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/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.h>
52 #include <linux/vmalloc.h>
53 #include <linux/vmpressure.h>
54 #include <linux/mm_inline.h>
55 #include <linux/page_cgroup.h>
56 #include <linux/cpu.h>
57 #include <linux/oom.h>
58 #include <linux/lockdep.h>
59 #include <linux/file.h>
63 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
70 EXPORT_SYMBOL(mem_cgroup_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata = 0;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct mem_cgroup_reclaim_iter {
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup *last_visited;
152 unsigned long last_dead_count;
154 /* scan generation, increased every round-trip */
155 unsigned int generation;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone {
162 struct lruvec lruvec;
163 unsigned long lru_size[NR_LRU_LISTS];
165 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
167 struct rb_node tree_node; /* RB tree node */
168 unsigned long long usage_in_excess;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup *memcg; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node {
176 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone {
185 struct rb_root rb_root;
189 struct mem_cgroup_tree_per_node {
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
193 struct mem_cgroup_tree {
194 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
199 struct mem_cgroup_threshold {
200 struct eventfd_ctx *eventfd;
205 struct mem_cgroup_threshold_ary {
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries[0];
214 struct mem_cgroup_thresholds {
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary *primary;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary *spare;
226 struct mem_cgroup_eventfd_list {
227 struct list_head list;
228 struct eventfd_ctx *eventfd;
232 * cgroup_event represents events which userspace want to receive.
234 struct cgroup_event {
236 * css which the event belongs to.
238 struct cgroup_subsys_state *css;
240 * Control file which the event associated.
244 * eventfd to signal userspace about the event.
246 struct eventfd_ctx *eventfd;
248 * Each of these stored in a list by the cgroup.
250 struct list_head list;
252 * register_event() callback will be used to add new userspace
253 * waiter for changes related to this event. Use eventfd_signal()
254 * on eventfd to send notification to userspace.
256 int (*register_event)(struct cgroup_subsys_state *css,
257 struct cftype *cft, struct eventfd_ctx *eventfd,
260 * unregister_event() callback will be called when userspace closes
261 * the eventfd or on cgroup removing. This callback must be set,
262 * if you want provide notification functionality.
264 void (*unregister_event)(struct cgroup_subsys_state *css,
266 struct eventfd_ctx *eventfd);
268 * All fields below needed to unregister event when
269 * userspace closes eventfd.
272 wait_queue_head_t *wqh;
274 struct work_struct remove;
277 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
278 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
281 * The memory controller data structure. The memory controller controls both
282 * page cache and RSS per cgroup. We would eventually like to provide
283 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
284 * to help the administrator determine what knobs to tune.
286 * TODO: Add a water mark for the memory controller. Reclaim will begin when
287 * we hit the water mark. May be even add a low water mark, such that
288 * no reclaim occurs from a cgroup at it's low water mark, this is
289 * a feature that will be implemented much later in the future.
292 struct cgroup_subsys_state css;
294 * the counter to account for memory usage
296 struct res_counter res;
298 /* vmpressure notifications */
299 struct vmpressure vmpressure;
302 * the counter to account for mem+swap usage.
304 struct res_counter memsw;
307 * the counter to account for kernel memory usage.
309 struct res_counter kmem;
311 * Should the accounting and control be hierarchical, per subtree?
314 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
318 atomic_t oom_wakeups;
321 /* OOM-Killer disable */
322 int oom_kill_disable;
324 /* set when res.limit == memsw.limit */
325 bool memsw_is_minimum;
327 /* protect arrays of thresholds */
328 struct mutex thresholds_lock;
330 /* thresholds for memory usage. RCU-protected */
331 struct mem_cgroup_thresholds thresholds;
333 /* thresholds for mem+swap usage. RCU-protected */
334 struct mem_cgroup_thresholds memsw_thresholds;
336 /* For oom notifier event fd */
337 struct list_head oom_notify;
340 * Should we move charges of a task when a task is moved into this
341 * mem_cgroup ? And what type of charges should we move ?
343 unsigned long move_charge_at_immigrate;
345 * set > 0 if pages under this cgroup are moving to other cgroup.
347 atomic_t moving_account;
348 /* taken only while moving_account > 0 */
349 spinlock_t move_lock;
353 struct mem_cgroup_stat_cpu __percpu *stat;
355 * used when a cpu is offlined or other synchronizations
356 * See mem_cgroup_read_stat().
358 struct mem_cgroup_stat_cpu nocpu_base;
359 spinlock_t pcp_counter_lock;
362 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
363 struct tcp_memcontrol tcp_mem;
365 #if defined(CONFIG_MEMCG_KMEM)
366 /* analogous to slab_common's slab_caches list. per-memcg */
367 struct list_head memcg_slab_caches;
368 /* Not a spinlock, we can take a lot of time walking the list */
369 struct mutex slab_caches_mutex;
370 /* Index in the kmem_cache->memcg_params->memcg_caches array */
374 int last_scanned_node;
376 nodemask_t scan_nodes;
377 atomic_t numainfo_events;
378 atomic_t numainfo_updating;
381 /* List of events which userspace want to receive */
382 struct list_head event_list;
383 spinlock_t event_list_lock;
385 struct mem_cgroup_per_node *nodeinfo[0];
386 /* WARNING: nodeinfo must be the last member here */
389 static size_t memcg_size(void)
391 return sizeof(struct mem_cgroup) +
392 nr_node_ids * sizeof(struct mem_cgroup_per_node);
395 /* internal only representation about the status of kmem accounting. */
397 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
398 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
399 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
402 /* We account when limit is on, but only after call sites are patched */
403 #define KMEM_ACCOUNTED_MASK \
404 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
406 #ifdef CONFIG_MEMCG_KMEM
407 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
409 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
412 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
414 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
417 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
419 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
422 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
424 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
427 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
430 * Our caller must use css_get() first, because memcg_uncharge_kmem()
431 * will call css_put() if it sees the memcg is dead.
434 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
435 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
438 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
440 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
441 &memcg->kmem_account_flags);
445 /* Stuffs for move charges at task migration. */
447 * Types of charges to be moved. "move_charge_at_immitgrate" and
448 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
451 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
452 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
456 /* "mc" and its members are protected by cgroup_mutex */
457 static struct move_charge_struct {
458 spinlock_t lock; /* for from, to */
459 struct mem_cgroup *from;
460 struct mem_cgroup *to;
461 unsigned long immigrate_flags;
462 unsigned long precharge;
463 unsigned long moved_charge;
464 unsigned long moved_swap;
465 struct task_struct *moving_task; /* a task moving charges */
466 wait_queue_head_t waitq; /* a waitq for other context */
468 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
469 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
472 static bool move_anon(void)
474 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
477 static bool move_file(void)
479 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
483 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
484 * limit reclaim to prevent infinite loops, if they ever occur.
486 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
487 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
490 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
491 MEM_CGROUP_CHARGE_TYPE_ANON,
492 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
493 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
497 /* for encoding cft->private value on file */
505 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
506 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
507 #define MEMFILE_ATTR(val) ((val) & 0xffff)
508 /* Used for OOM nofiier */
509 #define OOM_CONTROL (0)
512 * Reclaim flags for mem_cgroup_hierarchical_reclaim
514 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
515 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
516 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
517 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
520 * The memcg_create_mutex will be held whenever a new cgroup is created.
521 * As a consequence, any change that needs to protect against new child cgroups
522 * appearing has to hold it as well.
524 static DEFINE_MUTEX(memcg_create_mutex);
526 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
528 return s ? container_of(s, struct mem_cgroup, css) : NULL;
531 /* Some nice accessors for the vmpressure. */
532 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
535 memcg = root_mem_cgroup;
536 return &memcg->vmpressure;
539 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
541 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
544 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
546 return &mem_cgroup_from_css(css)->vmpressure;
549 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
551 return (memcg == root_mem_cgroup);
554 /* Writing them here to avoid exposing memcg's inner layout */
555 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
557 void sock_update_memcg(struct sock *sk)
559 if (mem_cgroup_sockets_enabled) {
560 struct mem_cgroup *memcg;
561 struct cg_proto *cg_proto;
563 BUG_ON(!sk->sk_prot->proto_cgroup);
565 /* Socket cloning can throw us here with sk_cgrp already
566 * filled. It won't however, necessarily happen from
567 * process context. So the test for root memcg given
568 * the current task's memcg won't help us in this case.
570 * Respecting the original socket's memcg is a better
571 * decision in this case.
574 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
575 css_get(&sk->sk_cgrp->memcg->css);
580 memcg = mem_cgroup_from_task(current);
581 cg_proto = sk->sk_prot->proto_cgroup(memcg);
582 if (!mem_cgroup_is_root(memcg) &&
583 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
584 sk->sk_cgrp = cg_proto;
589 EXPORT_SYMBOL(sock_update_memcg);
591 void sock_release_memcg(struct sock *sk)
593 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
594 struct mem_cgroup *memcg;
595 WARN_ON(!sk->sk_cgrp->memcg);
596 memcg = sk->sk_cgrp->memcg;
597 css_put(&sk->sk_cgrp->memcg->css);
601 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
603 if (!memcg || mem_cgroup_is_root(memcg))
606 return &memcg->tcp_mem.cg_proto;
608 EXPORT_SYMBOL(tcp_proto_cgroup);
610 static void disarm_sock_keys(struct mem_cgroup *memcg)
612 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
614 static_key_slow_dec(&memcg_socket_limit_enabled);
617 static void disarm_sock_keys(struct mem_cgroup *memcg)
622 #ifdef CONFIG_MEMCG_KMEM
624 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
625 * There are two main reasons for not using the css_id for this:
626 * 1) this works better in sparse environments, where we have a lot of memcgs,
627 * but only a few kmem-limited. Or also, if we have, for instance, 200
628 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
629 * 200 entry array for that.
631 * 2) In order not to violate the cgroup API, we would like to do all memory
632 * allocation in ->create(). At that point, we haven't yet allocated the
633 * css_id. Having a separate index prevents us from messing with the cgroup
636 * The current size of the caches array is stored in
637 * memcg_limited_groups_array_size. It will double each time we have to
640 static DEFINE_IDA(kmem_limited_groups);
641 int memcg_limited_groups_array_size;
644 * MIN_SIZE is different than 1, because we would like to avoid going through
645 * the alloc/free process all the time. In a small machine, 4 kmem-limited
646 * cgroups is a reasonable guess. In the future, it could be a parameter or
647 * tunable, but that is strictly not necessary.
649 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
650 * this constant directly from cgroup, but it is understandable that this is
651 * better kept as an internal representation in cgroup.c. In any case, the
652 * css_id space is not getting any smaller, and we don't have to necessarily
653 * increase ours as well if it increases.
655 #define MEMCG_CACHES_MIN_SIZE 4
656 #define MEMCG_CACHES_MAX_SIZE 65535
659 * A lot of the calls to the cache allocation functions are expected to be
660 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
661 * conditional to this static branch, we'll have to allow modules that does
662 * kmem_cache_alloc and the such to see this symbol as well
664 struct static_key memcg_kmem_enabled_key;
665 EXPORT_SYMBOL(memcg_kmem_enabled_key);
667 static void disarm_kmem_keys(struct mem_cgroup *memcg)
669 if (memcg_kmem_is_active(memcg)) {
670 static_key_slow_dec(&memcg_kmem_enabled_key);
671 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
674 * This check can't live in kmem destruction function,
675 * since the charges will outlive the cgroup
677 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
680 static void disarm_kmem_keys(struct mem_cgroup *memcg)
683 #endif /* CONFIG_MEMCG_KMEM */
685 static void disarm_static_keys(struct mem_cgroup *memcg)
687 disarm_sock_keys(memcg);
688 disarm_kmem_keys(memcg);
691 static void drain_all_stock_async(struct mem_cgroup *memcg);
693 static struct mem_cgroup_per_zone *
694 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
696 VM_BUG_ON((unsigned)nid >= nr_node_ids);
697 return &memcg->nodeinfo[nid]->zoneinfo[zid];
700 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
705 static struct mem_cgroup_per_zone *
706 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
708 int nid = page_to_nid(page);
709 int zid = page_zonenum(page);
711 return mem_cgroup_zoneinfo(memcg, nid, zid);
714 static struct mem_cgroup_tree_per_zone *
715 soft_limit_tree_node_zone(int nid, int zid)
717 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
720 static struct mem_cgroup_tree_per_zone *
721 soft_limit_tree_from_page(struct page *page)
723 int nid = page_to_nid(page);
724 int zid = page_zonenum(page);
726 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
730 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
731 struct mem_cgroup_per_zone *mz,
732 struct mem_cgroup_tree_per_zone *mctz,
733 unsigned long long new_usage_in_excess)
735 struct rb_node **p = &mctz->rb_root.rb_node;
736 struct rb_node *parent = NULL;
737 struct mem_cgroup_per_zone *mz_node;
742 mz->usage_in_excess = new_usage_in_excess;
743 if (!mz->usage_in_excess)
747 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
749 if (mz->usage_in_excess < mz_node->usage_in_excess)
752 * We can't avoid mem cgroups that are over their soft
753 * limit by the same amount
755 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
758 rb_link_node(&mz->tree_node, parent, p);
759 rb_insert_color(&mz->tree_node, &mctz->rb_root);
764 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
765 struct mem_cgroup_per_zone *mz,
766 struct mem_cgroup_tree_per_zone *mctz)
770 rb_erase(&mz->tree_node, &mctz->rb_root);
775 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
776 struct mem_cgroup_per_zone *mz,
777 struct mem_cgroup_tree_per_zone *mctz)
779 spin_lock(&mctz->lock);
780 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
781 spin_unlock(&mctz->lock);
785 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
787 unsigned long long excess;
788 struct mem_cgroup_per_zone *mz;
789 struct mem_cgroup_tree_per_zone *mctz;
790 int nid = page_to_nid(page);
791 int zid = page_zonenum(page);
792 mctz = soft_limit_tree_from_page(page);
795 * Necessary to update all ancestors when hierarchy is used.
796 * because their event counter is not touched.
798 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
799 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
800 excess = res_counter_soft_limit_excess(&memcg->res);
802 * We have to update the tree if mz is on RB-tree or
803 * mem is over its softlimit.
805 if (excess || mz->on_tree) {
806 spin_lock(&mctz->lock);
807 /* if on-tree, remove it */
809 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
811 * Insert again. mz->usage_in_excess will be updated.
812 * If excess is 0, no tree ops.
814 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
815 spin_unlock(&mctz->lock);
820 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
823 struct mem_cgroup_per_zone *mz;
824 struct mem_cgroup_tree_per_zone *mctz;
826 for_each_node(node) {
827 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
828 mz = mem_cgroup_zoneinfo(memcg, node, zone);
829 mctz = soft_limit_tree_node_zone(node, zone);
830 mem_cgroup_remove_exceeded(memcg, mz, mctz);
835 static struct mem_cgroup_per_zone *
836 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
838 struct rb_node *rightmost = NULL;
839 struct mem_cgroup_per_zone *mz;
843 rightmost = rb_last(&mctz->rb_root);
845 goto done; /* Nothing to reclaim from */
847 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
849 * Remove the node now but someone else can add it back,
850 * we will to add it back at the end of reclaim to its correct
851 * position in the tree.
853 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
854 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
855 !css_tryget(&mz->memcg->css))
861 static struct mem_cgroup_per_zone *
862 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
864 struct mem_cgroup_per_zone *mz;
866 spin_lock(&mctz->lock);
867 mz = __mem_cgroup_largest_soft_limit_node(mctz);
868 spin_unlock(&mctz->lock);
873 * Implementation Note: reading percpu statistics for memcg.
875 * Both of vmstat[] and percpu_counter has threshold and do periodic
876 * synchronization to implement "quick" read. There are trade-off between
877 * reading cost and precision of value. Then, we may have a chance to implement
878 * a periodic synchronizion of counter in memcg's counter.
880 * But this _read() function is used for user interface now. The user accounts
881 * memory usage by memory cgroup and he _always_ requires exact value because
882 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
883 * have to visit all online cpus and make sum. So, for now, unnecessary
884 * synchronization is not implemented. (just implemented for cpu hotplug)
886 * If there are kernel internal actions which can make use of some not-exact
887 * value, and reading all cpu value can be performance bottleneck in some
888 * common workload, threashold and synchonization as vmstat[] should be
891 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
892 enum mem_cgroup_stat_index idx)
898 for_each_online_cpu(cpu)
899 val += per_cpu(memcg->stat->count[idx], cpu);
900 #ifdef CONFIG_HOTPLUG_CPU
901 spin_lock(&memcg->pcp_counter_lock);
902 val += memcg->nocpu_base.count[idx];
903 spin_unlock(&memcg->pcp_counter_lock);
909 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
912 int val = (charge) ? 1 : -1;
913 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
916 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
917 enum mem_cgroup_events_index idx)
919 unsigned long val = 0;
923 for_each_online_cpu(cpu)
924 val += per_cpu(memcg->stat->events[idx], cpu);
925 #ifdef CONFIG_HOTPLUG_CPU
926 spin_lock(&memcg->pcp_counter_lock);
927 val += memcg->nocpu_base.events[idx];
928 spin_unlock(&memcg->pcp_counter_lock);
934 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
936 bool anon, int nr_pages)
941 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
942 * counted as CACHE even if it's on ANON LRU.
