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/mm_inline.h>
53 #include <linux/page_cgroup.h>
54 #include <linux/cpu.h>
55 #include <linux/oom.h>
59 #include <net/tcp_memcontrol.h>
61 #include <asm/uaccess.h>
63 #include <trace/events/vmscan.h>
65 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
66 #define MEM_CGROUP_RECLAIM_RETRIES 5
67 static struct mem_cgroup *root_mem_cgroup __read_mostly;
69 #ifdef CONFIG_MEMCG_SWAP
70 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
71 int do_swap_account __read_mostly;
73 /* for remember boot option*/
74 #ifdef CONFIG_MEMCG_SWAP_ENABLED
75 static int really_do_swap_account __initdata = 1;
77 static int really_do_swap_account __initdata = 0;
81 #define do_swap_account 0
86 * Statistics for memory cgroup.
88 enum mem_cgroup_stat_index {
90 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
92 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
93 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
94 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
95 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
96 MEM_CGROUP_STAT_NSTATS,
99 static const char * const mem_cgroup_stat_names[] = {
106 enum mem_cgroup_events_index {
107 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
108 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
109 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
110 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
111 MEM_CGROUP_EVENTS_NSTATS,
114 static const char * const mem_cgroup_events_names[] = {
122 * Per memcg event counter is incremented at every pagein/pageout. With THP,
123 * it will be incremated by the number of pages. This counter is used for
124 * for trigger some periodic events. This is straightforward and better
125 * than using jiffies etc. to handle periodic memcg event.
127 enum mem_cgroup_events_target {
128 MEM_CGROUP_TARGET_THRESH,
129 MEM_CGROUP_TARGET_SOFTLIMIT,
130 MEM_CGROUP_TARGET_NUMAINFO,
133 #define THRESHOLDS_EVENTS_TARGET 128
134 #define SOFTLIMIT_EVENTS_TARGET 1024
135 #define NUMAINFO_EVENTS_TARGET 1024
137 struct mem_cgroup_stat_cpu {
138 long count[MEM_CGROUP_STAT_NSTATS];
139 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
140 unsigned long nr_page_events;
141 unsigned long targets[MEM_CGROUP_NTARGETS];
144 struct mem_cgroup_reclaim_iter {
145 /* css_id of the last scanned hierarchy member */
147 /* scan generation, increased every round-trip */
148 unsigned int generation;
152 * per-zone information in memory controller.
154 struct mem_cgroup_per_zone {
155 struct lruvec lruvec;
156 unsigned long lru_size[NR_LRU_LISTS];
158 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
160 struct rb_node tree_node; /* RB tree node */
161 unsigned long long usage_in_excess;/* Set to the value by which */
162 /* the soft limit is exceeded*/
164 struct mem_cgroup *memcg; /* Back pointer, we cannot */
165 /* use container_of */
168 struct mem_cgroup_per_node {
169 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
172 struct mem_cgroup_lru_info {
173 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
177 * Cgroups above their limits are maintained in a RB-Tree, independent of
178 * their hierarchy representation
181 struct mem_cgroup_tree_per_zone {
182 struct rb_root rb_root;
186 struct mem_cgroup_tree_per_node {
187 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
190 struct mem_cgroup_tree {
191 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
194 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
196 struct mem_cgroup_threshold {
197 struct eventfd_ctx *eventfd;
202 struct mem_cgroup_threshold_ary {
203 /* An array index points to threshold just below or equal to usage. */
204 int current_threshold;
205 /* Size of entries[] */
207 /* Array of thresholds */
208 struct mem_cgroup_threshold entries[0];
211 struct mem_cgroup_thresholds {
212 /* Primary thresholds array */
213 struct mem_cgroup_threshold_ary *primary;
215 * Spare threshold array.
216 * This is needed to make mem_cgroup_unregister_event() "never fail".
217 * It must be able to store at least primary->size - 1 entries.
219 struct mem_cgroup_threshold_ary *spare;
223 struct mem_cgroup_eventfd_list {
224 struct list_head list;
225 struct eventfd_ctx *eventfd;
228 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
229 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
232 * The memory controller data structure. The memory controller controls both
233 * page cache and RSS per cgroup. We would eventually like to provide
234 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
235 * to help the administrator determine what knobs to tune.
237 * TODO: Add a water mark for the memory controller. Reclaim will begin when
238 * we hit the water mark. May be even add a low water mark, such that
239 * no reclaim occurs from a cgroup at it's low water mark, this is
240 * a feature that will be implemented much later in the future.
243 struct cgroup_subsys_state css;
245 * the counter to account for memory usage
247 struct res_counter res;
251 * the counter to account for mem+swap usage.
253 struct res_counter memsw;
256 * rcu_freeing is used only when freeing struct mem_cgroup,
257 * so put it into a union to avoid wasting more memory.
258 * It must be disjoint from the css field. It could be
259 * in a union with the res field, but res plays a much
260 * larger part in mem_cgroup life than memsw, and might
261 * be of interest, even at time of free, when debugging.
262 * So share rcu_head with the less interesting memsw.
264 struct rcu_head rcu_freeing;
266 * We also need some space for a worker in deferred freeing.
267 * By the time we call it, rcu_freeing is no longer in use.
269 struct work_struct work_freeing;
273 * the counter to account for kernel memory usage.
275 struct res_counter kmem;
277 * Per cgroup active and inactive list, similar to the
278 * per zone LRU lists.
280 struct mem_cgroup_lru_info info;
281 int last_scanned_node;
283 nodemask_t scan_nodes;
284 atomic_t numainfo_events;
285 atomic_t numainfo_updating;
288 * Should the accounting and control be hierarchical, per subtree?
291 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
299 /* OOM-Killer disable */
300 int oom_kill_disable;
302 /* set when res.limit == memsw.limit */
303 bool memsw_is_minimum;
305 /* protect arrays of thresholds */
306 struct mutex thresholds_lock;
308 /* thresholds for memory usage. RCU-protected */
309 struct mem_cgroup_thresholds thresholds;
311 /* thresholds for mem+swap usage. RCU-protected */
312 struct mem_cgroup_thresholds memsw_thresholds;
314 /* For oom notifier event fd */
315 struct list_head oom_notify;
318 * Should we move charges of a task when a task is moved into this
319 * mem_cgroup ? And what type of charges should we move ?
321 unsigned long move_charge_at_immigrate;
323 * set > 0 if pages under this cgroup are moving to other cgroup.
325 atomic_t moving_account;
326 /* taken only while moving_account > 0 */
327 spinlock_t move_lock;
331 struct mem_cgroup_stat_cpu __percpu *stat;
333 * used when a cpu is offlined or other synchronizations
334 * See mem_cgroup_read_stat().
336 struct mem_cgroup_stat_cpu nocpu_base;
337 spinlock_t pcp_counter_lock;
339 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
340 struct tcp_memcontrol tcp_mem;
342 #if defined(CONFIG_MEMCG_KMEM)
343 /* analogous to slab_common's slab_caches list. per-memcg */
344 struct list_head memcg_slab_caches;
345 /* Not a spinlock, we can take a lot of time walking the list */
346 struct mutex slab_caches_mutex;
347 /* Index in the kmem_cache->memcg_params->memcg_caches array */
352 /* internal only representation about the status of kmem accounting. */
354 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
355 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
356 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
359 /* We account when limit is on, but only after call sites are patched */
360 #define KMEM_ACCOUNTED_MASK \
361 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
363 #ifdef CONFIG_MEMCG_KMEM
364 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
366 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
369 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
371 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
374 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
376 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
379 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
381 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
384 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
386 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
387 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
390 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
392 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
393 &memcg->kmem_account_flags);
397 /* Stuffs for move charges at task migration. */
399 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
400 * left-shifted bitmap of these types.
403 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
404 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
408 /* "mc" and its members are protected by cgroup_mutex */
409 static struct move_charge_struct {
410 spinlock_t lock; /* for from, to */
411 struct mem_cgroup *from;
412 struct mem_cgroup *to;
413 unsigned long precharge;
414 unsigned long moved_charge;
415 unsigned long moved_swap;
416 struct task_struct *moving_task; /* a task moving charges */
417 wait_queue_head_t waitq; /* a waitq for other context */
419 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
420 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
423 static bool move_anon(void)
425 return test_bit(MOVE_CHARGE_TYPE_ANON,
426 &mc.to->move_charge_at_immigrate);
429 static bool move_file(void)
431 return test_bit(MOVE_CHARGE_TYPE_FILE,
432 &mc.to->move_charge_at_immigrate);
436 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
437 * limit reclaim to prevent infinite loops, if they ever occur.
439 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
440 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
443 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
444 MEM_CGROUP_CHARGE_TYPE_ANON,
445 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
446 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
450 /* for encoding cft->private value on file */
458 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
459 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
460 #define MEMFILE_ATTR(val) ((val) & 0xffff)
461 /* Used for OOM nofiier */
462 #define OOM_CONTROL (0)
465 * Reclaim flags for mem_cgroup_hierarchical_reclaim
467 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
468 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
469 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
470 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
472 static void mem_cgroup_get(struct mem_cgroup *memcg);
473 static void mem_cgroup_put(struct mem_cgroup *memcg);
476 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
478 return container_of(s, struct mem_cgroup, css);
481 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
483 return (memcg == root_mem_cgroup);
486 /* Writing them here to avoid exposing memcg's inner layout */
487 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
489 void sock_update_memcg(struct sock *sk)
491 if (mem_cgroup_sockets_enabled) {
492 struct mem_cgroup *memcg;
493 struct cg_proto *cg_proto;
495 BUG_ON(!sk->sk_prot->proto_cgroup);
497 /* Socket cloning can throw us here with sk_cgrp already
498 * filled. It won't however, necessarily happen from
499 * process context. So the test for root memcg given
500 * the current task's memcg won't help us in this case.
502 * Respecting the original socket's memcg is a better
503 * decision in this case.
506 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
507 mem_cgroup_get(sk->sk_cgrp->memcg);
512 memcg = mem_cgroup_from_task(current);
513 cg_proto = sk->sk_prot->proto_cgroup(memcg);
514 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
515 mem_cgroup_get(memcg);
516 sk->sk_cgrp = cg_proto;
521 EXPORT_SYMBOL(sock_update_memcg);
523 void sock_release_memcg(struct sock *sk)
525 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
526 struct mem_cgroup *memcg;
527 WARN_ON(!sk->sk_cgrp->memcg);
528 memcg = sk->sk_cgrp->memcg;
529 mem_cgroup_put(memcg);
533 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
535 if (!memcg || mem_cgroup_is_root(memcg))
538 return &memcg->tcp_mem.cg_proto;
540 EXPORT_SYMBOL(tcp_proto_cgroup);
542 static void disarm_sock_keys(struct mem_cgroup *memcg)
544 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
546 static_key_slow_dec(&memcg_socket_limit_enabled);
549 static void disarm_sock_keys(struct mem_cgroup *memcg)
554 #ifdef CONFIG_MEMCG_KMEM
556 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
557 * There are two main reasons for not using the css_id for this:
558 * 1) this works better in sparse environments, where we have a lot of memcgs,
559 * but only a few kmem-limited. Or also, if we have, for instance, 200
560 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
561 * 200 entry array for that.
563 * 2) In order not to violate the cgroup API, we would like to do all memory
564 * allocation in ->create(). At that point, we haven't yet allocated the
565 * css_id. Having a separate index prevents us from messing with the cgroup
568 * The current size of the caches array is stored in
569 * memcg_limited_groups_array_size. It will double each time we have to
572 static struct ida kmem_limited_groups;
573 static int memcg_limited_groups_array_size;
575 * MIN_SIZE is different than 1, because we would like to avoid going through
576 * the alloc/free process all the time. In a small machine, 4 kmem-limited
577 * cgroups is a reasonable guess. In the future, it could be a parameter or
578 * tunable, but that is strictly not necessary.
580 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
581 * this constant directly from cgroup, but it is understandable that this is
582 * better kept as an internal representation in cgroup.c. In any case, the
583 * css_id space is not getting any smaller, and we don't have to necessarily
584 * increase ours as well if it increases.
586 #define MEMCG_CACHES_MIN_SIZE 4
587 #define MEMCG_CACHES_MAX_SIZE 65535
589 struct static_key memcg_kmem_enabled_key;
591 static void disarm_kmem_keys(struct mem_cgroup *memcg)
593 if (memcg_kmem_is_active(memcg)) {
594 static_key_slow_dec(&memcg_kmem_enabled_key);
595 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
598 * This check can't live in kmem destruction function,
599 * since the charges will outlive the cgroup
601 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
604 static void disarm_kmem_keys(struct mem_cgroup *memcg)
607 #endif /* CONFIG_MEMCG_KMEM */
609 static void disarm_static_keys(struct mem_cgroup *memcg)
611 disarm_sock_keys(memcg);
612 disarm_kmem_keys(memcg);
615 static void drain_all_stock_async(struct mem_cgroup *memcg);
617 static struct mem_cgroup_per_zone *
618 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
620 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
623 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
628 static struct mem_cgroup_per_zone *
629 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
631 int nid = page_to_nid(page);
632 int zid = page_zonenum(page);
634 return mem_cgroup_zoneinfo(memcg, nid, zid);
637 static struct mem_cgroup_tree_per_zone *
638 soft_limit_tree_node_zone(int nid, int zid)
640 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
643 static struct mem_cgroup_tree_per_zone *
644 soft_limit_tree_from_page(struct page *page)
646 int nid = page_to_nid(page);
647 int zid = page_zonenum(page);
649 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
653 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
654 struct mem_cgroup_per_zone *mz,
655 struct mem_cgroup_tree_per_zone *mctz,
656 unsigned long long new_usage_in_excess)
658 struct rb_node **p = &mctz->rb_root.rb_node;
659 struct rb_node *parent = NULL;
660 struct mem_cgroup_per_zone *mz_node;
665 mz->usage_in_excess = new_usage_in_excess;
666 if (!mz->usage_in_excess)
670 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
672 if (mz->usage_in_excess < mz_node->usage_in_excess)
675 * We can't avoid mem cgroups that are over their soft
676 * limit by the same amount
678 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
681 rb_link_node(&mz->tree_node, parent, p);
682 rb_insert_color(&mz->tree_node, &mctz->rb_root);
687 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
688 struct mem_cgroup_per_zone *mz,
689 struct mem_cgroup_tree_per_zone *mctz)
693 rb_erase(&mz->tree_node, &mctz->rb_root);
698 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
699 struct mem_cgroup_per_zone *mz,
700 struct mem_cgroup_tree_per_zone *mctz)
702 spin_lock(&mctz->lock);
703 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
704 spin_unlock(&mctz->lock);
708 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
710 unsigned long long excess;
711 struct mem_cgroup_per_zone *mz;
712 struct mem_cgroup_tree_per_zone *mctz;
713 int nid = page_to_nid(page);
714 int zid = page_zonenum(page);
715 mctz = soft_limit_tree_from_page(page);
718 * Necessary to update all ancestors when hierarchy is used.
719 * because their event counter is not touched.
721 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
722 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
723 excess = res_counter_soft_limit_excess(&memcg->res);
725 * We have to update the tree if mz is on RB-tree or
726 * mem is over its softlimit.
728 if (excess || mz->on_tree) {
729 spin_lock(&mctz->lock);
730 /* if on-tree, remove it */
732 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
734 * Insert again. mz->usage_in_excess will be updated.
735 * If excess is 0, no tree ops.
737 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
738 spin_unlock(&mctz->lock);
743 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
746 struct mem_cgroup_per_zone *mz;
747 struct mem_cgroup_tree_per_zone *mctz;
749 for_each_node(node) {
750 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
751 mz = mem_cgroup_zoneinfo(memcg, node, zone);
752 mctz = soft_limit_tree_node_zone(node, zone);
753 mem_cgroup_remove_exceeded(memcg, mz, mctz);
758 static struct mem_cgroup_per_zone *
759 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
761 struct rb_node *rightmost = NULL;
762 struct mem_cgroup_per_zone *mz;
766 rightmost = rb_last(&mctz->rb_root);
768 goto done; /* Nothing to reclaim from */
770 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
772 * Remove the node now but someone else can add it back,
773 * we will to add it back at the end of reclaim to its correct
774 * position in the tree.
776 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
777 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
778 !css_tryget(&mz->memcg->css))
784 static struct mem_cgroup_per_zone *
785 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
787 struct mem_cgroup_per_zone *mz;
789 spin_lock(&mctz->lock);
790 mz = __mem_cgroup_largest_soft_limit_node(mctz);
791 spin_unlock(&mctz->lock);
796 * Implementation Note: reading percpu statistics for memcg.
798 * Both of vmstat[] and percpu_counter has threshold and do periodic
799 * synchronization to implement "quick" read. There are trade-off between
800 * reading cost and precision of value. Then, we may have a chance to implement
801 * a periodic synchronizion of counter in memcg's counter.
