1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/sort.h>
50 #include <linux/seq_file.h>
51 #include <linux/vmalloc.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
60 #include <net/tcp_memcontrol.h>
62 #include <asm/uaccess.h>
64 #include <trace/events/vmscan.h>
66 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
67 EXPORT_SYMBOL(mem_cgroup_subsys);
69 #define MEM_CGROUP_RECLAIM_RETRIES 5
70 static struct mem_cgroup *root_mem_cgroup __read_mostly;
72 #ifdef CONFIG_MEMCG_SWAP
73 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
74 int do_swap_account __read_mostly;
76 /* for remember boot option*/
77 #ifdef CONFIG_MEMCG_SWAP_ENABLED
78 static int really_do_swap_account __initdata = 1;
80 static int really_do_swap_account __initdata = 0;
84 #define do_swap_account 0
89 * Statistics for memory cgroup.
91 enum mem_cgroup_stat_index {
93 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
95 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
96 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
97 MEM_CGROUP_STAT_RSS_HUGE, /* # of pages charged as anon huge */
98 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
99 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
100 MEM_CGROUP_STAT_NSTATS,
103 static const char * const mem_cgroup_stat_names[] = {
111 enum mem_cgroup_events_index {
112 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
113 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
114 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
115 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
116 MEM_CGROUP_EVENTS_NSTATS,
119 static const char * const mem_cgroup_events_names[] = {
126 static const char * const mem_cgroup_lru_names[] = {
135 * Per memcg event counter is incremented at every pagein/pageout. With THP,
136 * it will be incremated by the number of pages. This counter is used for
137 * for trigger some periodic events. This is straightforward and better
138 * than using jiffies etc. to handle periodic memcg event.
140 enum mem_cgroup_events_target {
141 MEM_CGROUP_TARGET_THRESH,
142 MEM_CGROUP_TARGET_SOFTLIMIT,
143 MEM_CGROUP_TARGET_NUMAINFO,
146 #define THRESHOLDS_EVENTS_TARGET 128
147 #define SOFTLIMIT_EVENTS_TARGET 1024
148 #define NUMAINFO_EVENTS_TARGET 1024
150 struct mem_cgroup_stat_cpu {
151 long count[MEM_CGROUP_STAT_NSTATS];
152 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
153 unsigned long nr_page_events;
154 unsigned long targets[MEM_CGROUP_NTARGETS];
157 struct mem_cgroup_reclaim_iter {
159 * last scanned hierarchy member. Valid only if last_dead_count
160 * matches memcg->dead_count of the hierarchy root group.
162 struct mem_cgroup *last_visited;
163 unsigned long last_dead_count;
165 /* scan generation, increased every round-trip */
166 unsigned int generation;
170 * per-zone information in memory controller.
172 struct mem_cgroup_per_zone {
173 struct lruvec lruvec;
174 unsigned long lru_size[NR_LRU_LISTS];
176 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
178 struct rb_node tree_node; /* RB tree node */
179 unsigned long long usage_in_excess;/* Set to the value by which */
180 /* the soft limit is exceeded*/
182 struct mem_cgroup *memcg; /* Back pointer, we cannot */
183 /* use container_of */
186 struct mem_cgroup_per_node {
187 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
191 * Cgroups above their limits are maintained in a RB-Tree, independent of
192 * their hierarchy representation
195 struct mem_cgroup_tree_per_zone {
196 struct rb_root rb_root;
200 struct mem_cgroup_tree_per_node {
201 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
204 struct mem_cgroup_tree {
205 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
208 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
210 struct mem_cgroup_threshold {
211 struct eventfd_ctx *eventfd;
216 struct mem_cgroup_threshold_ary {
217 /* An array index points to threshold just below or equal to usage. */
218 int current_threshold;
219 /* Size of entries[] */
221 /* Array of thresholds */
222 struct mem_cgroup_threshold entries[0];
225 struct mem_cgroup_thresholds {
226 /* Primary thresholds array */
227 struct mem_cgroup_threshold_ary *primary;
229 * Spare threshold array.
230 * This is needed to make mem_cgroup_unregister_event() "never fail".
231 * It must be able to store at least primary->size - 1 entries.
233 struct mem_cgroup_threshold_ary *spare;
237 struct mem_cgroup_eventfd_list {
238 struct list_head list;
239 struct eventfd_ctx *eventfd;
242 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
243 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
246 * The memory controller data structure. The memory controller controls both
247 * page cache and RSS per cgroup. We would eventually like to provide
248 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
249 * to help the administrator determine what knobs to tune.
251 * TODO: Add a water mark for the memory controller. Reclaim will begin when
252 * we hit the water mark. May be even add a low water mark, such that
253 * no reclaim occurs from a cgroup at it's low water mark, this is
254 * a feature that will be implemented much later in the future.
257 struct cgroup_subsys_state css;
259 * the counter to account for memory usage
261 struct res_counter res;
263 /* vmpressure notifications */
264 struct vmpressure vmpressure;
268 * the counter to account for mem+swap usage.
270 struct res_counter memsw;
273 * rcu_freeing is used only when freeing struct mem_cgroup,
274 * so put it into a union to avoid wasting more memory.
275 * It must be disjoint from the css field. It could be
276 * in a union with the res field, but res plays a much
277 * larger part in mem_cgroup life than memsw, and might
278 * be of interest, even at time of free, when debugging.
279 * So share rcu_head with the less interesting memsw.
281 struct rcu_head rcu_freeing;
283 * We also need some space for a worker in deferred freeing.
284 * By the time we call it, rcu_freeing is no longer in use.
286 struct work_struct work_freeing;
290 * the counter to account for kernel memory usage.
292 struct res_counter kmem;
294 * Should the accounting and control be hierarchical, per subtree?
297 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
305 /* OOM-Killer disable */
306 int oom_kill_disable;
308 /* set when res.limit == memsw.limit */
309 bool memsw_is_minimum;
311 /* protect arrays of thresholds */
312 struct mutex thresholds_lock;
314 /* thresholds for memory usage. RCU-protected */
315 struct mem_cgroup_thresholds thresholds;
317 /* thresholds for mem+swap usage. RCU-protected */
318 struct mem_cgroup_thresholds memsw_thresholds;
320 /* For oom notifier event fd */
321 struct list_head oom_notify;
324 * Should we move charges of a task when a task is moved into this
325 * mem_cgroup ? And what type of charges should we move ?
327 unsigned long move_charge_at_immigrate;
329 * set > 0 if pages under this cgroup are moving to other cgroup.
331 atomic_t moving_account;
332 /* taken only while moving_account > 0 */
333 spinlock_t move_lock;
337 struct mem_cgroup_stat_cpu __percpu *stat;
339 * used when a cpu is offlined or other synchronizations
340 * See mem_cgroup_read_stat().
342 struct mem_cgroup_stat_cpu nocpu_base;
343 spinlock_t pcp_counter_lock;
346 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
347 struct tcp_memcontrol tcp_mem;
349 #if defined(CONFIG_MEMCG_KMEM)
350 /* analogous to slab_common's slab_caches list. per-memcg */
351 struct list_head memcg_slab_caches;
352 /* Not a spinlock, we can take a lot of time walking the list */
353 struct mutex slab_caches_mutex;
354 /* Index in the kmem_cache->memcg_params->memcg_caches array */
358 int last_scanned_node;
360 nodemask_t scan_nodes;
361 atomic_t numainfo_events;
362 atomic_t numainfo_updating;
365 struct mem_cgroup_per_node *nodeinfo[0];
366 /* WARNING: nodeinfo must be the last member here */
369 static size_t memcg_size(void)
371 return sizeof(struct mem_cgroup) +
372 nr_node_ids * sizeof(struct mem_cgroup_per_node);
375 /* internal only representation about the status of kmem accounting. */
377 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
378 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
379 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
382 /* We account when limit is on, but only after call sites are patched */
383 #define KMEM_ACCOUNTED_MASK \
384 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
386 #ifdef CONFIG_MEMCG_KMEM
387 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
389 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
392 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
394 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
397 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
399 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
402 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
404 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
407 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
410 * Our caller must use css_get() first, because memcg_uncharge_kmem()
411 * will call css_put() if it sees the memcg is dead.
414 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
415 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
418 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
420 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
421 &memcg->kmem_account_flags);
425 /* Stuffs for move charges at task migration. */
427 * Types of charges to be moved. "move_charge_at_immitgrate" and
428 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
431 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
432 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
436 /* "mc" and its members are protected by cgroup_mutex */
437 static struct move_charge_struct {
438 spinlock_t lock; /* for from, to */
439 struct mem_cgroup *from;
440 struct mem_cgroup *to;
441 unsigned long immigrate_flags;
442 unsigned long precharge;
443 unsigned long moved_charge;
444 unsigned long moved_swap;
445 struct task_struct *moving_task; /* a task moving charges */
446 wait_queue_head_t waitq; /* a waitq for other context */
448 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
449 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
452 static bool move_anon(void)
454 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
457 static bool move_file(void)
459 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
463 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
464 * limit reclaim to prevent infinite loops, if they ever occur.
466 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
467 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
470 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
471 MEM_CGROUP_CHARGE_TYPE_ANON,
472 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
473 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
477 /* for encoding cft->private value on file */
485 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
486 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
487 #define MEMFILE_ATTR(val) ((val) & 0xffff)
488 /* Used for OOM nofiier */
489 #define OOM_CONTROL (0)
492 * Reclaim flags for mem_cgroup_hierarchical_reclaim
494 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
495 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
496 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
497 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
500 * The memcg_create_mutex will be held whenever a new cgroup is created.
501 * As a consequence, any change that needs to protect against new child cgroups
502 * appearing has to hold it as well.
504 static DEFINE_MUTEX(memcg_create_mutex);
506 static void mem_cgroup_get(struct mem_cgroup *memcg);
507 static void mem_cgroup_put(struct mem_cgroup *memcg);
510 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
512 return container_of(s, struct mem_cgroup, css);
515 /* Some nice accessors for the vmpressure. */
516 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
519 memcg = root_mem_cgroup;
520 return &memcg->vmpressure;
523 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
525 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
528 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
530 return &mem_cgroup_from_css(css)->vmpressure;
533 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
535 return (memcg == root_mem_cgroup);
538 /* Writing them here to avoid exposing memcg's inner layout */
539 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
541 void sock_update_memcg(struct sock *sk)
543 if (mem_cgroup_sockets_enabled) {
544 struct mem_cgroup *memcg;
545 struct cg_proto *cg_proto;
547 BUG_ON(!sk->sk_prot->proto_cgroup);
549 /* Socket cloning can throw us here with sk_cgrp already
550 * filled. It won't however, necessarily happen from
551 * process context. So the test for root memcg given
552 * the current task's memcg won't help us in this case.
554 * Respecting the original socket's memcg is a better
555 * decision in this case.
558 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
559 css_get(&sk->sk_cgrp->memcg->css);
564 memcg = mem_cgroup_from_task(current);
565 cg_proto = sk->sk_prot->proto_cgroup(memcg);
566 if (!mem_cgroup_is_root(memcg) &&
567 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
568 sk->sk_cgrp = cg_proto;
573 EXPORT_SYMBOL(sock_update_memcg);
575 void sock_release_memcg(struct sock *sk)
577 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
578 struct mem_cgroup *memcg;
579 WARN_ON(!sk->sk_cgrp->memcg);
580 memcg = sk->sk_cgrp->memcg;
581 css_put(&sk->sk_cgrp->memcg->css);
585 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
587 if (!memcg || mem_cgroup_is_root(memcg))
590 return &memcg->tcp_mem.cg_proto;
592 EXPORT_SYMBOL(tcp_proto_cgroup);
594 static void disarm_sock_keys(struct mem_cgroup *memcg)
596 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
598 static_key_slow_dec(&memcg_socket_limit_enabled);
601 static void disarm_sock_keys(struct mem_cgroup *memcg)
606 #ifdef CONFIG_MEMCG_KMEM
608 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
609 * There are two main reasons for not using the css_id for this:
610 * 1) this works better in sparse environments, where we have a lot of memcgs,
611 * but only a few kmem-limited. Or also, if we have, for instance, 200
612 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
613 * 200 entry array for that.
615 * 2) In order not to violate the cgroup API, we would like to do all memory
616 * allocation in ->create(). At that point, we haven't yet allocated the
617 * css_id. Having a separate index prevents us from messing with the cgroup
620 * The current size of the caches array is stored in
621 * memcg_limited_groups_array_size. It will double each time we have to
624 static DEFINE_IDA(kmem_limited_groups);
625 int memcg_limited_groups_array_size;
628 * MIN_SIZE is different than 1, because we would like to avoid going through
629 * the alloc/free process all the time. In a small machine, 4 kmem-limited
630 * cgroups is a reasonable guess. In the future, it could be a parameter or
631 * tunable, but that is strictly not necessary.
633 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
634 * this constant directly from cgroup, but it is understandable that this is
635 * better kept as an internal representation in cgroup.c. In any case, the
636 * css_id space is not getting any smaller, and we don't have to necessarily
637 * increase ours as well if it increases.
639 #define MEMCG_CACHES_MIN_SIZE 4
640 #define MEMCG_CACHES_MAX_SIZE 65535
643 * A lot of the calls to the cache allocation functions are expected to be
644 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
645 * conditional to this static branch, we'll have to allow modules that does
646 * kmem_cache_alloc and the such to see this symbol as well
648 struct static_key memcg_kmem_enabled_key;
649 EXPORT_SYMBOL(memcg_kmem_enabled_key);
651 static void disarm_kmem_keys(struct mem_cgroup *memcg)
653 if (memcg_kmem_is_active(memcg)) {
654 static_key_slow_dec(&memcg_kmem_enabled_key);
655 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
658 * This check can't live in kmem destruction function,
659 * since the charges will outlive the cgroup
661 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
664 static void disarm_kmem_keys(struct mem_cgroup *memcg)
667 #endif /* CONFIG_MEMCG_KMEM */
669 static void disarm_static_keys(struct mem_cgroup *memcg)
671 disarm_sock_keys(memcg);
672 disarm_kmem_keys(memcg);
675 static void drain_all_stock_async(struct mem_cgroup *memcg);
677 static struct mem_cgroup_per_zone *
678 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
680 VM_BUG_ON((unsigned)nid >= nr_node_ids);
681 return &memcg->nodeinfo[nid]->zoneinfo[zid];
684 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
689 static struct mem_cgroup_per_zone *
690 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
692 int nid = page_to_nid(page);
693 int zid = page_zonenum(page);
695 return mem_cgroup_zoneinfo(memcg, nid, zid);
698 static struct mem_cgroup_tree_per_zone *
699 soft_limit_tree_node_zone(int nid, int zid)
701 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
704 static struct mem_cgroup_tree_per_zone *
705 soft_limit_tree_from_page(struct page *page)
707 int nid = page_to_nid(page);
708 int zid = page_zonenum(page);
710 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
714 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
715 struct mem_cgroup_per_zone *mz,
716 struct mem_cgroup_tree_per_zone *mctz,
717 unsigned long long new_usage_in_excess)
719 struct rb_node **p = &mctz->rb_root.rb_node;
720 struct rb_node *parent = NULL;
721 struct mem_cgroup_per_zone *mz_node;
726 mz->usage_in_excess = new_usage_in_excess;
727 if (!mz->usage_in_excess)
731 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
733 if (mz->usage_in_excess < mz_node->usage_in_excess)
736 * We can't avoid mem cgroups that are over their soft
737 * limit by the same amount
739 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
742 rb_link_node(&mz->tree_node, parent, p);
743 rb_insert_color(&mz->tree_node, &mctz->rb_root);
748 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
749 struct mem_cgroup_per_zone *mz,
750 struct mem_cgroup_tree_per_zone *mctz)
754 rb_erase(&mz->tree_node, &mctz->rb_root);
759 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
760 struct mem_cgroup_per_zone *mz,
761 struct mem_cgroup_tree_per_zone *mctz)
763 spin_lock(&mctz->lock);
764 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
765 spin_unlock(&mctz->lock);
769 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
771 unsigned long long excess;
772 struct mem_cgroup_per_zone *mz;
773 struct mem_cgroup_tree_per_zone *mctz;
774 int nid = page_to_nid(page);
775 int zid = page_zonenum(page);
776 mctz = soft_limit_tree_from_page(page);
779 * Necessary to update all ancestors when hierarchy is used.
780 * because their event counter is not touched.
782 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
783 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
784 excess = res_counter_soft_limit_excess(&memcg->res);
786 * We have to update the tree if mz is on RB-tree or
787 * mem is over its softlimit.
789 if (excess || mz->on_tree) {
790 spin_lock(&mctz->lock);
791 /* if on-tree, remove it */
793 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
795 * Insert again. mz->usage_in_excess will be updated.
796 * If excess is 0, no tree ops.
798 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
799 spin_unlock(&mctz->lock);
804 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
807 struct mem_cgroup_per_zone *mz;
808 struct mem_cgroup_tree_per_zone *mctz;
810 for_each_node(node) {
811 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
812 mz = mem_cgroup_zoneinfo(memcg, node, zone);
813 mctz = soft_limit_tree_node_zone(node, zone);
814 mem_cgroup_remove_exceeded(memcg, mz, mctz);
819 static struct mem_cgroup_per_zone *
820 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
822 struct rb_node *rightmost = NULL;
823 struct mem_cgroup_per_zone *mz;
827 rightmost = rb_last(&mctz->rb_root);
829 goto done; /* Nothing to reclaim from */
831 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
833 * Remove the node now but someone else can add it back,
834 * we will to add it back at the end of reclaim to its correct
835 * position in the tree.
837 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
838 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
839 !css_tryget(&mz->memcg->css))
845 static struct mem_cgroup_per_zone *
846 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
848 struct mem_cgroup_per_zone *mz;
850 spin_lock(&mctz->lock);
851 mz = __mem_cgroup_largest_soft_limit_node(mctz);
852 spin_unlock(&mctz->lock);
857 * Implementation Note: reading percpu statistics for memcg.
859 * Both of vmstat[] and percpu_counter has threshold and do periodic
860 * synchronization to implement "quick" read. There are trade-off between
861 * reading cost and precision of value. Then, we may have a chance to implement
862 * a periodic synchronizion of counter in memcg's counter.
864 * But this _read() function is used for user interface now. The user accounts
865 * memory usage by memory cgroup and he _always_ requires exact value because
866 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
867 * have to visit all online cpus and make sum. So, for now, unnecessary
868 * synchronization is not implemented. (just implemented for cpu hotplug)
870 * If there are kernel internal actions which can make use of some not-exact
871 * value, and reading all cpu value can be performance bottleneck in some
872 * common workload, threashold and synchonization as vmstat[] should be
875 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
876 enum mem_cgroup_stat_index idx)
882 for_each_online_cpu(cpu)
883 val += per_cpu(memcg->stat->count[idx], cpu);
884 #ifdef CONFIG_HOTPLUG_CPU
885 spin_lock(&memcg->pcp_counter_lock);
886 val += memcg->nocpu_base.count[idx];
887 spin_unlock(&memcg->pcp_counter_lock);
893 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
896 int val = (charge) ? 1 : -1;
897 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
900 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
901 enum mem_cgroup_events_index idx)
903 unsigned long val = 0;
906 for_each_online_cpu(cpu)
907 val += per_cpu(memcg->stat->events[idx], cpu);
908 #ifdef CONFIG_HOTPLUG_CPU
909 spin_lock(&memcg->pcp_counter_lock);
910 val += memcg->nocpu_base.events[idx];
911 spin_unlock(&memcg->pcp_counter_lock);
916 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
918 bool anon, int nr_pages)
923 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
924 * counted as CACHE even if it's on ANON LRU.
