charge/uncharge will be never happen and move_account() is done under
2895 * compound_lock(), so we don't have to take care of races.
2897 void mem_cgroup_split_huge_fixup(struct page *head)
2901 if (mem_cgroup_disabled())
2904 for (i = 1; i < HPAGE_PMD_NR; i++)
2905 head[i].mem_cgroup = head->mem_cgroup;
2907 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2910 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2913 * mem_cgroup_move_account - move account of the page
2915 * @nr_pages: number of regular pages (>1 for huge pages)
2916 * @from: mem_cgroup which the page is moved from.
2917 * @to: mem_cgroup which the page is moved to. @from != @to.
2919 * The caller must confirm following.
2920 * - page is not on LRU (isolate_page() is useful.)
2921 * - compound_lock is held when nr_pages > 1
2923 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2926 static int mem_cgroup_move_account(struct page *page,
2927 unsigned int nr_pages,
2928 struct mem_cgroup *from,
2929 struct mem_cgroup *to)
2931 unsigned long flags;
2934 VM_BUG_ON(from == to);
2935 VM_BUG_ON_PAGE(PageLRU(page), page);
2937 * The page is isolated from LRU. So, collapse function
2938 * will not handle this page. But page splitting can happen.
2939 * Do this check under compound_page_lock(). The caller should
2943 if (nr_pages > 1 && !PageTransHuge(page))
2947 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
2948 * of its source page while we change it: page migration takes
2949 * both pages off the LRU, but page cache replacement doesn't.
2951 if (!trylock_page(page))
2955 if (page->mem_cgroup != from)
2958 spin_lock_irqsave(&from->move_lock, flags);
2960 if (!PageAnon(page) && page_mapped(page)) {
2961 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2963 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
2967 if (PageWriteback(page)) {
2968 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2970 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
2975 * It is safe to change page->mem_cgroup here because the page
2976 * is referenced, charged, and isolated - we can't race with
2977 * uncharging, charging, migration, or LRU putback.
2980 /* caller should have done css_get */
2981 page->mem_cgroup = to;
2982 spin_unlock_irqrestore(&from->move_lock, flags);
2986 local_irq_disable();
2987 mem_cgroup_charge_statistics(to, page, nr_pages);
2988 memcg_check_events(to, page);
2989 mem_cgroup_charge_statistics(from, page, -nr_pages);
2990 memcg_check_events(from, page);
2998 #ifdef CONFIG_MEMCG_SWAP
2999 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
3002 int val = (charge) ? 1 : -1;
3003 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
3007 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3008 * @entry: swap entry to be moved
3009 * @from: mem_cgroup which the entry is moved from
3010 * @to: mem_cgroup which the entry is moved to
3012 * It succeeds only when the swap_cgroup's record for this entry is the same
3013 * as the mem_cgroup's id of @from.
3015 * Returns 0 on success, -EINVAL on failure.
3017 * The caller must have charged to @to, IOW, called page_counter_charge() about
3018 * both res and memsw, and called css_get().
3020 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3021 struct mem_cgroup *from, struct mem_cgroup *to)
3023 unsigned short old_id, new_id;
3025 old_id = mem_cgroup_id(from);
3026 new_id = mem_cgroup_id(to);
3028 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3029 mem_cgroup_swap_statistics(from, false);
3030 mem_cgroup_swap_statistics(to, true);
3036 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3037 struct mem_cgroup *from, struct mem_cgroup *to)
3043 static DEFINE_MUTEX(memcg_limit_mutex);
3045 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3046 unsigned long limit)
3048 unsigned long curusage;
3049 unsigned long oldusage;
3050 bool enlarge = false;
3055 * For keeping hierarchical_reclaim simple, how long we should retry
3056 * is depends on callers. We set our retry-count to be function
3057 * of # of children which we should visit in this loop.
3059 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3060 mem_cgroup_count_children(memcg);
3062 oldusage = page_counter_read(&memcg->memory);
3065 if (signal_pending(current)) {
3070 mutex_lock(&memcg_limit_mutex);
3071 if (limit > memcg->memsw.limit) {
3072 mutex_unlock(&memcg_limit_mutex);
3076 if (limit > memcg->memory.limit)
3078 ret = page_counter_limit(&memcg->memory, limit);
3079 mutex_unlock(&memcg_limit_mutex);
3084 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
3086 curusage = page_counter_read(&memcg->memory);
3087 /* Usage is reduced ? */
3088 if (curusage >= oldusage)
3091 oldusage = curusage;
3092 } while (retry_count);
3094 if (!ret && enlarge)
3095 memcg_oom_recover(memcg);
3100 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3101 unsigned long limit)
3103 unsigned long curusage;
3104 unsigned long oldusage;
3105 bool enlarge = false;
3109 /* see mem_cgroup_resize_res_limit */
3110 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3111 mem_cgroup_count_children(memcg);
3113 oldusage = page_counter_read(&memcg->memsw);
3116 if (signal_pending(current)) {
3121 mutex_lock(&memcg_limit_mutex);
3122 if (limit < memcg->memory.limit) {
3123 mutex_unlock(&memcg_limit_mutex);
3127 if (limit > memcg->memsw.limit)
3129 ret = page_counter_limit(&memcg->memsw, limit);
3130 mutex_unlock(&memcg_limit_mutex);
3135 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3137 curusage = page_counter_read(&memcg->memsw);
3138 /* Usage is reduced ? */
3139 if (curusage >= oldusage)
3142 oldusage = curusage;
3143 } while (retry_count);
3145 if (!ret && enlarge)
3146 memcg_oom_recover(memcg);
3151 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3153 unsigned long *total_scanned)
3155 unsigned long nr_reclaimed = 0;
3156 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3157 unsigned long reclaimed;
3159 struct mem_cgroup_tree_per_zone *mctz;
3160 unsigned long excess;
3161 unsigned long nr_scanned;
3166 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3168 * This loop can run a while, specially if mem_cgroup's continuously
3169 * keep exceeding their soft limit and putting the system under
3176 mz = mem_cgroup_largest_soft_limit_node(mctz);
3181 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3182 gfp_mask, &nr_scanned);
3183 nr_reclaimed += reclaimed;
3184 *total_scanned += nr_scanned;
3185 spin_lock_irq(&mctz->lock);
3186 __mem_cgroup_remove_exceeded(mz, mctz);
3189 * If we failed to reclaim anything from this memory cgroup
3190 * it is time to move on to the next cgroup
3194 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3196 excess = soft_limit_excess(mz->memcg);
3198 * One school of thought says that we should not add
3199 * back the node to the tree if reclaim returns 0.
3200 * But our reclaim could return 0, simply because due
3201 * to priority we are exposing a smaller subset of
3202 * memory to reclaim from. Consider this as a longer
3205 /* If excess == 0, no tree ops */
3206 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3207 spin_unlock_irq(&mctz->lock);
3208 css_put(&mz->memcg->css);
3211 * Could not reclaim anything and there are no more
3212 * mem cgroups to try or we seem to be looping without
3213 * reclaiming anything.
3215 if (!nr_reclaimed &&
3217 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3219 } while (!nr_reclaimed);
3221 css_put(&next_mz->memcg->css);
3222 return nr_reclaimed;
3226 * Test whether @memcg has children, dead or alive. Note that this
3227 * function doesn't care whether @memcg has use_hierarchy enabled and
3228 * returns %true if there are child csses according to the cgroup
3229 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3231 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3236 * The lock does not prevent addition or deletion of children, but
3237 * it prevents a new child from being initialized based on this
3238 * parent in css_online(), so it's enough to decide whether
3239 * hierarchically inherited attributes can still be changed or not.
3241 lockdep_assert_held(&memcg_create_mutex);
3244 ret = css_next_child(NULL, &memcg->css);
3250 * Reclaims as many pages from the given memcg as possible and moves
3251 * the rest to the parent.
3253 * Caller is responsible for holding css reference for memcg.
