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];
190 struct mem_cgroup_lru_info {
191 struct mem_cgroup_per_node *nodeinfo[0];
195 * Cgroups above their limits are maintained in a RB-Tree, independent of
196 * their hierarchy representation
199 struct mem_cgroup_tree_per_zone {
200 struct rb_root rb_root;
204 struct mem_cgroup_tree_per_node {
205 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
208 struct mem_cgroup_tree {
209 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
212 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
214 struct mem_cgroup_threshold {
215 struct eventfd_ctx *eventfd;
220 struct mem_cgroup_threshold_ary {
221 /* An array index points to threshold just below or equal to usage. */
222 int current_threshold;
223 /* Size of entries[] */
225 /* Array of thresholds */
226 struct mem_cgroup_threshold entries[0];
229 struct mem_cgroup_thresholds {
230 /* Primary thresholds array */
231 struct mem_cgroup_threshold_ary *primary;
233 * Spare threshold array.
234 * This is needed to make mem_cgroup_unregister_event() "never fail".
235 * It must be able to store at least primary->size - 1 entries.
237 struct mem_cgroup_threshold_ary *spare;
241 struct mem_cgroup_eventfd_list {
242 struct list_head list;
243 struct eventfd_ctx *eventfd;
246 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
247 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
250 * The memory controller data structure. The memory controller controls both
251 * page cache and RSS per cgroup. We would eventually like to provide
252 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
253 * to help the administrator determine what knobs to tune.
255 * TODO: Add a water mark for the memory controller. Reclaim will begin when
256 * we hit the water mark. May be even add a low water mark, such that
257 * no reclaim occurs from a cgroup at it's low water mark, this is
258 * a feature that will be implemented much later in the future.
261 struct cgroup_subsys_state css;
263 * the counter to account for memory usage
265 struct res_counter res;
267 /* vmpressure notifications */
268 struct vmpressure vmpressure;
272 * the counter to account for mem+swap usage.
274 struct res_counter memsw;
277 * rcu_freeing is used only when freeing struct mem_cgroup,
278 * so put it into a union to avoid wasting more memory.
279 * It must be disjoint from the css field. It could be
280 * in a union with the res field, but res plays a much
281 * larger part in mem_cgroup life than memsw, and might
282 * be of interest, even at time of free, when debugging.
283 * So share rcu_head with the less interesting memsw.
285 struct rcu_head rcu_freeing;
287 * We also need some space for a worker in deferred freeing.
288 * By the time we call it, rcu_freeing is no longer in use.
290 struct work_struct work_freeing;
294 * the counter to account for kernel memory usage.
296 struct res_counter kmem;
298 * Should the accounting and control be hierarchical, per subtree?
301 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
309 /* OOM-Killer disable */
310 int oom_kill_disable;
312 /* set when res.limit == memsw.limit */
313 bool memsw_is_minimum;
315 /* protect arrays of thresholds */
316 struct mutex thresholds_lock;
318 /* thresholds for memory usage. RCU-protected */
319 struct mem_cgroup_thresholds thresholds;
321 /* thresholds for mem+swap usage. RCU-protected */
322 struct mem_cgroup_thresholds memsw_thresholds;
324 /* For oom notifier event fd */
325 struct list_head oom_notify;
328 * Should we move charges of a task when a task is moved into this
329 * mem_cgroup ? And what type of charges should we move ?
331 unsigned long move_charge_at_immigrate;
333 * set > 0 if pages under this cgroup are moving to other cgroup.
335 atomic_t moving_account;
336 /* taken only while moving_account > 0 */
337 spinlock_t move_lock;
341 struct mem_cgroup_stat_cpu __percpu *stat;
343 * used when a cpu is offlined or other synchronizations
344 * See mem_cgroup_read_stat().
346 struct mem_cgroup_stat_cpu nocpu_base;
347 spinlock_t pcp_counter_lock;
350 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
351 struct tcp_memcontrol tcp_mem;
353 #if defined(CONFIG_MEMCG_KMEM)
354 /* analogous to slab_common's slab_caches list. per-memcg */
355 struct list_head memcg_slab_caches;
356 /* Not a spinlock, we can take a lot of time walking the list */
357 struct mutex slab_caches_mutex;
358 /* Index in the kmem_cache->memcg_params->memcg_caches array */
362 int last_scanned_node;
364 nodemask_t scan_nodes;
365 atomic_t numainfo_events;
366 atomic_t numainfo_updating;
370 * Per cgroup active and inactive list, similar to the
371 * per zone LRU lists.
373 * WARNING: This has to be the last element of the struct. Don't
374 * add new fields after this point.
376 struct mem_cgroup_lru_info info;
379 static size_t memcg_size(void)
381 return sizeof(struct mem_cgroup) +
382 nr_node_ids * sizeof(struct mem_cgroup_per_node);
385 /* internal only representation about the status of kmem accounting. */
387 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
388 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
389 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
392 /* We account when limit is on, but only after call sites are patched */
393 #define KMEM_ACCOUNTED_MASK \
394 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
396 #ifdef CONFIG_MEMCG_KMEM
397 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
399 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
402 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
404 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
407 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
409 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
412 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
414 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
417 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
419 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
420 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
423 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
425 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
426 &memcg->kmem_account_flags);
430 /* Stuffs for move charges at task migration. */
432 * Types of charges to be moved. "move_charge_at_immitgrate" and
433 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
436 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
437 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
441 /* "mc" and its members are protected by cgroup_mutex */
442 static struct move_charge_struct {
443 spinlock_t lock; /* for from, to */
444 struct mem_cgroup *from;
445 struct mem_cgroup *to;
446 unsigned long immigrate_flags;
447 unsigned long precharge;
448 unsigned long moved_charge;
449 unsigned long moved_swap;
450 struct task_struct *moving_task; /* a task moving charges */
451 wait_queue_head_t waitq; /* a waitq for other context */
453 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
454 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
457 static bool move_anon(void)
459 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
462 static bool move_file(void)
464 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
468 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
469 * limit reclaim to prevent infinite loops, if they ever occur.
471 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
472 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
475 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
476 MEM_CGROUP_CHARGE_TYPE_ANON,
477 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
478 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
482 /* for encoding cft->private value on file */
490 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
491 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
492 #define MEMFILE_ATTR(val) ((val) & 0xffff)
493 /* Used for OOM nofiier */
494 #define OOM_CONTROL (0)
497 * Reclaim flags for mem_cgroup_hierarchical_reclaim
499 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
500 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
501 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
502 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
505 * The memcg_create_mutex will be held whenever a new cgroup is created.
506 * As a consequence, any change that needs to protect against new child cgroups
507 * appearing has to hold it as well.
509 static DEFINE_MUTEX(memcg_create_mutex);
511 static void mem_cgroup_get(struct mem_cgroup *memcg);
512 static void mem_cgroup_put(struct mem_cgroup *memcg);
515 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
517 return container_of(s, struct mem_cgroup, css);
520 /* Some nice accessors for the vmpressure. */
521 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
524 memcg = root_mem_cgroup;
525 return &memcg->vmpressure;
528 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
530 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
533 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
535 return &mem_cgroup_from_css(css)->vmpressure;
538 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
540 return (memcg == root_mem_cgroup);
543 /* Writing them here to avoid exposing memcg's inner layout */
544 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
546 void sock_update_memcg(struct sock *sk)
548 if (mem_cgroup_sockets_enabled) {
549 struct mem_cgroup *memcg;
550 struct cg_proto *cg_proto;
552 BUG_ON(!sk->sk_prot->proto_cgroup);
554 /* Socket cloning can throw us here with sk_cgrp already
555 * filled. It won't however, necessarily happen from
556 * process context. So the test for root memcg given
557 * the current task's memcg won't help us in this case.
559 * Respecting the original socket's memcg is a better
560 * decision in this case.
563 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
564 mem_cgroup_get(sk->sk_cgrp->memcg);
569 memcg = mem_cgroup_from_task(current);
570 cg_proto = sk->sk_prot->proto_cgroup(memcg);
571 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
572 mem_cgroup_get(memcg);
573 sk->sk_cgrp = cg_proto;
578 EXPORT_SYMBOL(sock_update_memcg);
580 void sock_release_memcg(struct sock *sk)
582 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
583 struct mem_cgroup *memcg;
584 WARN_ON(!sk->sk_cgrp->memcg);
585 memcg = sk->sk_cgrp->memcg;
586 mem_cgroup_put(memcg);
590 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
592 if (!memcg || mem_cgroup_is_root(memcg))
595 return &memcg->tcp_mem.cg_proto;
597 EXPORT_SYMBOL(tcp_proto_cgroup);
599 static void disarm_sock_keys(struct mem_cgroup *memcg)
601 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
603 static_key_slow_dec(&memcg_socket_limit_enabled);
606 static void disarm_sock_keys(struct mem_cgroup *memcg)
611 #ifdef CONFIG_MEMCG_KMEM
613 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
614 * There are two main reasons for not using the css_id for this:
615 * 1) this works better in sparse environments, where we have a lot of memcgs,
616 * but only a few kmem-limited. Or also, if we have, for instance, 200
617 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
618 * 200 entry array for that.
620 * 2) In order not to violate the cgroup API, we would like to do all memory
621 * allocation in ->create(). At that point, we haven't yet allocated the
622 * css_id. Having a separate index prevents us from messing with the cgroup
625 * The current size of the caches array is stored in
626 * memcg_limited_groups_array_size. It will double each time we have to
629 static DEFINE_IDA(kmem_limited_groups);
630 int memcg_limited_groups_array_size;
633 * MIN_SIZE is different than 1, because we would like to avoid going through
634 * the alloc/free process all the time. In a small machine, 4 kmem-limited
635 * cgroups is a reasonable guess. In the future, it could be a parameter or
636 * tunable, but that is strictly not necessary.
638 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
639 * this constant directly from cgroup, but it is understandable that this is
640 * better kept as an internal representation in cgroup.c. In any case, the
641 * css_id space is not getting any smaller, and we don't have to necessarily
642 * increase ours as well if it increases.
644 #define MEMCG_CACHES_MIN_SIZE 4
645 #define MEMCG_CACHES_MAX_SIZE 65535
648 * A lot of the calls to the cache allocation functions are expected to be
649 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
650 * conditional to this static branch, we'll have to allow modules that does
651 * kmem_cache_alloc and the such to see this symbol as well
653 struct static_key memcg_kmem_enabled_key;
654 EXPORT_SYMBOL(memcg_kmem_enabled_key);
656 static void disarm_kmem_keys(struct mem_cgroup *memcg)
658 if (memcg_kmem_is_active(memcg)) {
659 static_key_slow_dec(&memcg_kmem_enabled_key);
660 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
663 * This check can't live in kmem destruction function,
664 * since the charges will outlive the cgroup
666 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
669 static void disarm_kmem_keys(struct mem_cgroup *memcg)
672 #endif /* CONFIG_MEMCG_KMEM */
674 static void disarm_static_keys(struct mem_cgroup *memcg)
676 disarm_sock_keys(memcg);
677 disarm_kmem_keys(memcg);
680 static void drain_all_stock_async(struct mem_cgroup *memcg);
682 static struct mem_cgroup_per_zone *
683 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
685 VM_BUG_ON((unsigned)nid >= nr_node_ids);
686 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
689 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
694 static struct mem_cgroup_per_zone *
695 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
697 int nid = page_to_nid(page);
698 int zid = page_zonenum(page);
700 return mem_cgroup_zoneinfo(memcg, nid, zid);
703 static struct mem_cgroup_tree_per_zone *
704 soft_limit_tree_node_zone(int nid, int zid)
706 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
709 static struct mem_cgroup_tree_per_zone *
710 soft_limit_tree_from_page(struct page *page)
712 int nid = page_to_nid(page);
713 int zid = page_zonenum(page);
715 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
719 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
720 struct mem_cgroup_per_zone *mz,
721 struct mem_cgroup_tree_per_zone *mctz,
722 unsigned long long new_usage_in_excess)
724 struct rb_node **p = &mctz->rb_root.rb_node;
725 struct rb_node *parent = NULL;
726 struct mem_cgroup_per_zone *mz_node;
731 mz->usage_in_excess = new_usage_in_excess;
732 if (!mz->usage_in_excess)
736 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
738 if (mz->usage_in_excess < mz_node->usage_in_excess)
741 * We can't avoid mem cgroups that are over their soft
742 * limit by the same amount
744 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
747 rb_link_node(&mz->tree_node, parent, p);
748 rb_insert_color(&mz->tree_node, &mctz->rb_root);
753 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
754 struct mem_cgroup_per_zone *mz,
755 struct mem_cgroup_tree_per_zone *mctz)
759 rb_erase(&mz->tree_node, &mctz->rb_root);
764 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
765 struct mem_cgroup_per_zone *mz,
766 struct mem_cgroup_tree_per_zone *mctz)
768 spin_lock(&mctz->lock);
769 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
770 spin_unlock(&mctz->lock);
774 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
776 unsigned long long excess;
777 struct mem_cgroup_per_zone *mz;
778 struct mem_cgroup_tree_per_zone *mctz;
779 int nid = page_to_nid(page);
780 int zid = page_zonenum(page);
781 mctz = soft_limit_tree_from_page(page);
784 * Necessary to update all ancestors when hierarchy is used.
785 * because their event counter is not touched.
787 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
788 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
789 excess = res_counter_soft_limit_excess(&memcg->res);
791 * We have to update the tree if mz is on RB-tree or
792 * mem is over its softlimit.
794 if (excess || mz->on_tree) {
795 spin_lock(&mctz->lock);
796 /* if on-tree, remove it */
798 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
800 * Insert again. mz->usage_in_excess will be updated.
801 * If excess is 0, no tree ops.
803 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
804 spin_unlock(&mctz->lock);
809 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
812 struct mem_cgroup_per_zone *mz;
813 struct mem_cgroup_tree_per_zone *mctz;
815 for_each_node(node) {
816 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
817 mz = mem_cgroup_zoneinfo(memcg, node, zone);
818 mctz = soft_limit_tree_node_zone(node, zone);
819 mem_cgroup_remove_exceeded(memcg, mz, mctz);
824 static struct mem_cgroup_per_zone *
825 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
827 struct rb_node *rightmost = NULL;
828 struct mem_cgroup_per_zone *mz;
832 rightmost = rb_last(&mctz->rb_root);
834 goto done; /* Nothing to reclaim from */
836 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
838 * Remove the node now but someone else can add it back,
839 * we will to add it back at the end of reclaim to its correct
840 * position in the tree.
842 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
843 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
844 !css_tryget(&mz->memcg->css))
850 static struct mem_cgroup_per_zone *
851 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
853 struct mem_cgroup_per_zone *mz;
855 spin_lock(&mctz->lock);
856 mz = __mem_cgroup_largest_soft_limit_node(mctz);
857 spin_unlock(&mctz->lock);
862 * Implementation Note: reading percpu statistics for memcg.
864 * Both of vmstat[] and percpu_counter has threshold and do periodic
865 * synchronization to implement "quick" read. There are trade-off between
866 * reading cost and precision of value. Then, we may have a chance to implement
867 * a periodic synchronizion of counter in memcg's counter.
869 * But this _read() function is used for user interface now. The user accounts
870 * memory usage by memory cgroup and he _always_ requires exact value because
871 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
872 * have to visit all online cpus and make sum. So, for now, unnecessary
873 * synchronization is not implemented. (just implemented for cpu hotplug)
875 * If there are kernel internal actions which can make use of some not-exact
876 * value, and reading all cpu value can be performance bottleneck in some
877 * common workload, threashold and synchonization as vmstat[] should be
880 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
881 enum mem_cgroup_stat_index idx)
887 for_each_online_cpu(cpu)
888 val += per_cpu(memcg->stat->count[idx], cpu);
889 #ifdef CONFIG_HOTPLUG_CPU
890 spin_lock(&memcg->pcp_counter_lock);
891 val += memcg->nocpu_base.count[idx];
892 spin_unlock(&memcg->pcp_counter_lock);
898 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
901 int val = (charge) ? 1 : -1;
902 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
905 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
906 enum mem_cgroup_events_index idx)
908 unsigned long val = 0;
911 for_each_online_cpu(cpu)
912 val += per_cpu(memcg->stat->events[idx], cpu);
913 #ifdef CONFIG_HOTPLUG_CPU
914 spin_lock(&memcg->pcp_counter_lock);
915 val += memcg->nocpu_base.events[idx];
916 spin_unlock(&memcg->pcp_counter_lock);
921 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
923 bool anon, int nr_pages)
928 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
929 * counted as CACHE even if it's on ANON LRU.
932 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
935 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
938 if (PageTransHuge(page))
939 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
942 /* pagein of a big page is an event. So, ignore page size */
944 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
946 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
947 nr_pages = -nr_pages; /* for event */
950 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
956 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
958 struct mem_cgroup_per_zone *mz;
960 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
961 return mz->lru_size[lru];
965 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
966 unsigned int lru_mask)
968 struct mem_cgroup_per_zone *mz;
970 unsigned long ret = 0;
972 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
975 if (BIT(lru) & lru_mask)
976 ret += mz->lru_size[lru];
982 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
983 int nid, unsigned int lru_mask)
988 for (zid = 0; zid < MAX_NR_ZONES; zid++)
989 total += mem_cgroup_zone_nr_lru_pages(memcg,
995 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
996 unsigned int lru_mask)
1001 for_each_node_state(nid, N_MEMORY)
1002 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1006 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1007 enum mem_cgroup_events_target target)
1009 unsigned long val, next;
1011 val = __this_cpu_read(memcg->stat->nr_page_events);
1012 next = __this_cpu_read(memcg->stat->targets[target]);
1013 /* from time_after() in jiffies.h */
1014 if ((long)next - (long)val < 0) {
1016 case MEM_CGROUP_TARGET_THRESH:
1017 next = val + THRESHOLDS_EVENTS_TARGET;
1019 case MEM_CGROUP_TARGET_SOFTLIMIT:
1020 next = val + SOFTLIMIT_EVENTS_TARGET;
1022 case MEM_CGROUP_TARGET_NUMAINFO:
1023 next = val + NUMAINFO_EVENTS_TARGET;
1028 __this_cpu_write(memcg->stat->targets[target], next);
1035 * Check events in order.
1038 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1041 /* threshold event is triggered in finer grain than soft limit */
1042 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1043 MEM_CGROUP_TARGET_THRESH))) {
1045 bool do_numainfo __maybe_unused;
1047 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1048 MEM_CGROUP_TARGET_SOFTLIMIT);
1049 #if MAX_NUMNODES > 1
1050 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1051 MEM_CGROUP_TARGET_NUMAINFO);
1055 mem_cgroup_threshold(memcg);
1056 if (unlikely(do_softlimit))
1057 mem_cgroup_update_tree(memcg, page);
1058 #if MAX_NUMNODES > 1
1059 if (unlikely(do_numainfo))
1060 atomic_inc(&memcg->numainfo_events);
1066 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1068 return mem_cgroup_from_css(
1069 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1072 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1075 * mm_update_next_owner() may clear mm->owner to NULL
1076 * if it races with swapoff, page migration, etc.
