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)
409 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
410 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
413 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
415 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
416 &memcg->kmem_account_flags);
420 /* Stuffs for move charges at task migration. */
422 * Types of charges to be moved. "move_charge_at_immitgrate" and
423 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
426 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
427 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
431 /* "mc" and its members are protected by cgroup_mutex */
432 static struct move_charge_struct {
433 spinlock_t lock; /* for from, to */
434 struct mem_cgroup *from;
435 struct mem_cgroup *to;
436 unsigned long immigrate_flags;
437 unsigned long precharge;
438 unsigned long moved_charge;
439 unsigned long moved_swap;
440 struct task_struct *moving_task; /* a task moving charges */
441 wait_queue_head_t waitq; /* a waitq for other context */
443 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
444 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
447 static bool move_anon(void)
449 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
452 static bool move_file(void)
454 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
458 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
459 * limit reclaim to prevent infinite loops, if they ever occur.
461 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
462 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
465 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
466 MEM_CGROUP_CHARGE_TYPE_ANON,
467 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
468 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
472 /* for encoding cft->private value on file */
480 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
481 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
482 #define MEMFILE_ATTR(val) ((val) & 0xffff)
483 /* Used for OOM nofiier */
484 #define OOM_CONTROL (0)
487 * Reclaim flags for mem_cgroup_hierarchical_reclaim
489 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
490 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
491 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
492 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
495 * The memcg_create_mutex will be held whenever a new cgroup is created.
496 * As a consequence, any change that needs to protect against new child cgroups
497 * appearing has to hold it as well.
499 static DEFINE_MUTEX(memcg_create_mutex);
501 static void mem_cgroup_get(struct mem_cgroup *memcg);
502 static void mem_cgroup_put(struct mem_cgroup *memcg);
505 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
507 return container_of(s, struct mem_cgroup, css);
510 /* Some nice accessors for the vmpressure. */
511 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
514 memcg = root_mem_cgroup;
515 return &memcg->vmpressure;
518 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
520 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
523 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
525 return &mem_cgroup_from_css(css)->vmpressure;
528 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
530 return (memcg == root_mem_cgroup);
533 /* Writing them here to avoid exposing memcg's inner layout */
534 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
536 void sock_update_memcg(struct sock *sk)
538 if (mem_cgroup_sockets_enabled) {
539 struct mem_cgroup *memcg;
540 struct cg_proto *cg_proto;
542 BUG_ON(!sk->sk_prot->proto_cgroup);
544 /* Socket cloning can throw us here with sk_cgrp already
545 * filled. It won't however, necessarily happen from
546 * process context. So the test for root memcg given
547 * the current task's memcg won't help us in this case.
549 * Respecting the original socket's memcg is a better
550 * decision in this case.
553 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
554 mem_cgroup_get(sk->sk_cgrp->memcg);
559 memcg = mem_cgroup_from_task(current);
560 cg_proto = sk->sk_prot->proto_cgroup(memcg);
561 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
562 mem_cgroup_get(memcg);
563 sk->sk_cgrp = cg_proto;
568 EXPORT_SYMBOL(sock_update_memcg);
570 void sock_release_memcg(struct sock *sk)
572 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
573 struct mem_cgroup *memcg;
574 WARN_ON(!sk->sk_cgrp->memcg);
575 memcg = sk->sk_cgrp->memcg;
576 mem_cgroup_put(memcg);
580 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
582 if (!memcg || mem_cgroup_is_root(memcg))
585 return &memcg->tcp_mem.cg_proto;
587 EXPORT_SYMBOL(tcp_proto_cgroup);
589 static void disarm_sock_keys(struct mem_cgroup *memcg)
591 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
593 static_key_slow_dec(&memcg_socket_limit_enabled);
596 static void disarm_sock_keys(struct mem_cgroup *memcg)
601 #ifdef CONFIG_MEMCG_KMEM
603 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
604 * There are two main reasons for not using the css_id for this:
605 * 1) this works better in sparse environments, where we have a lot of memcgs,
606 * but only a few kmem-limited. Or also, if we have, for instance, 200
607 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
608 * 200 entry array for that.
610 * 2) In order not to violate the cgroup API, we would like to do all memory
611 * allocation in ->create(). At that point, we haven't yet allocated the
612 * css_id. Having a separate index prevents us from messing with the cgroup
615 * The current size of the caches array is stored in
616 * memcg_limited_groups_array_size. It will double each time we have to
619 static DEFINE_IDA(kmem_limited_groups);
620 int memcg_limited_groups_array_size;
623 * MIN_SIZE is different than 1, because we would like to avoid going through
624 * the alloc/free process all the time. In a small machine, 4 kmem-limited
625 * cgroups is a reasonable guess. In the future, it could be a parameter or
626 * tunable, but that is strictly not necessary.
628 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
629 * this constant directly from cgroup, but it is understandable that this is
630 * better kept as an internal representation in cgroup.c. In any case, the
631 * css_id space is not getting any smaller, and we don't have to necessarily
632 * increase ours as well if it increases.
634 #define MEMCG_CACHES_MIN_SIZE 4
635 #define MEMCG_CACHES_MAX_SIZE 65535
638 * A lot of the calls to the cache allocation functions are expected to be
639 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
640 * conditional to this static branch, we'll have to allow modules that does
641 * kmem_cache_alloc and the such to see this symbol as well
643 struct static_key memcg_kmem_enabled_key;
644 EXPORT_SYMBOL(memcg_kmem_enabled_key);
646 static void disarm_kmem_keys(struct mem_cgroup *memcg)
648 if (memcg_kmem_is_active(memcg)) {
649 static_key_slow_dec(&memcg_kmem_enabled_key);
650 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
653 * This check can't live in kmem destruction function,
654 * since the charges will outlive the cgroup
656 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
659 static void disarm_kmem_keys(struct mem_cgroup *memcg)
662 #endif /* CONFIG_MEMCG_KMEM */
664 static void disarm_static_keys(struct mem_cgroup *memcg)
666 disarm_sock_keys(memcg);
667 disarm_kmem_keys(memcg);
670 static void drain_all_stock_async(struct mem_cgroup *memcg);
672 static struct mem_cgroup_per_zone *
673 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
675 VM_BUG_ON((unsigned)nid >= nr_node_ids);
676 return &memcg->nodeinfo[nid]->zoneinfo[zid];
679 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
684 static struct mem_cgroup_per_zone *
685 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
687 int nid = page_to_nid(page);
688 int zid = page_zonenum(page);
690 return mem_cgroup_zoneinfo(memcg, nid, zid);
693 static struct mem_cgroup_tree_per_zone *
694 soft_limit_tree_node_zone(int nid, int zid)
696 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
699 static struct mem_cgroup_tree_per_zone *
700 soft_limit_tree_from_page(struct page *page)
702 int nid = page_to_nid(page);
703 int zid = page_zonenum(page);
705 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
709 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
710 struct mem_cgroup_per_zone *mz,
711 struct mem_cgroup_tree_per_zone *mctz,
712 unsigned long long new_usage_in_excess)
714 struct rb_node **p = &mctz->rb_root.rb_node;
715 struct rb_node *parent = NULL;
716 struct mem_cgroup_per_zone *mz_node;
721 mz->usage_in_excess = new_usage_in_excess;
722 if (!mz->usage_in_excess)
726 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
728 if (mz->usage_in_excess < mz_node->usage_in_excess)
731 * We can't avoid mem cgroups that are over their soft
732 * limit by the same amount
734 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
737 rb_link_node(&mz->tree_node, parent, p);
738 rb_insert_color(&mz->tree_node, &mctz->rb_root);
743 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
744 struct mem_cgroup_per_zone *mz,
745 struct mem_cgroup_tree_per_zone *mctz)
749 rb_erase(&mz->tree_node, &mctz->rb_root);
754 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
755 struct mem_cgroup_per_zone *mz,
756 struct mem_cgroup_tree_per_zone *mctz)
758 spin_lock(&mctz->lock);
759 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
760 spin_unlock(&mctz->lock);
764 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
766 unsigned long long excess;
767 struct mem_cgroup_per_zone *mz;
768 struct mem_cgroup_tree_per_zone *mctz;
769 int nid = page_to_nid(page);
770 int zid = page_zonenum(page);
771 mctz = soft_limit_tree_from_page(page);
774 * Necessary to update all ancestors when hierarchy is used.
775 * because their event counter is not touched.
777 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
778 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
779 excess = res_counter_soft_limit_excess(&memcg->res);
781 * We have to update the tree if mz is on RB-tree or
782 * mem is over its softlimit.
784 if (excess || mz->on_tree) {
785 spin_lock(&mctz->lock);
786 /* if on-tree, remove it */
788 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
790 * Insert again. mz->usage_in_excess will be updated.
791 * If excess is 0, no tree ops.
793 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
794 spin_unlock(&mctz->lock);
799 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
802 struct mem_cgroup_per_zone *mz;
803 struct mem_cgroup_tree_per_zone *mctz;
805 for_each_node(node) {
806 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
807 mz = mem_cgroup_zoneinfo(memcg, node, zone);
808 mctz = soft_limit_tree_node_zone(node, zone);
809 mem_cgroup_remove_exceeded(memcg, mz, mctz);
814 static struct mem_cgroup_per_zone *
815 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
817 struct rb_node *rightmost = NULL;
818 struct mem_cgroup_per_zone *mz;
822 rightmost = rb_last(&mctz->rb_root);
824 goto done; /* Nothing to reclaim from */
826 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
828 * Remove the node now but someone else can add it back,
829 * we will to add it back at the end of reclaim to its correct
830 * position in the tree.
832 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
833 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
834 !css_tryget(&mz->memcg->css))
840 static struct mem_cgroup_per_zone *
841 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
843 struct mem_cgroup_per_zone *mz;
845 spin_lock(&mctz->lock);
846 mz = __mem_cgroup_largest_soft_limit_node(mctz);
847 spin_unlock(&mctz->lock);
852 * Implementation Note: reading percpu statistics for memcg.
854 * Both of vmstat[] and percpu_counter has threshold and do periodic
855 * synchronization to implement "quick" read. There are trade-off between
856 * reading cost and precision of value. Then, we may have a chance to implement
857 * a periodic synchronizion of counter in memcg's counter.
859 * But this _read() function is used for user interface now. The user accounts
860 * memory usage by memory cgroup and he _always_ requires exact value because
861 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
862 * have to visit all online cpus and make sum. So, for now, unnecessary
863 * synchronization is not implemented. (just implemented for cpu hotplug)
865 * If there are kernel internal actions which can make use of some not-exact
866 * value, and reading all cpu value can be performance bottleneck in some
867 * common workload, threashold and synchonization as vmstat[] should be
870 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
871 enum mem_cgroup_stat_index idx)
877 for_each_online_cpu(cpu)
878 val += per_cpu(memcg->stat->count[idx], cpu);
879 #ifdef CONFIG_HOTPLUG_CPU
880 spin_lock(&memcg->pcp_counter_lock);
881 val += memcg->nocpu_base.count[idx];
882 spin_unlock(&memcg->pcp_counter_lock);
888 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
891 int val = (charge) ? 1 : -1;
892 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
895 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
896 enum mem_cgroup_events_index idx)
898 unsigned long val = 0;
901 for_each_online_cpu(cpu)
902 val += per_cpu(memcg->stat->events[idx], cpu);
903 #ifdef CONFIG_HOTPLUG_CPU
904 spin_lock(&memcg->pcp_counter_lock);
905 val += memcg->nocpu_base.events[idx];
906 spin_unlock(&memcg->pcp_counter_lock);
911 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
913 bool anon, int nr_pages)
918 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
919 * counted as CACHE even if it's on ANON LRU.
922 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
925 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
928 if (PageTransHuge(page))
929 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
932 /* pagein of a big page is an event. So, ignore page size */
934 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
936 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
937 nr_pages = -nr_pages; /* for event */
940 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
946 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
948 struct mem_cgroup_per_zone *mz;
950 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
951 return mz->lru_size[lru];
955 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
956 unsigned int lru_mask)
958 struct mem_cgroup_per_zone *mz;
960 unsigned long ret = 0;
962 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
965 if (BIT(lru) & lru_mask)
966 ret += mz->lru_size[lru];
972 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
973 int nid, unsigned int lru_mask)
978 for (zid = 0; zid < MAX_NR_ZONES; zid++)
979 total += mem_cgroup_zone_nr_lru_pages(memcg,
985 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
986 unsigned int lru_mask)
991 for_each_node_state(nid, N_MEMORY)
992 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
996 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
997 enum mem_cgroup_events_target target)
999 unsigned long val, next;
1001 val = __this_cpu_read(memcg->stat->nr_page_events);
1002 next = __this_cpu_read(memcg->stat->targets[target]);
1003 /* from time_after() in jiffies.h */
1004 if ((long)next - (long)val < 0) {
1006 case MEM_CGROUP_TARGET_THRESH:
1007 next = val + THRESHOLDS_EVENTS_TARGET;
1009 case MEM_CGROUP_TARGET_SOFTLIMIT:
1010 next = val + SOFTLIMIT_EVENTS_TARGET;
1012 case MEM_CGROUP_TARGET_NUMAINFO:
1013 next = val + NUMAINFO_EVENTS_TARGET;
1018 __this_cpu_write(memcg->stat->targets[target], next);
1025 * Check events in order.
1028 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1031 /* threshold event is triggered in finer grain than soft limit */
1032 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1033 MEM_CGROUP_TARGET_THRESH))) {
1035 bool do_numainfo __maybe_unused;
1037 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1038 MEM_CGROUP_TARGET_SOFTLIMIT);
1039 #if MAX_NUMNODES > 1
1040 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1041 MEM_CGROUP_TARGET_NUMAINFO);
1045 mem_cgroup_threshold(memcg);
1046 if (unlikely(do_softlimit))
1047 mem_cgroup_update_tree(memcg, page);
1048 #if MAX_NUMNODES > 1
1049 if (unlikely(do_numainfo))
1050 atomic_inc(&memcg->numainfo_events);
1056 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1058 return mem_cgroup_from_css(
1059 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1062 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1065 * mm_update_next_owner() may clear mm->owner to NULL
1066 * if it races with swapoff, page migration, etc.
1067 * So this can be called with p == NULL.
1072 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1075 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1077 struct mem_cgroup *memcg = NULL;
1082 * Because we have no locks, mm->owner's may be being moved to other
1083 * cgroup. We use css_tryget() here even if this looks
1084 * pessimistic (rather than adding locks here).
1088 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1089 if (unlikely(!memcg))
1091 } while (!css_tryget(&memcg->css));
1097 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1098 * ref. count) or NULL if the whole root's subtree has been visited.
1100 * helper function to be used by mem_cgroup_iter
1102 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1103 struct mem_cgroup *last_visited)
1105 struct cgroup *prev_cgroup, *next_cgroup;
1108 * Root is not visited by cgroup iterators so it needs an
1114 prev_cgroup = (last_visited == root) ? NULL
1115 : last_visited->css.cgroup;
1117 next_cgroup = cgroup_next_descendant_pre(
1118 prev_cgroup, root->css.cgroup);
1121 * Even if we found a group we have to make sure it is
1122 * alive. css && !memcg means that the groups should be
1123 * skipped and we should continue the tree walk.
1124 * last_visited css is safe to use because it is
1125 * protected by css_get and the tree walk is rcu safe.
1128 struct mem_cgroup *mem = mem_cgroup_from_cont(
1130 if (css_tryget(&mem->css))
1133 prev_cgroup = next_cgroup;
1141 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1144 * When a group in the hierarchy below root is destroyed, the
1145 * hierarchy iterator can no longer be trusted since it might
1146 * have pointed to the destroyed group. Invalidate it.
1148 atomic_inc(&root->dead_count);
1151 static struct mem_cgroup *
1152 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1153 struct mem_cgroup *root,
1156 struct mem_cgroup *position = NULL;
1158 * A cgroup destruction happens in two stages: offlining and
1159 * release. They are separated by a RCU grace period.
1161 * If the iterator is valid, we may still race with an
1162 * offlining. The RCU lock ensures the object won't be
1163 * released, tryget will fail if we lost the race.
1165 *sequence = atomic_read(&root->dead_count);
1166 if (iter->last_dead_count == *sequence) {
1168 position = iter->last_visited;
1169 if (position && !css_tryget(&position->css))
1175 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1176 struct mem_cgroup *last_visited,
1177 struct mem_cgroup *new_position,
1181 css_put(&last_visited->css);
1183 * We store the sequence count from the time @last_visited was
1184 * loaded successfully instead of rereading it here so that we
1185 * don't lose destruction events in between. We could have
1186 * raced with the destruction of @new_position after all.
1188 iter->last_visited = new_position;
1190 iter->last_dead_count = sequence;
1194 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1195 * @root: hierarchy root
1196 * @prev: previously returned memcg, NULL on first invocation
1197 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1199 * Returns references to children of the hierarchy below @root, or
1200 * @root itself, or %NULL after a full round-trip.
1202 * Caller must pass the return value in @prev on subsequent
1203 * invocations for reference counting, or use mem_cgroup_iter_break()
1204 * to cancel a hierarchy walk before the round-trip is complete.
1206 * Reclaimers can specify a zone and a priority level in @reclaim to
1207 * divide up the memcgs in the hierarchy among all concurrent
1208 * reclaimers operating on the same zone and priority.
1210 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1211 struct mem_cgroup *prev,
1212 struct mem_cgroup_reclaim_cookie *reclaim)
1214 struct mem_cgroup *memcg = NULL;
1215 struct mem_cgroup *last_visited = NULL;
1217 if (mem_cgroup_disabled())
1221 root = root_mem_cgroup;
1223 if (prev && !reclaim)
1224 last_visited = prev;
1226 if (!root->use_hierarchy && root != root_mem_cgroup) {
1234 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1235 int uninitialized_var(seq);
1238 int nid = zone_to_nid(reclaim->zone);
1239 int zid = zone_idx(reclaim->zone);
1240 struct mem_cgroup_per_zone *mz;
1242 mz = mem_cgroup_zoneinfo(root, nid, zid);
1243 iter = &mz->reclaim_iter[reclaim->priority];
1244 if (prev && reclaim->generation != iter->generation) {
1245 iter->last_visited = NULL;
1249 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1252 memcg = __mem_cgroup_iter_next(root, last_visited);
1255 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1259 else if (!prev && memcg)
1260 reclaim->generation = iter->generation;
1269 if (prev && prev != root)
1270 css_put(&prev->css);
1276 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1277 * @root: hierarchy root
1278 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1280 void mem_cgroup_iter_break(struct mem_cgroup *root,
1281 struct mem_cgroup *prev)
1284 root = root_mem_cgroup;
1285 if (prev && prev != root)
1286 css_put(&prev->css);
1290 * Iteration constructs for visiting all cgroups (under a tree). If
1291 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1292 * be used for reference counting.
