1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/sort.h>
50 #include <linux/seq_file.h>
51 #include <linux/vmalloc.h>
52 #include <linux/mm_inline.h>
53 #include <linux/page_cgroup.h>
54 #include <linux/cpu.h>
55 #include <linux/oom.h>
59 #include <net/tcp_memcontrol.h>
61 #include <asm/uaccess.h>
63 #include <trace/events/vmscan.h>
65 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
66 EXPORT_SYMBOL(mem_cgroup_subsys);
68 #define MEM_CGROUP_RECLAIM_RETRIES 5
69 static struct mem_cgroup *root_mem_cgroup __read_mostly;
71 #ifdef CONFIG_MEMCG_SWAP
72 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
73 int do_swap_account __read_mostly;
75 /* for remember boot option*/
76 #ifdef CONFIG_MEMCG_SWAP_ENABLED
77 static int really_do_swap_account __initdata = 1;
79 static int really_do_swap_account __initdata = 0;
83 #define do_swap_account 0
88 * Statistics for memory cgroup.
90 enum mem_cgroup_stat_index {
92 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
94 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
95 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
96 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
97 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
98 MEM_CGROUP_STAT_NSTATS,
101 static const char * const mem_cgroup_stat_names[] = {
108 enum mem_cgroup_events_index {
109 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
110 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
111 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
112 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
113 MEM_CGROUP_EVENTS_NSTATS,
116 static const char * const mem_cgroup_events_names[] = {
123 static const char * const mem_cgroup_lru_names[] = {
132 * Per memcg event counter is incremented at every pagein/pageout. With THP,
133 * it will be incremated by the number of pages. This counter is used for
134 * for trigger some periodic events. This is straightforward and better
135 * than using jiffies etc. to handle periodic memcg event.
137 enum mem_cgroup_events_target {
138 MEM_CGROUP_TARGET_THRESH,
139 MEM_CGROUP_TARGET_SOFTLIMIT,
140 MEM_CGROUP_TARGET_NUMAINFO,
143 #define THRESHOLDS_EVENTS_TARGET 128
144 #define SOFTLIMIT_EVENTS_TARGET 1024
145 #define NUMAINFO_EVENTS_TARGET 1024
147 struct mem_cgroup_stat_cpu {
148 long count[MEM_CGROUP_STAT_NSTATS];
149 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
150 unsigned long nr_page_events;
151 unsigned long targets[MEM_CGROUP_NTARGETS];
154 struct mem_cgroup_reclaim_iter {
155 /* css_id of the last scanned hierarchy member */
157 /* scan generation, increased every round-trip */
158 unsigned int generation;
162 * per-zone information in memory controller.
164 struct mem_cgroup_per_zone {
165 struct lruvec lruvec;
166 unsigned long lru_size[NR_LRU_LISTS];
168 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
170 struct rb_node tree_node; /* RB tree node */
171 unsigned long long usage_in_excess;/* Set to the value by which */
172 /* the soft limit is exceeded*/
174 struct mem_cgroup *memcg; /* Back pointer, we cannot */
175 /* use container_of */
178 struct mem_cgroup_per_node {
179 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
182 struct mem_cgroup_lru_info {
183 struct mem_cgroup_per_node *nodeinfo[0];
187 * Cgroups above their limits are maintained in a RB-Tree, independent of
188 * their hierarchy representation
191 struct mem_cgroup_tree_per_zone {
192 struct rb_root rb_root;
196 struct mem_cgroup_tree_per_node {
197 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
200 struct mem_cgroup_tree {
201 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
204 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
206 struct mem_cgroup_threshold {
207 struct eventfd_ctx *eventfd;
212 struct mem_cgroup_threshold_ary {
213 /* An array index points to threshold just below or equal to usage. */
214 int current_threshold;
215 /* Size of entries[] */
217 /* Array of thresholds */
218 struct mem_cgroup_threshold entries[0];
221 struct mem_cgroup_thresholds {
222 /* Primary thresholds array */
223 struct mem_cgroup_threshold_ary *primary;
225 * Spare threshold array.
226 * This is needed to make mem_cgroup_unregister_event() "never fail".
227 * It must be able to store at least primary->size - 1 entries.
229 struct mem_cgroup_threshold_ary *spare;
233 struct mem_cgroup_eventfd_list {
234 struct list_head list;
235 struct eventfd_ctx *eventfd;
238 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
239 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
242 * The memory controller data structure. The memory controller controls both
243 * page cache and RSS per cgroup. We would eventually like to provide
244 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
245 * to help the administrator determine what knobs to tune.
247 * TODO: Add a water mark for the memory controller. Reclaim will begin when
248 * we hit the water mark. May be even add a low water mark, such that
249 * no reclaim occurs from a cgroup at it's low water mark, this is
250 * a feature that will be implemented much later in the future.
253 struct cgroup_subsys_state css;
255 * the counter to account for memory usage
257 struct res_counter res;
261 * the counter to account for mem+swap usage.
263 struct res_counter memsw;
266 * rcu_freeing is used only when freeing struct mem_cgroup,
267 * so put it into a union to avoid wasting more memory.
268 * It must be disjoint from the css field. It could be
269 * in a union with the res field, but res plays a much
270 * larger part in mem_cgroup life than memsw, and might
271 * be of interest, even at time of free, when debugging.
272 * So share rcu_head with the less interesting memsw.
274 struct rcu_head rcu_freeing;
276 * We also need some space for a worker in deferred freeing.
277 * By the time we call it, rcu_freeing is no longer in use.
279 struct work_struct work_freeing;
283 * the counter to account for kernel memory usage.
285 struct res_counter kmem;
287 * Should the accounting and control be hierarchical, per subtree?
290 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
298 /* OOM-Killer disable */
299 int oom_kill_disable;
301 /* set when res.limit == memsw.limit */
302 bool memsw_is_minimum;
304 /* protect arrays of thresholds */
305 struct mutex thresholds_lock;
307 /* thresholds for memory usage. RCU-protected */
308 struct mem_cgroup_thresholds thresholds;
310 /* thresholds for mem+swap usage. RCU-protected */
311 struct mem_cgroup_thresholds memsw_thresholds;
313 /* For oom notifier event fd */
314 struct list_head oom_notify;
317 * Should we move charges of a task when a task is moved into this
318 * mem_cgroup ? And what type of charges should we move ?
320 unsigned long move_charge_at_immigrate;
322 * set > 0 if pages under this cgroup are moving to other cgroup.
324 atomic_t moving_account;
325 /* taken only while moving_account > 0 */
326 spinlock_t move_lock;
330 struct mem_cgroup_stat_cpu __percpu *stat;
332 * used when a cpu is offlined or other synchronizations
333 * See mem_cgroup_read_stat().
335 struct mem_cgroup_stat_cpu nocpu_base;
336 spinlock_t pcp_counter_lock;
338 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
339 struct tcp_memcontrol tcp_mem;
341 #if defined(CONFIG_MEMCG_KMEM)
342 /* analogous to slab_common's slab_caches list. per-memcg */
343 struct list_head memcg_slab_caches;
344 /* Not a spinlock, we can take a lot of time walking the list */
345 struct mutex slab_caches_mutex;
346 /* Index in the kmem_cache->memcg_params->memcg_caches array */
350 int last_scanned_node;
352 nodemask_t scan_nodes;
353 atomic_t numainfo_events;
354 atomic_t numainfo_updating;
357 * Per cgroup active and inactive list, similar to the
358 * per zone LRU lists.
360 * WARNING: This has to be the last element of the struct. Don't
361 * add new fields after this point.
363 struct mem_cgroup_lru_info info;
366 static size_t memcg_size(void)
368 return sizeof(struct mem_cgroup) +
369 nr_node_ids * sizeof(struct mem_cgroup_per_node);
372 /* internal only representation about the status of kmem accounting. */
374 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
375 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
376 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
379 /* We account when limit is on, but only after call sites are patched */
380 #define KMEM_ACCOUNTED_MASK \
381 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
383 #ifdef CONFIG_MEMCG_KMEM
384 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
386 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
389 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
391 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
394 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
396 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
399 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
401 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
404 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
406 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
407 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
410 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
412 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
413 &memcg->kmem_account_flags);
417 /* Stuffs for move charges at task migration. */
419 * Types of charges to be moved. "move_charge_at_immitgrate" and
420 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
423 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
424 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
428 /* "mc" and its members are protected by cgroup_mutex */
429 static struct move_charge_struct {
430 spinlock_t lock; /* for from, to */
431 struct mem_cgroup *from;
432 struct mem_cgroup *to;
433 unsigned long immigrate_flags;
434 unsigned long precharge;
435 unsigned long moved_charge;
436 unsigned long moved_swap;
437 struct task_struct *moving_task; /* a task moving charges */
438 wait_queue_head_t waitq; /* a waitq for other context */
440 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
441 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
444 static bool move_anon(void)
446 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
449 static bool move_file(void)
451 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
455 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
456 * limit reclaim to prevent infinite loops, if they ever occur.
458 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
459 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
462 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
463 MEM_CGROUP_CHARGE_TYPE_ANON,
464 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
465 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
469 /* for encoding cft->private value on file */
477 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
478 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
479 #define MEMFILE_ATTR(val) ((val) & 0xffff)
480 /* Used for OOM nofiier */
481 #define OOM_CONTROL (0)
484 * Reclaim flags for mem_cgroup_hierarchical_reclaim
486 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
487 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
488 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
489 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
491 static void mem_cgroup_get(struct mem_cgroup *memcg);
492 static void mem_cgroup_put(struct mem_cgroup *memcg);
495 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
497 return container_of(s, struct mem_cgroup, css);
500 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
502 return (memcg == root_mem_cgroup);
505 /* Writing them here to avoid exposing memcg's inner layout */
506 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
508 void sock_update_memcg(struct sock *sk)
510 if (mem_cgroup_sockets_enabled) {
511 struct mem_cgroup *memcg;
512 struct cg_proto *cg_proto;
514 BUG_ON(!sk->sk_prot->proto_cgroup);
516 /* Socket cloning can throw us here with sk_cgrp already
517 * filled. It won't however, necessarily happen from
518 * process context. So the test for root memcg given
519 * the current task's memcg won't help us in this case.
521 * Respecting the original socket's memcg is a better
522 * decision in this case.
525 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
526 mem_cgroup_get(sk->sk_cgrp->memcg);
531 memcg = mem_cgroup_from_task(current);
532 cg_proto = sk->sk_prot->proto_cgroup(memcg);
533 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
534 mem_cgroup_get(memcg);
535 sk->sk_cgrp = cg_proto;
540 EXPORT_SYMBOL(sock_update_memcg);
542 void sock_release_memcg(struct sock *sk)
544 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
545 struct mem_cgroup *memcg;
546 WARN_ON(!sk->sk_cgrp->memcg);
547 memcg = sk->sk_cgrp->memcg;
548 mem_cgroup_put(memcg);
552 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
554 if (!memcg || mem_cgroup_is_root(memcg))
557 return &memcg->tcp_mem.cg_proto;
559 EXPORT_SYMBOL(tcp_proto_cgroup);
561 static void disarm_sock_keys(struct mem_cgroup *memcg)
563 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
565 static_key_slow_dec(&memcg_socket_limit_enabled);
568 static void disarm_sock_keys(struct mem_cgroup *memcg)
573 #ifdef CONFIG_MEMCG_KMEM
575 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
576 * There are two main reasons for not using the css_id for this:
577 * 1) this works better in sparse environments, where we have a lot of memcgs,
578 * but only a few kmem-limited. Or also, if we have, for instance, 200
579 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
580 * 200 entry array for that.
582 * 2) In order not to violate the cgroup API, we would like to do all memory
583 * allocation in ->create(). At that point, we haven't yet allocated the
584 * css_id. Having a separate index prevents us from messing with the cgroup
587 * The current size of the caches array is stored in
588 * memcg_limited_groups_array_size. It will double each time we have to
591 static DEFINE_IDA(kmem_limited_groups);
592 int memcg_limited_groups_array_size;
595 * MIN_SIZE is different than 1, because we would like to avoid going through
596 * the alloc/free process all the time. In a small machine, 4 kmem-limited
597 * cgroups is a reasonable guess. In the future, it could be a parameter or
598 * tunable, but that is strictly not necessary.
600 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
601 * this constant directly from cgroup, but it is understandable that this is
602 * better kept as an internal representation in cgroup.c. In any case, the
603 * css_id space is not getting any smaller, and we don't have to necessarily
604 * increase ours as well if it increases.
606 #define MEMCG_CACHES_MIN_SIZE 4
607 #define MEMCG_CACHES_MAX_SIZE 65535
610 * A lot of the calls to the cache allocation functions are expected to be
611 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
612 * conditional to this static branch, we'll have to allow modules that does
613 * kmem_cache_alloc and the such to see this symbol as well
615 struct static_key memcg_kmem_enabled_key;
616 EXPORT_SYMBOL(memcg_kmem_enabled_key);
618 static void disarm_kmem_keys(struct mem_cgroup *memcg)
620 if (memcg_kmem_is_active(memcg)) {
621 static_key_slow_dec(&memcg_kmem_enabled_key);
622 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
625 * This check can't live in kmem destruction function,
626 * since the charges will outlive the cgroup
628 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
631 static void disarm_kmem_keys(struct mem_cgroup *memcg)
634 #endif /* CONFIG_MEMCG_KMEM */
636 static void disarm_static_keys(struct mem_cgroup *memcg)
638 disarm_sock_keys(memcg);
639 disarm_kmem_keys(memcg);
642 static void drain_all_stock_async(struct mem_cgroup *memcg);
644 static struct mem_cgroup_per_zone *
645 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
647 VM_BUG_ON((unsigned)nid >= nr_node_ids);
648 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
651 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
656 static struct mem_cgroup_per_zone *
657 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
659 int nid = page_to_nid(page);
660 int zid = page_zonenum(page);
662 return mem_cgroup_zoneinfo(memcg, nid, zid);
665 static struct mem_cgroup_tree_per_zone *
666 soft_limit_tree_node_zone(int nid, int zid)
668 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
671 static struct mem_cgroup_tree_per_zone *
672 soft_limit_tree_from_page(struct page *page)
674 int nid = page_to_nid(page);
675 int zid = page_zonenum(page);
677 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
681 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
682 struct mem_cgroup_per_zone *mz,
683 struct mem_cgroup_tree_per_zone *mctz,
684 unsigned long long new_usage_in_excess)
686 struct rb_node **p = &mctz->rb_root.rb_node;
687 struct rb_node *parent = NULL;
688 struct mem_cgroup_per_zone *mz_node;
693 mz->usage_in_excess = new_usage_in_excess;
694 if (!mz->usage_in_excess)
698 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
700 if (mz->usage_in_excess < mz_node->usage_in_excess)
703 * We can't avoid mem cgroups that are over their soft
704 * limit by the same amount
706 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
709 rb_link_node(&mz->tree_node, parent, p);
710 rb_insert_color(&mz->tree_node, &mctz->rb_root);
715 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
716 struct mem_cgroup_per_zone *mz,
717 struct mem_cgroup_tree_per_zone *mctz)
721 rb_erase(&mz->tree_node, &mctz->rb_root);
726 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
727 struct mem_cgroup_per_zone *mz,
728 struct mem_cgroup_tree_per_zone *mctz)
730 spin_lock(&mctz->lock);
731 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
732 spin_unlock(&mctz->lock);
736 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
738 unsigned long long excess;
739 struct mem_cgroup_per_zone *mz;
740 struct mem_cgroup_tree_per_zone *mctz;
741 int nid = page_to_nid(page);
742 int zid = page_zonenum(page);
743 mctz = soft_limit_tree_from_page(page);
746 * Necessary to update all ancestors when hierarchy is used.
747 * because their event counter is not touched.
749 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
750 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
751 excess = res_counter_soft_limit_excess(&memcg->res);
753 * We have to update the tree if mz is on RB-tree or
754 * mem is over its softlimit.
756 if (excess || mz->on_tree) {
757 spin_lock(&mctz->lock);
758 /* if on-tree, remove it */
760 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
762 * Insert again. mz->usage_in_excess will be updated.
763 * If excess is 0, no tree ops.
765 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
766 spin_unlock(&mctz->lock);
771 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
774 struct mem_cgroup_per_zone *mz;
775 struct mem_cgroup_tree_per_zone *mctz;
777 for_each_node(node) {
778 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
779 mz = mem_cgroup_zoneinfo(memcg, node, zone);
780 mctz = soft_limit_tree_node_zone(node, zone);
781 mem_cgroup_remove_exceeded(memcg, mz, mctz);
786 static struct mem_cgroup_per_zone *
787 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
789 struct rb_node *rightmost = NULL;
790 struct mem_cgroup_per_zone *mz;
794 rightmost = rb_last(&mctz->rb_root);
796 goto done; /* Nothing to reclaim from */
798 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
800 * Remove the node now but someone else can add it back,
801 * we will to add it back at the end of reclaim to its correct
802 * position in the tree.
804 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
805 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
806 !css_tryget(&mz->memcg->css))
812 static struct mem_cgroup_per_zone *
813 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
815 struct mem_cgroup_per_zone *mz;
817 spin_lock(&mctz->lock);
818 mz = __mem_cgroup_largest_soft_limit_node(mctz);
819 spin_unlock(&mctz->lock);
824 * Implementation Note: reading percpu statistics for memcg.
826 * Both of vmstat[] and percpu_counter has threshold and do periodic
827 * synchronization to implement "quick" read. There are trade-off between
828 * reading cost and precision of value. Then, we may have a chance to implement
829 * a periodic synchronizion of counter in memcg's counter.
831 * But this _read() function is used for user interface now. The user accounts
832 * memory usage by memory cgroup and he _always_ requires exact value because
833 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
834 * have to visit all online cpus and make sum. So, for now, unnecessary
835 * synchronization is not implemented. (just implemented for cpu hotplug)
837 * If there are kernel internal actions which can make use of some not-exact
838 * value, and reading all cpu value can be performance bottleneck in some
839 * common workload, threashold and synchonization as vmstat[] should be
842 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
843 enum mem_cgroup_stat_index idx)
849 for_each_online_cpu(cpu)
850 val += per_cpu(memcg->stat->count[idx], cpu);
851 #ifdef CONFIG_HOTPLUG_CPU
852 spin_lock(&memcg->pcp_counter_lock);
853 val += memcg->nocpu_base.count[idx];
854 spin_unlock(&memcg->pcp_counter_lock);
860 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
863 int val = (charge) ? 1 : -1;
864 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
867 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
868 enum mem_cgroup_events_index idx)
870 unsigned long val = 0;
873 for_each_online_cpu(cpu)
874 val += per_cpu(memcg->stat->events[idx], cpu);
875 #ifdef CONFIG_HOTPLUG_CPU
876 spin_lock(&memcg->pcp_counter_lock);
877 val += memcg->nocpu_base.events[idx];
878 spin_unlock(&memcg->pcp_counter_lock);
883 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
884 bool anon, int nr_pages)
889 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
890 * counted as CACHE even if it's on ANON LRU.
