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 /* last scanned hierarchy member with elevated css ref count */
156 struct mem_cgroup *last_visited;
157 /* scan generation, increased every round-trip */
158 unsigned int generation;
159 /* lock to protect the position and generation */
160 spinlock_t iter_lock;
164 * per-zone information in memory controller.
166 struct mem_cgroup_per_zone {
167 struct lruvec lruvec;
168 unsigned long lru_size[NR_LRU_LISTS];
170 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
172 struct rb_node tree_node; /* RB tree node */
173 unsigned long long usage_in_excess;/* Set to the value by which */
174 /* the soft limit is exceeded*/
176 struct mem_cgroup *memcg; /* Back pointer, we cannot */
177 /* use container_of */
180 struct mem_cgroup_per_node {
181 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
184 struct mem_cgroup_lru_info {
185 struct mem_cgroup_per_node *nodeinfo[0];
189 * Cgroups above their limits are maintained in a RB-Tree, independent of
190 * their hierarchy representation
193 struct mem_cgroup_tree_per_zone {
194 struct rb_root rb_root;
198 struct mem_cgroup_tree_per_node {
199 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
202 struct mem_cgroup_tree {
203 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
206 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
208 struct mem_cgroup_threshold {
209 struct eventfd_ctx *eventfd;
214 struct mem_cgroup_threshold_ary {
215 /* An array index points to threshold just below or equal to usage. */
216 int current_threshold;
217 /* Size of entries[] */
219 /* Array of thresholds */
220 struct mem_cgroup_threshold entries[0];
223 struct mem_cgroup_thresholds {
224 /* Primary thresholds array */
225 struct mem_cgroup_threshold_ary *primary;
227 * Spare threshold array.
228 * This is needed to make mem_cgroup_unregister_event() "never fail".
229 * It must be able to store at least primary->size - 1 entries.
231 struct mem_cgroup_threshold_ary *spare;
235 struct mem_cgroup_eventfd_list {
236 struct list_head list;
237 struct eventfd_ctx *eventfd;
240 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
241 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
244 * The memory controller data structure. The memory controller controls both
245 * page cache and RSS per cgroup. We would eventually like to provide
246 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
247 * to help the administrator determine what knobs to tune.
249 * TODO: Add a water mark for the memory controller. Reclaim will begin when
250 * we hit the water mark. May be even add a low water mark, such that
251 * no reclaim occurs from a cgroup at it's low water mark, this is
252 * a feature that will be implemented much later in the future.
255 struct cgroup_subsys_state css;
257 * the counter to account for memory usage
259 struct res_counter res;
263 * the counter to account for mem+swap usage.
265 struct res_counter memsw;
268 * rcu_freeing is used only when freeing struct mem_cgroup,
269 * so put it into a union to avoid wasting more memory.
270 * It must be disjoint from the css field. It could be
271 * in a union with the res field, but res plays a much
272 * larger part in mem_cgroup life than memsw, and might
273 * be of interest, even at time of free, when debugging.
274 * So share rcu_head with the less interesting memsw.
276 struct rcu_head rcu_freeing;
278 * We also need some space for a worker in deferred freeing.
279 * By the time we call it, rcu_freeing is no longer in use.
281 struct work_struct work_freeing;
285 * the counter to account for kernel memory usage.
287 struct res_counter kmem;
289 * Should the accounting and control be hierarchical, per subtree?
292 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
300 /* OOM-Killer disable */
301 int oom_kill_disable;
303 /* set when res.limit == memsw.limit */
304 bool memsw_is_minimum;
306 /* protect arrays of thresholds */
307 struct mutex thresholds_lock;
309 /* thresholds for memory usage. RCU-protected */
310 struct mem_cgroup_thresholds thresholds;
312 /* thresholds for mem+swap usage. RCU-protected */
313 struct mem_cgroup_thresholds memsw_thresholds;
315 /* For oom notifier event fd */
316 struct list_head oom_notify;
319 * Should we move charges of a task when a task is moved into this
320 * mem_cgroup ? And what type of charges should we move ?
322 unsigned long move_charge_at_immigrate;
324 * set > 0 if pages under this cgroup are moving to other cgroup.
326 atomic_t moving_account;
327 /* taken only while moving_account > 0 */
328 spinlock_t move_lock;
332 struct mem_cgroup_stat_cpu __percpu *stat;
334 * used when a cpu is offlined or other synchronizations
335 * See mem_cgroup_read_stat().
337 struct mem_cgroup_stat_cpu nocpu_base;
338 spinlock_t pcp_counter_lock;
340 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
341 struct tcp_memcontrol tcp_mem;
343 #if defined(CONFIG_MEMCG_KMEM)
344 /* analogous to slab_common's slab_caches list. per-memcg */
345 struct list_head memcg_slab_caches;
346 /* Not a spinlock, we can take a lot of time walking the list */
347 struct mutex slab_caches_mutex;
348 /* Index in the kmem_cache->memcg_params->memcg_caches array */
352 int last_scanned_node;
354 nodemask_t scan_nodes;
355 atomic_t numainfo_events;
356 atomic_t numainfo_updating;
359 * Per cgroup active and inactive list, similar to the
360 * per zone LRU lists.
362 * WARNING: This has to be the last element of the struct. Don't
363 * add new fields after this point.
365 struct mem_cgroup_lru_info info;
368 static size_t memcg_size(void)
370 return sizeof(struct mem_cgroup) +
371 nr_node_ids * sizeof(struct mem_cgroup_per_node);
374 /* internal only representation about the status of kmem accounting. */
376 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
377 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
378 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
381 /* We account when limit is on, but only after call sites are patched */
382 #define KMEM_ACCOUNTED_MASK \
383 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
385 #ifdef CONFIG_MEMCG_KMEM
386 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
388 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
391 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
393 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
396 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
398 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
401 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
403 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
406 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
408 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
409 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
412 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
414 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
415 &memcg->kmem_account_flags);
419 /* Stuffs for move charges at task migration. */
421 * Types of charges to be moved. "move_charge_at_immitgrate" and
422 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
425 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
426 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
430 /* "mc" and its members are protected by cgroup_mutex */
431 static struct move_charge_struct {
432 spinlock_t lock; /* for from, to */
433 struct mem_cgroup *from;
434 struct mem_cgroup *to;
435 unsigned long immigrate_flags;
436 unsigned long precharge;
437 unsigned long moved_charge;
438 unsigned long moved_swap;
439 struct task_struct *moving_task; /* a task moving charges */
440 wait_queue_head_t waitq; /* a waitq for other context */
442 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
443 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
446 static bool move_anon(void)
448 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
451 static bool move_file(void)
453 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
457 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
458 * limit reclaim to prevent infinite loops, if they ever occur.
460 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
461 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
464 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
465 MEM_CGROUP_CHARGE_TYPE_ANON,
466 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
467 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
471 /* for encoding cft->private value on file */
479 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
480 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
481 #define MEMFILE_ATTR(val) ((val) & 0xffff)
482 /* Used for OOM nofiier */
483 #define OOM_CONTROL (0)
486 * Reclaim flags for mem_cgroup_hierarchical_reclaim
488 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
489 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
490 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
491 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
494 * The memcg_create_mutex will be held whenever a new cgroup is created.
495 * As a consequence, any change that needs to protect against new child cgroups
496 * appearing has to hold it as well.
498 static DEFINE_MUTEX(memcg_create_mutex);
500 static void mem_cgroup_get(struct mem_cgroup *memcg);
501 static void mem_cgroup_put(struct mem_cgroup *memcg);
504 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
506 return container_of(s, struct mem_cgroup, css);
509 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
511 return (memcg == root_mem_cgroup);
514 /* Writing them here to avoid exposing memcg's inner layout */
515 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
517 void sock_update_memcg(struct sock *sk)
519 if (mem_cgroup_sockets_enabled) {
520 struct mem_cgroup *memcg;
521 struct cg_proto *cg_proto;
523 BUG_ON(!sk->sk_prot->proto_cgroup);
525 /* Socket cloning can throw us here with sk_cgrp already
526 * filled. It won't however, necessarily happen from
527 * process context. So the test for root memcg given
528 * the current task's memcg won't help us in this case.
530 * Respecting the original socket's memcg is a better
531 * decision in this case.
534 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
535 mem_cgroup_get(sk->sk_cgrp->memcg);
540 memcg = mem_cgroup_from_task(current);
541 cg_proto = sk->sk_prot->proto_cgroup(memcg);
542 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
543 mem_cgroup_get(memcg);
544 sk->sk_cgrp = cg_proto;
549 EXPORT_SYMBOL(sock_update_memcg);
551 void sock_release_memcg(struct sock *sk)
553 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
554 struct mem_cgroup *memcg;
555 WARN_ON(!sk->sk_cgrp->memcg);
556 memcg = sk->sk_cgrp->memcg;
557 mem_cgroup_put(memcg);
561 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
563 if (!memcg || mem_cgroup_is_root(memcg))
566 return &memcg->tcp_mem.cg_proto;
568 EXPORT_SYMBOL(tcp_proto_cgroup);
570 static void disarm_sock_keys(struct mem_cgroup *memcg)
572 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
574 static_key_slow_dec(&memcg_socket_limit_enabled);
577 static void disarm_sock_keys(struct mem_cgroup *memcg)
582 #ifdef CONFIG_MEMCG_KMEM
584 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
585 * There are two main reasons for not using the css_id for this:
586 * 1) this works better in sparse environments, where we have a lot of memcgs,
587 * but only a few kmem-limited. Or also, if we have, for instance, 200
588 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
589 * 200 entry array for that.
591 * 2) In order not to violate the cgroup API, we would like to do all memory
592 * allocation in ->create(). At that point, we haven't yet allocated the
593 * css_id. Having a separate index prevents us from messing with the cgroup
596 * The current size of the caches array is stored in
597 * memcg_limited_groups_array_size. It will double each time we have to
600 static DEFINE_IDA(kmem_limited_groups);
601 int memcg_limited_groups_array_size;
604 * MIN_SIZE is different than 1, because we would like to avoid going through
605 * the alloc/free process all the time. In a small machine, 4 kmem-limited
606 * cgroups is a reasonable guess. In the future, it could be a parameter or
607 * tunable, but that is strictly not necessary.
609 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
610 * this constant directly from cgroup, but it is understandable that this is
611 * better kept as an internal representation in cgroup.c. In any case, the
612 * css_id space is not getting any smaller, and we don't have to necessarily
613 * increase ours as well if it increases.
615 #define MEMCG_CACHES_MIN_SIZE 4
616 #define MEMCG_CACHES_MAX_SIZE 65535
619 * A lot of the calls to the cache allocation functions are expected to be
620 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
621 * conditional to this static branch, we'll have to allow modules that does
622 * kmem_cache_alloc and the such to see this symbol as well
624 struct static_key memcg_kmem_enabled_key;
625 EXPORT_SYMBOL(memcg_kmem_enabled_key);
627 static void disarm_kmem_keys(struct mem_cgroup *memcg)
629 if (memcg_kmem_is_active(memcg)) {
630 static_key_slow_dec(&memcg_kmem_enabled_key);
631 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
634 * This check can't live in kmem destruction function,
635 * since the charges will outlive the cgroup
637 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
640 static void disarm_kmem_keys(struct mem_cgroup *memcg)
643 #endif /* CONFIG_MEMCG_KMEM */
645 static void disarm_static_keys(struct mem_cgroup *memcg)
647 disarm_sock_keys(memcg);
648 disarm_kmem_keys(memcg);
651 static void drain_all_stock_async(struct mem_cgroup *memcg);
653 static struct mem_cgroup_per_zone *
654 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
656 VM_BUG_ON((unsigned)nid >= nr_node_ids);
657 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
660 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
665 static struct mem_cgroup_per_zone *
666 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
668 int nid = page_to_nid(page);
669 int zid = page_zonenum(page);
671 return mem_cgroup_zoneinfo(memcg, nid, zid);
674 static struct mem_cgroup_tree_per_zone *
675 soft_limit_tree_node_zone(int nid, int zid)
677 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
680 static struct mem_cgroup_tree_per_zone *
681 soft_limit_tree_from_page(struct page *page)
683 int nid = page_to_nid(page);
684 int zid = page_zonenum(page);
686 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
690 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
691 struct mem_cgroup_per_zone *mz,
692 struct mem_cgroup_tree_per_zone *mctz,
693 unsigned long long new_usage_in_excess)
695 struct rb_node **p = &mctz->rb_root.rb_node;
696 struct rb_node *parent = NULL;
697 struct mem_cgroup_per_zone *mz_node;
702 mz->usage_in_excess = new_usage_in_excess;
703 if (!mz->usage_in_excess)
707 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
709 if (mz->usage_in_excess < mz_node->usage_in_excess)
712 * We can't avoid mem cgroups that are over their soft
713 * limit by the same amount
715 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
718 rb_link_node(&mz->tree_node, parent, p);
719 rb_insert_color(&mz->tree_node, &mctz->rb_root);
724 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
725 struct mem_cgroup_per_zone *mz,
726 struct mem_cgroup_tree_per_zone *mctz)
730 rb_erase(&mz->tree_node, &mctz->rb_root);
735 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
736 struct mem_cgroup_per_zone *mz,
737 struct mem_cgroup_tree_per_zone *mctz)
739 spin_lock(&mctz->lock);
740 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
741 spin_unlock(&mctz->lock);
745 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
747 unsigned long long excess;
748 struct mem_cgroup_per_zone *mz;
749 struct mem_cgroup_tree_per_zone *mctz;
750 int nid = page_to_nid(page);
751 int zid = page_zonenum(page);
752 mctz = soft_limit_tree_from_page(page);
755 * Necessary to update all ancestors when hierarchy is used.
756 * because their event counter is not touched.
758 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
759 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
760 excess = res_counter_soft_limit_excess(&memcg->res);
762 * We have to update the tree if mz is on RB-tree or
763 * mem is over its softlimit.
765 if (excess || mz->on_tree) {
766 spin_lock(&mctz->lock);
767 /* if on-tree, remove it */
769 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
771 * Insert again. mz->usage_in_excess will be updated.
772 * If excess is 0, no tree ops.
774 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
775 spin_unlock(&mctz->lock);
780 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
783 struct mem_cgroup_per_zone *mz;
784 struct mem_cgroup_tree_per_zone *mctz;
786 for_each_node(node) {
787 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
788 mz = mem_cgroup_zoneinfo(memcg, node, zone);
789 mctz = soft_limit_tree_node_zone(node, zone);
790 mem_cgroup_remove_exceeded(memcg, mz, mctz);
795 static struct mem_cgroup_per_zone *
796 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
798 struct rb_node *rightmost = NULL;
799 struct mem_cgroup_per_zone *mz;
803 rightmost = rb_last(&mctz->rb_root);
805 goto done; /* Nothing to reclaim from */
807 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
809 * Remove the node now but someone else can add it back,
810 * we will to add it back at the end of reclaim to its correct
811 * position in the tree.
813 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
814 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
815 !css_tryget(&mz->memcg->css))
821 static struct mem_cgroup_per_zone *
822 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
824 struct mem_cgroup_per_zone *mz;
826 spin_lock(&mctz->lock);
827 mz = __mem_cgroup_largest_soft_limit_node(mctz);
828 spin_unlock(&mctz->lock);
833 * Implementation Note: reading percpu statistics for memcg.
835 * Both of vmstat[] and percpu_counter has threshold and do periodic
836 * synchronization to implement "quick" read. There are trade-off between
837 * reading cost and precision of value. Then, we may have a chance to implement
838 * a periodic synchronizion of counter in memcg's counter.
840 * But this _read() function is used for user interface now. The user accounts
841 * memory usage by memory cgroup and he _always_ requires exact value because
842 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
843 * have to visit all online cpus and make sum. So, for now, unnecessary
844 * synchronization is not implemented. (just implemented for cpu hotplug)
846 * If there are kernel internal actions which can make use of some not-exact
847 * value, and reading all cpu value can be performance bottleneck in some
848 * common workload, threashold and synchonization as vmstat[] should be
851 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
852 enum mem_cgroup_stat_index idx)
858 for_each_online_cpu(cpu)
859 val += per_cpu(memcg->stat->count[idx], cpu);
860 #ifdef CONFIG_HOTPLUG_CPU
861 spin_lock(&memcg->pcp_counter_lock);
862 val += memcg->nocpu_base.count[idx];
863 spin_unlock(&memcg->pcp_counter_lock);
869 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
872 int val = (charge) ? 1 : -1;
873 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
876 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
877 enum mem_cgroup_events_index idx)
879 unsigned long val = 0;
882 for_each_online_cpu(cpu)
883 val += per_cpu(memcg->stat->events[idx], cpu);
884 #ifdef CONFIG_HOTPLUG_CPU
885 spin_lock(&memcg->pcp_counter_lock);
886 val += memcg->nocpu_base.events[idx];
887 spin_unlock(&memcg->pcp_counter_lock);
892 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
893 bool anon, int nr_pages)
898 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
899 * counted as CACHE even if it's on ANON LRU.
902 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
905 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
908 /* pagein of a big page is an event. So, ignore page size */
910 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
912 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
913 nr_pages = -nr_pages; /* for event */
916 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
922 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
924 struct mem_cgroup_per_zone *mz;
926 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
927 return mz->lru_size[lru];
931 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
932 unsigned int lru_mask)
934 struct mem_cgroup_per_zone *mz;
936 unsigned long ret = 0;
938 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
941 if (BIT(lru) & lru_mask)
942 ret += mz->lru_size[lru];
948 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
949 int nid, unsigned int lru_mask)
954 for (zid = 0; zid < MAX_NR_ZONES; zid++)
955 total += mem_cgroup_zone_nr_lru_pages(memcg,
961 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
962 unsigned int lru_mask)
967 for_each_node_state(nid, N_MEMORY)
968 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
972 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
973 enum mem_cgroup_events_target target)
975 unsigned long val, next;
977 val = __this_cpu_read(memcg->stat->nr_page_events);
978 next = __this_cpu_read(memcg->stat->targets[target]);
979 /* from time_after() in jiffies.h */
980 if ((long)next - (long)val < 0) {
982 case MEM_CGROUP_TARGET_THRESH:
983 next = val + THRESHOLDS_EVENTS_TARGET;
985 case MEM_CGROUP_TARGET_SOFTLIMIT:
986 next = val + SOFTLIMIT_EVENTS_TARGET;
988 case MEM_CGROUP_TARGET_NUMAINFO:
989 next = val + NUMAINFO_EVENTS_TARGET;
994 __this_cpu_write(memcg->stat->targets[target], next);
1001 * Check events in order.