927 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
930 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
933 if (PageTransHuge(page))
934 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
937 /* pagein of a big page is an event. So, ignore page size */
939 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
941 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
942 nr_pages = -nr_pages; /* for event */
945 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
951 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
953 struct mem_cgroup_per_zone *mz;
955 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
956 return mz->lru_size[lru];
960 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
961 unsigned int lru_mask)
963 struct mem_cgroup_per_zone *mz;
965 unsigned long ret = 0;
967 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
970 if (BIT(lru) & lru_mask)
971 ret += mz->lru_size[lru];
977 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
978 int nid, unsigned int lru_mask)
983 for (zid = 0; zid < MAX_NR_ZONES; zid++)
984 total += mem_cgroup_zone_nr_lru_pages(memcg,
990 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
991 unsigned int lru_mask)
996 for_each_node_state(nid, N_MEMORY)
997 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1001 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1002 enum mem_cgroup_events_target target)
1004 unsigned long val, next;
1006 val = __this_cpu_read(memcg->stat->nr_page_events);
1007 next = __this_cpu_read(memcg->stat->targets[target]);
1008 /* from time_after() in jiffies.h */
1009 if ((long)next - (long)val < 0) {
1011 case MEM_CGROUP_TARGET_THRESH:
1012 next = val + THRESHOLDS_EVENTS_TARGET;
1014 case MEM_CGROUP_TARGET_SOFTLIMIT:
1015 next = val + SOFTLIMIT_EVENTS_TARGET;
1017 case MEM_CGROUP_TARGET_NUMAINFO:
1018 next = val + NUMAINFO_EVENTS_TARGET;
1023 __this_cpu_write(memcg->stat->targets[target], next);
1030 * Check events in order.
1033 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1036 /* threshold event is triggered in finer grain than soft limit */
1037 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1038 MEM_CGROUP_TARGET_THRESH))) {
1040 bool do_numainfo __maybe_unused;
1042 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1043 MEM_CGROUP_TARGET_SOFTLIMIT);
1044 #if MAX_NUMNODES > 1
1045 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1046 MEM_CGROUP_TARGET_NUMAINFO);
1050 mem_cgroup_threshold(memcg);
1051 if (unlikely(do_softlimit))
1052 mem_cgroup_update_tree(memcg, page);
1053 #if MAX_NUMNODES > 1
1054 if (unlikely(do_numainfo))
1055 atomic_inc(&memcg->numainfo_events);
1061 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1063 return mem_cgroup_from_css(
1064 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1067 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1070 * mm_update_next_owner() may clear mm->owner to NULL
1071 * if it races with swapoff, page migration, etc.
1072 * So this can be called with p == NULL.
1077 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1080 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1082 struct mem_cgroup *memcg = NULL;
1087 * Because we have no locks, mm->owner's may be being moved to other
1088 * cgroup. We use css_tryget() here even if this looks
1089 * pessimistic (rather than adding locks here).
1093 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1094 if (unlikely(!memcg))
1096 } while (!css_tryget(&memcg->css));
1102 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1103 * ref. count) or NULL if the whole root's subtree has been visited.
1105 * helper function to be used by mem_cgroup_iter
1107 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1108 struct mem_cgroup *last_visited)
1110 struct cgroup *prev_cgroup, *next_cgroup;
1113 * Root is not visited by cgroup iterators so it needs an
1119 prev_cgroup = (last_visited == root) ? NULL
1120 : last_visited->css.cgroup;
1122 next_cgroup = cgroup_next_descendant_pre(
1123 prev_cgroup, root->css.cgroup);
1126 * Even if we found a group we have to make sure it is
1127 * alive. css && !memcg means that the groups should be
1128 * skipped and we should continue the tree walk.
1129 * last_visited css is safe to use because it is
1130 * protected by css_get and the tree walk is rcu safe.
1133 struct mem_cgroup *mem = mem_cgroup_from_cont(
1135 if (css_tryget(&mem->css))
1138 prev_cgroup = next_cgroup;
1146 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1149 * When a group in the hierarchy below root is destroyed, the
1150 * hierarchy iterator can no longer be trusted since it might
1151 * have pointed to the destroyed group. Invalidate it.
1153 atomic_inc(&root->dead_count);
1156 static struct mem_cgroup *
1157 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1158 struct mem_cgroup *root,
1161 struct mem_cgroup *position = NULL;
1163 * A cgroup destruction happens in two stages: offlining and
1164 * release. They are separated by a RCU grace period.
1166 * If the iterator is valid, we may still race with an
1167 * offlining. The RCU lock ensures the object won't be
1168 * released, tryget will fail if we lost the race.
1170 *sequence = atomic_read(&root->dead_count);
1171 if (iter->last_dead_count == *sequence) {
1173 position = iter->last_visited;
1174 if (position && !css_tryget(&position->css))
1180 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1181 struct mem_cgroup *last_visited,
1182 struct mem_cgroup *new_position,
1186 css_put(&last_visited->css);
1188 * We store the sequence count from the time @last_visited was
1189 * loaded successfully instead of rereading it here so that we
1190 * don't lose destruction events in between. We could have
1191 * raced with the destruction of @new_position after all.
1193 iter->last_visited = new_position;
1195 iter->last_dead_count = sequence;
1199 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1200 * @root: hierarchy root
1201 * @prev: previously returned memcg, NULL on first invocation
1202 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1204 * Returns references to children of the hierarchy below @root, or
1205 * @root itself, or %NULL after a full round-trip.
1207 * Caller must pass the return value in @prev on subsequent
1208 * invocations for reference counting, or use mem_cgroup_iter_break()
1209 * to cancel a hierarchy walk before the round-trip is complete.
1211 * Reclaimers can specify a zone and a priority level in @reclaim to
1212 * divide up the memcgs in the hierarchy among all concurrent
1213 * reclaimers operating on the same zone and priority.
1215 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1216 struct mem_cgroup *prev,
1217 struct mem_cgroup_reclaim_cookie *reclaim)
1219 struct mem_cgroup *memcg = NULL;
1220 struct mem_cgroup *last_visited = NULL;
1222 if (mem_cgroup_disabled())
1226 root = root_mem_cgroup;
1228 if (prev && !reclaim)
1229 last_visited = prev;
1231 if (!root->use_hierarchy && root != root_mem_cgroup) {
1239 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1240 int uninitialized_var(seq);
1243 int nid = zone_to_nid(reclaim->zone);
1244 int zid = zone_idx(reclaim->zone);
1245 struct mem_cgroup_per_zone *mz;
1247 mz = mem_cgroup_zoneinfo(root, nid, zid);
1248 iter = &mz->reclaim_iter[reclaim->priority];
1249 if (prev && reclaim->generation != iter->generation) {
1250 iter->last_visited = NULL;
1254 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1257 memcg = __mem_cgroup_iter_next(root, last_visited);
1260 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1264 else if (!prev && memcg)
1265 reclaim->generation = iter->generation;
1274 if (prev && prev != root)
1275 css_put(&prev->css);
1281 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1282 * @root: hierarchy root
1283 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1285 void mem_cgroup_iter_break(struct mem_cgroup *root,
1286 struct mem_cgroup *prev)
1289 root = root_mem_cgroup;
1290 if (prev && prev != root)
1291 css_put(&prev->css);
1295 * Iteration constructs for visiting all cgroups (under a tree). If
1296 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1297 * be used for reference counting.
1299 #define for_each_mem_cgroup_tree(iter, root) \
1300 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1302 iter = mem_cgroup_iter(root, iter, NULL))
1304 #define for_each_mem_cgroup(iter) \
1305 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1307 iter = mem_cgroup_iter(NULL, iter, NULL))
1309 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1311 struct mem_cgroup *memcg;
1314 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1315 if (unlikely(!memcg))
1320 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1323 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1331 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1334 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1335 * @zone: zone of the wanted lruvec
1336 * @memcg: memcg of the wanted lruvec
1338 * Returns the lru list vector holding pages for the given @zone and
1339 * @mem. This can be the global zone lruvec, if the memory controller
1342 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1343 struct mem_cgroup *memcg)
1345 struct mem_cgroup_per_zone *mz;
1346 struct lruvec *lruvec;
1348 if (mem_cgroup_disabled()) {
1349 lruvec = &zone->lruvec;
1353 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1354 lruvec = &mz->lruvec;
1357 * Since a node can be onlined after the mem_cgroup was created,
1358 * we have to be prepared to initialize lruvec->zone here;
1359 * and if offlined then reonlined, we need to reinitialize it.
1361 if (unlikely(lruvec->zone != zone))
1362 lruvec->zone = zone;
1367 * Following LRU functions are allowed to be used without PCG_LOCK.
1368 * Operations are called by routine of global LRU independently from memcg.
1369 * What we have to take care of here is validness of pc->mem_cgroup.
1371 * Changes to pc->mem_cgroup happens when
1374 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1375 * It is added to LRU before charge.
1376 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1377 * When moving account, the page is not on LRU. It's isolated.
1381 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1383 * @zone: zone of the page
1385 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1387 struct mem_cgroup_per_zone *mz;
1388 struct mem_cgroup *memcg;
1389 struct page_cgroup *pc;
1390 struct lruvec *lruvec;
1392 if (mem_cgroup_disabled()) {
1393 lruvec = &zone->lruvec;
1397 pc = lookup_page_cgroup(page);
1398 memcg = pc->mem_cgroup;
1401 * Surreptitiously switch any uncharged offlist page to root:
1402 * an uncharged page off lru does nothing to secure
1403 * its former mem_cgroup from sudden removal.
1405 * Our caller holds lru_lock, and PageCgroupUsed is updated
1406 * under page_cgroup lock: between them, they make all uses
1407 * of pc->mem_cgroup safe.
1409 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1410 pc->mem_cgroup = memcg = root_mem_cgroup;
1412 mz = page_cgroup_zoneinfo(memcg, page);
1413 lruvec = &mz->lruvec;
1416 * Since a node can be onlined after the mem_cgroup was created,
1417 * we have to be prepared to initialize lruvec->zone here;
1418 * and if offlined then reonlined, we need to reinitialize it.
1420 if (unlikely(lruvec->zone != zone))
1421 lruvec->zone = zone;
1426 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1427 * @lruvec: mem_cgroup per zone lru vector
1428 * @lru: index of lru list the page is sitting on
1429 * @nr_pages: positive when adding or negative when removing
1431 * This function must be called when a page is added to or removed from an
1434 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1437 struct mem_cgroup_per_zone *mz;
1438 unsigned long *lru_size;
1440 if (mem_cgroup_disabled())
1443 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1444 lru_size = mz->lru_size + lru;
1445 *lru_size += nr_pages;
1446 VM_BUG_ON((long)(*lru_size) < 0);
1450 * Checks whether given mem is same or in the root_mem_cgroup's
1453 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1454 struct mem_cgroup *memcg)
1456 if (root_memcg == memcg)
1458 if (!root_memcg->use_hierarchy || !memcg)
1460 return css_is_ancestor(&memcg->css, &root_memcg->css);
1463 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1464 struct mem_cgroup *memcg)
1469 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1474 bool task_in_mem_cgroup(struct task_struct *task,
1475 const struct mem_cgroup *memcg)
1477 struct mem_cgroup *curr = NULL;
1478 struct task_struct *p;
1481 p = find_lock_task_mm(task);
1483 curr = try_get_mem_cgroup_from_mm(p->mm);
1487 * All threads may have already detached their mm's, but the oom
1488 * killer still needs to detect if they have already been oom
1489 * killed to prevent needlessly killing additional tasks.
1492 curr = mem_cgroup_from_task(task);
1494 css_get(&curr->css);
1500 * We should check use_hierarchy of "memcg" not "curr". Because checking
1501 * use_hierarchy of "curr" here make this function true if hierarchy is
1502 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1503 * hierarchy(even if use_hierarchy is disabled in "memcg").
1505 ret = mem_cgroup_same_or_subtree(memcg, curr);
1506 css_put(&curr->css);
1510 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1512 unsigned long inactive_ratio;
1513 unsigned long inactive;
1514 unsigned long active;
1517 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1518 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1520 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1522 inactive_ratio = int_sqrt(10 * gb);
1526 return inactive * inactive_ratio < active;
1529 #define mem_cgroup_from_res_counter(counter, member) \
1530 container_of(counter, struct mem_cgroup, member)
1533 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1534 * @memcg: the memory cgroup
1536 * Returns the maximum amount of memory @mem can be charged with, in
1539 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1541 unsigned long long margin;
1543 margin = res_counter_margin(&memcg->res);
1544 if (do_swap_account)
1545 margin = min(margin, res_counter_margin(&memcg->memsw));
1546 return margin >> PAGE_SHIFT;
1549 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1551 struct cgroup *cgrp = memcg->css.cgroup;
1554 if (cgrp->parent == NULL)
1555 return vm_swappiness;
1557 return memcg->swappiness;
1561 * memcg->moving_account is used for checking possibility that some thread is
1562 * calling move_account(). When a thread on CPU-A starts moving pages under
1563 * a memcg, other threads should check memcg->moving_account under
1564 * rcu_read_lock(), like this:
1568 * memcg->moving_account+1 if (memcg->mocing_account)
1570 * synchronize_rcu() update something.
1575 /* for quick checking without looking up memcg */
1576 atomic_t memcg_moving __read_mostly;
1578 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1580 atomic_inc(&memcg_moving);
1581 atomic_inc(&memcg->moving_account);
1585 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1588 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1589 * We check NULL in callee rather than caller.
1592 atomic_dec(&memcg_moving);
1593 atomic_dec(&memcg->moving_account);
1598 * 2 routines for checking "mem" is under move_account() or not.
1600 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1601 * is used for avoiding races in accounting. If true,
1602 * pc->mem_cgroup may be overwritten.
1604 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1605 * under hierarchy of moving cgroups. This is for
1606 * waiting at hith-memory prressure caused by "move".
1609 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1611 VM_BUG_ON(!rcu_read_lock_held());
1612 return atomic_read(&memcg->moving_account) > 0;
1615 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1617 struct mem_cgroup *from;
1618 struct mem_cgroup *to;
1621 * Unlike task_move routines, we access mc.to, mc.from not under
1622 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1624 spin_lock(&mc.lock);
1630 ret = mem_cgroup_same_or_subtree(memcg, from)
1631 || mem_cgroup_same_or_subtree(memcg, to);
1633 spin_unlock(&mc.lock);
1637 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1639 if (mc.moving_task && current != mc.moving_task) {
1640 if (mem_cgroup_under_move(memcg)) {
1642 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1643 /* moving charge context might have finished. */
1646 finish_wait(&mc.waitq, &wait);
1654 * Take this lock when
1655 * - a code tries to modify page's memcg while it's USED.
1656 * - a code tries to modify page state accounting in a memcg.
1657 * see mem_cgroup_stolen(), too.
1659 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1660 unsigned long *flags)
1662 spin_lock_irqsave(&memcg->move_lock, *flags);
1665 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1666 unsigned long *flags)
1668 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1671 #define K(x) ((x) << (PAGE_SHIFT-10))
1673 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1674 * @memcg: The memory cgroup that went over limit
1675 * @p: Task that is going to be killed
1677 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1680 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1682 struct cgroup *task_cgrp;
1683 struct cgroup *mem_cgrp;
1685 * Need a buffer in BSS, can't rely on allocations. The code relies
1686 * on the assumption that OOM is serialized for memory controller.
1687 * If this assumption is broken, revisit this code.
1689 static char memcg_name[PATH_MAX];
1691 struct mem_cgroup *iter;
1699 mem_cgrp = memcg->css.cgroup;
1700 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1702 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1705 * Unfortunately, we are unable to convert to a useful name
1706 * But we'll still print out the usage information
1713 pr_info("Task in %s killed", memcg_name);
1716 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1724 * Continues from above, so we don't need an KERN_ level
1726 pr_cont(" as a result of limit of %s\n", memcg_name);
1729 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1730 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1731 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1732 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1733 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1734 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1735 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1736 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1737 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1738 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1739 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1740 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1742 for_each_mem_cgroup_tree(iter, memcg) {
1743 pr_info("Memory cgroup stats");
1746 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1748 pr_cont(" for %s", memcg_name);
1752 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1753 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1755 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1756 K(mem_cgroup_read_stat(iter, i)));
1759 for (i = 0; i < NR_LRU_LISTS; i++)
1760 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1761 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1768 * This function returns the number of memcg under hierarchy tree. Returns
1769 * 1(self count) if no children.
1771 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1774 struct mem_cgroup *iter;
1776 for_each_mem_cgroup_tree(iter, memcg)
1782 * Return the memory (and swap, if configured) limit for a memcg.
1784 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1788 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1791 * Do not consider swap space if we cannot swap due to swappiness
1793 if (mem_cgroup_swappiness(memcg)) {
1796 limit += total_swap_pages << PAGE_SHIFT;
1797 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1800 * If memsw is finite and limits the amount of swap space
1801 * available to this memcg, return that limit.
1803 limit = min(limit, memsw);
1809 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1812 struct mem_cgroup *iter;
1813 unsigned long chosen_points = 0;
1814 unsigned long totalpages;
1815 unsigned int points = 0;
1816 struct task_struct *chosen = NULL;
1819 * If current has a pending SIGKILL or is exiting, then automatically
1820 * select it. The goal is to allow it to allocate so that it may
1821 * quickly exit and free its memory.
1823 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1824 set_thread_flag(TIF_MEMDIE);
1828 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1829 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1830 for_each_mem_cgroup_tree(iter, memcg) {
1831 struct cgroup *cgroup = iter->css.cgroup;
1832 struct cgroup_iter it;
1833 struct task_struct *task;
1835 cgroup_iter_start(cgroup, &it);
1836 while ((task = cgroup_iter_next(cgroup, &it))) {
1837 switch (oom_scan_process_thread(task, totalpages, NULL,
1839 case OOM_SCAN_SELECT:
1841 put_task_struct(chosen);
1843 chosen_points = ULONG_MAX;
1844 get_task_struct(chosen);
1846 case OOM_SCAN_CONTINUE:
1848 case OOM_SCAN_ABORT:
1849 cgroup_iter_end(cgroup, &it);
1850 mem_cgroup_iter_break(memcg, iter);
1852 put_task_struct(chosen);
1857 points = oom_badness(task, memcg, NULL, totalpages);
1858 if (points > chosen_points) {
1860 put_task_struct(chosen);
1862 chosen_points = points;
1863 get_task_struct(chosen);
1866 cgroup_iter_end(cgroup, &it);
1871 points = chosen_points * 1000 / totalpages;
1872 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1873 NULL, "Memory cgroup out of memory");
1876 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1878 unsigned long flags)
1880 unsigned long total = 0;
1881 bool noswap = false;
1884 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1886 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1889 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1891 drain_all_stock_async(memcg);
1892 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1894 * Allow limit shrinkers, which are triggered directly
1895 * by userspace, to catch signals and stop reclaim
1896 * after minimal progress, regardless of the margin.
1898 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1900 if (mem_cgroup_margin(memcg))
1903 * If nothing was reclaimed after two attempts, there
1904 * may be no reclaimable pages in this hierarchy.
1913 * test_mem_cgroup_node_reclaimable
1914 * @memcg: the target memcg
1915 * @nid: the node ID to be checked.
1916 * @noswap : specify true here if the user wants flle only information.
1918 * This function returns whether the specified memcg contains any
1919 * reclaimable pages on a node. Returns true if there are any reclaimable
1920 * pages in the node.