893 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
896 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
899 /* pagein of a big page is an event. So, ignore page size */
901 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
903 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
904 nr_pages = -nr_pages; /* for event */
907 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
913 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
915 struct mem_cgroup_per_zone *mz;
917 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
918 return mz->lru_size[lru];
922 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
923 unsigned int lru_mask)
925 struct mem_cgroup_per_zone *mz;
927 unsigned long ret = 0;
929 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
932 if (BIT(lru) & lru_mask)
933 ret += mz->lru_size[lru];
939 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
940 int nid, unsigned int lru_mask)
945 for (zid = 0; zid < MAX_NR_ZONES; zid++)
946 total += mem_cgroup_zone_nr_lru_pages(memcg,
952 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
953 unsigned int lru_mask)
958 for_each_node_state(nid, N_MEMORY)
959 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
963 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
964 enum mem_cgroup_events_target target)
966 unsigned long val, next;
968 val = __this_cpu_read(memcg->stat->nr_page_events);
969 next = __this_cpu_read(memcg->stat->targets[target]);
970 /* from time_after() in jiffies.h */
971 if ((long)next - (long)val < 0) {
973 case MEM_CGROUP_TARGET_THRESH:
974 next = val + THRESHOLDS_EVENTS_TARGET;
976 case MEM_CGROUP_TARGET_SOFTLIMIT:
977 next = val + SOFTLIMIT_EVENTS_TARGET;
979 case MEM_CGROUP_TARGET_NUMAINFO:
980 next = val + NUMAINFO_EVENTS_TARGET;
985 __this_cpu_write(memcg->stat->targets[target], next);
992 * Check events in order.
995 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
998 /* threshold event is triggered in finer grain than soft limit */
999 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1000 MEM_CGROUP_TARGET_THRESH))) {
1002 bool do_numainfo __maybe_unused;
1004 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1005 MEM_CGROUP_TARGET_SOFTLIMIT);
1006 #if MAX_NUMNODES > 1
1007 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1008 MEM_CGROUP_TARGET_NUMAINFO);
1012 mem_cgroup_threshold(memcg);
1013 if (unlikely(do_softlimit))
1014 mem_cgroup_update_tree(memcg, page);
1015 #if MAX_NUMNODES > 1
1016 if (unlikely(do_numainfo))
1017 atomic_inc(&memcg->numainfo_events);
1023 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1025 return mem_cgroup_from_css(
1026 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1029 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1032 * mm_update_next_owner() may clear mm->owner to NULL
1033 * if it races with swapoff, page migration, etc.
1034 * So this can be called with p == NULL.
1039 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1042 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1044 struct mem_cgroup *memcg = NULL;
1049 * Because we have no locks, mm->owner's may be being moved to other
1050 * cgroup. We use css_tryget() here even if this looks
1051 * pessimistic (rather than adding locks here).
1055 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1056 if (unlikely(!memcg))
1058 } while (!css_tryget(&memcg->css));
1064 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1065 * @root: hierarchy root
1066 * @prev: previously returned memcg, NULL on first invocation
1067 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1069 * Returns references to children of the hierarchy below @root, or
1070 * @root itself, or %NULL after a full round-trip.
1072 * Caller must pass the return value in @prev on subsequent
1073 * invocations for reference counting, or use mem_cgroup_iter_break()
1074 * to cancel a hierarchy walk before the round-trip is complete.
1076 * Reclaimers can specify a zone and a priority level in @reclaim to
1077 * divide up the memcgs in the hierarchy among all concurrent
1078 * reclaimers operating on the same zone and priority.
1080 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1081 struct mem_cgroup *prev,
1082 struct mem_cgroup_reclaim_cookie *reclaim)
1084 struct mem_cgroup *memcg = NULL;
1087 if (mem_cgroup_disabled())
1091 root = root_mem_cgroup;
1093 if (prev && !reclaim)
1094 id = css_id(&prev->css);
1096 if (prev && prev != root)
1097 css_put(&prev->css);
1099 if (!root->use_hierarchy && root != root_mem_cgroup) {
1106 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1107 struct cgroup_subsys_state *css;
1110 int nid = zone_to_nid(reclaim->zone);
1111 int zid = zone_idx(reclaim->zone);
1112 struct mem_cgroup_per_zone *mz;
1114 mz = mem_cgroup_zoneinfo(root, nid, zid);
1115 iter = &mz->reclaim_iter[reclaim->priority];
1116 if (prev && reclaim->generation != iter->generation)
1118 id = iter->position;
1122 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
1124 if (css == &root->css || css_tryget(css))
1125 memcg = mem_cgroup_from_css(css);
1131 iter->position = id;
1134 else if (!prev && memcg)
1135 reclaim->generation = iter->generation;
1145 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1146 * @root: hierarchy root
1147 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1149 void mem_cgroup_iter_break(struct mem_cgroup *root,
1150 struct mem_cgroup *prev)
1153 root = root_mem_cgroup;
1154 if (prev && prev != root)
1155 css_put(&prev->css);
1159 * Iteration constructs for visiting all cgroups (under a tree). If
1160 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1161 * be used for reference counting.
1163 #define for_each_mem_cgroup_tree(iter, root) \
1164 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1166 iter = mem_cgroup_iter(root, iter, NULL))
1168 #define for_each_mem_cgroup(iter) \
1169 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1171 iter = mem_cgroup_iter(NULL, iter, NULL))
1173 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1175 struct mem_cgroup *memcg;
1178 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1179 if (unlikely(!memcg))
1184 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1187 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1195 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1198 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1199 * @zone: zone of the wanted lruvec
1200 * @memcg: memcg of the wanted lruvec
1202 * Returns the lru list vector holding pages for the given @zone and
1203 * @mem. This can be the global zone lruvec, if the memory controller
1206 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1207 struct mem_cgroup *memcg)
1209 struct mem_cgroup_per_zone *mz;
1210 struct lruvec *lruvec;
1212 if (mem_cgroup_disabled()) {
1213 lruvec = &zone->lruvec;
1217 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1218 lruvec = &mz->lruvec;
1221 * Since a node can be onlined after the mem_cgroup was created,
1222 * we have to be prepared to initialize lruvec->zone here;
1223 * and if offlined then reonlined, we need to reinitialize it.
1225 if (unlikely(lruvec->zone != zone))
1226 lruvec->zone = zone;
1231 * Following LRU functions are allowed to be used without PCG_LOCK.
1232 * Operations are called by routine of global LRU independently from memcg.
1233 * What we have to take care of here is validness of pc->mem_cgroup.
1235 * Changes to pc->mem_cgroup happens when
1238 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1239 * It is added to LRU before charge.
1240 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1241 * When moving account, the page is not on LRU. It's isolated.
1245 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1247 * @zone: zone of the page
1249 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1251 struct mem_cgroup_per_zone *mz;
1252 struct mem_cgroup *memcg;
1253 struct page_cgroup *pc;
1254 struct lruvec *lruvec;
1256 if (mem_cgroup_disabled()) {
1257 lruvec = &zone->lruvec;
1261 pc = lookup_page_cgroup(page);
1262 memcg = pc->mem_cgroup;
1265 * Surreptitiously switch any uncharged offlist page to root:
1266 * an uncharged page off lru does nothing to secure
1267 * its former mem_cgroup from sudden removal.
1269 * Our caller holds lru_lock, and PageCgroupUsed is updated
1270 * under page_cgroup lock: between them, they make all uses
1271 * of pc->mem_cgroup safe.
1273 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1274 pc->mem_cgroup = memcg = root_mem_cgroup;
1276 mz = page_cgroup_zoneinfo(memcg, page);
1277 lruvec = &mz->lruvec;
1280 * Since a node can be onlined after the mem_cgroup was created,
1281 * we have to be prepared to initialize lruvec->zone here;
1282 * and if offlined then reonlined, we need to reinitialize it.
1284 if (unlikely(lruvec->zone != zone))
1285 lruvec->zone = zone;
1290 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1291 * @lruvec: mem_cgroup per zone lru vector
1292 * @lru: index of lru list the page is sitting on
1293 * @nr_pages: positive when adding or negative when removing
1295 * This function must be called when a page is added to or removed from an
1298 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1301 struct mem_cgroup_per_zone *mz;
1302 unsigned long *lru_size;
1304 if (mem_cgroup_disabled())
1307 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1308 lru_size = mz->lru_size + lru;
1309 *lru_size += nr_pages;
1310 VM_BUG_ON((long)(*lru_size) < 0);
1314 * Checks whether given mem is same or in the root_mem_cgroup's
1317 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1318 struct mem_cgroup *memcg)
1320 if (root_memcg == memcg)
1322 if (!root_memcg->use_hierarchy || !memcg)
1324 return css_is_ancestor(&memcg->css, &root_memcg->css);
1327 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1328 struct mem_cgroup *memcg)
1333 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1338 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1341 struct mem_cgroup *curr = NULL;
1342 struct task_struct *p;
1344 p = find_lock_task_mm(task);
1346 curr = try_get_mem_cgroup_from_mm(p->mm);
1350 * All threads may have already detached their mm's, but the oom
1351 * killer still needs to detect if they have already been oom
1352 * killed to prevent needlessly killing additional tasks.
1355 curr = mem_cgroup_from_task(task);
1357 css_get(&curr->css);
1363 * We should check use_hierarchy of "memcg" not "curr". Because checking
1364 * use_hierarchy of "curr" here make this function true if hierarchy is
1365 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1366 * hierarchy(even if use_hierarchy is disabled in "memcg").
1368 ret = mem_cgroup_same_or_subtree(memcg, curr);
1369 css_put(&curr->css);
1373 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1375 unsigned long inactive_ratio;
1376 unsigned long inactive;
1377 unsigned long active;
1380 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1381 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1383 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1385 inactive_ratio = int_sqrt(10 * gb);
1389 return inactive * inactive_ratio < active;
1392 int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1394 unsigned long active;
1395 unsigned long inactive;
1397 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_FILE);
1398 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_FILE);
1400 return (active > inactive);
1403 #define mem_cgroup_from_res_counter(counter, member) \
1404 container_of(counter, struct mem_cgroup, member)
1407 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1408 * @memcg: the memory cgroup
1410 * Returns the maximum amount of memory @mem can be charged with, in
1413 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1415 unsigned long long margin;
1417 margin = res_counter_margin(&memcg->res);
1418 if (do_swap_account)
1419 margin = min(margin, res_counter_margin(&memcg->memsw));
1420 return margin >> PAGE_SHIFT;
1423 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1425 struct cgroup *cgrp = memcg->css.cgroup;
1428 if (cgrp->parent == NULL)
1429 return vm_swappiness;
1431 return memcg->swappiness;
1435 * memcg->moving_account is used for checking possibility that some thread is
1436 * calling move_account(). When a thread on CPU-A starts moving pages under
1437 * a memcg, other threads should check memcg->moving_account under
1438 * rcu_read_lock(), like this:
1442 * memcg->moving_account+1 if (memcg->mocing_account)
1444 * synchronize_rcu() update something.
1449 /* for quick checking without looking up memcg */
1450 atomic_t memcg_moving __read_mostly;
1452 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1454 atomic_inc(&memcg_moving);
1455 atomic_inc(&memcg->moving_account);
1459 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1462 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1463 * We check NULL in callee rather than caller.
1466 atomic_dec(&memcg_moving);
1467 atomic_dec(&memcg->moving_account);
1472 * 2 routines for checking "mem" is under move_account() or not.
1474 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1475 * is used for avoiding races in accounting. If true,
1476 * pc->mem_cgroup may be overwritten.
1478 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1479 * under hierarchy of moving cgroups. This is for
1480 * waiting at hith-memory prressure caused by "move".
1483 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1485 VM_BUG_ON(!rcu_read_lock_held());
1486 return atomic_read(&memcg->moving_account) > 0;
1489 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1491 struct mem_cgroup *from;
1492 struct mem_cgroup *to;
1495 * Unlike task_move routines, we access mc.to, mc.from not under
1496 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1498 spin_lock(&mc.lock);
1504 ret = mem_cgroup_same_or_subtree(memcg, from)
1505 || mem_cgroup_same_or_subtree(memcg, to);
1507 spin_unlock(&mc.lock);
1511 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1513 if (mc.moving_task && current != mc.moving_task) {
1514 if (mem_cgroup_under_move(memcg)) {
1516 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1517 /* moving charge context might have finished. */
1520 finish_wait(&mc.waitq, &wait);
1528 * Take this lock when
1529 * - a code tries to modify page's memcg while it's USED.
1530 * - a code tries to modify page state accounting in a memcg.
1531 * see mem_cgroup_stolen(), too.
1533 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1534 unsigned long *flags)
1536 spin_lock_irqsave(&memcg->move_lock, *flags);
1539 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1540 unsigned long *flags)
1542 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1545 #define K(x) ((x) << (PAGE_SHIFT-10))
1547 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1548 * @memcg: The memory cgroup that went over limit
1549 * @p: Task that is going to be killed
1551 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1554 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1556 struct cgroup *task_cgrp;
1557 struct cgroup *mem_cgrp;
1559 * Need a buffer in BSS, can't rely on allocations. The code relies
1560 * on the assumption that OOM is serialized for memory controller.
1561 * If this assumption is broken, revisit this code.
1563 static char memcg_name[PATH_MAX];
1565 struct mem_cgroup *iter;
1573 mem_cgrp = memcg->css.cgroup;
1574 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1576 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1579 * Unfortunately, we are unable to convert to a useful name
1580 * But we'll still print out the usage information
1587 pr_info("Task in %s killed", memcg_name);
1590 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1598 * Continues from above, so we don't need an KERN_ level
1600 pr_cont(" as a result of limit of %s\n", memcg_name);
1603 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1604 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1605 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1606 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1607 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1608 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1609 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1610 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1611 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1612 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1613 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1614 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1616 for_each_mem_cgroup_tree(iter, memcg) {
1617 pr_info("Memory cgroup stats");
1620 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1622 pr_cont(" for %s", memcg_name);
1626 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1627 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1629 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1630 K(mem_cgroup_read_stat(iter, i)));
1633 for (i = 0; i < NR_LRU_LISTS; i++)
1634 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1635 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1642 * This function returns the number of memcg under hierarchy tree. Returns
1643 * 1(self count) if no children.
1645 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1648 struct mem_cgroup *iter;
1650 for_each_mem_cgroup_tree(iter, memcg)
1656 * Return the memory (and swap, if configured) limit for a memcg.
1658 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1662 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1665 * Do not consider swap space if we cannot swap due to swappiness
1667 if (mem_cgroup_swappiness(memcg)) {
1670 limit += total_swap_pages << PAGE_SHIFT;
1671 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1674 * If memsw is finite and limits the amount of swap space
1675 * available to this memcg, return that limit.
1677 limit = min(limit, memsw);
1683 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1686 struct mem_cgroup *iter;
1687 unsigned long chosen_points = 0;
1688 unsigned long totalpages;
1689 unsigned int points = 0;
1690 struct task_struct *chosen = NULL;
1693 * If current has a pending SIGKILL, then automatically select it. The
1694 * goal is to allow it to allocate so that it may quickly exit and free
1697 if (fatal_signal_pending(current)) {
1698 set_thread_flag(TIF_MEMDIE);
1702 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1703 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1704 for_each_mem_cgroup_tree(iter, memcg) {
1705 struct cgroup *cgroup = iter->css.cgroup;
1706 struct cgroup_iter it;
1707 struct task_struct *task;
1709 cgroup_iter_start(cgroup, &it);
1710 while ((task = cgroup_iter_next(cgroup, &it))) {
1711 switch (oom_scan_process_thread(task, totalpages, NULL,
1713 case OOM_SCAN_SELECT:
1715 put_task_struct(chosen);
1717 chosen_points = ULONG_MAX;
1718 get_task_struct(chosen);
1720 case OOM_SCAN_CONTINUE:
1722 case OOM_SCAN_ABORT:
1723 cgroup_iter_end(cgroup, &it);
1724 mem_cgroup_iter_break(memcg, iter);
1726 put_task_struct(chosen);
1731 points = oom_badness(task, memcg, NULL, totalpages);
1732 if (points > chosen_points) {
1734 put_task_struct(chosen);
1736 chosen_points = points;
1737 get_task_struct(chosen);
1740 cgroup_iter_end(cgroup, &it);
1745 points = chosen_points * 1000 / totalpages;
1746 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1747 NULL, "Memory cgroup out of memory");
1750 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1752 unsigned long flags)
1754 unsigned long total = 0;
1755 bool noswap = false;
1758 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1760 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1763 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1765 drain_all_stock_async(memcg);
1766 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1768 * Allow limit shrinkers, which are triggered directly
1769 * by userspace, to catch signals and stop reclaim
1770 * after minimal progress, regardless of the margin.
1772 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1774 if (mem_cgroup_margin(memcg))
1777 * If nothing was reclaimed after two attempts, there
1778 * may be no reclaimable pages in this hierarchy.
1787 * test_mem_cgroup_node_reclaimable
1788 * @memcg: the target memcg
1789 * @nid: the node ID to be checked.
1790 * @noswap : specify true here if the user wants flle only information.
1792 * This function returns whether the specified memcg contains any
1793 * reclaimable pages on a node. Returns true if there are any reclaimable
1794 * pages in the node.
1796 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1797 int nid, bool noswap)
1799 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1801 if (noswap || !total_swap_pages)
1803 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1808 #if MAX_NUMNODES > 1
1811 * Always updating the nodemask is not very good - even if we have an empty
1812 * list or the wrong list here, we can start from some node and traverse all
1813 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1816 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1820 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1821 * pagein/pageout changes since the last update.
1823 if (!atomic_read(&memcg->numainfo_events))
1825 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1828 /* make a nodemask where this memcg uses memory from */
1829 memcg->scan_nodes = node_states[N_MEMORY];
1831 for_each_node_mask(nid, node_states[N_MEMORY]) {
1833 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1834 node_clear(nid, memcg->scan_nodes);
1837 atomic_set(&memcg->numainfo_events, 0);
1838 atomic_set(&memcg->numainfo_updating, 0);
1842 * Selecting a node where we start reclaim from. Because what we need is just
1843 * reducing usage counter, start from anywhere is O,K. Considering
1844 * memory reclaim from current node, there are pros. and cons.
1846 * Freeing memory from current node means freeing memory from a node which
1847 * we'll use or we've used. So, it may make LRU bad. And if several threads
1848 * hit limits, it will see a contention on a node. But freeing from remote
1849 * node means more costs for memory reclaim because of memory latency.
1851 * Now, we use round-robin. Better algorithm is welcomed.
1853 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1857 mem_cgroup_may_update_nodemask(memcg);
1858 node = memcg->last_scanned_node;
1860 node = next_node(node, memcg->scan_nodes);
1861 if (node == MAX_NUMNODES)
1862 node = first_node(memcg->scan_nodes);
1864 * We call this when we hit limit, not when pages are added to LRU.