803 * But this _read() function is used for user interface now. The user accounts
804 * memory usage by memory cgroup and he _always_ requires exact value because
805 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
806 * have to visit all online cpus and make sum. So, for now, unnecessary
807 * synchronization is not implemented. (just implemented for cpu hotplug)
809 * If there are kernel internal actions which can make use of some not-exact
810 * value, and reading all cpu value can be performance bottleneck in some
811 * common workload, threashold and synchonization as vmstat[] should be
814 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
815 enum mem_cgroup_stat_index idx)
821 for_each_online_cpu(cpu)
822 val += per_cpu(memcg->stat->count[idx], cpu);
823 #ifdef CONFIG_HOTPLUG_CPU
824 spin_lock(&memcg->pcp_counter_lock);
825 val += memcg->nocpu_base.count[idx];
826 spin_unlock(&memcg->pcp_counter_lock);
832 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
835 int val = (charge) ? 1 : -1;
836 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
839 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
840 enum mem_cgroup_events_index idx)
842 unsigned long val = 0;
845 for_each_online_cpu(cpu)
846 val += per_cpu(memcg->stat->events[idx], cpu);
847 #ifdef CONFIG_HOTPLUG_CPU
848 spin_lock(&memcg->pcp_counter_lock);
849 val += memcg->nocpu_base.events[idx];
850 spin_unlock(&memcg->pcp_counter_lock);
855 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
856 bool anon, int nr_pages)
861 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
862 * counted as CACHE even if it's on ANON LRU.
865 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
868 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
871 /* pagein of a big page is an event. So, ignore page size */
873 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
875 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
876 nr_pages = -nr_pages; /* for event */
879 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
885 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
887 struct mem_cgroup_per_zone *mz;
889 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
890 return mz->lru_size[lru];
894 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
895 unsigned int lru_mask)
897 struct mem_cgroup_per_zone *mz;
899 unsigned long ret = 0;
901 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
904 if (BIT(lru) & lru_mask)
905 ret += mz->lru_size[lru];
911 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
912 int nid, unsigned int lru_mask)
917 for (zid = 0; zid < MAX_NR_ZONES; zid++)
918 total += mem_cgroup_zone_nr_lru_pages(memcg,
924 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
925 unsigned int lru_mask)
930 for_each_node_state(nid, N_HIGH_MEMORY)
931 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
935 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
936 enum mem_cgroup_events_target target)
938 unsigned long val, next;
940 val = __this_cpu_read(memcg->stat->nr_page_events);
941 next = __this_cpu_read(memcg->stat->targets[target]);
942 /* from time_after() in jiffies.h */
943 if ((long)next - (long)val < 0) {
945 case MEM_CGROUP_TARGET_THRESH:
946 next = val + THRESHOLDS_EVENTS_TARGET;
948 case MEM_CGROUP_TARGET_SOFTLIMIT:
949 next = val + SOFTLIMIT_EVENTS_TARGET;
951 case MEM_CGROUP_TARGET_NUMAINFO:
952 next = val + NUMAINFO_EVENTS_TARGET;
957 __this_cpu_write(memcg->stat->targets[target], next);
964 * Check events in order.
967 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
970 /* threshold event is triggered in finer grain than soft limit */
971 if (unlikely(mem_cgroup_event_ratelimit(memcg,
972 MEM_CGROUP_TARGET_THRESH))) {
974 bool do_numainfo __maybe_unused;
976 do_softlimit = mem_cgroup_event_ratelimit(memcg,
977 MEM_CGROUP_TARGET_SOFTLIMIT);
979 do_numainfo = mem_cgroup_event_ratelimit(memcg,
980 MEM_CGROUP_TARGET_NUMAINFO);
984 mem_cgroup_threshold(memcg);
985 if (unlikely(do_softlimit))
986 mem_cgroup_update_tree(memcg, page);
988 if (unlikely(do_numainfo))
989 atomic_inc(&memcg->numainfo_events);
995 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
997 return mem_cgroup_from_css(
998 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1001 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1004 * mm_update_next_owner() may clear mm->owner to NULL
1005 * if it races with swapoff, page migration, etc.
1006 * So this can be called with p == NULL.
1011 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1014 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1016 struct mem_cgroup *memcg = NULL;
1021 * Because we have no locks, mm->owner's may be being moved to other
1022 * cgroup. We use css_tryget() here even if this looks
1023 * pessimistic (rather than adding locks here).
1027 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1028 if (unlikely(!memcg))
1030 } while (!css_tryget(&memcg->css));
1036 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1037 * @root: hierarchy root
1038 * @prev: previously returned memcg, NULL on first invocation
1039 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1041 * Returns references to children of the hierarchy below @root, or
1042 * @root itself, or %NULL after a full round-trip.
1044 * Caller must pass the return value in @prev on subsequent
1045 * invocations for reference counting, or use mem_cgroup_iter_break()
1046 * to cancel a hierarchy walk before the round-trip is complete.
1048 * Reclaimers can specify a zone and a priority level in @reclaim to
1049 * divide up the memcgs in the hierarchy among all concurrent
1050 * reclaimers operating on the same zone and priority.
1052 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1053 struct mem_cgroup *prev,
1054 struct mem_cgroup_reclaim_cookie *reclaim)
1056 struct mem_cgroup *memcg = NULL;
1059 if (mem_cgroup_disabled())
1063 root = root_mem_cgroup;
1065 if (prev && !reclaim)
1066 id = css_id(&prev->css);
1068 if (prev && prev != root)
1069 css_put(&prev->css);
1071 if (!root->use_hierarchy && root != root_mem_cgroup) {
1078 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1079 struct cgroup_subsys_state *css;
1082 int nid = zone_to_nid(reclaim->zone);
1083 int zid = zone_idx(reclaim->zone);
1084 struct mem_cgroup_per_zone *mz;
1086 mz = mem_cgroup_zoneinfo(root, nid, zid);
1087 iter = &mz->reclaim_iter[reclaim->priority];
1088 if (prev && reclaim->generation != iter->generation)
1090 id = iter->position;
1094 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
1096 if (css == &root->css || css_tryget(css))
1097 memcg = mem_cgroup_from_css(css);
1103 iter->position = id;
1106 else if (!prev && memcg)
1107 reclaim->generation = iter->generation;
1117 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1118 * @root: hierarchy root
1119 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1121 void mem_cgroup_iter_break(struct mem_cgroup *root,
1122 struct mem_cgroup *prev)
1125 root = root_mem_cgroup;
1126 if (prev && prev != root)
1127 css_put(&prev->css);
1131 * Iteration constructs for visiting all cgroups (under a tree). If
1132 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1133 * be used for reference counting.
1135 #define for_each_mem_cgroup_tree(iter, root) \
1136 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1138 iter = mem_cgroup_iter(root, iter, NULL))
1140 #define for_each_mem_cgroup(iter) \
1141 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1143 iter = mem_cgroup_iter(NULL, iter, NULL))
1145 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1147 struct mem_cgroup *memcg;
1153 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1154 if (unlikely(!memcg))
1159 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1162 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1170 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1173 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1174 * @zone: zone of the wanted lruvec
1175 * @memcg: memcg of the wanted lruvec
1177 * Returns the lru list vector holding pages for the given @zone and
1178 * @mem. This can be the global zone lruvec, if the memory controller
1181 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1182 struct mem_cgroup *memcg)
1184 struct mem_cgroup_per_zone *mz;
1186 if (mem_cgroup_disabled())
1187 return &zone->lruvec;
1189 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1194 * Following LRU functions are allowed to be used without PCG_LOCK.
1195 * Operations are called by routine of global LRU independently from memcg.
1196 * What we have to take care of here is validness of pc->mem_cgroup.
1198 * Changes to pc->mem_cgroup happens when
1201 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1202 * It is added to LRU before charge.
1203 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1204 * When moving account, the page is not on LRU. It's isolated.
1208 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1210 * @zone: zone of the page
1212 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1214 struct mem_cgroup_per_zone *mz;
1215 struct mem_cgroup *memcg;
1216 struct page_cgroup *pc;
1218 if (mem_cgroup_disabled())
1219 return &zone->lruvec;
1221 pc = lookup_page_cgroup(page);
1222 memcg = pc->mem_cgroup;
1225 * Surreptitiously switch any uncharged offlist page to root:
1226 * an uncharged page off lru does nothing to secure
1227 * its former mem_cgroup from sudden removal.
1229 * Our caller holds lru_lock, and PageCgroupUsed is updated
1230 * under page_cgroup lock: between them, they make all uses
1231 * of pc->mem_cgroup safe.
1233 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1234 pc->mem_cgroup = memcg = root_mem_cgroup;
1236 mz = page_cgroup_zoneinfo(memcg, page);
1241 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1242 * @lruvec: mem_cgroup per zone lru vector
1243 * @lru: index of lru list the page is sitting on
1244 * @nr_pages: positive when adding or negative when removing
1246 * This function must be called when a page is added to or removed from an
1249 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1252 struct mem_cgroup_per_zone *mz;
1253 unsigned long *lru_size;
1255 if (mem_cgroup_disabled())
1258 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1259 lru_size = mz->lru_size + lru;
1260 *lru_size += nr_pages;
1261 VM_BUG_ON((long)(*lru_size) < 0);
1265 * Checks whether given mem is same or in the root_mem_cgroup's
1268 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1269 struct mem_cgroup *memcg)
1271 if (root_memcg == memcg)
1273 if (!root_memcg->use_hierarchy || !memcg)
1275 return css_is_ancestor(&memcg->css, &root_memcg->css);
1278 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1279 struct mem_cgroup *memcg)
1284 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1289 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1292 struct mem_cgroup *curr = NULL;
1293 struct task_struct *p;
1295 p = find_lock_task_mm(task);
1297 curr = try_get_mem_cgroup_from_mm(p->mm);
1301 * All threads may have already detached their mm's, but the oom
1302 * killer still needs to detect if they have already been oom
1303 * killed to prevent needlessly killing additional tasks.
1306 curr = mem_cgroup_from_task(task);
1308 css_get(&curr->css);
1314 * We should check use_hierarchy of "memcg" not "curr". Because checking
1315 * use_hierarchy of "curr" here make this function true if hierarchy is
1316 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1317 * hierarchy(even if use_hierarchy is disabled in "memcg").
1319 ret = mem_cgroup_same_or_subtree(memcg, curr);
1320 css_put(&curr->css);
1324 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1326 unsigned long inactive_ratio;
1327 unsigned long inactive;
1328 unsigned long active;
1331 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1332 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1334 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1336 inactive_ratio = int_sqrt(10 * gb);
1340 return inactive * inactive_ratio < active;
1343 int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1345 unsigned long active;
1346 unsigned long inactive;
1348 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_FILE);
1349 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_FILE);
1351 return (active > inactive);
1354 #define mem_cgroup_from_res_counter(counter, member) \
1355 container_of(counter, struct mem_cgroup, member)
1358 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1359 * @memcg: the memory cgroup
1361 * Returns the maximum amount of memory @mem can be charged with, in
1364 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1366 unsigned long long margin;
1368 margin = res_counter_margin(&memcg->res);
1369 if (do_swap_account)
1370 margin = min(margin, res_counter_margin(&memcg->memsw));
1371 return margin >> PAGE_SHIFT;
1374 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1376 struct cgroup *cgrp = memcg->css.cgroup;
1379 if (cgrp->parent == NULL)
1380 return vm_swappiness;
1382 return memcg->swappiness;
1386 * memcg->moving_account is used for checking possibility that some thread is
1387 * calling move_account(). When a thread on CPU-A starts moving pages under
1388 * a memcg, other threads should check memcg->moving_account under
1389 * rcu_read_lock(), like this:
1393 * memcg->moving_account+1 if (memcg->mocing_account)
1395 * synchronize_rcu() update something.
1400 /* for quick checking without looking up memcg */
1401 atomic_t memcg_moving __read_mostly;
1403 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1405 atomic_inc(&memcg_moving);
1406 atomic_inc(&memcg->moving_account);
1410 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1413 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1414 * We check NULL in callee rather than caller.
1417 atomic_dec(&memcg_moving);
1418 atomic_dec(&memcg->moving_account);
1423 * 2 routines for checking "mem" is under move_account() or not.
1425 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1426 * is used for avoiding races in accounting. If true,
1427 * pc->mem_cgroup may be overwritten.
1429 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1430 * under hierarchy of moving cgroups. This is for
1431 * waiting at hith-memory prressure caused by "move".
1434 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1436 VM_BUG_ON(!rcu_read_lock_held());
1437 return atomic_read(&memcg->moving_account) > 0;
1440 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1442 struct mem_cgroup *from;
1443 struct mem_cgroup *to;
1446 * Unlike task_move routines, we access mc.to, mc.from not under
1447 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1449 spin_lock(&mc.lock);
1455 ret = mem_cgroup_same_or_subtree(memcg, from)
1456 || mem_cgroup_same_or_subtree(memcg, to);
1458 spin_unlock(&mc.lock);
1462 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1464 if (mc.moving_task && current != mc.moving_task) {
1465 if (mem_cgroup_under_move(memcg)) {
1467 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1468 /* moving charge context might have finished. */
1471 finish_wait(&mc.waitq, &wait);
1479 * Take this lock when
1480 * - a code tries to modify page's memcg while it's USED.
1481 * - a code tries to modify page state accounting in a memcg.
1482 * see mem_cgroup_stolen(), too.
1484 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1485 unsigned long *flags)
1487 spin_lock_irqsave(&memcg->move_lock, *flags);
1490 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1491 unsigned long *flags)
1493 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1497 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1498 * @memcg: The memory cgroup that went over limit
1499 * @p: Task that is going to be killed
1501 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1504 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1506 struct cgroup *task_cgrp;
1507 struct cgroup *mem_cgrp;
1509 * Need a buffer in BSS, can't rely on allocations. The code relies
1510 * on the assumption that OOM is serialized for memory controller.
1511 * If this assumption is broken, revisit this code.
1513 static char memcg_name[PATH_MAX];
1521 mem_cgrp = memcg->css.cgroup;
1522 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1524 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1527 * Unfortunately, we are unable to convert to a useful name
1528 * But we'll still print out the usage information
1535 printk(KERN_INFO "Task in %s killed", memcg_name);
1538 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1546 * Continues from above, so we don't need an KERN_ level
1548 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1551 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1552 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1553 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1554 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1555 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1557 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1558 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1559 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1560 printk(KERN_INFO "kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1561 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1562 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1563 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1567 * This function returns the number of memcg under hierarchy tree. Returns
1568 * 1(self count) if no children.
1570 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1573 struct mem_cgroup *iter;
1575 for_each_mem_cgroup_tree(iter, memcg)
1581 * Return the memory (and swap, if configured) limit for a memcg.
1583 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1587 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1590 * Do not consider swap space if we cannot swap due to swappiness
1592 if (mem_cgroup_swappiness(memcg)) {
1595 limit += total_swap_pages << PAGE_SHIFT;
1596 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1599 * If memsw is finite and limits the amount of swap space
1600 * available to this memcg, return that limit.
1602 limit = min(limit, memsw);
1608 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1611 struct mem_cgroup *iter;
1612 unsigned long chosen_points = 0;
1613 unsigned long totalpages;
1614 unsigned int points = 0;
1615 struct task_struct *chosen = NULL;
1618 * If current has a pending SIGKILL, then automatically select it. The
1619 * goal is to allow it to allocate so that it may quickly exit and free
1622 if (fatal_signal_pending(current)) {
1623 set_thread_flag(TIF_MEMDIE);
1627 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1628 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1629 for_each_mem_cgroup_tree(iter, memcg) {
1630 struct cgroup *cgroup = iter->css.cgroup;
1631 struct cgroup_iter it;
1632 struct task_struct *task;
1634 cgroup_iter_start(cgroup, &it);
1635 while ((task = cgroup_iter_next(cgroup, &it))) {
1636 switch (oom_scan_process_thread(task, totalpages, NULL,
1638 case OOM_SCAN_SELECT:
1640 put_task_struct(chosen);
1642 chosen_points = ULONG_MAX;
1643 get_task_struct(chosen);
1645 case OOM_SCAN_CONTINUE:
1647 case OOM_SCAN_ABORT:
1648 cgroup_iter_end(cgroup, &it);
1649 mem_cgroup_iter_break(memcg, iter);
1651 put_task_struct(chosen);
1656 points = oom_badness(task, memcg, NULL, totalpages);
1657 if (points > chosen_points) {
1659 put_task_struct(chosen);
1661 chosen_points = points;
1662 get_task_struct(chosen);
1665 cgroup_iter_end(cgroup, &it);
1670 points = chosen_points * 1000 / totalpages;
1671 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1672 NULL, "Memory cgroup out of memory");
1675 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1677 unsigned long flags)
1679 unsigned long total = 0;
1680 bool noswap = false;
1683 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1685 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1688 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1690 drain_all_stock_async(memcg);
1691 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1693 * Allow limit shrinkers, which are triggered directly
1694 * by userspace, to catch signals and stop reclaim
1695 * after minimal progress, regardless of the margin.
1697 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1699 if (mem_cgroup_margin(memcg))
1702 * If nothing was reclaimed after two attempts, there
1703 * may be no reclaimable pages in this hierarchy.
1712 * test_mem_cgroup_node_reclaimable
1713 * @memcg: the target memcg
1714 * @nid: the node ID to be checked.
1715 * @noswap : specify true here if the user wants flle only information.