1922 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1923 int nid, bool noswap)
1925 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1927 if (noswap || !total_swap_pages)
1929 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1934 #if MAX_NUMNODES > 1
1937 * Always updating the nodemask is not very good - even if we have an empty
1938 * list or the wrong list here, we can start from some node and traverse all
1939 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1942 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1946 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1947 * pagein/pageout changes since the last update.
1949 if (!atomic_read(&memcg->numainfo_events))
1951 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1954 /* make a nodemask where this memcg uses memory from */
1955 memcg->scan_nodes = node_states[N_MEMORY];
1957 for_each_node_mask(nid, node_states[N_MEMORY]) {
1959 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1960 node_clear(nid, memcg->scan_nodes);
1963 atomic_set(&memcg->numainfo_events, 0);
1964 atomic_set(&memcg->numainfo_updating, 0);
1968 * Selecting a node where we start reclaim from. Because what we need is just
1969 * reducing usage counter, start from anywhere is O,K. Considering
1970 * memory reclaim from current node, there are pros. and cons.
1972 * Freeing memory from current node means freeing memory from a node which
1973 * we'll use or we've used. So, it may make LRU bad. And if several threads
1974 * hit limits, it will see a contention on a node. But freeing from remote
1975 * node means more costs for memory reclaim because of memory latency.
1977 * Now, we use round-robin. Better algorithm is welcomed.
1979 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1983 mem_cgroup_may_update_nodemask(memcg);
1984 node = memcg->last_scanned_node;
1986 node = next_node(node, memcg->scan_nodes);
1987 if (node == MAX_NUMNODES)
1988 node = first_node(memcg->scan_nodes);
1990 * We call this when we hit limit, not when pages are added to LRU.
1991 * No LRU may hold pages because all pages are UNEVICTABLE or
1992 * memcg is too small and all pages are not on LRU. In that case,
1993 * we use curret node.
1995 if (unlikely(node == MAX_NUMNODES))
1996 node = numa_node_id();
1998 memcg->last_scanned_node = node;
2003 * Check all nodes whether it contains reclaimable pages or not.
2004 * For quick scan, we make use of scan_nodes. This will allow us to skip
2005 * unused nodes. But scan_nodes is lazily updated and may not cotain
2006 * enough new information. We need to do double check.
2008 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2013 * quick check...making use of scan_node.
2014 * We can skip unused nodes.
2016 if (!nodes_empty(memcg->scan_nodes)) {
2017 for (nid = first_node(memcg->scan_nodes);
2019 nid = next_node(nid, memcg->scan_nodes)) {
2021 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2026 * Check rest of nodes.
2028 for_each_node_state(nid, N_MEMORY) {
2029 if (node_isset(nid, memcg->scan_nodes))
2031 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2038 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2043 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2045 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2049 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2052 unsigned long *total_scanned)
2054 struct mem_cgroup *victim = NULL;
2057 unsigned long excess;
2058 unsigned long nr_scanned;
2059 struct mem_cgroup_reclaim_cookie reclaim = {
2064 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2067 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2072 * If we have not been able to reclaim
2073 * anything, it might because there are
2074 * no reclaimable pages under this hierarchy
2079 * We want to do more targeted reclaim.
2080 * excess >> 2 is not to excessive so as to
2081 * reclaim too much, nor too less that we keep
2082 * coming back to reclaim from this cgroup
2084 if (total >= (excess >> 2) ||
2085 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2090 if (!mem_cgroup_reclaimable(victim, false))
2092 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2094 *total_scanned += nr_scanned;
2095 if (!res_counter_soft_limit_excess(&root_memcg->res))
2098 mem_cgroup_iter_break(root_memcg, victim);
2103 * Check OOM-Killer is already running under our hierarchy.
2104 * If someone is running, return false.
2105 * Has to be called with memcg_oom_lock
2107 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
2109 struct mem_cgroup *iter, *failed = NULL;
2111 for_each_mem_cgroup_tree(iter, memcg) {
2112 if (iter->oom_lock) {
2114 * this subtree of our hierarchy is already locked
2115 * so we cannot give a lock.
2118 mem_cgroup_iter_break(memcg, iter);
2121 iter->oom_lock = true;
2128 * OK, we failed to lock the whole subtree so we have to clean up
2129 * what we set up to the failing subtree
2131 for_each_mem_cgroup_tree(iter, memcg) {
2132 if (iter == failed) {
2133 mem_cgroup_iter_break(memcg, iter);
2136 iter->oom_lock = false;
2142 * Has to be called with memcg_oom_lock
2144 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2146 struct mem_cgroup *iter;
2148 for_each_mem_cgroup_tree(iter, memcg)
2149 iter->oom_lock = false;
2153 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2155 struct mem_cgroup *iter;
2157 for_each_mem_cgroup_tree(iter, memcg)
2158 atomic_inc(&iter->under_oom);
2161 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2163 struct mem_cgroup *iter;
2166 * When a new child is created while the hierarchy is under oom,
2167 * mem_cgroup_oom_lock() may not be called. We have to use
2168 * atomic_add_unless() here.
2170 for_each_mem_cgroup_tree(iter, memcg)
2171 atomic_add_unless(&iter->under_oom, -1, 0);
2174 static DEFINE_SPINLOCK(memcg_oom_lock);
2175 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2177 struct oom_wait_info {
2178 struct mem_cgroup *memcg;
2182 static int memcg_oom_wake_function(wait_queue_t *wait,
2183 unsigned mode, int sync, void *arg)
2185 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2186 struct mem_cgroup *oom_wait_memcg;
2187 struct oom_wait_info *oom_wait_info;
2189 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2190 oom_wait_memcg = oom_wait_info->memcg;
2193 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2194 * Then we can use css_is_ancestor without taking care of RCU.
2196 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2197 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2199 return autoremove_wake_function(wait, mode, sync, arg);
2202 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2204 /* for filtering, pass "memcg" as argument. */
2205 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2208 static void memcg_oom_recover(struct mem_cgroup *memcg)
2210 if (memcg && atomic_read(&memcg->under_oom))
2211 memcg_wakeup_oom(memcg);
2215 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2217 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2220 struct oom_wait_info owait;
2221 bool locked, need_to_kill;
2223 owait.memcg = memcg;
2224 owait.wait.flags = 0;
2225 owait.wait.func = memcg_oom_wake_function;
2226 owait.wait.private = current;
2227 INIT_LIST_HEAD(&owait.wait.task_list);
2228 need_to_kill = true;
2229 mem_cgroup_mark_under_oom(memcg);
2231 /* At first, try to OOM lock hierarchy under memcg.*/
2232 spin_lock(&memcg_oom_lock);
2233 locked = mem_cgroup_oom_lock(memcg);
2235 * Even if signal_pending(), we can't quit charge() loop without
2236 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2237 * under OOM is always welcomed, use TASK_KILLABLE here.
2239 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2240 if (!locked || memcg->oom_kill_disable)
2241 need_to_kill = false;
2243 mem_cgroup_oom_notify(memcg);
2244 spin_unlock(&memcg_oom_lock);
2247 finish_wait(&memcg_oom_waitq, &owait.wait);
2248 mem_cgroup_out_of_memory(memcg, mask, order);
2251 finish_wait(&memcg_oom_waitq, &owait.wait);
2253 spin_lock(&memcg_oom_lock);
2255 mem_cgroup_oom_unlock(memcg);
2256 memcg_wakeup_oom(memcg);
2257 spin_unlock(&memcg_oom_lock);
2259 mem_cgroup_unmark_under_oom(memcg);
2261 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2263 /* Give chance to dying process */
2264 schedule_timeout_uninterruptible(1);
2269 * Currently used to update mapped file statistics, but the routine can be
2270 * generalized to update other statistics as well.
2272 * Notes: Race condition
2274 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2275 * it tends to be costly. But considering some conditions, we doesn't need
2276 * to do so _always_.
2278 * Considering "charge", lock_page_cgroup() is not required because all
2279 * file-stat operations happen after a page is attached to radix-tree. There
2280 * are no race with "charge".
2282 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2283 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2284 * if there are race with "uncharge". Statistics itself is properly handled
2287 * Considering "move", this is an only case we see a race. To make the race
2288 * small, we check mm->moving_account and detect there are possibility of race
2289 * If there is, we take a lock.
2292 void __mem_cgroup_begin_update_page_stat(struct page *page,
2293 bool *locked, unsigned long *flags)
2295 struct mem_cgroup *memcg;
2296 struct page_cgroup *pc;
2298 pc = lookup_page_cgroup(page);
2300 memcg = pc->mem_cgroup;
2301 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2304 * If this memory cgroup is not under account moving, we don't
2305 * need to take move_lock_mem_cgroup(). Because we already hold
2306 * rcu_read_lock(), any calls to move_account will be delayed until
2307 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2309 if (!mem_cgroup_stolen(memcg))
2312 move_lock_mem_cgroup(memcg, flags);
2313 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2314 move_unlock_mem_cgroup(memcg, flags);
2320 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2322 struct page_cgroup *pc = lookup_page_cgroup(page);
2325 * It's guaranteed that pc->mem_cgroup never changes while
2326 * lock is held because a routine modifies pc->mem_cgroup
2327 * should take move_lock_mem_cgroup().
2329 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2332 void mem_cgroup_update_page_stat(struct page *page,
2333 enum mem_cgroup_page_stat_item idx, int val)
2335 struct mem_cgroup *memcg;
2336 struct page_cgroup *pc = lookup_page_cgroup(page);
2337 unsigned long uninitialized_var(flags);
2339 if (mem_cgroup_disabled())
2342 memcg = pc->mem_cgroup;
2343 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2347 case MEMCG_NR_FILE_MAPPED:
2348 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2354 this_cpu_add(memcg->stat->count[idx], val);
2358 * size of first charge trial. "32" comes from vmscan.c's magic value.
2359 * TODO: maybe necessary to use big numbers in big irons.
2361 #define CHARGE_BATCH 32U
2362 struct memcg_stock_pcp {
2363 struct mem_cgroup *cached; /* this never be root cgroup */
2364 unsigned int nr_pages;
2365 struct work_struct work;
2366 unsigned long flags;
2367 #define FLUSHING_CACHED_CHARGE 0
2369 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2370 static DEFINE_MUTEX(percpu_charge_mutex);
2373 * consume_stock: Try to consume stocked charge on this cpu.
2374 * @memcg: memcg to consume from.
2375 * @nr_pages: how many pages to charge.
2377 * The charges will only happen if @memcg matches the current cpu's memcg
2378 * stock, and at least @nr_pages are available in that stock. Failure to
2379 * service an allocation will refill the stock.
2381 * returns true if successful, false otherwise.
2383 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2385 struct memcg_stock_pcp *stock;
2388 if (nr_pages > CHARGE_BATCH)
2391 stock = &get_cpu_var(memcg_stock);
2392 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2393 stock->nr_pages -= nr_pages;
2394 else /* need to call res_counter_charge */
2396 put_cpu_var(memcg_stock);
2401 * Returns stocks cached in percpu to res_counter and reset cached information.
2403 static void drain_stock(struct memcg_stock_pcp *stock)
2405 struct mem_cgroup *old = stock->cached;
2407 if (stock->nr_pages) {
2408 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2410 res_counter_uncharge(&old->res, bytes);
2411 if (do_swap_account)
2412 res_counter_uncharge(&old->memsw, bytes);
2413 stock->nr_pages = 0;
2415 stock->cached = NULL;
2419 * This must be called under preempt disabled or must be called by
2420 * a thread which is pinned to local cpu.
2422 static void drain_local_stock(struct work_struct *dummy)
2424 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2426 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2429 static void __init memcg_stock_init(void)
2433 for_each_possible_cpu(cpu) {
2434 struct memcg_stock_pcp *stock =
2435 &per_cpu(memcg_stock, cpu);
2436 INIT_WORK(&stock->work, drain_local_stock);
2441 * Cache charges(val) which is from res_counter, to local per_cpu area.
2442 * This will be consumed by consume_stock() function, later.
2444 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2446 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2448 if (stock->cached != memcg) { /* reset if necessary */
2450 stock->cached = memcg;
2452 stock->nr_pages += nr_pages;
2453 put_cpu_var(memcg_stock);
2457 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2458 * of the hierarchy under it. sync flag says whether we should block
2459 * until the work is done.
2461 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2465 /* Notify other cpus that system-wide "drain" is running */
2468 for_each_online_cpu(cpu) {
2469 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2470 struct mem_cgroup *memcg;
2472 memcg = stock->cached;
2473 if (!memcg || !stock->nr_pages)
2475 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2477 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2479 drain_local_stock(&stock->work);
2481 schedule_work_on(cpu, &stock->work);
2489 for_each_online_cpu(cpu) {
2490 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2491 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2492 flush_work(&stock->work);
2499 * Tries to drain stocked charges in other cpus. This function is asynchronous
2500 * and just put a work per cpu for draining localy on each cpu. Caller can
2501 * expects some charges will be back to res_counter later but cannot wait for
2504 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2507 * If someone calls draining, avoid adding more kworker runs.
2509 if (!mutex_trylock(&percpu_charge_mutex))
2511 drain_all_stock(root_memcg, false);
2512 mutex_unlock(&percpu_charge_mutex);
2515 /* This is a synchronous drain interface. */
2516 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2518 /* called when force_empty is called */
2519 mutex_lock(&percpu_charge_mutex);
2520 drain_all_stock(root_memcg, true);
2521 mutex_unlock(&percpu_charge_mutex);
2525 * This function drains percpu counter value from DEAD cpu and
2526 * move it to local cpu. Note that this function can be preempted.
2528 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2532 spin_lock(&memcg->pcp_counter_lock);
2533 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2534 long x = per_cpu(memcg->stat->count[i], cpu);
2536 per_cpu(memcg->stat->count[i], cpu) = 0;
2537 memcg->nocpu_base.count[i] += x;
2539 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2540 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2542 per_cpu(memcg->stat->events[i], cpu) = 0;
2543 memcg->nocpu_base.events[i] += x;
2545 spin_unlock(&memcg->pcp_counter_lock);
2548 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2549 unsigned long action,
2552 int cpu = (unsigned long)hcpu;
2553 struct memcg_stock_pcp *stock;
2554 struct mem_cgroup *iter;
2556 if (action == CPU_ONLINE)
2559 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2562 for_each_mem_cgroup(iter)
2563 mem_cgroup_drain_pcp_counter(iter, cpu);
2565 stock = &per_cpu(memcg_stock, cpu);
2571 /* See __mem_cgroup_try_charge() for details */
2573 CHARGE_OK, /* success */
2574 CHARGE_RETRY, /* need to retry but retry is not bad */
2575 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2576 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2577 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2580 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2581 unsigned int nr_pages, unsigned int min_pages,
2584 unsigned long csize = nr_pages * PAGE_SIZE;
2585 struct mem_cgroup *mem_over_limit;
2586 struct res_counter *fail_res;
2587 unsigned long flags = 0;
2590 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2593 if (!do_swap_account)
2595 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2599 res_counter_uncharge(&memcg->res, csize);
2600 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2601 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2603 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2605 * Never reclaim on behalf of optional batching, retry with a
2606 * single page instead.
2608 if (nr_pages > min_pages)
2609 return CHARGE_RETRY;
2611 if (!(gfp_mask & __GFP_WAIT))
2612 return CHARGE_WOULDBLOCK;
2614 if (gfp_mask & __GFP_NORETRY)
2615 return CHARGE_NOMEM;
2617 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2618 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2619 return CHARGE_RETRY;
2621 * Even though the limit is exceeded at this point, reclaim
2622 * may have been able to free some pages. Retry the charge
2623 * before killing the task.
2625 * Only for regular pages, though: huge pages are rather
2626 * unlikely to succeed so close to the limit, and we fall back
2627 * to regular pages anyway in case of failure.
2629 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2630 return CHARGE_RETRY;
2633 * At task move, charge accounts can be doubly counted. So, it's
2634 * better to wait until the end of task_move if something is going on.
2636 if (mem_cgroup_wait_acct_move(mem_over_limit))
2637 return CHARGE_RETRY;
2639 /* If we don't need to call oom-killer at el, return immediately */
2641 return CHARGE_NOMEM;
2643 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2644 return CHARGE_OOM_DIE;
2646 return CHARGE_RETRY;
2650 * __mem_cgroup_try_charge() does
2651 * 1. detect memcg to be charged against from passed *mm and *ptr,
2652 * 2. update res_counter
2653 * 3. call memory reclaim if necessary.
2655 * In some special case, if the task is fatal, fatal_signal_pending() or
2656 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2657 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2658 * as possible without any hazards. 2: all pages should have a valid
2659 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2660 * pointer, that is treated as a charge to root_mem_cgroup.
2662 * So __mem_cgroup_try_charge() will return
2663 * 0 ... on success, filling *ptr with a valid memcg pointer.
2664 * -ENOMEM ... charge failure because of resource limits.
2665 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2667 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2668 * the oom-killer can be invoked.
2670 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2672 unsigned int nr_pages,
2673 struct mem_cgroup **ptr,
2676 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2677 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2678 struct mem_cgroup *memcg = NULL;
2682 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2683 * in system level. So, allow to go ahead dying process in addition to
2686 if (unlikely(test_thread_flag(TIF_MEMDIE)
2687 || fatal_signal_pending(current)))
2691 * We always charge the cgroup the mm_struct belongs to.
2692 * The mm_struct's mem_cgroup changes on task migration if the
2693 * thread group leader migrates. It's possible that mm is not
2694 * set, if so charge the root memcg (happens for pagecache usage).
2697 *ptr = root_mem_cgroup;
2699 if (*ptr) { /* css should be a valid one */
2701 if (mem_cgroup_is_root(memcg))
2703 if (consume_stock(memcg, nr_pages))
2705 css_get(&memcg->css);
2707 struct task_struct *p;
2710 p = rcu_dereference(mm->owner);
2712 * Because we don't have task_lock(), "p" can exit.
2713 * In that case, "memcg" can point to root or p can be NULL with
2714 * race with swapoff. Then, we have small risk of mis-accouning.
2715 * But such kind of mis-account by race always happens because
2716 * we don't have cgroup_mutex(). It's overkill and we allo that
2718 * (*) swapoff at el will charge against mm-struct not against
2719 * task-struct. So, mm->owner can be NULL.
2721 memcg = mem_cgroup_from_task(p);
2723 memcg = root_mem_cgroup;
2724 if (mem_cgroup_is_root(memcg)) {
2728 if (consume_stock(memcg, nr_pages)) {
2730 * It seems dagerous to access memcg without css_get().
2731 * But considering how consume_stok works, it's not
2732 * necessary. If consume_stock success, some charges
2733 * from this memcg are cached on this cpu. So, we
2734 * don't need to call css_get()/css_tryget() before
2735 * calling consume_stock().
2740 /* after here, we may be blocked. we need to get refcnt */
2741 if (!css_tryget(&memcg->css)) {
2751 /* If killed, bypass charge */
2752 if (fatal_signal_pending(current)) {
2753 css_put(&memcg->css);
2758 if (oom && !nr_oom_retries) {
2760 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2763 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2768 case CHARGE_RETRY: /* not in OOM situation but retry */
2770 css_put(&memcg->css);
2773 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2774 css_put(&memcg->css);
2776 case CHARGE_NOMEM: /* OOM routine works */
2778 css_put(&memcg->css);
2781 /* If oom, we never return -ENOMEM */
2784 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2785 css_put(&memcg->css);
2788 } while (ret != CHARGE_OK);
2790 if (batch > nr_pages)
2791 refill_stock(memcg, batch - nr_pages);
2792 css_put(&memcg->css);
2800 *ptr = root_mem_cgroup;
2805 * Somemtimes we have to undo a charge we got by try_charge().
