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_info(" 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) || current->flags & PF_EXITING) {
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");
1620 * test_mem_cgroup_node_reclaimable
1621 * @memcg: the target memcg
1622 * @nid: the node ID to be checked.
1623 * @noswap : specify true here if the user wants flle only information.
1625 * This function returns whether the specified memcg contains any
1626 * reclaimable pages on a node. Returns true if there are any reclaimable
1627 * pages in the node.
1629 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1630 int nid, bool noswap)
1632 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1634 if (noswap || !total_swap_pages)
1636 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1641 #if MAX_NUMNODES > 1
1644 * Always updating the nodemask is not very good - even if we have an empty
1645 * list or the wrong list here, we can start from some node and traverse all
1646 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1649 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1653 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1654 * pagein/pageout changes since the last update.
1656 if (!atomic_read(&memcg->numainfo_events))
1658 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1661 /* make a nodemask where this memcg uses memory from */
1662 memcg->scan_nodes = node_states[N_MEMORY];
1664 for_each_node_mask(nid, node_states[N_MEMORY]) {
1666 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1667 node_clear(nid, memcg->scan_nodes);
1670 atomic_set(&memcg->numainfo_events, 0);
1671 atomic_set(&memcg->numainfo_updating, 0);
1675 * Selecting a node where we start reclaim from. Because what we need is just
1676 * reducing usage counter, start from anywhere is O,K. Considering
1677 * memory reclaim from current node, there are pros. and cons.
1679 * Freeing memory from current node means freeing memory from a node which
1680 * we'll use or we've used. So, it may make LRU bad. And if several threads
1681 * hit limits, it will see a contention on a node. But freeing from remote
1682 * node means more costs for memory reclaim because of memory latency.
1684 * Now, we use round-robin. Better algorithm is welcomed.
1686 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1690 mem_cgroup_may_update_nodemask(memcg);
1691 node = memcg->last_scanned_node;
1693 node = next_node(node, memcg->scan_nodes);
1694 if (node == MAX_NUMNODES)
1695 node = first_node(memcg->scan_nodes);
1697 * We call this when we hit limit, not when pages are added to LRU.
1698 * No LRU may hold pages because all pages are UNEVICTABLE or
1699 * memcg is too small and all pages are not on LRU. In that case,
1700 * we use curret node.
1702 if (unlikely(node == MAX_NUMNODES))
1703 node = numa_node_id();
1705 memcg->last_scanned_node = node;
1709 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1715 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1718 unsigned long *total_scanned)
1720 struct mem_cgroup *victim = NULL;
1723 unsigned long excess;
1724 unsigned long nr_scanned;
1725 struct mem_cgroup_reclaim_cookie reclaim = {
1730 excess = soft_limit_excess(root_memcg);
1733 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1738 * If we have not been able to reclaim
1739 * anything, it might because there are
1740 * no reclaimable pages under this hierarchy
1745 * We want to do more targeted reclaim.
1746 * excess >> 2 is not to excessive so as to
1747 * reclaim too much, nor too less that we keep
1748 * coming back to reclaim from this cgroup
1750 if (total >= (excess >> 2) ||
1751 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1756 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1758 *total_scanned += nr_scanned;
1759 if (!soft_limit_excess(root_memcg))
1762 mem_cgroup_iter_break(root_memcg, victim);
1766 #ifdef CONFIG_LOCKDEP
1767 static struct lockdep_map memcg_oom_lock_dep_map = {
1768 .name = "memcg_oom_lock",
1772 static DEFINE_SPINLOCK(memcg_oom_lock);
1775 * Check OOM-Killer is already running under our hierarchy.
1776 * If someone is running, return false.
1778 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1780 struct mem_cgroup *iter, *failed = NULL;
1782 spin_lock(&memcg_oom_lock);
1784 for_each_mem_cgroup_tree(iter, memcg) {
1785 if (iter->oom_lock) {
1787 * this subtree of our hierarchy is already locked
1788 * so we cannot give a lock.
1791 mem_cgroup_iter_break(memcg, iter);
1794 iter->oom_lock = true;
1799 * OK, we failed to lock the whole subtree so we have
1800 * to clean up what we set up to the failing subtree
1802 for_each_mem_cgroup_tree(iter, memcg) {
1803 if (iter == failed) {
1804 mem_cgroup_iter_break(memcg, iter);
1807 iter->oom_lock = false;
1810 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
1812 spin_unlock(&memcg_oom_lock);
1817 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1819 struct mem_cgroup *iter;
1821 spin_lock(&memcg_oom_lock);
1822 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
1823 for_each_mem_cgroup_tree(iter, memcg)
1824 iter->oom_lock = false;
1825 spin_unlock(&memcg_oom_lock);
1828 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1830 struct mem_cgroup *iter;
1832 for_each_mem_cgroup_tree(iter, memcg)
1833 atomic_inc(&iter->under_oom);
1836 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1838 struct mem_cgroup *iter;
1841 * When a new child is created while the hierarchy is under oom,
1842 * mem_cgroup_oom_lock() may not be called. We have to use
1843 * atomic_add_unless() here.
1845 for_each_mem_cgroup_tree(iter, memcg)
1846 atomic_add_unless(&iter->under_oom, -1, 0);
1849 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1851 struct oom_wait_info {
1852 struct mem_cgroup *memcg;
1856 static int memcg_oom_wake_function(wait_queue_t *wait,
1857 unsigned mode, int sync, void *arg)
1859 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1860 struct mem_cgroup *oom_wait_memcg;
1861 struct oom_wait_info *oom_wait_info;
1863 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1864 oom_wait_memcg = oom_wait_info->memcg;
1866 if (!mem_cgroup_is_descendant(wake_memcg, oom_wait_memcg) &&
1867 !mem_cgroup_is_descendant(oom_wait_memcg, wake_memcg))
1869 return autoremove_wake_function(wait, mode, sync, arg);
1872 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1874 atomic_inc(&memcg->oom_wakeups);
1875 /* for filtering, pass "memcg" as argument. */
1876 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1879 static void memcg_oom_recover(struct mem_cgroup *memcg)
1881 if (memcg && atomic_read(&memcg->under_oom))
1882 memcg_wakeup_oom(memcg);
1885 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1887 if (!current->memcg_oom.may_oom)
1890 * We are in the middle of the charge context here, so we
1891 * don't want to block when potentially sitting on a callstack
1892 * that holds all kinds of filesystem and mm locks.
1894 * Also, the caller may handle a failed allocation gracefully
1895 * (like optional page cache readahead) and so an OOM killer
1896 * invocation might not even be necessary.
1898 * That's why we don't do anything here except remember the
1899 * OOM context and then deal with it at the end of the page
1900 * fault when the stack is unwound, the locks are released,
1901 * and when we know whether the fault was overall successful.
1903 css_get(&memcg->css);
1904 current->memcg_oom.memcg = memcg;
1905 current->memcg_oom.gfp_mask = mask;
1906 current->memcg_oom.order = order;
1910 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1911 * @handle: actually kill/wait or just clean up the OOM state
1913 * This has to be called at the end of a page fault if the memcg OOM
1914 * handler was enabled.
1916 * Memcg supports userspace OOM handling where failed allocations must
1917 * sleep on a waitqueue until the userspace task resolves the
1918 * situation. Sleeping directly in the charge context with all kinds
1919 * of locks held is not a good idea, instead we remember an OOM state
1920 * in the task and mem_cgroup_oom_synchronize() has to be called at
1921 * the end of the page fault to complete the OOM handling.
1923 * Returns %true if an ongoing memcg OOM situation was detected and
1924 * completed, %false otherwise.
1926 bool mem_cgroup_oom_synchronize(bool handle)
1928 struct mem_cgroup *memcg = current->memcg_oom.memcg;
1929 struct oom_wait_info owait;
1932 /* OOM is global, do not handle */
1939 owait.memcg = memcg;
1940 owait.wait.flags = 0;
1941 owait.wait.func = memcg_oom_wake_function;
1942 owait.wait.private = current;
1943 INIT_LIST_HEAD(&owait.wait.task_list);
1945 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1946 mem_cgroup_mark_under_oom(memcg);
1948 locked = mem_cgroup_oom_trylock(memcg);
1951 mem_cgroup_oom_notify(memcg);
1953 if (locked && !memcg->oom_kill_disable) {
1954 mem_cgroup_unmark_under_oom(memcg);
1955 finish_wait(&memcg_oom_waitq, &owait.wait);
1956 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
1957 current->memcg_oom.order);
1960 mem_cgroup_unmark_under_oom(memcg);
1961 finish_wait(&memcg_oom_waitq, &owait.wait);
1965 mem_cgroup_oom_unlock(memcg);
1967 * There is no guarantee that an OOM-lock contender
1968 * sees the wakeups triggered by the OOM kill
1969 * uncharges. Wake any sleepers explicitely.
1971 memcg_oom_recover(memcg);
1974 current->memcg_oom.memcg = NULL;
1975 css_put(&memcg->css);
1980 * mem_cgroup_begin_page_stat - begin a page state statistics transaction
1981 * @page: page that is going to change accounted state
1982 * @locked: &memcg->move_lock slowpath was taken
1983 * @flags: IRQ-state flags for &memcg->move_lock
1985 * This function must mark the beginning of an accounted page state
1986 * change to prevent double accounting when the page is concurrently
1987 * being moved to another memcg:
1989 * memcg = mem_cgroup_begin_page_stat(page, &locked, &flags);
1990 * if (TestClearPageState(page))
1991 * mem_cgroup_update_page_stat(memcg, state, -1);
1992 * mem_cgroup_end_page_stat(memcg, locked, flags);
1994 * The RCU lock is held throughout the transaction. The fast path can
1995 * get away without acquiring the memcg->move_lock (@locked is false)
1996 * because page moving starts with an RCU grace period.
1998 * The RCU lock also protects the memcg from being freed when the page
1999 * state that is going to change is the only thing preventing the page
2000 * from being uncharged. E.g. end-writeback clearing PageWriteback(),
2001 * which allows migration to go ahead and uncharge the page before the
2002 * account transaction might be complete.
2004 struct mem_cgroup *mem_cgroup_begin_page_stat(struct page *page,
2006 unsigned long *flags)
2008 struct mem_cgroup *memcg;
2012 if (mem_cgroup_disabled())
2015 memcg = page->mem_cgroup;
2016 if (unlikely(!memcg))
2020 if (atomic_read(&memcg->moving_account) <= 0)
2023 spin_lock_irqsave(&memcg->move_lock, *flags);
2024 if (memcg != page->mem_cgroup) {
2025 spin_unlock_irqrestore(&memcg->move_lock, *flags);
2034 * mem_cgroup_end_page_stat - finish a page state statistics transaction
2035 * @memcg: the memcg that was accounted against
2036 * @locked: value received from mem_cgroup_begin_page_stat()
2037 * @flags: value received from mem_cgroup_begin_page_stat()
2039 void mem_cgroup_end_page_stat(struct mem_cgroup *memcg, bool *locked,
2040 unsigned long *flags)
2042 if (memcg && *locked)
2043 spin_unlock_irqrestore(&memcg->move_lock, *flags);
2049 * mem_cgroup_update_page_stat - update page state statistics
2050 * @memcg: memcg to account against
2051 * @idx: page state item to account
2052 * @val: number of pages (positive or negative)
2054 * See mem_cgroup_begin_page_stat() for locking requirements.
2056 void mem_cgroup_update_page_stat(struct mem_cgroup *memcg,
2057 enum mem_cgroup_stat_index idx, int val)
2059 VM_BUG_ON(!rcu_read_lock_held());
2062 this_cpu_add(memcg->stat->count[idx], val);
2066 * size of first charge trial. "32" comes from vmscan.c's magic value.
2067 * TODO: maybe necessary to use big numbers in big irons.
2069 #define CHARGE_BATCH 32U
2070 struct memcg_stock_pcp {
2071 struct mem_cgroup *cached; /* this never be root cgroup */
2072 unsigned int nr_pages;
2073 struct work_struct work;
2074 unsigned long flags;
2075 #define FLUSHING_CACHED_CHARGE 0
2077 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2078 static DEFINE_MUTEX(percpu_charge_mutex);
2081 * consume_stock: Try to consume stocked charge on this cpu.
2082 * @memcg: memcg to consume from.
2083 * @nr_pages: how many pages to charge.