3255 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3257 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3259 /* we call try-to-free pages for make this cgroup empty */
3260 lru_add_drain_all();
3261 /* try to free all pages in this cgroup */
3262 while (nr_retries && page_counter_read(&memcg->memory)) {
3265 if (signal_pending(current))
3268 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3272 /* maybe some writeback is necessary */
3273 congestion_wait(BLK_RW_ASYNC, HZ/10);
3281 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3282 char *buf, size_t nbytes,
3285 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3287 if (mem_cgroup_is_root(memcg))
3289 return mem_cgroup_force_empty(memcg) ?: nbytes;
3292 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3295 return mem_cgroup_from_css(css)->use_hierarchy;
3298 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3299 struct cftype *cft, u64 val)
3302 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3303 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3305 mutex_lock(&memcg_create_mutex);
3307 if (memcg->use_hierarchy == val)
3311 * If parent's use_hierarchy is set, we can't make any modifications
3312 * in the child subtrees. If it is unset, then the change can
3313 * occur, provided the current cgroup has no children.
3315 * For the root cgroup, parent_mem is NULL, we allow value to be
3316 * set if there are no children.
3318 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3319 (val == 1 || val == 0)) {
3320 if (!memcg_has_children(memcg))
3321 memcg->use_hierarchy = val;
3328 mutex_unlock(&memcg_create_mutex);
3333 static unsigned long tree_stat(struct mem_cgroup *memcg,
3334 enum mem_cgroup_stat_index idx)
3336 struct mem_cgroup *iter;
3339 /* Per-cpu values can be negative, use a signed accumulator */
3340 for_each_mem_cgroup_tree(iter, memcg)
3341 val += mem_cgroup_read_stat(iter, idx);
3343 if (val < 0) /* race ? */
3348 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3352 if (mem_cgroup_is_root(memcg)) {
3353 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3354 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3356 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3359 val = page_counter_read(&memcg->memory);
3361 val = page_counter_read(&memcg->memsw);
3363 return val << PAGE_SHIFT;
3374 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3377 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3378 struct page_counter *counter;
3380 switch (MEMFILE_TYPE(cft->private)) {
3382 counter = &memcg->memory;
3385 counter = &memcg->memsw;
3388 counter = &memcg->kmem;
3394 switch (MEMFILE_ATTR(cft->private)) {
3396 if (counter == &memcg->memory)
3397 return mem_cgroup_usage(memcg, false);
3398 if (counter == &memcg->memsw)
3399 return mem_cgroup_usage(memcg, true);
3400 return (u64)page_counter_read(counter) * PAGE_SIZE;
3402 return (u64)counter->limit * PAGE_SIZE;
3404 return (u64)counter->watermark * PAGE_SIZE;
3406 return counter->failcnt;
3407 case RES_SOFT_LIMIT:
3408 return (u64)memcg->soft_limit * PAGE_SIZE;
3414 #ifdef CONFIG_MEMCG_KMEM
3415 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3416 unsigned long nr_pages)
3421 if (memcg_kmem_is_active(memcg))
3425 * For simplicity, we won't allow this to be disabled. It also can't
3426 * be changed if the cgroup has children already, or if tasks had
3429 * If tasks join before we set the limit, a person looking at
3430 * kmem.usage_in_bytes will have no way to determine when it took
3431 * place, which makes the value quite meaningless.
3433 * After it first became limited, changes in the value of the limit are
3434 * of course permitted.
3436 mutex_lock(&memcg_create_mutex);
3437 if (cgroup_has_tasks(memcg->css.cgroup) ||
3438 (memcg->use_hierarchy && memcg_has_children(memcg)))
3440 mutex_unlock(&memcg_create_mutex);
3444 memcg_id = memcg_alloc_cache_id();
3451 * We couldn't have accounted to this cgroup, because it hasn't got
3452 * activated yet, so this should succeed.
3454 err = page_counter_limit(&memcg->kmem, nr_pages);
3457 static_key_slow_inc(&memcg_kmem_enabled_key);
3459 * A memory cgroup is considered kmem-active as soon as it gets
3460 * kmemcg_id. Setting the id after enabling static branching will
3461 * guarantee no one starts accounting before all call sites are
3464 memcg->kmemcg_id = memcg_id;
3469 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3470 unsigned long limit)
3474 mutex_lock(&memcg_limit_mutex);
3475 if (!memcg_kmem_is_active(memcg))
3476 ret = memcg_activate_kmem(memcg, limit);
3478 ret = page_counter_limit(&memcg->kmem, limit);
3479 mutex_unlock(&memcg_limit_mutex);
3483 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3486 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3491 mutex_lock(&memcg_limit_mutex);
3493 * If the parent cgroup is not kmem-active now, it cannot be activated
3494 * after this point, because it has at least one child already.
3496 if (memcg_kmem_is_active(parent))
3497 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3498 mutex_unlock(&memcg_limit_mutex);
3502 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3503 unsigned long limit)
3507 #endif /* CONFIG_MEMCG_KMEM */
3510 * The user of this function is...
3513 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3514 char *buf, size_t nbytes, loff_t off)
3516 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3517 unsigned long nr_pages;
3520 buf = strstrip(buf);
3521 ret = page_counter_memparse(buf, &nr_pages);
3525 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3527 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3531 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3533 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3536 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3539 ret = memcg_update_kmem_limit(memcg, nr_pages);
3543 case RES_SOFT_LIMIT:
3544 memcg->soft_limit = nr_pages;
3548 return ret ?: nbytes;
3551 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3552 size_t nbytes, loff_t off)
3554 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3555 struct page_counter *counter;
3557 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3559 counter = &memcg->memory;
3562 counter = &memcg->memsw;
3565 counter = &memcg->kmem;
3571 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3573 page_counter_reset_watermark(counter);
3576 counter->failcnt = 0;
3585 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3588 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3592 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3593 struct cftype *cft, u64 val)
3595 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3597 if (val >= (1 << NR_MOVE_TYPE))
3601 * No kind of locking is needed in here, because ->can_attach() will
3602 * check this value once in the beginning of the process, and then carry
3603 * on with stale data. This means that changes to this value will only
3604 * affect task migrations starting after the change.