1717 * This function returns whether the specified memcg contains any
1718 * reclaimable pages on a node. Returns true if there are any reclaimable
1719 * pages in the node.
1721 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1722 int nid, bool noswap)
1724 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1726 if (noswap || !total_swap_pages)
1728 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1733 #if MAX_NUMNODES > 1
1736 * Always updating the nodemask is not very good - even if we have an empty
1737 * list or the wrong list here, we can start from some node and traverse all
1738 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1741 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1745 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1746 * pagein/pageout changes since the last update.
1748 if (!atomic_read(&memcg->numainfo_events))
1750 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1753 /* make a nodemask where this memcg uses memory from */
1754 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1756 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1758 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1759 node_clear(nid, memcg->scan_nodes);
1762 atomic_set(&memcg->numainfo_events, 0);
1763 atomic_set(&memcg->numainfo_updating, 0);
1767 * Selecting a node where we start reclaim from. Because what we need is just
1768 * reducing usage counter, start from anywhere is O,K. Considering
1769 * memory reclaim from current node, there are pros. and cons.
1771 * Freeing memory from current node means freeing memory from a node which
1772 * we'll use or we've used. So, it may make LRU bad. And if several threads
1773 * hit limits, it will see a contention on a node. But freeing from remote
1774 * node means more costs for memory reclaim because of memory latency.
1776 * Now, we use round-robin. Better algorithm is welcomed.
1778 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1782 mem_cgroup_may_update_nodemask(memcg);
1783 node = memcg->last_scanned_node;
1785 node = next_node(node, memcg->scan_nodes);
1786 if (node == MAX_NUMNODES)
1787 node = first_node(memcg->scan_nodes);
1789 * We call this when we hit limit, not when pages are added to LRU.
1790 * No LRU may hold pages because all pages are UNEVICTABLE or
1791 * memcg is too small and all pages are not on LRU. In that case,
1792 * we use curret node.
1794 if (unlikely(node == MAX_NUMNODES))
1795 node = numa_node_id();
1797 memcg->last_scanned_node = node;
1802 * Check all nodes whether it contains reclaimable pages or not.
1803 * For quick scan, we make use of scan_nodes. This will allow us to skip
1804 * unused nodes. But scan_nodes is lazily updated and may not cotain
1805 * enough new information. We need to do double check.
1807 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1812 * quick check...making use of scan_node.
1813 * We can skip unused nodes.
1815 if (!nodes_empty(memcg->scan_nodes)) {
1816 for (nid = first_node(memcg->scan_nodes);
1818 nid = next_node(nid, memcg->scan_nodes)) {
1820 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1825 * Check rest of nodes.
1827 for_each_node_state(nid, N_HIGH_MEMORY) {
1828 if (node_isset(nid, memcg->scan_nodes))
1830 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1837 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1842 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1844 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1848 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1851 unsigned long *total_scanned)
1853 struct mem_cgroup *victim = NULL;
1856 unsigned long excess;
1857 unsigned long nr_scanned;
1858 struct mem_cgroup_reclaim_cookie reclaim = {
1863 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1866 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1871 * If we have not been able to reclaim
1872 * anything, it might because there are
1873 * no reclaimable pages under this hierarchy
1878 * We want to do more targeted reclaim.
1879 * excess >> 2 is not to excessive so as to
1880 * reclaim too much, nor too less that we keep
1881 * coming back to reclaim from this cgroup
1883 if (total >= (excess >> 2) ||
1884 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1889 if (!mem_cgroup_reclaimable(victim, false))
1891 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1893 *total_scanned += nr_scanned;
1894 if (!res_counter_soft_limit_excess(&root_memcg->res))
1897 mem_cgroup_iter_break(root_memcg, victim);
1902 * Check OOM-Killer is already running under our hierarchy.
1903 * If someone is running, return false.
1904 * Has to be called with memcg_oom_lock
1906 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1908 struct mem_cgroup *iter, *failed = NULL;
1910 for_each_mem_cgroup_tree(iter, memcg) {
1911 if (iter->oom_lock) {
1913 * this subtree of our hierarchy is already locked
1914 * so we cannot give a lock.
1917 mem_cgroup_iter_break(memcg, iter);
1920 iter->oom_lock = true;
1927 * OK, we failed to lock the whole subtree so we have to clean up
1928 * what we set up to the failing subtree
1930 for_each_mem_cgroup_tree(iter, memcg) {
1931 if (iter == failed) {
1932 mem_cgroup_iter_break(memcg, iter);
1935 iter->oom_lock = false;
1941 * Has to be called with memcg_oom_lock
1943 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1945 struct mem_cgroup *iter;
1947 for_each_mem_cgroup_tree(iter, memcg)
1948 iter->oom_lock = false;
1952 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1954 struct mem_cgroup *iter;
1956 for_each_mem_cgroup_tree(iter, memcg)
1957 atomic_inc(&iter->under_oom);
1960 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1962 struct mem_cgroup *iter;
1965 * When a new child is created while the hierarchy is under oom,
1966 * mem_cgroup_oom_lock() may not be called. We have to use
1967 * atomic_add_unless() here.
1969 for_each_mem_cgroup_tree(iter, memcg)
1970 atomic_add_unless(&iter->under_oom, -1, 0);
1973 static DEFINE_SPINLOCK(memcg_oom_lock);
1974 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1976 struct oom_wait_info {
1977 struct mem_cgroup *memcg;
1981 static int memcg_oom_wake_function(wait_queue_t *wait,
1982 unsigned mode, int sync, void *arg)
1984 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1985 struct mem_cgroup *oom_wait_memcg;
1986 struct oom_wait_info *oom_wait_info;
1988 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1989 oom_wait_memcg = oom_wait_info->memcg;
1992 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1993 * Then we can use css_is_ancestor without taking care of RCU.
1995 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1996 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1998 return autoremove_wake_function(wait, mode, sync, arg);
2001 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2003 /* for filtering, pass "memcg" as argument. */
2004 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2007 static void memcg_oom_recover(struct mem_cgroup *memcg)
2009 if (memcg && atomic_read(&memcg->under_oom))
2010 memcg_wakeup_oom(memcg);
2014 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2016 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2019 struct oom_wait_info owait;
2020 bool locked, need_to_kill;
2022 owait.memcg = memcg;
2023 owait.wait.flags = 0;
2024 owait.wait.func = memcg_oom_wake_function;
2025 owait.wait.private = current;
2026 INIT_LIST_HEAD(&owait.wait.task_list);
2027 need_to_kill = true;
2028 mem_cgroup_mark_under_oom(memcg);
2030 /* At first, try to OOM lock hierarchy under memcg.*/
2031 spin_lock(&memcg_oom_lock);
2032 locked = mem_cgroup_oom_lock(memcg);
2034 * Even if signal_pending(), we can't quit charge() loop without
2035 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2036 * under OOM is always welcomed, use TASK_KILLABLE here.
2038 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2039 if (!locked || memcg->oom_kill_disable)
2040 need_to_kill = false;
2042 mem_cgroup_oom_notify(memcg);
2043 spin_unlock(&memcg_oom_lock);
2046 finish_wait(&memcg_oom_waitq, &owait.wait);
2047 mem_cgroup_out_of_memory(memcg, mask, order);
2050 finish_wait(&memcg_oom_waitq, &owait.wait);
2052 spin_lock(&memcg_oom_lock);
2054 mem_cgroup_oom_unlock(memcg);
2055 memcg_wakeup_oom(memcg);
2056 spin_unlock(&memcg_oom_lock);
2058 mem_cgroup_unmark_under_oom(memcg);
2060 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2062 /* Give chance to dying process */
2063 schedule_timeout_uninterruptible(1);
2068 * Currently used to update mapped file statistics, but the routine can be
2069 * generalized to update other statistics as well.
2071 * Notes: Race condition
2073 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2074 * it tends to be costly. But considering some conditions, we doesn't need
2075 * to do so _always_.
2077 * Considering "charge", lock_page_cgroup() is not required because all
2078 * file-stat operations happen after a page is attached to radix-tree. There
2079 * are no race with "charge".
2081 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2082 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2083 * if there are race with "uncharge". Statistics itself is properly handled
2086 * Considering "move", this is an only case we see a race. To make the race
2087 * small, we check mm->moving_account and detect there are possibility of race
2088 * If there is, we take a lock.
2091 void __mem_cgroup_begin_update_page_stat(struct page *page,
2092 bool *locked, unsigned long *flags)
2094 struct mem_cgroup *memcg;
2095 struct page_cgroup *pc;
2097 pc = lookup_page_cgroup(page);
2099 memcg = pc->mem_cgroup;
2100 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2103 * If this memory cgroup is not under account moving, we don't
2104 * need to take move_lock_mem_cgroup(). Because we already hold
2105 * rcu_read_lock(), any calls to move_account will be delayed until
2106 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2108 if (!mem_cgroup_stolen(memcg))
2111 move_lock_mem_cgroup(memcg, flags);
2112 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2113 move_unlock_mem_cgroup(memcg, flags);
2119 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2121 struct page_cgroup *pc = lookup_page_cgroup(page);
2124 * It's guaranteed that pc->mem_cgroup never changes while
2125 * lock is held because a routine modifies pc->mem_cgroup
2126 * should take move_lock_mem_cgroup().
2128 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2131 void mem_cgroup_update_page_stat(struct page *page,
2132 enum mem_cgroup_page_stat_item idx, int val)
2134 struct mem_cgroup *memcg;
2135 struct page_cgroup *pc = lookup_page_cgroup(page);
2136 unsigned long uninitialized_var(flags);
2138 if (mem_cgroup_disabled())
2141 memcg = pc->mem_cgroup;
2142 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2146 case MEMCG_NR_FILE_MAPPED:
2147 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2153 this_cpu_add(memcg->stat->count[idx], val);
2157 * size of first charge trial. "32" comes from vmscan.c's magic value.
2158 * TODO: maybe necessary to use big numbers in big irons.
2160 #define CHARGE_BATCH 32U
2161 struct memcg_stock_pcp {
2162 struct mem_cgroup *cached; /* this never be root cgroup */
2163 unsigned int nr_pages;
2164 struct work_struct work;
2165 unsigned long flags;
2166 #define FLUSHING_CACHED_CHARGE 0
2168 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2169 static DEFINE_MUTEX(percpu_charge_mutex);
2172 * consume_stock: Try to consume stocked charge on this cpu.
2173 * @memcg: memcg to consume from.
2174 * @nr_pages: how many pages to charge.
2176 * The charges will only happen if @memcg matches the current cpu's memcg
2177 * stock, and at least @nr_pages are available in that stock. Failure to
2178 * service an allocation will refill the stock.
2180 * returns true if successful, false otherwise.
2182 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2184 struct memcg_stock_pcp *stock;
2187 if (nr_pages > CHARGE_BATCH)
2190 stock = &get_cpu_var(memcg_stock);
2191 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2192 stock->nr_pages -= nr_pages;
2193 else /* need to call res_counter_charge */
2195 put_cpu_var(memcg_stock);
2200 * Returns stocks cached in percpu to res_counter and reset cached information.
2202 static void drain_stock(struct memcg_stock_pcp *stock)
2204 struct mem_cgroup *old = stock->cached;
2206 if (stock->nr_pages) {
2207 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2209 res_counter_uncharge(&old->res, bytes);
2210 if (do_swap_account)
2211 res_counter_uncharge(&old->memsw, bytes);
2212 stock->nr_pages = 0;
2214 stock->cached = NULL;
2218 * This must be called under preempt disabled or must be called by
2219 * a thread which is pinned to local cpu.
2221 static void drain_local_stock(struct work_struct *dummy)
2223 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2225 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2229 * Cache charges(val) which is from res_counter, to local per_cpu area.
2230 * This will be consumed by consume_stock() function, later.
2232 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2234 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2236 if (stock->cached != memcg) { /* reset if necessary */
2238 stock->cached = memcg;
2240 stock->nr_pages += nr_pages;
2241 put_cpu_var(memcg_stock);
2245 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2246 * of the hierarchy under it. sync flag says whether we should block
2247 * until the work is done.
2249 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2253 /* Notify other cpus that system-wide "drain" is running */
2256 for_each_online_cpu(cpu) {
2257 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2258 struct mem_cgroup *memcg;
2260 memcg = stock->cached;
2261 if (!memcg || !stock->nr_pages)
2263 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2265 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2267 drain_local_stock(&stock->work);
2269 schedule_work_on(cpu, &stock->work);
2277 for_each_online_cpu(cpu) {
2278 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2279 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2280 flush_work(&stock->work);
2287 * Tries to drain stocked charges in other cpus. This function is asynchronous
2288 * and just put a work per cpu for draining localy on each cpu. Caller can
2289 * expects some charges will be back to res_counter later but cannot wait for
2292 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2295 * If someone calls draining, avoid adding more kworker runs.
2297 if (!mutex_trylock(&percpu_charge_mutex))
2299 drain_all_stock(root_memcg, false);
2300 mutex_unlock(&percpu_charge_mutex);
2303 /* This is a synchronous drain interface. */
2304 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2306 /* called when force_empty is called */
2307 mutex_lock(&percpu_charge_mutex);
2308 drain_all_stock(root_memcg, true);
2309 mutex_unlock(&percpu_charge_mutex);
2313 * This function drains percpu counter value from DEAD cpu and
2314 * move it to local cpu. Note that this function can be preempted.
2316 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2320 spin_lock(&memcg->pcp_counter_lock);
2321 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2322 long x = per_cpu(memcg->stat->count[i], cpu);
2324 per_cpu(memcg->stat->count[i], cpu) = 0;
2325 memcg->nocpu_base.count[i] += x;
2327 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2328 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2330 per_cpu(memcg->stat->events[i], cpu) = 0;
2331 memcg->nocpu_base.events[i] += x;
2333 spin_unlock(&memcg->pcp_counter_lock);
2336 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2337 unsigned long action,
2340 int cpu = (unsigned long)hcpu;
2341 struct memcg_stock_pcp *stock;
2342 struct mem_cgroup *iter;
2344 if (action == CPU_ONLINE)
2347 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2350 for_each_mem_cgroup(iter)
2351 mem_cgroup_drain_pcp_counter(iter, cpu);
2353 stock = &per_cpu(memcg_stock, cpu);
2359 /* See __mem_cgroup_try_charge() for details */
2361 CHARGE_OK, /* success */
2362 CHARGE_RETRY, /* need to retry but retry is not bad */
2363 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2364 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2365 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2368 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2369 unsigned int nr_pages, unsigned int min_pages,
2372 unsigned long csize = nr_pages * PAGE_SIZE;
2373 struct mem_cgroup *mem_over_limit;
2374 struct res_counter *fail_res;
2375 unsigned long flags = 0;
2378 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2381 if (!do_swap_account)
2383 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2387 res_counter_uncharge(&memcg->res, csize);
2388 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2389 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2391 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2393 * Never reclaim on behalf of optional batching, retry with a
2394 * single page instead.
2396 if (nr_pages > min_pages)
2397 return CHARGE_RETRY;
2399 if (!(gfp_mask & __GFP_WAIT))
2400 return CHARGE_WOULDBLOCK;
2402 if (gfp_mask & __GFP_NORETRY)
2403 return CHARGE_NOMEM;
2405 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2406 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2407 return CHARGE_RETRY;
2409 * Even though the limit is exceeded at this point, reclaim
2410 * may have been able to free some pages. Retry the charge
2411 * before killing the task.
2413 * Only for regular pages, though: huge pages are rather
2414 * unlikely to succeed so close to the limit, and we fall back
2415 * to regular pages anyway in case of failure.
2417 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2418 return CHARGE_RETRY;
2421 * At task move, charge accounts can be doubly counted. So, it's
2422 * better to wait until the end of task_move if something is going on.
2424 if (mem_cgroup_wait_acct_move(mem_over_limit))
2425 return CHARGE_RETRY;
2427 /* If we don't need to call oom-killer at el, return immediately */
2429 return CHARGE_NOMEM;
2431 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2432 return CHARGE_OOM_DIE;
2434 return CHARGE_RETRY;
2438 * __mem_cgroup_try_charge() does
2439 * 1. detect memcg to be charged against from passed *mm and *ptr,
2440 * 2. update res_counter
2441 * 3. call memory reclaim if necessary.
2443 * In some special case, if the task is fatal, fatal_signal_pending() or
2444 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2445 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2446 * as possible without any hazards. 2: all pages should have a valid
2447 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2448 * pointer, that is treated as a charge to root_mem_cgroup.
2450 * So __mem_cgroup_try_charge() will return
2451 * 0 ... on success, filling *ptr with a valid memcg pointer.
2452 * -ENOMEM ... charge failure because of resource limits.
2453 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2455 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2456 * the oom-killer can be invoked.
2458 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2460 unsigned int nr_pages,
2461 struct mem_cgroup **ptr,
2464 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2465 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2466 struct mem_cgroup *memcg = NULL;
2470 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2471 * in system level. So, allow to go ahead dying process in addition to
2474 if (unlikely(test_thread_flag(TIF_MEMDIE)
2475 || fatal_signal_pending(current)))
2479 * We always charge the cgroup the mm_struct belongs to.