1004 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1007 /* threshold event is triggered in finer grain than soft limit */
1008 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1009 MEM_CGROUP_TARGET_THRESH))) {
1011 bool do_numainfo __maybe_unused;
1013 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1014 MEM_CGROUP_TARGET_SOFTLIMIT);
1015 #if MAX_NUMNODES > 1
1016 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1017 MEM_CGROUP_TARGET_NUMAINFO);
1021 mem_cgroup_threshold(memcg);
1022 if (unlikely(do_softlimit))
1023 mem_cgroup_update_tree(memcg, page);
1024 #if MAX_NUMNODES > 1
1025 if (unlikely(do_numainfo))
1026 atomic_inc(&memcg->numainfo_events);
1032 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1034 return mem_cgroup_from_css(
1035 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
1038 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1041 * mm_update_next_owner() may clear mm->owner to NULL
1042 * if it races with swapoff, page migration, etc.
1043 * So this can be called with p == NULL.
1048 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
1051 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1053 struct mem_cgroup *memcg = NULL;
1058 * Because we have no locks, mm->owner's may be being moved to other
1059 * cgroup. We use css_tryget() here even if this looks
1060 * pessimistic (rather than adding locks here).
1064 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1065 if (unlikely(!memcg))
1067 } while (!css_tryget(&memcg->css));
1073 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1074 * @root: hierarchy root
1075 * @prev: previously returned memcg, NULL on first invocation
1076 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1078 * Returns references to children of the hierarchy below @root, or
1079 * @root itself, or %NULL after a full round-trip.
1081 * Caller must pass the return value in @prev on subsequent
1082 * invocations for reference counting, or use mem_cgroup_iter_break()
1083 * to cancel a hierarchy walk before the round-trip is complete.
1085 * Reclaimers can specify a zone and a priority level in @reclaim to
1086 * divide up the memcgs in the hierarchy among all concurrent
1087 * reclaimers operating on the same zone and priority.
1089 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1090 struct mem_cgroup *prev,
1091 struct mem_cgroup_reclaim_cookie *reclaim)
1093 struct mem_cgroup *memcg = NULL;
1094 struct mem_cgroup *last_visited = NULL;
1096 if (mem_cgroup_disabled())
1100 root = root_mem_cgroup;
1102 if (prev && !reclaim)
1103 last_visited = prev;
1105 if (!root->use_hierarchy && root != root_mem_cgroup) {
1113 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1114 struct cgroup_subsys_state *css = NULL;
1117 int nid = zone_to_nid(reclaim->zone);
1118 int zid = zone_idx(reclaim->zone);
1119 struct mem_cgroup_per_zone *mz;
1121 mz = mem_cgroup_zoneinfo(root, nid, zid);
1122 iter = &mz->reclaim_iter[reclaim->priority];
1123 spin_lock(&iter->iter_lock);
1124 last_visited = iter->last_visited;
1125 if (prev && reclaim->generation != iter->generation) {
1127 css_put(&last_visited->css);
1128 iter->last_visited = NULL;
1130 spin_unlock(&iter->iter_lock);
1136 * Root is not visited by cgroup iterators so it needs an
1139 if (!last_visited) {
1142 struct cgroup *prev_cgroup, *next_cgroup;
1144 prev_cgroup = (last_visited == root) ? NULL
1145 : last_visited->css.cgroup;
1146 next_cgroup = cgroup_next_descendant_pre(prev_cgroup,
1149 css = cgroup_subsys_state(next_cgroup,
1150 mem_cgroup_subsys_id);
1154 * Even if we found a group we have to make sure it is alive.
1155 * css && !memcg means that the groups should be skipped and
1156 * we should continue the tree walk.
1157 * last_visited css is safe to use because it is protected by
1158 * css_get and the tree walk is rcu safe.
1160 if (css == &root->css || (css && css_tryget(css)))
1161 memcg = mem_cgroup_from_css(css);
1164 struct mem_cgroup *curr = memcg;
1167 css_put(&last_visited->css);
1170 curr = mem_cgroup_from_css(css);
1172 /* make sure that the cached memcg is not removed */
1174 css_get(&curr->css);
1175 iter->last_visited = curr;
1179 else if (!prev && memcg)
1180 reclaim->generation = iter->generation;
1181 spin_unlock(&iter->iter_lock);
1182 } else if (css && !memcg) {
1183 last_visited = mem_cgroup_from_css(css);
1192 if (prev && prev != root)
1193 css_put(&prev->css);
1199 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1200 * @root: hierarchy root
1201 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1203 void mem_cgroup_iter_break(struct mem_cgroup *root,
1204 struct mem_cgroup *prev)
1207 root = root_mem_cgroup;
1208 if (prev && prev != root)
1209 css_put(&prev->css);
1213 * Iteration constructs for visiting all cgroups (under a tree). If
1214 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1215 * be used for reference counting.
1217 #define for_each_mem_cgroup_tree(iter, root) \
1218 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1220 iter = mem_cgroup_iter(root, iter, NULL))
1222 #define for_each_mem_cgroup(iter) \
1223 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1225 iter = mem_cgroup_iter(NULL, iter, NULL))
1227 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1229 struct mem_cgroup *memcg;
1232 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1233 if (unlikely(!memcg))
1238 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1241 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1249 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1252 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1253 * @zone: zone of the wanted lruvec
1254 * @memcg: memcg of the wanted lruvec
1256 * Returns the lru list vector holding pages for the given @zone and
1257 * @mem. This can be the global zone lruvec, if the memory controller
1260 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1261 struct mem_cgroup *memcg)
1263 struct mem_cgroup_per_zone *mz;
1264 struct lruvec *lruvec;
1266 if (mem_cgroup_disabled()) {
1267 lruvec = &zone->lruvec;
1271 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1272 lruvec = &mz->lruvec;
1275 * Since a node can be onlined after the mem_cgroup was created,
1276 * we have to be prepared to initialize lruvec->zone here;
1277 * and if offlined then reonlined, we need to reinitialize it.
1279 if (unlikely(lruvec->zone != zone))
1280 lruvec->zone = zone;
1285 * Following LRU functions are allowed to be used without PCG_LOCK.
1286 * Operations are called by routine of global LRU independently from memcg.
1287 * What we have to take care of here is validness of pc->mem_cgroup.
1289 * Changes to pc->mem_cgroup happens when
1292 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1293 * It is added to LRU before charge.
1294 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1295 * When moving account, the page is not on LRU. It's isolated.
1299 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1301 * @zone: zone of the page
1303 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1305 struct mem_cgroup_per_zone *mz;
1306 struct mem_cgroup *memcg;
1307 struct page_cgroup *pc;
1308 struct lruvec *lruvec;
1310 if (mem_cgroup_disabled()) {
1311 lruvec = &zone->lruvec;
1315 pc = lookup_page_cgroup(page);
1316 memcg = pc->mem_cgroup;
1319 * Surreptitiously switch any uncharged offlist page to root:
1320 * an uncharged page off lru does nothing to secure
1321 * its former mem_cgroup from sudden removal.
1323 * Our caller holds lru_lock, and PageCgroupUsed is updated
1324 * under page_cgroup lock: between them, they make all uses
1325 * of pc->mem_cgroup safe.
1327 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1328 pc->mem_cgroup = memcg = root_mem_cgroup;
1330 mz = page_cgroup_zoneinfo(memcg, page);
1331 lruvec = &mz->lruvec;
1334 * Since a node can be onlined after the mem_cgroup was created,
1335 * we have to be prepared to initialize lruvec->zone here;
1336 * and if offlined then reonlined, we need to reinitialize it.
1338 if (unlikely(lruvec->zone != zone))
1339 lruvec->zone = zone;
1344 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1345 * @lruvec: mem_cgroup per zone lru vector
1346 * @lru: index of lru list the page is sitting on
1347 * @nr_pages: positive when adding or negative when removing
1349 * This function must be called when a page is added to or removed from an
1352 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1355 struct mem_cgroup_per_zone *mz;
1356 unsigned long *lru_size;
1358 if (mem_cgroup_disabled())
1361 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1362 lru_size = mz->lru_size + lru;
1363 *lru_size += nr_pages;
1364 VM_BUG_ON((long)(*lru_size) < 0);
1368 * Checks whether given mem is same or in the root_mem_cgroup's
1371 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1372 struct mem_cgroup *memcg)
1374 if (root_memcg == memcg)
1376 if (!root_memcg->use_hierarchy || !memcg)
1378 return css_is_ancestor(&memcg->css, &root_memcg->css);
1381 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1382 struct mem_cgroup *memcg)
1387 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1392 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1395 struct mem_cgroup *curr = NULL;
1396 struct task_struct *p;
1398 p = find_lock_task_mm(task);
1400 curr = try_get_mem_cgroup_from_mm(p->mm);
1404 * All threads may have already detached their mm's, but the oom
1405 * killer still needs to detect if they have already been oom
1406 * killed to prevent needlessly killing additional tasks.
1409 curr = mem_cgroup_from_task(task);
1411 css_get(&curr->css);
1417 * We should check use_hierarchy of "memcg" not "curr". Because checking
1418 * use_hierarchy of "curr" here make this function true if hierarchy is
1419 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1420 * hierarchy(even if use_hierarchy is disabled in "memcg").
1422 ret = mem_cgroup_same_or_subtree(memcg, curr);
1423 css_put(&curr->css);
1427 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1429 unsigned long inactive_ratio;
1430 unsigned long inactive;
1431 unsigned long active;
1434 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1435 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1437 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1439 inactive_ratio = int_sqrt(10 * gb);
1443 return inactive * inactive_ratio < active;
1446 #define mem_cgroup_from_res_counter(counter, member) \
1447 container_of(counter, struct mem_cgroup, member)
1450 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1451 * @memcg: the memory cgroup
1453 * Returns the maximum amount of memory @mem can be charged with, in
1456 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1458 unsigned long long margin;
1460 margin = res_counter_margin(&memcg->res);
1461 if (do_swap_account)
1462 margin = min(margin, res_counter_margin(&memcg->memsw));
1463 return margin >> PAGE_SHIFT;
1466 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1468 struct cgroup *cgrp = memcg->css.cgroup;
1471 if (cgrp->parent == NULL)
1472 return vm_swappiness;
1474 return memcg->swappiness;
1478 * memcg->moving_account is used for checking possibility that some thread is
1479 * calling move_account(). When a thread on CPU-A starts moving pages under
1480 * a memcg, other threads should check memcg->moving_account under
1481 * rcu_read_lock(), like this:
1485 * memcg->moving_account+1 if (memcg->mocing_account)
1487 * synchronize_rcu() update something.
1492 /* for quick checking without looking up memcg */
1493 atomic_t memcg_moving __read_mostly;
1495 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1497 atomic_inc(&memcg_moving);
1498 atomic_inc(&memcg->moving_account);
1502 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1505 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1506 * We check NULL in callee rather than caller.
1509 atomic_dec(&memcg_moving);
1510 atomic_dec(&memcg->moving_account);
1515 * 2 routines for checking "mem" is under move_account() or not.
1517 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1518 * is used for avoiding races in accounting. If true,
1519 * pc->mem_cgroup may be overwritten.
1521 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1522 * under hierarchy of moving cgroups. This is for
1523 * waiting at hith-memory prressure caused by "move".
1526 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1528 VM_BUG_ON(!rcu_read_lock_held());
1529 return atomic_read(&memcg->moving_account) > 0;
1532 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1534 struct mem_cgroup *from;
1535 struct mem_cgroup *to;
1538 * Unlike task_move routines, we access mc.to, mc.from not under
1539 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1541 spin_lock(&mc.lock);
1547 ret = mem_cgroup_same_or_subtree(memcg, from)
1548 || mem_cgroup_same_or_subtree(memcg, to);
1550 spin_unlock(&mc.lock);
1554 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1556 if (mc.moving_task && current != mc.moving_task) {
1557 if (mem_cgroup_under_move(memcg)) {
1559 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1560 /* moving charge context might have finished. */
1563 finish_wait(&mc.waitq, &wait);
1571 * Take this lock when
1572 * - a code tries to modify page's memcg while it's USED.
1573 * - a code tries to modify page state accounting in a memcg.
1574 * see mem_cgroup_stolen(), too.
1576 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1577 unsigned long *flags)
1579 spin_lock_irqsave(&memcg->move_lock, *flags);
1582 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1583 unsigned long *flags)
1585 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1588 #define K(x) ((x) << (PAGE_SHIFT-10))
1590 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1591 * @memcg: The memory cgroup that went over limit
1592 * @p: Task that is going to be killed
1594 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1597 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1599 struct cgroup *task_cgrp;
1600 struct cgroup *mem_cgrp;
1602 * Need a buffer in BSS, can't rely on allocations. The code relies
1603 * on the assumption that OOM is serialized for memory controller.
1604 * If this assumption is broken, revisit this code.
1606 static char memcg_name[PATH_MAX];
1608 struct mem_cgroup *iter;
1616 mem_cgrp = memcg->css.cgroup;
1617 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1619 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1622 * Unfortunately, we are unable to convert to a useful name
1623 * But we'll still print out the usage information
1630 pr_info("Task in %s killed", memcg_name);
1633 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1641 * Continues from above, so we don't need an KERN_ level
1643 pr_cont(" as a result of limit of %s\n", memcg_name);
1646 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1647 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1648 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1649 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1650 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1651 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1652 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1653 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1654 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1655 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1656 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1657 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1659 for_each_mem_cgroup_tree(iter, memcg) {
1660 pr_info("Memory cgroup stats");
1663 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1665 pr_cont(" for %s", memcg_name);
1669 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1670 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1672 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1673 K(mem_cgroup_read_stat(iter, i)));
1676 for (i = 0; i < NR_LRU_LISTS; i++)
1677 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1678 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1685 * This function returns the number of memcg under hierarchy tree. Returns
1686 * 1(self count) if no children.
1688 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1691 struct mem_cgroup *iter;
1693 for_each_mem_cgroup_tree(iter, memcg)
1699 * Return the memory (and swap, if configured) limit for a memcg.
1701 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1705 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1708 * Do not consider swap space if we cannot swap due to swappiness
1710 if (mem_cgroup_swappiness(memcg)) {
1713 limit += total_swap_pages << PAGE_SHIFT;
1714 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1717 * If memsw is finite and limits the amount of swap space
1718 * available to this memcg, return that limit.
1720 limit = min(limit, memsw);
1726 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1729 struct mem_cgroup *iter;
1730 unsigned long chosen_points = 0;
1731 unsigned long totalpages;
1732 unsigned int points = 0;
1733 struct task_struct *chosen = NULL;
1736 * If current has a pending SIGKILL, then automatically select it. The
1737 * goal is to allow it to allocate so that it may quickly exit and free
1740 if (fatal_signal_pending(current)) {
1741 set_thread_flag(TIF_MEMDIE);
1745 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1746 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1747 for_each_mem_cgroup_tree(iter, memcg) {
1748 struct cgroup *cgroup = iter->css.cgroup;
1749 struct cgroup_iter it;
1750 struct task_struct *task;
1752 cgroup_iter_start(cgroup, &it);
1753 while ((task = cgroup_iter_next(cgroup, &it))) {
1754 switch (oom_scan_process_thread(task, totalpages, NULL,
1756 case OOM_SCAN_SELECT:
1758 put_task_struct(chosen);
1760 chosen_points = ULONG_MAX;
1761 get_task_struct(chosen);
1763 case OOM_SCAN_CONTINUE:
1765 case OOM_SCAN_ABORT:
1766 cgroup_iter_end(cgroup, &it);
1767 mem_cgroup_iter_break(memcg, iter);
1769 put_task_struct(chosen);
1774 points = oom_badness(task, memcg, NULL, totalpages);
1775 if (points > chosen_points) {
1777 put_task_struct(chosen);
1779 chosen_points = points;
1780 get_task_struct(chosen);
1783 cgroup_iter_end(cgroup, &it);
1788 points = chosen_points * 1000 / totalpages;
1789 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1790 NULL, "Memory cgroup out of memory");
1793 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1795 unsigned long flags)
1797 unsigned long total = 0;
1798 bool noswap = false;
1801 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1803 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1806 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1808 drain_all_stock_async(memcg);
1809 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1811 * Allow limit shrinkers, which are triggered directly
1812 * by userspace, to catch signals and stop reclaim
1813 * after minimal progress, regardless of the margin.
1815 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1817 if (mem_cgroup_margin(memcg))
1820 * If nothing was reclaimed after two attempts, there
1821 * may be no reclaimable pages in this hierarchy.
1830 * test_mem_cgroup_node_reclaimable
1831 * @memcg: the target memcg
1832 * @nid: the node ID to be checked.
1833 * @noswap : specify true here if the user wants flle only information.
1835 * This function returns whether the specified memcg contains any
1836 * reclaimable pages on a node. Returns true if there are any reclaimable
1837 * pages in the node.
1839 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1840 int nid, bool noswap)
1842 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1844 if (noswap || !total_swap_pages)
1846 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1851 #if MAX_NUMNODES > 1
1854 * Always updating the nodemask is not very good - even if we have an empty
1855 * list or the wrong list here, we can start from some node and traverse all
1856 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1859 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1863 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1864 * pagein/pageout changes since the last update.
1866 if (!atomic_read(&memcg->numainfo_events))
1868 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1871 /* make a nodemask where this memcg uses memory from */
1872 memcg->scan_nodes = node_states[N_MEMORY];
1874 for_each_node_mask(nid, node_states[N_MEMORY]) {
1876 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1877 node_clear(nid, memcg->scan_nodes);
1880 atomic_set(&memcg->numainfo_events, 0);
1881 atomic_set(&memcg->numainfo_updating, 0);
1885 * Selecting a node where we start reclaim from. Because what we need is just
1886 * reducing usage counter, start from anywhere is O,K. Considering
1887 * memory reclaim from current node, there are pros. and cons.
1889 * Freeing memory from current node means freeing memory from a node which
1890 * we'll use or we've used. So, it may make LRU bad. And if several threads
1891 * hit limits, it will see a contention on a node. But freeing from remote
1892 * node means more costs for memory reclaim because of memory latency.
1894 * Now, we use round-robin. Better algorithm is welcomed.
1896 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1900 mem_cgroup_may_update_nodemask(memcg);
1901 node = memcg->last_scanned_node;
1903 node = next_node(node, memcg->scan_nodes);
1904 if (node == MAX_NUMNODES)
1905 node = first_node(memcg->scan_nodes);
1907 * We call this when we hit limit, not when pages are added to LRU.
1908 * No LRU may hold pages because all pages are UNEVICTABLE or
1909 * memcg is too small and all pages are not on LRU. In that case,
1910 * we use curret node.
1912 if (unlikely(node == MAX_NUMNODES))
1913 node = numa_node_id();
1915 memcg->last_scanned_node = node;
1920 * Check all nodes whether it contains reclaimable pages or not.
1921 * For quick scan, we make use of scan_nodes. This will allow us to skip
1922 * unused nodes. But scan_nodes is lazily updated and may not cotain
1923 * enough new information. We need to do double check.
1925 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1930 * quick check...making use of scan_node.