2806 * This function is for that and do uncharge, put css's refcnt.
2807 * gotten by try_charge().
2809 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2810 unsigned int nr_pages)
2812 if (!mem_cgroup_is_root(memcg)) {
2813 unsigned long bytes = nr_pages * PAGE_SIZE;
2815 res_counter_uncharge(&memcg->res, bytes);
2816 if (do_swap_account)
2817 res_counter_uncharge(&memcg->memsw, bytes);
2822 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2823 * This is useful when moving usage to parent cgroup.
2825 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2826 unsigned int nr_pages)
2828 unsigned long bytes = nr_pages * PAGE_SIZE;
2830 if (mem_cgroup_is_root(memcg))
2833 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2834 if (do_swap_account)
2835 res_counter_uncharge_until(&memcg->memsw,
2836 memcg->memsw.parent, bytes);
2840 * A helper function to get mem_cgroup from ID. must be called under
2841 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2842 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2843 * called against removed memcg.)
2845 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2847 struct cgroup_subsys_state *css;
2849 /* ID 0 is unused ID */
2852 css = css_lookup(&mem_cgroup_subsys, id);
2855 return mem_cgroup_from_css(css);
2858 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2860 struct mem_cgroup *memcg = NULL;
2861 struct page_cgroup *pc;
2865 VM_BUG_ON(!PageLocked(page));
2867 pc = lookup_page_cgroup(page);
2868 lock_page_cgroup(pc);
2869 if (PageCgroupUsed(pc)) {
2870 memcg = pc->mem_cgroup;
2871 if (memcg && !css_tryget(&memcg->css))
2873 } else if (PageSwapCache(page)) {
2874 ent.val = page_private(page);
2875 id = lookup_swap_cgroup_id(ent);
2877 memcg = mem_cgroup_lookup(id);
2878 if (memcg && !css_tryget(&memcg->css))
2882 unlock_page_cgroup(pc);
2886 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2888 unsigned int nr_pages,
2889 enum charge_type ctype,
2892 struct page_cgroup *pc = lookup_page_cgroup(page);
2893 struct zone *uninitialized_var(zone);
2894 struct lruvec *lruvec;
2895 bool was_on_lru = false;
2898 lock_page_cgroup(pc);
2899 VM_BUG_ON(PageCgroupUsed(pc));
2901 * we don't need page_cgroup_lock about tail pages, becase they are not
2902 * accessed by any other context at this point.
2906 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2907 * may already be on some other mem_cgroup's LRU. Take care of it.
2910 zone = page_zone(page);
2911 spin_lock_irq(&zone->lru_lock);
2912 if (PageLRU(page)) {
2913 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2915 del_page_from_lru_list(page, lruvec, page_lru(page));
2920 pc->mem_cgroup = memcg;
2922 * We access a page_cgroup asynchronously without lock_page_cgroup().
2923 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2924 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2925 * before USED bit, we need memory barrier here.
2926 * See mem_cgroup_add_lru_list(), etc.
2929 SetPageCgroupUsed(pc);
2933 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2934 VM_BUG_ON(PageLRU(page));
2936 add_page_to_lru_list(page, lruvec, page_lru(page));
2938 spin_unlock_irq(&zone->lru_lock);
2941 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2946 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2947 unlock_page_cgroup(pc);
2950 * "charge_statistics" updated event counter. Then, check it.
2951 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2952 * if they exceeds softlimit.
2954 memcg_check_events(memcg, page);
2957 static DEFINE_MUTEX(set_limit_mutex);
2959 #ifdef CONFIG_MEMCG_KMEM
2960 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2962 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2963 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2967 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2968 * in the memcg_cache_params struct.
2970 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2972 struct kmem_cache *cachep;
2974 VM_BUG_ON(p->is_root_cache);
2975 cachep = p->root_cache;
2976 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2979 #ifdef CONFIG_SLABINFO
2980 static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2983 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2984 struct memcg_cache_params *params;
2986 if (!memcg_can_account_kmem(memcg))
2989 print_slabinfo_header(m);
2991 mutex_lock(&memcg->slab_caches_mutex);
2992 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2993 cache_show(memcg_params_to_cache(params), m);
2994 mutex_unlock(&memcg->slab_caches_mutex);
3000 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3002 struct res_counter *fail_res;
3003 struct mem_cgroup *_memcg;
3007 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3012 * Conditions under which we can wait for the oom_killer. Those are
3013 * the same conditions tested by the core page allocator
3015 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
3018 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3021 if (ret == -EINTR) {
3023 * __mem_cgroup_try_charge() chosed to bypass to root due to
3024 * OOM kill or fatal signal. Since our only options are to
3025 * either fail the allocation or charge it to this cgroup, do
3026 * it as a temporary condition. But we can't fail. From a
3027 * kmem/slab perspective, the cache has already been selected,
3028 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3031 * This condition will only trigger if the task entered
3032 * memcg_charge_kmem in a sane state, but was OOM-killed during
3033 * __mem_cgroup_try_charge() above. Tasks that were already
3034 * dying when the allocation triggers should have been already
3035 * directed to the root cgroup in memcontrol.h
3037 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3038 if (do_swap_account)
3039 res_counter_charge_nofail(&memcg->memsw, size,
3043 res_counter_uncharge(&memcg->kmem, size);
3048 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3050 res_counter_uncharge(&memcg->res, size);
3051 if (do_swap_account)
3052 res_counter_uncharge(&memcg->memsw, size);
3055 if (res_counter_uncharge(&memcg->kmem, size))
3059 * Releases a reference taken in kmem_cgroup_css_offline in case
3060 * this last uncharge is racing with the offlining code or it is
3061 * outliving the memcg existence.
3063 * The memory barrier imposed by test&clear is paired with the
3064 * explicit one in memcg_kmem_mark_dead().
3066 if (memcg_kmem_test_and_clear_dead(memcg))
3067 css_put(&memcg->css);
3070 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3075 mutex_lock(&memcg->slab_caches_mutex);
3076 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3077 mutex_unlock(&memcg->slab_caches_mutex);
3081 * helper for acessing a memcg's index. It will be used as an index in the
3082 * child cache array in kmem_cache, and also to derive its name. This function
3083 * will return -1 when this is not a kmem-limited memcg.
3085 int memcg_cache_id(struct mem_cgroup *memcg)
3087 return memcg ? memcg->kmemcg_id : -1;
3091 * This ends up being protected by the set_limit mutex, during normal
3092 * operation, because that is its main call site.
3094 * But when we create a new cache, we can call this as well if its parent
3095 * is kmem-limited. That will have to hold set_limit_mutex as well.
3097 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3101 num = ida_simple_get(&kmem_limited_groups,
3102 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3106 * After this point, kmem_accounted (that we test atomically in
3107 * the beginning of this conditional), is no longer 0. This
3108 * guarantees only one process will set the following boolean
3109 * to true. We don't need test_and_set because we're protected
3110 * by the set_limit_mutex anyway.
3112 memcg_kmem_set_activated(memcg);
3114 ret = memcg_update_all_caches(num+1);
3116 ida_simple_remove(&kmem_limited_groups, num);
3117 memcg_kmem_clear_activated(memcg);
3121 memcg->kmemcg_id = num;
3122 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3123 mutex_init(&memcg->slab_caches_mutex);
3127 static size_t memcg_caches_array_size(int num_groups)
3130 if (num_groups <= 0)
3133 size = 2 * num_groups;
3134 if (size < MEMCG_CACHES_MIN_SIZE)
3135 size = MEMCG_CACHES_MIN_SIZE;
3136 else if (size > MEMCG_CACHES_MAX_SIZE)
3137 size = MEMCG_CACHES_MAX_SIZE;
3143 * We should update the current array size iff all caches updates succeed. This
3144 * can only be done from the slab side. The slab mutex needs to be held when
3147 void memcg_update_array_size(int num)
3149 if (num > memcg_limited_groups_array_size)
3150 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3153 static void kmem_cache_destroy_work_func(struct work_struct *w);
3155 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3157 struct memcg_cache_params *cur_params = s->memcg_params;
3159 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3161 if (num_groups > memcg_limited_groups_array_size) {
3163 ssize_t size = memcg_caches_array_size(num_groups);
3165 size *= sizeof(void *);
3166 size += sizeof(struct memcg_cache_params);
3168 s->memcg_params = kzalloc(size, GFP_KERNEL);
3169 if (!s->memcg_params) {
3170 s->memcg_params = cur_params;
3174 s->memcg_params->is_root_cache = true;
3177 * There is the chance it will be bigger than
3178 * memcg_limited_groups_array_size, if we failed an allocation
3179 * in a cache, in which case all caches updated before it, will
3180 * have a bigger array.
3182 * But if that is the case, the data after
3183 * memcg_limited_groups_array_size is certainly unused
3185 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3186 if (!cur_params->memcg_caches[i])
3188 s->memcg_params->memcg_caches[i] =
3189 cur_params->memcg_caches[i];
3193 * Ideally, we would wait until all caches succeed, and only
3194 * then free the old one. But this is not worth the extra
3195 * pointer per-cache we'd have to have for this.
3197 * It is not a big deal if some caches are left with a size
3198 * bigger than the others. And all updates will reset this
3206 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3207 struct kmem_cache *root_cache)
3209 size_t size = sizeof(struct memcg_cache_params);
3211 if (!memcg_kmem_enabled())
3215 size += memcg_limited_groups_array_size * sizeof(void *);
3217 s->memcg_params = kzalloc(size, GFP_KERNEL);
3218 if (!s->memcg_params)
3221 INIT_WORK(&s->memcg_params->destroy,
3222 kmem_cache_destroy_work_func);
3224 s->memcg_params->memcg = memcg;
3225 s->memcg_params->root_cache = root_cache;
3227 s->memcg_params->is_root_cache = true;
3232 void memcg_release_cache(struct kmem_cache *s)
3234 struct kmem_cache *root;
3235 struct mem_cgroup *memcg;
3239 * This happens, for instance, when a root cache goes away before we
3242 if (!s->memcg_params)
3245 if (s->memcg_params->is_root_cache)
3248 memcg = s->memcg_params->memcg;
3249 id = memcg_cache_id(memcg);
3251 root = s->memcg_params->root_cache;
3252 root->memcg_params->memcg_caches[id] = NULL;
3254 mutex_lock(&memcg->slab_caches_mutex);
3255 list_del(&s->memcg_params->list);
3256 mutex_unlock(&memcg->slab_caches_mutex);
3258 css_put(&memcg->css);
3260 kfree(s->memcg_params);
3264 * During the creation a new cache, we need to disable our accounting mechanism
3265 * altogether. This is true even if we are not creating, but rather just
3266 * enqueing new caches to be created.
3268 * This is because that process will trigger allocations; some visible, like
3269 * explicit kmallocs to auxiliary data structures, name strings and internal
3270 * cache structures; some well concealed, like INIT_WORK() that can allocate
3271 * objects during debug.
3273 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3274 * to it. This may not be a bounded recursion: since the first cache creation
3275 * failed to complete (waiting on the allocation), we'll just try to create the
3276 * cache again, failing at the same point.
3278 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3279 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3280 * inside the following two functions.
3282 static inline void memcg_stop_kmem_account(void)
3284 VM_BUG_ON(!current->mm);
3285 current->memcg_kmem_skip_account++;
3288 static inline void memcg_resume_kmem_account(void)
3290 VM_BUG_ON(!current->mm);
3291 current->memcg_kmem_skip_account--;
3294 static void kmem_cache_destroy_work_func(struct work_struct *w)
3296 struct kmem_cache *cachep;
3297 struct memcg_cache_params *p;
3299 p = container_of(w, struct memcg_cache_params, destroy);
3301 cachep = memcg_params_to_cache(p);
3304 * If we get down to 0 after shrink, we could delete right away.
3305 * However, memcg_release_pages() already puts us back in the workqueue
3306 * in that case. If we proceed deleting, we'll get a dangling
3307 * reference, and removing the object from the workqueue in that case
3308 * is unnecessary complication. We are not a fast path.
3310 * Note that this case is fundamentally different from racing with
3311 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3312 * kmem_cache_shrink, not only we would be reinserting a dead cache
3313 * into the queue, but doing so from inside the worker racing to
3316 * So if we aren't down to zero, we'll just schedule a worker and try
3319 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3320 kmem_cache_shrink(cachep);
3321 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3324 kmem_cache_destroy(cachep);
3327 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3329 if (!cachep->memcg_params->dead)
3333 * There are many ways in which we can get here.
3335 * We can get to a memory-pressure situation while the delayed work is
3336 * still pending to run. The vmscan shrinkers can then release all
3337 * cache memory and get us to destruction. If this is the case, we'll
3338 * be executed twice, which is a bug (the second time will execute over
3339 * bogus data). In this case, cancelling the work should be fine.
3341 * But we can also get here from the worker itself, if
3342 * kmem_cache_shrink is enough to shake all the remaining objects and
3343 * get the page count to 0. In this case, we'll deadlock if we try to
3344 * cancel the work (the worker runs with an internal lock held, which
3345 * is the same lock we would hold for cancel_work_sync().)
3347 * Since we can't possibly know who got us here, just refrain from
3348 * running if there is already work pending
3350 if (work_pending(&cachep->memcg_params->destroy))
3353 * We have to defer the actual destroying to a workqueue, because
3354 * we might currently be in a context that cannot sleep.
3356 schedule_work(&cachep->memcg_params->destroy);
3360 * This lock protects updaters, not readers. We want readers to be as fast as
3361 * they can, and they will either see NULL or a valid cache value. Our model
3362 * allow them to see NULL, in which case the root memcg will be selected.
3364 * We need this lock because multiple allocations to the same cache from a non
3365 * will span more than one worker. Only one of them can create the cache.
3367 static DEFINE_MUTEX(memcg_cache_mutex);
3370 * Called with memcg_cache_mutex held
3372 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3373 struct kmem_cache *s)
3375 struct kmem_cache *new;
3376 static char *tmp_name = NULL;
3378 lockdep_assert_held(&memcg_cache_mutex);
3381 * kmem_cache_create_memcg duplicates the given name and
3382 * cgroup_name for this name requires RCU context.
3383 * This static temporary buffer is used to prevent from
3384 * pointless shortliving allocation.
3387 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3393 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3394 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3397 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3398 (s->flags & ~SLAB_PANIC), s->ctor, s);
3401 new->allocflags |= __GFP_KMEMCG;
3406 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3407 struct kmem_cache *cachep)
3409 struct kmem_cache *new_cachep;
3412 BUG_ON(!memcg_can_account_kmem(memcg));
3414 idx = memcg_cache_id(memcg);
3416 mutex_lock(&memcg_cache_mutex);
3417 new_cachep = cachep->memcg_params->memcg_caches[idx];
3419 css_put(&memcg->css);
3423 new_cachep = kmem_cache_dup(memcg, cachep);
3424 if (new_cachep == NULL) {
3425 new_cachep = cachep;
3426 css_put(&memcg->css);
3430 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3432 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3434 * the readers won't lock, make sure everybody sees the updated value,
3435 * so they won't put stuff in the queue again for no reason
3439 mutex_unlock(&memcg_cache_mutex);
3443 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3445 struct kmem_cache *c;
3448 if (!s->memcg_params)
3450 if (!s->memcg_params->is_root_cache)
3454 * If the cache is being destroyed, we trust that there is no one else
3455 * requesting objects from it. Even if there are, the sanity checks in
3456 * kmem_cache_destroy should caught this ill-case.
3458 * Still, we don't want anyone else freeing memcg_caches under our
3459 * noses, which can happen if a new memcg comes to life. As usual,
3460 * we'll take the set_limit_mutex to protect ourselves against this.
3462 mutex_lock(&set_limit_mutex);
3463 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3464 c = s->memcg_params->memcg_caches[i];
3469 * We will now manually delete the caches, so to avoid races
3470 * we need to cancel all pending destruction workers and
3471 * proceed with destruction ourselves.
3473 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3474 * and that could spawn the workers again: it is likely that
3475 * the cache still have active pages until this very moment.
3476 * This would lead us back to mem_cgroup_destroy_cache.
3478 * But that will not execute at all if the "dead" flag is not
3479 * set, so flip it down to guarantee we are in control.
3481 c->memcg_params->dead = false;
3482 cancel_work_sync(&c->memcg_params->destroy);
3483 kmem_cache_destroy(c);
3485 mutex_unlock(&set_limit_mutex);
3488 struct create_work {
3489 struct mem_cgroup *memcg;
3490 struct kmem_cache *cachep;
3491 struct work_struct work;
3494 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3496 struct kmem_cache *cachep;
3497 struct memcg_cache_params *params;
3499 if (!memcg_kmem_is_active(memcg))
3502 mutex_lock(&memcg->slab_caches_mutex);
3503 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3504 cachep = memcg_params_to_cache(params);
3505 cachep->memcg_params->dead = true;
3506 schedule_work(&cachep->memcg_params->destroy);
3508 mutex_unlock(&memcg->slab_caches_mutex);
3511 static void memcg_create_cache_work_func(struct work_struct *w)
3513 struct create_work *cw;
3515 cw = container_of(w, struct create_work, work);
3516 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3521 * Enqueue the creation of a per-memcg kmem_cache.
3523 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3524 struct kmem_cache *cachep)
3526 struct create_work *cw;
3528 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3530 css_put(&memcg->css);
3535 cw->cachep = cachep;
3537 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3538 schedule_work(&cw->work);
3541 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3542 struct kmem_cache *cachep)
3545 * We need to stop accounting when we kmalloc, because if the
3546 * corresponding kmalloc cache is not yet created, the first allocation
3547 * in __memcg_create_cache_enqueue will recurse.
3549 * However, it is better to enclose the whole function. Depending on
3550 * the debugging options enabled, INIT_WORK(), for instance, can
3551 * trigger an allocation. This too, will make us recurse. Because at
3552 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3553 * the safest choice is to do it like this, wrapping the whole function.
3555 memcg_stop_kmem_account();
3556 __memcg_create_cache_enqueue(memcg, cachep);
3557 memcg_resume_kmem_account();
3560 * Return the kmem_cache we're supposed to use for a slab allocation.
3561 * We try to use the current memcg's version of the cache.
3563 * If the cache does not exist yet, if we are the first user of it,
3564 * we either create it immediately, if possible, or create it asynchronously
3566 * In the latter case, we will let the current allocation go through with
3567 * the original cache.
3569 * Can't be called in interrupt context or from kernel threads.
3570 * This function needs to be called with rcu_read_lock() held.
3572 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3575 struct mem_cgroup *memcg;
3578 VM_BUG_ON(!cachep->memcg_params);
3579 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3581 if (!current->mm || current->memcg_kmem_skip_account)
3585 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3587 if (!memcg_can_account_kmem(memcg))
3590 idx = memcg_cache_id(memcg);
3593 * barrier to mare sure we're always seeing the up to date value. The
3594 * code updating memcg_caches will issue a write barrier to match this.
3596 read_barrier_depends();
3597 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3598 cachep = cachep->memcg_params->memcg_caches[idx];
3602 /* The corresponding put will be done in the workqueue. */
3603 if (!css_tryget(&memcg->css))
3608 * If we are in a safe context (can wait, and not in interrupt
3609 * context), we could be be predictable and return right away.
3610 * This would guarantee that the allocation being performed
3611 * already belongs in the new cache.