1865 * No LRU may hold pages because all pages are UNEVICTABLE or
1866 * memcg is too small and all pages are not on LRU. In that case,
1867 * we use curret node.
1869 if (unlikely(node == MAX_NUMNODES))
1870 node = numa_node_id();
1872 memcg->last_scanned_node = node;
1877 * Check all nodes whether it contains reclaimable pages or not.
1878 * For quick scan, we make use of scan_nodes. This will allow us to skip
1879 * unused nodes. But scan_nodes is lazily updated and may not cotain
1880 * enough new information. We need to do double check.
1882 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1887 * quick check...making use of scan_node.
1888 * We can skip unused nodes.
1890 if (!nodes_empty(memcg->scan_nodes)) {
1891 for (nid = first_node(memcg->scan_nodes);
1893 nid = next_node(nid, memcg->scan_nodes)) {
1895 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1900 * Check rest of nodes.
1902 for_each_node_state(nid, N_MEMORY) {
1903 if (node_isset(nid, memcg->scan_nodes))
1905 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1912 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1917 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1919 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1923 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1926 unsigned long *total_scanned)
1928 struct mem_cgroup *victim = NULL;
1931 unsigned long excess;
1932 unsigned long nr_scanned;
1933 struct mem_cgroup_reclaim_cookie reclaim = {
1938 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1941 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1946 * If we have not been able to reclaim
1947 * anything, it might because there are
1948 * no reclaimable pages under this hierarchy
1953 * We want to do more targeted reclaim.
1954 * excess >> 2 is not to excessive so as to
1955 * reclaim too much, nor too less that we keep
1956 * coming back to reclaim from this cgroup
1958 if (total >= (excess >> 2) ||
1959 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1964 if (!mem_cgroup_reclaimable(victim, false))
1966 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1968 *total_scanned += nr_scanned;
1969 if (!res_counter_soft_limit_excess(&root_memcg->res))
1972 mem_cgroup_iter_break(root_memcg, victim);
1977 * Check OOM-Killer is already running under our hierarchy.
1978 * If someone is running, return false.
1979 * Has to be called with memcg_oom_lock
1981 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1983 struct mem_cgroup *iter, *failed = NULL;
1985 for_each_mem_cgroup_tree(iter, memcg) {
1986 if (iter->oom_lock) {
1988 * this subtree of our hierarchy is already locked
1989 * so we cannot give a lock.
1992 mem_cgroup_iter_break(memcg, iter);
1995 iter->oom_lock = true;
2002 * OK, we failed to lock the whole subtree so we have to clean up
2003 * what we set up to the failing subtree
2005 for_each_mem_cgroup_tree(iter, memcg) {
2006 if (iter == failed) {
2007 mem_cgroup_iter_break(memcg, iter);
2010 iter->oom_lock = false;
2016 * Has to be called with memcg_oom_lock
2018 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2020 struct mem_cgroup *iter;
2022 for_each_mem_cgroup_tree(iter, memcg)
2023 iter->oom_lock = false;
2027 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2029 struct mem_cgroup *iter;
2031 for_each_mem_cgroup_tree(iter, memcg)
2032 atomic_inc(&iter->under_oom);
2035 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2037 struct mem_cgroup *iter;
2040 * When a new child is created while the hierarchy is under oom,
2041 * mem_cgroup_oom_lock() may not be called. We have to use
2042 * atomic_add_unless() here.
2044 for_each_mem_cgroup_tree(iter, memcg)
2045 atomic_add_unless(&iter->under_oom, -1, 0);
2048 static DEFINE_SPINLOCK(memcg_oom_lock);
2049 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2051 struct oom_wait_info {
2052 struct mem_cgroup *memcg;
2056 static int memcg_oom_wake_function(wait_queue_t *wait,
2057 unsigned mode, int sync, void *arg)
2059 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2060 struct mem_cgroup *oom_wait_memcg;
2061 struct oom_wait_info *oom_wait_info;
2063 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2064 oom_wait_memcg = oom_wait_info->memcg;
2067 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2068 * Then we can use css_is_ancestor without taking care of RCU.
2070 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2071 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2073 return autoremove_wake_function(wait, mode, sync, arg);
2076 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2078 /* for filtering, pass "memcg" as argument. */
2079 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2082 static void memcg_oom_recover(struct mem_cgroup *memcg)
2084 if (memcg && atomic_read(&memcg->under_oom))
2085 memcg_wakeup_oom(memcg);
2089 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2091 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2094 struct oom_wait_info owait;
2095 bool locked, need_to_kill;
2097 owait.memcg = memcg;
2098 owait.wait.flags = 0;
2099 owait.wait.func = memcg_oom_wake_function;
2100 owait.wait.private = current;
2101 INIT_LIST_HEAD(&owait.wait.task_list);
2102 need_to_kill = true;
2103 mem_cgroup_mark_under_oom(memcg);
2105 /* At first, try to OOM lock hierarchy under memcg.*/
2106 spin_lock(&memcg_oom_lock);
2107 locked = mem_cgroup_oom_lock(memcg);
2109 * Even if signal_pending(), we can't quit charge() loop without
2110 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2111 * under OOM is always welcomed, use TASK_KILLABLE here.
2113 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2114 if (!locked || memcg->oom_kill_disable)
2115 need_to_kill = false;
2117 mem_cgroup_oom_notify(memcg);
2118 spin_unlock(&memcg_oom_lock);
2121 finish_wait(&memcg_oom_waitq, &owait.wait);
2122 mem_cgroup_out_of_memory(memcg, mask, order);
2125 finish_wait(&memcg_oom_waitq, &owait.wait);
2127 spin_lock(&memcg_oom_lock);
2129 mem_cgroup_oom_unlock(memcg);
2130 memcg_wakeup_oom(memcg);
2131 spin_unlock(&memcg_oom_lock);
2133 mem_cgroup_unmark_under_oom(memcg);
2135 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2137 /* Give chance to dying process */
2138 schedule_timeout_uninterruptible(1);
2143 * Currently used to update mapped file statistics, but the routine can be
2144 * generalized to update other statistics as well.
2146 * Notes: Race condition
2148 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2149 * it tends to be costly. But considering some conditions, we doesn't need
2150 * to do so _always_.
2152 * Considering "charge", lock_page_cgroup() is not required because all
2153 * file-stat operations happen after a page is attached to radix-tree. There
2154 * are no race with "charge".
2156 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2157 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2158 * if there are race with "uncharge". Statistics itself is properly handled
2161 * Considering "move", this is an only case we see a race. To make the race
2162 * small, we check mm->moving_account and detect there are possibility of race
2163 * If there is, we take a lock.
2166 void __mem_cgroup_begin_update_page_stat(struct page *page,
2167 bool *locked, unsigned long *flags)
2169 struct mem_cgroup *memcg;
2170 struct page_cgroup *pc;
2172 pc = lookup_page_cgroup(page);
2174 memcg = pc->mem_cgroup;
2175 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2178 * If this memory cgroup is not under account moving, we don't
2179 * need to take move_lock_mem_cgroup(). Because we already hold
2180 * rcu_read_lock(), any calls to move_account will be delayed until
2181 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2183 if (!mem_cgroup_stolen(memcg))
2186 move_lock_mem_cgroup(memcg, flags);
2187 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2188 move_unlock_mem_cgroup(memcg, flags);
2194 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2196 struct page_cgroup *pc = lookup_page_cgroup(page);
2199 * It's guaranteed that pc->mem_cgroup never changes while
2200 * lock is held because a routine modifies pc->mem_cgroup
2201 * should take move_lock_mem_cgroup().
2203 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2206 void mem_cgroup_update_page_stat(struct page *page,
2207 enum mem_cgroup_page_stat_item idx, int val)
2209 struct mem_cgroup *memcg;
2210 struct page_cgroup *pc = lookup_page_cgroup(page);
2211 unsigned long uninitialized_var(flags);
2213 if (mem_cgroup_disabled())
2216 memcg = pc->mem_cgroup;
2217 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2221 case MEMCG_NR_FILE_MAPPED:
2222 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2228 this_cpu_add(memcg->stat->count[idx], val);
2232 * size of first charge trial. "32" comes from vmscan.c's magic value.
2233 * TODO: maybe necessary to use big numbers in big irons.
2235 #define CHARGE_BATCH 32U
2236 struct memcg_stock_pcp {
2237 struct mem_cgroup *cached; /* this never be root cgroup */
2238 unsigned int nr_pages;
2239 struct work_struct work;
2240 unsigned long flags;
2241 #define FLUSHING_CACHED_CHARGE 0
2243 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2244 static DEFINE_MUTEX(percpu_charge_mutex);
2247 * consume_stock: Try to consume stocked charge on this cpu.
2248 * @memcg: memcg to consume from.
2249 * @nr_pages: how many pages to charge.
2251 * The charges will only happen if @memcg matches the current cpu's memcg
2252 * stock, and at least @nr_pages are available in that stock. Failure to
2253 * service an allocation will refill the stock.
2255 * returns true if successful, false otherwise.
2257 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2259 struct memcg_stock_pcp *stock;
2262 if (nr_pages > CHARGE_BATCH)
2265 stock = &get_cpu_var(memcg_stock);
2266 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2267 stock->nr_pages -= nr_pages;
2268 else /* need to call res_counter_charge */
2270 put_cpu_var(memcg_stock);
2275 * Returns stocks cached in percpu to res_counter and reset cached information.
2277 static void drain_stock(struct memcg_stock_pcp *stock)
2279 struct mem_cgroup *old = stock->cached;
2281 if (stock->nr_pages) {
2282 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2284 res_counter_uncharge(&old->res, bytes);
2285 if (do_swap_account)
2286 res_counter_uncharge(&old->memsw, bytes);
2287 stock->nr_pages = 0;
2289 stock->cached = NULL;
2293 * This must be called under preempt disabled or must be called by
2294 * a thread which is pinned to local cpu.
2296 static void drain_local_stock(struct work_struct *dummy)
2298 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2300 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2304 * Cache charges(val) which is from res_counter, to local per_cpu area.
2305 * This will be consumed by consume_stock() function, later.
2307 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2309 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2311 if (stock->cached != memcg) { /* reset if necessary */
2313 stock->cached = memcg;
2315 stock->nr_pages += nr_pages;
2316 put_cpu_var(memcg_stock);
2320 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2321 * of the hierarchy under it. sync flag says whether we should block
2322 * until the work is done.
2324 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2328 /* Notify other cpus that system-wide "drain" is running */
2331 for_each_online_cpu(cpu) {
2332 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2333 struct mem_cgroup *memcg;
2335 memcg = stock->cached;
2336 if (!memcg || !stock->nr_pages)
2338 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2340 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2342 drain_local_stock(&stock->work);
2344 schedule_work_on(cpu, &stock->work);
2352 for_each_online_cpu(cpu) {
2353 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2354 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2355 flush_work(&stock->work);
2362 * Tries to drain stocked charges in other cpus. This function is asynchronous
2363 * and just put a work per cpu for draining localy on each cpu. Caller can
2364 * expects some charges will be back to res_counter later but cannot wait for
2367 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2370 * If someone calls draining, avoid adding more kworker runs.
2372 if (!mutex_trylock(&percpu_charge_mutex))
2374 drain_all_stock(root_memcg, false);
2375 mutex_unlock(&percpu_charge_mutex);
2378 /* This is a synchronous drain interface. */
2379 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2381 /* called when force_empty is called */
2382 mutex_lock(&percpu_charge_mutex);
2383 drain_all_stock(root_memcg, true);
2384 mutex_unlock(&percpu_charge_mutex);
2388 * This function drains percpu counter value from DEAD cpu and
2389 * move it to local cpu. Note that this function can be preempted.
2391 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2395 spin_lock(&memcg->pcp_counter_lock);
2396 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2397 long x = per_cpu(memcg->stat->count[i], cpu);
2399 per_cpu(memcg->stat->count[i], cpu) = 0;
2400 memcg->nocpu_base.count[i] += x;
2402 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2403 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2405 per_cpu(memcg->stat->events[i], cpu) = 0;
2406 memcg->nocpu_base.events[i] += x;
2408 spin_unlock(&memcg->pcp_counter_lock);
2411 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2412 unsigned long action,
2415 int cpu = (unsigned long)hcpu;
2416 struct memcg_stock_pcp *stock;
2417 struct mem_cgroup *iter;
2419 if (action == CPU_ONLINE)
2422 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2425 for_each_mem_cgroup(iter)
2426 mem_cgroup_drain_pcp_counter(iter, cpu);
2428 stock = &per_cpu(memcg_stock, cpu);
2434 /* See __mem_cgroup_try_charge() for details */
2436 CHARGE_OK, /* success */
2437 CHARGE_RETRY, /* need to retry but retry is not bad */
2438 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2439 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2440 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2443 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2444 unsigned int nr_pages, unsigned int min_pages,
2447 unsigned long csize = nr_pages * PAGE_SIZE;
2448 struct mem_cgroup *mem_over_limit;
2449 struct res_counter *fail_res;
2450 unsigned long flags = 0;
2453 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2456 if (!do_swap_account)
2458 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2462 res_counter_uncharge(&memcg->res, csize);
2463 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2464 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2466 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2468 * Never reclaim on behalf of optional batching, retry with a
2469 * single page instead.
2471 if (nr_pages > min_pages)
2472 return CHARGE_RETRY;
2474 if (!(gfp_mask & __GFP_WAIT))
2475 return CHARGE_WOULDBLOCK;
2477 if (gfp_mask & __GFP_NORETRY)
2478 return CHARGE_NOMEM;
2480 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2481 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2482 return CHARGE_RETRY;
2484 * Even though the limit is exceeded at this point, reclaim
2485 * may have been able to free some pages. Retry the charge
2486 * before killing the task.
2488 * Only for regular pages, though: huge pages are rather
2489 * unlikely to succeed so close to the limit, and we fall back
2490 * to regular pages anyway in case of failure.
2492 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2493 return CHARGE_RETRY;
2496 * At task move, charge accounts can be doubly counted. So, it's
2497 * better to wait until the end of task_move if something is going on.
2499 if (mem_cgroup_wait_acct_move(mem_over_limit))
2500 return CHARGE_RETRY;
2502 /* If we don't need to call oom-killer at el, return immediately */
2504 return CHARGE_NOMEM;
2506 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2507 return CHARGE_OOM_DIE;
2509 return CHARGE_RETRY;
2513 * __mem_cgroup_try_charge() does
2514 * 1. detect memcg to be charged against from passed *mm and *ptr,
2515 * 2. update res_counter
2516 * 3. call memory reclaim if necessary.
2518 * In some special case, if the task is fatal, fatal_signal_pending() or
2519 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2520 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2521 * as possible without any hazards. 2: all pages should have a valid
2522 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2523 * pointer, that is treated as a charge to root_mem_cgroup.
2525 * So __mem_cgroup_try_charge() will return
2526 * 0 ... on success, filling *ptr with a valid memcg pointer.
2527 * -ENOMEM ... charge failure because of resource limits.
2528 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2530 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2531 * the oom-killer can be invoked.
2533 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2535 unsigned int nr_pages,
2536 struct mem_cgroup **ptr,
2539 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2540 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2541 struct mem_cgroup *memcg = NULL;
2545 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2546 * in system level. So, allow to go ahead dying process in addition to
2549 if (unlikely(test_thread_flag(TIF_MEMDIE)
2550 || fatal_signal_pending(current)))
2554 * We always charge the cgroup the mm_struct belongs to.
2555 * The mm_struct's mem_cgroup changes on task migration if the
2556 * thread group leader migrates. It's possible that mm is not
2557 * set, if so charge the root memcg (happens for pagecache usage).
2560 *ptr = root_mem_cgroup;
2562 if (*ptr) { /* css should be a valid one */
2564 if (mem_cgroup_is_root(memcg))
2566 if (consume_stock(memcg, nr_pages))
2568 css_get(&memcg->css);
2570 struct task_struct *p;
2573 p = rcu_dereference(mm->owner);
2575 * Because we don't have task_lock(), "p" can exit.
2576 * In that case, "memcg" can point to root or p can be NULL with
2577 * race with swapoff. Then, we have small risk of mis-accouning.
2578 * But such kind of mis-account by race always happens because
2579 * we don't have cgroup_mutex(). It's overkill and we allo that
2581 * (*) swapoff at el will charge against mm-struct not against
2582 * task-struct. So, mm->owner can be NULL.
2584 memcg = mem_cgroup_from_task(p);
2586 memcg = root_mem_cgroup;
2587 if (mem_cgroup_is_root(memcg)) {
2591 if (consume_stock(memcg, nr_pages)) {
2593 * It seems dagerous to access memcg without css_get().
2594 * But considering how consume_stok works, it's not
2595 * necessary. If consume_stock success, some charges
2596 * from this memcg are cached on this cpu. So, we
2597 * don't need to call css_get()/css_tryget() before
2598 * calling consume_stock().
2603 /* after here, we may be blocked. we need to get refcnt */
2604 if (!css_tryget(&memcg->css)) {
2614 /* If killed, bypass charge */
2615 if (fatal_signal_pending(current)) {
2616 css_put(&memcg->css);
2621 if (oom && !nr_oom_retries) {
2623 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2626 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2631 case CHARGE_RETRY: /* not in OOM situation but retry */
2633 css_put(&memcg->css);
2636 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2637 css_put(&memcg->css);
2639 case CHARGE_NOMEM: /* OOM routine works */
2641 css_put(&memcg->css);
2644 /* If oom, we never return -ENOMEM */
2647 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2648 css_put(&memcg->css);
2651 } while (ret != CHARGE_OK);
2653 if (batch > nr_pages)
2654 refill_stock(memcg, batch - nr_pages);
2655 css_put(&memcg->css);
2663 *ptr = root_mem_cgroup;
2668 * Somemtimes we have to undo a charge we got by try_charge().
2669 * This function is for that and do uncharge, put css's refcnt.
2670 * gotten by try_charge().
2672 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2673 unsigned int nr_pages)
2675 if (!mem_cgroup_is_root(memcg)) {
2676 unsigned long bytes = nr_pages * PAGE_SIZE;
2678 res_counter_uncharge(&memcg->res, bytes);
2679 if (do_swap_account)
2680 res_counter_uncharge(&memcg->memsw, bytes);
2685 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2686 * This is useful when moving usage to parent cgroup.
2688 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2689 unsigned int nr_pages)
2691 unsigned long bytes = nr_pages * PAGE_SIZE;
2693 if (mem_cgroup_is_root(memcg))
2696 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2697 if (do_swap_account)
2698 res_counter_uncharge_until(&memcg->memsw,
2699 memcg->memsw.parent, bytes);
2703 * A helper function to get mem_cgroup from ID. must be called under
2704 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2705 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2706 * called against removed memcg.)