2085 * The charges will only happen if @memcg matches the current cpu's memcg
2086 * stock, and at least @nr_pages are available in that stock. Failure to
2087 * service an allocation will refill the stock.
2089 * returns true if successful, false otherwise.
2091 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2093 struct memcg_stock_pcp *stock;
2096 if (nr_pages > CHARGE_BATCH)
2099 stock = &get_cpu_var(memcg_stock);
2100 if (memcg == stock->cached && stock->nr_pages >= nr_pages) {
2101 stock->nr_pages -= nr_pages;
2104 put_cpu_var(memcg_stock);
2109 * Returns stocks cached in percpu and reset cached information.
2111 static void drain_stock(struct memcg_stock_pcp *stock)
2113 struct mem_cgroup *old = stock->cached;
2115 if (stock->nr_pages) {
2116 page_counter_uncharge(&old->memory, stock->nr_pages);
2117 if (do_swap_account)
2118 page_counter_uncharge(&old->memsw, stock->nr_pages);
2119 css_put_many(&old->css, stock->nr_pages);
2120 stock->nr_pages = 0;
2122 stock->cached = NULL;
2126 * This must be called under preempt disabled or must be called by
2127 * a thread which is pinned to local cpu.
2129 static void drain_local_stock(struct work_struct *dummy)
2131 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2133 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2136 static void __init memcg_stock_init(void)
2140 for_each_possible_cpu(cpu) {
2141 struct memcg_stock_pcp *stock =
2142 &per_cpu(memcg_stock, cpu);
2143 INIT_WORK(&stock->work, drain_local_stock);
2148 * Cache charges(val) to local per_cpu area.
2149 * This will be consumed by consume_stock() function, later.
2151 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2153 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2155 if (stock->cached != memcg) { /* reset if necessary */
2157 stock->cached = memcg;
2159 stock->nr_pages += nr_pages;
2160 put_cpu_var(memcg_stock);
2164 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2165 * of the hierarchy under it.
2167 static void drain_all_stock(struct mem_cgroup *root_memcg)
2171 /* If someone's already draining, avoid adding running more workers. */
2172 if (!mutex_trylock(&percpu_charge_mutex))
2174 /* Notify other cpus that system-wide "drain" is running */
2177 for_each_online_cpu(cpu) {
2178 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2179 struct mem_cgroup *memcg;
2181 memcg = stock->cached;
2182 if (!memcg || !stock->nr_pages)
2184 if (!mem_cgroup_is_descendant(memcg, root_memcg))
2186 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2188 drain_local_stock(&stock->work);
2190 schedule_work_on(cpu, &stock->work);
2195 mutex_unlock(&percpu_charge_mutex);
2199 * This function drains percpu counter value from DEAD cpu and
2200 * move it to local cpu. Note that this function can be preempted.
2202 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2206 spin_lock(&memcg->pcp_counter_lock);
2207 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2208 long x = per_cpu(memcg->stat->count[i], cpu);
2210 per_cpu(memcg->stat->count[i], cpu) = 0;
2211 memcg->nocpu_base.count[i] += x;
2213 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2214 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2216 per_cpu(memcg->stat->events[i], cpu) = 0;
2217 memcg->nocpu_base.events[i] += x;
2219 spin_unlock(&memcg->pcp_counter_lock);
2222 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2223 unsigned long action,
2226 int cpu = (unsigned long)hcpu;
2227 struct memcg_stock_pcp *stock;
2228 struct mem_cgroup *iter;
2230 if (action == CPU_ONLINE)
2233 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2236 for_each_mem_cgroup(iter)
2237 mem_cgroup_drain_pcp_counter(iter, cpu);
2239 stock = &per_cpu(memcg_stock, cpu);
2244 static int try_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2245 unsigned int nr_pages)
2247 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2248 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2249 struct mem_cgroup *mem_over_limit;
2250 struct page_counter *counter;
2251 unsigned long nr_reclaimed;
2252 bool may_swap = true;
2253 bool drained = false;
2256 if (mem_cgroup_is_root(memcg))
2259 if (consume_stock(memcg, nr_pages))
2262 if (!do_swap_account ||
2263 !page_counter_try_charge(&memcg->memsw, batch, &counter)) {
2264 if (!page_counter_try_charge(&memcg->memory, batch, &counter))
2266 if (do_swap_account)
2267 page_counter_uncharge(&memcg->memsw, batch);
2268 mem_over_limit = mem_cgroup_from_counter(counter, memory);
2270 mem_over_limit = mem_cgroup_from_counter(counter, memsw);
2274 if (batch > nr_pages) {
2280 * Unlike in global OOM situations, memcg is not in a physical
2281 * memory shortage. Allow dying and OOM-killed tasks to
2282 * bypass the last charges so that they can exit quickly and
2283 * free their memory.
2285 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2286 fatal_signal_pending(current) ||
2287 current->flags & PF_EXITING))
2290 if (unlikely(task_in_memcg_oom(current)))
2293 if (!(gfp_mask & __GFP_WAIT))
2296 nr_reclaimed = try_to_free_mem_cgroup_pages(mem_over_limit, nr_pages,
2297 gfp_mask, may_swap);
2299 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2303 drain_all_stock(mem_over_limit);
2308 if (gfp_mask & __GFP_NORETRY)
2311 * Even though the limit is exceeded at this point, reclaim
2312 * may have been able to free some pages. Retry the charge
2313 * before killing the task.
2315 * Only for regular pages, though: huge pages are rather
2316 * unlikely to succeed so close to the limit, and we fall back
2317 * to regular pages anyway in case of failure.
2319 if (nr_reclaimed && nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2322 * At task move, charge accounts can be doubly counted. So, it's
2323 * better to wait until the end of task_move if something is going on.
2325 if (mem_cgroup_wait_acct_move(mem_over_limit))
2331 if (gfp_mask & __GFP_NOFAIL)
2334 if (fatal_signal_pending(current))
2337 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(nr_pages));
2339 if (!(gfp_mask & __GFP_NOFAIL))
2345 css_get_many(&memcg->css, batch);
2346 if (batch > nr_pages)
2347 refill_stock(memcg, batch - nr_pages);
2352 static void cancel_charge(struct mem_cgroup *memcg, unsigned int nr_pages)
2354 if (mem_cgroup_is_root(memcg))
2357 page_counter_uncharge(&memcg->memory, nr_pages);
2358 if (do_swap_account)
2359 page_counter_uncharge(&memcg->memsw, nr_pages);
2361 css_put_many(&memcg->css, nr_pages);
2365 * A helper function to get mem_cgroup from ID. must be called under
2366 * rcu_read_lock(). The caller is responsible for calling
2367 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2368 * refcnt from swap can be called against removed memcg.)
2370 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2372 /* ID 0 is unused ID */
2375 return mem_cgroup_from_id(id);
2379 * try_get_mem_cgroup_from_page - look up page's memcg association
2382 * Look up, get a css reference, and return the memcg that owns @page.
2384 * The page must be locked to prevent racing with swap-in and page
2385 * cache charges. If coming from an unlocked page table, the caller
2386 * must ensure the page is on the LRU or this can race with charging.
2388 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2390 struct mem_cgroup *memcg;
2394 VM_BUG_ON_PAGE(!PageLocked(page), page);
2396 memcg = page->mem_cgroup;
2398 if (!css_tryget_online(&memcg->css))
2400 } else if (PageSwapCache(page)) {
2401 ent.val = page_private(page);
2402 id = lookup_swap_cgroup_id(ent);
2404 memcg = mem_cgroup_lookup(id);
2405 if (memcg && !css_tryget_online(&memcg->css))
2412 static void lock_page_lru(struct page *page, int *isolated)
2414 struct zone *zone = page_zone(page);
2416 spin_lock_irq(&zone->lru_lock);
2417 if (PageLRU(page)) {
2418 struct lruvec *lruvec;
2420 lruvec = mem_cgroup_page_lruvec(page, zone);
2422 del_page_from_lru_list(page, lruvec, page_lru(page));
2428 static void unlock_page_lru(struct page *page, int isolated)
2430 struct zone *zone = page_zone(page);
2433 struct lruvec *lruvec;
2435 lruvec = mem_cgroup_page_lruvec(page, zone);
2436 VM_BUG_ON_PAGE(PageLRU(page), page);
2438 add_page_to_lru_list(page, lruvec, page_lru(page));
2440 spin_unlock_irq(&zone->lru_lock);
2443 static void commit_charge(struct page *page, struct mem_cgroup *memcg,
2448 VM_BUG_ON_PAGE(page->mem_cgroup, page);
2451 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2452 * may already be on some other mem_cgroup's LRU. Take care of it.
2455 lock_page_lru(page, &isolated);
2458 * Nobody should be changing or seriously looking at
2459 * page->mem_cgroup at this point:
2461 * - the page is uncharged
2463 * - the page is off-LRU
2465 * - an anonymous fault has exclusive page access, except for
2466 * a locked page table
2468 * - a page cache insertion, a swapin fault, or a migration
2469 * have the page locked
2471 page->mem_cgroup = memcg;
2474 unlock_page_lru(page, isolated);
2477 #ifdef CONFIG_MEMCG_KMEM
2479 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2480 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2482 static DEFINE_MUTEX(memcg_slab_mutex);
2485 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2486 * in the memcg_cache_params struct.
2488 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2490 struct kmem_cache *cachep;
2492 VM_BUG_ON(p->is_root_cache);
2493 cachep = p->root_cache;
2494 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2497 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp,
2498 unsigned long nr_pages)
2500 struct page_counter *counter;
2503 ret = page_counter_try_charge(&memcg->kmem, nr_pages, &counter);
2507 ret = try_charge(memcg, gfp, nr_pages);
2508 if (ret == -EINTR) {
2510 * try_charge() chose to bypass to root due to OOM kill or
2511 * fatal signal. Since our only options are to either fail
2512 * the allocation or charge it to this cgroup, do it as a
2513 * temporary condition. But we can't fail. From a kmem/slab
2514 * perspective, the cache has already been selected, by
2515 * mem_cgroup_kmem_get_cache(), so it is too late to change
2518 * This condition will only trigger if the task entered
2519 * memcg_charge_kmem in a sane state, but was OOM-killed
2520 * during try_charge() above. Tasks that were already dying
2521 * when the allocation triggers should have been already
2522 * directed to the root cgroup in memcontrol.h
2524 page_counter_charge(&memcg->memory, nr_pages);
2525 if (do_swap_account)
2526 page_counter_charge(&memcg->memsw, nr_pages);
2527 css_get_many(&memcg->css, nr_pages);
2530 page_counter_uncharge(&memcg->kmem, nr_pages);
2535 static void memcg_uncharge_kmem(struct mem_cgroup *memcg,
2536 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 css_get(&memcg->css);
2638 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2641 * Since readers won't lock (see cache_from_memcg_idx()), we need a
2642 * barrier here to ensure nobody will see the kmem_cache partially
2647 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
2648 root_cache->memcg_params->memcg_caches[id] = cachep;
2651 static void memcg_unregister_cache(struct kmem_cache *cachep)
2653 struct kmem_cache *root_cache;
2654 struct mem_cgroup *memcg;
2657 lockdep_assert_held(&memcg_slab_mutex);
2659 BUG_ON(is_root_cache(cachep));
2661 root_cache = cachep->memcg_params->root_cache;
2662 memcg = cachep->memcg_params->memcg;
2663 id = memcg_cache_id(memcg);
2665 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
2666 root_cache->memcg_params->memcg_caches[id] = NULL;
2668 list_del(&cachep->memcg_params->list);
2670 kmem_cache_destroy(cachep);
2672 /* drop the reference taken in memcg_register_cache */
2673 css_put(&memcg->css);
2677 * During the creation a new cache, we need to disable our accounting mechanism
2678 * altogether. This is true even if we are not creating, but rather just
2679 * enqueing new caches to be created.
2681 * This is because that process will trigger allocations; some visible, like
2682 * explicit kmallocs to auxiliary data structures, name strings and internal
2683 * cache structures; some well concealed, like INIT_WORK() that can allocate
2684 * objects during debug.
2686 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
2687 * to it. This may not be a bounded recursion: since the first cache creation
2688 * failed to complete (waiting on the allocation), we'll just try to create the
2689 * cache again, failing at the same point.
2691 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
2692 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
2693 * inside the following two functions.