945 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
948 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
951 if (PageTransHuge(page))
952 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
955 /* pagein of a big page is an event. So, ignore page size */
957 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
959 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
960 nr_pages = -nr_pages; /* for event */
963 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
969 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
971 struct mem_cgroup_per_zone *mz;
973 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
974 return mz->lru_size[lru];
978 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
979 unsigned int lru_mask)
981 struct mem_cgroup_per_zone *mz;
983 unsigned long ret = 0;
985 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
988 if (BIT(lru) & lru_mask)
989 ret += mz->lru_size[lru];
995 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
996 int nid, unsigned int lru_mask)
1001 for (zid = 0; zid < MAX_NR_ZONES; zid++)
1002 total += mem_cgroup_zone_nr_lru_pages(memcg,
1003 nid, zid, lru_mask);
1008 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
1009 unsigned int lru_mask)
1014 for_each_node_state(nid, N_MEMORY)
1015 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
1019 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1020 enum mem_cgroup_events_target target)
1022 unsigned long val, next;
1024 val = __this_cpu_read(memcg->stat->nr_page_events);
1025 next = __this_cpu_read(memcg->stat->targets[target]);
1026 /* from time_after() in jiffies.h */
1027 if ((long)next - (long)val < 0) {
1029 case MEM_CGROUP_TARGET_THRESH:
1030 next = val + THRESHOLDS_EVENTS_TARGET;
1032 case MEM_CGROUP_TARGET_SOFTLIMIT:
1033 next = val + SOFTLIMIT_EVENTS_TARGET;
1035 case MEM_CGROUP_TARGET_NUMAINFO:
1036 next = val + NUMAINFO_EVENTS_TARGET;
1041 __this_cpu_write(memcg->stat->targets[target], next);
1048 * Check events in order.
1051 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1054 /* threshold event is triggered in finer grain than soft limit */
1055 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1056 MEM_CGROUP_TARGET_THRESH))) {
1058 bool do_numainfo __maybe_unused;
1060 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1061 MEM_CGROUP_TARGET_SOFTLIMIT);
1062 #if MAX_NUMNODES > 1
1063 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1064 MEM_CGROUP_TARGET_NUMAINFO);
1068 mem_cgroup_threshold(memcg);
1069 if (unlikely(do_softlimit))
1070 mem_cgroup_update_tree(memcg, page);
1071 #if MAX_NUMNODES > 1
1072 if (unlikely(do_numainfo))
1073 atomic_inc(&memcg->numainfo_events);
1079 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1082 * mm_update_next_owner() may clear mm->owner to NULL
1083 * if it races with swapoff, page migration, etc.
1084 * So this can be called with p == NULL.
1089 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1092 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1094 struct mem_cgroup *memcg = NULL;
1099 * Because we have no locks, mm->owner's may be being moved to other
1100 * cgroup. We use css_tryget() here even if this looks
1101 * pessimistic (rather than adding locks here).
1105 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1106 if (unlikely(!memcg))
1108 } while (!css_tryget(&memcg->css));
1114 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1115 * ref. count) or NULL if the whole root's subtree has been visited.
1117 * helper function to be used by mem_cgroup_iter
1119 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1120 struct mem_cgroup *last_visited)
1122 struct cgroup_subsys_state *prev_css, *next_css;
1124 prev_css = last_visited ? &last_visited->css : NULL;
1126 next_css = css_next_descendant_pre(prev_css, &root->css);
1129 * Even if we found a group we have to make sure it is
1130 * alive. css && !memcg means that the groups should be
1131 * skipped and we should continue the tree walk.
1132 * last_visited css is safe to use because it is
1133 * protected by css_get and the tree walk is rcu safe.
1136 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
1138 if (css_tryget(&mem->css))
1141 prev_css = next_css;
1149 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1152 * When a group in the hierarchy below root is destroyed, the
1153 * hierarchy iterator can no longer be trusted since it might
1154 * have pointed to the destroyed group. Invalidate it.
1156 atomic_inc(&root->dead_count);
1159 static struct mem_cgroup *
1160 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1161 struct mem_cgroup *root,
1164 struct mem_cgroup *position = NULL;
1166 * A cgroup destruction happens in two stages: offlining and
1167 * release. They are separated by a RCU grace period.
1169 * If the iterator is valid, we may still race with an
1170 * offlining. The RCU lock ensures the object won't be
1171 * released, tryget will fail if we lost the race.
1173 *sequence = atomic_read(&root->dead_count);
1174 if (iter->last_dead_count == *sequence) {
1176 position = iter->last_visited;
1177 if (position && !css_tryget(&position->css))
1183 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1184 struct mem_cgroup *last_visited,
1185 struct mem_cgroup *new_position,
1189 css_put(&last_visited->css);
1191 * We store the sequence count from the time @last_visited was
1192 * loaded successfully instead of rereading it here so that we
1193 * don't lose destruction events in between. We could have
1194 * raced with the destruction of @new_position after all.
1196 iter->last_visited = new_position;
1198 iter->last_dead_count = sequence;
1202 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1203 * @root: hierarchy root
1204 * @prev: previously returned memcg, NULL on first invocation
1205 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1207 * Returns references to children of the hierarchy below @root, or
1208 * @root itself, or %NULL after a full round-trip.
1210 * Caller must pass the return value in @prev on subsequent
1211 * invocations for reference counting, or use mem_cgroup_iter_break()
1212 * to cancel a hierarchy walk before the round-trip is complete.
1214 * Reclaimers can specify a zone and a priority level in @reclaim to
1215 * divide up the memcgs in the hierarchy among all concurrent
1216 * reclaimers operating on the same zone and priority.
1218 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1219 struct mem_cgroup *prev,
1220 struct mem_cgroup_reclaim_cookie *reclaim)
1222 struct mem_cgroup *memcg = NULL;
1223 struct mem_cgroup *last_visited = NULL;
1225 if (mem_cgroup_disabled())
1229 root = root_mem_cgroup;
1231 if (prev && !reclaim)
1232 last_visited = prev;
1234 if (!root->use_hierarchy && root != root_mem_cgroup) {
1242 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1243 int uninitialized_var(seq);
1246 int nid = zone_to_nid(reclaim->zone);
1247 int zid = zone_idx(reclaim->zone);
1248 struct mem_cgroup_per_zone *mz;
1250 mz = mem_cgroup_zoneinfo(root, nid, zid);
1251 iter = &mz->reclaim_iter[reclaim->priority];
1252 if (prev && reclaim->generation != iter->generation) {
1253 iter->last_visited = NULL;
1257 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1260 memcg = __mem_cgroup_iter_next(root, last_visited);
1263 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1267 else if (!prev && memcg)
1268 reclaim->generation = iter->generation;
1277 if (prev && prev != root)
1278 css_put(&prev->css);
1284 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1285 * @root: hierarchy root
1286 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1288 void mem_cgroup_iter_break(struct mem_cgroup *root,
1289 struct mem_cgroup *prev)
1292 root = root_mem_cgroup;
1293 if (prev && prev != root)
1294 css_put(&prev->css);
1298 * Iteration constructs for visiting all cgroups (under a tree). If
1299 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1300 * be used for reference counting.
1302 #define for_each_mem_cgroup_tree(iter, root) \
1303 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1305 iter = mem_cgroup_iter(root, iter, NULL))
1307 #define for_each_mem_cgroup(iter) \
1308 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1310 iter = mem_cgroup_iter(NULL, iter, NULL))
1312 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1314 struct mem_cgroup *memcg;
1317 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1318 if (unlikely(!memcg))
1323 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1326 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1334 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1337 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1338 * @zone: zone of the wanted lruvec
1339 * @memcg: memcg of the wanted lruvec
1341 * Returns the lru list vector holding pages for the given @zone and
1342 * @mem. This can be the global zone lruvec, if the memory controller
1345 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1346 struct mem_cgroup *memcg)
1348 struct mem_cgroup_per_zone *mz;
1349 struct lruvec *lruvec;
1351 if (mem_cgroup_disabled()) {
1352 lruvec = &zone->lruvec;
1356 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1357 lruvec = &mz->lruvec;
1360 * Since a node can be onlined after the mem_cgroup was created,
1361 * we have to be prepared to initialize lruvec->zone here;
1362 * and if offlined then reonlined, we need to reinitialize it.
1364 if (unlikely(lruvec->zone != zone))
1365 lruvec->zone = zone;
1370 * Following LRU functions are allowed to be used without PCG_LOCK.
1371 * Operations are called by routine of global LRU independently from memcg.
1372 * What we have to take care of here is validness of pc->mem_cgroup.
1374 * Changes to pc->mem_cgroup happens when
1377 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1378 * It is added to LRU before charge.
1379 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1380 * When moving account, the page is not on LRU. It's isolated.
1384 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1386 * @zone: zone of the page
1388 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1390 struct mem_cgroup_per_zone *mz;
1391 struct mem_cgroup *memcg;
1392 struct page_cgroup *pc;
1393 struct lruvec *lruvec;
1395 if (mem_cgroup_disabled()) {
1396 lruvec = &zone->lruvec;
1400 pc = lookup_page_cgroup(page);
1401 memcg = pc->mem_cgroup;
1404 * Surreptitiously switch any uncharged offlist page to root:
1405 * an uncharged page off lru does nothing to secure
1406 * its former mem_cgroup from sudden removal.
1408 * Our caller holds lru_lock, and PageCgroupUsed is updated
1409 * under page_cgroup lock: between them, they make all uses
1410 * of pc->mem_cgroup safe.
1412 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1413 pc->mem_cgroup = memcg = root_mem_cgroup;
1415 mz = page_cgroup_zoneinfo(memcg, page);
1416 lruvec = &mz->lruvec;
1419 * Since a node can be onlined after the mem_cgroup was created,
1420 * we have to be prepared to initialize lruvec->zone here;
1421 * and if offlined then reonlined, we need to reinitialize it.
1423 if (unlikely(lruvec->zone != zone))
1424 lruvec->zone = zone;
1429 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1430 * @lruvec: mem_cgroup per zone lru vector
1431 * @lru: index of lru list the page is sitting on
1432 * @nr_pages: positive when adding or negative when removing
1434 * This function must be called when a page is added to or removed from an
1437 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1440 struct mem_cgroup_per_zone *mz;
1441 unsigned long *lru_size;
1443 if (mem_cgroup_disabled())
1446 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1447 lru_size = mz->lru_size + lru;
1448 *lru_size += nr_pages;
1449 VM_BUG_ON((long)(*lru_size) < 0);
1453 * Checks whether given mem is same or in the root_mem_cgroup's
1456 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1457 struct mem_cgroup *memcg)
1459 if (root_memcg == memcg)
1461 if (!root_memcg->use_hierarchy || !memcg)
1463 return css_is_ancestor(&memcg->css, &root_memcg->css);
1466 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1467 struct mem_cgroup *memcg)
1472 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1477 bool task_in_mem_cgroup(struct task_struct *task,
1478 const struct mem_cgroup *memcg)
1480 struct mem_cgroup *curr = NULL;
1481 struct task_struct *p;
1484 p = find_lock_task_mm(task);
1486 curr = try_get_mem_cgroup_from_mm(p->mm);
1490 * All threads may have already detached their mm's, but the oom
1491 * killer still needs to detect if they have already been oom
1492 * killed to prevent needlessly killing additional tasks.
1495 curr = mem_cgroup_from_task(task);
1497 css_get(&curr->css);
1503 * We should check use_hierarchy of "memcg" not "curr". Because checking
1504 * use_hierarchy of "curr" here make this function true if hierarchy is
1505 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1506 * hierarchy(even if use_hierarchy is disabled in "memcg").
1508 ret = mem_cgroup_same_or_subtree(memcg, curr);
1509 css_put(&curr->css);
1513 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1515 unsigned long inactive_ratio;
1516 unsigned long inactive;
1517 unsigned long active;
1520 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1521 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1523 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1525 inactive_ratio = int_sqrt(10 * gb);
1529 return inactive * inactive_ratio < active;
1532 #define mem_cgroup_from_res_counter(counter, member) \
1533 container_of(counter, struct mem_cgroup, member)
1536 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1537 * @memcg: the memory cgroup
1539 * Returns the maximum amount of memory @mem can be charged with, in
1542 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1544 unsigned long long margin;
1546 margin = res_counter_margin(&memcg->res);
1547 if (do_swap_account)
1548 margin = min(margin, res_counter_margin(&memcg->memsw));
1549 return margin >> PAGE_SHIFT;
1552 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1555 if (!css_parent(&memcg->css))
1556 return vm_swappiness;
1558 return memcg->swappiness;
1562 * memcg->moving_account is used for checking possibility that some thread is
1563 * calling move_account(). When a thread on CPU-A starts moving pages under
1564 * a memcg, other threads should check memcg->moving_account under
1565 * rcu_read_lock(), like this:
1569 * memcg->moving_account+1 if (memcg->mocing_account)
1571 * synchronize_rcu() update something.
1576 /* for quick checking without looking up memcg */
1577 atomic_t memcg_moving __read_mostly;
1579 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1581 atomic_inc(&memcg_moving);
1582 atomic_inc(&memcg->moving_account);
1586 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1589 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1590 * We check NULL in callee rather than caller.
1593 atomic_dec(&memcg_moving);
1594 atomic_dec(&memcg->moving_account);
1599 * 2 routines for checking "mem" is under move_account() or not.
1601 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1602 * is used for avoiding races in accounting. If true,
1603 * pc->mem_cgroup may be overwritten.
1605 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1606 * under hierarchy of moving cgroups. This is for
1607 * waiting at hith-memory prressure caused by "move".
1610 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1612 VM_BUG_ON(!rcu_read_lock_held());
1613 return atomic_read(&memcg->moving_account) > 0;
1616 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1618 struct mem_cgroup *from;
1619 struct mem_cgroup *to;
1622 * Unlike task_move routines, we access mc.to, mc.from not under
1623 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1625 spin_lock(&mc.lock);
1631 ret = mem_cgroup_same_or_subtree(memcg, from)
1632 || mem_cgroup_same_or_subtree(memcg, to);
1634 spin_unlock(&mc.lock);
1638 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1640 if (mc.moving_task && current != mc.moving_task) {
1641 if (mem_cgroup_under_move(memcg)) {
1643 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1644 /* moving charge context might have finished. */
1647 finish_wait(&mc.waitq, &wait);
1655 * Take this lock when
1656 * - a code tries to modify page's memcg while it's USED.
1657 * - a code tries to modify page state accounting in a memcg.
1658 * see mem_cgroup_stolen(), too.
1660 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1661 unsigned long *flags)
1663 spin_lock_irqsave(&memcg->move_lock, *flags);
1666 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1667 unsigned long *flags)
1669 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1672 #define K(x) ((x) << (PAGE_SHIFT-10))
1674 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1675 * @memcg: The memory cgroup that went over limit
1676 * @p: Task that is going to be killed
1678 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1681 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1683 struct cgroup *task_cgrp;
1684 struct cgroup *mem_cgrp;
1686 * Need a buffer in BSS, can't rely on allocations. The code relies
1687 * on the assumption that OOM is serialized for memory controller.
1688 * If this assumption is broken, revisit this code.
1690 static char memcg_name[PATH_MAX];
1692 struct mem_cgroup *iter;
1700 mem_cgrp = memcg->css.cgroup;
1701 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1703 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1706 * Unfortunately, we are unable to convert to a useful name
1707 * But we'll still print out the usage information
1714 pr_info("Task in %s killed", memcg_name);
1717 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1725 * Continues from above, so we don't need an KERN_ level
1727 pr_cont(" as a result of limit of %s\n", memcg_name);
1730 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1731 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1732 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1733 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1734 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1735 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1736 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1737 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1738 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1739 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1740 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1741 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1743 for_each_mem_cgroup_tree(iter, memcg) {
1744 pr_info("Memory cgroup stats");
1747 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1749 pr_cont(" for %s", memcg_name);
1753 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1754 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1756 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1757 K(mem_cgroup_read_stat(iter, i)));
1760 for (i = 0; i < NR_LRU_LISTS; i++)
1761 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1762 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1769 * This function returns the number of memcg under hierarchy tree. Returns
1770 * 1(self count) if no children.
1772 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1775 struct mem_cgroup *iter;
1777 for_each_mem_cgroup_tree(iter, memcg)
1783 * Return the memory (and swap, if configured) limit for a memcg.
1785 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1789 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1792 * Do not consider swap space if we cannot swap due to swappiness
1794 if (mem_cgroup_swappiness(memcg)) {
1797 limit += total_swap_pages << PAGE_SHIFT;
1798 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1801 * If memsw is finite and limits the amount of swap space
1802 * available to this memcg, return that limit.
1804 limit = min(limit, memsw);
1810 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1813 struct mem_cgroup *iter;
1814 unsigned long chosen_points = 0;
1815 unsigned long totalpages;
1816 unsigned int points = 0;
1817 struct task_struct *chosen = NULL;
1820 * If current has a pending SIGKILL or is exiting, then automatically
1821 * select it. The goal is to allow it to allocate so that it may
1822 * quickly exit and free its memory.
1824 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1825 set_thread_flag(TIF_MEMDIE);
1829 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1830 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1831 for_each_mem_cgroup_tree(iter, memcg) {
1832 struct css_task_iter it;
1833 struct task_struct *task;
1835 css_task_iter_start(&iter->css, &it);
1836 while ((task = css_task_iter_next(&it))) {
1837 switch (oom_scan_process_thread(task, totalpages, NULL,
1839 case OOM_SCAN_SELECT:
1841 put_task_struct(chosen);
1843 chosen_points = ULONG_MAX;
1844 get_task_struct(chosen);
1846 case OOM_SCAN_CONTINUE:
1848 case OOM_SCAN_ABORT:
1849 css_task_iter_end(&it);
1850 mem_cgroup_iter_break(memcg, iter);
1852 put_task_struct(chosen);
1857 points = oom_badness(task, memcg, NULL, totalpages);
1858 if (points > chosen_points) {
1860 put_task_struct(chosen);
1862 chosen_points = points;
1863 get_task_struct(chosen);
1866 css_task_iter_end(&it);
1871 points = chosen_points * 1000 / totalpages;
1872 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1873 NULL, "Memory cgroup out of memory");
1876 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1878 unsigned long flags)
1880 unsigned long total = 0;
1881 bool noswap = false;
1884 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1886 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1889 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1891 drain_all_stock_async(memcg);
1892 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1894 * Allow limit shrinkers, which are triggered directly
1895 * by userspace, to catch signals and stop reclaim
1896 * after minimal progress, regardless of the margin.
1898 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1900 if (mem_cgroup_margin(memcg))
1903 * If nothing was reclaimed after two attempts, there
1904 * may be no reclaimable pages in this hierarchy.
1913 * test_mem_cgroup_node_reclaimable
1914 * @memcg: the target memcg
1915 * @nid: the node ID to be checked.
1916 * @noswap : specify true here if the user wants flle only information.
