1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/sort.h>
50 #include <linux/seq_file.h>
51 #include <linux/vmalloc.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
60 #include <net/tcp_memcontrol.h>
62 #include <asm/uaccess.h>
64 #include <trace/events/vmscan.h>
66 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
67 EXPORT_SYMBOL(mem_cgroup_subsys);
69 #define MEM_CGROUP_RECLAIM_RETRIES 5
70 static struct mem_cgroup *root_mem_cgroup __read_mostly;
72 #ifdef CONFIG_MEMCG_SWAP
73 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
74 int do_swap_account __read_mostly;
76 /* for remember boot option*/
77 #ifdef CONFIG_MEMCG_SWAP_ENABLED
78 static int really_do_swap_account __initdata = 1;
80 static int really_do_swap_account __initdata = 0;
84 #define do_swap_account 0
89 * Statistics for memory cgroup.
91 enum mem_cgroup_stat_index {
93 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
95 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
96 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
97 MEM_CGROUP_STAT_RSS_HUGE, /* # of pages charged as anon huge */
98 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
99 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
100 MEM_CGROUP_STAT_NSTATS,
103 static const char * const mem_cgroup_stat_names[] = {
111 enum mem_cgroup_events_index {
112 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
113 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
114 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
115 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
116 MEM_CGROUP_EVENTS_NSTATS,
119 static const char * const mem_cgroup_events_names[] = {
126 static const char * const mem_cgroup_lru_names[] = {
135 * Per memcg event counter is incremented at every pagein/pageout. With THP,
136 * it will be incremated by the number of pages. This counter is used for
137 * for trigger some periodic events. This is straightforward and better
138 * than using jiffies etc. to handle periodic memcg event.
140 enum mem_cgroup_events_target {
141 MEM_CGROUP_TARGET_THRESH,
142 MEM_CGROUP_TARGET_SOFTLIMIT,
143 MEM_CGROUP_TARGET_NUMAINFO,
146 #define THRESHOLDS_EVENTS_TARGET 128
147 #define SOFTLIMIT_EVENTS_TARGET 1024
148 #define NUMAINFO_EVENTS_TARGET 1024
150 struct mem_cgroup_stat_cpu {
151 long count[MEM_CGROUP_STAT_NSTATS];
152 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
153 unsigned long nr_page_events;
154 unsigned long targets[MEM_CGROUP_NTARGETS];
157 struct mem_cgroup_reclaim_iter {
159 * last scanned hierarchy member. Valid only if last_dead_count
160 * matches memcg->dead_count of the hierarchy root group.
162 struct mem_cgroup *last_visited;
163 unsigned long last_dead_count;
165 /* scan generation, increased every round-trip */
166 unsigned int generation;
170 * per-zone information in memory controller.
172 struct mem_cgroup_per_zone {
173 struct lruvec lruvec;
174 unsigned long lru_size[NR_LRU_LISTS];
176 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
178 struct rb_node tree_node; /* RB tree node */
179 unsigned long long usage_in_excess;/* Set to the value by which */
180 /* the soft limit is exceeded*/
182 struct mem_cgroup *memcg; /* Back pointer, we cannot */
183 /* use container_of */
186 struct mem_cgroup_per_node {
187 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
191 * Cgroups above their limits are maintained in a RB-Tree, independent of
192 * their hierarchy representation
195 struct mem_cgroup_tree_per_zone {
196 struct rb_root rb_root;
200 struct mem_cgroup_tree_per_node {
201 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
204 struct mem_cgroup_tree {
205 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
208 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
210 struct mem_cgroup_threshold {
211 struct eventfd_ctx *eventfd;
216 struct mem_cgroup_threshold_ary {
217 /* An array index points to threshold just below or equal to usage. */
218 int current_threshold;
219 /* Size of entries[] */
221 /* Array of thresholds */
222 struct mem_cgroup_threshold entries[0];
225 struct mem_cgroup_thresholds {
226 /* Primary thresholds array */
227 struct mem_cgroup_threshold_ary *primary;
229 * Spare threshold array.
230 * This is needed to make mem_cgroup_unregister_event() "never fail".
231 * It must be able to store at least primary->size - 1 entries.
233 struct mem_cgroup_threshold_ary *spare;
237 struct mem_cgroup_eventfd_list {
238 struct list_head list;
239 struct eventfd_ctx *eventfd;
242 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
243 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
246 * The memory controller data structure. The memory controller controls both
247 * page cache and RSS per cgroup. We would eventually like to provide
248 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
249 * to help the administrator determine what knobs to tune.
251 * TODO: Add a water mark for the memory controller. Reclaim will begin when
252 * we hit the water mark. May be even add a low water mark, such that
253 * no reclaim occurs from a cgroup at it's low water mark, this is
254 * a feature that will be implemented much later in the future.
257 struct cgroup_subsys_state css;
259 * the counter to account for memory usage
261 struct res_counter res;
263 /* vmpressure notifications */
264 struct vmpressure vmpressure;
268 * the counter to account for mem+swap usage.
270 struct res_counter memsw;
273 * rcu_freeing is used only when freeing struct mem_cgroup,
274 * so put it into a union to avoid wasting more memory.
275 * It must be disjoint from the css field. It could be
276 * in a union with the res field, but res plays a much
277 * larger part in mem_cgroup life than memsw, and might
278 * be of interest, even at time of free, when debugging.
279 * So share rcu_head with the less interesting memsw.
281 struct rcu_head rcu_freeing;
283 * We also need some space for a worker in deferred freeing.
284 * By the time we call it, rcu_freeing is no longer in use.
286 struct work_struct work_freeing;
290 * the counter to account for kernel memory usage.
292 struct res_counter kmem;
294 * Should the accounting and control be hierarchical, per subtree?
297 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
305 /* OOM-Killer disable */
306 int oom_kill_disable;
308 /* set when res.limit == memsw.limit */
309 bool memsw_is_minimum;
311 /* protect arrays of thresholds */
312 struct mutex thresholds_lock;
314 /* thresholds for memory usage. RCU-protected */
315 struct mem_cgroup_thresholds thresholds;
317 /* thresholds for mem+swap usage. RCU-protected */
318 struct mem_cgroup_thresholds memsw_thresholds;
320 /* For oom notifier event fd */
321 struct list_head oom_notify;
324 * Should we move charges of a task when a task is moved into this
325 * mem_cgroup ? And what type of charges should we move ?
327 unsigned long move_charge_at_immigrate;
329 * set > 0 if pages under this cgroup are moving to other cgroup.
331 atomic_t moving_account;
332 /* taken only while moving_account > 0 */
333 spinlock_t move_lock;
337 struct mem_cgroup_stat_cpu __percpu *stat;
339 * used when a cpu is offlined or other synchronizations
340 * See mem_cgroup_read_stat().
342 struct mem_cgroup_stat_cpu nocpu_base;
343 spinlock_t pcp_counter_lock;
346 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
347 struct tcp_memcontrol tcp_mem;
349 #if defined(CONFIG_MEMCG_KMEM)
350 /* analogous to slab_common's slab_caches list. per-memcg */
351 struct list_head memcg_slab_caches;
352 /* Not a spinlock, we can take a lot of time walking the list */
353 struct mutex slab_caches_mutex;
354 /* Index in the kmem_cache->memcg_params->memcg_caches array */
358 int last_scanned_node;
360 nodemask_t scan_nodes;
361 atomic_t numainfo_events;
362 atomic_t numainfo_updating;
365 struct mem_cgroup_per_node *nodeinfo[0];
366 /* WARNING: nodeinfo must be the last member here */
369 static size_t memcg_size(void)
371 return sizeof(struct mem_cgroup) +
372 nr_node_ids * sizeof(struct mem_cgroup_per_node);
375 /* internal only representation about the status of kmem accounting. */
377 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
378 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
379 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
382 /* We account when limit is on, but only after call sites are patched */
383 #define KMEM_ACCOUNTED_MASK \
384 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
386 #ifdef CONFIG_MEMCG_KMEM
387 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
389 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
392 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
394 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
397 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
399 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
402 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
404 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
407 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
410 * Our caller must use css_get() first, because memcg_uncharge_kmem()
411 * will call css_put() if it sees the memcg is dead.
414 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
415 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
418 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
420 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
421 &memcg->kmem_account_flags);
425 /* Stuffs for move charges at task migration. */
427 * Types of charges to be moved. "move_charge_at_immitgrate" and
428 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
431 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
432 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
436 /* "mc" and its members are protected by cgroup_mutex */
437 static struct move_charge_struct {
438 spinlock_t lock; /* for from, to */
439 struct mem_cgroup *from;
440 struct mem_cgroup *to;
441 unsigned long immigrate_flags;
442 unsigned long precharge;
443 unsigned long moved_charge;
444 unsigned long moved_swap;
445 struct task_struct *moving_task; /* a task moving charges */
446 wait_queue_head_t waitq; /* a waitq for other context */
448 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
449 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
452 static bool move_anon(void)
454 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
457 static bool move_file(void)
459 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
463 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
464 * limit reclaim to prevent infinite loops, if they ever occur.
466 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
467 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
470 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
471 MEM_CGROUP_CHARGE_TYPE_ANON,
472 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
473 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
477 /* for encoding cft->private value on file */
485 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
486 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
487 #define MEMFILE_ATTR(val) ((val) & 0xffff)
488 /* Used for OOM nofiier */
489 #define OOM_CONTROL (0)
492 * Reclaim flags for mem_cgroup_hierarchical_reclaim
494 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
495 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
496 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
497 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
500 * The memcg_create_mutex will be held whenever a new cgroup is created.
501 * As a consequence, any change that needs to protect against new child cgroups
502 * appearing has to hold it as well.
504 static DEFINE_MUTEX(memcg_create_mutex);
506 static void mem_cgroup_put(struct mem_cgroup *memcg);
509 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
511 return container_of(s, struct mem_cgroup, css);
514 /* Some nice accessors for the vmpressure. */
515 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
518 memcg = root_mem_cgroup;
519 return &memcg->vmpressure;
522 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
524 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
527 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
529 return &mem_cgroup_from_css(css)->vmpressure;
532 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
534 return (memcg == root_mem_cgroup);
537 /* Writing them here to avoid exposing memcg's inner layout */
538 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
540 void sock_update_memcg(struct sock *sk)
542 if (mem_cgroup_sockets_enabled) {
543 struct mem_cgroup *memcg;
544 struct cg_proto *cg_proto;
546 BUG_ON(!sk->sk_prot->proto_cgroup);
548 /* Socket cloning can throw us here with sk_cgrp already
549 * filled. It won't however, necessarily happen from
550 * process context. So the test for root memcg given
551 * the current task's memcg won't help us in this case.
553 * Respecting the original socket's memcg is a better
554 * decision in this case.
557 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
558 css_get(&sk->sk_cgrp->memcg->css);
563 memcg = mem_cgroup_from_task(current);
564 cg_proto = sk->sk_prot->proto_cgroup(memcg);
565 if (!mem_cgroup_is_root(memcg) &&
566 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
567 sk->sk_cgrp = cg_proto;
572 EXPORT_SYMBOL(sock_update_memcg);
574 void sock_release_memcg(struct sock *sk)
576 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
577 struct mem_cgroup *memcg;
578 WARN_ON(!sk->sk_cgrp->memcg);
579 memcg = sk->sk_cgrp->memcg;
580 css_put(&sk->sk_cgrp->memcg->css);
584 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
586 if (!memcg || mem_cgroup_is_root(memcg))
589 return &memcg->tcp_mem.cg_proto;
591 EXPORT_SYMBOL(tcp_proto_cgroup);
593 static void disarm_sock_keys(struct mem_cgroup *memcg)
595 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
597 static_key_slow_dec(&memcg_socket_limit_enabled);
600 static void disarm_sock_keys(struct mem_cgroup *memcg)
605 #ifdef CONFIG_MEMCG_KMEM
607 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
608 * There are two main reasons for not using the css_id for this:
609 * 1) this works better in sparse environments, where we have a lot of memcgs,
610 * but only a few kmem-limited. Or also, if we have, for instance, 200
611 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
612 * 200 entry array for that.
614 * 2) In order not to violate the cgroup API, we would like to do all memory
615 * allocation in ->create(). At that point, we haven't yet allocated the
616 * css_id. Having a separate index prevents us from messing with the cgroup
619 * The current size of the caches array is stored in
620 * memcg_limited_groups_array_size. It will double each time we have to
623 static DEFINE_IDA(kmem_limited_groups);
624 int memcg_limited_groups_array_size;
627 * MIN_SIZE is different than 1, because we would like to avoid going through
628 * the alloc/free process all the time. In a small machine, 4 kmem-limited
629 * cgroups is a reasonable guess. In the future, it could be a parameter or
630 * tunable, but that is strictly not necessary.
632 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
633 * this constant directly from cgroup, but it is understandable that this is
634 * better kept as an internal representation in cgroup.c. In any case, the
635 * css_id space is not getting any smaller, and we don't have to necessarily
636 * increase ours as well if it increases.
638 #define MEMCG_CACHES_MIN_SIZE 4
639 #define MEMCG_CACHES_MAX_SIZE 65535
642 * A lot of the calls to the cache allocation functions are expected to be
643 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
644 * conditional to this static branch, we'll have to allow modules that does
645 * kmem_cache_alloc and the such to see this symbol as well
647 struct static_key memcg_kmem_enabled_key;
648 EXPORT_SYMBOL(memcg_kmem_enabled_key);
650 static void disarm_kmem_keys(struct mem_cgroup *memcg)
652 if (memcg_kmem_is_active(memcg)) {
653 static_key_slow_dec(&memcg_kmem_enabled_key);
654 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
657 * This check can't live in kmem destruction function,
658 * since the charges will outlive the cgroup
660 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
663 static void disarm_kmem_keys(struct mem_cgroup *memcg)
666 #endif /* CONFIG_MEMCG_KMEM */
668 static void disarm_static_keys(struct mem_cgroup *memcg)
670 disarm_sock_keys(memcg);
671 disarm_kmem_keys(memcg);
674 static void drain_all_stock_async(struct mem_cgroup *memcg);
676 static struct mem_cgroup_per_zone *
677 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
679 VM_BUG_ON((unsigned)nid >= nr_node_ids);
680 return &memcg->nodeinfo[nid]->zoneinfo[zid];
683 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
688 static struct mem_cgroup_per_zone *
689 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
691 int nid = page_to_nid(page);
692 int zid = page_zonenum(page);
694 return mem_cgroup_zoneinfo(memcg, nid, zid);
697 static struct mem_cgroup_tree_per_zone *
698 soft_limit_tree_node_zone(int nid, int zid)
700 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
703 static struct mem_cgroup_tree_per_zone *
704 soft_limit_tree_from_page(struct page *page)
706 int nid = page_to_nid(page);
707 int zid = page_zonenum(page);
709 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
713 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
714 struct mem_cgroup_per_zone *mz,
715 struct mem_cgroup_tree_per_zone *mctz,
716 unsigned long long new_usage_in_excess)
718 struct rb_node **p = &mctz->rb_root.rb_node;
719 struct rb_node *parent = NULL;
720 struct mem_cgroup_per_zone *mz_node;
725 mz->usage_in_excess = new_usage_in_excess;
726 if (!mz->usage_in_excess)
730 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
732 if (mz->usage_in_excess < mz_node->usage_in_excess)
735 * We can't avoid mem cgroups that are over their soft
736 * limit by the same amount
738 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
741 rb_link_node(&mz->tree_node, parent, p);
742 rb_insert_color(&mz->tree_node, &mctz->rb_root);
747 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
748 struct mem_cgroup_per_zone *mz,
749 struct mem_cgroup_tree_per_zone *mctz)
753 rb_erase(&mz->tree_node, &mctz->rb_root);
758 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
759 struct mem_cgroup_per_zone *mz,
760 struct mem_cgroup_tree_per_zone *mctz)
762 spin_lock(&mctz->lock);
763 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
764 spin_unlock(&mctz->lock);
768 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
770 unsigned long long excess;
771 struct mem_cgroup_per_zone *mz;
772 struct mem_cgroup_tree_per_zone *mctz;
773 int nid = page_to_nid(page);
774 int zid = page_zonenum(page);
775 mctz = soft_limit_tree_from_page(page);
778 * Necessary to update all ancestors when hierarchy is used.
779 * because their event counter is not touched.
781 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
782 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
783 excess = res_counter_soft_limit_excess(&memcg->res);
785 * We have to update the tree if mz is on RB-tree or
786 * mem is over its softlimit.
788 if (excess || mz->on_tree) {
789 spin_lock(&mctz->lock);
790 /* if on-tree, remove it */
792 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
794 * Insert again. mz->usage_in_excess will be updated.
795 * If excess is 0, no tree ops.
797 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
798 spin_unlock(&mctz->lock);
803 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
806 struct mem_cgroup_per_zone *mz;
807 struct mem_cgroup_tree_per_zone *mctz;
809 for_each_node(node) {
810 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
811 mz = mem_cgroup_zoneinfo(memcg, node, zone);
812 mctz = soft_limit_tree_node_zone(node, zone);
813 mem_cgroup_remove_exceeded(memcg, mz, mctz);
818 static struct mem_cgroup_per_zone *
819 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
821 struct rb_node *rightmost = NULL;
822 struct mem_cgroup_per_zone *mz;
826 rightmost = rb_last(&mctz->rb_root);
828 goto done; /* Nothing to reclaim from */
830 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
832 * Remove the node now but someone else can add it back,
833 * we will to add it back at the end of reclaim to its correct
834 * position in the tree.
836 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
837 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
838 !css_tryget(&mz->memcg->css))
844 static struct mem_cgroup_per_zone *
845 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
847 struct mem_cgroup_per_zone *mz;
849 spin_lock(&mctz->lock);
850 mz = __mem_cgroup_largest_soft_limit_node(mctz);
851 spin_unlock(&mctz->lock);
856 * Implementation Note: reading percpu statistics for memcg.
858 * Both of vmstat[] and percpu_counter has threshold and do periodic
859 * synchronization to implement "quick" read. There are trade-off between
860 * reading cost and precision of value. Then, we may have a chance to implement
861 * a periodic synchronizion of counter in memcg's counter.
863 * But this _read() function is used for user interface now. The user accounts
864 * memory usage by memory cgroup and he _always_ requires exact value because
865 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
866 * have to visit all online cpus and make sum. So, for now, unnecessary
867 * synchronization is not implemented. (just implemented for cpu hotplug)
869 * If there are kernel internal actions which can make use of some not-exact
870 * value, and reading all cpu value can be performance bottleneck in some
871 * common workload, threashold and synchonization as vmstat[] should be
874 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
875 enum mem_cgroup_stat_index idx)
881 for_each_online_cpu(cpu)
882 val += per_cpu(memcg->stat->count[idx], cpu);
883 #ifdef CONFIG_HOTPLUG_CPU
884 spin_lock(&memcg->pcp_counter_lock);
885 val += memcg->nocpu_base.count[idx];
886 spin_unlock(&memcg->pcp_counter_lock);
892 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
895 int val = (charge) ? 1 : -1;
896 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
899 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
900 enum mem_cgroup_events_index idx)
902 unsigned long val = 0;
905 for_each_online_cpu(cpu)
906 val += per_cpu(memcg->stat->events[idx], cpu);
907 #ifdef CONFIG_HOTPLUG_CPU
908 spin_lock(&memcg->pcp_counter_lock);
909 val += memcg->nocpu_base.events[idx];
910 spin_unlock(&memcg->pcp_counter_lock);
915 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
917 bool anon, int nr_pages)
922 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
923 * counted as CACHE even if it's on ANON LRU.