2695 static inline void memcg_stop_kmem_account(void)
2697 VM_BUG_ON(!current->mm);
2698 current->memcg_kmem_skip_account++;
2701 static inline void memcg_resume_kmem_account(void)
2703 VM_BUG_ON(!current->mm);
2704 current->memcg_kmem_skip_account--;
2707 int __memcg_cleanup_cache_params(struct kmem_cache *s)
2709 struct kmem_cache *c;
2712 mutex_lock(&memcg_slab_mutex);
2713 for_each_memcg_cache_index(i) {
2714 c = cache_from_memcg_idx(s, i);
2718 memcg_unregister_cache(c);
2720 if (cache_from_memcg_idx(s, i))
2723 mutex_unlock(&memcg_slab_mutex);
2727 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
2729 struct kmem_cache *cachep;
2730 struct memcg_cache_params *params, *tmp;
2732 if (!memcg_kmem_is_active(memcg))
2735 mutex_lock(&memcg_slab_mutex);
2736 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
2737 cachep = memcg_params_to_cache(params);
2738 kmem_cache_shrink(cachep);
2739 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
2740 memcg_unregister_cache(cachep);
2742 mutex_unlock(&memcg_slab_mutex);
2745 struct memcg_register_cache_work {
2746 struct mem_cgroup *memcg;
2747 struct kmem_cache *cachep;
2748 struct work_struct work;
2751 static void memcg_register_cache_func(struct work_struct *w)
2753 struct memcg_register_cache_work *cw =
2754 container_of(w, struct memcg_register_cache_work, work);
2755 struct mem_cgroup *memcg = cw->memcg;
2756 struct kmem_cache *cachep = cw->cachep;
2758 mutex_lock(&memcg_slab_mutex);
2759 memcg_register_cache(memcg, cachep);
2760 mutex_unlock(&memcg_slab_mutex);
2762 css_put(&memcg->css);
2767 * Enqueue the creation of a per-memcg kmem_cache.
2769 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
2770 struct kmem_cache *cachep)
2772 struct memcg_register_cache_work *cw;
2774 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
2776 css_put(&memcg->css);
2781 cw->cachep = cachep;
2783 INIT_WORK(&cw->work, memcg_register_cache_func);
2784 schedule_work(&cw->work);
2787 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
2788 struct kmem_cache *cachep)
2791 * We need to stop accounting when we kmalloc, because if the
2792 * corresponding kmalloc cache is not yet created, the first allocation
2793 * in __memcg_schedule_register_cache will recurse.
2795 * However, it is better to enclose the whole function. Depending on
2796 * the debugging options enabled, INIT_WORK(), for instance, can
2797 * trigger an allocation. This too, will make us recurse. Because at
2798 * this point we can't allow ourselves back into memcg_kmem_get_cache,
2799 * the safest choice is to do it like this, wrapping the whole function.
2801 memcg_stop_kmem_account();
2802 __memcg_schedule_register_cache(memcg, cachep);
2803 memcg_resume_kmem_account();
2806 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
2808 unsigned int nr_pages = 1 << order;
2811 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp, nr_pages);
2813 atomic_add(nr_pages, &cachep->memcg_params->nr_pages);
2817 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
2819 unsigned int nr_pages = 1 << order;
2821 memcg_uncharge_kmem(cachep->memcg_params->memcg, nr_pages);
2822 atomic_sub(nr_pages, &cachep->memcg_params->nr_pages);
2826 * Return the kmem_cache we're supposed to use for a slab allocation.
2827 * We try to use the current memcg's version of the cache.
2829 * If the cache does not exist yet, if we are the first user of it,
2830 * we either create it immediately, if possible, or create it asynchronously
2832 * In the latter case, we will let the current allocation go through with
2833 * the original cache.
2835 * Can't be called in interrupt context or from kernel threads.
2836 * This function needs to be called with rcu_read_lock() held.
2838 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
2841 struct mem_cgroup *memcg;
2842 struct kmem_cache *memcg_cachep;
2844 VM_BUG_ON(!cachep->memcg_params);
2845 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
2847 if (current->memcg_kmem_skip_account)
2851 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
2853 if (!memcg_kmem_is_active(memcg))
2856 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
2857 if (likely(memcg_cachep)) {
2858 cachep = memcg_cachep;
2862 /* The corresponding put will be done in the workqueue. */
2863 if (!css_tryget_online(&memcg->css))
2868 * If we are in a safe context (can wait, and not in interrupt
2869 * context), we could be be predictable and return right away.
2870 * This would guarantee that the allocation being performed
2871 * already belongs in the new cache.
2873 * However, there are some clashes that can arrive from locking.
2874 * For instance, because we acquire the slab_mutex while doing
2875 * memcg_create_kmem_cache, this means no further allocation
2876 * could happen with the slab_mutex held. So it's better to
2879 memcg_schedule_register_cache(memcg, cachep);
2887 * We need to verify if the allocation against current->mm->owner's memcg is
2888 * possible for the given order. But the page is not allocated yet, so we'll
2889 * need a further commit step to do the final arrangements.
2891 * It is possible for the task to switch cgroups in this mean time, so at
2892 * commit time, we can't rely on task conversion any longer. We'll then use
2893 * the handle argument to return to the caller which cgroup we should commit
2894 * against. We could also return the memcg directly and avoid the pointer
2895 * passing, but a boolean return value gives better semantics considering
2896 * the compiled-out case as well.
2898 * Returning true means the allocation is possible.
2901 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2903 struct mem_cgroup *memcg;
2908 memcg = get_mem_cgroup_from_mm(current->mm);
2910 if (!memcg_kmem_is_active(memcg)) {
2911 css_put(&memcg->css);
2915 ret = memcg_charge_kmem(memcg, gfp, 1 << order);
2919 css_put(&memcg->css);
2923 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
2926 VM_BUG_ON(mem_cgroup_is_root(memcg));
2928 /* The page allocation failed. Revert */
2930 memcg_uncharge_kmem(memcg, 1 << order);
2933 page->mem_cgroup = memcg;
2936 void __memcg_kmem_uncharge_pages(struct page *page, int order)
2938 struct mem_cgroup *memcg = page->mem_cgroup;
2943 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
2945 memcg_uncharge_kmem(memcg, 1 << order);
2946 page->mem_cgroup = NULL;
2949 static inline void memcg_unregister_all_caches(struct mem_cgroup *memcg)
2952 #endif /* CONFIG_MEMCG_KMEM */
2954 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2957 * Because tail pages are not marked as "used", set it. We're under
2958 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2959 * charge/uncharge will be never happen and move_account() is done under
2960 * compound_lock(), so we don't have to take care of races.
2962 void mem_cgroup_split_huge_fixup(struct page *head)
2966 if (mem_cgroup_disabled())
2969 for (i = 1; i < HPAGE_PMD_NR; i++)
2970 head[i].mem_cgroup = head->mem_cgroup;
2972 __this_cpu_sub(head->mem_cgroup->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
2975 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2978 * mem_cgroup_move_account - move account of the page
2980 * @nr_pages: number of regular pages (>1 for huge pages)
2981 * @from: mem_cgroup which the page is moved from.
2982 * @to: mem_cgroup which the page is moved to. @from != @to.
2984 * The caller must confirm following.
2985 * - page is not on LRU (isolate_page() is useful.)
2986 * - compound_lock is held when nr_pages > 1
2988 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2991 static int mem_cgroup_move_account(struct page *page,
2992 unsigned int nr_pages,
2993 struct mem_cgroup *from,
2994 struct mem_cgroup *to)
2996 unsigned long flags;
2999 VM_BUG_ON(from == to);
3000 VM_BUG_ON_PAGE(PageLRU(page), page);
3002 * The page is isolated from LRU. So, collapse function
3003 * will not handle this page. But page splitting can happen.
3004 * Do this check under compound_page_lock(). The caller should
3008 if (nr_pages > 1 && !PageTransHuge(page))
3012 * Prevent mem_cgroup_migrate() from looking at page->mem_cgroup
3013 * of its source page while we change it: page migration takes
3014 * both pages off the LRU, but page cache replacement doesn't.
3016 if (!trylock_page(page))
3020 if (page->mem_cgroup != from)
3023 spin_lock_irqsave(&from->move_lock, flags);
3025 if (!PageAnon(page) && page_mapped(page)) {
3026 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3028 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3032 if (PageWriteback(page)) {
3033 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3035 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3040 * It is safe to change page->mem_cgroup here because the page
3041 * is referenced, charged, and isolated - we can't race with
3042 * uncharging, charging, migration, or LRU putback.
3045 /* caller should have done css_get */
3046 page->mem_cgroup = to;
3047 spin_unlock_irqrestore(&from->move_lock, flags);
3051 local_irq_disable();
3052 mem_cgroup_charge_statistics(to, page, nr_pages);
3053 memcg_check_events(to, page);
3054 mem_cgroup_charge_statistics(from, page, -nr_pages);
3055 memcg_check_events(from, page);
3063 #ifdef CONFIG_MEMCG_SWAP
3064 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
3067 int val = (charge) ? 1 : -1;
3068 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
3072 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3073 * @entry: swap entry to be moved
3074 * @from: mem_cgroup which the entry is moved from
3075 * @to: mem_cgroup which the entry is moved to
3077 * It succeeds only when the swap_cgroup's record for this entry is the same
3078 * as the mem_cgroup's id of @from.
3080 * Returns 0 on success, -EINVAL on failure.
3082 * The caller must have charged to @to, IOW, called page_counter_charge() about
3083 * both res and memsw, and called css_get().
3085 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3086 struct mem_cgroup *from, struct mem_cgroup *to)
3088 unsigned short old_id, new_id;
3090 old_id = mem_cgroup_id(from);
3091 new_id = mem_cgroup_id(to);
3093 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3094 mem_cgroup_swap_statistics(from, false);
3095 mem_cgroup_swap_statistics(to, true);
3097 * This function is only called from task migration context now.
3098 * It postpones page_counter and refcount handling till the end
3099 * of task migration(mem_cgroup_clear_mc()) for performance
3100 * improvement. But we cannot postpone css_get(to) because if
3101 * the process that has been moved to @to does swap-in, the
3102 * refcount of @to might be decreased to 0.
3104 * We are in attach() phase, so the cgroup is guaranteed to be
3105 * alive, so we can just call css_get().
3113 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3114 struct mem_cgroup *from, struct mem_cgroup *to)
3120 static DEFINE_MUTEX(memcg_limit_mutex);
3122 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3123 unsigned long limit)
3125 unsigned long curusage;
3126 unsigned long oldusage;
3127 bool enlarge = false;
3132 * For keeping hierarchical_reclaim simple, how long we should retry
3133 * is depends on callers. We set our retry-count to be function
3134 * of # of children which we should visit in this loop.
3136 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3137 mem_cgroup_count_children(memcg);
3139 oldusage = page_counter_read(&memcg->memory);
3142 if (signal_pending(current)) {
3147 mutex_lock(&memcg_limit_mutex);
3148 if (limit > memcg->memsw.limit) {
3149 mutex_unlock(&memcg_limit_mutex);
3153 if (limit > memcg->memory.limit)
3155 ret = page_counter_limit(&memcg->memory, limit);
3156 mutex_unlock(&memcg_limit_mutex);
3161 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, true);
3163 curusage = page_counter_read(&memcg->memory);
3164 /* Usage is reduced ? */
3165 if (curusage >= oldusage)
3168 oldusage = curusage;
3169 } while (retry_count);
3171 if (!ret && enlarge)
3172 memcg_oom_recover(memcg);
3177 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3178 unsigned long limit)
3180 unsigned long curusage;
3181 unsigned long oldusage;
3182 bool enlarge = false;
3186 /* see mem_cgroup_resize_res_limit */
3187 retry_count = MEM_CGROUP_RECLAIM_RETRIES *
3188 mem_cgroup_count_children(memcg);
3190 oldusage = page_counter_read(&memcg->memsw);
3193 if (signal_pending(current)) {
3198 mutex_lock(&memcg_limit_mutex);
3199 if (limit < memcg->memory.limit) {
3200 mutex_unlock(&memcg_limit_mutex);
3204 if (limit > memcg->memsw.limit)
3206 ret = page_counter_limit(&memcg->memsw, limit);
3207 mutex_unlock(&memcg_limit_mutex);
3212 try_to_free_mem_cgroup_pages(memcg, 1, GFP_KERNEL, false);
3214 curusage = page_counter_read(&memcg->memsw);
3215 /* Usage is reduced ? */
3216 if (curusage >= oldusage)
3219 oldusage = curusage;
3220 } while (retry_count);
3222 if (!ret && enlarge)
3223 memcg_oom_recover(memcg);
3228 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3230 unsigned long *total_scanned)
3232 unsigned long nr_reclaimed = 0;
3233 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3234 unsigned long reclaimed;
3236 struct mem_cgroup_tree_per_zone *mctz;
3237 unsigned long excess;
3238 unsigned long nr_scanned;
3243 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3245 * This loop can run a while, specially if mem_cgroup's continuously
3246 * keep exceeding their soft limit and putting the system under
3253 mz = mem_cgroup_largest_soft_limit_node(mctz);
3258 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3259 gfp_mask, &nr_scanned);
3260 nr_reclaimed += reclaimed;
3261 *total_scanned += nr_scanned;
3262 spin_lock_irq(&mctz->lock);
3263 __mem_cgroup_remove_exceeded(mz, mctz);
3266 * If we failed to reclaim anything from this memory cgroup
3267 * it is time to move on to the next cgroup
3271 next_mz = __mem_cgroup_largest_soft_limit_node(mctz);
3273 excess = soft_limit_excess(mz->memcg);
3275 * One school of thought says that we should not add
3276 * back the node to the tree if reclaim returns 0.