2708 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2710 struct cgroup_subsys_state *css;
2712 /* ID 0 is unused ID */
2715 css = css_lookup(&mem_cgroup_subsys, id);
2718 return mem_cgroup_from_css(css);
2721 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2723 struct mem_cgroup *memcg = NULL;
2724 struct page_cgroup *pc;
2728 VM_BUG_ON(!PageLocked(page));
2730 pc = lookup_page_cgroup(page);
2731 lock_page_cgroup(pc);
2732 if (PageCgroupUsed(pc)) {
2733 memcg = pc->mem_cgroup;
2734 if (memcg && !css_tryget(&memcg->css))
2736 } else if (PageSwapCache(page)) {
2737 ent.val = page_private(page);
2738 id = lookup_swap_cgroup_id(ent);
2740 memcg = mem_cgroup_lookup(id);
2741 if (memcg && !css_tryget(&memcg->css))
2745 unlock_page_cgroup(pc);
2749 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2751 unsigned int nr_pages,
2752 enum charge_type ctype,
2755 struct page_cgroup *pc = lookup_page_cgroup(page);
2756 struct zone *uninitialized_var(zone);
2757 struct lruvec *lruvec;
2758 bool was_on_lru = false;
2761 lock_page_cgroup(pc);
2762 VM_BUG_ON(PageCgroupUsed(pc));
2764 * we don't need page_cgroup_lock about tail pages, becase they are not
2765 * accessed by any other context at this point.
2769 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2770 * may already be on some other mem_cgroup's LRU. Take care of it.
2773 zone = page_zone(page);
2774 spin_lock_irq(&zone->lru_lock);
2775 if (PageLRU(page)) {
2776 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2778 del_page_from_lru_list(page, lruvec, page_lru(page));
2783 pc->mem_cgroup = memcg;
2785 * We access a page_cgroup asynchronously without lock_page_cgroup().
2786 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2787 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2788 * before USED bit, we need memory barrier here.
2789 * See mem_cgroup_add_lru_list(), etc.
2792 SetPageCgroupUsed(pc);
2796 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2797 VM_BUG_ON(PageLRU(page));
2799 add_page_to_lru_list(page, lruvec, page_lru(page));
2801 spin_unlock_irq(&zone->lru_lock);
2804 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2809 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2810 unlock_page_cgroup(pc);
2813 * "charge_statistics" updated event counter. Then, check it.
2814 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2815 * if they exceeds softlimit.
2817 memcg_check_events(memcg, page);
2820 static DEFINE_MUTEX(set_limit_mutex);
2822 #ifdef CONFIG_MEMCG_KMEM
2823 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2825 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2826 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2830 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2831 * in the memcg_cache_params struct.
2833 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2835 struct kmem_cache *cachep;
2837 VM_BUG_ON(p->is_root_cache);
2838 cachep = p->root_cache;
2839 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2842 #ifdef CONFIG_SLABINFO
2843 static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2846 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2847 struct memcg_cache_params *params;
2849 if (!memcg_can_account_kmem(memcg))
2852 print_slabinfo_header(m);
2854 mutex_lock(&memcg->slab_caches_mutex);
2855 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2856 cache_show(memcg_params_to_cache(params), m);
2857 mutex_unlock(&memcg->slab_caches_mutex);
2863 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2865 struct res_counter *fail_res;
2866 struct mem_cgroup *_memcg;
2870 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2875 * Conditions under which we can wait for the oom_killer. Those are
2876 * the same conditions tested by the core page allocator
2878 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2881 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2884 if (ret == -EINTR) {
2886 * __mem_cgroup_try_charge() chosed to bypass to root due to
2887 * OOM kill or fatal signal. Since our only options are to
2888 * either fail the allocation or charge it to this cgroup, do
2889 * it as a temporary condition. But we can't fail. From a
2890 * kmem/slab perspective, the cache has already been selected,
2891 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2894 * This condition will only trigger if the task entered
2895 * memcg_charge_kmem in a sane state, but was OOM-killed during
2896 * __mem_cgroup_try_charge() above. Tasks that were already
2897 * dying when the allocation triggers should have been already
2898 * directed to the root cgroup in memcontrol.h
2900 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2901 if (do_swap_account)
2902 res_counter_charge_nofail(&memcg->memsw, size,
2906 res_counter_uncharge(&memcg->kmem, size);
2911 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2913 res_counter_uncharge(&memcg->res, size);
2914 if (do_swap_account)
2915 res_counter_uncharge(&memcg->memsw, size);
2918 if (res_counter_uncharge(&memcg->kmem, size))
2921 if (memcg_kmem_test_and_clear_dead(memcg))
2922 mem_cgroup_put(memcg);
2925 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2930 mutex_lock(&memcg->slab_caches_mutex);
2931 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2932 mutex_unlock(&memcg->slab_caches_mutex);
2936 * helper for acessing a memcg's index. It will be used as an index in the
2937 * child cache array in kmem_cache, and also to derive its name. This function
2938 * will return -1 when this is not a kmem-limited memcg.
2940 int memcg_cache_id(struct mem_cgroup *memcg)
2942 return memcg ? memcg->kmemcg_id : -1;
2946 * This ends up being protected by the set_limit mutex, during normal
2947 * operation, because that is its main call site.
2949 * But when we create a new cache, we can call this as well if its parent
2950 * is kmem-limited. That will have to hold set_limit_mutex as well.
2952 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2956 num = ida_simple_get(&kmem_limited_groups,
2957 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2961 * After this point, kmem_accounted (that we test atomically in
2962 * the beginning of this conditional), is no longer 0. This
2963 * guarantees only one process will set the following boolean
2964 * to true. We don't need test_and_set because we're protected
2965 * by the set_limit_mutex anyway.
2967 memcg_kmem_set_activated(memcg);
2969 ret = memcg_update_all_caches(num+1);
2971 ida_simple_remove(&kmem_limited_groups, num);
2972 memcg_kmem_clear_activated(memcg);
2976 memcg->kmemcg_id = num;
2977 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2978 mutex_init(&memcg->slab_caches_mutex);
2982 static size_t memcg_caches_array_size(int num_groups)
2985 if (num_groups <= 0)
2988 size = 2 * num_groups;
2989 if (size < MEMCG_CACHES_MIN_SIZE)
2990 size = MEMCG_CACHES_MIN_SIZE;
2991 else if (size > MEMCG_CACHES_MAX_SIZE)
2992 size = MEMCG_CACHES_MAX_SIZE;
2998 * We should update the current array size iff all caches updates succeed. This
2999 * can only be done from the slab side. The slab mutex needs to be held when
3002 void memcg_update_array_size(int num)
3004 if (num > memcg_limited_groups_array_size)
3005 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3008 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3010 struct memcg_cache_params *cur_params = s->memcg_params;
3012 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3014 if (num_groups > memcg_limited_groups_array_size) {
3016 ssize_t size = memcg_caches_array_size(num_groups);
3018 size *= sizeof(void *);
3019 size += sizeof(struct memcg_cache_params);
3021 s->memcg_params = kzalloc(size, GFP_KERNEL);
3022 if (!s->memcg_params) {
3023 s->memcg_params = cur_params;
3027 s->memcg_params->is_root_cache = true;
3030 * There is the chance it will be bigger than
3031 * memcg_limited_groups_array_size, if we failed an allocation
3032 * in a cache, in which case all caches updated before it, will
3033 * have a bigger array.
3035 * But if that is the case, the data after
3036 * memcg_limited_groups_array_size is certainly unused
3038 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3039 if (!cur_params->memcg_caches[i])
3041 s->memcg_params->memcg_caches[i] =
3042 cur_params->memcg_caches[i];
3046 * Ideally, we would wait until all caches succeed, and only
3047 * then free the old one. But this is not worth the extra
3048 * pointer per-cache we'd have to have for this.
3050 * It is not a big deal if some caches are left with a size
3051 * bigger than the others. And all updates will reset this
3059 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3060 struct kmem_cache *root_cache)
3062 size_t size = sizeof(struct memcg_cache_params);
3064 if (!memcg_kmem_enabled())
3068 size += memcg_limited_groups_array_size * sizeof(void *);
3070 s->memcg_params = kzalloc(size, GFP_KERNEL);
3071 if (!s->memcg_params)
3075 s->memcg_params->memcg = memcg;
3076 s->memcg_params->root_cache = root_cache;
3078 s->memcg_params->is_root_cache = true;
3083 void memcg_release_cache(struct kmem_cache *s)
3085 struct kmem_cache *root;
3086 struct mem_cgroup *memcg;
3090 * This happens, for instance, when a root cache goes away before we
3093 if (!s->memcg_params)
3096 if (s->memcg_params->is_root_cache)
3099 memcg = s->memcg_params->memcg;
3100 id = memcg_cache_id(memcg);
3102 root = s->memcg_params->root_cache;
3103 root->memcg_params->memcg_caches[id] = NULL;
3104 mem_cgroup_put(memcg);
3106 mutex_lock(&memcg->slab_caches_mutex);
3107 list_del(&s->memcg_params->list);
3108 mutex_unlock(&memcg->slab_caches_mutex);
3111 kfree(s->memcg_params);
3115 * During the creation a new cache, we need to disable our accounting mechanism
3116 * altogether. This is true even if we are not creating, but rather just
3117 * enqueing new caches to be created.
3119 * This is because that process will trigger allocations; some visible, like
3120 * explicit kmallocs to auxiliary data structures, name strings and internal
3121 * cache structures; some well concealed, like INIT_WORK() that can allocate
3122 * objects during debug.
3124 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3125 * to it. This may not be a bounded recursion: since the first cache creation
3126 * failed to complete (waiting on the allocation), we'll just try to create the
3127 * cache again, failing at the same point.
3129 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3130 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3131 * inside the following two functions.
3133 static inline void memcg_stop_kmem_account(void)
3135 VM_BUG_ON(!current->mm);
3136 current->memcg_kmem_skip_account++;
3139 static inline void memcg_resume_kmem_account(void)
3141 VM_BUG_ON(!current->mm);
3142 current->memcg_kmem_skip_account--;
3145 static void kmem_cache_destroy_work_func(struct work_struct *w)
3147 struct kmem_cache *cachep;
3148 struct memcg_cache_params *p;
3150 p = container_of(w, struct memcg_cache_params, destroy);
3152 cachep = memcg_params_to_cache(p);
3155 * If we get down to 0 after shrink, we could delete right away.
3156 * However, memcg_release_pages() already puts us back in the workqueue
3157 * in that case. If we proceed deleting, we'll get a dangling
3158 * reference, and removing the object from the workqueue in that case
3159 * is unnecessary complication. We are not a fast path.
3161 * Note that this case is fundamentally different from racing with
3162 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3163 * kmem_cache_shrink, not only we would be reinserting a dead cache
3164 * into the queue, but doing so from inside the worker racing to
3167 * So if we aren't down to zero, we'll just schedule a worker and try
3170 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3171 kmem_cache_shrink(cachep);
3172 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3175 kmem_cache_destroy(cachep);
3178 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3180 if (!cachep->memcg_params->dead)
3184 * There are many ways in which we can get here.
3186 * We can get to a memory-pressure situation while the delayed work is
3187 * still pending to run. The vmscan shrinkers can then release all
3188 * cache memory and get us to destruction. If this is the case, we'll
3189 * be executed twice, which is a bug (the second time will execute over
3190 * bogus data). In this case, cancelling the work should be fine.
3192 * But we can also get here from the worker itself, if
3193 * kmem_cache_shrink is enough to shake all the remaining objects and
3194 * get the page count to 0. In this case, we'll deadlock if we try to
3195 * cancel the work (the worker runs with an internal lock held, which
3196 * is the same lock we would hold for cancel_work_sync().)
3198 * Since we can't possibly know who got us here, just refrain from
3199 * running if there is already work pending
3201 if (work_pending(&cachep->memcg_params->destroy))
3204 * We have to defer the actual destroying to a workqueue, because
3205 * we might currently be in a context that cannot sleep.
3207 schedule_work(&cachep->memcg_params->destroy);
3210 static char *memcg_cache_name(struct mem_cgroup *memcg, struct kmem_cache *s)
3213 struct dentry *dentry;
3216 dentry = rcu_dereference(memcg->css.cgroup->dentry);
3219 BUG_ON(dentry == NULL);
3221 name = kasprintf(GFP_KERNEL, "%s(%d:%s)", s->name,
3222 memcg_cache_id(memcg), dentry->d_name.name);
3227 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3228 struct kmem_cache *s)
3231 struct kmem_cache *new;
3233 name = memcg_cache_name(memcg, s);
3237 new = kmem_cache_create_memcg(memcg, name, s->object_size, s->align,
3238 (s->flags & ~SLAB_PANIC), s->ctor, s);
3241 new->allocflags |= __GFP_KMEMCG;
3248 * This lock protects updaters, not readers. We want readers to be as fast as
3249 * they can, and they will either see NULL or a valid cache value. Our model
3250 * allow them to see NULL, in which case the root memcg will be selected.
3252 * We need this lock because multiple allocations to the same cache from a non
3253 * will span more than one worker. Only one of them can create the cache.
3255 static DEFINE_MUTEX(memcg_cache_mutex);
3256 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3257 struct kmem_cache *cachep)
3259 struct kmem_cache *new_cachep;
3262 BUG_ON(!memcg_can_account_kmem(memcg));
3264 idx = memcg_cache_id(memcg);
3266 mutex_lock(&memcg_cache_mutex);
3267 new_cachep = cachep->memcg_params->memcg_caches[idx];
3271 new_cachep = kmem_cache_dup(memcg, cachep);
3272 if (new_cachep == NULL) {
3273 new_cachep = cachep;
3277 mem_cgroup_get(memcg);
3278 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3280 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3282 * the readers won't lock, make sure everybody sees the updated value,
3283 * so they won't put stuff in the queue again for no reason
3287 mutex_unlock(&memcg_cache_mutex);
3291 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3293 struct kmem_cache *c;
3296 if (!s->memcg_params)
3298 if (!s->memcg_params->is_root_cache)
3302 * If the cache is being destroyed, we trust that there is no one else
3303 * requesting objects from it. Even if there are, the sanity checks in
3304 * kmem_cache_destroy should caught this ill-case.
3306 * Still, we don't want anyone else freeing memcg_caches under our
3307 * noses, which can happen if a new memcg comes to life. As usual,
3308 * we'll take the set_limit_mutex to protect ourselves against this.
3310 mutex_lock(&set_limit_mutex);
3311 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3312 c = s->memcg_params->memcg_caches[i];
3317 * We will now manually delete the caches, so to avoid races
3318 * we need to cancel all pending destruction workers and
3319 * proceed with destruction ourselves.
3321 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3322 * and that could spawn the workers again: it is likely that
3323 * the cache still have active pages until this very moment.
3324 * This would lead us back to mem_cgroup_destroy_cache.
3326 * But that will not execute at all if the "dead" flag is not
3327 * set, so flip it down to guarantee we are in control.
3329 c->memcg_params->dead = false;
3330 cancel_work_sync(&c->memcg_params->destroy);
3331 kmem_cache_destroy(c);
3333 mutex_unlock(&set_limit_mutex);
3336 struct create_work {
3337 struct mem_cgroup *memcg;
3338 struct kmem_cache *cachep;
3339 struct work_struct work;
3342 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3344 struct kmem_cache *cachep;
3345 struct memcg_cache_params *params;
3347 if (!memcg_kmem_is_active(memcg))
3350 mutex_lock(&memcg->slab_caches_mutex);
3351 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3352 cachep = memcg_params_to_cache(params);
3353 cachep->memcg_params->dead = true;
3354 INIT_WORK(&cachep->memcg_params->destroy,
3355 kmem_cache_destroy_work_func);
3356 schedule_work(&cachep->memcg_params->destroy);
3358 mutex_unlock(&memcg->slab_caches_mutex);
3361 static void memcg_create_cache_work_func(struct work_struct *w)
3363 struct create_work *cw;
3365 cw = container_of(w, struct create_work, work);
3366 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3367 /* Drop the reference gotten when we enqueued. */
3368 css_put(&cw->memcg->css);
3373 * Enqueue the creation of a per-memcg kmem_cache.
3374 * Called with rcu_read_lock.
3376 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3377 struct kmem_cache *cachep)
3379 struct create_work *cw;
3381 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3385 /* The corresponding put will be done in the workqueue. */
3386 if (!css_tryget(&memcg->css)) {
3392 cw->cachep = cachep;
3394 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3395 schedule_work(&cw->work);
3398 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3399 struct kmem_cache *cachep)
3402 * We need to stop accounting when we kmalloc, because if the
3403 * corresponding kmalloc cache is not yet created, the first allocation
3404 * in __memcg_create_cache_enqueue will recurse.
3406 * However, it is better to enclose the whole function. Depending on
3407 * the debugging options enabled, INIT_WORK(), for instance, can
3408 * trigger an allocation. This too, will make us recurse. Because at
3409 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3410 * the safest choice is to do it like this, wrapping the whole function.
3412 memcg_stop_kmem_account();
3413 __memcg_create_cache_enqueue(memcg, cachep);
3414 memcg_resume_kmem_account();
3417 * Return the kmem_cache we're supposed to use for a slab allocation.
3418 * We try to use the current memcg's version of the cache.
3420 * If the cache does not exist yet, if we are the first user of it,
3421 * we either create it immediately, if possible, or create it asynchronously
3423 * In the latter case, we will let the current allocation go through with
3424 * the original cache.
3426 * Can't be called in interrupt context or from kernel threads.
3427 * This function needs to be called with rcu_read_lock() held.
3429 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3432 struct mem_cgroup *memcg;
3435 VM_BUG_ON(!cachep->memcg_params);
3436 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3438 if (!current->mm || current->memcg_kmem_skip_account)
3442 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3445 if (!memcg_can_account_kmem(memcg))
3448 idx = memcg_cache_id(memcg);
3451 * barrier to mare sure we're always seeing the up to date value. The
3452 * code updating memcg_caches will issue a write barrier to match this.
3454 read_barrier_depends();
3455 if (unlikely(cachep->memcg_params->memcg_caches[idx] == NULL)) {
3457 * If we are in a safe context (can wait, and not in interrupt
3458 * context), we could be be predictable and return right away.
3459 * This would guarantee that the allocation being performed
3460 * already belongs in the new cache.
3462 * However, there are some clashes that can arrive from locking.
3463 * For instance, because we acquire the slab_mutex while doing
3464 * kmem_cache_dup, this means no further allocation could happen
3465 * with the slab_mutex held.
3467 * Also, because cache creation issue get_online_cpus(), this
3468 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3469 * that ends up reversed during cpu hotplug. (cpuset allocates
3470 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3471 * better to defer everything.
3473 memcg_create_cache_enqueue(memcg, cachep);
3477 return cachep->memcg_params->memcg_caches[idx];
3479 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3482 * We need to verify if the allocation against current->mm->owner's memcg is
3483 * possible for the given order. But the page is not allocated yet, so we'll
3484 * need a further commit step to do the final arrangements.