3606 memcg->move_charge_at_immigrate = val;
3610 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3611 struct cftype *cft, u64 val)
3618 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3622 unsigned int lru_mask;
3625 static const struct numa_stat stats[] = {
3626 { "total", LRU_ALL },
3627 { "file", LRU_ALL_FILE },
3628 { "anon", LRU_ALL_ANON },
3629 { "unevictable", BIT(LRU_UNEVICTABLE) },
3631 const struct numa_stat *stat;
3634 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3636 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3637 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3638 seq_printf(m, "%s=%lu", stat->name, nr);
3639 for_each_node_state(nid, N_MEMORY) {
3640 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3642 seq_printf(m, " N%d=%lu", nid, nr);
3647 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3648 struct mem_cgroup *iter;
3651 for_each_mem_cgroup_tree(iter, memcg)
3652 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3653 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3654 for_each_node_state(nid, N_MEMORY) {
3656 for_each_mem_cgroup_tree(iter, memcg)
3657 nr += mem_cgroup_node_nr_lru_pages(
3658 iter, nid, stat->lru_mask);
3659 seq_printf(m, " N%d=%lu", nid, nr);
3666 #endif /* CONFIG_NUMA */
3668 static int memcg_stat_show(struct seq_file *m, void *v)
3670 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3671 unsigned long memory, memsw;
3672 struct mem_cgroup *mi;
3675 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3677 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3678 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3680 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3681 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3684 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3685 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3686 mem_cgroup_read_events(memcg, i));
3688 for (i = 0; i < NR_LRU_LISTS; i++)
3689 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3690 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3692 /* Hierarchical information */
3693 memory = memsw = PAGE_COUNTER_MAX;
3694 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3695 memory = min(memory, mi->memory.limit);
3696 memsw = min(memsw, mi->memsw.limit);
3698 seq_printf(m, "hierarchical_memory_limit %llu\n",
3699 (u64)memory * PAGE_SIZE);
3700 if (do_swap_account)
3701 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3702 (u64)memsw * PAGE_SIZE);
3704 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3707 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3709 for_each_mem_cgroup_tree(mi, memcg)
3710 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3711 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3714 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3715 unsigned long long val = 0;
3717 for_each_mem_cgroup_tree(mi, memcg)
3718 val += mem_cgroup_read_events(mi, i);
3719 seq_printf(m, "total_%s %llu\n",
3720 mem_cgroup_events_names[i], val);
3723 for (i = 0; i < NR_LRU_LISTS; i++) {
3724 unsigned long long val = 0;
3726 for_each_mem_cgroup_tree(mi, memcg)
3727 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3728 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3731 #ifdef CONFIG_DEBUG_VM
3734 struct mem_cgroup_per_zone *mz;
3735 struct zone_reclaim_stat *rstat;
3736 unsigned long recent_rotated[2] = {0, 0};
3737 unsigned long recent_scanned[2] = {0, 0};
3739 for_each_online_node(nid)
3740 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3741 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3742 rstat = &mz->lruvec.reclaim_stat;
3744 recent_rotated[0] += rstat->recent_rotated[0];
3745 recent_rotated[1] += rstat->recent_rotated[1];
3746 recent_scanned[0] += rstat->recent_scanned[0];
3747 recent_scanned[1] += rstat->recent_scanned[1];
3749 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3750 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3751 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3752 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3759 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3762 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3764 return mem_cgroup_swappiness(memcg);
3767 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3768 struct cftype *cft, u64 val)
3770 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3776 memcg->swappiness = val;
3778 vm_swappiness = val;
3783 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3785 struct mem_cgroup_threshold_ary *t;
3786 unsigned long usage;
3791 t = rcu_dereference(memcg->thresholds.primary);
3793 t = rcu_dereference(memcg->memsw_thresholds.primary);
3798 usage = mem_cgroup_usage(memcg, swap);
3801 * current_threshold points to threshold just below or equal to usage.
3802 * If it's not true, a threshold was crossed after last
3803 * call of __mem_cgroup_threshold().
3805 i = t->current_threshold;
3808 * Iterate backward over array of thresholds starting from
3809 * current_threshold and check if a threshold is crossed.
3810 * If none of thresholds below usage is crossed, we read
3811 * only one element of the array here.
3813 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3814 eventfd_signal(t->entries[i].eventfd, 1);
3816 /* i = current_threshold + 1 */
3820 * Iterate forward over array of thresholds starting from
3821 * current_threshold+1 and check if a threshold is crossed.
3822 * If none of thresholds above usage is crossed, we read
3823 * only one element of the array here.
3825 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3826 eventfd_signal(t->entries[i].eventfd, 1);
3828 /* Update current_threshold */
3829 t->current_threshold = i - 1;
3834 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3837 __mem_cgroup_threshold(memcg, false);
3838 if (do_swap_account)
3839 __mem_cgroup_threshold(memcg, true);
3841 memcg = parent_mem_cgroup(memcg);
3845 static int compare_thresholds(const void *a, const void *b)
3847 const struct mem_cgroup_threshold *_a = a;
3848 const struct mem_cgroup_threshold *_b = b;
3850 if (_a->threshold > _b->threshold)
3853 if (_a->threshold < _b->threshold)
3859 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3861 struct mem_cgroup_eventfd_list *ev;
3863 spin_lock(&memcg_oom_lock);
3865 list_for_each_entry(ev, &memcg->oom_notify, list)
3866 eventfd_signal(ev->eventfd, 1);
3868 spin_unlock(&memcg_oom_lock);
3872 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3874 struct mem_cgroup *iter;
3876 for_each_mem_cgroup_tree(iter, memcg)
3877 mem_cgroup_oom_notify_cb(iter);
3880 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3881 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3883 struct mem_cgroup_thresholds *thresholds;
3884 struct mem_cgroup_threshold_ary *new;
3885 unsigned long threshold;
3886 unsigned long usage;
3889 ret = page_counter_memparse(args, &threshold);
3893 mutex_lock(&memcg->thresholds_lock);
3896 thresholds = &memcg->thresholds;
3897 usage = mem_cgroup_usage(memcg, false);
3898 } else if (type == _MEMSWAP) {
3899 thresholds = &memcg->memsw_thresholds;
3900 usage = mem_cgroup_usage(memcg, true);
3904 /* Check if a threshold crossed before adding a new one */
3905 if (thresholds->primary)
3906 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3908 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3910 /* Allocate memory for new array of thresholds */
3911 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3919 /* Copy thresholds (if any) to new array */
3920 if (thresholds->primary) {
3921 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3922 sizeof(struct mem_cgroup_threshold));
3925 /* Add new threshold */
3926 new->entries[size - 1].eventfd = eventfd;
3927 new->entries[size - 1].threshold = threshold;
3929 /* Sort thresholds. Registering of new threshold isn't time-critical */
3930 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3931 compare_thresholds, NULL);
3933 /* Find current threshold */
3934 new->current_threshold = -1;
3935 for (i = 0; i < size; i++) {
3936 if (new->entries[i].threshold <= usage) {
3938 * new->current_threshold will not be used until
3939 * rcu_assign_pointer(), so it's safe to increment
3942 ++new->current_threshold;
3947 /* Free old spare buffer and save old primary buffer as spare */
3948 kfree(thresholds->spare);
3949 thresholds->spare = thresholds->primary;
3951 rcu_assign_pointer(thresholds->primary, new);
3953 /* To be sure that nobody uses thresholds */
3957 mutex_unlock(&memcg->thresholds_lock);
3962 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3963 struct eventfd_ctx *eventfd, const char *args)
3965 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
3968 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
3969 struct eventfd_ctx *eventfd, const char *args)
3971 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
3974 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
3975 struct eventfd_ctx *eventfd, enum res_type type)
3977 struct mem_cgroup_thresholds *thresholds;
3978 struct mem_cgroup_threshold_ary *new;
3979 unsigned long usage;
3982 mutex_lock(&memcg->thresholds_lock);
3985 thresholds = &memcg->thresholds;
3986 usage = mem_cgroup_usage(memcg, false);
3987 } else if (type == _MEMSWAP) {
3988 thresholds = &memcg->memsw_thresholds;
3989 usage = mem_cgroup_usage(memcg, true);
3993 if (!thresholds->primary)
3996 /* Check if a threshold crossed before removing */
3997 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3999 /* Calculate new number of threshold */
4001 for (i = 0; i < thresholds->primary->size; i++) {
4002 if (thresholds->primary->entries[i].eventfd != eventfd)
4006 new = thresholds->spare;
4008 /* Set thresholds array to NULL if we don't have thresholds */
4017 /* Copy thresholds and find current threshold */
4018 new->current_threshold = -1;
4019 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4020 if (thresholds->primary->entries[i].eventfd == eventfd)
4023 new->entries[j] = thresholds->primary->entries[i];
4024 if (new->entries[j].threshold <= usage) {
4026 * new->current_threshold will not be used
4027 * until rcu_assign_pointer(), so it's safe to increment
4030 ++new->current_threshold;
4036 /* Swap primary and spare array */
4037 thresholds->spare = thresholds->primary;
4038 /* If all events are unregistered, free the spare array */
4040 kfree(thresholds->spare);
4041 thresholds->spare = NULL;
4044 rcu_assign_pointer(thresholds->primary, new);
4046 /* To be sure that nobody uses thresholds */
4049 mutex_unlock(&memcg->thresholds_lock);
4052 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4053 struct eventfd_ctx *eventfd)
4055 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4058 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4059 struct eventfd_ctx *eventfd)
4061 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4064 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4065 struct eventfd_ctx *eventfd, const char *args)
4067 struct mem_cgroup_eventfd_list *event;
4069 event = kmalloc(sizeof(*event), GFP_KERNEL);
4073 spin_lock(&memcg_oom_lock);
4075 event->eventfd = eventfd;
4076 list_add(&event->list, &memcg->oom_notify);
4078 /* already in OOM ? */
4079 if (atomic_read(&memcg->under_oom))
4080 eventfd_signal(eventfd, 1);
4081 spin_unlock(&memcg_oom_lock);
4086 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4087 struct eventfd_ctx *eventfd)
4089 struct mem_cgroup_eventfd_list *ev, *tmp;
4091 spin_lock(&memcg_oom_lock);
4093 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4094 if (ev->eventfd == eventfd) {
4095 list_del(&ev->list);
4100 spin_unlock(&memcg_oom_lock);
4103 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4105 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
4107 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4108 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
4112 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4113 struct cftype *cft, u64 val)
4115 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4117 /* cannot set to root cgroup and only 0 and 1 are allowed */
4118 if (!css->parent || !((val == 0) || (val == 1)))
4121 memcg->oom_kill_disable = val;
4123 memcg_oom_recover(memcg);
4128 #ifdef CONFIG_MEMCG_KMEM
4129 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4133 ret = memcg_propagate_kmem(memcg);
4137 return mem_cgroup_sockets_init(memcg, ss);
4140 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4142 memcg_unregister_all_caches(memcg);
4143 mem_cgroup_sockets_destroy(memcg);
4146 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4151 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4157 * DO NOT USE IN NEW FILES.