3277 * But our reclaim could return 0, simply because due
3278 * to priority we are exposing a smaller subset of
3279 * memory to reclaim from. Consider this as a longer
3282 /* If excess == 0, no tree ops */
3283 __mem_cgroup_insert_exceeded(mz, mctz, excess);
3284 spin_unlock_irq(&mctz->lock);
3285 css_put(&mz->memcg->css);
3288 * Could not reclaim anything and there are no more
3289 * mem cgroups to try or we seem to be looping without
3290 * reclaiming anything.
3292 if (!nr_reclaimed &&
3294 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3296 } while (!nr_reclaimed);
3298 css_put(&next_mz->memcg->css);
3299 return nr_reclaimed;
3303 * Test whether @memcg has children, dead or alive. Note that this
3304 * function doesn't care whether @memcg has use_hierarchy enabled and
3305 * returns %true if there are child csses according to the cgroup
3306 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
3308 static inline bool memcg_has_children(struct mem_cgroup *memcg)
3313 * The lock does not prevent addition or deletion of children, but
3314 * it prevents a new child from being initialized based on this
3315 * parent in css_online(), so it's enough to decide whether
3316 * hierarchically inherited attributes can still be changed or not.
3318 lockdep_assert_held(&memcg_create_mutex);
3321 ret = css_next_child(NULL, &memcg->css);
3327 * Reclaims as many pages from the given memcg as possible and moves
3328 * the rest to the parent.
3330 * Caller is responsible for holding css reference for memcg.
3332 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3334 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3336 /* we call try-to-free pages for make this cgroup empty */
3337 lru_add_drain_all();
3338 /* try to free all pages in this cgroup */
3339 while (nr_retries && page_counter_read(&memcg->memory)) {
3342 if (signal_pending(current))
3345 progress = try_to_free_mem_cgroup_pages(memcg, 1,
3349 /* maybe some writeback is necessary */
3350 congestion_wait(BLK_RW_ASYNC, HZ/10);
3358 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
3359 char *buf, size_t nbytes,
3362 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3364 if (mem_cgroup_is_root(memcg))
3366 return mem_cgroup_force_empty(memcg) ?: nbytes;
3369 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
3372 return mem_cgroup_from_css(css)->use_hierarchy;
3375 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
3376 struct cftype *cft, u64 val)
3379 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3380 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
3382 mutex_lock(&memcg_create_mutex);
3384 if (memcg->use_hierarchy == val)
3388 * If parent's use_hierarchy is set, we can't make any modifications
3389 * in the child subtrees. If it is unset, then the change can
3390 * occur, provided the current cgroup has no children.
3392 * For the root cgroup, parent_mem is NULL, we allow value to be
3393 * set if there are no children.
3395 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3396 (val == 1 || val == 0)) {
3397 if (!memcg_has_children(memcg))
3398 memcg->use_hierarchy = val;
3405 mutex_unlock(&memcg_create_mutex);
3410 static unsigned long tree_stat(struct mem_cgroup *memcg,
3411 enum mem_cgroup_stat_index idx)
3413 struct mem_cgroup *iter;
3416 /* Per-cpu values can be negative, use a signed accumulator */
3417 for_each_mem_cgroup_tree(iter, memcg)
3418 val += mem_cgroup_read_stat(iter, idx);
3420 if (val < 0) /* race ? */
3425 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3429 if (mem_cgroup_is_root(memcg)) {
3430 val = tree_stat(memcg, MEM_CGROUP_STAT_CACHE);
3431 val += tree_stat(memcg, MEM_CGROUP_STAT_RSS);
3433 val += tree_stat(memcg, MEM_CGROUP_STAT_SWAP);
3436 val = page_counter_read(&memcg->memory);
3438 val = page_counter_read(&memcg->memsw);
3440 return val << PAGE_SHIFT;
3451 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
3454 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3455 struct page_counter *counter;
3457 switch (MEMFILE_TYPE(cft->private)) {
3459 counter = &memcg->memory;
3462 counter = &memcg->memsw;
3465 counter = &memcg->kmem;
3471 switch (MEMFILE_ATTR(cft->private)) {
3473 if (counter == &memcg->memory)
3474 return mem_cgroup_usage(memcg, false);
3475 if (counter == &memcg->memsw)
3476 return mem_cgroup_usage(memcg, true);
3477 return (u64)page_counter_read(counter) * PAGE_SIZE;
3479 return (u64)counter->limit * PAGE_SIZE;
3481 return (u64)counter->watermark * PAGE_SIZE;
3483 return counter->failcnt;
3484 case RES_SOFT_LIMIT:
3485 return (u64)memcg->soft_limit * PAGE_SIZE;
3491 #ifdef CONFIG_MEMCG_KMEM
3492 static int memcg_activate_kmem(struct mem_cgroup *memcg,
3493 unsigned long nr_pages)
3498 if (memcg_kmem_is_active(memcg))
3502 * For simplicity, we won't allow this to be disabled. It also can't
3503 * be changed if the cgroup has children already, or if tasks had
3506 * If tasks join before we set the limit, a person looking at
3507 * kmem.usage_in_bytes will have no way to determine when it took
3508 * place, which makes the value quite meaningless.
3510 * After it first became limited, changes in the value of the limit are
3511 * of course permitted.
3513 mutex_lock(&memcg_create_mutex);
3514 if (cgroup_has_tasks(memcg->css.cgroup) ||
3515 (memcg->use_hierarchy && memcg_has_children(memcg)))
3517 mutex_unlock(&memcg_create_mutex);
3521 memcg_id = memcg_alloc_cache_id();
3528 * We couldn't have accounted to this cgroup, because it hasn't got
3529 * activated yet, so this should succeed.
3531 err = page_counter_limit(&memcg->kmem, nr_pages);
3534 static_key_slow_inc(&memcg_kmem_enabled_key);
3536 * A memory cgroup is considered kmem-active as soon as it gets
3537 * kmemcg_id. Setting the id after enabling static branching will
3538 * guarantee no one starts accounting before all call sites are
3541 memcg->kmemcg_id = memcg_id;
3546 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3547 unsigned long limit)
3551 mutex_lock(&memcg_limit_mutex);
3552 if (!memcg_kmem_is_active(memcg))
3553 ret = memcg_activate_kmem(memcg, limit);
3555 ret = page_counter_limit(&memcg->kmem, limit);
3556 mutex_unlock(&memcg_limit_mutex);
3560 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
3563 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
3568 mutex_lock(&memcg_limit_mutex);
3570 * If the parent cgroup is not kmem-active now, it cannot be activated
3571 * after this point, because it has at least one child already.
3573 if (memcg_kmem_is_active(parent))
3574 ret = memcg_activate_kmem(memcg, PAGE_COUNTER_MAX);
3575 mutex_unlock(&memcg_limit_mutex);
3579 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
3580 unsigned long limit)
3584 #endif /* CONFIG_MEMCG_KMEM */
3587 * The user of this function is...
3590 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
3591 char *buf, size_t nbytes, loff_t off)
3593 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3594 unsigned long nr_pages;
3597 buf = strstrip(buf);
3598 ret = page_counter_memparse(buf, &nr_pages);
3602 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3604 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3608 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3610 ret = mem_cgroup_resize_limit(memcg, nr_pages);
3613 ret = mem_cgroup_resize_memsw_limit(memcg, nr_pages);
3616 ret = memcg_update_kmem_limit(memcg, nr_pages);
3620 case RES_SOFT_LIMIT:
3621 memcg->soft_limit = nr_pages;
3625 return ret ?: nbytes;
3628 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
3629 size_t nbytes, loff_t off)
3631 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
3632 struct page_counter *counter;
3634 switch (MEMFILE_TYPE(of_cft(of)->private)) {
3636 counter = &memcg->memory;
3639 counter = &memcg->memsw;
3642 counter = &memcg->kmem;
3648 switch (MEMFILE_ATTR(of_cft(of)->private)) {
3650 page_counter_reset_watermark(counter);
3653 counter->failcnt = 0;
3662 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
3665 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
3669 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3670 struct cftype *cft, u64 val)
3672 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3674 if (val >= (1 << NR_MOVE_TYPE))
3678 * No kind of locking is needed in here, because ->can_attach() will
3679 * check this value once in the beginning of the process, and then carry
3680 * on with stale data. This means that changes to this value will only
3681 * affect task migrations starting after the change.
3683 memcg->move_charge_at_immigrate = val;
3687 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
3688 struct cftype *cft, u64 val)
3695 static int memcg_numa_stat_show(struct seq_file *m, void *v)
3699 unsigned int lru_mask;
3702 static const struct numa_stat stats[] = {
3703 { "total", LRU_ALL },
3704 { "file", LRU_ALL_FILE },
3705 { "anon", LRU_ALL_ANON },
3706 { "unevictable", BIT(LRU_UNEVICTABLE) },
3708 const struct numa_stat *stat;
3711 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3713 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3714 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
3715 seq_printf(m, "%s=%lu", stat->name, nr);
3716 for_each_node_state(nid, N_MEMORY) {
3717 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
3719 seq_printf(m, " N%d=%lu", nid, nr);
3724 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
3725 struct mem_cgroup *iter;
3728 for_each_mem_cgroup_tree(iter, memcg)
3729 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
3730 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
3731 for_each_node_state(nid, N_MEMORY) {
3733 for_each_mem_cgroup_tree(iter, memcg)
3734 nr += mem_cgroup_node_nr_lru_pages(
3735 iter, nid, stat->lru_mask);
3736 seq_printf(m, " N%d=%lu", nid, nr);
3743 #endif /* CONFIG_NUMA */
3745 static inline void mem_cgroup_lru_names_not_uptodate(void)
3747 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
3750 static int memcg_stat_show(struct seq_file *m, void *v)
3752 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
3753 unsigned long memory, memsw;
3754 struct mem_cgroup *mi;
3757 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3758 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3760 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
3761 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
3764 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
3765 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
3766 mem_cgroup_read_events(memcg, i));
3768 for (i = 0; i < NR_LRU_LISTS; i++)
3769 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
3770 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
3772 /* Hierarchical information */
3773 memory = memsw = PAGE_COUNTER_MAX;
3774 for (mi = memcg; mi; mi = parent_mem_cgroup(mi)) {
3775 memory = min(memory, mi->memory.limit);
3776 memsw = min(memsw, mi->memsw.limit);
3778 seq_printf(m, "hierarchical_memory_limit %llu\n",
3779 (u64)memory * PAGE_SIZE);
3780 if (do_swap_account)
3781 seq_printf(m, "hierarchical_memsw_limit %llu\n",
3782 (u64)memsw * PAGE_SIZE);
3784 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
3787 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
3789 for_each_mem_cgroup_tree(mi, memcg)
3790 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
3791 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
3794 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
3795 unsigned long long val = 0;
3797 for_each_mem_cgroup_tree(mi, memcg)
3798 val += mem_cgroup_read_events(mi, i);
3799 seq_printf(m, "total_%s %llu\n",
3800 mem_cgroup_events_names[i], val);
3803 for (i = 0; i < NR_LRU_LISTS; i++) {
3804 unsigned long long val = 0;
3806 for_each_mem_cgroup_tree(mi, memcg)
3807 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
3808 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
3811 #ifdef CONFIG_DEBUG_VM
3814 struct mem_cgroup_per_zone *mz;
3815 struct zone_reclaim_stat *rstat;
3816 unsigned long recent_rotated[2] = {0, 0};
3817 unsigned long recent_scanned[2] = {0, 0};
3819 for_each_online_node(nid)
3820 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3821 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
3822 rstat = &mz->lruvec.reclaim_stat;
3824 recent_rotated[0] += rstat->recent_rotated[0];
3825 recent_rotated[1] += rstat->recent_rotated[1];
3826 recent_scanned[0] += rstat->recent_scanned[0];
3827 recent_scanned[1] += rstat->recent_scanned[1];
3829 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
3830 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
3831 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
3832 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
3839 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
3842 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3844 return mem_cgroup_swappiness(memcg);
3847 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
3848 struct cftype *cft, u64 val)
3850 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3856 memcg->swappiness = val;
3858 vm_swappiness = val;
3863 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3865 struct mem_cgroup_threshold_ary *t;
3866 unsigned long usage;
3871 t = rcu_dereference(memcg->thresholds.primary);
3873 t = rcu_dereference(memcg->memsw_thresholds.primary);
3878 usage = mem_cgroup_usage(memcg, swap);
3881 * current_threshold points to threshold just below or equal to usage.