926 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
929 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
932 if (PageTransHuge(page))
933 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
936 /* pagein of a big page is an event. So, ignore page size */
938 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
940 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
941 nr_pages = -nr_pages; /* for event */
944 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
950 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
952 struct mem_cgroup_per_zone *mz;
954 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
955 return mz->lru_size[lru];
959 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
960 unsigned int lru_mask)
962 struct mem_cgroup_per_zone *mz;
964 unsigned long ret = 0;
966 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
969 if (BIT(lru) & lru_mask)
970 ret += mz->lru_size[lru];
976 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
977 int nid, unsigned int lru_mask)
982 for (zid = 0; zid < MAX_NR_ZONES; zid++)
983 total += mem_cgroup_zone_nr_lru_pages(memcg,
989 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
990 unsigned int lru_mask)
995 for_each_node_state(nid, N_MEMORY)
996 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1000 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1001 enum mem_cgroup_events_target target)
1003 unsigned long val, next;
1005 val = __this_cpu_read(memcg->stat->nr_page_events);
1006 next = __this_cpu_read(memcg->stat->targets[target]);
1007 /* from time_after() in jiffies.h */
1008 if ((long)next - (long)val < 0) {
1010 case MEM_CGROUP_TARGET_THRESH:
1011 next = val + THRESHOLDS_EVENTS_TARGET;
1013 case MEM_CGROUP_TARGET_SOFTLIMIT:
1014 next = val + SOFTLIMIT_EVENTS_TARGET;
1016 case MEM_CGROUP_TARGET_NUMAINFO:
1017 next = val + NUMAINFO_EVENTS_TARGET;
1022 __this_cpu_write(memcg->stat->targets[target], next);
1029 * Check events in order.
1032 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1035 /* threshold event is triggered in finer grain than soft limit */
1036 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1037 MEM_CGROUP_TARGET_THRESH))) {
1039 bool do_numainfo __maybe_unused;
1041 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1042 MEM_CGROUP_TARGET_SOFTLIMIT);
1043 #if MAX_NUMNODES > 1
1044 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1045 MEM_CGROUP_TARGET_NUMAINFO);
1049 mem_cgroup_threshold(memcg);
1050 if (unlikely(do_softlimit))
1051 mem_cgroup_update_tree(memcg, page);
1052 #if MAX_NUMNODES > 1
1053 if (unlikely(do_numainfo))
1054 atomic_inc(&memcg->numainfo_events);
1060 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1062 return mem_cgroup_from_css(
1063 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1066 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1069 * mm_update_next_owner() may clear mm->owner to NULL
1070 * if it races with swapoff, page migration, etc.
1071 * So this can be called with p == NULL.
1076 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1079 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1081 struct mem_cgroup *memcg = NULL;
1086 * Because we have no locks, mm->owner's may be being moved to other
1087 * cgroup. We use css_tryget() here even if this looks
1088 * pessimistic (rather than adding locks here).
1092 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1093 if (unlikely(!memcg))
1095 } while (!css_tryget(&memcg->css));
1101 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1102 * ref. count) or NULL if the whole root's subtree has been visited.
1104 * helper function to be used by mem_cgroup_iter
1106 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1107 struct mem_cgroup *last_visited)
1109 struct cgroup *prev_cgroup, *next_cgroup;
1112 * Root is not visited by cgroup iterators so it needs an
1118 prev_cgroup = (last_visited == root) ? NULL
1119 : last_visited->css.cgroup;
1121 next_cgroup = cgroup_next_descendant_pre(
1122 prev_cgroup, root->css.cgroup);
1125 * Even if we found a group we have to make sure it is
1126 * alive. css && !memcg means that the groups should be
1127 * skipped and we should continue the tree walk.
1128 * last_visited css is safe to use because it is
1129 * protected by css_get and the tree walk is rcu safe.
1132 struct mem_cgroup *mem = mem_cgroup_from_cont(
1134 if (css_tryget(&mem->css))
1137 prev_cgroup = next_cgroup;
1145 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1148 * When a group in the hierarchy below root is destroyed, the
1149 * hierarchy iterator can no longer be trusted since it might
1150 * have pointed to the destroyed group. Invalidate it.
1152 atomic_inc(&root->dead_count);
1155 static struct mem_cgroup *
1156 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1157 struct mem_cgroup *root,
1160 struct mem_cgroup *position = NULL;
1162 * A cgroup destruction happens in two stages: offlining and
1163 * release. They are separated by a RCU grace period.
1165 * If the iterator is valid, we may still race with an
1166 * offlining. The RCU lock ensures the object won't be
1167 * released, tryget will fail if we lost the race.
1169 *sequence = atomic_read(&root->dead_count);
1170 if (iter->last_dead_count == *sequence) {
1172 position = iter->last_visited;
1173 if (position && !css_tryget(&position->css))
1179 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1180 struct mem_cgroup *last_visited,
1181 struct mem_cgroup *new_position,
1185 css_put(&last_visited->css);
1187 * We store the sequence count from the time @last_visited was
1188 * loaded successfully instead of rereading it here so that we
1189 * don't lose destruction events in between. We could have
1190 * raced with the destruction of @new_position after all.
1192 iter->last_visited = new_position;
1194 iter->last_dead_count = sequence;
1198 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1199 * @root: hierarchy root
1200 * @prev: previously returned memcg, NULL on first invocation
1201 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1203 * Returns references to children of the hierarchy below @root, or
1204 * @root itself, or %NULL after a full round-trip.
1206 * Caller must pass the return value in @prev on subsequent
1207 * invocations for reference counting, or use mem_cgroup_iter_break()
1208 * to cancel a hierarchy walk before the round-trip is complete.
1210 * Reclaimers can specify a zone and a priority level in @reclaim to
1211 * divide up the memcgs in the hierarchy among all concurrent
1212 * reclaimers operating on the same zone and priority.
1214 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1215 struct mem_cgroup *prev,
1216 struct mem_cgroup_reclaim_cookie *reclaim)
1218 struct mem_cgroup *memcg = NULL;
1219 struct mem_cgroup *last_visited = NULL;
1221 if (mem_cgroup_disabled())
1225 root = root_mem_cgroup;
1227 if (prev && !reclaim)
1228 last_visited = prev;
1230 if (!root->use_hierarchy && root != root_mem_cgroup) {
1238 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1239 int uninitialized_var(seq);
1242 int nid = zone_to_nid(reclaim->zone);
1243 int zid = zone_idx(reclaim->zone);
1244 struct mem_cgroup_per_zone *mz;
1246 mz = mem_cgroup_zoneinfo(root, nid, zid);
1247 iter = &mz->reclaim_iter[reclaim->priority];
1248 if (prev && reclaim->generation != iter->generation) {
1249 iter->last_visited = NULL;
1253 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1256 memcg = __mem_cgroup_iter_next(root, last_visited);
1259 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1263 else if (!prev && memcg)
1264 reclaim->generation = iter->generation;
1273 if (prev && prev != root)
1274 css_put(&prev->css);
1280 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1281 * @root: hierarchy root
1282 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1284 void mem_cgroup_iter_break(struct mem_cgroup *root,
1285 struct mem_cgroup *prev)
1288 root = root_mem_cgroup;
1289 if (prev && prev != root)
1290 css_put(&prev->css);
1294 * Iteration constructs for visiting all cgroups (under a tree). If
1295 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1296 * be used for reference counting.
1298 #define for_each_mem_cgroup_tree(iter, root) \
1299 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1301 iter = mem_cgroup_iter(root, iter, NULL))
1303 #define for_each_mem_cgroup(iter) \
1304 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1306 iter = mem_cgroup_iter(NULL, iter, NULL))
1308 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1310 struct mem_cgroup *memcg;
1313 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1314 if (unlikely(!memcg))
1319 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1322 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1330 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1333 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1334 * @zone: zone of the wanted lruvec
1335 * @memcg: memcg of the wanted lruvec
1337 * Returns the lru list vector holding pages for the given @zone and
1338 * @mem. This can be the global zone lruvec, if the memory controller
1341 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1342 struct mem_cgroup *memcg)
1344 struct mem_cgroup_per_zone *mz;
1345 struct lruvec *lruvec;
1347 if (mem_cgroup_disabled()) {
1348 lruvec = &zone->lruvec;
1352 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1353 lruvec = &mz->lruvec;
1356 * Since a node can be onlined after the mem_cgroup was created,
1357 * we have to be prepared to initialize lruvec->zone here;
1358 * and if offlined then reonlined, we need to reinitialize it.
1360 if (unlikely(lruvec->zone != zone))
1361 lruvec->zone = zone;
1366 * Following LRU functions are allowed to be used without PCG_LOCK.
1367 * Operations are called by routine of global LRU independently from memcg.
1368 * What we have to take care of here is validness of pc->mem_cgroup.
1370 * Changes to pc->mem_cgroup happens when
1373 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1374 * It is added to LRU before charge.
1375 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1376 * When moving account, the page is not on LRU. It's isolated.
1380 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1382 * @zone: zone of the page
1384 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1386 struct mem_cgroup_per_zone *mz;
1387 struct mem_cgroup *memcg;
1388 struct page_cgroup *pc;
1389 struct lruvec *lruvec;
1391 if (mem_cgroup_disabled()) {
1392 lruvec = &zone->lruvec;
1396 pc = lookup_page_cgroup(page);
1397 memcg = pc->mem_cgroup;
1400 * Surreptitiously switch any uncharged offlist page to root:
1401 * an uncharged page off lru does nothing to secure
1402 * its former mem_cgroup from sudden removal.
1404 * Our caller holds lru_lock, and PageCgroupUsed is updated
1405 * under page_cgroup lock: between them, they make all uses
1406 * of pc->mem_cgroup safe.
1408 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1409 pc->mem_cgroup = memcg = root_mem_cgroup;
1411 mz = page_cgroup_zoneinfo(memcg, page);
1412 lruvec = &mz->lruvec;
1415 * Since a node can be onlined after the mem_cgroup was created,
1416 * we have to be prepared to initialize lruvec->zone here;
1417 * and if offlined then reonlined, we need to reinitialize it.
1419 if (unlikely(lruvec->zone != zone))
1420 lruvec->zone = zone;
1425 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1426 * @lruvec: mem_cgroup per zone lru vector
1427 * @lru: index of lru list the page is sitting on
1428 * @nr_pages: positive when adding or negative when removing
1430 * This function must be called when a page is added to or removed from an
1433 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1436 struct mem_cgroup_per_zone *mz;
1437 unsigned long *lru_size;
1439 if (mem_cgroup_disabled())
1442 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1443 lru_size = mz->lru_size + lru;
1444 *lru_size += nr_pages;
1445 VM_BUG_ON((long)(*lru_size) < 0);
1449 * Checks whether given mem is same or in the root_mem_cgroup's
1452 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1453 struct mem_cgroup *memcg)
1455 if (root_memcg == memcg)
1457 if (!root_memcg->use_hierarchy || !memcg)
1459 return css_is_ancestor(&memcg->css, &root_memcg->css);
1462 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1463 struct mem_cgroup *memcg)
1468 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1473 bool task_in_mem_cgroup(struct task_struct *task,
1474 const struct mem_cgroup *memcg)
1476 struct mem_cgroup *curr = NULL;
1477 struct task_struct *p;
1480 p = find_lock_task_mm(task);
1482 curr = try_get_mem_cgroup_from_mm(p->mm);
1486 * All threads may have already detached their mm's, but the oom
1487 * killer still needs to detect if they have already been oom
1488 * killed to prevent needlessly killing additional tasks.
1491 curr = mem_cgroup_from_task(task);
1493 css_get(&curr->css);
1499 * We should check use_hierarchy of "memcg" not "curr". Because checking
1500 * use_hierarchy of "curr" here make this function true if hierarchy is
1501 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1502 * hierarchy(even if use_hierarchy is disabled in "memcg").
1504 ret = mem_cgroup_same_or_subtree(memcg, curr);
1505 css_put(&curr->css);
1509 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1511 unsigned long inactive_ratio;
1512 unsigned long inactive;
1513 unsigned long active;
1516 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1517 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1519 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1521 inactive_ratio = int_sqrt(10 * gb);
1525 return inactive * inactive_ratio < active;
1528 #define mem_cgroup_from_res_counter(counter, member) \
1529 container_of(counter, struct mem_cgroup, member)
1532 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1533 * @memcg: the memory cgroup
1535 * Returns the maximum amount of memory @mem can be charged with, in
1538 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1540 unsigned long long margin;
1542 margin = res_counter_margin(&memcg->res);
1543 if (do_swap_account)
1544 margin = min(margin, res_counter_margin(&memcg->memsw));
1545 return margin >> PAGE_SHIFT;
1548 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1550 struct cgroup *cgrp = memcg->css.cgroup;
1553 if (cgrp->parent == NULL)
1554 return vm_swappiness;
1556 return memcg->swappiness;
1560 * memcg->moving_account is used for checking possibility that some thread is
1561 * calling move_account(). When a thread on CPU-A starts moving pages under
1562 * a memcg, other threads should check memcg->moving_account under
1563 * rcu_read_lock(), like this:
1567 * memcg->moving_account+1 if (memcg->mocing_account)
1569 * synchronize_rcu() update something.
1574 /* for quick checking without looking up memcg */
1575 atomic_t memcg_moving __read_mostly;
1577 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1579 atomic_inc(&memcg_moving);
1580 atomic_inc(&memcg->moving_account);
1584 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1587 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1588 * We check NULL in callee rather than caller.
1591 atomic_dec(&memcg_moving);
1592 atomic_dec(&memcg->moving_account);
1597 * 2 routines for checking "mem" is under move_account() or not.
1599 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1600 * is used for avoiding races in accounting. If true,
1601 * pc->mem_cgroup may be overwritten.
1603 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1604 * under hierarchy of moving cgroups. This is for
1605 * waiting at hith-memory prressure caused by "move".
1608 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1610 VM_BUG_ON(!rcu_read_lock_held());
1611 return atomic_read(&memcg->moving_account) > 0;
1614 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1616 struct mem_cgroup *from;
1617 struct mem_cgroup *to;
1620 * Unlike task_move routines, we access mc.to, mc.from not under
1621 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1623 spin_lock(&mc.lock);
1629 ret = mem_cgroup_same_or_subtree(memcg, from)
1630 || mem_cgroup_same_or_subtree(memcg, to);
1632 spin_unlock(&mc.lock);
1636 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1638 if (mc.moving_task && current != mc.moving_task) {
1639 if (mem_cgroup_under_move(memcg)) {
1641 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1642 /* moving charge context might have finished. */
1645 finish_wait(&mc.waitq, &wait);
1653 * Take this lock when
1654 * - a code tries to modify page's memcg while it's USED.
1655 * - a code tries to modify page state accounting in a memcg.
1656 * see mem_cgroup_stolen(), too.
1658 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1659 unsigned long *flags)
1661 spin_lock_irqsave(&memcg->move_lock, *flags);
1664 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1665 unsigned long *flags)
1667 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1670 #define K(x) ((x) << (PAGE_SHIFT-10))
1672 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1673 * @memcg: The memory cgroup that went over limit
1674 * @p: Task that is going to be killed
1676 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1679 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1681 struct cgroup *task_cgrp;
1682 struct cgroup *mem_cgrp;
1684 * Need a buffer in BSS, can't rely on allocations. The code relies
1685 * on the assumption that OOM is serialized for memory controller.
1686 * If this assumption is broken, revisit this code.
1688 static char memcg_name[PATH_MAX];
1690 struct mem_cgroup *iter;
1698 mem_cgrp = memcg->css.cgroup;
1699 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1701 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1704 * Unfortunately, we are unable to convert to a useful name
1705 * But we'll still print out the usage information
1712 pr_info("Task in %s killed", memcg_name);
1715 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1723 * Continues from above, so we don't need an KERN_ level
1725 pr_cont(" as a result of limit of %s\n", memcg_name);
1728 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1729 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1730 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1731 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1732 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1733 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1734 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1735 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1736 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1737 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1738 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1739 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1741 for_each_mem_cgroup_tree(iter, memcg) {
1742 pr_info("Memory cgroup stats");
1745 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1747 pr_cont(" for %s", memcg_name);
1751 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1752 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1754 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1755 K(mem_cgroup_read_stat(iter, i)));
1758 for (i = 0; i < NR_LRU_LISTS; i++)
1759 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1760 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1767 * This function returns the number of memcg under hierarchy tree. Returns
1768 * 1(self count) if no children.
1770 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1773 struct mem_cgroup *iter;
1775 for_each_mem_cgroup_tree(iter, memcg)
1781 * Return the memory (and swap, if configured) limit for a memcg.
1783 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1787 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1790 * Do not consider swap space if we cannot swap due to swappiness
1792 if (mem_cgroup_swappiness(memcg)) {
1795 limit += total_swap_pages << PAGE_SHIFT;
1796 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1799 * If memsw is finite and limits the amount of swap space
1800 * available to this memcg, return that limit.
1802 limit = min(limit, memsw);
1808 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1811 struct mem_cgroup *iter;
1812 unsigned long chosen_points = 0;
1813 unsigned long totalpages;
1814 unsigned int points = 0;
1815 struct task_struct *chosen = NULL;
1818 * If current has a pending SIGKILL or is exiting, then automatically
1819 * select it. The goal is to allow it to allocate so that it may
1820 * quickly exit and free its memory.
1822 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1823 set_thread_flag(TIF_MEMDIE);
1827 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1828 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1829 for_each_mem_cgroup_tree(iter, memcg) {
1830 struct cgroup *cgroup = iter->css.cgroup;
1831 struct cgroup_iter it;
1832 struct task_struct *task;
1834 cgroup_iter_start(cgroup, &it);
1835 while ((task = cgroup_iter_next(cgroup, &it))) {
1836 switch (oom_scan_process_thread(task, totalpages, NULL,
1838 case OOM_SCAN_SELECT:
1840 put_task_struct(chosen);
1842 chosen_points = ULONG_MAX;
1843 get_task_struct(chosen);
1845 case OOM_SCAN_CONTINUE:
1847 case OOM_SCAN_ABORT:
1848 cgroup_iter_end(cgroup, &it);
1849 mem_cgroup_iter_break(memcg, iter);
1851 put_task_struct(chosen);
1856 points = oom_badness(task, memcg, NULL, totalpages);
1857 if (points > chosen_points) {
1859 put_task_struct(chosen);
1861 chosen_points = points;
1862 get_task_struct(chosen);
1865 cgroup_iter_end(cgroup, &it);
1870 points = chosen_points * 1000 / totalpages;
1871 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1872 NULL, "Memory cgroup out of memory");
1875 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1877 unsigned long flags)
1879 unsigned long total = 0;
1880 bool noswap = false;
1883 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1885 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1888 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1890 drain_all_stock_async(memcg);
1891 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1893 * Allow limit shrinkers, which are triggered directly
1894 * by userspace, to catch signals and stop reclaim
1895 * after minimal progress, regardless of the margin.
1897 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1899 if (mem_cgroup_margin(memcg))
1902 * If nothing was reclaimed after two attempts, there
1903 * may be no reclaimable pages in this hierarchy.
1912 * test_mem_cgroup_node_reclaimable
1913 * @memcg: the target memcg
1914 * @nid: the node ID to be checked.
1915 * @noswap : specify true here if the user wants flle only information.
1917 * This function returns whether the specified memcg contains any
1918 * reclaimable pages on a node. Returns true if there are any reclaimable
1919 * pages in the node.