4159 * "cgroup.event_control" implementation.
4161 * This is way over-engineered. It tries to support fully configurable
4162 * events for each user. Such level of flexibility is completely
4163 * unnecessary especially in the light of the planned unified hierarchy.
4165 * Please deprecate this and replace with something simpler if at all
4170 * Unregister event and free resources.
4172 * Gets called from workqueue.
4174 static void memcg_event_remove(struct work_struct *work)
4176 struct mem_cgroup_event *event =
4177 container_of(work, struct mem_cgroup_event, remove);
4178 struct mem_cgroup *memcg = event->memcg;
4180 remove_wait_queue(event->wqh, &event->wait);
4182 event->unregister_event(memcg, event->eventfd);
4184 /* Notify userspace the event is going away. */
4185 eventfd_signal(event->eventfd, 1);
4187 eventfd_ctx_put(event->eventfd);
4189 css_put(&memcg->css);
4193 * Gets called on POLLHUP on eventfd when user closes it.
4195 * Called with wqh->lock held and interrupts disabled.
4197 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4198 int sync, void *key)
4200 struct mem_cgroup_event *event =
4201 container_of(wait, struct mem_cgroup_event, wait);
4202 struct mem_cgroup *memcg = event->memcg;
4203 unsigned long flags = (unsigned long)key;
4205 if (flags & POLLHUP) {
4207 * If the event has been detached at cgroup removal, we
4208 * can simply return knowing the other side will cleanup
4211 * We can't race against event freeing since the other
4212 * side will require wqh->lock via remove_wait_queue(),
4215 spin_lock(&memcg->event_list_lock);
4216 if (!list_empty(&event->list)) {
4217 list_del_init(&event->list);
4219 * We are in atomic context, but cgroup_event_remove()
4220 * may sleep, so we have to call it in workqueue.
4222 schedule_work(&event->remove);
4224 spin_unlock(&memcg->event_list_lock);
4230 static void memcg_event_ptable_queue_proc(struct file *file,
4231 wait_queue_head_t *wqh, poll_table *pt)
4233 struct mem_cgroup_event *event =
4234 container_of(pt, struct mem_cgroup_event, pt);
4237 add_wait_queue(wqh, &event->wait);
4241 * DO NOT USE IN NEW FILES.
4243 * Parse input and register new cgroup event handler.
4245 * Input must be in format '<event_fd> <control_fd> <args>'.
4246 * Interpretation of args is defined by control file implementation.
4248 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4249 char *buf, size_t nbytes, loff_t off)
4251 struct cgroup_subsys_state *css = of_css(of);
4252 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4253 struct mem_cgroup_event *event;
4254 struct cgroup_subsys_state *cfile_css;
4255 unsigned int efd, cfd;
4262 buf = strstrip(buf);
4264 efd = simple_strtoul(buf, &endp, 10);
4269 cfd = simple_strtoul(buf, &endp, 10);
4270 if ((*endp != ' ') && (*endp != '\0'))
4274 event = kzalloc(sizeof(*event), GFP_KERNEL);
4278 event->memcg = memcg;
4279 INIT_LIST_HEAD(&event->list);
4280 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4281 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4282 INIT_WORK(&event->remove, memcg_event_remove);
4290 event->eventfd = eventfd_ctx_fileget(efile.file);
4291 if (IS_ERR(event->eventfd)) {
4292 ret = PTR_ERR(event->eventfd);
4299 goto out_put_eventfd;
4302 /* the process need read permission on control file */
4303 /* AV: shouldn't we check that it's been opened for read instead? */
4304 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4309 * Determine the event callbacks and set them in @event. This used
4310 * to be done via struct cftype but cgroup core no longer knows
4311 * about these events. The following is crude but the whole thing
4312 * is for compatibility anyway.
4314 * DO NOT ADD NEW FILES.
4316 name = cfile.file->f_path.dentry->d_name.name;
4318 if (!strcmp(name, "memory.usage_in_bytes")) {
4319 event->register_event = mem_cgroup_usage_register_event;
4320 event->unregister_event = mem_cgroup_usage_unregister_event;
4321 } else if (!strcmp(name, "memory.oom_control")) {
4322 event->register_event = mem_cgroup_oom_register_event;
4323 event->unregister_event = mem_cgroup_oom_unregister_event;
4324 } else if (!strcmp(name, "memory.pressure_level")) {
4325 event->register_event = vmpressure_register_event;
4326 event->unregister_event = vmpressure_unregister_event;
4327 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4328 event->register_event = memsw_cgroup_usage_register_event;
4329 event->unregister_event = memsw_cgroup_usage_unregister_event;
4336 * Verify @cfile should belong to @css. Also, remaining events are
4337 * automatically removed on cgroup destruction but the removal is
4338 * asynchronous, so take an extra ref on @css.