1931 * We can skip unused nodes.
1933 if (!nodes_empty(memcg->scan_nodes)) {
1934 for (nid = first_node(memcg->scan_nodes);
1936 nid = next_node(nid, memcg->scan_nodes)) {
1938 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1943 * Check rest of nodes.
1945 for_each_node_state(nid, N_MEMORY) {
1946 if (node_isset(nid, memcg->scan_nodes))
1948 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1955 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1960 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1962 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1966 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1969 unsigned long *total_scanned)
1971 struct mem_cgroup *victim = NULL;
1974 unsigned long excess;
1975 unsigned long nr_scanned;
1976 struct mem_cgroup_reclaim_cookie reclaim = {
1981 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1984 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1989 * If we have not been able to reclaim
1990 * anything, it might because there are
1991 * no reclaimable pages under this hierarchy
1996 * We want to do more targeted reclaim.
1997 * excess >> 2 is not to excessive so as to
1998 * reclaim too much, nor too less that we keep
1999 * coming back to reclaim from this cgroup
2001 if (total >= (excess >> 2) ||
2002 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2007 if (!mem_cgroup_reclaimable(victim, false))
2009 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2011 *total_scanned += nr_scanned;
2012 if (!res_counter_soft_limit_excess(&root_memcg->res))
2015 mem_cgroup_iter_break(root_memcg, victim);
2020 * Check OOM-Killer is already running under our hierarchy.
2021 * If someone is running, return false.
2022 * Has to be called with memcg_oom_lock
2024 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
2026 struct mem_cgroup *iter, *failed = NULL;
2028 for_each_mem_cgroup_tree(iter, memcg) {
2029 if (iter->oom_lock) {
2031 * this subtree of our hierarchy is already locked
2032 * so we cannot give a lock.
2035 mem_cgroup_iter_break(memcg, iter);
2038 iter->oom_lock = true;
2045 * OK, we failed to lock the whole subtree so we have to clean up
2046 * what we set up to the failing subtree
2048 for_each_mem_cgroup_tree(iter, memcg) {
2049 if (iter == failed) {
2050 mem_cgroup_iter_break(memcg, iter);
2053 iter->oom_lock = false;
2059 * Has to be called with memcg_oom_lock
2061 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2063 struct mem_cgroup *iter;
2065 for_each_mem_cgroup_tree(iter, memcg)
2066 iter->oom_lock = false;
2070 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2072 struct mem_cgroup *iter;
2074 for_each_mem_cgroup_tree(iter, memcg)
2075 atomic_inc(&iter->under_oom);
2078 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2080 struct mem_cgroup *iter;
2083 * When a new child is created while the hierarchy is under oom,
2084 * mem_cgroup_oom_lock() may not be called. We have to use
2085 * atomic_add_unless() here.
2087 for_each_mem_cgroup_tree(iter, memcg)
2088 atomic_add_unless(&iter->under_oom, -1, 0);
2091 static DEFINE_SPINLOCK(memcg_oom_lock);
2092 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2094 struct oom_wait_info {
2095 struct mem_cgroup *memcg;
2099 static int memcg_oom_wake_function(wait_queue_t *wait,
2100 unsigned mode, int sync, void *arg)
2102 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2103 struct mem_cgroup *oom_wait_memcg;
2104 struct oom_wait_info *oom_wait_info;
2106 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2107 oom_wait_memcg = oom_wait_info->memcg;
2110 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2111 * Then we can use css_is_ancestor without taking care of RCU.
2113 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2114 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2116 return autoremove_wake_function(wait, mode, sync, arg);
2119 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2121 /* for filtering, pass "memcg" as argument. */
2122 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2125 static void memcg_oom_recover(struct mem_cgroup *memcg)
2127 if (memcg && atomic_read(&memcg->under_oom))
2128 memcg_wakeup_oom(memcg);
2132 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2134 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2137 struct oom_wait_info owait;
2138 bool locked, need_to_kill;
2140 owait.memcg = memcg;
2141 owait.wait.flags = 0;
2142 owait.wait.func = memcg_oom_wake_function;
2143 owait.wait.private = current;
2144 INIT_LIST_HEAD(&owait.wait.task_list);
2145 need_to_kill = true;
2146 mem_cgroup_mark_under_oom(memcg);
2148 /* At first, try to OOM lock hierarchy under memcg.*/
2149 spin_lock(&memcg_oom_lock);
2150 locked = mem_cgroup_oom_lock(memcg);
2152 * Even if signal_pending(), we can't quit charge() loop without
2153 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2154 * under OOM is always welcomed, use TASK_KILLABLE here.
2156 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2157 if (!locked || memcg->oom_kill_disable)
2158 need_to_kill = false;
2160 mem_cgroup_oom_notify(memcg);
2161 spin_unlock(&memcg_oom_lock);
2164 finish_wait(&memcg_oom_waitq, &owait.wait);
2165 mem_cgroup_out_of_memory(memcg, mask, order);
2168 finish_wait(&memcg_oom_waitq, &owait.wait);
2170 spin_lock(&memcg_oom_lock);
2172 mem_cgroup_oom_unlock(memcg);
2173 memcg_wakeup_oom(memcg);
2174 spin_unlock(&memcg_oom_lock);
2176 mem_cgroup_unmark_under_oom(memcg);
2178 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2180 /* Give chance to dying process */
2181 schedule_timeout_uninterruptible(1);
2186 * Currently used to update mapped file statistics, but the routine can be
2187 * generalized to update other statistics as well.
2189 * Notes: Race condition
2191 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2192 * it tends to be costly. But considering some conditions, we doesn't need
2193 * to do so _always_.
2195 * Considering "charge", lock_page_cgroup() is not required because all
2196 * file-stat operations happen after a page is attached to radix-tree. There
2197 * are no race with "charge".
2199 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2200 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2201 * if there are race with "uncharge". Statistics itself is properly handled
2204 * Considering "move", this is an only case we see a race. To make the race
2205 * small, we check mm->moving_account and detect there are possibility of race
2206 * If there is, we take a lock.
2209 void __mem_cgroup_begin_update_page_stat(struct page *page,
2210 bool *locked, unsigned long *flags)
2212 struct mem_cgroup *memcg;
2213 struct page_cgroup *pc;
2215 pc = lookup_page_cgroup(page);
2217 memcg = pc->mem_cgroup;
2218 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2221 * If this memory cgroup is not under account moving, we don't
2222 * need to take move_lock_mem_cgroup(). Because we already hold
2223 * rcu_read_lock(), any calls to move_account will be delayed until
2224 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2226 if (!mem_cgroup_stolen(memcg))
2229 move_lock_mem_cgroup(memcg, flags);
2230 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2231 move_unlock_mem_cgroup(memcg, flags);
2237 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2239 struct page_cgroup *pc = lookup_page_cgroup(page);
2242 * It's guaranteed that pc->mem_cgroup never changes while
2243 * lock is held because a routine modifies pc->mem_cgroup
2244 * should take move_lock_mem_cgroup().
2246 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2249 void mem_cgroup_update_page_stat(struct page *page,
2250 enum mem_cgroup_page_stat_item idx, int val)
2252 struct mem_cgroup *memcg;
2253 struct page_cgroup *pc = lookup_page_cgroup(page);
2254 unsigned long uninitialized_var(flags);
2256 if (mem_cgroup_disabled())
2259 memcg = pc->mem_cgroup;
2260 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2264 case MEMCG_NR_FILE_MAPPED:
2265 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2271 this_cpu_add(memcg->stat->count[idx], val);
2275 * size of first charge trial. "32" comes from vmscan.c's magic value.
2276 * TODO: maybe necessary to use big numbers in big irons.
2278 #define CHARGE_BATCH 32U
2279 struct memcg_stock_pcp {
2280 struct mem_cgroup *cached; /* this never be root cgroup */
2281 unsigned int nr_pages;
2282 struct work_struct work;
2283 unsigned long flags;
2284 #define FLUSHING_CACHED_CHARGE 0
2286 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2287 static DEFINE_MUTEX(percpu_charge_mutex);
2290 * consume_stock: Try to consume stocked charge on this cpu.
2291 * @memcg: memcg to consume from.
2292 * @nr_pages: how many pages to charge.
2294 * The charges will only happen if @memcg matches the current cpu's memcg
2295 * stock, and at least @nr_pages are available in that stock. Failure to
2296 * service an allocation will refill the stock.
2298 * returns true if successful, false otherwise.
2300 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2302 struct memcg_stock_pcp *stock;
2305 if (nr_pages > CHARGE_BATCH)
2308 stock = &get_cpu_var(memcg_stock);
2309 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2310 stock->nr_pages -= nr_pages;
2311 else /* need to call res_counter_charge */
2313 put_cpu_var(memcg_stock);
2318 * Returns stocks cached in percpu to res_counter and reset cached information.
2320 static void drain_stock(struct memcg_stock_pcp *stock)
2322 struct mem_cgroup *old = stock->cached;
2324 if (stock->nr_pages) {
2325 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2327 res_counter_uncharge(&old->res, bytes);
2328 if (do_swap_account)
2329 res_counter_uncharge(&old->memsw, bytes);
2330 stock->nr_pages = 0;
2332 stock->cached = NULL;
2336 * This must be called under preempt disabled or must be called by
2337 * a thread which is pinned to local cpu.
2339 static void drain_local_stock(struct work_struct *dummy)
2341 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2343 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2346 static void __init memcg_stock_init(void)
2350 for_each_possible_cpu(cpu) {
2351 struct memcg_stock_pcp *stock =
2352 &per_cpu(memcg_stock, cpu);
2353 INIT_WORK(&stock->work, drain_local_stock);
2358 * Cache charges(val) which is from res_counter, to local per_cpu area.
2359 * This will be consumed by consume_stock() function, later.
2361 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2363 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2365 if (stock->cached != memcg) { /* reset if necessary */
2367 stock->cached = memcg;
2369 stock->nr_pages += nr_pages;
2370 put_cpu_var(memcg_stock);
2374 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2375 * of the hierarchy under it. sync flag says whether we should block
2376 * until the work is done.
2378 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2382 /* Notify other cpus that system-wide "drain" is running */
2385 for_each_online_cpu(cpu) {
2386 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2387 struct mem_cgroup *memcg;
2389 memcg = stock->cached;
2390 if (!memcg || !stock->nr_pages)
2392 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2394 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2396 drain_local_stock(&stock->work);
2398 schedule_work_on(cpu, &stock->work);
2406 for_each_online_cpu(cpu) {
2407 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2408 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2409 flush_work(&stock->work);
2416 * Tries to drain stocked charges in other cpus. This function is asynchronous
2417 * and just put a work per cpu for draining localy on each cpu. Caller can
2418 * expects some charges will be back to res_counter later but cannot wait for
2421 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2424 * If someone calls draining, avoid adding more kworker runs.
2426 if (!mutex_trylock(&percpu_charge_mutex))
2428 drain_all_stock(root_memcg, false);
2429 mutex_unlock(&percpu_charge_mutex);
2432 /* This is a synchronous drain interface. */
2433 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2435 /* called when force_empty is called */
2436 mutex_lock(&percpu_charge_mutex);
2437 drain_all_stock(root_memcg, true);
2438 mutex_unlock(&percpu_charge_mutex);
2442 * This function drains percpu counter value from DEAD cpu and
2443 * move it to local cpu. Note that this function can be preempted.
2445 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2449 spin_lock(&memcg->pcp_counter_lock);
2450 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2451 long x = per_cpu(memcg->stat->count[i], cpu);
2453 per_cpu(memcg->stat->count[i], cpu) = 0;
2454 memcg->nocpu_base.count[i] += x;
2456 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2457 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2459 per_cpu(memcg->stat->events[i], cpu) = 0;
2460 memcg->nocpu_base.events[i] += x;
2462 spin_unlock(&memcg->pcp_counter_lock);
2465 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2466 unsigned long action,
2469 int cpu = (unsigned long)hcpu;
2470 struct memcg_stock_pcp *stock;
2471 struct mem_cgroup *iter;
2473 if (action == CPU_ONLINE)
2476 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2479 for_each_mem_cgroup(iter)
2480 mem_cgroup_drain_pcp_counter(iter, cpu);
2482 stock = &per_cpu(memcg_stock, cpu);
2488 /* See __mem_cgroup_try_charge() for details */
2490 CHARGE_OK, /* success */
2491 CHARGE_RETRY, /* need to retry but retry is not bad */
2492 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2493 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2494 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2497 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2498 unsigned int nr_pages, unsigned int min_pages,
2501 unsigned long csize = nr_pages * PAGE_SIZE;
2502 struct mem_cgroup *mem_over_limit;
2503 struct res_counter *fail_res;
2504 unsigned long flags = 0;
2507 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2510 if (!do_swap_account)
2512 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2516 res_counter_uncharge(&memcg->res, csize);
2517 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2518 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2520 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2522 * Never reclaim on behalf of optional batching, retry with a
2523 * single page instead.
2525 if (nr_pages > min_pages)
2526 return CHARGE_RETRY;
2528 if (!(gfp_mask & __GFP_WAIT))
2529 return CHARGE_WOULDBLOCK;
2531 if (gfp_mask & __GFP_NORETRY)
2532 return CHARGE_NOMEM;
2534 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2535 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2536 return CHARGE_RETRY;
2538 * Even though the limit is exceeded at this point, reclaim
2539 * may have been able to free some pages. Retry the charge
2540 * before killing the task.
2542 * Only for regular pages, though: huge pages are rather
2543 * unlikely to succeed so close to the limit, and we fall back
2544 * to regular pages anyway in case of failure.
2546 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2547 return CHARGE_RETRY;
2550 * At task move, charge accounts can be doubly counted. So, it's
2551 * better to wait until the end of task_move if something is going on.
2553 if (mem_cgroup_wait_acct_move(mem_over_limit))
2554 return CHARGE_RETRY;
2556 /* If we don't need to call oom-killer at el, return immediately */
2558 return CHARGE_NOMEM;
2560 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2561 return CHARGE_OOM_DIE;
2563 return CHARGE_RETRY;
2567 * __mem_cgroup_try_charge() does
2568 * 1. detect memcg to be charged against from passed *mm and *ptr,
2569 * 2. update res_counter
2570 * 3. call memory reclaim if necessary.
2572 * In some special case, if the task is fatal, fatal_signal_pending() or
2573 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2574 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2575 * as possible without any hazards. 2: all pages should have a valid
2576 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2577 * pointer, that is treated as a charge to root_mem_cgroup.
2579 * So __mem_cgroup_try_charge() will return
2580 * 0 ... on success, filling *ptr with a valid memcg pointer.
2581 * -ENOMEM ... charge failure because of resource limits.
2582 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2584 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2585 * the oom-killer can be invoked.
2587 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2589 unsigned int nr_pages,
2590 struct mem_cgroup **ptr,
2593 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2594 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2595 struct mem_cgroup *memcg = NULL;
2599 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2600 * in system level. So, allow to go ahead dying process in addition to
2603 if (unlikely(test_thread_flag(TIF_MEMDIE)
2604 || fatal_signal_pending(current)))
2608 * We always charge the cgroup the mm_struct belongs to.
2609 * The mm_struct's mem_cgroup changes on task migration if the
2610 * thread group leader migrates. It's possible that mm is not
2611 * set, if so charge the root memcg (happens for pagecache usage).
2614 *ptr = root_mem_cgroup;
2616 if (*ptr) { /* css should be a valid one */
2618 if (mem_cgroup_is_root(memcg))
2620 if (consume_stock(memcg, nr_pages))
2622 css_get(&memcg->css);
2624 struct task_struct *p;
2627 p = rcu_dereference(mm->owner);
2629 * Because we don't have task_lock(), "p" can exit.
2630 * In that case, "memcg" can point to root or p can be NULL with
2631 * race with swapoff. Then, we have small risk of mis-accouning.
2632 * But such kind of mis-account by race always happens because
2633 * we don't have cgroup_mutex(). It's overkill and we allo that
2635 * (*) swapoff at el will charge against mm-struct not against
2636 * task-struct. So, mm->owner can be NULL.
2638 memcg = mem_cgroup_from_task(p);
2640 memcg = root_mem_cgroup;
2641 if (mem_cgroup_is_root(memcg)) {
2645 if (consume_stock(memcg, nr_pages)) {
2647 * It seems dagerous to access memcg without css_get().
2648 * But considering how consume_stok works, it's not
2649 * necessary. If consume_stock success, some charges
2650 * from this memcg are cached on this cpu. So, we
2651 * don't need to call css_get()/css_tryget() before
2652 * calling consume_stock().
2657 /* after here, we may be blocked. we need to get refcnt */
2658 if (!css_tryget(&memcg->css)) {
2668 /* If killed, bypass charge */
2669 if (fatal_signal_pending(current)) {
2670 css_put(&memcg->css);
2675 if (oom && !nr_oom_retries) {
2677 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2680 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2685 case CHARGE_RETRY: /* not in OOM situation but retry */
2687 css_put(&memcg->css);
2690 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2691 css_put(&memcg->css);
2693 case CHARGE_NOMEM: /* OOM routine works */
2695 css_put(&memcg->css);
2698 /* If oom, we never return -ENOMEM */
2701 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2702 css_put(&memcg->css);
2705 } while (ret != CHARGE_OK);
2707 if (batch > nr_pages)
2708 refill_stock(memcg, batch - nr_pages);
2709 css_put(&memcg->css);
2717 *ptr = root_mem_cgroup;
2722 * Somemtimes we have to undo a charge we got by try_charge().
2723 * This function is for that and do uncharge, put css's refcnt.
2724 * gotten by try_charge().
2726 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2727 unsigned int nr_pages)
2729 if (!mem_cgroup_is_root(memcg)) {
2730 unsigned long bytes = nr_pages * PAGE_SIZE;
2732 res_counter_uncharge(&memcg->res, bytes);
2733 if (do_swap_account)
2734 res_counter_uncharge(&memcg->memsw, bytes);
2739 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2740 * This is useful when moving usage to parent cgroup.
2742 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2743 unsigned int nr_pages)
2745 unsigned long bytes = nr_pages * PAGE_SIZE;
2747 if (mem_cgroup_is_root(memcg))
2750 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2751 if (do_swap_account)
2752 res_counter_uncharge_until(&memcg->memsw,
2753 memcg->memsw.parent, bytes);
2757 * A helper function to get mem_cgroup from ID. must be called under
2758 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2759 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2760 * called against removed memcg.)