2480 * The mm_struct's mem_cgroup changes on task migration if the
2481 * thread group leader migrates. It's possible that mm is not
2482 * set, if so charge the root memcg (happens for pagecache usage).
2485 *ptr = root_mem_cgroup;
2487 if (*ptr) { /* css should be a valid one */
2489 if (mem_cgroup_is_root(memcg))
2491 if (consume_stock(memcg, nr_pages))
2493 css_get(&memcg->css);
2495 struct task_struct *p;
2498 p = rcu_dereference(mm->owner);
2500 * Because we don't have task_lock(), "p" can exit.
2501 * In that case, "memcg" can point to root or p can be NULL with
2502 * race with swapoff. Then, we have small risk of mis-accouning.
2503 * But such kind of mis-account by race always happens because
2504 * we don't have cgroup_mutex(). It's overkill and we allo that
2506 * (*) swapoff at el will charge against mm-struct not against
2507 * task-struct. So, mm->owner can be NULL.
2509 memcg = mem_cgroup_from_task(p);
2511 memcg = root_mem_cgroup;
2512 if (mem_cgroup_is_root(memcg)) {
2516 if (consume_stock(memcg, nr_pages)) {
2518 * It seems dagerous to access memcg without css_get().
2519 * But considering how consume_stok works, it's not
2520 * necessary. If consume_stock success, some charges
2521 * from this memcg are cached on this cpu. So, we
2522 * don't need to call css_get()/css_tryget() before
2523 * calling consume_stock().
2528 /* after here, we may be blocked. we need to get refcnt */
2529 if (!css_tryget(&memcg->css)) {
2539 /* If killed, bypass charge */
2540 if (fatal_signal_pending(current)) {
2541 css_put(&memcg->css);
2546 if (oom && !nr_oom_retries) {
2548 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2551 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2556 case CHARGE_RETRY: /* not in OOM situation but retry */
2558 css_put(&memcg->css);
2561 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2562 css_put(&memcg->css);
2564 case CHARGE_NOMEM: /* OOM routine works */
2566 css_put(&memcg->css);
2569 /* If oom, we never return -ENOMEM */
2572 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2573 css_put(&memcg->css);
2576 } while (ret != CHARGE_OK);
2578 if (batch > nr_pages)
2579 refill_stock(memcg, batch - nr_pages);
2580 css_put(&memcg->css);
2588 *ptr = root_mem_cgroup;
2593 * Somemtimes we have to undo a charge we got by try_charge().
2594 * This function is for that and do uncharge, put css's refcnt.
2595 * gotten by try_charge().
2597 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2598 unsigned int nr_pages)
2600 if (!mem_cgroup_is_root(memcg)) {
2601 unsigned long bytes = nr_pages * PAGE_SIZE;
2603 res_counter_uncharge(&memcg->res, bytes);
2604 if (do_swap_account)
2605 res_counter_uncharge(&memcg->memsw, bytes);
2610 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2611 * This is useful when moving usage to parent cgroup.
2613 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2614 unsigned int nr_pages)
2616 unsigned long bytes = nr_pages * PAGE_SIZE;
2618 if (mem_cgroup_is_root(memcg))
2621 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2622 if (do_swap_account)
2623 res_counter_uncharge_until(&memcg->memsw,
2624 memcg->memsw.parent, bytes);
2628 * A helper function to get mem_cgroup from ID. must be called under
2629 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2630 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2631 * called against removed memcg.)
2633 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2635 struct cgroup_subsys_state *css;
2637 /* ID 0 is unused ID */
2640 css = css_lookup(&mem_cgroup_subsys, id);
2643 return mem_cgroup_from_css(css);
2646 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2648 struct mem_cgroup *memcg = NULL;
2649 struct page_cgroup *pc;
2653 VM_BUG_ON(!PageLocked(page));
2655 pc = lookup_page_cgroup(page);
2656 lock_page_cgroup(pc);
2657 if (PageCgroupUsed(pc)) {
2658 memcg = pc->mem_cgroup;
2659 if (memcg && !css_tryget(&memcg->css))
2661 } else if (PageSwapCache(page)) {
2662 ent.val = page_private(page);
2663 id = lookup_swap_cgroup_id(ent);
2665 memcg = mem_cgroup_lookup(id);
2666 if (memcg && !css_tryget(&memcg->css))
2670 unlock_page_cgroup(pc);
2674 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2676 unsigned int nr_pages,
2677 enum charge_type ctype,
2680 struct page_cgroup *pc = lookup_page_cgroup(page);
2681 struct zone *uninitialized_var(zone);
2682 struct lruvec *lruvec;
2683 bool was_on_lru = false;
2686 lock_page_cgroup(pc);
2687 VM_BUG_ON(PageCgroupUsed(pc));
2689 * we don't need page_cgroup_lock about tail pages, becase they are not
2690 * accessed by any other context at this point.
2694 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2695 * may already be on some other mem_cgroup's LRU. Take care of it.
2698 zone = page_zone(page);
2699 spin_lock_irq(&zone->lru_lock);
2700 if (PageLRU(page)) {
2701 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2703 del_page_from_lru_list(page, lruvec, page_lru(page));
2708 pc->mem_cgroup = memcg;
2710 * We access a page_cgroup asynchronously without lock_page_cgroup().
2711 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2712 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2713 * before USED bit, we need memory barrier here.
2714 * See mem_cgroup_add_lru_list(), etc.
2717 SetPageCgroupUsed(pc);
2721 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2722 VM_BUG_ON(PageLRU(page));
2724 add_page_to_lru_list(page, lruvec, page_lru(page));
2726 spin_unlock_irq(&zone->lru_lock);
2729 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2734 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2735 unlock_page_cgroup(pc);
2738 * "charge_statistics" updated event counter. Then, check it.
2739 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2740 * if they exceeds softlimit.
2742 memcg_check_events(memcg, page);
2745 #ifdef CONFIG_MEMCG_KMEM
2746 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2748 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2749 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2752 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2754 struct res_counter *fail_res;
2755 struct mem_cgroup *_memcg;
2759 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2764 * Conditions under which we can wait for the oom_killer. Those are
2765 * the same conditions tested by the core page allocator
2767 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2770 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2773 if (ret == -EINTR) {
2775 * __mem_cgroup_try_charge() chosed to bypass to root due to
2776 * OOM kill or fatal signal. Since our only options are to
2777 * either fail the allocation or charge it to this cgroup, do
2778 * it as a temporary condition. But we can't fail. From a
2779 * kmem/slab perspective, the cache has already been selected,
2780 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2783 * This condition will only trigger if the task entered
2784 * memcg_charge_kmem in a sane state, but was OOM-killed during
2785 * __mem_cgroup_try_charge() above. Tasks that were already
2786 * dying when the allocation triggers should have been already
2787 * directed to the root cgroup in memcontrol.h
2789 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2790 if (do_swap_account)
2791 res_counter_charge_nofail(&memcg->memsw, size,
2795 res_counter_uncharge(&memcg->kmem, size);
2800 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2802 res_counter_uncharge(&memcg->res, size);
2803 if (do_swap_account)
2804 res_counter_uncharge(&memcg->memsw, size);
2807 if (res_counter_uncharge(&memcg->kmem, size))
2810 if (memcg_kmem_test_and_clear_dead(memcg))
2811 mem_cgroup_put(memcg);
2814 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2819 mutex_lock(&memcg->slab_caches_mutex);
2820 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2821 mutex_unlock(&memcg->slab_caches_mutex);
2825 * helper for acessing a memcg's index. It will be used as an index in the
2826 * child cache array in kmem_cache, and also to derive its name. This function
2827 * will return -1 when this is not a kmem-limited memcg.
2829 int memcg_cache_id(struct mem_cgroup *memcg)
2831 return memcg ? memcg->kmemcg_id : -1;
2835 * This ends up being protected by the set_limit mutex, during normal
2836 * operation, because that is its main call site.
2838 * But when we create a new cache, we can call this as well if its parent
2839 * is kmem-limited. That will have to hold set_limit_mutex as well.
2841 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2845 num = ida_simple_get(&kmem_limited_groups,
2846 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2850 * After this point, kmem_accounted (that we test atomically in
2851 * the beginning of this conditional), is no longer 0. This
2852 * guarantees only one process will set the following boolean
2853 * to true. We don't need test_and_set because we're protected
2854 * by the set_limit_mutex anyway.
2856 memcg_kmem_set_activated(memcg);
2858 ret = memcg_update_all_caches(num+1);
2860 ida_simple_remove(&kmem_limited_groups, num);
2861 memcg_kmem_clear_activated(memcg);
2865 memcg->kmemcg_id = num;
2866 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2867 mutex_init(&memcg->slab_caches_mutex);
2871 static size_t memcg_caches_array_size(int num_groups)
2874 if (num_groups <= 0)
2877 size = 2 * num_groups;
2878 if (size < MEMCG_CACHES_MIN_SIZE)
2879 size = MEMCG_CACHES_MIN_SIZE;
2880 else if (size > MEMCG_CACHES_MAX_SIZE)
2881 size = MEMCG_CACHES_MAX_SIZE;
2887 * We should update the current array size iff all caches updates succeed. This
2888 * can only be done from the slab side. The slab mutex needs to be held when
2891 void memcg_update_array_size(int num)
2893 if (num > memcg_limited_groups_array_size)
2894 memcg_limited_groups_array_size = memcg_caches_array_size(num);
2897 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
2899 struct memcg_cache_params *cur_params = s->memcg_params;
2901 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
2903 if (num_groups > memcg_limited_groups_array_size) {
2905 ssize_t size = memcg_caches_array_size(num_groups);
2907 size *= sizeof(void *);
2908 size += sizeof(struct memcg_cache_params);
2910 s->memcg_params = kzalloc(size, GFP_KERNEL);
2911 if (!s->memcg_params) {
2912 s->memcg_params = cur_params;
2916 s->memcg_params->is_root_cache = true;
2919 * There is the chance it will be bigger than
2920 * memcg_limited_groups_array_size, if we failed an allocation
2921 * in a cache, in which case all caches updated before it, will
2922 * have a bigger array.
2924 * But if that is the case, the data after
2925 * memcg_limited_groups_array_size is certainly unused
2927 for (i = 0; i < memcg_limited_groups_array_size; i++) {
2928 if (!cur_params->memcg_caches[i])
2930 s->memcg_params->memcg_caches[i] =
2931 cur_params->memcg_caches[i];
2935 * Ideally, we would wait until all caches succeed, and only
2936 * then free the old one. But this is not worth the extra
2937 * pointer per-cache we'd have to have for this.
2939 * It is not a big deal if some caches are left with a size
2940 * bigger than the others. And all updates will reset this
2948 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s)
2950 size_t size = sizeof(struct memcg_cache_params);
2952 if (!memcg_kmem_enabled())
2956 size += memcg_limited_groups_array_size * sizeof(void *);
2958 s->memcg_params = kzalloc(size, GFP_KERNEL);
2959 if (!s->memcg_params)
2963 s->memcg_params->memcg = memcg;
2967 void memcg_release_cache(struct kmem_cache *s)
2969 kfree(s->memcg_params);
2973 * We need to verify if the allocation against current->mm->owner's memcg is
2974 * possible for the given order. But the page is not allocated yet, so we'll
2975 * need a further commit step to do the final arrangements.
2977 * It is possible for the task to switch cgroups in this mean time, so at
2978 * commit time, we can't rely on task conversion any longer. We'll then use
2979 * the handle argument to return to the caller which cgroup we should commit
2980 * against. We could also return the memcg directly and avoid the pointer
2981 * passing, but a boolean return value gives better semantics considering
2982 * the compiled-out case as well.
2984 * Returning true means the allocation is possible.
2987 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
2989 struct mem_cgroup *memcg;
2993 memcg = try_get_mem_cgroup_from_mm(current->mm);
2996 * very rare case described in mem_cgroup_from_task. Unfortunately there
2997 * isn't much we can do without complicating this too much, and it would
2998 * be gfp-dependent anyway. Just let it go
3000 if (unlikely(!memcg))
3003 if (!memcg_can_account_kmem(memcg)) {
3004 css_put(&memcg->css);
3008 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3012 css_put(&memcg->css);
3016 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3019 struct page_cgroup *pc;
3021 VM_BUG_ON(mem_cgroup_is_root(memcg));
3023 /* The page allocation failed. Revert */
3025 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3029 pc = lookup_page_cgroup(page);
3030 lock_page_cgroup(pc);
3031 pc->mem_cgroup = memcg;
3032 SetPageCgroupUsed(pc);
3033 unlock_page_cgroup(pc);
3036 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3038 struct mem_cgroup *memcg = NULL;
3039 struct page_cgroup *pc;
3042 pc = lookup_page_cgroup(page);
3044 * Fast unlocked return. Theoretically might have changed, have to
3045 * check again after locking.
3047 if (!PageCgroupUsed(pc))
3050 lock_page_cgroup(pc);
3051 if (PageCgroupUsed(pc)) {
3052 memcg = pc->mem_cgroup;
3053 ClearPageCgroupUsed(pc);
3055 unlock_page_cgroup(pc);
3058 * We trust that only if there is a memcg associated with the page, it
3059 * is a valid allocation
3064 VM_BUG_ON(mem_cgroup_is_root(memcg));
3065 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3067 #endif /* CONFIG_MEMCG_KMEM */
3069 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3071 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3073 * Because tail pages are not marked as "used", set it. We're under
3074 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3075 * charge/uncharge will be never happen and move_account() is done under
3076 * compound_lock(), so we don't have to take care of races.
3078 void mem_cgroup_split_huge_fixup(struct page *head)
3080 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3081 struct page_cgroup *pc;
3084 if (mem_cgroup_disabled())
3086 for (i = 1; i < HPAGE_PMD_NR; i++) {
3088 pc->mem_cgroup = head_pc->mem_cgroup;
3089 smp_wmb();/* see __commit_charge() */
3090 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3093 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3096 * mem_cgroup_move_account - move account of the page
3098 * @nr_pages: number of regular pages (>1 for huge pages)
3099 * @pc: page_cgroup of the page.
3100 * @from: mem_cgroup which the page is moved from.
3101 * @to: mem_cgroup which the page is moved to. @from != @to.
3103 * The caller must confirm following.
3104 * - page is not on LRU (isolate_page() is useful.)
3105 * - compound_lock is held when nr_pages > 1
3107 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3110 static int mem_cgroup_move_account(struct page *page,
3111 unsigned int nr_pages,
3112 struct page_cgroup *pc,
3113 struct mem_cgroup *from,
3114 struct mem_cgroup *to)
3116 unsigned long flags;
3118 bool anon = PageAnon(page);
3120 VM_BUG_ON(from == to);
3121 VM_BUG_ON(PageLRU(page));
3123 * The page is isolated from LRU. So, collapse function
3124 * will not handle this page. But page splitting can happen.
3125 * Do this check under compound_page_lock(). The caller should
3129 if (nr_pages > 1 && !PageTransHuge(page))
3132 lock_page_cgroup(pc);
3135 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3138 move_lock_mem_cgroup(from, &flags);
3140 if (!anon && page_mapped(page)) {
3141 /* Update mapped_file data for mem_cgroup */
3143 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3144 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3147 mem_cgroup_charge_statistics(from, anon, -nr_pages);
3149 /* caller should have done css_get */
3150 pc->mem_cgroup = to;
3151 mem_cgroup_charge_statistics(to, anon, nr_pages);
3152 move_unlock_mem_cgroup(from, &flags);
3155 unlock_page_cgroup(pc);
3159 memcg_check_events(to, page);
3160 memcg_check_events(from, page);
3166 * mem_cgroup_move_parent - moves page to the parent group
3167 * @page: the page to move
3168 * @pc: page_cgroup of the page
3169 * @child: page's cgroup
3171 * move charges to its parent or the root cgroup if the group has no
3172 * parent (aka use_hierarchy==0).
3173 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3174 * mem_cgroup_move_account fails) the failure is always temporary and
3175 * it signals a race with a page removal/uncharge or migration. In the
3176 * first case the page is on the way out and it will vanish from the LRU
3177 * on the next attempt and the call should be retried later.
3178 * Isolation from the LRU fails only if page has been isolated from
3179 * the LRU since we looked at it and that usually means either global
3180 * reclaim or migration going on. The page will either get back to the
3182 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3183 * (!PageCgroupUsed) or moved to a different group. The page will
3184 * disappear in the next attempt.
3186 static int mem_cgroup_move_parent(struct page *page,
3187 struct page_cgroup *pc,
3188 struct mem_cgroup *child)
3190 struct mem_cgroup *parent;
3191 unsigned int nr_pages;
3192 unsigned long uninitialized_var(flags);
3195 VM_BUG_ON(mem_cgroup_is_root(child));
3198 if (!get_page_unless_zero(page))
3200 if (isolate_lru_page(page))
3203 nr_pages = hpage_nr_pages(page);
3205 parent = parent_mem_cgroup(child);
3207 * If no parent, move charges to root cgroup.