3882 * If it's not true, a threshold was crossed after last
3883 * call of __mem_cgroup_threshold().
3885 i = t->current_threshold;
3888 * Iterate backward over array of thresholds starting from
3889 * current_threshold and check if a threshold is crossed.
3890 * If none of thresholds below usage is crossed, we read
3891 * only one element of the array here.
3893 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3894 eventfd_signal(t->entries[i].eventfd, 1);
3896 /* i = current_threshold + 1 */
3900 * Iterate forward over array of thresholds starting from
3901 * current_threshold+1 and check if a threshold is crossed.
3902 * If none of thresholds above usage is crossed, we read
3903 * only one element of the array here.
3905 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3906 eventfd_signal(t->entries[i].eventfd, 1);
3908 /* Update current_threshold */
3909 t->current_threshold = i - 1;
3914 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3917 __mem_cgroup_threshold(memcg, false);
3918 if (do_swap_account)
3919 __mem_cgroup_threshold(memcg, true);
3921 memcg = parent_mem_cgroup(memcg);
3925 static int compare_thresholds(const void *a, const void *b)
3927 const struct mem_cgroup_threshold *_a = a;
3928 const struct mem_cgroup_threshold *_b = b;
3930 if (_a->threshold > _b->threshold)
3933 if (_a->threshold < _b->threshold)
3939 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
3941 struct mem_cgroup_eventfd_list *ev;
3943 spin_lock(&memcg_oom_lock);
3945 list_for_each_entry(ev, &memcg->oom_notify, list)
3946 eventfd_signal(ev->eventfd, 1);
3948 spin_unlock(&memcg_oom_lock);
3952 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
3954 struct mem_cgroup *iter;
3956 for_each_mem_cgroup_tree(iter, memcg)
3957 mem_cgroup_oom_notify_cb(iter);
3960 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
3961 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
3963 struct mem_cgroup_thresholds *thresholds;
3964 struct mem_cgroup_threshold_ary *new;
3965 unsigned long threshold;
3966 unsigned long usage;
3969 ret = page_counter_memparse(args, &threshold);
3973 mutex_lock(&memcg->thresholds_lock);
3976 thresholds = &memcg->thresholds;
3977 usage = mem_cgroup_usage(memcg, false);
3978 } else if (type == _MEMSWAP) {
3979 thresholds = &memcg->memsw_thresholds;
3980 usage = mem_cgroup_usage(memcg, true);
3984 /* Check if a threshold crossed before adding a new one */
3985 if (thresholds->primary)
3986 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3988 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3990 /* Allocate memory for new array of thresholds */
3991 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3999 /* Copy thresholds (if any) to new array */
4000 if (thresholds->primary) {
4001 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4002 sizeof(struct mem_cgroup_threshold));
4005 /* Add new threshold */
4006 new->entries[size - 1].eventfd = eventfd;
4007 new->entries[size - 1].threshold = threshold;
4009 /* Sort thresholds. Registering of new threshold isn't time-critical */
4010 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4011 compare_thresholds, NULL);
4013 /* Find current threshold */
4014 new->current_threshold = -1;
4015 for (i = 0; i < size; i++) {
4016 if (new->entries[i].threshold <= usage) {
4018 * new->current_threshold will not be used until
4019 * rcu_assign_pointer(), so it's safe to increment
4022 ++new->current_threshold;
4027 /* Free old spare buffer and save old primary buffer as spare */
4028 kfree(thresholds->spare);
4029 thresholds->spare = thresholds->primary;
4031 rcu_assign_pointer(thresholds->primary, new);
4033 /* To be sure that nobody uses thresholds */
4037 mutex_unlock(&memcg->thresholds_lock);
4042 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
4043 struct eventfd_ctx *eventfd, const char *args)
4045 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
4048 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
4049 struct eventfd_ctx *eventfd, const char *args)
4051 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
4054 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4055 struct eventfd_ctx *eventfd, enum res_type type)
4057 struct mem_cgroup_thresholds *thresholds;
4058 struct mem_cgroup_threshold_ary *new;
4059 unsigned long usage;
4062 mutex_lock(&memcg->thresholds_lock);
4065 thresholds = &memcg->thresholds;
4066 usage = mem_cgroup_usage(memcg, false);
4067 } else if (type == _MEMSWAP) {
4068 thresholds = &memcg->memsw_thresholds;
4069 usage = mem_cgroup_usage(memcg, true);
4073 if (!thresholds->primary)
4076 /* Check if a threshold crossed before removing */
4077 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4079 /* Calculate new number of threshold */
4081 for (i = 0; i < thresholds->primary->size; i++) {
4082 if (thresholds->primary->entries[i].eventfd != eventfd)
4086 new = thresholds->spare;
4088 /* Set thresholds array to NULL if we don't have thresholds */
4097 /* Copy thresholds and find current threshold */
4098 new->current_threshold = -1;
4099 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4100 if (thresholds->primary->entries[i].eventfd == eventfd)
4103 new->entries[j] = thresholds->primary->entries[i];
4104 if (new->entries[j].threshold <= usage) {
4106 * new->current_threshold will not be used
4107 * until rcu_assign_pointer(), so it's safe to increment
4110 ++new->current_threshold;
4116 /* Swap primary and spare array */
4117 thresholds->spare = thresholds->primary;
4118 /* If all events are unregistered, free the spare array */
4120 kfree(thresholds->spare);
4121 thresholds->spare = NULL;
4124 rcu_assign_pointer(thresholds->primary, new);
4126 /* To be sure that nobody uses thresholds */
4129 mutex_unlock(&memcg->thresholds_lock);
4132 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4133 struct eventfd_ctx *eventfd)
4135 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
4138 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
4139 struct eventfd_ctx *eventfd)
4141 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
4144 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
4145 struct eventfd_ctx *eventfd, const char *args)
4147 struct mem_cgroup_eventfd_list *event;
4149 event = kmalloc(sizeof(*event), GFP_KERNEL);
4153 spin_lock(&memcg_oom_lock);
4155 event->eventfd = eventfd;
4156 list_add(&event->list, &memcg->oom_notify);
4158 /* already in OOM ? */
4159 if (atomic_read(&memcg->under_oom))
4160 eventfd_signal(eventfd, 1);
4161 spin_unlock(&memcg_oom_lock);
4166 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
4167 struct eventfd_ctx *eventfd)
4169 struct mem_cgroup_eventfd_list *ev, *tmp;
4171 spin_lock(&memcg_oom_lock);
4173 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4174 if (ev->eventfd == eventfd) {
4175 list_del(&ev->list);
4180 spin_unlock(&memcg_oom_lock);
4183 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
4185 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
4187 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
4188 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
4192 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
4193 struct cftype *cft, u64 val)
4195 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4197 /* cannot set to root cgroup and only 0 and 1 are allowed */
4198 if (!css->parent || !((val == 0) || (val == 1)))
4201 memcg->oom_kill_disable = val;
4203 memcg_oom_recover(memcg);
4208 #ifdef CONFIG_MEMCG_KMEM
4209 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4213 ret = memcg_propagate_kmem(memcg);
4217 return mem_cgroup_sockets_init(memcg, ss);
4220 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4222 mem_cgroup_sockets_destroy(memcg);
4225 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4230 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
4236 * DO NOT USE IN NEW FILES.
4238 * "cgroup.event_control" implementation.
4240 * This is way over-engineered. It tries to support fully configurable
4241 * events for each user. Such level of flexibility is completely
4242 * unnecessary especially in the light of the planned unified hierarchy.
4244 * Please deprecate this and replace with something simpler if at all
4249 * Unregister event and free resources.
4251 * Gets called from workqueue.
4253 static void memcg_event_remove(struct work_struct *work)
4255 struct mem_cgroup_event *event =
4256 container_of(work, struct mem_cgroup_event, remove);
4257 struct mem_cgroup *memcg = event->memcg;
4259 remove_wait_queue(event->wqh, &event->wait);
4261 event->unregister_event(memcg, event->eventfd);
4263 /* Notify userspace the event is going away. */
4264 eventfd_signal(event->eventfd, 1);
4266 eventfd_ctx_put(event->eventfd);
4268 css_put(&memcg->css);
4272 * Gets called on POLLHUP on eventfd when user closes it.
4274 * Called with wqh->lock held and interrupts disabled.
4276 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
4277 int sync, void *key)
4279 struct mem_cgroup_event *event =
4280 container_of(wait, struct mem_cgroup_event, wait);
4281 struct mem_cgroup *memcg = event->memcg;
4282 unsigned long flags = (unsigned long)key;
4284 if (flags & POLLHUP) {
4286 * If the event has been detached at cgroup removal, we
4287 * can simply return knowing the other side will cleanup
4290 * We can't race against event freeing since the other
4291 * side will require wqh->lock via remove_wait_queue(),
4294 spin_lock(&memcg->event_list_lock);
4295 if (!list_empty(&event->list)) {
4296 list_del_init(&event->list);
4298 * We are in atomic context, but cgroup_event_remove()
4299 * may sleep, so we have to call it in workqueue.
4301 schedule_work(&event->remove);
4303 spin_unlock(&memcg->event_list_lock);
4309 static void memcg_event_ptable_queue_proc(struct file *file,
4310 wait_queue_head_t *wqh, poll_table *pt)
4312 struct mem_cgroup_event *event =
4313 container_of(pt, struct mem_cgroup_event, pt);
4316 add_wait_queue(wqh, &event->wait);
4320 * DO NOT USE IN NEW FILES.
4322 * Parse input and register new cgroup event handler.
4324 * Input must be in format '<event_fd> <control_fd> <args>'.
4325 * Interpretation of args is defined by control file implementation.
4327 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
4328 char *buf, size_t nbytes, loff_t off)
4330 struct cgroup_subsys_state *css = of_css(of);
4331 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4332 struct mem_cgroup_event *event;
4333 struct cgroup_subsys_state *cfile_css;
4334 unsigned int efd, cfd;
4341 buf = strstrip(buf);
4343 efd = simple_strtoul(buf, &endp, 10);
4348 cfd = simple_strtoul(buf, &endp, 10);
4349 if ((*endp != ' ') && (*endp != '\0'))
4353 event = kzalloc(sizeof(*event), GFP_KERNEL);
4357 event->memcg = memcg;
4358 INIT_LIST_HEAD(&event->list);
4359 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
4360 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
4361 INIT_WORK(&event->remove, memcg_event_remove);
4369 event->eventfd = eventfd_ctx_fileget(efile.file);
4370 if (IS_ERR(event->eventfd)) {
4371 ret = PTR_ERR(event->eventfd);
4378 goto out_put_eventfd;
4381 /* the process need read permission on control file */
4382 /* AV: shouldn't we check that it's been opened for read instead? */
4383 ret = inode_permission(file_inode(cfile.file), MAY_READ);
4388 * Determine the event callbacks and set them in @event. This used
4389 * to be done via struct cftype but cgroup core no longer knows
4390 * about these events. The following is crude but the whole thing
4391 * is for compatibility anyway.
4393 * DO NOT ADD NEW FILES.