1077 * So this can be called with p == NULL.
1082 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1085 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1087 struct mem_cgroup *memcg = NULL;
1092 * Because we have no locks, mm->owner's may be being moved to other
1093 * cgroup. We use css_tryget() here even if this looks
1094 * pessimistic (rather than adding locks here).
1098 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1099 if (unlikely(!memcg))
1101 } while (!css_tryget(&memcg->css));
1107 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1108 * ref. count) or NULL if the whole root's subtree has been visited.
1110 * helper function to be used by mem_cgroup_iter
1112 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1113 struct mem_cgroup *last_visited)
1115 struct cgroup *prev_cgroup, *next_cgroup;
1118 * Root is not visited by cgroup iterators so it needs an
1124 prev_cgroup = (last_visited == root) ? NULL
1125 : last_visited->css.cgroup;
1127 next_cgroup = cgroup_next_descendant_pre(
1128 prev_cgroup, root->css.cgroup);
1131 * Even if we found a group we have to make sure it is
1132 * alive. css && !memcg means that the groups should be
1133 * skipped and we should continue the tree walk.
1134 * last_visited css is safe to use because it is
1135 * protected by css_get and the tree walk is rcu safe.
1138 struct mem_cgroup *mem = mem_cgroup_from_cont(
1140 if (css_tryget(&mem->css))
1143 prev_cgroup = next_cgroup;
1152 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1153 * @root: hierarchy root
1154 * @prev: previously returned memcg, NULL on first invocation
1155 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1157 * Returns references to children of the hierarchy below @root, or
1158 * @root itself, or %NULL after a full round-trip.
1160 * Caller must pass the return value in @prev on subsequent
1161 * invocations for reference counting, or use mem_cgroup_iter_break()
1162 * to cancel a hierarchy walk before the round-trip is complete.
1164 * Reclaimers can specify a zone and a priority level in @reclaim to
1165 * divide up the memcgs in the hierarchy among all concurrent
1166 * reclaimers operating on the same zone and priority.
1168 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1169 struct mem_cgroup *prev,
1170 struct mem_cgroup_reclaim_cookie *reclaim)
1172 struct mem_cgroup *memcg = NULL;
1173 struct mem_cgroup *last_visited = NULL;
1174 unsigned long uninitialized_var(dead_count);
1176 if (mem_cgroup_disabled())
1180 root = root_mem_cgroup;
1182 if (prev && !reclaim)
1183 last_visited = prev;
1185 if (!root->use_hierarchy && root != root_mem_cgroup) {
1193 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1196 int nid = zone_to_nid(reclaim->zone);
1197 int zid = zone_idx(reclaim->zone);
1198 struct mem_cgroup_per_zone *mz;
1200 mz = mem_cgroup_zoneinfo(root, nid, zid);
1201 iter = &mz->reclaim_iter[reclaim->priority];
1202 if (prev && reclaim->generation != iter->generation) {
1203 iter->last_visited = NULL;
1208 * If the dead_count mismatches, a destruction
1209 * has happened or is happening concurrently.
1210 * If the dead_count matches, a destruction
1211 * might still happen concurrently, but since
1212 * we checked under RCU, that destruction
1213 * won't free the object until we release the
1214 * RCU reader lock. Thus, the dead_count
1215 * check verifies the pointer is still valid,
1216 * css_tryget() verifies the cgroup pointed to
1219 dead_count = atomic_read(&root->dead_count);
1220 if (dead_count == iter->last_dead_count) {
1222 last_visited = iter->last_visited;
1224 !css_tryget(&last_visited->css))
1225 last_visited = NULL;
1229 memcg = __mem_cgroup_iter_next(root, last_visited);
1233 css_put(&last_visited->css);
1235 iter->last_visited = memcg;
1237 iter->last_dead_count = dead_count;
1241 else if (!prev && memcg)
1242 reclaim->generation = iter->generation;
1251 if (prev && prev != root)
1252 css_put(&prev->css);
1258 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1259 * @root: hierarchy root
1260 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1262 void mem_cgroup_iter_break(struct mem_cgroup *root,
1263 struct mem_cgroup *prev)
1266 root = root_mem_cgroup;
1267 if (prev && prev != root)
1268 css_put(&prev->css);
1272 * Iteration constructs for visiting all cgroups (under a tree). If
1273 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1274 * be used for reference counting.
1276 #define for_each_mem_cgroup_tree(iter, root) \
1277 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1279 iter = mem_cgroup_iter(root, iter, NULL))
1281 #define for_each_mem_cgroup(iter) \
1282 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1284 iter = mem_cgroup_iter(NULL, iter, NULL))
1286 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1288 struct mem_cgroup *memcg;
1291 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1292 if (unlikely(!memcg))
1297 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1300 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1308 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1311 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1312 * @zone: zone of the wanted lruvec
1313 * @memcg: memcg of the wanted lruvec
1315 * Returns the lru list vector holding pages for the given @zone and
1316 * @mem. This can be the global zone lruvec, if the memory controller
1319 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1320 struct mem_cgroup *memcg)
1322 struct mem_cgroup_per_zone *mz;
1323 struct lruvec *lruvec;
1325 if (mem_cgroup_disabled()) {
1326 lruvec = &zone->lruvec;
1330 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1331 lruvec = &mz->lruvec;
1334 * Since a node can be onlined after the mem_cgroup was created,
1335 * we have to be prepared to initialize lruvec->zone here;
1336 * and if offlined then reonlined, we need to reinitialize it.
1338 if (unlikely(lruvec->zone != zone))
1339 lruvec->zone = zone;
1344 * Following LRU functions are allowed to be used without PCG_LOCK.
1345 * Operations are called by routine of global LRU independently from memcg.
1346 * What we have to take care of here is validness of pc->mem_cgroup.
1348 * Changes to pc->mem_cgroup happens when
1351 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1352 * It is added to LRU before charge.
1353 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1354 * When moving account, the page is not on LRU. It's isolated.
1358 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1360 * @zone: zone of the page
1362 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1364 struct mem_cgroup_per_zone *mz;
1365 struct mem_cgroup *memcg;
1366 struct page_cgroup *pc;
1367 struct lruvec *lruvec;
1369 if (mem_cgroup_disabled()) {
1370 lruvec = &zone->lruvec;
1374 pc = lookup_page_cgroup(page);
1375 memcg = pc->mem_cgroup;
1378 * Surreptitiously switch any uncharged offlist page to root:
1379 * an uncharged page off lru does nothing to secure
1380 * its former mem_cgroup from sudden removal.
1382 * Our caller holds lru_lock, and PageCgroupUsed is updated
1383 * under page_cgroup lock: between them, they make all uses
1384 * of pc->mem_cgroup safe.
1386 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1387 pc->mem_cgroup = memcg = root_mem_cgroup;
1389 mz = page_cgroup_zoneinfo(memcg, page);
1390 lruvec = &mz->lruvec;
1393 * Since a node can be onlined after the mem_cgroup was created,
1394 * we have to be prepared to initialize lruvec->zone here;
1395 * and if offlined then reonlined, we need to reinitialize it.
1397 if (unlikely(lruvec->zone != zone))
1398 lruvec->zone = zone;
1403 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1404 * @lruvec: mem_cgroup per zone lru vector
1405 * @lru: index of lru list the page is sitting on
1406 * @nr_pages: positive when adding or negative when removing
1408 * This function must be called when a page is added to or removed from an
1411 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1414 struct mem_cgroup_per_zone *mz;
1415 unsigned long *lru_size;
1417 if (mem_cgroup_disabled())
1420 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1421 lru_size = mz->lru_size + lru;
1422 *lru_size += nr_pages;
1423 VM_BUG_ON((long)(*lru_size) < 0);
1427 * Checks whether given mem is same or in the root_mem_cgroup's
1430 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1431 struct mem_cgroup *memcg)
1433 if (root_memcg == memcg)
1435 if (!root_memcg->use_hierarchy || !memcg)
1437 return css_is_ancestor(&memcg->css, &root_memcg->css);
1440 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1441 struct mem_cgroup *memcg)
1446 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1451 bool task_in_mem_cgroup(struct task_struct *task,
1452 const struct mem_cgroup *memcg)
1454 struct mem_cgroup *curr = NULL;
1455 struct task_struct *p;
1458 p = find_lock_task_mm(task);
1460 curr = try_get_mem_cgroup_from_mm(p->mm);
1464 * All threads may have already detached their mm's, but the oom
1465 * killer still needs to detect if they have already been oom
1466 * killed to prevent needlessly killing additional tasks.
1469 curr = mem_cgroup_from_task(task);
1471 css_get(&curr->css);
1477 * We should check use_hierarchy of "memcg" not "curr". Because checking
1478 * use_hierarchy of "curr" here make this function true if hierarchy is
1479 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1480 * hierarchy(even if use_hierarchy is disabled in "memcg").
1482 ret = mem_cgroup_same_or_subtree(memcg, curr);
1483 css_put(&curr->css);
1487 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1489 unsigned long inactive_ratio;
1490 unsigned long inactive;
1491 unsigned long active;
1494 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1495 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1497 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1499 inactive_ratio = int_sqrt(10 * gb);
1503 return inactive * inactive_ratio < active;
1506 #define mem_cgroup_from_res_counter(counter, member) \
1507 container_of(counter, struct mem_cgroup, member)
1510 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1511 * @memcg: the memory cgroup
1513 * Returns the maximum amount of memory @mem can be charged with, in
1516 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1518 unsigned long long margin;
1520 margin = res_counter_margin(&memcg->res);
1521 if (do_swap_account)
1522 margin = min(margin, res_counter_margin(&memcg->memsw));
1523 return margin >> PAGE_SHIFT;
1526 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1528 struct cgroup *cgrp = memcg->css.cgroup;
1531 if (cgrp->parent == NULL)
1532 return vm_swappiness;
1534 return memcg->swappiness;
1538 * memcg->moving_account is used for checking possibility that some thread is
1539 * calling move_account(). When a thread on CPU-A starts moving pages under
1540 * a memcg, other threads should check memcg->moving_account under
1541 * rcu_read_lock(), like this:
1545 * memcg->moving_account+1 if (memcg->mocing_account)
1547 * synchronize_rcu() update something.
1552 /* for quick checking without looking up memcg */
1553 atomic_t memcg_moving __read_mostly;
1555 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1557 atomic_inc(&memcg_moving);
1558 atomic_inc(&memcg->moving_account);
1562 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1565 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1566 * We check NULL in callee rather than caller.
1569 atomic_dec(&memcg_moving);
1570 atomic_dec(&memcg->moving_account);
1575 * 2 routines for checking "mem" is under move_account() or not.
1577 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1578 * is used for avoiding races in accounting. If true,
1579 * pc->mem_cgroup may be overwritten.
1581 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1582 * under hierarchy of moving cgroups. This is for
1583 * waiting at hith-memory prressure caused by "move".
1586 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1588 VM_BUG_ON(!rcu_read_lock_held());
1589 return atomic_read(&memcg->moving_account) > 0;
1592 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1594 struct mem_cgroup *from;
1595 struct mem_cgroup *to;
1598 * Unlike task_move routines, we access mc.to, mc.from not under
1599 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1601 spin_lock(&mc.lock);
1607 ret = mem_cgroup_same_or_subtree(memcg, from)
1608 || mem_cgroup_same_or_subtree(memcg, to);
1610 spin_unlock(&mc.lock);
1614 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1616 if (mc.moving_task && current != mc.moving_task) {
1617 if (mem_cgroup_under_move(memcg)) {
1619 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1620 /* moving charge context might have finished. */
1623 finish_wait(&mc.waitq, &wait);
1631 * Take this lock when
1632 * - a code tries to modify page's memcg while it's USED.
1633 * - a code tries to modify page state accounting in a memcg.
1634 * see mem_cgroup_stolen(), too.
1636 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1637 unsigned long *flags)
1639 spin_lock_irqsave(&memcg->move_lock, *flags);
1642 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1643 unsigned long *flags)
1645 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1648 #define K(x) ((x) << (PAGE_SHIFT-10))
1650 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1651 * @memcg: The memory cgroup that went over limit
1652 * @p: Task that is going to be killed
1654 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1657 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1659 struct cgroup *task_cgrp;
1660 struct cgroup *mem_cgrp;
1662 * Need a buffer in BSS, can't rely on allocations. The code relies
1663 * on the assumption that OOM is serialized for memory controller.
1664 * If this assumption is broken, revisit this code.
1666 static char memcg_name[PATH_MAX];
1668 struct mem_cgroup *iter;
1676 mem_cgrp = memcg->css.cgroup;
1677 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1679 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1682 * Unfortunately, we are unable to convert to a useful name
1683 * But we'll still print out the usage information
1690 pr_info("Task in %s killed", memcg_name);
1693 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1701 * Continues from above, so we don't need an KERN_ level
1703 pr_cont(" as a result of limit of %s\n", memcg_name);
1706 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1707 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1708 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1709 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1710 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1711 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1712 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1713 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1714 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1715 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1716 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1717 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1719 for_each_mem_cgroup_tree(iter, memcg) {
1720 pr_info("Memory cgroup stats");
1723 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1725 pr_cont(" for %s", memcg_name);
1729 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1730 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1732 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1733 K(mem_cgroup_read_stat(iter, i)));
1736 for (i = 0; i < NR_LRU_LISTS; i++)
1737 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1738 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1745 * This function returns the number of memcg under hierarchy tree. Returns
1746 * 1(self count) if no children.
1748 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1751 struct mem_cgroup *iter;
1753 for_each_mem_cgroup_tree(iter, memcg)
1759 * Return the memory (and swap, if configured) limit for a memcg.
1761 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1765 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1768 * Do not consider swap space if we cannot swap due to swappiness
1770 if (mem_cgroup_swappiness(memcg)) {
1773 limit += total_swap_pages << PAGE_SHIFT;
1774 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1777 * If memsw is finite and limits the amount of swap space
1778 * available to this memcg, return that limit.
1780 limit = min(limit, memsw);
1786 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1789 struct mem_cgroup *iter;
1790 unsigned long chosen_points = 0;
1791 unsigned long totalpages;
1792 unsigned int points = 0;
1793 struct task_struct *chosen = NULL;
1796 * If current has a pending SIGKILL or is exiting, then automatically
1797 * select it. The goal is to allow it to allocate so that it may
1798 * quickly exit and free its memory.
1800 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1801 set_thread_flag(TIF_MEMDIE);
1805 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1806 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1807 for_each_mem_cgroup_tree(iter, memcg) {
1808 struct cgroup *cgroup = iter->css.cgroup;
1809 struct cgroup_iter it;
1810 struct task_struct *task;
1812 cgroup_iter_start(cgroup, &it);
1813 while ((task = cgroup_iter_next(cgroup, &it))) {
1814 switch (oom_scan_process_thread(task, totalpages, NULL,
1816 case OOM_SCAN_SELECT:
1818 put_task_struct(chosen);
1820 chosen_points = ULONG_MAX;
1821 get_task_struct(chosen);
1823 case OOM_SCAN_CONTINUE:
1825 case OOM_SCAN_ABORT:
1826 cgroup_iter_end(cgroup, &it);
1827 mem_cgroup_iter_break(memcg, iter);
1829 put_task_struct(chosen);
1834 points = oom_badness(task, memcg, NULL, totalpages);
1835 if (points > chosen_points) {
1837 put_task_struct(chosen);
1839 chosen_points = points;
1840 get_task_struct(chosen);
1843 cgroup_iter_end(cgroup, &it);
1848 points = chosen_points * 1000 / totalpages;
1849 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1850 NULL, "Memory cgroup out of memory");
1853 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1855 unsigned long flags)
1857 unsigned long total = 0;
1858 bool noswap = false;
1861 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1863 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1866 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1868 drain_all_stock_async(memcg);
1869 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1871 * Allow limit shrinkers, which are triggered directly
1872 * by userspace, to catch signals and stop reclaim
1873 * after minimal progress, regardless of the margin.
1875 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1877 if (mem_cgroup_margin(memcg))
1880 * If nothing was reclaimed after two attempts, there
1881 * may be no reclaimable pages in this hierarchy.
1890 * test_mem_cgroup_node_reclaimable
1891 * @memcg: the target memcg
1892 * @nid: the node ID to be checked.
1893 * @noswap : specify true here if the user wants flle only information.
1895 * This function returns whether the specified memcg contains any
1896 * reclaimable pages on a node. Returns true if there are any reclaimable
1897 * pages in the node.
1899 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1900 int nid, bool noswap)
1902 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1904 if (noswap || !total_swap_pages)
1906 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1911 #if MAX_NUMNODES > 1
1914 * Always updating the nodemask is not very good - even if we have an empty
1915 * list or the wrong list here, we can start from some node and traverse all
1916 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1919 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1923 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1924 * pagein/pageout changes since the last update.
1926 if (!atomic_read(&memcg->numainfo_events))
1928 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1931 /* make a nodemask where this memcg uses memory from */
1932 memcg->scan_nodes = node_states[N_MEMORY];
1934 for_each_node_mask(nid, node_states[N_MEMORY]) {
1936 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1937 node_clear(nid, memcg->scan_nodes);
1940 atomic_set(&memcg->numainfo_events, 0);
1941 atomic_set(&memcg->numainfo_updating, 0);
1945 * Selecting a node where we start reclaim from. Because what we need is just
1946 * reducing usage counter, start from anywhere is O,K. Considering
1947 * memory reclaim from current node, there are pros. and cons.
1949 * Freeing memory from current node means freeing memory from a node which
1950 * we'll use or we've used. So, it may make LRU bad. And if several threads
1951 * hit limits, it will see a contention on a node. But freeing from remote
1952 * node means more costs for memory reclaim because of memory latency.
1954 * Now, we use round-robin. Better algorithm is welcomed.
1956 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1960 mem_cgroup_may_update_nodemask(memcg);
1961 node = memcg->last_scanned_node;
1963 node = next_node(node, memcg->scan_nodes);
1964 if (node == MAX_NUMNODES)
1965 node = first_node(memcg->scan_nodes);
1967 * We call this when we hit limit, not when pages are added to LRU.
1968 * No LRU may hold pages because all pages are UNEVICTABLE or
1969 * memcg is too small and all pages are not on LRU. In that case,
1970 * we use curret node.
1972 if (unlikely(node == MAX_NUMNODES))
1973 node = numa_node_id();
1975 memcg->last_scanned_node = node;
1980 * Check all nodes whether it contains reclaimable pages or not.
1981 * For quick scan, we make use of scan_nodes. This will allow us to skip
1982 * unused nodes. But scan_nodes is lazily updated and may not cotain
1983 * enough new information. We need to do double check.
1985 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1990 * quick check...making use of scan_node.
1991 * We can skip unused nodes.