3210 parent = root_mem_cgroup;
3213 VM_BUG_ON(!PageTransHuge(page));
3214 flags = compound_lock_irqsave(page);
3217 ret = mem_cgroup_move_account(page, nr_pages,
3220 __mem_cgroup_cancel_local_charge(child, nr_pages);
3223 compound_unlock_irqrestore(page, flags);
3224 putback_lru_page(page);
3232 * Charge the memory controller for page usage.
3234 * 0 if the charge was successful
3235 * < 0 if the cgroup is over its limit
3237 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3238 gfp_t gfp_mask, enum charge_type ctype)
3240 struct mem_cgroup *memcg = NULL;
3241 unsigned int nr_pages = 1;
3245 if (PageTransHuge(page)) {
3246 nr_pages <<= compound_order(page);
3247 VM_BUG_ON(!PageTransHuge(page));
3249 * Never OOM-kill a process for a huge page. The
3250 * fault handler will fall back to regular pages.
3255 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3258 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3262 int mem_cgroup_newpage_charge(struct page *page,
3263 struct mm_struct *mm, gfp_t gfp_mask)
3265 if (mem_cgroup_disabled())
3267 VM_BUG_ON(page_mapped(page));
3268 VM_BUG_ON(page->mapping && !PageAnon(page));
3270 return mem_cgroup_charge_common(page, mm, gfp_mask,
3271 MEM_CGROUP_CHARGE_TYPE_ANON);
3275 * While swap-in, try_charge -> commit or cancel, the page is locked.
3276 * And when try_charge() successfully returns, one refcnt to memcg without
3277 * struct page_cgroup is acquired. This refcnt will be consumed by
3278 * "commit()" or removed by "cancel()"
3280 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3283 struct mem_cgroup **memcgp)
3285 struct mem_cgroup *memcg;
3286 struct page_cgroup *pc;
3289 pc = lookup_page_cgroup(page);
3291 * Every swap fault against a single page tries to charge the
3292 * page, bail as early as possible. shmem_unuse() encounters
3293 * already charged pages, too. The USED bit is protected by
3294 * the page lock, which serializes swap cache removal, which
3295 * in turn serializes uncharging.
3297 if (PageCgroupUsed(pc))
3299 if (!do_swap_account)
3301 memcg = try_get_mem_cgroup_from_page(page);
3305 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3306 css_put(&memcg->css);
3311 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3317 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3318 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3321 if (mem_cgroup_disabled())
3324 * A racing thread's fault, or swapoff, may have already
3325 * updated the pte, and even removed page from swap cache: in
3326 * those cases unuse_pte()'s pte_same() test will fail; but
3327 * there's also a KSM case which does need to charge the page.
3329 if (!PageSwapCache(page)) {
3332 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3337 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3340 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3342 if (mem_cgroup_disabled())
3346 __mem_cgroup_cancel_charge(memcg, 1);
3350 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3351 enum charge_type ctype)
3353 if (mem_cgroup_disabled())
3358 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3360 * Now swap is on-memory. This means this page may be
3361 * counted both as mem and swap....double count.
3362 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3363 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3364 * may call delete_from_swap_cache() before reach here.
3366 if (do_swap_account && PageSwapCache(page)) {
3367 swp_entry_t ent = {.val = page_private(page)};
3368 mem_cgroup_uncharge_swap(ent);
3372 void mem_cgroup_commit_charge_swapin(struct page *page,
3373 struct mem_cgroup *memcg)
3375 __mem_cgroup_commit_charge_swapin(page, memcg,
3376 MEM_CGROUP_CHARGE_TYPE_ANON);
3379 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3382 struct mem_cgroup *memcg = NULL;
3383 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3386 if (mem_cgroup_disabled())
3388 if (PageCompound(page))
3391 if (!PageSwapCache(page))
3392 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3393 else { /* page is swapcache/shmem */
3394 ret = __mem_cgroup_try_charge_swapin(mm, page,
3397 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3402 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3403 unsigned int nr_pages,
3404 const enum charge_type ctype)
3406 struct memcg_batch_info *batch = NULL;
3407 bool uncharge_memsw = true;
3409 /* If swapout, usage of swap doesn't decrease */
3410 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3411 uncharge_memsw = false;
3413 batch = ¤t->memcg_batch;
3415 * In usual, we do css_get() when we remember memcg pointer.
3416 * But in this case, we keep res->usage until end of a series of
3417 * uncharges. Then, it's ok to ignore memcg's refcnt.
3420 batch->memcg = memcg;
3422 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3423 * In those cases, all pages freed continuously can be expected to be in
3424 * the same cgroup and we have chance to coalesce uncharges.
3425 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3426 * because we want to do uncharge as soon as possible.
3429 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3430 goto direct_uncharge;
3433 goto direct_uncharge;
3436 * In typical case, batch->memcg == mem. This means we can
3437 * merge a series of uncharges to an uncharge of res_counter.
3438 * If not, we uncharge res_counter ony by one.
3440 if (batch->memcg != memcg)
3441 goto direct_uncharge;
3442 /* remember freed charge and uncharge it later */
3445 batch->memsw_nr_pages++;
3448 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3450 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3451 if (unlikely(batch->memcg != memcg))
3452 memcg_oom_recover(memcg);
3456 * uncharge if !page_mapped(page)
3458 static struct mem_cgroup *
3459 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3462 struct mem_cgroup *memcg = NULL;
3463 unsigned int nr_pages = 1;
3464 struct page_cgroup *pc;
3467 if (mem_cgroup_disabled())
3470 VM_BUG_ON(PageSwapCache(page));
3472 if (PageTransHuge(page)) {
3473 nr_pages <<= compound_order(page);
3474 VM_BUG_ON(!PageTransHuge(page));
3477 * Check if our page_cgroup is valid
3479 pc = lookup_page_cgroup(page);
3480 if (unlikely(!PageCgroupUsed(pc)))
3483 lock_page_cgroup(pc);
3485 memcg = pc->mem_cgroup;
3487 if (!PageCgroupUsed(pc))
3490 anon = PageAnon(page);
3493 case MEM_CGROUP_CHARGE_TYPE_ANON:
3495 * Generally PageAnon tells if it's the anon statistics to be
3496 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3497 * used before page reached the stage of being marked PageAnon.
3501 case MEM_CGROUP_CHARGE_TYPE_DROP:
3502 /* See mem_cgroup_prepare_migration() */
3503 if (page_mapped(page))
3506 * Pages under migration may not be uncharged. But
3507 * end_migration() /must/ be the one uncharging the
3508 * unused post-migration page and so it has to call
3509 * here with the migration bit still set. See the
3510 * res_counter handling below.
3512 if (!end_migration && PageCgroupMigration(pc))
3515 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3516 if (!PageAnon(page)) { /* Shared memory */
3517 if (page->mapping && !page_is_file_cache(page))
3519 } else if (page_mapped(page)) /* Anon */
3526 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3528 ClearPageCgroupUsed(pc);
3530 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3531 * freed from LRU. This is safe because uncharged page is expected not
3532 * to be reused (freed soon). Exception is SwapCache, it's handled by
3533 * special functions.
3536 unlock_page_cgroup(pc);
3538 * even after unlock, we have memcg->res.usage here and this memcg
3539 * will never be freed.
3541 memcg_check_events(memcg, page);
3542 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3543 mem_cgroup_swap_statistics(memcg, true);
3544 mem_cgroup_get(memcg);
3547 * Migration does not charge the res_counter for the
3548 * replacement page, so leave it alone when phasing out the
3549 * page that is unused after the migration.
3551 if (!end_migration && !mem_cgroup_is_root(memcg))
3552 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3557 unlock_page_cgroup(pc);
3561 void mem_cgroup_uncharge_page(struct page *page)
3564 if (page_mapped(page))
3566 VM_BUG_ON(page->mapping && !PageAnon(page));
3567 if (PageSwapCache(page))
3569 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
3572 void mem_cgroup_uncharge_cache_page(struct page *page)
3574 VM_BUG_ON(page_mapped(page));
3575 VM_BUG_ON(page->mapping);
3576 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
3580 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3581 * In that cases, pages are freed continuously and we can expect pages
3582 * are in the same memcg. All these calls itself limits the number of
3583 * pages freed at once, then uncharge_start/end() is called properly.
3584 * This may be called prural(2) times in a context,
3587 void mem_cgroup_uncharge_start(void)
3589 current->memcg_batch.do_batch++;
3590 /* We can do nest. */
3591 if (current->memcg_batch.do_batch == 1) {
3592 current->memcg_batch.memcg = NULL;
3593 current->memcg_batch.nr_pages = 0;
3594 current->memcg_batch.memsw_nr_pages = 0;
3598 void mem_cgroup_uncharge_end(void)
3600 struct memcg_batch_info *batch = ¤t->memcg_batch;
3602 if (!batch->do_batch)
3606 if (batch->do_batch) /* If stacked, do nothing. */
3612 * This "batch->memcg" is valid without any css_get/put etc...
3613 * bacause we hide charges behind us.
3615 if (batch->nr_pages)
3616 res_counter_uncharge(&batch->memcg->res,
3617 batch->nr_pages * PAGE_SIZE);
3618 if (batch->memsw_nr_pages)
3619 res_counter_uncharge(&batch->memcg->memsw,
3620 batch->memsw_nr_pages * PAGE_SIZE);
3621 memcg_oom_recover(batch->memcg);
3622 /* forget this pointer (for sanity check) */
3623 batch->memcg = NULL;
3628 * called after __delete_from_swap_cache() and drop "page" account.
3629 * memcg information is recorded to swap_cgroup of "ent"
3632 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3634 struct mem_cgroup *memcg;
3635 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3637 if (!swapout) /* this was a swap cache but the swap is unused ! */
3638 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3640 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
3643 * record memcg information, if swapout && memcg != NULL,
3644 * mem_cgroup_get() was called in uncharge().
3646 if (do_swap_account && swapout && memcg)
3647 swap_cgroup_record(ent, css_id(&memcg->css));
3651 #ifdef CONFIG_MEMCG_SWAP
3653 * called from swap_entry_free(). remove record in swap_cgroup and
3654 * uncharge "memsw" account.
3656 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3658 struct mem_cgroup *memcg;
3661 if (!do_swap_account)
3664 id = swap_cgroup_record(ent, 0);
3666 memcg = mem_cgroup_lookup(id);
3669 * We uncharge this because swap is freed.
3670 * This memcg can be obsolete one. We avoid calling css_tryget
3672 if (!mem_cgroup_is_root(memcg))
3673 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3674 mem_cgroup_swap_statistics(memcg, false);
3675 mem_cgroup_put(memcg);
3681 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3682 * @entry: swap entry to be moved
3683 * @from: mem_cgroup which the entry is moved from
3684 * @to: mem_cgroup which the entry is moved to
3686 * It succeeds only when the swap_cgroup's record for this entry is the same
3687 * as the mem_cgroup's id of @from.
3689 * Returns 0 on success, -EINVAL on failure.
3691 * The caller must have charged to @to, IOW, called res_counter_charge() about
3692 * both res and memsw, and called css_get().
3694 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3695 struct mem_cgroup *from, struct mem_cgroup *to)
3697 unsigned short old_id, new_id;
3699 old_id = css_id(&from->css);
3700 new_id = css_id(&to->css);
3702 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3703 mem_cgroup_swap_statistics(from, false);
3704 mem_cgroup_swap_statistics(to, true);
3706 * This function is only called from task migration context now.
3707 * It postpones res_counter and refcount handling till the end
3708 * of task migration(mem_cgroup_clear_mc()) for performance
3709 * improvement. But we cannot postpone mem_cgroup_get(to)
3710 * because if the process that has been moved to @to does
3711 * swap-in, the refcount of @to might be decreased to 0.
3719 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3720 struct mem_cgroup *from, struct mem_cgroup *to)
3727 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3730 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
3731 struct mem_cgroup **memcgp)
3733 struct mem_cgroup *memcg = NULL;
3734 struct page_cgroup *pc;
3735 enum charge_type ctype;
3739 VM_BUG_ON(PageTransHuge(page));
3740 if (mem_cgroup_disabled())
3743 pc = lookup_page_cgroup(page);
3744 lock_page_cgroup(pc);
3745 if (PageCgroupUsed(pc)) {
3746 memcg = pc->mem_cgroup;
3747 css_get(&memcg->css);
3749 * At migrating an anonymous page, its mapcount goes down
3750 * to 0 and uncharge() will be called. But, even if it's fully
3751 * unmapped, migration may fail and this page has to be
3752 * charged again. We set MIGRATION flag here and delay uncharge
3753 * until end_migration() is called
3755 * Corner Case Thinking
3757 * When the old page was mapped as Anon and it's unmap-and-freed
3758 * while migration was ongoing.
3759 * If unmap finds the old page, uncharge() of it will be delayed
3760 * until end_migration(). If unmap finds a new page, it's
3761 * uncharged when it make mapcount to be 1->0. If unmap code
3762 * finds swap_migration_entry, the new page will not be mapped
3763 * and end_migration() will find it(mapcount==0).
3766 * When the old page was mapped but migraion fails, the kernel
3767 * remaps it. A charge for it is kept by MIGRATION flag even
3768 * if mapcount goes down to 0. We can do remap successfully
3769 * without charging it again.
3772 * The "old" page is under lock_page() until the end of
3773 * migration, so, the old page itself will not be swapped-out.
3774 * If the new page is swapped out before end_migraton, our
3775 * hook to usual swap-out path will catch the event.
3778 SetPageCgroupMigration(pc);
3780 unlock_page_cgroup(pc);
3782 * If the page is not charged at this point,
3790 * We charge new page before it's used/mapped. So, even if unlock_page()
3791 * is called before end_migration, we can catch all events on this new
3792 * page. In the case new page is migrated but not remapped, new page's
3793 * mapcount will be finally 0 and we call uncharge in end_migration().
3796 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
3798 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3800 * The page is committed to the memcg, but it's not actually
3801 * charged to the res_counter since we plan on replacing the
3802 * old one and only one page is going to be left afterwards.
3804 __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false);
3807 /* remove redundant charge if migration failed*/
3808 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3809 struct page *oldpage, struct page *newpage, bool migration_ok)
3811 struct page *used, *unused;
3812 struct page_cgroup *pc;
3818 if (!migration_ok) {
3825 anon = PageAnon(used);
3826 __mem_cgroup_uncharge_common(unused,
3827 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
3828 : MEM_CGROUP_CHARGE_TYPE_CACHE,
3830 css_put(&memcg->css);
3832 * We disallowed uncharge of pages under migration because mapcount
3833 * of the page goes down to zero, temporarly.
3834 * Clear the flag and check the page should be charged.
3836 pc = lookup_page_cgroup(oldpage);
3837 lock_page_cgroup(pc);
3838 ClearPageCgroupMigration(pc);
3839 unlock_page_cgroup(pc);
3842 * If a page is a file cache, radix-tree replacement is very atomic
3843 * and we can skip this check. When it was an Anon page, its mapcount
3844 * goes down to 0. But because we added MIGRATION flage, it's not
3845 * uncharged yet. There are several case but page->mapcount check
3846 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3847 * check. (see prepare_charge() also)
3850 mem_cgroup_uncharge_page(used);
3854 * At replace page cache, newpage is not under any memcg but it's on
3855 * LRU. So, this function doesn't touch res_counter but handles LRU
3856 * in correct way. Both pages are locked so we cannot race with uncharge.
3858 void mem_cgroup_replace_page_cache(struct page *oldpage,
3859 struct page *newpage)
3861 struct mem_cgroup *memcg = NULL;
3862 struct page_cgroup *pc;
3863 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3865 if (mem_cgroup_disabled())
3868 pc = lookup_page_cgroup(oldpage);
3869 /* fix accounting on old pages */
3870 lock_page_cgroup(pc);
3871 if (PageCgroupUsed(pc)) {
3872 memcg = pc->mem_cgroup;
3873 mem_cgroup_charge_statistics(memcg, false, -1);
3874 ClearPageCgroupUsed(pc);
3876 unlock_page_cgroup(pc);
3879 * When called from shmem_replace_page(), in some cases the
3880 * oldpage has already been charged, and in some cases not.
3885 * Even if newpage->mapping was NULL before starting replacement,
3886 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3887 * LRU while we overwrite pc->mem_cgroup.
3889 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3892 #ifdef CONFIG_DEBUG_VM
3893 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3895 struct page_cgroup *pc;
3897 pc = lookup_page_cgroup(page);
3899 * Can be NULL while feeding pages into the page allocator for
3900 * the first time, i.e. during boot or memory hotplug;
3901 * or when mem_cgroup_disabled().
3903 if (likely(pc) && PageCgroupUsed(pc))
3908 bool mem_cgroup_bad_page_check(struct page *page)
3910 if (mem_cgroup_disabled())
3913 return lookup_page_cgroup_used(page) != NULL;
3916 void mem_cgroup_print_bad_page(struct page *page)
3918 struct page_cgroup *pc;
3920 pc = lookup_page_cgroup_used(page);
3922 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3923 pc, pc->flags, pc->mem_cgroup);
3928 static DEFINE_MUTEX(set_limit_mutex);
3930 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3931 unsigned long long val)
3934 u64 memswlimit, memlimit;
3936 int children = mem_cgroup_count_children(memcg);
3937 u64 curusage, oldusage;
3941 * For keeping hierarchical_reclaim simple, how long we should retry
3942 * is depends on callers. We set our retry-count to be function
3943 * of # of children which we should visit in this loop.