1918 * This function returns whether the specified memcg contains any
1919 * reclaimable pages on a node. Returns true if there are any reclaimable
1920 * pages in the node.
1922 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1923 int nid, bool noswap)
1925 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1927 if (noswap || !total_swap_pages)
1929 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1934 #if MAX_NUMNODES > 1
1937 * Always updating the nodemask is not very good - even if we have an empty
1938 * list or the wrong list here, we can start from some node and traverse all
1939 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1942 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1946 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1947 * pagein/pageout changes since the last update.
1949 if (!atomic_read(&memcg->numainfo_events))
1951 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1954 /* make a nodemask where this memcg uses memory from */
1955 memcg->scan_nodes = node_states[N_MEMORY];
1957 for_each_node_mask(nid, node_states[N_MEMORY]) {
1959 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1960 node_clear(nid, memcg->scan_nodes);
1963 atomic_set(&memcg->numainfo_events, 0);
1964 atomic_set(&memcg->numainfo_updating, 0);
1968 * Selecting a node where we start reclaim from. Because what we need is just
1969 * reducing usage counter, start from anywhere is O,K. Considering
1970 * memory reclaim from current node, there are pros. and cons.
1972 * Freeing memory from current node means freeing memory from a node which
1973 * we'll use or we've used. So, it may make LRU bad. And if several threads
1974 * hit limits, it will see a contention on a node. But freeing from remote
1975 * node means more costs for memory reclaim because of memory latency.
1977 * Now, we use round-robin. Better algorithm is welcomed.
1979 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1983 mem_cgroup_may_update_nodemask(memcg);
1984 node = memcg->last_scanned_node;
1986 node = next_node(node, memcg->scan_nodes);
1987 if (node == MAX_NUMNODES)
1988 node = first_node(memcg->scan_nodes);
1990 * We call this when we hit limit, not when pages are added to LRU.
1991 * No LRU may hold pages because all pages are UNEVICTABLE or
1992 * memcg is too small and all pages are not on LRU. In that case,
1993 * we use curret node.
1995 if (unlikely(node == MAX_NUMNODES))
1996 node = numa_node_id();
1998 memcg->last_scanned_node = node;
2003 * Check all nodes whether it contains reclaimable pages or not.
2004 * For quick scan, we make use of scan_nodes. This will allow us to skip
2005 * unused nodes. But scan_nodes is lazily updated and may not cotain
2006 * enough new information. We need to do double check.
2008 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2013 * quick check...making use of scan_node.
2014 * We can skip unused nodes.
2016 if (!nodes_empty(memcg->scan_nodes)) {
2017 for (nid = first_node(memcg->scan_nodes);
2019 nid = next_node(nid, memcg->scan_nodes)) {
2021 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2026 * Check rest of nodes.
2028 for_each_node_state(nid, N_MEMORY) {
2029 if (node_isset(nid, memcg->scan_nodes))
2031 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2038 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2043 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2045 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2049 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2052 unsigned long *total_scanned)
2054 struct mem_cgroup *victim = NULL;
2057 unsigned long excess;
2058 unsigned long nr_scanned;
2059 struct mem_cgroup_reclaim_cookie reclaim = {
2064 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2067 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2072 * If we have not been able to reclaim
2073 * anything, it might because there are
2074 * no reclaimable pages under this hierarchy
2079 * We want to do more targeted reclaim.
2080 * excess >> 2 is not to excessive so as to
2081 * reclaim too much, nor too less that we keep
2082 * coming back to reclaim from this cgroup
2084 if (total >= (excess >> 2) ||
2085 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2090 if (!mem_cgroup_reclaimable(victim, false))
2092 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2094 *total_scanned += nr_scanned;
2095 if (!res_counter_soft_limit_excess(&root_memcg->res))
2098 mem_cgroup_iter_break(root_memcg, victim);
2102 #ifdef CONFIG_LOCKDEP
2103 static struct lockdep_map memcg_oom_lock_dep_map = {
2104 .name = "memcg_oom_lock",
2108 static DEFINE_SPINLOCK(memcg_oom_lock);
2111 * Check OOM-Killer is already running under our hierarchy.
2112 * If someone is running, return false.
2114 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2116 struct mem_cgroup *iter, *failed = NULL;
2118 spin_lock(&memcg_oom_lock);
2120 for_each_mem_cgroup_tree(iter, memcg) {
2121 if (iter->oom_lock) {
2123 * this subtree of our hierarchy is already locked
2124 * so we cannot give a lock.
2127 mem_cgroup_iter_break(memcg, iter);
2130 iter->oom_lock = true;
2135 * OK, we failed to lock the whole subtree so we have
2136 * to clean up what we set up to the failing subtree
2138 for_each_mem_cgroup_tree(iter, memcg) {
2139 if (iter == failed) {
2140 mem_cgroup_iter_break(memcg, iter);
2143 iter->oom_lock = false;
2146 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2148 spin_unlock(&memcg_oom_lock);
2153 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2155 struct mem_cgroup *iter;
2157 spin_lock(&memcg_oom_lock);
2158 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2159 for_each_mem_cgroup_tree(iter, memcg)
2160 iter->oom_lock = false;
2161 spin_unlock(&memcg_oom_lock);
2164 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2166 struct mem_cgroup *iter;
2168 for_each_mem_cgroup_tree(iter, memcg)
2169 atomic_inc(&iter->under_oom);
2172 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2174 struct mem_cgroup *iter;
2177 * When a new child is created while the hierarchy is under oom,
2178 * mem_cgroup_oom_lock() may not be called. We have to use
2179 * atomic_add_unless() here.
2181 for_each_mem_cgroup_tree(iter, memcg)
2182 atomic_add_unless(&iter->under_oom, -1, 0);
2185 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2187 struct oom_wait_info {
2188 struct mem_cgroup *memcg;
2192 static int memcg_oom_wake_function(wait_queue_t *wait,
2193 unsigned mode, int sync, void *arg)
2195 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2196 struct mem_cgroup *oom_wait_memcg;
2197 struct oom_wait_info *oom_wait_info;
2199 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2200 oom_wait_memcg = oom_wait_info->memcg;
2203 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2204 * Then we can use css_is_ancestor without taking care of RCU.
2206 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2207 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2209 return autoremove_wake_function(wait, mode, sync, arg);
2212 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2214 atomic_inc(&memcg->oom_wakeups);
2215 /* for filtering, pass "memcg" as argument. */
2216 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2219 static void memcg_oom_recover(struct mem_cgroup *memcg)
2221 if (memcg && atomic_read(&memcg->under_oom))
2222 memcg_wakeup_oom(memcg);
2225 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2227 if (!current->memcg_oom.may_oom)
2230 * We are in the middle of the charge context here, so we
2231 * don't want to block when potentially sitting on a callstack
2232 * that holds all kinds of filesystem and mm locks.
2234 * Also, the caller may handle a failed allocation gracefully
2235 * (like optional page cache readahead) and so an OOM killer
2236 * invocation might not even be necessary.
2238 * That's why we don't do anything here except remember the
2239 * OOM context and then deal with it at the end of the page
2240 * fault when the stack is unwound, the locks are released,
2241 * and when we know whether the fault was overall successful.
2243 css_get(&memcg->css);
2244 current->memcg_oom.memcg = memcg;
2245 current->memcg_oom.gfp_mask = mask;
2246 current->memcg_oom.order = order;
2250 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2251 * @handle: actually kill/wait or just clean up the OOM state
2253 * This has to be called at the end of a page fault if the memcg OOM
2254 * handler was enabled.
2256 * Memcg supports userspace OOM handling where failed allocations must
2257 * sleep on a waitqueue until the userspace task resolves the
2258 * situation. Sleeping directly in the charge context with all kinds
2259 * of locks held is not a good idea, instead we remember an OOM state
2260 * in the task and mem_cgroup_oom_synchronize() has to be called at
2261 * the end of the page fault to complete the OOM handling.
2263 * Returns %true if an ongoing memcg OOM situation was detected and
2264 * completed, %false otherwise.
2266 bool mem_cgroup_oom_synchronize(bool handle)
2268 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2269 struct oom_wait_info owait;
2272 /* OOM is global, do not handle */
2279 owait.memcg = memcg;
2280 owait.wait.flags = 0;
2281 owait.wait.func = memcg_oom_wake_function;
2282 owait.wait.private = current;
2283 INIT_LIST_HEAD(&owait.wait.task_list);
2285 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2286 mem_cgroup_mark_under_oom(memcg);
2288 locked = mem_cgroup_oom_trylock(memcg);
2291 mem_cgroup_oom_notify(memcg);
2293 if (locked && !memcg->oom_kill_disable) {
2294 mem_cgroup_unmark_under_oom(memcg);
2295 finish_wait(&memcg_oom_waitq, &owait.wait);
2296 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2297 current->memcg_oom.order);
2300 mem_cgroup_unmark_under_oom(memcg);
2301 finish_wait(&memcg_oom_waitq, &owait.wait);
2305 mem_cgroup_oom_unlock(memcg);
2307 * There is no guarantee that an OOM-lock contender
2308 * sees the wakeups triggered by the OOM kill
2309 * uncharges. Wake any sleepers explicitely.
2311 memcg_oom_recover(memcg);
2314 current->memcg_oom.memcg = NULL;
2315 css_put(&memcg->css);
2320 * Currently used to update mapped file statistics, but the routine can be
2321 * generalized to update other statistics as well.
2323 * Notes: Race condition
2325 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2326 * it tends to be costly. But considering some conditions, we doesn't need
2327 * to do so _always_.
2329 * Considering "charge", lock_page_cgroup() is not required because all
2330 * file-stat operations happen after a page is attached to radix-tree. There
2331 * are no race with "charge".
2333 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2334 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2335 * if there are race with "uncharge". Statistics itself is properly handled
2338 * Considering "move", this is an only case we see a race. To make the race
2339 * small, we check mm->moving_account and detect there are possibility of race
2340 * If there is, we take a lock.
2343 void __mem_cgroup_begin_update_page_stat(struct page *page,
2344 bool *locked, unsigned long *flags)
2346 struct mem_cgroup *memcg;
2347 struct page_cgroup *pc;
2349 pc = lookup_page_cgroup(page);
2351 memcg = pc->mem_cgroup;
2352 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2355 * If this memory cgroup is not under account moving, we don't
2356 * need to take move_lock_mem_cgroup(). Because we already hold
2357 * rcu_read_lock(), any calls to move_account will be delayed until
2358 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2360 if (!mem_cgroup_stolen(memcg))
2363 move_lock_mem_cgroup(memcg, flags);
2364 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2365 move_unlock_mem_cgroup(memcg, flags);
2371 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2373 struct page_cgroup *pc = lookup_page_cgroup(page);
2376 * It's guaranteed that pc->mem_cgroup never changes while
2377 * lock is held because a routine modifies pc->mem_cgroup
2378 * should take move_lock_mem_cgroup().
2380 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2383 void mem_cgroup_update_page_stat(struct page *page,
2384 enum mem_cgroup_stat_index idx, int val)
2386 struct mem_cgroup *memcg;
2387 struct page_cgroup *pc = lookup_page_cgroup(page);
2388 unsigned long uninitialized_var(flags);
2390 if (mem_cgroup_disabled())
2393 VM_BUG_ON(!rcu_read_lock_held());
2394 memcg = pc->mem_cgroup;
2395 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2398 this_cpu_add(memcg->stat->count[idx], val);
2402 * size of first charge trial. "32" comes from vmscan.c's magic value.
2403 * TODO: maybe necessary to use big numbers in big irons.
2405 #define CHARGE_BATCH 32U
2406 struct memcg_stock_pcp {
2407 struct mem_cgroup *cached; /* this never be root cgroup */
2408 unsigned int nr_pages;
2409 struct work_struct work;
2410 unsigned long flags;
2411 #define FLUSHING_CACHED_CHARGE 0
2413 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2414 static DEFINE_MUTEX(percpu_charge_mutex);
2417 * consume_stock: Try to consume stocked charge on this cpu.
2418 * @memcg: memcg to consume from.
2419 * @nr_pages: how many pages to charge.
2421 * The charges will only happen if @memcg matches the current cpu's memcg
2422 * stock, and at least @nr_pages are available in that stock. Failure to
2423 * service an allocation will refill the stock.
2425 * returns true if successful, false otherwise.
2427 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2429 struct memcg_stock_pcp *stock;
2432 if (nr_pages > CHARGE_BATCH)
2435 stock = &get_cpu_var(memcg_stock);
2436 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2437 stock->nr_pages -= nr_pages;
2438 else /* need to call res_counter_charge */
2440 put_cpu_var(memcg_stock);
2445 * Returns stocks cached in percpu to res_counter and reset cached information.
2447 static void drain_stock(struct memcg_stock_pcp *stock)
2449 struct mem_cgroup *old = stock->cached;
2451 if (stock->nr_pages) {
2452 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2454 res_counter_uncharge(&old->res, bytes);
2455 if (do_swap_account)
2456 res_counter_uncharge(&old->memsw, bytes);
2457 stock->nr_pages = 0;
2459 stock->cached = NULL;
2463 * This must be called under preempt disabled or must be called by
2464 * a thread which is pinned to local cpu.
2466 static void drain_local_stock(struct work_struct *dummy)
2468 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2470 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2473 static void __init memcg_stock_init(void)
2477 for_each_possible_cpu(cpu) {
2478 struct memcg_stock_pcp *stock =
2479 &per_cpu(memcg_stock, cpu);
2480 INIT_WORK(&stock->work, drain_local_stock);
2485 * Cache charges(val) which is from res_counter, to local per_cpu area.
2486 * This will be consumed by consume_stock() function, later.
2488 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2490 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2492 if (stock->cached != memcg) { /* reset if necessary */
2494 stock->cached = memcg;
2496 stock->nr_pages += nr_pages;
2497 put_cpu_var(memcg_stock);
2501 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2502 * of the hierarchy under it. sync flag says whether we should block
2503 * until the work is done.
2505 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2509 /* Notify other cpus that system-wide "drain" is running */
2512 for_each_online_cpu(cpu) {
2513 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2514 struct mem_cgroup *memcg;
2516 memcg = stock->cached;
2517 if (!memcg || !stock->nr_pages)
2519 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2521 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2523 drain_local_stock(&stock->work);
2525 schedule_work_on(cpu, &stock->work);
2533 for_each_online_cpu(cpu) {
2534 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2535 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2536 flush_work(&stock->work);
2543 * Tries to drain stocked charges in other cpus. This function is asynchronous
2544 * and just put a work per cpu for draining localy on each cpu. Caller can
2545 * expects some charges will be back to res_counter later but cannot wait for
2548 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2551 * If someone calls draining, avoid adding more kworker runs.
2553 if (!mutex_trylock(&percpu_charge_mutex))
2555 drain_all_stock(root_memcg, false);
2556 mutex_unlock(&percpu_charge_mutex);
2559 /* This is a synchronous drain interface. */
2560 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2562 /* called when force_empty is called */
2563 mutex_lock(&percpu_charge_mutex);
2564 drain_all_stock(root_memcg, true);
2565 mutex_unlock(&percpu_charge_mutex);
2569 * This function drains percpu counter value from DEAD cpu and
2570 * move it to local cpu. Note that this function can be preempted.
2572 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2576 spin_lock(&memcg->pcp_counter_lock);
2577 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2578 long x = per_cpu(memcg->stat->count[i], cpu);
2580 per_cpu(memcg->stat->count[i], cpu) = 0;
2581 memcg->nocpu_base.count[i] += x;
2583 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2584 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2586 per_cpu(memcg->stat->events[i], cpu) = 0;
2587 memcg->nocpu_base.events[i] += x;
2589 spin_unlock(&memcg->pcp_counter_lock);
2592 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2593 unsigned long action,
2596 int cpu = (unsigned long)hcpu;
2597 struct memcg_stock_pcp *stock;
2598 struct mem_cgroup *iter;
2600 if (action == CPU_ONLINE)
2603 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2606 for_each_mem_cgroup(iter)
2607 mem_cgroup_drain_pcp_counter(iter, cpu);
2609 stock = &per_cpu(memcg_stock, cpu);
2615 /* See __mem_cgroup_try_charge() for details */
2617 CHARGE_OK, /* success */
2618 CHARGE_RETRY, /* need to retry but retry is not bad */
2619 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2620 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2623 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2624 unsigned int nr_pages, unsigned int min_pages,
2627 unsigned long csize = nr_pages * PAGE_SIZE;
2628 struct mem_cgroup *mem_over_limit;
2629 struct res_counter *fail_res;
2630 unsigned long flags = 0;
2633 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2636 if (!do_swap_account)
2638 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2642 res_counter_uncharge(&memcg->res, csize);
2643 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2644 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2646 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2648 * Never reclaim on behalf of optional batching, retry with a
2649 * single page instead.
2651 if (nr_pages > min_pages)
2652 return CHARGE_RETRY;
2654 if (!(gfp_mask & __GFP_WAIT))
2655 return CHARGE_WOULDBLOCK;
2657 if (gfp_mask & __GFP_NORETRY)
2658 return CHARGE_NOMEM;
2660 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2661 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2662 return CHARGE_RETRY;
2664 * Even though the limit is exceeded at this point, reclaim
2665 * may have been able to free some pages. Retry the charge
2666 * before killing the task.
2668 * Only for regular pages, though: huge pages are rather
2669 * unlikely to succeed so close to the limit, and we fall back
2670 * to regular pages anyway in case of failure.
2672 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2673 return CHARGE_RETRY;
2676 * At task move, charge accounts can be doubly counted. So, it's
2677 * better to wait until the end of task_move if something is going on.
2679 if (mem_cgroup_wait_acct_move(mem_over_limit))
2680 return CHARGE_RETRY;
2683 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2685 return CHARGE_NOMEM;
2689 * __mem_cgroup_try_charge() does
2690 * 1. detect memcg to be charged against from passed *mm and *ptr,
2691 * 2. update res_counter
2692 * 3. call memory reclaim if necessary.
2694 * In some special case, if the task is fatal, fatal_signal_pending() or
2695 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2696 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2697 * as possible without any hazards. 2: all pages should have a valid
2698 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2699 * pointer, that is treated as a charge to root_mem_cgroup.