4340 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4341 &memory_cgrp_subsys);
4343 if (IS_ERR(cfile_css))
4345 if (cfile_css != css) {
4350 ret = event->register_event(memcg, event->eventfd, buf);
4354 efile.file->f_op->poll(efile.file, &event->pt);
4356 spin_lock(&memcg->event_list_lock);
4357 list_add(&event->list, &memcg->event_list);
4358 spin_unlock(&memcg->event_list_lock);
4370 eventfd_ctx_put(event->eventfd);
4379 static struct cftype mem_cgroup_files[] = {
4381 .name = "usage_in_bytes",
4382 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4383 .read_u64 = mem_cgroup_read_u64,
4386 .name = "max_usage_in_bytes",
4387 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4388 .write = mem_cgroup_reset,
4389 .read_u64 = mem_cgroup_read_u64,
4392 .name = "limit_in_bytes",
4393 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4394 .write = mem_cgroup_write,
4395 .read_u64 = mem_cgroup_read_u64,
4398 .name = "soft_limit_in_bytes",
4399 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4400 .write = mem_cgroup_write,
4401 .read_u64 = mem_cgroup_read_u64,
4405 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4406 .write = mem_cgroup_reset,
4407 .read_u64 = mem_cgroup_read_u64,
4411 .seq_show = memcg_stat_show,
4414 .name = "force_empty",
4415 .write = mem_cgroup_force_empty_write,
4418 .name = "use_hierarchy",
4419 .write_u64 = mem_cgroup_hierarchy_write,
4420 .read_u64 = mem_cgroup_hierarchy_read,
4423 .name = "cgroup.event_control", /* XXX: for compat */
4424 .write = memcg_write_event_control,
4425 .flags = CFTYPE_NO_PREFIX,
4429 .name = "swappiness",
4430 .read_u64 = mem_cgroup_swappiness_read,
4431 .write_u64 = mem_cgroup_swappiness_write,
4434 .name = "move_charge_at_immigrate",
4435 .read_u64 = mem_cgroup_move_charge_read,
4436 .write_u64 = mem_cgroup_move_charge_write,
4439 .name = "oom_control",
4440 .seq_show = mem_cgroup_oom_control_read,
4441 .write_u64 = mem_cgroup_oom_control_write,
4442 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4445 .name = "pressure_level",
4449 .name = "numa_stat",
4450 .seq_show = memcg_numa_stat_show,
4453 #ifdef CONFIG_MEMCG_KMEM
4455 .name = "kmem.limit_in_bytes",
4456 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4457 .write = mem_cgroup_write,
4458 .read_u64 = mem_cgroup_read_u64,
4461 .name = "kmem.usage_in_bytes",
4462 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4463 .read_u64 = mem_cgroup_read_u64,
4466 .name = "kmem.failcnt",
4467 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4468 .write = mem_cgroup_reset,
4469 .read_u64 = mem_cgroup_read_u64,
4472 .name = "kmem.max_usage_in_bytes",
4473 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4474 .write = mem_cgroup_reset,
4475 .read_u64 = mem_cgroup_read_u64,
4477 #ifdef CONFIG_SLABINFO
4479 .name = "kmem.slabinfo",
4480 .seq_start = slab_start,
4481 .seq_next = slab_next,
4482 .seq_stop = slab_stop,
4483 .seq_show = memcg_slab_show,
4487 { }, /* terminate */
4490 #ifdef CONFIG_MEMCG_SWAP
4491 static struct cftype memsw_cgroup_files[] = {
4493 .name = "memsw.usage_in_bytes",
4494 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4495 .read_u64 = mem_cgroup_read_u64,
4498 .name = "memsw.max_usage_in_bytes",
4499 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4500 .write = mem_cgroup_reset,
4501 .read_u64 = mem_cgroup_read_u64,
4504 .name = "memsw.limit_in_bytes",
4505 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4506 .write = mem_cgroup_write,
4507 .read_u64 = mem_cgroup_read_u64,
4510 .name = "memsw.failcnt",
4511 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4512 .write = mem_cgroup_reset,
4513 .read_u64 = mem_cgroup_read_u64,
4515 { }, /* terminate */
4518 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4520 struct mem_cgroup_per_node *pn;
4521 struct mem_cgroup_per_zone *mz;
4522 int zone, tmp = node;
4524 * This routine is called against possible nodes.
4525 * But it's BUG to call kmalloc() against offline node.
4527 * TODO: this routine can waste much memory for nodes which will
4528 * never be onlined. It's better to use memory hotplug callback
4531 if (!node_state(node, N_NORMAL_MEMORY))
4533 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4537 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4538 mz = &pn->zoneinfo[zone];
4539 lruvec_init(&mz->lruvec);
4540 mz->usage_in_excess = 0;
4541 mz->on_tree = false;
4544 memcg->nodeinfo[node] = pn;
4548 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4550 kfree(memcg->nodeinfo[node]);
4553 static struct mem_cgroup *mem_cgroup_alloc(void)
4555 struct mem_cgroup *memcg;
4558 size = sizeof(struct mem_cgroup);
4559 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4561 memcg = kzalloc(size, GFP_KERNEL);
4565 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4568 spin_lock_init(&memcg->pcp_counter_lock);
4577 * At destroying mem_cgroup, references from swap_cgroup can remain.
4578 * (scanning all at force_empty is too costly...)
4580 * Instead of clearing all references at force_empty, we remember
4581 * the number of reference from swap_cgroup and free mem_cgroup when
4582 * it goes down to 0.
4584 * Removal of cgroup itself succeeds regardless of refs from swap.
4587 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4591 mem_cgroup_remove_from_trees(memcg);
4594 free_mem_cgroup_per_zone_info(memcg, node);
4596 free_percpu(memcg->stat);
4598 disarm_static_keys(memcg);
4603 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4605 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4607 if (!memcg->memory.parent)
4609 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4611 EXPORT_SYMBOL(parent_mem_cgroup);
4613 static void __init mem_cgroup_soft_limit_tree_init(void)
4615 struct mem_cgroup_tree_per_node *rtpn;
4616 struct mem_cgroup_tree_per_zone *rtpz;
4617 int tmp, node, zone;
4619 for_each_node(node) {
4621 if (!node_state(node, N_NORMAL_MEMORY))
4623 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4626 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4628 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4629 rtpz = &rtpn->rb_tree_per_zone[zone];
4630 rtpz->rb_root = RB_ROOT;
4631 spin_lock_init(&rtpz->lock);
4636 static struct cgroup_subsys_state * __ref
4637 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4639 struct mem_cgroup *memcg;
4640 long error = -ENOMEM;
4643 memcg = mem_cgroup_alloc();
4645 return ERR_PTR(error);
4648 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4652 if (parent_css == NULL) {
4653 root_mem_cgroup = memcg;
4654 page_counter_init(&memcg->memory, NULL);
4655 memcg->soft_limit = PAGE_COUNTER_MAX;
4656 page_counter_init(&memcg->memsw, NULL);
4657 page_counter_init(&memcg->kmem, NULL);
4660 memcg->last_scanned_node = MAX_NUMNODES;
4661 INIT_LIST_HEAD(&memcg->oom_notify);
4662 memcg->move_charge_at_immigrate = 0;
4663 mutex_init(&memcg->thresholds_lock);
4664 spin_lock_init(&memcg->move_lock);
4665 vmpressure_init(&memcg->vmpressure);
4666 INIT_LIST_HEAD(&memcg->event_list);
4667 spin_lock_init(&memcg->event_list_lock);
4668 #ifdef CONFIG_MEMCG_KMEM
4669 memcg->kmemcg_id = -1;
4670 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
4676 __mem_cgroup_free(memcg);
4677 return ERR_PTR(error);
4681 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4683 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4684 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4687 if (css->id > MEM_CGROUP_ID_MAX)
4693 mutex_lock(&memcg_create_mutex);
4695 memcg->use_hierarchy = parent->use_hierarchy;
4696 memcg->oom_kill_disable = parent->oom_kill_disable;
4697 memcg->swappiness = mem_cgroup_swappiness(parent);
4699 if (parent->use_hierarchy) {
4700 page_counter_init(&memcg->memory, &parent->memory);
4701 memcg->soft_limit = PAGE_COUNTER_MAX;
4702 page_counter_init(&memcg->memsw, &parent->memsw);
4703 page_counter_init(&memcg->kmem, &parent->kmem);
4706 * No need to take a reference to the parent because cgroup
4707 * core guarantees its existence.
4710 page_counter_init(&memcg->memory, NULL);
4711 memcg->soft_limit = PAGE_COUNTER_MAX;
4712 page_counter_init(&memcg->memsw, NULL);
4713 page_counter_init(&memcg->kmem, NULL);
4715 * Deeper hierachy with use_hierarchy == false doesn't make
4716 * much sense so let cgroup subsystem know about this
4717 * unfortunate state in our controller.
4719 if (parent != root_mem_cgroup)
4720 memory_cgrp_subsys.broken_hierarchy = true;
4722 mutex_unlock(&memcg_create_mutex);
4724 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4729 * Make sure the memcg is initialized: mem_cgroup_iter()
4730 * orders reading memcg->initialized against its callers
4731 * reading the memcg members.
4733 smp_store_release(&memcg->initialized, 1);
4738 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4740 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4741 struct mem_cgroup_event *event, *tmp;
4744 * Unregister events and notify userspace.
4745 * Notify userspace about cgroup removing only after rmdir of cgroup
4746 * directory to avoid race between userspace and kernelspace.