3945 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3947 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3950 while (retry_count) {
3951 if (signal_pending(current)) {
3956 * Rather than hide all in some function, I do this in
3957 * open coded manner. You see what this really does.
3958 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3960 mutex_lock(&set_limit_mutex);
3961 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3962 if (memswlimit < val) {
3964 mutex_unlock(&set_limit_mutex);
3968 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3972 ret = res_counter_set_limit(&memcg->res, val);
3974 if (memswlimit == val)
3975 memcg->memsw_is_minimum = true;
3977 memcg->memsw_is_minimum = false;
3979 mutex_unlock(&set_limit_mutex);
3984 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3985 MEM_CGROUP_RECLAIM_SHRINK);
3986 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3987 /* Usage is reduced ? */
3988 if (curusage >= oldusage)
3991 oldusage = curusage;
3993 if (!ret && enlarge)
3994 memcg_oom_recover(memcg);
3999 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4000 unsigned long long val)
4003 u64 memlimit, memswlimit, oldusage, curusage;
4004 int children = mem_cgroup_count_children(memcg);
4008 /* see mem_cgroup_resize_res_limit */
4009 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4010 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4011 while (retry_count) {
4012 if (signal_pending(current)) {
4017 * Rather than hide all in some function, I do this in
4018 * open coded manner. You see what this really does.
4019 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4021 mutex_lock(&set_limit_mutex);
4022 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4023 if (memlimit > val) {
4025 mutex_unlock(&set_limit_mutex);
4028 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4029 if (memswlimit < val)
4031 ret = res_counter_set_limit(&memcg->memsw, val);
4033 if (memlimit == val)
4034 memcg->memsw_is_minimum = true;
4036 memcg->memsw_is_minimum = false;
4038 mutex_unlock(&set_limit_mutex);
4043 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4044 MEM_CGROUP_RECLAIM_NOSWAP |
4045 MEM_CGROUP_RECLAIM_SHRINK);
4046 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4047 /* Usage is reduced ? */
4048 if (curusage >= oldusage)
4051 oldusage = curusage;
4053 if (!ret && enlarge)
4054 memcg_oom_recover(memcg);
4058 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4060 unsigned long *total_scanned)
4062 unsigned long nr_reclaimed = 0;
4063 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4064 unsigned long reclaimed;
4066 struct mem_cgroup_tree_per_zone *mctz;
4067 unsigned long long excess;
4068 unsigned long nr_scanned;
4073 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4075 * This loop can run a while, specially if mem_cgroup's continuously
4076 * keep exceeding their soft limit and putting the system under
4083 mz = mem_cgroup_largest_soft_limit_node(mctz);
4088 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4089 gfp_mask, &nr_scanned);
4090 nr_reclaimed += reclaimed;
4091 *total_scanned += nr_scanned;
4092 spin_lock(&mctz->lock);
4095 * If we failed to reclaim anything from this memory cgroup
4096 * it is time to move on to the next cgroup
4102 * Loop until we find yet another one.
4104 * By the time we get the soft_limit lock
4105 * again, someone might have aded the
4106 * group back on the RB tree. Iterate to
4107 * make sure we get a different mem.
4108 * mem_cgroup_largest_soft_limit_node returns
4109 * NULL if no other cgroup is present on
4113 __mem_cgroup_largest_soft_limit_node(mctz);
4115 css_put(&next_mz->memcg->css);
4116 else /* next_mz == NULL or other memcg */
4120 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4121 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4123 * One school of thought says that we should not add
4124 * back the node to the tree if reclaim returns 0.
4125 * But our reclaim could return 0, simply because due
4126 * to priority we are exposing a smaller subset of
4127 * memory to reclaim from. Consider this as a longer
4130 /* If excess == 0, no tree ops */
4131 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4132 spin_unlock(&mctz->lock);
4133 css_put(&mz->memcg->css);
4136 * Could not reclaim anything and there are no more
4137 * mem cgroups to try or we seem to be looping without
4138 * reclaiming anything.
4140 if (!nr_reclaimed &&
4142 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4144 } while (!nr_reclaimed);
4146 css_put(&next_mz->memcg->css);
4147 return nr_reclaimed;
4151 * mem_cgroup_force_empty_list - clears LRU of a group
4152 * @memcg: group to clear
4155 * @lru: lru to to clear
4157 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4158 * reclaim the pages page themselves - pages are moved to the parent (or root)
4161 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4162 int node, int zid, enum lru_list lru)
4164 struct mem_cgroup_per_zone *mz;
4165 unsigned long flags;
4166 struct list_head *list;
4170 zone = &NODE_DATA(node)->node_zones[zid];
4171 mz = mem_cgroup_zoneinfo(memcg, node, zid);
4172 list = &mz->lruvec.lists[lru];
4176 struct page_cgroup *pc;
4179 spin_lock_irqsave(&zone->lru_lock, flags);
4180 if (list_empty(list)) {
4181 spin_unlock_irqrestore(&zone->lru_lock, flags);
4184 page = list_entry(list->prev, struct page, lru);
4186 list_move(&page->lru, list);
4188 spin_unlock_irqrestore(&zone->lru_lock, flags);
4191 spin_unlock_irqrestore(&zone->lru_lock, flags);
4193 pc = lookup_page_cgroup(page);
4195 if (mem_cgroup_move_parent(page, pc, memcg)) {
4196 /* found lock contention or "pc" is obsolete. */
4201 } while (!list_empty(list));
4205 * make mem_cgroup's charge to be 0 if there is no task by moving
4206 * all the charges and pages to the parent.
4207 * This enables deleting this mem_cgroup.
4209 * Caller is responsible for holding css reference on the memcg.
4211 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4217 /* This is for making all *used* pages to be on LRU. */
4218 lru_add_drain_all();
4219 drain_all_stock_sync(memcg);
4220 mem_cgroup_start_move(memcg);
4221 for_each_node_state(node, N_HIGH_MEMORY) {
4222 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4225 mem_cgroup_force_empty_list(memcg,
4230 mem_cgroup_end_move(memcg);
4231 memcg_oom_recover(memcg);
4235 * Kernel memory may not necessarily be trackable to a specific
4236 * process. So they are not migrated, and therefore we can't
4237 * expect their value to drop to 0 here.
4238 * Having res filled up with kmem only is enough.
4240 * This is a safety check because mem_cgroup_force_empty_list
4241 * could have raced with mem_cgroup_replace_page_cache callers
4242 * so the lru seemed empty but the page could have been added
4243 * right after the check. RES_USAGE should be safe as we always
4244 * charge before adding to the LRU.
4246 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4247 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4248 } while (usage > 0);
4252 * Reclaims as many pages from the given memcg as possible and moves
4253 * the rest to the parent.
4255 * Caller is responsible for holding css reference for memcg.
4257 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4259 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4260 struct cgroup *cgrp = memcg->css.cgroup;
4262 /* returns EBUSY if there is a task or if we come here twice. */
4263 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4266 /* we call try-to-free pages for make this cgroup empty */
4267 lru_add_drain_all();
4268 /* try to free all pages in this cgroup */
4269 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4272 if (signal_pending(current))
4275 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4279 /* maybe some writeback is necessary */
4280 congestion_wait(BLK_RW_ASYNC, HZ/10);
4285 mem_cgroup_reparent_charges(memcg);
4290 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4292 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4295 if (mem_cgroup_is_root(memcg))
4297 css_get(&memcg->css);
4298 ret = mem_cgroup_force_empty(memcg);
4299 css_put(&memcg->css);
4305 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4307 return mem_cgroup_from_cont(cont)->use_hierarchy;
4310 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4314 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4315 struct cgroup *parent = cont->parent;
4316 struct mem_cgroup *parent_memcg = NULL;
4319 parent_memcg = mem_cgroup_from_cont(parent);
4323 if (memcg->use_hierarchy == val)
4327 * If parent's use_hierarchy is set, we can't make any modifications
4328 * in the child subtrees. If it is unset, then the change can
4329 * occur, provided the current cgroup has no children.
4331 * For the root cgroup, parent_mem is NULL, we allow value to be
4332 * set if there are no children.
4334 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4335 (val == 1 || val == 0)) {
4336 if (list_empty(&cont->children))
4337 memcg->use_hierarchy = val;
4350 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4351 enum mem_cgroup_stat_index idx)
4353 struct mem_cgroup *iter;
4356 /* Per-cpu values can be negative, use a signed accumulator */
4357 for_each_mem_cgroup_tree(iter, memcg)
4358 val += mem_cgroup_read_stat(iter, idx);
4360 if (val < 0) /* race ? */
4365 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4369 if (!mem_cgroup_is_root(memcg)) {
4371 return res_counter_read_u64(&memcg->res, RES_USAGE);
4373 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4376 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4377 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4380 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4382 return val << PAGE_SHIFT;
4385 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
4386 struct file *file, char __user *buf,
4387 size_t nbytes, loff_t *ppos)
4389 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4395 type = MEMFILE_TYPE(cft->private);
4396 name = MEMFILE_ATTR(cft->private);
4398 if (!do_swap_account && type == _MEMSWAP)
4403 if (name == RES_USAGE)
4404 val = mem_cgroup_usage(memcg, false);
4406 val = res_counter_read_u64(&memcg->res, name);
4409 if (name == RES_USAGE)
4410 val = mem_cgroup_usage(memcg, true);
4412 val = res_counter_read_u64(&memcg->memsw, name);
4415 val = res_counter_read_u64(&memcg->kmem, name);
4421 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4422 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4425 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
4428 #ifdef CONFIG_MEMCG_KMEM
4429 bool must_inc_static_branch = false;
4431 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4433 * For simplicity, we won't allow this to be disabled. It also can't
4434 * be changed if the cgroup has children already, or if tasks had
4437 * If tasks join before we set the limit, a person looking at
4438 * kmem.usage_in_bytes will have no way to determine when it took
4439 * place, which makes the value quite meaningless.
4441 * After it first became limited, changes in the value of the limit are
4442 * of course permitted.
4444 * Taking the cgroup_lock is really offensive, but it is so far the only
4445 * way to guarantee that no children will appear. There are plenty of
4446 * other offenders, and they should all go away. Fine grained locking
4447 * is probably the way to go here. When we are fully hierarchical, we
4448 * can also get rid of the use_hierarchy check.
4451 mutex_lock(&set_limit_mutex);
4452 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
4453 if (cgroup_task_count(cont) || (memcg->use_hierarchy &&
4454 !list_empty(&cont->children))) {
4458 ret = res_counter_set_limit(&memcg->kmem, val);
4461 ret = memcg_update_cache_sizes(memcg);
4463 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
4466 must_inc_static_branch = true;
4468 * kmem charges can outlive the cgroup. In the case of slab
4469 * pages, for instance, a page contain objects from various
4470 * processes, so it is unfeasible to migrate them away. We
4471 * need to reference count the memcg because of that.
4473 mem_cgroup_get(memcg);
4475 ret = res_counter_set_limit(&memcg->kmem, val);
4477 mutex_unlock(&set_limit_mutex);
4481 * We are by now familiar with the fact that we can't inc the static
4482 * branch inside cgroup_lock. See disarm functions for details. A
4483 * worker here is overkill, but also wrong: After the limit is set, we
4484 * must start accounting right away. Since this operation can't fail,
4485 * we can safely defer it to here - no rollback will be needed.
4487 * The boolean used to control this is also safe, because
4488 * KMEM_ACCOUNTED_ACTIVATED guarantees that only one process will be
4489 * able to set it to true;
4491 if (must_inc_static_branch) {
4492 static_key_slow_inc(&memcg_kmem_enabled_key);
4494 * setting the active bit after the inc will guarantee no one
4495 * starts accounting before all call sites are patched
4497 memcg_kmem_set_active(memcg);
4504 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4507 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4511 memcg->kmem_account_flags = parent->kmem_account_flags;
4512 #ifdef CONFIG_MEMCG_KMEM
4514 * When that happen, we need to disable the static branch only on those
4515 * memcgs that enabled it. To achieve this, we would be forced to
4516 * complicate the code by keeping track of which memcgs were the ones
4517 * that actually enabled limits, and which ones got it from its
4520 * It is a lot simpler just to do static_key_slow_inc() on every child
4521 * that is accounted.
4523 if (!memcg_kmem_is_active(memcg))
4527 * destroy(), called if we fail, will issue static_key_slow_inc() and
4528 * mem_cgroup_put() if kmem is enabled. We have to either call them
4529 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
4530 * this more consistent, since it always leads to the same destroy path
4532 mem_cgroup_get(memcg);
4533 static_key_slow_inc(&memcg_kmem_enabled_key);
4535 mutex_lock(&set_limit_mutex);
4536 ret = memcg_update_cache_sizes(memcg);
4537 mutex_unlock(&set_limit_mutex);
4544 * The user of this function is...
4547 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4550 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4553 unsigned long long val;
4556 type = MEMFILE_TYPE(cft->private);
4557 name = MEMFILE_ATTR(cft->private);
4559 if (!do_swap_account && type == _MEMSWAP)
4564 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4568 /* This function does all necessary parse...reuse it */
4569 ret = res_counter_memparse_write_strategy(buffer, &val);
4573 ret = mem_cgroup_resize_limit(memcg, val);
4574 else if (type == _MEMSWAP)
4575 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4576 else if (type == _KMEM)
4577 ret = memcg_update_kmem_limit(cont, val);
4581 case RES_SOFT_LIMIT:
4582 ret = res_counter_memparse_write_strategy(buffer, &val);
4586 * For memsw, soft limits are hard to implement in terms
4587 * of semantics, for now, we support soft limits for
4588 * control without swap
4591 ret = res_counter_set_soft_limit(&memcg->res, val);
4596 ret = -EINVAL; /* should be BUG() ? */
4602 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4603 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4605 struct cgroup *cgroup;
4606 unsigned long long min_limit, min_memsw_limit, tmp;
4608 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4609 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4610 cgroup = memcg->css.cgroup;
4611 if (!memcg->use_hierarchy)
4614 while (cgroup->parent) {
4615 cgroup = cgroup->parent;
4616 memcg = mem_cgroup_from_cont(cgroup);
4617 if (!memcg->use_hierarchy)
4619 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4620 min_limit = min(min_limit, tmp);
4621 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4622 min_memsw_limit = min(min_memsw_limit, tmp);
4625 *mem_limit = min_limit;
4626 *memsw_limit = min_memsw_limit;
4629 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4631 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4635 type = MEMFILE_TYPE(event);
4636 name = MEMFILE_ATTR(event);
4638 if (!do_swap_account && type == _MEMSWAP)
4644 res_counter_reset_max(&memcg->res);
4645 else if (type == _MEMSWAP)
4646 res_counter_reset_max(&memcg->memsw);
4647 else if (type == _KMEM)
4648 res_counter_reset_max(&memcg->kmem);
4654 res_counter_reset_failcnt(&memcg->res);
4655 else if (type == _MEMSWAP)
4656 res_counter_reset_failcnt(&memcg->memsw);
4657 else if (type == _KMEM)
4658 res_counter_reset_failcnt(&memcg->kmem);
4667 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4670 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4674 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4675 struct cftype *cft, u64 val)
4677 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4679 if (val >= (1 << NR_MOVE_TYPE))
4682 * We check this value several times in both in can_attach() and
4683 * attach(), so we need cgroup lock to prevent this value from being
4687 memcg->move_charge_at_immigrate = val;
4693 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4694 struct cftype *cft, u64 val)
4701 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
4705 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4706 unsigned long node_nr;
4707 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4709 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4710 seq_printf(m, "total=%lu", total_nr);
4711 for_each_node_state(nid, N_HIGH_MEMORY) {
4712 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4713 seq_printf(m, " N%d=%lu", nid, node_nr);
4717 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4718 seq_printf(m, "file=%lu", file_nr);
4719 for_each_node_state(nid, N_HIGH_MEMORY) {
4720 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4722 seq_printf(m, " N%d=%lu", nid, node_nr);
4726 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4727 seq_printf(m, "anon=%lu", anon_nr);
4728 for_each_node_state(nid, N_HIGH_MEMORY) {
4729 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4731 seq_printf(m, " N%d=%lu", nid, node_nr);
4735 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4736 seq_printf(m, "unevictable=%lu", unevictable_nr);
4737 for_each_node_state(nid, N_HIGH_MEMORY) {
4738 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4739 BIT(LRU_UNEVICTABLE));
4740 seq_printf(m, " N%d=%lu", nid, node_nr);
4745 #endif /* CONFIG_NUMA */
4747 static const char * const mem_cgroup_lru_names[] = {
4755 static inline void mem_cgroup_lru_names_not_uptodate(void)
4757 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4760 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
4763 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4764 struct mem_cgroup *mi;
4767 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4768 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4770 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4771 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4774 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4775 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4776 mem_cgroup_read_events(memcg, i));
4778 for (i = 0; i < NR_LRU_LISTS; i++)
4779 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4780 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4782 /* Hierarchical information */
4784 unsigned long long limit, memsw_limit;
4785 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4786 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
4787 if (do_swap_account)
4788 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4792 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4795 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4797 for_each_mem_cgroup_tree(mi, memcg)
4798 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4799 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4802 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4803 unsigned long long val = 0;
4805 for_each_mem_cgroup_tree(mi, memcg)
4806 val += mem_cgroup_read_events(mi, i);
4807 seq_printf(m, "total_%s %llu\n",
4808 mem_cgroup_events_names[i], val);
4811 for (i = 0; i < NR_LRU_LISTS; i++) {
4812 unsigned long long val = 0;
4814 for_each_mem_cgroup_tree(mi, memcg)
4815 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4816 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4819 #ifdef CONFIG_DEBUG_VM
4822 struct mem_cgroup_per_zone *mz;
4823 struct zone_reclaim_stat *rstat;
4824 unsigned long recent_rotated[2] = {0, 0};
4825 unsigned long recent_scanned[2] = {0, 0};
4827 for_each_online_node(nid)
4828 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4829 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4830 rstat = &mz->lruvec.reclaim_stat;
4832 recent_rotated[0] += rstat->recent_rotated[0];
4833 recent_rotated[1] += rstat->recent_rotated[1];
4834 recent_scanned[0] += rstat->recent_scanned[0];
4835 recent_scanned[1] += rstat->recent_scanned[1];
4837 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4838 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4839 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4840 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4847 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4849 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4851 return mem_cgroup_swappiness(memcg);
4854 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4857 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4858 struct mem_cgroup *parent;
4863 if (cgrp->parent == NULL)
4866 parent = mem_cgroup_from_cont(cgrp->parent);
4870 /* If under hierarchy, only empty-root can set this value */
4871 if ((parent->use_hierarchy) ||
4872 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4877 memcg->swappiness = val;
4884 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4886 struct mem_cgroup_threshold_ary *t;
4892 t = rcu_dereference(memcg->thresholds.primary);
4894 t = rcu_dereference(memcg->memsw_thresholds.primary);
4899 usage = mem_cgroup_usage(memcg, swap);
4902 * current_threshold points to threshold just below or equal to usage.