2762 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2764 struct cgroup_subsys_state *css;
2766 /* ID 0 is unused ID */
2769 css = css_lookup(&mem_cgroup_subsys, id);
2772 return mem_cgroup_from_css(css);
2775 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2777 struct mem_cgroup *memcg = NULL;
2778 struct page_cgroup *pc;
2782 VM_BUG_ON(!PageLocked(page));
2784 pc = lookup_page_cgroup(page);
2785 lock_page_cgroup(pc);
2786 if (PageCgroupUsed(pc)) {
2787 memcg = pc->mem_cgroup;
2788 if (memcg && !css_tryget(&memcg->css))
2790 } else if (PageSwapCache(page)) {
2791 ent.val = page_private(page);
2792 id = lookup_swap_cgroup_id(ent);
2794 memcg = mem_cgroup_lookup(id);
2795 if (memcg && !css_tryget(&memcg->css))
2799 unlock_page_cgroup(pc);
2803 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2805 unsigned int nr_pages,
2806 enum charge_type ctype,
2809 struct page_cgroup *pc = lookup_page_cgroup(page);
2810 struct zone *uninitialized_var(zone);
2811 struct lruvec *lruvec;
2812 bool was_on_lru = false;
2815 lock_page_cgroup(pc);
2816 VM_BUG_ON(PageCgroupUsed(pc));
2818 * we don't need page_cgroup_lock about tail pages, becase they are not
2819 * accessed by any other context at this point.
2823 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2824 * may already be on some other mem_cgroup's LRU. Take care of it.
2827 zone = page_zone(page);
2828 spin_lock_irq(&zone->lru_lock);
2829 if (PageLRU(page)) {
2830 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2832 del_page_from_lru_list(page, lruvec, page_lru(page));
2837 pc->mem_cgroup = memcg;
2839 * We access a page_cgroup asynchronously without lock_page_cgroup().
2840 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2841 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2842 * before USED bit, we need memory barrier here.
2843 * See mem_cgroup_add_lru_list(), etc.
2846 SetPageCgroupUsed(pc);
2850 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2851 VM_BUG_ON(PageLRU(page));
2853 add_page_to_lru_list(page, lruvec, page_lru(page));
2855 spin_unlock_irq(&zone->lru_lock);
2858 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2863 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2864 unlock_page_cgroup(pc);
2867 * "charge_statistics" updated event counter. Then, check it.
2868 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2869 * if they exceeds softlimit.
2871 memcg_check_events(memcg, page);
2874 static DEFINE_MUTEX(set_limit_mutex);
2876 #ifdef CONFIG_MEMCG_KMEM
2877 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2879 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2880 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2884 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2885 * in the memcg_cache_params struct.
2887 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2889 struct kmem_cache *cachep;
2891 VM_BUG_ON(p->is_root_cache);
2892 cachep = p->root_cache;
2893 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2896 #ifdef CONFIG_SLABINFO
2897 static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2900 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2901 struct memcg_cache_params *params;
2903 if (!memcg_can_account_kmem(memcg))
2906 print_slabinfo_header(m);
2908 mutex_lock(&memcg->slab_caches_mutex);
2909 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2910 cache_show(memcg_params_to_cache(params), m);
2911 mutex_unlock(&memcg->slab_caches_mutex);
2917 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2919 struct res_counter *fail_res;
2920 struct mem_cgroup *_memcg;
2924 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2929 * Conditions under which we can wait for the oom_killer. Those are
2930 * the same conditions tested by the core page allocator
2932 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2935 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2938 if (ret == -EINTR) {
2940 * __mem_cgroup_try_charge() chosed to bypass to root due to
2941 * OOM kill or fatal signal. Since our only options are to
2942 * either fail the allocation or charge it to this cgroup, do
2943 * it as a temporary condition. But we can't fail. From a
2944 * kmem/slab perspective, the cache has already been selected,
2945 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2948 * This condition will only trigger if the task entered
2949 * memcg_charge_kmem in a sane state, but was OOM-killed during
2950 * __mem_cgroup_try_charge() above. Tasks that were already
2951 * dying when the allocation triggers should have been already
2952 * directed to the root cgroup in memcontrol.h
2954 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2955 if (do_swap_account)
2956 res_counter_charge_nofail(&memcg->memsw, size,
2960 res_counter_uncharge(&memcg->kmem, size);
2965 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2967 res_counter_uncharge(&memcg->res, size);
2968 if (do_swap_account)
2969 res_counter_uncharge(&memcg->memsw, size);
2972 if (res_counter_uncharge(&memcg->kmem, size))
2975 if (memcg_kmem_test_and_clear_dead(memcg))
2976 mem_cgroup_put(memcg);
2979 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2984 mutex_lock(&memcg->slab_caches_mutex);
2985 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2986 mutex_unlock(&memcg->slab_caches_mutex);
2990 * helper for acessing a memcg's index. It will be used as an index in the
2991 * child cache array in kmem_cache, and also to derive its name. This function
2992 * will return -1 when this is not a kmem-limited memcg.
2994 int memcg_cache_id(struct mem_cgroup *memcg)
2996 return memcg ? memcg->kmemcg_id : -1;
3000 * This ends up being protected by the set_limit mutex, during normal
3001 * operation, because that is its main call site.
3003 * But when we create a new cache, we can call this as well if its parent
3004 * is kmem-limited. That will have to hold set_limit_mutex as well.
3006 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3010 num = ida_simple_get(&kmem_limited_groups,
3011 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3015 * After this point, kmem_accounted (that we test atomically in
3016 * the beginning of this conditional), is no longer 0. This
3017 * guarantees only one process will set the following boolean
3018 * to true. We don't need test_and_set because we're protected
3019 * by the set_limit_mutex anyway.
3021 memcg_kmem_set_activated(memcg);
3023 ret = memcg_update_all_caches(num+1);
3025 ida_simple_remove(&kmem_limited_groups, num);
3026 memcg_kmem_clear_activated(memcg);
3030 memcg->kmemcg_id = num;
3031 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3032 mutex_init(&memcg->slab_caches_mutex);
3036 static size_t memcg_caches_array_size(int num_groups)
3039 if (num_groups <= 0)
3042 size = 2 * num_groups;
3043 if (size < MEMCG_CACHES_MIN_SIZE)
3044 size = MEMCG_CACHES_MIN_SIZE;
3045 else if (size > MEMCG_CACHES_MAX_SIZE)
3046 size = MEMCG_CACHES_MAX_SIZE;
3052 * We should update the current array size iff all caches updates succeed. This
3053 * can only be done from the slab side. The slab mutex needs to be held when
3056 void memcg_update_array_size(int num)
3058 if (num > memcg_limited_groups_array_size)
3059 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3062 static void kmem_cache_destroy_work_func(struct work_struct *w);
3064 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3066 struct memcg_cache_params *cur_params = s->memcg_params;
3068 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3070 if (num_groups > memcg_limited_groups_array_size) {
3072 ssize_t size = memcg_caches_array_size(num_groups);
3074 size *= sizeof(void *);
3075 size += sizeof(struct memcg_cache_params);
3077 s->memcg_params = kzalloc(size, GFP_KERNEL);
3078 if (!s->memcg_params) {
3079 s->memcg_params = cur_params;
3083 INIT_WORK(&s->memcg_params->destroy,
3084 kmem_cache_destroy_work_func);
3085 s->memcg_params->is_root_cache = true;
3088 * There is the chance it will be bigger than
3089 * memcg_limited_groups_array_size, if we failed an allocation
3090 * in a cache, in which case all caches updated before it, will
3091 * have a bigger array.
3093 * But if that is the case, the data after
3094 * memcg_limited_groups_array_size is certainly unused
3096 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3097 if (!cur_params->memcg_caches[i])
3099 s->memcg_params->memcg_caches[i] =
3100 cur_params->memcg_caches[i];
3104 * Ideally, we would wait until all caches succeed, and only
3105 * then free the old one. But this is not worth the extra
3106 * pointer per-cache we'd have to have for this.
3108 * It is not a big deal if some caches are left with a size
3109 * bigger than the others. And all updates will reset this
3117 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3118 struct kmem_cache *root_cache)
3120 size_t size = sizeof(struct memcg_cache_params);
3122 if (!memcg_kmem_enabled())
3126 size += memcg_limited_groups_array_size * sizeof(void *);
3128 s->memcg_params = kzalloc(size, GFP_KERNEL);
3129 if (!s->memcg_params)
3132 INIT_WORK(&s->memcg_params->destroy,
3133 kmem_cache_destroy_work_func);
3135 s->memcg_params->memcg = memcg;
3136 s->memcg_params->root_cache = root_cache;
3138 s->memcg_params->is_root_cache = true;
3143 void memcg_release_cache(struct kmem_cache *s)
3145 struct kmem_cache *root;
3146 struct mem_cgroup *memcg;
3150 * This happens, for instance, when a root cache goes away before we
3153 if (!s->memcg_params)
3156 if (s->memcg_params->is_root_cache)
3159 memcg = s->memcg_params->memcg;
3160 id = memcg_cache_id(memcg);
3162 root = s->memcg_params->root_cache;
3163 root->memcg_params->memcg_caches[id] = NULL;
3164 mem_cgroup_put(memcg);
3166 mutex_lock(&memcg->slab_caches_mutex);
3167 list_del(&s->memcg_params->list);
3168 mutex_unlock(&memcg->slab_caches_mutex);
3171 kfree(s->memcg_params);
3175 * During the creation a new cache, we need to disable our accounting mechanism
3176 * altogether. This is true even if we are not creating, but rather just
3177 * enqueing new caches to be created.
3179 * This is because that process will trigger allocations; some visible, like
3180 * explicit kmallocs to auxiliary data structures, name strings and internal
3181 * cache structures; some well concealed, like INIT_WORK() that can allocate
3182 * objects during debug.
3184 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3185 * to it. This may not be a bounded recursion: since the first cache creation
3186 * failed to complete (waiting on the allocation), we'll just try to create the
3187 * cache again, failing at the same point.
3189 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3190 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3191 * inside the following two functions.
3193 static inline void memcg_stop_kmem_account(void)
3195 VM_BUG_ON(!current->mm);
3196 current->memcg_kmem_skip_account++;
3199 static inline void memcg_resume_kmem_account(void)
3201 VM_BUG_ON(!current->mm);
3202 current->memcg_kmem_skip_account--;
3205 static void kmem_cache_destroy_work_func(struct work_struct *w)
3207 struct kmem_cache *cachep;
3208 struct memcg_cache_params *p;
3210 p = container_of(w, struct memcg_cache_params, destroy);
3212 cachep = memcg_params_to_cache(p);
3215 * If we get down to 0 after shrink, we could delete right away.
3216 * However, memcg_release_pages() already puts us back in the workqueue
3217 * in that case. If we proceed deleting, we'll get a dangling
3218 * reference, and removing the object from the workqueue in that case
3219 * is unnecessary complication. We are not a fast path.
3221 * Note that this case is fundamentally different from racing with
3222 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3223 * kmem_cache_shrink, not only we would be reinserting a dead cache
3224 * into the queue, but doing so from inside the worker racing to
3227 * So if we aren't down to zero, we'll just schedule a worker and try
3230 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3231 kmem_cache_shrink(cachep);
3232 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3235 kmem_cache_destroy(cachep);
3238 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3240 if (!cachep->memcg_params->dead)
3244 * There are many ways in which we can get here.
3246 * We can get to a memory-pressure situation while the delayed work is
3247 * still pending to run. The vmscan shrinkers can then release all
3248 * cache memory and get us to destruction. If this is the case, we'll
3249 * be executed twice, which is a bug (the second time will execute over
3250 * bogus data). In this case, cancelling the work should be fine.
3252 * But we can also get here from the worker itself, if
3253 * kmem_cache_shrink is enough to shake all the remaining objects and
3254 * get the page count to 0. In this case, we'll deadlock if we try to
3255 * cancel the work (the worker runs with an internal lock held, which
3256 * is the same lock we would hold for cancel_work_sync().)
3258 * Since we can't possibly know who got us here, just refrain from
3259 * running if there is already work pending
3261 if (work_pending(&cachep->memcg_params->destroy))
3264 * We have to defer the actual destroying to a workqueue, because
3265 * we might currently be in a context that cannot sleep.
3267 schedule_work(&cachep->memcg_params->destroy);
3270 static char *memcg_cache_name(struct mem_cgroup *memcg, struct kmem_cache *s)
3273 struct dentry *dentry;
3276 dentry = rcu_dereference(memcg->css.cgroup->dentry);
3279 BUG_ON(dentry == NULL);
3281 name = kasprintf(GFP_KERNEL, "%s(%d:%s)", s->name,
3282 memcg_cache_id(memcg), dentry->d_name.name);
3287 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3288 struct kmem_cache *s)
3291 struct kmem_cache *new;
3293 name = memcg_cache_name(memcg, s);
3297 new = kmem_cache_create_memcg(memcg, name, s->object_size, s->align,
3298 (s->flags & ~SLAB_PANIC), s->ctor, s);
3301 new->allocflags |= __GFP_KMEMCG;
3308 * This lock protects updaters, not readers. We want readers to be as fast as
3309 * they can, and they will either see NULL or a valid cache value. Our model
3310 * allow them to see NULL, in which case the root memcg will be selected.
3312 * We need this lock because multiple allocations to the same cache from a non
3313 * will span more than one worker. Only one of them can create the cache.
3315 static DEFINE_MUTEX(memcg_cache_mutex);
3316 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3317 struct kmem_cache *cachep)
3319 struct kmem_cache *new_cachep;
3322 BUG_ON(!memcg_can_account_kmem(memcg));
3324 idx = memcg_cache_id(memcg);
3326 mutex_lock(&memcg_cache_mutex);
3327 new_cachep = cachep->memcg_params->memcg_caches[idx];
3331 new_cachep = kmem_cache_dup(memcg, cachep);
3332 if (new_cachep == NULL) {
3333 new_cachep = cachep;
3337 mem_cgroup_get(memcg);
3338 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3340 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3342 * the readers won't lock, make sure everybody sees the updated value,
3343 * so they won't put stuff in the queue again for no reason
3347 mutex_unlock(&memcg_cache_mutex);
3351 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3353 struct kmem_cache *c;
3356 if (!s->memcg_params)
3358 if (!s->memcg_params->is_root_cache)
3362 * If the cache is being destroyed, we trust that there is no one else
3363 * requesting objects from it. Even if there are, the sanity checks in
3364 * kmem_cache_destroy should caught this ill-case.
3366 * Still, we don't want anyone else freeing memcg_caches under our
3367 * noses, which can happen if a new memcg comes to life. As usual,
3368 * we'll take the set_limit_mutex to protect ourselves against this.
3370 mutex_lock(&set_limit_mutex);
3371 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3372 c = s->memcg_params->memcg_caches[i];
3377 * We will now manually delete the caches, so to avoid races
3378 * we need to cancel all pending destruction workers and
3379 * proceed with destruction ourselves.
3381 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3382 * and that could spawn the workers again: it is likely that
3383 * the cache still have active pages until this very moment.
3384 * This would lead us back to mem_cgroup_destroy_cache.
3386 * But that will not execute at all if the "dead" flag is not
3387 * set, so flip it down to guarantee we are in control.
3389 c->memcg_params->dead = false;
3390 cancel_work_sync(&c->memcg_params->destroy);
3391 kmem_cache_destroy(c);
3393 mutex_unlock(&set_limit_mutex);
3396 struct create_work {
3397 struct mem_cgroup *memcg;
3398 struct kmem_cache *cachep;
3399 struct work_struct work;
3402 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3404 struct kmem_cache *cachep;
3405 struct memcg_cache_params *params;
3407 if (!memcg_kmem_is_active(memcg))
3410 mutex_lock(&memcg->slab_caches_mutex);
3411 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3412 cachep = memcg_params_to_cache(params);
3413 cachep->memcg_params->dead = true;
3414 schedule_work(&cachep->memcg_params->destroy);
3416 mutex_unlock(&memcg->slab_caches_mutex);
3419 static void memcg_create_cache_work_func(struct work_struct *w)
3421 struct create_work *cw;
3423 cw = container_of(w, struct create_work, work);
3424 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3425 /* Drop the reference gotten when we enqueued. */
3426 css_put(&cw->memcg->css);
3431 * Enqueue the creation of a per-memcg kmem_cache.
3432 * Called with rcu_read_lock.
3434 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3435 struct kmem_cache *cachep)
3437 struct create_work *cw;
3439 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3443 /* The corresponding put will be done in the workqueue. */
3444 if (!css_tryget(&memcg->css)) {
3450 cw->cachep = cachep;
3452 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3453 schedule_work(&cw->work);
3456 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3457 struct kmem_cache *cachep)
3460 * We need to stop accounting when we kmalloc, because if the
3461 * corresponding kmalloc cache is not yet created, the first allocation
3462 * in __memcg_create_cache_enqueue will recurse.
3464 * However, it is better to enclose the whole function. Depending on
3465 * the debugging options enabled, INIT_WORK(), for instance, can
3466 * trigger an allocation. This too, will make us recurse. Because at
3467 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3468 * the safest choice is to do it like this, wrapping the whole function.
3470 memcg_stop_kmem_account();
3471 __memcg_create_cache_enqueue(memcg, cachep);
3472 memcg_resume_kmem_account();
3475 * Return the kmem_cache we're supposed to use for a slab allocation.
3476 * We try to use the current memcg's version of the cache.
3478 * If the cache does not exist yet, if we are the first user of it,
3479 * we either create it immediately, if possible, or create it asynchronously
3481 * In the latter case, we will let the current allocation go through with
3482 * the original cache.
3484 * Can't be called in interrupt context or from kernel threads.
3485 * This function needs to be called with rcu_read_lock() held.
3487 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3490 struct mem_cgroup *memcg;
3493 VM_BUG_ON(!cachep->memcg_params);
3494 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3496 if (!current->mm || current->memcg_kmem_skip_account)
3500 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3503 if (!memcg_can_account_kmem(memcg))
3506 idx = memcg_cache_id(memcg);
3509 * barrier to mare sure we're always seeing the up to date value. The
3510 * code updating memcg_caches will issue a write barrier to match this.
3512 read_barrier_depends();
3513 if (unlikely(cachep->memcg_params->memcg_caches[idx] == NULL)) {
3515 * If we are in a safe context (can wait, and not in interrupt
3516 * context), we could be be predictable and return right away.
3517 * This would guarantee that the allocation being performed
3518 * already belongs in the new cache.
3520 * However, there are some clashes that can arrive from locking.
3521 * For instance, because we acquire the slab_mutex while doing
3522 * kmem_cache_dup, this means no further allocation could happen
3523 * with the slab_mutex held.
3525 * Also, because cache creation issue get_online_cpus(), this
3526 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3527 * that ends up reversed during cpu hotplug. (cpuset allocates
3528 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3529 * better to defer everything.
3531 memcg_create_cache_enqueue(memcg, cachep);
3535 return cachep->memcg_params->memcg_caches[idx];
3537 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3540 * We need to verify if the allocation against current->mm->owner's memcg is
3541 * possible for the given order. But the page is not allocated yet, so we'll
3542 * need a further commit step to do the final arrangements.