3486 * It is possible for the task to switch cgroups in this mean time, so at
3487 * commit time, we can't rely on task conversion any longer. We'll then use
3488 * the handle argument to return to the caller which cgroup we should commit
3489 * against. We could also return the memcg directly and avoid the pointer
3490 * passing, but a boolean return value gives better semantics considering
3491 * the compiled-out case as well.
3493 * Returning true means the allocation is possible.
3496 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3498 struct mem_cgroup *memcg;
3502 memcg = try_get_mem_cgroup_from_mm(current->mm);
3505 * very rare case described in mem_cgroup_from_task. Unfortunately there
3506 * isn't much we can do without complicating this too much, and it would
3507 * be gfp-dependent anyway. Just let it go
3509 if (unlikely(!memcg))
3512 if (!memcg_can_account_kmem(memcg)) {
3513 css_put(&memcg->css);
3517 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3521 css_put(&memcg->css);
3525 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3528 struct page_cgroup *pc;
3530 VM_BUG_ON(mem_cgroup_is_root(memcg));
3532 /* The page allocation failed. Revert */
3534 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3538 pc = lookup_page_cgroup(page);
3539 lock_page_cgroup(pc);
3540 pc->mem_cgroup = memcg;
3541 SetPageCgroupUsed(pc);
3542 unlock_page_cgroup(pc);
3545 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3547 struct mem_cgroup *memcg = NULL;
3548 struct page_cgroup *pc;
3551 pc = lookup_page_cgroup(page);
3553 * Fast unlocked return. Theoretically might have changed, have to
3554 * check again after locking.
3556 if (!PageCgroupUsed(pc))
3559 lock_page_cgroup(pc);
3560 if (PageCgroupUsed(pc)) {
3561 memcg = pc->mem_cgroup;
3562 ClearPageCgroupUsed(pc);
3564 unlock_page_cgroup(pc);
3567 * We trust that only if there is a memcg associated with the page, it
3568 * is a valid allocation
3573 VM_BUG_ON(mem_cgroup_is_root(memcg));
3574 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3577 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3580 #endif /* CONFIG_MEMCG_KMEM */
3582 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3584 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3586 * Because tail pages are not marked as "used", set it. We're under
3587 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3588 * charge/uncharge will be never happen and move_account() is done under
3589 * compound_lock(), so we don't have to take care of races.
3591 void mem_cgroup_split_huge_fixup(struct page *head)
3593 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3594 struct page_cgroup *pc;
3597 if (mem_cgroup_disabled())
3599 for (i = 1; i < HPAGE_PMD_NR; i++) {
3601 pc->mem_cgroup = head_pc->mem_cgroup;
3602 smp_wmb();/* see __commit_charge() */
3603 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3606 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3609 * mem_cgroup_move_account - move account of the page
3611 * @nr_pages: number of regular pages (>1 for huge pages)
3612 * @pc: page_cgroup of the page.
3613 * @from: mem_cgroup which the page is moved from.
3614 * @to: mem_cgroup which the page is moved to. @from != @to.
3616 * The caller must confirm following.
3617 * - page is not on LRU (isolate_page() is useful.)
3618 * - compound_lock is held when nr_pages > 1
3620 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3623 static int mem_cgroup_move_account(struct page *page,
3624 unsigned int nr_pages,
3625 struct page_cgroup *pc,
3626 struct mem_cgroup *from,
3627 struct mem_cgroup *to)
3629 unsigned long flags;
3631 bool anon = PageAnon(page);
3633 VM_BUG_ON(from == to);
3634 VM_BUG_ON(PageLRU(page));
3636 * The page is isolated from LRU. So, collapse function
3637 * will not handle this page. But page splitting can happen.
3638 * Do this check under compound_page_lock(). The caller should
3642 if (nr_pages > 1 && !PageTransHuge(page))
3645 lock_page_cgroup(pc);
3648 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3651 move_lock_mem_cgroup(from, &flags);
3653 if (!anon && page_mapped(page)) {
3654 /* Update mapped_file data for mem_cgroup */
3656 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3657 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3660 mem_cgroup_charge_statistics(from, anon, -nr_pages);
3662 /* caller should have done css_get */
3663 pc->mem_cgroup = to;
3664 mem_cgroup_charge_statistics(to, anon, nr_pages);
3665 move_unlock_mem_cgroup(from, &flags);
3668 unlock_page_cgroup(pc);
3672 memcg_check_events(to, page);
3673 memcg_check_events(from, page);
3679 * mem_cgroup_move_parent - moves page to the parent group
3680 * @page: the page to move
3681 * @pc: page_cgroup of the page
3682 * @child: page's cgroup
3684 * move charges to its parent or the root cgroup if the group has no
3685 * parent (aka use_hierarchy==0).
3686 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3687 * mem_cgroup_move_account fails) the failure is always temporary and
3688 * it signals a race with a page removal/uncharge or migration. In the
3689 * first case the page is on the way out and it will vanish from the LRU
3690 * on the next attempt and the call should be retried later.
3691 * Isolation from the LRU fails only if page has been isolated from
3692 * the LRU since we looked at it and that usually means either global
3693 * reclaim or migration going on. The page will either get back to the
3695 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3696 * (!PageCgroupUsed) or moved to a different group. The page will
3697 * disappear in the next attempt.
3699 static int mem_cgroup_move_parent(struct page *page,
3700 struct page_cgroup *pc,
3701 struct mem_cgroup *child)
3703 struct mem_cgroup *parent;
3704 unsigned int nr_pages;
3705 unsigned long uninitialized_var(flags);
3708 VM_BUG_ON(mem_cgroup_is_root(child));
3711 if (!get_page_unless_zero(page))
3713 if (isolate_lru_page(page))
3716 nr_pages = hpage_nr_pages(page);
3718 parent = parent_mem_cgroup(child);
3720 * If no parent, move charges to root cgroup.
3723 parent = root_mem_cgroup;
3726 VM_BUG_ON(!PageTransHuge(page));
3727 flags = compound_lock_irqsave(page);
3730 ret = mem_cgroup_move_account(page, nr_pages,
3733 __mem_cgroup_cancel_local_charge(child, nr_pages);
3736 compound_unlock_irqrestore(page, flags);
3737 putback_lru_page(page);
3745 * Charge the memory controller for page usage.
3747 * 0 if the charge was successful
3748 * < 0 if the cgroup is over its limit
3750 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3751 gfp_t gfp_mask, enum charge_type ctype)
3753 struct mem_cgroup *memcg = NULL;
3754 unsigned int nr_pages = 1;
3758 if (PageTransHuge(page)) {
3759 nr_pages <<= compound_order(page);
3760 VM_BUG_ON(!PageTransHuge(page));
3762 * Never OOM-kill a process for a huge page. The
3763 * fault handler will fall back to regular pages.
3768 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3771 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3775 int mem_cgroup_newpage_charge(struct page *page,
3776 struct mm_struct *mm, gfp_t gfp_mask)
3778 if (mem_cgroup_disabled())
3780 VM_BUG_ON(page_mapped(page));
3781 VM_BUG_ON(page->mapping && !PageAnon(page));
3783 return mem_cgroup_charge_common(page, mm, gfp_mask,
3784 MEM_CGROUP_CHARGE_TYPE_ANON);
3788 * While swap-in, try_charge -> commit or cancel, the page is locked.
3789 * And when try_charge() successfully returns, one refcnt to memcg without
3790 * struct page_cgroup is acquired. This refcnt will be consumed by
3791 * "commit()" or removed by "cancel()"
3793 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3796 struct mem_cgroup **memcgp)
3798 struct mem_cgroup *memcg;
3799 struct page_cgroup *pc;
3802 pc = lookup_page_cgroup(page);
3804 * Every swap fault against a single page tries to charge the
3805 * page, bail as early as possible. shmem_unuse() encounters
3806 * already charged pages, too. The USED bit is protected by
3807 * the page lock, which serializes swap cache removal, which
3808 * in turn serializes uncharging.
3810 if (PageCgroupUsed(pc))
3812 if (!do_swap_account)
3814 memcg = try_get_mem_cgroup_from_page(page);
3818 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3819 css_put(&memcg->css);
3824 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3830 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3831 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3834 if (mem_cgroup_disabled())
3837 * A racing thread's fault, or swapoff, may have already
3838 * updated the pte, and even removed page from swap cache: in
3839 * those cases unuse_pte()'s pte_same() test will fail; but
3840 * there's also a KSM case which does need to charge the page.
3842 if (!PageSwapCache(page)) {
3845 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3850 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3853 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3855 if (mem_cgroup_disabled())
3859 __mem_cgroup_cancel_charge(memcg, 1);
3863 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3864 enum charge_type ctype)
3866 if (mem_cgroup_disabled())
3871 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3873 * Now swap is on-memory. This means this page may be
3874 * counted both as mem and swap....double count.
3875 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3876 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3877 * may call delete_from_swap_cache() before reach here.
3879 if (do_swap_account && PageSwapCache(page)) {
3880 swp_entry_t ent = {.val = page_private(page)};
3881 mem_cgroup_uncharge_swap(ent);
3885 void mem_cgroup_commit_charge_swapin(struct page *page,
3886 struct mem_cgroup *memcg)
3888 __mem_cgroup_commit_charge_swapin(page, memcg,
3889 MEM_CGROUP_CHARGE_TYPE_ANON);
3892 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3895 struct mem_cgroup *memcg = NULL;
3896 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3899 if (mem_cgroup_disabled())
3901 if (PageCompound(page))
3904 if (!PageSwapCache(page))
3905 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3906 else { /* page is swapcache/shmem */
3907 ret = __mem_cgroup_try_charge_swapin(mm, page,
3910 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3915 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3916 unsigned int nr_pages,
3917 const enum charge_type ctype)
3919 struct memcg_batch_info *batch = NULL;
3920 bool uncharge_memsw = true;
3922 /* If swapout, usage of swap doesn't decrease */
3923 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3924 uncharge_memsw = false;
3926 batch = ¤t->memcg_batch;
3928 * In usual, we do css_get() when we remember memcg pointer.
3929 * But in this case, we keep res->usage until end of a series of
3930 * uncharges. Then, it's ok to ignore memcg's refcnt.
3933 batch->memcg = memcg;
3935 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3936 * In those cases, all pages freed continuously can be expected to be in
3937 * the same cgroup and we have chance to coalesce uncharges.
3938 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3939 * because we want to do uncharge as soon as possible.
3942 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3943 goto direct_uncharge;
3946 goto direct_uncharge;
3949 * In typical case, batch->memcg == mem. This means we can
3950 * merge a series of uncharges to an uncharge of res_counter.
3951 * If not, we uncharge res_counter ony by one.
3953 if (batch->memcg != memcg)
3954 goto direct_uncharge;
3955 /* remember freed charge and uncharge it later */
3958 batch->memsw_nr_pages++;
3961 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3963 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3964 if (unlikely(batch->memcg != memcg))
3965 memcg_oom_recover(memcg);
3969 * uncharge if !page_mapped(page)
3971 static struct mem_cgroup *
3972 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3975 struct mem_cgroup *memcg = NULL;
3976 unsigned int nr_pages = 1;
3977 struct page_cgroup *pc;
3980 if (mem_cgroup_disabled())
3983 VM_BUG_ON(PageSwapCache(page));
3985 if (PageTransHuge(page)) {
3986 nr_pages <<= compound_order(page);
3987 VM_BUG_ON(!PageTransHuge(page));
3990 * Check if our page_cgroup is valid
3992 pc = lookup_page_cgroup(page);
3993 if (unlikely(!PageCgroupUsed(pc)))
3996 lock_page_cgroup(pc);
3998 memcg = pc->mem_cgroup;
4000 if (!PageCgroupUsed(pc))
4003 anon = PageAnon(page);
4006 case MEM_CGROUP_CHARGE_TYPE_ANON:
4008 * Generally PageAnon tells if it's the anon statistics to be
4009 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4010 * used before page reached the stage of being marked PageAnon.
4014 case MEM_CGROUP_CHARGE_TYPE_DROP:
4015 /* See mem_cgroup_prepare_migration() */
4016 if (page_mapped(page))
4019 * Pages under migration may not be uncharged. But
4020 * end_migration() /must/ be the one uncharging the
4021 * unused post-migration page and so it has to call
4022 * here with the migration bit still set. See the
4023 * res_counter handling below.
4025 if (!end_migration && PageCgroupMigration(pc))
4028 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4029 if (!PageAnon(page)) { /* Shared memory */
4030 if (page->mapping && !page_is_file_cache(page))
4032 } else if (page_mapped(page)) /* Anon */
4039 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
4041 ClearPageCgroupUsed(pc);
4043 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4044 * freed from LRU. This is safe because uncharged page is expected not
4045 * to be reused (freed soon). Exception is SwapCache, it's handled by
4046 * special functions.
4049 unlock_page_cgroup(pc);
4051 * even after unlock, we have memcg->res.usage here and this memcg
4052 * will never be freed.
4054 memcg_check_events(memcg, page);
4055 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4056 mem_cgroup_swap_statistics(memcg, true);
4057 mem_cgroup_get(memcg);
4060 * Migration does not charge the res_counter for the
4061 * replacement page, so leave it alone when phasing out the
4062 * page that is unused after the migration.
4064 if (!end_migration && !mem_cgroup_is_root(memcg))
4065 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4070 unlock_page_cgroup(pc);
4074 void mem_cgroup_uncharge_page(struct page *page)
4077 if (page_mapped(page))
4079 VM_BUG_ON(page->mapping && !PageAnon(page));
4080 if (PageSwapCache(page))
4082 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4085 void mem_cgroup_uncharge_cache_page(struct page *page)
4087 VM_BUG_ON(page_mapped(page));
4088 VM_BUG_ON(page->mapping);
4089 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4093 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4094 * In that cases, pages are freed continuously and we can expect pages
4095 * are in the same memcg. All these calls itself limits the number of
4096 * pages freed at once, then uncharge_start/end() is called properly.
4097 * This may be called prural(2) times in a context,
4100 void mem_cgroup_uncharge_start(void)
4102 current->memcg_batch.do_batch++;
4103 /* We can do nest. */
4104 if (current->memcg_batch.do_batch == 1) {
4105 current->memcg_batch.memcg = NULL;
4106 current->memcg_batch.nr_pages = 0;
4107 current->memcg_batch.memsw_nr_pages = 0;
4111 void mem_cgroup_uncharge_end(void)
4113 struct memcg_batch_info *batch = ¤t->memcg_batch;
4115 if (!batch->do_batch)
4119 if (batch->do_batch) /* If stacked, do nothing. */
4125 * This "batch->memcg" is valid without any css_get/put etc...
4126 * bacause we hide charges behind us.
4128 if (batch->nr_pages)
4129 res_counter_uncharge(&batch->memcg->res,
4130 batch->nr_pages * PAGE_SIZE);
4131 if (batch->memsw_nr_pages)
4132 res_counter_uncharge(&batch->memcg->memsw,
4133 batch->memsw_nr_pages * PAGE_SIZE);
4134 memcg_oom_recover(batch->memcg);
4135 /* forget this pointer (for sanity check) */
4136 batch->memcg = NULL;
4141 * called after __delete_from_swap_cache() and drop "page" account.
4142 * memcg information is recorded to swap_cgroup of "ent"
4145 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4147 struct mem_cgroup *memcg;
4148 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4150 if (!swapout) /* this was a swap cache but the swap is unused ! */
4151 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4153 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4156 * record memcg information, if swapout && memcg != NULL,
4157 * mem_cgroup_get() was called in uncharge().
4159 if (do_swap_account && swapout && memcg)
4160 swap_cgroup_record(ent, css_id(&memcg->css));
4164 #ifdef CONFIG_MEMCG_SWAP
4166 * called from swap_entry_free(). remove record in swap_cgroup and
4167 * uncharge "memsw" account.
4169 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4171 struct mem_cgroup *memcg;
4174 if (!do_swap_account)
4177 id = swap_cgroup_record(ent, 0);
4179 memcg = mem_cgroup_lookup(id);
4182 * We uncharge this because swap is freed.
4183 * This memcg can be obsolete one. We avoid calling css_tryget
4185 if (!mem_cgroup_is_root(memcg))
4186 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4187 mem_cgroup_swap_statistics(memcg, false);
4188 mem_cgroup_put(memcg);
4194 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4195 * @entry: swap entry to be moved
4196 * @from: mem_cgroup which the entry is moved from
4197 * @to: mem_cgroup which the entry is moved to
4199 * It succeeds only when the swap_cgroup's record for this entry is the same
4200 * as the mem_cgroup's id of @from.
4202 * Returns 0 on success, -EINVAL on failure.
4204 * The caller must have charged to @to, IOW, called res_counter_charge() about
4205 * both res and memsw, and called css_get().
4207 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4208 struct mem_cgroup *from, struct mem_cgroup *to)
4210 unsigned short old_id, new_id;
4212 old_id = css_id(&from->css);
4213 new_id = css_id(&to->css);
4215 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4216 mem_cgroup_swap_statistics(from, false);
4217 mem_cgroup_swap_statistics(to, true);
4219 * This function is only called from task migration context now.
4220 * It postpones res_counter and refcount handling till the end
4221 * of task migration(mem_cgroup_clear_mc()) for performance
4222 * improvement. But we cannot postpone mem_cgroup_get(to)
4223 * because if the process that has been moved to @to does
4224 * swap-in, the refcount of @to might be decreased to 0.
4232 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4233 struct mem_cgroup *from, struct mem_cgroup *to)
4240 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4243 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4244 struct mem_cgroup **memcgp)
4246 struct mem_cgroup *memcg = NULL;
4247 unsigned int nr_pages = 1;
4248 struct page_cgroup *pc;
4249 enum charge_type ctype;
4253 if (mem_cgroup_disabled())
4256 if (PageTransHuge(page))
4257 nr_pages <<= compound_order(page);
4259 pc = lookup_page_cgroup(page);
4260 lock_page_cgroup(pc);
4261 if (PageCgroupUsed(pc)) {
4262 memcg = pc->mem_cgroup;
4263 css_get(&memcg->css);
4265 * At migrating an anonymous page, its mapcount goes down
4266 * to 0 and uncharge() will be called. But, even if it's fully
4267 * unmapped, migration may fail and this page has to be
4268 * charged again. We set MIGRATION flag here and delay uncharge
4269 * until end_migration() is called
4271 * Corner Case Thinking
4273 * When the old page was mapped as Anon and it's unmap-and-freed
4274 * while migration was ongoing.
4275 * If unmap finds the old page, uncharge() of it will be delayed
4276 * until end_migration(). If unmap finds a new page, it's
4277 * uncharged when it make mapcount to be 1->0. If unmap code
4278 * finds swap_migration_entry, the new page will not be mapped
4279 * and end_migration() will find it(mapcount==0).
4282 * When the old page was mapped but migraion fails, the kernel
4283 * remaps it. A charge for it is kept by MIGRATION flag even
4284 * if mapcount goes down to 0. We can do remap successfully
4285 * without charging it again.