1921 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1922 int nid, bool noswap)
1924 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1926 if (noswap || !total_swap_pages)
1928 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1933 #if MAX_NUMNODES > 1
1936 * Always updating the nodemask is not very good - even if we have an empty
1937 * list or the wrong list here, we can start from some node and traverse all
1938 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1941 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1945 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1946 * pagein/pageout changes since the last update.
1948 if (!atomic_read(&memcg->numainfo_events))
1950 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1953 /* make a nodemask where this memcg uses memory from */
1954 memcg->scan_nodes = node_states[N_MEMORY];
1956 for_each_node_mask(nid, node_states[N_MEMORY]) {
1958 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1959 node_clear(nid, memcg->scan_nodes);
1962 atomic_set(&memcg->numainfo_events, 0);
1963 atomic_set(&memcg->numainfo_updating, 0);
1967 * Selecting a node where we start reclaim from. Because what we need is just
1968 * reducing usage counter, start from anywhere is O,K. Considering
1969 * memory reclaim from current node, there are pros. and cons.
1971 * Freeing memory from current node means freeing memory from a node which
1972 * we'll use or we've used. So, it may make LRU bad. And if several threads
1973 * hit limits, it will see a contention on a node. But freeing from remote
1974 * node means more costs for memory reclaim because of memory latency.
1976 * Now, we use round-robin. Better algorithm is welcomed.
1978 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1982 mem_cgroup_may_update_nodemask(memcg);
1983 node = memcg->last_scanned_node;
1985 node = next_node(node, memcg->scan_nodes);
1986 if (node == MAX_NUMNODES)
1987 node = first_node(memcg->scan_nodes);
1989 * We call this when we hit limit, not when pages are added to LRU.
1990 * No LRU may hold pages because all pages are UNEVICTABLE or
1991 * memcg is too small and all pages are not on LRU. In that case,
1992 * we use curret node.
1994 if (unlikely(node == MAX_NUMNODES))
1995 node = numa_node_id();
1997 memcg->last_scanned_node = node;
2002 * Check all nodes whether it contains reclaimable pages or not.
2003 * For quick scan, we make use of scan_nodes. This will allow us to skip
2004 * unused nodes. But scan_nodes is lazily updated and may not cotain
2005 * enough new information. We need to do double check.
2007 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2012 * quick check...making use of scan_node.
2013 * We can skip unused nodes.
2015 if (!nodes_empty(memcg->scan_nodes)) {
2016 for (nid = first_node(memcg->scan_nodes);
2018 nid = next_node(nid, memcg->scan_nodes)) {
2020 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2025 * Check rest of nodes.
2027 for_each_node_state(nid, N_MEMORY) {
2028 if (node_isset(nid, memcg->scan_nodes))
2030 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2037 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2042 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2044 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2048 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2051 unsigned long *total_scanned)
2053 struct mem_cgroup *victim = NULL;
2056 unsigned long excess;
2057 unsigned long nr_scanned;
2058 struct mem_cgroup_reclaim_cookie reclaim = {
2063 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2066 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2071 * If we have not been able to reclaim
2072 * anything, it might because there are
2073 * no reclaimable pages under this hierarchy
2078 * We want to do more targeted reclaim.
2079 * excess >> 2 is not to excessive so as to
2080 * reclaim too much, nor too less that we keep
2081 * coming back to reclaim from this cgroup
2083 if (total >= (excess >> 2) ||
2084 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2089 if (!mem_cgroup_reclaimable(victim, false))
2091 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2093 *total_scanned += nr_scanned;
2094 if (!res_counter_soft_limit_excess(&root_memcg->res))
2097 mem_cgroup_iter_break(root_memcg, victim);
2102 * Check OOM-Killer is already running under our hierarchy.
2103 * If someone is running, return false.
2104 * Has to be called with memcg_oom_lock
2106 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
2108 struct mem_cgroup *iter, *failed = NULL;
2110 for_each_mem_cgroup_tree(iter, memcg) {
2111 if (iter->oom_lock) {
2113 * this subtree of our hierarchy is already locked
2114 * so we cannot give a lock.
2117 mem_cgroup_iter_break(memcg, iter);
2120 iter->oom_lock = true;
2127 * OK, we failed to lock the whole subtree so we have to clean up
2128 * what we set up to the failing subtree
2130 for_each_mem_cgroup_tree(iter, memcg) {
2131 if (iter == failed) {
2132 mem_cgroup_iter_break(memcg, iter);
2135 iter->oom_lock = false;
2141 * Has to be called with memcg_oom_lock
2143 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2145 struct mem_cgroup *iter;
2147 for_each_mem_cgroup_tree(iter, memcg)
2148 iter->oom_lock = false;
2152 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2154 struct mem_cgroup *iter;
2156 for_each_mem_cgroup_tree(iter, memcg)
2157 atomic_inc(&iter->under_oom);
2160 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2162 struct mem_cgroup *iter;
2165 * When a new child is created while the hierarchy is under oom,
2166 * mem_cgroup_oom_lock() may not be called. We have to use
2167 * atomic_add_unless() here.
2169 for_each_mem_cgroup_tree(iter, memcg)
2170 atomic_add_unless(&iter->under_oom, -1, 0);
2173 static DEFINE_SPINLOCK(memcg_oom_lock);
2174 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2176 struct oom_wait_info {
2177 struct mem_cgroup *memcg;
2181 static int memcg_oom_wake_function(wait_queue_t *wait,
2182 unsigned mode, int sync, void *arg)
2184 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2185 struct mem_cgroup *oom_wait_memcg;
2186 struct oom_wait_info *oom_wait_info;
2188 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2189 oom_wait_memcg = oom_wait_info->memcg;
2192 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2193 * Then we can use css_is_ancestor without taking care of RCU.
2195 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2196 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2198 return autoremove_wake_function(wait, mode, sync, arg);
2201 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2203 /* for filtering, pass "memcg" as argument. */
2204 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2207 static void memcg_oom_recover(struct mem_cgroup *memcg)
2209 if (memcg && atomic_read(&memcg->under_oom))
2210 memcg_wakeup_oom(memcg);
2214 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2216 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2219 struct oom_wait_info owait;
2220 bool locked, need_to_kill;
2222 owait.memcg = memcg;
2223 owait.wait.flags = 0;
2224 owait.wait.func = memcg_oom_wake_function;
2225 owait.wait.private = current;
2226 INIT_LIST_HEAD(&owait.wait.task_list);
2227 need_to_kill = true;
2228 mem_cgroup_mark_under_oom(memcg);
2230 /* At first, try to OOM lock hierarchy under memcg.*/
2231 spin_lock(&memcg_oom_lock);
2232 locked = mem_cgroup_oom_lock(memcg);
2234 * Even if signal_pending(), we can't quit charge() loop without
2235 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2236 * under OOM is always welcomed, use TASK_KILLABLE here.
2238 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2239 if (!locked || memcg->oom_kill_disable)
2240 need_to_kill = false;
2242 mem_cgroup_oom_notify(memcg);
2243 spin_unlock(&memcg_oom_lock);
2246 finish_wait(&memcg_oom_waitq, &owait.wait);
2247 mem_cgroup_out_of_memory(memcg, mask, order);
2250 finish_wait(&memcg_oom_waitq, &owait.wait);
2252 spin_lock(&memcg_oom_lock);
2254 mem_cgroup_oom_unlock(memcg);
2255 memcg_wakeup_oom(memcg);
2256 spin_unlock(&memcg_oom_lock);
2258 mem_cgroup_unmark_under_oom(memcg);
2260 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2262 /* Give chance to dying process */
2263 schedule_timeout_uninterruptible(1);
2268 * Currently used to update mapped file statistics, but the routine can be
2269 * generalized to update other statistics as well.
2271 * Notes: Race condition
2273 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2274 * it tends to be costly. But considering some conditions, we doesn't need
2275 * to do so _always_.
2277 * Considering "charge", lock_page_cgroup() is not required because all
2278 * file-stat operations happen after a page is attached to radix-tree. There
2279 * are no race with "charge".
2281 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2282 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2283 * if there are race with "uncharge". Statistics itself is properly handled
2286 * Considering "move", this is an only case we see a race. To make the race
2287 * small, we check mm->moving_account and detect there are possibility of race
2288 * If there is, we take a lock.
2291 void __mem_cgroup_begin_update_page_stat(struct page *page,
2292 bool *locked, unsigned long *flags)
2294 struct mem_cgroup *memcg;
2295 struct page_cgroup *pc;
2297 pc = lookup_page_cgroup(page);
2299 memcg = pc->mem_cgroup;
2300 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2303 * If this memory cgroup is not under account moving, we don't
2304 * need to take move_lock_mem_cgroup(). Because we already hold
2305 * rcu_read_lock(), any calls to move_account will be delayed until
2306 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2308 if (!mem_cgroup_stolen(memcg))
2311 move_lock_mem_cgroup(memcg, flags);
2312 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2313 move_unlock_mem_cgroup(memcg, flags);
2319 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2321 struct page_cgroup *pc = lookup_page_cgroup(page);
2324 * It's guaranteed that pc->mem_cgroup never changes while
2325 * lock is held because a routine modifies pc->mem_cgroup
2326 * should take move_lock_mem_cgroup().
2328 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2331 void mem_cgroup_update_page_stat(struct page *page,
2332 enum mem_cgroup_page_stat_item idx, int val)
2334 struct mem_cgroup *memcg;
2335 struct page_cgroup *pc = lookup_page_cgroup(page);
2336 unsigned long uninitialized_var(flags);
2338 if (mem_cgroup_disabled())
2341 memcg = pc->mem_cgroup;
2342 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2346 case MEMCG_NR_FILE_MAPPED:
2347 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2353 this_cpu_add(memcg->stat->count[idx], val);
2357 * size of first charge trial. "32" comes from vmscan.c's magic value.
2358 * TODO: maybe necessary to use big numbers in big irons.
2360 #define CHARGE_BATCH 32U
2361 struct memcg_stock_pcp {
2362 struct mem_cgroup *cached; /* this never be root cgroup */
2363 unsigned int nr_pages;
2364 struct work_struct work;
2365 unsigned long flags;
2366 #define FLUSHING_CACHED_CHARGE 0
2368 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2369 static DEFINE_MUTEX(percpu_charge_mutex);
2372 * consume_stock: Try to consume stocked charge on this cpu.
2373 * @memcg: memcg to consume from.
2374 * @nr_pages: how many pages to charge.
2376 * The charges will only happen if @memcg matches the current cpu's memcg
2377 * stock, and at least @nr_pages are available in that stock. Failure to
2378 * service an allocation will refill the stock.
2380 * returns true if successful, false otherwise.
2382 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2384 struct memcg_stock_pcp *stock;
2387 if (nr_pages > CHARGE_BATCH)
2390 stock = &get_cpu_var(memcg_stock);
2391 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2392 stock->nr_pages -= nr_pages;
2393 else /* need to call res_counter_charge */
2395 put_cpu_var(memcg_stock);
2400 * Returns stocks cached in percpu to res_counter and reset cached information.
2402 static void drain_stock(struct memcg_stock_pcp *stock)
2404 struct mem_cgroup *old = stock->cached;
2406 if (stock->nr_pages) {
2407 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2409 res_counter_uncharge(&old->res, bytes);
2410 if (do_swap_account)
2411 res_counter_uncharge(&old->memsw, bytes);
2412 stock->nr_pages = 0;
2414 stock->cached = NULL;
2418 * This must be called under preempt disabled or must be called by
2419 * a thread which is pinned to local cpu.
2421 static void drain_local_stock(struct work_struct *dummy)
2423 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2425 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2428 static void __init memcg_stock_init(void)
2432 for_each_possible_cpu(cpu) {
2433 struct memcg_stock_pcp *stock =
2434 &per_cpu(memcg_stock, cpu);
2435 INIT_WORK(&stock->work, drain_local_stock);
2440 * Cache charges(val) which is from res_counter, to local per_cpu area.
2441 * This will be consumed by consume_stock() function, later.
2443 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2445 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2447 if (stock->cached != memcg) { /* reset if necessary */
2449 stock->cached = memcg;
2451 stock->nr_pages += nr_pages;
2452 put_cpu_var(memcg_stock);
2456 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2457 * of the hierarchy under it. sync flag says whether we should block
2458 * until the work is done.
2460 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2464 /* Notify other cpus that system-wide "drain" is running */
2467 for_each_online_cpu(cpu) {
2468 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2469 struct mem_cgroup *memcg;
2471 memcg = stock->cached;
2472 if (!memcg || !stock->nr_pages)
2474 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2476 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2478 drain_local_stock(&stock->work);
2480 schedule_work_on(cpu, &stock->work);
2488 for_each_online_cpu(cpu) {
2489 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2490 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2491 flush_work(&stock->work);
2498 * Tries to drain stocked charges in other cpus. This function is asynchronous
2499 * and just put a work per cpu for draining localy on each cpu. Caller can
2500 * expects some charges will be back to res_counter later but cannot wait for
2503 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2506 * If someone calls draining, avoid adding more kworker runs.
2508 if (!mutex_trylock(&percpu_charge_mutex))
2510 drain_all_stock(root_memcg, false);
2511 mutex_unlock(&percpu_charge_mutex);
2514 /* This is a synchronous drain interface. */
2515 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2517 /* called when force_empty is called */
2518 mutex_lock(&percpu_charge_mutex);
2519 drain_all_stock(root_memcg, true);
2520 mutex_unlock(&percpu_charge_mutex);
2524 * This function drains percpu counter value from DEAD cpu and
2525 * move it to local cpu. Note that this function can be preempted.
2527 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2531 spin_lock(&memcg->pcp_counter_lock);
2532 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2533 long x = per_cpu(memcg->stat->count[i], cpu);
2535 per_cpu(memcg->stat->count[i], cpu) = 0;
2536 memcg->nocpu_base.count[i] += x;
2538 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2539 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2541 per_cpu(memcg->stat->events[i], cpu) = 0;
2542 memcg->nocpu_base.events[i] += x;
2544 spin_unlock(&memcg->pcp_counter_lock);
2547 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2548 unsigned long action,
2551 int cpu = (unsigned long)hcpu;
2552 struct memcg_stock_pcp *stock;
2553 struct mem_cgroup *iter;
2555 if (action == CPU_ONLINE)
2558 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2561 for_each_mem_cgroup(iter)
2562 mem_cgroup_drain_pcp_counter(iter, cpu);
2564 stock = &per_cpu(memcg_stock, cpu);
2570 /* See __mem_cgroup_try_charge() for details */
2572 CHARGE_OK, /* success */
2573 CHARGE_RETRY, /* need to retry but retry is not bad */
2574 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2575 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2576 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2579 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2580 unsigned int nr_pages, unsigned int min_pages,
2583 unsigned long csize = nr_pages * PAGE_SIZE;
2584 struct mem_cgroup *mem_over_limit;
2585 struct res_counter *fail_res;
2586 unsigned long flags = 0;
2589 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2592 if (!do_swap_account)
2594 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2598 res_counter_uncharge(&memcg->res, csize);
2599 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2600 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2602 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2604 * Never reclaim on behalf of optional batching, retry with a
2605 * single page instead.
2607 if (nr_pages > min_pages)
2608 return CHARGE_RETRY;
2610 if (!(gfp_mask & __GFP_WAIT))
2611 return CHARGE_WOULDBLOCK;
2613 if (gfp_mask & __GFP_NORETRY)
2614 return CHARGE_NOMEM;
2616 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2617 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2618 return CHARGE_RETRY;
2620 * Even though the limit is exceeded at this point, reclaim
2621 * may have been able to free some pages. Retry the charge
2622 * before killing the task.
2624 * Only for regular pages, though: huge pages are rather
2625 * unlikely to succeed so close to the limit, and we fall back
2626 * to regular pages anyway in case of failure.
2628 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2629 return CHARGE_RETRY;
2632 * At task move, charge accounts can be doubly counted. So, it's
2633 * better to wait until the end of task_move if something is going on.
2635 if (mem_cgroup_wait_acct_move(mem_over_limit))
2636 return CHARGE_RETRY;
2638 /* If we don't need to call oom-killer at el, return immediately */
2640 return CHARGE_NOMEM;
2642 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2643 return CHARGE_OOM_DIE;
2645 return CHARGE_RETRY;
2649 * __mem_cgroup_try_charge() does
2650 * 1. detect memcg to be charged against from passed *mm and *ptr,
2651 * 2. update res_counter
2652 * 3. call memory reclaim if necessary.
2654 * In some special case, if the task is fatal, fatal_signal_pending() or
2655 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2656 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2657 * as possible without any hazards. 2: all pages should have a valid
2658 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2659 * pointer, that is treated as a charge to root_mem_cgroup.
2661 * So __mem_cgroup_try_charge() will return
2662 * 0 ... on success, filling *ptr with a valid memcg pointer.
2663 * -ENOMEM ... charge failure because of resource limits.
2664 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2666 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2667 * the oom-killer can be invoked.
2669 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2671 unsigned int nr_pages,
2672 struct mem_cgroup **ptr,
2675 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2676 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2677 struct mem_cgroup *memcg = NULL;
2681 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2682 * in system level. So, allow to go ahead dying process in addition to
2685 if (unlikely(test_thread_flag(TIF_MEMDIE)
2686 || fatal_signal_pending(current)))
2690 * We always charge the cgroup the mm_struct belongs to.
2691 * The mm_struct's mem_cgroup changes on task migration if the
2692 * thread group leader migrates. It's possible that mm is not
2693 * set, if so charge the root memcg (happens for pagecache usage).
2696 *ptr = root_mem_cgroup;
2698 if (*ptr) { /* css should be a valid one */
2700 if (mem_cgroup_is_root(memcg))
2702 if (consume_stock(memcg, nr_pages))
2704 css_get(&memcg->css);
2706 struct task_struct *p;
2709 p = rcu_dereference(mm->owner);
2711 * Because we don't have task_lock(), "p" can exit.
2712 * In that case, "memcg" can point to root or p can be NULL with
2713 * race with swapoff. Then, we have small risk of mis-accouning.
2714 * But such kind of mis-account by race always happens because
2715 * we don't have cgroup_mutex(). It's overkill and we allo that
2717 * (*) swapoff at el will charge against mm-struct not against
2718 * task-struct. So, mm->owner can be NULL.
2720 memcg = mem_cgroup_from_task(p);
2722 memcg = root_mem_cgroup;
2723 if (mem_cgroup_is_root(memcg)) {
2727 if (consume_stock(memcg, nr_pages)) {
2729 * It seems dagerous to access memcg without css_get().
2730 * But considering how consume_stok works, it's not
2731 * necessary. If consume_stock success, some charges
2732 * from this memcg are cached on this cpu. So, we
2733 * don't need to call css_get()/css_tryget() before
2734 * calling consume_stock().
2739 /* after here, we may be blocked. we need to get refcnt */
2740 if (!css_tryget(&memcg->css)) {
2750 /* If killed, bypass charge */
2751 if (fatal_signal_pending(current)) {
2752 css_put(&memcg->css);
2757 if (oom && !nr_oom_retries) {
2759 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2762 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2767 case CHARGE_RETRY: /* not in OOM situation but retry */
2769 css_put(&memcg->css);
2772 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2773 css_put(&memcg->css);
2775 case CHARGE_NOMEM: /* OOM routine works */
2777 css_put(&memcg->css);
2780 /* If oom, we never return -ENOMEM */
2783 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2784 css_put(&memcg->css);
2787 } while (ret != CHARGE_OK);
2789 if (batch > nr_pages)
2790 refill_stock(memcg, batch - nr_pages);
2791 css_put(&memcg->css);
2799 *ptr = root_mem_cgroup;
2804 * Somemtimes we have to undo a charge we got by try_charge().