4395 name = cfile.file->f_path.dentry->d_name.name;
4397 if (!strcmp(name, "memory.usage_in_bytes")) {
4398 event->register_event = mem_cgroup_usage_register_event;
4399 event->unregister_event = mem_cgroup_usage_unregister_event;
4400 } else if (!strcmp(name, "memory.oom_control")) {
4401 event->register_event = mem_cgroup_oom_register_event;
4402 event->unregister_event = mem_cgroup_oom_unregister_event;
4403 } else if (!strcmp(name, "memory.pressure_level")) {
4404 event->register_event = vmpressure_register_event;
4405 event->unregister_event = vmpressure_unregister_event;
4406 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
4407 event->register_event = memsw_cgroup_usage_register_event;
4408 event->unregister_event = memsw_cgroup_usage_unregister_event;
4415 * Verify @cfile should belong to @css. Also, remaining events are
4416 * automatically removed on cgroup destruction but the removal is
4417 * asynchronous, so take an extra ref on @css.
4419 cfile_css = css_tryget_online_from_dir(cfile.file->f_path.dentry->d_parent,
4420 &memory_cgrp_subsys);
4422 if (IS_ERR(cfile_css))
4424 if (cfile_css != css) {
4429 ret = event->register_event(memcg, event->eventfd, buf);
4433 efile.file->f_op->poll(efile.file, &event->pt);
4435 spin_lock(&memcg->event_list_lock);
4436 list_add(&event->list, &memcg->event_list);
4437 spin_unlock(&memcg->event_list_lock);
4449 eventfd_ctx_put(event->eventfd);
4458 static struct cftype mem_cgroup_files[] = {
4460 .name = "usage_in_bytes",
4461 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4462 .read_u64 = mem_cgroup_read_u64,
4465 .name = "max_usage_in_bytes",
4466 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4467 .write = mem_cgroup_reset,
4468 .read_u64 = mem_cgroup_read_u64,
4471 .name = "limit_in_bytes",
4472 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4473 .write = mem_cgroup_write,
4474 .read_u64 = mem_cgroup_read_u64,
4477 .name = "soft_limit_in_bytes",
4478 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4479 .write = mem_cgroup_write,
4480 .read_u64 = mem_cgroup_read_u64,
4484 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4485 .write = mem_cgroup_reset,
4486 .read_u64 = mem_cgroup_read_u64,
4490 .seq_show = memcg_stat_show,
4493 .name = "force_empty",
4494 .write = mem_cgroup_force_empty_write,
4497 .name = "use_hierarchy",
4498 .write_u64 = mem_cgroup_hierarchy_write,
4499 .read_u64 = mem_cgroup_hierarchy_read,
4502 .name = "cgroup.event_control", /* XXX: for compat */
4503 .write = memcg_write_event_control,
4504 .flags = CFTYPE_NO_PREFIX,
4508 .name = "swappiness",
4509 .read_u64 = mem_cgroup_swappiness_read,
4510 .write_u64 = mem_cgroup_swappiness_write,
4513 .name = "move_charge_at_immigrate",
4514 .read_u64 = mem_cgroup_move_charge_read,
4515 .write_u64 = mem_cgroup_move_charge_write,
4518 .name = "oom_control",
4519 .seq_show = mem_cgroup_oom_control_read,
4520 .write_u64 = mem_cgroup_oom_control_write,
4521 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4524 .name = "pressure_level",
4528 .name = "numa_stat",
4529 .seq_show = memcg_numa_stat_show,
4532 #ifdef CONFIG_MEMCG_KMEM
4534 .name = "kmem.limit_in_bytes",
4535 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
4536 .write = mem_cgroup_write,
4537 .read_u64 = mem_cgroup_read_u64,
4540 .name = "kmem.usage_in_bytes",
4541 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
4542 .read_u64 = mem_cgroup_read_u64,
4545 .name = "kmem.failcnt",
4546 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
4547 .write = mem_cgroup_reset,
4548 .read_u64 = mem_cgroup_read_u64,
4551 .name = "kmem.max_usage_in_bytes",
4552 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
4553 .write = mem_cgroup_reset,
4554 .read_u64 = mem_cgroup_read_u64,
4556 #ifdef CONFIG_SLABINFO
4558 .name = "kmem.slabinfo",
4559 .seq_start = slab_start,
4560 .seq_next = slab_next,
4561 .seq_stop = slab_stop,
4562 .seq_show = memcg_slab_show,
4566 { }, /* terminate */
4569 #ifdef CONFIG_MEMCG_SWAP
4570 static struct cftype memsw_cgroup_files[] = {
4572 .name = "memsw.usage_in_bytes",
4573 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4574 .read_u64 = mem_cgroup_read_u64,
4577 .name = "memsw.max_usage_in_bytes",
4578 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4579 .write = mem_cgroup_reset,
4580 .read_u64 = mem_cgroup_read_u64,
4583 .name = "memsw.limit_in_bytes",
4584 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4585 .write = mem_cgroup_write,
4586 .read_u64 = mem_cgroup_read_u64,
4589 .name = "memsw.failcnt",
4590 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4591 .write = mem_cgroup_reset,
4592 .read_u64 = mem_cgroup_read_u64,
4594 { }, /* terminate */
4597 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4599 struct mem_cgroup_per_node *pn;
4600 struct mem_cgroup_per_zone *mz;
4601 int zone, tmp = node;
4603 * This routine is called against possible nodes.
4604 * But it's BUG to call kmalloc() against offline node.
4606 * TODO: this routine can waste much memory for nodes which will
4607 * never be onlined. It's better to use memory hotplug callback
4610 if (!node_state(node, N_NORMAL_MEMORY))
4612 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4616 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4617 mz = &pn->zoneinfo[zone];
4618 lruvec_init(&mz->lruvec);
4619 mz->usage_in_excess = 0;
4620 mz->on_tree = false;
4623 memcg->nodeinfo[node] = pn;
4627 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4629 kfree(memcg->nodeinfo[node]);
4632 static struct mem_cgroup *mem_cgroup_alloc(void)
4634 struct mem_cgroup *memcg;
4637 size = sizeof(struct mem_cgroup);
4638 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
4640 memcg = kzalloc(size, GFP_KERNEL);
4644 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4647 spin_lock_init(&memcg->pcp_counter_lock);
4656 * At destroying mem_cgroup, references from swap_cgroup can remain.
4657 * (scanning all at force_empty is too costly...)
4659 * Instead of clearing all references at force_empty, we remember
4660 * the number of reference from swap_cgroup and free mem_cgroup when
4661 * it goes down to 0.
4663 * Removal of cgroup itself succeeds regardless of refs from swap.
4666 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4670 mem_cgroup_remove_from_trees(memcg);
4673 free_mem_cgroup_per_zone_info(memcg, node);
4675 free_percpu(memcg->stat);
4677 disarm_static_keys(memcg);
4682 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4684 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4686 if (!memcg->memory.parent)
4688 return mem_cgroup_from_counter(memcg->memory.parent, memory);
4690 EXPORT_SYMBOL(parent_mem_cgroup);
4692 static void __init mem_cgroup_soft_limit_tree_init(void)
4694 struct mem_cgroup_tree_per_node *rtpn;
4695 struct mem_cgroup_tree_per_zone *rtpz;
4696 int tmp, node, zone;
4698 for_each_node(node) {
4700 if (!node_state(node, N_NORMAL_MEMORY))
4702 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4705 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4707 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4708 rtpz = &rtpn->rb_tree_per_zone[zone];
4709 rtpz->rb_root = RB_ROOT;
4710 spin_lock_init(&rtpz->lock);
4715 static struct cgroup_subsys_state * __ref
4716 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
4718 struct mem_cgroup *memcg;
4719 long error = -ENOMEM;
4722 memcg = mem_cgroup_alloc();
4724 return ERR_PTR(error);
4727 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4731 if (parent_css == NULL) {
4732 root_mem_cgroup = memcg;
4733 page_counter_init(&memcg->memory, NULL);
4734 page_counter_init(&memcg->memsw, NULL);
4735 page_counter_init(&memcg->kmem, NULL);
4738 memcg->last_scanned_node = MAX_NUMNODES;
4739 INIT_LIST_HEAD(&memcg->oom_notify);
4740 memcg->move_charge_at_immigrate = 0;
4741 mutex_init(&memcg->thresholds_lock);
4742 spin_lock_init(&memcg->move_lock);
4743 vmpressure_init(&memcg->vmpressure);
4744 INIT_LIST_HEAD(&memcg->event_list);
4745 spin_lock_init(&memcg->event_list_lock);
4746 #ifdef CONFIG_MEMCG_KMEM
4747 memcg->kmemcg_id = -1;
4748 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
4754 __mem_cgroup_free(memcg);
4755 return ERR_PTR(error);
4759 mem_cgroup_css_online(struct cgroup_subsys_state *css)
4761 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4762 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
4765 if (css->id > MEM_CGROUP_ID_MAX)
4771 mutex_lock(&memcg_create_mutex);
4773 memcg->use_hierarchy = parent->use_hierarchy;
4774 memcg->oom_kill_disable = parent->oom_kill_disable;
4775 memcg->swappiness = mem_cgroup_swappiness(parent);
4777 if (parent->use_hierarchy) {
4778 page_counter_init(&memcg->memory, &parent->memory);
4779 page_counter_init(&memcg->memsw, &parent->memsw);
4780 page_counter_init(&memcg->kmem, &parent->kmem);
4783 * No need to take a reference to the parent because cgroup
4784 * core guarantees its existence.
4787 page_counter_init(&memcg->memory, NULL);
4788 page_counter_init(&memcg->memsw, NULL);
4789 page_counter_init(&memcg->kmem, NULL);
4791 * Deeper hierachy with use_hierarchy == false doesn't make
4792 * much sense so let cgroup subsystem know about this
4793 * unfortunate state in our controller.
4795 if (parent != root_mem_cgroup)
4796 memory_cgrp_subsys.broken_hierarchy = true;
4798 mutex_unlock(&memcg_create_mutex);
4800 ret = memcg_init_kmem(memcg, &memory_cgrp_subsys);
4805 * Make sure the memcg is initialized: mem_cgroup_iter()
4806 * orders reading memcg->initialized against its callers
4807 * reading the memcg members.
4809 smp_store_release(&memcg->initialized, 1);
4814 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
4816 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4817 struct mem_cgroup_event *event, *tmp;
4820 * Unregister events and notify userspace.
4821 * Notify userspace about cgroup removing only after rmdir of cgroup
4822 * directory to avoid race between userspace and kernelspace.
4824 spin_lock(&memcg->event_list_lock);
4825 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
4826 list_del_init(&event->list);
4827 schedule_work(&event->remove);
4829 spin_unlock(&memcg->event_list_lock);
4831 memcg_unregister_all_caches(memcg);
4832 vmpressure_cleanup(&memcg->vmpressure);
4835 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
4837 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4839 memcg_destroy_kmem(memcg);
4840 __mem_cgroup_free(memcg);
4844 * mem_cgroup_css_reset - reset the states of a mem_cgroup
4845 * @css: the target css
4847 * Reset the states of the mem_cgroup associated with @css. This is
4848 * invoked when the userland requests disabling on the default hierarchy
4849 * but the memcg is pinned through dependency. The memcg should stop
4850 * applying policies and should revert to the vanilla state as it may be
4851 * made visible again.
4853 * The current implementation only resets the essential configurations.
4854 * This needs to be expanded to cover all the visible parts.
4856 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
4858 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4860 mem_cgroup_resize_limit(memcg, PAGE_COUNTER_MAX);
4861 mem_cgroup_resize_memsw_limit(memcg, PAGE_COUNTER_MAX);
4862 memcg_update_kmem_limit(memcg, PAGE_COUNTER_MAX);
4863 memcg->soft_limit = 0;
4867 /* Handlers for move charge at task migration. */
4868 static int mem_cgroup_do_precharge(unsigned long count)
4872 /* Try a single bulk charge without reclaim first */
4873 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_WAIT, count);
4875 mc.precharge += count;
4878 if (ret == -EINTR) {
4879 cancel_charge(root_mem_cgroup, count);
4883 /* Try charges one by one with reclaim */
4885 ret = try_charge(mc.to, GFP_KERNEL & ~__GFP_NORETRY, 1);
4887 * In case of failure, any residual charges against
4888 * mc.to will be dropped by mem_cgroup_clear_mc()
4889 * later on. However, cancel any charges that are
4890 * bypassed to root right away or they'll be lost.
