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 last_visited = iter->last_visited;
1203 if (prev && reclaim->generation != iter->generation) {
1204 iter->last_visited = NULL;
1209 * If the dead_count mismatches, a destruction
1210 * has happened or is happening concurrently.
1211 * If the dead_count matches, a destruction
1212 * might still happen concurrently, but since
1213 * we checked under RCU, that destruction
1214 * won't free the object until we release the
1215 * RCU reader lock. Thus, the dead_count
1216 * check verifies the pointer is still valid,
1217 * css_tryget() verifies the cgroup pointed to
1220 dead_count = atomic_read(&root->dead_count);
1222 last_visited = iter->last_visited;
1224 if ((dead_count != iter->last_dead_count) ||
1225 !css_tryget(&last_visited->css)) {
1226 last_visited = NULL;
1231 memcg = __mem_cgroup_iter_next(root, last_visited);
1235 css_put(&last_visited->css);
1237 iter->last_visited = memcg;
1239 iter->last_dead_count = dead_count;
1243 else if (!prev && memcg)
1244 reclaim->generation = iter->generation;
1253 if (prev && prev != root)
1254 css_put(&prev->css);
1260 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1261 * @root: hierarchy root
1262 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1264 void mem_cgroup_iter_break(struct mem_cgroup *root,
1265 struct mem_cgroup *prev)
1268 root = root_mem_cgroup;
1269 if (prev && prev != root)
1270 css_put(&prev->css);
1274 * Iteration constructs for visiting all cgroups (under a tree). If
1275 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1276 * be used for reference counting.
1278 #define for_each_mem_cgroup_tree(iter, root) \
1279 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1281 iter = mem_cgroup_iter(root, iter, NULL))
1283 #define for_each_mem_cgroup(iter) \
1284 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1286 iter = mem_cgroup_iter(NULL, iter, NULL))
1288 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1290 struct mem_cgroup *memcg;
1293 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1294 if (unlikely(!memcg))
1299 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1302 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1310 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1313 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1314 * @zone: zone of the wanted lruvec
1315 * @memcg: memcg of the wanted lruvec
1317 * Returns the lru list vector holding pages for the given @zone and
1318 * @mem. This can be the global zone lruvec, if the memory controller
1321 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1322 struct mem_cgroup *memcg)
1324 struct mem_cgroup_per_zone *mz;
1325 struct lruvec *lruvec;
1327 if (mem_cgroup_disabled()) {
1328 lruvec = &zone->lruvec;
1332 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1333 lruvec = &mz->lruvec;
1336 * Since a node can be onlined after the mem_cgroup was created,
1337 * we have to be prepared to initialize lruvec->zone here;
1338 * and if offlined then reonlined, we need to reinitialize it.
1340 if (unlikely(lruvec->zone != zone))
1341 lruvec->zone = zone;
1346 * Following LRU functions are allowed to be used without PCG_LOCK.
1347 * Operations are called by routine of global LRU independently from memcg.
1348 * What we have to take care of here is validness of pc->mem_cgroup.
1350 * Changes to pc->mem_cgroup happens when
1353 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1354 * It is added to LRU before charge.
1355 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1356 * When moving account, the page is not on LRU. It's isolated.
1360 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1362 * @zone: zone of the page
1364 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1366 struct mem_cgroup_per_zone *mz;
1367 struct mem_cgroup *memcg;
1368 struct page_cgroup *pc;
1369 struct lruvec *lruvec;
1371 if (mem_cgroup_disabled()) {
1372 lruvec = &zone->lruvec;
1376 pc = lookup_page_cgroup(page);
1377 memcg = pc->mem_cgroup;
1380 * Surreptitiously switch any uncharged offlist page to root:
1381 * an uncharged page off lru does nothing to secure
1382 * its former mem_cgroup from sudden removal.
1384 * Our caller holds lru_lock, and PageCgroupUsed is updated
1385 * under page_cgroup lock: between them, they make all uses
1386 * of pc->mem_cgroup safe.
1388 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1389 pc->mem_cgroup = memcg = root_mem_cgroup;
1391 mz = page_cgroup_zoneinfo(memcg, page);
1392 lruvec = &mz->lruvec;
1395 * Since a node can be onlined after the mem_cgroup was created,
1396 * we have to be prepared to initialize lruvec->zone here;
1397 * and if offlined then reonlined, we need to reinitialize it.
1399 if (unlikely(lruvec->zone != zone))
1400 lruvec->zone = zone;
1405 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1406 * @lruvec: mem_cgroup per zone lru vector
1407 * @lru: index of lru list the page is sitting on
1408 * @nr_pages: positive when adding or negative when removing
1410 * This function must be called when a page is added to or removed from an
1413 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1416 struct mem_cgroup_per_zone *mz;
1417 unsigned long *lru_size;
1419 if (mem_cgroup_disabled())
1422 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1423 lru_size = mz->lru_size + lru;
1424 *lru_size += nr_pages;
1425 VM_BUG_ON((long)(*lru_size) < 0);
1429 * Checks whether given mem is same or in the root_mem_cgroup's
1432 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1433 struct mem_cgroup *memcg)
1435 if (root_memcg == memcg)
1437 if (!root_memcg->use_hierarchy || !memcg)
1439 return css_is_ancestor(&memcg->css, &root_memcg->css);
1442 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1443 struct mem_cgroup *memcg)
1448 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1453 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1456 struct mem_cgroup *curr = NULL;
1457 struct task_struct *p;
1459 p = find_lock_task_mm(task);
1461 curr = try_get_mem_cgroup_from_mm(p->mm);
1465 * All threads may have already detached their mm's, but the oom
1466 * killer still needs to detect if they have already been oom
1467 * killed to prevent needlessly killing additional tasks.
1470 curr = mem_cgroup_from_task(task);
1472 css_get(&curr->css);
1478 * We should check use_hierarchy of "memcg" not "curr". Because checking
1479 * use_hierarchy of "curr" here make this function true if hierarchy is
1480 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1481 * hierarchy(even if use_hierarchy is disabled in "memcg").
1483 ret = mem_cgroup_same_or_subtree(memcg, curr);
1484 css_put(&curr->css);
1488 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1490 unsigned long inactive_ratio;
1491 unsigned long inactive;
1492 unsigned long active;
1495 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1496 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1498 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1500 inactive_ratio = int_sqrt(10 * gb);
1504 return inactive * inactive_ratio < active;
1507 #define mem_cgroup_from_res_counter(counter, member) \
1508 container_of(counter, struct mem_cgroup, member)
1511 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1512 * @memcg: the memory cgroup
1514 * Returns the maximum amount of memory @mem can be charged with, in
1517 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1519 unsigned long long margin;
1521 margin = res_counter_margin(&memcg->res);
1522 if (do_swap_account)
1523 margin = min(margin, res_counter_margin(&memcg->memsw));
1524 return margin >> PAGE_SHIFT;
1527 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1529 struct cgroup *cgrp = memcg->css.cgroup;
1532 if (cgrp->parent == NULL)
1533 return vm_swappiness;
1535 return memcg->swappiness;
1539 * memcg->moving_account is used for checking possibility that some thread is
1540 * calling move_account(). When a thread on CPU-A starts moving pages under
1541 * a memcg, other threads should check memcg->moving_account under
1542 * rcu_read_lock(), like this:
1546 * memcg->moving_account+1 if (memcg->mocing_account)
1548 * synchronize_rcu() update something.
1553 /* for quick checking without looking up memcg */
1554 atomic_t memcg_moving __read_mostly;
1556 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1558 atomic_inc(&memcg_moving);
1559 atomic_inc(&memcg->moving_account);
1563 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1566 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1567 * We check NULL in callee rather than caller.
1570 atomic_dec(&memcg_moving);
1571 atomic_dec(&memcg->moving_account);
1576 * 2 routines for checking "mem" is under move_account() or not.
1578 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1579 * is used for avoiding races in accounting. If true,
1580 * pc->mem_cgroup may be overwritten.
1582 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1583 * under hierarchy of moving cgroups. This is for
1584 * waiting at hith-memory prressure caused by "move".
1587 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1589 VM_BUG_ON(!rcu_read_lock_held());
1590 return atomic_read(&memcg->moving_account) > 0;
1593 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1595 struct mem_cgroup *from;
1596 struct mem_cgroup *to;
1599 * Unlike task_move routines, we access mc.to, mc.from not under
1600 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1602 spin_lock(&mc.lock);
1608 ret = mem_cgroup_same_or_subtree(memcg, from)
1609 || mem_cgroup_same_or_subtree(memcg, to);
1611 spin_unlock(&mc.lock);
1615 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1617 if (mc.moving_task && current != mc.moving_task) {
1618 if (mem_cgroup_under_move(memcg)) {
1620 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1621 /* moving charge context might have finished. */
1624 finish_wait(&mc.waitq, &wait);
1632 * Take this lock when
1633 * - a code tries to modify page's memcg while it's USED.
1634 * - a code tries to modify page state accounting in a memcg.
1635 * see mem_cgroup_stolen(), too.
1637 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1638 unsigned long *flags)
1640 spin_lock_irqsave(&memcg->move_lock, *flags);
1643 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1644 unsigned long *flags)
1646 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1649 #define K(x) ((x) << (PAGE_SHIFT-10))
1651 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1652 * @memcg: The memory cgroup that went over limit
1653 * @p: Task that is going to be killed
1655 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1658 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1660 struct cgroup *task_cgrp;
1661 struct cgroup *mem_cgrp;
1663 * Need a buffer in BSS, can't rely on allocations. The code relies
1664 * on the assumption that OOM is serialized for memory controller.
1665 * If this assumption is broken, revisit this code.
1667 static char memcg_name[PATH_MAX];
1669 struct mem_cgroup *iter;
1677 mem_cgrp = memcg->css.cgroup;
1678 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1680 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1683 * Unfortunately, we are unable to convert to a useful name
1684 * But we'll still print out the usage information
1691 pr_info("Task in %s killed", memcg_name);
1694 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1702 * Continues from above, so we don't need an KERN_ level
1704 pr_cont(" as a result of limit of %s\n", memcg_name);
1707 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1708 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1709 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1710 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1711 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1712 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1713 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1714 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1715 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1716 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1717 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1718 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1720 for_each_mem_cgroup_tree(iter, memcg) {
1721 pr_info("Memory cgroup stats");
1724 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1726 pr_cont(" for %s", memcg_name);
1730 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1731 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1733 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1734 K(mem_cgroup_read_stat(iter, i)));
1737 for (i = 0; i < NR_LRU_LISTS; i++)
1738 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1739 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1746 * This function returns the number of memcg under hierarchy tree. Returns
1747 * 1(self count) if no children.
1749 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1752 struct mem_cgroup *iter;
1754 for_each_mem_cgroup_tree(iter, memcg)
1760 * Return the memory (and swap, if configured) limit for a memcg.
1762 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1766 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1769 * Do not consider swap space if we cannot swap due to swappiness
1771 if (mem_cgroup_swappiness(memcg)) {
1774 limit += total_swap_pages << PAGE_SHIFT;
1775 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1778 * If memsw is finite and limits the amount of swap space
1779 * available to this memcg, return that limit.
1781 limit = min(limit, memsw);
1787 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1790 struct mem_cgroup *iter;
1791 unsigned long chosen_points = 0;
1792 unsigned long totalpages;
1793 unsigned int points = 0;
1794 struct task_struct *chosen = NULL;
1797 * If current has a pending SIGKILL or is exiting, then automatically
1798 * select it. The goal is to allow it to allocate so that it may
1799 * quickly exit and free its memory.
1801 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1802 set_thread_flag(TIF_MEMDIE);
1806 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1807 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1808 for_each_mem_cgroup_tree(iter, memcg) {
1809 struct cgroup *cgroup = iter->css.cgroup;
1810 struct cgroup_iter it;
1811 struct task_struct *task;
1813 cgroup_iter_start(cgroup, &it);
1814 while ((task = cgroup_iter_next(cgroup, &it))) {
1815 switch (oom_scan_process_thread(task, totalpages, NULL,
1817 case OOM_SCAN_SELECT:
1819 put_task_struct(chosen);
1821 chosen_points = ULONG_MAX;
1822 get_task_struct(chosen);
1824 case OOM_SCAN_CONTINUE:
1826 case OOM_SCAN_ABORT:
1827 cgroup_iter_end(cgroup, &it);
1828 mem_cgroup_iter_break(memcg, iter);
1830 put_task_struct(chosen);
1835 points = oom_badness(task, memcg, NULL, totalpages);
1836 if (points > chosen_points) {
1838 put_task_struct(chosen);
1840 chosen_points = points;
1841 get_task_struct(chosen);
1844 cgroup_iter_end(cgroup, &it);
1849 points = chosen_points * 1000 / totalpages;
1850 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1851 NULL, "Memory cgroup out of memory");
1854 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1856 unsigned long flags)
1858 unsigned long total = 0;
1859 bool noswap = false;
1862 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1864 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1867 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1869 drain_all_stock_async(memcg);
1870 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1872 * Allow limit shrinkers, which are triggered directly
1873 * by userspace, to catch signals and stop reclaim
1874 * after minimal progress, regardless of the margin.
1876 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1878 if (mem_cgroup_margin(memcg))
1881 * If nothing was reclaimed after two attempts, there
1882 * may be no reclaimable pages in this hierarchy.
1891 * test_mem_cgroup_node_reclaimable
1892 * @memcg: the target memcg
1893 * @nid: the node ID to be checked.
1894 * @noswap : specify true here if the user wants flle only information.
1896 * This function returns whether the specified memcg contains any
1897 * reclaimable pages on a node. Returns true if there are any reclaimable
1898 * pages in the node.
1900 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1901 int nid, bool noswap)
1903 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1905 if (noswap || !total_swap_pages)
1907 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1912 #if MAX_NUMNODES > 1
1915 * Always updating the nodemask is not very good - even if we have an empty
1916 * list or the wrong list here, we can start from some node and traverse all
1917 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1920 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1924 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1925 * pagein/pageout changes since the last update.
1927 if (!atomic_read(&memcg->numainfo_events))
1929 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1932 /* make a nodemask where this memcg uses memory from */
1933 memcg->scan_nodes = node_states[N_MEMORY];
1935 for_each_node_mask(nid, node_states[N_MEMORY]) {
1937 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1938 node_clear(nid, memcg->scan_nodes);
1941 atomic_set(&memcg->numainfo_events, 0);
1942 atomic_set(&memcg->numainfo_updating, 0);
1946 * Selecting a node where we start reclaim from. Because what we need is just
1947 * reducing usage counter, start from anywhere is O,K. Considering
1948 * memory reclaim from current node, there are pros. and cons.
1950 * Freeing memory from current node means freeing memory from a node which
1951 * we'll use or we've used. So, it may make LRU bad. And if several threads
1952 * hit limits, it will see a contention on a node. But freeing from remote
1953 * node means more costs for memory reclaim because of memory latency.
1955 * Now, we use round-robin. Better algorithm is welcomed.
1957 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1961 mem_cgroup_may_update_nodemask(memcg);
1962 node = memcg->last_scanned_node;
1964 node = next_node(node, memcg->scan_nodes);
1965 if (node == MAX_NUMNODES)
1966 node = first_node(memcg->scan_nodes);
1968 * We call this when we hit limit, not when pages are added to LRU.
1969 * No LRU may hold pages because all pages are UNEVICTABLE or
1970 * memcg is too small and all pages are not on LRU. In that case,
1971 * we use curret node.
1973 if (unlikely(node == MAX_NUMNODES))
1974 node = numa_node_id();
1976 memcg->last_scanned_node = node;
1981 * Check all nodes whether it contains reclaimable pages or not.
1982 * For quick scan, we make use of scan_nodes. This will allow us to skip
1983 * unused nodes. But scan_nodes is lazily updated and may not cotain
1984 * enough new information. We need to do double check.
1986 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1991 * quick check...making use of scan_node.
1992 * We can skip unused nodes.
1994 if (!nodes_empty(memcg->scan_nodes)) {
1995 for (nid = first_node(memcg->scan_nodes);
1997 nid = next_node(nid, memcg->scan_nodes)) {
1999 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2004 * Check rest of nodes.