2805 * This function is for that and do uncharge, put css's refcnt.
2806 * gotten by try_charge().
2808 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2809 unsigned int nr_pages)
2811 if (!mem_cgroup_is_root(memcg)) {
2812 unsigned long bytes = nr_pages * PAGE_SIZE;
2814 res_counter_uncharge(&memcg->res, bytes);
2815 if (do_swap_account)
2816 res_counter_uncharge(&memcg->memsw, bytes);
2821 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2822 * This is useful when moving usage to parent cgroup.
2824 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2825 unsigned int nr_pages)
2827 unsigned long bytes = nr_pages * PAGE_SIZE;
2829 if (mem_cgroup_is_root(memcg))
2832 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2833 if (do_swap_account)
2834 res_counter_uncharge_until(&memcg->memsw,
2835 memcg->memsw.parent, bytes);
2839 * A helper function to get mem_cgroup from ID. must be called under
2840 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2841 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2842 * called against removed memcg.)
2844 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2846 struct cgroup_subsys_state *css;
2848 /* ID 0 is unused ID */
2851 css = css_lookup(&mem_cgroup_subsys, id);
2854 return mem_cgroup_from_css(css);
2857 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2859 struct mem_cgroup *memcg = NULL;
2860 struct page_cgroup *pc;
2864 VM_BUG_ON(!PageLocked(page));
2866 pc = lookup_page_cgroup(page);
2867 lock_page_cgroup(pc);
2868 if (PageCgroupUsed(pc)) {
2869 memcg = pc->mem_cgroup;
2870 if (memcg && !css_tryget(&memcg->css))
2872 } else if (PageSwapCache(page)) {
2873 ent.val = page_private(page);
2874 id = lookup_swap_cgroup_id(ent);
2876 memcg = mem_cgroup_lookup(id);
2877 if (memcg && !css_tryget(&memcg->css))
2881 unlock_page_cgroup(pc);
2885 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2887 unsigned int nr_pages,
2888 enum charge_type ctype,
2891 struct page_cgroup *pc = lookup_page_cgroup(page);
2892 struct zone *uninitialized_var(zone);
2893 struct lruvec *lruvec;
2894 bool was_on_lru = false;
2897 lock_page_cgroup(pc);
2898 VM_BUG_ON(PageCgroupUsed(pc));
2900 * we don't need page_cgroup_lock about tail pages, becase they are not
2901 * accessed by any other context at this point.
2905 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2906 * may already be on some other mem_cgroup's LRU. Take care of it.
2909 zone = page_zone(page);
2910 spin_lock_irq(&zone->lru_lock);
2911 if (PageLRU(page)) {
2912 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2914 del_page_from_lru_list(page, lruvec, page_lru(page));
2919 pc->mem_cgroup = memcg;
2921 * We access a page_cgroup asynchronously without lock_page_cgroup().
2922 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2923 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2924 * before USED bit, we need memory barrier here.
2925 * See mem_cgroup_add_lru_list(), etc.
2928 SetPageCgroupUsed(pc);
2932 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2933 VM_BUG_ON(PageLRU(page));
2935 add_page_to_lru_list(page, lruvec, page_lru(page));
2937 spin_unlock_irq(&zone->lru_lock);
2940 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2945 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2946 unlock_page_cgroup(pc);
2949 * "charge_statistics" updated event counter. Then, check it.
2950 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2951 * if they exceeds softlimit.
2953 memcg_check_events(memcg, page);
2956 static DEFINE_MUTEX(set_limit_mutex);
2958 #ifdef CONFIG_MEMCG_KMEM
2959 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2961 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2962 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2966 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2967 * in the memcg_cache_params struct.
2969 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2971 struct kmem_cache *cachep;
2973 VM_BUG_ON(p->is_root_cache);
2974 cachep = p->root_cache;
2975 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2978 #ifdef CONFIG_SLABINFO
2979 static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2982 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2983 struct memcg_cache_params *params;
2985 if (!memcg_can_account_kmem(memcg))
2988 print_slabinfo_header(m);
2990 mutex_lock(&memcg->slab_caches_mutex);
2991 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2992 cache_show(memcg_params_to_cache(params), m);
2993 mutex_unlock(&memcg->slab_caches_mutex);
2999 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3001 struct res_counter *fail_res;
3002 struct mem_cgroup *_memcg;
3006 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3011 * Conditions under which we can wait for the oom_killer. Those are
3012 * the same conditions tested by the core page allocator
3014 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
3017 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3020 if (ret == -EINTR) {
3022 * __mem_cgroup_try_charge() chosed to bypass to root due to
3023 * OOM kill or fatal signal. Since our only options are to
3024 * either fail the allocation or charge it to this cgroup, do
3025 * it as a temporary condition. But we can't fail. From a
3026 * kmem/slab perspective, the cache has already been selected,
3027 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3030 * This condition will only trigger if the task entered
3031 * memcg_charge_kmem in a sane state, but was OOM-killed during
3032 * __mem_cgroup_try_charge() above. Tasks that were already
3033 * dying when the allocation triggers should have been already
3034 * directed to the root cgroup in memcontrol.h
3036 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3037 if (do_swap_account)
3038 res_counter_charge_nofail(&memcg->memsw, size,
3042 res_counter_uncharge(&memcg->kmem, size);
3047 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3049 res_counter_uncharge(&memcg->res, size);
3050 if (do_swap_account)
3051 res_counter_uncharge(&memcg->memsw, size);
3054 if (res_counter_uncharge(&memcg->kmem, size))
3058 * Releases a reference taken in kmem_cgroup_css_offline in case
3059 * this last uncharge is racing with the offlining code or it is
3060 * outliving the memcg existence.
3062 * The memory barrier imposed by test&clear is paired with the
3063 * explicit one in memcg_kmem_mark_dead().
3065 if (memcg_kmem_test_and_clear_dead(memcg))
3066 css_put(&memcg->css);
3069 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3074 mutex_lock(&memcg->slab_caches_mutex);
3075 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3076 mutex_unlock(&memcg->slab_caches_mutex);
3080 * helper for acessing a memcg's index. It will be used as an index in the
3081 * child cache array in kmem_cache, and also to derive its name. This function
3082 * will return -1 when this is not a kmem-limited memcg.
3084 int memcg_cache_id(struct mem_cgroup *memcg)
3086 return memcg ? memcg->kmemcg_id : -1;
3090 * This ends up being protected by the set_limit mutex, during normal
3091 * operation, because that is its main call site.
3093 * But when we create a new cache, we can call this as well if its parent
3094 * is kmem-limited. That will have to hold set_limit_mutex as well.
3096 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3100 num = ida_simple_get(&kmem_limited_groups,
3101 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3105 * After this point, kmem_accounted (that we test atomically in
3106 * the beginning of this conditional), is no longer 0. This
3107 * guarantees only one process will set the following boolean
3108 * to true. We don't need test_and_set because we're protected
3109 * by the set_limit_mutex anyway.
3111 memcg_kmem_set_activated(memcg);
3113 ret = memcg_update_all_caches(num+1);
3115 ida_simple_remove(&kmem_limited_groups, num);
3116 memcg_kmem_clear_activated(memcg);
3120 memcg->kmemcg_id = num;
3121 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3122 mutex_init(&memcg->slab_caches_mutex);
3126 static size_t memcg_caches_array_size(int num_groups)
3129 if (num_groups <= 0)
3132 size = 2 * num_groups;
3133 if (size < MEMCG_CACHES_MIN_SIZE)
3134 size = MEMCG_CACHES_MIN_SIZE;
3135 else if (size > MEMCG_CACHES_MAX_SIZE)
3136 size = MEMCG_CACHES_MAX_SIZE;
3142 * We should update the current array size iff all caches updates succeed. This
3143 * can only be done from the slab side. The slab mutex needs to be held when
3146 void memcg_update_array_size(int num)
3148 if (num > memcg_limited_groups_array_size)
3149 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3152 static void kmem_cache_destroy_work_func(struct work_struct *w);
3154 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3156 struct memcg_cache_params *cur_params = s->memcg_params;
3158 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3160 if (num_groups > memcg_limited_groups_array_size) {
3162 ssize_t size = memcg_caches_array_size(num_groups);
3164 size *= sizeof(void *);
3165 size += sizeof(struct memcg_cache_params);
3167 s->memcg_params = kzalloc(size, GFP_KERNEL);
3168 if (!s->memcg_params) {
3169 s->memcg_params = cur_params;
3173 s->memcg_params->is_root_cache = true;
3176 * There is the chance it will be bigger than
3177 * memcg_limited_groups_array_size, if we failed an allocation
3178 * in a cache, in which case all caches updated before it, will
3179 * have a bigger array.
3181 * But if that is the case, the data after
3182 * memcg_limited_groups_array_size is certainly unused
3184 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3185 if (!cur_params->memcg_caches[i])
3187 s->memcg_params->memcg_caches[i] =
3188 cur_params->memcg_caches[i];
3192 * Ideally, we would wait until all caches succeed, and only
3193 * then free the old one. But this is not worth the extra
3194 * pointer per-cache we'd have to have for this.
3196 * It is not a big deal if some caches are left with a size
3197 * bigger than the others. And all updates will reset this
3205 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3206 struct kmem_cache *root_cache)
3208 size_t size = sizeof(struct memcg_cache_params);
3210 if (!memcg_kmem_enabled())
3214 size += memcg_limited_groups_array_size * sizeof(void *);
3216 s->memcg_params = kzalloc(size, GFP_KERNEL);
3217 if (!s->memcg_params)
3220 INIT_WORK(&s->memcg_params->destroy,
3221 kmem_cache_destroy_work_func);
3223 s->memcg_params->memcg = memcg;
3224 s->memcg_params->root_cache = root_cache;
3226 s->memcg_params->is_root_cache = true;
3231 void memcg_release_cache(struct kmem_cache *s)
3233 struct kmem_cache *root;
3234 struct mem_cgroup *memcg;
3238 * This happens, for instance, when a root cache goes away before we
3241 if (!s->memcg_params)
3244 if (s->memcg_params->is_root_cache)
3247 memcg = s->memcg_params->memcg;
3248 id = memcg_cache_id(memcg);
3250 root = s->memcg_params->root_cache;
3251 root->memcg_params->memcg_caches[id] = NULL;
3253 mutex_lock(&memcg->slab_caches_mutex);
3254 list_del(&s->memcg_params->list);
3255 mutex_unlock(&memcg->slab_caches_mutex);
3257 css_put(&memcg->css);
3259 kfree(s->memcg_params);
3263 * During the creation a new cache, we need to disable our accounting mechanism
3264 * altogether. This is true even if we are not creating, but rather just
3265 * enqueing new caches to be created.
3267 * This is because that process will trigger allocations; some visible, like
3268 * explicit kmallocs to auxiliary data structures, name strings and internal
3269 * cache structures; some well concealed, like INIT_WORK() that can allocate
3270 * objects during debug.
3272 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3273 * to it. This may not be a bounded recursion: since the first cache creation
3274 * failed to complete (waiting on the allocation), we'll just try to create the
3275 * cache again, failing at the same point.
3277 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3278 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3279 * inside the following two functions.
3281 static inline void memcg_stop_kmem_account(void)
3283 VM_BUG_ON(!current->mm);
3284 current->memcg_kmem_skip_account++;
3287 static inline void memcg_resume_kmem_account(void)
3289 VM_BUG_ON(!current->mm);
3290 current->memcg_kmem_skip_account--;
3293 static void kmem_cache_destroy_work_func(struct work_struct *w)
3295 struct kmem_cache *cachep;
3296 struct memcg_cache_params *p;
3298 p = container_of(w, struct memcg_cache_params, destroy);
3300 cachep = memcg_params_to_cache(p);
3303 * If we get down to 0 after shrink, we could delete right away.
3304 * However, memcg_release_pages() already puts us back in the workqueue
3305 * in that case. If we proceed deleting, we'll get a dangling
3306 * reference, and removing the object from the workqueue in that case
3307 * is unnecessary complication. We are not a fast path.
3309 * Note that this case is fundamentally different from racing with
3310 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3311 * kmem_cache_shrink, not only we would be reinserting a dead cache
3312 * into the queue, but doing so from inside the worker racing to
3315 * So if we aren't down to zero, we'll just schedule a worker and try
3318 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3319 kmem_cache_shrink(cachep);
3320 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3323 kmem_cache_destroy(cachep);
3326 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3328 if (!cachep->memcg_params->dead)
3332 * There are many ways in which we can get here.
3334 * We can get to a memory-pressure situation while the delayed work is
3335 * still pending to run. The vmscan shrinkers can then release all
3336 * cache memory and get us to destruction. If this is the case, we'll
3337 * be executed twice, which is a bug (the second time will execute over
3338 * bogus data). In this case, cancelling the work should be fine.
3340 * But we can also get here from the worker itself, if
3341 * kmem_cache_shrink is enough to shake all the remaining objects and
3342 * get the page count to 0. In this case, we'll deadlock if we try to
3343 * cancel the work (the worker runs with an internal lock held, which
3344 * is the same lock we would hold for cancel_work_sync().)
3346 * Since we can't possibly know who got us here, just refrain from
3347 * running if there is already work pending
3349 if (work_pending(&cachep->memcg_params->destroy))
3352 * We have to defer the actual destroying to a workqueue, because
3353 * we might currently be in a context that cannot sleep.
3355 schedule_work(&cachep->memcg_params->destroy);
3359 * This lock protects updaters, not readers. We want readers to be as fast as
3360 * they can, and they will either see NULL or a valid cache value. Our model
3361 * allow them to see NULL, in which case the root memcg will be selected.
3363 * We need this lock because multiple allocations to the same cache from a non
3364 * will span more than one worker. Only one of them can create the cache.
3366 static DEFINE_MUTEX(memcg_cache_mutex);
3369 * Called with memcg_cache_mutex held
3371 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3372 struct kmem_cache *s)
3374 struct kmem_cache *new;
3375 static char *tmp_name = NULL;
3377 lockdep_assert_held(&memcg_cache_mutex);
3380 * kmem_cache_create_memcg duplicates the given name and
3381 * cgroup_name for this name requires RCU context.
3382 * This static temporary buffer is used to prevent from
3383 * pointless shortliving allocation.
3386 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3392 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3393 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3396 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3397 (s->flags & ~SLAB_PANIC), s->ctor, s);
3400 new->allocflags |= __GFP_KMEMCG;
3405 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3406 struct kmem_cache *cachep)
3408 struct kmem_cache *new_cachep;
3411 BUG_ON(!memcg_can_account_kmem(memcg));
3413 idx = memcg_cache_id(memcg);
3415 mutex_lock(&memcg_cache_mutex);
3416 new_cachep = cachep->memcg_params->memcg_caches[idx];
3418 css_put(&memcg->css);
3422 new_cachep = kmem_cache_dup(memcg, cachep);
3423 if (new_cachep == NULL) {
3424 new_cachep = cachep;
3425 css_put(&memcg->css);
3429 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3431 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3433 * the readers won't lock, make sure everybody sees the updated value,
3434 * so they won't put stuff in the queue again for no reason
3438 mutex_unlock(&memcg_cache_mutex);
3442 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3444 struct kmem_cache *c;
3447 if (!s->memcg_params)
3449 if (!s->memcg_params->is_root_cache)
3453 * If the cache is being destroyed, we trust that there is no one else
3454 * requesting objects from it. Even if there are, the sanity checks in
3455 * kmem_cache_destroy should caught this ill-case.
3457 * Still, we don't want anyone else freeing memcg_caches under our
3458 * noses, which can happen if a new memcg comes to life. As usual,
3459 * we'll take the set_limit_mutex to protect ourselves against this.
3461 mutex_lock(&set_limit_mutex);
3462 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3463 c = s->memcg_params->memcg_caches[i];
3468 * We will now manually delete the caches, so to avoid races
3469 * we need to cancel all pending destruction workers and
3470 * proceed with destruction ourselves.
3472 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3473 * and that could spawn the workers again: it is likely that
3474 * the cache still have active pages until this very moment.
3475 * This would lead us back to mem_cgroup_destroy_cache.
3477 * But that will not execute at all if the "dead" flag is not
3478 * set, so flip it down to guarantee we are in control.
3480 c->memcg_params->dead = false;
3481 cancel_work_sync(&c->memcg_params->destroy);
3482 kmem_cache_destroy(c);
3484 mutex_unlock(&set_limit_mutex);
3487 struct create_work {
3488 struct mem_cgroup *memcg;
3489 struct kmem_cache *cachep;
3490 struct work_struct work;
3493 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3495 struct kmem_cache *cachep;
3496 struct memcg_cache_params *params;
3498 if (!memcg_kmem_is_active(memcg))
3501 mutex_lock(&memcg->slab_caches_mutex);
3502 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3503 cachep = memcg_params_to_cache(params);
3504 cachep->memcg_params->dead = true;
3505 schedule_work(&cachep->memcg_params->destroy);
3507 mutex_unlock(&memcg->slab_caches_mutex);
3510 static void memcg_create_cache_work_func(struct work_struct *w)
3512 struct create_work *cw;
3514 cw = container_of(w, struct create_work, work);
3515 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3520 * Enqueue the creation of a per-memcg kmem_cache.
3522 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3523 struct kmem_cache *cachep)
3525 struct create_work *cw;
3527 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3529 css_put(&memcg->css);
3534 cw->cachep = cachep;
3536 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3537 schedule_work(&cw->work);
3540 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3541 struct kmem_cache *cachep)
3544 * We need to stop accounting when we kmalloc, because if the
3545 * corresponding kmalloc cache is not yet created, the first allocation
3546 * in __memcg_create_cache_enqueue will recurse.
3548 * However, it is better to enclose the whole function. Depending on
3549 * the debugging options enabled, INIT_WORK(), for instance, can
3550 * trigger an allocation. This too, will make us recurse. Because at
3551 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3552 * the safest choice is to do it like this, wrapping the whole function.
3554 memcg_stop_kmem_account();
3555 __memcg_create_cache_enqueue(memcg, cachep);
3556 memcg_resume_kmem_account();
3559 * Return the kmem_cache we're supposed to use for a slab allocation.
3560 * We try to use the current memcg's version of the cache.
3562 * If the cache does not exist yet, if we are the first user of it,
3563 * we either create it immediately, if possible, or create it asynchronously
3565 * In the latter case, we will let the current allocation go through with
3566 * the original cache.
3568 * Can't be called in interrupt context or from kernel threads.
3569 * This function needs to be called with rcu_read_lock() held.
3571 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3574 struct mem_cgroup *memcg;
3577 VM_BUG_ON(!cachep->memcg_params);
3578 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3580 if (!current->mm || current->memcg_kmem_skip_account)
3584 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3586 if (!memcg_can_account_kmem(memcg))
3589 idx = memcg_cache_id(memcg);
3592 * barrier to mare sure we're always seeing the up to date value. The
3593 * code updating memcg_caches will issue a write barrier to match this.
3595 read_barrier_depends();
3596 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3597 cachep = cachep->memcg_params->memcg_caches[idx];
3601 /* The corresponding put will be done in the workqueue. */
3602 if (!css_tryget(&memcg->css))
3607 * If we are in a safe context (can wait, and not in interrupt
3608 * context), we could be be predictable and return right away.