4893 cancel_charge(root_mem_cgroup, 1);
4903 * get_mctgt_type - get target type of moving charge
4904 * @vma: the vma the pte to be checked belongs
4905 * @addr: the address corresponding to the pte to be checked
4906 * @ptent: the pte to be checked
4907 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4910 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4911 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4912 * move charge. if @target is not NULL, the page is stored in target->page
4913 * with extra refcnt got(Callers should handle it).
4914 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4915 * target for charge migration. if @target is not NULL, the entry is stored
4918 * Called with pte lock held.
4925 enum mc_target_type {
4931 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4932 unsigned long addr, pte_t ptent)
4934 struct page *page = vm_normal_page(vma, addr, ptent);
4936 if (!page || !page_mapped(page))
4938 if (PageAnon(page)) {
4939 /* we don't move shared anon */
4942 } else if (!move_file())
4943 /* we ignore mapcount for file pages */
4945 if (!get_page_unless_zero(page))
4952 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4953 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4955 struct page *page = NULL;
4956 swp_entry_t ent = pte_to_swp_entry(ptent);
4958 if (!move_anon() || non_swap_entry(ent))
4961 * Because lookup_swap_cache() updates some statistics counter,
4962 * we call find_get_page() with swapper_space directly.
4964 page = find_get_page(swap_address_space(ent), ent.val);
4965 if (do_swap_account)
4966 entry->val = ent.val;
4971 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4972 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4978 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4979 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4981 struct page *page = NULL;
4982 struct address_space *mapping;
4985 if (!vma->vm_file) /* anonymous vma */
4990 mapping = vma->vm_file->f_mapping;
4991 if (pte_none(ptent))
4992 pgoff = linear_page_index(vma, addr);
4993 else /* pte_file(ptent) is true */
4994 pgoff = pte_to_pgoff(ptent);
4996 /* page is moved even if it's not RSS of this task(page-faulted). */
4998 /* shmem/tmpfs may report page out on swap: account for that too. */
4999 if (shmem_mapping(mapping)) {
5000 page = find_get_entry(mapping, pgoff);
5001 if (radix_tree_exceptional_entry(page)) {
5002 swp_entry_t swp = radix_to_swp_entry(page);
5003 if (do_swap_account)
5005 page = find_get_page(swap_address_space(swp), swp.val);
5008 page = find_get_page(mapping, pgoff);
5010 page = find_get_page(mapping, pgoff);
5015 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5016 unsigned long addr, pte_t ptent, union mc_target *target)
5018 struct page *page = NULL;
5019 enum mc_target_type ret = MC_TARGET_NONE;
5020 swp_entry_t ent = { .val = 0 };
5022 if (pte_present(ptent))
5023 page = mc_handle_present_pte(vma, addr, ptent);
5024 else if (is_swap_pte(ptent))
5025 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5026 else if (pte_none(ptent) || pte_file(ptent))
5027 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5029 if (!page && !ent.val)
5033 * Do only loose check w/o serialization.
5034 * mem_cgroup_move_account() checks the page is valid or
5035 * not under LRU exclusion.
5037 if (page->mem_cgroup == mc.from) {
5038 ret = MC_TARGET_PAGE;
5040 target->page = page;
5042 if (!ret || !target)
5045 /* There is a swap entry and a page doesn't exist or isn't charged */
5046 if (ent.val && !ret &&
5047 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
5048 ret = MC_TARGET_SWAP;
5055 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5057 * We don't consider swapping or file mapped pages because THP does not
5058 * support them for now.
5059 * Caller should make sure that pmd_trans_huge(pmd) is true.
5061 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5062 unsigned long addr, pmd_t pmd, union mc_target *target)
5064 struct page *page = NULL;
5065 enum mc_target_type ret = MC_TARGET_NONE;
5067 page = pmd_page(pmd);
5068 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
5071 if (page->mem_cgroup == mc.from) {
5072 ret = MC_TARGET_PAGE;
5075 target->page = page;
5081 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5082 unsigned long addr, pmd_t pmd, union mc_target *target)
5084 return MC_TARGET_NONE;
5088 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5089 unsigned long addr, unsigned long end,
5090 struct mm_walk *walk)
5092 struct vm_area_struct *vma = walk->private;
5096 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5097 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5098 mc.precharge += HPAGE_PMD_NR;
5103 if (pmd_trans_unstable(pmd))
5105 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5106 for (; addr != end; pte++, addr += PAGE_SIZE)
5107 if (get_mctgt_type(vma, addr, *pte, NULL))
5108 mc.precharge++; /* increment precharge temporarily */
5109 pte_unmap_unlock(pte - 1, ptl);
5115 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5117 unsigned long precharge;
5118 struct vm_area_struct *vma;
5120 down_read(&mm->mmap_sem);
5121 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5122 struct mm_walk mem_cgroup_count_precharge_walk = {
5123 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5127 if (is_vm_hugetlb_page(vma))
5129 walk_page_range(vma->vm_start, vma->vm_end,
5130 &mem_cgroup_count_precharge_walk);
5132 up_read(&mm->mmap_sem);
5134 precharge = mc.precharge;
5140 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5142 unsigned long precharge = mem_cgroup_count_precharge(mm);
5144 VM_BUG_ON(mc.moving_task);
5145 mc.moving_task = current;
5146 return mem_cgroup_do_precharge(precharge);
5149 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5150 static void __mem_cgroup_clear_mc(void)
5152 struct mem_cgroup *from = mc.from;
5153 struct mem_cgroup *to = mc.to;
5155 /* we must uncharge all the leftover precharges from mc.to */
5157 cancel_charge(mc.to, mc.precharge);
5161 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5162 * we must uncharge here.
5164 if (mc.moved_charge) {
5165 cancel_charge(mc.from, mc.moved_charge);
5166 mc.moved_charge = 0;
5168 /* we must fixup refcnts and charges */
5169 if (mc.moved_swap) {
5170 /* uncharge swap account from the old cgroup */
5171 if (!mem_cgroup_is_root(mc.from))
5172 page_counter_uncharge(&mc.from->memsw, mc.moved_swap);
5175 * we charged both to->memory and to->memsw, so we
5176 * should uncharge to->memory.
5178 if (!mem_cgroup_is_root(mc.to))
5179 page_counter_uncharge(&mc.to->memory, mc.moved_swap);
5181 css_put_many(&mc.from->css, mc.moved_swap);
5183 /* we've already done css_get(mc.to) */
5186 memcg_oom_recover(from);
5187 memcg_oom_recover(to);
5188 wake_up_all(&mc.waitq);
5191 static void mem_cgroup_clear_mc(void)
5194 * we must clear moving_task before waking up waiters at the end of
5197 mc.moving_task = NULL;
5198 __mem_cgroup_clear_mc();
5199 spin_lock(&mc.lock);
5202 spin_unlock(&mc.lock);
5205 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5206 struct cgroup_taskset *tset)
5208 struct task_struct *p = cgroup_taskset_first(tset);
5210 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5211 unsigned long move_charge_at_immigrate;
5214 * We are now commited to this value whatever it is. Changes in this
5215 * tunable will only affect upcoming migrations, not the current one.
5216 * So we need to save it, and keep it going.
5218 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
5219 if (move_charge_at_immigrate) {
5220 struct mm_struct *mm;
5221 struct mem_cgroup *from = mem_cgroup_from_task(p);
5223 VM_BUG_ON(from == memcg);
5225 mm = get_task_mm(p);
5228 /* We move charges only when we move a owner of the mm */
5229 if (mm->owner == p) {
5232 VM_BUG_ON(mc.precharge);
5233 VM_BUG_ON(mc.moved_charge);
5234 VM_BUG_ON(mc.moved_swap);
5236 spin_lock(&mc.lock);
5239 mc.immigrate_flags = move_charge_at_immigrate;
5240 spin_unlock(&mc.lock);
5241 /* We set mc.moving_task later */
5243 ret = mem_cgroup_precharge_mc(mm);
5245 mem_cgroup_clear_mc();
5252 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5253 struct cgroup_taskset *tset)
5256 mem_cgroup_clear_mc();
5259 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5260 unsigned long addr, unsigned long end,
5261 struct mm_walk *walk)
5264 struct vm_area_struct *vma = walk->private;
5267 enum mc_target_type target_type;
5268 union mc_target target;
5272 * We don't take compound_lock() here but no race with splitting thp
5274 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5275 * under splitting, which means there's no concurrent thp split,
5276 * - if another thread runs into split_huge_page() just after we
5277 * entered this if-block, the thread must wait for page table lock
5278 * to be unlocked in __split_huge_page_splitting(), where the main
5279 * part of thp split is not executed yet.
5281 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
5282 if (mc.precharge < HPAGE_PMD_NR) {
5286 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5287 if (target_type == MC_TARGET_PAGE) {
5289 if (!isolate_lru_page(page)) {
5290 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5292 mc.precharge -= HPAGE_PMD_NR;
5293 mc.moved_charge += HPAGE_PMD_NR;
5295 putback_lru_page(page);
5303 if (pmd_trans_unstable(pmd))
5306 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5307 for (; addr != end; addr += PAGE_SIZE) {
5308 pte_t ptent = *(pte++);
5314 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5315 case MC_TARGET_PAGE:
5317 if (isolate_lru_page(page))
5319 if (!mem_cgroup_move_account(page, 1, mc.from, mc.to)) {
5321 /* we uncharge from mc.from later. */
5324 putback_lru_page(page);
5325 put: /* get_mctgt_type() gets the page */
5328 case MC_TARGET_SWAP:
5330 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5332 /* we fixup refcnts and charges later. */
5340 pte_unmap_unlock(pte - 1, ptl);
5345 * We have consumed all precharges we got in can_attach().
5346 * We try charge one by one, but don't do any additional
5347 * charges to mc.to if we have failed in charge once in attach()
5350 ret = mem_cgroup_do_precharge(1);
5358 static void mem_cgroup_move_charge(struct mm_struct *mm)
5360 struct vm_area_struct *vma;
5362 lru_add_drain_all();
5364 * Signal mem_cgroup_begin_page_stat() to take the memcg's
5365 * move_lock while we're moving its pages to another memcg.
5366 * Then wait for already started RCU-only updates to finish.
5368 atomic_inc(&mc.from->moving_account);
5371 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5373 * Someone who are holding the mmap_sem might be waiting in
5374 * waitq. So we cancel all extra charges, wake up all waiters,
5375 * and retry. Because we cancel precharges, we might not be able
5376 * to move enough charges, but moving charge is a best-effort
5377 * feature anyway, so it wouldn't be a big problem.
5379 __mem_cgroup_clear_mc();
5383 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5385 struct mm_walk mem_cgroup_move_charge_walk = {
5386 .pmd_entry = mem_cgroup_move_charge_pte_range,
5390 if (is_vm_hugetlb_page(vma))
5392 ret = walk_page_range(vma->vm_start, vma->vm_end,
5393 &mem_cgroup_move_charge_walk);
5396 * means we have consumed all precharges and failed in
5397 * doing additional charge. Just abandon here.
5401 up_read(&mm->mmap_sem);
5402 atomic_dec(&mc.from->moving_account);
5405 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5406 struct cgroup_taskset *tset)
5408 struct task_struct *p = cgroup_taskset_first(tset);
5409 struct mm_struct *mm = get_task_mm(p);
5413 mem_cgroup_move_charge(mm);
5417 mem_cgroup_clear_mc();
5419 #else /* !CONFIG_MMU */
5420 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
5421 struct cgroup_taskset *tset)
5425 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
5426 struct cgroup_taskset *tset)
5429 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
5430 struct cgroup_taskset *tset)
5436 * Cgroup retains root cgroups across [un]mount cycles making it necessary
5437 * to verify whether we're attached to the default hierarchy on each mount
5440 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
5443 * use_hierarchy is forced on the default hierarchy. cgroup core
5444 * guarantees that @root doesn't have any children, so turning it
5445 * on for the root memcg is enough.