2701 * So __mem_cgroup_try_charge() will return
2702 * 0 ... on success, filling *ptr with a valid memcg pointer.
2703 * -ENOMEM ... charge failure because of resource limits.
2704 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2706 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2707 * the oom-killer can be invoked.
2709 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2711 unsigned int nr_pages,
2712 struct mem_cgroup **ptr,
2715 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2716 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2717 struct mem_cgroup *memcg = NULL;
2721 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2722 * in system level. So, allow to go ahead dying process in addition to
2725 if (unlikely(test_thread_flag(TIF_MEMDIE)
2726 || fatal_signal_pending(current)))
2729 if (unlikely(task_in_memcg_oom(current)))
2733 * We always charge the cgroup the mm_struct belongs to.
2734 * The mm_struct's mem_cgroup changes on task migration if the
2735 * thread group leader migrates. It's possible that mm is not
2736 * set, if so charge the root memcg (happens for pagecache usage).
2739 *ptr = root_mem_cgroup;
2741 if (*ptr) { /* css should be a valid one */
2743 if (mem_cgroup_is_root(memcg))
2745 if (consume_stock(memcg, nr_pages))
2747 css_get(&memcg->css);
2749 struct task_struct *p;
2752 p = rcu_dereference(mm->owner);
2754 * Because we don't have task_lock(), "p" can exit.
2755 * In that case, "memcg" can point to root or p can be NULL with
2756 * race with swapoff. Then, we have small risk of mis-accouning.
2757 * But such kind of mis-account by race always happens because
2758 * we don't have cgroup_mutex(). It's overkill and we allo that
2760 * (*) swapoff at el will charge against mm-struct not against
2761 * task-struct. So, mm->owner can be NULL.
2763 memcg = mem_cgroup_from_task(p);
2765 memcg = root_mem_cgroup;
2766 if (mem_cgroup_is_root(memcg)) {
2770 if (consume_stock(memcg, nr_pages)) {
2772 * It seems dagerous to access memcg without css_get().
2773 * But considering how consume_stok works, it's not
2774 * necessary. If consume_stock success, some charges
2775 * from this memcg are cached on this cpu. So, we
2776 * don't need to call css_get()/css_tryget() before
2777 * calling consume_stock().
2782 /* after here, we may be blocked. we need to get refcnt */
2783 if (!css_tryget(&memcg->css)) {
2791 bool invoke_oom = oom && !nr_oom_retries;
2793 /* If killed, bypass charge */
2794 if (fatal_signal_pending(current)) {
2795 css_put(&memcg->css);
2799 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2800 nr_pages, invoke_oom);
2804 case CHARGE_RETRY: /* not in OOM situation but retry */
2806 css_put(&memcg->css);
2809 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2810 css_put(&memcg->css);
2812 case CHARGE_NOMEM: /* OOM routine works */
2813 if (!oom || invoke_oom) {
2814 css_put(&memcg->css);
2820 } while (ret != CHARGE_OK);
2822 if (batch > nr_pages)
2823 refill_stock(memcg, batch - nr_pages);
2824 css_put(&memcg->css);
2829 if (!(gfp_mask & __GFP_NOFAIL)) {
2834 *ptr = root_mem_cgroup;
2839 * Somemtimes we have to undo a charge we got by try_charge().
2840 * This function is for that and do uncharge, put css's refcnt.
2841 * gotten by try_charge().
2843 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2844 unsigned int nr_pages)
2846 if (!mem_cgroup_is_root(memcg)) {
2847 unsigned long bytes = nr_pages * PAGE_SIZE;
2849 res_counter_uncharge(&memcg->res, bytes);
2850 if (do_swap_account)
2851 res_counter_uncharge(&memcg->memsw, bytes);
2856 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2857 * This is useful when moving usage to parent cgroup.
2859 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2860 unsigned int nr_pages)
2862 unsigned long bytes = nr_pages * PAGE_SIZE;
2864 if (mem_cgroup_is_root(memcg))
2867 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2868 if (do_swap_account)
2869 res_counter_uncharge_until(&memcg->memsw,
2870 memcg->memsw.parent, bytes);
2874 * A helper function to get mem_cgroup from ID. must be called under
2875 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2876 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2877 * called against removed memcg.)
2879 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2881 struct cgroup_subsys_state *css;
2883 /* ID 0 is unused ID */
2886 css = css_lookup(&mem_cgroup_subsys, id);
2889 return mem_cgroup_from_css(css);
2892 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2894 struct mem_cgroup *memcg = NULL;
2895 struct page_cgroup *pc;
2899 VM_BUG_ON(!PageLocked(page));
2901 pc = lookup_page_cgroup(page);
2902 lock_page_cgroup(pc);
2903 if (PageCgroupUsed(pc)) {
2904 memcg = pc->mem_cgroup;
2905 if (memcg && !css_tryget(&memcg->css))
2907 } else if (PageSwapCache(page)) {
2908 ent.val = page_private(page);
2909 id = lookup_swap_cgroup_id(ent);
2911 memcg = mem_cgroup_lookup(id);
2912 if (memcg && !css_tryget(&memcg->css))
2916 unlock_page_cgroup(pc);
2920 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2922 unsigned int nr_pages,
2923 enum charge_type ctype,
2926 struct page_cgroup *pc = lookup_page_cgroup(page);
2927 struct zone *uninitialized_var(zone);
2928 struct lruvec *lruvec;
2929 bool was_on_lru = false;
2932 lock_page_cgroup(pc);
2933 VM_BUG_ON(PageCgroupUsed(pc));
2935 * we don't need page_cgroup_lock about tail pages, becase they are not
2936 * accessed by any other context at this point.
2940 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2941 * may already be on some other mem_cgroup's LRU. Take care of it.
2944 zone = page_zone(page);
2945 spin_lock_irq(&zone->lru_lock);
2946 if (PageLRU(page)) {
2947 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2949 del_page_from_lru_list(page, lruvec, page_lru(page));
2954 pc->mem_cgroup = memcg;
2956 * We access a page_cgroup asynchronously without lock_page_cgroup().
2957 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2958 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2959 * before USED bit, we need memory barrier here.
2960 * See mem_cgroup_add_lru_list(), etc.
2963 SetPageCgroupUsed(pc);
2967 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2968 VM_BUG_ON(PageLRU(page));
2970 add_page_to_lru_list(page, lruvec, page_lru(page));
2972 spin_unlock_irq(&zone->lru_lock);
2975 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2980 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2981 unlock_page_cgroup(pc);
2984 * "charge_statistics" updated event counter. Then, check it.
2985 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2986 * if they exceeds softlimit.
2988 memcg_check_events(memcg, page);
2991 static DEFINE_MUTEX(set_limit_mutex);
2993 #ifdef CONFIG_MEMCG_KMEM
2994 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2996 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2997 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
3001 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
3002 * in the memcg_cache_params struct.
3004 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
3006 struct kmem_cache *cachep;
3008 VM_BUG_ON(p->is_root_cache);
3009 cachep = p->root_cache;
3010 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
3013 #ifdef CONFIG_SLABINFO
3014 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
3015 struct cftype *cft, struct seq_file *m)
3017 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
3018 struct memcg_cache_params *params;
3020 if (!memcg_can_account_kmem(memcg))
3023 print_slabinfo_header(m);
3025 mutex_lock(&memcg->slab_caches_mutex);
3026 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
3027 cache_show(memcg_params_to_cache(params), m);
3028 mutex_unlock(&memcg->slab_caches_mutex);
3034 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3036 struct res_counter *fail_res;
3037 struct mem_cgroup *_memcg;
3041 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3046 * Conditions under which we can wait for the oom_killer. Those are
3047 * the same conditions tested by the core page allocator
3049 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
3052 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3055 if (ret == -EINTR) {
3057 * __mem_cgroup_try_charge() chosed to bypass to root due to
3058 * OOM kill or fatal signal. Since our only options are to
3059 * either fail the allocation or charge it to this cgroup, do
3060 * it as a temporary condition. But we can't fail. From a
3061 * kmem/slab perspective, the cache has already been selected,
3062 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3065 * This condition will only trigger if the task entered
3066 * memcg_charge_kmem in a sane state, but was OOM-killed during
3067 * __mem_cgroup_try_charge() above. Tasks that were already
3068 * dying when the allocation triggers should have been already
3069 * directed to the root cgroup in memcontrol.h
3071 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3072 if (do_swap_account)
3073 res_counter_charge_nofail(&memcg->memsw, size,
3077 res_counter_uncharge(&memcg->kmem, size);
3082 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3084 res_counter_uncharge(&memcg->res, size);
3085 if (do_swap_account)
3086 res_counter_uncharge(&memcg->memsw, size);
3089 if (res_counter_uncharge(&memcg->kmem, size))
3093 * Releases a reference taken in kmem_cgroup_css_offline in case
3094 * this last uncharge is racing with the offlining code or it is
3095 * outliving the memcg existence.
3097 * The memory barrier imposed by test&clear is paired with the
3098 * explicit one in memcg_kmem_mark_dead().
3100 if (memcg_kmem_test_and_clear_dead(memcg))
3101 css_put(&memcg->css);
3104 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3109 mutex_lock(&memcg->slab_caches_mutex);
3110 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3111 mutex_unlock(&memcg->slab_caches_mutex);
3115 * helper for acessing a memcg's index. It will be used as an index in the
3116 * child cache array in kmem_cache, and also to derive its name. This function
3117 * will return -1 when this is not a kmem-limited memcg.
3119 int memcg_cache_id(struct mem_cgroup *memcg)
3121 return memcg ? memcg->kmemcg_id : -1;
3125 * This ends up being protected by the set_limit mutex, during normal
3126 * operation, because that is its main call site.
3128 * But when we create a new cache, we can call this as well if its parent
3129 * is kmem-limited. That will have to hold set_limit_mutex as well.
3131 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3135 num = ida_simple_get(&kmem_limited_groups,
3136 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3140 * After this point, kmem_accounted (that we test atomically in
3141 * the beginning of this conditional), is no longer 0. This
3142 * guarantees only one process will set the following boolean
3143 * to true. We don't need test_and_set because we're protected
3144 * by the set_limit_mutex anyway.
3146 memcg_kmem_set_activated(memcg);
3148 ret = memcg_update_all_caches(num+1);
3150 ida_simple_remove(&kmem_limited_groups, num);
3151 memcg_kmem_clear_activated(memcg);
3155 memcg->kmemcg_id = num;
3156 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3157 mutex_init(&memcg->slab_caches_mutex);
3161 static size_t memcg_caches_array_size(int num_groups)
3164 if (num_groups <= 0)
3167 size = 2 * num_groups;
3168 if (size < MEMCG_CACHES_MIN_SIZE)
3169 size = MEMCG_CACHES_MIN_SIZE;
3170 else if (size > MEMCG_CACHES_MAX_SIZE)
3171 size = MEMCG_CACHES_MAX_SIZE;
3177 * We should update the current array size iff all caches updates succeed. This
3178 * can only be done from the slab side. The slab mutex needs to be held when
3181 void memcg_update_array_size(int num)
3183 if (num > memcg_limited_groups_array_size)
3184 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3187 static void kmem_cache_destroy_work_func(struct work_struct *w);
3189 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3191 struct memcg_cache_params *cur_params = s->memcg_params;
3193 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3195 if (num_groups > memcg_limited_groups_array_size) {
3197 ssize_t size = memcg_caches_array_size(num_groups);
3199 size *= sizeof(void *);
3200 size += offsetof(struct memcg_cache_params, memcg_caches);
3202 s->memcg_params = kzalloc(size, GFP_KERNEL);
3203 if (!s->memcg_params) {
3204 s->memcg_params = cur_params;
3208 s->memcg_params->is_root_cache = true;
3211 * There is the chance it will be bigger than
3212 * memcg_limited_groups_array_size, if we failed an allocation
3213 * in a cache, in which case all caches updated before it, will
3214 * have a bigger array.
3216 * But if that is the case, the data after
3217 * memcg_limited_groups_array_size is certainly unused
3219 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3220 if (!cur_params->memcg_caches[i])
3222 s->memcg_params->memcg_caches[i] =
3223 cur_params->memcg_caches[i];
3227 * Ideally, we would wait until all caches succeed, and only
3228 * then free the old one. But this is not worth the extra
3229 * pointer per-cache we'd have to have for this.
3231 * It is not a big deal if some caches are left with a size
3232 * bigger than the others. And all updates will reset this
3240 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3241 struct kmem_cache *root_cache)
3245 if (!memcg_kmem_enabled())
3249 size = offsetof(struct memcg_cache_params, memcg_caches);
3250 size += memcg_limited_groups_array_size * sizeof(void *);
3252 size = sizeof(struct memcg_cache_params);
3254 s->memcg_params = kzalloc(size, GFP_KERNEL);
3255 if (!s->memcg_params)
3259 s->memcg_params->memcg = memcg;
3260 s->memcg_params->root_cache = root_cache;
3261 INIT_WORK(&s->memcg_params->destroy,
3262 kmem_cache_destroy_work_func);
3264 s->memcg_params->is_root_cache = true;
3269 void memcg_release_cache(struct kmem_cache *s)
3271 struct kmem_cache *root;
3272 struct mem_cgroup *memcg;
3276 * This happens, for instance, when a root cache goes away before we
3279 if (!s->memcg_params)
3282 if (s->memcg_params->is_root_cache)
3285 memcg = s->memcg_params->memcg;
3286 id = memcg_cache_id(memcg);
3288 root = s->memcg_params->root_cache;
3289 root->memcg_params->memcg_caches[id] = NULL;
3291 mutex_lock(&memcg->slab_caches_mutex);
3292 list_del(&s->memcg_params->list);
3293 mutex_unlock(&memcg->slab_caches_mutex);
3295 css_put(&memcg->css);
3297 kfree(s->memcg_params);
3301 * During the creation a new cache, we need to disable our accounting mechanism
3302 * altogether. This is true even if we are not creating, but rather just
3303 * enqueing new caches to be created.
3305 * This is because that process will trigger allocations; some visible, like
3306 * explicit kmallocs to auxiliary data structures, name strings and internal
3307 * cache structures; some well concealed, like INIT_WORK() that can allocate
3308 * objects during debug.
3310 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3311 * to it. This may not be a bounded recursion: since the first cache creation
3312 * failed to complete (waiting on the allocation), we'll just try to create the
3313 * cache again, failing at the same point.
3315 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3316 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3317 * inside the following two functions.
3319 static inline void memcg_stop_kmem_account(void)
3321 VM_BUG_ON(!current->mm);
3322 current->memcg_kmem_skip_account++;
3325 static inline void memcg_resume_kmem_account(void)
3327 VM_BUG_ON(!current->mm);
3328 current->memcg_kmem_skip_account--;
3331 static void kmem_cache_destroy_work_func(struct work_struct *w)
3333 struct kmem_cache *cachep;
3334 struct memcg_cache_params *p;
3336 p = container_of(w, struct memcg_cache_params, destroy);
3338 cachep = memcg_params_to_cache(p);
3341 * If we get down to 0 after shrink, we could delete right away.
3342 * However, memcg_release_pages() already puts us back in the workqueue
3343 * in that case. If we proceed deleting, we'll get a dangling
3344 * reference, and removing the object from the workqueue in that case
3345 * is unnecessary complication. We are not a fast path.
3347 * Note that this case is fundamentally different from racing with
3348 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3349 * kmem_cache_shrink, not only we would be reinserting a dead cache
3350 * into the queue, but doing so from inside the worker racing to
3353 * So if we aren't down to zero, we'll just schedule a worker and try
3356 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3357 kmem_cache_shrink(cachep);
3358 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3361 kmem_cache_destroy(cachep);
3364 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3366 if (!cachep->memcg_params->dead)
3370 * There are many ways in which we can get here.
3372 * We can get to a memory-pressure situation while the delayed work is
3373 * still pending to run. The vmscan shrinkers can then release all
3374 * cache memory and get us to destruction. If this is the case, we'll
3375 * be executed twice, which is a bug (the second time will execute over
3376 * bogus data). In this case, cancelling the work should be fine.
3378 * But we can also get here from the worker itself, if
3379 * kmem_cache_shrink is enough to shake all the remaining objects and
3380 * get the page count to 0. In this case, we'll deadlock if we try to
3381 * cancel the work (the worker runs with an internal lock held, which
3382 * is the same lock we would hold for cancel_work_sync().)
3384 * Since we can't possibly know who got us here, just refrain from
3385 * running if there is already work pending
3387 if (work_pending(&cachep->memcg_params->destroy))
3390 * We have to defer the actual destroying to a workqueue, because
3391 * we might currently be in a context that cannot sleep.
3393 schedule_work(&cachep->memcg_params->destroy);
3397 * This lock protects updaters, not readers. We want readers to be as fast as
3398 * they can, and they will either see NULL or a valid cache value. Our model
3399 * allow them to see NULL, in which case the root memcg will be selected.
3401 * We need this lock because multiple allocations to the same cache from a non
3402 * will span more than one worker. Only one of them can create the cache.
3404 static DEFINE_MUTEX(memcg_cache_mutex);
3407 * Called with memcg_cache_mutex held
3409 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3410 struct kmem_cache *s)
3412 struct kmem_cache *new;
3413 static char *tmp_name = NULL;
3415 lockdep_assert_held(&memcg_cache_mutex);
3418 * kmem_cache_create_memcg duplicates the given name and
3419 * cgroup_name for this name requires RCU context.
3420 * This static temporary buffer is used to prevent from
3421 * pointless shortliving allocation.
3424 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3430 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3431 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3434 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3435 (s->flags & ~SLAB_PANIC), s->ctor, s);
3438 new->allocflags |= __GFP_KMEMCG;
3443 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3444 struct kmem_cache *cachep)
3446 struct kmem_cache *new_cachep;
3449 BUG_ON(!memcg_can_account_kmem(memcg));
3451 idx = memcg_cache_id(memcg);
3453 mutex_lock(&memcg_cache_mutex);
3454 new_cachep = cachep->memcg_params->memcg_caches[idx];
3456 css_put(&memcg->css);
3460 new_cachep = kmem_cache_dup(memcg, cachep);
3461 if (new_cachep == NULL) {
3462 new_cachep = cachep;
3463 css_put(&memcg->css);
3467 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3469 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3471 * the readers won't lock, make sure everybody sees the updated value,
3472 * so they won't put stuff in the queue again for no reason
3476 mutex_unlock(&memcg_cache_mutex);
3480 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3482 struct kmem_cache *c;
3485 if (!s->memcg_params)
3487 if (!s->memcg_params->is_root_cache)
3491 * If the cache is being destroyed, we trust that there is no one else
3492 * requesting objects from it. Even if there are, the sanity checks in
3493 * kmem_cache_destroy should caught this ill-case.