3544 * It is possible for the task to switch cgroups in this mean time, so at
3545 * commit time, we can't rely on task conversion any longer. We'll then use
3546 * the handle argument to return to the caller which cgroup we should commit
3547 * against. We could also return the memcg directly and avoid the pointer
3548 * passing, but a boolean return value gives better semantics considering
3549 * the compiled-out case as well.
3551 * Returning true means the allocation is possible.
3554 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3556 struct mem_cgroup *memcg;
3560 memcg = try_get_mem_cgroup_from_mm(current->mm);
3563 * very rare case described in mem_cgroup_from_task. Unfortunately there
3564 * isn't much we can do without complicating this too much, and it would
3565 * be gfp-dependent anyway. Just let it go
3567 if (unlikely(!memcg))
3570 if (!memcg_can_account_kmem(memcg)) {
3571 css_put(&memcg->css);
3575 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3579 css_put(&memcg->css);
3583 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3586 struct page_cgroup *pc;
3588 VM_BUG_ON(mem_cgroup_is_root(memcg));
3590 /* The page allocation failed. Revert */
3592 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3596 pc = lookup_page_cgroup(page);
3597 lock_page_cgroup(pc);
3598 pc->mem_cgroup = memcg;
3599 SetPageCgroupUsed(pc);
3600 unlock_page_cgroup(pc);
3603 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3605 struct mem_cgroup *memcg = NULL;
3606 struct page_cgroup *pc;
3609 pc = lookup_page_cgroup(page);
3611 * Fast unlocked return. Theoretically might have changed, have to
3612 * check again after locking.
3614 if (!PageCgroupUsed(pc))
3617 lock_page_cgroup(pc);
3618 if (PageCgroupUsed(pc)) {
3619 memcg = pc->mem_cgroup;
3620 ClearPageCgroupUsed(pc);
3622 unlock_page_cgroup(pc);
3625 * We trust that only if there is a memcg associated with the page, it
3626 * is a valid allocation
3631 VM_BUG_ON(mem_cgroup_is_root(memcg));
3632 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3635 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3638 #endif /* CONFIG_MEMCG_KMEM */
3640 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3642 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3644 * Because tail pages are not marked as "used", set it. We're under
3645 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3646 * charge/uncharge will be never happen and move_account() is done under
3647 * compound_lock(), so we don't have to take care of races.
3649 void mem_cgroup_split_huge_fixup(struct page *head)
3651 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3652 struct page_cgroup *pc;
3655 if (mem_cgroup_disabled())
3657 for (i = 1; i < HPAGE_PMD_NR; i++) {
3659 pc->mem_cgroup = head_pc->mem_cgroup;
3660 smp_wmb();/* see __commit_charge() */
3661 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3664 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3667 * mem_cgroup_move_account - move account of the page
3669 * @nr_pages: number of regular pages (>1 for huge pages)
3670 * @pc: page_cgroup of the page.
3671 * @from: mem_cgroup which the page is moved from.
3672 * @to: mem_cgroup which the page is moved to. @from != @to.
3674 * The caller must confirm following.
3675 * - page is not on LRU (isolate_page() is useful.)
3676 * - compound_lock is held when nr_pages > 1
3678 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3681 static int mem_cgroup_move_account(struct page *page,
3682 unsigned int nr_pages,
3683 struct page_cgroup *pc,
3684 struct mem_cgroup *from,
3685 struct mem_cgroup *to)
3687 unsigned long flags;
3689 bool anon = PageAnon(page);
3691 VM_BUG_ON(from == to);
3692 VM_BUG_ON(PageLRU(page));
3694 * The page is isolated from LRU. So, collapse function
3695 * will not handle this page. But page splitting can happen.
3696 * Do this check under compound_page_lock(). The caller should
3700 if (nr_pages > 1 && !PageTransHuge(page))
3703 lock_page_cgroup(pc);
3706 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3709 move_lock_mem_cgroup(from, &flags);
3711 if (!anon && page_mapped(page)) {
3712 /* Update mapped_file data for mem_cgroup */
3714 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3715 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3718 mem_cgroup_charge_statistics(from, anon, -nr_pages);
3720 /* caller should have done css_get */
3721 pc->mem_cgroup = to;
3722 mem_cgroup_charge_statistics(to, anon, nr_pages);
3723 move_unlock_mem_cgroup(from, &flags);
3726 unlock_page_cgroup(pc);
3730 memcg_check_events(to, page);
3731 memcg_check_events(from, page);
3737 * mem_cgroup_move_parent - moves page to the parent group
3738 * @page: the page to move
3739 * @pc: page_cgroup of the page
3740 * @child: page's cgroup
3742 * move charges to its parent or the root cgroup if the group has no
3743 * parent (aka use_hierarchy==0).
3744 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3745 * mem_cgroup_move_account fails) the failure is always temporary and
3746 * it signals a race with a page removal/uncharge or migration. In the
3747 * first case the page is on the way out and it will vanish from the LRU
3748 * on the next attempt and the call should be retried later.
3749 * Isolation from the LRU fails only if page has been isolated from
3750 * the LRU since we looked at it and that usually means either global
3751 * reclaim or migration going on. The page will either get back to the
3753 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3754 * (!PageCgroupUsed) or moved to a different group. The page will
3755 * disappear in the next attempt.
3757 static int mem_cgroup_move_parent(struct page *page,
3758 struct page_cgroup *pc,
3759 struct mem_cgroup *child)
3761 struct mem_cgroup *parent;
3762 unsigned int nr_pages;
3763 unsigned long uninitialized_var(flags);
3766 VM_BUG_ON(mem_cgroup_is_root(child));
3769 if (!get_page_unless_zero(page))
3771 if (isolate_lru_page(page))
3774 nr_pages = hpage_nr_pages(page);
3776 parent = parent_mem_cgroup(child);
3778 * If no parent, move charges to root cgroup.
3781 parent = root_mem_cgroup;
3784 VM_BUG_ON(!PageTransHuge(page));
3785 flags = compound_lock_irqsave(page);
3788 ret = mem_cgroup_move_account(page, nr_pages,
3791 __mem_cgroup_cancel_local_charge(child, nr_pages);
3794 compound_unlock_irqrestore(page, flags);
3795 putback_lru_page(page);
3803 * Charge the memory controller for page usage.
3805 * 0 if the charge was successful
3806 * < 0 if the cgroup is over its limit
3808 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3809 gfp_t gfp_mask, enum charge_type ctype)
3811 struct mem_cgroup *memcg = NULL;
3812 unsigned int nr_pages = 1;
3816 if (PageTransHuge(page)) {
3817 nr_pages <<= compound_order(page);
3818 VM_BUG_ON(!PageTransHuge(page));
3820 * Never OOM-kill a process for a huge page. The
3821 * fault handler will fall back to regular pages.
3826 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3829 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3833 int mem_cgroup_newpage_charge(struct page *page,
3834 struct mm_struct *mm, gfp_t gfp_mask)
3836 if (mem_cgroup_disabled())
3838 VM_BUG_ON(page_mapped(page));
3839 VM_BUG_ON(page->mapping && !PageAnon(page));
3841 return mem_cgroup_charge_common(page, mm, gfp_mask,
3842 MEM_CGROUP_CHARGE_TYPE_ANON);
3846 * While swap-in, try_charge -> commit or cancel, the page is locked.
3847 * And when try_charge() successfully returns, one refcnt to memcg without
3848 * struct page_cgroup is acquired. This refcnt will be consumed by
3849 * "commit()" or removed by "cancel()"
3851 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3854 struct mem_cgroup **memcgp)
3856 struct mem_cgroup *memcg;
3857 struct page_cgroup *pc;
3860 pc = lookup_page_cgroup(page);
3862 * Every swap fault against a single page tries to charge the
3863 * page, bail as early as possible. shmem_unuse() encounters
3864 * already charged pages, too. The USED bit is protected by
3865 * the page lock, which serializes swap cache removal, which
3866 * in turn serializes uncharging.
3868 if (PageCgroupUsed(pc))
3870 if (!do_swap_account)
3872 memcg = try_get_mem_cgroup_from_page(page);
3876 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3877 css_put(&memcg->css);
3882 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3888 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3889 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3892 if (mem_cgroup_disabled())
3895 * A racing thread's fault, or swapoff, may have already
3896 * updated the pte, and even removed page from swap cache: in
3897 * those cases unuse_pte()'s pte_same() test will fail; but
3898 * there's also a KSM case which does need to charge the page.
3900 if (!PageSwapCache(page)) {
3903 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3908 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3911 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3913 if (mem_cgroup_disabled())
3917 __mem_cgroup_cancel_charge(memcg, 1);
3921 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3922 enum charge_type ctype)
3924 if (mem_cgroup_disabled())
3929 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3931 * Now swap is on-memory. This means this page may be
3932 * counted both as mem and swap....double count.
3933 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3934 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3935 * may call delete_from_swap_cache() before reach here.
3937 if (do_swap_account && PageSwapCache(page)) {
3938 swp_entry_t ent = {.val = page_private(page)};
3939 mem_cgroup_uncharge_swap(ent);
3943 void mem_cgroup_commit_charge_swapin(struct page *page,
3944 struct mem_cgroup *memcg)
3946 __mem_cgroup_commit_charge_swapin(page, memcg,
3947 MEM_CGROUP_CHARGE_TYPE_ANON);
3950 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3953 struct mem_cgroup *memcg = NULL;
3954 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3957 if (mem_cgroup_disabled())
3959 if (PageCompound(page))
3962 if (!PageSwapCache(page))
3963 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3964 else { /* page is swapcache/shmem */
3965 ret = __mem_cgroup_try_charge_swapin(mm, page,
3968 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3973 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3974 unsigned int nr_pages,
3975 const enum charge_type ctype)
3977 struct memcg_batch_info *batch = NULL;
3978 bool uncharge_memsw = true;
3980 /* If swapout, usage of swap doesn't decrease */
3981 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3982 uncharge_memsw = false;
3984 batch = ¤t->memcg_batch;
3986 * In usual, we do css_get() when we remember memcg pointer.
3987 * But in this case, we keep res->usage until end of a series of
3988 * uncharges. Then, it's ok to ignore memcg's refcnt.
3991 batch->memcg = memcg;
3993 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3994 * In those cases, all pages freed continuously can be expected to be in
3995 * the same cgroup and we have chance to coalesce uncharges.
3996 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3997 * because we want to do uncharge as soon as possible.
4000 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4001 goto direct_uncharge;
4004 goto direct_uncharge;
4007 * In typical case, batch->memcg == mem. This means we can
4008 * merge a series of uncharges to an uncharge of res_counter.
4009 * If not, we uncharge res_counter ony by one.
4011 if (batch->memcg != memcg)
4012 goto direct_uncharge;
4013 /* remember freed charge and uncharge it later */
4016 batch->memsw_nr_pages++;
4019 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4021 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4022 if (unlikely(batch->memcg != memcg))
4023 memcg_oom_recover(memcg);
4027 * uncharge if !page_mapped(page)
4029 static struct mem_cgroup *
4030 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4033 struct mem_cgroup *memcg = NULL;
4034 unsigned int nr_pages = 1;
4035 struct page_cgroup *pc;
4038 if (mem_cgroup_disabled())
4041 VM_BUG_ON(PageSwapCache(page));
4043 if (PageTransHuge(page)) {
4044 nr_pages <<= compound_order(page);
4045 VM_BUG_ON(!PageTransHuge(page));
4048 * Check if our page_cgroup is valid
4050 pc = lookup_page_cgroup(page);
4051 if (unlikely(!PageCgroupUsed(pc)))
4054 lock_page_cgroup(pc);
4056 memcg = pc->mem_cgroup;
4058 if (!PageCgroupUsed(pc))
4061 anon = PageAnon(page);
4064 case MEM_CGROUP_CHARGE_TYPE_ANON:
4066 * Generally PageAnon tells if it's the anon statistics to be
4067 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4068 * used before page reached the stage of being marked PageAnon.
4072 case MEM_CGROUP_CHARGE_TYPE_DROP:
4073 /* See mem_cgroup_prepare_migration() */
4074 if (page_mapped(page))
4077 * Pages under migration may not be uncharged. But
4078 * end_migration() /must/ be the one uncharging the
4079 * unused post-migration page and so it has to call
4080 * here with the migration bit still set. See the
4081 * res_counter handling below.
4083 if (!end_migration && PageCgroupMigration(pc))
4086 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4087 if (!PageAnon(page)) { /* Shared memory */
4088 if (page->mapping && !page_is_file_cache(page))
4090 } else if (page_mapped(page)) /* Anon */
4097 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
4099 ClearPageCgroupUsed(pc);
4101 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4102 * freed from LRU. This is safe because uncharged page is expected not
4103 * to be reused (freed soon). Exception is SwapCache, it's handled by
4104 * special functions.
4107 unlock_page_cgroup(pc);
4109 * even after unlock, we have memcg->res.usage here and this memcg
4110 * will never be freed.
4112 memcg_check_events(memcg, page);
4113 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4114 mem_cgroup_swap_statistics(memcg, true);
4115 mem_cgroup_get(memcg);
4118 * Migration does not charge the res_counter for the
4119 * replacement page, so leave it alone when phasing out the
4120 * page that is unused after the migration.
4122 if (!end_migration && !mem_cgroup_is_root(memcg))
4123 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4128 unlock_page_cgroup(pc);
4132 void mem_cgroup_uncharge_page(struct page *page)
4135 if (page_mapped(page))
4137 VM_BUG_ON(page->mapping && !PageAnon(page));
4138 if (PageSwapCache(page))
4140 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4143 void mem_cgroup_uncharge_cache_page(struct page *page)
4145 VM_BUG_ON(page_mapped(page));
4146 VM_BUG_ON(page->mapping);
4147 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4151 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4152 * In that cases, pages are freed continuously and we can expect pages
4153 * are in the same memcg. All these calls itself limits the number of
4154 * pages freed at once, then uncharge_start/end() is called properly.
4155 * This may be called prural(2) times in a context,
4158 void mem_cgroup_uncharge_start(void)
4160 current->memcg_batch.do_batch++;
4161 /* We can do nest. */
4162 if (current->memcg_batch.do_batch == 1) {
4163 current->memcg_batch.memcg = NULL;
4164 current->memcg_batch.nr_pages = 0;
4165 current->memcg_batch.memsw_nr_pages = 0;
4169 void mem_cgroup_uncharge_end(void)
4171 struct memcg_batch_info *batch = ¤t->memcg_batch;
4173 if (!batch->do_batch)
4177 if (batch->do_batch) /* If stacked, do nothing. */
4183 * This "batch->memcg" is valid without any css_get/put etc...
4184 * bacause we hide charges behind us.
4186 if (batch->nr_pages)
4187 res_counter_uncharge(&batch->memcg->res,
4188 batch->nr_pages * PAGE_SIZE);
4189 if (batch->memsw_nr_pages)
4190 res_counter_uncharge(&batch->memcg->memsw,
4191 batch->memsw_nr_pages * PAGE_SIZE);
4192 memcg_oom_recover(batch->memcg);
4193 /* forget this pointer (for sanity check) */
4194 batch->memcg = NULL;
4199 * called after __delete_from_swap_cache() and drop "page" account.
4200 * memcg information is recorded to swap_cgroup of "ent"
4203 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4205 struct mem_cgroup *memcg;
4206 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4208 if (!swapout) /* this was a swap cache but the swap is unused ! */
4209 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4211 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4214 * record memcg information, if swapout && memcg != NULL,
4215 * mem_cgroup_get() was called in uncharge().
4217 if (do_swap_account && swapout && memcg)
4218 swap_cgroup_record(ent, css_id(&memcg->css));
4222 #ifdef CONFIG_MEMCG_SWAP
4224 * called from swap_entry_free(). remove record in swap_cgroup and
4225 * uncharge "memsw" account.
4227 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4229 struct mem_cgroup *memcg;
4232 if (!do_swap_account)
4235 id = swap_cgroup_record(ent, 0);
4237 memcg = mem_cgroup_lookup(id);
4240 * We uncharge this because swap is freed.
4241 * This memcg can be obsolete one. We avoid calling css_tryget
4243 if (!mem_cgroup_is_root(memcg))
4244 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4245 mem_cgroup_swap_statistics(memcg, false);
4246 mem_cgroup_put(memcg);
4252 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4253 * @entry: swap entry to be moved
4254 * @from: mem_cgroup which the entry is moved from
4255 * @to: mem_cgroup which the entry is moved to
4257 * It succeeds only when the swap_cgroup's record for this entry is the same
4258 * as the mem_cgroup's id of @from.
4260 * Returns 0 on success, -EINVAL on failure.
4262 * The caller must have charged to @to, IOW, called res_counter_charge() about
4263 * both res and memsw, and called css_get().
4265 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4266 struct mem_cgroup *from, struct mem_cgroup *to)
4268 unsigned short old_id, new_id;
4270 old_id = css_id(&from->css);
4271 new_id = css_id(&to->css);
4273 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4274 mem_cgroup_swap_statistics(from, false);
4275 mem_cgroup_swap_statistics(to, true);
4277 * This function is only called from task migration context now.
4278 * It postpones res_counter and refcount handling till the end
4279 * of task migration(mem_cgroup_clear_mc()) for performance
4280 * improvement. But we cannot postpone mem_cgroup_get(to)
4281 * because if the process that has been moved to @to does
4282 * swap-in, the refcount of @to might be decreased to 0.
4290 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4291 struct mem_cgroup *from, struct mem_cgroup *to)
4298 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4301 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4302 struct mem_cgroup **memcgp)
4304 struct mem_cgroup *memcg = NULL;
4305 unsigned int nr_pages = 1;
4306 struct page_cgroup *pc;
4307 enum charge_type ctype;
4311 if (mem_cgroup_disabled())
4314 if (PageTransHuge(page))
4315 nr_pages <<= compound_order(page);
4317 pc = lookup_page_cgroup(page);
4318 lock_page_cgroup(pc);
4319 if (PageCgroupUsed(pc)) {
4320 memcg = pc->mem_cgroup;
4321 css_get(&memcg->css);
4323 * At migrating an anonymous page, its mapcount goes down
4324 * to 0 and uncharge() will be called. But, even if it's fully
4325 * unmapped, migration may fail and this page has to be
4326 * charged again. We set MIGRATION flag here and delay uncharge
4327 * until end_migration() is called
4329 * Corner Case Thinking
4331 * When the old page was mapped as Anon and it's unmap-and-freed
4332 * while migration was ongoing.
4333 * If unmap finds the old page, uncharge() of it will be delayed
4334 * until end_migration(). If unmap finds a new page, it's
4335 * uncharged when it make mapcount to be 1->0. If unmap code
4336 * finds swap_migration_entry, the new page will not be mapped
4337 * and end_migration() will find it(mapcount==0).