1294 #define for_each_mem_cgroup_tree(iter, root) \
1295 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1297 iter = mem_cgroup_iter(root, iter, NULL))
1299 #define for_each_mem_cgroup(iter) \
1300 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1302 iter = mem_cgroup_iter(NULL, iter, NULL))
1304 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1306 struct mem_cgroup *memcg;
1309 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1310 if (unlikely(!memcg))
1315 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1318 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1326 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1329 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1330 * @zone: zone of the wanted lruvec
1331 * @memcg: memcg of the wanted lruvec
1333 * Returns the lru list vector holding pages for the given @zone and
1334 * @mem. This can be the global zone lruvec, if the memory controller
1337 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1338 struct mem_cgroup *memcg)
1340 struct mem_cgroup_per_zone *mz;
1341 struct lruvec *lruvec;
1343 if (mem_cgroup_disabled()) {
1344 lruvec = &zone->lruvec;
1348 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1349 lruvec = &mz->lruvec;
1352 * Since a node can be onlined after the mem_cgroup was created,
1353 * we have to be prepared to initialize lruvec->zone here;
1354 * and if offlined then reonlined, we need to reinitialize it.
1356 if (unlikely(lruvec->zone != zone))
1357 lruvec->zone = zone;
1362 * Following LRU functions are allowed to be used without PCG_LOCK.
1363 * Operations are called by routine of global LRU independently from memcg.
1364 * What we have to take care of here is validness of pc->mem_cgroup.
1366 * Changes to pc->mem_cgroup happens when
1369 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1370 * It is added to LRU before charge.
1371 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1372 * When moving account, the page is not on LRU. It's isolated.
1376 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1378 * @zone: zone of the page
1380 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1382 struct mem_cgroup_per_zone *mz;
1383 struct mem_cgroup *memcg;
1384 struct page_cgroup *pc;
1385 struct lruvec *lruvec;
1387 if (mem_cgroup_disabled()) {
1388 lruvec = &zone->lruvec;
1392 pc = lookup_page_cgroup(page);
1393 memcg = pc->mem_cgroup;
1396 * Surreptitiously switch any uncharged offlist page to root:
1397 * an uncharged page off lru does nothing to secure
1398 * its former mem_cgroup from sudden removal.
1400 * Our caller holds lru_lock, and PageCgroupUsed is updated
1401 * under page_cgroup lock: between them, they make all uses
1402 * of pc->mem_cgroup safe.
1404 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1405 pc->mem_cgroup = memcg = root_mem_cgroup;
1407 mz = page_cgroup_zoneinfo(memcg, page);
1408 lruvec = &mz->lruvec;
1411 * Since a node can be onlined after the mem_cgroup was created,
1412 * we have to be prepared to initialize lruvec->zone here;
1413 * and if offlined then reonlined, we need to reinitialize it.
1415 if (unlikely(lruvec->zone != zone))
1416 lruvec->zone = zone;
1421 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1422 * @lruvec: mem_cgroup per zone lru vector
1423 * @lru: index of lru list the page is sitting on
1424 * @nr_pages: positive when adding or negative when removing
1426 * This function must be called when a page is added to or removed from an
1429 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1432 struct mem_cgroup_per_zone *mz;
1433 unsigned long *lru_size;
1435 if (mem_cgroup_disabled())
1438 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1439 lru_size = mz->lru_size + lru;
1440 *lru_size += nr_pages;
1441 VM_BUG_ON((long)(*lru_size) < 0);
1445 * Checks whether given mem is same or in the root_mem_cgroup's
1448 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1449 struct mem_cgroup *memcg)
1451 if (root_memcg == memcg)
1453 if (!root_memcg->use_hierarchy || !memcg)
1455 return css_is_ancestor(&memcg->css, &root_memcg->css);
1458 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1459 struct mem_cgroup *memcg)
1464 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1469 bool task_in_mem_cgroup(struct task_struct *task,
1470 const struct mem_cgroup *memcg)
1472 struct mem_cgroup *curr = NULL;
1473 struct task_struct *p;
1476 p = find_lock_task_mm(task);
1478 curr = try_get_mem_cgroup_from_mm(p->mm);
1482 * All threads may have already detached their mm's, but the oom
1483 * killer still needs to detect if they have already been oom
1484 * killed to prevent needlessly killing additional tasks.
1487 curr = mem_cgroup_from_task(task);
1489 css_get(&curr->css);
1495 * We should check use_hierarchy of "memcg" not "curr". Because checking
1496 * use_hierarchy of "curr" here make this function true if hierarchy is
1497 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1498 * hierarchy(even if use_hierarchy is disabled in "memcg").
1500 ret = mem_cgroup_same_or_subtree(memcg, curr);
1501 css_put(&curr->css);
1505 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1507 unsigned long inactive_ratio;
1508 unsigned long inactive;
1509 unsigned long active;
1512 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1513 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1515 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1517 inactive_ratio = int_sqrt(10 * gb);
1521 return inactive * inactive_ratio < active;
1524 #define mem_cgroup_from_res_counter(counter, member) \
1525 container_of(counter, struct mem_cgroup, member)
1528 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1529 * @memcg: the memory cgroup
1531 * Returns the maximum amount of memory @mem can be charged with, in
1534 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1536 unsigned long long margin;
1538 margin = res_counter_margin(&memcg->res);
1539 if (do_swap_account)
1540 margin = min(margin, res_counter_margin(&memcg->memsw));
1541 return margin >> PAGE_SHIFT;
1544 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1546 struct cgroup *cgrp = memcg->css.cgroup;
1549 if (cgrp->parent == NULL)
1550 return vm_swappiness;
1552 return memcg->swappiness;
1556 * memcg->moving_account is used for checking possibility that some thread is
1557 * calling move_account(). When a thread on CPU-A starts moving pages under
1558 * a memcg, other threads should check memcg->moving_account under
1559 * rcu_read_lock(), like this:
1563 * memcg->moving_account+1 if (memcg->mocing_account)
1565 * synchronize_rcu() update something.
1570 /* for quick checking without looking up memcg */
1571 atomic_t memcg_moving __read_mostly;
1573 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1575 atomic_inc(&memcg_moving);
1576 atomic_inc(&memcg->moving_account);
1580 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1583 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1584 * We check NULL in callee rather than caller.
1587 atomic_dec(&memcg_moving);
1588 atomic_dec(&memcg->moving_account);
1593 * 2 routines for checking "mem" is under move_account() or not.
1595 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1596 * is used for avoiding races in accounting. If true,
1597 * pc->mem_cgroup may be overwritten.
1599 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1600 * under hierarchy of moving cgroups. This is for
1601 * waiting at hith-memory prressure caused by "move".
1604 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1606 VM_BUG_ON(!rcu_read_lock_held());
1607 return atomic_read(&memcg->moving_account) > 0;
1610 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1612 struct mem_cgroup *from;
1613 struct mem_cgroup *to;
1616 * Unlike task_move routines, we access mc.to, mc.from not under
1617 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1619 spin_lock(&mc.lock);
1625 ret = mem_cgroup_same_or_subtree(memcg, from)
1626 || mem_cgroup_same_or_subtree(memcg, to);
1628 spin_unlock(&mc.lock);
1632 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1634 if (mc.moving_task && current != mc.moving_task) {
1635 if (mem_cgroup_under_move(memcg)) {
1637 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1638 /* moving charge context might have finished. */
1641 finish_wait(&mc.waitq, &wait);
1649 * Take this lock when
1650 * - a code tries to modify page's memcg while it's USED.
1651 * - a code tries to modify page state accounting in a memcg.
1652 * see mem_cgroup_stolen(), too.
1654 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1655 unsigned long *flags)
1657 spin_lock_irqsave(&memcg->move_lock, *flags);
1660 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1661 unsigned long *flags)
1663 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1666 #define K(x) ((x) << (PAGE_SHIFT-10))
1668 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1669 * @memcg: The memory cgroup that went over limit
1670 * @p: Task that is going to be killed
1672 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1675 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1677 struct cgroup *task_cgrp;
1678 struct cgroup *mem_cgrp;
1680 * Need a buffer in BSS, can't rely on allocations. The code relies
1681 * on the assumption that OOM is serialized for memory controller.
1682 * If this assumption is broken, revisit this code.
1684 static char memcg_name[PATH_MAX];
1686 struct mem_cgroup *iter;
1694 mem_cgrp = memcg->css.cgroup;
1695 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1697 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1700 * Unfortunately, we are unable to convert to a useful name
1701 * But we'll still print out the usage information
1708 pr_info("Task in %s killed", memcg_name);
1711 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1719 * Continues from above, so we don't need an KERN_ level
1721 pr_cont(" as a result of limit of %s\n", memcg_name);
1724 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1725 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1726 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1727 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1728 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1729 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1730 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1731 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1732 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1733 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1734 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1735 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1737 for_each_mem_cgroup_tree(iter, memcg) {
1738 pr_info("Memory cgroup stats");
1741 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1743 pr_cont(" for %s", memcg_name);
1747 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1748 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1750 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1751 K(mem_cgroup_read_stat(iter, i)));
1754 for (i = 0; i < NR_LRU_LISTS; i++)
1755 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1756 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1763 * This function returns the number of memcg under hierarchy tree. Returns
1764 * 1(self count) if no children.
1766 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1769 struct mem_cgroup *iter;
1771 for_each_mem_cgroup_tree(iter, memcg)
1777 * Return the memory (and swap, if configured) limit for a memcg.
1779 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1783 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1786 * Do not consider swap space if we cannot swap due to swappiness
1788 if (mem_cgroup_swappiness(memcg)) {
1791 limit += total_swap_pages << PAGE_SHIFT;
1792 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1795 * If memsw is finite and limits the amount of swap space
1796 * available to this memcg, return that limit.
1798 limit = min(limit, memsw);
1804 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1807 struct mem_cgroup *iter;
1808 unsigned long chosen_points = 0;
1809 unsigned long totalpages;
1810 unsigned int points = 0;
1811 struct task_struct *chosen = NULL;
1814 * If current has a pending SIGKILL or is exiting, then automatically
1815 * select it. The goal is to allow it to allocate so that it may
1816 * quickly exit and free its memory.
1818 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1819 set_thread_flag(TIF_MEMDIE);
1823 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1824 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1825 for_each_mem_cgroup_tree(iter, memcg) {
1826 struct cgroup *cgroup = iter->css.cgroup;
1827 struct cgroup_iter it;
1828 struct task_struct *task;
1830 cgroup_iter_start(cgroup, &it);
1831 while ((task = cgroup_iter_next(cgroup, &it))) {
1832 switch (oom_scan_process_thread(task, totalpages, NULL,
1834 case OOM_SCAN_SELECT:
1836 put_task_struct(chosen);
1838 chosen_points = ULONG_MAX;
1839 get_task_struct(chosen);
1841 case OOM_SCAN_CONTINUE:
1843 case OOM_SCAN_ABORT:
1844 cgroup_iter_end(cgroup, &it);
1845 mem_cgroup_iter_break(memcg, iter);
1847 put_task_struct(chosen);
1852 points = oom_badness(task, memcg, NULL, totalpages);
1853 if (points > chosen_points) {
1855 put_task_struct(chosen);
1857 chosen_points = points;
1858 get_task_struct(chosen);
1861 cgroup_iter_end(cgroup, &it);
1866 points = chosen_points * 1000 / totalpages;
1867 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1868 NULL, "Memory cgroup out of memory");
1871 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1873 unsigned long flags)
1875 unsigned long total = 0;
1876 bool noswap = false;
1879 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1881 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1884 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1886 drain_all_stock_async(memcg);
1887 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1889 * Allow limit shrinkers, which are triggered directly
1890 * by userspace, to catch signals and stop reclaim
1891 * after minimal progress, regardless of the margin.
1893 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1895 if (mem_cgroup_margin(memcg))
1898 * If nothing was reclaimed after two attempts, there
1899 * may be no reclaimable pages in this hierarchy.
1908 * test_mem_cgroup_node_reclaimable
1909 * @memcg: the target memcg
1910 * @nid: the node ID to be checked.
1911 * @noswap : specify true here if the user wants flle only information.
1913 * This function returns whether the specified memcg contains any
1914 * reclaimable pages on a node. Returns true if there are any reclaimable
1915 * pages in the node.
1917 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1918 int nid, bool noswap)
1920 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1922 if (noswap || !total_swap_pages)
1924 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1929 #if MAX_NUMNODES > 1
1932 * Always updating the nodemask is not very good - even if we have an empty
1933 * list or the wrong list here, we can start from some node and traverse all
1934 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1937 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1941 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1942 * pagein/pageout changes since the last update.
1944 if (!atomic_read(&memcg->numainfo_events))
1946 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1949 /* make a nodemask where this memcg uses memory from */
1950 memcg->scan_nodes = node_states[N_MEMORY];
1952 for_each_node_mask(nid, node_states[N_MEMORY]) {
1954 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1955 node_clear(nid, memcg->scan_nodes);
1958 atomic_set(&memcg->numainfo_events, 0);
1959 atomic_set(&memcg->numainfo_updating, 0);
1963 * Selecting a node where we start reclaim from. Because what we need is just
1964 * reducing usage counter, start from anywhere is O,K. Considering
1965 * memory reclaim from current node, there are pros. and cons.
1967 * Freeing memory from current node means freeing memory from a node which
1968 * we'll use or we've used. So, it may make LRU bad. And if several threads
1969 * hit limits, it will see a contention on a node. But freeing from remote
1970 * node means more costs for memory reclaim because of memory latency.
1972 * Now, we use round-robin. Better algorithm is welcomed.
1974 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1978 mem_cgroup_may_update_nodemask(memcg);
1979 node = memcg->last_scanned_node;
1981 node = next_node(node, memcg->scan_nodes);
1982 if (node == MAX_NUMNODES)
1983 node = first_node(memcg->scan_nodes);
1985 * We call this when we hit limit, not when pages are added to LRU.
1986 * No LRU may hold pages because all pages are UNEVICTABLE or
1987 * memcg is too small and all pages are not on LRU. In that case,
1988 * we use curret node.
1990 if (unlikely(node == MAX_NUMNODES))
1991 node = numa_node_id();
1993 memcg->last_scanned_node = node;
1998 * Check all nodes whether it contains reclaimable pages or not.
1999 * For quick scan, we make use of scan_nodes. This will allow us to skip
2000 * unused nodes. But scan_nodes is lazily updated and may not cotain
2001 * enough new information. We need to do double check.
2003 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2008 * quick check...making use of scan_node.
2009 * We can skip unused nodes.
2011 if (!nodes_empty(memcg->scan_nodes)) {
2012 for (nid = first_node(memcg->scan_nodes);
2014 nid = next_node(nid, memcg->scan_nodes)) {
2016 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2021 * Check rest of nodes.
2023 for_each_node_state(nid, N_MEMORY) {
2024 if (node_isset(nid, memcg->scan_nodes))
2026 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2033 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2038 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2040 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2044 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2047 unsigned long *total_scanned)
2049 struct mem_cgroup *victim = NULL;
2052 unsigned long excess;
2053 unsigned long nr_scanned;
2054 struct mem_cgroup_reclaim_cookie reclaim = {
2059 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2062 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2067 * If we have not been able to reclaim
2068 * anything, it might because there are
2069 * no reclaimable pages under this hierarchy
2074 * We want to do more targeted reclaim.
2075 * excess >> 2 is not to excessive so as to
2076 * reclaim too much, nor too less that we keep
2077 * coming back to reclaim from this cgroup
2079 if (total >= (excess >> 2) ||
2080 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2085 if (!mem_cgroup_reclaimable(victim, false))
2087 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2089 *total_scanned += nr_scanned;
2090 if (!res_counter_soft_limit_excess(&root_memcg->res))
2093 mem_cgroup_iter_break(root_memcg, victim);
2098 * Check OOM-Killer is already running under our hierarchy.
2099 * If someone is running, return false.
2100 * Has to be called with memcg_oom_lock
2102 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
2104 struct mem_cgroup *iter, *failed = NULL;
2106 for_each_mem_cgroup_tree(iter, memcg) {
2107 if (iter->oom_lock) {
2109 * this subtree of our hierarchy is already locked
2110 * so we cannot give a lock.
2113 mem_cgroup_iter_break(memcg, iter);
2116 iter->oom_lock = true;
2123 * OK, we failed to lock the whole subtree so we have to clean up
2124 * what we set up to the failing subtree
2126 for_each_mem_cgroup_tree(iter, memcg) {
2127 if (iter == failed) {
2128 mem_cgroup_iter_break(memcg, iter);
2131 iter->oom_lock = false;
2137 * Has to be called with memcg_oom_lock
2139 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2141 struct mem_cgroup *iter;
2143 for_each_mem_cgroup_tree(iter, memcg)
2144 iter->oom_lock = false;
2148 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2150 struct mem_cgroup *iter;
2152 for_each_mem_cgroup_tree(iter, memcg)
2153 atomic_inc(&iter->under_oom);
2156 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2158 struct mem_cgroup *iter;
2161 * When a new child is created while the hierarchy is under oom,
2162 * mem_cgroup_oom_lock() may not be called. We have to use
2163 * atomic_add_unless() here.
2165 for_each_mem_cgroup_tree(iter, memcg)
2166 atomic_add_unless(&iter->under_oom, -1, 0);
2169 static DEFINE_SPINLOCK(memcg_oom_lock);
2170 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2172 struct oom_wait_info {
2173 struct mem_cgroup *memcg;
2177 static int memcg_oom_wake_function(wait_queue_t *wait,
2178 unsigned mode, int sync, void *arg)
2180 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2181 struct mem_cgroup *oom_wait_memcg;
2182 struct oom_wait_info *oom_wait_info;
2184 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2185 oom_wait_memcg = oom_wait_info->memcg;
2188 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2189 * Then we can use css_is_ancestor without taking care of RCU.