1993 if (!nodes_empty(memcg->scan_nodes)) {
1994 for (nid = first_node(memcg->scan_nodes);
1996 nid = next_node(nid, memcg->scan_nodes)) {
1998 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2003 * Check rest of nodes.
2005 for_each_node_state(nid, N_MEMORY) {
2006 if (node_isset(nid, memcg->scan_nodes))
2008 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2015 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2020 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2022 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2026 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2029 unsigned long *total_scanned)
2031 struct mem_cgroup *victim = NULL;
2034 unsigned long excess;
2035 unsigned long nr_scanned;
2036 struct mem_cgroup_reclaim_cookie reclaim = {
2041 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2044 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2049 * If we have not been able to reclaim
2050 * anything, it might because there are
2051 * no reclaimable pages under this hierarchy
2056 * We want to do more targeted reclaim.
2057 * excess >> 2 is not to excessive so as to
2058 * reclaim too much, nor too less that we keep
2059 * coming back to reclaim from this cgroup
2061 if (total >= (excess >> 2) ||
2062 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2067 if (!mem_cgroup_reclaimable(victim, false))
2069 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2071 *total_scanned += nr_scanned;
2072 if (!res_counter_soft_limit_excess(&root_memcg->res))
2075 mem_cgroup_iter_break(root_memcg, victim);
2080 * Check OOM-Killer is already running under our hierarchy.
2081 * If someone is running, return false.
2082 * Has to be called with memcg_oom_lock
2084 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
2086 struct mem_cgroup *iter, *failed = NULL;
2088 for_each_mem_cgroup_tree(iter, memcg) {
2089 if (iter->oom_lock) {
2091 * this subtree of our hierarchy is already locked
2092 * so we cannot give a lock.
2095 mem_cgroup_iter_break(memcg, iter);
2098 iter->oom_lock = true;
2105 * OK, we failed to lock the whole subtree so we have to clean up
2106 * what we set up to the failing subtree
2108 for_each_mem_cgroup_tree(iter, memcg) {
2109 if (iter == failed) {
2110 mem_cgroup_iter_break(memcg, iter);
2113 iter->oom_lock = false;
2119 * Has to be called with memcg_oom_lock
2121 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2123 struct mem_cgroup *iter;
2125 for_each_mem_cgroup_tree(iter, memcg)
2126 iter->oom_lock = false;
2130 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2132 struct mem_cgroup *iter;
2134 for_each_mem_cgroup_tree(iter, memcg)
2135 atomic_inc(&iter->under_oom);
2138 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2140 struct mem_cgroup *iter;
2143 * When a new child is created while the hierarchy is under oom,
2144 * mem_cgroup_oom_lock() may not be called. We have to use
2145 * atomic_add_unless() here.
2147 for_each_mem_cgroup_tree(iter, memcg)
2148 atomic_add_unless(&iter->under_oom, -1, 0);
2151 static DEFINE_SPINLOCK(memcg_oom_lock);
2152 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2154 struct oom_wait_info {
2155 struct mem_cgroup *memcg;
2159 static int memcg_oom_wake_function(wait_queue_t *wait,
2160 unsigned mode, int sync, void *arg)
2162 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2163 struct mem_cgroup *oom_wait_memcg;
2164 struct oom_wait_info *oom_wait_info;
2166 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2167 oom_wait_memcg = oom_wait_info->memcg;
2170 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2171 * Then we can use css_is_ancestor without taking care of RCU.
2173 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2174 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2176 return autoremove_wake_function(wait, mode, sync, arg);
2179 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2181 /* for filtering, pass "memcg" as argument. */
2182 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2185 static void memcg_oom_recover(struct mem_cgroup *memcg)
2187 if (memcg && atomic_read(&memcg->under_oom))
2188 memcg_wakeup_oom(memcg);
2192 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2194 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2197 struct oom_wait_info owait;
2198 bool locked, need_to_kill;
2200 owait.memcg = memcg;
2201 owait.wait.flags = 0;
2202 owait.wait.func = memcg_oom_wake_function;
2203 owait.wait.private = current;
2204 INIT_LIST_HEAD(&owait.wait.task_list);
2205 need_to_kill = true;
2206 mem_cgroup_mark_under_oom(memcg);
2208 /* At first, try to OOM lock hierarchy under memcg.*/
2209 spin_lock(&memcg_oom_lock);
2210 locked = mem_cgroup_oom_lock(memcg);
2212 * Even if signal_pending(), we can't quit charge() loop without
2213 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2214 * under OOM is always welcomed, use TASK_KILLABLE here.
2216 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2217 if (!locked || memcg->oom_kill_disable)
2218 need_to_kill = false;
2220 mem_cgroup_oom_notify(memcg);
2221 spin_unlock(&memcg_oom_lock);
2224 finish_wait(&memcg_oom_waitq, &owait.wait);
2225 mem_cgroup_out_of_memory(memcg, mask, order);
2228 finish_wait(&memcg_oom_waitq, &owait.wait);
2230 spin_lock(&memcg_oom_lock);
2232 mem_cgroup_oom_unlock(memcg);
2233 memcg_wakeup_oom(memcg);
2234 spin_unlock(&memcg_oom_lock);
2236 mem_cgroup_unmark_under_oom(memcg);
2238 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2240 /* Give chance to dying process */
2241 schedule_timeout_uninterruptible(1);
2246 * Currently used to update mapped file statistics, but the routine can be
2247 * generalized to update other statistics as well.
2249 * Notes: Race condition
2251 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2252 * it tends to be costly. But considering some conditions, we doesn't need
2253 * to do so _always_.
2255 * Considering "charge", lock_page_cgroup() is not required because all
2256 * file-stat operations happen after a page is attached to radix-tree. There
2257 * are no race with "charge".
2259 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2260 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2261 * if there are race with "uncharge". Statistics itself is properly handled
2264 * Considering "move", this is an only case we see a race. To make the race
2265 * small, we check mm->moving_account and detect there are possibility of race
2266 * If there is, we take a lock.
2269 void __mem_cgroup_begin_update_page_stat(struct page *page,
2270 bool *locked, unsigned long *flags)
2272 struct mem_cgroup *memcg;
2273 struct page_cgroup *pc;
2275 pc = lookup_page_cgroup(page);
2277 memcg = pc->mem_cgroup;
2278 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2281 * If this memory cgroup is not under account moving, we don't
2282 * need to take move_lock_mem_cgroup(). Because we already hold
2283 * rcu_read_lock(), any calls to move_account will be delayed until
2284 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2286 if (!mem_cgroup_stolen(memcg))
2289 move_lock_mem_cgroup(memcg, flags);
2290 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2291 move_unlock_mem_cgroup(memcg, flags);
2297 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2299 struct page_cgroup *pc = lookup_page_cgroup(page);
2302 * It's guaranteed that pc->mem_cgroup never changes while
2303 * lock is held because a routine modifies pc->mem_cgroup
2304 * should take move_lock_mem_cgroup().
2306 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2309 void mem_cgroup_update_page_stat(struct page *page,
2310 enum mem_cgroup_page_stat_item idx, int val)
2312 struct mem_cgroup *memcg;
2313 struct page_cgroup *pc = lookup_page_cgroup(page);
2314 unsigned long uninitialized_var(flags);
2316 if (mem_cgroup_disabled())
2319 memcg = pc->mem_cgroup;
2320 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2324 case MEMCG_NR_FILE_MAPPED:
2325 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2331 this_cpu_add(memcg->stat->count[idx], val);
2335 * size of first charge trial. "32" comes from vmscan.c's magic value.
2336 * TODO: maybe necessary to use big numbers in big irons.
2338 #define CHARGE_BATCH 32U
2339 struct memcg_stock_pcp {
2340 struct mem_cgroup *cached; /* this never be root cgroup */
2341 unsigned int nr_pages;
2342 struct work_struct work;
2343 unsigned long flags;
2344 #define FLUSHING_CACHED_CHARGE 0
2346 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2347 static DEFINE_MUTEX(percpu_charge_mutex);
2350 * consume_stock: Try to consume stocked charge on this cpu.
2351 * @memcg: memcg to consume from.
2352 * @nr_pages: how many pages to charge.
2354 * The charges will only happen if @memcg matches the current cpu's memcg
2355 * stock, and at least @nr_pages are available in that stock. Failure to
2356 * service an allocation will refill the stock.
2358 * returns true if successful, false otherwise.
2360 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2362 struct memcg_stock_pcp *stock;
2365 if (nr_pages > CHARGE_BATCH)
2368 stock = &get_cpu_var(memcg_stock);
2369 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2370 stock->nr_pages -= nr_pages;
2371 else /* need to call res_counter_charge */
2373 put_cpu_var(memcg_stock);
2378 * Returns stocks cached in percpu to res_counter and reset cached information.
2380 static void drain_stock(struct memcg_stock_pcp *stock)
2382 struct mem_cgroup *old = stock->cached;
2384 if (stock->nr_pages) {
2385 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2387 res_counter_uncharge(&old->res, bytes);
2388 if (do_swap_account)
2389 res_counter_uncharge(&old->memsw, bytes);
2390 stock->nr_pages = 0;
2392 stock->cached = NULL;
2396 * This must be called under preempt disabled or must be called by
2397 * a thread which is pinned to local cpu.
2399 static void drain_local_stock(struct work_struct *dummy)
2401 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2403 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2406 static void __init memcg_stock_init(void)
2410 for_each_possible_cpu(cpu) {
2411 struct memcg_stock_pcp *stock =
2412 &per_cpu(memcg_stock, cpu);
2413 INIT_WORK(&stock->work, drain_local_stock);
2418 * Cache charges(val) which is from res_counter, to local per_cpu area.
2419 * This will be consumed by consume_stock() function, later.
2421 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2423 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2425 if (stock->cached != memcg) { /* reset if necessary */
2427 stock->cached = memcg;
2429 stock->nr_pages += nr_pages;
2430 put_cpu_var(memcg_stock);
2434 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2435 * of the hierarchy under it. sync flag says whether we should block
2436 * until the work is done.
2438 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2442 /* Notify other cpus that system-wide "drain" is running */
2445 for_each_online_cpu(cpu) {
2446 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2447 struct mem_cgroup *memcg;
2449 memcg = stock->cached;
2450 if (!memcg || !stock->nr_pages)
2452 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2454 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2456 drain_local_stock(&stock->work);
2458 schedule_work_on(cpu, &stock->work);
2466 for_each_online_cpu(cpu) {
2467 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2468 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2469 flush_work(&stock->work);
2476 * Tries to drain stocked charges in other cpus. This function is asynchronous
2477 * and just put a work per cpu for draining localy on each cpu. Caller can
2478 * expects some charges will be back to res_counter later but cannot wait for
2481 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2484 * If someone calls draining, avoid adding more kworker runs.
2486 if (!mutex_trylock(&percpu_charge_mutex))
2488 drain_all_stock(root_memcg, false);
2489 mutex_unlock(&percpu_charge_mutex);
2492 /* This is a synchronous drain interface. */
2493 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2495 /* called when force_empty is called */
2496 mutex_lock(&percpu_charge_mutex);
2497 drain_all_stock(root_memcg, true);
2498 mutex_unlock(&percpu_charge_mutex);
2502 * This function drains percpu counter value from DEAD cpu and
2503 * move it to local cpu. Note that this function can be preempted.
2505 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2509 spin_lock(&memcg->pcp_counter_lock);
2510 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2511 long x = per_cpu(memcg->stat->count[i], cpu);
2513 per_cpu(memcg->stat->count[i], cpu) = 0;
2514 memcg->nocpu_base.count[i] += x;
2516 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2517 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2519 per_cpu(memcg->stat->events[i], cpu) = 0;
2520 memcg->nocpu_base.events[i] += x;
2522 spin_unlock(&memcg->pcp_counter_lock);
2525 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2526 unsigned long action,
2529 int cpu = (unsigned long)hcpu;
2530 struct memcg_stock_pcp *stock;
2531 struct mem_cgroup *iter;
2533 if (action == CPU_ONLINE)
2536 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2539 for_each_mem_cgroup(iter)
2540 mem_cgroup_drain_pcp_counter(iter, cpu);
2542 stock = &per_cpu(memcg_stock, cpu);
2548 /* See __mem_cgroup_try_charge() for details */
2550 CHARGE_OK, /* success */
2551 CHARGE_RETRY, /* need to retry but retry is not bad */
2552 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2553 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2554 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2557 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2558 unsigned int nr_pages, unsigned int min_pages,
2561 unsigned long csize = nr_pages * PAGE_SIZE;
2562 struct mem_cgroup *mem_over_limit;
2563 struct res_counter *fail_res;
2564 unsigned long flags = 0;
2567 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2570 if (!do_swap_account)
2572 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2576 res_counter_uncharge(&memcg->res, csize);
2577 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2578 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2580 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2582 * Never reclaim on behalf of optional batching, retry with a
2583 * single page instead.
2585 if (nr_pages > min_pages)
2586 return CHARGE_RETRY;
2588 if (!(gfp_mask & __GFP_WAIT))
2589 return CHARGE_WOULDBLOCK;
2591 if (gfp_mask & __GFP_NORETRY)
2592 return CHARGE_NOMEM;
2594 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2595 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2596 return CHARGE_RETRY;
2598 * Even though the limit is exceeded at this point, reclaim
2599 * may have been able to free some pages. Retry the charge
2600 * before killing the task.
2602 * Only for regular pages, though: huge pages are rather
2603 * unlikely to succeed so close to the limit, and we fall back
2604 * to regular pages anyway in case of failure.
2606 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2607 return CHARGE_RETRY;
2610 * At task move, charge accounts can be doubly counted. So, it's
2611 * better to wait until the end of task_move if something is going on.
2613 if (mem_cgroup_wait_acct_move(mem_over_limit))
2614 return CHARGE_RETRY;
2616 /* If we don't need to call oom-killer at el, return immediately */
2618 return CHARGE_NOMEM;
2620 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2621 return CHARGE_OOM_DIE;
2623 return CHARGE_RETRY;
2627 * __mem_cgroup_try_charge() does
2628 * 1. detect memcg to be charged against from passed *mm and *ptr,
2629 * 2. update res_counter
2630 * 3. call memory reclaim if necessary.
2632 * In some special case, if the task is fatal, fatal_signal_pending() or
2633 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2634 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2635 * as possible without any hazards. 2: all pages should have a valid
2636 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2637 * pointer, that is treated as a charge to root_mem_cgroup.
2639 * So __mem_cgroup_try_charge() will return
2640 * 0 ... on success, filling *ptr with a valid memcg pointer.
2641 * -ENOMEM ... charge failure because of resource limits.
2642 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2644 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2645 * the oom-killer can be invoked.
2647 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2649 unsigned int nr_pages,
2650 struct mem_cgroup **ptr,
2653 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2654 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2655 struct mem_cgroup *memcg = NULL;
2659 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2660 * in system level. So, allow to go ahead dying process in addition to
2663 if (unlikely(test_thread_flag(TIF_MEMDIE)
2664 || fatal_signal_pending(current)))
2668 * We always charge the cgroup the mm_struct belongs to.
2669 * The mm_struct's mem_cgroup changes on task migration if the
2670 * thread group leader migrates. It's possible that mm is not
2671 * set, if so charge the root memcg (happens for pagecache usage).
2674 *ptr = root_mem_cgroup;
2676 if (*ptr) { /* css should be a valid one */
2678 if (mem_cgroup_is_root(memcg))
2680 if (consume_stock(memcg, nr_pages))
2682 css_get(&memcg->css);
2684 struct task_struct *p;
2687 p = rcu_dereference(mm->owner);
2689 * Because we don't have task_lock(), "p" can exit.
2690 * In that case, "memcg" can point to root or p can be NULL with
2691 * race with swapoff. Then, we have small risk of mis-accouning.
2692 * But such kind of mis-account by race always happens because
2693 * we don't have cgroup_mutex(). It's overkill and we allo that
2695 * (*) swapoff at el will charge against mm-struct not against
2696 * task-struct. So, mm->owner can be NULL.
2698 memcg = mem_cgroup_from_task(p);
2700 memcg = root_mem_cgroup;
2701 if (mem_cgroup_is_root(memcg)) {
2705 if (consume_stock(memcg, nr_pages)) {
2707 * It seems dagerous to access memcg without css_get().
2708 * But considering how consume_stok works, it's not
2709 * necessary. If consume_stock success, some charges
2710 * from this memcg are cached on this cpu. So, we
2711 * don't need to call css_get()/css_tryget() before
2712 * calling consume_stock().
2717 /* after here, we may be blocked. we need to get refcnt */
2718 if (!css_tryget(&memcg->css)) {
2728 /* If killed, bypass charge */
2729 if (fatal_signal_pending(current)) {
2730 css_put(&memcg->css);
2735 if (oom && !nr_oom_retries) {
2737 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2740 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2745 case CHARGE_RETRY: /* not in OOM situation but retry */
2747 css_put(&memcg->css);
2750 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2751 css_put(&memcg->css);
2753 case CHARGE_NOMEM: /* OOM routine works */
2755 css_put(&memcg->css);
2758 /* If oom, we never return -ENOMEM */
2761 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2762 css_put(&memcg->css);
2765 } while (ret != CHARGE_OK);
2767 if (batch > nr_pages)
2768 refill_stock(memcg, batch - nr_pages);
2769 css_put(&memcg->css);
2777 *ptr = root_mem_cgroup;
2782 * Somemtimes we have to undo a charge we got by try_charge().
2783 * This function is for that and do uncharge, put css's refcnt.
2784 * gotten by try_charge().
2786 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2787 unsigned int nr_pages)
2789 if (!mem_cgroup_is_root(memcg)) {
2790 unsigned long bytes = nr_pages * PAGE_SIZE;
2792 res_counter_uncharge(&memcg->res, bytes);
2793 if (do_swap_account)
2794 res_counter_uncharge(&memcg->memsw, bytes);
2799 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2800 * This is useful when moving usage to parent cgroup.
2802 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2803 unsigned int nr_pages)
2805 unsigned long bytes = nr_pages * PAGE_SIZE;
2807 if (mem_cgroup_is_root(memcg))
2810 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2811 if (do_swap_account)
2812 res_counter_uncharge_until(&memcg->memsw,
2813 memcg->memsw.parent, bytes);
2817 * A helper function to get mem_cgroup from ID. must be called under
2818 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2819 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2820 * called against removed memcg.)