4340 * When the old page was mapped but migraion fails, the kernel
4341 * remaps it. A charge for it is kept by MIGRATION flag even
4342 * if mapcount goes down to 0. We can do remap successfully
4343 * without charging it again.
4346 * The "old" page is under lock_page() until the end of
4347 * migration, so, the old page itself will not be swapped-out.
4348 * If the new page is swapped out before end_migraton, our
4349 * hook to usual swap-out path will catch the event.
4352 SetPageCgroupMigration(pc);
4354 unlock_page_cgroup(pc);
4356 * If the page is not charged at this point,
4364 * We charge new page before it's used/mapped. So, even if unlock_page()
4365 * is called before end_migration, we can catch all events on this new
4366 * page. In the case new page is migrated but not remapped, new page's
4367 * mapcount will be finally 0 and we call uncharge in end_migration().
4370 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4372 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4374 * The page is committed to the memcg, but it's not actually
4375 * charged to the res_counter since we plan on replacing the
4376 * old one and only one page is going to be left afterwards.
4378 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4381 /* remove redundant charge if migration failed*/
4382 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4383 struct page *oldpage, struct page *newpage, bool migration_ok)
4385 struct page *used, *unused;
4386 struct page_cgroup *pc;
4392 if (!migration_ok) {
4399 anon = PageAnon(used);
4400 __mem_cgroup_uncharge_common(unused,
4401 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4402 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4404 css_put(&memcg->css);
4406 * We disallowed uncharge of pages under migration because mapcount
4407 * of the page goes down to zero, temporarly.
4408 * Clear the flag and check the page should be charged.
4410 pc = lookup_page_cgroup(oldpage);
4411 lock_page_cgroup(pc);
4412 ClearPageCgroupMigration(pc);
4413 unlock_page_cgroup(pc);
4416 * If a page is a file cache, radix-tree replacement is very atomic
4417 * and we can skip this check. When it was an Anon page, its mapcount
4418 * goes down to 0. But because we added MIGRATION flage, it's not
4419 * uncharged yet. There are several case but page->mapcount check
4420 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4421 * check. (see prepare_charge() also)
4424 mem_cgroup_uncharge_page(used);
4428 * At replace page cache, newpage is not under any memcg but it's on
4429 * LRU. So, this function doesn't touch res_counter but handles LRU
4430 * in correct way. Both pages are locked so we cannot race with uncharge.
4432 void mem_cgroup_replace_page_cache(struct page *oldpage,
4433 struct page *newpage)
4435 struct mem_cgroup *memcg = NULL;
4436 struct page_cgroup *pc;
4437 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4439 if (mem_cgroup_disabled())
4442 pc = lookup_page_cgroup(oldpage);
4443 /* fix accounting on old pages */
4444 lock_page_cgroup(pc);
4445 if (PageCgroupUsed(pc)) {
4446 memcg = pc->mem_cgroup;
4447 mem_cgroup_charge_statistics(memcg, false, -1);
4448 ClearPageCgroupUsed(pc);
4450 unlock_page_cgroup(pc);
4453 * When called from shmem_replace_page(), in some cases the
4454 * oldpage has already been charged, and in some cases not.
4459 * Even if newpage->mapping was NULL before starting replacement,
4460 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4461 * LRU while we overwrite pc->mem_cgroup.
4463 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4466 #ifdef CONFIG_DEBUG_VM
4467 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4469 struct page_cgroup *pc;
4471 pc = lookup_page_cgroup(page);
4473 * Can be NULL while feeding pages into the page allocator for
4474 * the first time, i.e. during boot or memory hotplug;
4475 * or when mem_cgroup_disabled().
4477 if (likely(pc) && PageCgroupUsed(pc))
4482 bool mem_cgroup_bad_page_check(struct page *page)
4484 if (mem_cgroup_disabled())
4487 return lookup_page_cgroup_used(page) != NULL;
4490 void mem_cgroup_print_bad_page(struct page *page)
4492 struct page_cgroup *pc;
4494 pc = lookup_page_cgroup_used(page);
4496 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4497 pc, pc->flags, pc->mem_cgroup);
4502 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4503 unsigned long long val)
4506 u64 memswlimit, memlimit;
4508 int children = mem_cgroup_count_children(memcg);
4509 u64 curusage, oldusage;
4513 * For keeping hierarchical_reclaim simple, how long we should retry
4514 * is depends on callers. We set our retry-count to be function
4515 * of # of children which we should visit in this loop.
4517 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4519 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4522 while (retry_count) {
4523 if (signal_pending(current)) {
4528 * Rather than hide all in some function, I do this in
4529 * open coded manner. You see what this really does.
4530 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4532 mutex_lock(&set_limit_mutex);
4533 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4534 if (memswlimit < val) {
4536 mutex_unlock(&set_limit_mutex);
4540 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4544 ret = res_counter_set_limit(&memcg->res, val);
4546 if (memswlimit == val)
4547 memcg->memsw_is_minimum = true;
4549 memcg->memsw_is_minimum = false;
4551 mutex_unlock(&set_limit_mutex);
4556 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4557 MEM_CGROUP_RECLAIM_SHRINK);
4558 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4559 /* Usage is reduced ? */
4560 if (curusage >= oldusage)
4563 oldusage = curusage;
4565 if (!ret && enlarge)
4566 memcg_oom_recover(memcg);
4571 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4572 unsigned long long val)
4575 u64 memlimit, memswlimit, oldusage, curusage;
4576 int children = mem_cgroup_count_children(memcg);
4580 /* see mem_cgroup_resize_res_limit */
4581 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4582 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4583 while (retry_count) {
4584 if (signal_pending(current)) {
4589 * Rather than hide all in some function, I do this in
4590 * open coded manner. You see what this really does.
4591 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4593 mutex_lock(&set_limit_mutex);
4594 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4595 if (memlimit > val) {
4597 mutex_unlock(&set_limit_mutex);
4600 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4601 if (memswlimit < val)
4603 ret = res_counter_set_limit(&memcg->memsw, val);
4605 if (memlimit == val)
4606 memcg->memsw_is_minimum = true;
4608 memcg->memsw_is_minimum = false;
4610 mutex_unlock(&set_limit_mutex);
4615 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4616 MEM_CGROUP_RECLAIM_NOSWAP |
4617 MEM_CGROUP_RECLAIM_SHRINK);
4618 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4619 /* Usage is reduced ? */
4620 if (curusage >= oldusage)
4623 oldusage = curusage;
4625 if (!ret && enlarge)
4626 memcg_oom_recover(memcg);
4630 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4632 unsigned long *total_scanned)
4634 unsigned long nr_reclaimed = 0;
4635 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4636 unsigned long reclaimed;
4638 struct mem_cgroup_tree_per_zone *mctz;
4639 unsigned long long excess;
4640 unsigned long nr_scanned;
4645 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4647 * This loop can run a while, specially if mem_cgroup's continuously
4648 * keep exceeding their soft limit and putting the system under
4655 mz = mem_cgroup_largest_soft_limit_node(mctz);
4660 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4661 gfp_mask, &nr_scanned);
4662 nr_reclaimed += reclaimed;
4663 *total_scanned += nr_scanned;
4664 spin_lock(&mctz->lock);
4667 * If we failed to reclaim anything from this memory cgroup
4668 * it is time to move on to the next cgroup
4674 * Loop until we find yet another one.
4676 * By the time we get the soft_limit lock
4677 * again, someone might have aded the
4678 * group back on the RB tree. Iterate to
4679 * make sure we get a different mem.
4680 * mem_cgroup_largest_soft_limit_node returns
4681 * NULL if no other cgroup is present on
4685 __mem_cgroup_largest_soft_limit_node(mctz);
4687 css_put(&next_mz->memcg->css);
4688 else /* next_mz == NULL or other memcg */
4692 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4693 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4695 * One school of thought says that we should not add
4696 * back the node to the tree if reclaim returns 0.
4697 * But our reclaim could return 0, simply because due
4698 * to priority we are exposing a smaller subset of
4699 * memory to reclaim from. Consider this as a longer
4702 /* If excess == 0, no tree ops */
4703 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4704 spin_unlock(&mctz->lock);
4705 css_put(&mz->memcg->css);
4708 * Could not reclaim anything and there are no more
4709 * mem cgroups to try or we seem to be looping without
4710 * reclaiming anything.
4712 if (!nr_reclaimed &&
4714 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4716 } while (!nr_reclaimed);
4718 css_put(&next_mz->memcg->css);
4719 return nr_reclaimed;
4723 * mem_cgroup_force_empty_list - clears LRU of a group
4724 * @memcg: group to clear
4727 * @lru: lru to to clear
4729 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4730 * reclaim the pages page themselves - pages are moved to the parent (or root)
4733 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4734 int node, int zid, enum lru_list lru)
4736 struct lruvec *lruvec;
4737 unsigned long flags;
4738 struct list_head *list;
4742 zone = &NODE_DATA(node)->node_zones[zid];
4743 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4744 list = &lruvec->lists[lru];
4748 struct page_cgroup *pc;
4751 spin_lock_irqsave(&zone->lru_lock, flags);
4752 if (list_empty(list)) {
4753 spin_unlock_irqrestore(&zone->lru_lock, flags);
4756 page = list_entry(list->prev, struct page, lru);
4758 list_move(&page->lru, list);
4760 spin_unlock_irqrestore(&zone->lru_lock, flags);
4763 spin_unlock_irqrestore(&zone->lru_lock, flags);
4765 pc = lookup_page_cgroup(page);
4767 if (mem_cgroup_move_parent(page, pc, memcg)) {
4768 /* found lock contention or "pc" is obsolete. */
4773 } while (!list_empty(list));
4777 * make mem_cgroup's charge to be 0 if there is no task by moving
4778 * all the charges and pages to the parent.
4779 * This enables deleting this mem_cgroup.
4781 * Caller is responsible for holding css reference on the memcg.
4783 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4789 /* This is for making all *used* pages to be on LRU. */
4790 lru_add_drain_all();
4791 drain_all_stock_sync(memcg);
4792 mem_cgroup_start_move(memcg);
4793 for_each_node_state(node, N_MEMORY) {
4794 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4797 mem_cgroup_force_empty_list(memcg,
4802 mem_cgroup_end_move(memcg);
4803 memcg_oom_recover(memcg);
4807 * Kernel memory may not necessarily be trackable to a specific
4808 * process. So they are not migrated, and therefore we can't
4809 * expect their value to drop to 0 here.
4810 * Having res filled up with kmem only is enough.
4812 * This is a safety check because mem_cgroup_force_empty_list
4813 * could have raced with mem_cgroup_replace_page_cache callers
4814 * so the lru seemed empty but the page could have been added
4815 * right after the check. RES_USAGE should be safe as we always
4816 * charge before adding to the LRU.
4818 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4819 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4820 } while (usage > 0);
4824 * This mainly exists for tests during the setting of set of use_hierarchy.
4825 * Since this is the very setting we are changing, the current hierarchy value
4828 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4832 /* bounce at first found */
4833 cgroup_for_each_child(pos, memcg->css.cgroup)
4839 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4840 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4841 * from mem_cgroup_count_children(), in the sense that we don't really care how
4842 * many children we have; we only need to know if we have any. It also counts
4843 * any memcg without hierarchy as infertile.
4845 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4847 return memcg->use_hierarchy && __memcg_has_children(memcg);
4851 * Reclaims as many pages from the given memcg as possible and moves
4852 * the rest to the parent.
4854 * Caller is responsible for holding css reference for memcg.
4856 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4858 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4859 struct cgroup *cgrp = memcg->css.cgroup;
4861 /* returns EBUSY if there is a task or if we come here twice. */
4862 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4865 /* we call try-to-free pages for make this cgroup empty */
4866 lru_add_drain_all();
4867 /* try to free all pages in this cgroup */
4868 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4871 if (signal_pending(current))
4874 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4878 /* maybe some writeback is necessary */
4879 congestion_wait(BLK_RW_ASYNC, HZ/10);
4884 mem_cgroup_reparent_charges(memcg);
4889 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
4891 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4894 if (mem_cgroup_is_root(memcg))
4896 css_get(&memcg->css);
4897 ret = mem_cgroup_force_empty(memcg);
4898 css_put(&memcg->css);
4904 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
4906 return mem_cgroup_from_cont(cont)->use_hierarchy;
4909 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
4913 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4914 struct cgroup *parent = cont->parent;
4915 struct mem_cgroup *parent_memcg = NULL;
4918 parent_memcg = mem_cgroup_from_cont(parent);
4920 mutex_lock(&memcg_create_mutex);
4922 if (memcg->use_hierarchy == val)
4926 * If parent's use_hierarchy is set, we can't make any modifications
4927 * in the child subtrees. If it is unset, then the change can
4928 * occur, provided the current cgroup has no children.
4930 * For the root cgroup, parent_mem is NULL, we allow value to be
4931 * set if there are no children.
4933 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4934 (val == 1 || val == 0)) {
4935 if (!__memcg_has_children(memcg))
4936 memcg->use_hierarchy = val;
4943 mutex_unlock(&memcg_create_mutex);
4949 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4950 enum mem_cgroup_stat_index idx)
4952 struct mem_cgroup *iter;
4955 /* Per-cpu values can be negative, use a signed accumulator */
4956 for_each_mem_cgroup_tree(iter, memcg)
4957 val += mem_cgroup_read_stat(iter, idx);
4959 if (val < 0) /* race ? */
4964 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4968 if (!mem_cgroup_is_root(memcg)) {
4970 return res_counter_read_u64(&memcg->res, RES_USAGE);
4972 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4975 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4976 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4979 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4981 return val << PAGE_SHIFT;
4984 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
4985 struct file *file, char __user *buf,
4986 size_t nbytes, loff_t *ppos)
4988 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4994 type = MEMFILE_TYPE(cft->private);
4995 name = MEMFILE_ATTR(cft->private);
4997 if (!do_swap_account && type == _MEMSWAP)
5002 if (name == RES_USAGE)
5003 val = mem_cgroup_usage(memcg, false);
5005 val = res_counter_read_u64(&memcg->res, name);
5008 if (name == RES_USAGE)
5009 val = mem_cgroup_usage(memcg, true);
5011 val = res_counter_read_u64(&memcg->memsw, name);
5014 val = res_counter_read_u64(&memcg->kmem, name);
5020 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5021 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5024 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
5027 #ifdef CONFIG_MEMCG_KMEM
5028 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5030 * For simplicity, we won't allow this to be disabled. It also can't
5031 * be changed if the cgroup has children already, or if tasks had
5034 * If tasks join before we set the limit, a person looking at
5035 * kmem.usage_in_bytes will have no way to determine when it took
5036 * place, which makes the value quite meaningless.
5038 * After it first became limited, changes in the value of the limit are
5039 * of course permitted.
5041 mutex_lock(&memcg_create_mutex);
5042 mutex_lock(&set_limit_mutex);
5043 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
5044 if (cgroup_task_count(cont) || memcg_has_children(memcg)) {
5048 ret = res_counter_set_limit(&memcg->kmem, val);
5051 ret = memcg_update_cache_sizes(memcg);
5053 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5056 static_key_slow_inc(&memcg_kmem_enabled_key);
5058 * setting the active bit after the inc will guarantee no one
5059 * starts accounting before all call sites are patched
5061 memcg_kmem_set_active(memcg);
5064 * kmem charges can outlive the cgroup. In the case of slab
5065 * pages, for instance, a page contain objects from various
5066 * processes, so it is unfeasible to migrate them away. We
5067 * need to reference count the memcg because of that.
5069 mem_cgroup_get(memcg);
5071 ret = res_counter_set_limit(&memcg->kmem, val);
5073 mutex_unlock(&set_limit_mutex);
5074 mutex_unlock(&memcg_create_mutex);
5079 #ifdef CONFIG_MEMCG_KMEM
5080 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5083 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5087 memcg->kmem_account_flags = parent->kmem_account_flags;
5089 * When that happen, we need to disable the static branch only on those
5090 * memcgs that enabled it. To achieve this, we would be forced to
5091 * complicate the code by keeping track of which memcgs were the ones
5092 * that actually enabled limits, and which ones got it from its
5095 * It is a lot simpler just to do static_key_slow_inc() on every child
5096 * that is accounted.
5098 if (!memcg_kmem_is_active(memcg))
5102 * destroy(), called if we fail, will issue static_key_slow_inc() and
5103 * mem_cgroup_put() if kmem is enabled. We have to either call them
5104 * unconditionally, or clear the KMEM_ACTIVE flag. I personally find
5105 * this more consistent, since it always leads to the same destroy path
5107 mem_cgroup_get(memcg);
5108 static_key_slow_inc(&memcg_kmem_enabled_key);
5110 mutex_lock(&set_limit_mutex);
5111 ret = memcg_update_cache_sizes(memcg);
5112 mutex_unlock(&set_limit_mutex);
5116 #endif /* CONFIG_MEMCG_KMEM */
5119 * The user of this function is...
5122 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5125 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5128 unsigned long long val;
5131 type = MEMFILE_TYPE(cft->private);
5132 name = MEMFILE_ATTR(cft->private);
5134 if (!do_swap_account && type == _MEMSWAP)
5139 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5143 /* This function does all necessary parse...reuse it */
5144 ret = res_counter_memparse_write_strategy(buffer, &val);
5148 ret = mem_cgroup_resize_limit(memcg, val);
5149 else if (type == _MEMSWAP)
5150 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5151 else if (type == _KMEM)
5152 ret = memcg_update_kmem_limit(cont, val);
5156 case RES_SOFT_LIMIT:
5157 ret = res_counter_memparse_write_strategy(buffer, &val);
5161 * For memsw, soft limits are hard to implement in terms
5162 * of semantics, for now, we support soft limits for
5163 * control without swap
5166 ret = res_counter_set_soft_limit(&memcg->res, val);
5171 ret = -EINVAL; /* should be BUG() ? */
5177 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5178 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5180 struct cgroup *cgroup;
5181 unsigned long long min_limit, min_memsw_limit, tmp;
5183 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5184 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5185 cgroup = memcg->css.cgroup;
5186 if (!memcg->use_hierarchy)
5189 while (cgroup->parent) {
5190 cgroup = cgroup->parent;
5191 memcg = mem_cgroup_from_cont(cgroup);
5192 if (!memcg->use_hierarchy)
5194 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5195 min_limit = min(min_limit, tmp);
5196 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5197 min_memsw_limit = min(min_memsw_limit, tmp);
5200 *mem_limit = min_limit;
5201 *memsw_limit = min_memsw_limit;
5204 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5206 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5210 type = MEMFILE_TYPE(event);
5211 name = MEMFILE_ATTR(event);
5213 if (!do_swap_account && type == _MEMSWAP)
5219 res_counter_reset_max(&memcg->res);
5220 else if (type == _MEMSWAP)
5221 res_counter_reset_max(&memcg->memsw);
5222 else if (type == _KMEM)
5223 res_counter_reset_max(&memcg->kmem);
5229 res_counter_reset_failcnt(&memcg->res);
5230 else if (type == _MEMSWAP)
5231 res_counter_reset_failcnt(&memcg->memsw);
5232 else if (type == _KMEM)
5233 res_counter_reset_failcnt(&memcg->kmem);
5242 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5245 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5249 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5250 struct cftype *cft, u64 val)
5252 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5254 if (val >= (1 << NR_MOVE_TYPE))
5258 * No kind of locking is needed in here, because ->can_attach() will
5259 * check this value once in the beginning of the process, and then carry
5260 * on with stale data. This means that changes to this value will only
5261 * affect task migrations starting after the change.