4903 * If it's not true, a threshold was crossed after last
4904 * call of __mem_cgroup_threshold().
4906 i = t->current_threshold;
4909 * Iterate backward over array of thresholds starting from
4910 * current_threshold and check if a threshold is crossed.
4911 * If none of thresholds below usage is crossed, we read
4912 * only one element of the array here.
4914 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4915 eventfd_signal(t->entries[i].eventfd, 1);
4917 /* i = current_threshold + 1 */
4921 * Iterate forward over array of thresholds starting from
4922 * current_threshold+1 and check if a threshold is crossed.
4923 * If none of thresholds above usage is crossed, we read
4924 * only one element of the array here.
4926 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4927 eventfd_signal(t->entries[i].eventfd, 1);
4929 /* Update current_threshold */
4930 t->current_threshold = i - 1;
4935 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4938 __mem_cgroup_threshold(memcg, false);
4939 if (do_swap_account)
4940 __mem_cgroup_threshold(memcg, true);
4942 memcg = parent_mem_cgroup(memcg);
4946 static int compare_thresholds(const void *a, const void *b)
4948 const struct mem_cgroup_threshold *_a = a;
4949 const struct mem_cgroup_threshold *_b = b;
4951 return _a->threshold - _b->threshold;
4954 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4956 struct mem_cgroup_eventfd_list *ev;
4958 list_for_each_entry(ev, &memcg->oom_notify, list)
4959 eventfd_signal(ev->eventfd, 1);
4963 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4965 struct mem_cgroup *iter;
4967 for_each_mem_cgroup_tree(iter, memcg)
4968 mem_cgroup_oom_notify_cb(iter);
4971 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4972 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4974 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4975 struct mem_cgroup_thresholds *thresholds;
4976 struct mem_cgroup_threshold_ary *new;
4977 enum res_type type = MEMFILE_TYPE(cft->private);
4978 u64 threshold, usage;
4981 ret = res_counter_memparse_write_strategy(args, &threshold);
4985 mutex_lock(&memcg->thresholds_lock);
4988 thresholds = &memcg->thresholds;
4989 else if (type == _MEMSWAP)
4990 thresholds = &memcg->memsw_thresholds;
4994 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4996 /* Check if a threshold crossed before adding a new one */
4997 if (thresholds->primary)
4998 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5000 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5002 /* Allocate memory for new array of thresholds */
5003 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5011 /* Copy thresholds (if any) to new array */
5012 if (thresholds->primary) {
5013 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5014 sizeof(struct mem_cgroup_threshold));
5017 /* Add new threshold */
5018 new->entries[size - 1].eventfd = eventfd;
5019 new->entries[size - 1].threshold = threshold;
5021 /* Sort thresholds. Registering of new threshold isn't time-critical */
5022 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5023 compare_thresholds, NULL);
5025 /* Find current threshold */
5026 new->current_threshold = -1;
5027 for (i = 0; i < size; i++) {
5028 if (new->entries[i].threshold <= usage) {
5030 * new->current_threshold will not be used until
5031 * rcu_assign_pointer(), so it's safe to increment
5034 ++new->current_threshold;
5039 /* Free old spare buffer and save old primary buffer as spare */
5040 kfree(thresholds->spare);
5041 thresholds->spare = thresholds->primary;
5043 rcu_assign_pointer(thresholds->primary, new);
5045 /* To be sure that nobody uses thresholds */
5049 mutex_unlock(&memcg->thresholds_lock);
5054 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5055 struct cftype *cft, struct eventfd_ctx *eventfd)
5057 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5058 struct mem_cgroup_thresholds *thresholds;
5059 struct mem_cgroup_threshold_ary *new;
5060 enum res_type type = MEMFILE_TYPE(cft->private);
5064 mutex_lock(&memcg->thresholds_lock);
5066 thresholds = &memcg->thresholds;
5067 else if (type == _MEMSWAP)
5068 thresholds = &memcg->memsw_thresholds;
5072 if (!thresholds->primary)
5075 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5077 /* Check if a threshold crossed before removing */
5078 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5080 /* Calculate new number of threshold */
5082 for (i = 0; i < thresholds->primary->size; i++) {
5083 if (thresholds->primary->entries[i].eventfd != eventfd)
5087 new = thresholds->spare;
5089 /* Set thresholds array to NULL if we don't have thresholds */
5098 /* Copy thresholds and find current threshold */
5099 new->current_threshold = -1;
5100 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5101 if (thresholds->primary->entries[i].eventfd == eventfd)
5104 new->entries[j] = thresholds->primary->entries[i];
5105 if (new->entries[j].threshold <= usage) {
5107 * new->current_threshold will not be used
5108 * until rcu_assign_pointer(), so it's safe to increment
5111 ++new->current_threshold;
5117 /* Swap primary and spare array */
5118 thresholds->spare = thresholds->primary;
5119 /* If all events are unregistered, free the spare array */
5121 kfree(thresholds->spare);
5122 thresholds->spare = NULL;
5125 rcu_assign_pointer(thresholds->primary, new);
5127 /* To be sure that nobody uses thresholds */
5130 mutex_unlock(&memcg->thresholds_lock);
5133 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5134 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5136 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5137 struct mem_cgroup_eventfd_list *event;
5138 enum res_type type = MEMFILE_TYPE(cft->private);
5140 BUG_ON(type != _OOM_TYPE);
5141 event = kmalloc(sizeof(*event), GFP_KERNEL);
5145 spin_lock(&memcg_oom_lock);
5147 event->eventfd = eventfd;
5148 list_add(&event->list, &memcg->oom_notify);
5150 /* already in OOM ? */
5151 if (atomic_read(&memcg->under_oom))
5152 eventfd_signal(eventfd, 1);
5153 spin_unlock(&memcg_oom_lock);
5158 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5159 struct cftype *cft, struct eventfd_ctx *eventfd)
5161 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5162 struct mem_cgroup_eventfd_list *ev, *tmp;
5163 enum res_type type = MEMFILE_TYPE(cft->private);
5165 BUG_ON(type != _OOM_TYPE);
5167 spin_lock(&memcg_oom_lock);
5169 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5170 if (ev->eventfd == eventfd) {
5171 list_del(&ev->list);
5176 spin_unlock(&memcg_oom_lock);
5179 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5180 struct cftype *cft, struct cgroup_map_cb *cb)
5182 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5184 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5186 if (atomic_read(&memcg->under_oom))
5187 cb->fill(cb, "under_oom", 1);
5189 cb->fill(cb, "under_oom", 0);
5193 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5194 struct cftype *cft, u64 val)
5196 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5197 struct mem_cgroup *parent;
5199 /* cannot set to root cgroup and only 0 and 1 are allowed */
5200 if (!cgrp->parent || !((val == 0) || (val == 1)))
5203 parent = mem_cgroup_from_cont(cgrp->parent);
5206 /* oom-kill-disable is a flag for subhierarchy. */
5207 if ((parent->use_hierarchy) ||
5208 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
5212 memcg->oom_kill_disable = val;
5214 memcg_oom_recover(memcg);
5219 #ifdef CONFIG_MEMCG_KMEM
5220 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5224 memcg->kmemcg_id = -1;
5225 ret = memcg_propagate_kmem(memcg);
5229 if (mem_cgroup_is_root(memcg))
5230 ida_init(&kmem_limited_groups);
5232 return mem_cgroup_sockets_init(memcg, ss);
5235 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5237 mem_cgroup_sockets_destroy(memcg);
5239 memcg_kmem_mark_dead(memcg);
5241 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5245 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5246 * path here, being careful not to race with memcg_uncharge_kmem: it is
5247 * possible that the charges went down to 0 between mark_dead and the
5248 * res_counter read, so in that case, we don't need the put
5250 if (memcg_kmem_test_and_clear_dead(memcg))
5251 mem_cgroup_put(memcg);
5254 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5259 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5264 static struct cftype mem_cgroup_files[] = {
5266 .name = "usage_in_bytes",
5267 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5268 .read = mem_cgroup_read,
5269 .register_event = mem_cgroup_usage_register_event,
5270 .unregister_event = mem_cgroup_usage_unregister_event,
5273 .name = "max_usage_in_bytes",
5274 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5275 .trigger = mem_cgroup_reset,
5276 .read = mem_cgroup_read,
5279 .name = "limit_in_bytes",
5280 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5281 .write_string = mem_cgroup_write,
5282 .read = mem_cgroup_read,
5285 .name = "soft_limit_in_bytes",
5286 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5287 .write_string = mem_cgroup_write,
5288 .read = mem_cgroup_read,
5292 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5293 .trigger = mem_cgroup_reset,
5294 .read = mem_cgroup_read,
5298 .read_seq_string = memcg_stat_show,
5301 .name = "force_empty",
5302 .trigger = mem_cgroup_force_empty_write,
5305 .name = "use_hierarchy",
5306 .write_u64 = mem_cgroup_hierarchy_write,
5307 .read_u64 = mem_cgroup_hierarchy_read,
5310 .name = "swappiness",
5311 .read_u64 = mem_cgroup_swappiness_read,
5312 .write_u64 = mem_cgroup_swappiness_write,
5315 .name = "move_charge_at_immigrate",
5316 .read_u64 = mem_cgroup_move_charge_read,
5317 .write_u64 = mem_cgroup_move_charge_write,
5320 .name = "oom_control",
5321 .read_map = mem_cgroup_oom_control_read,
5322 .write_u64 = mem_cgroup_oom_control_write,
5323 .register_event = mem_cgroup_oom_register_event,
5324 .unregister_event = mem_cgroup_oom_unregister_event,
5325 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5329 .name = "numa_stat",
5330 .read_seq_string = memcg_numa_stat_show,
5333 #ifdef CONFIG_MEMCG_SWAP
5335 .name = "memsw.usage_in_bytes",
5336 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5337 .read = mem_cgroup_read,
5338 .register_event = mem_cgroup_usage_register_event,
5339 .unregister_event = mem_cgroup_usage_unregister_event,
5342 .name = "memsw.max_usage_in_bytes",
5343 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5344 .trigger = mem_cgroup_reset,
5345 .read = mem_cgroup_read,
5348 .name = "memsw.limit_in_bytes",
5349 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5350 .write_string = mem_cgroup_write,
5351 .read = mem_cgroup_read,
5354 .name = "memsw.failcnt",
5355 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5356 .trigger = mem_cgroup_reset,
5357 .read = mem_cgroup_read,
5360 #ifdef CONFIG_MEMCG_KMEM
5362 .name = "kmem.limit_in_bytes",
5363 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5364 .write_string = mem_cgroup_write,
5365 .read = mem_cgroup_read,
5368 .name = "kmem.usage_in_bytes",
5369 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5370 .read = mem_cgroup_read,
5373 .name = "kmem.failcnt",
5374 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5375 .trigger = mem_cgroup_reset,
5376 .read = mem_cgroup_read,
5379 .name = "kmem.max_usage_in_bytes",
5380 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5381 .trigger = mem_cgroup_reset,
5382 .read = mem_cgroup_read,
5385 { }, /* terminate */
5388 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5390 struct mem_cgroup_per_node *pn;
5391 struct mem_cgroup_per_zone *mz;
5392 int zone, tmp = node;
5394 * This routine is called against possible nodes.
5395 * But it's BUG to call kmalloc() against offline node.
5397 * TODO: this routine can waste much memory for nodes which will
5398 * never be onlined. It's better to use memory hotplug callback
5401 if (!node_state(node, N_NORMAL_MEMORY))
5403 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5407 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5408 mz = &pn->zoneinfo[zone];
5409 lruvec_init(&mz->lruvec, &NODE_DATA(node)->node_zones[zone]);
5410 mz->usage_in_excess = 0;
5411 mz->on_tree = false;
5414 memcg->info.nodeinfo[node] = pn;
5418 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5420 kfree(memcg->info.nodeinfo[node]);
5423 static struct mem_cgroup *mem_cgroup_alloc(void)
5425 struct mem_cgroup *memcg;
5426 int size = sizeof(struct mem_cgroup);
5428 /* Can be very big if MAX_NUMNODES is very big */
5429 if (size < PAGE_SIZE)
5430 memcg = kzalloc(size, GFP_KERNEL);
5432 memcg = vzalloc(size);
5437 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5440 spin_lock_init(&memcg->pcp_counter_lock);
5444 if (size < PAGE_SIZE)
5452 * At destroying mem_cgroup, references from swap_cgroup can remain.
5453 * (scanning all at force_empty is too costly...)
5455 * Instead of clearing all references at force_empty, we remember
5456 * the number of reference from swap_cgroup and free mem_cgroup when
5457 * it goes down to 0.
5459 * Removal of cgroup itself succeeds regardless of refs from swap.
5462 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5465 int size = sizeof(struct mem_cgroup);
5467 mem_cgroup_remove_from_trees(memcg);
5468 free_css_id(&mem_cgroup_subsys, &memcg->css);
5471 free_mem_cgroup_per_zone_info(memcg, node);
5473 free_percpu(memcg->stat);
5476 * We need to make sure that (at least for now), the jump label
5477 * destruction code runs outside of the cgroup lock. This is because
5478 * get_online_cpus(), which is called from the static_branch update,
5479 * can't be called inside the cgroup_lock. cpusets are the ones
5480 * enforcing this dependency, so if they ever change, we might as well.
5482 * schedule_work() will guarantee this happens. Be careful if you need
5483 * to move this code around, and make sure it is outside
5486 disarm_static_keys(memcg);
5487 if (size < PAGE_SIZE)
5495 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
5496 * but in process context. The work_freeing structure is overlaid
5497 * on the rcu_freeing structure, which itself is overlaid on memsw.