3495 * Still, we don't want anyone else freeing memcg_caches under our
3496 * noses, which can happen if a new memcg comes to life. As usual,
3497 * we'll take the set_limit_mutex to protect ourselves against this.
3499 mutex_lock(&set_limit_mutex);
3500 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3501 c = s->memcg_params->memcg_caches[i];
3506 * We will now manually delete the caches, so to avoid races
3507 * we need to cancel all pending destruction workers and
3508 * proceed with destruction ourselves.
3510 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3511 * and that could spawn the workers again: it is likely that
3512 * the cache still have active pages until this very moment.
3513 * This would lead us back to mem_cgroup_destroy_cache.
3515 * But that will not execute at all if the "dead" flag is not
3516 * set, so flip it down to guarantee we are in control.
3518 c->memcg_params->dead = false;
3519 cancel_work_sync(&c->memcg_params->destroy);
3520 kmem_cache_destroy(c);
3522 mutex_unlock(&set_limit_mutex);
3525 struct create_work {
3526 struct mem_cgroup *memcg;
3527 struct kmem_cache *cachep;
3528 struct work_struct work;
3531 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3533 struct kmem_cache *cachep;
3534 struct memcg_cache_params *params;
3536 if (!memcg_kmem_is_active(memcg))
3539 mutex_lock(&memcg->slab_caches_mutex);
3540 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3541 cachep = memcg_params_to_cache(params);
3542 cachep->memcg_params->dead = true;
3543 schedule_work(&cachep->memcg_params->destroy);
3545 mutex_unlock(&memcg->slab_caches_mutex);
3548 static void memcg_create_cache_work_func(struct work_struct *w)
3550 struct create_work *cw;
3552 cw = container_of(w, struct create_work, work);
3553 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3558 * Enqueue the creation of a per-memcg kmem_cache.
3560 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3561 struct kmem_cache *cachep)
3563 struct create_work *cw;
3565 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3567 css_put(&memcg->css);
3572 cw->cachep = cachep;
3574 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3575 schedule_work(&cw->work);
3578 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3579 struct kmem_cache *cachep)
3582 * We need to stop accounting when we kmalloc, because if the
3583 * corresponding kmalloc cache is not yet created, the first allocation
3584 * in __memcg_create_cache_enqueue will recurse.
3586 * However, it is better to enclose the whole function. Depending on
3587 * the debugging options enabled, INIT_WORK(), for instance, can
3588 * trigger an allocation. This too, will make us recurse. Because at
3589 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3590 * the safest choice is to do it like this, wrapping the whole function.
3592 memcg_stop_kmem_account();
3593 __memcg_create_cache_enqueue(memcg, cachep);
3594 memcg_resume_kmem_account();
3597 * Return the kmem_cache we're supposed to use for a slab allocation.
3598 * We try to use the current memcg's version of the cache.
3600 * If the cache does not exist yet, if we are the first user of it,
3601 * we either create it immediately, if possible, or create it asynchronously
3603 * In the latter case, we will let the current allocation go through with
3604 * the original cache.
3606 * Can't be called in interrupt context or from kernel threads.
3607 * This function needs to be called with rcu_read_lock() held.
3609 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3612 struct mem_cgroup *memcg;
3615 VM_BUG_ON(!cachep->memcg_params);
3616 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3618 if (!current->mm || current->memcg_kmem_skip_account)
3622 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3624 if (!memcg_can_account_kmem(memcg))
3627 idx = memcg_cache_id(memcg);
3630 * barrier to mare sure we're always seeing the up to date value. The
3631 * code updating memcg_caches will issue a write barrier to match this.
3633 read_barrier_depends();
3634 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3635 cachep = cachep->memcg_params->memcg_caches[idx];
3639 /* The corresponding put will be done in the workqueue. */
3640 if (!css_tryget(&memcg->css))
3645 * If we are in a safe context (can wait, and not in interrupt
3646 * context), we could be be predictable and return right away.
3647 * This would guarantee that the allocation being performed
3648 * already belongs in the new cache.
3650 * However, there are some clashes that can arrive from locking.
3651 * For instance, because we acquire the slab_mutex while doing
3652 * kmem_cache_dup, this means no further allocation could happen
3653 * with the slab_mutex held.
3655 * Also, because cache creation issue get_online_cpus(), this
3656 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3657 * that ends up reversed during cpu hotplug. (cpuset allocates
3658 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3659 * better to defer everything.
3661 memcg_create_cache_enqueue(memcg, cachep);
3667 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3670 * We need to verify if the allocation against current->mm->owner's memcg is
3671 * possible for the given order. But the page is not allocated yet, so we'll
3672 * need a further commit step to do the final arrangements.
3674 * It is possible for the task to switch cgroups in this mean time, so at
3675 * commit time, we can't rely on task conversion any longer. We'll then use
3676 * the handle argument to return to the caller which cgroup we should commit
3677 * against. We could also return the memcg directly and avoid the pointer
3678 * passing, but a boolean return value gives better semantics considering
3679 * the compiled-out case as well.
3681 * Returning true means the allocation is possible.
3684 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3686 struct mem_cgroup *memcg;
3692 * Disabling accounting is only relevant for some specific memcg
3693 * internal allocations. Therefore we would initially not have such
3694 * check here, since direct calls to the page allocator that are marked
3695 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3696 * concerned with cache allocations, and by having this test at
3697 * memcg_kmem_get_cache, we are already able to relay the allocation to
3698 * the root cache and bypass the memcg cache altogether.
3700 * There is one exception, though: the SLUB allocator does not create
3701 * large order caches, but rather service large kmallocs directly from
3702 * the page allocator. Therefore, the following sequence when backed by
3703 * the SLUB allocator:
3705 * memcg_stop_kmem_account();
3706 * kmalloc(<large_number>)
3707 * memcg_resume_kmem_account();
3709 * would effectively ignore the fact that we should skip accounting,
3710 * since it will drive us directly to this function without passing
3711 * through the cache selector memcg_kmem_get_cache. Such large
3712 * allocations are extremely rare but can happen, for instance, for the
3713 * cache arrays. We bring this test here.
3715 if (!current->mm || current->memcg_kmem_skip_account)
3718 memcg = try_get_mem_cgroup_from_mm(current->mm);
3721 * very rare case described in mem_cgroup_from_task. Unfortunately there
3722 * isn't much we can do without complicating this too much, and it would
3723 * be gfp-dependent anyway. Just let it go
3725 if (unlikely(!memcg))
3728 if (!memcg_can_account_kmem(memcg)) {
3729 css_put(&memcg->css);
3733 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3737 css_put(&memcg->css);
3741 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3744 struct page_cgroup *pc;
3746 VM_BUG_ON(mem_cgroup_is_root(memcg));
3748 /* The page allocation failed. Revert */
3750 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3754 pc = lookup_page_cgroup(page);
3755 lock_page_cgroup(pc);
3756 pc->mem_cgroup = memcg;
3757 SetPageCgroupUsed(pc);
3758 unlock_page_cgroup(pc);
3761 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3763 struct mem_cgroup *memcg = NULL;
3764 struct page_cgroup *pc;
3767 pc = lookup_page_cgroup(page);
3769 * Fast unlocked return. Theoretically might have changed, have to
3770 * check again after locking.
3772 if (!PageCgroupUsed(pc))
3775 lock_page_cgroup(pc);
3776 if (PageCgroupUsed(pc)) {
3777 memcg = pc->mem_cgroup;
3778 ClearPageCgroupUsed(pc);
3780 unlock_page_cgroup(pc);
3783 * We trust that only if there is a memcg associated with the page, it
3784 * is a valid allocation
3789 VM_BUG_ON(mem_cgroup_is_root(memcg));
3790 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3793 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3796 #endif /* CONFIG_MEMCG_KMEM */
3798 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3800 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3802 * Because tail pages are not marked as "used", set it. We're under
3803 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3804 * charge/uncharge will be never happen and move_account() is done under
3805 * compound_lock(), so we don't have to take care of races.
3807 void mem_cgroup_split_huge_fixup(struct page *head)
3809 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3810 struct page_cgroup *pc;
3811 struct mem_cgroup *memcg;
3814 if (mem_cgroup_disabled())
3817 memcg = head_pc->mem_cgroup;
3818 for (i = 1; i < HPAGE_PMD_NR; i++) {
3820 pc->mem_cgroup = memcg;
3821 smp_wmb();/* see __commit_charge() */
3822 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3824 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3827 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3830 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3831 struct mem_cgroup *to,
3832 unsigned int nr_pages,
3833 enum mem_cgroup_stat_index idx)
3835 /* Update stat data for mem_cgroup */
3837 __this_cpu_sub(from->stat->count[idx], nr_pages);
3838 __this_cpu_add(to->stat->count[idx], nr_pages);
3843 * mem_cgroup_move_account - move account of the page
3845 * @nr_pages: number of regular pages (>1 for huge pages)
3846 * @pc: page_cgroup of the page.
3847 * @from: mem_cgroup which the page is moved from.
3848 * @to: mem_cgroup which the page is moved to. @from != @to.
3850 * The caller must confirm following.
3851 * - page is not on LRU (isolate_page() is useful.)
3852 * - compound_lock is held when nr_pages > 1
3854 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3857 static int mem_cgroup_move_account(struct page *page,
3858 unsigned int nr_pages,
3859 struct page_cgroup *pc,
3860 struct mem_cgroup *from,
3861 struct mem_cgroup *to)
3863 unsigned long flags;
3865 bool anon = PageAnon(page);
3867 VM_BUG_ON(from == to);
3868 VM_BUG_ON(PageLRU(page));
3870 * The page is isolated from LRU. So, collapse function
3871 * will not handle this page. But page splitting can happen.
3872 * Do this check under compound_page_lock(). The caller should
3876 if (nr_pages > 1 && !PageTransHuge(page))
3879 lock_page_cgroup(pc);
3882 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3885 move_lock_mem_cgroup(from, &flags);
3887 if (!anon && page_mapped(page))
3888 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3889 MEM_CGROUP_STAT_FILE_MAPPED);
3891 if (PageWriteback(page))
3892 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3893 MEM_CGROUP_STAT_WRITEBACK);
3895 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3897 /* caller should have done css_get */
3898 pc->mem_cgroup = to;
3899 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3900 move_unlock_mem_cgroup(from, &flags);
3903 unlock_page_cgroup(pc);
3907 memcg_check_events(to, page);
3908 memcg_check_events(from, page);
3914 * mem_cgroup_move_parent - moves page to the parent group
3915 * @page: the page to move
3916 * @pc: page_cgroup of the page
3917 * @child: page's cgroup
3919 * move charges to its parent or the root cgroup if the group has no
3920 * parent (aka use_hierarchy==0).
3921 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3922 * mem_cgroup_move_account fails) the failure is always temporary and
3923 * it signals a race with a page removal/uncharge or migration. In the
3924 * first case the page is on the way out and it will vanish from the LRU
3925 * on the next attempt and the call should be retried later.
3926 * Isolation from the LRU fails only if page has been isolated from
3927 * the LRU since we looked at it and that usually means either global
3928 * reclaim or migration going on. The page will either get back to the
3930 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3931 * (!PageCgroupUsed) or moved to a different group. The page will
3932 * disappear in the next attempt.
3934 static int mem_cgroup_move_parent(struct page *page,
3935 struct page_cgroup *pc,
3936 struct mem_cgroup *child)
3938 struct mem_cgroup *parent;
3939 unsigned int nr_pages;
3940 unsigned long uninitialized_var(flags);
3943 VM_BUG_ON(mem_cgroup_is_root(child));
3946 if (!get_page_unless_zero(page))
3948 if (isolate_lru_page(page))
3951 nr_pages = hpage_nr_pages(page);
3953 parent = parent_mem_cgroup(child);
3955 * If no parent, move charges to root cgroup.
3958 parent = root_mem_cgroup;
3961 VM_BUG_ON(!PageTransHuge(page));
3962 flags = compound_lock_irqsave(page);
3965 ret = mem_cgroup_move_account(page, nr_pages,
3968 __mem_cgroup_cancel_local_charge(child, nr_pages);
3971 compound_unlock_irqrestore(page, flags);
3972 putback_lru_page(page);
3980 * Charge the memory controller for page usage.
3982 * 0 if the charge was successful
3983 * < 0 if the cgroup is over its limit
3985 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3986 gfp_t gfp_mask, enum charge_type ctype)
3988 struct mem_cgroup *memcg = NULL;
3989 unsigned int nr_pages = 1;
3993 if (PageTransHuge(page)) {
3994 nr_pages <<= compound_order(page);
3995 VM_BUG_ON(!PageTransHuge(page));
3997 * Never OOM-kill a process for a huge page. The
3998 * fault handler will fall back to regular pages.
4003 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
4006 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
4010 int mem_cgroup_newpage_charge(struct page *page,
4011 struct mm_struct *mm, gfp_t gfp_mask)
4013 if (mem_cgroup_disabled())
4015 VM_BUG_ON(page_mapped(page));
4016 VM_BUG_ON(page->mapping && !PageAnon(page));
4018 return mem_cgroup_charge_common(page, mm, gfp_mask,
4019 MEM_CGROUP_CHARGE_TYPE_ANON);
4023 * While swap-in, try_charge -> commit or cancel, the page is locked.
4024 * And when try_charge() successfully returns, one refcnt to memcg without
4025 * struct page_cgroup is acquired. This refcnt will be consumed by
4026 * "commit()" or removed by "cancel()"
4028 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
4031 struct mem_cgroup **memcgp)
4033 struct mem_cgroup *memcg;
4034 struct page_cgroup *pc;
4037 pc = lookup_page_cgroup(page);
4039 * Every swap fault against a single page tries to charge the
4040 * page, bail as early as possible. shmem_unuse() encounters
4041 * already charged pages, too. The USED bit is protected by
4042 * the page lock, which serializes swap cache removal, which
4043 * in turn serializes uncharging.
4045 if (PageCgroupUsed(pc))
4047 if (!do_swap_account)
4049 memcg = try_get_mem_cgroup_from_page(page);
4053 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4054 css_put(&memcg->css);
4059 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4065 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4066 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4069 if (mem_cgroup_disabled())
4072 * A racing thread's fault, or swapoff, may have already
4073 * updated the pte, and even removed page from swap cache: in
4074 * those cases unuse_pte()'s pte_same() test will fail; but
4075 * there's also a KSM case which does need to charge the page.
4077 if (!PageSwapCache(page)) {
4080 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4085 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4088 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4090 if (mem_cgroup_disabled())
4094 __mem_cgroup_cancel_charge(memcg, 1);
4098 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4099 enum charge_type ctype)
4101 if (mem_cgroup_disabled())
4106 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4108 * Now swap is on-memory. This means this page may be
4109 * counted both as mem and swap....double count.
4110 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4111 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4112 * may call delete_from_swap_cache() before reach here.
4114 if (do_swap_account && PageSwapCache(page)) {
4115 swp_entry_t ent = {.val = page_private(page)};
4116 mem_cgroup_uncharge_swap(ent);
4120 void mem_cgroup_commit_charge_swapin(struct page *page,
4121 struct mem_cgroup *memcg)
4123 __mem_cgroup_commit_charge_swapin(page, memcg,
4124 MEM_CGROUP_CHARGE_TYPE_ANON);
4127 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4130 struct mem_cgroup *memcg = NULL;
4131 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4134 if (mem_cgroup_disabled())
4136 if (PageCompound(page))
4139 if (!PageSwapCache(page))
4140 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4141 else { /* page is swapcache/shmem */
4142 ret = __mem_cgroup_try_charge_swapin(mm, page,
4145 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4150 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4151 unsigned int nr_pages,
4152 const enum charge_type ctype)
4154 struct memcg_batch_info *batch = NULL;
4155 bool uncharge_memsw = true;
4157 /* If swapout, usage of swap doesn't decrease */
4158 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4159 uncharge_memsw = false;
4161 batch = ¤t->memcg_batch;
4163 * In usual, we do css_get() when we remember memcg pointer.
4164 * But in this case, we keep res->usage until end of a series of
4165 * uncharges. Then, it's ok to ignore memcg's refcnt.
4168 batch->memcg = memcg;
4170 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4171 * In those cases, all pages freed continuously can be expected to be in
4172 * the same cgroup and we have chance to coalesce uncharges.
4173 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4174 * because we want to do uncharge as soon as possible.
4177 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4178 goto direct_uncharge;
4181 goto direct_uncharge;
4184 * In typical case, batch->memcg == mem. This means we can
4185 * merge a series of uncharges to an uncharge of res_counter.
4186 * If not, we uncharge res_counter ony by one.
4188 if (batch->memcg != memcg)
4189 goto direct_uncharge;
4190 /* remember freed charge and uncharge it later */
4193 batch->memsw_nr_pages++;
4196 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4198 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4199 if (unlikely(batch->memcg != memcg))
4200 memcg_oom_recover(memcg);
4204 * uncharge if !page_mapped(page)
4206 static struct mem_cgroup *
4207 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4210 struct mem_cgroup *memcg = NULL;
4211 unsigned int nr_pages = 1;
4212 struct page_cgroup *pc;
4215 if (mem_cgroup_disabled())
4218 if (PageTransHuge(page)) {
4219 nr_pages <<= compound_order(page);
4220 VM_BUG_ON(!PageTransHuge(page));
4223 * Check if our page_cgroup is valid
4225 pc = lookup_page_cgroup(page);
4226 if (unlikely(!PageCgroupUsed(pc)))
4229 lock_page_cgroup(pc);
4231 memcg = pc->mem_cgroup;
4233 if (!PageCgroupUsed(pc))
4236 anon = PageAnon(page);
4239 case MEM_CGROUP_CHARGE_TYPE_ANON:
4241 * Generally PageAnon tells if it's the anon statistics to be
4242 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4243 * used before page reached the stage of being marked PageAnon.