3613 * However, there are some clashes that can arrive from locking.
3614 * For instance, because we acquire the slab_mutex while doing
3615 * kmem_cache_dup, this means no further allocation could happen
3616 * with the slab_mutex held.
3618 * Also, because cache creation issue get_online_cpus(), this
3619 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3620 * that ends up reversed during cpu hotplug. (cpuset allocates
3621 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3622 * better to defer everything.
3624 memcg_create_cache_enqueue(memcg, cachep);
3630 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3633 * We need to verify if the allocation against current->mm->owner's memcg is
3634 * possible for the given order. But the page is not allocated yet, so we'll
3635 * need a further commit step to do the final arrangements.
3637 * It is possible for the task to switch cgroups in this mean time, so at
3638 * commit time, we can't rely on task conversion any longer. We'll then use
3639 * the handle argument to return to the caller which cgroup we should commit
3640 * against. We could also return the memcg directly and avoid the pointer
3641 * passing, but a boolean return value gives better semantics considering
3642 * the compiled-out case as well.
3644 * Returning true means the allocation is possible.
3647 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3649 struct mem_cgroup *memcg;
3655 * Disabling accounting is only relevant for some specific memcg
3656 * internal allocations. Therefore we would initially not have such
3657 * check here, since direct calls to the page allocator that are marked
3658 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3659 * concerned with cache allocations, and by having this test at
3660 * memcg_kmem_get_cache, we are already able to relay the allocation to
3661 * the root cache and bypass the memcg cache altogether.
3663 * There is one exception, though: the SLUB allocator does not create
3664 * large order caches, but rather service large kmallocs directly from
3665 * the page allocator. Therefore, the following sequence when backed by
3666 * the SLUB allocator:
3668 * memcg_stop_kmem_account();
3669 * kmalloc(<large_number>)
3670 * memcg_resume_kmem_account();
3672 * would effectively ignore the fact that we should skip accounting,
3673 * since it will drive us directly to this function without passing
3674 * through the cache selector memcg_kmem_get_cache. Such large
3675 * allocations are extremely rare but can happen, for instance, for the
3676 * cache arrays. We bring this test here.
3678 if (!current->mm || current->memcg_kmem_skip_account)
3681 memcg = try_get_mem_cgroup_from_mm(current->mm);
3684 * very rare case described in mem_cgroup_from_task. Unfortunately there
3685 * isn't much we can do without complicating this too much, and it would
3686 * be gfp-dependent anyway. Just let it go
3688 if (unlikely(!memcg))
3691 if (!memcg_can_account_kmem(memcg)) {
3692 css_put(&memcg->css);
3696 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3700 css_put(&memcg->css);
3704 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3707 struct page_cgroup *pc;
3709 VM_BUG_ON(mem_cgroup_is_root(memcg));
3711 /* The page allocation failed. Revert */
3713 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3717 pc = lookup_page_cgroup(page);
3718 lock_page_cgroup(pc);
3719 pc->mem_cgroup = memcg;
3720 SetPageCgroupUsed(pc);
3721 unlock_page_cgroup(pc);
3724 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3726 struct mem_cgroup *memcg = NULL;
3727 struct page_cgroup *pc;
3730 pc = lookup_page_cgroup(page);
3732 * Fast unlocked return. Theoretically might have changed, have to
3733 * check again after locking.
3735 if (!PageCgroupUsed(pc))
3738 lock_page_cgroup(pc);
3739 if (PageCgroupUsed(pc)) {
3740 memcg = pc->mem_cgroup;
3741 ClearPageCgroupUsed(pc);
3743 unlock_page_cgroup(pc);
3746 * We trust that only if there is a memcg associated with the page, it
3747 * is a valid allocation
3752 VM_BUG_ON(mem_cgroup_is_root(memcg));
3753 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3756 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3759 #endif /* CONFIG_MEMCG_KMEM */
3761 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3763 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3765 * Because tail pages are not marked as "used", set it. We're under
3766 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3767 * charge/uncharge will be never happen and move_account() is done under
3768 * compound_lock(), so we don't have to take care of races.
3770 void mem_cgroup_split_huge_fixup(struct page *head)
3772 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3773 struct page_cgroup *pc;
3774 struct mem_cgroup *memcg;
3777 if (mem_cgroup_disabled())
3780 memcg = head_pc->mem_cgroup;
3781 for (i = 1; i < HPAGE_PMD_NR; i++) {
3783 pc->mem_cgroup = memcg;
3784 smp_wmb();/* see __commit_charge() */
3785 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3787 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3790 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3793 * mem_cgroup_move_account - move account of the page
3795 * @nr_pages: number of regular pages (>1 for huge pages)
3796 * @pc: page_cgroup of the page.
3797 * @from: mem_cgroup which the page is moved from.
3798 * @to: mem_cgroup which the page is moved to. @from != @to.
3800 * The caller must confirm following.
3801 * - page is not on LRU (isolate_page() is useful.)
3802 * - compound_lock is held when nr_pages > 1
3804 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3807 static int mem_cgroup_move_account(struct page *page,
3808 unsigned int nr_pages,
3809 struct page_cgroup *pc,
3810 struct mem_cgroup *from,
3811 struct mem_cgroup *to)
3813 unsigned long flags;
3815 bool anon = PageAnon(page);
3817 VM_BUG_ON(from == to);
3818 VM_BUG_ON(PageLRU(page));
3820 * The page is isolated from LRU. So, collapse function
3821 * will not handle this page. But page splitting can happen.
3822 * Do this check under compound_page_lock(). The caller should
3826 if (nr_pages > 1 && !PageTransHuge(page))
3829 lock_page_cgroup(pc);
3832 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3835 move_lock_mem_cgroup(from, &flags);
3837 if (!anon && page_mapped(page)) {
3838 /* Update mapped_file data for mem_cgroup */
3840 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3841 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3844 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3846 /* caller should have done css_get */
3847 pc->mem_cgroup = to;
3848 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3849 move_unlock_mem_cgroup(from, &flags);
3852 unlock_page_cgroup(pc);
3856 memcg_check_events(to, page);
3857 memcg_check_events(from, page);
3863 * mem_cgroup_move_parent - moves page to the parent group
3864 * @page: the page to move
3865 * @pc: page_cgroup of the page
3866 * @child: page's cgroup
3868 * move charges to its parent or the root cgroup if the group has no
3869 * parent (aka use_hierarchy==0).
3870 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3871 * mem_cgroup_move_account fails) the failure is always temporary and
3872 * it signals a race with a page removal/uncharge or migration. In the
3873 * first case the page is on the way out and it will vanish from the LRU
3874 * on the next attempt and the call should be retried later.
3875 * Isolation from the LRU fails only if page has been isolated from
3876 * the LRU since we looked at it and that usually means either global
3877 * reclaim or migration going on. The page will either get back to the
3879 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3880 * (!PageCgroupUsed) or moved to a different group. The page will
3881 * disappear in the next attempt.
3883 static int mem_cgroup_move_parent(struct page *page,
3884 struct page_cgroup *pc,
3885 struct mem_cgroup *child)
3887 struct mem_cgroup *parent;
3888 unsigned int nr_pages;
3889 unsigned long uninitialized_var(flags);
3892 VM_BUG_ON(mem_cgroup_is_root(child));
3895 if (!get_page_unless_zero(page))
3897 if (isolate_lru_page(page))
3900 nr_pages = hpage_nr_pages(page);
3902 parent = parent_mem_cgroup(child);
3904 * If no parent, move charges to root cgroup.
3907 parent = root_mem_cgroup;
3910 VM_BUG_ON(!PageTransHuge(page));
3911 flags = compound_lock_irqsave(page);
3914 ret = mem_cgroup_move_account(page, nr_pages,
3917 __mem_cgroup_cancel_local_charge(child, nr_pages);
3920 compound_unlock_irqrestore(page, flags);
3921 putback_lru_page(page);
3929 * Charge the memory controller for page usage.
3931 * 0 if the charge was successful
3932 * < 0 if the cgroup is over its limit
3934 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3935 gfp_t gfp_mask, enum charge_type ctype)
3937 struct mem_cgroup *memcg = NULL;
3938 unsigned int nr_pages = 1;
3942 if (PageTransHuge(page)) {
3943 nr_pages <<= compound_order(page);
3944 VM_BUG_ON(!PageTransHuge(page));
3946 * Never OOM-kill a process for a huge page. The
3947 * fault handler will fall back to regular pages.
3952 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3955 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3959 int mem_cgroup_newpage_charge(struct page *page,
3960 struct mm_struct *mm, gfp_t gfp_mask)
3962 if (mem_cgroup_disabled())
3964 VM_BUG_ON(page_mapped(page));
3965 VM_BUG_ON(page->mapping && !PageAnon(page));
3967 return mem_cgroup_charge_common(page, mm, gfp_mask,
3968 MEM_CGROUP_CHARGE_TYPE_ANON);
3972 * While swap-in, try_charge -> commit or cancel, the page is locked.
3973 * And when try_charge() successfully returns, one refcnt to memcg without
3974 * struct page_cgroup is acquired. This refcnt will be consumed by
3975 * "commit()" or removed by "cancel()"
3977 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3980 struct mem_cgroup **memcgp)
3982 struct mem_cgroup *memcg;
3983 struct page_cgroup *pc;
3986 pc = lookup_page_cgroup(page);
3988 * Every swap fault against a single page tries to charge the
3989 * page, bail as early as possible. shmem_unuse() encounters
3990 * already charged pages, too. The USED bit is protected by
3991 * the page lock, which serializes swap cache removal, which
3992 * in turn serializes uncharging.
3994 if (PageCgroupUsed(pc))
3996 if (!do_swap_account)
3998 memcg = try_get_mem_cgroup_from_page(page);
4002 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4003 css_put(&memcg->css);
4008 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4014 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4015 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4018 if (mem_cgroup_disabled())
4021 * A racing thread's fault, or swapoff, may have already
4022 * updated the pte, and even removed page from swap cache: in
4023 * those cases unuse_pte()'s pte_same() test will fail; but
4024 * there's also a KSM case which does need to charge the page.
4026 if (!PageSwapCache(page)) {
4029 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4034 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4037 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4039 if (mem_cgroup_disabled())
4043 __mem_cgroup_cancel_charge(memcg, 1);
4047 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4048 enum charge_type ctype)
4050 if (mem_cgroup_disabled())
4055 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4057 * Now swap is on-memory. This means this page may be
4058 * counted both as mem and swap....double count.
4059 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4060 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4061 * may call delete_from_swap_cache() before reach here.
4063 if (do_swap_account && PageSwapCache(page)) {
4064 swp_entry_t ent = {.val = page_private(page)};
4065 mem_cgroup_uncharge_swap(ent);
4069 void mem_cgroup_commit_charge_swapin(struct page *page,
4070 struct mem_cgroup *memcg)
4072 __mem_cgroup_commit_charge_swapin(page, memcg,
4073 MEM_CGROUP_CHARGE_TYPE_ANON);
4076 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4079 struct mem_cgroup *memcg = NULL;
4080 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4083 if (mem_cgroup_disabled())
4085 if (PageCompound(page))
4088 if (!PageSwapCache(page))
4089 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4090 else { /* page is swapcache/shmem */
4091 ret = __mem_cgroup_try_charge_swapin(mm, page,
4094 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4099 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4100 unsigned int nr_pages,
4101 const enum charge_type ctype)
4103 struct memcg_batch_info *batch = NULL;
4104 bool uncharge_memsw = true;
4106 /* If swapout, usage of swap doesn't decrease */
4107 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4108 uncharge_memsw = false;
4110 batch = ¤t->memcg_batch;
4112 * In usual, we do css_get() when we remember memcg pointer.
4113 * But in this case, we keep res->usage until end of a series of
4114 * uncharges. Then, it's ok to ignore memcg's refcnt.
4117 batch->memcg = memcg;
4119 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4120 * In those cases, all pages freed continuously can be expected to be in
4121 * the same cgroup and we have chance to coalesce uncharges.
4122 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4123 * because we want to do uncharge as soon as possible.
4126 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4127 goto direct_uncharge;
4130 goto direct_uncharge;
4133 * In typical case, batch->memcg == mem. This means we can
4134 * merge a series of uncharges to an uncharge of res_counter.
4135 * If not, we uncharge res_counter ony by one.
4137 if (batch->memcg != memcg)
4138 goto direct_uncharge;
4139 /* remember freed charge and uncharge it later */
4142 batch->memsw_nr_pages++;
4145 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4147 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4148 if (unlikely(batch->memcg != memcg))
4149 memcg_oom_recover(memcg);
4153 * uncharge if !page_mapped(page)
4155 static struct mem_cgroup *
4156 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4159 struct mem_cgroup *memcg = NULL;
4160 unsigned int nr_pages = 1;
4161 struct page_cgroup *pc;
4164 if (mem_cgroup_disabled())
4167 if (PageTransHuge(page)) {
4168 nr_pages <<= compound_order(page);
4169 VM_BUG_ON(!PageTransHuge(page));
4172 * Check if our page_cgroup is valid
4174 pc = lookup_page_cgroup(page);
4175 if (unlikely(!PageCgroupUsed(pc)))
4178 lock_page_cgroup(pc);
4180 memcg = pc->mem_cgroup;
4182 if (!PageCgroupUsed(pc))
4185 anon = PageAnon(page);
4188 case MEM_CGROUP_CHARGE_TYPE_ANON:
4190 * Generally PageAnon tells if it's the anon statistics to be
4191 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4192 * used before page reached the stage of being marked PageAnon.
4196 case MEM_CGROUP_CHARGE_TYPE_DROP:
4197 /* See mem_cgroup_prepare_migration() */
4198 if (page_mapped(page))
4201 * Pages under migration may not be uncharged. But
4202 * end_migration() /must/ be the one uncharging the
4203 * unused post-migration page and so it has to call
4204 * here with the migration bit still set. See the
4205 * res_counter handling below.
4207 if (!end_migration && PageCgroupMigration(pc))
4210 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4211 if (!PageAnon(page)) { /* Shared memory */
4212 if (page->mapping && !page_is_file_cache(page))
4214 } else if (page_mapped(page)) /* Anon */
4221 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4223 ClearPageCgroupUsed(pc);
4225 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4226 * freed from LRU. This is safe because uncharged page is expected not
4227 * to be reused (freed soon). Exception is SwapCache, it's handled by
4228 * special functions.
4231 unlock_page_cgroup(pc);
4233 * even after unlock, we have memcg->res.usage here and this memcg
4234 * will never be freed, so it's safe to call css_get().
4236 memcg_check_events(memcg, page);
4237 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4238 mem_cgroup_swap_statistics(memcg, true);
4239 css_get(&memcg->css);
4242 * Migration does not charge the res_counter for the
4243 * replacement page, so leave it alone when phasing out the
4244 * page that is unused after the migration.
4246 if (!end_migration && !mem_cgroup_is_root(memcg))
4247 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4252 unlock_page_cgroup(pc);
4256 void mem_cgroup_uncharge_page(struct page *page)
4259 if (page_mapped(page))
4261 VM_BUG_ON(page->mapping && !PageAnon(page));
4263 * If the page is in swap cache, uncharge should be deferred
4264 * to the swap path, which also properly accounts swap usage
4265 * and handles memcg lifetime.
4267 * Note that this check is not stable and reclaim may add the
4268 * page to swap cache at any time after this. However, if the
4269 * page is not in swap cache by the time page->mapcount hits
4270 * 0, there won't be any page table references to the swap
4271 * slot, and reclaim will free it and not actually write the
4274 if (PageSwapCache(page))
4276 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4279 void mem_cgroup_uncharge_cache_page(struct page *page)
4281 VM_BUG_ON(page_mapped(page));
4282 VM_BUG_ON(page->mapping);
4283 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4287 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4288 * In that cases, pages are freed continuously and we can expect pages
4289 * are in the same memcg. All these calls itself limits the number of
4290 * pages freed at once, then uncharge_start/end() is called properly.
4291 * This may be called prural(2) times in a context,
4294 void mem_cgroup_uncharge_start(void)
4296 current->memcg_batch.do_batch++;
4297 /* We can do nest. */
4298 if (current->memcg_batch.do_batch == 1) {
4299 current->memcg_batch.memcg = NULL;
4300 current->memcg_batch.nr_pages = 0;
4301 current->memcg_batch.memsw_nr_pages = 0;
4305 void mem_cgroup_uncharge_end(void)
4307 struct memcg_batch_info *batch = ¤t->memcg_batch;
4309 if (!batch->do_batch)
4313 if (batch->do_batch) /* If stacked, do nothing. */
4319 * This "batch->memcg" is valid without any css_get/put etc...
4320 * bacause we hide charges behind us.
4322 if (batch->nr_pages)
4323 res_counter_uncharge(&batch->memcg->res,
4324 batch->nr_pages * PAGE_SIZE);
4325 if (batch->memsw_nr_pages)
4326 res_counter_uncharge(&batch->memcg->memsw,
4327 batch->memsw_nr_pages * PAGE_SIZE);
4328 memcg_oom_recover(batch->memcg);
4329 /* forget this pointer (for sanity check) */
4330 batch->memcg = NULL;
4335 * called after __delete_from_swap_cache() and drop "page" account.
4336 * memcg information is recorded to swap_cgroup of "ent"
4339 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4341 struct mem_cgroup *memcg;
4342 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4344 if (!swapout) /* this was a swap cache but the swap is unused ! */
4345 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4347 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4350 * record memcg information, if swapout && memcg != NULL,
4351 * css_get() was called in uncharge().
4353 if (do_swap_account && swapout && memcg)
4354 swap_cgroup_record(ent, css_id(&memcg->css));
4358 #ifdef CONFIG_MEMCG_SWAP
4360 * called from swap_entry_free(). remove record in swap_cgroup and
4361 * uncharge "memsw" account.
4363 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4365 struct mem_cgroup *memcg;
4368 if (!do_swap_account)
4371 id = swap_cgroup_record(ent, 0);
4373 memcg = mem_cgroup_lookup(id);
4376 * We uncharge this because swap is freed.
4377 * This memcg can be obsolete one. We avoid calling css_tryget
4379 if (!mem_cgroup_is_root(memcg))
4380 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4381 mem_cgroup_swap_statistics(memcg, false);
4382 css_put(&memcg->css);
4388 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4389 * @entry: swap entry to be moved
4390 * @from: mem_cgroup which the entry is moved from
4391 * @to: mem_cgroup which the entry is moved to
4393 * It succeeds only when the swap_cgroup's record for this entry is the same
4394 * as the mem_cgroup's id of @from.
4396 * Returns 0 on success, -EINVAL on failure.
4398 * The caller must have charged to @to, IOW, called res_counter_charge() about
4399 * both res and memsw, and called css_get().
4401 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4402 struct mem_cgroup *from, struct mem_cgroup *to)
4404 unsigned short old_id, new_id;
4406 old_id = css_id(&from->css);
4407 new_id = css_id(&to->css);
4409 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4410 mem_cgroup_swap_statistics(from, false);
4411 mem_cgroup_swap_statistics(to, true);
4413 * This function is only called from task migration context now.
4414 * It postpones res_counter and refcount handling till the end
4415 * of task migration(mem_cgroup_clear_mc()) for performance
4416 * improvement. But we cannot postpone css_get(to) because if
4417 * the process that has been moved to @to does swap-in, the
4418 * refcount of @to might be decreased to 0.
4420 * We are in attach() phase, so the cgroup is guaranteed to be
4421 * alive, so we can just call css_get().