3609 * This would guarantee that the allocation being performed
3610 * already belongs in the new cache.
3612 * However, there are some clashes that can arrive from locking.
3613 * For instance, because we acquire the slab_mutex while doing
3614 * kmem_cache_dup, this means no further allocation could happen
3615 * with the slab_mutex held.
3617 * Also, because cache creation issue get_online_cpus(), this
3618 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3619 * that ends up reversed during cpu hotplug. (cpuset allocates
3620 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3621 * better to defer everything.
3623 memcg_create_cache_enqueue(memcg, cachep);
3629 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3632 * We need to verify if the allocation against current->mm->owner's memcg is
3633 * possible for the given order. But the page is not allocated yet, so we'll
3634 * need a further commit step to do the final arrangements.
3636 * It is possible for the task to switch cgroups in this mean time, so at
3637 * commit time, we can't rely on task conversion any longer. We'll then use
3638 * the handle argument to return to the caller which cgroup we should commit
3639 * against. We could also return the memcg directly and avoid the pointer
3640 * passing, but a boolean return value gives better semantics considering
3641 * the compiled-out case as well.
3643 * Returning true means the allocation is possible.
3646 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3648 struct mem_cgroup *memcg;
3654 * Disabling accounting is only relevant for some specific memcg
3655 * internal allocations. Therefore we would initially not have such
3656 * check here, since direct calls to the page allocator that are marked
3657 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3658 * concerned with cache allocations, and by having this test at
3659 * memcg_kmem_get_cache, we are already able to relay the allocation to
3660 * the root cache and bypass the memcg cache altogether.
3662 * There is one exception, though: the SLUB allocator does not create
3663 * large order caches, but rather service large kmallocs directly from
3664 * the page allocator. Therefore, the following sequence when backed by
3665 * the SLUB allocator:
3667 * memcg_stop_kmem_account();
3668 * kmalloc(<large_number>)
3669 * memcg_resume_kmem_account();
3671 * would effectively ignore the fact that we should skip accounting,
3672 * since it will drive us directly to this function without passing
3673 * through the cache selector memcg_kmem_get_cache. Such large
3674 * allocations are extremely rare but can happen, for instance, for the
3675 * cache arrays. We bring this test here.
3677 if (!current->mm || current->memcg_kmem_skip_account)
3680 memcg = try_get_mem_cgroup_from_mm(current->mm);
3683 * very rare case described in mem_cgroup_from_task. Unfortunately there
3684 * isn't much we can do without complicating this too much, and it would
3685 * be gfp-dependent anyway. Just let it go
3687 if (unlikely(!memcg))
3690 if (!memcg_can_account_kmem(memcg)) {
3691 css_put(&memcg->css);
3695 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3699 css_put(&memcg->css);
3703 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3706 struct page_cgroup *pc;
3708 VM_BUG_ON(mem_cgroup_is_root(memcg));
3710 /* The page allocation failed. Revert */
3712 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3716 pc = lookup_page_cgroup(page);
3717 lock_page_cgroup(pc);
3718 pc->mem_cgroup = memcg;
3719 SetPageCgroupUsed(pc);
3720 unlock_page_cgroup(pc);
3723 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3725 struct mem_cgroup *memcg = NULL;
3726 struct page_cgroup *pc;
3729 pc = lookup_page_cgroup(page);
3731 * Fast unlocked return. Theoretically might have changed, have to
3732 * check again after locking.
3734 if (!PageCgroupUsed(pc))
3737 lock_page_cgroup(pc);
3738 if (PageCgroupUsed(pc)) {
3739 memcg = pc->mem_cgroup;
3740 ClearPageCgroupUsed(pc);
3742 unlock_page_cgroup(pc);
3745 * We trust that only if there is a memcg associated with the page, it
3746 * is a valid allocation
3751 VM_BUG_ON(mem_cgroup_is_root(memcg));
3752 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3755 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3758 #endif /* CONFIG_MEMCG_KMEM */
3760 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3762 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3764 * Because tail pages are not marked as "used", set it. We're under
3765 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3766 * charge/uncharge will be never happen and move_account() is done under
3767 * compound_lock(), so we don't have to take care of races.
3769 void mem_cgroup_split_huge_fixup(struct page *head)
3771 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3772 struct page_cgroup *pc;
3773 struct mem_cgroup *memcg;
3776 if (mem_cgroup_disabled())
3779 memcg = head_pc->mem_cgroup;
3780 for (i = 1; i < HPAGE_PMD_NR; i++) {
3782 pc->mem_cgroup = memcg;
3783 smp_wmb();/* see __commit_charge() */
3784 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3786 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3789 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3792 * mem_cgroup_move_account - move account of the page
3794 * @nr_pages: number of regular pages (>1 for huge pages)
3795 * @pc: page_cgroup of the page.
3796 * @from: mem_cgroup which the page is moved from.
3797 * @to: mem_cgroup which the page is moved to. @from != @to.
3799 * The caller must confirm following.
3800 * - page is not on LRU (isolate_page() is useful.)
3801 * - compound_lock is held when nr_pages > 1
3803 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3806 static int mem_cgroup_move_account(struct page *page,
3807 unsigned int nr_pages,
3808 struct page_cgroup *pc,
3809 struct mem_cgroup *from,
3810 struct mem_cgroup *to)
3812 unsigned long flags;
3814 bool anon = PageAnon(page);
3816 VM_BUG_ON(from == to);
3817 VM_BUG_ON(PageLRU(page));
3819 * The page is isolated from LRU. So, collapse function
3820 * will not handle this page. But page splitting can happen.
3821 * Do this check under compound_page_lock(). The caller should
3825 if (nr_pages > 1 && !PageTransHuge(page))
3828 lock_page_cgroup(pc);
3831 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3834 move_lock_mem_cgroup(from, &flags);
3836 if (!anon && page_mapped(page)) {
3837 /* Update mapped_file data for mem_cgroup */
3839 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3840 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3843 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3845 /* caller should have done css_get */
3846 pc->mem_cgroup = to;
3847 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3848 move_unlock_mem_cgroup(from, &flags);
3851 unlock_page_cgroup(pc);
3855 memcg_check_events(to, page);
3856 memcg_check_events(from, page);
3862 * mem_cgroup_move_parent - moves page to the parent group
3863 * @page: the page to move
3864 * @pc: page_cgroup of the page
3865 * @child: page's cgroup
3867 * move charges to its parent or the root cgroup if the group has no
3868 * parent (aka use_hierarchy==0).
3869 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3870 * mem_cgroup_move_account fails) the failure is always temporary and
3871 * it signals a race with a page removal/uncharge or migration. In the
3872 * first case the page is on the way out and it will vanish from the LRU
3873 * on the next attempt and the call should be retried later.
3874 * Isolation from the LRU fails only if page has been isolated from
3875 * the LRU since we looked at it and that usually means either global
3876 * reclaim or migration going on. The page will either get back to the
3878 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3879 * (!PageCgroupUsed) or moved to a different group. The page will
3880 * disappear in the next attempt.
3882 static int mem_cgroup_move_parent(struct page *page,
3883 struct page_cgroup *pc,
3884 struct mem_cgroup *child)
3886 struct mem_cgroup *parent;
3887 unsigned int nr_pages;
3888 unsigned long uninitialized_var(flags);
3891 VM_BUG_ON(mem_cgroup_is_root(child));
3894 if (!get_page_unless_zero(page))
3896 if (isolate_lru_page(page))
3899 nr_pages = hpage_nr_pages(page);
3901 parent = parent_mem_cgroup(child);
3903 * If no parent, move charges to root cgroup.
3906 parent = root_mem_cgroup;
3909 VM_BUG_ON(!PageTransHuge(page));
3910 flags = compound_lock_irqsave(page);
3913 ret = mem_cgroup_move_account(page, nr_pages,
3916 __mem_cgroup_cancel_local_charge(child, nr_pages);
3919 compound_unlock_irqrestore(page, flags);
3920 putback_lru_page(page);
3928 * Charge the memory controller for page usage.
3930 * 0 if the charge was successful
3931 * < 0 if the cgroup is over its limit
3933 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3934 gfp_t gfp_mask, enum charge_type ctype)
3936 struct mem_cgroup *memcg = NULL;
3937 unsigned int nr_pages = 1;
3941 if (PageTransHuge(page)) {
3942 nr_pages <<= compound_order(page);
3943 VM_BUG_ON(!PageTransHuge(page));
3945 * Never OOM-kill a process for a huge page. The
3946 * fault handler will fall back to regular pages.
3951 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3954 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3958 int mem_cgroup_newpage_charge(struct page *page,
3959 struct mm_struct *mm, gfp_t gfp_mask)
3961 if (mem_cgroup_disabled())
3963 VM_BUG_ON(page_mapped(page));
3964 VM_BUG_ON(page->mapping && !PageAnon(page));
3966 return mem_cgroup_charge_common(page, mm, gfp_mask,
3967 MEM_CGROUP_CHARGE_TYPE_ANON);
3971 * While swap-in, try_charge -> commit or cancel, the page is locked.
3972 * And when try_charge() successfully returns, one refcnt to memcg without
3973 * struct page_cgroup is acquired. This refcnt will be consumed by
3974 * "commit()" or removed by "cancel()"
3976 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3979 struct mem_cgroup **memcgp)
3981 struct mem_cgroup *memcg;
3982 struct page_cgroup *pc;
3985 pc = lookup_page_cgroup(page);
3987 * Every swap fault against a single page tries to charge the
3988 * page, bail as early as possible. shmem_unuse() encounters
3989 * already charged pages, too. The USED bit is protected by
3990 * the page lock, which serializes swap cache removal, which
3991 * in turn serializes uncharging.
3993 if (PageCgroupUsed(pc))
3995 if (!do_swap_account)
3997 memcg = try_get_mem_cgroup_from_page(page);
4001 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4002 css_put(&memcg->css);
4007 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4013 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4014 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4017 if (mem_cgroup_disabled())
4020 * A racing thread's fault, or swapoff, may have already
4021 * updated the pte, and even removed page from swap cache: in
4022 * those cases unuse_pte()'s pte_same() test will fail; but
4023 * there's also a KSM case which does need to charge the page.
4025 if (!PageSwapCache(page)) {
4028 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4033 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4036 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4038 if (mem_cgroup_disabled())
4042 __mem_cgroup_cancel_charge(memcg, 1);
4046 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4047 enum charge_type ctype)
4049 if (mem_cgroup_disabled())
4054 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4056 * Now swap is on-memory. This means this page may be
4057 * counted both as mem and swap....double count.
4058 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4059 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4060 * may call delete_from_swap_cache() before reach here.
4062 if (do_swap_account && PageSwapCache(page)) {
4063 swp_entry_t ent = {.val = page_private(page)};
4064 mem_cgroup_uncharge_swap(ent);
4068 void mem_cgroup_commit_charge_swapin(struct page *page,
4069 struct mem_cgroup *memcg)
4071 __mem_cgroup_commit_charge_swapin(page, memcg,
4072 MEM_CGROUP_CHARGE_TYPE_ANON);
4075 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4078 struct mem_cgroup *memcg = NULL;
4079 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4082 if (mem_cgroup_disabled())
4084 if (PageCompound(page))
4087 if (!PageSwapCache(page))
4088 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4089 else { /* page is swapcache/shmem */
4090 ret = __mem_cgroup_try_charge_swapin(mm, page,
4093 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4098 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4099 unsigned int nr_pages,
4100 const enum charge_type ctype)
4102 struct memcg_batch_info *batch = NULL;
4103 bool uncharge_memsw = true;
4105 /* If swapout, usage of swap doesn't decrease */
4106 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4107 uncharge_memsw = false;
4109 batch = ¤t->memcg_batch;
4111 * In usual, we do css_get() when we remember memcg pointer.
4112 * But in this case, we keep res->usage until end of a series of
4113 * uncharges. Then, it's ok to ignore memcg's refcnt.
4116 batch->memcg = memcg;
4118 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4119 * In those cases, all pages freed continuously can be expected to be in
4120 * the same cgroup and we have chance to coalesce uncharges.
4121 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4122 * because we want to do uncharge as soon as possible.
4125 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4126 goto direct_uncharge;
4129 goto direct_uncharge;
4132 * In typical case, batch->memcg == mem. This means we can
4133 * merge a series of uncharges to an uncharge of res_counter.
4134 * If not, we uncharge res_counter ony by one.
4136 if (batch->memcg != memcg)
4137 goto direct_uncharge;
4138 /* remember freed charge and uncharge it later */
4141 batch->memsw_nr_pages++;
4144 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4146 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4147 if (unlikely(batch->memcg != memcg))
4148 memcg_oom_recover(memcg);
4152 * uncharge if !page_mapped(page)
4154 static struct mem_cgroup *
4155 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4158 struct mem_cgroup *memcg = NULL;
4159 unsigned int nr_pages = 1;
4160 struct page_cgroup *pc;
4163 if (mem_cgroup_disabled())
4166 if (PageTransHuge(page)) {
4167 nr_pages <<= compound_order(page);
4168 VM_BUG_ON(!PageTransHuge(page));
4171 * Check if our page_cgroup is valid
4173 pc = lookup_page_cgroup(page);
4174 if (unlikely(!PageCgroupUsed(pc)))
4177 lock_page_cgroup(pc);
4179 memcg = pc->mem_cgroup;
4181 if (!PageCgroupUsed(pc))
4184 anon = PageAnon(page);
4187 case MEM_CGROUP_CHARGE_TYPE_ANON:
4189 * Generally PageAnon tells if it's the anon statistics to be
4190 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4191 * used before page reached the stage of being marked PageAnon.
4195 case MEM_CGROUP_CHARGE_TYPE_DROP:
4196 /* See mem_cgroup_prepare_migration() */
4197 if (page_mapped(page))
4200 * Pages under migration may not be uncharged. But
4201 * end_migration() /must/ be the one uncharging the
4202 * unused post-migration page and so it has to call
4203 * here with the migration bit still set. See the
4204 * res_counter handling below.
4206 if (!end_migration && PageCgroupMigration(pc))
4209 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4210 if (!PageAnon(page)) { /* Shared memory */
4211 if (page->mapping && !page_is_file_cache(page))
4213 } else if (page_mapped(page)) /* Anon */
4220 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4222 ClearPageCgroupUsed(pc);
4224 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4225 * freed from LRU. This is safe because uncharged page is expected not
4226 * to be reused (freed soon). Exception is SwapCache, it's handled by
4227 * special functions.
4230 unlock_page_cgroup(pc);
4232 * even after unlock, we have memcg->res.usage here and this memcg
4233 * will never be freed, so it's safe to call css_get().
4235 memcg_check_events(memcg, page);
4236 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4237 mem_cgroup_swap_statistics(memcg, true);
4238 css_get(&memcg->css);
4241 * Migration does not charge the res_counter for the
4242 * replacement page, so leave it alone when phasing out the
4243 * page that is unused after the migration.
4245 if (!end_migration && !mem_cgroup_is_root(memcg))
4246 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4251 unlock_page_cgroup(pc);
4255 void mem_cgroup_uncharge_page(struct page *page)
4258 if (page_mapped(page))
4260 VM_BUG_ON(page->mapping && !PageAnon(page));
4262 * If the page is in swap cache, uncharge should be deferred
4263 * to the swap path, which also properly accounts swap usage
4264 * and handles memcg lifetime.
4266 * Note that this check is not stable and reclaim may add the
4267 * page to swap cache at any time after this. However, if the
4268 * page is not in swap cache by the time page->mapcount hits
4269 * 0, there won't be any page table references to the swap
4270 * slot, and reclaim will free it and not actually write the
4273 if (PageSwapCache(page))
4275 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4278 void mem_cgroup_uncharge_cache_page(struct page *page)
4280 VM_BUG_ON(page_mapped(page));
4281 VM_BUG_ON(page->mapping);
4282 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4286 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4287 * In that cases, pages are freed continuously and we can expect pages
4288 * are in the same memcg. All these calls itself limits the number of
4289 * pages freed at once, then uncharge_start/end() is called properly.
4290 * This may be called prural(2) times in a context,
4293 void mem_cgroup_uncharge_start(void)
4295 current->memcg_batch.do_batch++;
4296 /* We can do nest. */
4297 if (current->memcg_batch.do_batch == 1) {
4298 current->memcg_batch.memcg = NULL;
4299 current->memcg_batch.nr_pages = 0;
4300 current->memcg_batch.memsw_nr_pages = 0;
4304 void mem_cgroup_uncharge_end(void)
4306 struct memcg_batch_info *batch = ¤t->memcg_batch;
4308 if (!batch->do_batch)
4312 if (batch->do_batch) /* If stacked, do nothing. */
4318 * This "batch->memcg" is valid without any css_get/put etc...
4319 * bacause we hide charges behind us.
4321 if (batch->nr_pages)
4322 res_counter_uncharge(&batch->memcg->res,
4323 batch->nr_pages * PAGE_SIZE);
4324 if (batch->memsw_nr_pages)
4325 res_counter_uncharge(&batch->memcg->memsw,
4326 batch->memsw_nr_pages * PAGE_SIZE);
4327 memcg_oom_recover(batch->memcg);
4328 /* forget this pointer (for sanity check) */
4329 batch->memcg = NULL;
4334 * called after __delete_from_swap_cache() and drop "page" account.
4335 * memcg information is recorded to swap_cgroup of "ent"
4338 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4340 struct mem_cgroup *memcg;
4341 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4343 if (!swapout) /* this was a swap cache but the swap is unused ! */
4344 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4346 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4349 * record memcg information, if swapout && memcg != NULL,
4350 * css_get() was called in uncharge().
4352 if (do_swap_account && swapout && memcg)
4353 swap_cgroup_record(ent, css_id(&memcg->css));
4357 #ifdef CONFIG_MEMCG_SWAP
4359 * called from swap_entry_free(). remove record in swap_cgroup and
4360 * uncharge "memsw" account.
4362 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4364 struct mem_cgroup *memcg;
4367 if (!do_swap_account)
4370 id = swap_cgroup_record(ent, 0);
4372 memcg = mem_cgroup_lookup(id);
4375 * We uncharge this because swap is freed.
4376 * This memcg can be obsolete one. We avoid calling css_tryget
4378 if (!mem_cgroup_is_root(memcg))
4379 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4380 mem_cgroup_swap_statistics(memcg, false);
4381 css_put(&memcg->css);
4387 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4388 * @entry: swap entry to be moved
4389 * @from: mem_cgroup which the entry is moved from
4390 * @to: mem_cgroup which the entry is moved to
4392 * It succeeds only when the swap_cgroup's record for this entry is the same
4393 * as the mem_cgroup's id of @from.
4395 * Returns 0 on success, -EINVAL on failure.
4397 * The caller must have charged to @to, IOW, called res_counter_charge() about
4398 * both res and memsw, and called css_get().
4400 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4401 struct mem_cgroup *from, struct mem_cgroup *to)
4403 unsigned short old_id, new_id;
4405 old_id = css_id(&from->css);
4406 new_id = css_id(&to->css);
4408 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4409 mem_cgroup_swap_statistics(from, false);
4410 mem_cgroup_swap_statistics(to, true);
4412 * This function is only called from task migration context now.
4413 * It postpones res_counter and refcount handling till the end
4414 * of task migration(mem_cgroup_clear_mc()) for performance
4415 * improvement. But we cannot postpone css_get(to) because if
4416 * the process that has been moved to @to does swap-in, the
4417 * refcount of @to might be decreased to 0.