5263 memcg->move_charge_at_immigrate = val;
5267 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5268 struct cftype *cft, u64 val)
5275 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5279 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5280 unsigned long node_nr;
5281 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5283 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5284 seq_printf(m, "total=%lu", total_nr);
5285 for_each_node_state(nid, N_MEMORY) {
5286 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5287 seq_printf(m, " N%d=%lu", nid, node_nr);
5291 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5292 seq_printf(m, "file=%lu", file_nr);
5293 for_each_node_state(nid, N_MEMORY) {
5294 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5296 seq_printf(m, " N%d=%lu", nid, node_nr);
5300 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5301 seq_printf(m, "anon=%lu", anon_nr);
5302 for_each_node_state(nid, N_MEMORY) {
5303 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5305 seq_printf(m, " N%d=%lu", nid, node_nr);
5309 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5310 seq_printf(m, "unevictable=%lu", unevictable_nr);
5311 for_each_node_state(nid, N_MEMORY) {
5312 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5313 BIT(LRU_UNEVICTABLE));
5314 seq_printf(m, " N%d=%lu", nid, node_nr);
5319 #endif /* CONFIG_NUMA */
5321 static inline void mem_cgroup_lru_names_not_uptodate(void)
5323 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5326 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5329 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5330 struct mem_cgroup *mi;
5333 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5334 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5336 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5337 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5340 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5341 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5342 mem_cgroup_read_events(memcg, i));
5344 for (i = 0; i < NR_LRU_LISTS; i++)
5345 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5346 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5348 /* Hierarchical information */
5350 unsigned long long limit, memsw_limit;
5351 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5352 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5353 if (do_swap_account)
5354 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5358 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5361 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5363 for_each_mem_cgroup_tree(mi, memcg)
5364 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5365 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5368 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5369 unsigned long long val = 0;
5371 for_each_mem_cgroup_tree(mi, memcg)
5372 val += mem_cgroup_read_events(mi, i);
5373 seq_printf(m, "total_%s %llu\n",
5374 mem_cgroup_events_names[i], val);
5377 for (i = 0; i < NR_LRU_LISTS; i++) {
5378 unsigned long long val = 0;
5380 for_each_mem_cgroup_tree(mi, memcg)
5381 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5382 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5385 #ifdef CONFIG_DEBUG_VM
5388 struct mem_cgroup_per_zone *mz;
5389 struct zone_reclaim_stat *rstat;
5390 unsigned long recent_rotated[2] = {0, 0};
5391 unsigned long recent_scanned[2] = {0, 0};
5393 for_each_online_node(nid)
5394 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5395 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5396 rstat = &mz->lruvec.reclaim_stat;
5398 recent_rotated[0] += rstat->recent_rotated[0];
5399 recent_rotated[1] += rstat->recent_rotated[1];
5400 recent_scanned[0] += rstat->recent_scanned[0];
5401 recent_scanned[1] += rstat->recent_scanned[1];
5403 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5404 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5405 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5406 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5413 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5415 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5417 return mem_cgroup_swappiness(memcg);
5420 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5423 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5424 struct mem_cgroup *parent;
5429 if (cgrp->parent == NULL)
5432 parent = mem_cgroup_from_cont(cgrp->parent);
5434 mutex_lock(&memcg_create_mutex);
5436 /* If under hierarchy, only empty-root can set this value */
5437 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5438 mutex_unlock(&memcg_create_mutex);
5442 memcg->swappiness = val;
5444 mutex_unlock(&memcg_create_mutex);
5449 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5451 struct mem_cgroup_threshold_ary *t;
5457 t = rcu_dereference(memcg->thresholds.primary);
5459 t = rcu_dereference(memcg->memsw_thresholds.primary);
5464 usage = mem_cgroup_usage(memcg, swap);
5467 * current_threshold points to threshold just below or equal to usage.
5468 * If it's not true, a threshold was crossed after last
5469 * call of __mem_cgroup_threshold().
5471 i = t->current_threshold;
5474 * Iterate backward over array of thresholds starting from
5475 * current_threshold and check if a threshold is crossed.
5476 * If none of thresholds below usage is crossed, we read
5477 * only one element of the array here.
5479 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5480 eventfd_signal(t->entries[i].eventfd, 1);
5482 /* i = current_threshold + 1 */
5486 * Iterate forward over array of thresholds starting from
5487 * current_threshold+1 and check if a threshold is crossed.
5488 * If none of thresholds above usage is crossed, we read
5489 * only one element of the array here.
5491 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5492 eventfd_signal(t->entries[i].eventfd, 1);
5494 /* Update current_threshold */
5495 t->current_threshold = i - 1;
5500 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5503 __mem_cgroup_threshold(memcg, false);
5504 if (do_swap_account)
5505 __mem_cgroup_threshold(memcg, true);
5507 memcg = parent_mem_cgroup(memcg);
5511 static int compare_thresholds(const void *a, const void *b)
5513 const struct mem_cgroup_threshold *_a = a;
5514 const struct mem_cgroup_threshold *_b = b;
5516 return _a->threshold - _b->threshold;
5519 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5521 struct mem_cgroup_eventfd_list *ev;
5523 list_for_each_entry(ev, &memcg->oom_notify, list)
5524 eventfd_signal(ev->eventfd, 1);
5528 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5530 struct mem_cgroup *iter;
5532 for_each_mem_cgroup_tree(iter, memcg)
5533 mem_cgroup_oom_notify_cb(iter);
5536 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5537 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5539 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5540 struct mem_cgroup_thresholds *thresholds;
5541 struct mem_cgroup_threshold_ary *new;
5542 enum res_type type = MEMFILE_TYPE(cft->private);
5543 u64 threshold, usage;
5546 ret = res_counter_memparse_write_strategy(args, &threshold);
5550 mutex_lock(&memcg->thresholds_lock);
5553 thresholds = &memcg->thresholds;
5554 else if (type == _MEMSWAP)
5555 thresholds = &memcg->memsw_thresholds;
5559 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5561 /* Check if a threshold crossed before adding a new one */
5562 if (thresholds->primary)
5563 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5565 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5567 /* Allocate memory for new array of thresholds */
5568 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5576 /* Copy thresholds (if any) to new array */
5577 if (thresholds->primary) {
5578 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5579 sizeof(struct mem_cgroup_threshold));
5582 /* Add new threshold */
5583 new->entries[size - 1].eventfd = eventfd;
5584 new->entries[size - 1].threshold = threshold;
5586 /* Sort thresholds. Registering of new threshold isn't time-critical */
5587 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5588 compare_thresholds, NULL);
5590 /* Find current threshold */
5591 new->current_threshold = -1;
5592 for (i = 0; i < size; i++) {
5593 if (new->entries[i].threshold <= usage) {
5595 * new->current_threshold will not be used until
5596 * rcu_assign_pointer(), so it's safe to increment
5599 ++new->current_threshold;
5604 /* Free old spare buffer and save old primary buffer as spare */
5605 kfree(thresholds->spare);
5606 thresholds->spare = thresholds->primary;
5608 rcu_assign_pointer(thresholds->primary, new);
5610 /* To be sure that nobody uses thresholds */
5614 mutex_unlock(&memcg->thresholds_lock);
5619 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5620 struct cftype *cft, struct eventfd_ctx *eventfd)
5622 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5623 struct mem_cgroup_thresholds *thresholds;
5624 struct mem_cgroup_threshold_ary *new;
5625 enum res_type type = MEMFILE_TYPE(cft->private);
5629 mutex_lock(&memcg->thresholds_lock);
5631 thresholds = &memcg->thresholds;
5632 else if (type == _MEMSWAP)
5633 thresholds = &memcg->memsw_thresholds;
5637 if (!thresholds->primary)
5640 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5642 /* Check if a threshold crossed before removing */
5643 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5645 /* Calculate new number of threshold */
5647 for (i = 0; i < thresholds->primary->size; i++) {
5648 if (thresholds->primary->entries[i].eventfd != eventfd)
5652 new = thresholds->spare;
5654 /* Set thresholds array to NULL if we don't have thresholds */
5663 /* Copy thresholds and find current threshold */
5664 new->current_threshold = -1;
5665 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5666 if (thresholds->primary->entries[i].eventfd == eventfd)
5669 new->entries[j] = thresholds->primary->entries[i];
5670 if (new->entries[j].threshold <= usage) {
5672 * new->current_threshold will not be used
5673 * until rcu_assign_pointer(), so it's safe to increment
5676 ++new->current_threshold;
5682 /* Swap primary and spare array */
5683 thresholds->spare = thresholds->primary;
5684 /* If all events are unregistered, free the spare array */
5686 kfree(thresholds->spare);
5687 thresholds->spare = NULL;
5690 rcu_assign_pointer(thresholds->primary, new);
5692 /* To be sure that nobody uses thresholds */
5695 mutex_unlock(&memcg->thresholds_lock);
5698 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5699 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5701 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5702 struct mem_cgroup_eventfd_list *event;
5703 enum res_type type = MEMFILE_TYPE(cft->private);
5705 BUG_ON(type != _OOM_TYPE);
5706 event = kmalloc(sizeof(*event), GFP_KERNEL);
5710 spin_lock(&memcg_oom_lock);
5712 event->eventfd = eventfd;
5713 list_add(&event->list, &memcg->oom_notify);
5715 /* already in OOM ? */
5716 if (atomic_read(&memcg->under_oom))
5717 eventfd_signal(eventfd, 1);
5718 spin_unlock(&memcg_oom_lock);
5723 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5724 struct cftype *cft, struct eventfd_ctx *eventfd)
5726 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5727 struct mem_cgroup_eventfd_list *ev, *tmp;
5728 enum res_type type = MEMFILE_TYPE(cft->private);
5730 BUG_ON(type != _OOM_TYPE);
5732 spin_lock(&memcg_oom_lock);
5734 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5735 if (ev->eventfd == eventfd) {
5736 list_del(&ev->list);
5741 spin_unlock(&memcg_oom_lock);
5744 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5745 struct cftype *cft, struct cgroup_map_cb *cb)
5747 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5749 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5751 if (atomic_read(&memcg->under_oom))
5752 cb->fill(cb, "under_oom", 1);
5754 cb->fill(cb, "under_oom", 0);
5758 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5759 struct cftype *cft, u64 val)
5761 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5762 struct mem_cgroup *parent;
5764 /* cannot set to root cgroup and only 0 and 1 are allowed */
5765 if (!cgrp->parent || !((val == 0) || (val == 1)))
5768 parent = mem_cgroup_from_cont(cgrp->parent);
5770 mutex_lock(&memcg_create_mutex);
5771 /* oom-kill-disable is a flag for subhierarchy. */
5772 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5773 mutex_unlock(&memcg_create_mutex);
5776 memcg->oom_kill_disable = val;
5778 memcg_oom_recover(memcg);
5779 mutex_unlock(&memcg_create_mutex);
5783 #ifdef CONFIG_MEMCG_KMEM
5784 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5788 memcg->kmemcg_id = -1;
5789 ret = memcg_propagate_kmem(memcg);
5793 return mem_cgroup_sockets_init(memcg, ss);
5796 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5798 mem_cgroup_sockets_destroy(memcg);
5800 memcg_kmem_mark_dead(memcg);
5802 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5806 * Charges already down to 0, undo mem_cgroup_get() done in the charge
5807 * path here, being careful not to race with memcg_uncharge_kmem: it is
5808 * possible that the charges went down to 0 between mark_dead and the
5809 * res_counter read, so in that case, we don't need the put
5811 if (memcg_kmem_test_and_clear_dead(memcg))
5812 mem_cgroup_put(memcg);
5815 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5820 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
5825 static struct cftype mem_cgroup_files[] = {
5827 .name = "usage_in_bytes",
5828 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5829 .read = mem_cgroup_read,
5830 .register_event = mem_cgroup_usage_register_event,
5831 .unregister_event = mem_cgroup_usage_unregister_event,
5834 .name = "max_usage_in_bytes",
5835 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5836 .trigger = mem_cgroup_reset,
5837 .read = mem_cgroup_read,
5840 .name = "limit_in_bytes",
5841 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5842 .write_string = mem_cgroup_write,
5843 .read = mem_cgroup_read,
5846 .name = "soft_limit_in_bytes",
5847 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5848 .write_string = mem_cgroup_write,
5849 .read = mem_cgroup_read,
5853 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5854 .trigger = mem_cgroup_reset,
5855 .read = mem_cgroup_read,
5859 .read_seq_string = memcg_stat_show,
5862 .name = "force_empty",
5863 .trigger = mem_cgroup_force_empty_write,
5866 .name = "use_hierarchy",
5867 .write_u64 = mem_cgroup_hierarchy_write,
5868 .read_u64 = mem_cgroup_hierarchy_read,
5871 .name = "swappiness",
5872 .read_u64 = mem_cgroup_swappiness_read,
5873 .write_u64 = mem_cgroup_swappiness_write,
5876 .name = "move_charge_at_immigrate",
5877 .read_u64 = mem_cgroup_move_charge_read,
5878 .write_u64 = mem_cgroup_move_charge_write,
5881 .name = "oom_control",
5882 .read_map = mem_cgroup_oom_control_read,
5883 .write_u64 = mem_cgroup_oom_control_write,
5884 .register_event = mem_cgroup_oom_register_event,
5885 .unregister_event = mem_cgroup_oom_unregister_event,
5886 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5890 .name = "numa_stat",
5891 .read_seq_string = memcg_numa_stat_show,
5894 #ifdef CONFIG_MEMCG_KMEM
5896 .name = "kmem.limit_in_bytes",
5897 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5898 .write_string = mem_cgroup_write,
5899 .read = mem_cgroup_read,
5902 .name = "kmem.usage_in_bytes",
5903 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5904 .read = mem_cgroup_read,
5907 .name = "kmem.failcnt",
5908 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5909 .trigger = mem_cgroup_reset,
5910 .read = mem_cgroup_read,
5913 .name = "kmem.max_usage_in_bytes",
5914 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5915 .trigger = mem_cgroup_reset,
5916 .read = mem_cgroup_read,
5918 #ifdef CONFIG_SLABINFO
5920 .name = "kmem.slabinfo",
5921 .read_seq_string = mem_cgroup_slabinfo_read,
5925 { }, /* terminate */
5928 #ifdef CONFIG_MEMCG_SWAP
5929 static struct cftype memsw_cgroup_files[] = {
5931 .name = "memsw.usage_in_bytes",
5932 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5933 .read = mem_cgroup_read,
5934 .register_event = mem_cgroup_usage_register_event,
5935 .unregister_event = mem_cgroup_usage_unregister_event,
5938 .name = "memsw.max_usage_in_bytes",
5939 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5940 .trigger = mem_cgroup_reset,
5941 .read = mem_cgroup_read,
5944 .name = "memsw.limit_in_bytes",
5945 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5946 .write_string = mem_cgroup_write,
5947 .read = mem_cgroup_read,
5950 .name = "memsw.failcnt",
5951 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5952 .trigger = mem_cgroup_reset,
5953 .read = mem_cgroup_read,
5955 { }, /* terminate */
5958 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5960 struct mem_cgroup_per_node *pn;
5961 struct mem_cgroup_per_zone *mz;
5962 int zone, tmp = node;
5964 * This routine is called against possible nodes.
5965 * But it's BUG to call kmalloc() against offline node.
5967 * TODO: this routine can waste much memory for nodes which will
5968 * never be onlined. It's better to use memory hotplug callback
5971 if (!node_state(node, N_NORMAL_MEMORY))
5973 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5977 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5980 mz = &pn->zoneinfo[zone];
5981 lruvec_init(&mz->lruvec);
5982 for (prio = 0; prio < DEF_PRIORITY + 1; prio++)
5983 spin_lock_init(&mz->reclaim_iter[prio].iter_lock);
5984 mz->usage_in_excess = 0;
5985 mz->on_tree = false;
5988 memcg->info.nodeinfo[node] = pn;
5992 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5994 kfree(memcg->info.nodeinfo[node]);
5997 static struct mem_cgroup *mem_cgroup_alloc(void)
5999 struct mem_cgroup *memcg;
6000 size_t size = memcg_size();
6002 /* Can be very big if nr_node_ids is very big */
6003 if (size < PAGE_SIZE)
6004 memcg = kzalloc(size, GFP_KERNEL);
6006 memcg = vzalloc(size);
6011 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6014 spin_lock_init(&memcg->pcp_counter_lock);
6018 if (size < PAGE_SIZE)
6026 * At destroying mem_cgroup, references from swap_cgroup can remain.
6027 * (scanning all at force_empty is too costly...)
6029 * Instead of clearing all references at force_empty, we remember
6030 * the number of reference from swap_cgroup and free mem_cgroup when
6031 * it goes down to 0.
6033 * Removal of cgroup itself succeeds regardless of refs from swap.
6036 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6039 size_t size = memcg_size();
6041 mem_cgroup_remove_from_trees(memcg);
6042 free_css_id(&mem_cgroup_subsys, &memcg->css);
6045 free_mem_cgroup_per_zone_info(memcg, node);
6047 free_percpu(memcg->stat);
6050 * We need to make sure that (at least for now), the jump label
6051 * destruction code runs outside of the cgroup lock. This is because
6052 * get_online_cpus(), which is called from the static_branch update,
6053 * can't be called inside the cgroup_lock. cpusets are the ones
6054 * enforcing this dependency, so if they ever change, we might as well.
6056 * schedule_work() will guarantee this happens. Be careful if you need
6057 * to move this code around, and make sure it is outside
6060 disarm_static_keys(memcg);
6061 if (size < PAGE_SIZE)
6069 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
6070 * but in process context. The work_freeing structure is overlaid
6071 * on the rcu_freeing structure, which itself is overlaid on memsw.