2191 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2192 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2194 return autoremove_wake_function(wait, mode, sync, arg);
2197 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2199 /* for filtering, pass "memcg" as argument. */
2200 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2203 static void memcg_oom_recover(struct mem_cgroup *memcg)
2205 if (memcg && atomic_read(&memcg->under_oom))
2206 memcg_wakeup_oom(memcg);
2210 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2212 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2215 struct oom_wait_info owait;
2216 bool locked, need_to_kill;
2218 owait.memcg = memcg;
2219 owait.wait.flags = 0;
2220 owait.wait.func = memcg_oom_wake_function;
2221 owait.wait.private = current;
2222 INIT_LIST_HEAD(&owait.wait.task_list);
2223 need_to_kill = true;
2224 mem_cgroup_mark_under_oom(memcg);
2226 /* At first, try to OOM lock hierarchy under memcg.*/
2227 spin_lock(&memcg_oom_lock);
2228 locked = mem_cgroup_oom_lock(memcg);
2230 * Even if signal_pending(), we can't quit charge() loop without
2231 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2232 * under OOM is always welcomed, use TASK_KILLABLE here.
2234 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2235 if (!locked || memcg->oom_kill_disable)
2236 need_to_kill = false;
2238 mem_cgroup_oom_notify(memcg);
2239 spin_unlock(&memcg_oom_lock);
2242 finish_wait(&memcg_oom_waitq, &owait.wait);
2243 mem_cgroup_out_of_memory(memcg, mask, order);
2246 finish_wait(&memcg_oom_waitq, &owait.wait);
2248 spin_lock(&memcg_oom_lock);
2250 mem_cgroup_oom_unlock(memcg);
2251 memcg_wakeup_oom(memcg);
2252 spin_unlock(&memcg_oom_lock);
2254 mem_cgroup_unmark_under_oom(memcg);
2256 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2258 /* Give chance to dying process */
2259 schedule_timeout_uninterruptible(1);
2264 * Currently used to update mapped file statistics, but the routine can be
2265 * generalized to update other statistics as well.
2267 * Notes: Race condition
2269 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2270 * it tends to be costly. But considering some conditions, we doesn't need
2271 * to do so _always_.
2273 * Considering "charge", lock_page_cgroup() is not required because all
2274 * file-stat operations happen after a page is attached to radix-tree. There
2275 * are no race with "charge".
2277 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2278 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2279 * if there are race with "uncharge". Statistics itself is properly handled
2282 * Considering "move", this is an only case we see a race. To make the race
2283 * small, we check mm->moving_account and detect there are possibility of race
2284 * If there is, we take a lock.
2287 void __mem_cgroup_begin_update_page_stat(struct page *page,
2288 bool *locked, unsigned long *flags)
2290 struct mem_cgroup *memcg;
2291 struct page_cgroup *pc;
2293 pc = lookup_page_cgroup(page);
2295 memcg = pc->mem_cgroup;
2296 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2299 * If this memory cgroup is not under account moving, we don't
2300 * need to take move_lock_mem_cgroup(). Because we already hold
2301 * rcu_read_lock(), any calls to move_account will be delayed until
2302 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2304 if (!mem_cgroup_stolen(memcg))
2307 move_lock_mem_cgroup(memcg, flags);
2308 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2309 move_unlock_mem_cgroup(memcg, flags);
2315 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2317 struct page_cgroup *pc = lookup_page_cgroup(page);
2320 * It's guaranteed that pc->mem_cgroup never changes while
2321 * lock is held because a routine modifies pc->mem_cgroup
2322 * should take move_lock_mem_cgroup().
2324 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2327 void mem_cgroup_update_page_stat(struct page *page,
2328 enum mem_cgroup_page_stat_item idx, int val)
2330 struct mem_cgroup *memcg;
2331 struct page_cgroup *pc = lookup_page_cgroup(page);
2332 unsigned long uninitialized_var(flags);
2334 if (mem_cgroup_disabled())
2337 memcg = pc->mem_cgroup;
2338 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2342 case MEMCG_NR_FILE_MAPPED:
2343 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2349 this_cpu_add(memcg->stat->count[idx], val);
2353 * size of first charge trial. "32" comes from vmscan.c's magic value.
2354 * TODO: maybe necessary to use big numbers in big irons.
2356 #define CHARGE_BATCH 32U
2357 struct memcg_stock_pcp {
2358 struct mem_cgroup *cached; /* this never be root cgroup */
2359 unsigned int nr_pages;
2360 struct work_struct work;
2361 unsigned long flags;
2362 #define FLUSHING_CACHED_CHARGE 0
2364 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2365 static DEFINE_MUTEX(percpu_charge_mutex);
2368 * consume_stock: Try to consume stocked charge on this cpu.
2369 * @memcg: memcg to consume from.
2370 * @nr_pages: how many pages to charge.
2372 * The charges will only happen if @memcg matches the current cpu's memcg
2373 * stock, and at least @nr_pages are available in that stock. Failure to
2374 * service an allocation will refill the stock.
2376 * returns true if successful, false otherwise.
2378 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2380 struct memcg_stock_pcp *stock;
2383 if (nr_pages > CHARGE_BATCH)
2386 stock = &get_cpu_var(memcg_stock);
2387 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2388 stock->nr_pages -= nr_pages;
2389 else /* need to call res_counter_charge */
2391 put_cpu_var(memcg_stock);
2396 * Returns stocks cached in percpu to res_counter and reset cached information.
2398 static void drain_stock(struct memcg_stock_pcp *stock)
2400 struct mem_cgroup *old = stock->cached;
2402 if (stock->nr_pages) {
2403 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2405 res_counter_uncharge(&old->res, bytes);
2406 if (do_swap_account)
2407 res_counter_uncharge(&old->memsw, bytes);
2408 stock->nr_pages = 0;
2410 stock->cached = NULL;
2414 * This must be called under preempt disabled or must be called by
2415 * a thread which is pinned to local cpu.
2417 static void drain_local_stock(struct work_struct *dummy)
2419 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2421 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2424 static void __init memcg_stock_init(void)
2428 for_each_possible_cpu(cpu) {
2429 struct memcg_stock_pcp *stock =
2430 &per_cpu(memcg_stock, cpu);
2431 INIT_WORK(&stock->work, drain_local_stock);
2436 * Cache charges(val) which is from res_counter, to local per_cpu area.
2437 * This will be consumed by consume_stock() function, later.
2439 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2441 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2443 if (stock->cached != memcg) { /* reset if necessary */
2445 stock->cached = memcg;
2447 stock->nr_pages += nr_pages;
2448 put_cpu_var(memcg_stock);
2452 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2453 * of the hierarchy under it. sync flag says whether we should block
2454 * until the work is done.
2456 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2460 /* Notify other cpus that system-wide "drain" is running */
2463 for_each_online_cpu(cpu) {
2464 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2465 struct mem_cgroup *memcg;
2467 memcg = stock->cached;
2468 if (!memcg || !stock->nr_pages)
2470 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2472 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2474 drain_local_stock(&stock->work);
2476 schedule_work_on(cpu, &stock->work);
2484 for_each_online_cpu(cpu) {
2485 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2486 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2487 flush_work(&stock->work);
2494 * Tries to drain stocked charges in other cpus. This function is asynchronous
2495 * and just put a work per cpu for draining localy on each cpu. Caller can
2496 * expects some charges will be back to res_counter later but cannot wait for
2499 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2502 * If someone calls draining, avoid adding more kworker runs.
2504 if (!mutex_trylock(&percpu_charge_mutex))
2506 drain_all_stock(root_memcg, false);
2507 mutex_unlock(&percpu_charge_mutex);
2510 /* This is a synchronous drain interface. */
2511 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2513 /* called when force_empty is called */
2514 mutex_lock(&percpu_charge_mutex);
2515 drain_all_stock(root_memcg, true);
2516 mutex_unlock(&percpu_charge_mutex);
2520 * This function drains percpu counter value from DEAD cpu and
2521 * move it to local cpu. Note that this function can be preempted.
2523 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2527 spin_lock(&memcg->pcp_counter_lock);
2528 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2529 long x = per_cpu(memcg->stat->count[i], cpu);
2531 per_cpu(memcg->stat->count[i], cpu) = 0;
2532 memcg->nocpu_base.count[i] += x;
2534 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2535 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2537 per_cpu(memcg->stat->events[i], cpu) = 0;
2538 memcg->nocpu_base.events[i] += x;
2540 spin_unlock(&memcg->pcp_counter_lock);
2543 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2544 unsigned long action,
2547 int cpu = (unsigned long)hcpu;
2548 struct memcg_stock_pcp *stock;
2549 struct mem_cgroup *iter;
2551 if (action == CPU_ONLINE)
2554 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2557 for_each_mem_cgroup(iter)
2558 mem_cgroup_drain_pcp_counter(iter, cpu);
2560 stock = &per_cpu(memcg_stock, cpu);
2566 /* See __mem_cgroup_try_charge() for details */
2568 CHARGE_OK, /* success */
2569 CHARGE_RETRY, /* need to retry but retry is not bad */
2570 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2571 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2572 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2575 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2576 unsigned int nr_pages, unsigned int min_pages,
2579 unsigned long csize = nr_pages * PAGE_SIZE;
2580 struct mem_cgroup *mem_over_limit;
2581 struct res_counter *fail_res;
2582 unsigned long flags = 0;
2585 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2588 if (!do_swap_account)
2590 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2594 res_counter_uncharge(&memcg->res, csize);
2595 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2596 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2598 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2600 * Never reclaim on behalf of optional batching, retry with a
2601 * single page instead.
2603 if (nr_pages > min_pages)
2604 return CHARGE_RETRY;
2606 if (!(gfp_mask & __GFP_WAIT))
2607 return CHARGE_WOULDBLOCK;
2609 if (gfp_mask & __GFP_NORETRY)
2610 return CHARGE_NOMEM;
2612 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2613 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2614 return CHARGE_RETRY;
2616 * Even though the limit is exceeded at this point, reclaim
2617 * may have been able to free some pages. Retry the charge
2618 * before killing the task.
2620 * Only for regular pages, though: huge pages are rather
2621 * unlikely to succeed so close to the limit, and we fall back
2622 * to regular pages anyway in case of failure.
2624 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2625 return CHARGE_RETRY;
2628 * At task move, charge accounts can be doubly counted. So, it's
2629 * better to wait until the end of task_move if something is going on.
2631 if (mem_cgroup_wait_acct_move(mem_over_limit))
2632 return CHARGE_RETRY;
2634 /* If we don't need to call oom-killer at el, return immediately */
2636 return CHARGE_NOMEM;
2638 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2639 return CHARGE_OOM_DIE;
2641 return CHARGE_RETRY;
2645 * __mem_cgroup_try_charge() does
2646 * 1. detect memcg to be charged against from passed *mm and *ptr,
2647 * 2. update res_counter
2648 * 3. call memory reclaim if necessary.
2650 * In some special case, if the task is fatal, fatal_signal_pending() or
2651 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2652 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2653 * as possible without any hazards. 2: all pages should have a valid
2654 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2655 * pointer, that is treated as a charge to root_mem_cgroup.
2657 * So __mem_cgroup_try_charge() will return
2658 * 0 ... on success, filling *ptr with a valid memcg pointer.
2659 * -ENOMEM ... charge failure because of resource limits.
2660 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2662 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2663 * the oom-killer can be invoked.
2665 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2667 unsigned int nr_pages,
2668 struct mem_cgroup **ptr,
2671 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2672 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2673 struct mem_cgroup *memcg = NULL;
2677 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2678 * in system level. So, allow to go ahead dying process in addition to
2681 if (unlikely(test_thread_flag(TIF_MEMDIE)
2682 || fatal_signal_pending(current)))
2686 * We always charge the cgroup the mm_struct belongs to.
2687 * The mm_struct's mem_cgroup changes on task migration if the
2688 * thread group leader migrates. It's possible that mm is not
2689 * set, if so charge the root memcg (happens for pagecache usage).
2692 *ptr = root_mem_cgroup;
2694 if (*ptr) { /* css should be a valid one */
2696 if (mem_cgroup_is_root(memcg))
2698 if (consume_stock(memcg, nr_pages))
2700 css_get(&memcg->css);
2702 struct task_struct *p;
2705 p = rcu_dereference(mm->owner);
2707 * Because we don't have task_lock(), "p" can exit.
2708 * In that case, "memcg" can point to root or p can be NULL with
2709 * race with swapoff. Then, we have small risk of mis-accouning.
2710 * But such kind of mis-account by race always happens because
2711 * we don't have cgroup_mutex(). It's overkill and we allo that
2713 * (*) swapoff at el will charge against mm-struct not against
2714 * task-struct. So, mm->owner can be NULL.
2716 memcg = mem_cgroup_from_task(p);
2718 memcg = root_mem_cgroup;
2719 if (mem_cgroup_is_root(memcg)) {
2723 if (consume_stock(memcg, nr_pages)) {
2725 * It seems dagerous to access memcg without css_get().
2726 * But considering how consume_stok works, it's not
2727 * necessary. If consume_stock success, some charges
2728 * from this memcg are cached on this cpu. So, we
2729 * don't need to call css_get()/css_tryget() before
2730 * calling consume_stock().
2735 /* after here, we may be blocked. we need to get refcnt */
2736 if (!css_tryget(&memcg->css)) {
2746 /* If killed, bypass charge */
2747 if (fatal_signal_pending(current)) {
2748 css_put(&memcg->css);
2753 if (oom && !nr_oom_retries) {
2755 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2758 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2763 case CHARGE_RETRY: /* not in OOM situation but retry */
2765 css_put(&memcg->css);
2768 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2769 css_put(&memcg->css);
2771 case CHARGE_NOMEM: /* OOM routine works */
2773 css_put(&memcg->css);
2776 /* If oom, we never return -ENOMEM */
2779 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2780 css_put(&memcg->css);
2783 } while (ret != CHARGE_OK);
2785 if (batch > nr_pages)
2786 refill_stock(memcg, batch - nr_pages);
2787 css_put(&memcg->css);
2795 *ptr = root_mem_cgroup;
2800 * Somemtimes we have to undo a charge we got by try_charge().
2801 * This function is for that and do uncharge, put css's refcnt.
2802 * gotten by try_charge().
2804 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2805 unsigned int nr_pages)
2807 if (!mem_cgroup_is_root(memcg)) {
2808 unsigned long bytes = nr_pages * PAGE_SIZE;
2810 res_counter_uncharge(&memcg->res, bytes);
2811 if (do_swap_account)
2812 res_counter_uncharge(&memcg->memsw, bytes);
2817 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2818 * This is useful when moving usage to parent cgroup.
2820 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2821 unsigned int nr_pages)
2823 unsigned long bytes = nr_pages * PAGE_SIZE;
2825 if (mem_cgroup_is_root(memcg))
2828 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2829 if (do_swap_account)
2830 res_counter_uncharge_until(&memcg->memsw,
2831 memcg->memsw.parent, bytes);
2835 * A helper function to get mem_cgroup from ID. must be called under
2836 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2837 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2838 * called against removed memcg.)
2840 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2842 struct cgroup_subsys_state *css;
2844 /* ID 0 is unused ID */
2847 css = css_lookup(&mem_cgroup_subsys, id);
2850 return mem_cgroup_from_css(css);
2853 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2855 struct mem_cgroup *memcg = NULL;
2856 struct page_cgroup *pc;
2860 VM_BUG_ON(!PageLocked(page));
2862 pc = lookup_page_cgroup(page);
2863 lock_page_cgroup(pc);
2864 if (PageCgroupUsed(pc)) {
2865 memcg = pc->mem_cgroup;
2866 if (memcg && !css_tryget(&memcg->css))
2868 } else if (PageSwapCache(page)) {
2869 ent.val = page_private(page);
2870 id = lookup_swap_cgroup_id(ent);
2872 memcg = mem_cgroup_lookup(id);
2873 if (memcg && !css_tryget(&memcg->css))
2877 unlock_page_cgroup(pc);
2881 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2883 unsigned int nr_pages,
2884 enum charge_type ctype,
2887 struct page_cgroup *pc = lookup_page_cgroup(page);
2888 struct zone *uninitialized_var(zone);
2889 struct lruvec *lruvec;
2890 bool was_on_lru = false;
2893 lock_page_cgroup(pc);
2894 VM_BUG_ON(PageCgroupUsed(pc));
2896 * we don't need page_cgroup_lock about tail pages, becase they are not
2897 * accessed by any other context at this point.
2901 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2902 * may already be on some other mem_cgroup's LRU. Take care of it.
2905 zone = page_zone(page);
2906 spin_lock_irq(&zone->lru_lock);
2907 if (PageLRU(page)) {
2908 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2910 del_page_from_lru_list(page, lruvec, page_lru(page));
2915 pc->mem_cgroup = memcg;
2917 * We access a page_cgroup asynchronously without lock_page_cgroup().
2918 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2919 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2920 * before USED bit, we need memory barrier here.
2921 * See mem_cgroup_add_lru_list(), etc.
2924 SetPageCgroupUsed(pc);
2928 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2929 VM_BUG_ON(PageLRU(page));
2931 add_page_to_lru_list(page, lruvec, page_lru(page));
2933 spin_unlock_irq(&zone->lru_lock);
2936 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2941 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2942 unlock_page_cgroup(pc);
2945 * "charge_statistics" updated event counter. Then, check it.
2946 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2947 * if they exceeds softlimit.
2949 memcg_check_events(memcg, page);
2952 static DEFINE_MUTEX(set_limit_mutex);
2954 #ifdef CONFIG_MEMCG_KMEM
2955 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2957 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2958 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2962 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2963 * in the memcg_cache_params struct.
2965 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2967 struct kmem_cache *cachep;
2969 VM_BUG_ON(p->is_root_cache);
2970 cachep = p->root_cache;
2971 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2974 #ifdef CONFIG_SLABINFO
2975 static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2978 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2979 struct memcg_cache_params *params;
2981 if (!memcg_can_account_kmem(memcg))
2984 print_slabinfo_header(m);
2986 mutex_lock(&memcg->slab_caches_mutex);
2987 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2988 cache_show(memcg_params_to_cache(params), m);
2989 mutex_unlock(&memcg->slab_caches_mutex);
2995 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2997 struct res_counter *fail_res;
2998 struct mem_cgroup *_memcg;
3002 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3007 * Conditions under which we can wait for the oom_killer. Those are
3008 * the same conditions tested by the core page allocator
3010 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
3013 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3016 if (ret == -EINTR) {
3018 * __mem_cgroup_try_charge() chosed to bypass to root due to
3019 * OOM kill or fatal signal. Since our only options are to
3020 * either fail the allocation or charge it to this cgroup, do
3021 * it as a temporary condition. But we can't fail. From a
3022 * kmem/slab perspective, the cache has already been selected,
3023 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3026 * This condition will only trigger if the task entered
3027 * memcg_charge_kmem in a sane state, but was OOM-killed during
3028 * __mem_cgroup_try_charge() above. Tasks that were already
3029 * dying when the allocation triggers should have been already
3030 * directed to the root cgroup in memcontrol.h
3032 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3033 if (do_swap_account)
3034 res_counter_charge_nofail(&memcg->memsw, size,
3038 res_counter_uncharge(&memcg->kmem, size);
3043 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3045 res_counter_uncharge(&memcg->res, size);
3046 if (do_swap_account)
3047 res_counter_uncharge(&memcg->memsw, size);
3050 if (res_counter_uncharge(&memcg->kmem, size))
3053 if (memcg_kmem_test_and_clear_dead(memcg))
3054 mem_cgroup_put(memcg);
3057 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3062 mutex_lock(&memcg->slab_caches_mutex);
3063 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3064 mutex_unlock(&memcg->slab_caches_mutex);
3068 * helper for acessing a memcg's index. It will be used as an index in the
3069 * child cache array in kmem_cache, and also to derive its name. This function
3070 * will return -1 when this is not a kmem-limited memcg.