4748 spin_lock(&memcg->event_list_lock);
4749 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4750 list_del_init(&event->list);
4751 schedule_work(&event->remove);
4753 spin_unlock(&memcg->event_list_lock);
4755 vmpressure_cleanup(&memcg->vmpressure);
4758 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4760 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4762 memcg_destroy_kmem(memcg);
4763 __mem_cgroup_free(memcg);
4767 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4768 * @css: the target css
4770 * Reset the states of the mem_cgroup associated with @css. This is
4771 * invoked when the userland requests disabling on the default hierarchy
4772 * but the memcg is pinned through dependency. The memcg should stop
4773 * applying policies and should revert to the vanilla state as it may be
4774 * made visible again.
4776 * The current implementation only resets the essential configurations.
4777 * This needs to be expanded to cover all the visible parts.
4779 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4781 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4783 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4784 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4785 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4786 memcg->soft_limit = PAGE_COUNTER_MAX;
4790 /* Handlers for move charge at task migration. */
4791 static int mem_cgroup_do_precharge(unsigned long count)
4795 /* Try a single bulk charge without reclaim first */
4796 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4798 mc.precharge += count;
4801 if (ret == -EINTR) {
4802 cancel_charge(root_mem_cgroup, count);
4806 /* Try charges one by one with reclaim */
4808 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4810 * In case of failure, any residual charges against
4811 * mc.to will be dropped by mem_cgroup_clear_mc()
4812 * later on. However, cancel any charges that are
4813 * bypassed to root right away or they'll be lost.
4816 cancel_charge(root_mem_cgroup, 1);
4826 * get_mctgt_type - get target type of moving charge
4827 * @vma: the vma the pte to be checked belongs
4828 * @addr: the address corresponding to the pte to be checked
4829 * @ptent: the pte to be checked
4830 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4833 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4834 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4835 * move charge. if @target is not NULL, the page is stored in target->page
4836 * with extra refcnt got(Callers should handle it).
4837 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4838 * target for charge migration. if @target is not NULL, the entry is stored
4841 * Called with pte lock held.
4848 enum mc_target_type {
4854 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4855 unsigned long addr, pte_t ptent)
4857 struct page *page = vm_normal_page(vma, addr, ptent);
4859 if (!page || !page_mapped(page))
4861 if (PageAnon(page)) {
4862 /* we don't move shared anon */
4865 } else if (!move_file())
4866 /* we ignore mapcount for file pages */
4868 if (!get_page_unless_zero(page))
4875 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4876 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4878 struct page *page = NULL;
4879 swp_entry_t ent = pte_to_swp_entry(ptent);
4881 if (!move_anon() || non_swap_entry(ent))
4884 * Because lookup_swap_cache() updates some statistics counter,
4885 * we call find_get_page() with swapper_space directly.
4887 page = find_get_page(swap_address_space(ent), ent.val);
4888 if (do_swap_account)
4889 entry->val = ent.val;
4894 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4895 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4901 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4902 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4904 struct page *page = NULL;
4905 struct address_space *mapping;
4908 if (!vma->vm_file) /* anonymous vma */
4913 mapping = vma->vm_file->f_mapping;
4914 pgoff = linear_page_index(vma, addr);
4916 /* page is moved even if it's not RSS of this task(page-faulted). */
4918 /* shmem/tmpfs may report page out on swap: account for that too. */
4919 if (shmem_mapping(mapping)) {
4920 page = find_get_entry(mapping, pgoff);
4921 if (radix_tree_exceptional_entry(page)) {
4922 swp_entry_t swp = radix_to_swp_entry(page);
4923 if (do_swap_account)
4925 page = find_get_page(swap_address_space(swp), swp.val);
4928 page = find_get_page(mapping, pgoff);
4930 page = find_get_page(mapping, pgoff);
4935 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
4936 unsigned long addr, pte_t ptent, union mc_target *target)
4938 struct page *page = NULL;
4939 enum mc_target_type ret = MC_TARGET_NONE;
4940 swp_entry_t ent = { .val = 0 };
4942 if (pte_present(ptent))
4943 page = mc_handle_present_pte(vma, addr, ptent);
4944 else if (is_swap_pte(ptent))
4945 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4946 else if (pte_none(ptent))
4947 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4949 if (!page && !ent.val)
4953 * Do only loose check w/o serialization.
4954 * mem_cgroup_move_account() checks the page is valid or
4955 * not under LRU exclusion.
4957 if (page->mem_cgroup == mc.from) {
4958 ret = MC_TARGET_PAGE;
4960 target->page = page;
4962 if (!ret || !target)
4965 /* There is a swap entry and a page doesn't exist or isn't charged */
4966 if (ent.val && !ret &&
4967 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
4968 ret = MC_TARGET_SWAP;
4975 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4977 * We don't consider swapping or file mapped pages because THP does not
4978 * support them for now.
4979 * Caller should make sure that pmd_trans_huge(pmd) is true.
4981 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
4982 unsigned long addr, pmd_t pmd, union mc_target *target)
4984 struct page *page = NULL;
4985 enum mc_target_type ret = MC_TARGET_NONE;
4987 page = pmd_page(pmd);
4988 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
4991 if (page->mem_cgroup == mc.from) {
4992 ret = MC_TARGET_PAGE;
4995 target->page = page;
5001 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5002 unsigned long addr, pmd_t pmd, union mc_target *target)
5004 return MC_TARGET_NONE;
5008 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5009 unsigned long addr, unsigned long end,
5010 struct mm_walk *walk)
5012 struct vm_area_struct *vma = walk->private;
5016 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5017 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5018 mc.precharge += HPAGE_PMD_NR;
5023 if (pmd_trans_unstable(pmd))
5025 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5026 for (; addr != end; pte++, addr += PAGE_SIZE)
5027 if (get_mctgt_type(vma, addr, *pte, NULL))
5028 mc.precharge++; /* increment precharge temporarily */
5029 pte_unmap_unlock(pte - 1, ptl);
5035 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5037 unsigned long precharge;
5038 struct vm_area_struct *vma;
5040 down_read(&mm->mmap_sem);
5041 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5042 struct mm_walk mem_cgroup_count_precharge_walk = {
5043 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5047 if (is_vm_hugetlb_page(vma))
5049 walk_page_range(vma->vm_start, vma->vm_end,
5050 &mem_cgroup_count_precharge_walk);
5052 up_read(&mm->mmap_sem);
5054 precharge = mc.precharge;
5060 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5062 unsigned long precharge = mem_cgroup_count_precharge(mm);
5064 VM_BUG_ON(mc.moving_task);
5065 mc.moving_task = current;
5066 return mem_cgroup_do_precharge(precharge);
5069 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5070 static void __mem_cgroup_clear_mc(void)
5072 struct mem_cgroup *from = mc.from;
5073 struct mem_cgroup *to = mc.to;
5075 /* we must uncharge all the leftover precharges from mc.to */
5077 cancel_charge(mc.to, mc.precharge);
5081 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5082 * we must uncharge here.
5084 if (mc.moved_charge) {
5085 cancel_charge(mc.from, mc.moved_charge);
5086 mc.moved_charge = 0;
5088 /* we must fixup refcnts and charges */
5089 if (mc.moved_swap) {
5090 /* uncharge swap account from the old cgroup */
5091 if (!mem_cgroup_is_root(mc.from))
5092 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5095 * we charged both to->memory and to->memsw, so we
5096 * should uncharge to->memory.
5098 if (!mem_cgroup_is_root(mc.to))
5099 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5101 css_put_many(&mc.from->css, mc.moved_swap);
5103 /* we've already done css_get(mc.to) */
5106 memcg_oom_recover(from);
5107 memcg_oom_recover(to);
5108 wake_up_all(&mc.waitq);
5111 static void mem_cgroup_clear_mc(void)
5114 * we must clear moving_task before waking up waiters at the end of
5117 mc.moving_task = NULL;
5118 __mem_cgroup_clear_mc();
5119 spin_lock(&mc.lock);
5122 spin_unlock(&mc.lock);
5125 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5126 struct cgroup_taskset *tset)
5128 struct task_struct *p = cgroup_taskset_first(tset);
5130 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5131 unsigned long move_charge_at_immigrate;
5134 * We are now commited to this value whatever it is. Changes in this
5135 * tunable will only affect upcoming migrations, not the current one.