4419 * We are in attach() phase, so the cgroup is guaranteed to be
4420 * alive, so we can just call css_get().
4428 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4429 struct mem_cgroup *from, struct mem_cgroup *to)
4436 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4439 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4440 struct mem_cgroup **memcgp)
4442 struct mem_cgroup *memcg = NULL;
4443 unsigned int nr_pages = 1;
4444 struct page_cgroup *pc;
4445 enum charge_type ctype;
4449 if (mem_cgroup_disabled())
4452 if (PageTransHuge(page))
4453 nr_pages <<= compound_order(page);
4455 pc = lookup_page_cgroup(page);
4456 lock_page_cgroup(pc);
4457 if (PageCgroupUsed(pc)) {
4458 memcg = pc->mem_cgroup;
4459 css_get(&memcg->css);
4461 * At migrating an anonymous page, its mapcount goes down
4462 * to 0 and uncharge() will be called. But, even if it's fully
4463 * unmapped, migration may fail and this page has to be
4464 * charged again. We set MIGRATION flag here and delay uncharge
4465 * until end_migration() is called
4467 * Corner Case Thinking
4469 * When the old page was mapped as Anon and it's unmap-and-freed
4470 * while migration was ongoing.
4471 * If unmap finds the old page, uncharge() of it will be delayed
4472 * until end_migration(). If unmap finds a new page, it's
4473 * uncharged when it make mapcount to be 1->0. If unmap code
4474 * finds swap_migration_entry, the new page will not be mapped
4475 * and end_migration() will find it(mapcount==0).
4478 * When the old page was mapped but migraion fails, the kernel
4479 * remaps it. A charge for it is kept by MIGRATION flag even
4480 * if mapcount goes down to 0. We can do remap successfully
4481 * without charging it again.
4484 * The "old" page is under lock_page() until the end of
4485 * migration, so, the old page itself will not be swapped-out.
4486 * If the new page is swapped out before end_migraton, our
4487 * hook to usual swap-out path will catch the event.
4490 SetPageCgroupMigration(pc);
4492 unlock_page_cgroup(pc);
4494 * If the page is not charged at this point,
4502 * We charge new page before it's used/mapped. So, even if unlock_page()
4503 * is called before end_migration, we can catch all events on this new
4504 * page. In the case new page is migrated but not remapped, new page's
4505 * mapcount will be finally 0 and we call uncharge in end_migration().
4508 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4510 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4512 * The page is committed to the memcg, but it's not actually
4513 * charged to the res_counter since we plan on replacing the
4514 * old one and only one page is going to be left afterwards.
4516 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4519 /* remove redundant charge if migration failed*/
4520 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4521 struct page *oldpage, struct page *newpage, bool migration_ok)
4523 struct page *used, *unused;
4524 struct page_cgroup *pc;
4530 if (!migration_ok) {
4537 anon = PageAnon(used);
4538 __mem_cgroup_uncharge_common(unused,
4539 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4540 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4542 css_put(&memcg->css);
4544 * We disallowed uncharge of pages under migration because mapcount
4545 * of the page goes down to zero, temporarly.
4546 * Clear the flag and check the page should be charged.
4548 pc = lookup_page_cgroup(oldpage);
4549 lock_page_cgroup(pc);
4550 ClearPageCgroupMigration(pc);
4551 unlock_page_cgroup(pc);
4554 * If a page is a file cache, radix-tree replacement is very atomic
4555 * and we can skip this check. When it was an Anon page, its mapcount
4556 * goes down to 0. But because we added MIGRATION flage, it's not
4557 * uncharged yet. There are several case but page->mapcount check
4558 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4559 * check. (see prepare_charge() also)
4562 mem_cgroup_uncharge_page(used);
4566 * At replace page cache, newpage is not under any memcg but it's on
4567 * LRU. So, this function doesn't touch res_counter but handles LRU
4568 * in correct way. Both pages are locked so we cannot race with uncharge.
4570 void mem_cgroup_replace_page_cache(struct page *oldpage,
4571 struct page *newpage)
4573 struct mem_cgroup *memcg = NULL;
4574 struct page_cgroup *pc;
4575 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4577 if (mem_cgroup_disabled())
4580 pc = lookup_page_cgroup(oldpage);
4581 /* fix accounting on old pages */
4582 lock_page_cgroup(pc);
4583 if (PageCgroupUsed(pc)) {
4584 memcg = pc->mem_cgroup;
4585 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4586 ClearPageCgroupUsed(pc);
4588 unlock_page_cgroup(pc);
4591 * When called from shmem_replace_page(), in some cases the
4592 * oldpage has already been charged, and in some cases not.
4597 * Even if newpage->mapping was NULL before starting replacement,
4598 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4599 * LRU while we overwrite pc->mem_cgroup.
4601 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4604 #ifdef CONFIG_DEBUG_VM
4605 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4607 struct page_cgroup *pc;
4609 pc = lookup_page_cgroup(page);
4611 * Can be NULL while feeding pages into the page allocator for
4612 * the first time, i.e. during boot or memory hotplug;
4613 * or when mem_cgroup_disabled().
4615 if (likely(pc) && PageCgroupUsed(pc))
4620 bool mem_cgroup_bad_page_check(struct page *page)
4622 if (mem_cgroup_disabled())
4625 return lookup_page_cgroup_used(page) != NULL;
4628 void mem_cgroup_print_bad_page(struct page *page)
4630 struct page_cgroup *pc;
4632 pc = lookup_page_cgroup_used(page);
4634 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4635 pc, pc->flags, pc->mem_cgroup);
4640 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4641 unsigned long long val)
4644 u64 memswlimit, memlimit;
4646 int children = mem_cgroup_count_children(memcg);
4647 u64 curusage, oldusage;
4651 * For keeping hierarchical_reclaim simple, how long we should retry
4652 * is depends on callers. We set our retry-count to be function
4653 * of # of children which we should visit in this loop.
4655 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4657 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4660 while (retry_count) {
4661 if (signal_pending(current)) {
4666 * Rather than hide all in some function, I do this in
4667 * open coded manner. You see what this really does.
4668 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4670 mutex_lock(&set_limit_mutex);
4671 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4672 if (memswlimit < val) {
4674 mutex_unlock(&set_limit_mutex);
4678 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4682 ret = res_counter_set_limit(&memcg->res, val);
4684 if (memswlimit == val)
4685 memcg->memsw_is_minimum = true;
4687 memcg->memsw_is_minimum = false;
4689 mutex_unlock(&set_limit_mutex);
4694 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4695 MEM_CGROUP_RECLAIM_SHRINK);
4696 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4697 /* Usage is reduced ? */
4698 if (curusage >= oldusage)
4701 oldusage = curusage;
4703 if (!ret && enlarge)
4704 memcg_oom_recover(memcg);
4709 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4710 unsigned long long val)
4713 u64 memlimit, memswlimit, oldusage, curusage;
4714 int children = mem_cgroup_count_children(memcg);
4718 /* see mem_cgroup_resize_res_limit */
4719 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4720 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4721 while (retry_count) {
4722 if (signal_pending(current)) {
4727 * Rather than hide all in some function, I do this in
4728 * open coded manner. You see what this really does.
4729 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4731 mutex_lock(&set_limit_mutex);
4732 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4733 if (memlimit > val) {
4735 mutex_unlock(&set_limit_mutex);
4738 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4739 if (memswlimit < val)
4741 ret = res_counter_set_limit(&memcg->memsw, val);
4743 if (memlimit == val)
4744 memcg->memsw_is_minimum = true;
4746 memcg->memsw_is_minimum = false;
4748 mutex_unlock(&set_limit_mutex);
4753 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4754 MEM_CGROUP_RECLAIM_NOSWAP |
4755 MEM_CGROUP_RECLAIM_SHRINK);
4756 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4757 /* Usage is reduced ? */
4758 if (curusage >= oldusage)
4761 oldusage = curusage;
4763 if (!ret && enlarge)
4764 memcg_oom_recover(memcg);
4768 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4770 unsigned long *total_scanned)
4772 unsigned long nr_reclaimed = 0;
4773 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4774 unsigned long reclaimed;
4776 struct mem_cgroup_tree_per_zone *mctz;
4777 unsigned long long excess;
4778 unsigned long nr_scanned;
4783 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4785 * This loop can run a while, specially if mem_cgroup's continuously
4786 * keep exceeding their soft limit and putting the system under
4793 mz = mem_cgroup_largest_soft_limit_node(mctz);
4798 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4799 gfp_mask, &nr_scanned);
4800 nr_reclaimed += reclaimed;
4801 *total_scanned += nr_scanned;
4802 spin_lock(&mctz->lock);
4805 * If we failed to reclaim anything from this memory cgroup
4806 * it is time to move on to the next cgroup
4812 * Loop until we find yet another one.
4814 * By the time we get the soft_limit lock
4815 * again, someone might have aded the
4816 * group back on the RB tree. Iterate to
4817 * make sure we get a different mem.
4818 * mem_cgroup_largest_soft_limit_node returns
4819 * NULL if no other cgroup is present on
4823 __mem_cgroup_largest_soft_limit_node(mctz);
4825 css_put(&next_mz->memcg->css);
4826 else /* next_mz == NULL or other memcg */
4830 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4831 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4833 * One school of thought says that we should not add
4834 * back the node to the tree if reclaim returns 0.
4835 * But our reclaim could return 0, simply because due
4836 * to priority we are exposing a smaller subset of
4837 * memory to reclaim from. Consider this as a longer
4840 /* If excess == 0, no tree ops */
4841 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4842 spin_unlock(&mctz->lock);
4843 css_put(&mz->memcg->css);
4846 * Could not reclaim anything and there are no more
4847 * mem cgroups to try or we seem to be looping without
4848 * reclaiming anything.
4850 if (!nr_reclaimed &&
4852 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4854 } while (!nr_reclaimed);
4856 css_put(&next_mz->memcg->css);
4857 return nr_reclaimed;
4861 * mem_cgroup_force_empty_list - clears LRU of a group
4862 * @memcg: group to clear
4865 * @lru: lru to to clear
4867 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4868 * reclaim the pages page themselves - pages are moved to the parent (or root)
4871 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4872 int node, int zid, enum lru_list lru)
4874 struct lruvec *lruvec;
4875 unsigned long flags;
4876 struct list_head *list;
4880 zone = &NODE_DATA(node)->node_zones[zid];
4881 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4882 list = &lruvec->lists[lru];
4886 struct page_cgroup *pc;
4889 spin_lock_irqsave(&zone->lru_lock, flags);
4890 if (list_empty(list)) {
4891 spin_unlock_irqrestore(&zone->lru_lock, flags);
4894 page = list_entry(list->prev, struct page, lru);
4896 list_move(&page->lru, list);
4898 spin_unlock_irqrestore(&zone->lru_lock, flags);
4901 spin_unlock_irqrestore(&zone->lru_lock, flags);
4903 pc = lookup_page_cgroup(page);
4905 if (mem_cgroup_move_parent(page, pc, memcg)) {
4906 /* found lock contention or "pc" is obsolete. */
4911 } while (!list_empty(list));
4915 * make mem_cgroup's charge to be 0 if there is no task by moving
4916 * all the charges and pages to the parent.
4917 * This enables deleting this mem_cgroup.
4919 * Caller is responsible for holding css reference on the memcg.
4921 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4927 /* This is for making all *used* pages to be on LRU. */
4928 lru_add_drain_all();
4929 drain_all_stock_sync(memcg);
4930 mem_cgroup_start_move(memcg);
4931 for_each_node_state(node, N_MEMORY) {
4932 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4935 mem_cgroup_force_empty_list(memcg,
4940 mem_cgroup_end_move(memcg);
4941 memcg_oom_recover(memcg);
4945 * Kernel memory may not necessarily be trackable to a specific
4946 * process. So they are not migrated, and therefore we can't
4947 * expect their value to drop to 0 here.
4948 * Having res filled up with kmem only is enough.
4950 * This is a safety check because mem_cgroup_force_empty_list
4951 * could have raced with mem_cgroup_replace_page_cache callers
4952 * so the lru seemed empty but the page could have been added
4953 * right after the check. RES_USAGE should be safe as we always
4954 * charge before adding to the LRU.
4956 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4957 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4958 } while (usage > 0);
4962 * This mainly exists for tests during the setting of set of use_hierarchy.
4963 * Since this is the very setting we are changing, the current hierarchy value
4966 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4970 /* bounce at first found */
4971 cgroup_for_each_child(pos, memcg->css.cgroup)
4977 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4978 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4979 * from mem_cgroup_count_children(), in the sense that we don't really care how
4980 * many children we have; we only need to know if we have any. It also counts
4981 * any memcg without hierarchy as infertile.
4983 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4985 return memcg->use_hierarchy && __memcg_has_children(memcg);
4989 * Reclaims as many pages from the given memcg as possible and moves
4990 * the rest to the parent.
4992 * Caller is responsible for holding css reference for memcg.
4994 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4996 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4997 struct cgroup *cgrp = memcg->css.cgroup;
4999 /* returns EBUSY if there is a task or if we come here twice. */
5000 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
5003 /* we call try-to-free pages for make this cgroup empty */
5004 lru_add_drain_all();
5005 /* try to free all pages in this cgroup */
5006 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
5009 if (signal_pending(current))
5012 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5016 /* maybe some writeback is necessary */
5017 congestion_wait(BLK_RW_ASYNC, HZ/10);
5022 mem_cgroup_reparent_charges(memcg);
5027 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
5029 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5032 if (mem_cgroup_is_root(memcg))
5034 css_get(&memcg->css);
5035 ret = mem_cgroup_force_empty(memcg);
5036 css_put(&memcg->css);
5042 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
5044 return mem_cgroup_from_cont(cont)->use_hierarchy;
5047 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
5051 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5052 struct cgroup *parent = cont->parent;
5053 struct mem_cgroup *parent_memcg = NULL;
5056 parent_memcg = mem_cgroup_from_cont(parent);
5058 mutex_lock(&memcg_create_mutex);
5060 if (memcg->use_hierarchy == val)
5064 * If parent's use_hierarchy is set, we can't make any modifications
5065 * in the child subtrees. If it is unset, then the change can
5066 * occur, provided the current cgroup has no children.
5068 * For the root cgroup, parent_mem is NULL, we allow value to be
5069 * set if there are no children.
5071 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5072 (val == 1 || val == 0)) {
5073 if (!__memcg_has_children(memcg))
5074 memcg->use_hierarchy = val;
5081 mutex_unlock(&memcg_create_mutex);
5087 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5088 enum mem_cgroup_stat_index idx)
5090 struct mem_cgroup *iter;
5093 /* Per-cpu values can be negative, use a signed accumulator */
5094 for_each_mem_cgroup_tree(iter, memcg)
5095 val += mem_cgroup_read_stat(iter, idx);
5097 if (val < 0) /* race ? */
5102 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5106 if (!mem_cgroup_is_root(memcg)) {
5108 return res_counter_read_u64(&memcg->res, RES_USAGE);
5110 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5114 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5115 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5117 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5118 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5121 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5123 return val << PAGE_SHIFT;
5126 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
5127 struct file *file, char __user *buf,
5128 size_t nbytes, loff_t *ppos)
5130 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5136 type = MEMFILE_TYPE(cft->private);
5137 name = MEMFILE_ATTR(cft->private);
5141 if (name == RES_USAGE)
5142 val = mem_cgroup_usage(memcg, false);
5144 val = res_counter_read_u64(&memcg->res, name);
5147 if (name == RES_USAGE)
5148 val = mem_cgroup_usage(memcg, true);
5150 val = res_counter_read_u64(&memcg->memsw, name);
5153 val = res_counter_read_u64(&memcg->kmem, name);
5159 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5160 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5163 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
5166 #ifdef CONFIG_MEMCG_KMEM
5167 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5169 * For simplicity, we won't allow this to be disabled. It also can't
5170 * be changed if the cgroup has children already, or if tasks had
5173 * If tasks join before we set the limit, a person looking at
5174 * kmem.usage_in_bytes will have no way to determine when it took
5175 * place, which makes the value quite meaningless.
5177 * After it first became limited, changes in the value of the limit are
5178 * of course permitted.
5180 mutex_lock(&memcg_create_mutex);
5181 mutex_lock(&set_limit_mutex);
5182 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
5183 if (cgroup_task_count(cont) || memcg_has_children(memcg)) {
5187 ret = res_counter_set_limit(&memcg->kmem, val);
5190 ret = memcg_update_cache_sizes(memcg);
5192 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5195 static_key_slow_inc(&memcg_kmem_enabled_key);
5197 * setting the active bit after the inc will guarantee no one
5198 * starts accounting before all call sites are patched
5200 memcg_kmem_set_active(memcg);
5202 ret = res_counter_set_limit(&memcg->kmem, val);
5204 mutex_unlock(&set_limit_mutex);
5205 mutex_unlock(&memcg_create_mutex);
5210 #ifdef CONFIG_MEMCG_KMEM
5211 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5214 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5218 memcg->kmem_account_flags = parent->kmem_account_flags;
5220 * When that happen, we need to disable the static branch only on those
5221 * memcgs that enabled it. To achieve this, we would be forced to
5222 * complicate the code by keeping track of which memcgs were the ones
5223 * that actually enabled limits, and which ones got it from its
5226 * It is a lot simpler just to do static_key_slow_inc() on every child
5227 * that is accounted.
5229 if (!memcg_kmem_is_active(memcg))
5233 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5234 * memcg is active already. If the later initialization fails then the
5235 * cgroup core triggers the cleanup so we do not have to do it here.
5237 static_key_slow_inc(&memcg_kmem_enabled_key);
5239 mutex_lock(&set_limit_mutex);
5240 memcg_stop_kmem_account();
5241 ret = memcg_update_cache_sizes(memcg);
5242 memcg_resume_kmem_account();
5243 mutex_unlock(&set_limit_mutex);
5247 #endif /* CONFIG_MEMCG_KMEM */
5250 * The user of this function is...
5253 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5256 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5259 unsigned long long val;
5262 type = MEMFILE_TYPE(cft->private);
5263 name = MEMFILE_ATTR(cft->private);
5267 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5271 /* This function does all necessary parse...reuse it */
5272 ret = res_counter_memparse_write_strategy(buffer, &val);
5276 ret = mem_cgroup_resize_limit(memcg, val);
5277 else if (type == _MEMSWAP)
5278 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5279 else if (type == _KMEM)
5280 ret = memcg_update_kmem_limit(cont, val);
5284 case RES_SOFT_LIMIT:
5285 ret = res_counter_memparse_write_strategy(buffer, &val);
5289 * For memsw, soft limits are hard to implement in terms
5290 * of semantics, for now, we support soft limits for
5291 * control without swap
5294 ret = res_counter_set_soft_limit(&memcg->res, val);
5299 ret = -EINVAL; /* should be BUG() ? */
5305 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5306 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5308 struct cgroup *cgroup;
5309 unsigned long long min_limit, min_memsw_limit, tmp;
5311 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5312 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5313 cgroup = memcg->css.cgroup;
5314 if (!memcg->use_hierarchy)
5317 while (cgroup->parent) {
5318 cgroup = cgroup->parent;
5319 memcg = mem_cgroup_from_cont(cgroup);
5320 if (!memcg->use_hierarchy)
5322 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5323 min_limit = min(min_limit, tmp);
5324 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5325 min_memsw_limit = min(min_memsw_limit, tmp);
5328 *mem_limit = min_limit;
5329 *memsw_limit = min_memsw_limit;
5332 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5334 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5338 type = MEMFILE_TYPE(event);
5339 name = MEMFILE_ATTR(event);
5344 res_counter_reset_max(&memcg->res);
5345 else if (type == _MEMSWAP)
5346 res_counter_reset_max(&memcg->memsw);
5347 else if (type == _KMEM)
5348 res_counter_reset_max(&memcg->kmem);
5354 res_counter_reset_failcnt(&memcg->res);
5355 else if (type == _MEMSWAP)
5356 res_counter_reset_failcnt(&memcg->memsw);
5357 else if (type == _KMEM)
5358 res_counter_reset_failcnt(&memcg->kmem);
5367 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5370 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5374 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5375 struct cftype *cft, u64 val)
5377 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5379 if (val >= (1 << NR_MOVE_TYPE))
5383 * No kind of locking is needed in here, because ->can_attach() will
5384 * check this value once in the beginning of the process, and then carry
5385 * on with stale data. This means that changes to this value will only
5386 * affect task migrations starting after the change.