6073 static void free_work(struct work_struct *work)
6075 struct mem_cgroup *memcg;
6077 memcg = container_of(work, struct mem_cgroup, work_freeing);
6078 __mem_cgroup_free(memcg);
6081 static void free_rcu(struct rcu_head *rcu_head)
6083 struct mem_cgroup *memcg;
6085 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
6086 INIT_WORK(&memcg->work_freeing, free_work);
6087 schedule_work(&memcg->work_freeing);
6090 static void mem_cgroup_get(struct mem_cgroup *memcg)
6092 atomic_inc(&memcg->refcnt);
6095 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
6097 if (atomic_sub_and_test(count, &memcg->refcnt)) {
6098 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
6099 call_rcu(&memcg->rcu_freeing, free_rcu);
6101 mem_cgroup_put(parent);
6105 static void mem_cgroup_put(struct mem_cgroup *memcg)
6107 __mem_cgroup_put(memcg, 1);
6111 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6113 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6115 if (!memcg->res.parent)
6117 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6119 EXPORT_SYMBOL(parent_mem_cgroup);
6121 static void __init mem_cgroup_soft_limit_tree_init(void)
6123 struct mem_cgroup_tree_per_node *rtpn;
6124 struct mem_cgroup_tree_per_zone *rtpz;
6125 int tmp, node, zone;
6127 for_each_node(node) {
6129 if (!node_state(node, N_NORMAL_MEMORY))
6131 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6134 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6136 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6137 rtpz = &rtpn->rb_tree_per_zone[zone];
6138 rtpz->rb_root = RB_ROOT;
6139 spin_lock_init(&rtpz->lock);
6144 static struct cgroup_subsys_state * __ref
6145 mem_cgroup_css_alloc(struct cgroup *cont)
6147 struct mem_cgroup *memcg;
6148 long error = -ENOMEM;
6151 memcg = mem_cgroup_alloc();
6153 return ERR_PTR(error);
6156 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6160 if (cont->parent == NULL) {
6161 root_mem_cgroup = memcg;
6162 res_counter_init(&memcg->res, NULL);
6163 res_counter_init(&memcg->memsw, NULL);
6164 res_counter_init(&memcg->kmem, NULL);
6167 memcg->last_scanned_node = MAX_NUMNODES;
6168 INIT_LIST_HEAD(&memcg->oom_notify);
6169 atomic_set(&memcg->refcnt, 1);
6170 memcg->move_charge_at_immigrate = 0;
6171 mutex_init(&memcg->thresholds_lock);
6172 spin_lock_init(&memcg->move_lock);
6177 __mem_cgroup_free(memcg);
6178 return ERR_PTR(error);
6182 mem_cgroup_css_online(struct cgroup *cont)
6184 struct mem_cgroup *memcg, *parent;
6190 mutex_lock(&memcg_create_mutex);
6191 memcg = mem_cgroup_from_cont(cont);
6192 parent = mem_cgroup_from_cont(cont->parent);
6194 memcg->use_hierarchy = parent->use_hierarchy;
6195 memcg->oom_kill_disable = parent->oom_kill_disable;
6196 memcg->swappiness = mem_cgroup_swappiness(parent);
6198 if (parent->use_hierarchy) {
6199 res_counter_init(&memcg->res, &parent->res);
6200 res_counter_init(&memcg->memsw, &parent->memsw);
6201 res_counter_init(&memcg->kmem, &parent->kmem);
6204 * We increment refcnt of the parent to ensure that we can
6205 * safely access it on res_counter_charge/uncharge.
6206 * This refcnt will be decremented when freeing this
6207 * mem_cgroup(see mem_cgroup_put).
6209 mem_cgroup_get(parent);
6211 res_counter_init(&memcg->res, NULL);
6212 res_counter_init(&memcg->memsw, NULL);
6213 res_counter_init(&memcg->kmem, NULL);
6215 * Deeper hierachy with use_hierarchy == false doesn't make
6216 * much sense so let cgroup subsystem know about this
6217 * unfortunate state in our controller.
6219 if (parent != root_mem_cgroup)
6220 mem_cgroup_subsys.broken_hierarchy = true;
6223 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6224 mutex_unlock(&memcg_create_mutex);
6227 * We call put now because our (and parent's) refcnts
6228 * are already in place. mem_cgroup_put() will internally
6229 * call __mem_cgroup_free, so return directly
6231 mem_cgroup_put(memcg);
6232 if (parent->use_hierarchy)
6233 mem_cgroup_put(parent);
6238 static void mem_cgroup_css_offline(struct cgroup *cont)
6240 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6242 mem_cgroup_reparent_charges(memcg);
6243 mem_cgroup_destroy_all_caches(memcg);
6246 static void mem_cgroup_css_free(struct cgroup *cont)
6248 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6250 kmem_cgroup_destroy(memcg);
6252 mem_cgroup_put(memcg);
6256 /* Handlers for move charge at task migration. */
6257 #define PRECHARGE_COUNT_AT_ONCE 256
6258 static int mem_cgroup_do_precharge(unsigned long count)
6261 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6262 struct mem_cgroup *memcg = mc.to;
6264 if (mem_cgroup_is_root(memcg)) {
6265 mc.precharge += count;
6266 /* we don't need css_get for root */
6269 /* try to charge at once */
6271 struct res_counter *dummy;
6273 * "memcg" cannot be under rmdir() because we've already checked
6274 * by cgroup_lock_live_cgroup() that it is not removed and we
6275 * are still under the same cgroup_mutex. So we can postpone
6278 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6280 if (do_swap_account && res_counter_charge(&memcg->memsw,
6281 PAGE_SIZE * count, &dummy)) {
6282 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6285 mc.precharge += count;
6289 /* fall back to one by one charge */
6291 if (signal_pending(current)) {
6295 if (!batch_count--) {
6296 batch_count = PRECHARGE_COUNT_AT_ONCE;
6299 ret = __mem_cgroup_try_charge(NULL,
6300 GFP_KERNEL, 1, &memcg, false);
6302 /* mem_cgroup_clear_mc() will do uncharge later */
6310 * get_mctgt_type - get target type of moving charge
6311 * @vma: the vma the pte to be checked belongs
6312 * @addr: the address corresponding to the pte to be checked
6313 * @ptent: the pte to be checked
6314 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6317 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6318 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6319 * move charge. if @target is not NULL, the page is stored in target->page
6320 * with extra refcnt got(Callers should handle it).
6321 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6322 * target for charge migration. if @target is not NULL, the entry is stored
6325 * Called with pte lock held.
6332 enum mc_target_type {
6338 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6339 unsigned long addr, pte_t ptent)
6341 struct page *page = vm_normal_page(vma, addr, ptent);
6343 if (!page || !page_mapped(page))
6345 if (PageAnon(page)) {
6346 /* we don't move shared anon */
6349 } else if (!move_file())
6350 /* we ignore mapcount for file pages */
6352 if (!get_page_unless_zero(page))
6359 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6360 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6362 struct page *page = NULL;
6363 swp_entry_t ent = pte_to_swp_entry(ptent);
6365 if (!move_anon() || non_swap_entry(ent))
6368 * Because lookup_swap_cache() updates some statistics counter,
6369 * we call find_get_page() with swapper_space directly.
6371 page = find_get_page(swap_address_space(ent), ent.val);
6372 if (do_swap_account)
6373 entry->val = ent.val;
6378 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6379 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6385 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6386 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6388 struct page *page = NULL;
6389 struct address_space *mapping;
6392 if (!vma->vm_file) /* anonymous vma */
6397 mapping = vma->vm_file->f_mapping;
6398 if (pte_none(ptent))
6399 pgoff = linear_page_index(vma, addr);
6400 else /* pte_file(ptent) is true */
6401 pgoff = pte_to_pgoff(ptent);
6403 /* page is moved even if it's not RSS of this task(page-faulted). */
6404 page = find_get_page(mapping, pgoff);
6407 /* shmem/tmpfs may report page out on swap: account for that too. */
6408 if (radix_tree_exceptional_entry(page)) {
6409 swp_entry_t swap = radix_to_swp_entry(page);
6410 if (do_swap_account)
6412 page = find_get_page(swap_address_space(swap), swap.val);
6418 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6419 unsigned long addr, pte_t ptent, union mc_target *target)
6421 struct page *page = NULL;
6422 struct page_cgroup *pc;
6423 enum mc_target_type ret = MC_TARGET_NONE;
6424 swp_entry_t ent = { .val = 0 };
6426 if (pte_present(ptent))
6427 page = mc_handle_present_pte(vma, addr, ptent);
6428 else if (is_swap_pte(ptent))
6429 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6430 else if (pte_none(ptent) || pte_file(ptent))
6431 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6433 if (!page && !ent.val)
6436 pc = lookup_page_cgroup(page);
6438 * Do only loose check w/o page_cgroup lock.
6439 * mem_cgroup_move_account() checks the pc is valid or not under
6442 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6443 ret = MC_TARGET_PAGE;
6445 target->page = page;
6447 if (!ret || !target)
6450 /* There is a swap entry and a page doesn't exist or isn't charged */
6451 if (ent.val && !ret &&
6452 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6453 ret = MC_TARGET_SWAP;
6460 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6462 * We don't consider swapping or file mapped pages because THP does not
6463 * support them for now.
6464 * Caller should make sure that pmd_trans_huge(pmd) is true.
6466 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6467 unsigned long addr, pmd_t pmd, union mc_target *target)
6469 struct page *page = NULL;
6470 struct page_cgroup *pc;
6471 enum mc_target_type ret = MC_TARGET_NONE;
6473 page = pmd_page(pmd);
6474 VM_BUG_ON(!page || !PageHead(page));
6477 pc = lookup_page_cgroup(page);
6478 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6479 ret = MC_TARGET_PAGE;
6482 target->page = page;
6488 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6489 unsigned long addr, pmd_t pmd, union mc_target *target)
6491 return MC_TARGET_NONE;
6495 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6496 unsigned long addr, unsigned long end,
6497 struct mm_walk *walk)
6499 struct vm_area_struct *vma = walk->private;
6503 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6504 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6505 mc.precharge += HPAGE_PMD_NR;
6506 spin_unlock(&vma->vm_mm->page_table_lock);
6510 if (pmd_trans_unstable(pmd))
6512 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6513 for (; addr != end; pte++, addr += PAGE_SIZE)
6514 if (get_mctgt_type(vma, addr, *pte, NULL))
6515 mc.precharge++; /* increment precharge temporarily */
6516 pte_unmap_unlock(pte - 1, ptl);
6522 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6524 unsigned long precharge;
6525 struct vm_area_struct *vma;
6527 down_read(&mm->mmap_sem);
6528 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6529 struct mm_walk mem_cgroup_count_precharge_walk = {
6530 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6534 if (is_vm_hugetlb_page(vma))
6536 walk_page_range(vma->vm_start, vma->vm_end,
6537 &mem_cgroup_count_precharge_walk);
6539 up_read(&mm->mmap_sem);
6541 precharge = mc.precharge;
6547 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6549 unsigned long precharge = mem_cgroup_count_precharge(mm);
6551 VM_BUG_ON(mc.moving_task);
6552 mc.moving_task = current;
6553 return mem_cgroup_do_precharge(precharge);
6556 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6557 static void __mem_cgroup_clear_mc(void)
6559 struct mem_cgroup *from = mc.from;
6560 struct mem_cgroup *to = mc.to;
6562 /* we must uncharge all the leftover precharges from mc.to */
6564 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6568 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6569 * we must uncharge here.
6571 if (mc.moved_charge) {
6572 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6573 mc.moved_charge = 0;
6575 /* we must fixup refcnts and charges */
6576 if (mc.moved_swap) {
6577 /* uncharge swap account from the old cgroup */
6578 if (!mem_cgroup_is_root(mc.from))
6579 res_counter_uncharge(&mc.from->memsw,
6580 PAGE_SIZE * mc.moved_swap);
6581 __mem_cgroup_put(mc.from, mc.moved_swap);
6583 if (!mem_cgroup_is_root(mc.to)) {
6585 * we charged both to->res and to->memsw, so we should
6588 res_counter_uncharge(&mc.to->res,
6589 PAGE_SIZE * mc.moved_swap);
6591 /* we've already done mem_cgroup_get(mc.to) */
6594 memcg_oom_recover(from);
6595 memcg_oom_recover(to);
6596 wake_up_all(&mc.waitq);
6599 static void mem_cgroup_clear_mc(void)
6601 struct mem_cgroup *from = mc.from;
6604 * we must clear moving_task before waking up waiters at the end of
6607 mc.moving_task = NULL;
6608 __mem_cgroup_clear_mc();
6609 spin_lock(&mc.lock);
6612 spin_unlock(&mc.lock);
6613 mem_cgroup_end_move(from);
6616 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6617 struct cgroup_taskset *tset)
6619 struct task_struct *p = cgroup_taskset_first(tset);
6621 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6622 unsigned long move_charge_at_immigrate;
6625 * We are now commited to this value whatever it is. Changes in this
6626 * tunable will only affect upcoming migrations, not the current one.
6627 * So we need to save it, and keep it going.
6629 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6630 if (move_charge_at_immigrate) {
6631 struct mm_struct *mm;
6632 struct mem_cgroup *from = mem_cgroup_from_task(p);
6634 VM_BUG_ON(from == memcg);
6636 mm = get_task_mm(p);
6639 /* We move charges only when we move a owner of the mm */
6640 if (mm->owner == p) {
6643 VM_BUG_ON(mc.precharge);
6644 VM_BUG_ON(mc.moved_charge);
6645 VM_BUG_ON(mc.moved_swap);
6646 mem_cgroup_start_move(from);
6647 spin_lock(&mc.lock);
6650 mc.immigrate_flags = move_charge_at_immigrate;
6651 spin_unlock(&mc.lock);
6652 /* We set mc.moving_task later */
6654 ret = mem_cgroup_precharge_mc(mm);
6656 mem_cgroup_clear_mc();
6663 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6664 struct cgroup_taskset *tset)
6666 mem_cgroup_clear_mc();
6669 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6670 unsigned long addr, unsigned long end,
6671 struct mm_walk *walk)
6674 struct vm_area_struct *vma = walk->private;
6677 enum mc_target_type target_type;
6678 union mc_target target;
6680 struct page_cgroup *pc;
6683 * We don't take compound_lock() here but no race with splitting thp
6685 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6686 * under splitting, which means there's no concurrent thp split,
6687 * - if another thread runs into split_huge_page() just after we
6688 * entered this if-block, the thread must wait for page table lock
6689 * to be unlocked in __split_huge_page_splitting(), where the main
6690 * part of thp split is not executed yet.
6692 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6693 if (mc.precharge < HPAGE_PMD_NR) {
6694 spin_unlock(&vma->vm_mm->page_table_lock);
6697 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6698 if (target_type == MC_TARGET_PAGE) {
6700 if (!isolate_lru_page(page)) {
6701 pc = lookup_page_cgroup(page);
6702 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6703 pc, mc.from, mc.to)) {
6704 mc.precharge -= HPAGE_PMD_NR;
6705 mc.moved_charge += HPAGE_PMD_NR;
6707 putback_lru_page(page);
6711 spin_unlock(&vma->vm_mm->page_table_lock);
6715 if (pmd_trans_unstable(pmd))
6718 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6719 for (; addr != end; addr += PAGE_SIZE) {
6720 pte_t ptent = *(pte++);
6726 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6727 case MC_TARGET_PAGE:
6729 if (isolate_lru_page(page))
6731 pc = lookup_page_cgroup(page);
6732 if (!mem_cgroup_move_account(page, 1, pc,
6735 /* we uncharge from mc.from later. */
6738 putback_lru_page(page);
6739 put: /* get_mctgt_type() gets the page */
6742 case MC_TARGET_SWAP:
6744 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6746 /* we fixup refcnts and charges later. */
6754 pte_unmap_unlock(pte - 1, ptl);
6759 * We have consumed all precharges we got in can_attach().
6760 * We try charge one by one, but don't do any additional
6761 * charges to mc.to if we have failed in charge once in attach()
6764 ret = mem_cgroup_do_precharge(1);
6772 static void mem_cgroup_move_charge(struct mm_struct *mm)
6774 struct vm_area_struct *vma;
6776 lru_add_drain_all();
6778 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6780 * Someone who are holding the mmap_sem might be waiting in
6781 * waitq. So we cancel all extra charges, wake up all waiters,
6782 * and retry. Because we cancel precharges, we might not be able
6783 * to move enough charges, but moving charge is a best-effort
6784 * feature anyway, so it wouldn't be a big problem.
6786 __mem_cgroup_clear_mc();
6790 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6792 struct mm_walk mem_cgroup_move_charge_walk = {
6793 .pmd_entry = mem_cgroup_move_charge_pte_range,
6797 if (is_vm_hugetlb_page(vma))
6799 ret = walk_page_range(vma->vm_start, vma->vm_end,
6800 &mem_cgroup_move_charge_walk);
6803 * means we have consumed all precharges and failed in
6804 * doing additional charge. Just abandon here.
6808 up_read(&mm->mmap_sem);
6811 static void mem_cgroup_move_task(struct cgroup *cont,
6812 struct cgroup_taskset *tset)
6814 struct task_struct *p = cgroup_taskset_first(tset);
6815 struct mm_struct *mm = get_task_mm(p);
6819 mem_cgroup_move_charge(mm);
6823 mem_cgroup_clear_mc();
6825 #else /* !CONFIG_MMU */
6826 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6827 struct cgroup_taskset *tset)
6831 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6832 struct cgroup_taskset *tset)
6835 static void mem_cgroup_move_task(struct cgroup *cont,
6836 struct cgroup_taskset *tset)
6841 struct cgroup_subsys mem_cgroup_subsys = {
6843 .subsys_id = mem_cgroup_subsys_id,
6844 .css_alloc = mem_cgroup_css_alloc,
6845 .css_online = mem_cgroup_css_online,
6846 .css_offline = mem_cgroup_css_offline,
6847 .css_free = mem_cgroup_css_free,
6848 .can_attach = mem_cgroup_can_attach,
6849 .cancel_attach = mem_cgroup_cancel_attach,
6850 .attach = mem_cgroup_move_task,
6851 .base_cftypes = mem_cgroup_files,
6856 #ifdef CONFIG_MEMCG_SWAP
6857 static int __init enable_swap_account(char *s)
6859 /* consider enabled if no parameter or 1 is given */
6860 if (!strcmp(s, "1"))
6861 really_do_swap_account = 1;
6862 else if (!strcmp(s, "0"))
6863 really_do_swap_account = 0;
6866 __setup("swapaccount=", enable_swap_account);
6868 static void __init memsw_file_init(void)
6870 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6873 static void __init enable_swap_cgroup(void)
6875 if (!mem_cgroup_disabled() && really_do_swap_account) {
6876 do_swap_account = 1;
6882 static void __init enable_swap_cgroup(void)
6888 * subsys_initcall() for memory controller.
6890 * Some parts like hotcpu_notifier() have to be initialized from this context
6891 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6892 * everything that doesn't depend on a specific mem_cgroup structure should
6893 * be initialized from here.
6895 static int __init mem_cgroup_init(void)
6897 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6898 enable_swap_cgroup();
6899 mem_cgroup_soft_limit_tree_init();
6903 subsys_initcall(mem_cgroup_init);