2822 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2824 struct cgroup_subsys_state *css;
2826 /* ID 0 is unused ID */
2829 css = css_lookup(&mem_cgroup_subsys, id);
2832 return mem_cgroup_from_css(css);
2835 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2837 struct mem_cgroup *memcg = NULL;
2838 struct page_cgroup *pc;
2842 VM_BUG_ON(!PageLocked(page));
2844 pc = lookup_page_cgroup(page);
2845 lock_page_cgroup(pc);
2846 if (PageCgroupUsed(pc)) {
2847 memcg = pc->mem_cgroup;
2848 if (memcg && !css_tryget(&memcg->css))
2850 } else if (PageSwapCache(page)) {
2851 ent.val = page_private(page);
2852 id = lookup_swap_cgroup_id(ent);
2854 memcg = mem_cgroup_lookup(id);
2855 if (memcg && !css_tryget(&memcg->css))
2859 unlock_page_cgroup(pc);
2863 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2865 unsigned int nr_pages,
2866 enum charge_type ctype,
2869 struct page_cgroup *pc = lookup_page_cgroup(page);
2870 struct zone *uninitialized_var(zone);
2871 struct lruvec *lruvec;
2872 bool was_on_lru = false;
2875 lock_page_cgroup(pc);
2876 VM_BUG_ON(PageCgroupUsed(pc));
2878 * we don't need page_cgroup_lock about tail pages, becase they are not
2879 * accessed by any other context at this point.
2883 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2884 * may already be on some other mem_cgroup's LRU. Take care of it.
2887 zone = page_zone(page);
2888 spin_lock_irq(&zone->lru_lock);
2889 if (PageLRU(page)) {
2890 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2892 del_page_from_lru_list(page, lruvec, page_lru(page));
2897 pc->mem_cgroup = memcg;
2899 * We access a page_cgroup asynchronously without lock_page_cgroup().
2900 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2901 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2902 * before USED bit, we need memory barrier here.
2903 * See mem_cgroup_add_lru_list(), etc.
2906 SetPageCgroupUsed(pc);
2910 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2911 VM_BUG_ON(PageLRU(page));
2913 add_page_to_lru_list(page, lruvec, page_lru(page));
2915 spin_unlock_irq(&zone->lru_lock);
2918 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2923 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2924 unlock_page_cgroup(pc);
2927 * "charge_statistics" updated event counter. Then, check it.
2928 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2929 * if they exceeds softlimit.
2931 memcg_check_events(memcg, page);
2934 static DEFINE_MUTEX(set_limit_mutex);
2936 #ifdef CONFIG_MEMCG_KMEM
2937 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2939 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2940 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2944 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2945 * in the memcg_cache_params struct.
2947 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2949 struct kmem_cache *cachep;
2951 VM_BUG_ON(p->is_root_cache);
2952 cachep = p->root_cache;
2953 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2956 #ifdef CONFIG_SLABINFO
2957 static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2960 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2961 struct memcg_cache_params *params;
2963 if (!memcg_can_account_kmem(memcg))
2966 print_slabinfo_header(m);
2968 mutex_lock(&memcg->slab_caches_mutex);
2969 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2970 cache_show(memcg_params_to_cache(params), m);
2971 mutex_unlock(&memcg->slab_caches_mutex);
2977 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2979 struct res_counter *fail_res;
2980 struct mem_cgroup *_memcg;
2984 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2989 * Conditions under which we can wait for the oom_killer. Those are
2990 * the same conditions tested by the core page allocator
2992 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2995 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2998 if (ret == -EINTR) {
3000 * __mem_cgroup_try_charge() chosed to bypass to root due to
3001 * OOM kill or fatal signal. Since our only options are to
3002 * either fail the allocation or charge it to this cgroup, do
3003 * it as a temporary condition. But we can't fail. From a
3004 * kmem/slab perspective, the cache has already been selected,
3005 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3008 * This condition will only trigger if the task entered
3009 * memcg_charge_kmem in a sane state, but was OOM-killed during
3010 * __mem_cgroup_try_charge() above. Tasks that were already
3011 * dying when the allocation triggers should have been already
3012 * directed to the root cgroup in memcontrol.h
3014 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3015 if (do_swap_account)
3016 res_counter_charge_nofail(&memcg->memsw, size,
3020 res_counter_uncharge(&memcg->kmem, size);
3025 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3027 res_counter_uncharge(&memcg->res, size);
3028 if (do_swap_account)
3029 res_counter_uncharge(&memcg->memsw, size);
3032 if (res_counter_uncharge(&memcg->kmem, size))
3035 if (memcg_kmem_test_and_clear_dead(memcg))
3036 mem_cgroup_put(memcg);
3039 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3044 mutex_lock(&memcg->slab_caches_mutex);
3045 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3046 mutex_unlock(&memcg->slab_caches_mutex);
3050 * helper for acessing a memcg's index. It will be used as an index in the
3051 * child cache array in kmem_cache, and also to derive its name. This function
3052 * will return -1 when this is not a kmem-limited memcg.
3054 int memcg_cache_id(struct mem_cgroup *memcg)
3056 return memcg ? memcg->kmemcg_id : -1;
3060 * This ends up being protected by the set_limit mutex, during normal
3061 * operation, because that is its main call site.
3063 * But when we create a new cache, we can call this as well if its parent
3064 * is kmem-limited. That will have to hold set_limit_mutex as well.
3066 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3070 num = ida_simple_get(&kmem_limited_groups,
3071 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3075 * After this point, kmem_accounted (that we test atomically in
3076 * the beginning of this conditional), is no longer 0. This
3077 * guarantees only one process will set the following boolean
3078 * to true. We don't need test_and_set because we're protected
3079 * by the set_limit_mutex anyway.
3081 memcg_kmem_set_activated(memcg);
3083 ret = memcg_update_all_caches(num+1);
3085 ida_simple_remove(&kmem_limited_groups, num);
3086 memcg_kmem_clear_activated(memcg);
3090 memcg->kmemcg_id = num;
3091 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3092 mutex_init(&memcg->slab_caches_mutex);
3096 static size_t memcg_caches_array_size(int num_groups)
3099 if (num_groups <= 0)
3102 size = 2 * num_groups;
3103 if (size < MEMCG_CACHES_MIN_SIZE)
3104 size = MEMCG_CACHES_MIN_SIZE;
3105 else if (size > MEMCG_CACHES_MAX_SIZE)
3106 size = MEMCG_CACHES_MAX_SIZE;
3112 * We should update the current array size iff all caches updates succeed. This
3113 * can only be done from the slab side. The slab mutex needs to be held when
3116 void memcg_update_array_size(int num)
3118 if (num > memcg_limited_groups_array_size)
3119 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3122 static void kmem_cache_destroy_work_func(struct work_struct *w);
3124 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3126 struct memcg_cache_params *cur_params = s->memcg_params;
3128 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3130 if (num_groups > memcg_limited_groups_array_size) {
3132 ssize_t size = memcg_caches_array_size(num_groups);
3134 size *= sizeof(void *);
3135 size += sizeof(struct memcg_cache_params);
3137 s->memcg_params = kzalloc(size, GFP_KERNEL);
3138 if (!s->memcg_params) {
3139 s->memcg_params = cur_params;
3143 s->memcg_params->is_root_cache = true;
3146 * There is the chance it will be bigger than
3147 * memcg_limited_groups_array_size, if we failed an allocation
3148 * in a cache, in which case all caches updated before it, will
3149 * have a bigger array.
3151 * But if that is the case, the data after
3152 * memcg_limited_groups_array_size is certainly unused
3154 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3155 if (!cur_params->memcg_caches[i])
3157 s->memcg_params->memcg_caches[i] =
3158 cur_params->memcg_caches[i];
3162 * Ideally, we would wait until all caches succeed, and only
3163 * then free the old one. But this is not worth the extra
3164 * pointer per-cache we'd have to have for this.
3166 * It is not a big deal if some caches are left with a size
3167 * bigger than the others. And all updates will reset this
3175 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3176 struct kmem_cache *root_cache)
3178 size_t size = sizeof(struct memcg_cache_params);
3180 if (!memcg_kmem_enabled())
3184 size += memcg_limited_groups_array_size * sizeof(void *);
3186 s->memcg_params = kzalloc(size, GFP_KERNEL);
3187 if (!s->memcg_params)
3190 INIT_WORK(&s->memcg_params->destroy,
3191 kmem_cache_destroy_work_func);
3193 s->memcg_params->memcg = memcg;
3194 s->memcg_params->root_cache = root_cache;
3196 s->memcg_params->is_root_cache = true;
3201 void memcg_release_cache(struct kmem_cache *s)
3203 struct kmem_cache *root;
3204 struct mem_cgroup *memcg;
3208 * This happens, for instance, when a root cache goes away before we
3211 if (!s->memcg_params)
3214 if (s->memcg_params->is_root_cache)
3217 memcg = s->memcg_params->memcg;
3218 id = memcg_cache_id(memcg);
3220 root = s->memcg_params->root_cache;
3221 root->memcg_params->memcg_caches[id] = NULL;
3223 mutex_lock(&memcg->slab_caches_mutex);
3224 list_del(&s->memcg_params->list);
3225 mutex_unlock(&memcg->slab_caches_mutex);
3227 mem_cgroup_put(memcg);
3229 kfree(s->memcg_params);
3233 * During the creation a new cache, we need to disable our accounting mechanism
3234 * altogether. This is true even if we are not creating, but rather just
3235 * enqueing new caches to be created.
3237 * This is because that process will trigger allocations; some visible, like
3238 * explicit kmallocs to auxiliary data structures, name strings and internal
3239 * cache structures; some well concealed, like INIT_WORK() that can allocate
3240 * objects during debug.
3242 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3243 * to it. This may not be a bounded recursion: since the first cache creation
3244 * failed to complete (waiting on the allocation), we'll just try to create the
3245 * cache again, failing at the same point.
3247 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3248 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3249 * inside the following two functions.
3251 static inline void memcg_stop_kmem_account(void)
3253 VM_BUG_ON(!current->mm);
3254 current->memcg_kmem_skip_account++;
3257 static inline void memcg_resume_kmem_account(void)
3259 VM_BUG_ON(!current->mm);
3260 current->memcg_kmem_skip_account--;
3263 static void kmem_cache_destroy_work_func(struct work_struct *w)
3265 struct kmem_cache *cachep;
3266 struct memcg_cache_params *p;
3268 p = container_of(w, struct memcg_cache_params, destroy);
3270 cachep = memcg_params_to_cache(p);
3273 * If we get down to 0 after shrink, we could delete right away.
3274 * However, memcg_release_pages() already puts us back in the workqueue
3275 * in that case. If we proceed deleting, we'll get a dangling
3276 * reference, and removing the object from the workqueue in that case
3277 * is unnecessary complication. We are not a fast path.
3279 * Note that this case is fundamentally different from racing with
3280 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3281 * kmem_cache_shrink, not only we would be reinserting a dead cache
3282 * into the queue, but doing so from inside the worker racing to
3285 * So if we aren't down to zero, we'll just schedule a worker and try
3288 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3289 kmem_cache_shrink(cachep);
3290 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3293 kmem_cache_destroy(cachep);
3296 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3298 if (!cachep->memcg_params->dead)
3302 * There are many ways in which we can get here.
3304 * We can get to a memory-pressure situation while the delayed work is
3305 * still pending to run. The vmscan shrinkers can then release all
3306 * cache memory and get us to destruction. If this is the case, we'll
3307 * be executed twice, which is a bug (the second time will execute over
3308 * bogus data). In this case, cancelling the work should be fine.
3310 * But we can also get here from the worker itself, if
3311 * kmem_cache_shrink is enough to shake all the remaining objects and
3312 * get the page count to 0. In this case, we'll deadlock if we try to
3313 * cancel the work (the worker runs with an internal lock held, which
3314 * is the same lock we would hold for cancel_work_sync().)
3316 * Since we can't possibly know who got us here, just refrain from
3317 * running if there is already work pending
3319 if (work_pending(&cachep->memcg_params->destroy))
3322 * We have to defer the actual destroying to a workqueue, because
3323 * we might currently be in a context that cannot sleep.
3325 schedule_work(&cachep->memcg_params->destroy);
3329 * This lock protects updaters, not readers. We want readers to be as fast as
3330 * they can, and they will either see NULL or a valid cache value. Our model
3331 * allow them to see NULL, in which case the root memcg will be selected.
3333 * We need this lock because multiple allocations to the same cache from a non
3334 * will span more than one worker. Only one of them can create the cache.
3336 static DEFINE_MUTEX(memcg_cache_mutex);
3339 * Called with memcg_cache_mutex held
3341 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3342 struct kmem_cache *s)
3344 struct kmem_cache *new;
3345 static char *tmp_name = NULL;
3347 lockdep_assert_held(&memcg_cache_mutex);
3350 * kmem_cache_create_memcg duplicates the given name and
3351 * cgroup_name for this name requires RCU context.
3352 * This static temporary buffer is used to prevent from
3353 * pointless shortliving allocation.
3356 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3362 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3363 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3366 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3367 (s->flags & ~SLAB_PANIC), s->ctor, s);
3370 new->allocflags |= __GFP_KMEMCG;
3375 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3376 struct kmem_cache *cachep)
3378 struct kmem_cache *new_cachep;
3381 BUG_ON(!memcg_can_account_kmem(memcg));
3383 idx = memcg_cache_id(memcg);
3385 mutex_lock(&memcg_cache_mutex);
3386 new_cachep = cachep->memcg_params->memcg_caches[idx];
3390 new_cachep = kmem_cache_dup(memcg, cachep);
3391 if (new_cachep == NULL) {
3392 new_cachep = cachep;
3396 mem_cgroup_get(memcg);
3397 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3399 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3401 * the readers won't lock, make sure everybody sees the updated value,
3402 * so they won't put stuff in the queue again for no reason
3406 mutex_unlock(&memcg_cache_mutex);
3410 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3412 struct kmem_cache *c;
3415 if (!s->memcg_params)
3417 if (!s->memcg_params->is_root_cache)
3421 * If the cache is being destroyed, we trust that there is no one else
3422 * requesting objects from it. Even if there are, the sanity checks in
3423 * kmem_cache_destroy should caught this ill-case.
3425 * Still, we don't want anyone else freeing memcg_caches under our
3426 * noses, which can happen if a new memcg comes to life. As usual,
3427 * we'll take the set_limit_mutex to protect ourselves against this.
3429 mutex_lock(&set_limit_mutex);
3430 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3431 c = s->memcg_params->memcg_caches[i];
3436 * We will now manually delete the caches, so to avoid races
3437 * we need to cancel all pending destruction workers and
3438 * proceed with destruction ourselves.
3440 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3441 * and that could spawn the workers again: it is likely that
3442 * the cache still have active pages until this very moment.
3443 * This would lead us back to mem_cgroup_destroy_cache.
3445 * But that will not execute at all if the "dead" flag is not
3446 * set, so flip it down to guarantee we are in control.
3448 c->memcg_params->dead = false;
3449 cancel_work_sync(&c->memcg_params->destroy);
3450 kmem_cache_destroy(c);
3452 mutex_unlock(&set_limit_mutex);
3455 struct create_work {
3456 struct mem_cgroup *memcg;
3457 struct kmem_cache *cachep;
3458 struct work_struct work;
3461 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3463 struct kmem_cache *cachep;
3464 struct memcg_cache_params *params;
3466 if (!memcg_kmem_is_active(memcg))
3469 mutex_lock(&memcg->slab_caches_mutex);
3470 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3471 cachep = memcg_params_to_cache(params);
3472 cachep->memcg_params->dead = true;
3473 schedule_work(&cachep->memcg_params->destroy);
3475 mutex_unlock(&memcg->slab_caches_mutex);
3478 static void memcg_create_cache_work_func(struct work_struct *w)
3480 struct create_work *cw;
3482 cw = container_of(w, struct create_work, work);
3483 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3484 /* Drop the reference gotten when we enqueued. */
3485 css_put(&cw->memcg->css);
3490 * Enqueue the creation of a per-memcg kmem_cache.
3492 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3493 struct kmem_cache *cachep)
3495 struct create_work *cw;
3497 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3499 css_put(&memcg->css);
3504 cw->cachep = cachep;
3506 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3507 schedule_work(&cw->work);
3510 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3511 struct kmem_cache *cachep)
3514 * We need to stop accounting when we kmalloc, because if the
3515 * corresponding kmalloc cache is not yet created, the first allocation
3516 * in __memcg_create_cache_enqueue will recurse.
3518 * However, it is better to enclose the whole function. Depending on
3519 * the debugging options enabled, INIT_WORK(), for instance, can
3520 * trigger an allocation. This too, will make us recurse. Because at
3521 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3522 * the safest choice is to do it like this, wrapping the whole function.
3524 memcg_stop_kmem_account();
3525 __memcg_create_cache_enqueue(memcg, cachep);
3526 memcg_resume_kmem_account();
3529 * Return the kmem_cache we're supposed to use for a slab allocation.
3530 * We try to use the current memcg's version of the cache.
3532 * If the cache does not exist yet, if we are the first user of it,
3533 * we either create it immediately, if possible, or create it asynchronously
3535 * In the latter case, we will let the current allocation go through with
3536 * the original cache.
3538 * Can't be called in interrupt context or from kernel threads.
3539 * This function needs to be called with rcu_read_lock() held.
3541 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3544 struct mem_cgroup *memcg;
3547 VM_BUG_ON(!cachep->memcg_params);
3548 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3550 if (!current->mm || current->memcg_kmem_skip_account)
3554 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3556 if (!memcg_can_account_kmem(memcg))
3559 idx = memcg_cache_id(memcg);
3562 * barrier to mare sure we're always seeing the up to date value. The
3563 * code updating memcg_caches will issue a write barrier to match this.
3565 read_barrier_depends();
3566 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3567 cachep = cachep->memcg_params->memcg_caches[idx];
3571 /* The corresponding put will be done in the workqueue. */
3572 if (!css_tryget(&memcg->css))
3577 * If we are in a safe context (can wait, and not in interrupt
3578 * context), we could be be predictable and return right away.
3579 * This would guarantee that the allocation being performed
3580 * already belongs in the new cache.
3582 * However, there are some clashes that can arrive from locking.
3583 * For instance, because we acquire the slab_mutex while doing
3584 * kmem_cache_dup, this means no further allocation could happen
3585 * with the slab_mutex held.
3587 * Also, because cache creation issue get_online_cpus(), this
3588 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3589 * that ends up reversed during cpu hotplug. (cpuset allocates
3590 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3591 * better to defer everything.
3593 memcg_create_cache_enqueue(memcg, cachep);
3599 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3602 * We need to verify if the allocation against current->mm->owner's memcg is
3603 * possible for the given order. But the page is not allocated yet, so we'll
3604 * need a further commit step to do the final arrangements.
3606 * It is possible for the task to switch cgroups in this mean time, so at
3607 * commit time, we can't rely on task conversion any longer. We'll then use
3608 * the handle argument to return to the caller which cgroup we should commit
3609 * against. We could also return the memcg directly and avoid the pointer
3610 * passing, but a boolean return value gives better semantics considering
3611 * the compiled-out case as well.
3613 * Returning true means the allocation is possible.