4247 case MEM_CGROUP_CHARGE_TYPE_DROP:
4248 /* See mem_cgroup_prepare_migration() */
4249 if (page_mapped(page))
4252 * Pages under migration may not be uncharged. But
4253 * end_migration() /must/ be the one uncharging the
4254 * unused post-migration page and so it has to call
4255 * here with the migration bit still set. See the
4256 * res_counter handling below.
4258 if (!end_migration && PageCgroupMigration(pc))
4261 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4262 if (!PageAnon(page)) { /* Shared memory */
4263 if (page->mapping && !page_is_file_cache(page))
4265 } else if (page_mapped(page)) /* Anon */
4272 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4274 ClearPageCgroupUsed(pc);
4276 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4277 * freed from LRU. This is safe because uncharged page is expected not
4278 * to be reused (freed soon). Exception is SwapCache, it's handled by
4279 * special functions.
4282 unlock_page_cgroup(pc);
4284 * even after unlock, we have memcg->res.usage here and this memcg
4285 * will never be freed, so it's safe to call css_get().
4287 memcg_check_events(memcg, page);
4288 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4289 mem_cgroup_swap_statistics(memcg, true);
4290 css_get(&memcg->css);
4293 * Migration does not charge the res_counter for the
4294 * replacement page, so leave it alone when phasing out the
4295 * page that is unused after the migration.
4297 if (!end_migration && !mem_cgroup_is_root(memcg))
4298 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4303 unlock_page_cgroup(pc);
4307 void mem_cgroup_uncharge_page(struct page *page)
4310 if (page_mapped(page))
4312 VM_BUG_ON(page->mapping && !PageAnon(page));
4314 * If the page is in swap cache, uncharge should be deferred
4315 * to the swap path, which also properly accounts swap usage
4316 * and handles memcg lifetime.
4318 * Note that this check is not stable and reclaim may add the
4319 * page to swap cache at any time after this. However, if the
4320 * page is not in swap cache by the time page->mapcount hits
4321 * 0, there won't be any page table references to the swap
4322 * slot, and reclaim will free it and not actually write the
4325 if (PageSwapCache(page))
4327 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4330 void mem_cgroup_uncharge_cache_page(struct page *page)
4332 VM_BUG_ON(page_mapped(page));
4333 VM_BUG_ON(page->mapping);
4334 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4338 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4339 * In that cases, pages are freed continuously and we can expect pages
4340 * are in the same memcg. All these calls itself limits the number of
4341 * pages freed at once, then uncharge_start/end() is called properly.
4342 * This may be called prural(2) times in a context,
4345 void mem_cgroup_uncharge_start(void)
4347 current->memcg_batch.do_batch++;
4348 /* We can do nest. */
4349 if (current->memcg_batch.do_batch == 1) {
4350 current->memcg_batch.memcg = NULL;
4351 current->memcg_batch.nr_pages = 0;
4352 current->memcg_batch.memsw_nr_pages = 0;
4356 void mem_cgroup_uncharge_end(void)
4358 struct memcg_batch_info *batch = ¤t->memcg_batch;
4360 if (!batch->do_batch)
4364 if (batch->do_batch) /* If stacked, do nothing. */
4370 * This "batch->memcg" is valid without any css_get/put etc...
4371 * bacause we hide charges behind us.
4373 if (batch->nr_pages)
4374 res_counter_uncharge(&batch->memcg->res,
4375 batch->nr_pages * PAGE_SIZE);
4376 if (batch->memsw_nr_pages)
4377 res_counter_uncharge(&batch->memcg->memsw,
4378 batch->memsw_nr_pages * PAGE_SIZE);
4379 memcg_oom_recover(batch->memcg);
4380 /* forget this pointer (for sanity check) */
4381 batch->memcg = NULL;
4386 * called after __delete_from_swap_cache() and drop "page" account.
4387 * memcg information is recorded to swap_cgroup of "ent"
4390 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4392 struct mem_cgroup *memcg;
4393 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4395 if (!swapout) /* this was a swap cache but the swap is unused ! */
4396 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4398 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4401 * record memcg information, if swapout && memcg != NULL,
4402 * css_get() was called in uncharge().
4404 if (do_swap_account && swapout && memcg)
4405 swap_cgroup_record(ent, css_id(&memcg->css));
4409 #ifdef CONFIG_MEMCG_SWAP
4411 * called from swap_entry_free(). remove record in swap_cgroup and
4412 * uncharge "memsw" account.
4414 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4416 struct mem_cgroup *memcg;
4419 if (!do_swap_account)
4422 id = swap_cgroup_record(ent, 0);
4424 memcg = mem_cgroup_lookup(id);
4427 * We uncharge this because swap is freed.
4428 * This memcg can be obsolete one. We avoid calling css_tryget
4430 if (!mem_cgroup_is_root(memcg))
4431 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4432 mem_cgroup_swap_statistics(memcg, false);
4433 css_put(&memcg->css);
4439 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4440 * @entry: swap entry to be moved
4441 * @from: mem_cgroup which the entry is moved from
4442 * @to: mem_cgroup which the entry is moved to
4444 * It succeeds only when the swap_cgroup's record for this entry is the same
4445 * as the mem_cgroup's id of @from.
4447 * Returns 0 on success, -EINVAL on failure.
4449 * The caller must have charged to @to, IOW, called res_counter_charge() about
4450 * both res and memsw, and called css_get().
4452 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4453 struct mem_cgroup *from, struct mem_cgroup *to)
4455 unsigned short old_id, new_id;
4457 old_id = css_id(&from->css);
4458 new_id = css_id(&to->css);
4460 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4461 mem_cgroup_swap_statistics(from, false);
4462 mem_cgroup_swap_statistics(to, true);
4464 * This function is only called from task migration context now.
4465 * It postpones res_counter and refcount handling till the end
4466 * of task migration(mem_cgroup_clear_mc()) for performance
4467 * improvement. But we cannot postpone css_get(to) because if
4468 * the process that has been moved to @to does swap-in, the
4469 * refcount of @to might be decreased to 0.
4471 * We are in attach() phase, so the cgroup is guaranteed to be
4472 * alive, so we can just call css_get().
4480 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4481 struct mem_cgroup *from, struct mem_cgroup *to)
4488 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4491 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4492 struct mem_cgroup **memcgp)
4494 struct mem_cgroup *memcg = NULL;
4495 unsigned int nr_pages = 1;
4496 struct page_cgroup *pc;
4497 enum charge_type ctype;
4501 if (mem_cgroup_disabled())
4504 if (PageTransHuge(page))
4505 nr_pages <<= compound_order(page);
4507 pc = lookup_page_cgroup(page);
4508 lock_page_cgroup(pc);
4509 if (PageCgroupUsed(pc)) {
4510 memcg = pc->mem_cgroup;
4511 css_get(&memcg->css);
4513 * At migrating an anonymous page, its mapcount goes down
4514 * to 0 and uncharge() will be called. But, even if it's fully
4515 * unmapped, migration may fail and this page has to be
4516 * charged again. We set MIGRATION flag here and delay uncharge
4517 * until end_migration() is called
4519 * Corner Case Thinking
4521 * When the old page was mapped as Anon and it's unmap-and-freed
4522 * while migration was ongoing.
4523 * If unmap finds the old page, uncharge() of it will be delayed
4524 * until end_migration(). If unmap finds a new page, it's
4525 * uncharged when it make mapcount to be 1->0. If unmap code
4526 * finds swap_migration_entry, the new page will not be mapped
4527 * and end_migration() will find it(mapcount==0).
4530 * When the old page was mapped but migraion fails, the kernel
4531 * remaps it. A charge for it is kept by MIGRATION flag even
4532 * if mapcount goes down to 0. We can do remap successfully
4533 * without charging it again.
4536 * The "old" page is under lock_page() until the end of
4537 * migration, so, the old page itself will not be swapped-out.
4538 * If the new page is swapped out before end_migraton, our
4539 * hook to usual swap-out path will catch the event.
4542 SetPageCgroupMigration(pc);
4544 unlock_page_cgroup(pc);
4546 * If the page is not charged at this point,
4554 * We charge new page before it's used/mapped. So, even if unlock_page()
4555 * is called before end_migration, we can catch all events on this new
4556 * page. In the case new page is migrated but not remapped, new page's
4557 * mapcount will be finally 0 and we call uncharge in end_migration().
4560 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4562 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4564 * The page is committed to the memcg, but it's not actually
4565 * charged to the res_counter since we plan on replacing the
4566 * old one and only one page is going to be left afterwards.
4568 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4571 /* remove redundant charge if migration failed*/
4572 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4573 struct page *oldpage, struct page *newpage, bool migration_ok)
4575 struct page *used, *unused;
4576 struct page_cgroup *pc;
4582 if (!migration_ok) {
4589 anon = PageAnon(used);
4590 __mem_cgroup_uncharge_common(unused,
4591 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4592 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4594 css_put(&memcg->css);
4596 * We disallowed uncharge of pages under migration because mapcount
4597 * of the page goes down to zero, temporarly.
4598 * Clear the flag and check the page should be charged.
4600 pc = lookup_page_cgroup(oldpage);
4601 lock_page_cgroup(pc);
4602 ClearPageCgroupMigration(pc);
4603 unlock_page_cgroup(pc);
4606 * If a page is a file cache, radix-tree replacement is very atomic
4607 * and we can skip this check. When it was an Anon page, its mapcount
4608 * goes down to 0. But because we added MIGRATION flage, it's not
4609 * uncharged yet. There are several case but page->mapcount check
4610 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4611 * check. (see prepare_charge() also)
4614 mem_cgroup_uncharge_page(used);
4618 * At replace page cache, newpage is not under any memcg but it's on
4619 * LRU. So, this function doesn't touch res_counter but handles LRU
4620 * in correct way. Both pages are locked so we cannot race with uncharge.
4622 void mem_cgroup_replace_page_cache(struct page *oldpage,
4623 struct page *newpage)
4625 struct mem_cgroup *memcg = NULL;
4626 struct page_cgroup *pc;
4627 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4629 if (mem_cgroup_disabled())
4632 pc = lookup_page_cgroup(oldpage);
4633 /* fix accounting on old pages */
4634 lock_page_cgroup(pc);
4635 if (PageCgroupUsed(pc)) {
4636 memcg = pc->mem_cgroup;
4637 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4638 ClearPageCgroupUsed(pc);
4640 unlock_page_cgroup(pc);
4643 * When called from shmem_replace_page(), in some cases the
4644 * oldpage has already been charged, and in some cases not.
4649 * Even if newpage->mapping was NULL before starting replacement,
4650 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4651 * LRU while we overwrite pc->mem_cgroup.
4653 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4656 #ifdef CONFIG_DEBUG_VM
4657 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4659 struct page_cgroup *pc;
4661 pc = lookup_page_cgroup(page);
4663 * Can be NULL while feeding pages into the page allocator for
4664 * the first time, i.e. during boot or memory hotplug;
4665 * or when mem_cgroup_disabled().
4667 if (likely(pc) && PageCgroupUsed(pc))
4672 bool mem_cgroup_bad_page_check(struct page *page)
4674 if (mem_cgroup_disabled())
4677 return lookup_page_cgroup_used(page) != NULL;
4680 void mem_cgroup_print_bad_page(struct page *page)
4682 struct page_cgroup *pc;
4684 pc = lookup_page_cgroup_used(page);
4686 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4687 pc, pc->flags, pc->mem_cgroup);
4692 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4693 unsigned long long val)
4696 u64 memswlimit, memlimit;
4698 int children = mem_cgroup_count_children(memcg);
4699 u64 curusage, oldusage;
4703 * For keeping hierarchical_reclaim simple, how long we should retry
4704 * is depends on callers. We set our retry-count to be function
4705 * of # of children which we should visit in this loop.
4707 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4709 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4712 while (retry_count) {
4713 if (signal_pending(current)) {
4718 * Rather than hide all in some function, I do this in
4719 * open coded manner. You see what this really does.
4720 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4722 mutex_lock(&set_limit_mutex);
4723 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4724 if (memswlimit < val) {
4726 mutex_unlock(&set_limit_mutex);
4730 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4734 ret = res_counter_set_limit(&memcg->res, val);
4736 if (memswlimit == val)
4737 memcg->memsw_is_minimum = true;
4739 memcg->memsw_is_minimum = false;
4741 mutex_unlock(&set_limit_mutex);
4746 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4747 MEM_CGROUP_RECLAIM_SHRINK);
4748 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4749 /* Usage is reduced ? */
4750 if (curusage >= oldusage)
4753 oldusage = curusage;
4755 if (!ret && enlarge)
4756 memcg_oom_recover(memcg);
4761 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4762 unsigned long long val)
4765 u64 memlimit, memswlimit, oldusage, curusage;
4766 int children = mem_cgroup_count_children(memcg);
4770 /* see mem_cgroup_resize_res_limit */
4771 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4772 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4773 while (retry_count) {
4774 if (signal_pending(current)) {
4779 * Rather than hide all in some function, I do this in
4780 * open coded manner. You see what this really does.
4781 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4783 mutex_lock(&set_limit_mutex);
4784 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4785 if (memlimit > val) {
4787 mutex_unlock(&set_limit_mutex);
4790 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4791 if (memswlimit < val)
4793 ret = res_counter_set_limit(&memcg->memsw, val);
4795 if (memlimit == val)
4796 memcg->memsw_is_minimum = true;
4798 memcg->memsw_is_minimum = false;
4800 mutex_unlock(&set_limit_mutex);
4805 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4806 MEM_CGROUP_RECLAIM_NOSWAP |
4807 MEM_CGROUP_RECLAIM_SHRINK);
4808 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4809 /* Usage is reduced ? */
4810 if (curusage >= oldusage)
4813 oldusage = curusage;
4815 if (!ret && enlarge)
4816 memcg_oom_recover(memcg);
4820 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4822 unsigned long *total_scanned)
4824 unsigned long nr_reclaimed = 0;
4825 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4826 unsigned long reclaimed;
4828 struct mem_cgroup_tree_per_zone *mctz;
4829 unsigned long long excess;
4830 unsigned long nr_scanned;
4835 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4837 * This loop can run a while, specially if mem_cgroup's continuously
4838 * keep exceeding their soft limit and putting the system under
4845 mz = mem_cgroup_largest_soft_limit_node(mctz);
4850 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4851 gfp_mask, &nr_scanned);
4852 nr_reclaimed += reclaimed;
4853 *total_scanned += nr_scanned;
4854 spin_lock(&mctz->lock);
4857 * If we failed to reclaim anything from this memory cgroup
4858 * it is time to move on to the next cgroup
4864 * Loop until we find yet another one.
4866 * By the time we get the soft_limit lock
4867 * again, someone might have aded the
4868 * group back on the RB tree. Iterate to
4869 * make sure we get a different mem.
4870 * mem_cgroup_largest_soft_limit_node returns
4871 * NULL if no other cgroup is present on
4875 __mem_cgroup_largest_soft_limit_node(mctz);
4877 css_put(&next_mz->memcg->css);
4878 else /* next_mz == NULL or other memcg */
4882 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4883 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4885 * One school of thought says that we should not add
4886 * back the node to the tree if reclaim returns 0.
4887 * But our reclaim could return 0, simply because due
4888 * to priority we are exposing a smaller subset of
4889 * memory to reclaim from. Consider this as a longer
4892 /* If excess == 0, no tree ops */
4893 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4894 spin_unlock(&mctz->lock);
4895 css_put(&mz->memcg->css);
4898 * Could not reclaim anything and there are no more
4899 * mem cgroups to try or we seem to be looping without
4900 * reclaiming anything.
4902 if (!nr_reclaimed &&
4904 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4906 } while (!nr_reclaimed);
4908 css_put(&next_mz->memcg->css);
4909 return nr_reclaimed;
4913 * mem_cgroup_force_empty_list - clears LRU of a group
4914 * @memcg: group to clear
4917 * @lru: lru to to clear
4919 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4920 * reclaim the pages page themselves - pages are moved to the parent (or root)
4923 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4924 int node, int zid, enum lru_list lru)
4926 struct lruvec *lruvec;
4927 unsigned long flags;
4928 struct list_head *list;
4932 zone = &NODE_DATA(node)->node_zones[zid];
4933 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4934 list = &lruvec->lists[lru];
4938 struct page_cgroup *pc;
4941 spin_lock_irqsave(&zone->lru_lock, flags);
4942 if (list_empty(list)) {
4943 spin_unlock_irqrestore(&zone->lru_lock, flags);
4946 page = list_entry(list->prev, struct page, lru);
4948 list_move(&page->lru, list);
4950 spin_unlock_irqrestore(&zone->lru_lock, flags);
4953 spin_unlock_irqrestore(&zone->lru_lock, flags);
4955 pc = lookup_page_cgroup(page);
4957 if (mem_cgroup_move_parent(page, pc, memcg)) {
4958 /* found lock contention or "pc" is obsolete. */
4963 } while (!list_empty(list));
4967 * make mem_cgroup's charge to be 0 if there is no task by moving
4968 * all the charges and pages to the parent.
4969 * This enables deleting this mem_cgroup.
4971 * Caller is responsible for holding css reference on the memcg.
4973 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4979 /* This is for making all *used* pages to be on LRU. */
4980 lru_add_drain_all();
4981 drain_all_stock_sync(memcg);
4982 mem_cgroup_start_move(memcg);
4983 for_each_node_state(node, N_MEMORY) {
4984 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4987 mem_cgroup_force_empty_list(memcg,
4992 mem_cgroup_end_move(memcg);
4993 memcg_oom_recover(memcg);
4997 * Kernel memory may not necessarily be trackable to a specific
4998 * process. So they are not migrated, and therefore we can't
4999 * expect their value to drop to 0 here.
5000 * Having res filled up with kmem only is enough.