4429 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4430 struct mem_cgroup *from, struct mem_cgroup *to)
4437 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4440 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4441 struct mem_cgroup **memcgp)
4443 struct mem_cgroup *memcg = NULL;
4444 unsigned int nr_pages = 1;
4445 struct page_cgroup *pc;
4446 enum charge_type ctype;
4450 if (mem_cgroup_disabled())
4453 if (PageTransHuge(page))
4454 nr_pages <<= compound_order(page);
4456 pc = lookup_page_cgroup(page);
4457 lock_page_cgroup(pc);
4458 if (PageCgroupUsed(pc)) {
4459 memcg = pc->mem_cgroup;
4460 css_get(&memcg->css);
4462 * At migrating an anonymous page, its mapcount goes down
4463 * to 0 and uncharge() will be called. But, even if it's fully
4464 * unmapped, migration may fail and this page has to be
4465 * charged again. We set MIGRATION flag here and delay uncharge
4466 * until end_migration() is called
4468 * Corner Case Thinking
4470 * When the old page was mapped as Anon and it's unmap-and-freed
4471 * while migration was ongoing.
4472 * If unmap finds the old page, uncharge() of it will be delayed
4473 * until end_migration(). If unmap finds a new page, it's
4474 * uncharged when it make mapcount to be 1->0. If unmap code
4475 * finds swap_migration_entry, the new page will not be mapped
4476 * and end_migration() will find it(mapcount==0).
4479 * When the old page was mapped but migraion fails, the kernel
4480 * remaps it. A charge for it is kept by MIGRATION flag even
4481 * if mapcount goes down to 0. We can do remap successfully
4482 * without charging it again.
4485 * The "old" page is under lock_page() until the end of
4486 * migration, so, the old page itself will not be swapped-out.
4487 * If the new page is swapped out before end_migraton, our
4488 * hook to usual swap-out path will catch the event.
4491 SetPageCgroupMigration(pc);
4493 unlock_page_cgroup(pc);
4495 * If the page is not charged at this point,
4503 * We charge new page before it's used/mapped. So, even if unlock_page()
4504 * is called before end_migration, we can catch all events on this new
4505 * page. In the case new page is migrated but not remapped, new page's
4506 * mapcount will be finally 0 and we call uncharge in end_migration().
4509 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4511 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4513 * The page is committed to the memcg, but it's not actually
4514 * charged to the res_counter since we plan on replacing the
4515 * old one and only one page is going to be left afterwards.
4517 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4520 /* remove redundant charge if migration failed*/
4521 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4522 struct page *oldpage, struct page *newpage, bool migration_ok)
4524 struct page *used, *unused;
4525 struct page_cgroup *pc;
4531 if (!migration_ok) {
4538 anon = PageAnon(used);
4539 __mem_cgroup_uncharge_common(unused,
4540 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4541 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4543 css_put(&memcg->css);
4545 * We disallowed uncharge of pages under migration because mapcount
4546 * of the page goes down to zero, temporarly.
4547 * Clear the flag and check the page should be charged.
4549 pc = lookup_page_cgroup(oldpage);
4550 lock_page_cgroup(pc);
4551 ClearPageCgroupMigration(pc);
4552 unlock_page_cgroup(pc);
4555 * If a page is a file cache, radix-tree replacement is very atomic
4556 * and we can skip this check. When it was an Anon page, its mapcount
4557 * goes down to 0. But because we added MIGRATION flage, it's not
4558 * uncharged yet. There are several case but page->mapcount check
4559 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4560 * check. (see prepare_charge() also)
4563 mem_cgroup_uncharge_page(used);
4567 * At replace page cache, newpage is not under any memcg but it's on
4568 * LRU. So, this function doesn't touch res_counter but handles LRU
4569 * in correct way. Both pages are locked so we cannot race with uncharge.
4571 void mem_cgroup_replace_page_cache(struct page *oldpage,
4572 struct page *newpage)
4574 struct mem_cgroup *memcg = NULL;
4575 struct page_cgroup *pc;
4576 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4578 if (mem_cgroup_disabled())
4581 pc = lookup_page_cgroup(oldpage);
4582 /* fix accounting on old pages */
4583 lock_page_cgroup(pc);
4584 if (PageCgroupUsed(pc)) {
4585 memcg = pc->mem_cgroup;
4586 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4587 ClearPageCgroupUsed(pc);
4589 unlock_page_cgroup(pc);
4592 * When called from shmem_replace_page(), in some cases the
4593 * oldpage has already been charged, and in some cases not.
4598 * Even if newpage->mapping was NULL before starting replacement,
4599 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4600 * LRU while we overwrite pc->mem_cgroup.
4602 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4605 #ifdef CONFIG_DEBUG_VM
4606 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4608 struct page_cgroup *pc;
4610 pc = lookup_page_cgroup(page);
4612 * Can be NULL while feeding pages into the page allocator for
4613 * the first time, i.e. during boot or memory hotplug;
4614 * or when mem_cgroup_disabled().
4616 if (likely(pc) && PageCgroupUsed(pc))
4621 bool mem_cgroup_bad_page_check(struct page *page)
4623 if (mem_cgroup_disabled())
4626 return lookup_page_cgroup_used(page) != NULL;
4629 void mem_cgroup_print_bad_page(struct page *page)
4631 struct page_cgroup *pc;
4633 pc = lookup_page_cgroup_used(page);
4635 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4636 pc, pc->flags, pc->mem_cgroup);
4641 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4642 unsigned long long val)
4645 u64 memswlimit, memlimit;
4647 int children = mem_cgroup_count_children(memcg);
4648 u64 curusage, oldusage;
4652 * For keeping hierarchical_reclaim simple, how long we should retry
4653 * is depends on callers. We set our retry-count to be function
4654 * of # of children which we should visit in this loop.
4656 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4658 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4661 while (retry_count) {
4662 if (signal_pending(current)) {
4667 * Rather than hide all in some function, I do this in
4668 * open coded manner. You see what this really does.
4669 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4671 mutex_lock(&set_limit_mutex);
4672 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4673 if (memswlimit < val) {
4675 mutex_unlock(&set_limit_mutex);
4679 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4683 ret = res_counter_set_limit(&memcg->res, val);
4685 if (memswlimit == val)
4686 memcg->memsw_is_minimum = true;
4688 memcg->memsw_is_minimum = false;
4690 mutex_unlock(&set_limit_mutex);
4695 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4696 MEM_CGROUP_RECLAIM_SHRINK);
4697 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4698 /* Usage is reduced ? */
4699 if (curusage >= oldusage)
4702 oldusage = curusage;
4704 if (!ret && enlarge)
4705 memcg_oom_recover(memcg);
4710 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4711 unsigned long long val)
4714 u64 memlimit, memswlimit, oldusage, curusage;
4715 int children = mem_cgroup_count_children(memcg);
4719 /* see mem_cgroup_resize_res_limit */
4720 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4721 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4722 while (retry_count) {
4723 if (signal_pending(current)) {
4728 * Rather than hide all in some function, I do this in
4729 * open coded manner. You see what this really does.
4730 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4732 mutex_lock(&set_limit_mutex);
4733 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4734 if (memlimit > val) {
4736 mutex_unlock(&set_limit_mutex);
4739 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4740 if (memswlimit < val)
4742 ret = res_counter_set_limit(&memcg->memsw, val);
4744 if (memlimit == val)
4745 memcg->memsw_is_minimum = true;
4747 memcg->memsw_is_minimum = false;
4749 mutex_unlock(&set_limit_mutex);
4754 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4755 MEM_CGROUP_RECLAIM_NOSWAP |
4756 MEM_CGROUP_RECLAIM_SHRINK);
4757 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4758 /* Usage is reduced ? */
4759 if (curusage >= oldusage)
4762 oldusage = curusage;
4764 if (!ret && enlarge)
4765 memcg_oom_recover(memcg);
4769 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4771 unsigned long *total_scanned)
4773 unsigned long nr_reclaimed = 0;
4774 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4775 unsigned long reclaimed;
4777 struct mem_cgroup_tree_per_zone *mctz;
4778 unsigned long long excess;
4779 unsigned long nr_scanned;
4784 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4786 * This loop can run a while, specially if mem_cgroup's continuously
4787 * keep exceeding their soft limit and putting the system under
4794 mz = mem_cgroup_largest_soft_limit_node(mctz);
4799 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4800 gfp_mask, &nr_scanned);
4801 nr_reclaimed += reclaimed;
4802 *total_scanned += nr_scanned;
4803 spin_lock(&mctz->lock);
4806 * If we failed to reclaim anything from this memory cgroup
4807 * it is time to move on to the next cgroup
4813 * Loop until we find yet another one.
4815 * By the time we get the soft_limit lock
4816 * again, someone might have aded the
4817 * group back on the RB tree. Iterate to
4818 * make sure we get a different mem.
4819 * mem_cgroup_largest_soft_limit_node returns
4820 * NULL if no other cgroup is present on
4824 __mem_cgroup_largest_soft_limit_node(mctz);
4826 css_put(&next_mz->memcg->css);
4827 else /* next_mz == NULL or other memcg */
4831 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4832 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4834 * One school of thought says that we should not add
4835 * back the node to the tree if reclaim returns 0.
4836 * But our reclaim could return 0, simply because due
4837 * to priority we are exposing a smaller subset of
4838 * memory to reclaim from. Consider this as a longer
4841 /* If excess == 0, no tree ops */
4842 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4843 spin_unlock(&mctz->lock);
4844 css_put(&mz->memcg->css);
4847 * Could not reclaim anything and there are no more
4848 * mem cgroups to try or we seem to be looping without
4849 * reclaiming anything.
4851 if (!nr_reclaimed &&
4853 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4855 } while (!nr_reclaimed);
4857 css_put(&next_mz->memcg->css);
4858 return nr_reclaimed;
4862 * mem_cgroup_force_empty_list - clears LRU of a group
4863 * @memcg: group to clear
4866 * @lru: lru to to clear
4868 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4869 * reclaim the pages page themselves - pages are moved to the parent (or root)
4872 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4873 int node, int zid, enum lru_list lru)
4875 struct lruvec *lruvec;
4876 unsigned long flags;
4877 struct list_head *list;
4881 zone = &NODE_DATA(node)->node_zones[zid];
4882 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4883 list = &lruvec->lists[lru];
4887 struct page_cgroup *pc;
4890 spin_lock_irqsave(&zone->lru_lock, flags);
4891 if (list_empty(list)) {
4892 spin_unlock_irqrestore(&zone->lru_lock, flags);
4895 page = list_entry(list->prev, struct page, lru);
4897 list_move(&page->lru, list);
4899 spin_unlock_irqrestore(&zone->lru_lock, flags);
4902 spin_unlock_irqrestore(&zone->lru_lock, flags);
4904 pc = lookup_page_cgroup(page);
4906 if (mem_cgroup_move_parent(page, pc, memcg)) {
4907 /* found lock contention or "pc" is obsolete. */
4912 } while (!list_empty(list));
4916 * make mem_cgroup's charge to be 0 if there is no task by moving
4917 * all the charges and pages to the parent.
4918 * This enables deleting this mem_cgroup.
4920 * Caller is responsible for holding css reference on the memcg.
4922 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4928 /* This is for making all *used* pages to be on LRU. */
4929 lru_add_drain_all();
4930 drain_all_stock_sync(memcg);
4931 mem_cgroup_start_move(memcg);
4932 for_each_node_state(node, N_MEMORY) {
4933 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4936 mem_cgroup_force_empty_list(memcg,
4941 mem_cgroup_end_move(memcg);
4942 memcg_oom_recover(memcg);
4946 * Kernel memory may not necessarily be trackable to a specific
4947 * process. So they are not migrated, and therefore we can't
4948 * expect their value to drop to 0 here.
4949 * Having res filled up with kmem only is enough.
4951 * This is a safety check because mem_cgroup_force_empty_list
4952 * could have raced with mem_cgroup_replace_page_cache callers
4953 * so the lru seemed empty but the page could have been added
4954 * right after the check. RES_USAGE should be safe as we always
4955 * charge before adding to the LRU.
4957 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4958 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4959 } while (usage > 0);
4963 * This mainly exists for tests during the setting of set of use_hierarchy.
4964 * Since this is the very setting we are changing, the current hierarchy value
4967 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4971 /* bounce at first found */
4972 cgroup_for_each_child(pos, memcg->css.cgroup)
4978 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4979 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4980 * from mem_cgroup_count_children(), in the sense that we don't really care how
4981 * many children we have; we only need to know if we have any. It also counts
4982 * any memcg without hierarchy as infertile.
4984 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4986 return memcg->use_hierarchy && __memcg_has_children(memcg);
4990 * Reclaims as many pages from the given memcg as possible and moves
4991 * the rest to the parent.
4993 * Caller is responsible for holding css reference for memcg.
4995 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4997 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4998 struct cgroup *cgrp = memcg->css.cgroup;
5000 /* returns EBUSY if there is a task or if we come here twice. */
5001 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
5004 /* we call try-to-free pages for make this cgroup empty */
5005 lru_add_drain_all();
5006 /* try to free all pages in this cgroup */
5007 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
5010 if (signal_pending(current))
5013 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5017 /* maybe some writeback is necessary */
5018 congestion_wait(BLK_RW_ASYNC, HZ/10);
5023 mem_cgroup_reparent_charges(memcg);
5028 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
5030 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5033 if (mem_cgroup_is_root(memcg))
5035 css_get(&memcg->css);
5036 ret = mem_cgroup_force_empty(memcg);
5037 css_put(&memcg->css);
5043 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
5045 return mem_cgroup_from_cont(cont)->use_hierarchy;
5048 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
5052 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5053 struct cgroup *parent = cont->parent;
5054 struct mem_cgroup *parent_memcg = NULL;
5057 parent_memcg = mem_cgroup_from_cont(parent);
5059 mutex_lock(&memcg_create_mutex);
5061 if (memcg->use_hierarchy == val)
5065 * If parent's use_hierarchy is set, we can't make any modifications
5066 * in the child subtrees. If it is unset, then the change can
5067 * occur, provided the current cgroup has no children.
5069 * For the root cgroup, parent_mem is NULL, we allow value to be
5070 * set if there are no children.
5072 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5073 (val == 1 || val == 0)) {
5074 if (!__memcg_has_children(memcg))
5075 memcg->use_hierarchy = val;
5082 mutex_unlock(&memcg_create_mutex);
5088 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5089 enum mem_cgroup_stat_index idx)
5091 struct mem_cgroup *iter;
5094 /* Per-cpu values can be negative, use a signed accumulator */
5095 for_each_mem_cgroup_tree(iter, memcg)
5096 val += mem_cgroup_read_stat(iter, idx);
5098 if (val < 0) /* race ? */
5103 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5107 if (!mem_cgroup_is_root(memcg)) {
5109 return res_counter_read_u64(&memcg->res, RES_USAGE);
5111 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5115 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5116 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5118 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5119 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5122 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5124 return val << PAGE_SHIFT;
5127 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
5128 struct file *file, char __user *buf,
5129 size_t nbytes, loff_t *ppos)
5131 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5137 type = MEMFILE_TYPE(cft->private);
5138 name = MEMFILE_ATTR(cft->private);
5142 if (name == RES_USAGE)
5143 val = mem_cgroup_usage(memcg, false);
5145 val = res_counter_read_u64(&memcg->res, name);
5148 if (name == RES_USAGE)
5149 val = mem_cgroup_usage(memcg, true);
5151 val = res_counter_read_u64(&memcg->memsw, name);
5154 val = res_counter_read_u64(&memcg->kmem, name);
5160 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5161 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5164 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
5167 #ifdef CONFIG_MEMCG_KMEM
5168 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5170 * For simplicity, we won't allow this to be disabled. It also can't
5171 * be changed if the cgroup has children already, or if tasks had
5174 * If tasks join before we set the limit, a person looking at
5175 * kmem.usage_in_bytes will have no way to determine when it took
5176 * place, which makes the value quite meaningless.
5178 * After it first became limited, changes in the value of the limit are
5179 * of course permitted.
5181 mutex_lock(&memcg_create_mutex);
5182 mutex_lock(&set_limit_mutex);
5183 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
5184 if (cgroup_task_count(cont) || memcg_has_children(memcg)) {
5188 ret = res_counter_set_limit(&memcg->kmem, val);
5191 ret = memcg_update_cache_sizes(memcg);
5193 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5196 static_key_slow_inc(&memcg_kmem_enabled_key);
5198 * setting the active bit after the inc will guarantee no one
5199 * starts accounting before all call sites are patched
5201 memcg_kmem_set_active(memcg);
5203 ret = res_counter_set_limit(&memcg->kmem, val);
5205 mutex_unlock(&set_limit_mutex);
5206 mutex_unlock(&memcg_create_mutex);
5211 #ifdef CONFIG_MEMCG_KMEM
5212 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5215 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5219 memcg->kmem_account_flags = parent->kmem_account_flags;
5221 * When that happen, we need to disable the static branch only on those
5222 * memcgs that enabled it. To achieve this, we would be forced to
5223 * complicate the code by keeping track of which memcgs were the ones
5224 * that actually enabled limits, and which ones got it from its
5227 * It is a lot simpler just to do static_key_slow_inc() on every child
5228 * that is accounted.
5230 if (!memcg_kmem_is_active(memcg))
5234 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5235 * memcg is active already. If the later initialization fails then the
5236 * cgroup core triggers the cleanup so we do not have to do it here.
5238 static_key_slow_inc(&memcg_kmem_enabled_key);
5240 mutex_lock(&set_limit_mutex);
5241 memcg_stop_kmem_account();
5242 ret = memcg_update_cache_sizes(memcg);
5243 memcg_resume_kmem_account();
5244 mutex_unlock(&set_limit_mutex);
5248 #endif /* CONFIG_MEMCG_KMEM */
5251 * The user of this function is...
5254 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5257 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5260 unsigned long long val;
5263 type = MEMFILE_TYPE(cft->private);
5264 name = MEMFILE_ATTR(cft->private);
5268 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5272 /* This function does all necessary parse...reuse it */
5273 ret = res_counter_memparse_write_strategy(buffer, &val);
5277 ret = mem_cgroup_resize_limit(memcg, val);
5278 else if (type == _MEMSWAP)
5279 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5280 else if (type == _KMEM)
5281 ret = memcg_update_kmem_limit(cont, val);
5285 case RES_SOFT_LIMIT:
5286 ret = res_counter_memparse_write_strategy(buffer, &val);
5290 * For memsw, soft limits are hard to implement in terms
5291 * of semantics, for now, we support soft limits for
5292 * control without swap
5295 ret = res_counter_set_soft_limit(&memcg->res, val);
5300 ret = -EINVAL; /* should be BUG() ? */
5306 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5307 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5309 struct cgroup *cgroup;
5310 unsigned long long min_limit, min_memsw_limit, tmp;
5312 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5313 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5314 cgroup = memcg->css.cgroup;
5315 if (!memcg->use_hierarchy)
5318 while (cgroup->parent) {
5319 cgroup = cgroup->parent;
5320 memcg = mem_cgroup_from_cont(cgroup);
5321 if (!memcg->use_hierarchy)
5323 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5324 min_limit = min(min_limit, tmp);
5325 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5326 min_memsw_limit = min(min_memsw_limit, tmp);
5329 *mem_limit = min_limit;
5330 *memsw_limit = min_memsw_limit;
5333 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5335 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5339 type = MEMFILE_TYPE(event);
5340 name = MEMFILE_ATTR(event);
5345 res_counter_reset_max(&memcg->res);
5346 else if (type == _MEMSWAP)
5347 res_counter_reset_max(&memcg->memsw);
5348 else if (type == _KMEM)
5349 res_counter_reset_max(&memcg->kmem);
5355 res_counter_reset_failcnt(&memcg->res);
5356 else if (type == _MEMSWAP)
5357 res_counter_reset_failcnt(&memcg->memsw);
5358 else if (type == _KMEM)
5359 res_counter_reset_failcnt(&memcg->kmem);
5368 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5371 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5375 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5376 struct cftype *cft, u64 val)
5378 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5380 if (val >= (1 << NR_MOVE_TYPE))
5384 * No kind of locking is needed in here, because ->can_attach() will
5385 * check this value once in the beginning of the process, and then carry
5386 * on with stale data. This means that changes to this value will only
5387 * affect task migrations starting after the change.