5447 if (cgroup_on_dfl(root_css->cgroup))
5448 mem_cgroup_from_css(root_css)->use_hierarchy = true;
5451 struct cgroup_subsys memory_cgrp_subsys = {
5452 .css_alloc = mem_cgroup_css_alloc,
5453 .css_online = mem_cgroup_css_online,
5454 .css_offline = mem_cgroup_css_offline,
5455 .css_free = mem_cgroup_css_free,
5456 .css_reset = mem_cgroup_css_reset,
5457 .can_attach = mem_cgroup_can_attach,
5458 .cancel_attach = mem_cgroup_cancel_attach,
5459 .attach = mem_cgroup_move_task,
5460 .bind = mem_cgroup_bind,
5461 .legacy_cftypes = mem_cgroup_files,
5465 #ifdef CONFIG_MEMCG_SWAP
5466 static int __init enable_swap_account(char *s)
5468 if (!strcmp(s, "1"))
5469 really_do_swap_account = 1;
5470 else if (!strcmp(s, "0"))
5471 really_do_swap_account = 0;
5474 __setup("swapaccount=", enable_swap_account);
5476 static void __init memsw_file_init(void)
5478 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
5479 memsw_cgroup_files));
5482 static void __init enable_swap_cgroup(void)
5484 if (!mem_cgroup_disabled() && really_do_swap_account) {
5485 do_swap_account = 1;
5491 static void __init enable_swap_cgroup(void)
5496 #ifdef CONFIG_MEMCG_SWAP
5498 * mem_cgroup_swapout - transfer a memsw charge to swap
5499 * @page: page whose memsw charge to transfer
5500 * @entry: swap entry to move the charge to
5502 * Transfer the memsw charge of @page to @entry.
5504 void mem_cgroup_swapout(struct page *page, swp_entry_t entry)
5506 struct mem_cgroup *memcg;
5507 unsigned short oldid;
5509 VM_BUG_ON_PAGE(PageLRU(page), page);
5510 VM_BUG_ON_PAGE(page_count(page), page);
5512 if (!do_swap_account)
5515 memcg = page->mem_cgroup;
5517 /* Readahead page, never charged */
5521 oldid = swap_cgroup_record(entry, mem_cgroup_id(memcg));
5522 VM_BUG_ON_PAGE(oldid, page);
5523 mem_cgroup_swap_statistics(memcg, true);
5525 page->mem_cgroup = NULL;
5527 if (!mem_cgroup_is_root(memcg))
5528 page_counter_uncharge(&memcg->memory, 1);
5530 /* XXX: caller holds IRQ-safe mapping->tree_lock */
5531 VM_BUG_ON(!irqs_disabled());
5533 mem_cgroup_charge_statistics(memcg, page, -1);
5534 memcg_check_events(memcg, page);
5538 * mem_cgroup_uncharge_swap - uncharge a swap entry
5539 * @entry: swap entry to uncharge
5541 * Drop the memsw charge associated with @entry.
5543 void mem_cgroup_uncharge_swap(swp_entry_t entry)
5545 struct mem_cgroup *memcg;
5548 if (!do_swap_account)
5551 id = swap_cgroup_record(entry, 0);
5553 memcg = mem_cgroup_lookup(id);
5555 if (!mem_cgroup_is_root(memcg))
5556 page_counter_uncharge(&memcg->memsw, 1);
5557 mem_cgroup_swap_statistics(memcg, false);
5558 css_put(&memcg->css);
5565 * mem_cgroup_try_charge - try charging a page
5566 * @page: page to charge
5567 * @mm: mm context of the victim
5568 * @gfp_mask: reclaim mode
5569 * @memcgp: charged memcg return
5571 * Try to charge @page to the memcg that @mm belongs to, reclaiming
5572 * pages according to @gfp_mask if necessary.
5574 * Returns 0 on success, with *@memcgp pointing to the charged memcg.
5575 * Otherwise, an error code is returned.
5577 * After page->mapping has been set up, the caller must finalize the
5578 * charge with mem_cgroup_commit_charge(). Or abort the transaction
5579 * with mem_cgroup_cancel_charge() in case page instantiation fails.
5581 int mem_cgroup_try_charge(struct page *page, struct mm_struct *mm,
5582 gfp_t gfp_mask, struct mem_cgroup **memcgp)
5584 struct mem_cgroup *memcg = NULL;
5585 unsigned int nr_pages = 1;
5588 if (mem_cgroup_disabled())
5591 if (PageSwapCache(page)) {
5593 * Every swap fault against a single page tries to charge the
5594 * page, bail as early as possible. shmem_unuse() encounters
5595 * already charged pages, too. The USED bit is protected by
5596 * the page lock, which serializes swap cache removal, which
5597 * in turn serializes uncharging.
5599 if (page->mem_cgroup)
5603 if (PageTransHuge(page)) {
5604 nr_pages <<= compound_order(page);
5605 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5608 if (do_swap_account && PageSwapCache(page))
5609 memcg = try_get_mem_cgroup_from_page(page);
5611 memcg = get_mem_cgroup_from_mm(mm);
5613 ret = try_charge(memcg, gfp_mask, nr_pages);
5615 css_put(&memcg->css);
5617 if (ret == -EINTR) {
5618 memcg = root_mem_cgroup;
5627 * mem_cgroup_commit_charge - commit a page charge
5628 * @page: page to charge
5629 * @memcg: memcg to charge the page to
5630 * @lrucare: page might be on LRU already
5632 * Finalize a charge transaction started by mem_cgroup_try_charge(),
5633 * after page->mapping has been set up. This must happen atomically
5634 * as part of the page instantiation, i.e. under the page table lock
5635 * for anonymous pages, under the page lock for page and swap cache.
5637 * In addition, the page must not be on the LRU during the commit, to
5638 * prevent racing with task migration. If it might be, use @lrucare.
5640 * Use mem_cgroup_cancel_charge() to cancel the transaction instead.
5642 void mem_cgroup_commit_charge(struct page *page, struct mem_cgroup *memcg,
5645 unsigned int nr_pages = 1;
5647 VM_BUG_ON_PAGE(!page->mapping, page);
5648 VM_BUG_ON_PAGE(PageLRU(page) && !lrucare, page);
5650 if (mem_cgroup_disabled())
5653 * Swap faults will attempt to charge the same page multiple
5654 * times. But reuse_swap_page() might have removed the page
5655 * from swapcache already, so we can't check PageSwapCache().
5660 commit_charge(page, memcg, lrucare);
5662 if (PageTransHuge(page)) {
5663 nr_pages <<= compound_order(page);
5664 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5667 local_irq_disable();
5668 mem_cgroup_charge_statistics(memcg, page, nr_pages);
5669 memcg_check_events(memcg, page);
5672 if (do_swap_account && PageSwapCache(page)) {
5673 swp_entry_t entry = { .val = page_private(page) };
5675 * The swap entry might not get freed for a long time,
5676 * let's not wait for it. The page already received a
5677 * memory+swap charge, drop the swap entry duplicate.
5679 mem_cgroup_uncharge_swap(entry);
5684 * mem_cgroup_cancel_charge - cancel a page charge
5685 * @page: page to charge
5686 * @memcg: memcg to charge the page to
5688 * Cancel a charge transaction started by mem_cgroup_try_charge().
5690 void mem_cgroup_cancel_charge(struct page *page, struct mem_cgroup *memcg)
5692 unsigned int nr_pages = 1;
5694 if (mem_cgroup_disabled())
5697 * Swap faults will attempt to charge the same page multiple
5698 * times. But reuse_swap_page() might have removed the page
5699 * from swapcache already, so we can't check PageSwapCache().
5704 if (PageTransHuge(page)) {
5705 nr_pages <<= compound_order(page);
5706 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5709 cancel_charge(memcg, nr_pages);
5712 static void uncharge_batch(struct mem_cgroup *memcg, unsigned long pgpgout,
5713 unsigned long nr_anon, unsigned long nr_file,
5714 unsigned long nr_huge, struct page *dummy_page)
5716 unsigned long nr_pages = nr_anon + nr_file;
5717 unsigned long flags;
5719 if (!mem_cgroup_is_root(memcg)) {
5720 page_counter_uncharge(&memcg->memory, nr_pages);
5721 if (do_swap_account)
5722 page_counter_uncharge(&memcg->memsw, nr_pages);
5723 memcg_oom_recover(memcg);
5726 local_irq_save(flags);
5727 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS], nr_anon);
5728 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_CACHE], nr_file);
5729 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE], nr_huge);
5730 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT], pgpgout);
5731 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
5732 memcg_check_events(memcg, dummy_page);
5733 local_irq_restore(flags);
5735 if (!mem_cgroup_is_root(memcg))
5736 css_put_many(&memcg->css, nr_pages);
5739 static void uncharge_list(struct list_head *page_list)
5741 struct mem_cgroup *memcg = NULL;
5742 unsigned long nr_anon = 0;
5743 unsigned long nr_file = 0;
5744 unsigned long nr_huge = 0;
5745 unsigned long pgpgout = 0;
5746 struct list_head *next;
5749 next = page_list->next;
5751 unsigned int nr_pages = 1;
5753 page = list_entry(next, struct page, lru);
5754 next = page->lru.next;
5756 VM_BUG_ON_PAGE(PageLRU(page), page);
5757 VM_BUG_ON_PAGE(page_count(page), page);
5759 if (!page->mem_cgroup)
5763 * Nobody should be changing or seriously looking at
5764 * page->mem_cgroup at this point, we have fully
5765 * exclusive access to the page.
5768 if (memcg != page->mem_cgroup) {
5770 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5772 pgpgout = nr_anon = nr_file = nr_huge = 0;
5774 memcg = page->mem_cgroup;
5777 if (PageTransHuge(page)) {
5778 nr_pages <<= compound_order(page);
5779 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
5780 nr_huge += nr_pages;
5784 nr_anon += nr_pages;
5786 nr_file += nr_pages;
5788 page->mem_cgroup = NULL;
5791 } while (next != page_list);
5794 uncharge_batch(memcg, pgpgout, nr_anon, nr_file,
5799 * mem_cgroup_uncharge - uncharge a page
5800 * @page: page to uncharge
5802 * Uncharge a page previously charged with mem_cgroup_try_charge() and
5803 * mem_cgroup_commit_charge().
5805 void mem_cgroup_uncharge(struct page *page)
5807 if (mem_cgroup_disabled())
5810 /* Don't touch page->lru of any random page, pre-check: */
5811 if (!page->mem_cgroup)
5814 INIT_LIST_HEAD(&page->lru);
5815 uncharge_list(&page->lru);
5819 * mem_cgroup_uncharge_list - uncharge a list of page
5820 * @page_list: list of pages to uncharge
5822 * Uncharge a list of pages previously charged with
5823 * mem_cgroup_try_charge() and mem_cgroup_commit_charge().
5825 void mem_cgroup_uncharge_list(struct list_head *page_list)
5827 if (mem_cgroup_disabled())
5830 if (!list_empty(page_list))
5831 uncharge_list(page_list);
5835 * mem_cgroup_migrate - migrate a charge to another page
5836 * @oldpage: currently charged page
5837 * @newpage: page to transfer the charge to
5838 * @lrucare: both pages might be on the LRU already
5840 * Migrate the charge from @oldpage to @newpage.
5842 * Both pages must be locked, @newpage->mapping must be set up.
5844 void mem_cgroup_migrate(struct page *oldpage, struct page *newpage,
5847 struct mem_cgroup *memcg;
5850 VM_BUG_ON_PAGE(!PageLocked(oldpage), oldpage);
5851 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
5852 VM_BUG_ON_PAGE(!lrucare && PageLRU(oldpage), oldpage);
5853 VM_BUG_ON_PAGE(!lrucare && PageLRU(newpage), newpage);
5854 VM_BUG_ON_PAGE(PageAnon(oldpage) != PageAnon(newpage), newpage);
5855 VM_BUG_ON_PAGE(PageTransHuge(oldpage) != PageTransHuge(newpage),
5858 if (mem_cgroup_disabled())
5861 /* Page cache replacement: new page already charged? */
5862 if (newpage->mem_cgroup)
5866 * Swapcache readahead pages can get migrated before being
5867 * charged, and migration from compaction can happen to an
5868 * uncharged page when the PFN walker finds a page that
5869 * reclaim just put back on the LRU but has not released yet.
5871 memcg = oldpage->mem_cgroup;
5876 lock_page_lru(oldpage, &isolated);
5878 oldpage->mem_cgroup = NULL;
5881 unlock_page_lru(oldpage, isolated);
5883 commit_charge(newpage, memcg, lrucare);
5887 * subsys_initcall() for memory controller.
5889 * Some parts like hotcpu_notifier() have to be initialized from this context
5890 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
5891 * everything that doesn't depend on a specific mem_cgroup structure should
5892 * be initialized from here.
5894 static int __init mem_cgroup_init(void)
5896 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5897 enable_swap_cgroup();
5898 mem_cgroup_soft_limit_tree_init();
5902 subsys_initcall(mem_cgroup_init);