5136 * So we need to save it, and keep it going.
5138 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
5139 if (move_charge_at_immigrate) {
5140 struct mm_struct *mm;
5141 struct mem_cgroup *from = mem_cgroup_from_task(p);
5143 VM_BUG_ON(from == memcg);
5145 mm = get_task_mm(p);
5148 /* We move charges only when we move a owner of the mm */
5149 if (mm->owner == p) {
5152 VM_BUG_ON(mc.precharge);
5153 VM_BUG_ON(mc.moved_charge);
5154 VM_BUG_ON(mc.moved_swap);
5156 spin_lock(&mc.lock);
5159 mc.immigrate_flags = move_charge_at_immigrate;
5160 spin_unlock(&mc.lock);
5161 /* We set mc.moving_task later */
5163 ret = mem_cgroup_precharge_mc(mm);
5165 mem_cgroup_clear_mc();
5172 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5173 struct cgroup_taskset *tset)
5176 mem_cgroup_clear_mc();
5179 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5180 unsigned long addr, unsigned long end,
5181 struct mm_walk *walk)
5184 struct vm_area_struct *vma = walk->private;
5187 enum mc_target_type target_type;
5188 union mc_target target;
5192 * We don't take compound_lock() here but no race with splitting thp
5194 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5195 * under splitting, which means there's no concurrent thp split,
5196 * - if another thread runs into split_huge_page() just after we
5197 * entered this if-block, the thread must wait for page table lock
5198 * to be unlocked in __split_huge_page_splitting(), where the main
5199 * part of thp split is not executed yet.
5201 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5202 if (mc.precharge < HPAGE_PMD_NR) {
5206 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5207 if (target_type == MC_TARGET_PAGE) {
5209 if (!isolate_lru_page(page)) {
5210 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5212 mc.precharge -= HPAGE_PMD_NR;
5213 mc.moved_charge += HPAGE_PMD_NR;
5215 putback_lru_page(page);
5223 if (pmd_trans_unstable(pmd))
5226 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5227 for (; addr != end; addr += PAGE_SIZE) {
5228 pte_t ptent = *(pte++);
5234 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5235 case MC_TARGET_PAGE:
5237 if (isolate_lru_page(page))
5239 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5241 /* we uncharge from mc.from later. */
5244 putback_lru_page(page);
5245 put: /* get_mctgt_type() gets the page */
5248 case MC_TARGET_SWAP:
5250 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5252 /* we fixup refcnts and charges later. */
5260 pte_unmap_unlock(pte - 1, ptl);
5265 * We have consumed all precharges we got in can_attach().
5266 * We try charge one by one, but don't do any additional
5267 * charges to mc.to if we have failed in charge once in attach()
5270 ret = mem_cgroup_do_precharge(1);
5278 static void mem_cgroup_move_charge(struct mm_struct *mm)
5280 struct vm_area_struct *vma;
5282 lru_add_drain_all();
5284 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5285 * move_lock while we're moving its pages to another memcg.
5286 * Then wait for already started RCU-only updates to finish.
5288 atomic_inc(&mc.from->moving_account);
5291 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5293 * Someone who are holding the mmap_sem might be waiting in
5294 * waitq. So we cancel all extra charges, wake up all waiters,
5295 * and retry. Because we cancel precharges, we might not be able
5296 * to move enough charges, but moving charge is a best-effort
5297 * feature anyway, so it wouldn't be a big problem.
5299 __mem_cgroup_clear_mc();
5303 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5305 struct mm_walk mem_cgroup_move_charge_walk = {
5306 .pmd_entry = mem_cgroup_move_charge_pte_range,
5310 if (is_vm_hugetlb_page(vma))
5312 ret = walk_page_range(vma->vm_start, vma->vm_end,
5313 &mem_cgroup_move_charge_walk);
5316 * means we have consumed all precharges and failed in
5317 * doing additional charge. Just abandon here.
5321 up_read(&mm->mmap_sem);
5322 atomic_dec(&mc.from->moving_account);
5325 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5326 struct cgroup_taskset *tset)
5328 struct task_struct *p = cgroup_taskset_first(tset);
5329 struct mm_struct *mm = get_task_mm(p);
5333 mem_cgroup_move_charge(mm);
5337 mem_cgroup_clear_mc();
5339 #else /* !CONFIG_MMU */
5340 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5341 struct cgroup_taskset *tset)
5345 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5346 struct cgroup_taskset *tset)
5349 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5350 struct cgroup_taskset *tset)
5356 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5357 * to verify whether we're attached to the default hierarchy on each mount
5360 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5363 * use_hierarchy is forced on the default hierarchy. cgroup core
5364 * guarantees that @root doesn't have any children, so turning it
5365 * on for the root memcg is enough.
5367 if (cgroup_on_dfl(root_css->cgroup))
5368 mem_cgroup_from_css(root_css)->use_hierarchy = true;
5371 struct cgroup_subsys memory_cgrp_subsys = {
5372 .css_alloc = mem_cgroup_css_alloc,
5373 .css_online = mem_cgroup_css_online,
5374 .css_offline = mem_cgroup_css_offline,
5375 .css_free = mem_cgroup_css_free,
5376 .css_reset = mem_cgroup_css_reset,
5377 .can_attach = mem_cgroup_can_attach,
5378 .cancel_attach = mem_cgroup_cancel_attach,
5379 .attach = mem_cgroup_move_task,
5380 .bind = mem_cgroup_bind,
5381 .legacy_cftypes = mem_cgroup_files,
5385 #ifdef CONFIG_MEMCG_SWAP
5386 static int __init enable_swap_account(char *s)
5388 if (!strcmp(s, "1"))
5389 really_do_swap_account = 1;
5390 else if (!strcmp(s, "0"))
5391 really_do_swap_account = 0;
5394 __setup("swapaccount=", enable_swap_account);
5396 static void __init memsw_file_init(void)
5398 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5399 memsw_cgroup_files));
5402 static void __init enable_swap_cgroup(void)
5404 if (!mem_cgroup_disabled() && really_do_swap_account) {
5405 do_swap_account = 1;
5411 static void __init enable_swap_cgroup(void)
5416 #ifdef CONFIG_MEMCG_SWAP
5418 * mem_cgroup_swapout - transfer a memsw charge to swap
5419 * @page: page whose memsw charge to transfer
5420 * @entry: swap entry to move the charge to
5422 * Transfer the memsw charge of @page to @entry.
5424 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5426 struct mem_cgroup *memcg;
5427 unsigned short oldid;
5429 VM_BUG_ON_PAGE(PageLRU(page), page);
5430 VM_BUG_ON_PAGE(page_count(page), page);
5432 if (!do_swap_account)
5435 memcg = page->mem_cgroup;
5437 /* Readahead page, never charged */
5441 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5442 VM_BUG_ON_PAGE(oldid, page);
5443 mem_cgroup_swap_statistics(memcg, true);
5445 page->mem_cgroup = NULL;
5447 if (!mem_cgroup_is_root(memcg))
5448 page_counter_uncharge(&memcg->memory, 1);
5450 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5451 VM_BUG_ON(!irqs_disabled());
5453 mem_cgroup_charge_statistics(memcg, page, -1);
5454 memcg_check_events(memcg, page);
5458 * mem_cgroup_uncharge_swap - uncharge a swap entry
5459 * @entry: swap entry to uncharge
5461 * Drop the memsw charge associated with @entry.
5463 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5465 struct mem_cgroup *memcg;
5468 if (!do_swap_account)
5471 id = swap_cgroup_record(entry, 0);
5473 memcg = mem_cgroup_lookup(id);
5475 if (!mem_cgroup_is_root(memcg))
5476 page_counter_uncharge(&memcg->memsw, 1);
5477 mem_cgroup_swap_statistics(memcg, false);
5478 css_put(&memcg->css);
5485 * mem_cgroup_try_charge - try charging a page
5486 * @page: page to charge
5487 * @mm: mm context of the victim
5488 * @gfp_mask: reclaim mode
5489 * @memcgp: charged memcg return
5491 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5492 * pages according to @gfp_mask if necessary.