5388 memcg->move_charge_at_immigrate = val;
5392 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5393 struct cftype *cft, u64 val)
5400 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5404 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5405 unsigned long node_nr;
5406 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5408 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5409 seq_printf(m, "total=%lu", total_nr);
5410 for_each_node_state(nid, N_MEMORY) {
5411 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5412 seq_printf(m, " N%d=%lu", nid, node_nr);
5416 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5417 seq_printf(m, "file=%lu", file_nr);
5418 for_each_node_state(nid, N_MEMORY) {
5419 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5421 seq_printf(m, " N%d=%lu", nid, node_nr);
5425 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5426 seq_printf(m, "anon=%lu", anon_nr);
5427 for_each_node_state(nid, N_MEMORY) {
5428 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5430 seq_printf(m, " N%d=%lu", nid, node_nr);
5434 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5435 seq_printf(m, "unevictable=%lu", unevictable_nr);
5436 for_each_node_state(nid, N_MEMORY) {
5437 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5438 BIT(LRU_UNEVICTABLE));
5439 seq_printf(m, " N%d=%lu", nid, node_nr);
5444 #endif /* CONFIG_NUMA */
5446 static inline void mem_cgroup_lru_names_not_uptodate(void)
5448 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5451 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5454 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5455 struct mem_cgroup *mi;
5458 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5459 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5461 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5462 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5465 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5466 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5467 mem_cgroup_read_events(memcg, i));
5469 for (i = 0; i < NR_LRU_LISTS; i++)
5470 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5471 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5473 /* Hierarchical information */
5475 unsigned long long limit, memsw_limit;
5476 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5477 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5478 if (do_swap_account)
5479 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5483 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5486 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5488 for_each_mem_cgroup_tree(mi, memcg)
5489 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5490 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5493 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5494 unsigned long long val = 0;
5496 for_each_mem_cgroup_tree(mi, memcg)
5497 val += mem_cgroup_read_events(mi, i);
5498 seq_printf(m, "total_%s %llu\n",
5499 mem_cgroup_events_names[i], val);
5502 for (i = 0; i < NR_LRU_LISTS; i++) {
5503 unsigned long long val = 0;
5505 for_each_mem_cgroup_tree(mi, memcg)
5506 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5507 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5510 #ifdef CONFIG_DEBUG_VM
5513 struct mem_cgroup_per_zone *mz;
5514 struct zone_reclaim_stat *rstat;
5515 unsigned long recent_rotated[2] = {0, 0};
5516 unsigned long recent_scanned[2] = {0, 0};
5518 for_each_online_node(nid)
5519 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5520 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5521 rstat = &mz->lruvec.reclaim_stat;
5523 recent_rotated[0] += rstat->recent_rotated[0];
5524 recent_rotated[1] += rstat->recent_rotated[1];
5525 recent_scanned[0] += rstat->recent_scanned[0];
5526 recent_scanned[1] += rstat->recent_scanned[1];
5528 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5529 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5530 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5531 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5538 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5540 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5542 return mem_cgroup_swappiness(memcg);
5545 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5548 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5549 struct mem_cgroup *parent;
5554 if (cgrp->parent == NULL)
5557 parent = mem_cgroup_from_cont(cgrp->parent);
5559 mutex_lock(&memcg_create_mutex);
5561 /* If under hierarchy, only empty-root can set this value */
5562 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5563 mutex_unlock(&memcg_create_mutex);
5567 memcg->swappiness = val;
5569 mutex_unlock(&memcg_create_mutex);
5574 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5576 struct mem_cgroup_threshold_ary *t;
5582 t = rcu_dereference(memcg->thresholds.primary);
5584 t = rcu_dereference(memcg->memsw_thresholds.primary);
5589 usage = mem_cgroup_usage(memcg, swap);
5592 * current_threshold points to threshold just below or equal to usage.
5593 * If it's not true, a threshold was crossed after last
5594 * call of __mem_cgroup_threshold().
5596 i = t->current_threshold;
5599 * Iterate backward over array of thresholds starting from
5600 * current_threshold and check if a threshold is crossed.
5601 * If none of thresholds below usage is crossed, we read
5602 * only one element of the array here.
5604 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5605 eventfd_signal(t->entries[i].eventfd, 1);
5607 /* i = current_threshold + 1 */
5611 * Iterate forward over array of thresholds starting from
5612 * current_threshold+1 and check if a threshold is crossed.
5613 * If none of thresholds above usage is crossed, we read
5614 * only one element of the array here.
5616 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5617 eventfd_signal(t->entries[i].eventfd, 1);
5619 /* Update current_threshold */
5620 t->current_threshold = i - 1;
5625 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5628 __mem_cgroup_threshold(memcg, false);
5629 if (do_swap_account)
5630 __mem_cgroup_threshold(memcg, true);
5632 memcg = parent_mem_cgroup(memcg);
5636 static int compare_thresholds(const void *a, const void *b)
5638 const struct mem_cgroup_threshold *_a = a;
5639 const struct mem_cgroup_threshold *_b = b;
5641 return _a->threshold - _b->threshold;
5644 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5646 struct mem_cgroup_eventfd_list *ev;
5648 list_for_each_entry(ev, &memcg->oom_notify, list)
5649 eventfd_signal(ev->eventfd, 1);
5653 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5655 struct mem_cgroup *iter;
5657 for_each_mem_cgroup_tree(iter, memcg)
5658 mem_cgroup_oom_notify_cb(iter);
5661 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5662 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5664 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5665 struct mem_cgroup_thresholds *thresholds;
5666 struct mem_cgroup_threshold_ary *new;
5667 enum res_type type = MEMFILE_TYPE(cft->private);
5668 u64 threshold, usage;
5671 ret = res_counter_memparse_write_strategy(args, &threshold);
5675 mutex_lock(&memcg->thresholds_lock);
5678 thresholds = &memcg->thresholds;
5679 else if (type == _MEMSWAP)
5680 thresholds = &memcg->memsw_thresholds;
5684 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5686 /* Check if a threshold crossed before adding a new one */
5687 if (thresholds->primary)
5688 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5690 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5692 /* Allocate memory for new array of thresholds */
5693 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5701 /* Copy thresholds (if any) to new array */
5702 if (thresholds->primary) {
5703 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5704 sizeof(struct mem_cgroup_threshold));
5707 /* Add new threshold */
5708 new->entries[size - 1].eventfd = eventfd;
5709 new->entries[size - 1].threshold = threshold;
5711 /* Sort thresholds. Registering of new threshold isn't time-critical */
5712 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5713 compare_thresholds, NULL);
5715 /* Find current threshold */
5716 new->current_threshold = -1;
5717 for (i = 0; i < size; i++) {
5718 if (new->entries[i].threshold <= usage) {
5720 * new->current_threshold will not be used until
5721 * rcu_assign_pointer(), so it's safe to increment
5724 ++new->current_threshold;
5729 /* Free old spare buffer and save old primary buffer as spare */
5730 kfree(thresholds->spare);
5731 thresholds->spare = thresholds->primary;
5733 rcu_assign_pointer(thresholds->primary, new);
5735 /* To be sure that nobody uses thresholds */
5739 mutex_unlock(&memcg->thresholds_lock);
5744 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5745 struct cftype *cft, struct eventfd_ctx *eventfd)
5747 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5748 struct mem_cgroup_thresholds *thresholds;
5749 struct mem_cgroup_threshold_ary *new;
5750 enum res_type type = MEMFILE_TYPE(cft->private);
5754 mutex_lock(&memcg->thresholds_lock);
5756 thresholds = &memcg->thresholds;
5757 else if (type == _MEMSWAP)
5758 thresholds = &memcg->memsw_thresholds;
5762 if (!thresholds->primary)
5765 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5767 /* Check if a threshold crossed before removing */
5768 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5770 /* Calculate new number of threshold */
5772 for (i = 0; i < thresholds->primary->size; i++) {
5773 if (thresholds->primary->entries[i].eventfd != eventfd)
5777 new = thresholds->spare;
5779 /* Set thresholds array to NULL if we don't have thresholds */
5788 /* Copy thresholds and find current threshold */
5789 new->current_threshold = -1;
5790 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5791 if (thresholds->primary->entries[i].eventfd == eventfd)
5794 new->entries[j] = thresholds->primary->entries[i];
5795 if (new->entries[j].threshold <= usage) {
5797 * new->current_threshold will not be used
5798 * until rcu_assign_pointer(), so it's safe to increment
5801 ++new->current_threshold;
5807 /* Swap primary and spare array */
5808 thresholds->spare = thresholds->primary;
5809 /* If all events are unregistered, free the spare array */
5811 kfree(thresholds->spare);
5812 thresholds->spare = NULL;
5815 rcu_assign_pointer(thresholds->primary, new);
5817 /* To be sure that nobody uses thresholds */
5820 mutex_unlock(&memcg->thresholds_lock);
5823 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5824 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5826 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5827 struct mem_cgroup_eventfd_list *event;
5828 enum res_type type = MEMFILE_TYPE(cft->private);
5830 BUG_ON(type != _OOM_TYPE);
5831 event = kmalloc(sizeof(*event), GFP_KERNEL);
5835 spin_lock(&memcg_oom_lock);
5837 event->eventfd = eventfd;
5838 list_add(&event->list, &memcg->oom_notify);
5840 /* already in OOM ? */
5841 if (atomic_read(&memcg->under_oom))
5842 eventfd_signal(eventfd, 1);
5843 spin_unlock(&memcg_oom_lock);
5848 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5849 struct cftype *cft, struct eventfd_ctx *eventfd)
5851 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5852 struct mem_cgroup_eventfd_list *ev, *tmp;
5853 enum res_type type = MEMFILE_TYPE(cft->private);
5855 BUG_ON(type != _OOM_TYPE);
5857 spin_lock(&memcg_oom_lock);
5859 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5860 if (ev->eventfd == eventfd) {
5861 list_del(&ev->list);
5866 spin_unlock(&memcg_oom_lock);
5869 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5870 struct cftype *cft, struct cgroup_map_cb *cb)
5872 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5874 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5876 if (atomic_read(&memcg->under_oom))
5877 cb->fill(cb, "under_oom", 1);
5879 cb->fill(cb, "under_oom", 0);
5883 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5884 struct cftype *cft, u64 val)
5886 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5887 struct mem_cgroup *parent;
5889 /* cannot set to root cgroup and only 0 and 1 are allowed */
5890 if (!cgrp->parent || !((val == 0) || (val == 1)))
5893 parent = mem_cgroup_from_cont(cgrp->parent);
5895 mutex_lock(&memcg_create_mutex);
5896 /* oom-kill-disable is a flag for subhierarchy. */
5897 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5898 mutex_unlock(&memcg_create_mutex);
5901 memcg->oom_kill_disable = val;
5903 memcg_oom_recover(memcg);
5904 mutex_unlock(&memcg_create_mutex);
5908 #ifdef CONFIG_MEMCG_KMEM
5909 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5913 memcg->kmemcg_id = -1;
5914 ret = memcg_propagate_kmem(memcg);
5918 return mem_cgroup_sockets_init(memcg, ss);
5921 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5923 mem_cgroup_sockets_destroy(memcg);
5926 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5928 if (!memcg_kmem_is_active(memcg))
5932 * kmem charges can outlive the cgroup. In the case of slab
5933 * pages, for instance, a page contain objects from various
5934 * processes. As we prevent from taking a reference for every
5935 * such allocation we have to be careful when doing uncharge
5936 * (see memcg_uncharge_kmem) and here during offlining.
5938 * The idea is that that only the _last_ uncharge which sees
5939 * the dead memcg will drop the last reference. An additional
5940 * reference is taken here before the group is marked dead
5941 * which is then paired with css_put during uncharge resp. here.
5943 * Although this might sound strange as this path is called from
5944 * css_offline() when the referencemight have dropped down to 0
5945 * and shouldn't be incremented anymore (css_tryget would fail)
5946 * we do not have other options because of the kmem allocations
5949 css_get(&memcg->css);
5951 memcg_kmem_mark_dead(memcg);
5953 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5956 if (memcg_kmem_test_and_clear_dead(memcg))
5957 css_put(&memcg->css);
5960 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5965 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5969 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5974 static struct cftype mem_cgroup_files[] = {
5976 .name = "usage_in_bytes",
5977 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5978 .read = mem_cgroup_read,
5979 .register_event = mem_cgroup_usage_register_event,
5980 .unregister_event = mem_cgroup_usage_unregister_event,
5983 .name = "max_usage_in_bytes",
5984 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5985 .trigger = mem_cgroup_reset,
5986 .read = mem_cgroup_read,
5989 .name = "limit_in_bytes",
5990 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5991 .write_string = mem_cgroup_write,
5992 .read = mem_cgroup_read,
5995 .name = "soft_limit_in_bytes",
5996 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5997 .write_string = mem_cgroup_write,
5998 .read = mem_cgroup_read,
6002 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6003 .trigger = mem_cgroup_reset,
6004 .read = mem_cgroup_read,
6008 .read_seq_string = memcg_stat_show,
6011 .name = "force_empty",
6012 .trigger = mem_cgroup_force_empty_write,
6015 .name = "use_hierarchy",
6016 .flags = CFTYPE_INSANE,
6017 .write_u64 = mem_cgroup_hierarchy_write,
6018 .read_u64 = mem_cgroup_hierarchy_read,
6021 .name = "swappiness",
6022 .read_u64 = mem_cgroup_swappiness_read,
6023 .write_u64 = mem_cgroup_swappiness_write,
6026 .name = "move_charge_at_immigrate",
6027 .read_u64 = mem_cgroup_move_charge_read,
6028 .write_u64 = mem_cgroup_move_charge_write,
6031 .name = "oom_control",
6032 .read_map = mem_cgroup_oom_control_read,
6033 .write_u64 = mem_cgroup_oom_control_write,
6034 .register_event = mem_cgroup_oom_register_event,
6035 .unregister_event = mem_cgroup_oom_unregister_event,
6036 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6039 .name = "pressure_level",
6040 .register_event = vmpressure_register_event,
6041 .unregister_event = vmpressure_unregister_event,
6045 .name = "numa_stat",
6046 .read_seq_string = memcg_numa_stat_show,
6049 #ifdef CONFIG_MEMCG_KMEM
6051 .name = "kmem.limit_in_bytes",
6052 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6053 .write_string = mem_cgroup_write,
6054 .read = mem_cgroup_read,
6057 .name = "kmem.usage_in_bytes",
6058 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6059 .read = mem_cgroup_read,
6062 .name = "kmem.failcnt",
6063 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6064 .trigger = mem_cgroup_reset,
6065 .read = mem_cgroup_read,
6068 .name = "kmem.max_usage_in_bytes",
6069 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6070 .trigger = mem_cgroup_reset,
6071 .read = mem_cgroup_read,
6073 #ifdef CONFIG_SLABINFO
6075 .name = "kmem.slabinfo",
6076 .read_seq_string = mem_cgroup_slabinfo_read,
6080 { }, /* terminate */
6083 #ifdef CONFIG_MEMCG_SWAP
6084 static struct cftype memsw_cgroup_files[] = {
6086 .name = "memsw.usage_in_bytes",
6087 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6088 .read = mem_cgroup_read,
6089 .register_event = mem_cgroup_usage_register_event,
6090 .unregister_event = mem_cgroup_usage_unregister_event,
6093 .name = "memsw.max_usage_in_bytes",
6094 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6095 .trigger = mem_cgroup_reset,
6096 .read = mem_cgroup_read,
6099 .name = "memsw.limit_in_bytes",
6100 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6101 .write_string = mem_cgroup_write,
6102 .read = mem_cgroup_read,
6105 .name = "memsw.failcnt",
6106 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6107 .trigger = mem_cgroup_reset,
6108 .read = mem_cgroup_read,
6110 { }, /* terminate */
6113 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6115 struct mem_cgroup_per_node *pn;
6116 struct mem_cgroup_per_zone *mz;
6117 int zone, tmp = node;
6119 * This routine is called against possible nodes.
6120 * But it's BUG to call kmalloc() against offline node.
6122 * TODO: this routine can waste much memory for nodes which will
6123 * never be onlined. It's better to use memory hotplug callback
6126 if (!node_state(node, N_NORMAL_MEMORY))
6128 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6132 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6133 mz = &pn->zoneinfo[zone];
6134 lruvec_init(&mz->lruvec);
6135 mz->usage_in_excess = 0;
6136 mz->on_tree = false;
6139 memcg->nodeinfo[node] = pn;
6143 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6145 kfree(memcg->nodeinfo[node]);
6148 static struct mem_cgroup *mem_cgroup_alloc(void)
6150 struct mem_cgroup *memcg;
6151 size_t size = memcg_size();
6153 /* Can be very big if nr_node_ids is very big */
6154 if (size < PAGE_SIZE)
6155 memcg = kzalloc(size, GFP_KERNEL);
6157 memcg = vzalloc(size);
6162 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6165 spin_lock_init(&memcg->pcp_counter_lock);
6169 if (size < PAGE_SIZE)
6177 * At destroying mem_cgroup, references from swap_cgroup can remain.
6178 * (scanning all at force_empty is too costly...)
6180 * Instead of clearing all references at force_empty, we remember
6181 * the number of reference from swap_cgroup and free mem_cgroup when
6182 * it goes down to 0.
6184 * Removal of cgroup itself succeeds regardless of refs from swap.
6187 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6190 size_t size = memcg_size();
6192 mem_cgroup_remove_from_trees(memcg);
6193 free_css_id(&mem_cgroup_subsys, &memcg->css);
6196 free_mem_cgroup_per_zone_info(memcg, node);
6198 free_percpu(memcg->stat);
6201 * We need to make sure that (at least for now), the jump label
6202 * destruction code runs outside of the cgroup lock. This is because
6203 * get_online_cpus(), which is called from the static_branch update,
6204 * can't be called inside the cgroup_lock. cpusets are the ones
6205 * enforcing this dependency, so if they ever change, we might as well.
6207 * schedule_work() will guarantee this happens. Be careful if you need
6208 * to move this code around, and make sure it is outside
6211 disarm_static_keys(memcg);
6212 if (size < PAGE_SIZE)
6220 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
6221 * but in process context. The work_freeing structure is overlaid
6222 * on the rcu_freeing structure, which itself is overlaid on memsw.