4288 * The "old" page is under lock_page() until the end of
4289 * migration, so, the old page itself will not be swapped-out.
4290 * If the new page is swapped out before end_migraton, our
4291 * hook to usual swap-out path will catch the event.
4294 SetPageCgroupMigration(pc);
4296 unlock_page_cgroup(pc);
4298 * If the page is not charged at this point,
4306 * We charge new page before it's used/mapped. So, even if unlock_page()
4307 * is called before end_migration, we can catch all events on this new
4308 * page. In the case new page is migrated but not remapped, new page's
4309 * mapcount will be finally 0 and we call uncharge in end_migration().
4312 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4314 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4316 * The page is committed to the memcg, but it's not actually
4317 * charged to the res_counter since we plan on replacing the
4318 * old one and only one page is going to be left afterwards.
4320 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4323 /* remove redundant charge if migration failed*/
4324 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4325 struct page *oldpage, struct page *newpage, bool migration_ok)
4327 struct page *used, *unused;
4328 struct page_cgroup *pc;
4334 if (!migration_ok) {
4341 anon = PageAnon(used);
4342 __mem_cgroup_uncharge_common(unused,
4343 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4344 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4346 css_put(&memcg->css);
4348 * We disallowed uncharge of pages under migration because mapcount
4349 * of the page goes down to zero, temporarly.
4350 * Clear the flag and check the page should be charged.
4352 pc = lookup_page_cgroup(oldpage);
4353 lock_page_cgroup(pc);
4354 ClearPageCgroupMigration(pc);
4355 unlock_page_cgroup(pc);
4358 * If a page is a file cache, radix-tree replacement is very atomic
4359 * and we can skip this check. When it was an Anon page, its mapcount
4360 * goes down to 0. But because we added MIGRATION flage, it's not
4361 * uncharged yet. There are several case but page->mapcount check
4362 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4363 * check. (see prepare_charge() also)
4366 mem_cgroup_uncharge_page(used);
4370 * At replace page cache, newpage is not under any memcg but it's on
4371 * LRU. So, this function doesn't touch res_counter but handles LRU
4372 * in correct way. Both pages are locked so we cannot race with uncharge.
4374 void mem_cgroup_replace_page_cache(struct page *oldpage,
4375 struct page *newpage)
4377 struct mem_cgroup *memcg = NULL;
4378 struct page_cgroup *pc;
4379 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4381 if (mem_cgroup_disabled())
4384 pc = lookup_page_cgroup(oldpage);
4385 /* fix accounting on old pages */
4386 lock_page_cgroup(pc);
4387 if (PageCgroupUsed(pc)) {
4388 memcg = pc->mem_cgroup;
4389 mem_cgroup_charge_statistics(memcg, false, -1);
4390 ClearPageCgroupUsed(pc);
4392 unlock_page_cgroup(pc);
4395 * When called from shmem_replace_page(), in some cases the
4396 * oldpage has already been charged, and in some cases not.
4401 * Even if newpage->mapping was NULL before starting replacement,
4402 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4403 * LRU while we overwrite pc->mem_cgroup.
4405 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4408 #ifdef CONFIG_DEBUG_VM
4409 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4411 struct page_cgroup *pc;
4413 pc = lookup_page_cgroup(page);
4415 * Can be NULL while feeding pages into the page allocator for
4416 * the first time, i.e. during boot or memory hotplug;
4417 * or when mem_cgroup_disabled().
4419 if (likely(pc) && PageCgroupUsed(pc))
4424 bool mem_cgroup_bad_page_check(struct page *page)
4426 if (mem_cgroup_disabled())
4429 return lookup_page_cgroup_used(page) != NULL;
4432 void mem_cgroup_print_bad_page(struct page *page)
4434 struct page_cgroup *pc;
4436 pc = lookup_page_cgroup_used(page);
4438 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4439 pc, pc->flags, pc->mem_cgroup);
4444 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4445 unsigned long long val)
4448 u64 memswlimit, memlimit;
4450 int children = mem_cgroup_count_children(memcg);
4451 u64 curusage, oldusage;
4455 * For keeping hierarchical_reclaim simple, how long we should retry
4456 * is depends on callers. We set our retry-count to be function
4457 * of # of children which we should visit in this loop.
4459 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4461 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4464 while (retry_count) {
4465 if (signal_pending(current)) {
4470 * Rather than hide all in some function, I do this in
4471 * open coded manner. You see what this really does.
4472 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4474 mutex_lock(&set_limit_mutex);
4475 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4476 if (memswlimit < val) {
4478 mutex_unlock(&set_limit_mutex);
4482 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4486 ret = res_counter_set_limit(&memcg->res, val);
4488 if (memswlimit == val)
4489 memcg->memsw_is_minimum = true;
4491 memcg->memsw_is_minimum = false;
4493 mutex_unlock(&set_limit_mutex);
4498 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4499 MEM_CGROUP_RECLAIM_SHRINK);
4500 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4501 /* Usage is reduced ? */
4502 if (curusage >= oldusage)
4505 oldusage = curusage;
4507 if (!ret && enlarge)
4508 memcg_oom_recover(memcg);
4513 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4514 unsigned long long val)
4517 u64 memlimit, memswlimit, oldusage, curusage;
4518 int children = mem_cgroup_count_children(memcg);
4522 /* see mem_cgroup_resize_res_limit */
4523 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4524 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4525 while (retry_count) {
4526 if (signal_pending(current)) {
4531 * Rather than hide all in some function, I do this in
4532 * open coded manner. You see what this really does.
4533 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4535 mutex_lock(&set_limit_mutex);
4536 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4537 if (memlimit > val) {
4539 mutex_unlock(&set_limit_mutex);
4542 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4543 if (memswlimit < val)
4545 ret = res_counter_set_limit(&memcg->memsw, val);
4547 if (memlimit == val)
4548 memcg->memsw_is_minimum = true;
4550 memcg->memsw_is_minimum = false;
4552 mutex_unlock(&set_limit_mutex);
4557 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4558 MEM_CGROUP_RECLAIM_NOSWAP |
4559 MEM_CGROUP_RECLAIM_SHRINK);
4560 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4561 /* Usage is reduced ? */
4562 if (curusage >= oldusage)
4565 oldusage = curusage;
4567 if (!ret && enlarge)
4568 memcg_oom_recover(memcg);
4572 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4574 unsigned long *total_scanned)
4576 unsigned long nr_reclaimed = 0;
4577 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4578 unsigned long reclaimed;
4580 struct mem_cgroup_tree_per_zone *mctz;
4581 unsigned long long excess;
4582 unsigned long nr_scanned;
4587 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4589 * This loop can run a while, specially if mem_cgroup's continuously
4590 * keep exceeding their soft limit and putting the system under
4597 mz = mem_cgroup_largest_soft_limit_node(mctz);
4602 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4603 gfp_mask, &nr_scanned);
4604 nr_reclaimed += reclaimed;
4605 *total_scanned += nr_scanned;
4606 spin_lock(&mctz->lock);
4609 * If we failed to reclaim anything from this memory cgroup
4610 * it is time to move on to the next cgroup
4616 * Loop until we find yet another one.
4618 * By the time we get the soft_limit lock
4619 * again, someone might have aded the
4620 * group back on the RB tree. Iterate to
4621 * make sure we get a different mem.
4622 * mem_cgroup_largest_soft_limit_node returns
4623 * NULL if no other cgroup is present on
4627 __mem_cgroup_largest_soft_limit_node(mctz);
4629 css_put(&next_mz->memcg->css);
4630 else /* next_mz == NULL or other memcg */
4634 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4635 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4637 * One school of thought says that we should not add
4638 * back the node to the tree if reclaim returns 0.
4639 * But our reclaim could return 0, simply because due
4640 * to priority we are exposing a smaller subset of
4641 * memory to reclaim from. Consider this as a longer
4644 /* If excess == 0, no tree ops */
4645 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4646 spin_unlock(&mctz->lock);
4647 css_put(&mz->memcg->css);
4650 * Could not reclaim anything and there are no more
4651 * mem cgroups to try or we seem to be looping without
4652 * reclaiming anything.
4654 if (!nr_reclaimed &&
4656 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4658 } while (!nr_reclaimed);
4660 css_put(&next_mz->memcg->css);
4661 return nr_reclaimed;
4665 * mem_cgroup_force_empty_list - clears LRU of a group
4666 * @memcg: group to clear
4669 * @lru: lru to to clear
4671 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4672 * reclaim the pages page themselves - pages are moved to the parent (or root)
4675 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4676 int node, int zid, enum lru_list lru)
4678 struct lruvec *lruvec;
4679 unsigned long flags;
4680 struct list_head *list;
4684 zone = &NODE_DATA(node)->node_zones[zid];
4685 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4686 list = &lruvec->lists[lru];
4690 struct page_cgroup *pc;
4693 spin_lock_irqsave(&zone->lru_lock, flags);
4694 if (list_empty(list)) {
4695 spin_unlock_irqrestore(&zone->lru_lock, flags);
4698 page = list_entry(list->prev, struct page, lru);
4700 list_move(&page->lru, list);
4702 spin_unlock_irqrestore(&zone->lru_lock, flags);
4705 spin_unlock_irqrestore(&zone->lru_lock, flags);
4707 pc = lookup_page_cgroup(page);
4709 if (mem_cgroup_move_parent(page, pc, memcg)) {
4710 /* found lock contention or "pc" is obsolete. */
4715 } while (!list_empty(list));
4719 * make mem_cgroup's charge to be 0 if there is no task by moving
4720 * all the charges and pages to the parent.
4721 * This enables deleting this mem_cgroup.
4723 * Caller is responsible for holding css reference on the memcg.
4725 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4731 /* This is for making all *used* pages to be on LRU. */
4732 lru_add_drain_all();
4733 drain_all_stock_sync(memcg);
4734 mem_cgroup_start_move(memcg);
4735 for_each_node_state(node, N_MEMORY) {
4736 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4739 mem_cgroup_force_empty_list(memcg,
4744 mem_cgroup_end_move(memcg);
4745 memcg_oom_recover(memcg);
4749 * Kernel memory may not necessarily be trackable to a specific
4750 * process. So they are not migrated, and therefore we can't
4751 * expect their value to drop to 0 here.
4752 * Having res filled up with kmem only is enough.
4754 * This is a safety check because mem_cgroup_force_empty_list
4755 * could have raced with mem_cgroup_replace_page_cache callers
4756 * so the lru seemed empty but the page could have been added
4757 * right after the check. RES_USAGE should be safe as we always
4758 * charge before adding to the LRU.
4760 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4761 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4762 } while (usage > 0);
4766 * This mainly exists for tests during the setting of set of use_hierarchy.
4767 * Since this is the very setting we are changing, the current hierarchy value
4770 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4774 /* bounce at first found */
4775 cgroup_for_each_child(pos, memcg->css.cgroup)
4781 * Must be called with cgroup_lock held, unless the cgroup is guaranteed to be
4782 * already dead (in mem_cgroup_force_empty(), for instance). This is different
4783 * from mem_cgroup_count_children(), in the sense that we don't really care how
4784 * many children we have; we only need to know if we have any. It also counts
4785 * any memcg without hierarchy as infertile.
4787 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4789 return memcg->use_hierarchy && __memcg_has_children(memcg);
4793 * Reclaims as many pages from the given memcg as possible and moves
4794 * the rest to the parent.
4796 * Caller is responsible for holding css reference for memcg.
4798 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4800 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4801 struct cgroup *cgrp = memcg->css.cgroup;
4803 /* returns EBUSY if there is a task or if we come here twice. */
4804 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4807 /* we call try-to-free pages for make this cgroup empty */
4808 lru_add_drain_all();
4809 /* try to free all pages in this cgroup */
4810 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4813 if (signal_pending(current))
4816 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4820 /* maybe some writeback is necessary */
4821 congestion_wait(BLK_RW_ASYNC, HZ/10);
4826 mem_cgroup_reparent_charges(memcg);
4831 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4833 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4836 if (mem_cgroup_is_root(memcg))
4838 css_get(&memcg->css);
4839 ret = mem_cgroup_force_empty(memcg);
4840 css_put(&memcg->css);
4846 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4848 return mem_cgroup_from_cont(cont)->use_hierarchy;
4851 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4855 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4856 struct cgroup *parent = cont->parent;
4857 struct mem_cgroup *parent_memcg = NULL;
4860 parent_memcg = mem_cgroup_from_cont(parent);
4864 if (memcg->use_hierarchy == val)
4868 * If parent's use_hierarchy is set, we can't make any modifications
4869 * in the child subtrees. If it is unset, then the change can
4870 * occur, provided the current cgroup has no children.
4872 * For the root cgroup, parent_mem is NULL, we allow value to be
4873 * set if there are no children.
4875 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4876 (val == 1 || val == 0)) {
4877 if (!__memcg_has_children(memcg))
4878 memcg->use_hierarchy = val;
4891 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4892 enum mem_cgroup_stat_index idx)
4894 struct mem_cgroup *iter;
4897 /* Per-cpu values can be negative, use a signed accumulator */
4898 for_each_mem_cgroup_tree(iter, memcg)
4899 val += mem_cgroup_read_stat(iter, idx);
4901 if (val < 0) /* race ? */
4906 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4910 if (!mem_cgroup_is_root(memcg)) {
4912 return res_counter_read_u64(&memcg->res, RES_USAGE);
4914 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4917 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4918 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4921 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4923 return val << PAGE_SHIFT;
4926 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
4927 struct file *file, char __user *buf,
4928 size_t nbytes, loff_t *ppos)
4930 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4936 type = MEMFILE_TYPE(cft->private);
4937 name = MEMFILE_ATTR(cft->private);
4939 if (!do_swap_account && type == _MEMSWAP)
4944 if (name == RES_USAGE)
4945 val = mem_cgroup_usage(memcg, false);
4947 val = res_counter_read_u64(&memcg->res, name);
4950 if (name == RES_USAGE)
4951 val = mem_cgroup_usage(memcg, true);
4953 val = res_counter_read_u64(&memcg->memsw, name);
4956 val = res_counter_read_u64(&memcg->kmem, name);
4962 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4963 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4966 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
4969 #ifdef CONFIG_MEMCG_KMEM
4970 bool must_inc_static_branch = false;
4972 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4974 * For simplicity, we won't allow this to be disabled. It also can't
4975 * be changed if the cgroup has children already, or if tasks had
4978 * If tasks join before we set the limit, a person looking at
4979 * kmem.usage_in_bytes will have no way to determine when it took
4980 * place, which makes the value quite meaningless.
4982 * After it first became limited, changes in the value of the limit are
4983 * of course permitted.
4985 * Taking the cgroup_lock is really offensive, but it is so far the only
4986 * way to guarantee that no children will appear. There are plenty of
4987 * other offenders, and they should all go away. Fine grained locking
4988 * is probably the way to go here. When we are fully hierarchical, we
4989 * can also get rid of the use_hierarchy check.
4992 mutex_lock(&set_limit_mutex);
4993 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
4994 if (cgroup_task_count(cont) || memcg_has_children(memcg)) {
4998 ret = res_counter_set_limit(&memcg->kmem, val);
5001 ret = memcg_update_cache_sizes(memcg);
5003 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5006 must_inc_static_branch = true;
5008 * kmem charges can outlive the cgroup. In the case of slab
5009 * pages, for instance, a page contain objects from various
5010 * processes, so it is unfeasible to migrate them away. We
5011 * need to reference count the memcg because of that.
5013 mem_cgroup_get(memcg);
5015 ret = res_counter_set_limit(&memcg->kmem, val);
5017 mutex_unlock(&set_limit_mutex);
5021 * We are by now familiar with the fact that we can't inc the static
5022 * branch inside cgroup_lock. See disarm functions for details. A
5023 * worker here is overkill, but also wrong: After the limit is set, we
5024 * must start accounting right away. Since this operation can't fail,
5025 * we can safely defer it to here - no rollback will be needed.
5027 * The boolean used to control this is also safe, because
5028 * KMEM_ACCOUNTED_ACTIVATED guarantees that only one process will be
5029 * able to set it to true;
5031 if (must_inc_static_branch) {
5032 static_key_slow_inc(&memcg_kmem_enabled_key);
5034 * setting the active bit after the inc will guarantee no one
5035 * starts accounting before all call sites are patched
5037 memcg_kmem_set_active(memcg);
5044 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5047 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5051 memcg->kmem_account_flags = parent->kmem_account_flags;
5052 #ifdef CONFIG_MEMCG_KMEM
5054 * When that happen, we need to disable the static branch only on those
5055 * memcgs that enabled it. To achieve this, we would be forced to
5056 * complicate the code by keeping track of which memcgs were the ones
5057 * that actually enabled limits, and which ones got it from its
5060 * It is a lot simpler just to do static_key_slow_inc() on every child
5061 * that is accounted.
5063 if (!memcg_kmem_is_active(memcg))
5067 * destroy(), called if we fail, will issue static_key_slow_inc() and
5068 * mem_cgroup_put() if kmem is enabled. We have to either call them
5069 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5070 * this more consistent, since it always leads to the same destroy path
5072 mem_cgroup_get(memcg);
5073 static_key_slow_inc(&memcg_kmem_enabled_key);
5075 mutex_lock(&set_limit_mutex);
5076 ret = memcg_update_cache_sizes(memcg);
5077 mutex_unlock(&set_limit_mutex);
5084 * The user of this function is...
5087 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5090 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5093 unsigned long long val;
5096 type = MEMFILE_TYPE(cft->private);
5097 name = MEMFILE_ATTR(cft->private);
5099 if (!do_swap_account && type == _MEMSWAP)
5104 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5108 /* This function does all necessary parse...reuse it */
5109 ret = res_counter_memparse_write_strategy(buffer, &val);
5113 ret = mem_cgroup_resize_limit(memcg, val);
5114 else if (type == _MEMSWAP)
5115 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5116 else if (type == _KMEM)
5117 ret = memcg_update_kmem_limit(cont, val);
5121 case RES_SOFT_LIMIT:
5122 ret = res_counter_memparse_write_strategy(buffer, &val);
5126 * For memsw, soft limits are hard to implement in terms
5127 * of semantics, for now, we support soft limits for
5128 * control without swap
5131 ret = res_counter_set_soft_limit(&memcg->res, val);
5136 ret = -EINVAL; /* should be BUG() ? */
5142 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5143 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5145 struct cgroup *cgroup;
5146 unsigned long long min_limit, min_memsw_limit, tmp;
5148 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5149 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5150 cgroup = memcg->css.cgroup;
5151 if (!memcg->use_hierarchy)
5154 while (cgroup->parent) {
5155 cgroup = cgroup->parent;
5156 memcg = mem_cgroup_from_cont(cgroup);
5157 if (!memcg->use_hierarchy)
5159 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5160 min_limit = min(min_limit, tmp);
5161 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5162 min_memsw_limit = min(min_memsw_limit, tmp);
5165 *mem_limit = min_limit;
5166 *memsw_limit = min_memsw_limit;
5169 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5171 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5175 type = MEMFILE_TYPE(event);
5176 name = MEMFILE_ATTR(event);
5178 if (!do_swap_account && type == _MEMSWAP)
5184 res_counter_reset_max(&memcg->res);
5185 else if (type == _MEMSWAP)
5186 res_counter_reset_max(&memcg->memsw);
5187 else if (type == _KMEM)
5188 res_counter_reset_max(&memcg->kmem);
5194 res_counter_reset_failcnt(&memcg->res);
5195 else if (type == _MEMSWAP)
5196 res_counter_reset_failcnt(&memcg->memsw);
5197 else if (type == _KMEM)
5198 res_counter_reset_failcnt(&memcg->kmem);
5207 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5210 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5214 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5215 struct cftype *cft, u64 val)
5217 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5219 if (val >= (1 << NR_MOVE_TYPE))
5223 * No kind of locking is needed in here, because ->can_attach() will
5224 * check this value once in the beginning of the process, and then carry
5225 * on with stale data. This means that changes to this value will only
5226 * affect task migrations starting after the change.