3072 int memcg_cache_id(struct mem_cgroup *memcg)
3074 return memcg ? memcg->kmemcg_id : -1;
3078 * This ends up being protected by the set_limit mutex, during normal
3079 * operation, because that is its main call site.
3081 * But when we create a new cache, we can call this as well if its parent
3082 * is kmem-limited. That will have to hold set_limit_mutex as well.
3084 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3088 num = ida_simple_get(&kmem_limited_groups,
3089 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3093 * After this point, kmem_accounted (that we test atomically in
3094 * the beginning of this conditional), is no longer 0. This
3095 * guarantees only one process will set the following boolean
3096 * to true. We don't need test_and_set because we're protected
3097 * by the set_limit_mutex anyway.
3099 memcg_kmem_set_activated(memcg);
3101 ret = memcg_update_all_caches(num+1);
3103 ida_simple_remove(&kmem_limited_groups, num);
3104 memcg_kmem_clear_activated(memcg);
3108 memcg->kmemcg_id = num;
3109 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3110 mutex_init(&memcg->slab_caches_mutex);
3114 static size_t memcg_caches_array_size(int num_groups)
3117 if (num_groups <= 0)
3120 size = 2 * num_groups;
3121 if (size < MEMCG_CACHES_MIN_SIZE)
3122 size = MEMCG_CACHES_MIN_SIZE;
3123 else if (size > MEMCG_CACHES_MAX_SIZE)
3124 size = MEMCG_CACHES_MAX_SIZE;
3130 * We should update the current array size iff all caches updates succeed. This
3131 * can only be done from the slab side. The slab mutex needs to be held when
3134 void memcg_update_array_size(int num)
3136 if (num > memcg_limited_groups_array_size)
3137 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3140 static void kmem_cache_destroy_work_func(struct work_struct *w);
3142 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3144 struct memcg_cache_params *cur_params = s->memcg_params;
3146 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3148 if (num_groups > memcg_limited_groups_array_size) {
3150 ssize_t size = memcg_caches_array_size(num_groups);
3152 size *= sizeof(void *);
3153 size += sizeof(struct memcg_cache_params);
3155 s->memcg_params = kzalloc(size, GFP_KERNEL);
3156 if (!s->memcg_params) {
3157 s->memcg_params = cur_params;
3161 s->memcg_params->is_root_cache = true;
3164 * There is the chance it will be bigger than
3165 * memcg_limited_groups_array_size, if we failed an allocation
3166 * in a cache, in which case all caches updated before it, will
3167 * have a bigger array.
3169 * But if that is the case, the data after
3170 * memcg_limited_groups_array_size is certainly unused
3172 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3173 if (!cur_params->memcg_caches[i])
3175 s->memcg_params->memcg_caches[i] =
3176 cur_params->memcg_caches[i];
3180 * Ideally, we would wait until all caches succeed, and only
3181 * then free the old one. But this is not worth the extra
3182 * pointer per-cache we'd have to have for this.
3184 * It is not a big deal if some caches are left with a size
3185 * bigger than the others. And all updates will reset this
3193 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3194 struct kmem_cache *root_cache)
3196 size_t size = sizeof(struct memcg_cache_params);
3198 if (!memcg_kmem_enabled())
3202 size += memcg_limited_groups_array_size * sizeof(void *);
3204 s->memcg_params = kzalloc(size, GFP_KERNEL);
3205 if (!s->memcg_params)
3208 INIT_WORK(&s->memcg_params->destroy,
3209 kmem_cache_destroy_work_func);
3211 s->memcg_params->memcg = memcg;
3212 s->memcg_params->root_cache = root_cache;
3214 s->memcg_params->is_root_cache = true;
3219 void memcg_release_cache(struct kmem_cache *s)
3221 struct kmem_cache *root;
3222 struct mem_cgroup *memcg;
3226 * This happens, for instance, when a root cache goes away before we
3229 if (!s->memcg_params)
3232 if (s->memcg_params->is_root_cache)
3235 memcg = s->memcg_params->memcg;
3236 id = memcg_cache_id(memcg);
3238 root = s->memcg_params->root_cache;
3239 root->memcg_params->memcg_caches[id] = NULL;
3241 mutex_lock(&memcg->slab_caches_mutex);
3242 list_del(&s->memcg_params->list);
3243 mutex_unlock(&memcg->slab_caches_mutex);
3245 mem_cgroup_put(memcg);
3247 kfree(s->memcg_params);
3251 * During the creation a new cache, we need to disable our accounting mechanism
3252 * altogether. This is true even if we are not creating, but rather just
3253 * enqueing new caches to be created.
3255 * This is because that process will trigger allocations; some visible, like
3256 * explicit kmallocs to auxiliary data structures, name strings and internal
3257 * cache structures; some well concealed, like INIT_WORK() that can allocate
3258 * objects during debug.
3260 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3261 * to it. This may not be a bounded recursion: since the first cache creation
3262 * failed to complete (waiting on the allocation), we'll just try to create the
3263 * cache again, failing at the same point.
3265 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3266 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3267 * inside the following two functions.
3269 static inline void memcg_stop_kmem_account(void)
3271 VM_BUG_ON(!current->mm);
3272 current->memcg_kmem_skip_account++;
3275 static inline void memcg_resume_kmem_account(void)
3277 VM_BUG_ON(!current->mm);
3278 current->memcg_kmem_skip_account--;
3281 static void kmem_cache_destroy_work_func(struct work_struct *w)
3283 struct kmem_cache *cachep;
3284 struct memcg_cache_params *p;
3286 p = container_of(w, struct memcg_cache_params, destroy);
3288 cachep = memcg_params_to_cache(p);
3291 * If we get down to 0 after shrink, we could delete right away.
3292 * However, memcg_release_pages() already puts us back in the workqueue
3293 * in that case. If we proceed deleting, we'll get a dangling
3294 * reference, and removing the object from the workqueue in that case
3295 * is unnecessary complication. We are not a fast path.
3297 * Note that this case is fundamentally different from racing with
3298 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3299 * kmem_cache_shrink, not only we would be reinserting a dead cache
3300 * into the queue, but doing so from inside the worker racing to
3303 * So if we aren't down to zero, we'll just schedule a worker and try
3306 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3307 kmem_cache_shrink(cachep);
3308 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3311 kmem_cache_destroy(cachep);
3314 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3316 if (!cachep->memcg_params->dead)
3320 * There are many ways in which we can get here.
3322 * We can get to a memory-pressure situation while the delayed work is
3323 * still pending to run. The vmscan shrinkers can then release all
3324 * cache memory and get us to destruction. If this is the case, we'll
3325 * be executed twice, which is a bug (the second time will execute over
3326 * bogus data). In this case, cancelling the work should be fine.
3328 * But we can also get here from the worker itself, if
3329 * kmem_cache_shrink is enough to shake all the remaining objects and
3330 * get the page count to 0. In this case, we'll deadlock if we try to
3331 * cancel the work (the worker runs with an internal lock held, which
3332 * is the same lock we would hold for cancel_work_sync().)
3334 * Since we can't possibly know who got us here, just refrain from
3335 * running if there is already work pending
3337 if (work_pending(&cachep->memcg_params->destroy))
3340 * We have to defer the actual destroying to a workqueue, because
3341 * we might currently be in a context that cannot sleep.
3343 schedule_work(&cachep->memcg_params->destroy);
3347 * This lock protects updaters, not readers. We want readers to be as fast as
3348 * they can, and they will either see NULL or a valid cache value. Our model
3349 * allow them to see NULL, in which case the root memcg will be selected.
3351 * We need this lock because multiple allocations to the same cache from a non
3352 * will span more than one worker. Only one of them can create the cache.
3354 static DEFINE_MUTEX(memcg_cache_mutex);
3357 * Called with memcg_cache_mutex held
3359 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3360 struct kmem_cache *s)
3362 struct kmem_cache *new;
3363 static char *tmp_name = NULL;
3365 lockdep_assert_held(&memcg_cache_mutex);
3368 * kmem_cache_create_memcg duplicates the given name and
3369 * cgroup_name for this name requires RCU context.
3370 * This static temporary buffer is used to prevent from
3371 * pointless shortliving allocation.
3374 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3380 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3381 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3384 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3385 (s->flags & ~SLAB_PANIC), s->ctor, s);
3388 new->allocflags |= __GFP_KMEMCG;
3393 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3394 struct kmem_cache *cachep)
3396 struct kmem_cache *new_cachep;
3399 BUG_ON(!memcg_can_account_kmem(memcg));
3401 idx = memcg_cache_id(memcg);
3403 mutex_lock(&memcg_cache_mutex);
3404 new_cachep = cachep->memcg_params->memcg_caches[idx];
3408 new_cachep = kmem_cache_dup(memcg, cachep);
3409 if (new_cachep == NULL) {
3410 new_cachep = cachep;
3414 mem_cgroup_get(memcg);
3415 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3417 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3419 * the readers won't lock, make sure everybody sees the updated value,
3420 * so they won't put stuff in the queue again for no reason
3424 mutex_unlock(&memcg_cache_mutex);
3428 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3430 struct kmem_cache *c;
3433 if (!s->memcg_params)
3435 if (!s->memcg_params->is_root_cache)
3439 * If the cache is being destroyed, we trust that there is no one else
3440 * requesting objects from it. Even if there are, the sanity checks in
3441 * kmem_cache_destroy should caught this ill-case.
3443 * Still, we don't want anyone else freeing memcg_caches under our
3444 * noses, which can happen if a new memcg comes to life. As usual,
3445 * we'll take the set_limit_mutex to protect ourselves against this.
3447 mutex_lock(&set_limit_mutex);
3448 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3449 c = s->memcg_params->memcg_caches[i];
3454 * We will now manually delete the caches, so to avoid races
3455 * we need to cancel all pending destruction workers and
3456 * proceed with destruction ourselves.
3458 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3459 * and that could spawn the workers again: it is likely that
3460 * the cache still have active pages until this very moment.
3461 * This would lead us back to mem_cgroup_destroy_cache.
3463 * But that will not execute at all if the "dead" flag is not
3464 * set, so flip it down to guarantee we are in control.
3466 c->memcg_params->dead = false;
3467 cancel_work_sync(&c->memcg_params->destroy);
3468 kmem_cache_destroy(c);
3470 mutex_unlock(&set_limit_mutex);
3473 struct create_work {
3474 struct mem_cgroup *memcg;
3475 struct kmem_cache *cachep;
3476 struct work_struct work;
3479 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3481 struct kmem_cache *cachep;
3482 struct memcg_cache_params *params;
3484 if (!memcg_kmem_is_active(memcg))
3487 mutex_lock(&memcg->slab_caches_mutex);
3488 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3489 cachep = memcg_params_to_cache(params);
3490 cachep->memcg_params->dead = true;
3491 schedule_work(&cachep->memcg_params->destroy);
3493 mutex_unlock(&memcg->slab_caches_mutex);
3496 static void memcg_create_cache_work_func(struct work_struct *w)
3498 struct create_work *cw;
3500 cw = container_of(w, struct create_work, work);
3501 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3502 /* Drop the reference gotten when we enqueued. */
3503 css_put(&cw->memcg->css);
3508 * Enqueue the creation of a per-memcg kmem_cache.
3510 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3511 struct kmem_cache *cachep)
3513 struct create_work *cw;
3515 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3517 css_put(&memcg->css);
3522 cw->cachep = cachep;
3524 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3525 schedule_work(&cw->work);
3528 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3529 struct kmem_cache *cachep)
3532 * We need to stop accounting when we kmalloc, because if the
3533 * corresponding kmalloc cache is not yet created, the first allocation
3534 * in __memcg_create_cache_enqueue will recurse.
3536 * However, it is better to enclose the whole function. Depending on
3537 * the debugging options enabled, INIT_WORK(), for instance, can
3538 * trigger an allocation. This too, will make us recurse. Because at
3539 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3540 * the safest choice is to do it like this, wrapping the whole function.
3542 memcg_stop_kmem_account();
3543 __memcg_create_cache_enqueue(memcg, cachep);
3544 memcg_resume_kmem_account();
3547 * Return the kmem_cache we're supposed to use for a slab allocation.
3548 * We try to use the current memcg's version of the cache.
3550 * If the cache does not exist yet, if we are the first user of it,
3551 * we either create it immediately, if possible, or create it asynchronously
3553 * In the latter case, we will let the current allocation go through with
3554 * the original cache.
3556 * Can't be called in interrupt context or from kernel threads.
3557 * This function needs to be called with rcu_read_lock() held.
3559 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3562 struct mem_cgroup *memcg;
3565 VM_BUG_ON(!cachep->memcg_params);
3566 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3568 if (!current->mm || current->memcg_kmem_skip_account)
3572 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3574 if (!memcg_can_account_kmem(memcg))
3577 idx = memcg_cache_id(memcg);
3580 * barrier to mare sure we're always seeing the up to date value. The
3581 * code updating memcg_caches will issue a write barrier to match this.
3583 read_barrier_depends();
3584 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3585 cachep = cachep->memcg_params->memcg_caches[idx];
3589 /* The corresponding put will be done in the workqueue. */
3590 if (!css_tryget(&memcg->css))
3595 * If we are in a safe context (can wait, and not in interrupt
3596 * context), we could be be predictable and return right away.
3597 * This would guarantee that the allocation being performed
3598 * already belongs in the new cache.
3600 * However, there are some clashes that can arrive from locking.
3601 * For instance, because we acquire the slab_mutex while doing
3602 * kmem_cache_dup, this means no further allocation could happen
3603 * with the slab_mutex held.
3605 * Also, because cache creation issue get_online_cpus(), this
3606 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3607 * that ends up reversed during cpu hotplug. (cpuset allocates
3608 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3609 * better to defer everything.
3611 memcg_create_cache_enqueue(memcg, cachep);
3617 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3620 * We need to verify if the allocation against current->mm->owner's memcg is
3621 * possible for the given order. But the page is not allocated yet, so we'll
3622 * need a further commit step to do the final arrangements.
3624 * It is possible for the task to switch cgroups in this mean time, so at
3625 * commit time, we can't rely on task conversion any longer. We'll then use
3626 * the handle argument to return to the caller which cgroup we should commit
3627 * against. We could also return the memcg directly and avoid the pointer
3628 * passing, but a boolean return value gives better semantics considering
3629 * the compiled-out case as well.
3631 * Returning true means the allocation is possible.
3634 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3636 struct mem_cgroup *memcg;
3642 * Disabling accounting is only relevant for some specific memcg
3643 * internal allocations. Therefore we would initially not have such
3644 * check here, since direct calls to the page allocator that are marked
3645 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3646 * concerned with cache allocations, and by having this test at
3647 * memcg_kmem_get_cache, we are already able to relay the allocation to
3648 * the root cache and bypass the memcg cache altogether.
3650 * There is one exception, though: the SLUB allocator does not create
3651 * large order caches, but rather service large kmallocs directly from
3652 * the page allocator. Therefore, the following sequence when backed by
3653 * the SLUB allocator:
3655 * memcg_stop_kmem_account();
3656 * kmalloc(<large_number>)
3657 * memcg_resume_kmem_account();
3659 * would effectively ignore the fact that we should skip accounting,
3660 * since it will drive us directly to this function without passing
3661 * through the cache selector memcg_kmem_get_cache. Such large
3662 * allocations are extremely rare but can happen, for instance, for the
3663 * cache arrays. We bring this test here.
3665 if (!current->mm || current->memcg_kmem_skip_account)
3668 memcg = try_get_mem_cgroup_from_mm(current->mm);
3671 * very rare case described in mem_cgroup_from_task. Unfortunately there
3672 * isn't much we can do without complicating this too much, and it would
3673 * be gfp-dependent anyway. Just let it go
3675 if (unlikely(!memcg))
3678 if (!memcg_can_account_kmem(memcg)) {
3679 css_put(&memcg->css);
3683 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3687 css_put(&memcg->css);
3691 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3694 struct page_cgroup *pc;
3696 VM_BUG_ON(mem_cgroup_is_root(memcg));
3698 /* The page allocation failed. Revert */
3700 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3704 pc = lookup_page_cgroup(page);
3705 lock_page_cgroup(pc);
3706 pc->mem_cgroup = memcg;
3707 SetPageCgroupUsed(pc);
3708 unlock_page_cgroup(pc);
3711 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3713 struct mem_cgroup *memcg = NULL;
3714 struct page_cgroup *pc;
3717 pc = lookup_page_cgroup(page);
3719 * Fast unlocked return. Theoretically might have changed, have to
3720 * check again after locking.
3722 if (!PageCgroupUsed(pc))
3725 lock_page_cgroup(pc);
3726 if (PageCgroupUsed(pc)) {
3727 memcg = pc->mem_cgroup;
3728 ClearPageCgroupUsed(pc);
3730 unlock_page_cgroup(pc);
3733 * We trust that only if there is a memcg associated with the page, it
3734 * is a valid allocation
3739 VM_BUG_ON(mem_cgroup_is_root(memcg));
3740 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3743 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3746 #endif /* CONFIG_MEMCG_KMEM */
3748 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3750 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3752 * Because tail pages are not marked as "used", set it. We're under
3753 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3754 * charge/uncharge will be never happen and move_account() is done under
3755 * compound_lock(), so we don't have to take care of races.