2006 for_each_node_state(nid, N_MEMORY) {
2007 if (node_isset(nid, memcg->scan_nodes))
2009 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2016 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2021 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2023 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2027 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2030 unsigned long *total_scanned)
2032 struct mem_cgroup *victim = NULL;
2035 unsigned long excess;
2036 unsigned long nr_scanned;
2037 struct mem_cgroup_reclaim_cookie reclaim = {
2042 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2045 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2050 * If we have not been able to reclaim
2051 * anything, it might because there are
2052 * no reclaimable pages under this hierarchy
2057 * We want to do more targeted reclaim.
2058 * excess >> 2 is not to excessive so as to
2059 * reclaim too much, nor too less that we keep
2060 * coming back to reclaim from this cgroup
2062 if (total >= (excess >> 2) ||
2063 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2068 if (!mem_cgroup_reclaimable(victim, false))
2070 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2072 *total_scanned += nr_scanned;
2073 if (!res_counter_soft_limit_excess(&root_memcg->res))
2076 mem_cgroup_iter_break(root_memcg, victim);
2081 * Check OOM-Killer is already running under our hierarchy.
2082 * If someone is running, return false.
2083 * Has to be called with memcg_oom_lock
2085 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
2087 struct mem_cgroup *iter, *failed = NULL;
2089 for_each_mem_cgroup_tree(iter, memcg) {
2090 if (iter->oom_lock) {
2092 * this subtree of our hierarchy is already locked
2093 * so we cannot give a lock.
2096 mem_cgroup_iter_break(memcg, iter);
2099 iter->oom_lock = true;
2106 * OK, we failed to lock the whole subtree so we have to clean up
2107 * what we set up to the failing subtree
2109 for_each_mem_cgroup_tree(iter, memcg) {
2110 if (iter == failed) {
2111 mem_cgroup_iter_break(memcg, iter);
2114 iter->oom_lock = false;
2120 * Has to be called with memcg_oom_lock
2122 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2124 struct mem_cgroup *iter;
2126 for_each_mem_cgroup_tree(iter, memcg)
2127 iter->oom_lock = false;
2131 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2133 struct mem_cgroup *iter;
2135 for_each_mem_cgroup_tree(iter, memcg)
2136 atomic_inc(&iter->under_oom);
2139 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2141 struct mem_cgroup *iter;
2144 * When a new child is created while the hierarchy is under oom,
2145 * mem_cgroup_oom_lock() may not be called. We have to use
2146 * atomic_add_unless() here.
2148 for_each_mem_cgroup_tree(iter, memcg)
2149 atomic_add_unless(&iter->under_oom, -1, 0);
2152 static DEFINE_SPINLOCK(memcg_oom_lock);
2153 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2155 struct oom_wait_info {
2156 struct mem_cgroup *memcg;
2160 static int memcg_oom_wake_function(wait_queue_t *wait,
2161 unsigned mode, int sync, void *arg)
2163 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2164 struct mem_cgroup *oom_wait_memcg;
2165 struct oom_wait_info *oom_wait_info;
2167 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2168 oom_wait_memcg = oom_wait_info->memcg;
2171 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2172 * Then we can use css_is_ancestor without taking care of RCU.
2174 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2175 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2177 return autoremove_wake_function(wait, mode, sync, arg);
2180 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2182 /* for filtering, pass "memcg" as argument. */
2183 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2186 static void memcg_oom_recover(struct mem_cgroup *memcg)
2188 if (memcg && atomic_read(&memcg->under_oom))
2189 memcg_wakeup_oom(memcg);
2193 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2195 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2198 struct oom_wait_info owait;
2199 bool locked, need_to_kill;
2201 owait.memcg = memcg;
2202 owait.wait.flags = 0;
2203 owait.wait.func = memcg_oom_wake_function;
2204 owait.wait.private = current;
2205 INIT_LIST_HEAD(&owait.wait.task_list);
2206 need_to_kill = true;
2207 mem_cgroup_mark_under_oom(memcg);
2209 /* At first, try to OOM lock hierarchy under memcg.*/
2210 spin_lock(&memcg_oom_lock);
2211 locked = mem_cgroup_oom_lock(memcg);
2213 * Even if signal_pending(), we can't quit charge() loop without
2214 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2215 * under OOM is always welcomed, use TASK_KILLABLE here.
2217 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2218 if (!locked || memcg->oom_kill_disable)
2219 need_to_kill = false;
2221 mem_cgroup_oom_notify(memcg);
2222 spin_unlock(&memcg_oom_lock);
2225 finish_wait(&memcg_oom_waitq, &owait.wait);
2226 mem_cgroup_out_of_memory(memcg, mask, order);
2229 finish_wait(&memcg_oom_waitq, &owait.wait);
2231 spin_lock(&memcg_oom_lock);
2233 mem_cgroup_oom_unlock(memcg);
2234 memcg_wakeup_oom(memcg);
2235 spin_unlock(&memcg_oom_lock);
2237 mem_cgroup_unmark_under_oom(memcg);
2239 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2241 /* Give chance to dying process */
2242 schedule_timeout_uninterruptible(1);
2247 * Currently used to update mapped file statistics, but the routine can be
2248 * generalized to update other statistics as well.
2250 * Notes: Race condition
2252 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2253 * it tends to be costly. But considering some conditions, we doesn't need
2254 * to do so _always_.
2256 * Considering "charge", lock_page_cgroup() is not required because all
2257 * file-stat operations happen after a page is attached to radix-tree. There
2258 * are no race with "charge".
2260 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2261 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2262 * if there are race with "uncharge". Statistics itself is properly handled
2265 * Considering "move", this is an only case we see a race. To make the race
2266 * small, we check mm->moving_account and detect there are possibility of race
2267 * If there is, we take a lock.
2270 void __mem_cgroup_begin_update_page_stat(struct page *page,
2271 bool *locked, unsigned long *flags)
2273 struct mem_cgroup *memcg;
2274 struct page_cgroup *pc;
2276 pc = lookup_page_cgroup(page);
2278 memcg = pc->mem_cgroup;
2279 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2282 * If this memory cgroup is not under account moving, we don't
2283 * need to take move_lock_mem_cgroup(). Because we already hold
2284 * rcu_read_lock(), any calls to move_account will be delayed until
2285 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2287 if (!mem_cgroup_stolen(memcg))
2290 move_lock_mem_cgroup(memcg, flags);
2291 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2292 move_unlock_mem_cgroup(memcg, flags);
2298 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2300 struct page_cgroup *pc = lookup_page_cgroup(page);
2303 * It's guaranteed that pc->mem_cgroup never changes while
2304 * lock is held because a routine modifies pc->mem_cgroup
2305 * should take move_lock_mem_cgroup().
2307 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2310 void mem_cgroup_update_page_stat(struct page *page,
2311 enum mem_cgroup_page_stat_item idx, int val)
2313 struct mem_cgroup *memcg;
2314 struct page_cgroup *pc = lookup_page_cgroup(page);
2315 unsigned long uninitialized_var(flags);
2317 if (mem_cgroup_disabled())
2320 memcg = pc->mem_cgroup;
2321 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2325 case MEMCG_NR_FILE_MAPPED:
2326 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2332 this_cpu_add(memcg->stat->count[idx], val);
2336 * size of first charge trial. "32" comes from vmscan.c's magic value.
2337 * TODO: maybe necessary to use big numbers in big irons.
2339 #define CHARGE_BATCH 32U
2340 struct memcg_stock_pcp {
2341 struct mem_cgroup *cached; /* this never be root cgroup */
2342 unsigned int nr_pages;
2343 struct work_struct work;
2344 unsigned long flags;
2345 #define FLUSHING_CACHED_CHARGE 0
2347 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2348 static DEFINE_MUTEX(percpu_charge_mutex);
2351 * consume_stock: Try to consume stocked charge on this cpu.
2352 * @memcg: memcg to consume from.
2353 * @nr_pages: how many pages to charge.
2355 * The charges will only happen if @memcg matches the current cpu's memcg
2356 * stock, and at least @nr_pages are available in that stock. Failure to
2357 * service an allocation will refill the stock.
2359 * returns true if successful, false otherwise.
2361 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2363 struct memcg_stock_pcp *stock;
2366 if (nr_pages > CHARGE_BATCH)
2369 stock = &get_cpu_var(memcg_stock);
2370 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2371 stock->nr_pages -= nr_pages;
2372 else /* need to call res_counter_charge */
2374 put_cpu_var(memcg_stock);
2379 * Returns stocks cached in percpu to res_counter and reset cached information.
2381 static void drain_stock(struct memcg_stock_pcp *stock)
2383 struct mem_cgroup *old = stock->cached;
2385 if (stock->nr_pages) {
2386 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2388 res_counter_uncharge(&old->res, bytes);
2389 if (do_swap_account)
2390 res_counter_uncharge(&old->memsw, bytes);
2391 stock->nr_pages = 0;
2393 stock->cached = NULL;
2397 * This must be called under preempt disabled or must be called by
2398 * a thread which is pinned to local cpu.
2400 static void drain_local_stock(struct work_struct *dummy)
2402 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2404 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2407 static void __init memcg_stock_init(void)
2411 for_each_possible_cpu(cpu) {
2412 struct memcg_stock_pcp *stock =
2413 &per_cpu(memcg_stock, cpu);
2414 INIT_WORK(&stock->work, drain_local_stock);
2419 * Cache charges(val) which is from res_counter, to local per_cpu area.
2420 * This will be consumed by consume_stock() function, later.
2422 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2424 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2426 if (stock->cached != memcg) { /* reset if necessary */
2428 stock->cached = memcg;
2430 stock->nr_pages += nr_pages;
2431 put_cpu_var(memcg_stock);
2435 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2436 * of the hierarchy under it. sync flag says whether we should block
2437 * until the work is done.
2439 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2443 /* Notify other cpus that system-wide "drain" is running */
2446 for_each_online_cpu(cpu) {
2447 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2448 struct mem_cgroup *memcg;
2450 memcg = stock->cached;
2451 if (!memcg || !stock->nr_pages)
2453 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2455 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2457 drain_local_stock(&stock->work);
2459 schedule_work_on(cpu, &stock->work);
2467 for_each_online_cpu(cpu) {
2468 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2469 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2470 flush_work(&stock->work);
2477 * Tries to drain stocked charges in other cpus. This function is asynchronous
2478 * and just put a work per cpu for draining localy on each cpu. Caller can
2479 * expects some charges will be back to res_counter later but cannot wait for
2482 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2485 * If someone calls draining, avoid adding more kworker runs.
2487 if (!mutex_trylock(&percpu_charge_mutex))
2489 drain_all_stock(root_memcg, false);
2490 mutex_unlock(&percpu_charge_mutex);
2493 /* This is a synchronous drain interface. */
2494 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2496 /* called when force_empty is called */
2497 mutex_lock(&percpu_charge_mutex);
2498 drain_all_stock(root_memcg, true);
2499 mutex_unlock(&percpu_charge_mutex);
2503 * This function drains percpu counter value from DEAD cpu and
2504 * move it to local cpu. Note that this function can be preempted.
2506 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2510 spin_lock(&memcg->pcp_counter_lock);
2511 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2512 long x = per_cpu(memcg->stat->count[i], cpu);
2514 per_cpu(memcg->stat->count[i], cpu) = 0;
2515 memcg->nocpu_base.count[i] += x;
2517 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2518 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2520 per_cpu(memcg->stat->events[i], cpu) = 0;
2521 memcg->nocpu_base.events[i] += x;
2523 spin_unlock(&memcg->pcp_counter_lock);
2526 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2527 unsigned long action,
2530 int cpu = (unsigned long)hcpu;
2531 struct memcg_stock_pcp *stock;
2532 struct mem_cgroup *iter;
2534 if (action == CPU_ONLINE)
2537 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2540 for_each_mem_cgroup(iter)
2541 mem_cgroup_drain_pcp_counter(iter, cpu);
2543 stock = &per_cpu(memcg_stock, cpu);
2549 /* See __mem_cgroup_try_charge() for details */
2551 CHARGE_OK, /* success */
2552 CHARGE_RETRY, /* need to retry but retry is not bad */
2553 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2554 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2555 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2558 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2559 unsigned int nr_pages, unsigned int min_pages,
2562 unsigned long csize = nr_pages * PAGE_SIZE;
2563 struct mem_cgroup *mem_over_limit;
2564 struct res_counter *fail_res;
2565 unsigned long flags = 0;
2568 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2571 if (!do_swap_account)
2573 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2577 res_counter_uncharge(&memcg->res, csize);
2578 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2579 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2581 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2583 * Never reclaim on behalf of optional batching, retry with a
2584 * single page instead.
2586 if (nr_pages > min_pages)
2587 return CHARGE_RETRY;
2589 if (!(gfp_mask & __GFP_WAIT))
2590 return CHARGE_WOULDBLOCK;
2592 if (gfp_mask & __GFP_NORETRY)
2593 return CHARGE_NOMEM;
2595 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2596 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2597 return CHARGE_RETRY;
2599 * Even though the limit is exceeded at this point, reclaim
2600 * may have been able to free some pages. Retry the charge
2601 * before killing the task.
2603 * Only for regular pages, though: huge pages are rather
2604 * unlikely to succeed so close to the limit, and we fall back
2605 * to regular pages anyway in case of failure.
2607 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2608 return CHARGE_RETRY;
2611 * At task move, charge accounts can be doubly counted. So, it's
2612 * better to wait until the end of task_move if something is going on.
2614 if (mem_cgroup_wait_acct_move(mem_over_limit))
2615 return CHARGE_RETRY;
2617 /* If we don't need to call oom-killer at el, return immediately */
2619 return CHARGE_NOMEM;
2621 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2622 return CHARGE_OOM_DIE;
2624 return CHARGE_RETRY;
2628 * __mem_cgroup_try_charge() does
2629 * 1. detect memcg to be charged against from passed *mm and *ptr,
2630 * 2. update res_counter
2631 * 3. call memory reclaim if necessary.
2633 * In some special case, if the task is fatal, fatal_signal_pending() or
2634 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2635 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2636 * as possible without any hazards. 2: all pages should have a valid
2637 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2638 * pointer, that is treated as a charge to root_mem_cgroup.
2640 * So __mem_cgroup_try_charge() will return
2641 * 0 ... on success, filling *ptr with a valid memcg pointer.
2642 * -ENOMEM ... charge failure because of resource limits.
2643 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2645 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2646 * the oom-killer can be invoked.
2648 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2650 unsigned int nr_pages,
2651 struct mem_cgroup **ptr,
2654 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2655 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2656 struct mem_cgroup *memcg = NULL;
2660 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2661 * in system level. So, allow to go ahead dying process in addition to
2664 if (unlikely(test_thread_flag(TIF_MEMDIE)
2665 || fatal_signal_pending(current)))
2669 * We always charge the cgroup the mm_struct belongs to.
2670 * The mm_struct's mem_cgroup changes on task migration if the
2671 * thread group leader migrates. It's possible that mm is not
2672 * set, if so charge the root memcg (happens for pagecache usage).
2675 *ptr = root_mem_cgroup;
2677 if (*ptr) { /* css should be a valid one */
2679 if (mem_cgroup_is_root(memcg))
2681 if (consume_stock(memcg, nr_pages))
2683 css_get(&memcg->css);
2685 struct task_struct *p;
2688 p = rcu_dereference(mm->owner);
2690 * Because we don't have task_lock(), "p" can exit.
2691 * In that case, "memcg" can point to root or p can be NULL with
2692 * race with swapoff. Then, we have small risk of mis-accouning.
2693 * But such kind of mis-account by race always happens because
2694 * we don't have cgroup_mutex(). It's overkill and we allo that
2696 * (*) swapoff at el will charge against mm-struct not against
2697 * task-struct. So, mm->owner can be NULL.
2699 memcg = mem_cgroup_from_task(p);
2701 memcg = root_mem_cgroup;
2702 if (mem_cgroup_is_root(memcg)) {
2706 if (consume_stock(memcg, nr_pages)) {
2708 * It seems dagerous to access memcg without css_get().
2709 * But considering how consume_stok works, it's not
2710 * necessary. If consume_stock success, some charges
2711 * from this memcg are cached on this cpu. So, we
2712 * don't need to call css_get()/css_tryget() before
2713 * calling consume_stock().