5389 memcg->move_charge_at_immigrate = val;
5393 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5394 struct cftype *cft, u64 val)
5401 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5405 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5406 unsigned long node_nr;
5407 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5409 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5410 seq_printf(m, "total=%lu", total_nr);
5411 for_each_node_state(nid, N_MEMORY) {
5412 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5413 seq_printf(m, " N%d=%lu", nid, node_nr);
5417 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5418 seq_printf(m, "file=%lu", file_nr);
5419 for_each_node_state(nid, N_MEMORY) {
5420 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5422 seq_printf(m, " N%d=%lu", nid, node_nr);
5426 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5427 seq_printf(m, "anon=%lu", anon_nr);
5428 for_each_node_state(nid, N_MEMORY) {
5429 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5431 seq_printf(m, " N%d=%lu", nid, node_nr);
5435 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5436 seq_printf(m, "unevictable=%lu", unevictable_nr);
5437 for_each_node_state(nid, N_MEMORY) {
5438 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5439 BIT(LRU_UNEVICTABLE));
5440 seq_printf(m, " N%d=%lu", nid, node_nr);
5445 #endif /* CONFIG_NUMA */
5447 static inline void mem_cgroup_lru_names_not_uptodate(void)
5449 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5452 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5455 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5456 struct mem_cgroup *mi;
5459 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5460 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5462 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5463 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5466 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5467 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5468 mem_cgroup_read_events(memcg, i));
5470 for (i = 0; i < NR_LRU_LISTS; i++)
5471 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5472 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5474 /* Hierarchical information */
5476 unsigned long long limit, memsw_limit;
5477 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5478 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5479 if (do_swap_account)
5480 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5484 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5487 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5489 for_each_mem_cgroup_tree(mi, memcg)
5490 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5491 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5494 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5495 unsigned long long val = 0;
5497 for_each_mem_cgroup_tree(mi, memcg)
5498 val += mem_cgroup_read_events(mi, i);
5499 seq_printf(m, "total_%s %llu\n",
5500 mem_cgroup_events_names[i], val);
5503 for (i = 0; i < NR_LRU_LISTS; i++) {
5504 unsigned long long val = 0;
5506 for_each_mem_cgroup_tree(mi, memcg)
5507 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5508 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5511 #ifdef CONFIG_DEBUG_VM
5514 struct mem_cgroup_per_zone *mz;
5515 struct zone_reclaim_stat *rstat;
5516 unsigned long recent_rotated[2] = {0, 0};
5517 unsigned long recent_scanned[2] = {0, 0};
5519 for_each_online_node(nid)
5520 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5521 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5522 rstat = &mz->lruvec.reclaim_stat;
5524 recent_rotated[0] += rstat->recent_rotated[0];
5525 recent_rotated[1] += rstat->recent_rotated[1];
5526 recent_scanned[0] += rstat->recent_scanned[0];
5527 recent_scanned[1] += rstat->recent_scanned[1];
5529 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5530 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5531 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5532 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5539 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5541 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5543 return mem_cgroup_swappiness(memcg);
5546 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5549 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5550 struct mem_cgroup *parent;
5555 if (cgrp->parent == NULL)
5558 parent = mem_cgroup_from_cont(cgrp->parent);
5560 mutex_lock(&memcg_create_mutex);
5562 /* If under hierarchy, only empty-root can set this value */
5563 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5564 mutex_unlock(&memcg_create_mutex);
5568 memcg->swappiness = val;
5570 mutex_unlock(&memcg_create_mutex);
5575 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5577 struct mem_cgroup_threshold_ary *t;
5583 t = rcu_dereference(memcg->thresholds.primary);
5585 t = rcu_dereference(memcg->memsw_thresholds.primary);
5590 usage = mem_cgroup_usage(memcg, swap);
5593 * current_threshold points to threshold just below or equal to usage.
5594 * If it's not true, a threshold was crossed after last
5595 * call of __mem_cgroup_threshold().
5597 i = t->current_threshold;
5600 * Iterate backward over array of thresholds starting from
5601 * current_threshold and check if a threshold is crossed.
5602 * If none of thresholds below usage is crossed, we read
5603 * only one element of the array here.
5605 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5606 eventfd_signal(t->entries[i].eventfd, 1);
5608 /* i = current_threshold + 1 */
5612 * Iterate forward over array of thresholds starting from
5613 * current_threshold+1 and check if a threshold is crossed.
5614 * If none of thresholds above usage is crossed, we read
5615 * only one element of the array here.
5617 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5618 eventfd_signal(t->entries[i].eventfd, 1);
5620 /* Update current_threshold */
5621 t->current_threshold = i - 1;
5626 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5629 __mem_cgroup_threshold(memcg, false);
5630 if (do_swap_account)
5631 __mem_cgroup_threshold(memcg, true);
5633 memcg = parent_mem_cgroup(memcg);
5637 static int compare_thresholds(const void *a, const void *b)
5639 const struct mem_cgroup_threshold *_a = a;
5640 const struct mem_cgroup_threshold *_b = b;
5642 return _a->threshold - _b->threshold;
5645 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5647 struct mem_cgroup_eventfd_list *ev;
5649 list_for_each_entry(ev, &memcg->oom_notify, list)
5650 eventfd_signal(ev->eventfd, 1);
5654 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5656 struct mem_cgroup *iter;
5658 for_each_mem_cgroup_tree(iter, memcg)
5659 mem_cgroup_oom_notify_cb(iter);
5662 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5663 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5665 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5666 struct mem_cgroup_thresholds *thresholds;
5667 struct mem_cgroup_threshold_ary *new;
5668 enum res_type type = MEMFILE_TYPE(cft->private);
5669 u64 threshold, usage;
5672 ret = res_counter_memparse_write_strategy(args, &threshold);
5676 mutex_lock(&memcg->thresholds_lock);
5679 thresholds = &memcg->thresholds;
5680 else if (type == _MEMSWAP)
5681 thresholds = &memcg->memsw_thresholds;
5685 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5687 /* Check if a threshold crossed before adding a new one */
5688 if (thresholds->primary)
5689 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5691 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5693 /* Allocate memory for new array of thresholds */
5694 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5702 /* Copy thresholds (if any) to new array */
5703 if (thresholds->primary) {
5704 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5705 sizeof(struct mem_cgroup_threshold));
5708 /* Add new threshold */
5709 new->entries[size - 1].eventfd = eventfd;
5710 new->entries[size - 1].threshold = threshold;
5712 /* Sort thresholds. Registering of new threshold isn't time-critical */
5713 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5714 compare_thresholds, NULL);
5716 /* Find current threshold */
5717 new->current_threshold = -1;
5718 for (i = 0; i < size; i++) {
5719 if (new->entries[i].threshold <= usage) {
5721 * new->current_threshold will not be used until
5722 * rcu_assign_pointer(), so it's safe to increment
5725 ++new->current_threshold;
5730 /* Free old spare buffer and save old primary buffer as spare */
5731 kfree(thresholds->spare);
5732 thresholds->spare = thresholds->primary;
5734 rcu_assign_pointer(thresholds->primary, new);
5736 /* To be sure that nobody uses thresholds */
5740 mutex_unlock(&memcg->thresholds_lock);
5745 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5746 struct cftype *cft, struct eventfd_ctx *eventfd)
5748 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5749 struct mem_cgroup_thresholds *thresholds;
5750 struct mem_cgroup_threshold_ary *new;
5751 enum res_type type = MEMFILE_TYPE(cft->private);
5755 mutex_lock(&memcg->thresholds_lock);
5757 thresholds = &memcg->thresholds;
5758 else if (type == _MEMSWAP)
5759 thresholds = &memcg->memsw_thresholds;
5763 if (!thresholds->primary)
5766 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5768 /* Check if a threshold crossed before removing */
5769 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5771 /* Calculate new number of threshold */
5773 for (i = 0; i < thresholds->primary->size; i++) {
5774 if (thresholds->primary->entries[i].eventfd != eventfd)
5778 new = thresholds->spare;
5780 /* Set thresholds array to NULL if we don't have thresholds */
5789 /* Copy thresholds and find current threshold */
5790 new->current_threshold = -1;
5791 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5792 if (thresholds->primary->entries[i].eventfd == eventfd)
5795 new->entries[j] = thresholds->primary->entries[i];
5796 if (new->entries[j].threshold <= usage) {
5798 * new->current_threshold will not be used
5799 * until rcu_assign_pointer(), so it's safe to increment
5802 ++new->current_threshold;
5808 /* Swap primary and spare array */
5809 thresholds->spare = thresholds->primary;
5810 /* If all events are unregistered, free the spare array */
5812 kfree(thresholds->spare);
5813 thresholds->spare = NULL;
5816 rcu_assign_pointer(thresholds->primary, new);
5818 /* To be sure that nobody uses thresholds */
5821 mutex_unlock(&memcg->thresholds_lock);
5824 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5825 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5827 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5828 struct mem_cgroup_eventfd_list *event;
5829 enum res_type type = MEMFILE_TYPE(cft->private);
5831 BUG_ON(type != _OOM_TYPE);
5832 event = kmalloc(sizeof(*event), GFP_KERNEL);
5836 spin_lock(&memcg_oom_lock);
5838 event->eventfd = eventfd;
5839 list_add(&event->list, &memcg->oom_notify);
5841 /* already in OOM ? */
5842 if (atomic_read(&memcg->under_oom))
5843 eventfd_signal(eventfd, 1);
5844 spin_unlock(&memcg_oom_lock);
5849 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5850 struct cftype *cft, struct eventfd_ctx *eventfd)
5852 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5853 struct mem_cgroup_eventfd_list *ev, *tmp;
5854 enum res_type type = MEMFILE_TYPE(cft->private);
5856 BUG_ON(type != _OOM_TYPE);
5858 spin_lock(&memcg_oom_lock);
5860 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5861 if (ev->eventfd == eventfd) {
5862 list_del(&ev->list);
5867 spin_unlock(&memcg_oom_lock);
5870 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5871 struct cftype *cft, struct cgroup_map_cb *cb)
5873 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5875 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5877 if (atomic_read(&memcg->under_oom))
5878 cb->fill(cb, "under_oom", 1);
5880 cb->fill(cb, "under_oom", 0);
5884 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5885 struct cftype *cft, u64 val)
5887 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5888 struct mem_cgroup *parent;
5890 /* cannot set to root cgroup and only 0 and 1 are allowed */
5891 if (!cgrp->parent || !((val == 0) || (val == 1)))
5894 parent = mem_cgroup_from_cont(cgrp->parent);
5896 mutex_lock(&memcg_create_mutex);
5897 /* oom-kill-disable is a flag for subhierarchy. */
5898 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5899 mutex_unlock(&memcg_create_mutex);
5902 memcg->oom_kill_disable = val;
5904 memcg_oom_recover(memcg);
5905 mutex_unlock(&memcg_create_mutex);
5909 #ifdef CONFIG_MEMCG_KMEM
5910 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5914 memcg->kmemcg_id = -1;
5915 ret = memcg_propagate_kmem(memcg);
5919 return mem_cgroup_sockets_init(memcg, ss);
5922 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5924 mem_cgroup_sockets_destroy(memcg);
5927 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5929 if (!memcg_kmem_is_active(memcg))
5933 * kmem charges can outlive the cgroup. In the case of slab
5934 * pages, for instance, a page contain objects from various
5935 * processes. As we prevent from taking a reference for every
5936 * such allocation we have to be careful when doing uncharge
5937 * (see memcg_uncharge_kmem) and here during offlining.
5939 * The idea is that that only the _last_ uncharge which sees
5940 * the dead memcg will drop the last reference. An additional
5941 * reference is taken here before the group is marked dead
5942 * which is then paired with css_put during uncharge resp. here.
5944 * Although this might sound strange as this path is called from
5945 * css_offline() when the referencemight have dropped down to 0
5946 * and shouldn't be incremented anymore (css_tryget would fail)
5947 * we do not have other options because of the kmem allocations
5950 css_get(&memcg->css);
5952 memcg_kmem_mark_dead(memcg);
5954 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5957 if (memcg_kmem_test_and_clear_dead(memcg))
5958 css_put(&memcg->css);
5961 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5966 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5970 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5975 static struct cftype mem_cgroup_files[] = {
5977 .name = "usage_in_bytes",
5978 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5979 .read = mem_cgroup_read,
5980 .register_event = mem_cgroup_usage_register_event,
5981 .unregister_event = mem_cgroup_usage_unregister_event,
5984 .name = "max_usage_in_bytes",
5985 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5986 .trigger = mem_cgroup_reset,
5987 .read = mem_cgroup_read,
5990 .name = "limit_in_bytes",
5991 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5992 .write_string = mem_cgroup_write,
5993 .read = mem_cgroup_read,
5996 .name = "soft_limit_in_bytes",
5997 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5998 .write_string = mem_cgroup_write,
5999 .read = mem_cgroup_read,
6003 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6004 .trigger = mem_cgroup_reset,
6005 .read = mem_cgroup_read,
6009 .read_seq_string = memcg_stat_show,
6012 .name = "force_empty",
6013 .trigger = mem_cgroup_force_empty_write,
6016 .name = "use_hierarchy",
6017 .flags = CFTYPE_INSANE,
6018 .write_u64 = mem_cgroup_hierarchy_write,
6019 .read_u64 = mem_cgroup_hierarchy_read,
6022 .name = "swappiness",
6023 .read_u64 = mem_cgroup_swappiness_read,
6024 .write_u64 = mem_cgroup_swappiness_write,
6027 .name = "move_charge_at_immigrate",
6028 .read_u64 = mem_cgroup_move_charge_read,
6029 .write_u64 = mem_cgroup_move_charge_write,
6032 .name = "oom_control",
6033 .read_map = mem_cgroup_oom_control_read,
6034 .write_u64 = mem_cgroup_oom_control_write,
6035 .register_event = mem_cgroup_oom_register_event,
6036 .unregister_event = mem_cgroup_oom_unregister_event,
6037 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6040 .name = "pressure_level",
6041 .register_event = vmpressure_register_event,
6042 .unregister_event = vmpressure_unregister_event,
6046 .name = "numa_stat",
6047 .read_seq_string = memcg_numa_stat_show,
6050 #ifdef CONFIG_MEMCG_KMEM
6052 .name = "kmem.limit_in_bytes",
6053 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6054 .write_string = mem_cgroup_write,
6055 .read = mem_cgroup_read,
6058 .name = "kmem.usage_in_bytes",
6059 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6060 .read = mem_cgroup_read,
6063 .name = "kmem.failcnt",
6064 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6065 .trigger = mem_cgroup_reset,
6066 .read = mem_cgroup_read,
6069 .name = "kmem.max_usage_in_bytes",
6070 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6071 .trigger = mem_cgroup_reset,
6072 .read = mem_cgroup_read,
6074 #ifdef CONFIG_SLABINFO
6076 .name = "kmem.slabinfo",
6077 .read_seq_string = mem_cgroup_slabinfo_read,
6081 { }, /* terminate */
6084 #ifdef CONFIG_MEMCG_SWAP
6085 static struct cftype memsw_cgroup_files[] = {
6087 .name = "memsw.usage_in_bytes",
6088 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6089 .read = mem_cgroup_read,
6090 .register_event = mem_cgroup_usage_register_event,
6091 .unregister_event = mem_cgroup_usage_unregister_event,
6094 .name = "memsw.max_usage_in_bytes",
6095 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6096 .trigger = mem_cgroup_reset,
6097 .read = mem_cgroup_read,
6100 .name = "memsw.limit_in_bytes",
6101 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6102 .write_string = mem_cgroup_write,
6103 .read = mem_cgroup_read,
6106 .name = "memsw.failcnt",
6107 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6108 .trigger = mem_cgroup_reset,
6109 .read = mem_cgroup_read,
6111 { }, /* terminate */
6114 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6116 struct mem_cgroup_per_node *pn;
6117 struct mem_cgroup_per_zone *mz;
6118 int zone, tmp = node;
6120 * This routine is called against possible nodes.
6121 * But it's BUG to call kmalloc() against offline node.
6123 * TODO: this routine can waste much memory for nodes which will
6124 * never be onlined. It's better to use memory hotplug callback
6127 if (!node_state(node, N_NORMAL_MEMORY))
6129 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6133 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6134 mz = &pn->zoneinfo[zone];
6135 lruvec_init(&mz->lruvec);
6136 mz->usage_in_excess = 0;
6137 mz->on_tree = false;
6140 memcg->nodeinfo[node] = pn;
6144 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6146 kfree(memcg->nodeinfo[node]);
6149 static struct mem_cgroup *mem_cgroup_alloc(void)
6151 struct mem_cgroup *memcg;
6152 size_t size = memcg_size();
6154 /* Can be very big if nr_node_ids is very big */
6155 if (size < PAGE_SIZE)
6156 memcg = kzalloc(size, GFP_KERNEL);
6158 memcg = vzalloc(size);
6163 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6166 spin_lock_init(&memcg->pcp_counter_lock);
6170 if (size < PAGE_SIZE)
6178 * At destroying mem_cgroup, references from swap_cgroup can remain.
6179 * (scanning all at force_empty is too costly...)
6181 * Instead of clearing all references at force_empty, we remember
6182 * the number of reference from swap_cgroup and free mem_cgroup when
6183 * it goes down to 0.
6185 * Removal of cgroup itself succeeds regardless of refs from swap.
6188 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6191 size_t size = memcg_size();
6193 mem_cgroup_remove_from_trees(memcg);
6194 free_css_id(&mem_cgroup_subsys, &memcg->css);
6197 free_mem_cgroup_per_zone_info(memcg, node);
6199 free_percpu(memcg->stat);
6202 * We need to make sure that (at least for now), the jump label
6203 * destruction code runs outside of the cgroup lock. This is because
6204 * get_online_cpus(), which is called from the static_branch update,
6205 * can't be called inside the cgroup_lock. cpusets are the ones
6206 * enforcing this dependency, so if they ever change, we might as well.
6208 * schedule_work() will guarantee this happens. Be careful if you need
6209 * to move this code around, and make sure it is outside
6212 disarm_static_keys(memcg);
6213 if (size < PAGE_SIZE)
6221 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
6222 * but in process context. The work_freeing structure is overlaid
6223 * on the rcu_freeing structure, which itself is overlaid on memsw.