3616 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3618 struct mem_cgroup *memcg;
3622 memcg = try_get_mem_cgroup_from_mm(current->mm);
3625 * very rare case described in mem_cgroup_from_task. Unfortunately there
3626 * isn't much we can do without complicating this too much, and it would
3627 * be gfp-dependent anyway. Just let it go
3629 if (unlikely(!memcg))
3632 if (!memcg_can_account_kmem(memcg)) {
3633 css_put(&memcg->css);
3637 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3641 css_put(&memcg->css);
3645 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3648 struct page_cgroup *pc;
3650 VM_BUG_ON(mem_cgroup_is_root(memcg));
3652 /* The page allocation failed. Revert */
3654 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3658 pc = lookup_page_cgroup(page);
3659 lock_page_cgroup(pc);
3660 pc->mem_cgroup = memcg;
3661 SetPageCgroupUsed(pc);
3662 unlock_page_cgroup(pc);
3665 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3667 struct mem_cgroup *memcg = NULL;
3668 struct page_cgroup *pc;
3671 pc = lookup_page_cgroup(page);
3673 * Fast unlocked return. Theoretically might have changed, have to
3674 * check again after locking.
3676 if (!PageCgroupUsed(pc))
3679 lock_page_cgroup(pc);
3680 if (PageCgroupUsed(pc)) {
3681 memcg = pc->mem_cgroup;
3682 ClearPageCgroupUsed(pc);
3684 unlock_page_cgroup(pc);
3687 * We trust that only if there is a memcg associated with the page, it
3688 * is a valid allocation
3693 VM_BUG_ON(mem_cgroup_is_root(memcg));
3694 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3697 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3700 #endif /* CONFIG_MEMCG_KMEM */
3702 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3704 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3706 * Because tail pages are not marked as "used", set it. We're under
3707 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3708 * charge/uncharge will be never happen and move_account() is done under
3709 * compound_lock(), so we don't have to take care of races.
3711 void mem_cgroup_split_huge_fixup(struct page *head)
3713 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3714 struct page_cgroup *pc;
3715 struct mem_cgroup *memcg;
3718 if (mem_cgroup_disabled())
3721 memcg = head_pc->mem_cgroup;
3722 for (i = 1; i < HPAGE_PMD_NR; i++) {
3724 pc->mem_cgroup = memcg;
3725 smp_wmb();/* see __commit_charge() */
3726 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3728 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3731 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3734 * mem_cgroup_move_account - move account of the page
3736 * @nr_pages: number of regular pages (>1 for huge pages)
3737 * @pc: page_cgroup of the page.
3738 * @from: mem_cgroup which the page is moved from.
3739 * @to: mem_cgroup which the page is moved to. @from != @to.
3741 * The caller must confirm following.
3742 * - page is not on LRU (isolate_page() is useful.)
3743 * - compound_lock is held when nr_pages > 1
3745 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3748 static int mem_cgroup_move_account(struct page *page,
3749 unsigned int nr_pages,
3750 struct page_cgroup *pc,
3751 struct mem_cgroup *from,
3752 struct mem_cgroup *to)
3754 unsigned long flags;
3756 bool anon = PageAnon(page);
3758 VM_BUG_ON(from == to);
3759 VM_BUG_ON(PageLRU(page));
3761 * The page is isolated from LRU. So, collapse function
3762 * will not handle this page. But page splitting can happen.
3763 * Do this check under compound_page_lock(). The caller should
3767 if (nr_pages > 1 && !PageTransHuge(page))
3770 lock_page_cgroup(pc);
3773 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3776 move_lock_mem_cgroup(from, &flags);
3778 if (!anon && page_mapped(page)) {
3779 /* Update mapped_file data for mem_cgroup */
3781 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3782 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3785 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3787 /* caller should have done css_get */
3788 pc->mem_cgroup = to;
3789 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3790 move_unlock_mem_cgroup(from, &flags);
3793 unlock_page_cgroup(pc);
3797 memcg_check_events(to, page);
3798 memcg_check_events(from, page);
3804 * mem_cgroup_move_parent - moves page to the parent group
3805 * @page: the page to move
3806 * @pc: page_cgroup of the page
3807 * @child: page's cgroup
3809 * move charges to its parent or the root cgroup if the group has no
3810 * parent (aka use_hierarchy==0).
3811 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3812 * mem_cgroup_move_account fails) the failure is always temporary and
3813 * it signals a race with a page removal/uncharge or migration. In the
3814 * first case the page is on the way out and it will vanish from the LRU
3815 * on the next attempt and the call should be retried later.
3816 * Isolation from the LRU fails only if page has been isolated from
3817 * the LRU since we looked at it and that usually means either global
3818 * reclaim or migration going on. The page will either get back to the
3820 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3821 * (!PageCgroupUsed) or moved to a different group. The page will
3822 * disappear in the next attempt.
3824 static int mem_cgroup_move_parent(struct page *page,
3825 struct page_cgroup *pc,
3826 struct mem_cgroup *child)
3828 struct mem_cgroup *parent;
3829 unsigned int nr_pages;
3830 unsigned long uninitialized_var(flags);
3833 VM_BUG_ON(mem_cgroup_is_root(child));
3836 if (!get_page_unless_zero(page))
3838 if (isolate_lru_page(page))
3841 nr_pages = hpage_nr_pages(page);
3843 parent = parent_mem_cgroup(child);
3845 * If no parent, move charges to root cgroup.
3848 parent = root_mem_cgroup;
3851 VM_BUG_ON(!PageTransHuge(page));
3852 flags = compound_lock_irqsave(page);
3855 ret = mem_cgroup_move_account(page, nr_pages,
3858 __mem_cgroup_cancel_local_charge(child, nr_pages);
3861 compound_unlock_irqrestore(page, flags);
3862 putback_lru_page(page);
3870 * Charge the memory controller for page usage.
3872 * 0 if the charge was successful
3873 * < 0 if the cgroup is over its limit
3875 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3876 gfp_t gfp_mask, enum charge_type ctype)
3878 struct mem_cgroup *memcg = NULL;
3879 unsigned int nr_pages = 1;
3883 if (PageTransHuge(page)) {
3884 nr_pages <<= compound_order(page);
3885 VM_BUG_ON(!PageTransHuge(page));
3887 * Never OOM-kill a process for a huge page. The
3888 * fault handler will fall back to regular pages.
3893 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3896 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3900 int mem_cgroup_newpage_charge(struct page *page,
3901 struct mm_struct *mm, gfp_t gfp_mask)
3903 if (mem_cgroup_disabled())
3905 VM_BUG_ON(page_mapped(page));
3906 VM_BUG_ON(page->mapping && !PageAnon(page));
3908 return mem_cgroup_charge_common(page, mm, gfp_mask,
3909 MEM_CGROUP_CHARGE_TYPE_ANON);
3913 * While swap-in, try_charge -> commit or cancel, the page is locked.
3914 * And when try_charge() successfully returns, one refcnt to memcg without
3915 * struct page_cgroup is acquired. This refcnt will be consumed by
3916 * "commit()" or removed by "cancel()"
3918 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3921 struct mem_cgroup **memcgp)
3923 struct mem_cgroup *memcg;
3924 struct page_cgroup *pc;
3927 pc = lookup_page_cgroup(page);
3929 * Every swap fault against a single page tries to charge the
3930 * page, bail as early as possible. shmem_unuse() encounters
3931 * already charged pages, too. The USED bit is protected by
3932 * the page lock, which serializes swap cache removal, which
3933 * in turn serializes uncharging.
3935 if (PageCgroupUsed(pc))
3937 if (!do_swap_account)
3939 memcg = try_get_mem_cgroup_from_page(page);
3943 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3944 css_put(&memcg->css);
3949 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3955 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3956 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3959 if (mem_cgroup_disabled())
3962 * A racing thread's fault, or swapoff, may have already
3963 * updated the pte, and even removed page from swap cache: in
3964 * those cases unuse_pte()'s pte_same() test will fail; but
3965 * there's also a KSM case which does need to charge the page.
3967 if (!PageSwapCache(page)) {
3970 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3975 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3978 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3980 if (mem_cgroup_disabled())
3984 __mem_cgroup_cancel_charge(memcg, 1);
3988 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3989 enum charge_type ctype)
3991 if (mem_cgroup_disabled())
3996 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3998 * Now swap is on-memory. This means this page may be
3999 * counted both as mem and swap....double count.
4000 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4001 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4002 * may call delete_from_swap_cache() before reach here.
4004 if (do_swap_account && PageSwapCache(page)) {
4005 swp_entry_t ent = {.val = page_private(page)};
4006 mem_cgroup_uncharge_swap(ent);
4010 void mem_cgroup_commit_charge_swapin(struct page *page,
4011 struct mem_cgroup *memcg)
4013 __mem_cgroup_commit_charge_swapin(page, memcg,
4014 MEM_CGROUP_CHARGE_TYPE_ANON);
4017 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4020 struct mem_cgroup *memcg = NULL;
4021 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4024 if (mem_cgroup_disabled())
4026 if (PageCompound(page))
4029 if (!PageSwapCache(page))
4030 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4031 else { /* page is swapcache/shmem */
4032 ret = __mem_cgroup_try_charge_swapin(mm, page,
4035 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4040 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4041 unsigned int nr_pages,
4042 const enum charge_type ctype)
4044 struct memcg_batch_info *batch = NULL;
4045 bool uncharge_memsw = true;
4047 /* If swapout, usage of swap doesn't decrease */
4048 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4049 uncharge_memsw = false;
4051 batch = ¤t->memcg_batch;
4053 * In usual, we do css_get() when we remember memcg pointer.
4054 * But in this case, we keep res->usage until end of a series of
4055 * uncharges. Then, it's ok to ignore memcg's refcnt.
4058 batch->memcg = memcg;
4060 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4061 * In those cases, all pages freed continuously can be expected to be in
4062 * the same cgroup and we have chance to coalesce uncharges.
4063 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4064 * because we want to do uncharge as soon as possible.
4067 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4068 goto direct_uncharge;
4071 goto direct_uncharge;
4074 * In typical case, batch->memcg == mem. This means we can
4075 * merge a series of uncharges to an uncharge of res_counter.
4076 * If not, we uncharge res_counter ony by one.
4078 if (batch->memcg != memcg)
4079 goto direct_uncharge;
4080 /* remember freed charge and uncharge it later */
4083 batch->memsw_nr_pages++;
4086 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4088 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4089 if (unlikely(batch->memcg != memcg))
4090 memcg_oom_recover(memcg);
4094 * uncharge if !page_mapped(page)
4096 static struct mem_cgroup *
4097 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4100 struct mem_cgroup *memcg = NULL;
4101 unsigned int nr_pages = 1;
4102 struct page_cgroup *pc;
4105 if (mem_cgroup_disabled())
4108 if (PageTransHuge(page)) {
4109 nr_pages <<= compound_order(page);
4110 VM_BUG_ON(!PageTransHuge(page));
4113 * Check if our page_cgroup is valid
4115 pc = lookup_page_cgroup(page);
4116 if (unlikely(!PageCgroupUsed(pc)))
4119 lock_page_cgroup(pc);
4121 memcg = pc->mem_cgroup;
4123 if (!PageCgroupUsed(pc))
4126 anon = PageAnon(page);
4129 case MEM_CGROUP_CHARGE_TYPE_ANON:
4131 * Generally PageAnon tells if it's the anon statistics to be
4132 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4133 * used before page reached the stage of being marked PageAnon.
4137 case MEM_CGROUP_CHARGE_TYPE_DROP:
4138 /* See mem_cgroup_prepare_migration() */
4139 if (page_mapped(page))
4142 * Pages under migration may not be uncharged. But
4143 * end_migration() /must/ be the one uncharging the
4144 * unused post-migration page and so it has to call
4145 * here with the migration bit still set. See the
4146 * res_counter handling below.
4148 if (!end_migration && PageCgroupMigration(pc))
4151 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4152 if (!PageAnon(page)) { /* Shared memory */
4153 if (page->mapping && !page_is_file_cache(page))
4155 } else if (page_mapped(page)) /* Anon */
4162 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4164 ClearPageCgroupUsed(pc);
4166 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4167 * freed from LRU. This is safe because uncharged page is expected not
4168 * to be reused (freed soon). Exception is SwapCache, it's handled by
4169 * special functions.
4172 unlock_page_cgroup(pc);
4174 * even after unlock, we have memcg->res.usage here and this memcg
4175 * will never be freed.
4177 memcg_check_events(memcg, page);
4178 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4179 mem_cgroup_swap_statistics(memcg, true);
4180 mem_cgroup_get(memcg);
4183 * Migration does not charge the res_counter for the
4184 * replacement page, so leave it alone when phasing out the
4185 * page that is unused after the migration.
4187 if (!end_migration && !mem_cgroup_is_root(memcg))
4188 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4193 unlock_page_cgroup(pc);
4197 void mem_cgroup_uncharge_page(struct page *page)
4200 if (page_mapped(page))
4202 VM_BUG_ON(page->mapping && !PageAnon(page));
4204 * If the page is in swap cache, uncharge should be deferred
4205 * to the swap path, which also properly accounts swap usage
4206 * and handles memcg lifetime.
4208 * Note that this check is not stable and reclaim may add the
4209 * page to swap cache at any time after this. However, if the
4210 * page is not in swap cache by the time page->mapcount hits
4211 * 0, there won't be any page table references to the swap
4212 * slot, and reclaim will free it and not actually write the
4215 if (PageSwapCache(page))
4217 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4220 void mem_cgroup_uncharge_cache_page(struct page *page)
4222 VM_BUG_ON(page_mapped(page));
4223 VM_BUG_ON(page->mapping);
4224 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4228 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4229 * In that cases, pages are freed continuously and we can expect pages
4230 * are in the same memcg. All these calls itself limits the number of
4231 * pages freed at once, then uncharge_start/end() is called properly.
4232 * This may be called prural(2) times in a context,
4235 void mem_cgroup_uncharge_start(void)
4237 current->memcg_batch.do_batch++;
4238 /* We can do nest. */
4239 if (current->memcg_batch.do_batch == 1) {
4240 current->memcg_batch.memcg = NULL;
4241 current->memcg_batch.nr_pages = 0;
4242 current->memcg_batch.memsw_nr_pages = 0;
4246 void mem_cgroup_uncharge_end(void)
4248 struct memcg_batch_info *batch = ¤t->memcg_batch;
4250 if (!batch->do_batch)
4254 if (batch->do_batch) /* If stacked, do nothing. */
4260 * This "batch->memcg" is valid without any css_get/put etc...
4261 * bacause we hide charges behind us.
4263 if (batch->nr_pages)
4264 res_counter_uncharge(&batch->memcg->res,
4265 batch->nr_pages * PAGE_SIZE);
4266 if (batch->memsw_nr_pages)
4267 res_counter_uncharge(&batch->memcg->memsw,
4268 batch->memsw_nr_pages * PAGE_SIZE);
4269 memcg_oom_recover(batch->memcg);
4270 /* forget this pointer (for sanity check) */
4271 batch->memcg = NULL;
4276 * called after __delete_from_swap_cache() and drop "page" account.
4277 * memcg information is recorded to swap_cgroup of "ent"
4280 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4282 struct mem_cgroup *memcg;
4283 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4285 if (!swapout) /* this was a swap cache but the swap is unused ! */
4286 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4288 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4291 * record memcg information, if swapout && memcg != NULL,
4292 * mem_cgroup_get() was called in uncharge().
4294 if (do_swap_account && swapout && memcg)
4295 swap_cgroup_record(ent, css_id(&memcg->css));
4299 #ifdef CONFIG_MEMCG_SWAP
4301 * called from swap_entry_free(). remove record in swap_cgroup and
4302 * uncharge "memsw" account.
4304 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4306 struct mem_cgroup *memcg;
4309 if (!do_swap_account)
4312 id = swap_cgroup_record(ent, 0);
4314 memcg = mem_cgroup_lookup(id);
4317 * We uncharge this because swap is freed.
4318 * This memcg can be obsolete one. We avoid calling css_tryget
4320 if (!mem_cgroup_is_root(memcg))
4321 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4322 mem_cgroup_swap_statistics(memcg, false);
4323 mem_cgroup_put(memcg);
4329 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4330 * @entry: swap entry to be moved
4331 * @from: mem_cgroup which the entry is moved from
4332 * @to: mem_cgroup which the entry is moved to
4334 * It succeeds only when the swap_cgroup's record for this entry is the same
4335 * as the mem_cgroup's id of @from.
4337 * Returns 0 on success, -EINVAL on failure.
4339 * The caller must have charged to @to, IOW, called res_counter_charge() about
4340 * both res and memsw, and called css_get().
4342 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4343 struct mem_cgroup *from, struct mem_cgroup *to)
4345 unsigned short old_id, new_id;
4347 old_id = css_id(&from->css);
4348 new_id = css_id(&to->css);
4350 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4351 mem_cgroup_swap_statistics(from, false);
4352 mem_cgroup_swap_statistics(to, true);
4354 * This function is only called from task migration context now.
4355 * It postpones res_counter and refcount handling till the end
4356 * of task migration(mem_cgroup_clear_mc()) for performance
4357 * improvement. But we cannot postpone mem_cgroup_get(to)
4358 * because if the process that has been moved to @to does
4359 * swap-in, the refcount of @to might be decreased to 0.
4367 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4368 struct mem_cgroup *from, struct mem_cgroup *to)
4375 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4378 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4379 struct mem_cgroup **memcgp)
4381 struct mem_cgroup *memcg = NULL;
4382 unsigned int nr_pages = 1;
4383 struct page_cgroup *pc;
4384 enum charge_type ctype;
4388 if (mem_cgroup_disabled())
4391 if (PageTransHuge(page))
4392 nr_pages <<= compound_order(page);
4394 pc = lookup_page_cgroup(page);
4395 lock_page_cgroup(pc);
4396 if (PageCgroupUsed(pc)) {
4397 memcg = pc->mem_cgroup;
4398 css_get(&memcg->css);
4400 * At migrating an anonymous page, its mapcount goes down
4401 * to 0 and uncharge() will be called. But, even if it's fully
4402 * unmapped, migration may fail and this page has to be
4403 * charged again. We set MIGRATION flag here and delay uncharge
4404 * until end_migration() is called
4406 * Corner Case Thinking
4408 * When the old page was mapped as Anon and it's unmap-and-freed
4409 * while migration was ongoing.
4410 * If unmap finds the old page, uncharge() of it will be delayed
4411 * until end_migration(). If unmap finds a new page, it's
4412 * uncharged when it make mapcount to be 1->0. If unmap code
4413 * finds swap_migration_entry, the new page will not be mapped
4414 * and end_migration() will find it(mapcount==0).
4417 * When the old page was mapped but migraion fails, the kernel
4418 * remaps it. A charge for it is kept by MIGRATION flag even
4419 * if mapcount goes down to 0. We can do remap successfully
4420 * without charging it again.
4423 * The "old" page is under lock_page() until the end of
4424 * migration, so, the old page itself will not be swapped-out.
4425 * If the new page is swapped out before end_migraton, our
4426 * hook to usual swap-out path will catch the event.