2718 /* after here, we may be blocked. we need to get refcnt */
2719 if (!css_tryget(&memcg->css)) {
2729 /* If killed, bypass charge */
2730 if (fatal_signal_pending(current)) {
2731 css_put(&memcg->css);
2736 if (oom && !nr_oom_retries) {
2738 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2741 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2746 case CHARGE_RETRY: /* not in OOM situation but retry */
2748 css_put(&memcg->css);
2751 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2752 css_put(&memcg->css);
2754 case CHARGE_NOMEM: /* OOM routine works */
2756 css_put(&memcg->css);
2759 /* If oom, we never return -ENOMEM */
2762 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2763 css_put(&memcg->css);
2766 } while (ret != CHARGE_OK);
2768 if (batch > nr_pages)
2769 refill_stock(memcg, batch - nr_pages);
2770 css_put(&memcg->css);
2778 *ptr = root_mem_cgroup;
2783 * Somemtimes we have to undo a charge we got by try_charge().
2784 * This function is for that and do uncharge, put css's refcnt.
2785 * gotten by try_charge().
2787 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2788 unsigned int nr_pages)
2790 if (!mem_cgroup_is_root(memcg)) {
2791 unsigned long bytes = nr_pages * PAGE_SIZE;
2793 res_counter_uncharge(&memcg->res, bytes);
2794 if (do_swap_account)
2795 res_counter_uncharge(&memcg->memsw, bytes);
2800 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2801 * This is useful when moving usage to parent cgroup.
2803 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2804 unsigned int nr_pages)
2806 unsigned long bytes = nr_pages * PAGE_SIZE;
2808 if (mem_cgroup_is_root(memcg))
2811 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2812 if (do_swap_account)
2813 res_counter_uncharge_until(&memcg->memsw,
2814 memcg->memsw.parent, bytes);
2818 * A helper function to get mem_cgroup from ID. must be called under
2819 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2820 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2821 * called against removed memcg.)
2823 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2825 struct cgroup_subsys_state *css;
2827 /* ID 0 is unused ID */
2830 css = css_lookup(&mem_cgroup_subsys, id);
2833 return mem_cgroup_from_css(css);
2836 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2838 struct mem_cgroup *memcg = NULL;
2839 struct page_cgroup *pc;
2843 VM_BUG_ON(!PageLocked(page));
2845 pc = lookup_page_cgroup(page);
2846 lock_page_cgroup(pc);
2847 if (PageCgroupUsed(pc)) {
2848 memcg = pc->mem_cgroup;
2849 if (memcg && !css_tryget(&memcg->css))
2851 } else if (PageSwapCache(page)) {
2852 ent.val = page_private(page);
2853 id = lookup_swap_cgroup_id(ent);
2855 memcg = mem_cgroup_lookup(id);
2856 if (memcg && !css_tryget(&memcg->css))
2860 unlock_page_cgroup(pc);
2864 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2866 unsigned int nr_pages,
2867 enum charge_type ctype,
2870 struct page_cgroup *pc = lookup_page_cgroup(page);
2871 struct zone *uninitialized_var(zone);
2872 struct lruvec *lruvec;
2873 bool was_on_lru = false;
2876 lock_page_cgroup(pc);
2877 VM_BUG_ON(PageCgroupUsed(pc));
2879 * we don't need page_cgroup_lock about tail pages, becase they are not
2880 * accessed by any other context at this point.
2884 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2885 * may already be on some other mem_cgroup's LRU. Take care of it.
2888 zone = page_zone(page);
2889 spin_lock_irq(&zone->lru_lock);
2890 if (PageLRU(page)) {
2891 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2893 del_page_from_lru_list(page, lruvec, page_lru(page));
2898 pc->mem_cgroup = memcg;
2900 * We access a page_cgroup asynchronously without lock_page_cgroup().
2901 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2902 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2903 * before USED bit, we need memory barrier here.
2904 * See mem_cgroup_add_lru_list(), etc.
2907 SetPageCgroupUsed(pc);
2911 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2912 VM_BUG_ON(PageLRU(page));
2914 add_page_to_lru_list(page, lruvec, page_lru(page));
2916 spin_unlock_irq(&zone->lru_lock);
2919 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2924 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2925 unlock_page_cgroup(pc);
2928 * "charge_statistics" updated event counter. Then, check it.
2929 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2930 * if they exceeds softlimit.
2932 memcg_check_events(memcg, page);
2935 static DEFINE_MUTEX(set_limit_mutex);
2937 #ifdef CONFIG_MEMCG_KMEM
2938 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2940 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2941 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2945 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2946 * in the memcg_cache_params struct.
2948 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2950 struct kmem_cache *cachep;
2952 VM_BUG_ON(p->is_root_cache);
2953 cachep = p->root_cache;
2954 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2957 #ifdef CONFIG_SLABINFO
2958 static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2961 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2962 struct memcg_cache_params *params;
2964 if (!memcg_can_account_kmem(memcg))
2967 print_slabinfo_header(m);
2969 mutex_lock(&memcg->slab_caches_mutex);
2970 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2971 cache_show(memcg_params_to_cache(params), m);
2972 mutex_unlock(&memcg->slab_caches_mutex);
2978 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2980 struct res_counter *fail_res;
2981 struct mem_cgroup *_memcg;
2985 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2990 * Conditions under which we can wait for the oom_killer. Those are
2991 * the same conditions tested by the core page allocator
2993 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2996 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2999 if (ret == -EINTR) {
3001 * __mem_cgroup_try_charge() chosed to bypass to root due to
3002 * OOM kill or fatal signal. Since our only options are to
3003 * either fail the allocation or charge it to this cgroup, do
3004 * it as a temporary condition. But we can't fail. From a
3005 * kmem/slab perspective, the cache has already been selected,
3006 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3009 * This condition will only trigger if the task entered
3010 * memcg_charge_kmem in a sane state, but was OOM-killed during
3011 * __mem_cgroup_try_charge() above. Tasks that were already
3012 * dying when the allocation triggers should have been already
3013 * directed to the root cgroup in memcontrol.h
3015 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3016 if (do_swap_account)
3017 res_counter_charge_nofail(&memcg->memsw, size,
3021 res_counter_uncharge(&memcg->kmem, size);
3026 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3028 res_counter_uncharge(&memcg->res, size);
3029 if (do_swap_account)
3030 res_counter_uncharge(&memcg->memsw, size);
3033 if (res_counter_uncharge(&memcg->kmem, size))
3036 if (memcg_kmem_test_and_clear_dead(memcg))
3037 mem_cgroup_put(memcg);
3040 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3045 mutex_lock(&memcg->slab_caches_mutex);
3046 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3047 mutex_unlock(&memcg->slab_caches_mutex);
3051 * helper for acessing a memcg's index. It will be used as an index in the
3052 * child cache array in kmem_cache, and also to derive its name. This function
3053 * will return -1 when this is not a kmem-limited memcg.
3055 int memcg_cache_id(struct mem_cgroup *memcg)
3057 return memcg ? memcg->kmemcg_id : -1;
3061 * This ends up being protected by the set_limit mutex, during normal
3062 * operation, because that is its main call site.
3064 * But when we create a new cache, we can call this as well if its parent
3065 * is kmem-limited. That will have to hold set_limit_mutex as well.
3067 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3071 num = ida_simple_get(&kmem_limited_groups,
3072 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3076 * After this point, kmem_accounted (that we test atomically in
3077 * the beginning of this conditional), is no longer 0. This
3078 * guarantees only one process will set the following boolean
3079 * to true. We don't need test_and_set because we're protected
3080 * by the set_limit_mutex anyway.
3082 memcg_kmem_set_activated(memcg);
3084 ret = memcg_update_all_caches(num+1);
3086 ida_simple_remove(&kmem_limited_groups, num);
3087 memcg_kmem_clear_activated(memcg);
3091 memcg->kmemcg_id = num;
3092 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3093 mutex_init(&memcg->slab_caches_mutex);
3097 static size_t memcg_caches_array_size(int num_groups)
3100 if (num_groups <= 0)
3103 size = 2 * num_groups;
3104 if (size < MEMCG_CACHES_MIN_SIZE)
3105 size = MEMCG_CACHES_MIN_SIZE;
3106 else if (size > MEMCG_CACHES_MAX_SIZE)
3107 size = MEMCG_CACHES_MAX_SIZE;
3113 * We should update the current array size iff all caches updates succeed. This
3114 * can only be done from the slab side. The slab mutex needs to be held when
3117 void memcg_update_array_size(int num)
3119 if (num > memcg_limited_groups_array_size)
3120 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3123 static void kmem_cache_destroy_work_func(struct work_struct *w);
3125 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3127 struct memcg_cache_params *cur_params = s->memcg_params;
3129 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3131 if (num_groups > memcg_limited_groups_array_size) {
3133 ssize_t size = memcg_caches_array_size(num_groups);
3135 size *= sizeof(void *);
3136 size += sizeof(struct memcg_cache_params);
3138 s->memcg_params = kzalloc(size, GFP_KERNEL);
3139 if (!s->memcg_params) {
3140 s->memcg_params = cur_params;
3144 s->memcg_params->is_root_cache = true;
3147 * There is the chance it will be bigger than
3148 * memcg_limited_groups_array_size, if we failed an allocation
3149 * in a cache, in which case all caches updated before it, will
3150 * have a bigger array.
3152 * But if that is the case, the data after
3153 * memcg_limited_groups_array_size is certainly unused
3155 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3156 if (!cur_params->memcg_caches[i])
3158 s->memcg_params->memcg_caches[i] =
3159 cur_params->memcg_caches[i];
3163 * Ideally, we would wait until all caches succeed, and only
3164 * then free the old one. But this is not worth the extra
3165 * pointer per-cache we'd have to have for this.
3167 * It is not a big deal if some caches are left with a size
3168 * bigger than the others. And all updates will reset this
3176 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3177 struct kmem_cache *root_cache)
3179 size_t size = sizeof(struct memcg_cache_params);
3181 if (!memcg_kmem_enabled())
3185 size += memcg_limited_groups_array_size * sizeof(void *);
3187 s->memcg_params = kzalloc(size, GFP_KERNEL);
3188 if (!s->memcg_params)
3191 INIT_WORK(&s->memcg_params->destroy,
3192 kmem_cache_destroy_work_func);
3194 s->memcg_params->memcg = memcg;
3195 s->memcg_params->root_cache = root_cache;
3197 s->memcg_params->is_root_cache = true;
3202 void memcg_release_cache(struct kmem_cache *s)
3204 struct kmem_cache *root;
3205 struct mem_cgroup *memcg;
3209 * This happens, for instance, when a root cache goes away before we
3212 if (!s->memcg_params)
3215 if (s->memcg_params->is_root_cache)
3218 memcg = s->memcg_params->memcg;
3219 id = memcg_cache_id(memcg);
3221 root = s->memcg_params->root_cache;
3222 root->memcg_params->memcg_caches[id] = NULL;
3224 mutex_lock(&memcg->slab_caches_mutex);
3225 list_del(&s->memcg_params->list);
3226 mutex_unlock(&memcg->slab_caches_mutex);
3228 mem_cgroup_put(memcg);
3230 kfree(s->memcg_params);
3234 * During the creation a new cache, we need to disable our accounting mechanism
3235 * altogether. This is true even if we are not creating, but rather just
3236 * enqueing new caches to be created.
3238 * This is because that process will trigger allocations; some visible, like
3239 * explicit kmallocs to auxiliary data structures, name strings and internal
3240 * cache structures; some well concealed, like INIT_WORK() that can allocate
3241 * objects during debug.
3243 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3244 * to it. This may not be a bounded recursion: since the first cache creation
3245 * failed to complete (waiting on the allocation), we'll just try to create the
3246 * cache again, failing at the same point.
3248 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3249 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3250 * inside the following two functions.
3252 static inline void memcg_stop_kmem_account(void)
3254 VM_BUG_ON(!current->mm);
3255 current->memcg_kmem_skip_account++;
3258 static inline void memcg_resume_kmem_account(void)
3260 VM_BUG_ON(!current->mm);
3261 current->memcg_kmem_skip_account--;
3264 static void kmem_cache_destroy_work_func(struct work_struct *w)
3266 struct kmem_cache *cachep;
3267 struct memcg_cache_params *p;
3269 p = container_of(w, struct memcg_cache_params, destroy);
3271 cachep = memcg_params_to_cache(p);
3274 * If we get down to 0 after shrink, we could delete right away.
3275 * However, memcg_release_pages() already puts us back in the workqueue
3276 * in that case. If we proceed deleting, we'll get a dangling
3277 * reference, and removing the object from the workqueue in that case
3278 * is unnecessary complication. We are not a fast path.
3280 * Note that this case is fundamentally different from racing with
3281 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3282 * kmem_cache_shrink, not only we would be reinserting a dead cache
3283 * into the queue, but doing so from inside the worker racing to
3286 * So if we aren't down to zero, we'll just schedule a worker and try
3289 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3290 kmem_cache_shrink(cachep);
3291 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3294 kmem_cache_destroy(cachep);
3297 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3299 if (!cachep->memcg_params->dead)
3303 * There are many ways in which we can get here.
3305 * We can get to a memory-pressure situation while the delayed work is
3306 * still pending to run. The vmscan shrinkers can then release all
3307 * cache memory and get us to destruction. If this is the case, we'll
3308 * be executed twice, which is a bug (the second time will execute over
3309 * bogus data). In this case, cancelling the work should be fine.
3311 * But we can also get here from the worker itself, if
3312 * kmem_cache_shrink is enough to shake all the remaining objects and
3313 * get the page count to 0. In this case, we'll deadlock if we try to
3314 * cancel the work (the worker runs with an internal lock held, which
3315 * is the same lock we would hold for cancel_work_sync().)
3317 * Since we can't possibly know who got us here, just refrain from
3318 * running if there is already work pending
3320 if (work_pending(&cachep->memcg_params->destroy))
3323 * We have to defer the actual destroying to a workqueue, because
3324 * we might currently be in a context that cannot sleep.
3326 schedule_work(&cachep->memcg_params->destroy);
3330 * This lock protects updaters, not readers. We want readers to be as fast as
3331 * they can, and they will either see NULL or a valid cache value. Our model
3332 * allow them to see NULL, in which case the root memcg will be selected.
3334 * We need this lock because multiple allocations to the same cache from a non
3335 * will span more than one worker. Only one of them can create the cache.
3337 static DEFINE_MUTEX(memcg_cache_mutex);
3340 * Called with memcg_cache_mutex held
3342 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3343 struct kmem_cache *s)
3345 struct kmem_cache *new;
3346 static char *tmp_name = NULL;
3348 lockdep_assert_held(&memcg_cache_mutex);
3351 * kmem_cache_create_memcg duplicates the given name and
3352 * cgroup_name for this name requires RCU context.
3353 * This static temporary buffer is used to prevent from
3354 * pointless shortliving allocation.
3357 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3363 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3364 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3367 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3368 (s->flags & ~SLAB_PANIC), s->ctor, s);
3371 new->allocflags |= __GFP_KMEMCG;
3376 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3377 struct kmem_cache *cachep)
3379 struct kmem_cache *new_cachep;
3382 BUG_ON(!memcg_can_account_kmem(memcg));
3384 idx = memcg_cache_id(memcg);
3386 mutex_lock(&memcg_cache_mutex);
3387 new_cachep = cachep->memcg_params->memcg_caches[idx];
3391 new_cachep = kmem_cache_dup(memcg, cachep);
3392 if (new_cachep == NULL) {
3393 new_cachep = cachep;
3397 mem_cgroup_get(memcg);
3398 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3400 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3402 * the readers won't lock, make sure everybody sees the updated value,
3403 * so they won't put stuff in the queue again for no reason
3407 mutex_unlock(&memcg_cache_mutex);
3411 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3413 struct kmem_cache *c;
3416 if (!s->memcg_params)
3418 if (!s->memcg_params->is_root_cache)
3422 * If the cache is being destroyed, we trust that there is no one else
3423 * requesting objects from it. Even if there are, the sanity checks in
3424 * kmem_cache_destroy should caught this ill-case.