5002 * This is a safety check because mem_cgroup_force_empty_list
5003 * could have raced with mem_cgroup_replace_page_cache callers
5004 * so the lru seemed empty but the page could have been added
5005 * right after the check. RES_USAGE should be safe as we always
5006 * charge before adding to the LRU.
5008 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
5009 res_counter_read_u64(&memcg->kmem, RES_USAGE);
5010 } while (usage > 0);
5013 static inline bool memcg_has_children(struct mem_cgroup *memcg)
5015 lockdep_assert_held(&memcg_create_mutex);
5017 * The lock does not prevent addition or deletion to the list
5018 * of children, but it prevents a new child from being
5019 * initialized based on this parent in css_online(), so it's
5020 * enough to decide whether hierarchically inherited
5021 * attributes can still be changed or not.
5023 return memcg->use_hierarchy &&
5024 !list_empty(&memcg->css.cgroup->children);
5028 * Reclaims as many pages from the given memcg as possible and moves
5029 * the rest to the parent.
5031 * Caller is responsible for holding css reference for memcg.
5033 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
5035 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
5036 struct cgroup *cgrp = memcg->css.cgroup;
5038 /* returns EBUSY if there is a task or if we come here twice. */
5039 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
5042 /* we call try-to-free pages for make this cgroup empty */
5043 lru_add_drain_all();
5044 /* try to free all pages in this cgroup */
5045 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
5048 if (signal_pending(current))
5051 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5055 /* maybe some writeback is necessary */
5056 congestion_wait(BLK_RW_ASYNC, HZ/10);
5061 mem_cgroup_reparent_charges(memcg);
5066 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
5069 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5071 if (mem_cgroup_is_root(memcg))
5073 return mem_cgroup_force_empty(memcg);
5076 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5079 return mem_cgroup_from_css(css)->use_hierarchy;
5082 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5083 struct cftype *cft, u64 val)
5086 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5087 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5089 mutex_lock(&memcg_create_mutex);
5091 if (memcg->use_hierarchy == val)
5095 * If parent's use_hierarchy is set, we can't make any modifications
5096 * in the child subtrees. If it is unset, then the change can
5097 * occur, provided the current cgroup has no children.
5099 * For the root cgroup, parent_mem is NULL, we allow value to be
5100 * set if there are no children.
5102 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5103 (val == 1 || val == 0)) {
5104 if (list_empty(&memcg->css.cgroup->children))
5105 memcg->use_hierarchy = val;
5112 mutex_unlock(&memcg_create_mutex);
5118 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5119 enum mem_cgroup_stat_index idx)
5121 struct mem_cgroup *iter;
5124 /* Per-cpu values can be negative, use a signed accumulator */
5125 for_each_mem_cgroup_tree(iter, memcg)
5126 val += mem_cgroup_read_stat(iter, idx);
5128 if (val < 0) /* race ? */
5133 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5137 if (!mem_cgroup_is_root(memcg)) {
5139 return res_counter_read_u64(&memcg->res, RES_USAGE);
5141 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5145 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5146 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5148 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5149 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5152 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5154 return val << PAGE_SHIFT;
5157 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
5158 struct cftype *cft, struct file *file,
5159 char __user *buf, size_t nbytes, loff_t *ppos)
5161 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5167 type = MEMFILE_TYPE(cft->private);
5168 name = MEMFILE_ATTR(cft->private);
5172 if (name == RES_USAGE)
5173 val = mem_cgroup_usage(memcg, false);
5175 val = res_counter_read_u64(&memcg->res, name);
5178 if (name == RES_USAGE)
5179 val = mem_cgroup_usage(memcg, true);
5181 val = res_counter_read_u64(&memcg->memsw, name);
5184 val = res_counter_read_u64(&memcg->kmem, name);
5190 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5191 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5194 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
5197 #ifdef CONFIG_MEMCG_KMEM
5198 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5200 * For simplicity, we won't allow this to be disabled. It also can't
5201 * be changed if the cgroup has children already, or if tasks had
5204 * If tasks join before we set the limit, a person looking at
5205 * kmem.usage_in_bytes will have no way to determine when it took
5206 * place, which makes the value quite meaningless.
5208 * After it first became limited, changes in the value of the limit are
5209 * of course permitted.
5211 mutex_lock(&memcg_create_mutex);
5212 mutex_lock(&set_limit_mutex);
5213 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
5214 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
5218 ret = res_counter_set_limit(&memcg->kmem, val);
5221 ret = memcg_update_cache_sizes(memcg);
5223 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
5226 static_key_slow_inc(&memcg_kmem_enabled_key);
5228 * setting the active bit after the inc will guarantee no one
5229 * starts accounting before all call sites are patched
5231 memcg_kmem_set_active(memcg);
5233 ret = res_counter_set_limit(&memcg->kmem, val);
5235 mutex_unlock(&set_limit_mutex);
5236 mutex_unlock(&memcg_create_mutex);
5241 #ifdef CONFIG_MEMCG_KMEM
5242 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5245 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5249 memcg->kmem_account_flags = parent->kmem_account_flags;
5251 * When that happen, we need to disable the static branch only on those
5252 * memcgs that enabled it. To achieve this, we would be forced to
5253 * complicate the code by keeping track of which memcgs were the ones
5254 * that actually enabled limits, and which ones got it from its
5257 * It is a lot simpler just to do static_key_slow_inc() on every child
5258 * that is accounted.
5260 if (!memcg_kmem_is_active(memcg))
5264 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5265 * memcg is active already. If the later initialization fails then the
5266 * cgroup core triggers the cleanup so we do not have to do it here.
5268 static_key_slow_inc(&memcg_kmem_enabled_key);
5270 mutex_lock(&set_limit_mutex);
5271 memcg_stop_kmem_account();
5272 ret = memcg_update_cache_sizes(memcg);
5273 memcg_resume_kmem_account();
5274 mutex_unlock(&set_limit_mutex);
5278 #endif /* CONFIG_MEMCG_KMEM */
5281 * The user of this function is...
5284 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5287 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5290 unsigned long long val;
5293 type = MEMFILE_TYPE(cft->private);
5294 name = MEMFILE_ATTR(cft->private);
5298 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5302 /* This function does all necessary parse...reuse it */
5303 ret = res_counter_memparse_write_strategy(buffer, &val);
5307 ret = mem_cgroup_resize_limit(memcg, val);
5308 else if (type == _MEMSWAP)
5309 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5310 else if (type == _KMEM)
5311 ret = memcg_update_kmem_limit(css, val);
5315 case RES_SOFT_LIMIT:
5316 ret = res_counter_memparse_write_strategy(buffer, &val);
5320 * For memsw, soft limits are hard to implement in terms
5321 * of semantics, for now, we support soft limits for
5322 * control without swap
5325 ret = res_counter_set_soft_limit(&memcg->res, val);
5330 ret = -EINVAL; /* should be BUG() ? */
5336 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5337 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5339 unsigned long long min_limit, min_memsw_limit, tmp;
5341 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5342 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5343 if (!memcg->use_hierarchy)
5346 while (css_parent(&memcg->css)) {
5347 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5348 if (!memcg->use_hierarchy)
5350 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5351 min_limit = min(min_limit, tmp);
5352 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5353 min_memsw_limit = min(min_memsw_limit, tmp);
5356 *mem_limit = min_limit;
5357 *memsw_limit = min_memsw_limit;
5360 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5362 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5366 type = MEMFILE_TYPE(event);
5367 name = MEMFILE_ATTR(event);
5372 res_counter_reset_max(&memcg->res);
5373 else if (type == _MEMSWAP)
5374 res_counter_reset_max(&memcg->memsw);
5375 else if (type == _KMEM)
5376 res_counter_reset_max(&memcg->kmem);
5382 res_counter_reset_failcnt(&memcg->res);
5383 else if (type == _MEMSWAP)
5384 res_counter_reset_failcnt(&memcg->memsw);
5385 else if (type == _KMEM)
5386 res_counter_reset_failcnt(&memcg->kmem);
5395 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5398 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5402 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5403 struct cftype *cft, u64 val)
5405 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5407 if (val >= (1 << NR_MOVE_TYPE))
5411 * No kind of locking is needed in here, because ->can_attach() will
5412 * check this value once in the beginning of the process, and then carry
5413 * on with stale data. This means that changes to this value will only
5414 * affect task migrations starting after the change.
5416 memcg->move_charge_at_immigrate = val;
5420 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5421 struct cftype *cft, u64 val)
5428 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5429 struct cftype *cft, struct seq_file *m)
5432 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5433 unsigned long node_nr;
5434 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5436 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5437 seq_printf(m, "total=%lu", total_nr);
5438 for_each_node_state(nid, N_MEMORY) {
5439 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5440 seq_printf(m, " N%d=%lu", nid, node_nr);
5444 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5445 seq_printf(m, "file=%lu", file_nr);
5446 for_each_node_state(nid, N_MEMORY) {
5447 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5449 seq_printf(m, " N%d=%lu", nid, node_nr);
5453 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5454 seq_printf(m, "anon=%lu", anon_nr);
5455 for_each_node_state(nid, N_MEMORY) {
5456 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5458 seq_printf(m, " N%d=%lu", nid, node_nr);
5462 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5463 seq_printf(m, "unevictable=%lu", unevictable_nr);
5464 for_each_node_state(nid, N_MEMORY) {
5465 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5466 BIT(LRU_UNEVICTABLE));
5467 seq_printf(m, " N%d=%lu", nid, node_nr);
5472 #endif /* CONFIG_NUMA */
5474 static inline void mem_cgroup_lru_names_not_uptodate(void)
5476 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5479 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5482 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5483 struct mem_cgroup *mi;
5486 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5487 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5489 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5490 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5493 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5494 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5495 mem_cgroup_read_events(memcg, i));
5497 for (i = 0; i < NR_LRU_LISTS; i++)
5498 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5499 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5501 /* Hierarchical information */
5503 unsigned long long limit, memsw_limit;
5504 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5505 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5506 if (do_swap_account)
5507 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5511 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5514 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5516 for_each_mem_cgroup_tree(mi, memcg)
5517 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5518 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5521 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5522 unsigned long long val = 0;
5524 for_each_mem_cgroup_tree(mi, memcg)
5525 val += mem_cgroup_read_events(mi, i);
5526 seq_printf(m, "total_%s %llu\n",
5527 mem_cgroup_events_names[i], val);
5530 for (i = 0; i < NR_LRU_LISTS; i++) {
5531 unsigned long long val = 0;
5533 for_each_mem_cgroup_tree(mi, memcg)
5534 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5535 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5538 #ifdef CONFIG_DEBUG_VM
5541 struct mem_cgroup_per_zone *mz;
5542 struct zone_reclaim_stat *rstat;
5543 unsigned long recent_rotated[2] = {0, 0};
5544 unsigned long recent_scanned[2] = {0, 0};
5546 for_each_online_node(nid)
5547 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5548 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5549 rstat = &mz->lruvec.reclaim_stat;
5551 recent_rotated[0] += rstat->recent_rotated[0];
5552 recent_rotated[1] += rstat->recent_rotated[1];
5553 recent_scanned[0] += rstat->recent_scanned[0];
5554 recent_scanned[1] += rstat->recent_scanned[1];
5556 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5557 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5558 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5559 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5566 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5569 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5571 return mem_cgroup_swappiness(memcg);
5574 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5575 struct cftype *cft, u64 val)
5577 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5578 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5580 if (val > 100 || !parent)
5583 mutex_lock(&memcg_create_mutex);
5585 /* If under hierarchy, only empty-root can set this value */
5586 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5587 mutex_unlock(&memcg_create_mutex);
5591 memcg->swappiness = val;
5593 mutex_unlock(&memcg_create_mutex);
5598 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5600 struct mem_cgroup_threshold_ary *t;
5606 t = rcu_dereference(memcg->thresholds.primary);
5608 t = rcu_dereference(memcg->memsw_thresholds.primary);
5613 usage = mem_cgroup_usage(memcg, swap);
5616 * current_threshold points to threshold just below or equal to usage.
5617 * If it's not true, a threshold was crossed after last
5618 * call of __mem_cgroup_threshold().
5620 i = t->current_threshold;
5623 * Iterate backward over array of thresholds starting from
5624 * current_threshold and check if a threshold is crossed.
5625 * If none of thresholds below usage is crossed, we read
5626 * only one element of the array here.
5628 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5629 eventfd_signal(t->entries[i].eventfd, 1);
5631 /* i = current_threshold + 1 */
5635 * Iterate forward over array of thresholds starting from
5636 * current_threshold+1 and check if a threshold is crossed.
5637 * If none of thresholds above usage is crossed, we read
5638 * only one element of the array here.
5640 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5641 eventfd_signal(t->entries[i].eventfd, 1);
5643 /* Update current_threshold */
5644 t->current_threshold = i - 1;
5649 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5652 __mem_cgroup_threshold(memcg, false);
5653 if (do_swap_account)
5654 __mem_cgroup_threshold(memcg, true);
5656 memcg = parent_mem_cgroup(memcg);
5660 static int compare_thresholds(const void *a, const void *b)
5662 const struct mem_cgroup_threshold *_a = a;
5663 const struct mem_cgroup_threshold *_b = b;
5665 if (_a->threshold > _b->threshold)
5668 if (_a->threshold < _b->threshold)
5674 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5676 struct mem_cgroup_eventfd_list *ev;
5678 list_for_each_entry(ev, &memcg->oom_notify, list)
5679 eventfd_signal(ev->eventfd, 1);
5683 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5685 struct mem_cgroup *iter;
5687 for_each_mem_cgroup_tree(iter, memcg)
5688 mem_cgroup_oom_notify_cb(iter);
5691 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5692 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5694 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5695 struct mem_cgroup_thresholds *thresholds;
5696 struct mem_cgroup_threshold_ary *new;
5697 enum res_type type = MEMFILE_TYPE(cft->private);
5698 u64 threshold, usage;
5701 ret = res_counter_memparse_write_strategy(args, &threshold);
5705 mutex_lock(&memcg->thresholds_lock);
5708 thresholds = &memcg->thresholds;
5709 else if (type == _MEMSWAP)
5710 thresholds = &memcg->memsw_thresholds;
5714 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5716 /* Check if a threshold crossed before adding a new one */
5717 if (thresholds->primary)
5718 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5720 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5722 /* Allocate memory for new array of thresholds */
5723 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5731 /* Copy thresholds (if any) to new array */
5732 if (thresholds->primary) {
5733 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5734 sizeof(struct mem_cgroup_threshold));
5737 /* Add new threshold */
5738 new->entries[size - 1].eventfd = eventfd;
5739 new->entries[size - 1].threshold = threshold;
5741 /* Sort thresholds. Registering of new threshold isn't time-critical */
5742 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5743 compare_thresholds, NULL);
5745 /* Find current threshold */
5746 new->current_threshold = -1;
5747 for (i = 0; i < size; i++) {
5748 if (new->entries[i].threshold <= usage) {
5750 * new->current_threshold will not be used until
5751 * rcu_assign_pointer(), so it's safe to increment
5754 ++new->current_threshold;
5759 /* Free old spare buffer and save old primary buffer as spare */
5760 kfree(thresholds->spare);
5761 thresholds->spare = thresholds->primary;
5763 rcu_assign_pointer(thresholds->primary, new);
5765 /* To be sure that nobody uses thresholds */
5769 mutex_unlock(&memcg->thresholds_lock);
5774 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5775 struct cftype *cft, struct eventfd_ctx *eventfd)
5777 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5778 struct mem_cgroup_thresholds *thresholds;
5779 struct mem_cgroup_threshold_ary *new;
5780 enum res_type type = MEMFILE_TYPE(cft->private);
5784 mutex_lock(&memcg->thresholds_lock);
5786 thresholds = &memcg->thresholds;
5787 else if (type == _MEMSWAP)
5788 thresholds = &memcg->memsw_thresholds;
5792 if (!thresholds->primary)
5795 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5797 /* Check if a threshold crossed before removing */
5798 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5800 /* Calculate new number of threshold */
5802 for (i = 0; i < thresholds->primary->size; i++) {
5803 if (thresholds->primary->entries[i].eventfd != eventfd)
5807 new = thresholds->spare;
5809 /* Set thresholds array to NULL if we don't have thresholds */
5818 /* Copy thresholds and find current threshold */
5819 new->current_threshold = -1;
5820 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5821 if (thresholds->primary->entries[i].eventfd == eventfd)
5824 new->entries[j] = thresholds->primary->entries[i];
5825 if (new->entries[j].threshold <= usage) {
5827 * new->current_threshold will not be used
5828 * until rcu_assign_pointer(), so it's safe to increment
5831 ++new->current_threshold;
5837 /* Swap primary and spare array */
5838 thresholds->spare = thresholds->primary;
5839 /* If all events are unregistered, free the spare array */
5841 kfree(thresholds->spare);
5842 thresholds->spare = NULL;
5845 rcu_assign_pointer(thresholds->primary, new);
5847 /* To be sure that nobody uses thresholds */
5850 mutex_unlock(&memcg->thresholds_lock);
5853 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5854 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5856 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5857 struct mem_cgroup_eventfd_list *event;
5858 enum res_type type = MEMFILE_TYPE(cft->private);
5860 BUG_ON(type != _OOM_TYPE);
5861 event = kmalloc(sizeof(*event), GFP_KERNEL);
5865 spin_lock(&memcg_oom_lock);
5867 event->eventfd = eventfd;
5868 list_add(&event->list, &memcg->oom_notify);
5870 /* already in OOM ? */
5871 if (atomic_read(&memcg->under_oom))
5872 eventfd_signal(eventfd, 1);
5873 spin_unlock(&memcg_oom_lock);
5878 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5879 struct cftype *cft, struct eventfd_ctx *eventfd)
5881 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5882 struct mem_cgroup_eventfd_list *ev, *tmp;
5883 enum res_type type = MEMFILE_TYPE(cft->private);
5885 BUG_ON(type != _OOM_TYPE);
5887 spin_lock(&memcg_oom_lock);
5889 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5890 if (ev->eventfd == eventfd) {
5891 list_del(&ev->list);
5896 spin_unlock(&memcg_oom_lock);
5899 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5900 struct cftype *cft, struct cgroup_map_cb *cb)
5902 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5904 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5906 if (atomic_read(&memcg->under_oom))
5907 cb->fill(cb, "under_oom", 1);
5909 cb->fill(cb, "under_oom", 0);
5913 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5914 struct cftype *cft, u64 val)
5916 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5917 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5919 /* cannot set to root cgroup and only 0 and 1 are allowed */
5920 if (!parent || !((val == 0) || (val == 1)))
5923 mutex_lock(&memcg_create_mutex);
5924 /* oom-kill-disable is a flag for subhierarchy. */
5925 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5926 mutex_unlock(&memcg_create_mutex);
5929 memcg->oom_kill_disable = val;
5931 memcg_oom_recover(memcg);
5932 mutex_unlock(&memcg_create_mutex);
5936 #ifdef CONFIG_MEMCG_KMEM
5937 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5941 memcg->kmemcg_id = -1;
5942 ret = memcg_propagate_kmem(memcg);
5946 return mem_cgroup_sockets_init(memcg, ss);
5949 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5951 mem_cgroup_sockets_destroy(memcg);
5954 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5956 if (!memcg_kmem_is_active(memcg))
5960 * kmem charges can outlive the cgroup. In the case of slab
5961 * pages, for instance, a page contain objects from various
5962 * processes. As we prevent from taking a reference for every
5963 * such allocation we have to be careful when doing uncharge
5964 * (see memcg_uncharge_kmem) and here during offlining.