4429 SetPageCgroupMigration(pc);
4431 unlock_page_cgroup(pc);
4433 * If the page is not charged at this point,
4441 * We charge new page before it's used/mapped. So, even if unlock_page()
4442 * is called before end_migration, we can catch all events on this new
4443 * page. In the case new page is migrated but not remapped, new page's
4444 * mapcount will be finally 0 and we call uncharge in end_migration().
4447 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4449 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4451 * The page is committed to the memcg, but it's not actually
4452 * charged to the res_counter since we plan on replacing the
4453 * old one and only one page is going to be left afterwards.
4455 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4458 /* remove redundant charge if migration failed*/
4459 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4460 struct page *oldpage, struct page *newpage, bool migration_ok)
4462 struct page *used, *unused;
4463 struct page_cgroup *pc;
4469 if (!migration_ok) {
4476 anon = PageAnon(used);
4477 __mem_cgroup_uncharge_common(unused,
4478 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4479 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4481 css_put(&memcg->css);
4483 * We disallowed uncharge of pages under migration because mapcount
4484 * of the page goes down to zero, temporarly.
4485 * Clear the flag and check the page should be charged.
4487 pc = lookup_page_cgroup(oldpage);
4488 lock_page_cgroup(pc);
4489 ClearPageCgroupMigration(pc);
4490 unlock_page_cgroup(pc);
4493 * If a page is a file cache, radix-tree replacement is very atomic
4494 * and we can skip this check. When it was an Anon page, its mapcount
4495 * goes down to 0. But because we added MIGRATION flage, it's not
4496 * uncharged yet. There are several case but page->mapcount check
4497 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4498 * check. (see prepare_charge() also)
4501 mem_cgroup_uncharge_page(used);
4505 * At replace page cache, newpage is not under any memcg but it's on
4506 * LRU. So, this function doesn't touch res_counter but handles LRU
4507 * in correct way. Both pages are locked so we cannot race with uncharge.
4509 void mem_cgroup_replace_page_cache(struct page *oldpage,
4510 struct page *newpage)
4512 struct mem_cgroup *memcg = NULL;
4513 struct page_cgroup *pc;
4514 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4516 if (mem_cgroup_disabled())
4519 pc = lookup_page_cgroup(oldpage);
4520 /* fix accounting on old pages */
4521 lock_page_cgroup(pc);
4522 if (PageCgroupUsed(pc)) {
4523 memcg = pc->mem_cgroup;
4524 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4525 ClearPageCgroupUsed(pc);
4527 unlock_page_cgroup(pc);
4530 * When called from shmem_replace_page(), in some cases the
4531 * oldpage has already been charged, and in some cases not.
4536 * Even if newpage->mapping was NULL before starting replacement,
4537 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4538 * LRU while we overwrite pc->mem_cgroup.
4540 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4543 #ifdef CONFIG_DEBUG_VM
4544 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4546 struct page_cgroup *pc;
4548 pc = lookup_page_cgroup(page);
4550 * Can be NULL while feeding pages into the page allocator for
4551 * the first time, i.e. during boot or memory hotplug;
4552 * or when mem_cgroup_disabled().
4554 if (likely(pc) && PageCgroupUsed(pc))
4559 bool mem_cgroup_bad_page_check(struct page *page)
4561 if (mem_cgroup_disabled())
4564 return lookup_page_cgroup_used(page) != NULL;
4567 void mem_cgroup_print_bad_page(struct page *page)
4569 struct page_cgroup *pc;
4571 pc = lookup_page_cgroup_used(page);
4573 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4574 pc, pc->flags, pc->mem_cgroup);
4579 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4580 unsigned long long val)
4583 u64 memswlimit, memlimit;
4585 int children = mem_cgroup_count_children(memcg);
4586 u64 curusage, oldusage;
4590 * For keeping hierarchical_reclaim simple, how long we should retry
4591 * is depends on callers. We set our retry-count to be function
4592 * of # of children which we should visit in this loop.
4594 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4596 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4599 while (retry_count) {
4600 if (signal_pending(current)) {
4605 * Rather than hide all in some function, I do this in
4606 * open coded manner. You see what this really does.
4607 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4609 mutex_lock(&set_limit_mutex);
4610 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4611 if (memswlimit < val) {
4613 mutex_unlock(&set_limit_mutex);
4617 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4621 ret = res_counter_set_limit(&memcg->res, val);
4623 if (memswlimit == val)
4624 memcg->memsw_is_minimum = true;
4626 memcg->memsw_is_minimum = false;
4628 mutex_unlock(&set_limit_mutex);
4633 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4634 MEM_CGROUP_RECLAIM_SHRINK);
4635 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4636 /* Usage is reduced ? */
4637 if (curusage >= oldusage)
4640 oldusage = curusage;
4642 if (!ret && enlarge)
4643 memcg_oom_recover(memcg);
4648 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4649 unsigned long long val)
4652 u64 memlimit, memswlimit, oldusage, curusage;
4653 int children = mem_cgroup_count_children(memcg);
4657 /* see mem_cgroup_resize_res_limit */
4658 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4659 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4660 while (retry_count) {
4661 if (signal_pending(current)) {
4666 * Rather than hide all in some function, I do this in
4667 * open coded manner. You see what this really does.
4668 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4670 mutex_lock(&set_limit_mutex);
4671 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4672 if (memlimit > val) {
4674 mutex_unlock(&set_limit_mutex);
4677 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4678 if (memswlimit < val)
4680 ret = res_counter_set_limit(&memcg->memsw, val);
4682 if (memlimit == val)
4683 memcg->memsw_is_minimum = true;
4685 memcg->memsw_is_minimum = false;
4687 mutex_unlock(&set_limit_mutex);
4692 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4693 MEM_CGROUP_RECLAIM_NOSWAP |
4694 MEM_CGROUP_RECLAIM_SHRINK);
4695 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4696 /* Usage is reduced ? */
4697 if (curusage >= oldusage)
4700 oldusage = curusage;
4702 if (!ret && enlarge)
4703 memcg_oom_recover(memcg);
4707 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4709 unsigned long *total_scanned)
4711 unsigned long nr_reclaimed = 0;
4712 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4713 unsigned long reclaimed;
4715 struct mem_cgroup_tree_per_zone *mctz;
4716 unsigned long long excess;
4717 unsigned long nr_scanned;
4722 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4724 * This loop can run a while, specially if mem_cgroup's continuously
4725 * keep exceeding their soft limit and putting the system under
4732 mz = mem_cgroup_largest_soft_limit_node(mctz);
4737 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4738 gfp_mask, &nr_scanned);
4739 nr_reclaimed += reclaimed;
4740 *total_scanned += nr_scanned;
4741 spin_lock(&mctz->lock);
4744 * If we failed to reclaim anything from this memory cgroup
4745 * it is time to move on to the next cgroup
4751 * Loop until we find yet another one.
4753 * By the time we get the soft_limit lock
4754 * again, someone might have aded the
4755 * group back on the RB tree. Iterate to
4756 * make sure we get a different mem.
4757 * mem_cgroup_largest_soft_limit_node returns
4758 * NULL if no other cgroup is present on
4762 __mem_cgroup_largest_soft_limit_node(mctz);
4764 css_put(&next_mz->memcg->css);
4765 else /* next_mz == NULL or other memcg */
4769 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4770 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4772 * One school of thought says that we should not add
4773 * back the node to the tree if reclaim returns 0.
4774 * But our reclaim could return 0, simply because due
4775 * to priority we are exposing a smaller subset of
4776 * memory to reclaim from. Consider this as a longer
4779 /* If excess == 0, no tree ops */
4780 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4781 spin_unlock(&mctz->lock);
4782 css_put(&mz->memcg->css);
4785 * Could not reclaim anything and there are no more
4786 * mem cgroups to try or we seem to be looping without
4787 * reclaiming anything.
4789 if (!nr_reclaimed &&
4791 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4793 } while (!nr_reclaimed);
4795 css_put(&next_mz->memcg->css);
4796 return nr_reclaimed;
4800 * mem_cgroup_force_empty_list - clears LRU of a group
4801 * @memcg: group to clear
4804 * @lru: lru to to clear
4806 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4807 * reclaim the pages page themselves - pages are moved to the parent (or root)
4810 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4811 int node, int zid, enum lru_list lru)
4813 struct lruvec *lruvec;
4814 unsigned long flags;
4815 struct list_head *list;
4819 zone = &NODE_DATA(node)->node_zones[zid];
4820 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4821 list = &lruvec->lists[lru];
4825 struct page_cgroup *pc;
4828 spin_lock_irqsave(&zone->lru_lock, flags);
4829 if (list_empty(list)) {
4830 spin_unlock_irqrestore(&zone->lru_lock, flags);
4833 page = list_entry(list->prev, struct page, lru);
4835 list_move(&page->lru, list);
4837 spin_unlock_irqrestore(&zone->lru_lock, flags);
4840 spin_unlock_irqrestore(&zone->lru_lock, flags);
4842 pc = lookup_page_cgroup(page);
4844 if (mem_cgroup_move_parent(page, pc, memcg)) {
4845 /* found lock contention or "pc" is obsolete. */
4850 } while (!list_empty(list));
4854 * make mem_cgroup's charge to be 0 if there is no task by moving
4855 * all the charges and pages to the parent.
4856 * This enables deleting this mem_cgroup.
4858 * Caller is responsible for holding css reference on the memcg.
4860 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4866 /* This is for making all *used* pages to be on LRU. */
4867 lru_add_drain_all();
4868 drain_all_stock_sync(memcg);
4869 mem_cgroup_start_move(memcg);
4870 for_each_node_state(node, N_MEMORY) {
4871 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4874 mem_cgroup_force_empty_list(memcg,
4879 mem_cgroup_end_move(memcg);
4880 memcg_oom_recover(memcg);
4884 * Kernel memory may not necessarily be trackable to a specific
4885 * process. So they are not migrated, and therefore we can't
4886 * expect their value to drop to 0 here.
4887 * Having res filled up with kmem only is enough.
4889 * This is a safety check because mem_cgroup_force_empty_list
4890 * could have raced with mem_cgroup_replace_page_cache callers
4891 * so the lru seemed empty but the page could have been added
4892 * right after the check. RES_USAGE should be safe as we always
4893 * charge before adding to the LRU.
4895 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4896 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4897 } while (usage > 0);
4901 * This mainly exists for tests during the setting of set of use_hierarchy.
4902 * Since this is the very setting we are changing, the current hierarchy value
4905 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4909 /* bounce at first found */
4910 cgroup_for_each_child(pos, memcg->css.cgroup)
4916 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4917 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4918 * from mem_cgroup_count_children(), in the sense that we don't really care how
4919 * many children we have; we only need to know if we have any. It also counts
4920 * any memcg without hierarchy as infertile.
4922 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4924 return memcg->use_hierarchy && __memcg_has_children(memcg);
4928 * Reclaims as many pages from the given memcg as possible and moves
4929 * the rest to the parent.
4931 * Caller is responsible for holding css reference for memcg.
4933 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4935 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4936 struct cgroup *cgrp = memcg->css.cgroup;
4938 /* returns EBUSY if there is a task or if we come here twice. */
4939 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4942 /* we call try-to-free pages for make this cgroup empty */
4943 lru_add_drain_all();
4944 /* try to free all pages in this cgroup */
4945 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4948 if (signal_pending(current))
4951 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4955 /* maybe some writeback is necessary */
4956 congestion_wait(BLK_RW_ASYNC, HZ/10);
4961 mem_cgroup_reparent_charges(memcg);
4966 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4968 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4971 if (mem_cgroup_is_root(memcg))
4973 css_get(&memcg->css);
4974 ret = mem_cgroup_force_empty(memcg);
4975 css_put(&memcg->css);
4981 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4983 return mem_cgroup_from_cont(cont)->use_hierarchy;
4986 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4990 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4991 struct cgroup *parent = cont->parent;
4992 struct mem_cgroup *parent_memcg = NULL;
4995 parent_memcg = mem_cgroup_from_cont(parent);
4997 mutex_lock(&memcg_create_mutex);
4999 if (memcg->use_hierarchy == val)
5003 * If parent's use_hierarchy is set, we can't make any modifications
5004 * in the child subtrees. If it is unset, then the change can
5005 * occur, provided the current cgroup has no children.
5007 * For the root cgroup, parent_mem is NULL, we allow value to be
5008 * set if there are no children.
5010 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5011 (val == 1 || val == 0)) {
5012 if (!__memcg_has_children(memcg))
5013 memcg->use_hierarchy = val;
5020 mutex_unlock(&memcg_create_mutex);
5026 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5027 enum mem_cgroup_stat_index idx)
5029 struct mem_cgroup *iter;
5032 /* Per-cpu values can be negative, use a signed accumulator */
5033 for_each_mem_cgroup_tree(iter, memcg)
5034 val += mem_cgroup_read_stat(iter, idx);
5036 if (val < 0) /* race ? */
5041 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5045 if (!mem_cgroup_is_root(memcg)) {
5047 return res_counter_read_u64(&memcg->res, RES_USAGE);
5049 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5053 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5054 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5056 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5057 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5060 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5062 return val << PAGE_SHIFT;
5065 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
5066 struct file *file, char __user *buf,
5067 size_t nbytes, loff_t *ppos)
5069 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5075 type = MEMFILE_TYPE(cft->private);
5076 name = MEMFILE_ATTR(cft->private);
5080 if (name == RES_USAGE)
5081 val = mem_cgroup_usage(memcg, false);
5083 val = res_counter_read_u64(&memcg->res, name);
5086 if (name == RES_USAGE)
5087 val = mem_cgroup_usage(memcg, true);
5089 val = res_counter_read_u64(&memcg->memsw, name);
5092 val = res_counter_read_u64(&memcg->kmem, name);
5098 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5099 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5102 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
5105 #ifdef CONFIG_MEMCG_KMEM
5106 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5108 * For simplicity, we won't allow this to be disabled. It also can't
5109 * be changed if the cgroup has children already, or if tasks had
5112 * If tasks join before we set the limit, a person looking at
5113 * kmem.usage_in_bytes will have no way to determine when it took
5114 * place, which makes the value quite meaningless.
5116 * After it first became limited, changes in the value of the limit are
5117 * of course permitted.
5119 mutex_lock(&memcg_create_mutex);
5120 mutex_lock(&set_limit_mutex);
5121 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
5122 if (cgroup_task_count(cont) || memcg_has_children(memcg)) {
5126 ret = res_counter_set_limit(&memcg->kmem, val);
5129 ret = memcg_update_cache_sizes(memcg);
5131 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5134 static_key_slow_inc(&memcg_kmem_enabled_key);
5136 * setting the active bit after the inc will guarantee no one
5137 * starts accounting before all call sites are patched
5139 memcg_kmem_set_active(memcg);
5142 * kmem charges can outlive the cgroup. In the case of slab
5143 * pages, for instance, a page contain objects from various
5144 * processes, so it is unfeasible to migrate them away. We
5145 * need to reference count the memcg because of that.
5147 mem_cgroup_get(memcg);
5149 ret = res_counter_set_limit(&memcg->kmem, val);
5151 mutex_unlock(&set_limit_mutex);
5152 mutex_unlock(&memcg_create_mutex);
5157 #ifdef CONFIG_MEMCG_KMEM
5158 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5161 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5165 memcg->kmem_account_flags = parent->kmem_account_flags;
5167 * When that happen, we need to disable the static branch only on those
5168 * memcgs that enabled it. To achieve this, we would be forced to
5169 * complicate the code by keeping track of which memcgs were the ones
5170 * that actually enabled limits, and which ones got it from its
5173 * It is a lot simpler just to do static_key_slow_inc() on every child
5174 * that is accounted.
5176 if (!memcg_kmem_is_active(memcg))
5180 * destroy(), called if we fail, will issue static_key_slow_inc() and
5181 * mem_cgroup_put() if kmem is enabled. We have to either call them
5182 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5183 * this more consistent, since it always leads to the same destroy path
5185 mem_cgroup_get(memcg);
5186 static_key_slow_inc(&memcg_kmem_enabled_key);
5188 mutex_lock(&set_limit_mutex);
5189 ret = memcg_update_cache_sizes(memcg);
5190 mutex_unlock(&set_limit_mutex);
5194 #endif /* CONFIG_MEMCG_KMEM */
5197 * The user of this function is...
5200 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5203 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5206 unsigned long long val;
5209 type = MEMFILE_TYPE(cft->private);
5210 name = MEMFILE_ATTR(cft->private);
5214 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5218 /* This function does all necessary parse...reuse it */
5219 ret = res_counter_memparse_write_strategy(buffer, &val);
5223 ret = mem_cgroup_resize_limit(memcg, val);
5224 else if (type == _MEMSWAP)
5225 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5226 else if (type == _KMEM)
5227 ret = memcg_update_kmem_limit(cont, val);
5231 case RES_SOFT_LIMIT:
5232 ret = res_counter_memparse_write_strategy(buffer, &val);
5236 * For memsw, soft limits are hard to implement in terms
5237 * of semantics, for now, we support soft limits for
5238 * control without swap
5241 ret = res_counter_set_soft_limit(&memcg->res, val);
5246 ret = -EINVAL; /* should be BUG() ? */
5252 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5253 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5255 struct cgroup *cgroup;
5256 unsigned long long min_limit, min_memsw_limit, tmp;
5258 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5259 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5260 cgroup = memcg->css.cgroup;
5261 if (!memcg->use_hierarchy)
5264 while (cgroup->parent) {
5265 cgroup = cgroup->parent;
5266 memcg = mem_cgroup_from_cont(cgroup);
5267 if (!memcg->use_hierarchy)
5269 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5270 min_limit = min(min_limit, tmp);
5271 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5272 min_memsw_limit = min(min_memsw_limit, tmp);
5275 *mem_limit = min_limit;
5276 *memsw_limit = min_memsw_limit;
5279 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5281 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5285 type = MEMFILE_TYPE(event);
5286 name = MEMFILE_ATTR(event);
5291 res_counter_reset_max(&memcg->res);
5292 else if (type == _MEMSWAP)
5293 res_counter_reset_max(&memcg->memsw);
5294 else if (type == _KMEM)
5295 res_counter_reset_max(&memcg->kmem);
5301 res_counter_reset_failcnt(&memcg->res);
5302 else if (type == _MEMSWAP)
5303 res_counter_reset_failcnt(&memcg->memsw);
5304 else if (type == _KMEM)
5305 res_counter_reset_failcnt(&memcg->kmem);
5314 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5317 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5321 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5322 struct cftype *cft, u64 val)
5324 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5326 if (val >= (1 << NR_MOVE_TYPE))
5330 * No kind of locking is needed in here, because ->can_attach() will
5331 * check this value once in the beginning of the process, and then carry
5332 * on with stale data. This means that changes to this value will only
5333 * affect task migrations starting after the change.