6225 static void free_work(struct work_struct *work)
6227 struct mem_cgroup *memcg;
6229 memcg = container_of(work, struct mem_cgroup, work_freeing);
6230 __mem_cgroup_free(memcg);
6233 static void free_rcu(struct rcu_head *rcu_head)
6235 struct mem_cgroup *memcg;
6237 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
6238 INIT_WORK(&memcg->work_freeing, free_work);
6239 schedule_work(&memcg->work_freeing);
6242 static void mem_cgroup_get(struct mem_cgroup *memcg)
6244 atomic_inc(&memcg->refcnt);
6247 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
6249 if (atomic_sub_and_test(count, &memcg->refcnt)) {
6250 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
6251 call_rcu(&memcg->rcu_freeing, free_rcu);
6253 mem_cgroup_put(parent);
6257 static void mem_cgroup_put(struct mem_cgroup *memcg)
6259 __mem_cgroup_put(memcg, 1);
6263 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6265 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6267 if (!memcg->res.parent)
6269 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6271 EXPORT_SYMBOL(parent_mem_cgroup);
6273 static void __init mem_cgroup_soft_limit_tree_init(void)
6275 struct mem_cgroup_tree_per_node *rtpn;
6276 struct mem_cgroup_tree_per_zone *rtpz;
6277 int tmp, node, zone;
6279 for_each_node(node) {
6281 if (!node_state(node, N_NORMAL_MEMORY))
6283 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6286 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6288 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6289 rtpz = &rtpn->rb_tree_per_zone[zone];
6290 rtpz->rb_root = RB_ROOT;
6291 spin_lock_init(&rtpz->lock);
6296 static struct cgroup_subsys_state * __ref
6297 mem_cgroup_css_alloc(struct cgroup *cont)
6299 struct mem_cgroup *memcg;
6300 long error = -ENOMEM;
6303 memcg = mem_cgroup_alloc();
6305 return ERR_PTR(error);
6308 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6312 if (cont->parent == NULL) {
6313 root_mem_cgroup = memcg;
6314 res_counter_init(&memcg->res, NULL);
6315 res_counter_init(&memcg->memsw, NULL);
6316 res_counter_init(&memcg->kmem, NULL);
6319 memcg->last_scanned_node = MAX_NUMNODES;
6320 INIT_LIST_HEAD(&memcg->oom_notify);
6321 atomic_set(&memcg->refcnt, 1);
6322 memcg->move_charge_at_immigrate = 0;
6323 mutex_init(&memcg->thresholds_lock);
6324 spin_lock_init(&memcg->move_lock);
6325 vmpressure_init(&memcg->vmpressure);
6330 __mem_cgroup_free(memcg);
6331 return ERR_PTR(error);
6335 mem_cgroup_css_online(struct cgroup *cont)
6337 struct mem_cgroup *memcg, *parent;
6343 mutex_lock(&memcg_create_mutex);
6344 memcg = mem_cgroup_from_cont(cont);
6345 parent = mem_cgroup_from_cont(cont->parent);
6347 memcg->use_hierarchy = parent->use_hierarchy;
6348 memcg->oom_kill_disable = parent->oom_kill_disable;
6349 memcg->swappiness = mem_cgroup_swappiness(parent);
6351 if (parent->use_hierarchy) {
6352 res_counter_init(&memcg->res, &parent->res);
6353 res_counter_init(&memcg->memsw, &parent->memsw);
6354 res_counter_init(&memcg->kmem, &parent->kmem);
6357 * We increment refcnt of the parent to ensure that we can
6358 * safely access it on res_counter_charge/uncharge.
6359 * This refcnt will be decremented when freeing this
6360 * mem_cgroup(see mem_cgroup_put).
6362 mem_cgroup_get(parent);
6364 res_counter_init(&memcg->res, NULL);
6365 res_counter_init(&memcg->memsw, NULL);
6366 res_counter_init(&memcg->kmem, NULL);
6368 * Deeper hierachy with use_hierarchy == false doesn't make
6369 * much sense so let cgroup subsystem know about this
6370 * unfortunate state in our controller.
6372 if (parent != root_mem_cgroup)
6373 mem_cgroup_subsys.broken_hierarchy = true;
6376 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6377 mutex_unlock(&memcg_create_mutex);
6382 * Announce all parents that a group from their hierarchy is gone.
6384 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6386 struct mem_cgroup *parent = memcg;
6388 while ((parent = parent_mem_cgroup(parent)))
6389 mem_cgroup_iter_invalidate(parent);
6392 * if the root memcg is not hierarchical we have to check it
6395 if (!root_mem_cgroup->use_hierarchy)
6396 mem_cgroup_iter_invalidate(root_mem_cgroup);
6399 static void mem_cgroup_css_offline(struct cgroup *cont)
6401 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6403 kmem_cgroup_css_offline(memcg);
6405 mem_cgroup_invalidate_reclaim_iterators(memcg);
6406 mem_cgroup_reparent_charges(memcg);
6407 mem_cgroup_destroy_all_caches(memcg);
6410 static void mem_cgroup_css_free(struct cgroup *cont)
6412 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6414 memcg_destroy_kmem(memcg);
6415 __mem_cgroup_free(memcg);
6419 /* Handlers for move charge at task migration. */
6420 #define PRECHARGE_COUNT_AT_ONCE 256
6421 static int mem_cgroup_do_precharge(unsigned long count)
6424 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6425 struct mem_cgroup *memcg = mc.to;
6427 if (mem_cgroup_is_root(memcg)) {
6428 mc.precharge += count;
6429 /* we don't need css_get for root */
6432 /* try to charge at once */
6434 struct res_counter *dummy;
6436 * "memcg" cannot be under rmdir() because we've already checked
6437 * by cgroup_lock_live_cgroup() that it is not removed and we
6438 * are still under the same cgroup_mutex. So we can postpone
6441 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6443 if (do_swap_account && res_counter_charge(&memcg->memsw,
6444 PAGE_SIZE * count, &dummy)) {
6445 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6448 mc.precharge += count;
6452 /* fall back to one by one charge */
6454 if (signal_pending(current)) {
6458 if (!batch_count--) {
6459 batch_count = PRECHARGE_COUNT_AT_ONCE;
6462 ret = __mem_cgroup_try_charge(NULL,
6463 GFP_KERNEL, 1, &memcg, false);
6465 /* mem_cgroup_clear_mc() will do uncharge later */
6473 * get_mctgt_type - get target type of moving charge
6474 * @vma: the vma the pte to be checked belongs
6475 * @addr: the address corresponding to the pte to be checked
6476 * @ptent: the pte to be checked
6477 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6480 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6481 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6482 * move charge. if @target is not NULL, the page is stored in target->page
6483 * with extra refcnt got(Callers should handle it).
6484 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6485 * target for charge migration. if @target is not NULL, the entry is stored
6488 * Called with pte lock held.
6495 enum mc_target_type {
6501 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6502 unsigned long addr, pte_t ptent)
6504 struct page *page = vm_normal_page(vma, addr, ptent);
6506 if (!page || !page_mapped(page))
6508 if (PageAnon(page)) {
6509 /* we don't move shared anon */
6512 } else if (!move_file())
6513 /* we ignore mapcount for file pages */
6515 if (!get_page_unless_zero(page))
6522 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6523 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6525 struct page *page = NULL;
6526 swp_entry_t ent = pte_to_swp_entry(ptent);
6528 if (!move_anon() || non_swap_entry(ent))
6531 * Because lookup_swap_cache() updates some statistics counter,
6532 * we call find_get_page() with swapper_space directly.
6534 page = find_get_page(swap_address_space(ent), ent.val);
6535 if (do_swap_account)
6536 entry->val = ent.val;
6541 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6542 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6548 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6549 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6551 struct page *page = NULL;
6552 struct address_space *mapping;
6555 if (!vma->vm_file) /* anonymous vma */
6560 mapping = vma->vm_file->f_mapping;
6561 if (pte_none(ptent))
6562 pgoff = linear_page_index(vma, addr);
6563 else /* pte_file(ptent) is true */
6564 pgoff = pte_to_pgoff(ptent);
6566 /* page is moved even if it's not RSS of this task(page-faulted). */
6567 page = find_get_page(mapping, pgoff);
6570 /* shmem/tmpfs may report page out on swap: account for that too. */
6571 if (radix_tree_exceptional_entry(page)) {
6572 swp_entry_t swap = radix_to_swp_entry(page);
6573 if (do_swap_account)
6575 page = find_get_page(swap_address_space(swap), swap.val);
6581 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6582 unsigned long addr, pte_t ptent, union mc_target *target)
6584 struct page *page = NULL;
6585 struct page_cgroup *pc;
6586 enum mc_target_type ret = MC_TARGET_NONE;
6587 swp_entry_t ent = { .val = 0 };
6589 if (pte_present(ptent))
6590 page = mc_handle_present_pte(vma, addr, ptent);
6591 else if (is_swap_pte(ptent))
6592 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6593 else if (pte_none(ptent) || pte_file(ptent))
6594 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6596 if (!page && !ent.val)
6599 pc = lookup_page_cgroup(page);
6601 * Do only loose check w/o page_cgroup lock.
6602 * mem_cgroup_move_account() checks the pc is valid or not under
6605 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6606 ret = MC_TARGET_PAGE;
6608 target->page = page;
6610 if (!ret || !target)
6613 /* There is a swap entry and a page doesn't exist or isn't charged */
6614 if (ent.val && !ret &&
6615 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6616 ret = MC_TARGET_SWAP;
6623 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6625 * We don't consider swapping or file mapped pages because THP does not
6626 * support them for now.
6627 * Caller should make sure that pmd_trans_huge(pmd) is true.
6629 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6630 unsigned long addr, pmd_t pmd, union mc_target *target)
6632 struct page *page = NULL;
6633 struct page_cgroup *pc;
6634 enum mc_target_type ret = MC_TARGET_NONE;
6636 page = pmd_page(pmd);
6637 VM_BUG_ON(!page || !PageHead(page));
6640 pc = lookup_page_cgroup(page);
6641 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6642 ret = MC_TARGET_PAGE;
6645 target->page = page;
6651 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6652 unsigned long addr, pmd_t pmd, union mc_target *target)
6654 return MC_TARGET_NONE;
6658 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6659 unsigned long addr, unsigned long end,
6660 struct mm_walk *walk)
6662 struct vm_area_struct *vma = walk->private;
6666 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6667 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6668 mc.precharge += HPAGE_PMD_NR;
6669 spin_unlock(&vma->vm_mm->page_table_lock);
6673 if (pmd_trans_unstable(pmd))
6675 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6676 for (; addr != end; pte++, addr += PAGE_SIZE)
6677 if (get_mctgt_type(vma, addr, *pte, NULL))
6678 mc.precharge++; /* increment precharge temporarily */
6679 pte_unmap_unlock(pte - 1, ptl);
6685 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6687 unsigned long precharge;
6688 struct vm_area_struct *vma;
6690 down_read(&mm->mmap_sem);
6691 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6692 struct mm_walk mem_cgroup_count_precharge_walk = {
6693 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6697 if (is_vm_hugetlb_page(vma))
6699 walk_page_range(vma->vm_start, vma->vm_end,
6700 &mem_cgroup_count_precharge_walk);
6702 up_read(&mm->mmap_sem);
6704 precharge = mc.precharge;
6710 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6712 unsigned long precharge = mem_cgroup_count_precharge(mm);
6714 VM_BUG_ON(mc.moving_task);
6715 mc.moving_task = current;
6716 return mem_cgroup_do_precharge(precharge);
6719 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6720 static void __mem_cgroup_clear_mc(void)
6722 struct mem_cgroup *from = mc.from;
6723 struct mem_cgroup *to = mc.to;
6726 /* we must uncharge all the leftover precharges from mc.to */
6728 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6732 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6733 * we must uncharge here.
6735 if (mc.moved_charge) {
6736 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6737 mc.moved_charge = 0;
6739 /* we must fixup refcnts and charges */
6740 if (mc.moved_swap) {
6741 /* uncharge swap account from the old cgroup */
6742 if (!mem_cgroup_is_root(mc.from))
6743 res_counter_uncharge(&mc.from->memsw,
6744 PAGE_SIZE * mc.moved_swap);
6746 for (i = 0; i < mc.moved_swap; i++)
6747 css_put(&mc.from->css);
6749 if (!mem_cgroup_is_root(mc.to)) {
6751 * we charged both to->res and to->memsw, so we should
6754 res_counter_uncharge(&mc.to->res,
6755 PAGE_SIZE * mc.moved_swap);
6757 /* we've already done css_get(mc.to) */
6760 memcg_oom_recover(from);
6761 memcg_oom_recover(to);
6762 wake_up_all(&mc.waitq);
6765 static void mem_cgroup_clear_mc(void)
6767 struct mem_cgroup *from = mc.from;
6770 * we must clear moving_task before waking up waiters at the end of
6773 mc.moving_task = NULL;
6774 __mem_cgroup_clear_mc();
6775 spin_lock(&mc.lock);
6778 spin_unlock(&mc.lock);
6779 mem_cgroup_end_move(from);
6782 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6783 struct cgroup_taskset *tset)
6785 struct task_struct *p = cgroup_taskset_first(tset);
6787 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6788 unsigned long move_charge_at_immigrate;
6791 * We are now commited to this value whatever it is. Changes in this
6792 * tunable will only affect upcoming migrations, not the current one.
6793 * So we need to save it, and keep it going.
6795 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6796 if (move_charge_at_immigrate) {
6797 struct mm_struct *mm;
6798 struct mem_cgroup *from = mem_cgroup_from_task(p);
6800 VM_BUG_ON(from == memcg);
6802 mm = get_task_mm(p);
6805 /* We move charges only when we move a owner of the mm */
6806 if (mm->owner == p) {
6809 VM_BUG_ON(mc.precharge);
6810 VM_BUG_ON(mc.moved_charge);
6811 VM_BUG_ON(mc.moved_swap);
6812 mem_cgroup_start_move(from);
6813 spin_lock(&mc.lock);
6816 mc.immigrate_flags = move_charge_at_immigrate;
6817 spin_unlock(&mc.lock);
6818 /* We set mc.moving_task later */
6820 ret = mem_cgroup_precharge_mc(mm);
6822 mem_cgroup_clear_mc();
6829 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6830 struct cgroup_taskset *tset)
6832 mem_cgroup_clear_mc();
6835 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6836 unsigned long addr, unsigned long end,
6837 struct mm_walk *walk)
6840 struct vm_area_struct *vma = walk->private;
6843 enum mc_target_type target_type;
6844 union mc_target target;
6846 struct page_cgroup *pc;
6849 * We don't take compound_lock() here but no race with splitting thp
6851 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6852 * under splitting, which means there's no concurrent thp split,
6853 * - if another thread runs into split_huge_page() just after we
6854 * entered this if-block, the thread must wait for page table lock
6855 * to be unlocked in __split_huge_page_splitting(), where the main
6856 * part of thp split is not executed yet.
6858 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6859 if (mc.precharge < HPAGE_PMD_NR) {
6860 spin_unlock(&vma->vm_mm->page_table_lock);
6863 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6864 if (target_type == MC_TARGET_PAGE) {
6866 if (!isolate_lru_page(page)) {
6867 pc = lookup_page_cgroup(page);
6868 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6869 pc, mc.from, mc.to)) {
6870 mc.precharge -= HPAGE_PMD_NR;
6871 mc.moved_charge += HPAGE_PMD_NR;
6873 putback_lru_page(page);
6877 spin_unlock(&vma->vm_mm->page_table_lock);
6881 if (pmd_trans_unstable(pmd))
6884 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6885 for (; addr != end; addr += PAGE_SIZE) {
6886 pte_t ptent = *(pte++);
6892 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6893 case MC_TARGET_PAGE:
6895 if (isolate_lru_page(page))
6897 pc = lookup_page_cgroup(page);
6898 if (!mem_cgroup_move_account(page, 1, pc,
6901 /* we uncharge from mc.from later. */
6904 putback_lru_page(page);
6905 put: /* get_mctgt_type() gets the page */
6908 case MC_TARGET_SWAP:
6910 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6912 /* we fixup refcnts and charges later. */
6920 pte_unmap_unlock(pte - 1, ptl);
6925 * We have consumed all precharges we got in can_attach().
6926 * We try charge one by one, but don't do any additional
6927 * charges to mc.to if we have failed in charge once in attach()
6930 ret = mem_cgroup_do_precharge(1);
6938 static void mem_cgroup_move_charge(struct mm_struct *mm)
6940 struct vm_area_struct *vma;
6942 lru_add_drain_all();
6944 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6946 * Someone who are holding the mmap_sem might be waiting in
6947 * waitq. So we cancel all extra charges, wake up all waiters,
6948 * and retry. Because we cancel precharges, we might not be able
6949 * to move enough charges, but moving charge is a best-effort
6950 * feature anyway, so it wouldn't be a big problem.
6952 __mem_cgroup_clear_mc();
6956 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6958 struct mm_walk mem_cgroup_move_charge_walk = {
6959 .pmd_entry = mem_cgroup_move_charge_pte_range,
6963 if (is_vm_hugetlb_page(vma))
6965 ret = walk_page_range(vma->vm_start, vma->vm_end,
6966 &mem_cgroup_move_charge_walk);
6969 * means we have consumed all precharges and failed in
6970 * doing additional charge. Just abandon here.
6974 up_read(&mm->mmap_sem);
6977 static void mem_cgroup_move_task(struct cgroup *cont,
6978 struct cgroup_taskset *tset)
6980 struct task_struct *p = cgroup_taskset_first(tset);
6981 struct mm_struct *mm = get_task_mm(p);
6985 mem_cgroup_move_charge(mm);
6989 mem_cgroup_clear_mc();
6991 #else /* !CONFIG_MMU */
6992 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6993 struct cgroup_taskset *tset)
6997 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6998 struct cgroup_taskset *tset)
7001 static void mem_cgroup_move_task(struct cgroup *cont,
7002 struct cgroup_taskset *tset)
7008 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7009 * to verify sane_behavior flag on each mount attempt.
7011 static void mem_cgroup_bind(struct cgroup *root)
7014 * use_hierarchy is forced with sane_behavior. cgroup core
7015 * guarantees that @root doesn't have any children, so turning it
7016 * on for the root memcg is enough.
7018 if (cgroup_sane_behavior(root))
7019 mem_cgroup_from_cont(root)->use_hierarchy = true;
7022 struct cgroup_subsys mem_cgroup_subsys = {
7024 .subsys_id = mem_cgroup_subsys_id,
7025 .css_alloc = mem_cgroup_css_alloc,
7026 .css_online = mem_cgroup_css_online,
7027 .css_offline = mem_cgroup_css_offline,
7028 .css_free = mem_cgroup_css_free,
7029 .can_attach = mem_cgroup_can_attach,
7030 .cancel_attach = mem_cgroup_cancel_attach,
7031 .attach = mem_cgroup_move_task,
7032 .bind = mem_cgroup_bind,
7033 .base_cftypes = mem_cgroup_files,
7038 #ifdef CONFIG_MEMCG_SWAP
7039 static int __init enable_swap_account(char *s)
7041 /* consider enabled if no parameter or 1 is given */
7042 if (!strcmp(s, "1"))
7043 really_do_swap_account = 1;
7044 else if (!strcmp(s, "0"))
7045 really_do_swap_account = 0;
7048 __setup("swapaccount=", enable_swap_account);
7050 static void __init memsw_file_init(void)
7052 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7055 static void __init enable_swap_cgroup(void)
7057 if (!mem_cgroup_disabled() && really_do_swap_account) {
7058 do_swap_account = 1;
7064 static void __init enable_swap_cgroup(void)
7070 * subsys_initcall() for memory controller.
7072 * Some parts like hotcpu_notifier() have to be initialized from this context
7073 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7074 * everything that doesn't depend on a specific mem_cgroup structure should
7075 * be initialized from here.
7077 static int __init mem_cgroup_init(void)
7079 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7080 enable_swap_cgroup();
7081 mem_cgroup_soft_limit_tree_init();
7085 subsys_initcall(mem_cgroup_init);