3757 void mem_cgroup_split_huge_fixup(struct page *head)
3759 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3760 struct page_cgroup *pc;
3761 struct mem_cgroup *memcg;
3764 if (mem_cgroup_disabled())
3767 memcg = head_pc->mem_cgroup;
3768 for (i = 1; i < HPAGE_PMD_NR; i++) {
3770 pc->mem_cgroup = memcg;
3771 smp_wmb();/* see __commit_charge() */
3772 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3774 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3777 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3780 * mem_cgroup_move_account - move account of the page
3782 * @nr_pages: number of regular pages (>1 for huge pages)
3783 * @pc: page_cgroup of the page.
3784 * @from: mem_cgroup which the page is moved from.
3785 * @to: mem_cgroup which the page is moved to. @from != @to.
3787 * The caller must confirm following.
3788 * - page is not on LRU (isolate_page() is useful.)
3789 * - compound_lock is held when nr_pages > 1
3791 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3794 static int mem_cgroup_move_account(struct page *page,
3795 unsigned int nr_pages,
3796 struct page_cgroup *pc,
3797 struct mem_cgroup *from,
3798 struct mem_cgroup *to)
3800 unsigned long flags;
3802 bool anon = PageAnon(page);
3804 VM_BUG_ON(from == to);
3805 VM_BUG_ON(PageLRU(page));
3807 * The page is isolated from LRU. So, collapse function
3808 * will not handle this page. But page splitting can happen.
3809 * Do this check under compound_page_lock(). The caller should
3813 if (nr_pages > 1 && !PageTransHuge(page))
3816 lock_page_cgroup(pc);
3819 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3822 move_lock_mem_cgroup(from, &flags);
3824 if (!anon && page_mapped(page)) {
3825 /* Update mapped_file data for mem_cgroup */
3827 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3828 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3831 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3833 /* caller should have done css_get */
3834 pc->mem_cgroup = to;
3835 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3836 move_unlock_mem_cgroup(from, &flags);
3839 unlock_page_cgroup(pc);
3843 memcg_check_events(to, page);
3844 memcg_check_events(from, page);
3850 * mem_cgroup_move_parent - moves page to the parent group
3851 * @page: the page to move
3852 * @pc: page_cgroup of the page
3853 * @child: page's cgroup
3855 * move charges to its parent or the root cgroup if the group has no
3856 * parent (aka use_hierarchy==0).
3857 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3858 * mem_cgroup_move_account fails) the failure is always temporary and
3859 * it signals a race with a page removal/uncharge or migration. In the
3860 * first case the page is on the way out and it will vanish from the LRU
3861 * on the next attempt and the call should be retried later.
3862 * Isolation from the LRU fails only if page has been isolated from
3863 * the LRU since we looked at it and that usually means either global
3864 * reclaim or migration going on. The page will either get back to the
3866 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3867 * (!PageCgroupUsed) or moved to a different group. The page will
3868 * disappear in the next attempt.
3870 static int mem_cgroup_move_parent(struct page *page,
3871 struct page_cgroup *pc,
3872 struct mem_cgroup *child)
3874 struct mem_cgroup *parent;
3875 unsigned int nr_pages;
3876 unsigned long uninitialized_var(flags);
3879 VM_BUG_ON(mem_cgroup_is_root(child));
3882 if (!get_page_unless_zero(page))
3884 if (isolate_lru_page(page))
3887 nr_pages = hpage_nr_pages(page);
3889 parent = parent_mem_cgroup(child);
3891 * If no parent, move charges to root cgroup.
3894 parent = root_mem_cgroup;
3897 VM_BUG_ON(!PageTransHuge(page));
3898 flags = compound_lock_irqsave(page);
3901 ret = mem_cgroup_move_account(page, nr_pages,
3904 __mem_cgroup_cancel_local_charge(child, nr_pages);
3907 compound_unlock_irqrestore(page, flags);
3908 putback_lru_page(page);
3916 * Charge the memory controller for page usage.
3918 * 0 if the charge was successful
3919 * < 0 if the cgroup is over its limit
3921 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3922 gfp_t gfp_mask, enum charge_type ctype)
3924 struct mem_cgroup *memcg = NULL;
3925 unsigned int nr_pages = 1;
3929 if (PageTransHuge(page)) {
3930 nr_pages <<= compound_order(page);
3931 VM_BUG_ON(!PageTransHuge(page));
3933 * Never OOM-kill a process for a huge page. The
3934 * fault handler will fall back to regular pages.
3939 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3942 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3946 int mem_cgroup_newpage_charge(struct page *page,
3947 struct mm_struct *mm, gfp_t gfp_mask)
3949 if (mem_cgroup_disabled())
3951 VM_BUG_ON(page_mapped(page));
3952 VM_BUG_ON(page->mapping && !PageAnon(page));
3954 return mem_cgroup_charge_common(page, mm, gfp_mask,
3955 MEM_CGROUP_CHARGE_TYPE_ANON);
3959 * While swap-in, try_charge -> commit or cancel, the page is locked.
3960 * And when try_charge() successfully returns, one refcnt to memcg without
3961 * struct page_cgroup is acquired. This refcnt will be consumed by
3962 * "commit()" or removed by "cancel()"
3964 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3967 struct mem_cgroup **memcgp)
3969 struct mem_cgroup *memcg;
3970 struct page_cgroup *pc;
3973 pc = lookup_page_cgroup(page);
3975 * Every swap fault against a single page tries to charge the
3976 * page, bail as early as possible. shmem_unuse() encounters
3977 * already charged pages, too. The USED bit is protected by
3978 * the page lock, which serializes swap cache removal, which
3979 * in turn serializes uncharging.
3981 if (PageCgroupUsed(pc))
3983 if (!do_swap_account)
3985 memcg = try_get_mem_cgroup_from_page(page);
3989 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3990 css_put(&memcg->css);
3995 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4001 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4002 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4005 if (mem_cgroup_disabled())
4008 * A racing thread's fault, or swapoff, may have already
4009 * updated the pte, and even removed page from swap cache: in
4010 * those cases unuse_pte()'s pte_same() test will fail; but
4011 * there's also a KSM case which does need to charge the page.
4013 if (!PageSwapCache(page)) {
4016 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4021 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4024 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4026 if (mem_cgroup_disabled())
4030 __mem_cgroup_cancel_charge(memcg, 1);
4034 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4035 enum charge_type ctype)
4037 if (mem_cgroup_disabled())
4042 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4044 * Now swap is on-memory. This means this page may be
4045 * counted both as mem and swap....double count.
4046 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4047 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4048 * may call delete_from_swap_cache() before reach here.
4050 if (do_swap_account && PageSwapCache(page)) {
4051 swp_entry_t ent = {.val = page_private(page)};
4052 mem_cgroup_uncharge_swap(ent);
4056 void mem_cgroup_commit_charge_swapin(struct page *page,
4057 struct mem_cgroup *memcg)
4059 __mem_cgroup_commit_charge_swapin(page, memcg,
4060 MEM_CGROUP_CHARGE_TYPE_ANON);
4063 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4066 struct mem_cgroup *memcg = NULL;
4067 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4070 if (mem_cgroup_disabled())
4072 if (PageCompound(page))
4075 if (!PageSwapCache(page))
4076 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4077 else { /* page is swapcache/shmem */
4078 ret = __mem_cgroup_try_charge_swapin(mm, page,
4081 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4086 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4087 unsigned int nr_pages,
4088 const enum charge_type ctype)
4090 struct memcg_batch_info *batch = NULL;
4091 bool uncharge_memsw = true;
4093 /* If swapout, usage of swap doesn't decrease */
4094 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4095 uncharge_memsw = false;
4097 batch = ¤t->memcg_batch;
4099 * In usual, we do css_get() when we remember memcg pointer.
4100 * But in this case, we keep res->usage until end of a series of
4101 * uncharges. Then, it's ok to ignore memcg's refcnt.
4104 batch->memcg = memcg;
4106 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4107 * In those cases, all pages freed continuously can be expected to be in
4108 * the same cgroup and we have chance to coalesce uncharges.
4109 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4110 * because we want to do uncharge as soon as possible.
4113 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4114 goto direct_uncharge;
4117 goto direct_uncharge;
4120 * In typical case, batch->memcg == mem. This means we can
4121 * merge a series of uncharges to an uncharge of res_counter.
4122 * If not, we uncharge res_counter ony by one.
4124 if (batch->memcg != memcg)
4125 goto direct_uncharge;
4126 /* remember freed charge and uncharge it later */
4129 batch->memsw_nr_pages++;
4132 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4134 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4135 if (unlikely(batch->memcg != memcg))
4136 memcg_oom_recover(memcg);
4140 * uncharge if !page_mapped(page)
4142 static struct mem_cgroup *
4143 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4146 struct mem_cgroup *memcg = NULL;
4147 unsigned int nr_pages = 1;
4148 struct page_cgroup *pc;
4151 if (mem_cgroup_disabled())
4154 if (PageTransHuge(page)) {
4155 nr_pages <<= compound_order(page);
4156 VM_BUG_ON(!PageTransHuge(page));
4159 * Check if our page_cgroup is valid
4161 pc = lookup_page_cgroup(page);
4162 if (unlikely(!PageCgroupUsed(pc)))
4165 lock_page_cgroup(pc);
4167 memcg = pc->mem_cgroup;
4169 if (!PageCgroupUsed(pc))
4172 anon = PageAnon(page);
4175 case MEM_CGROUP_CHARGE_TYPE_ANON:
4177 * Generally PageAnon tells if it's the anon statistics to be
4178 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4179 * used before page reached the stage of being marked PageAnon.
4183 case MEM_CGROUP_CHARGE_TYPE_DROP:
4184 /* See mem_cgroup_prepare_migration() */
4185 if (page_mapped(page))
4188 * Pages under migration may not be uncharged. But
4189 * end_migration() /must/ be the one uncharging the
4190 * unused post-migration page and so it has to call
4191 * here with the migration bit still set. See the
4192 * res_counter handling below.
4194 if (!end_migration && PageCgroupMigration(pc))
4197 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4198 if (!PageAnon(page)) { /* Shared memory */
4199 if (page->mapping && !page_is_file_cache(page))
4201 } else if (page_mapped(page)) /* Anon */
4208 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4210 ClearPageCgroupUsed(pc);
4212 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4213 * freed from LRU. This is safe because uncharged page is expected not
4214 * to be reused (freed soon). Exception is SwapCache, it's handled by
4215 * special functions.
4218 unlock_page_cgroup(pc);
4220 * even after unlock, we have memcg->res.usage here and this memcg
4221 * will never be freed.
4223 memcg_check_events(memcg, page);
4224 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4225 mem_cgroup_swap_statistics(memcg, true);
4226 mem_cgroup_get(memcg);
4229 * Migration does not charge the res_counter for the
4230 * replacement page, so leave it alone when phasing out the
4231 * page that is unused after the migration.
4233 if (!end_migration && !mem_cgroup_is_root(memcg))
4234 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4239 unlock_page_cgroup(pc);
4243 void mem_cgroup_uncharge_page(struct page *page)
4246 if (page_mapped(page))
4248 VM_BUG_ON(page->mapping && !PageAnon(page));
4250 * If the page is in swap cache, uncharge should be deferred
4251 * to the swap path, which also properly accounts swap usage
4252 * and handles memcg lifetime.
4254 * Note that this check is not stable and reclaim may add the
4255 * page to swap cache at any time after this. However, if the
4256 * page is not in swap cache by the time page->mapcount hits
4257 * 0, there won't be any page table references to the swap
4258 * slot, and reclaim will free it and not actually write the
4261 if (PageSwapCache(page))
4263 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4266 void mem_cgroup_uncharge_cache_page(struct page *page)
4268 VM_BUG_ON(page_mapped(page));
4269 VM_BUG_ON(page->mapping);
4270 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4274 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4275 * In that cases, pages are freed continuously and we can expect pages
4276 * are in the same memcg. All these calls itself limits the number of
4277 * pages freed at once, then uncharge_start/end() is called properly.
4278 * This may be called prural(2) times in a context,
4281 void mem_cgroup_uncharge_start(void)
4283 current->memcg_batch.do_batch++;
4284 /* We can do nest. */
4285 if (current->memcg_batch.do_batch == 1) {
4286 current->memcg_batch.memcg = NULL;
4287 current->memcg_batch.nr_pages = 0;
4288 current->memcg_batch.memsw_nr_pages = 0;
4292 void mem_cgroup_uncharge_end(void)
4294 struct memcg_batch_info *batch = ¤t->memcg_batch;
4296 if (!batch->do_batch)
4300 if (batch->do_batch) /* If stacked, do nothing. */
4306 * This "batch->memcg" is valid without any css_get/put etc...
4307 * bacause we hide charges behind us.
4309 if (batch->nr_pages)
4310 res_counter_uncharge(&batch->memcg->res,
4311 batch->nr_pages * PAGE_SIZE);
4312 if (batch->memsw_nr_pages)
4313 res_counter_uncharge(&batch->memcg->memsw,
4314 batch->memsw_nr_pages * PAGE_SIZE);
4315 memcg_oom_recover(batch->memcg);
4316 /* forget this pointer (for sanity check) */
4317 batch->memcg = NULL;
4322 * called after __delete_from_swap_cache() and drop "page" account.
4323 * memcg information is recorded to swap_cgroup of "ent"
4326 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4328 struct mem_cgroup *memcg;
4329 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4331 if (!swapout) /* this was a swap cache but the swap is unused ! */
4332 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4334 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4337 * record memcg information, if swapout && memcg != NULL,
4338 * mem_cgroup_get() was called in uncharge().
4340 if (do_swap_account && swapout && memcg)
4341 swap_cgroup_record(ent, css_id(&memcg->css));
4345 #ifdef CONFIG_MEMCG_SWAP
4347 * called from swap_entry_free(). remove record in swap_cgroup and
4348 * uncharge "memsw" account.
4350 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4352 struct mem_cgroup *memcg;
4355 if (!do_swap_account)
4358 id = swap_cgroup_record(ent, 0);
4360 memcg = mem_cgroup_lookup(id);
4363 * We uncharge this because swap is freed.
4364 * This memcg can be obsolete one. We avoid calling css_tryget
4366 if (!mem_cgroup_is_root(memcg))
4367 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4368 mem_cgroup_swap_statistics(memcg, false);
4369 mem_cgroup_put(memcg);
4375 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4376 * @entry: swap entry to be moved
4377 * @from: mem_cgroup which the entry is moved from
4378 * @to: mem_cgroup which the entry is moved to
4380 * It succeeds only when the swap_cgroup's record for this entry is the same
4381 * as the mem_cgroup's id of @from.
4383 * Returns 0 on success, -EINVAL on failure.
4385 * The caller must have charged to @to, IOW, called res_counter_charge() about
4386 * both res and memsw, and called css_get().
4388 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4389 struct mem_cgroup *from, struct mem_cgroup *to)
4391 unsigned short old_id, new_id;
4393 old_id = css_id(&from->css);
4394 new_id = css_id(&to->css);
4396 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4397 mem_cgroup_swap_statistics(from, false);
4398 mem_cgroup_swap_statistics(to, true);
4400 * This function is only called from task migration context now.
4401 * It postpones res_counter and refcount handling till the end
4402 * of task migration(mem_cgroup_clear_mc()) for performance
4403 * improvement. But we cannot postpone mem_cgroup_get(to)
4404 * because if the process that has been moved to @to does
4405 * swap-in, the refcount of @to might be decreased to 0.
4413 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4414 struct mem_cgroup *from, struct mem_cgroup *to)
4421 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4424 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4425 struct mem_cgroup **memcgp)
4427 struct mem_cgroup *memcg = NULL;
4428 unsigned int nr_pages = 1;
4429 struct page_cgroup *pc;
4430 enum charge_type ctype;
4434 if (mem_cgroup_disabled())
4437 if (PageTransHuge(page))
4438 nr_pages <<= compound_order(page);
4440 pc = lookup_page_cgroup(page);
4441 lock_page_cgroup(pc);
4442 if (PageCgroupUsed(pc)) {
4443 memcg = pc->mem_cgroup;
4444 css_get(&memcg->css);
4446 * At migrating an anonymous page, its mapcount goes down
4447 * to 0 and uncharge() will be called. But, even if it's fully
4448 * unmapped, migration may fail and this page has to be
4449 * charged again. We set MIGRATION flag here and delay uncharge
4450 * until end_migration() is called
4452 * Corner Case Thinking
4454 * When the old page was mapped as Anon and it's unmap-and-freed
4455 * while migration was ongoing.
4456 * If unmap finds the old page, uncharge() of it will be delayed
4457 * until end_migration(). If unmap finds a new page, it's
4458 * uncharged when it make mapcount to be 1->0. If unmap code
4459 * finds swap_migration_entry, the new page will not be mapped
4460 * and end_migration() will find it(mapcount==0).
4463 * When the old page was mapped but migraion fails, the kernel
4464 * remaps it. A charge for it is kept by MIGRATION flag even
4465 * if mapcount goes down to 0. We can do remap successfully
4466 * without charging it again.
4469 * The "old" page is under lock_page() until the end of
4470 * migration, so, the old page itself will not be swapped-out.
4471 * If the new page is swapped out before end_migraton, our
4472 * hook to usual swap-out path will catch the event.
4475 SetPageCgroupMigration(pc);
4477 unlock_page_cgroup(pc);
4479 * If the page is not charged at this point,
4487 * We charge new page before it's used/mapped. So, even if unlock_page()
4488 * is called before end_migration, we can catch all events on this new
4489 * page. In the case new page is migrated but not remapped, new page's
4490 * mapcount will be finally 0 and we call uncharge in end_migration().
4493 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4495 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4497 * The page is committed to the memcg, but it's not actually
4498 * charged to the res_counter since we plan on replacing the
4499 * old one and only one page is going to be left afterwards.
4501 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4504 /* remove redundant charge if migration failed*/
4505 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4506 struct page *oldpage, struct page *newpage, bool migration_ok)
4508 struct page *used, *unused;
4509 struct page_cgroup *pc;
4515 if (!migration_ok) {
4522 anon = PageAnon(used);
4523 __mem_cgroup_uncharge_common(unused,
4524 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4525 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4527 css_put(&memcg->css);
4529 * We disallowed uncharge of pages under migration because mapcount
4530 * of the page goes down to zero, temporarly.
4531 * Clear the flag and check the page should be charged.