5335 memcg->move_charge_at_immigrate = val;
5339 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5340 struct cftype *cft, u64 val)
5347 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5351 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5352 unsigned long node_nr;
5353 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5355 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5356 seq_printf(m, "total=%lu", total_nr);
5357 for_each_node_state(nid, N_MEMORY) {
5358 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5359 seq_printf(m, " N%d=%lu", nid, node_nr);
5363 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5364 seq_printf(m, "file=%lu", file_nr);
5365 for_each_node_state(nid, N_MEMORY) {
5366 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5368 seq_printf(m, " N%d=%lu", nid, node_nr);
5372 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5373 seq_printf(m, "anon=%lu", anon_nr);
5374 for_each_node_state(nid, N_MEMORY) {
5375 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5377 seq_printf(m, " N%d=%lu", nid, node_nr);
5381 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5382 seq_printf(m, "unevictable=%lu", unevictable_nr);
5383 for_each_node_state(nid, N_MEMORY) {
5384 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5385 BIT(LRU_UNEVICTABLE));
5386 seq_printf(m, " N%d=%lu", nid, node_nr);
5391 #endif /* CONFIG_NUMA */
5393 static inline void mem_cgroup_lru_names_not_uptodate(void)
5395 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5398 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5401 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5402 struct mem_cgroup *mi;
5405 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5406 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5408 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5409 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5412 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5413 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5414 mem_cgroup_read_events(memcg, i));
5416 for (i = 0; i < NR_LRU_LISTS; i++)
5417 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5418 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5420 /* Hierarchical information */
5422 unsigned long long limit, memsw_limit;
5423 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5424 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5425 if (do_swap_account)
5426 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5430 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5433 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5435 for_each_mem_cgroup_tree(mi, memcg)
5436 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5437 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5440 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5441 unsigned long long val = 0;
5443 for_each_mem_cgroup_tree(mi, memcg)
5444 val += mem_cgroup_read_events(mi, i);
5445 seq_printf(m, "total_%s %llu\n",
5446 mem_cgroup_events_names[i], val);
5449 for (i = 0; i < NR_LRU_LISTS; i++) {
5450 unsigned long long val = 0;
5452 for_each_mem_cgroup_tree(mi, memcg)
5453 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5454 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5457 #ifdef CONFIG_DEBUG_VM
5460 struct mem_cgroup_per_zone *mz;
5461 struct zone_reclaim_stat *rstat;
5462 unsigned long recent_rotated[2] = {0, 0};
5463 unsigned long recent_scanned[2] = {0, 0};
5465 for_each_online_node(nid)
5466 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5467 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5468 rstat = &mz->lruvec.reclaim_stat;
5470 recent_rotated[0] += rstat->recent_rotated[0];
5471 recent_rotated[1] += rstat->recent_rotated[1];
5472 recent_scanned[0] += rstat->recent_scanned[0];
5473 recent_scanned[1] += rstat->recent_scanned[1];
5475 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5476 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5477 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5478 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5485 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5487 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5489 return mem_cgroup_swappiness(memcg);
5492 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5495 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5496 struct mem_cgroup *parent;
5501 if (cgrp->parent == NULL)
5504 parent = mem_cgroup_from_cont(cgrp->parent);
5506 mutex_lock(&memcg_create_mutex);
5508 /* If under hierarchy, only empty-root can set this value */
5509 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5510 mutex_unlock(&memcg_create_mutex);
5514 memcg->swappiness = val;
5516 mutex_unlock(&memcg_create_mutex);
5521 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5523 struct mem_cgroup_threshold_ary *t;
5529 t = rcu_dereference(memcg->thresholds.primary);
5531 t = rcu_dereference(memcg->memsw_thresholds.primary);
5536 usage = mem_cgroup_usage(memcg, swap);
5539 * current_threshold points to threshold just below or equal to usage.
5540 * If it's not true, a threshold was crossed after last
5541 * call of __mem_cgroup_threshold().
5543 i = t->current_threshold;
5546 * Iterate backward over array of thresholds starting from
5547 * current_threshold and check if a threshold is crossed.
5548 * If none of thresholds below usage is crossed, we read
5549 * only one element of the array here.
5551 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5552 eventfd_signal(t->entries[i].eventfd, 1);
5554 /* i = current_threshold + 1 */
5558 * Iterate forward over array of thresholds starting from
5559 * current_threshold+1 and check if a threshold is crossed.
5560 * If none of thresholds above usage is crossed, we read
5561 * only one element of the array here.
5563 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5564 eventfd_signal(t->entries[i].eventfd, 1);
5566 /* Update current_threshold */
5567 t->current_threshold = i - 1;
5572 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5575 __mem_cgroup_threshold(memcg, false);
5576 if (do_swap_account)
5577 __mem_cgroup_threshold(memcg, true);
5579 memcg = parent_mem_cgroup(memcg);
5583 static int compare_thresholds(const void *a, const void *b)
5585 const struct mem_cgroup_threshold *_a = a;
5586 const struct mem_cgroup_threshold *_b = b;
5588 return _a->threshold - _b->threshold;
5591 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5593 struct mem_cgroup_eventfd_list *ev;
5595 list_for_each_entry(ev, &memcg->oom_notify, list)
5596 eventfd_signal(ev->eventfd, 1);
5600 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5602 struct mem_cgroup *iter;
5604 for_each_mem_cgroup_tree(iter, memcg)
5605 mem_cgroup_oom_notify_cb(iter);
5608 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5609 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5611 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5612 struct mem_cgroup_thresholds *thresholds;
5613 struct mem_cgroup_threshold_ary *new;
5614 enum res_type type = MEMFILE_TYPE(cft->private);
5615 u64 threshold, usage;
5618 ret = res_counter_memparse_write_strategy(args, &threshold);
5622 mutex_lock(&memcg->thresholds_lock);
5625 thresholds = &memcg->thresholds;
5626 else if (type == _MEMSWAP)
5627 thresholds = &memcg->memsw_thresholds;
5631 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5633 /* Check if a threshold crossed before adding a new one */
5634 if (thresholds->primary)
5635 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5637 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5639 /* Allocate memory for new array of thresholds */
5640 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5648 /* Copy thresholds (if any) to new array */
5649 if (thresholds->primary) {
5650 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5651 sizeof(struct mem_cgroup_threshold));
5654 /* Add new threshold */
5655 new->entries[size - 1].eventfd = eventfd;
5656 new->entries[size - 1].threshold = threshold;
5658 /* Sort thresholds. Registering of new threshold isn't time-critical */
5659 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5660 compare_thresholds, NULL);
5662 /* Find current threshold */
5663 new->current_threshold = -1;
5664 for (i = 0; i < size; i++) {
5665 if (new->entries[i].threshold <= usage) {
5667 * new->current_threshold will not be used until
5668 * rcu_assign_pointer(), so it's safe to increment
5671 ++new->current_threshold;
5676 /* Free old spare buffer and save old primary buffer as spare */
5677 kfree(thresholds->spare);
5678 thresholds->spare = thresholds->primary;
5680 rcu_assign_pointer(thresholds->primary, new);
5682 /* To be sure that nobody uses thresholds */
5686 mutex_unlock(&memcg->thresholds_lock);
5691 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5692 struct cftype *cft, struct eventfd_ctx *eventfd)
5694 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5695 struct mem_cgroup_thresholds *thresholds;
5696 struct mem_cgroup_threshold_ary *new;
5697 enum res_type type = MEMFILE_TYPE(cft->private);
5701 mutex_lock(&memcg->thresholds_lock);
5703 thresholds = &memcg->thresholds;
5704 else if (type == _MEMSWAP)
5705 thresholds = &memcg->memsw_thresholds;
5709 if (!thresholds->primary)
5712 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5714 /* Check if a threshold crossed before removing */
5715 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5717 /* Calculate new number of threshold */
5719 for (i = 0; i < thresholds->primary->size; i++) {
5720 if (thresholds->primary->entries[i].eventfd != eventfd)
5724 new = thresholds->spare;
5726 /* Set thresholds array to NULL if we don't have thresholds */
5735 /* Copy thresholds and find current threshold */
5736 new->current_threshold = -1;
5737 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5738 if (thresholds->primary->entries[i].eventfd == eventfd)
5741 new->entries[j] = thresholds->primary->entries[i];
5742 if (new->entries[j].threshold <= usage) {
5744 * new->current_threshold will not be used
5745 * until rcu_assign_pointer(), so it's safe to increment
5748 ++new->current_threshold;
5754 /* Swap primary and spare array */
5755 thresholds->spare = thresholds->primary;
5756 /* If all events are unregistered, free the spare array */
5758 kfree(thresholds->spare);
5759 thresholds->spare = NULL;
5762 rcu_assign_pointer(thresholds->primary, new);
5764 /* To be sure that nobody uses thresholds */
5767 mutex_unlock(&memcg->thresholds_lock);
5770 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5771 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5773 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5774 struct mem_cgroup_eventfd_list *event;
5775 enum res_type type = MEMFILE_TYPE(cft->private);
5777 BUG_ON(type != _OOM_TYPE);
5778 event = kmalloc(sizeof(*event), GFP_KERNEL);
5782 spin_lock(&memcg_oom_lock);
5784 event->eventfd = eventfd;
5785 list_add(&event->list, &memcg->oom_notify);
5787 /* already in OOM ? */
5788 if (atomic_read(&memcg->under_oom))
5789 eventfd_signal(eventfd, 1);
5790 spin_unlock(&memcg_oom_lock);
5795 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5796 struct cftype *cft, struct eventfd_ctx *eventfd)
5798 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5799 struct mem_cgroup_eventfd_list *ev, *tmp;
5800 enum res_type type = MEMFILE_TYPE(cft->private);
5802 BUG_ON(type != _OOM_TYPE);
5804 spin_lock(&memcg_oom_lock);
5806 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5807 if (ev->eventfd == eventfd) {
5808 list_del(&ev->list);
5813 spin_unlock(&memcg_oom_lock);
5816 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5817 struct cftype *cft, struct cgroup_map_cb *cb)
5819 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5821 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5823 if (atomic_read(&memcg->under_oom))
5824 cb->fill(cb, "under_oom", 1);
5826 cb->fill(cb, "under_oom", 0);
5830 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5831 struct cftype *cft, u64 val)
5833 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5834 struct mem_cgroup *parent;
5836 /* cannot set to root cgroup and only 0 and 1 are allowed */
5837 if (!cgrp->parent || !((val == 0) || (val == 1)))
5840 parent = mem_cgroup_from_cont(cgrp->parent);
5842 mutex_lock(&memcg_create_mutex);
5843 /* oom-kill-disable is a flag for subhierarchy. */
5844 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5845 mutex_unlock(&memcg_create_mutex);
5848 memcg->oom_kill_disable = val;
5850 memcg_oom_recover(memcg);
5851 mutex_unlock(&memcg_create_mutex);
5855 #ifdef CONFIG_MEMCG_KMEM
5856 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5860 memcg->kmemcg_id = -1;
5861 ret = memcg_propagate_kmem(memcg);
5865 return mem_cgroup_sockets_init(memcg, ss);
5868 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5870 mem_cgroup_sockets_destroy(memcg);
5872 memcg_kmem_mark_dead(memcg);
5874 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5878 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5879 * path here, being careful not to race with memcg_uncharge_kmem: it is
5880 * possible that the charges went down to 0 between mark_dead and the
5881 * res_counter read, so in that case, we don't need the put
5883 if (memcg_kmem_test_and_clear_dead(memcg))
5884 mem_cgroup_put(memcg);
5887 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5892 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5897 static struct cftype mem_cgroup_files[] = {
5899 .name = "usage_in_bytes",
5900 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5901 .read = mem_cgroup_read,
5902 .register_event = mem_cgroup_usage_register_event,
5903 .unregister_event = mem_cgroup_usage_unregister_event,
5906 .name = "max_usage_in_bytes",
5907 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5908 .trigger = mem_cgroup_reset,
5909 .read = mem_cgroup_read,
5912 .name = "limit_in_bytes",
5913 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5914 .write_string = mem_cgroup_write,
5915 .read = mem_cgroup_read,
5918 .name = "soft_limit_in_bytes",
5919 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5920 .write_string = mem_cgroup_write,
5921 .read = mem_cgroup_read,
5925 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5926 .trigger = mem_cgroup_reset,
5927 .read = mem_cgroup_read,
5931 .read_seq_string = memcg_stat_show,
5934 .name = "force_empty",
5935 .trigger = mem_cgroup_force_empty_write,
5938 .name = "use_hierarchy",
5939 .flags = CFTYPE_INSANE,
5940 .write_u64 = mem_cgroup_hierarchy_write,
5941 .read_u64 = mem_cgroup_hierarchy_read,
5944 .name = "swappiness",
5945 .read_u64 = mem_cgroup_swappiness_read,
5946 .write_u64 = mem_cgroup_swappiness_write,
5949 .name = "move_charge_at_immigrate",
5950 .read_u64 = mem_cgroup_move_charge_read,
5951 .write_u64 = mem_cgroup_move_charge_write,
5954 .name = "oom_control",
5955 .read_map = mem_cgroup_oom_control_read,
5956 .write_u64 = mem_cgroup_oom_control_write,
5957 .register_event = mem_cgroup_oom_register_event,
5958 .unregister_event = mem_cgroup_oom_unregister_event,
5959 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5962 .name = "pressure_level",
5963 .register_event = vmpressure_register_event,
5964 .unregister_event = vmpressure_unregister_event,
5968 .name = "numa_stat",
5969 .read_seq_string = memcg_numa_stat_show,
5972 #ifdef CONFIG_MEMCG_KMEM
5974 .name = "kmem.limit_in_bytes",
5975 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5976 .write_string = mem_cgroup_write,
5977 .read = mem_cgroup_read,
5980 .name = "kmem.usage_in_bytes",
5981 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5982 .read = mem_cgroup_read,
5985 .name = "kmem.failcnt",
5986 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5987 .trigger = mem_cgroup_reset,
5988 .read = mem_cgroup_read,
5991 .name = "kmem.max_usage_in_bytes",
5992 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5993 .trigger = mem_cgroup_reset,
5994 .read = mem_cgroup_read,
5996 #ifdef CONFIG_SLABINFO
5998 .name = "kmem.slabinfo",
5999 .read_seq_string = mem_cgroup_slabinfo_read,
6003 { }, /* terminate */
6006 #ifdef CONFIG_MEMCG_SWAP
6007 static struct cftype memsw_cgroup_files[] = {
6009 .name = "memsw.usage_in_bytes",
6010 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6011 .read = mem_cgroup_read,
6012 .register_event = mem_cgroup_usage_register_event,
6013 .unregister_event = mem_cgroup_usage_unregister_event,
6016 .name = "memsw.max_usage_in_bytes",
6017 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6018 .trigger = mem_cgroup_reset,
6019 .read = mem_cgroup_read,
6022 .name = "memsw.limit_in_bytes",
6023 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6024 .write_string = mem_cgroup_write,
6025 .read = mem_cgroup_read,
6028 .name = "memsw.failcnt",
6029 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6030 .trigger = mem_cgroup_reset,
6031 .read = mem_cgroup_read,
6033 { }, /* terminate */
6036 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6038 struct mem_cgroup_per_node *pn;
6039 struct mem_cgroup_per_zone *mz;
6040 int zone, tmp = node;
6042 * This routine is called against possible nodes.
6043 * But it's BUG to call kmalloc() against offline node.
6045 * TODO: this routine can waste much memory for nodes which will
6046 * never be onlined. It's better to use memory hotplug callback
6049 if (!node_state(node, N_NORMAL_MEMORY))
6051 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6055 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6056 mz = &pn->zoneinfo[zone];
6057 lruvec_init(&mz->lruvec);
6058 mz->usage_in_excess = 0;
6059 mz->on_tree = false;
6062 memcg->info.nodeinfo[node] = pn;
6066 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6068 kfree(memcg->info.nodeinfo[node]);
6071 static struct mem_cgroup *mem_cgroup_alloc(void)
6073 struct mem_cgroup *memcg;
6074 size_t size = memcg_size();
6076 /* Can be very big if nr_node_ids is very big */
6077 if (size < PAGE_SIZE)
6078 memcg = kzalloc(size, GFP_KERNEL);
6080 memcg = vzalloc(size);
6085 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6088 spin_lock_init(&memcg->pcp_counter_lock);
6092 if (size < PAGE_SIZE)
6100 * At destroying mem_cgroup, references from swap_cgroup can remain.
6101 * (scanning all at force_empty is too costly...)
6103 * Instead of clearing all references at force_empty, we remember
6104 * the number of reference from swap_cgroup and free mem_cgroup when
6105 * it goes down to 0.
6107 * Removal of cgroup itself succeeds regardless of refs from swap.
6110 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6113 size_t size = memcg_size();
6115 mem_cgroup_remove_from_trees(memcg);
6116 free_css_id(&mem_cgroup_subsys, &memcg->css);
6119 free_mem_cgroup_per_zone_info(memcg, node);
6121 free_percpu(memcg->stat);
6124 * We need to make sure that (at least for now), the jump label
6125 * destruction code runs outside of the cgroup lock. This is because
6126 * get_online_cpus(), which is called from the static_branch update,
6127 * can't be called inside the cgroup_lock. cpusets are the ones
6128 * enforcing this dependency, so if they ever change, we might as well.
6130 * schedule_work() will guarantee this happens. Be careful if you need
6131 * to move this code around, and make sure it is outside
6134 disarm_static_keys(memcg);
6135 if (size < PAGE_SIZE)
6143 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
6144 * but in process context. The work_freeing structure is overlaid
6145 * on the rcu_freeing structure, which itself is overlaid on memsw.