3426 * Still, we don't want anyone else freeing memcg_caches under our
3427 * noses, which can happen if a new memcg comes to life. As usual,
3428 * we'll take the set_limit_mutex to protect ourselves against this.
3430 mutex_lock(&set_limit_mutex);
3431 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3432 c = s->memcg_params->memcg_caches[i];
3437 * We will now manually delete the caches, so to avoid races
3438 * we need to cancel all pending destruction workers and
3439 * proceed with destruction ourselves.
3441 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3442 * and that could spawn the workers again: it is likely that
3443 * the cache still have active pages until this very moment.
3444 * This would lead us back to mem_cgroup_destroy_cache.
3446 * But that will not execute at all if the "dead" flag is not
3447 * set, so flip it down to guarantee we are in control.
3449 c->memcg_params->dead = false;
3450 cancel_work_sync(&c->memcg_params->destroy);
3451 kmem_cache_destroy(c);
3453 mutex_unlock(&set_limit_mutex);
3456 struct create_work {
3457 struct mem_cgroup *memcg;
3458 struct kmem_cache *cachep;
3459 struct work_struct work;
3462 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3464 struct kmem_cache *cachep;
3465 struct memcg_cache_params *params;
3467 if (!memcg_kmem_is_active(memcg))
3470 mutex_lock(&memcg->slab_caches_mutex);
3471 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3472 cachep = memcg_params_to_cache(params);
3473 cachep->memcg_params->dead = true;
3474 schedule_work(&cachep->memcg_params->destroy);
3476 mutex_unlock(&memcg->slab_caches_mutex);
3479 static void memcg_create_cache_work_func(struct work_struct *w)
3481 struct create_work *cw;
3483 cw = container_of(w, struct create_work, work);
3484 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3485 /* Drop the reference gotten when we enqueued. */
3486 css_put(&cw->memcg->css);
3491 * Enqueue the creation of a per-memcg kmem_cache.
3493 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3494 struct kmem_cache *cachep)
3496 struct create_work *cw;
3498 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3500 css_put(&memcg->css);
3505 cw->cachep = cachep;
3507 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3508 schedule_work(&cw->work);
3511 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3512 struct kmem_cache *cachep)
3515 * We need to stop accounting when we kmalloc, because if the
3516 * corresponding kmalloc cache is not yet created, the first allocation
3517 * in __memcg_create_cache_enqueue will recurse.
3519 * However, it is better to enclose the whole function. Depending on
3520 * the debugging options enabled, INIT_WORK(), for instance, can
3521 * trigger an allocation. This too, will make us recurse. Because at
3522 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3523 * the safest choice is to do it like this, wrapping the whole function.
3525 memcg_stop_kmem_account();
3526 __memcg_create_cache_enqueue(memcg, cachep);
3527 memcg_resume_kmem_account();
3530 * Return the kmem_cache we're supposed to use for a slab allocation.
3531 * We try to use the current memcg's version of the cache.
3533 * If the cache does not exist yet, if we are the first user of it,
3534 * we either create it immediately, if possible, or create it asynchronously
3536 * In the latter case, we will let the current allocation go through with
3537 * the original cache.
3539 * Can't be called in interrupt context or from kernel threads.
3540 * This function needs to be called with rcu_read_lock() held.
3542 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3545 struct mem_cgroup *memcg;
3548 VM_BUG_ON(!cachep->memcg_params);
3549 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3551 if (!current->mm || current->memcg_kmem_skip_account)
3555 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3557 if (!memcg_can_account_kmem(memcg))
3560 idx = memcg_cache_id(memcg);
3563 * barrier to mare sure we're always seeing the up to date value. The
3564 * code updating memcg_caches will issue a write barrier to match this.
3566 read_barrier_depends();
3567 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3568 cachep = cachep->memcg_params->memcg_caches[idx];
3572 /* The corresponding put will be done in the workqueue. */
3573 if (!css_tryget(&memcg->css))
3578 * If we are in a safe context (can wait, and not in interrupt
3579 * context), we could be be predictable and return right away.
3580 * This would guarantee that the allocation being performed
3581 * already belongs in the new cache.
3583 * However, there are some clashes that can arrive from locking.
3584 * For instance, because we acquire the slab_mutex while doing
3585 * kmem_cache_dup, this means no further allocation could happen
3586 * with the slab_mutex held.
3588 * Also, because cache creation issue get_online_cpus(), this
3589 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3590 * that ends up reversed during cpu hotplug. (cpuset allocates
3591 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3592 * better to defer everything.
3594 memcg_create_cache_enqueue(memcg, cachep);
3600 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3603 * We need to verify if the allocation against current->mm->owner's memcg is
3604 * possible for the given order. But the page is not allocated yet, so we'll
3605 * need a further commit step to do the final arrangements.
3607 * It is possible for the task to switch cgroups in this mean time, so at
3608 * commit time, we can't rely on task conversion any longer. We'll then use
3609 * the handle argument to return to the caller which cgroup we should commit
3610 * against. We could also return the memcg directly and avoid the pointer
3611 * passing, but a boolean return value gives better semantics considering
3612 * the compiled-out case as well.
3614 * Returning true means the allocation is possible.
3617 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3619 struct mem_cgroup *memcg;
3623 memcg = try_get_mem_cgroup_from_mm(current->mm);
3626 * very rare case described in mem_cgroup_from_task. Unfortunately there
3627 * isn't much we can do without complicating this too much, and it would
3628 * be gfp-dependent anyway. Just let it go
3630 if (unlikely(!memcg))
3633 if (!memcg_can_account_kmem(memcg)) {
3634 css_put(&memcg->css);
3638 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3642 css_put(&memcg->css);
3646 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3649 struct page_cgroup *pc;
3651 VM_BUG_ON(mem_cgroup_is_root(memcg));
3653 /* The page allocation failed. Revert */
3655 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3659 pc = lookup_page_cgroup(page);
3660 lock_page_cgroup(pc);
3661 pc->mem_cgroup = memcg;
3662 SetPageCgroupUsed(pc);
3663 unlock_page_cgroup(pc);
3666 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3668 struct mem_cgroup *memcg = NULL;
3669 struct page_cgroup *pc;
3672 pc = lookup_page_cgroup(page);
3674 * Fast unlocked return. Theoretically might have changed, have to
3675 * check again after locking.
3677 if (!PageCgroupUsed(pc))
3680 lock_page_cgroup(pc);
3681 if (PageCgroupUsed(pc)) {
3682 memcg = pc->mem_cgroup;
3683 ClearPageCgroupUsed(pc);
3685 unlock_page_cgroup(pc);
3688 * We trust that only if there is a memcg associated with the page, it
3689 * is a valid allocation
3694 VM_BUG_ON(mem_cgroup_is_root(memcg));
3695 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3698 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3701 #endif /* CONFIG_MEMCG_KMEM */
3703 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3705 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3707 * Because tail pages are not marked as "used", set it. We're under
3708 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3709 * charge/uncharge will be never happen and move_account() is done under
3710 * compound_lock(), so we don't have to take care of races.
3712 void mem_cgroup_split_huge_fixup(struct page *head)
3714 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3715 struct page_cgroup *pc;
3716 struct mem_cgroup *memcg;
3719 if (mem_cgroup_disabled())
3722 memcg = head_pc->mem_cgroup;
3723 for (i = 1; i < HPAGE_PMD_NR; i++) {
3725 pc->mem_cgroup = memcg;
3726 smp_wmb();/* see __commit_charge() */
3727 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3729 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3732 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3735 * mem_cgroup_move_account - move account of the page
3737 * @nr_pages: number of regular pages (>1 for huge pages)
3738 * @pc: page_cgroup of the page.
3739 * @from: mem_cgroup which the page is moved from.
3740 * @to: mem_cgroup which the page is moved to. @from != @to.
3742 * The caller must confirm following.
3743 * - page is not on LRU (isolate_page() is useful.)
3744 * - compound_lock is held when nr_pages > 1
3746 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3749 static int mem_cgroup_move_account(struct page *page,
3750 unsigned int nr_pages,
3751 struct page_cgroup *pc,
3752 struct mem_cgroup *from,
3753 struct mem_cgroup *to)
3755 unsigned long flags;
3757 bool anon = PageAnon(page);
3759 VM_BUG_ON(from == to);
3760 VM_BUG_ON(PageLRU(page));
3762 * The page is isolated from LRU. So, collapse function
3763 * will not handle this page. But page splitting can happen.
3764 * Do this check under compound_page_lock(). The caller should
3768 if (nr_pages > 1 && !PageTransHuge(page))
3771 lock_page_cgroup(pc);
3774 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3777 move_lock_mem_cgroup(from, &flags);
3779 if (!anon && page_mapped(page)) {
3780 /* Update mapped_file data for mem_cgroup */
3782 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3783 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3786 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3788 /* caller should have done css_get */
3789 pc->mem_cgroup = to;
3790 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3791 move_unlock_mem_cgroup(from, &flags);
3794 unlock_page_cgroup(pc);
3798 memcg_check_events(to, page);
3799 memcg_check_events(from, page);
3805 * mem_cgroup_move_parent - moves page to the parent group
3806 * @page: the page to move
3807 * @pc: page_cgroup of the page
3808 * @child: page's cgroup
3810 * move charges to its parent or the root cgroup if the group has no
3811 * parent (aka use_hierarchy==0).
3812 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3813 * mem_cgroup_move_account fails) the failure is always temporary and
3814 * it signals a race with a page removal/uncharge or migration. In the
3815 * first case the page is on the way out and it will vanish from the LRU
3816 * on the next attempt and the call should be retried later.
3817 * Isolation from the LRU fails only if page has been isolated from
3818 * the LRU since we looked at it and that usually means either global
3819 * reclaim or migration going on. The page will either get back to the
3821 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3822 * (!PageCgroupUsed) or moved to a different group. The page will
3823 * disappear in the next attempt.
3825 static int mem_cgroup_move_parent(struct page *page,
3826 struct page_cgroup *pc,
3827 struct mem_cgroup *child)
3829 struct mem_cgroup *parent;
3830 unsigned int nr_pages;
3831 unsigned long uninitialized_var(flags);
3834 VM_BUG_ON(mem_cgroup_is_root(child));
3837 if (!get_page_unless_zero(page))
3839 if (isolate_lru_page(page))
3842 nr_pages = hpage_nr_pages(page);
3844 parent = parent_mem_cgroup(child);
3846 * If no parent, move charges to root cgroup.
3849 parent = root_mem_cgroup;
3852 VM_BUG_ON(!PageTransHuge(page));
3853 flags = compound_lock_irqsave(page);
3856 ret = mem_cgroup_move_account(page, nr_pages,
3859 __mem_cgroup_cancel_local_charge(child, nr_pages);
3862 compound_unlock_irqrestore(page, flags);
3863 putback_lru_page(page);
3871 * Charge the memory controller for page usage.
3873 * 0 if the charge was successful
3874 * < 0 if the cgroup is over its limit
3876 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3877 gfp_t gfp_mask, enum charge_type ctype)
3879 struct mem_cgroup *memcg = NULL;
3880 unsigned int nr_pages = 1;
3884 if (PageTransHuge(page)) {
3885 nr_pages <<= compound_order(page);
3886 VM_BUG_ON(!PageTransHuge(page));
3888 * Never OOM-kill a process for a huge page. The
3889 * fault handler will fall back to regular pages.
3894 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3897 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3901 int mem_cgroup_newpage_charge(struct page *page,
3902 struct mm_struct *mm, gfp_t gfp_mask)
3904 if (mem_cgroup_disabled())
3906 VM_BUG_ON(page_mapped(page));
3907 VM_BUG_ON(page->mapping && !PageAnon(page));
3909 return mem_cgroup_charge_common(page, mm, gfp_mask,
3910 MEM_CGROUP_CHARGE_TYPE_ANON);
3914 * While swap-in, try_charge -> commit or cancel, the page is locked.
3915 * And when try_charge() successfully returns, one refcnt to memcg without
3916 * struct page_cgroup is acquired. This refcnt will be consumed by
3917 * "commit()" or removed by "cancel()"
3919 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3922 struct mem_cgroup **memcgp)
3924 struct mem_cgroup *memcg;
3925 struct page_cgroup *pc;
3928 pc = lookup_page_cgroup(page);
3930 * Every swap fault against a single page tries to charge the
3931 * page, bail as early as possible. shmem_unuse() encounters
3932 * already charged pages, too. The USED bit is protected by
3933 * the page lock, which serializes swap cache removal, which
3934 * in turn serializes uncharging.
3936 if (PageCgroupUsed(pc))
3938 if (!do_swap_account)
3940 memcg = try_get_mem_cgroup_from_page(page);
3944 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3945 css_put(&memcg->css);
3950 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3956 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3957 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3960 if (mem_cgroup_disabled())
3963 * A racing thread's fault, or swapoff, may have already
3964 * updated the pte, and even removed page from swap cache: in
3965 * those cases unuse_pte()'s pte_same() test will fail; but
3966 * there's also a KSM case which does need to charge the page.
3968 if (!PageSwapCache(page)) {
3971 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3976 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3979 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3981 if (mem_cgroup_disabled())
3985 __mem_cgroup_cancel_charge(memcg, 1);
3989 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3990 enum charge_type ctype)
3992 if (mem_cgroup_disabled())
3997 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3999 * Now swap is on-memory. This means this page may be
4000 * counted both as mem and swap....double count.
4001 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4002 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4003 * may call delete_from_swap_cache() before reach here.
4005 if (do_swap_account && PageSwapCache(page)) {
4006 swp_entry_t ent = {.val = page_private(page)};
4007 mem_cgroup_uncharge_swap(ent);
4011 void mem_cgroup_commit_charge_swapin(struct page *page,
4012 struct mem_cgroup *memcg)
4014 __mem_cgroup_commit_charge_swapin(page, memcg,
4015 MEM_CGROUP_CHARGE_TYPE_ANON);
4018 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4021 struct mem_cgroup *memcg = NULL;
4022 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4025 if (mem_cgroup_disabled())
4027 if (PageCompound(page))
4030 if (!PageSwapCache(page))
4031 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4032 else { /* page is swapcache/shmem */
4033 ret = __mem_cgroup_try_charge_swapin(mm, page,
4036 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4041 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4042 unsigned int nr_pages,
4043 const enum charge_type ctype)
4045 struct memcg_batch_info *batch = NULL;
4046 bool uncharge_memsw = true;
4048 /* If swapout, usage of swap doesn't decrease */
4049 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4050 uncharge_memsw = false;
4052 batch = ¤t->memcg_batch;
4054 * In usual, we do css_get() when we remember memcg pointer.
4055 * But in this case, we keep res->usage until end of a series of
4056 * uncharges. Then, it's ok to ignore memcg's refcnt.
4059 batch->memcg = memcg;
4061 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4062 * In those cases, all pages freed continuously can be expected to be in
4063 * the same cgroup and we have chance to coalesce uncharges.
4064 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4065 * because we want to do uncharge as soon as possible.
4068 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4069 goto direct_uncharge;
4072 goto direct_uncharge;
4075 * In typical case, batch->memcg == mem. This means we can
4076 * merge a series of uncharges to an uncharge of res_counter.
4077 * If not, we uncharge res_counter ony by one.
4079 if (batch->memcg != memcg)
4080 goto direct_uncharge;
4081 /* remember freed charge and uncharge it later */
4084 batch->memsw_nr_pages++;
4087 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4089 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4090 if (unlikely(batch->memcg != memcg))
4091 memcg_oom_recover(memcg);
4095 * uncharge if !page_mapped(page)
4097 static struct mem_cgroup *
4098 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4101 struct mem_cgroup *memcg = NULL;
4102 unsigned int nr_pages = 1;
4103 struct page_cgroup *pc;
4106 if (mem_cgroup_disabled())
4109 if (PageTransHuge(page)) {
4110 nr_pages <<= compound_order(page);
4111 VM_BUG_ON(!PageTransHuge(page));
4114 * Check if our page_cgroup is valid
4116 pc = lookup_page_cgroup(page);
4117 if (unlikely(!PageCgroupUsed(pc)))
4120 lock_page_cgroup(pc);
4122 memcg = pc->mem_cgroup;
4124 if (!PageCgroupUsed(pc))
4127 anon = PageAnon(page);
4130 case MEM_CGROUP_CHARGE_TYPE_ANON:
4132 * Generally PageAnon tells if it's the anon statistics to be
4133 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4134 * used before page reached the stage of being marked PageAnon.