5228 memcg->move_charge_at_immigrate = val;
5232 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5233 struct cftype *cft, u64 val)
5240 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5244 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5245 unsigned long node_nr;
5246 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5248 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5249 seq_printf(m, "total=%lu", total_nr);
5250 for_each_node_state(nid, N_MEMORY) {
5251 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5252 seq_printf(m, " N%d=%lu", nid, node_nr);
5256 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5257 seq_printf(m, "file=%lu", file_nr);
5258 for_each_node_state(nid, N_MEMORY) {
5259 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5261 seq_printf(m, " N%d=%lu", nid, node_nr);
5265 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5266 seq_printf(m, "anon=%lu", anon_nr);
5267 for_each_node_state(nid, N_MEMORY) {
5268 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5270 seq_printf(m, " N%d=%lu", nid, node_nr);
5274 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5275 seq_printf(m, "unevictable=%lu", unevictable_nr);
5276 for_each_node_state(nid, N_MEMORY) {
5277 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5278 BIT(LRU_UNEVICTABLE));
5279 seq_printf(m, " N%d=%lu", nid, node_nr);
5284 #endif /* CONFIG_NUMA */
5286 static inline void mem_cgroup_lru_names_not_uptodate(void)
5288 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5291 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5294 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5295 struct mem_cgroup *mi;
5298 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5299 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5301 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5302 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5305 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5306 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5307 mem_cgroup_read_events(memcg, i));
5309 for (i = 0; i < NR_LRU_LISTS; i++)
5310 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5311 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5313 /* Hierarchical information */
5315 unsigned long long limit, memsw_limit;
5316 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5317 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5318 if (do_swap_account)
5319 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5323 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5326 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5328 for_each_mem_cgroup_tree(mi, memcg)
5329 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5330 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5333 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5334 unsigned long long val = 0;
5336 for_each_mem_cgroup_tree(mi, memcg)
5337 val += mem_cgroup_read_events(mi, i);
5338 seq_printf(m, "total_%s %llu\n",
5339 mem_cgroup_events_names[i], val);
5342 for (i = 0; i < NR_LRU_LISTS; i++) {
5343 unsigned long long val = 0;
5345 for_each_mem_cgroup_tree(mi, memcg)
5346 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5347 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5350 #ifdef CONFIG_DEBUG_VM
5353 struct mem_cgroup_per_zone *mz;
5354 struct zone_reclaim_stat *rstat;
5355 unsigned long recent_rotated[2] = {0, 0};
5356 unsigned long recent_scanned[2] = {0, 0};
5358 for_each_online_node(nid)
5359 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5360 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5361 rstat = &mz->lruvec.reclaim_stat;
5363 recent_rotated[0] += rstat->recent_rotated[0];
5364 recent_rotated[1] += rstat->recent_rotated[1];
5365 recent_scanned[0] += rstat->recent_scanned[0];
5366 recent_scanned[1] += rstat->recent_scanned[1];
5368 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5369 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5370 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5371 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5378 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5380 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5382 return mem_cgroup_swappiness(memcg);
5385 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5388 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5389 struct mem_cgroup *parent;
5394 if (cgrp->parent == NULL)
5397 parent = mem_cgroup_from_cont(cgrp->parent);
5401 /* If under hierarchy, only empty-root can set this value */
5402 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5407 memcg->swappiness = val;
5414 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5416 struct mem_cgroup_threshold_ary *t;
5422 t = rcu_dereference(memcg->thresholds.primary);
5424 t = rcu_dereference(memcg->memsw_thresholds.primary);
5429 usage = mem_cgroup_usage(memcg, swap);
5432 * current_threshold points to threshold just below or equal to usage.
5433 * If it's not true, a threshold was crossed after last
5434 * call of __mem_cgroup_threshold().
5436 i = t->current_threshold;
5439 * Iterate backward over array of thresholds starting from
5440 * current_threshold and check if a threshold is crossed.
5441 * If none of thresholds below usage is crossed, we read
5442 * only one element of the array here.
5444 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5445 eventfd_signal(t->entries[i].eventfd, 1);
5447 /* i = current_threshold + 1 */
5451 * Iterate forward over array of thresholds starting from
5452 * current_threshold+1 and check if a threshold is crossed.
5453 * If none of thresholds above usage is crossed, we read
5454 * only one element of the array here.
5456 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5457 eventfd_signal(t->entries[i].eventfd, 1);
5459 /* Update current_threshold */
5460 t->current_threshold = i - 1;
5465 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5468 __mem_cgroup_threshold(memcg, false);
5469 if (do_swap_account)
5470 __mem_cgroup_threshold(memcg, true);
5472 memcg = parent_mem_cgroup(memcg);
5476 static int compare_thresholds(const void *a, const void *b)
5478 const struct mem_cgroup_threshold *_a = a;
5479 const struct mem_cgroup_threshold *_b = b;
5481 return _a->threshold - _b->threshold;
5484 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5486 struct mem_cgroup_eventfd_list *ev;
5488 list_for_each_entry(ev, &memcg->oom_notify, list)
5489 eventfd_signal(ev->eventfd, 1);
5493 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5495 struct mem_cgroup *iter;
5497 for_each_mem_cgroup_tree(iter, memcg)
5498 mem_cgroup_oom_notify_cb(iter);
5501 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5502 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5504 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5505 struct mem_cgroup_thresholds *thresholds;
5506 struct mem_cgroup_threshold_ary *new;
5507 enum res_type type = MEMFILE_TYPE(cft->private);
5508 u64 threshold, usage;
5511 ret = res_counter_memparse_write_strategy(args, &threshold);
5515 mutex_lock(&memcg->thresholds_lock);
5518 thresholds = &memcg->thresholds;
5519 else if (type == _MEMSWAP)
5520 thresholds = &memcg->memsw_thresholds;
5524 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5526 /* Check if a threshold crossed before adding a new one */
5527 if (thresholds->primary)
5528 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5530 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5532 /* Allocate memory for new array of thresholds */
5533 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5541 /* Copy thresholds (if any) to new array */
5542 if (thresholds->primary) {
5543 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5544 sizeof(struct mem_cgroup_threshold));
5547 /* Add new threshold */
5548 new->entries[size - 1].eventfd = eventfd;
5549 new->entries[size - 1].threshold = threshold;
5551 /* Sort thresholds. Registering of new threshold isn't time-critical */
5552 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5553 compare_thresholds, NULL);
5555 /* Find current threshold */
5556 new->current_threshold = -1;
5557 for (i = 0; i < size; i++) {
5558 if (new->entries[i].threshold <= usage) {
5560 * new->current_threshold will not be used until
5561 * rcu_assign_pointer(), so it's safe to increment
5564 ++new->current_threshold;
5569 /* Free old spare buffer and save old primary buffer as spare */
5570 kfree(thresholds->spare);
5571 thresholds->spare = thresholds->primary;
5573 rcu_assign_pointer(thresholds->primary, new);
5575 /* To be sure that nobody uses thresholds */
5579 mutex_unlock(&memcg->thresholds_lock);
5584 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5585 struct cftype *cft, struct eventfd_ctx *eventfd)
5587 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5588 struct mem_cgroup_thresholds *thresholds;
5589 struct mem_cgroup_threshold_ary *new;
5590 enum res_type type = MEMFILE_TYPE(cft->private);
5594 mutex_lock(&memcg->thresholds_lock);
5596 thresholds = &memcg->thresholds;
5597 else if (type == _MEMSWAP)
5598 thresholds = &memcg->memsw_thresholds;
5602 if (!thresholds->primary)
5605 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5607 /* Check if a threshold crossed before removing */
5608 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5610 /* Calculate new number of threshold */
5612 for (i = 0; i < thresholds->primary->size; i++) {
5613 if (thresholds->primary->entries[i].eventfd != eventfd)
5617 new = thresholds->spare;
5619 /* Set thresholds array to NULL if we don't have thresholds */
5628 /* Copy thresholds and find current threshold */
5629 new->current_threshold = -1;
5630 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5631 if (thresholds->primary->entries[i].eventfd == eventfd)
5634 new->entries[j] = thresholds->primary->entries[i];
5635 if (new->entries[j].threshold <= usage) {
5637 * new->current_threshold will not be used
5638 * until rcu_assign_pointer(), so it's safe to increment
5641 ++new->current_threshold;
5647 /* Swap primary and spare array */
5648 thresholds->spare = thresholds->primary;
5649 /* If all events are unregistered, free the spare array */
5651 kfree(thresholds->spare);
5652 thresholds->spare = NULL;
5655 rcu_assign_pointer(thresholds->primary, new);
5657 /* To be sure that nobody uses thresholds */
5660 mutex_unlock(&memcg->thresholds_lock);
5663 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5664 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5666 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5667 struct mem_cgroup_eventfd_list *event;
5668 enum res_type type = MEMFILE_TYPE(cft->private);
5670 BUG_ON(type != _OOM_TYPE);
5671 event = kmalloc(sizeof(*event), GFP_KERNEL);
5675 spin_lock(&memcg_oom_lock);
5677 event->eventfd = eventfd;
5678 list_add(&event->list, &memcg->oom_notify);
5680 /* already in OOM ? */
5681 if (atomic_read(&memcg->under_oom))
5682 eventfd_signal(eventfd, 1);
5683 spin_unlock(&memcg_oom_lock);
5688 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5689 struct cftype *cft, struct eventfd_ctx *eventfd)
5691 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5692 struct mem_cgroup_eventfd_list *ev, *tmp;
5693 enum res_type type = MEMFILE_TYPE(cft->private);
5695 BUG_ON(type != _OOM_TYPE);
5697 spin_lock(&memcg_oom_lock);
5699 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5700 if (ev->eventfd == eventfd) {
5701 list_del(&ev->list);
5706 spin_unlock(&memcg_oom_lock);
5709 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5710 struct cftype *cft, struct cgroup_map_cb *cb)
5712 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5714 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5716 if (atomic_read(&memcg->under_oom))
5717 cb->fill(cb, "under_oom", 1);
5719 cb->fill(cb, "under_oom", 0);
5723 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5724 struct cftype *cft, u64 val)
5726 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5727 struct mem_cgroup *parent;
5729 /* cannot set to root cgroup and only 0 and 1 are allowed */
5730 if (!cgrp->parent || !((val == 0) || (val == 1)))
5733 parent = mem_cgroup_from_cont(cgrp->parent);
5736 /* oom-kill-disable is a flag for subhierarchy. */
5737 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5741 memcg->oom_kill_disable = val;
5743 memcg_oom_recover(memcg);
5748 #ifdef CONFIG_MEMCG_KMEM
5749 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5753 memcg->kmemcg_id = -1;
5754 ret = memcg_propagate_kmem(memcg);
5758 return mem_cgroup_sockets_init(memcg, ss);
5761 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5763 mem_cgroup_sockets_destroy(memcg);
5765 memcg_kmem_mark_dead(memcg);
5767 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5771 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5772 * path here, being careful not to race with memcg_uncharge_kmem: it is
5773 * possible that the charges went down to 0 between mark_dead and the
5774 * res_counter read, so in that case, we don't need the put
5776 if (memcg_kmem_test_and_clear_dead(memcg))
5777 mem_cgroup_put(memcg);
5780 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5785 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5790 static struct cftype mem_cgroup_files[] = {
5792 .name = "usage_in_bytes",
5793 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5794 .read = mem_cgroup_read,
5795 .register_event = mem_cgroup_usage_register_event,
5796 .unregister_event = mem_cgroup_usage_unregister_event,
5799 .name = "max_usage_in_bytes",
5800 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5801 .trigger = mem_cgroup_reset,
5802 .read = mem_cgroup_read,
5805 .name = "limit_in_bytes",
5806 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5807 .write_string = mem_cgroup_write,
5808 .read = mem_cgroup_read,
5811 .name = "soft_limit_in_bytes",
5812 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5813 .write_string = mem_cgroup_write,
5814 .read = mem_cgroup_read,
5818 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5819 .trigger = mem_cgroup_reset,
5820 .read = mem_cgroup_read,
5824 .read_seq_string = memcg_stat_show,
5827 .name = "force_empty",
5828 .trigger = mem_cgroup_force_empty_write,
5831 .name = "use_hierarchy",
5832 .write_u64 = mem_cgroup_hierarchy_write,
5833 .read_u64 = mem_cgroup_hierarchy_read,
5836 .name = "swappiness",
5837 .read_u64 = mem_cgroup_swappiness_read,
5838 .write_u64 = mem_cgroup_swappiness_write,
5841 .name = "move_charge_at_immigrate",
5842 .read_u64 = mem_cgroup_move_charge_read,
5843 .write_u64 = mem_cgroup_move_charge_write,
5846 .name = "oom_control",
5847 .read_map = mem_cgroup_oom_control_read,
5848 .write_u64 = mem_cgroup_oom_control_write,
5849 .register_event = mem_cgroup_oom_register_event,
5850 .unregister_event = mem_cgroup_oom_unregister_event,
5851 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5855 .name = "numa_stat",
5856 .read_seq_string = memcg_numa_stat_show,
5859 #ifdef CONFIG_MEMCG_KMEM
5861 .name = "kmem.limit_in_bytes",
5862 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5863 .write_string = mem_cgroup_write,
5864 .read = mem_cgroup_read,
5867 .name = "kmem.usage_in_bytes",
5868 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5869 .read = mem_cgroup_read,
5872 .name = "kmem.failcnt",
5873 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5874 .trigger = mem_cgroup_reset,
5875 .read = mem_cgroup_read,
5878 .name = "kmem.max_usage_in_bytes",
5879 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5880 .trigger = mem_cgroup_reset,
5881 .read = mem_cgroup_read,
5883 #ifdef CONFIG_SLABINFO
5885 .name = "kmem.slabinfo",
5886 .read_seq_string = mem_cgroup_slabinfo_read,
5890 { }, /* terminate */
5893 #ifdef CONFIG_MEMCG_SWAP
5894 static struct cftype memsw_cgroup_files[] = {
5896 .name = "memsw.usage_in_bytes",
5897 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5898 .read = mem_cgroup_read,
5899 .register_event = mem_cgroup_usage_register_event,
5900 .unregister_event = mem_cgroup_usage_unregister_event,
5903 .name = "memsw.max_usage_in_bytes",
5904 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5905 .trigger = mem_cgroup_reset,
5906 .read = mem_cgroup_read,
5909 .name = "memsw.limit_in_bytes",
5910 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5911 .write_string = mem_cgroup_write,
5912 .read = mem_cgroup_read,
5915 .name = "memsw.failcnt",
5916 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5917 .trigger = mem_cgroup_reset,
5918 .read = mem_cgroup_read,
5920 { }, /* terminate */
5923 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5925 struct mem_cgroup_per_node *pn;
5926 struct mem_cgroup_per_zone *mz;
5927 int zone, tmp = node;
5929 * This routine is called against possible nodes.
5930 * But it's BUG to call kmalloc() against offline node.
5932 * TODO: this routine can waste much memory for nodes which will
5933 * never be onlined. It's better to use memory hotplug callback
5936 if (!node_state(node, N_NORMAL_MEMORY))
5938 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5942 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5943 mz = &pn->zoneinfo[zone];
5944 lruvec_init(&mz->lruvec);
5945 mz->usage_in_excess = 0;
5946 mz->on_tree = false;
5949 memcg->info.nodeinfo[node] = pn;
5953 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5955 kfree(memcg->info.nodeinfo[node]);
5958 static struct mem_cgroup *mem_cgroup_alloc(void)
5960 struct mem_cgroup *memcg;
5961 size_t size = memcg_size();
5963 /* Can be very big if nr_node_ids is very big */
5964 if (size < PAGE_SIZE)
5965 memcg = kzalloc(size, GFP_KERNEL);
5967 memcg = vzalloc(size);
5972 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5975 spin_lock_init(&memcg->pcp_counter_lock);
5979 if (size < PAGE_SIZE)
5987 * At destroying mem_cgroup, references from swap_cgroup can remain.
5988 * (scanning all at force_empty is too costly...)
5990 * Instead of clearing all references at force_empty, we remember
5991 * the number of reference from swap_cgroup and free mem_cgroup when
5992 * it goes down to 0.
5994 * Removal of cgroup itself succeeds regardless of refs from swap.
5997 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6000 size_t size = memcg_size();
6002 mem_cgroup_remove_from_trees(memcg);
6003 free_css_id(&mem_cgroup_subsys, &memcg->css);
6006 free_mem_cgroup_per_zone_info(memcg, node);
6008 free_percpu(memcg->stat);
6011 * We need to make sure that (at least for now), the jump label
6012 * destruction code runs outside of the cgroup lock. This is because
6013 * get_online_cpus(), which is called from the static_branch update,
6014 * can't be called inside the cgroup_lock. cpusets are the ones
6015 * enforcing this dependency, so if they ever change, we might as well.
6017 * schedule_work() will guarantee this happens. Be careful if you need
6018 * to move this code around, and make sure it is outside
6021 disarm_static_keys(memcg);
6022 if (size < PAGE_SIZE)
6030 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
6031 * but in process context. The work_freeing structure is overlaid
6032 * on the rcu_freeing structure, which itself is overlaid on memsw.