6147 static void free_work(struct work_struct *work)
6149 struct mem_cgroup *memcg;
6151 memcg = container_of(work, struct mem_cgroup, work_freeing);
6152 __mem_cgroup_free(memcg);
6155 static void free_rcu(struct rcu_head *rcu_head)
6157 struct mem_cgroup *memcg;
6159 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
6160 INIT_WORK(&memcg->work_freeing, free_work);
6161 schedule_work(&memcg->work_freeing);
6164 static void mem_cgroup_get(struct mem_cgroup *memcg)
6166 atomic_inc(&memcg->refcnt);
6169 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
6171 if (atomic_sub_and_test(count, &memcg->refcnt)) {
6172 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
6173 call_rcu(&memcg->rcu_freeing, free_rcu);
6175 mem_cgroup_put(parent);
6179 static void mem_cgroup_put(struct mem_cgroup *memcg)
6181 __mem_cgroup_put(memcg, 1);
6185 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6187 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6189 if (!memcg->res.parent)
6191 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6193 EXPORT_SYMBOL(parent_mem_cgroup);
6195 static void __init mem_cgroup_soft_limit_tree_init(void)
6197 struct mem_cgroup_tree_per_node *rtpn;
6198 struct mem_cgroup_tree_per_zone *rtpz;
6199 int tmp, node, zone;
6201 for_each_node(node) {
6203 if (!node_state(node, N_NORMAL_MEMORY))
6205 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6208 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6210 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6211 rtpz = &rtpn->rb_tree_per_zone[zone];
6212 rtpz->rb_root = RB_ROOT;
6213 spin_lock_init(&rtpz->lock);
6218 static struct cgroup_subsys_state * __ref
6219 mem_cgroup_css_alloc(struct cgroup *cont)
6221 struct mem_cgroup *memcg;
6222 long error = -ENOMEM;
6225 memcg = mem_cgroup_alloc();
6227 return ERR_PTR(error);
6230 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6234 if (cont->parent == NULL) {
6235 root_mem_cgroup = memcg;
6236 res_counter_init(&memcg->res, NULL);
6237 res_counter_init(&memcg->memsw, NULL);
6238 res_counter_init(&memcg->kmem, NULL);
6241 memcg->last_scanned_node = MAX_NUMNODES;
6242 INIT_LIST_HEAD(&memcg->oom_notify);
6243 atomic_set(&memcg->refcnt, 1);
6244 memcg->move_charge_at_immigrate = 0;
6245 mutex_init(&memcg->thresholds_lock);
6246 spin_lock_init(&memcg->move_lock);
6247 vmpressure_init(&memcg->vmpressure);
6252 __mem_cgroup_free(memcg);
6253 return ERR_PTR(error);
6257 mem_cgroup_css_online(struct cgroup *cont)
6259 struct mem_cgroup *memcg, *parent;
6265 mutex_lock(&memcg_create_mutex);
6266 memcg = mem_cgroup_from_cont(cont);
6267 parent = mem_cgroup_from_cont(cont->parent);
6269 memcg->use_hierarchy = parent->use_hierarchy;
6270 memcg->oom_kill_disable = parent->oom_kill_disable;
6271 memcg->swappiness = mem_cgroup_swappiness(parent);
6273 if (parent->use_hierarchy) {
6274 res_counter_init(&memcg->res, &parent->res);
6275 res_counter_init(&memcg->memsw, &parent->memsw);
6276 res_counter_init(&memcg->kmem, &parent->kmem);
6279 * We increment refcnt of the parent to ensure that we can
6280 * safely access it on res_counter_charge/uncharge.
6281 * This refcnt will be decremented when freeing this
6282 * mem_cgroup(see mem_cgroup_put).
6284 mem_cgroup_get(parent);
6286 res_counter_init(&memcg->res, NULL);
6287 res_counter_init(&memcg->memsw, NULL);
6288 res_counter_init(&memcg->kmem, NULL);
6290 * Deeper hierachy with use_hierarchy == false doesn't make
6291 * much sense so let cgroup subsystem know about this
6292 * unfortunate state in our controller.
6294 if (parent != root_mem_cgroup)
6295 mem_cgroup_subsys.broken_hierarchy = true;
6298 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6299 mutex_unlock(&memcg_create_mutex);
6302 * We call put now because our (and parent's) refcnts
6303 * are already in place. mem_cgroup_put() will internally
6304 * call __mem_cgroup_free, so return directly
6306 mem_cgroup_put(memcg);
6307 if (parent->use_hierarchy)
6308 mem_cgroup_put(parent);
6314 * Announce all parents that a group from their hierarchy is gone.
6316 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6318 struct mem_cgroup *parent = memcg;
6320 while ((parent = parent_mem_cgroup(parent)))
6321 atomic_inc(&parent->dead_count);
6324 * if the root memcg is not hierarchical we have to check it
6327 if (!root_mem_cgroup->use_hierarchy)
6328 atomic_inc(&root_mem_cgroup->dead_count);
6331 static void mem_cgroup_css_offline(struct cgroup *cont)
6333 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6335 mem_cgroup_invalidate_reclaim_iterators(memcg);
6336 mem_cgroup_reparent_charges(memcg);
6337 mem_cgroup_destroy_all_caches(memcg);
6340 static void mem_cgroup_css_free(struct cgroup *cont)
6342 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6344 kmem_cgroup_destroy(memcg);
6346 mem_cgroup_put(memcg);
6350 /* Handlers for move charge at task migration. */
6351 #define PRECHARGE_COUNT_AT_ONCE 256
6352 static int mem_cgroup_do_precharge(unsigned long count)
6355 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6356 struct mem_cgroup *memcg = mc.to;
6358 if (mem_cgroup_is_root(memcg)) {
6359 mc.precharge += count;
6360 /* we don't need css_get for root */
6363 /* try to charge at once */
6365 struct res_counter *dummy;
6367 * "memcg" cannot be under rmdir() because we've already checked
6368 * by cgroup_lock_live_cgroup() that it is not removed and we
6369 * are still under the same cgroup_mutex. So we can postpone
6372 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6374 if (do_swap_account && res_counter_charge(&memcg->memsw,
6375 PAGE_SIZE * count, &dummy)) {
6376 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6379 mc.precharge += count;
6383 /* fall back to one by one charge */
6385 if (signal_pending(current)) {
6389 if (!batch_count--) {
6390 batch_count = PRECHARGE_COUNT_AT_ONCE;
6393 ret = __mem_cgroup_try_charge(NULL,
6394 GFP_KERNEL, 1, &memcg, false);
6396 /* mem_cgroup_clear_mc() will do uncharge later */
6404 * get_mctgt_type - get target type of moving charge
6405 * @vma: the vma the pte to be checked belongs
6406 * @addr: the address corresponding to the pte to be checked
6407 * @ptent: the pte to be checked
6408 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6411 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6412 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6413 * move charge. if @target is not NULL, the page is stored in target->page
6414 * with extra refcnt got(Callers should handle it).
6415 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6416 * target for charge migration. if @target is not NULL, the entry is stored
6419 * Called with pte lock held.
6426 enum mc_target_type {
6432 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6433 unsigned long addr, pte_t ptent)
6435 struct page *page = vm_normal_page(vma, addr, ptent);
6437 if (!page || !page_mapped(page))
6439 if (PageAnon(page)) {
6440 /* we don't move shared anon */
6443 } else if (!move_file())
6444 /* we ignore mapcount for file pages */
6446 if (!get_page_unless_zero(page))
6453 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6454 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6456 struct page *page = NULL;
6457 swp_entry_t ent = pte_to_swp_entry(ptent);
6459 if (!move_anon() || non_swap_entry(ent))
6462 * Because lookup_swap_cache() updates some statistics counter,
6463 * we call find_get_page() with swapper_space directly.
6465 page = find_get_page(swap_address_space(ent), ent.val);
6466 if (do_swap_account)
6467 entry->val = ent.val;
6472 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6473 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6479 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6480 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6482 struct page *page = NULL;
6483 struct address_space *mapping;
6486 if (!vma->vm_file) /* anonymous vma */
6491 mapping = vma->vm_file->f_mapping;
6492 if (pte_none(ptent))
6493 pgoff = linear_page_index(vma, addr);
6494 else /* pte_file(ptent) is true */
6495 pgoff = pte_to_pgoff(ptent);
6497 /* page is moved even if it's not RSS of this task(page-faulted). */
6498 page = find_get_page(mapping, pgoff);
6501 /* shmem/tmpfs may report page out on swap: account for that too. */
6502 if (radix_tree_exceptional_entry(page)) {
6503 swp_entry_t swap = radix_to_swp_entry(page);
6504 if (do_swap_account)
6506 page = find_get_page(swap_address_space(swap), swap.val);
6512 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6513 unsigned long addr, pte_t ptent, union mc_target *target)
6515 struct page *page = NULL;
6516 struct page_cgroup *pc;
6517 enum mc_target_type ret = MC_TARGET_NONE;
6518 swp_entry_t ent = { .val = 0 };
6520 if (pte_present(ptent))
6521 page = mc_handle_present_pte(vma, addr, ptent);
6522 else if (is_swap_pte(ptent))
6523 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6524 else if (pte_none(ptent) || pte_file(ptent))
6525 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6527 if (!page && !ent.val)
6530 pc = lookup_page_cgroup(page);
6532 * Do only loose check w/o page_cgroup lock.
6533 * mem_cgroup_move_account() checks the pc is valid or not under
6536 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6537 ret = MC_TARGET_PAGE;
6539 target->page = page;
6541 if (!ret || !target)
6544 /* There is a swap entry and a page doesn't exist or isn't charged */
6545 if (ent.val && !ret &&
6546 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6547 ret = MC_TARGET_SWAP;
6554 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6556 * We don't consider swapping or file mapped pages because THP does not
6557 * support them for now.
6558 * Caller should make sure that pmd_trans_huge(pmd) is true.
6560 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6561 unsigned long addr, pmd_t pmd, union mc_target *target)
6563 struct page *page = NULL;
6564 struct page_cgroup *pc;
6565 enum mc_target_type ret = MC_TARGET_NONE;
6567 page = pmd_page(pmd);
6568 VM_BUG_ON(!page || !PageHead(page));
6571 pc = lookup_page_cgroup(page);
6572 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6573 ret = MC_TARGET_PAGE;
6576 target->page = page;
6582 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6583 unsigned long addr, pmd_t pmd, union mc_target *target)
6585 return MC_TARGET_NONE;
6589 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6590 unsigned long addr, unsigned long end,
6591 struct mm_walk *walk)
6593 struct vm_area_struct *vma = walk->private;
6597 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6598 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6599 mc.precharge += HPAGE_PMD_NR;
6600 spin_unlock(&vma->vm_mm->page_table_lock);
6604 if (pmd_trans_unstable(pmd))
6606 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6607 for (; addr != end; pte++, addr += PAGE_SIZE)
6608 if (get_mctgt_type(vma, addr, *pte, NULL))
6609 mc.precharge++; /* increment precharge temporarily */
6610 pte_unmap_unlock(pte - 1, ptl);
6616 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6618 unsigned long precharge;
6619 struct vm_area_struct *vma;
6621 down_read(&mm->mmap_sem);
6622 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6623 struct mm_walk mem_cgroup_count_precharge_walk = {
6624 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6628 if (is_vm_hugetlb_page(vma))
6630 walk_page_range(vma->vm_start, vma->vm_end,
6631 &mem_cgroup_count_precharge_walk);
6633 up_read(&mm->mmap_sem);
6635 precharge = mc.precharge;
6641 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6643 unsigned long precharge = mem_cgroup_count_precharge(mm);
6645 VM_BUG_ON(mc.moving_task);
6646 mc.moving_task = current;
6647 return mem_cgroup_do_precharge(precharge);
6650 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6651 static void __mem_cgroup_clear_mc(void)
6653 struct mem_cgroup *from = mc.from;
6654 struct mem_cgroup *to = mc.to;
6656 /* we must uncharge all the leftover precharges from mc.to */
6658 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6662 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6663 * we must uncharge here.
6665 if (mc.moved_charge) {
6666 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6667 mc.moved_charge = 0;
6669 /* we must fixup refcnts and charges */
6670 if (mc.moved_swap) {
6671 /* uncharge swap account from the old cgroup */
6672 if (!mem_cgroup_is_root(mc.from))
6673 res_counter_uncharge(&mc.from->memsw,
6674 PAGE_SIZE * mc.moved_swap);
6675 __mem_cgroup_put(mc.from, mc.moved_swap);
6677 if (!mem_cgroup_is_root(mc.to)) {
6679 * we charged both to->res and to->memsw, so we should
6682 res_counter_uncharge(&mc.to->res,
6683 PAGE_SIZE * mc.moved_swap);
6685 /* we've already done mem_cgroup_get(mc.to) */
6688 memcg_oom_recover(from);
6689 memcg_oom_recover(to);
6690 wake_up_all(&mc.waitq);
6693 static void mem_cgroup_clear_mc(void)
6695 struct mem_cgroup *from = mc.from;
6698 * we must clear moving_task before waking up waiters at the end of
6701 mc.moving_task = NULL;
6702 __mem_cgroup_clear_mc();
6703 spin_lock(&mc.lock);
6706 spin_unlock(&mc.lock);
6707 mem_cgroup_end_move(from);
6710 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6711 struct cgroup_taskset *tset)
6713 struct task_struct *p = cgroup_taskset_first(tset);
6715 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6716 unsigned long move_charge_at_immigrate;
6719 * We are now commited to this value whatever it is. Changes in this
6720 * tunable will only affect upcoming migrations, not the current one.
6721 * So we need to save it, and keep it going.
6723 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6724 if (move_charge_at_immigrate) {
6725 struct mm_struct *mm;
6726 struct mem_cgroup *from = mem_cgroup_from_task(p);
6728 VM_BUG_ON(from == memcg);
6730 mm = get_task_mm(p);
6733 /* We move charges only when we move a owner of the mm */
6734 if (mm->owner == p) {
6737 VM_BUG_ON(mc.precharge);
6738 VM_BUG_ON(mc.moved_charge);
6739 VM_BUG_ON(mc.moved_swap);
6740 mem_cgroup_start_move(from);
6741 spin_lock(&mc.lock);
6744 mc.immigrate_flags = move_charge_at_immigrate;
6745 spin_unlock(&mc.lock);
6746 /* We set mc.moving_task later */
6748 ret = mem_cgroup_precharge_mc(mm);
6750 mem_cgroup_clear_mc();
6757 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6758 struct cgroup_taskset *tset)
6760 mem_cgroup_clear_mc();
6763 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6764 unsigned long addr, unsigned long end,
6765 struct mm_walk *walk)
6768 struct vm_area_struct *vma = walk->private;
6771 enum mc_target_type target_type;
6772 union mc_target target;
6774 struct page_cgroup *pc;
6777 * We don't take compound_lock() here but no race with splitting thp
6779 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6780 * under splitting, which means there's no concurrent thp split,
6781 * - if another thread runs into split_huge_page() just after we
6782 * entered this if-block, the thread must wait for page table lock
6783 * to be unlocked in __split_huge_page_splitting(), where the main
6784 * part of thp split is not executed yet.
6786 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6787 if (mc.precharge < HPAGE_PMD_NR) {
6788 spin_unlock(&vma->vm_mm->page_table_lock);
6791 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6792 if (target_type == MC_TARGET_PAGE) {
6794 if (!isolate_lru_page(page)) {
6795 pc = lookup_page_cgroup(page);
6796 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6797 pc, mc.from, mc.to)) {
6798 mc.precharge -= HPAGE_PMD_NR;
6799 mc.moved_charge += HPAGE_PMD_NR;
6801 putback_lru_page(page);
6805 spin_unlock(&vma->vm_mm->page_table_lock);
6809 if (pmd_trans_unstable(pmd))
6812 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6813 for (; addr != end; addr += PAGE_SIZE) {
6814 pte_t ptent = *(pte++);
6820 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6821 case MC_TARGET_PAGE:
6823 if (isolate_lru_page(page))
6825 pc = lookup_page_cgroup(page);
6826 if (!mem_cgroup_move_account(page, 1, pc,
6829 /* we uncharge from mc.from later. */
6832 putback_lru_page(page);
6833 put: /* get_mctgt_type() gets the page */
6836 case MC_TARGET_SWAP:
6838 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6840 /* we fixup refcnts and charges later. */
6848 pte_unmap_unlock(pte - 1, ptl);
6853 * We have consumed all precharges we got in can_attach().
6854 * We try charge one by one, but don't do any additional
6855 * charges to mc.to if we have failed in charge once in attach()
6858 ret = mem_cgroup_do_precharge(1);
6866 static void mem_cgroup_move_charge(struct mm_struct *mm)
6868 struct vm_area_struct *vma;
6870 lru_add_drain_all();
6872 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6874 * Someone who are holding the mmap_sem might be waiting in
6875 * waitq. So we cancel all extra charges, wake up all waiters,
6876 * and retry. Because we cancel precharges, we might not be able
6877 * to move enough charges, but moving charge is a best-effort
6878 * feature anyway, so it wouldn't be a big problem.
6880 __mem_cgroup_clear_mc();
6884 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6886 struct mm_walk mem_cgroup_move_charge_walk = {
6887 .pmd_entry = mem_cgroup_move_charge_pte_range,
6891 if (is_vm_hugetlb_page(vma))
6893 ret = walk_page_range(vma->vm_start, vma->vm_end,
6894 &mem_cgroup_move_charge_walk);
6897 * means we have consumed all precharges and failed in
6898 * doing additional charge. Just abandon here.
6902 up_read(&mm->mmap_sem);
6905 static void mem_cgroup_move_task(struct cgroup *cont,
6906 struct cgroup_taskset *tset)
6908 struct task_struct *p = cgroup_taskset_first(tset);
6909 struct mm_struct *mm = get_task_mm(p);
6913 mem_cgroup_move_charge(mm);
6917 mem_cgroup_clear_mc();
6919 #else /* !CONFIG_MMU */
6920 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6921 struct cgroup_taskset *tset)
6925 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6926 struct cgroup_taskset *tset)
6929 static void mem_cgroup_move_task(struct cgroup *cont,
6930 struct cgroup_taskset *tset)
6936 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6937 * to verify sane_behavior flag on each mount attempt.
6939 static void mem_cgroup_bind(struct cgroup *root)
6942 * use_hierarchy is forced with sane_behavior. cgroup core
6943 * guarantees that @root doesn't have any children, so turning it
6944 * on for the root memcg is enough.
6946 if (cgroup_sane_behavior(root))
6947 mem_cgroup_from_cont(root)->use_hierarchy = true;
6950 struct cgroup_subsys mem_cgroup_subsys = {
6952 .subsys_id = mem_cgroup_subsys_id,
6953 .css_alloc = mem_cgroup_css_alloc,
6954 .css_online = mem_cgroup_css_online,
6955 .css_offline = mem_cgroup_css_offline,
6956 .css_free = mem_cgroup_css_free,
6957 .can_attach = mem_cgroup_can_attach,
6958 .cancel_attach = mem_cgroup_cancel_attach,
6959 .attach = mem_cgroup_move_task,
6960 .bind = mem_cgroup_bind,
6961 .base_cftypes = mem_cgroup_files,
6966 #ifdef CONFIG_MEMCG_SWAP
6967 static int __init enable_swap_account(char *s)
6969 /* consider enabled if no parameter or 1 is given */
6970 if (!strcmp(s, "1"))
6971 really_do_swap_account = 1;
6972 else if (!strcmp(s, "0"))
6973 really_do_swap_account = 0;
6976 __setup("swapaccount=", enable_swap_account);
6978 static void __init memsw_file_init(void)
6980 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6983 static void __init enable_swap_cgroup(void)
6985 if (!mem_cgroup_disabled() && really_do_swap_account) {
6986 do_swap_account = 1;
6992 static void __init enable_swap_cgroup(void)
6998 * subsys_initcall() for memory controller.
7000 * Some parts like hotcpu_notifier() have to be initialized from this context
7001 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7002 * everything that doesn't depend on a specific mem_cgroup structure should
7003 * be initialized from here.
7005 static int __init mem_cgroup_init(void)
7007 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7008 enable_swap_cgroup();
7009 mem_cgroup_soft_limit_tree_init();
7013 subsys_initcall(mem_cgroup_init);