4138 case MEM_CGROUP_CHARGE_TYPE_DROP:
4139 /* See mem_cgroup_prepare_migration() */
4140 if (page_mapped(page))
4143 * Pages under migration may not be uncharged. But
4144 * end_migration() /must/ be the one uncharging the
4145 * unused post-migration page and so it has to call
4146 * here with the migration bit still set. See the
4147 * res_counter handling below.
4149 if (!end_migration && PageCgroupMigration(pc))
4152 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4153 if (!PageAnon(page)) { /* Shared memory */
4154 if (page->mapping && !page_is_file_cache(page))
4156 } else if (page_mapped(page)) /* Anon */
4163 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4165 ClearPageCgroupUsed(pc);
4167 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4168 * freed from LRU. This is safe because uncharged page is expected not
4169 * to be reused (freed soon). Exception is SwapCache, it's handled by
4170 * special functions.
4173 unlock_page_cgroup(pc);
4175 * even after unlock, we have memcg->res.usage here and this memcg
4176 * will never be freed.
4178 memcg_check_events(memcg, page);
4179 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4180 mem_cgroup_swap_statistics(memcg, true);
4181 mem_cgroup_get(memcg);
4184 * Migration does not charge the res_counter for the
4185 * replacement page, so leave it alone when phasing out the
4186 * page that is unused after the migration.
4188 if (!end_migration && !mem_cgroup_is_root(memcg))
4189 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4194 unlock_page_cgroup(pc);
4198 void mem_cgroup_uncharge_page(struct page *page)
4201 if (page_mapped(page))
4203 VM_BUG_ON(page->mapping && !PageAnon(page));
4205 * If the page is in swap cache, uncharge should be deferred
4206 * to the swap path, which also properly accounts swap usage
4207 * and handles memcg lifetime.
4209 * Note that this check is not stable and reclaim may add the
4210 * page to swap cache at any time after this. However, if the
4211 * page is not in swap cache by the time page->mapcount hits
4212 * 0, there won't be any page table references to the swap
4213 * slot, and reclaim will free it and not actually write the
4216 if (PageSwapCache(page))
4218 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4221 void mem_cgroup_uncharge_cache_page(struct page *page)
4223 VM_BUG_ON(page_mapped(page));
4224 VM_BUG_ON(page->mapping);
4225 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4229 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4230 * In that cases, pages are freed continuously and we can expect pages
4231 * are in the same memcg. All these calls itself limits the number of
4232 * pages freed at once, then uncharge_start/end() is called properly.
4233 * This may be called prural(2) times in a context,
4236 void mem_cgroup_uncharge_start(void)
4238 current->memcg_batch.do_batch++;
4239 /* We can do nest. */
4240 if (current->memcg_batch.do_batch == 1) {
4241 current->memcg_batch.memcg = NULL;
4242 current->memcg_batch.nr_pages = 0;
4243 current->memcg_batch.memsw_nr_pages = 0;
4247 void mem_cgroup_uncharge_end(void)
4249 struct memcg_batch_info *batch = ¤t->memcg_batch;
4251 if (!batch->do_batch)
4255 if (batch->do_batch) /* If stacked, do nothing. */
4261 * This "batch->memcg" is valid without any css_get/put etc...
4262 * bacause we hide charges behind us.
4264 if (batch->nr_pages)
4265 res_counter_uncharge(&batch->memcg->res,
4266 batch->nr_pages * PAGE_SIZE);
4267 if (batch->memsw_nr_pages)
4268 res_counter_uncharge(&batch->memcg->memsw,
4269 batch->memsw_nr_pages * PAGE_SIZE);
4270 memcg_oom_recover(batch->memcg);
4271 /* forget this pointer (for sanity check) */
4272 batch->memcg = NULL;
4277 * called after __delete_from_swap_cache() and drop "page" account.
4278 * memcg information is recorded to swap_cgroup of "ent"
4281 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4283 struct mem_cgroup *memcg;
4284 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4286 if (!swapout) /* this was a swap cache but the swap is unused ! */
4287 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4289 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4292 * record memcg information, if swapout && memcg != NULL,
4293 * mem_cgroup_get() was called in uncharge().
4295 if (do_swap_account && swapout && memcg)
4296 swap_cgroup_record(ent, css_id(&memcg->css));
4300 #ifdef CONFIG_MEMCG_SWAP
4302 * called from swap_entry_free(). remove record in swap_cgroup and
4303 * uncharge "memsw" account.
4305 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4307 struct mem_cgroup *memcg;
4310 if (!do_swap_account)
4313 id = swap_cgroup_record(ent, 0);
4315 memcg = mem_cgroup_lookup(id);
4318 * We uncharge this because swap is freed.
4319 * This memcg can be obsolete one. We avoid calling css_tryget
4321 if (!mem_cgroup_is_root(memcg))
4322 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4323 mem_cgroup_swap_statistics(memcg, false);
4324 mem_cgroup_put(memcg);
4330 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4331 * @entry: swap entry to be moved
4332 * @from: mem_cgroup which the entry is moved from
4333 * @to: mem_cgroup which the entry is moved to
4335 * It succeeds only when the swap_cgroup's record for this entry is the same
4336 * as the mem_cgroup's id of @from.
4338 * Returns 0 on success, -EINVAL on failure.
4340 * The caller must have charged to @to, IOW, called res_counter_charge() about
4341 * both res and memsw, and called css_get().
4343 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4344 struct mem_cgroup *from, struct mem_cgroup *to)
4346 unsigned short old_id, new_id;
4348 old_id = css_id(&from->css);
4349 new_id = css_id(&to->css);
4351 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4352 mem_cgroup_swap_statistics(from, false);
4353 mem_cgroup_swap_statistics(to, true);
4355 * This function is only called from task migration context now.
4356 * It postpones res_counter and refcount handling till the end
4357 * of task migration(mem_cgroup_clear_mc()) for performance
4358 * improvement. But we cannot postpone mem_cgroup_get(to)
4359 * because if the process that has been moved to @to does
4360 * swap-in, the refcount of @to might be decreased to 0.
4368 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4369 struct mem_cgroup *from, struct mem_cgroup *to)
4376 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4379 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4380 struct mem_cgroup **memcgp)
4382 struct mem_cgroup *memcg = NULL;
4383 unsigned int nr_pages = 1;
4384 struct page_cgroup *pc;
4385 enum charge_type ctype;
4389 if (mem_cgroup_disabled())
4392 if (PageTransHuge(page))
4393 nr_pages <<= compound_order(page);
4395 pc = lookup_page_cgroup(page);
4396 lock_page_cgroup(pc);
4397 if (PageCgroupUsed(pc)) {
4398 memcg = pc->mem_cgroup;
4399 css_get(&memcg->css);
4401 * At migrating an anonymous page, its mapcount goes down
4402 * to 0 and uncharge() will be called. But, even if it's fully
4403 * unmapped, migration may fail and this page has to be
4404 * charged again. We set MIGRATION flag here and delay uncharge
4405 * until end_migration() is called
4407 * Corner Case Thinking
4409 * When the old page was mapped as Anon and it's unmap-and-freed
4410 * while migration was ongoing.
4411 * If unmap finds the old page, uncharge() of it will be delayed
4412 * until end_migration(). If unmap finds a new page, it's
4413 * uncharged when it make mapcount to be 1->0. If unmap code
4414 * finds swap_migration_entry, the new page will not be mapped
4415 * and end_migration() will find it(mapcount==0).
4418 * When the old page was mapped but migraion fails, the kernel
4419 * remaps it. A charge for it is kept by MIGRATION flag even
4420 * if mapcount goes down to 0. We can do remap successfully
4421 * without charging it again.
4424 * The "old" page is under lock_page() until the end of
4425 * migration, so, the old page itself will not be swapped-out.
4426 * If the new page is swapped out before end_migraton, our
4427 * hook to usual swap-out path will catch the event.
4430 SetPageCgroupMigration(pc);
4432 unlock_page_cgroup(pc);
4434 * If the page is not charged at this point,
4442 * We charge new page before it's used/mapped. So, even if unlock_page()
4443 * is called before end_migration, we can catch all events on this new
4444 * page. In the case new page is migrated but not remapped, new page's
4445 * mapcount will be finally 0 and we call uncharge in end_migration().
4448 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4450 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4452 * The page is committed to the memcg, but it's not actually
4453 * charged to the res_counter since we plan on replacing the
4454 * old one and only one page is going to be left afterwards.
4456 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4459 /* remove redundant charge if migration failed*/
4460 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4461 struct page *oldpage, struct page *newpage, bool migration_ok)
4463 struct page *used, *unused;
4464 struct page_cgroup *pc;
4470 if (!migration_ok) {
4477 anon = PageAnon(used);
4478 __mem_cgroup_uncharge_common(unused,
4479 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4480 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4482 css_put(&memcg->css);
4484 * We disallowed uncharge of pages under migration because mapcount
4485 * of the page goes down to zero, temporarly.
4486 * Clear the flag and check the page should be charged.
4488 pc = lookup_page_cgroup(oldpage);
4489 lock_page_cgroup(pc);
4490 ClearPageCgroupMigration(pc);
4491 unlock_page_cgroup(pc);
4494 * If a page is a file cache, radix-tree replacement is very atomic
4495 * and we can skip this check. When it was an Anon page, its mapcount
4496 * goes down to 0. But because we added MIGRATION flage, it's not
4497 * uncharged yet. There are several case but page->mapcount check
4498 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4499 * check. (see prepare_charge() also)
4502 mem_cgroup_uncharge_page(used);
4506 * At replace page cache, newpage is not under any memcg but it's on
4507 * LRU. So, this function doesn't touch res_counter but handles LRU
4508 * in correct way. Both pages are locked so we cannot race with uncharge.
4510 void mem_cgroup_replace_page_cache(struct page *oldpage,
4511 struct page *newpage)
4513 struct mem_cgroup *memcg = NULL;
4514 struct page_cgroup *pc;
4515 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4517 if (mem_cgroup_disabled())
4520 pc = lookup_page_cgroup(oldpage);
4521 /* fix accounting on old pages */
4522 lock_page_cgroup(pc);
4523 if (PageCgroupUsed(pc)) {
4524 memcg = pc->mem_cgroup;
4525 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4526 ClearPageCgroupUsed(pc);
4528 unlock_page_cgroup(pc);
4531 * When called from shmem_replace_page(), in some cases the
4532 * oldpage has already been charged, and in some cases not.
4537 * Even if newpage->mapping was NULL before starting replacement,
4538 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4539 * LRU while we overwrite pc->mem_cgroup.
4541 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4544 #ifdef CONFIG_DEBUG_VM
4545 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4547 struct page_cgroup *pc;
4549 pc = lookup_page_cgroup(page);
4551 * Can be NULL while feeding pages into the page allocator for
4552 * the first time, i.e. during boot or memory hotplug;
4553 * or when mem_cgroup_disabled().
4555 if (likely(pc) && PageCgroupUsed(pc))
4560 bool mem_cgroup_bad_page_check(struct page *page)
4562 if (mem_cgroup_disabled())
4565 return lookup_page_cgroup_used(page) != NULL;
4568 void mem_cgroup_print_bad_page(struct page *page)
4570 struct page_cgroup *pc;
4572 pc = lookup_page_cgroup_used(page);
4574 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4575 pc, pc->flags, pc->mem_cgroup);
4580 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4581 unsigned long long val)
4584 u64 memswlimit, memlimit;
4586 int children = mem_cgroup_count_children(memcg);
4587 u64 curusage, oldusage;
4591 * For keeping hierarchical_reclaim simple, how long we should retry
4592 * is depends on callers. We set our retry-count to be function
4593 * of # of children which we should visit in this loop.
4595 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4597 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4600 while (retry_count) {
4601 if (signal_pending(current)) {
4606 * Rather than hide all in some function, I do this in
4607 * open coded manner. You see what this really does.
4608 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4610 mutex_lock(&set_limit_mutex);
4611 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4612 if (memswlimit < val) {
4614 mutex_unlock(&set_limit_mutex);
4618 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4622 ret = res_counter_set_limit(&memcg->res, val);
4624 if (memswlimit == val)
4625 memcg->memsw_is_minimum = true;
4627 memcg->memsw_is_minimum = false;
4629 mutex_unlock(&set_limit_mutex);
4634 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4635 MEM_CGROUP_RECLAIM_SHRINK);
4636 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4637 /* Usage is reduced ? */
4638 if (curusage >= oldusage)
4641 oldusage = curusage;
4643 if (!ret && enlarge)
4644 memcg_oom_recover(memcg);
4649 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4650 unsigned long long val)
4653 u64 memlimit, memswlimit, oldusage, curusage;
4654 int children = mem_cgroup_count_children(memcg);
4658 /* see mem_cgroup_resize_res_limit */
4659 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4660 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4661 while (retry_count) {
4662 if (signal_pending(current)) {
4667 * Rather than hide all in some function, I do this in
4668 * open coded manner. You see what this really does.
4669 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4671 mutex_lock(&set_limit_mutex);
4672 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4673 if (memlimit > val) {
4675 mutex_unlock(&set_limit_mutex);
4678 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4679 if (memswlimit < val)
4681 ret = res_counter_set_limit(&memcg->memsw, val);
4683 if (memlimit == val)
4684 memcg->memsw_is_minimum = true;
4686 memcg->memsw_is_minimum = false;
4688 mutex_unlock(&set_limit_mutex);
4693 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4694 MEM_CGROUP_RECLAIM_NOSWAP |
4695 MEM_CGROUP_RECLAIM_SHRINK);
4696 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4697 /* Usage is reduced ? */
4698 if (curusage >= oldusage)
4701 oldusage = curusage;
4703 if (!ret && enlarge)
4704 memcg_oom_recover(memcg);
4708 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4710 unsigned long *total_scanned)
4712 unsigned long nr_reclaimed = 0;
4713 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4714 unsigned long reclaimed;
4716 struct mem_cgroup_tree_per_zone *mctz;
4717 unsigned long long excess;
4718 unsigned long nr_scanned;
4723 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4725 * This loop can run a while, specially if mem_cgroup's continuously
4726 * keep exceeding their soft limit and putting the system under
4733 mz = mem_cgroup_largest_soft_limit_node(mctz);
4738 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4739 gfp_mask, &nr_scanned);
4740 nr_reclaimed += reclaimed;
4741 *total_scanned += nr_scanned;
4742 spin_lock(&mctz->lock);
4745 * If we failed to reclaim anything from this memory cgroup
4746 * it is time to move on to the next cgroup
4752 * Loop until we find yet another one.
4754 * By the time we get the soft_limit lock
4755 * again, someone might have aded the
4756 * group back on the RB tree. Iterate to
4757 * make sure we get a different mem.
4758 * mem_cgroup_largest_soft_limit_node returns
4759 * NULL if no other cgroup is present on
4763 __mem_cgroup_largest_soft_limit_node(mctz);
4765 css_put(&next_mz->memcg->css);
4766 else /* next_mz == NULL or other memcg */
4770 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4771 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4773 * One school of thought says that we should not add
4774 * back the node to the tree if reclaim returns 0.
4775 * But our reclaim could return 0, simply because due
4776 * to priority we are exposing a smaller subset of
4777 * memory to reclaim from. Consider this as a longer
4780 /* If excess == 0, no tree ops */
4781 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4782 spin_unlock(&mctz->lock);
4783 css_put(&mz->memcg->css);
4786 * Could not reclaim anything and there are no more
4787 * mem cgroups to try or we seem to be looping without
4788 * reclaiming anything.