4533 pc = lookup_page_cgroup(oldpage);
4534 lock_page_cgroup(pc);
4535 ClearPageCgroupMigration(pc);
4536 unlock_page_cgroup(pc);
4539 * If a page is a file cache, radix-tree replacement is very atomic
4540 * and we can skip this check. When it was an Anon page, its mapcount
4541 * goes down to 0. But because we added MIGRATION flage, it's not
4542 * uncharged yet. There are several case but page->mapcount check
4543 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4544 * check. (see prepare_charge() also)
4547 mem_cgroup_uncharge_page(used);
4551 * At replace page cache, newpage is not under any memcg but it's on
4552 * LRU. So, this function doesn't touch res_counter but handles LRU
4553 * in correct way. Both pages are locked so we cannot race with uncharge.
4555 void mem_cgroup_replace_page_cache(struct page *oldpage,
4556 struct page *newpage)
4558 struct mem_cgroup *memcg = NULL;
4559 struct page_cgroup *pc;
4560 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4562 if (mem_cgroup_disabled())
4565 pc = lookup_page_cgroup(oldpage);
4566 /* fix accounting on old pages */
4567 lock_page_cgroup(pc);
4568 if (PageCgroupUsed(pc)) {
4569 memcg = pc->mem_cgroup;
4570 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4571 ClearPageCgroupUsed(pc);
4573 unlock_page_cgroup(pc);
4576 * When called from shmem_replace_page(), in some cases the
4577 * oldpage has already been charged, and in some cases not.
4582 * Even if newpage->mapping was NULL before starting replacement,
4583 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4584 * LRU while we overwrite pc->mem_cgroup.
4586 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4589 #ifdef CONFIG_DEBUG_VM
4590 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4592 struct page_cgroup *pc;
4594 pc = lookup_page_cgroup(page);
4596 * Can be NULL while feeding pages into the page allocator for
4597 * the first time, i.e. during boot or memory hotplug;
4598 * or when mem_cgroup_disabled().
4600 if (likely(pc) && PageCgroupUsed(pc))
4605 bool mem_cgroup_bad_page_check(struct page *page)
4607 if (mem_cgroup_disabled())
4610 return lookup_page_cgroup_used(page) != NULL;
4613 void mem_cgroup_print_bad_page(struct page *page)
4615 struct page_cgroup *pc;
4617 pc = lookup_page_cgroup_used(page);
4619 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4620 pc, pc->flags, pc->mem_cgroup);
4625 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4626 unsigned long long val)
4629 u64 memswlimit, memlimit;
4631 int children = mem_cgroup_count_children(memcg);
4632 u64 curusage, oldusage;
4636 * For keeping hierarchical_reclaim simple, how long we should retry
4637 * is depends on callers. We set our retry-count to be function
4638 * of # of children which we should visit in this loop.
4640 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4642 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4645 while (retry_count) {
4646 if (signal_pending(current)) {
4651 * Rather than hide all in some function, I do this in
4652 * open coded manner. You see what this really does.
4653 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4655 mutex_lock(&set_limit_mutex);
4656 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4657 if (memswlimit < val) {
4659 mutex_unlock(&set_limit_mutex);
4663 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4667 ret = res_counter_set_limit(&memcg->res, val);
4669 if (memswlimit == val)
4670 memcg->memsw_is_minimum = true;
4672 memcg->memsw_is_minimum = false;
4674 mutex_unlock(&set_limit_mutex);
4679 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4680 MEM_CGROUP_RECLAIM_SHRINK);
4681 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4682 /* Usage is reduced ? */
4683 if (curusage >= oldusage)
4686 oldusage = curusage;
4688 if (!ret && enlarge)
4689 memcg_oom_recover(memcg);
4694 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4695 unsigned long long val)
4698 u64 memlimit, memswlimit, oldusage, curusage;
4699 int children = mem_cgroup_count_children(memcg);
4703 /* see mem_cgroup_resize_res_limit */
4704 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4705 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4706 while (retry_count) {
4707 if (signal_pending(current)) {
4712 * Rather than hide all in some function, I do this in
4713 * open coded manner. You see what this really does.
4714 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4716 mutex_lock(&set_limit_mutex);
4717 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4718 if (memlimit > val) {
4720 mutex_unlock(&set_limit_mutex);
4723 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4724 if (memswlimit < val)
4726 ret = res_counter_set_limit(&memcg->memsw, val);
4728 if (memlimit == val)
4729 memcg->memsw_is_minimum = true;
4731 memcg->memsw_is_minimum = false;
4733 mutex_unlock(&set_limit_mutex);
4738 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4739 MEM_CGROUP_RECLAIM_NOSWAP |
4740 MEM_CGROUP_RECLAIM_SHRINK);
4741 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4742 /* Usage is reduced ? */
4743 if (curusage >= oldusage)
4746 oldusage = curusage;
4748 if (!ret && enlarge)
4749 memcg_oom_recover(memcg);
4753 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4755 unsigned long *total_scanned)
4757 unsigned long nr_reclaimed = 0;
4758 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4759 unsigned long reclaimed;
4761 struct mem_cgroup_tree_per_zone *mctz;
4762 unsigned long long excess;
4763 unsigned long nr_scanned;
4768 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4770 * This loop can run a while, specially if mem_cgroup's continuously
4771 * keep exceeding their soft limit and putting the system under
4778 mz = mem_cgroup_largest_soft_limit_node(mctz);
4783 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4784 gfp_mask, &nr_scanned);
4785 nr_reclaimed += reclaimed;
4786 *total_scanned += nr_scanned;
4787 spin_lock(&mctz->lock);
4790 * If we failed to reclaim anything from this memory cgroup
4791 * it is time to move on to the next cgroup
4797 * Loop until we find yet another one.
4799 * By the time we get the soft_limit lock
4800 * again, someone might have aded the
4801 * group back on the RB tree. Iterate to
4802 * make sure we get a different mem.
4803 * mem_cgroup_largest_soft_limit_node returns
4804 * NULL if no other cgroup is present on
4808 __mem_cgroup_largest_soft_limit_node(mctz);
4810 css_put(&next_mz->memcg->css);
4811 else /* next_mz == NULL or other memcg */
4815 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4816 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4818 * One school of thought says that we should not add
4819 * back the node to the tree if reclaim returns 0.
4820 * But our reclaim could return 0, simply because due
4821 * to priority we are exposing a smaller subset of
4822 * memory to reclaim from. Consider this as a longer
4825 /* If excess == 0, no tree ops */
4826 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4827 spin_unlock(&mctz->lock);
4828 css_put(&mz->memcg->css);
4831 * Could not reclaim anything and there are no more
4832 * mem cgroups to try or we seem to be looping without
4833 * reclaiming anything.
4835 if (!nr_reclaimed &&
4837 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4839 } while (!nr_reclaimed);
4841 css_put(&next_mz->memcg->css);
4842 return nr_reclaimed;
4846 * mem_cgroup_force_empty_list - clears LRU of a group
4847 * @memcg: group to clear
4850 * @lru: lru to to clear
4852 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4853 * reclaim the pages page themselves - pages are moved to the parent (or root)
4856 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4857 int node, int zid, enum lru_list lru)
4859 struct lruvec *lruvec;
4860 unsigned long flags;
4861 struct list_head *list;
4865 zone = &NODE_DATA(node)->node_zones[zid];
4866 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4867 list = &lruvec->lists[lru];
4871 struct page_cgroup *pc;
4874 spin_lock_irqsave(&zone->lru_lock, flags);
4875 if (list_empty(list)) {
4876 spin_unlock_irqrestore(&zone->lru_lock, flags);
4879 page = list_entry(list->prev, struct page, lru);
4881 list_move(&page->lru, list);
4883 spin_unlock_irqrestore(&zone->lru_lock, flags);
4886 spin_unlock_irqrestore(&zone->lru_lock, flags);
4888 pc = lookup_page_cgroup(page);
4890 if (mem_cgroup_move_parent(page, pc, memcg)) {
4891 /* found lock contention or "pc" is obsolete. */
4896 } while (!list_empty(list));
4900 * make mem_cgroup's charge to be 0 if there is no task by moving
4901 * all the charges and pages to the parent.
4902 * This enables deleting this mem_cgroup.
4904 * Caller is responsible for holding css reference on the memcg.
4906 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4912 /* This is for making all *used* pages to be on LRU. */
4913 lru_add_drain_all();
4914 drain_all_stock_sync(memcg);
4915 mem_cgroup_start_move(memcg);
4916 for_each_node_state(node, N_MEMORY) {
4917 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4920 mem_cgroup_force_empty_list(memcg,
4925 mem_cgroup_end_move(memcg);
4926 memcg_oom_recover(memcg);
4930 * Kernel memory may not necessarily be trackable to a specific
4931 * process. So they are not migrated, and therefore we can't
4932 * expect their value to drop to 0 here.
4933 * Having res filled up with kmem only is enough.
4935 * This is a safety check because mem_cgroup_force_empty_list
4936 * could have raced with mem_cgroup_replace_page_cache callers
4937 * so the lru seemed empty but the page could have been added
4938 * right after the check. RES_USAGE should be safe as we always
4939 * charge before adding to the LRU.
4941 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4942 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4943 } while (usage > 0);
4947 * This mainly exists for tests during the setting of set of use_hierarchy.
4948 * Since this is the very setting we are changing, the current hierarchy value
4951 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4955 /* bounce at first found */
4956 cgroup_for_each_child(pos, memcg->css.cgroup)
4962 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4963 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4964 * from mem_cgroup_count_children(), in the sense that we don't really care how
4965 * many children we have; we only need to know if we have any. It also counts
4966 * any memcg without hierarchy as infertile.
4968 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4970 return memcg->use_hierarchy && __memcg_has_children(memcg);
4974 * Reclaims as many pages from the given memcg as possible and moves
4975 * the rest to the parent.
4977 * Caller is responsible for holding css reference for memcg.
4979 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4981 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4982 struct cgroup *cgrp = memcg->css.cgroup;
4984 /* returns EBUSY if there is a task or if we come here twice. */
4985 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4988 /* we call try-to-free pages for make this cgroup empty */
4989 lru_add_drain_all();
4990 /* try to free all pages in this cgroup */
4991 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4994 if (signal_pending(current))
4997 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5001 /* maybe some writeback is necessary */
5002 congestion_wait(BLK_RW_ASYNC, HZ/10);
5007 mem_cgroup_reparent_charges(memcg);
5012 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
5014 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5017 if (mem_cgroup_is_root(memcg))
5019 css_get(&memcg->css);
5020 ret = mem_cgroup_force_empty(memcg);
5021 css_put(&memcg->css);
5027 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
5029 return mem_cgroup_from_cont(cont)->use_hierarchy;
5032 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
5036 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5037 struct cgroup *parent = cont->parent;
5038 struct mem_cgroup *parent_memcg = NULL;
5041 parent_memcg = mem_cgroup_from_cont(parent);
5043 mutex_lock(&memcg_create_mutex);
5045 if (memcg->use_hierarchy == val)
5049 * If parent's use_hierarchy is set, we can't make any modifications
5050 * in the child subtrees. If it is unset, then the change can
5051 * occur, provided the current cgroup has no children.
5053 * For the root cgroup, parent_mem is NULL, we allow value to be
5054 * set if there are no children.
5056 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5057 (val == 1 || val == 0)) {
5058 if (!__memcg_has_children(memcg))
5059 memcg->use_hierarchy = val;
5066 mutex_unlock(&memcg_create_mutex);
5072 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5073 enum mem_cgroup_stat_index idx)
5075 struct mem_cgroup *iter;
5078 /* Per-cpu values can be negative, use a signed accumulator */
5079 for_each_mem_cgroup_tree(iter, memcg)
5080 val += mem_cgroup_read_stat(iter, idx);
5082 if (val < 0) /* race ? */
5087 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5091 if (!mem_cgroup_is_root(memcg)) {
5093 return res_counter_read_u64(&memcg->res, RES_USAGE);
5095 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5099 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5100 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5102 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5103 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5106 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5108 return val << PAGE_SHIFT;
5111 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
5112 struct file *file, char __user *buf,
5113 size_t nbytes, loff_t *ppos)
5115 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5121 type = MEMFILE_TYPE(cft->private);
5122 name = MEMFILE_ATTR(cft->private);
5126 if (name == RES_USAGE)
5127 val = mem_cgroup_usage(memcg, false);
5129 val = res_counter_read_u64(&memcg->res, name);
5132 if (name == RES_USAGE)
5133 val = mem_cgroup_usage(memcg, true);
5135 val = res_counter_read_u64(&memcg->memsw, name);
5138 val = res_counter_read_u64(&memcg->kmem, name);
5144 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5145 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5148 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
5151 #ifdef CONFIG_MEMCG_KMEM
5152 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5154 * For simplicity, we won't allow this to be disabled. It also can't
5155 * be changed if the cgroup has children already, or if tasks had
5158 * If tasks join before we set the limit, a person looking at
5159 * kmem.usage_in_bytes will have no way to determine when it took
5160 * place, which makes the value quite meaningless.
5162 * After it first became limited, changes in the value of the limit are
5163 * of course permitted.
5165 mutex_lock(&memcg_create_mutex);
5166 mutex_lock(&set_limit_mutex);
5167 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
5168 if (cgroup_task_count(cont) || memcg_has_children(memcg)) {
5172 ret = res_counter_set_limit(&memcg->kmem, val);
5175 ret = memcg_update_cache_sizes(memcg);
5177 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5180 static_key_slow_inc(&memcg_kmem_enabled_key);
5182 * setting the active bit after the inc will guarantee no one
5183 * starts accounting before all call sites are patched
5185 memcg_kmem_set_active(memcg);
5188 * kmem charges can outlive the cgroup. In the case of slab
5189 * pages, for instance, a page contain objects from various
5190 * processes, so it is unfeasible to migrate them away. We
5191 * need to reference count the memcg because of that.
5193 mem_cgroup_get(memcg);
5195 ret = res_counter_set_limit(&memcg->kmem, val);
5197 mutex_unlock(&set_limit_mutex);
5198 mutex_unlock(&memcg_create_mutex);
5203 #ifdef CONFIG_MEMCG_KMEM
5204 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5207 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5211 memcg->kmem_account_flags = parent->kmem_account_flags;
5213 * When that happen, we need to disable the static branch only on those
5214 * memcgs that enabled it. To achieve this, we would be forced to
5215 * complicate the code by keeping track of which memcgs were the ones
5216 * that actually enabled limits, and which ones got it from its
5219 * It is a lot simpler just to do static_key_slow_inc() on every child
5220 * that is accounted.
5222 if (!memcg_kmem_is_active(memcg))
5226 * destroy(), called if we fail, will issue static_key_slow_inc() and
5227 * mem_cgroup_put() if kmem is enabled. We have to either call them
5228 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5229 * this more consistent, since it always leads to the same destroy path
5231 mem_cgroup_get(memcg);
5232 static_key_slow_inc(&memcg_kmem_enabled_key);
5234 mutex_lock(&set_limit_mutex);
5235 ret = memcg_update_cache_sizes(memcg);
5236 mutex_unlock(&set_limit_mutex);
5240 #endif /* CONFIG_MEMCG_KMEM */
5243 * The user of this function is...
5246 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5249 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5252 unsigned long long val;
5255 type = MEMFILE_TYPE(cft->private);
5256 name = MEMFILE_ATTR(cft->private);
5260 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5264 /* This function does all necessary parse...reuse it */
5265 ret = res_counter_memparse_write_strategy(buffer, &val);
5269 ret = mem_cgroup_resize_limit(memcg, val);
5270 else if (type == _MEMSWAP)
5271 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5272 else if (type == _KMEM)
5273 ret = memcg_update_kmem_limit(cont, val);
5277 case RES_SOFT_LIMIT:
5278 ret = res_counter_memparse_write_strategy(buffer, &val);
5282 * For memsw, soft limits are hard to implement in terms
5283 * of semantics, for now, we support soft limits for
5284 * control without swap
5287 ret = res_counter_set_soft_limit(&memcg->res, val);
5292 ret = -EINVAL; /* should be BUG() ? */
5298 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5299 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5301 struct cgroup *cgroup;
5302 unsigned long long min_limit, min_memsw_limit, tmp;
5304 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5305 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5306 cgroup = memcg->css.cgroup;
5307 if (!memcg->use_hierarchy)
5310 while (cgroup->parent) {
5311 cgroup = cgroup->parent;
5312 memcg = mem_cgroup_from_cont(cgroup);
5313 if (!memcg->use_hierarchy)
5315 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5316 min_limit = min(min_limit, tmp);
5317 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5318 min_memsw_limit = min(min_memsw_limit, tmp);
5321 *mem_limit = min_limit;
5322 *memsw_limit = min_memsw_limit;
5325 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5327 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5331 type = MEMFILE_TYPE(event);
5332 name = MEMFILE_ATTR(event);
5337 res_counter_reset_max(&memcg->res);
5338 else if (type == _MEMSWAP)
5339 res_counter_reset_max(&memcg->memsw);
5340 else if (type == _KMEM)
5341 res_counter_reset_max(&memcg->kmem);
5347 res_counter_reset_failcnt(&memcg->res);
5348 else if (type == _MEMSWAP)
5349 res_counter_reset_failcnt(&memcg->memsw);
5350 else if (type == _KMEM)
5351 res_counter_reset_failcnt(&memcg->kmem);
5360 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5363 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5367 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5368 struct cftype *cft, u64 val)
5370 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5372 if (val >= (1 << NR_MOVE_TYPE))
5376 * No kind of locking is needed in here, because ->can_attach() will
5377 * check this value once in the beginning of the process, and then carry
5378 * on with stale data. This means that changes to this value will only
5379 * affect task migrations starting after the change.