6224 static void free_work(struct work_struct *work)
6226 struct mem_cgroup *memcg;
6228 memcg = container_of(work, struct mem_cgroup, work_freeing);
6229 __mem_cgroup_free(memcg);
6232 static void free_rcu(struct rcu_head *rcu_head)
6234 struct mem_cgroup *memcg;
6236 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
6237 INIT_WORK(&memcg->work_freeing, free_work);
6238 schedule_work(&memcg->work_freeing);
6241 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
6243 if (atomic_sub_and_test(count, &memcg->refcnt))
6244 call_rcu(&memcg->rcu_freeing, free_rcu);
6247 static void mem_cgroup_put(struct mem_cgroup *memcg)
6249 __mem_cgroup_put(memcg, 1);
6253 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6255 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6257 if (!memcg->res.parent)
6259 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6261 EXPORT_SYMBOL(parent_mem_cgroup);
6263 static void __init mem_cgroup_soft_limit_tree_init(void)
6265 struct mem_cgroup_tree_per_node *rtpn;
6266 struct mem_cgroup_tree_per_zone *rtpz;
6267 int tmp, node, zone;
6269 for_each_node(node) {
6271 if (!node_state(node, N_NORMAL_MEMORY))
6273 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6276 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6278 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6279 rtpz = &rtpn->rb_tree_per_zone[zone];
6280 rtpz->rb_root = RB_ROOT;
6281 spin_lock_init(&rtpz->lock);
6286 static struct cgroup_subsys_state * __ref
6287 mem_cgroup_css_alloc(struct cgroup *cont)
6289 struct mem_cgroup *memcg;
6290 long error = -ENOMEM;
6293 memcg = mem_cgroup_alloc();
6295 return ERR_PTR(error);
6298 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6302 if (cont->parent == NULL) {
6303 root_mem_cgroup = memcg;
6304 res_counter_init(&memcg->res, NULL);
6305 res_counter_init(&memcg->memsw, NULL);
6306 res_counter_init(&memcg->kmem, NULL);
6309 memcg->last_scanned_node = MAX_NUMNODES;
6310 INIT_LIST_HEAD(&memcg->oom_notify);
6311 atomic_set(&memcg->refcnt, 1);
6312 memcg->move_charge_at_immigrate = 0;
6313 mutex_init(&memcg->thresholds_lock);
6314 spin_lock_init(&memcg->move_lock);
6315 vmpressure_init(&memcg->vmpressure);
6320 __mem_cgroup_free(memcg);
6321 return ERR_PTR(error);
6325 mem_cgroup_css_online(struct cgroup *cont)
6327 struct mem_cgroup *memcg, *parent;
6333 mutex_lock(&memcg_create_mutex);
6334 memcg = mem_cgroup_from_cont(cont);
6335 parent = mem_cgroup_from_cont(cont->parent);
6337 memcg->use_hierarchy = parent->use_hierarchy;
6338 memcg->oom_kill_disable = parent->oom_kill_disable;
6339 memcg->swappiness = mem_cgroup_swappiness(parent);
6341 if (parent->use_hierarchy) {
6342 res_counter_init(&memcg->res, &parent->res);
6343 res_counter_init(&memcg->memsw, &parent->memsw);
6344 res_counter_init(&memcg->kmem, &parent->kmem);
6347 * No need to take a reference to the parent because cgroup
6348 * core guarantees its existence.
6351 res_counter_init(&memcg->res, NULL);
6352 res_counter_init(&memcg->memsw, NULL);
6353 res_counter_init(&memcg->kmem, NULL);
6355 * Deeper hierachy with use_hierarchy == false doesn't make
6356 * much sense so let cgroup subsystem know about this
6357 * unfortunate state in our controller.
6359 if (parent != root_mem_cgroup)
6360 mem_cgroup_subsys.broken_hierarchy = true;
6363 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6364 mutex_unlock(&memcg_create_mutex);
6369 * Announce all parents that a group from their hierarchy is gone.
6371 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6373 struct mem_cgroup *parent = memcg;
6375 while ((parent = parent_mem_cgroup(parent)))
6376 mem_cgroup_iter_invalidate(parent);
6379 * if the root memcg is not hierarchical we have to check it
6382 if (!root_mem_cgroup->use_hierarchy)
6383 mem_cgroup_iter_invalidate(root_mem_cgroup);
6386 static void mem_cgroup_css_offline(struct cgroup *cont)
6388 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6390 kmem_cgroup_css_offline(memcg);
6392 mem_cgroup_invalidate_reclaim_iterators(memcg);
6393 mem_cgroup_reparent_charges(memcg);
6394 mem_cgroup_destroy_all_caches(memcg);
6397 static void mem_cgroup_css_free(struct cgroup *cont)
6399 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6401 memcg_destroy_kmem(memcg);
6402 __mem_cgroup_free(memcg);
6406 /* Handlers for move charge at task migration. */
6407 #define PRECHARGE_COUNT_AT_ONCE 256
6408 static int mem_cgroup_do_precharge(unsigned long count)
6411 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6412 struct mem_cgroup *memcg = mc.to;
6414 if (mem_cgroup_is_root(memcg)) {
6415 mc.precharge += count;
6416 /* we don't need css_get for root */
6419 /* try to charge at once */
6421 struct res_counter *dummy;
6423 * "memcg" cannot be under rmdir() because we've already checked
6424 * by cgroup_lock_live_cgroup() that it is not removed and we
6425 * are still under the same cgroup_mutex. So we can postpone
6428 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6430 if (do_swap_account && res_counter_charge(&memcg->memsw,
6431 PAGE_SIZE * count, &dummy)) {
6432 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6435 mc.precharge += count;
6439 /* fall back to one by one charge */
6441 if (signal_pending(current)) {
6445 if (!batch_count--) {
6446 batch_count = PRECHARGE_COUNT_AT_ONCE;
6449 ret = __mem_cgroup_try_charge(NULL,
6450 GFP_KERNEL, 1, &memcg, false);
6452 /* mem_cgroup_clear_mc() will do uncharge later */
6460 * get_mctgt_type - get target type of moving charge
6461 * @vma: the vma the pte to be checked belongs
6462 * @addr: the address corresponding to the pte to be checked
6463 * @ptent: the pte to be checked
6464 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6467 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6468 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6469 * move charge. if @target is not NULL, the page is stored in target->page
6470 * with extra refcnt got(Callers should handle it).
6471 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6472 * target for charge migration. if @target is not NULL, the entry is stored
6475 * Called with pte lock held.
6482 enum mc_target_type {
6488 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6489 unsigned long addr, pte_t ptent)
6491 struct page *page = vm_normal_page(vma, addr, ptent);
6493 if (!page || !page_mapped(page))
6495 if (PageAnon(page)) {
6496 /* we don't move shared anon */
6499 } else if (!move_file())
6500 /* we ignore mapcount for file pages */
6502 if (!get_page_unless_zero(page))
6509 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6510 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6512 struct page *page = NULL;
6513 swp_entry_t ent = pte_to_swp_entry(ptent);
6515 if (!move_anon() || non_swap_entry(ent))
6518 * Because lookup_swap_cache() updates some statistics counter,
6519 * we call find_get_page() with swapper_space directly.
6521 page = find_get_page(swap_address_space(ent), ent.val);
6522 if (do_swap_account)
6523 entry->val = ent.val;
6528 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6529 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6535 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6536 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6538 struct page *page = NULL;
6539 struct address_space *mapping;
6542 if (!vma->vm_file) /* anonymous vma */
6547 mapping = vma->vm_file->f_mapping;
6548 if (pte_none(ptent))
6549 pgoff = linear_page_index(vma, addr);
6550 else /* pte_file(ptent) is true */
6551 pgoff = pte_to_pgoff(ptent);
6553 /* page is moved even if it's not RSS of this task(page-faulted). */
6554 page = find_get_page(mapping, pgoff);
6557 /* shmem/tmpfs may report page out on swap: account for that too. */
6558 if (radix_tree_exceptional_entry(page)) {
6559 swp_entry_t swap = radix_to_swp_entry(page);
6560 if (do_swap_account)
6562 page = find_get_page(swap_address_space(swap), swap.val);
6568 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6569 unsigned long addr, pte_t ptent, union mc_target *target)
6571 struct page *page = NULL;
6572 struct page_cgroup *pc;
6573 enum mc_target_type ret = MC_TARGET_NONE;
6574 swp_entry_t ent = { .val = 0 };
6576 if (pte_present(ptent))
6577 page = mc_handle_present_pte(vma, addr, ptent);
6578 else if (is_swap_pte(ptent))
6579 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6580 else if (pte_none(ptent) || pte_file(ptent))
6581 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6583 if (!page && !ent.val)
6586 pc = lookup_page_cgroup(page);
6588 * Do only loose check w/o page_cgroup lock.
6589 * mem_cgroup_move_account() checks the pc is valid or not under
6592 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6593 ret = MC_TARGET_PAGE;
6595 target->page = page;
6597 if (!ret || !target)
6600 /* There is a swap entry and a page doesn't exist or isn't charged */
6601 if (ent.val && !ret &&
6602 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6603 ret = MC_TARGET_SWAP;
6610 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6612 * We don't consider swapping or file mapped pages because THP does not
6613 * support them for now.
6614 * Caller should make sure that pmd_trans_huge(pmd) is true.
6616 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6617 unsigned long addr, pmd_t pmd, union mc_target *target)
6619 struct page *page = NULL;
6620 struct page_cgroup *pc;
6621 enum mc_target_type ret = MC_TARGET_NONE;
6623 page = pmd_page(pmd);
6624 VM_BUG_ON(!page || !PageHead(page));
6627 pc = lookup_page_cgroup(page);
6628 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6629 ret = MC_TARGET_PAGE;
6632 target->page = page;
6638 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6639 unsigned long addr, pmd_t pmd, union mc_target *target)
6641 return MC_TARGET_NONE;
6645 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6646 unsigned long addr, unsigned long end,
6647 struct mm_walk *walk)
6649 struct vm_area_struct *vma = walk->private;
6653 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6654 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6655 mc.precharge += HPAGE_PMD_NR;
6656 spin_unlock(&vma->vm_mm->page_table_lock);
6660 if (pmd_trans_unstable(pmd))
6662 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6663 for (; addr != end; pte++, addr += PAGE_SIZE)
6664 if (get_mctgt_type(vma, addr, *pte, NULL))
6665 mc.precharge++; /* increment precharge temporarily */
6666 pte_unmap_unlock(pte - 1, ptl);
6672 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6674 unsigned long precharge;
6675 struct vm_area_struct *vma;
6677 down_read(&mm->mmap_sem);
6678 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6679 struct mm_walk mem_cgroup_count_precharge_walk = {
6680 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6684 if (is_vm_hugetlb_page(vma))
6686 walk_page_range(vma->vm_start, vma->vm_end,
6687 &mem_cgroup_count_precharge_walk);
6689 up_read(&mm->mmap_sem);
6691 precharge = mc.precharge;
6697 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6699 unsigned long precharge = mem_cgroup_count_precharge(mm);
6701 VM_BUG_ON(mc.moving_task);
6702 mc.moving_task = current;
6703 return mem_cgroup_do_precharge(precharge);
6706 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6707 static void __mem_cgroup_clear_mc(void)
6709 struct mem_cgroup *from = mc.from;
6710 struct mem_cgroup *to = mc.to;
6713 /* we must uncharge all the leftover precharges from mc.to */
6715 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6719 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6720 * we must uncharge here.
6722 if (mc.moved_charge) {
6723 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6724 mc.moved_charge = 0;
6726 /* we must fixup refcnts and charges */
6727 if (mc.moved_swap) {
6728 /* uncharge swap account from the old cgroup */
6729 if (!mem_cgroup_is_root(mc.from))
6730 res_counter_uncharge(&mc.from->memsw,
6731 PAGE_SIZE * mc.moved_swap);
6733 for (i = 0; i < mc.moved_swap; i++)
6734 css_put(&mc.from->css);
6736 if (!mem_cgroup_is_root(mc.to)) {
6738 * we charged both to->res and to->memsw, so we should
6741 res_counter_uncharge(&mc.to->res,
6742 PAGE_SIZE * mc.moved_swap);
6744 /* we've already done css_get(mc.to) */
6747 memcg_oom_recover(from);
6748 memcg_oom_recover(to);
6749 wake_up_all(&mc.waitq);
6752 static void mem_cgroup_clear_mc(void)
6754 struct mem_cgroup *from = mc.from;
6757 * we must clear moving_task before waking up waiters at the end of
6760 mc.moving_task = NULL;
6761 __mem_cgroup_clear_mc();
6762 spin_lock(&mc.lock);
6765 spin_unlock(&mc.lock);
6766 mem_cgroup_end_move(from);
6769 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6770 struct cgroup_taskset *tset)
6772 struct task_struct *p = cgroup_taskset_first(tset);
6774 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6775 unsigned long move_charge_at_immigrate;
6778 * We are now commited to this value whatever it is. Changes in this
6779 * tunable will only affect upcoming migrations, not the current one.
6780 * So we need to save it, and keep it going.
6782 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6783 if (move_charge_at_immigrate) {
6784 struct mm_struct *mm;
6785 struct mem_cgroup *from = mem_cgroup_from_task(p);
6787 VM_BUG_ON(from == memcg);
6789 mm = get_task_mm(p);
6792 /* We move charges only when we move a owner of the mm */
6793 if (mm->owner == p) {
6796 VM_BUG_ON(mc.precharge);
6797 VM_BUG_ON(mc.moved_charge);
6798 VM_BUG_ON(mc.moved_swap);
6799 mem_cgroup_start_move(from);
6800 spin_lock(&mc.lock);
6803 mc.immigrate_flags = move_charge_at_immigrate;
6804 spin_unlock(&mc.lock);
6805 /* We set mc.moving_task later */
6807 ret = mem_cgroup_precharge_mc(mm);
6809 mem_cgroup_clear_mc();
6816 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6817 struct cgroup_taskset *tset)
6819 mem_cgroup_clear_mc();
6822 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6823 unsigned long addr, unsigned long end,
6824 struct mm_walk *walk)
6827 struct vm_area_struct *vma = walk->private;
6830 enum mc_target_type target_type;
6831 union mc_target target;
6833 struct page_cgroup *pc;
6836 * We don't take compound_lock() here but no race with splitting thp
6838 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6839 * under splitting, which means there's no concurrent thp split,
6840 * - if another thread runs into split_huge_page() just after we
6841 * entered this if-block, the thread must wait for page table lock
6842 * to be unlocked in __split_huge_page_splitting(), where the main
6843 * part of thp split is not executed yet.
6845 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6846 if (mc.precharge < HPAGE_PMD_NR) {
6847 spin_unlock(&vma->vm_mm->page_table_lock);
6850 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6851 if (target_type == MC_TARGET_PAGE) {
6853 if (!isolate_lru_page(page)) {
6854 pc = lookup_page_cgroup(page);
6855 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6856 pc, mc.from, mc.to)) {
6857 mc.precharge -= HPAGE_PMD_NR;
6858 mc.moved_charge += HPAGE_PMD_NR;
6860 putback_lru_page(page);
6864 spin_unlock(&vma->vm_mm->page_table_lock);
6868 if (pmd_trans_unstable(pmd))
6871 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6872 for (; addr != end; addr += PAGE_SIZE) {
6873 pte_t ptent = *(pte++);
6879 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6880 case MC_TARGET_PAGE:
6882 if (isolate_lru_page(page))
6884 pc = lookup_page_cgroup(page);
6885 if (!mem_cgroup_move_account(page, 1, pc,
6888 /* we uncharge from mc.from later. */
6891 putback_lru_page(page);
6892 put: /* get_mctgt_type() gets the page */
6895 case MC_TARGET_SWAP:
6897 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6899 /* we fixup refcnts and charges later. */
6907 pte_unmap_unlock(pte - 1, ptl);
6912 * We have consumed all precharges we got in can_attach().
6913 * We try charge one by one, but don't do any additional
6914 * charges to mc.to if we have failed in charge once in attach()
6917 ret = mem_cgroup_do_precharge(1);
6925 static void mem_cgroup_move_charge(struct mm_struct *mm)
6927 struct vm_area_struct *vma;
6929 lru_add_drain_all();
6931 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6933 * Someone who are holding the mmap_sem might be waiting in
6934 * waitq. So we cancel all extra charges, wake up all waiters,
6935 * and retry. Because we cancel precharges, we might not be able
6936 * to move enough charges, but moving charge is a best-effort
6937 * feature anyway, so it wouldn't be a big problem.
6939 __mem_cgroup_clear_mc();
6943 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6945 struct mm_walk mem_cgroup_move_charge_walk = {
6946 .pmd_entry = mem_cgroup_move_charge_pte_range,
6950 if (is_vm_hugetlb_page(vma))
6952 ret = walk_page_range(vma->vm_start, vma->vm_end,
6953 &mem_cgroup_move_charge_walk);
6956 * means we have consumed all precharges and failed in
6957 * doing additional charge. Just abandon here.
6961 up_read(&mm->mmap_sem);
6964 static void mem_cgroup_move_task(struct cgroup *cont,
6965 struct cgroup_taskset *tset)
6967 struct task_struct *p = cgroup_taskset_first(tset);
6968 struct mm_struct *mm = get_task_mm(p);
6972 mem_cgroup_move_charge(mm);
6976 mem_cgroup_clear_mc();
6978 #else /* !CONFIG_MMU */
6979 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6980 struct cgroup_taskset *tset)
6984 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6985 struct cgroup_taskset *tset)
6988 static void mem_cgroup_move_task(struct cgroup *cont,
6989 struct cgroup_taskset *tset)
6995 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6996 * to verify sane_behavior flag on each mount attempt.
6998 static void mem_cgroup_bind(struct cgroup *root)
7001 * use_hierarchy is forced with sane_behavior. cgroup core
7002 * guarantees that @root doesn't have any children, so turning it
7003 * on for the root memcg is enough.
7005 if (cgroup_sane_behavior(root))
7006 mem_cgroup_from_cont(root)->use_hierarchy = true;
7009 struct cgroup_subsys mem_cgroup_subsys = {
7011 .subsys_id = mem_cgroup_subsys_id,
7012 .css_alloc = mem_cgroup_css_alloc,
7013 .css_online = mem_cgroup_css_online,
7014 .css_offline = mem_cgroup_css_offline,
7015 .css_free = mem_cgroup_css_free,
7016 .can_attach = mem_cgroup_can_attach,
7017 .cancel_attach = mem_cgroup_cancel_attach,
7018 .attach = mem_cgroup_move_task,
7019 .bind = mem_cgroup_bind,
7020 .base_cftypes = mem_cgroup_files,
7025 #ifdef CONFIG_MEMCG_SWAP
7026 static int __init enable_swap_account(char *s)
7028 /* consider enabled if no parameter or 1 is given */
7029 if (!strcmp(s, "1"))
7030 really_do_swap_account = 1;
7031 else if (!strcmp(s, "0"))
7032 really_do_swap_account = 0;
7035 __setup("swapaccount=", enable_swap_account);
7037 static void __init memsw_file_init(void)
7039 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7042 static void __init enable_swap_cgroup(void)
7044 if (!mem_cgroup_disabled() && really_do_swap_account) {
7045 do_swap_account = 1;
7051 static void __init enable_swap_cgroup(void)
7057 * subsys_initcall() for memory controller.
7059 * Some parts like hotcpu_notifier() have to be initialized from this context
7060 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7061 * everything that doesn't depend on a specific mem_cgroup structure should
7062 * be initialized from here.
7064 static int __init mem_cgroup_init(void)
7066 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7067 enable_swap_cgroup();
7068 mem_cgroup_soft_limit_tree_init();
7072 subsys_initcall(mem_cgroup_init);