4790 if (!nr_reclaimed &&
4792 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4794 } while (!nr_reclaimed);
4796 css_put(&next_mz->memcg->css);
4797 return nr_reclaimed;
4801 * mem_cgroup_force_empty_list - clears LRU of a group
4802 * @memcg: group to clear
4805 * @lru: lru to to clear
4807 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4808 * reclaim the pages page themselves - pages are moved to the parent (or root)
4811 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4812 int node, int zid, enum lru_list lru)
4814 struct lruvec *lruvec;
4815 unsigned long flags;
4816 struct list_head *list;
4820 zone = &NODE_DATA(node)->node_zones[zid];
4821 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4822 list = &lruvec->lists[lru];
4826 struct page_cgroup *pc;
4829 spin_lock_irqsave(&zone->lru_lock, flags);
4830 if (list_empty(list)) {
4831 spin_unlock_irqrestore(&zone->lru_lock, flags);
4834 page = list_entry(list->prev, struct page, lru);
4836 list_move(&page->lru, list);
4838 spin_unlock_irqrestore(&zone->lru_lock, flags);
4841 spin_unlock_irqrestore(&zone->lru_lock, flags);
4843 pc = lookup_page_cgroup(page);
4845 if (mem_cgroup_move_parent(page, pc, memcg)) {
4846 /* found lock contention or "pc" is obsolete. */
4851 } while (!list_empty(list));
4855 * make mem_cgroup's charge to be 0 if there is no task by moving
4856 * all the charges and pages to the parent.
4857 * This enables deleting this mem_cgroup.
4859 * Caller is responsible for holding css reference on the memcg.
4861 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4867 /* This is for making all *used* pages to be on LRU. */
4868 lru_add_drain_all();
4869 drain_all_stock_sync(memcg);
4870 mem_cgroup_start_move(memcg);
4871 for_each_node_state(node, N_MEMORY) {
4872 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4875 mem_cgroup_force_empty_list(memcg,
4880 mem_cgroup_end_move(memcg);
4881 memcg_oom_recover(memcg);
4885 * Kernel memory may not necessarily be trackable to a specific
4886 * process. So they are not migrated, and therefore we can't
4887 * expect their value to drop to 0 here.
4888 * Having res filled up with kmem only is enough.
4890 * This is a safety check because mem_cgroup_force_empty_list
4891 * could have raced with mem_cgroup_replace_page_cache callers
4892 * so the lru seemed empty but the page could have been added
4893 * right after the check. RES_USAGE should be safe as we always
4894 * charge before adding to the LRU.
4896 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4897 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4898 } while (usage > 0);
4902 * This mainly exists for tests during the setting of set of use_hierarchy.
4903 * Since this is the very setting we are changing, the current hierarchy value
4906 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4910 /* bounce at first found */
4911 cgroup_for_each_child(pos, memcg->css.cgroup)
4917 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4918 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4919 * from mem_cgroup_count_children(), in the sense that we don't really care how
4920 * many children we have; we only need to know if we have any. It also counts
4921 * any memcg without hierarchy as infertile.
4923 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4925 return memcg->use_hierarchy && __memcg_has_children(memcg);
4929 * Reclaims as many pages from the given memcg as possible and moves
4930 * the rest to the parent.
4932 * Caller is responsible for holding css reference for memcg.
4934 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4936 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4937 struct cgroup *cgrp = memcg->css.cgroup;
4939 /* returns EBUSY if there is a task or if we come here twice. */
4940 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4943 /* we call try-to-free pages for make this cgroup empty */
4944 lru_add_drain_all();
4945 /* try to free all pages in this cgroup */
4946 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4949 if (signal_pending(current))
4952 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4956 /* maybe some writeback is necessary */
4957 congestion_wait(BLK_RW_ASYNC, HZ/10);
4962 mem_cgroup_reparent_charges(memcg);
4967 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4969 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4972 if (mem_cgroup_is_root(memcg))
4974 css_get(&memcg->css);
4975 ret = mem_cgroup_force_empty(memcg);
4976 css_put(&memcg->css);
4982 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4984 return mem_cgroup_from_cont(cont)->use_hierarchy;
4987 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4991 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4992 struct cgroup *parent = cont->parent;
4993 struct mem_cgroup *parent_memcg = NULL;
4996 parent_memcg = mem_cgroup_from_cont(parent);
4998 mutex_lock(&memcg_create_mutex);
5000 if (memcg->use_hierarchy == val)
5004 * If parent's use_hierarchy is set, we can't make any modifications
5005 * in the child subtrees. If it is unset, then the change can
5006 * occur, provided the current cgroup has no children.
5008 * For the root cgroup, parent_mem is NULL, we allow value to be
5009 * set if there are no children.
5011 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5012 (val == 1 || val == 0)) {
5013 if (!__memcg_has_children(memcg))
5014 memcg->use_hierarchy = val;
5021 mutex_unlock(&memcg_create_mutex);
5027 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5028 enum mem_cgroup_stat_index idx)
5030 struct mem_cgroup *iter;
5033 /* Per-cpu values can be negative, use a signed accumulator */
5034 for_each_mem_cgroup_tree(iter, memcg)
5035 val += mem_cgroup_read_stat(iter, idx);
5037 if (val < 0) /* race ? */
5042 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5046 if (!mem_cgroup_is_root(memcg)) {
5048 return res_counter_read_u64(&memcg->res, RES_USAGE);
5050 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5054 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5055 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5057 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5058 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5061 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5063 return val << PAGE_SHIFT;
5066 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
5067 struct file *file, char __user *buf,
5068 size_t nbytes, loff_t *ppos)
5070 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5076 type = MEMFILE_TYPE(cft->private);
5077 name = MEMFILE_ATTR(cft->private);
5081 if (name == RES_USAGE)
5082 val = mem_cgroup_usage(memcg, false);
5084 val = res_counter_read_u64(&memcg->res, name);
5087 if (name == RES_USAGE)
5088 val = mem_cgroup_usage(memcg, true);
5090 val = res_counter_read_u64(&memcg->memsw, name);
5093 val = res_counter_read_u64(&memcg->kmem, name);
5099 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5100 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5103 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
5106 #ifdef CONFIG_MEMCG_KMEM
5107 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5109 * For simplicity, we won't allow this to be disabled. It also can't
5110 * be changed if the cgroup has children already, or if tasks had
5113 * If tasks join before we set the limit, a person looking at
5114 * kmem.usage_in_bytes will have no way to determine when it took
5115 * place, which makes the value quite meaningless.
5117 * After it first became limited, changes in the value of the limit are
5118 * of course permitted.
5120 mutex_lock(&memcg_create_mutex);
5121 mutex_lock(&set_limit_mutex);
5122 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
5123 if (cgroup_task_count(cont) || memcg_has_children(memcg)) {
5127 ret = res_counter_set_limit(&memcg->kmem, val);
5130 ret = memcg_update_cache_sizes(memcg);
5132 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5135 static_key_slow_inc(&memcg_kmem_enabled_key);
5137 * setting the active bit after the inc will guarantee no one
5138 * starts accounting before all call sites are patched
5140 memcg_kmem_set_active(memcg);
5143 * kmem charges can outlive the cgroup. In the case of slab
5144 * pages, for instance, a page contain objects from various
5145 * processes, so it is unfeasible to migrate them away. We
5146 * need to reference count the memcg because of that.
5148 mem_cgroup_get(memcg);
5150 ret = res_counter_set_limit(&memcg->kmem, val);
5152 mutex_unlock(&set_limit_mutex);
5153 mutex_unlock(&memcg_create_mutex);
5158 #ifdef CONFIG_MEMCG_KMEM
5159 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5162 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5166 memcg->kmem_account_flags = parent->kmem_account_flags;
5168 * When that happen, we need to disable the static branch only on those
5169 * memcgs that enabled it. To achieve this, we would be forced to
5170 * complicate the code by keeping track of which memcgs were the ones
5171 * that actually enabled limits, and which ones got it from its
5174 * It is a lot simpler just to do static_key_slow_inc() on every child
5175 * that is accounted.
5177 if (!memcg_kmem_is_active(memcg))
5181 * destroy(), called if we fail, will issue static_key_slow_inc() and
5182 * mem_cgroup_put() if kmem is enabled. We have to either call them
5183 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5184 * this more consistent, since it always leads to the same destroy path
5186 mem_cgroup_get(memcg);
5187 static_key_slow_inc(&memcg_kmem_enabled_key);
5189 mutex_lock(&set_limit_mutex);
5190 ret = memcg_update_cache_sizes(memcg);
5191 mutex_unlock(&set_limit_mutex);
5195 #endif /* CONFIG_MEMCG_KMEM */
5198 * The user of this function is...
5201 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5204 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5207 unsigned long long val;
5210 type = MEMFILE_TYPE(cft->private);
5211 name = MEMFILE_ATTR(cft->private);
5215 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5219 /* This function does all necessary parse...reuse it */
5220 ret = res_counter_memparse_write_strategy(buffer, &val);
5224 ret = mem_cgroup_resize_limit(memcg, val);
5225 else if (type == _MEMSWAP)
5226 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5227 else if (type == _KMEM)
5228 ret = memcg_update_kmem_limit(cont, val);
5232 case RES_SOFT_LIMIT:
5233 ret = res_counter_memparse_write_strategy(buffer, &val);
5237 * For memsw, soft limits are hard to implement in terms
5238 * of semantics, for now, we support soft limits for
5239 * control without swap
5242 ret = res_counter_set_soft_limit(&memcg->res, val);
5247 ret = -EINVAL; /* should be BUG() ? */
5253 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5254 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5256 struct cgroup *cgroup;
5257 unsigned long long min_limit, min_memsw_limit, tmp;
5259 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5260 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5261 cgroup = memcg->css.cgroup;
5262 if (!memcg->use_hierarchy)
5265 while (cgroup->parent) {
5266 cgroup = cgroup->parent;
5267 memcg = mem_cgroup_from_cont(cgroup);
5268 if (!memcg->use_hierarchy)
5270 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5271 min_limit = min(min_limit, tmp);
5272 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5273 min_memsw_limit = min(min_memsw_limit, tmp);
5276 *mem_limit = min_limit;
5277 *memsw_limit = min_memsw_limit;
5280 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5282 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5286 type = MEMFILE_TYPE(event);
5287 name = MEMFILE_ATTR(event);
5292 res_counter_reset_max(&memcg->res);
5293 else if (type == _MEMSWAP)
5294 res_counter_reset_max(&memcg->memsw);
5295 else if (type == _KMEM)
5296 res_counter_reset_max(&memcg->kmem);
5302 res_counter_reset_failcnt(&memcg->res);
5303 else if (type == _MEMSWAP)
5304 res_counter_reset_failcnt(&memcg->memsw);
5305 else if (type == _KMEM)
5306 res_counter_reset_failcnt(&memcg->kmem);
5315 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5318 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5322 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5323 struct cftype *cft, u64 val)
5325 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5327 if (val >= (1 << NR_MOVE_TYPE))
5331 * No kind of locking is needed in here, because ->can_attach() will
5332 * check this value once in the beginning of the process, and then carry
5333 * on with stale data. This means that changes to this value will only
5334 * affect task migrations starting after the change.
5336 memcg->move_charge_at_immigrate = val;
5340 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5341 struct cftype *cft, u64 val)
5348 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5352 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5353 unsigned long node_nr;
5354 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5356 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5357 seq_printf(m, "total=%lu", total_nr);
5358 for_each_node_state(nid, N_MEMORY) {
5359 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5360 seq_printf(m, " N%d=%lu", nid, node_nr);
5364 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5365 seq_printf(m, "file=%lu", file_nr);
5366 for_each_node_state(nid, N_MEMORY) {
5367 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5369 seq_printf(m, " N%d=%lu", nid, node_nr);
5373 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5374 seq_printf(m, "anon=%lu", anon_nr);
5375 for_each_node_state(nid, N_MEMORY) {
5376 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5378 seq_printf(m, " N%d=%lu", nid, node_nr);
5382 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5383 seq_printf(m, "unevictable=%lu", unevictable_nr);
5384 for_each_node_state(nid, N_MEMORY) {
5385 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5386 BIT(LRU_UNEVICTABLE));
5387 seq_printf(m, " N%d=%lu", nid, node_nr);
5392 #endif /* CONFIG_NUMA */
5394 static inline void mem_cgroup_lru_names_not_uptodate(void)
5396 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5399 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5402 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5403 struct mem_cgroup *mi;
5406 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5407 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5409 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5410 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5413 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5414 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5415 mem_cgroup_read_events(memcg, i));
5417 for (i = 0; i < NR_LRU_LISTS; i++)
5418 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5419 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5421 /* Hierarchical information */
5423 unsigned long long limit, memsw_limit;
5424 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5425 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5426 if (do_swap_account)
5427 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5431 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5434 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5436 for_each_mem_cgroup_tree(mi, memcg)
5437 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5438 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5441 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5442 unsigned long long val = 0;
5444 for_each_mem_cgroup_tree(mi, memcg)
5445 val += mem_cgroup_read_events(mi, i);
5446 seq_printf(m, "total_%s %llu\n",
5447 mem_cgroup_events_names[i], val);
5450 for (i = 0; i < NR_LRU_LISTS; i++) {
5451 unsigned long long val = 0;
5453 for_each_mem_cgroup_tree(mi, memcg)
5454 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5455 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5458 #ifdef CONFIG_DEBUG_VM
5461 struct mem_cgroup_per_zone *mz;
5462 struct zone_reclaim_stat *rstat;
5463 unsigned long recent_rotated[2] = {0, 0};
5464 unsigned long recent_scanned[2] = {0, 0};
5466 for_each_online_node(nid)
5467 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5468 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5469 rstat = &mz->lruvec.reclaim_stat;
5471 recent_rotated[0] += rstat->recent_rotated[0];
5472 recent_rotated[1] += rstat->recent_rotated[1];
5473 recent_scanned[0] += rstat->recent_scanned[0];
5474 recent_scanned[1] += rstat->recent_scanned[1];
5476 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5477 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5478 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5479 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5486 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5488 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5490 return mem_cgroup_swappiness(memcg);
5493 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5496 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5497 struct mem_cgroup *parent;
5502 if (cgrp->parent == NULL)
5505 parent = mem_cgroup_from_cont(cgrp->parent);
5507 mutex_lock(&memcg_create_mutex);
5509 /* If under hierarchy, only empty-root can set this value */
5510 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5511 mutex_unlock(&memcg_create_mutex);
5515 memcg->swappiness = val;
5517 mutex_unlock(&memcg_create_mutex);
5522 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5524 struct mem_cgroup_threshold_ary *t;
5530 t = rcu_dereference(memcg->thresholds.primary);
5532 t = rcu_dereference(memcg->memsw_thresholds.primary);
5537 usage = mem_cgroup_usage(memcg, swap);
5540 * current_threshold points to threshold just below or equal to usage.
5541 * If it's not true, a threshold was crossed after last
5542 * call of __mem_cgroup_threshold().
5544 i = t->current_threshold;
5547 * Iterate backward over array of thresholds starting from
5548 * current_threshold and check if a threshold is crossed.
5549 * If none of thresholds below usage is crossed, we read
5550 * only one element of the array here.
5552 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5553 eventfd_signal(t->entries[i].eventfd, 1);
5555 /* i = current_threshold + 1 */
5559 * Iterate forward over array of thresholds starting from
5560 * current_threshold+1 and check if a threshold is crossed.
5561 * If none of thresholds above usage is crossed, we read
5562 * only one element of the array here.