5381 memcg->move_charge_at_immigrate = val;
5385 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5386 struct cftype *cft, u64 val)
5393 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5397 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5398 unsigned long node_nr;
5399 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5401 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5402 seq_printf(m, "total=%lu", total_nr);
5403 for_each_node_state(nid, N_MEMORY) {
5404 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5405 seq_printf(m, " N%d=%lu", nid, node_nr);
5409 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5410 seq_printf(m, "file=%lu", file_nr);
5411 for_each_node_state(nid, N_MEMORY) {
5412 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5414 seq_printf(m, " N%d=%lu", nid, node_nr);
5418 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5419 seq_printf(m, "anon=%lu", anon_nr);
5420 for_each_node_state(nid, N_MEMORY) {
5421 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5423 seq_printf(m, " N%d=%lu", nid, node_nr);
5427 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5428 seq_printf(m, "unevictable=%lu", unevictable_nr);
5429 for_each_node_state(nid, N_MEMORY) {
5430 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5431 BIT(LRU_UNEVICTABLE));
5432 seq_printf(m, " N%d=%lu", nid, node_nr);
5437 #endif /* CONFIG_NUMA */
5439 static inline void mem_cgroup_lru_names_not_uptodate(void)
5441 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5444 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5447 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5448 struct mem_cgroup *mi;
5451 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5452 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5454 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5455 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5458 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5459 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5460 mem_cgroup_read_events(memcg, i));
5462 for (i = 0; i < NR_LRU_LISTS; i++)
5463 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5464 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5466 /* Hierarchical information */
5468 unsigned long long limit, memsw_limit;
5469 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5470 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5471 if (do_swap_account)
5472 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5476 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5479 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5481 for_each_mem_cgroup_tree(mi, memcg)
5482 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5483 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5486 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5487 unsigned long long val = 0;
5489 for_each_mem_cgroup_tree(mi, memcg)
5490 val += mem_cgroup_read_events(mi, i);
5491 seq_printf(m, "total_%s %llu\n",
5492 mem_cgroup_events_names[i], val);
5495 for (i = 0; i < NR_LRU_LISTS; i++) {
5496 unsigned long long val = 0;
5498 for_each_mem_cgroup_tree(mi, memcg)
5499 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5500 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5503 #ifdef CONFIG_DEBUG_VM
5506 struct mem_cgroup_per_zone *mz;
5507 struct zone_reclaim_stat *rstat;
5508 unsigned long recent_rotated[2] = {0, 0};
5509 unsigned long recent_scanned[2] = {0, 0};
5511 for_each_online_node(nid)
5512 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5513 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5514 rstat = &mz->lruvec.reclaim_stat;
5516 recent_rotated[0] += rstat->recent_rotated[0];
5517 recent_rotated[1] += rstat->recent_rotated[1];
5518 recent_scanned[0] += rstat->recent_scanned[0];
5519 recent_scanned[1] += rstat->recent_scanned[1];
5521 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5522 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5523 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5524 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5531 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5533 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5535 return mem_cgroup_swappiness(memcg);
5538 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5541 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5542 struct mem_cgroup *parent;
5547 if (cgrp->parent == NULL)
5550 parent = mem_cgroup_from_cont(cgrp->parent);
5552 mutex_lock(&memcg_create_mutex);
5554 /* If under hierarchy, only empty-root can set this value */
5555 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5556 mutex_unlock(&memcg_create_mutex);
5560 memcg->swappiness = val;
5562 mutex_unlock(&memcg_create_mutex);
5567 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5569 struct mem_cgroup_threshold_ary *t;
5575 t = rcu_dereference(memcg->thresholds.primary);
5577 t = rcu_dereference(memcg->memsw_thresholds.primary);
5582 usage = mem_cgroup_usage(memcg, swap);
5585 * current_threshold points to threshold just below or equal to usage.
5586 * If it's not true, a threshold was crossed after last
5587 * call of __mem_cgroup_threshold().
5589 i = t->current_threshold;
5592 * Iterate backward over array of thresholds starting from
5593 * current_threshold and check if a threshold is crossed.
5594 * If none of thresholds below usage is crossed, we read
5595 * only one element of the array here.
5597 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5598 eventfd_signal(t->entries[i].eventfd, 1);
5600 /* i = current_threshold + 1 */
5604 * Iterate forward over array of thresholds starting from
5605 * current_threshold+1 and check if a threshold is crossed.
5606 * If none of thresholds above usage is crossed, we read
5607 * only one element of the array here.
5609 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5610 eventfd_signal(t->entries[i].eventfd, 1);
5612 /* Update current_threshold */
5613 t->current_threshold = i - 1;
5618 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5621 __mem_cgroup_threshold(memcg, false);
5622 if (do_swap_account)
5623 __mem_cgroup_threshold(memcg, true);
5625 memcg = parent_mem_cgroup(memcg);
5629 static int compare_thresholds(const void *a, const void *b)
5631 const struct mem_cgroup_threshold *_a = a;
5632 const struct mem_cgroup_threshold *_b = b;
5634 return _a->threshold - _b->threshold;
5637 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5639 struct mem_cgroup_eventfd_list *ev;
5641 list_for_each_entry(ev, &memcg->oom_notify, list)
5642 eventfd_signal(ev->eventfd, 1);
5646 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5648 struct mem_cgroup *iter;
5650 for_each_mem_cgroup_tree(iter, memcg)
5651 mem_cgroup_oom_notify_cb(iter);
5654 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5655 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5657 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5658 struct mem_cgroup_thresholds *thresholds;
5659 struct mem_cgroup_threshold_ary *new;
5660 enum res_type type = MEMFILE_TYPE(cft->private);
5661 u64 threshold, usage;
5664 ret = res_counter_memparse_write_strategy(args, &threshold);
5668 mutex_lock(&memcg->thresholds_lock);
5671 thresholds = &memcg->thresholds;
5672 else if (type == _MEMSWAP)
5673 thresholds = &memcg->memsw_thresholds;
5677 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5679 /* Check if a threshold crossed before adding a new one */
5680 if (thresholds->primary)
5681 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5683 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5685 /* Allocate memory for new array of thresholds */
5686 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5694 /* Copy thresholds (if any) to new array */
5695 if (thresholds->primary) {
5696 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5697 sizeof(struct mem_cgroup_threshold));
5700 /* Add new threshold */
5701 new->entries[size - 1].eventfd = eventfd;
5702 new->entries[size - 1].threshold = threshold;
5704 /* Sort thresholds. Registering of new threshold isn't time-critical */
5705 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5706 compare_thresholds, NULL);
5708 /* Find current threshold */
5709 new->current_threshold = -1;
5710 for (i = 0; i < size; i++) {
5711 if (new->entries[i].threshold <= usage) {
5713 * new->current_threshold will not be used until
5714 * rcu_assign_pointer(), so it's safe to increment
5717 ++new->current_threshold;
5722 /* Free old spare buffer and save old primary buffer as spare */
5723 kfree(thresholds->spare);
5724 thresholds->spare = thresholds->primary;
5726 rcu_assign_pointer(thresholds->primary, new);
5728 /* To be sure that nobody uses thresholds */
5732 mutex_unlock(&memcg->thresholds_lock);
5737 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5738 struct cftype *cft, struct eventfd_ctx *eventfd)
5740 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5741 struct mem_cgroup_thresholds *thresholds;
5742 struct mem_cgroup_threshold_ary *new;
5743 enum res_type type = MEMFILE_TYPE(cft->private);
5747 mutex_lock(&memcg->thresholds_lock);
5749 thresholds = &memcg->thresholds;
5750 else if (type == _MEMSWAP)
5751 thresholds = &memcg->memsw_thresholds;
5755 if (!thresholds->primary)
5758 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5760 /* Check if a threshold crossed before removing */
5761 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5763 /* Calculate new number of threshold */
5765 for (i = 0; i < thresholds->primary->size; i++) {
5766 if (thresholds->primary->entries[i].eventfd != eventfd)
5770 new = thresholds->spare;
5772 /* Set thresholds array to NULL if we don't have thresholds */
5781 /* Copy thresholds and find current threshold */
5782 new->current_threshold = -1;
5783 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5784 if (thresholds->primary->entries[i].eventfd == eventfd)
5787 new->entries[j] = thresholds->primary->entries[i];
5788 if (new->entries[j].threshold <= usage) {
5790 * new->current_threshold will not be used
5791 * until rcu_assign_pointer(), so it's safe to increment
5794 ++new->current_threshold;
5800 /* Swap primary and spare array */
5801 thresholds->spare = thresholds->primary;
5802 /* If all events are unregistered, free the spare array */
5804 kfree(thresholds->spare);
5805 thresholds->spare = NULL;
5808 rcu_assign_pointer(thresholds->primary, new);
5810 /* To be sure that nobody uses thresholds */
5813 mutex_unlock(&memcg->thresholds_lock);
5816 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5817 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5819 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5820 struct mem_cgroup_eventfd_list *event;
5821 enum res_type type = MEMFILE_TYPE(cft->private);
5823 BUG_ON(type != _OOM_TYPE);
5824 event = kmalloc(sizeof(*event), GFP_KERNEL);
5828 spin_lock(&memcg_oom_lock);
5830 event->eventfd = eventfd;
5831 list_add(&event->list, &memcg->oom_notify);
5833 /* already in OOM ? */
5834 if (atomic_read(&memcg->under_oom))
5835 eventfd_signal(eventfd, 1);
5836 spin_unlock(&memcg_oom_lock);
5841 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5842 struct cftype *cft, struct eventfd_ctx *eventfd)
5844 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5845 struct mem_cgroup_eventfd_list *ev, *tmp;
5846 enum res_type type = MEMFILE_TYPE(cft->private);
5848 BUG_ON(type != _OOM_TYPE);
5850 spin_lock(&memcg_oom_lock);
5852 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5853 if (ev->eventfd == eventfd) {
5854 list_del(&ev->list);
5859 spin_unlock(&memcg_oom_lock);
5862 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5863 struct cftype *cft, struct cgroup_map_cb *cb)
5865 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5867 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5869 if (atomic_read(&memcg->under_oom))
5870 cb->fill(cb, "under_oom", 1);
5872 cb->fill(cb, "under_oom", 0);
5876 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5877 struct cftype *cft, u64 val)
5879 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5880 struct mem_cgroup *parent;
5882 /* cannot set to root cgroup and only 0 and 1 are allowed */
5883 if (!cgrp->parent || !((val == 0) || (val == 1)))
5886 parent = mem_cgroup_from_cont(cgrp->parent);
5888 mutex_lock(&memcg_create_mutex);
5889 /* oom-kill-disable is a flag for subhierarchy. */
5890 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5891 mutex_unlock(&memcg_create_mutex);
5894 memcg->oom_kill_disable = val;
5896 memcg_oom_recover(memcg);
5897 mutex_unlock(&memcg_create_mutex);
5901 #ifdef CONFIG_MEMCG_KMEM
5902 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5906 memcg->kmemcg_id = -1;
5907 ret = memcg_propagate_kmem(memcg);
5911 return mem_cgroup_sockets_init(memcg, ss);
5914 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5916 mem_cgroup_sockets_destroy(memcg);
5918 memcg_kmem_mark_dead(memcg);
5920 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5924 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5925 * path here, being careful not to race with memcg_uncharge_kmem: it is
5926 * possible that the charges went down to 0 between mark_dead and the
5927 * res_counter read, so in that case, we don't need the put
5929 if (memcg_kmem_test_and_clear_dead(memcg))
5930 mem_cgroup_put(memcg);
5933 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5938 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5943 static struct cftype mem_cgroup_files[] = {
5945 .name = "usage_in_bytes",
5946 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5947 .read = mem_cgroup_read,
5948 .register_event = mem_cgroup_usage_register_event,
5949 .unregister_event = mem_cgroup_usage_unregister_event,
5952 .name = "max_usage_in_bytes",
5953 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5954 .trigger = mem_cgroup_reset,
5955 .read = mem_cgroup_read,
5958 .name = "limit_in_bytes",
5959 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5960 .write_string = mem_cgroup_write,
5961 .read = mem_cgroup_read,
5964 .name = "soft_limit_in_bytes",
5965 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5966 .write_string = mem_cgroup_write,
5967 .read = mem_cgroup_read,
5971 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5972 .trigger = mem_cgroup_reset,
5973 .read = mem_cgroup_read,
5977 .read_seq_string = memcg_stat_show,
5980 .name = "force_empty",
5981 .trigger = mem_cgroup_force_empty_write,
5984 .name = "use_hierarchy",
5985 .flags = CFTYPE_INSANE,
5986 .write_u64 = mem_cgroup_hierarchy_write,
5987 .read_u64 = mem_cgroup_hierarchy_read,
5990 .name = "swappiness",
5991 .read_u64 = mem_cgroup_swappiness_read,
5992 .write_u64 = mem_cgroup_swappiness_write,
5995 .name = "move_charge_at_immigrate",
5996 .read_u64 = mem_cgroup_move_charge_read,
5997 .write_u64 = mem_cgroup_move_charge_write,
6000 .name = "oom_control",
6001 .read_map = mem_cgroup_oom_control_read,
6002 .write_u64 = mem_cgroup_oom_control_write,
6003 .register_event = mem_cgroup_oom_register_event,
6004 .unregister_event = mem_cgroup_oom_unregister_event,
6005 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6008 .name = "pressure_level",
6009 .register_event = vmpressure_register_event,
6010 .unregister_event = vmpressure_unregister_event,
6014 .name = "numa_stat",
6015 .read_seq_string = memcg_numa_stat_show,
6018 #ifdef CONFIG_MEMCG_KMEM
6020 .name = "kmem.limit_in_bytes",
6021 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6022 .write_string = mem_cgroup_write,
6023 .read = mem_cgroup_read,
6026 .name = "kmem.usage_in_bytes",
6027 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6028 .read = mem_cgroup_read,
6031 .name = "kmem.failcnt",
6032 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6033 .trigger = mem_cgroup_reset,
6034 .read = mem_cgroup_read,
6037 .name = "kmem.max_usage_in_bytes",
6038 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6039 .trigger = mem_cgroup_reset,
6040 .read = mem_cgroup_read,
6042 #ifdef CONFIG_SLABINFO
6044 .name = "kmem.slabinfo",
6045 .read_seq_string = mem_cgroup_slabinfo_read,
6049 { }, /* terminate */
6052 #ifdef CONFIG_MEMCG_SWAP
6053 static struct cftype memsw_cgroup_files[] = {
6055 .name = "memsw.usage_in_bytes",
6056 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6057 .read = mem_cgroup_read,
6058 .register_event = mem_cgroup_usage_register_event,
6059 .unregister_event = mem_cgroup_usage_unregister_event,
6062 .name = "memsw.max_usage_in_bytes",
6063 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6064 .trigger = mem_cgroup_reset,
6065 .read = mem_cgroup_read,
6068 .name = "memsw.limit_in_bytes",
6069 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6070 .write_string = mem_cgroup_write,
6071 .read = mem_cgroup_read,
6074 .name = "memsw.failcnt",
6075 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6076 .trigger = mem_cgroup_reset,
6077 .read = mem_cgroup_read,
6079 { }, /* terminate */
6082 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6084 struct mem_cgroup_per_node *pn;
6085 struct mem_cgroup_per_zone *mz;
6086 int zone, tmp = node;
6088 * This routine is called against possible nodes.
6089 * But it's BUG to call kmalloc() against offline node.
6091 * TODO: this routine can waste much memory for nodes which will
6092 * never be onlined. It's better to use memory hotplug callback
6095 if (!node_state(node, N_NORMAL_MEMORY))
6097 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6101 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6102 mz = &pn->zoneinfo[zone];
6103 lruvec_init(&mz->lruvec);
6104 mz->usage_in_excess = 0;
6105 mz->on_tree = false;
6108 memcg->nodeinfo[node] = pn;
6112 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6114 kfree(memcg->nodeinfo[node]);
6117 static struct mem_cgroup *mem_cgroup_alloc(void)
6119 struct mem_cgroup *memcg;
6120 size_t size = memcg_size();
6122 /* Can be very big if nr_node_ids is very big */
6123 if (size < PAGE_SIZE)
6124 memcg = kzalloc(size, GFP_KERNEL);
6126 memcg = vzalloc(size);
6131 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6134 spin_lock_init(&memcg->pcp_counter_lock);
6138 if (size < PAGE_SIZE)
6146 * At destroying mem_cgroup, references from swap_cgroup can remain.
6147 * (scanning all at force_empty is too costly...)
6149 * Instead of clearing all references at force_empty, we remember
6150 * the number of reference from swap_cgroup and free mem_cgroup when
6151 * it goes down to 0.
6153 * Removal of cgroup itself succeeds regardless of refs from swap.
6156 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6159 size_t size = memcg_size();
6161 mem_cgroup_remove_from_trees(memcg);
6162 free_css_id(&mem_cgroup_subsys, &memcg->css);
6165 free_mem_cgroup_per_zone_info(memcg, node);
6167 free_percpu(memcg->stat);
6170 * We need to make sure that (at least for now), the jump label
6171 * destruction code runs outside of the cgroup lock. This is because
6172 * get_online_cpus(), which is called from the static_branch update,
6173 * can't be called inside the cgroup_lock. cpusets are the ones
6174 * enforcing this dependency, so if they ever change, we might as well.
6176 * schedule_work() will guarantee this happens. Be careful if you need
6177 * to move this code around, and make sure it is outside
6180 disarm_static_keys(memcg);
6181 if (size < PAGE_SIZE)
6189 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
6190 * but in process context. The work_freeing structure is overlaid
6191 * on the rcu_freeing structure, which itself is overlaid on memsw.
6193 static void free_work(struct work_struct *work)
6195 struct mem_cgroup *memcg;
6197 memcg = container_of(work, struct mem_cgroup, work_freeing);
6198 __mem_cgroup_free(memcg);
6201 static void free_rcu(struct rcu_head *rcu_head)
6203 struct mem_cgroup *memcg;
6205 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
6206 INIT_WORK(&memcg->work_freeing, free_work);
6207 schedule_work(&memcg->work_freeing);
6210 static void mem_cgroup_get(struct mem_cgroup *memcg)
6212 atomic_inc(&memcg->refcnt);
6215 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
6217 if (atomic_sub_and_test(count, &memcg->refcnt)) {
6218 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
6219 call_rcu(&memcg->rcu_freeing, free_rcu);
6221 mem_cgroup_put(parent);
6225 static void mem_cgroup_put(struct mem_cgroup *memcg)
6227 __mem_cgroup_put(memcg, 1);
6231 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6233 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6235 if (!memcg->res.parent)
6237 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6239 EXPORT_SYMBOL(parent_mem_cgroup);
6241 static void __init mem_cgroup_soft_limit_tree_init(void)
6243 struct mem_cgroup_tree_per_node *rtpn;
6244 struct mem_cgroup_tree_per_zone *rtpz;
6245 int tmp, node, zone;
6247 for_each_node(node) {
6249 if (!node_state(node, N_NORMAL_MEMORY))
6251 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6254 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6256 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6257 rtpz = &rtpn->rb_tree_per_zone[zone];
6258 rtpz->rb_root = RB_ROOT;
6259 spin_lock_init(&rtpz->lock);
6264 static struct cgroup_subsys_state * __ref
6265 mem_cgroup_css_alloc(struct cgroup *cont)
6267 struct mem_cgroup *memcg;
6268 long error = -ENOMEM;
6271 memcg = mem_cgroup_alloc();
6273 return ERR_PTR(error);
6276 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6280 if (cont->parent == NULL) {
6281 root_mem_cgroup = memcg;
6282 res_counter_init(&memcg->res, NULL);
6283 res_counter_init(&memcg->memsw, NULL);
6284 res_counter_init(&memcg->kmem, NULL);
6287 memcg->last_scanned_node = MAX_NUMNODES;
6288 INIT_LIST_HEAD(&memcg->oom_notify);
6289 atomic_set(&memcg->refcnt, 1);
6290 memcg->move_charge_at_immigrate = 0;
6291 mutex_init(&memcg->thresholds_lock);
6292 spin_lock_init(&memcg->move_lock);
6293 vmpressure_init(&memcg->vmpressure);
6298 __mem_cgroup_free(memcg);
6299 return ERR_PTR(error);
6303 mem_cgroup_css_online(struct cgroup *cont)
6305 struct mem_cgroup *memcg, *parent;
6311 mutex_lock(&memcg_create_mutex);
6312 memcg = mem_cgroup_from_cont(cont);
6313 parent = mem_cgroup_from_cont(cont->parent);
6315 memcg->use_hierarchy = parent->use_hierarchy;
6316 memcg->oom_kill_disable = parent->oom_kill_disable;
6317 memcg->swappiness = mem_cgroup_swappiness(parent);
6319 if (parent->use_hierarchy) {
6320 res_counter_init(&memcg->res, &parent->res);
6321 res_counter_init(&memcg->memsw, &parent->memsw);
6322 res_counter_init(&memcg->kmem, &parent->kmem);
6325 * We increment refcnt of the parent to ensure that we can
6326 * safely access it on res_counter_charge/uncharge.