5499 static void free_work(struct work_struct *work)
5501 struct mem_cgroup *memcg;
5503 memcg = container_of(work, struct mem_cgroup, work_freeing);
5504 __mem_cgroup_free(memcg);
5507 static void free_rcu(struct rcu_head *rcu_head)
5509 struct mem_cgroup *memcg;
5511 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
5512 INIT_WORK(&memcg->work_freeing, free_work);
5513 schedule_work(&memcg->work_freeing);
5516 static void mem_cgroup_get(struct mem_cgroup *memcg)
5518 atomic_inc(&memcg->refcnt);
5521 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
5523 if (atomic_sub_and_test(count, &memcg->refcnt)) {
5524 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5525 call_rcu(&memcg->rcu_freeing, free_rcu);
5527 mem_cgroup_put(parent);
5531 static void mem_cgroup_put(struct mem_cgroup *memcg)
5533 __mem_cgroup_put(memcg, 1);
5537 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5539 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5541 if (!memcg->res.parent)
5543 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5545 EXPORT_SYMBOL(parent_mem_cgroup);
5547 #ifdef CONFIG_MEMCG_SWAP
5548 static void __init enable_swap_cgroup(void)
5550 if (!mem_cgroup_disabled() && really_do_swap_account)
5551 do_swap_account = 1;
5554 static void __init enable_swap_cgroup(void)
5559 static int mem_cgroup_soft_limit_tree_init(void)
5561 struct mem_cgroup_tree_per_node *rtpn;
5562 struct mem_cgroup_tree_per_zone *rtpz;
5563 int tmp, node, zone;
5565 for_each_node(node) {
5567 if (!node_state(node, N_NORMAL_MEMORY))
5569 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5573 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5575 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5576 rtpz = &rtpn->rb_tree_per_zone[zone];
5577 rtpz->rb_root = RB_ROOT;
5578 spin_lock_init(&rtpz->lock);
5584 for_each_node(node) {
5585 if (!soft_limit_tree.rb_tree_per_node[node])
5587 kfree(soft_limit_tree.rb_tree_per_node[node]);
5588 soft_limit_tree.rb_tree_per_node[node] = NULL;
5594 static struct cgroup_subsys_state * __ref
5595 mem_cgroup_create(struct cgroup *cont)
5597 struct mem_cgroup *memcg, *parent;
5598 long error = -ENOMEM;
5601 memcg = mem_cgroup_alloc();
5603 return ERR_PTR(error);
5606 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5610 if (cont->parent == NULL) {
5612 enable_swap_cgroup();
5614 if (mem_cgroup_soft_limit_tree_init())
5616 root_mem_cgroup = memcg;
5617 for_each_possible_cpu(cpu) {
5618 struct memcg_stock_pcp *stock =
5619 &per_cpu(memcg_stock, cpu);
5620 INIT_WORK(&stock->work, drain_local_stock);
5622 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5624 parent = mem_cgroup_from_cont(cont->parent);
5625 memcg->use_hierarchy = parent->use_hierarchy;
5626 memcg->oom_kill_disable = parent->oom_kill_disable;
5629 if (parent && parent->use_hierarchy) {
5630 res_counter_init(&memcg->res, &parent->res);
5631 res_counter_init(&memcg->memsw, &parent->memsw);
5632 res_counter_init(&memcg->kmem, &parent->kmem);
5635 * We increment refcnt of the parent to ensure that we can
5636 * safely access it on res_counter_charge/uncharge.
5637 * This refcnt will be decremented when freeing this
5638 * mem_cgroup(see mem_cgroup_put).
5640 mem_cgroup_get(parent);
5642 res_counter_init(&memcg->res, NULL);
5643 res_counter_init(&memcg->memsw, NULL);
5644 res_counter_init(&memcg->kmem, NULL);
5646 * Deeper hierachy with use_hierarchy == false doesn't make
5647 * much sense so let cgroup subsystem know about this
5648 * unfortunate state in our controller.
5650 if (parent && parent != root_mem_cgroup)
5651 mem_cgroup_subsys.broken_hierarchy = true;
5653 memcg->last_scanned_node = MAX_NUMNODES;
5654 INIT_LIST_HEAD(&memcg->oom_notify);
5657 memcg->swappiness = mem_cgroup_swappiness(parent);
5658 atomic_set(&memcg->refcnt, 1);
5659 memcg->move_charge_at_immigrate = 0;
5660 mutex_init(&memcg->thresholds_lock);
5661 spin_lock_init(&memcg->move_lock);
5663 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
5666 * We call put now because our (and parent's) refcnts
5667 * are already in place. mem_cgroup_put() will internally
5668 * call __mem_cgroup_free, so return directly
5670 mem_cgroup_put(memcg);
5671 return ERR_PTR(error);
5675 __mem_cgroup_free(memcg);
5676 return ERR_PTR(error);
5679 static void mem_cgroup_pre_destroy(struct cgroup *cont)
5681 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5683 mem_cgroup_reparent_charges(memcg);
5686 static void mem_cgroup_destroy(struct cgroup *cont)
5688 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5690 kmem_cgroup_destroy(memcg);
5692 mem_cgroup_put(memcg);
5696 /* Handlers for move charge at task migration. */
5697 #define PRECHARGE_COUNT_AT_ONCE 256
5698 static int mem_cgroup_do_precharge(unsigned long count)
5701 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5702 struct mem_cgroup *memcg = mc.to;
5704 if (mem_cgroup_is_root(memcg)) {
5705 mc.precharge += count;
5706 /* we don't need css_get for root */
5709 /* try to charge at once */
5711 struct res_counter *dummy;
5713 * "memcg" cannot be under rmdir() because we've already checked
5714 * by cgroup_lock_live_cgroup() that it is not removed and we
5715 * are still under the same cgroup_mutex. So we can postpone
5718 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5720 if (do_swap_account && res_counter_charge(&memcg->memsw,
5721 PAGE_SIZE * count, &dummy)) {
5722 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5725 mc.precharge += count;
5729 /* fall back to one by one charge */
5731 if (signal_pending(current)) {
5735 if (!batch_count--) {
5736 batch_count = PRECHARGE_COUNT_AT_ONCE;
5739 ret = __mem_cgroup_try_charge(NULL,
5740 GFP_KERNEL, 1, &memcg, false);
5742 /* mem_cgroup_clear_mc() will do uncharge later */
5750 * get_mctgt_type - get target type of moving charge
5751 * @vma: the vma the pte to be checked belongs
5752 * @addr: the address corresponding to the pte to be checked
5753 * @ptent: the pte to be checked
5754 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5757 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5758 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5759 * move charge. if @target is not NULL, the page is stored in target->page
5760 * with extra refcnt got(Callers should handle it).
5761 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5762 * target for charge migration. if @target is not NULL, the entry is stored
5765 * Called with pte lock held.
5772 enum mc_target_type {
5778 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5779 unsigned long addr, pte_t ptent)
5781 struct page *page = vm_normal_page(vma, addr, ptent);
5783 if (!page || !page_mapped(page))
5785 if (PageAnon(page)) {
5786 /* we don't move shared anon */
5789 } else if (!move_file())
5790 /* we ignore mapcount for file pages */
5792 if (!get_page_unless_zero(page))
5799 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5800 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5802 struct page *page = NULL;
5803 swp_entry_t ent = pte_to_swp_entry(ptent);
5805 if (!move_anon() || non_swap_entry(ent))
5808 * Because lookup_swap_cache() updates some statistics counter,
5809 * we call find_get_page() with swapper_space directly.
5811 page = find_get_page(&swapper_space, ent.val);
5812 if (do_swap_account)
5813 entry->val = ent.val;
5818 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5819 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5825 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5826 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5828 struct page *page = NULL;
5829 struct address_space *mapping;
5832 if (!vma->vm_file) /* anonymous vma */
5837 mapping = vma->vm_file->f_mapping;
5838 if (pte_none(ptent))
5839 pgoff = linear_page_index(vma, addr);
5840 else /* pte_file(ptent) is true */
5841 pgoff = pte_to_pgoff(ptent);
5843 /* page is moved even if it's not RSS of this task(page-faulted). */
5844 page = find_get_page(mapping, pgoff);
5847 /* shmem/tmpfs may report page out on swap: account for that too. */
5848 if (radix_tree_exceptional_entry(page)) {
5849 swp_entry_t swap = radix_to_swp_entry(page);
5850 if (do_swap_account)
5852 page = find_get_page(&swapper_space, swap.val);
5858 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5859 unsigned long addr, pte_t ptent, union mc_target *target)
5861 struct page *page = NULL;
5862 struct page_cgroup *pc;
5863 enum mc_target_type ret = MC_TARGET_NONE;
5864 swp_entry_t ent = { .val = 0 };
5866 if (pte_present(ptent))
5867 page = mc_handle_present_pte(vma, addr, ptent);
5868 else if (is_swap_pte(ptent))
5869 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5870 else if (pte_none(ptent) || pte_file(ptent))
5871 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5873 if (!page && !ent.val)
5876 pc = lookup_page_cgroup(page);
5878 * Do only loose check w/o page_cgroup lock.
5879 * mem_cgroup_move_account() checks the pc is valid or not under
5882 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5883 ret = MC_TARGET_PAGE;
5885 target->page = page;
5887 if (!ret || !target)
5890 /* There is a swap entry and a page doesn't exist or isn't charged */
5891 if (ent.val && !ret &&
5892 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5893 ret = MC_TARGET_SWAP;
5900 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5902 * We don't consider swapping or file mapped pages because THP does not
5903 * support them for now.
5904 * Caller should make sure that pmd_trans_huge(pmd) is true.
5906 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5907 unsigned long addr, pmd_t pmd, union mc_target *target)
5909 struct page *page = NULL;
5910 struct page_cgroup *pc;
5911 enum mc_target_type ret = MC_TARGET_NONE;
5913 page = pmd_page(pmd);
5914 VM_BUG_ON(!page || !PageHead(page));
5917 pc = lookup_page_cgroup(page);
5918 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5919 ret = MC_TARGET_PAGE;
5922 target->page = page;
5928 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5929 unsigned long addr, pmd_t pmd, union mc_target *target)
5931 return MC_TARGET_NONE;
5935 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5936 unsigned long addr, unsigned long end,
5937 struct mm_walk *walk)
5939 struct vm_area_struct *vma = walk->private;
5943 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5944 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5945 mc.precharge += HPAGE_PMD_NR;
5946 spin_unlock(&vma->vm_mm->page_table_lock);
5950 if (pmd_trans_unstable(pmd))
5952 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5953 for (; addr != end; pte++, addr += PAGE_SIZE)
5954 if (get_mctgt_type(vma, addr, *pte, NULL))
5955 mc.precharge++; /* increment precharge temporarily */
5956 pte_unmap_unlock(pte - 1, ptl);
5962 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5964 unsigned long precharge;
5965 struct vm_area_struct *vma;
5967 down_read(&mm->mmap_sem);
5968 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5969 struct mm_walk mem_cgroup_count_precharge_walk = {
5970 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5974 if (is_vm_hugetlb_page(vma))
5976 walk_page_range(vma->vm_start, vma->vm_end,
5977 &mem_cgroup_count_precharge_walk);
5979 up_read(&mm->mmap_sem);
5981 precharge = mc.precharge;
5987 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5989 unsigned long precharge = mem_cgroup_count_precharge(mm);
5991 VM_BUG_ON(mc.moving_task);
5992 mc.moving_task = current;
5993 return mem_cgroup_do_precharge(precharge);
5996 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5997 static void __mem_cgroup_clear_mc(void)
5999 struct mem_cgroup *from = mc.from;
6000 struct mem_cgroup *to = mc.to;
6002 /* we must uncharge all the leftover precharges from mc.to */
6004 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6008 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6009 * we must uncharge here.
6011 if (mc.moved_charge) {
6012 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6013 mc.moved_charge = 0;
6015 /* we must fixup refcnts and charges */
6016 if (mc.moved_swap) {
6017 /* uncharge swap account from the old cgroup */
6018 if (!mem_cgroup_is_root(mc.from))
6019 res_counter_uncharge(&mc.from->memsw,
6020 PAGE_SIZE * mc.moved_swap);
6021 __mem_cgroup_put(mc.from, mc.moved_swap);
6023 if (!mem_cgroup_is_root(mc.to)) {
6025 * we charged both to->res and to->memsw, so we should
6028 res_counter_uncharge(&mc.to->res,
6029 PAGE_SIZE * mc.moved_swap);
6031 /* we've already done mem_cgroup_get(mc.to) */
6034 memcg_oom_recover(from);
6035 memcg_oom_recover(to);
6036 wake_up_all(&mc.waitq);
6039 static void mem_cgroup_clear_mc(void)
6041 struct mem_cgroup *from = mc.from;
6044 * we must clear moving_task before waking up waiters at the end of
6047 mc.moving_task = NULL;
6048 __mem_cgroup_clear_mc();
6049 spin_lock(&mc.lock);
6052 spin_unlock(&mc.lock);
6053 mem_cgroup_end_move(from);
6056 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6057 struct cgroup_taskset *tset)
6059 struct task_struct *p = cgroup_taskset_first(tset);
6061 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6063 if (memcg->move_charge_at_immigrate) {
6064 struct mm_struct *mm;
6065 struct mem_cgroup *from = mem_cgroup_from_task(p);
6067 VM_BUG_ON(from == memcg);
6069 mm = get_task_mm(p);
6072 /* We move charges only when we move a owner of the mm */
6073 if (mm->owner == p) {
6076 VM_BUG_ON(mc.precharge);
6077 VM_BUG_ON(mc.moved_charge);
6078 VM_BUG_ON(mc.moved_swap);
6079 mem_cgroup_start_move(from);
6080 spin_lock(&mc.lock);
6083 spin_unlock(&mc.lock);
6084 /* We set mc.moving_task later */
6086 ret = mem_cgroup_precharge_mc(mm);
6088 mem_cgroup_clear_mc();
6095 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6096 struct cgroup_taskset *tset)
6098 mem_cgroup_clear_mc();
6101 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6102 unsigned long addr, unsigned long end,
6103 struct mm_walk *walk)
6106 struct vm_area_struct *vma = walk->private;
6109 enum mc_target_type target_type;
6110 union mc_target target;
6112 struct page_cgroup *pc;
6115 * We don't take compound_lock() here but no race with splitting thp
6117 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6118 * under splitting, which means there's no concurrent thp split,
6119 * - if another thread runs into split_huge_page() just after we
6120 * entered this if-block, the thread must wait for page table lock
6121 * to be unlocked in __split_huge_page_splitting(), where the main
6122 * part of thp split is not executed yet.
6124 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6125 if (mc.precharge < HPAGE_PMD_NR) {
6126 spin_unlock(&vma->vm_mm->page_table_lock);
6129 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6130 if (target_type == MC_TARGET_PAGE) {
6132 if (!isolate_lru_page(page)) {
6133 pc = lookup_page_cgroup(page);
6134 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6135 pc, mc.from, mc.to)) {
6136 mc.precharge -= HPAGE_PMD_NR;
6137 mc.moved_charge += HPAGE_PMD_NR;
6139 putback_lru_page(page);
6143 spin_unlock(&vma->vm_mm->page_table_lock);
6147 if (pmd_trans_unstable(pmd))
6150 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6151 for (; addr != end; addr += PAGE_SIZE) {
6152 pte_t ptent = *(pte++);
6158 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6159 case MC_TARGET_PAGE:
6161 if (isolate_lru_page(page))
6163 pc = lookup_page_cgroup(page);
6164 if (!mem_cgroup_move_account(page, 1, pc,
6167 /* we uncharge from mc.from later. */
6170 putback_lru_page(page);
6171 put: /* get_mctgt_type() gets the page */
6174 case MC_TARGET_SWAP:
6176 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6178 /* we fixup refcnts and charges later. */
6186 pte_unmap_unlock(pte - 1, ptl);
6191 * We have consumed all precharges we got in can_attach().
6192 * We try charge one by one, but don't do any additional
6193 * charges to mc.to if we have failed in charge once in attach()
6196 ret = mem_cgroup_do_precharge(1);
6204 static void mem_cgroup_move_charge(struct mm_struct *mm)
6206 struct vm_area_struct *vma;
6208 lru_add_drain_all();
6210 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6212 * Someone who are holding the mmap_sem might be waiting in
6213 * waitq. So we cancel all extra charges, wake up all waiters,
6214 * and retry. Because we cancel precharges, we might not be able
6215 * to move enough charges, but moving charge is a best-effort
6216 * feature anyway, so it wouldn't be a big problem.
6218 __mem_cgroup_clear_mc();
6222 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6224 struct mm_walk mem_cgroup_move_charge_walk = {
6225 .pmd_entry = mem_cgroup_move_charge_pte_range,
6229 if (is_vm_hugetlb_page(vma))
6231 ret = walk_page_range(vma->vm_start, vma->vm_end,
6232 &mem_cgroup_move_charge_walk);
6235 * means we have consumed all precharges and failed in
6236 * doing additional charge. Just abandon here.
6240 up_read(&mm->mmap_sem);
6243 static void mem_cgroup_move_task(struct cgroup *cont,
6244 struct cgroup_taskset *tset)
6246 struct task_struct *p = cgroup_taskset_first(tset);
6247 struct mm_struct *mm = get_task_mm(p);
6251 mem_cgroup_move_charge(mm);
6255 mem_cgroup_clear_mc();
6257 #else /* !CONFIG_MMU */
6258 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6259 struct cgroup_taskset *tset)
6263 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6264 struct cgroup_taskset *tset)
6267 static void mem_cgroup_move_task(struct cgroup *cont,
6268 struct cgroup_taskset *tset)
6273 struct cgroup_subsys mem_cgroup_subsys = {
6275 .subsys_id = mem_cgroup_subsys_id,
6276 .create = mem_cgroup_create,
6277 .pre_destroy = mem_cgroup_pre_destroy,
6278 .destroy = mem_cgroup_destroy,
6279 .can_attach = mem_cgroup_can_attach,
6280 .cancel_attach = mem_cgroup_cancel_attach,
6281 .attach = mem_cgroup_move_task,
6282 .base_cftypes = mem_cgroup_files,
6287 #ifdef CONFIG_MEMCG_SWAP
6288 static int __init enable_swap_account(char *s)
6290 /* consider enabled if no parameter or 1 is given */
6291 if (!strcmp(s, "1"))
6292 really_do_swap_account = 1;
6293 else if (!strcmp(s, "0"))
6294 really_do_swap_account = 0;
6297 __setup("swapaccount=", enable_swap_account);