5564 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5565 eventfd_signal(t->entries[i].eventfd, 1);
5567 /* Update current_threshold */
5568 t->current_threshold = i - 1;
5573 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5576 __mem_cgroup_threshold(memcg, false);
5577 if (do_swap_account)
5578 __mem_cgroup_threshold(memcg, true);
5580 memcg = parent_mem_cgroup(memcg);
5584 static int compare_thresholds(const void *a, const void *b)
5586 const struct mem_cgroup_threshold *_a = a;
5587 const struct mem_cgroup_threshold *_b = b;
5589 return _a->threshold - _b->threshold;
5592 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5594 struct mem_cgroup_eventfd_list *ev;
5596 list_for_each_entry(ev, &memcg->oom_notify, list)
5597 eventfd_signal(ev->eventfd, 1);
5601 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5603 struct mem_cgroup *iter;
5605 for_each_mem_cgroup_tree(iter, memcg)
5606 mem_cgroup_oom_notify_cb(iter);
5609 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5610 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5612 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5613 struct mem_cgroup_thresholds *thresholds;
5614 struct mem_cgroup_threshold_ary *new;
5615 enum res_type type = MEMFILE_TYPE(cft->private);
5616 u64 threshold, usage;
5619 ret = res_counter_memparse_write_strategy(args, &threshold);
5623 mutex_lock(&memcg->thresholds_lock);
5626 thresholds = &memcg->thresholds;
5627 else if (type == _MEMSWAP)
5628 thresholds = &memcg->memsw_thresholds;
5632 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5634 /* Check if a threshold crossed before adding a new one */
5635 if (thresholds->primary)
5636 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5638 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5640 /* Allocate memory for new array of thresholds */
5641 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5649 /* Copy thresholds (if any) to new array */
5650 if (thresholds->primary) {
5651 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5652 sizeof(struct mem_cgroup_threshold));
5655 /* Add new threshold */
5656 new->entries[size - 1].eventfd = eventfd;
5657 new->entries[size - 1].threshold = threshold;
5659 /* Sort thresholds. Registering of new threshold isn't time-critical */
5660 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5661 compare_thresholds, NULL);
5663 /* Find current threshold */
5664 new->current_threshold = -1;
5665 for (i = 0; i < size; i++) {
5666 if (new->entries[i].threshold <= usage) {
5668 * new->current_threshold will not be used until
5669 * rcu_assign_pointer(), so it's safe to increment
5672 ++new->current_threshold;
5677 /* Free old spare buffer and save old primary buffer as spare */
5678 kfree(thresholds->spare);
5679 thresholds->spare = thresholds->primary;
5681 rcu_assign_pointer(thresholds->primary, new);
5683 /* To be sure that nobody uses thresholds */
5687 mutex_unlock(&memcg->thresholds_lock);
5692 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5693 struct cftype *cft, struct eventfd_ctx *eventfd)
5695 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5696 struct mem_cgroup_thresholds *thresholds;
5697 struct mem_cgroup_threshold_ary *new;
5698 enum res_type type = MEMFILE_TYPE(cft->private);
5702 mutex_lock(&memcg->thresholds_lock);
5704 thresholds = &memcg->thresholds;
5705 else if (type == _MEMSWAP)
5706 thresholds = &memcg->memsw_thresholds;
5710 if (!thresholds->primary)
5713 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5715 /* Check if a threshold crossed before removing */
5716 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5718 /* Calculate new number of threshold */
5720 for (i = 0; i < thresholds->primary->size; i++) {
5721 if (thresholds->primary->entries[i].eventfd != eventfd)
5725 new = thresholds->spare;
5727 /* Set thresholds array to NULL if we don't have thresholds */
5736 /* Copy thresholds and find current threshold */
5737 new->current_threshold = -1;
5738 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5739 if (thresholds->primary->entries[i].eventfd == eventfd)
5742 new->entries[j] = thresholds->primary->entries[i];
5743 if (new->entries[j].threshold <= usage) {
5745 * new->current_threshold will not be used
5746 * until rcu_assign_pointer(), so it's safe to increment
5749 ++new->current_threshold;
5755 /* Swap primary and spare array */
5756 thresholds->spare = thresholds->primary;
5757 /* If all events are unregistered, free the spare array */
5759 kfree(thresholds->spare);
5760 thresholds->spare = NULL;
5763 rcu_assign_pointer(thresholds->primary, new);
5765 /* To be sure that nobody uses thresholds */
5768 mutex_unlock(&memcg->thresholds_lock);
5771 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5772 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5774 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5775 struct mem_cgroup_eventfd_list *event;
5776 enum res_type type = MEMFILE_TYPE(cft->private);
5778 BUG_ON(type != _OOM_TYPE);
5779 event = kmalloc(sizeof(*event), GFP_KERNEL);
5783 spin_lock(&memcg_oom_lock);
5785 event->eventfd = eventfd;
5786 list_add(&event->list, &memcg->oom_notify);
5788 /* already in OOM ? */
5789 if (atomic_read(&memcg->under_oom))
5790 eventfd_signal(eventfd, 1);
5791 spin_unlock(&memcg_oom_lock);
5796 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5797 struct cftype *cft, struct eventfd_ctx *eventfd)
5799 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5800 struct mem_cgroup_eventfd_list *ev, *tmp;
5801 enum res_type type = MEMFILE_TYPE(cft->private);
5803 BUG_ON(type != _OOM_TYPE);
5805 spin_lock(&memcg_oom_lock);
5807 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5808 if (ev->eventfd == eventfd) {
5809 list_del(&ev->list);
5814 spin_unlock(&memcg_oom_lock);
5817 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5818 struct cftype *cft, struct cgroup_map_cb *cb)
5820 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5822 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5824 if (atomic_read(&memcg->under_oom))
5825 cb->fill(cb, "under_oom", 1);
5827 cb->fill(cb, "under_oom", 0);
5831 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5832 struct cftype *cft, u64 val)
5834 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5835 struct mem_cgroup *parent;
5837 /* cannot set to root cgroup and only 0 and 1 are allowed */
5838 if (!cgrp->parent || !((val == 0) || (val == 1)))
5841 parent = mem_cgroup_from_cont(cgrp->parent);
5843 mutex_lock(&memcg_create_mutex);
5844 /* oom-kill-disable is a flag for subhierarchy. */
5845 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5846 mutex_unlock(&memcg_create_mutex);
5849 memcg->oom_kill_disable = val;
5851 memcg_oom_recover(memcg);
5852 mutex_unlock(&memcg_create_mutex);
5856 #ifdef CONFIG_MEMCG_KMEM
5857 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5861 memcg->kmemcg_id = -1;
5862 ret = memcg_propagate_kmem(memcg);
5866 return mem_cgroup_sockets_init(memcg, ss);
5869 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5871 mem_cgroup_sockets_destroy(memcg);
5873 memcg_kmem_mark_dead(memcg);
5875 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5879 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5880 * path here, being careful not to race with memcg_uncharge_kmem: it is
5881 * possible that the charges went down to 0 between mark_dead and the
5882 * res_counter read, so in that case, we don't need the put
5884 if (memcg_kmem_test_and_clear_dead(memcg))
5885 mem_cgroup_put(memcg);
5888 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5893 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5898 static struct cftype mem_cgroup_files[] = {
5900 .name = "usage_in_bytes",
5901 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5902 .read = mem_cgroup_read,
5903 .register_event = mem_cgroup_usage_register_event,
5904 .unregister_event = mem_cgroup_usage_unregister_event,
5907 .name = "max_usage_in_bytes",
5908 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5909 .trigger = mem_cgroup_reset,
5910 .read = mem_cgroup_read,
5913 .name = "limit_in_bytes",
5914 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5915 .write_string = mem_cgroup_write,
5916 .read = mem_cgroup_read,
5919 .name = "soft_limit_in_bytes",
5920 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5921 .write_string = mem_cgroup_write,
5922 .read = mem_cgroup_read,
5926 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5927 .trigger = mem_cgroup_reset,
5928 .read = mem_cgroup_read,
5932 .read_seq_string = memcg_stat_show,
5935 .name = "force_empty",
5936 .trigger = mem_cgroup_force_empty_write,
5939 .name = "use_hierarchy",
5940 .flags = CFTYPE_INSANE,
5941 .write_u64 = mem_cgroup_hierarchy_write,
5942 .read_u64 = mem_cgroup_hierarchy_read,
5945 .name = "swappiness",
5946 .read_u64 = mem_cgroup_swappiness_read,
5947 .write_u64 = mem_cgroup_swappiness_write,
5950 .name = "move_charge_at_immigrate",
5951 .read_u64 = mem_cgroup_move_charge_read,
5952 .write_u64 = mem_cgroup_move_charge_write,
5955 .name = "oom_control",
5956 .read_map = mem_cgroup_oom_control_read,
5957 .write_u64 = mem_cgroup_oom_control_write,
5958 .register_event = mem_cgroup_oom_register_event,
5959 .unregister_event = mem_cgroup_oom_unregister_event,
5960 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5963 .name = "pressure_level",
5964 .register_event = vmpressure_register_event,
5965 .unregister_event = vmpressure_unregister_event,
5969 .name = "numa_stat",
5970 .read_seq_string = memcg_numa_stat_show,
5973 #ifdef CONFIG_MEMCG_KMEM
5975 .name = "kmem.limit_in_bytes",
5976 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5977 .write_string = mem_cgroup_write,
5978 .read = mem_cgroup_read,
5981 .name = "kmem.usage_in_bytes",
5982 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5983 .read = mem_cgroup_read,
5986 .name = "kmem.failcnt",
5987 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5988 .trigger = mem_cgroup_reset,
5989 .read = mem_cgroup_read,
5992 .name = "kmem.max_usage_in_bytes",
5993 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5994 .trigger = mem_cgroup_reset,
5995 .read = mem_cgroup_read,
5997 #ifdef CONFIG_SLABINFO
5999 .name = "kmem.slabinfo",
6000 .read_seq_string = mem_cgroup_slabinfo_read,
6004 { }, /* terminate */
6007 #ifdef CONFIG_MEMCG_SWAP
6008 static struct cftype memsw_cgroup_files[] = {
6010 .name = "memsw.usage_in_bytes",
6011 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6012 .read = mem_cgroup_read,
6013 .register_event = mem_cgroup_usage_register_event,
6014 .unregister_event = mem_cgroup_usage_unregister_event,
6017 .name = "memsw.max_usage_in_bytes",
6018 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6019 .trigger = mem_cgroup_reset,
6020 .read = mem_cgroup_read,
6023 .name = "memsw.limit_in_bytes",
6024 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6025 .write_string = mem_cgroup_write,
6026 .read = mem_cgroup_read,
6029 .name = "memsw.failcnt",
6030 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6031 .trigger = mem_cgroup_reset,
6032 .read = mem_cgroup_read,
6034 { }, /* terminate */
6037 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6039 struct mem_cgroup_per_node *pn;
6040 struct mem_cgroup_per_zone *mz;
6041 int zone, tmp = node;
6043 * This routine is called against possible nodes.
6044 * But it's BUG to call kmalloc() against offline node.
6046 * TODO: this routine can waste much memory for nodes which will
6047 * never be onlined. It's better to use memory hotplug callback
6050 if (!node_state(node, N_NORMAL_MEMORY))
6052 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6056 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6057 mz = &pn->zoneinfo[zone];
6058 lruvec_init(&mz->lruvec);
6059 mz->usage_in_excess = 0;
6060 mz->on_tree = false;
6063 memcg->info.nodeinfo[node] = pn;
6067 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6069 kfree(memcg->info.nodeinfo[node]);
6072 static struct mem_cgroup *mem_cgroup_alloc(void)
6074 struct mem_cgroup *memcg;
6075 size_t size = memcg_size();
6077 /* Can be very big if nr_node_ids is very big */
6078 if (size < PAGE_SIZE)
6079 memcg = kzalloc(size, GFP_KERNEL);
6081 memcg = vzalloc(size);
6086 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6089 spin_lock_init(&memcg->pcp_counter_lock);
6093 if (size < PAGE_SIZE)
6101 * At destroying mem_cgroup, references from swap_cgroup can remain.
6102 * (scanning all at force_empty is too costly...)
6104 * Instead of clearing all references at force_empty, we remember
6105 * the number of reference from swap_cgroup and free mem_cgroup when
6106 * it goes down to 0.
6108 * Removal of cgroup itself succeeds regardless of refs from swap.
6111 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6114 size_t size = memcg_size();
6116 mem_cgroup_remove_from_trees(memcg);
6117 free_css_id(&mem_cgroup_subsys, &memcg->css);
6120 free_mem_cgroup_per_zone_info(memcg, node);
6122 free_percpu(memcg->stat);
6125 * We need to make sure that (at least for now), the jump label
6126 * destruction code runs outside of the cgroup lock. This is because
6127 * get_online_cpus(), which is called from the static_branch update,
6128 * can't be called inside the cgroup_lock. cpusets are the ones
6129 * enforcing this dependency, so if they ever change, we might as well.
6131 * schedule_work() will guarantee this happens. Be careful if you need
6132 * to move this code around, and make sure it is outside
6135 disarm_static_keys(memcg);
6136 if (size < PAGE_SIZE)
6144 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
6145 * but in process context. The work_freeing structure is overlaid
6146 * on the rcu_freeing structure, which itself is overlaid on memsw.
6148 static void free_work(struct work_struct *work)
6150 struct mem_cgroup *memcg;
6152 memcg = container_of(work, struct mem_cgroup, work_freeing);
6153 __mem_cgroup_free(memcg);
6156 static void free_rcu(struct rcu_head *rcu_head)
6158 struct mem_cgroup *memcg;
6160 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
6161 INIT_WORK(&memcg->work_freeing, free_work);
6162 schedule_work(&memcg->work_freeing);
6165 static void mem_cgroup_get(struct mem_cgroup *memcg)
6167 atomic_inc(&memcg->refcnt);
6170 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
6172 if (atomic_sub_and_test(count, &memcg->refcnt)) {
6173 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
6174 call_rcu(&memcg->rcu_freeing, free_rcu);
6176 mem_cgroup_put(parent);
6180 static void mem_cgroup_put(struct mem_cgroup *memcg)
6182 __mem_cgroup_put(memcg, 1);
6186 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6188 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6190 if (!memcg->res.parent)
6192 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6194 EXPORT_SYMBOL(parent_mem_cgroup);
6196 static void __init mem_cgroup_soft_limit_tree_init(void)
6198 struct mem_cgroup_tree_per_node *rtpn;
6199 struct mem_cgroup_tree_per_zone *rtpz;
6200 int tmp, node, zone;
6202 for_each_node(node) {
6204 if (!node_state(node, N_NORMAL_MEMORY))
6206 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6209 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6211 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6212 rtpz = &rtpn->rb_tree_per_zone[zone];
6213 rtpz->rb_root = RB_ROOT;
6214 spin_lock_init(&rtpz->lock);
6219 static struct cgroup_subsys_state * __ref
6220 mem_cgroup_css_alloc(struct cgroup *cont)
6222 struct mem_cgroup *memcg;
6223 long error = -ENOMEM;
6226 memcg = mem_cgroup_alloc();
6228 return ERR_PTR(error);
6231 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6235 if (cont->parent == NULL) {
6236 root_mem_cgroup = memcg;
6237 res_counter_init(&memcg->res, NULL);
6238 res_counter_init(&memcg->memsw, NULL);
6239 res_counter_init(&memcg->kmem, NULL);
6242 memcg->last_scanned_node = MAX_NUMNODES;
6243 INIT_LIST_HEAD(&memcg->oom_notify);
6244 atomic_set(&memcg->refcnt, 1);
6245 memcg->move_charge_at_immigrate = 0;
6246 mutex_init(&memcg->thresholds_lock);
6247 spin_lock_init(&memcg->move_lock);
6248 vmpressure_init(&memcg->vmpressure);
6253 __mem_cgroup_free(memcg);
6254 return ERR_PTR(error);
6258 mem_cgroup_css_online(struct cgroup *cont)
6260 struct mem_cgroup *memcg, *parent;
6266 mutex_lock(&memcg_create_mutex);
6267 memcg = mem_cgroup_from_cont(cont);
6268 parent = mem_cgroup_from_cont(cont->parent);
6270 memcg->use_hierarchy = parent->use_hierarchy;
6271 memcg->oom_kill_disable = parent->oom_kill_disable;
6272 memcg->swappiness = mem_cgroup_swappiness(parent);
6274 if (parent->use_hierarchy) {
6275 res_counter_init(&memcg->res, &parent->res);
6276 res_counter_init(&memcg->memsw, &parent->memsw);
6277 res_counter_init(&memcg->kmem, &parent->kmem);
6280 * We increment refcnt of the parent to ensure that we can
6281 * safely access it on res_counter_charge/uncharge.
6282 * This refcnt will be decremented when freeing this
6283 * mem_cgroup(see mem_cgroup_put).
6285 mem_cgroup_get(parent);
6287 res_counter_init(&memcg->res, NULL);
6288 res_counter_init(&memcg->memsw, NULL);
6289 res_counter_init(&memcg->kmem, NULL);
6291 * Deeper hierachy with use_hierarchy == false doesn't make
6292 * much sense so let cgroup subsystem know about this
6293 * unfortunate state in our controller.