6327 * This refcnt will be decremented when freeing this
6328 * mem_cgroup(see mem_cgroup_put).
6330 mem_cgroup_get(parent);
6332 res_counter_init(&memcg->res, NULL);
6333 res_counter_init(&memcg->memsw, NULL);
6334 res_counter_init(&memcg->kmem, NULL);
6336 * Deeper hierachy with use_hierarchy == false doesn't make
6337 * much sense so let cgroup subsystem know about this
6338 * unfortunate state in our controller.
6340 if (parent != root_mem_cgroup)
6341 mem_cgroup_subsys.broken_hierarchy = true;
6344 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6345 mutex_unlock(&memcg_create_mutex);
6348 * We call put now because our (and parent's) refcnts
6349 * are already in place. mem_cgroup_put() will internally
6350 * call __mem_cgroup_free, so return directly
6352 mem_cgroup_put(memcg);
6353 if (parent->use_hierarchy)
6354 mem_cgroup_put(parent);
6360 * Announce all parents that a group from their hierarchy is gone.
6362 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6364 struct mem_cgroup *parent = memcg;
6366 while ((parent = parent_mem_cgroup(parent)))
6367 mem_cgroup_iter_invalidate(parent);
6370 * if the root memcg is not hierarchical we have to check it
6373 if (!root_mem_cgroup->use_hierarchy)
6374 mem_cgroup_iter_invalidate(root_mem_cgroup);
6377 static void mem_cgroup_css_offline(struct cgroup *cont)
6379 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6381 mem_cgroup_invalidate_reclaim_iterators(memcg);
6382 mem_cgroup_reparent_charges(memcg);
6383 mem_cgroup_destroy_all_caches(memcg);
6386 static void mem_cgroup_css_free(struct cgroup *cont)
6388 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6390 kmem_cgroup_destroy(memcg);
6392 mem_cgroup_put(memcg);
6396 /* Handlers for move charge at task migration. */
6397 #define PRECHARGE_COUNT_AT_ONCE 256
6398 static int mem_cgroup_do_precharge(unsigned long count)
6401 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6402 struct mem_cgroup *memcg = mc.to;
6404 if (mem_cgroup_is_root(memcg)) {
6405 mc.precharge += count;
6406 /* we don't need css_get for root */
6409 /* try to charge at once */
6411 struct res_counter *dummy;
6413 * "memcg" cannot be under rmdir() because we've already checked
6414 * by cgroup_lock_live_cgroup() that it is not removed and we
6415 * are still under the same cgroup_mutex. So we can postpone
6418 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6420 if (do_swap_account && res_counter_charge(&memcg->memsw,
6421 PAGE_SIZE * count, &dummy)) {
6422 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6425 mc.precharge += count;
6429 /* fall back to one by one charge */
6431 if (signal_pending(current)) {
6435 if (!batch_count--) {
6436 batch_count = PRECHARGE_COUNT_AT_ONCE;
6439 ret = __mem_cgroup_try_charge(NULL,
6440 GFP_KERNEL, 1, &memcg, false);
6442 /* mem_cgroup_clear_mc() will do uncharge later */
6450 * get_mctgt_type - get target type of moving charge
6451 * @vma: the vma the pte to be checked belongs
6452 * @addr: the address corresponding to the pte to be checked
6453 * @ptent: the pte to be checked
6454 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6457 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6458 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6459 * move charge. if @target is not NULL, the page is stored in target->page
6460 * with extra refcnt got(Callers should handle it).
6461 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6462 * target for charge migration. if @target is not NULL, the entry is stored
6465 * Called with pte lock held.
6472 enum mc_target_type {
6478 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6479 unsigned long addr, pte_t ptent)
6481 struct page *page = vm_normal_page(vma, addr, ptent);
6483 if (!page || !page_mapped(page))
6485 if (PageAnon(page)) {
6486 /* we don't move shared anon */
6489 } else if (!move_file())
6490 /* we ignore mapcount for file pages */
6492 if (!get_page_unless_zero(page))
6499 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6500 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6502 struct page *page = NULL;
6503 swp_entry_t ent = pte_to_swp_entry(ptent);
6505 if (!move_anon() || non_swap_entry(ent))
6508 * Because lookup_swap_cache() updates some statistics counter,
6509 * we call find_get_page() with swapper_space directly.
6511 page = find_get_page(swap_address_space(ent), ent.val);
6512 if (do_swap_account)
6513 entry->val = ent.val;
6518 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6519 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6525 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6526 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6528 struct page *page = NULL;
6529 struct address_space *mapping;
6532 if (!vma->vm_file) /* anonymous vma */
6537 mapping = vma->vm_file->f_mapping;
6538 if (pte_none(ptent))
6539 pgoff = linear_page_index(vma, addr);
6540 else /* pte_file(ptent) is true */
6541 pgoff = pte_to_pgoff(ptent);
6543 /* page is moved even if it's not RSS of this task(page-faulted). */
6544 page = find_get_page(mapping, pgoff);
6547 /* shmem/tmpfs may report page out on swap: account for that too. */
6548 if (radix_tree_exceptional_entry(page)) {
6549 swp_entry_t swap = radix_to_swp_entry(page);
6550 if (do_swap_account)
6552 page = find_get_page(swap_address_space(swap), swap.val);
6558 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6559 unsigned long addr, pte_t ptent, union mc_target *target)
6561 struct page *page = NULL;
6562 struct page_cgroup *pc;
6563 enum mc_target_type ret = MC_TARGET_NONE;
6564 swp_entry_t ent = { .val = 0 };
6566 if (pte_present(ptent))
6567 page = mc_handle_present_pte(vma, addr, ptent);
6568 else if (is_swap_pte(ptent))
6569 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6570 else if (pte_none(ptent) || pte_file(ptent))
6571 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6573 if (!page && !ent.val)
6576 pc = lookup_page_cgroup(page);
6578 * Do only loose check w/o page_cgroup lock.
6579 * mem_cgroup_move_account() checks the pc is valid or not under
6582 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6583 ret = MC_TARGET_PAGE;
6585 target->page = page;
6587 if (!ret || !target)
6590 /* There is a swap entry and a page doesn't exist or isn't charged */
6591 if (ent.val && !ret &&
6592 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6593 ret = MC_TARGET_SWAP;
6600 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6602 * We don't consider swapping or file mapped pages because THP does not
6603 * support them for now.
6604 * Caller should make sure that pmd_trans_huge(pmd) is true.
6606 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6607 unsigned long addr, pmd_t pmd, union mc_target *target)
6609 struct page *page = NULL;
6610 struct page_cgroup *pc;
6611 enum mc_target_type ret = MC_TARGET_NONE;
6613 page = pmd_page(pmd);
6614 VM_BUG_ON(!page || !PageHead(page));
6617 pc = lookup_page_cgroup(page);
6618 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6619 ret = MC_TARGET_PAGE;
6622 target->page = page;
6628 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6629 unsigned long addr, pmd_t pmd, union mc_target *target)
6631 return MC_TARGET_NONE;
6635 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6636 unsigned long addr, unsigned long end,
6637 struct mm_walk *walk)
6639 struct vm_area_struct *vma = walk->private;
6643 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6644 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6645 mc.precharge += HPAGE_PMD_NR;
6646 spin_unlock(&vma->vm_mm->page_table_lock);
6650 if (pmd_trans_unstable(pmd))
6652 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6653 for (; addr != end; pte++, addr += PAGE_SIZE)
6654 if (get_mctgt_type(vma, addr, *pte, NULL))
6655 mc.precharge++; /* increment precharge temporarily */
6656 pte_unmap_unlock(pte - 1, ptl);
6662 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6664 unsigned long precharge;
6665 struct vm_area_struct *vma;
6667 down_read(&mm->mmap_sem);
6668 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6669 struct mm_walk mem_cgroup_count_precharge_walk = {
6670 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6674 if (is_vm_hugetlb_page(vma))
6676 walk_page_range(vma->vm_start, vma->vm_end,
6677 &mem_cgroup_count_precharge_walk);
6679 up_read(&mm->mmap_sem);
6681 precharge = mc.precharge;
6687 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6689 unsigned long precharge = mem_cgroup_count_precharge(mm);
6691 VM_BUG_ON(mc.moving_task);
6692 mc.moving_task = current;
6693 return mem_cgroup_do_precharge(precharge);
6696 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6697 static void __mem_cgroup_clear_mc(void)
6699 struct mem_cgroup *from = mc.from;
6700 struct mem_cgroup *to = mc.to;
6702 /* we must uncharge all the leftover precharges from mc.to */
6704 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6708 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6709 * we must uncharge here.
6711 if (mc.moved_charge) {
6712 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6713 mc.moved_charge = 0;
6715 /* we must fixup refcnts and charges */
6716 if (mc.moved_swap) {
6717 /* uncharge swap account from the old cgroup */
6718 if (!mem_cgroup_is_root(mc.from))
6719 res_counter_uncharge(&mc.from->memsw,
6720 PAGE_SIZE * mc.moved_swap);
6721 __mem_cgroup_put(mc.from, mc.moved_swap);
6723 if (!mem_cgroup_is_root(mc.to)) {
6725 * we charged both to->res and to->memsw, so we should
6728 res_counter_uncharge(&mc.to->res,
6729 PAGE_SIZE * mc.moved_swap);
6731 /* we've already done mem_cgroup_get(mc.to) */
6734 memcg_oom_recover(from);
6735 memcg_oom_recover(to);
6736 wake_up_all(&mc.waitq);
6739 static void mem_cgroup_clear_mc(void)
6741 struct mem_cgroup *from = mc.from;
6744 * we must clear moving_task before waking up waiters at the end of
6747 mc.moving_task = NULL;
6748 __mem_cgroup_clear_mc();
6749 spin_lock(&mc.lock);
6752 spin_unlock(&mc.lock);
6753 mem_cgroup_end_move(from);
6756 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6757 struct cgroup_taskset *tset)
6759 struct task_struct *p = cgroup_taskset_first(tset);
6761 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6762 unsigned long move_charge_at_immigrate;
6765 * We are now commited to this value whatever it is. Changes in this
6766 * tunable will only affect upcoming migrations, not the current one.
6767 * So we need to save it, and keep it going.
6769 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6770 if (move_charge_at_immigrate) {
6771 struct mm_struct *mm;
6772 struct mem_cgroup *from = mem_cgroup_from_task(p);
6774 VM_BUG_ON(from == memcg);
6776 mm = get_task_mm(p);
6779 /* We move charges only when we move a owner of the mm */
6780 if (mm->owner == p) {
6783 VM_BUG_ON(mc.precharge);
6784 VM_BUG_ON(mc.moved_charge);
6785 VM_BUG_ON(mc.moved_swap);
6786 mem_cgroup_start_move(from);
6787 spin_lock(&mc.lock);
6790 mc.immigrate_flags = move_charge_at_immigrate;
6791 spin_unlock(&mc.lock);
6792 /* We set mc.moving_task later */
6794 ret = mem_cgroup_precharge_mc(mm);
6796 mem_cgroup_clear_mc();
6803 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6804 struct cgroup_taskset *tset)
6806 mem_cgroup_clear_mc();
6809 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6810 unsigned long addr, unsigned long end,
6811 struct mm_walk *walk)
6814 struct vm_area_struct *vma = walk->private;
6817 enum mc_target_type target_type;
6818 union mc_target target;
6820 struct page_cgroup *pc;
6823 * We don't take compound_lock() here but no race with splitting thp
6825 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6826 * under splitting, which means there's no concurrent thp split,
6827 * - if another thread runs into split_huge_page() just after we
6828 * entered this if-block, the thread must wait for page table lock
6829 * to be unlocked in __split_huge_page_splitting(), where the main
6830 * part of thp split is not executed yet.
6832 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6833 if (mc.precharge < HPAGE_PMD_NR) {
6834 spin_unlock(&vma->vm_mm->page_table_lock);
6837 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6838 if (target_type == MC_TARGET_PAGE) {
6840 if (!isolate_lru_page(page)) {
6841 pc = lookup_page_cgroup(page);
6842 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6843 pc, mc.from, mc.to)) {
6844 mc.precharge -= HPAGE_PMD_NR;
6845 mc.moved_charge += HPAGE_PMD_NR;
6847 putback_lru_page(page);
6851 spin_unlock(&vma->vm_mm->page_table_lock);
6855 if (pmd_trans_unstable(pmd))
6858 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6859 for (; addr != end; addr += PAGE_SIZE) {
6860 pte_t ptent = *(pte++);
6866 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6867 case MC_TARGET_PAGE:
6869 if (isolate_lru_page(page))
6871 pc = lookup_page_cgroup(page);
6872 if (!mem_cgroup_move_account(page, 1, pc,
6875 /* we uncharge from mc.from later. */
6878 putback_lru_page(page);
6879 put: /* get_mctgt_type() gets the page */
6882 case MC_TARGET_SWAP:
6884 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6886 /* we fixup refcnts and charges later. */
6894 pte_unmap_unlock(pte - 1, ptl);
6899 * We have consumed all precharges we got in can_attach().
6900 * We try charge one by one, but don't do any additional
6901 * charges to mc.to if we have failed in charge once in attach()
6904 ret = mem_cgroup_do_precharge(1);
6912 static void mem_cgroup_move_charge(struct mm_struct *mm)
6914 struct vm_area_struct *vma;
6916 lru_add_drain_all();
6918 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6920 * Someone who are holding the mmap_sem might be waiting in
6921 * waitq. So we cancel all extra charges, wake up all waiters,
6922 * and retry. Because we cancel precharges, we might not be able
6923 * to move enough charges, but moving charge is a best-effort
6924 * feature anyway, so it wouldn't be a big problem.
6926 __mem_cgroup_clear_mc();
6930 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6932 struct mm_walk mem_cgroup_move_charge_walk = {
6933 .pmd_entry = mem_cgroup_move_charge_pte_range,
6937 if (is_vm_hugetlb_page(vma))
6939 ret = walk_page_range(vma->vm_start, vma->vm_end,
6940 &mem_cgroup_move_charge_walk);
6943 * means we have consumed all precharges and failed in
6944 * doing additional charge. Just abandon here.
6948 up_read(&mm->mmap_sem);
6951 static void mem_cgroup_move_task(struct cgroup *cont,
6952 struct cgroup_taskset *tset)
6954 struct task_struct *p = cgroup_taskset_first(tset);
6955 struct mm_struct *mm = get_task_mm(p);
6959 mem_cgroup_move_charge(mm);
6963 mem_cgroup_clear_mc();
6965 #else /* !CONFIG_MMU */
6966 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6967 struct cgroup_taskset *tset)
6971 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6972 struct cgroup_taskset *tset)
6975 static void mem_cgroup_move_task(struct cgroup *cont,
6976 struct cgroup_taskset *tset)
6982 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6983 * to verify sane_behavior flag on each mount attempt.
6985 static void mem_cgroup_bind(struct cgroup *root)
6988 * use_hierarchy is forced with sane_behavior. cgroup core
6989 * guarantees that @root doesn't have any children, so turning it
6990 * on for the root memcg is enough.
6992 if (cgroup_sane_behavior(root))
6993 mem_cgroup_from_cont(root)->use_hierarchy = true;
6996 struct cgroup_subsys mem_cgroup_subsys = {
6998 .subsys_id = mem_cgroup_subsys_id,
6999 .css_alloc = mem_cgroup_css_alloc,
7000 .css_online = mem_cgroup_css_online,
7001 .css_offline = mem_cgroup_css_offline,
7002 .css_free = mem_cgroup_css_free,
7003 .can_attach = mem_cgroup_can_attach,
7004 .cancel_attach = mem_cgroup_cancel_attach,
7005 .attach = mem_cgroup_move_task,
7006 .bind = mem_cgroup_bind,
7007 .base_cftypes = mem_cgroup_files,
7012 #ifdef CONFIG_MEMCG_SWAP
7013 static int __init enable_swap_account(char *s)
7015 /* consider enabled if no parameter or 1 is given */
7016 if (!strcmp(s, "1"))
7017 really_do_swap_account = 1;
7018 else if (!strcmp(s, "0"))
7019 really_do_swap_account = 0;
7022 __setup("swapaccount=", enable_swap_account);
7024 static void __init memsw_file_init(void)
7026 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7029 static void __init enable_swap_cgroup(void)
7031 if (!mem_cgroup_disabled() && really_do_swap_account) {
7032 do_swap_account = 1;
7038 static void __init enable_swap_cgroup(void)
7044 * subsys_initcall() for memory controller.
7046 * Some parts like hotcpu_notifier() have to be initialized from this context
7047 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7048 * everything that doesn't depend on a specific mem_cgroup structure should
7049 * be initialized from here.
7051 static int __init mem_cgroup_init(void)
7053 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7054 enable_swap_cgroup();
7055 mem_cgroup_soft_limit_tree_init();
7059 subsys_initcall(mem_cgroup_init);