5494 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5495 * Otherwise, an error code is returned.
5497 * After page->mapping has been set up, the caller must finalize the
5498 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5499 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5501 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5502 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5504 struct mem_cgroup *memcg = NULL;
5505 unsigned int nr_pages = 1;
5508 if (mem_cgroup_disabled())
5511 if (PageSwapCache(page)) {
5513 * Every swap fault against a single page tries to charge the
5514 * page, bail as early as possible. shmem_unuse() encounters
5515 * already charged pages, too. The USED bit is protected by
5516 * the page lock, which serializes swap cache removal, which
5517 * in turn serializes uncharging.
5519 if (page->mem_cgroup)
5523 if (PageTransHuge(page)) {
5524 nr_pages <<= compound_order(page);
5525 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5528 if (do_swap_account && PageSwapCache(page))
5529 memcg = try_get_mem_cgroup_from_page(page);
5531 memcg = get_mem_cgroup_from_mm(mm);
5533 ret = try_charge(memcg, gfp_mask, nr_pages);
5535 css_put(&memcg->css);
5537 if (ret == -EINTR) {
5538 memcg = root_mem_cgroup;
5547 * mem_cgroup_commit_charge - commit a page charge
5548 * @page: page to charge
5549 * @memcg: memcg to charge the page to
5550 * @lrucare: page might be on LRU already
5552 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5553 * after page->mapping has been set up. This must happen atomically
5554 * as part of the page instantiation, i.e. under the page table lock
5555 * for anonymous pages, under the page lock for page and swap cache.
5557 * In addition, the page must not be on the LRU during the commit, to
5558 * prevent racing with task migration. If it might be, use @lrucare.
5560 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5562 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5565 unsigned int nr_pages = 1;
5567 VM_BUG_ON_PAGE(!page->mapping, page);
5568 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5570 if (mem_cgroup_disabled())
5573 * Swap faults will attempt to charge the same page multiple
5574 * times. But reuse_swap_page() might have removed the page
5575 * from swapcache already, so we can't check PageSwapCache().
5580 commit_charge(page, memcg, lrucare);
5582 if (PageTransHuge(page)) {
5583 nr_pages <<= compound_order(page);
5584 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5587 local_irq_disable();
5588 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5589 memcg_check_events(memcg, page);
5592 if (do_swap_account && PageSwapCache(page)) {
5593 swp_entry_t entry = { .val = page_private(page) };
5595 * The swap entry might not get freed for a long time,
5596 * let's not wait for it. The page already received a
5597 * memory+swap charge, drop the swap entry duplicate.
5599 mem_cgroup_uncharge_swap(entry);
5604 * mem_cgroup_cancel_charge - cancel a page charge
5605 * @page: page to charge
5606 * @memcg: memcg to charge the page to
5608 * Cancel a charge transaction started by mem_cgroup_try_charge().
5610 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5612 unsigned int nr_pages = 1;
5614 if (mem_cgroup_disabled())
5617 * Swap faults will attempt to charge the same page multiple
5618 * times. But reuse_swap_page() might have removed the page
5619 * from swapcache already, so we can't check PageSwapCache().
5624 if (PageTransHuge(page)) {
5625 nr_pages <<= compound_order(page);
5626 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5629 cancel_charge(memcg, nr_pages);
5632 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5633 unsigned long nr_anon, unsigned long nr_file,
5634 unsigned long nr_huge, struct page *dummy_page)
5636 unsigned long nr_pages = nr_anon + nr_file;
5637 unsigned long flags;
5639 if (!mem_cgroup_is_root(memcg)) {
5640 page_counter_uncharge(&memcg->memory, nr_pages);
5641 if (do_swap_account)
5642 page_counter_uncharge(&memcg->memsw, nr_pages);
5643 memcg_oom_recover(memcg);
5646 local_irq_save(flags);
5647 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5648 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5649 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5650 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5651 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5652 memcg_check_events(memcg, dummy_page);
5653 local_irq_restore(flags);
5655 if (!mem_cgroup_is_root(memcg))
5656 css_put_many(&memcg->css, nr_pages);
5659 static void uncharge_list(struct list_head *page_list)
5661 struct mem_cgroup *memcg = NULL;
5662 unsigned long nr_anon = 0;
5663 unsigned long nr_file = 0;
5664 unsigned long nr_huge = 0;
5665 unsigned long pgpgout = 0;
5666 struct list_head *next;
5669 next = page_list->next;
5671 unsigned int nr_pages = 1;
5673 page = list_entry(next, struct page, lru);
5674 next = page->lru.next;
5676 VM_BUG_ON_PAGE(PageLRU(page), page);
5677 VM_BUG_ON_PAGE(page_count(page), page);
5679 if (!page->mem_cgroup)
5683 * Nobody should be changing or seriously looking at
5684 * page->mem_cgroup at this point, we have fully
5685 * exclusive access to the page.
5688 if (memcg != page->mem_cgroup) {
5690 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5692 pgpgout = nr_anon = nr_file = nr_huge = 0;
5694 memcg = page->mem_cgroup;
5697 if (PageTransHuge(page)) {
5698 nr_pages <<= compound_order(page);
5699 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5700 nr_huge += nr_pages;
5704 nr_anon += nr_pages;
5706 nr_file += nr_pages;
5708 page->mem_cgroup = NULL;
5711 } while (next != page_list);
5714 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5719 * mem_cgroup_uncharge - uncharge a page
5720 * @page: page to uncharge
5722 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5723 * mem_cgroup_commit_charge().
5725 void mem_cgroup_uncharge(struct page *page)
5727 if (mem_cgroup_disabled())
5730 /* Don't touch page->lru of any random page, pre-check: */
5731 if (!page->mem_cgroup)
5734 INIT_LIST_HEAD(&page->lru);
5735 uncharge_list(&page->lru);
5739 * mem_cgroup_uncharge_list - uncharge a list of page
5740 * @page_list: list of pages to uncharge
5742 * Uncharge a list of pages previously charged with
5743 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5745 void mem_cgroup_uncharge_list(struct list_head *page_list)
5747 if (mem_cgroup_disabled())
5750 if (!list_empty(page_list))
5751 uncharge_list(page_list);
5755 * mem_cgroup_migrate - migrate a charge to another page
5756 * @oldpage: currently charged page
5757 * @newpage: page to transfer the charge to
5758 * @lrucare: either or both pages might be on the LRU already
5760 * Migrate the charge from @oldpage to @newpage.
5762 * Both pages must be locked, @newpage->mapping must be set up.
5764 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5767 struct mem_cgroup *memcg;
5770 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5771 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5772 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5773 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5774 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5775 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5778 if (mem_cgroup_disabled())
5781 /* Page cache replacement: new page already charged? */
5782 if (newpage->mem_cgroup)
5786 * Swapcache readahead pages can get migrated before being
5787 * charged, and migration from compaction can happen to an
5788 * uncharged page when the PFN walker finds a page that
5789 * reclaim just put back on the LRU but has not released yet.
5791 memcg = oldpage->mem_cgroup;
5796 lock_page_lru(oldpage, &isolated);
5798 oldpage->mem_cgroup = NULL;
5801 unlock_page_lru(oldpage, isolated);
5803 commit_charge(newpage, memcg, lrucare);
5807 * subsys_initcall() for memory controller.
5809 * Some parts like hotcpu_notifier() have to be initialized from this context
5810 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5811 * everything that doesn't depend on a specific mem_cgroup structure should
5812 * be initialized from here.
5814 static int __init mem_cgroup_init(void)
5816 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5817 enable_swap_cgroup();
5818 mem_cgroup_soft_limit_tree_init();
5822 subsys_initcall(mem_cgroup_init);