6295 if (parent != root_mem_cgroup)
6296 mem_cgroup_subsys.broken_hierarchy = true;
6299 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6300 mutex_unlock(&memcg_create_mutex);
6303 * We call put now because our (and parent's) refcnts
6304 * are already in place. mem_cgroup_put() will internally
6305 * call __mem_cgroup_free, so return directly
6307 mem_cgroup_put(memcg);
6308 if (parent->use_hierarchy)
6309 mem_cgroup_put(parent);
6315 * Announce all parents that a group from their hierarchy is gone.
6317 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6319 struct mem_cgroup *parent = memcg;
6321 while ((parent = parent_mem_cgroup(parent)))
6322 atomic_inc(&parent->dead_count);
6325 * if the root memcg is not hierarchical we have to check it
6328 if (!root_mem_cgroup->use_hierarchy)
6329 atomic_inc(&root_mem_cgroup->dead_count);
6332 static void mem_cgroup_css_offline(struct cgroup *cont)
6334 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6336 mem_cgroup_invalidate_reclaim_iterators(memcg);
6337 mem_cgroup_reparent_charges(memcg);
6338 mem_cgroup_destroy_all_caches(memcg);
6341 static void mem_cgroup_css_free(struct cgroup *cont)
6343 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6345 kmem_cgroup_destroy(memcg);
6347 mem_cgroup_put(memcg);
6351 /* Handlers for move charge at task migration. */
6352 #define PRECHARGE_COUNT_AT_ONCE 256
6353 static int mem_cgroup_do_precharge(unsigned long count)
6356 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6357 struct mem_cgroup *memcg = mc.to;
6359 if (mem_cgroup_is_root(memcg)) {
6360 mc.precharge += count;
6361 /* we don't need css_get for root */
6364 /* try to charge at once */
6366 struct res_counter *dummy;
6368 * "memcg" cannot be under rmdir() because we've already checked
6369 * by cgroup_lock_live_cgroup() that it is not removed and we
6370 * are still under the same cgroup_mutex. So we can postpone
6373 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6375 if (do_swap_account && res_counter_charge(&memcg->memsw,
6376 PAGE_SIZE * count, &dummy)) {
6377 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6380 mc.precharge += count;
6384 /* fall back to one by one charge */
6386 if (signal_pending(current)) {
6390 if (!batch_count--) {
6391 batch_count = PRECHARGE_COUNT_AT_ONCE;
6394 ret = __mem_cgroup_try_charge(NULL,
6395 GFP_KERNEL, 1, &memcg, false);
6397 /* mem_cgroup_clear_mc() will do uncharge later */
6405 * get_mctgt_type - get target type of moving charge
6406 * @vma: the vma the pte to be checked belongs
6407 * @addr: the address corresponding to the pte to be checked
6408 * @ptent: the pte to be checked
6409 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6412 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6413 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6414 * move charge. if @target is not NULL, the page is stored in target->page
6415 * with extra refcnt got(Callers should handle it).
6416 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6417 * target for charge migration. if @target is not NULL, the entry is stored
6420 * Called with pte lock held.
6427 enum mc_target_type {
6433 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6434 unsigned long addr, pte_t ptent)
6436 struct page *page = vm_normal_page(vma, addr, ptent);
6438 if (!page || !page_mapped(page))
6440 if (PageAnon(page)) {
6441 /* we don't move shared anon */
6444 } else if (!move_file())
6445 /* we ignore mapcount for file pages */
6447 if (!get_page_unless_zero(page))
6454 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6455 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6457 struct page *page = NULL;
6458 swp_entry_t ent = pte_to_swp_entry(ptent);
6460 if (!move_anon() || non_swap_entry(ent))
6463 * Because lookup_swap_cache() updates some statistics counter,
6464 * we call find_get_page() with swapper_space directly.
6466 page = find_get_page(swap_address_space(ent), ent.val);
6467 if (do_swap_account)
6468 entry->val = ent.val;
6473 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6474 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6480 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6481 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6483 struct page *page = NULL;
6484 struct address_space *mapping;
6487 if (!vma->vm_file) /* anonymous vma */
6492 mapping = vma->vm_file->f_mapping;
6493 if (pte_none(ptent))
6494 pgoff = linear_page_index(vma, addr);
6495 else /* pte_file(ptent) is true */
6496 pgoff = pte_to_pgoff(ptent);
6498 /* page is moved even if it's not RSS of this task(page-faulted). */
6499 page = find_get_page(mapping, pgoff);
6502 /* shmem/tmpfs may report page out on swap: account for that too. */
6503 if (radix_tree_exceptional_entry(page)) {
6504 swp_entry_t swap = radix_to_swp_entry(page);
6505 if (do_swap_account)
6507 page = find_get_page(swap_address_space(swap), swap.val);
6513 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6514 unsigned long addr, pte_t ptent, union mc_target *target)
6516 struct page *page = NULL;
6517 struct page_cgroup *pc;
6518 enum mc_target_type ret = MC_TARGET_NONE;
6519 swp_entry_t ent = { .val = 0 };
6521 if (pte_present(ptent))
6522 page = mc_handle_present_pte(vma, addr, ptent);
6523 else if (is_swap_pte(ptent))
6524 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6525 else if (pte_none(ptent) || pte_file(ptent))
6526 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6528 if (!page && !ent.val)
6531 pc = lookup_page_cgroup(page);
6533 * Do only loose check w/o page_cgroup lock.
6534 * mem_cgroup_move_account() checks the pc is valid or not under
6537 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6538 ret = MC_TARGET_PAGE;
6540 target->page = page;
6542 if (!ret || !target)
6545 /* There is a swap entry and a page doesn't exist or isn't charged */
6546 if (ent.val && !ret &&
6547 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6548 ret = MC_TARGET_SWAP;
6555 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6557 * We don't consider swapping or file mapped pages because THP does not
6558 * support them for now.
6559 * Caller should make sure that pmd_trans_huge(pmd) is true.
6561 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6562 unsigned long addr, pmd_t pmd, union mc_target *target)
6564 struct page *page = NULL;
6565 struct page_cgroup *pc;
6566 enum mc_target_type ret = MC_TARGET_NONE;
6568 page = pmd_page(pmd);
6569 VM_BUG_ON(!page || !PageHead(page));
6572 pc = lookup_page_cgroup(page);
6573 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6574 ret = MC_TARGET_PAGE;
6577 target->page = page;
6583 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6584 unsigned long addr, pmd_t pmd, union mc_target *target)
6586 return MC_TARGET_NONE;
6590 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6591 unsigned long addr, unsigned long end,
6592 struct mm_walk *walk)
6594 struct vm_area_struct *vma = walk->private;
6598 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6599 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6600 mc.precharge += HPAGE_PMD_NR;
6601 spin_unlock(&vma->vm_mm->page_table_lock);
6605 if (pmd_trans_unstable(pmd))
6607 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6608 for (; addr != end; pte++, addr += PAGE_SIZE)
6609 if (get_mctgt_type(vma, addr, *pte, NULL))
6610 mc.precharge++; /* increment precharge temporarily */
6611 pte_unmap_unlock(pte - 1, ptl);
6617 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6619 unsigned long precharge;
6620 struct vm_area_struct *vma;
6622 down_read(&mm->mmap_sem);
6623 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6624 struct mm_walk mem_cgroup_count_precharge_walk = {
6625 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6629 if (is_vm_hugetlb_page(vma))
6631 walk_page_range(vma->vm_start, vma->vm_end,
6632 &mem_cgroup_count_precharge_walk);
6634 up_read(&mm->mmap_sem);
6636 precharge = mc.precharge;
6642 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6644 unsigned long precharge = mem_cgroup_count_precharge(mm);
6646 VM_BUG_ON(mc.moving_task);
6647 mc.moving_task = current;
6648 return mem_cgroup_do_precharge(precharge);
6651 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6652 static void __mem_cgroup_clear_mc(void)
6654 struct mem_cgroup *from = mc.from;
6655 struct mem_cgroup *to = mc.to;
6657 /* we must uncharge all the leftover precharges from mc.to */
6659 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6663 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6664 * we must uncharge here.
6666 if (mc.moved_charge) {
6667 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6668 mc.moved_charge = 0;
6670 /* we must fixup refcnts and charges */
6671 if (mc.moved_swap) {
6672 /* uncharge swap account from the old cgroup */
6673 if (!mem_cgroup_is_root(mc.from))
6674 res_counter_uncharge(&mc.from->memsw,
6675 PAGE_SIZE * mc.moved_swap);
6676 __mem_cgroup_put(mc.from, mc.moved_swap);
6678 if (!mem_cgroup_is_root(mc.to)) {
6680 * we charged both to->res and to->memsw, so we should
6683 res_counter_uncharge(&mc.to->res,
6684 PAGE_SIZE * mc.moved_swap);
6686 /* we've already done mem_cgroup_get(mc.to) */
6689 memcg_oom_recover(from);
6690 memcg_oom_recover(to);
6691 wake_up_all(&mc.waitq);
6694 static void mem_cgroup_clear_mc(void)
6696 struct mem_cgroup *from = mc.from;
6699 * we must clear moving_task before waking up waiters at the end of
6702 mc.moving_task = NULL;
6703 __mem_cgroup_clear_mc();
6704 spin_lock(&mc.lock);
6707 spin_unlock(&mc.lock);
6708 mem_cgroup_end_move(from);
6711 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6712 struct cgroup_taskset *tset)
6714 struct task_struct *p = cgroup_taskset_first(tset);
6716 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6717 unsigned long move_charge_at_immigrate;
6720 * We are now commited to this value whatever it is. Changes in this
6721 * tunable will only affect upcoming migrations, not the current one.
6722 * So we need to save it, and keep it going.
6724 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6725 if (move_charge_at_immigrate) {
6726 struct mm_struct *mm;
6727 struct mem_cgroup *from = mem_cgroup_from_task(p);
6729 VM_BUG_ON(from == memcg);
6731 mm = get_task_mm(p);
6734 /* We move charges only when we move a owner of the mm */
6735 if (mm->owner == p) {
6738 VM_BUG_ON(mc.precharge);
6739 VM_BUG_ON(mc.moved_charge);
6740 VM_BUG_ON(mc.moved_swap);
6741 mem_cgroup_start_move(from);
6742 spin_lock(&mc.lock);
6745 mc.immigrate_flags = move_charge_at_immigrate;
6746 spin_unlock(&mc.lock);
6747 /* We set mc.moving_task later */
6749 ret = mem_cgroup_precharge_mc(mm);
6751 mem_cgroup_clear_mc();
6758 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6759 struct cgroup_taskset *tset)
6761 mem_cgroup_clear_mc();
6764 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6765 unsigned long addr, unsigned long end,
6766 struct mm_walk *walk)
6769 struct vm_area_struct *vma = walk->private;
6772 enum mc_target_type target_type;
6773 union mc_target target;
6775 struct page_cgroup *pc;
6778 * We don't take compound_lock() here but no race with splitting thp
6780 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6781 * under splitting, which means there's no concurrent thp split,
6782 * - if another thread runs into split_huge_page() just after we
6783 * entered this if-block, the thread must wait for page table lock
6784 * to be unlocked in __split_huge_page_splitting(), where the main
6785 * part of thp split is not executed yet.
6787 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6788 if (mc.precharge < HPAGE_PMD_NR) {
6789 spin_unlock(&vma->vm_mm->page_table_lock);
6792 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6793 if (target_type == MC_TARGET_PAGE) {
6795 if (!isolate_lru_page(page)) {
6796 pc = lookup_page_cgroup(page);
6797 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6798 pc, mc.from, mc.to)) {
6799 mc.precharge -= HPAGE_PMD_NR;
6800 mc.moved_charge += HPAGE_PMD_NR;
6802 putback_lru_page(page);
6806 spin_unlock(&vma->vm_mm->page_table_lock);
6810 if (pmd_trans_unstable(pmd))
6813 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6814 for (; addr != end; addr += PAGE_SIZE) {
6815 pte_t ptent = *(pte++);
6821 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6822 case MC_TARGET_PAGE:
6824 if (isolate_lru_page(page))
6826 pc = lookup_page_cgroup(page);
6827 if (!mem_cgroup_move_account(page, 1, pc,
6830 /* we uncharge from mc.from later. */
6833 putback_lru_page(page);
6834 put: /* get_mctgt_type() gets the page */
6837 case MC_TARGET_SWAP:
6839 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6841 /* we fixup refcnts and charges later. */
6849 pte_unmap_unlock(pte - 1, ptl);
6854 * We have consumed all precharges we got in can_attach().
6855 * We try charge one by one, but don't do any additional
6856 * charges to mc.to if we have failed in charge once in attach()
6859 ret = mem_cgroup_do_precharge(1);
6867 static void mem_cgroup_move_charge(struct mm_struct *mm)
6869 struct vm_area_struct *vma;
6871 lru_add_drain_all();
6873 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6875 * Someone who are holding the mmap_sem might be waiting in
6876 * waitq. So we cancel all extra charges, wake up all waiters,
6877 * and retry. Because we cancel precharges, we might not be able
6878 * to move enough charges, but moving charge is a best-effort
6879 * feature anyway, so it wouldn't be a big problem.
6881 __mem_cgroup_clear_mc();
6885 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6887 struct mm_walk mem_cgroup_move_charge_walk = {
6888 .pmd_entry = mem_cgroup_move_charge_pte_range,
6892 if (is_vm_hugetlb_page(vma))
6894 ret = walk_page_range(vma->vm_start, vma->vm_end,
6895 &mem_cgroup_move_charge_walk);
6898 * means we have consumed all precharges and failed in
6899 * doing additional charge. Just abandon here.
6903 up_read(&mm->mmap_sem);
6906 static void mem_cgroup_move_task(struct cgroup *cont,
6907 struct cgroup_taskset *tset)
6909 struct task_struct *p = cgroup_taskset_first(tset);
6910 struct mm_struct *mm = get_task_mm(p);
6914 mem_cgroup_move_charge(mm);
6918 mem_cgroup_clear_mc();
6920 #else /* !CONFIG_MMU */
6921 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6922 struct cgroup_taskset *tset)
6926 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6927 struct cgroup_taskset *tset)
6930 static void mem_cgroup_move_task(struct cgroup *cont,
6931 struct cgroup_taskset *tset)
6937 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6938 * to verify sane_behavior flag on each mount attempt.
6940 static void mem_cgroup_bind(struct cgroup *root)
6943 * use_hierarchy is forced with sane_behavior. cgroup core
6944 * guarantees that @root doesn't have any children, so turning it
6945 * on for the root memcg is enough.
6947 if (cgroup_sane_behavior(root))
6948 mem_cgroup_from_cont(root)->use_hierarchy = true;
6951 struct cgroup_subsys mem_cgroup_subsys = {
6953 .subsys_id = mem_cgroup_subsys_id,
6954 .css_alloc = mem_cgroup_css_alloc,
6955 .css_online = mem_cgroup_css_online,
6956 .css_offline = mem_cgroup_css_offline,
6957 .css_free = mem_cgroup_css_free,
6958 .can_attach = mem_cgroup_can_attach,
6959 .cancel_attach = mem_cgroup_cancel_attach,
6960 .attach = mem_cgroup_move_task,
6961 .bind = mem_cgroup_bind,
6962 .base_cftypes = mem_cgroup_files,
6967 #ifdef CONFIG_MEMCG_SWAP
6968 static int __init enable_swap_account(char *s)
6970 /* consider enabled if no parameter or 1 is given */
6971 if (!strcmp(s, "1"))
6972 really_do_swap_account = 1;
6973 else if (!strcmp(s, "0"))
6974 really_do_swap_account = 0;
6977 __setup("swapaccount=", enable_swap_account);
6979 static void __init memsw_file_init(void)
6981 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6984 static void __init enable_swap_cgroup(void)
6986 if (!mem_cgroup_disabled() && really_do_swap_account) {
6987 do_swap_account = 1;
6993 static void __init enable_swap_cgroup(void)
6999 * subsys_initcall() for memory controller.
7001 * Some parts like hotcpu_notifier() have to be initialized from this context
7002 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7003 * everything that doesn't depend on a specific mem_cgroup structure should
7004 * be initialized from here.
7006 static int __init mem_cgroup_init(void)
7008 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7009 enable_swap_cgroup();
7010 mem_cgroup_soft_limit_tree_init();
7014 subsys_initcall(mem_cgroup_init);