6034 static void free_work(struct work_struct *work)
6036 struct mem_cgroup *memcg;
6038 memcg = container_of(work, struct mem_cgroup, work_freeing);
6039 __mem_cgroup_free(memcg);
6042 static void free_rcu(struct rcu_head *rcu_head)
6044 struct mem_cgroup *memcg;
6046 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
6047 INIT_WORK(&memcg->work_freeing, free_work);
6048 schedule_work(&memcg->work_freeing);
6051 static void mem_cgroup_get(struct mem_cgroup *memcg)
6053 atomic_inc(&memcg->refcnt);
6056 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
6058 if (atomic_sub_and_test(count, &memcg->refcnt)) {
6059 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
6060 call_rcu(&memcg->rcu_freeing, free_rcu);
6062 mem_cgroup_put(parent);
6066 static void mem_cgroup_put(struct mem_cgroup *memcg)
6068 __mem_cgroup_put(memcg, 1);
6072 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6074 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6076 if (!memcg->res.parent)
6078 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6080 EXPORT_SYMBOL(parent_mem_cgroup);
6082 static int mem_cgroup_soft_limit_tree_init(void)
6084 struct mem_cgroup_tree_per_node *rtpn;
6085 struct mem_cgroup_tree_per_zone *rtpz;
6086 int tmp, node, zone;
6088 for_each_node(node) {
6090 if (!node_state(node, N_NORMAL_MEMORY))
6092 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6096 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6098 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6099 rtpz = &rtpn->rb_tree_per_zone[zone];
6100 rtpz->rb_root = RB_ROOT;
6101 spin_lock_init(&rtpz->lock);
6107 for_each_node(node) {
6108 if (!soft_limit_tree.rb_tree_per_node[node])
6110 kfree(soft_limit_tree.rb_tree_per_node[node]);
6111 soft_limit_tree.rb_tree_per_node[node] = NULL;
6117 static struct cgroup_subsys_state * __ref
6118 mem_cgroup_css_alloc(struct cgroup *cont)
6120 struct mem_cgroup *memcg;
6121 long error = -ENOMEM;
6124 memcg = mem_cgroup_alloc();
6126 return ERR_PTR(error);
6129 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6133 if (cont->parent == NULL) {
6136 if (mem_cgroup_soft_limit_tree_init())
6138 root_mem_cgroup = memcg;
6139 for_each_possible_cpu(cpu) {
6140 struct memcg_stock_pcp *stock =
6141 &per_cpu(memcg_stock, cpu);
6142 INIT_WORK(&stock->work, drain_local_stock);
6145 res_counter_init(&memcg->res, NULL);
6146 res_counter_init(&memcg->memsw, NULL);
6147 res_counter_init(&memcg->kmem, NULL);
6150 memcg->last_scanned_node = MAX_NUMNODES;
6151 INIT_LIST_HEAD(&memcg->oom_notify);
6152 atomic_set(&memcg->refcnt, 1);
6153 memcg->move_charge_at_immigrate = 0;
6154 mutex_init(&memcg->thresholds_lock);
6155 spin_lock_init(&memcg->move_lock);
6160 __mem_cgroup_free(memcg);
6161 return ERR_PTR(error);
6165 mem_cgroup_css_online(struct cgroup *cont)
6167 struct mem_cgroup *memcg, *parent;
6173 memcg = mem_cgroup_from_cont(cont);
6174 parent = mem_cgroup_from_cont(cont->parent);
6176 memcg->use_hierarchy = parent->use_hierarchy;
6177 memcg->oom_kill_disable = parent->oom_kill_disable;
6178 memcg->swappiness = mem_cgroup_swappiness(parent);
6180 if (parent->use_hierarchy) {
6181 res_counter_init(&memcg->res, &parent->res);
6182 res_counter_init(&memcg->memsw, &parent->memsw);
6183 res_counter_init(&memcg->kmem, &parent->kmem);
6186 * We increment refcnt of the parent to ensure that we can
6187 * safely access it on res_counter_charge/uncharge.
6188 * This refcnt will be decremented when freeing this
6189 * mem_cgroup(see mem_cgroup_put).
6191 mem_cgroup_get(parent);
6193 res_counter_init(&memcg->res, NULL);
6194 res_counter_init(&memcg->memsw, NULL);
6195 res_counter_init(&memcg->kmem, NULL);
6197 * Deeper hierachy with use_hierarchy == false doesn't make
6198 * much sense so let cgroup subsystem know about this
6199 * unfortunate state in our controller.
6201 if (parent != root_mem_cgroup)
6202 mem_cgroup_subsys.broken_hierarchy = true;
6205 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6208 * We call put now because our (and parent's) refcnts
6209 * are already in place. mem_cgroup_put() will internally
6210 * call __mem_cgroup_free, so return directly
6212 mem_cgroup_put(memcg);
6217 static void mem_cgroup_css_offline(struct cgroup *cont)
6219 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6221 mem_cgroup_reparent_charges(memcg);
6222 mem_cgroup_destroy_all_caches(memcg);
6225 static void mem_cgroup_css_free(struct cgroup *cont)
6227 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6229 kmem_cgroup_destroy(memcg);
6231 mem_cgroup_put(memcg);
6235 /* Handlers for move charge at task migration. */
6236 #define PRECHARGE_COUNT_AT_ONCE 256
6237 static int mem_cgroup_do_precharge(unsigned long count)
6240 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6241 struct mem_cgroup *memcg = mc.to;
6243 if (mem_cgroup_is_root(memcg)) {
6244 mc.precharge += count;
6245 /* we don't need css_get for root */
6248 /* try to charge at once */
6250 struct res_counter *dummy;
6252 * "memcg" cannot be under rmdir() because we've already checked
6253 * by cgroup_lock_live_cgroup() that it is not removed and we
6254 * are still under the same cgroup_mutex. So we can postpone
6257 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6259 if (do_swap_account && res_counter_charge(&memcg->memsw,
6260 PAGE_SIZE * count, &dummy)) {
6261 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6264 mc.precharge += count;
6268 /* fall back to one by one charge */
6270 if (signal_pending(current)) {
6274 if (!batch_count--) {
6275 batch_count = PRECHARGE_COUNT_AT_ONCE;
6278 ret = __mem_cgroup_try_charge(NULL,
6279 GFP_KERNEL, 1, &memcg, false);
6281 /* mem_cgroup_clear_mc() will do uncharge later */
6289 * get_mctgt_type - get target type of moving charge
6290 * @vma: the vma the pte to be checked belongs
6291 * @addr: the address corresponding to the pte to be checked
6292 * @ptent: the pte to be checked
6293 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6296 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6297 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6298 * move charge. if @target is not NULL, the page is stored in target->page
6299 * with extra refcnt got(Callers should handle it).
6300 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6301 * target for charge migration. if @target is not NULL, the entry is stored
6304 * Called with pte lock held.
6311 enum mc_target_type {
6317 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6318 unsigned long addr, pte_t ptent)
6320 struct page *page = vm_normal_page(vma, addr, ptent);
6322 if (!page || !page_mapped(page))
6324 if (PageAnon(page)) {
6325 /* we don't move shared anon */
6328 } else if (!move_file())
6329 /* we ignore mapcount for file pages */
6331 if (!get_page_unless_zero(page))
6338 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6339 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6341 struct page *page = NULL;
6342 swp_entry_t ent = pte_to_swp_entry(ptent);
6344 if (!move_anon() || non_swap_entry(ent))
6347 * Because lookup_swap_cache() updates some statistics counter,
6348 * we call find_get_page() with swapper_space directly.
6350 page = find_get_page(swap_address_space(ent), ent.val);
6351 if (do_swap_account)
6352 entry->val = ent.val;
6357 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6358 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6364 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6365 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6367 struct page *page = NULL;
6368 struct address_space *mapping;
6371 if (!vma->vm_file) /* anonymous vma */
6376 mapping = vma->vm_file->f_mapping;
6377 if (pte_none(ptent))
6378 pgoff = linear_page_index(vma, addr);
6379 else /* pte_file(ptent) is true */
6380 pgoff = pte_to_pgoff(ptent);
6382 /* page is moved even if it's not RSS of this task(page-faulted). */
6383 page = find_get_page(mapping, pgoff);
6386 /* shmem/tmpfs may report page out on swap: account for that too. */
6387 if (radix_tree_exceptional_entry(page)) {
6388 swp_entry_t swap = radix_to_swp_entry(page);
6389 if (do_swap_account)
6391 page = find_get_page(swap_address_space(swap), swap.val);
6397 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6398 unsigned long addr, pte_t ptent, union mc_target *target)
6400 struct page *page = NULL;
6401 struct page_cgroup *pc;
6402 enum mc_target_type ret = MC_TARGET_NONE;
6403 swp_entry_t ent = { .val = 0 };
6405 if (pte_present(ptent))
6406 page = mc_handle_present_pte(vma, addr, ptent);
6407 else if (is_swap_pte(ptent))
6408 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6409 else if (pte_none(ptent) || pte_file(ptent))
6410 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6412 if (!page && !ent.val)
6415 pc = lookup_page_cgroup(page);
6417 * Do only loose check w/o page_cgroup lock.
6418 * mem_cgroup_move_account() checks the pc is valid or not under
6421 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6422 ret = MC_TARGET_PAGE;
6424 target->page = page;
6426 if (!ret || !target)
6429 /* There is a swap entry and a page doesn't exist or isn't charged */
6430 if (ent.val && !ret &&
6431 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6432 ret = MC_TARGET_SWAP;
6439 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6441 * We don't consider swapping or file mapped pages because THP does not
6442 * support them for now.
6443 * Caller should make sure that pmd_trans_huge(pmd) is true.
6445 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6446 unsigned long addr, pmd_t pmd, union mc_target *target)
6448 struct page *page = NULL;
6449 struct page_cgroup *pc;
6450 enum mc_target_type ret = MC_TARGET_NONE;
6452 page = pmd_page(pmd);
6453 VM_BUG_ON(!page || !PageHead(page));
6456 pc = lookup_page_cgroup(page);
6457 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6458 ret = MC_TARGET_PAGE;
6461 target->page = page;
6467 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6468 unsigned long addr, pmd_t pmd, union mc_target *target)
6470 return MC_TARGET_NONE;
6474 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6475 unsigned long addr, unsigned long end,
6476 struct mm_walk *walk)
6478 struct vm_area_struct *vma = walk->private;
6482 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6483 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6484 mc.precharge += HPAGE_PMD_NR;
6485 spin_unlock(&vma->vm_mm->page_table_lock);
6489 if (pmd_trans_unstable(pmd))
6491 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6492 for (; addr != end; pte++, addr += PAGE_SIZE)
6493 if (get_mctgt_type(vma, addr, *pte, NULL))
6494 mc.precharge++; /* increment precharge temporarily */
6495 pte_unmap_unlock(pte - 1, ptl);
6501 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6503 unsigned long precharge;
6504 struct vm_area_struct *vma;
6506 down_read(&mm->mmap_sem);
6507 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6508 struct mm_walk mem_cgroup_count_precharge_walk = {
6509 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6513 if (is_vm_hugetlb_page(vma))
6515 walk_page_range(vma->vm_start, vma->vm_end,
6516 &mem_cgroup_count_precharge_walk);
6518 up_read(&mm->mmap_sem);
6520 precharge = mc.precharge;
6526 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6528 unsigned long precharge = mem_cgroup_count_precharge(mm);
6530 VM_BUG_ON(mc.moving_task);
6531 mc.moving_task = current;
6532 return mem_cgroup_do_precharge(precharge);
6535 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6536 static void __mem_cgroup_clear_mc(void)
6538 struct mem_cgroup *from = mc.from;
6539 struct mem_cgroup *to = mc.to;
6541 /* we must uncharge all the leftover precharges from mc.to */
6543 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6547 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6548 * we must uncharge here.
6550 if (mc.moved_charge) {
6551 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6552 mc.moved_charge = 0;
6554 /* we must fixup refcnts and charges */
6555 if (mc.moved_swap) {
6556 /* uncharge swap account from the old cgroup */
6557 if (!mem_cgroup_is_root(mc.from))
6558 res_counter_uncharge(&mc.from->memsw,
6559 PAGE_SIZE * mc.moved_swap);
6560 __mem_cgroup_put(mc.from, mc.moved_swap);
6562 if (!mem_cgroup_is_root(mc.to)) {
6564 * we charged both to->res and to->memsw, so we should
6567 res_counter_uncharge(&mc.to->res,
6568 PAGE_SIZE * mc.moved_swap);
6570 /* we've already done mem_cgroup_get(mc.to) */
6573 memcg_oom_recover(from);
6574 memcg_oom_recover(to);
6575 wake_up_all(&mc.waitq);
6578 static void mem_cgroup_clear_mc(void)
6580 struct mem_cgroup *from = mc.from;
6583 * we must clear moving_task before waking up waiters at the end of
6586 mc.moving_task = NULL;
6587 __mem_cgroup_clear_mc();
6588 spin_lock(&mc.lock);
6591 spin_unlock(&mc.lock);
6592 mem_cgroup_end_move(from);
6595 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6596 struct cgroup_taskset *tset)
6598 struct task_struct *p = cgroup_taskset_first(tset);
6600 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6601 unsigned long move_charge_at_immigrate;
6604 * We are now commited to this value whatever it is. Changes in this
6605 * tunable will only affect upcoming migrations, not the current one.
6606 * So we need to save it, and keep it going.
6608 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6609 if (move_charge_at_immigrate) {
6610 struct mm_struct *mm;
6611 struct mem_cgroup *from = mem_cgroup_from_task(p);
6613 VM_BUG_ON(from == memcg);
6615 mm = get_task_mm(p);
6618 /* We move charges only when we move a owner of the mm */
6619 if (mm->owner == p) {
6622 VM_BUG_ON(mc.precharge);
6623 VM_BUG_ON(mc.moved_charge);
6624 VM_BUG_ON(mc.moved_swap);
6625 mem_cgroup_start_move(from);
6626 spin_lock(&mc.lock);
6629 mc.immigrate_flags = move_charge_at_immigrate;
6630 spin_unlock(&mc.lock);
6631 /* We set mc.moving_task later */
6633 ret = mem_cgroup_precharge_mc(mm);
6635 mem_cgroup_clear_mc();
6642 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6643 struct cgroup_taskset *tset)
6645 mem_cgroup_clear_mc();
6648 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6649 unsigned long addr, unsigned long end,
6650 struct mm_walk *walk)
6653 struct vm_area_struct *vma = walk->private;
6656 enum mc_target_type target_type;
6657 union mc_target target;
6659 struct page_cgroup *pc;
6662 * We don't take compound_lock() here but no race with splitting thp
6664 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6665 * under splitting, which means there's no concurrent thp split,
6666 * - if another thread runs into split_huge_page() just after we
6667 * entered this if-block, the thread must wait for page table lock
6668 * to be unlocked in __split_huge_page_splitting(), where the main
6669 * part of thp split is not executed yet.
6671 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6672 if (mc.precharge < HPAGE_PMD_NR) {
6673 spin_unlock(&vma->vm_mm->page_table_lock);
6676 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6677 if (target_type == MC_TARGET_PAGE) {
6679 if (!isolate_lru_page(page)) {
6680 pc = lookup_page_cgroup(page);
6681 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6682 pc, mc.from, mc.to)) {
6683 mc.precharge -= HPAGE_PMD_NR;
6684 mc.moved_charge += HPAGE_PMD_NR;
6686 putback_lru_page(page);
6690 spin_unlock(&vma->vm_mm->page_table_lock);
6694 if (pmd_trans_unstable(pmd))
6697 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6698 for (; addr != end; addr += PAGE_SIZE) {
6699 pte_t ptent = *(pte++);
6705 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6706 case MC_TARGET_PAGE:
6708 if (isolate_lru_page(page))
6710 pc = lookup_page_cgroup(page);
6711 if (!mem_cgroup_move_account(page, 1, pc,
6714 /* we uncharge from mc.from later. */
6717 putback_lru_page(page);
6718 put: /* get_mctgt_type() gets the page */
6721 case MC_TARGET_SWAP:
6723 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6725 /* we fixup refcnts and charges later. */
6733 pte_unmap_unlock(pte - 1, ptl);
6738 * We have consumed all precharges we got in can_attach().
6739 * We try charge one by one, but don't do any additional
6740 * charges to mc.to if we have failed in charge once in attach()
6743 ret = mem_cgroup_do_precharge(1);
6751 static void mem_cgroup_move_charge(struct mm_struct *mm)
6753 struct vm_area_struct *vma;
6755 lru_add_drain_all();
6757 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6759 * Someone who are holding the mmap_sem might be waiting in
6760 * waitq. So we cancel all extra charges, wake up all waiters,
6761 * and retry. Because we cancel precharges, we might not be able
6762 * to move enough charges, but moving charge is a best-effort
6763 * feature anyway, so it wouldn't be a big problem.
6765 __mem_cgroup_clear_mc();
6769 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6771 struct mm_walk mem_cgroup_move_charge_walk = {
6772 .pmd_entry = mem_cgroup_move_charge_pte_range,
6776 if (is_vm_hugetlb_page(vma))
6778 ret = walk_page_range(vma->vm_start, vma->vm_end,
6779 &mem_cgroup_move_charge_walk);
6782 * means we have consumed all precharges and failed in
6783 * doing additional charge. Just abandon here.
6787 up_read(&mm->mmap_sem);
6790 static void mem_cgroup_move_task(struct cgroup *cont,
6791 struct cgroup_taskset *tset)
6793 struct task_struct *p = cgroup_taskset_first(tset);
6794 struct mm_struct *mm = get_task_mm(p);
6798 mem_cgroup_move_charge(mm);
6802 mem_cgroup_clear_mc();
6804 #else /* !CONFIG_MMU */
6805 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6806 struct cgroup_taskset *tset)
6810 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6811 struct cgroup_taskset *tset)
6814 static void mem_cgroup_move_task(struct cgroup *cont,
6815 struct cgroup_taskset *tset)
6820 struct cgroup_subsys mem_cgroup_subsys = {
6822 .subsys_id = mem_cgroup_subsys_id,
6823 .css_alloc = mem_cgroup_css_alloc,
6824 .css_online = mem_cgroup_css_online,
6825 .css_offline = mem_cgroup_css_offline,
6826 .css_free = mem_cgroup_css_free,
6827 .can_attach = mem_cgroup_can_attach,
6828 .cancel_attach = mem_cgroup_cancel_attach,
6829 .attach = mem_cgroup_move_task,
6830 .base_cftypes = mem_cgroup_files,
6835 #ifdef CONFIG_MEMCG_SWAP
6836 static int __init enable_swap_account(char *s)
6838 /* consider enabled if no parameter or 1 is given */
6839 if (!strcmp(s, "1"))
6840 really_do_swap_account = 1;
6841 else if (!strcmp(s, "0"))
6842 really_do_swap_account = 0;
6845 __setup("swapaccount=", enable_swap_account);
6847 static void __init memsw_file_init(void)
6849 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6852 static void __init enable_swap_cgroup(void)
6854 if (!mem_cgroup_disabled() && really_do_swap_account) {
6855 do_swap_account = 1;
6861 static void __init enable_swap_cgroup(void)
6867 * The rest of init is performed during ->css_alloc() for root css which
6868 * happens before initcalls. hotcpu_notifier() can't be done together as
6869 * it would introduce circular locking by adding cgroup_lock -> cpu hotplug
6870 * dependency. Do it from a subsys_initcall().
6872 static int __init mem_cgroup_init(void)
6874 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6875 enable_swap_cgroup();
6878 subsys_initcall(mem_cgroup_init);