5966 * The idea is that that only the _last_ uncharge which sees
5967 * the dead memcg will drop the last reference. An additional
5968 * reference is taken here before the group is marked dead
5969 * which is then paired with css_put during uncharge resp. here.
5971 * Although this might sound strange as this path is called from
5972 * css_offline() when the referencemight have dropped down to 0
5973 * and shouldn't be incremented anymore (css_tryget would fail)
5974 * we do not have other options because of the kmem allocations
5977 css_get(&memcg->css);
5979 memcg_kmem_mark_dead(memcg);
5981 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5984 if (memcg_kmem_test_and_clear_dead(memcg))
5985 css_put(&memcg->css);
5988 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5993 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5997 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
6003 * Unregister event and free resources.
6005 * Gets called from workqueue.
6007 static void cgroup_event_remove(struct work_struct *work)
6009 struct cgroup_event *event = container_of(work, struct cgroup_event,
6011 struct cgroup_subsys_state *css = event->css;
6013 remove_wait_queue(event->wqh, &event->wait);
6015 event->unregister_event(css, event->cft, event->eventfd);
6017 /* Notify userspace the event is going away. */
6018 eventfd_signal(event->eventfd, 1);
6020 eventfd_ctx_put(event->eventfd);
6026 * Gets called on POLLHUP on eventfd when user closes it.
6028 * Called with wqh->lock held and interrupts disabled.
6030 static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
6031 int sync, void *key)
6033 struct cgroup_event *event = container_of(wait,
6034 struct cgroup_event, wait);
6035 struct mem_cgroup *memcg = mem_cgroup_from_css(event->css);
6036 unsigned long flags = (unsigned long)key;
6038 if (flags & POLLHUP) {
6040 * If the event has been detached at cgroup removal, we
6041 * can simply return knowing the other side will cleanup
6044 * We can't race against event freeing since the other
6045 * side will require wqh->lock via remove_wait_queue(),
6048 spin_lock(&memcg->event_list_lock);
6049 if (!list_empty(&event->list)) {
6050 list_del_init(&event->list);
6052 * We are in atomic context, but cgroup_event_remove()
6053 * may sleep, so we have to call it in workqueue.
6055 schedule_work(&event->remove);
6057 spin_unlock(&memcg->event_list_lock);
6063 static void cgroup_event_ptable_queue_proc(struct file *file,
6064 wait_queue_head_t *wqh, poll_table *pt)
6066 struct cgroup_event *event = container_of(pt,
6067 struct cgroup_event, pt);
6070 add_wait_queue(wqh, &event->wait);
6074 * Parse input and register new cgroup event handler.
6076 * Input must be in format '<event_fd> <control_fd> <args>'.
6077 * Interpretation of args is defined by control file implementation.
6079 static int cgroup_write_event_control(struct cgroup_subsys_state *css,
6080 struct cftype *cft, const char *buffer)
6082 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6083 struct cgroup_event *event;
6084 struct cgroup_subsys_state *cfile_css;
6085 unsigned int efd, cfd;
6092 efd = simple_strtoul(buffer, &endp, 10);
6097 cfd = simple_strtoul(buffer, &endp, 10);
6098 if ((*endp != ' ') && (*endp != '\0'))
6102 event = kzalloc(sizeof(*event), GFP_KERNEL);
6107 INIT_LIST_HEAD(&event->list);
6108 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
6109 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
6110 INIT_WORK(&event->remove, cgroup_event_remove);
6118 event->eventfd = eventfd_ctx_fileget(efile.file);
6119 if (IS_ERR(event->eventfd)) {
6120 ret = PTR_ERR(event->eventfd);
6127 goto out_put_eventfd;
6130 /* the process need read permission on control file */
6131 /* AV: shouldn't we check that it's been opened for read instead? */
6132 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6136 event->cft = __file_cft(cfile.file);
6137 if (IS_ERR(event->cft)) {
6138 ret = PTR_ERR(event->cft);
6143 * Determine the event callbacks and set them in @event. This used
6144 * to be done via struct cftype but cgroup core no longer knows
6145 * about these events. The following is crude but the whole thing
6146 * is for compatibility anyway.
6148 name = cfile.file->f_dentry->d_name.name;
6150 if (!strcmp(name, "memory.usage_in_bytes")) {
6151 event->register_event = mem_cgroup_usage_register_event;
6152 event->unregister_event = mem_cgroup_usage_unregister_event;
6153 } else if (!strcmp(name, "memory.oom_control")) {
6154 event->register_event = mem_cgroup_oom_register_event;
6155 event->unregister_event = mem_cgroup_oom_unregister_event;
6156 } else if (!strcmp(name, "memory.pressure_level")) {
6157 event->register_event = vmpressure_register_event;
6158 event->unregister_event = vmpressure_unregister_event;
6159 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
6160 event->register_event = mem_cgroup_usage_register_event;
6161 event->unregister_event = mem_cgroup_usage_unregister_event;
6168 * Verify @cfile should belong to @css. Also, remaining events are
6169 * automatically removed on cgroup destruction but the removal is
6170 * asynchronous, so take an extra ref on @css.
6175 cfile_css = css_from_dir(cfile.file->f_dentry->d_parent,
6176 &mem_cgroup_subsys);
6177 if (cfile_css == css && css_tryget(css))
6184 ret = event->register_event(css, event->cft, event->eventfd, buffer);
6188 efile.file->f_op->poll(efile.file, &event->pt);
6190 spin_lock(&memcg->event_list_lock);
6191 list_add(&event->list, &memcg->event_list);
6192 spin_unlock(&memcg->event_list_lock);
6204 eventfd_ctx_put(event->eventfd);
6213 static struct cftype mem_cgroup_files[] = {
6215 .name = "usage_in_bytes",
6216 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6217 .read = mem_cgroup_read,
6220 .name = "max_usage_in_bytes",
6221 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6222 .trigger = mem_cgroup_reset,
6223 .read = mem_cgroup_read,
6226 .name = "limit_in_bytes",
6227 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6228 .write_string = mem_cgroup_write,
6229 .read = mem_cgroup_read,
6232 .name = "soft_limit_in_bytes",
6233 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6234 .write_string = mem_cgroup_write,
6235 .read = mem_cgroup_read,
6239 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6240 .trigger = mem_cgroup_reset,
6241 .read = mem_cgroup_read,
6245 .read_seq_string = memcg_stat_show,
6248 .name = "force_empty",
6249 .trigger = mem_cgroup_force_empty_write,
6252 .name = "use_hierarchy",
6253 .flags = CFTYPE_INSANE,
6254 .write_u64 = mem_cgroup_hierarchy_write,
6255 .read_u64 = mem_cgroup_hierarchy_read,
6258 .name = "cgroup.event_control",
6259 .write_string = cgroup_write_event_control,
6260 .flags = CFTYPE_NO_PREFIX,
6264 .name = "swappiness",
6265 .read_u64 = mem_cgroup_swappiness_read,
6266 .write_u64 = mem_cgroup_swappiness_write,
6269 .name = "move_charge_at_immigrate",
6270 .read_u64 = mem_cgroup_move_charge_read,
6271 .write_u64 = mem_cgroup_move_charge_write,
6274 .name = "oom_control",
6275 .read_map = mem_cgroup_oom_control_read,
6276 .write_u64 = mem_cgroup_oom_control_write,
6277 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6280 .name = "pressure_level",
6284 .name = "numa_stat",
6285 .read_seq_string = memcg_numa_stat_show,
6288 #ifdef CONFIG_MEMCG_KMEM
6290 .name = "kmem.limit_in_bytes",
6291 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6292 .write_string = mem_cgroup_write,
6293 .read = mem_cgroup_read,
6296 .name = "kmem.usage_in_bytes",
6297 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6298 .read = mem_cgroup_read,
6301 .name = "kmem.failcnt",
6302 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6303 .trigger = mem_cgroup_reset,
6304 .read = mem_cgroup_read,
6307 .name = "kmem.max_usage_in_bytes",
6308 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6309 .trigger = mem_cgroup_reset,
6310 .read = mem_cgroup_read,
6312 #ifdef CONFIG_SLABINFO
6314 .name = "kmem.slabinfo",
6315 .read_seq_string = mem_cgroup_slabinfo_read,
6319 { }, /* terminate */
6322 #ifdef CONFIG_MEMCG_SWAP
6323 static struct cftype memsw_cgroup_files[] = {
6325 .name = "memsw.usage_in_bytes",
6326 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6327 .read = mem_cgroup_read,
6330 .name = "memsw.max_usage_in_bytes",
6331 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6332 .trigger = mem_cgroup_reset,
6333 .read = mem_cgroup_read,
6336 .name = "memsw.limit_in_bytes",
6337 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6338 .write_string = mem_cgroup_write,
6339 .read = mem_cgroup_read,
6342 .name = "memsw.failcnt",
6343 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6344 .trigger = mem_cgroup_reset,
6345 .read = mem_cgroup_read,
6347 { }, /* terminate */
6350 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6352 struct mem_cgroup_per_node *pn;
6353 struct mem_cgroup_per_zone *mz;
6354 int zone, tmp = node;
6356 * This routine is called against possible nodes.
6357 * But it's BUG to call kmalloc() against offline node.
6359 * TODO: this routine can waste much memory for nodes which will
6360 * never be onlined. It's better to use memory hotplug callback
6363 if (!node_state(node, N_NORMAL_MEMORY))
6365 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6369 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6370 mz = &pn->zoneinfo[zone];
6371 lruvec_init(&mz->lruvec);
6372 mz->usage_in_excess = 0;
6373 mz->on_tree = false;
6376 memcg->nodeinfo[node] = pn;
6380 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6382 kfree(memcg->nodeinfo[node]);
6385 static struct mem_cgroup *mem_cgroup_alloc(void)
6387 struct mem_cgroup *memcg;
6388 size_t size = memcg_size();
6390 /* Can be very big if nr_node_ids is very big */
6391 if (size < PAGE_SIZE)
6392 memcg = kzalloc(size, GFP_KERNEL);
6394 memcg = vzalloc(size);
6399 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6402 spin_lock_init(&memcg->pcp_counter_lock);
6406 if (size < PAGE_SIZE)
6414 * At destroying mem_cgroup, references from swap_cgroup can remain.
6415 * (scanning all at force_empty is too costly...)
6417 * Instead of clearing all references at force_empty, we remember
6418 * the number of reference from swap_cgroup and free mem_cgroup when
6419 * it goes down to 0.
6421 * Removal of cgroup itself succeeds regardless of refs from swap.
6424 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6427 size_t size = memcg_size();
6429 mem_cgroup_remove_from_trees(memcg);
6430 free_css_id(&mem_cgroup_subsys, &memcg->css);
6433 free_mem_cgroup_per_zone_info(memcg, node);
6435 free_percpu(memcg->stat);
6438 * We need to make sure that (at least for now), the jump label
6439 * destruction code runs outside of the cgroup lock. This is because
6440 * get_online_cpus(), which is called from the static_branch update,
6441 * can't be called inside the cgroup_lock. cpusets are the ones
6442 * enforcing this dependency, so if they ever change, we might as well.
6444 * schedule_work() will guarantee this happens. Be careful if you need
6445 * to move this code around, and make sure it is outside
6448 disarm_static_keys(memcg);
6449 if (size < PAGE_SIZE)
6456 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6458 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6460 if (!memcg->res.parent)
6462 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6464 EXPORT_SYMBOL(parent_mem_cgroup);
6466 static void __init mem_cgroup_soft_limit_tree_init(void)
6468 struct mem_cgroup_tree_per_node *rtpn;
6469 struct mem_cgroup_tree_per_zone *rtpz;
6470 int tmp, node, zone;
6472 for_each_node(node) {
6474 if (!node_state(node, N_NORMAL_MEMORY))
6476 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6479 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6481 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6482 rtpz = &rtpn->rb_tree_per_zone[zone];
6483 rtpz->rb_root = RB_ROOT;
6484 spin_lock_init(&rtpz->lock);
6489 static struct cgroup_subsys_state * __ref
6490 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6492 struct mem_cgroup *memcg;
6493 long error = -ENOMEM;
6496 memcg = mem_cgroup_alloc();
6498 return ERR_PTR(error);
6501 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6505 if (parent_css == NULL) {
6506 root_mem_cgroup = memcg;
6507 res_counter_init(&memcg->res, NULL);
6508 res_counter_init(&memcg->memsw, NULL);
6509 res_counter_init(&memcg->kmem, NULL);
6512 memcg->last_scanned_node = MAX_NUMNODES;
6513 INIT_LIST_HEAD(&memcg->oom_notify);
6514 memcg->move_charge_at_immigrate = 0;
6515 mutex_init(&memcg->thresholds_lock);
6516 spin_lock_init(&memcg->move_lock);
6517 vmpressure_init(&memcg->vmpressure);
6518 INIT_LIST_HEAD(&memcg->event_list);
6519 spin_lock_init(&memcg->event_list_lock);
6524 __mem_cgroup_free(memcg);
6525 return ERR_PTR(error);
6529 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6531 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6532 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6538 mutex_lock(&memcg_create_mutex);
6540 memcg->use_hierarchy = parent->use_hierarchy;
6541 memcg->oom_kill_disable = parent->oom_kill_disable;
6542 memcg->swappiness = mem_cgroup_swappiness(parent);
6544 if (parent->use_hierarchy) {
6545 res_counter_init(&memcg->res, &parent->res);
6546 res_counter_init(&memcg->memsw, &parent->memsw);
6547 res_counter_init(&memcg->kmem, &parent->kmem);
6550 * No need to take a reference to the parent because cgroup
6551 * core guarantees its existence.
6554 res_counter_init(&memcg->res, NULL);
6555 res_counter_init(&memcg->memsw, NULL);
6556 res_counter_init(&memcg->kmem, NULL);
6558 * Deeper hierachy with use_hierarchy == false doesn't make
6559 * much sense so let cgroup subsystem know about this
6560 * unfortunate state in our controller.
6562 if (parent != root_mem_cgroup)
6563 mem_cgroup_subsys.broken_hierarchy = true;
6566 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6567 mutex_unlock(&memcg_create_mutex);
6572 * Announce all parents that a group from their hierarchy is gone.
6574 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6576 struct mem_cgroup *parent = memcg;
6578 while ((parent = parent_mem_cgroup(parent)))
6579 mem_cgroup_iter_invalidate(parent);
6582 * if the root memcg is not hierarchical we have to check it
6585 if (!root_mem_cgroup->use_hierarchy)
6586 mem_cgroup_iter_invalidate(root_mem_cgroup);
6589 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6591 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6592 struct cgroup_event *event, *tmp;
6595 * Unregister events and notify userspace.
6596 * Notify userspace about cgroup removing only after rmdir of cgroup
6597 * directory to avoid race between userspace and kernelspace.
6599 spin_lock(&memcg->event_list_lock);
6600 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6601 list_del_init(&event->list);
6602 schedule_work(&event->remove);
6604 spin_unlock(&memcg->event_list_lock);
6606 kmem_cgroup_css_offline(memcg);
6608 mem_cgroup_invalidate_reclaim_iterators(memcg);
6609 mem_cgroup_reparent_charges(memcg);
6610 mem_cgroup_destroy_all_caches(memcg);
6611 vmpressure_cleanup(&memcg->vmpressure);
6614 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6616 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6618 memcg_destroy_kmem(memcg);
6619 __mem_cgroup_free(memcg);
6623 /* Handlers for move charge at task migration. */
6624 #define PRECHARGE_COUNT_AT_ONCE 256
6625 static int mem_cgroup_do_precharge(unsigned long count)
6628 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6629 struct mem_cgroup *memcg = mc.to;
6631 if (mem_cgroup_is_root(memcg)) {
6632 mc.precharge += count;
6633 /* we don't need css_get for root */
6636 /* try to charge at once */
6638 struct res_counter *dummy;
6640 * "memcg" cannot be under rmdir() because we've already checked
6641 * by cgroup_lock_live_cgroup() that it is not removed and we
6642 * are still under the same cgroup_mutex. So we can postpone
6645 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6647 if (do_swap_account && res_counter_charge(&memcg->memsw,
6648 PAGE_SIZE * count, &dummy)) {
6649 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6652 mc.precharge += count;
6656 /* fall back to one by one charge */
6658 if (signal_pending(current)) {
6662 if (!batch_count--) {
6663 batch_count = PRECHARGE_COUNT_AT_ONCE;
6666 ret = __mem_cgroup_try_charge(NULL,
6667 GFP_KERNEL, 1, &memcg, false);
6669 /* mem_cgroup_clear_mc() will do uncharge later */
6677 * get_mctgt_type - get target type of moving charge
6678 * @vma: the vma the pte to be checked belongs
6679 * @addr: the address corresponding to the pte to be checked
6680 * @ptent: the pte to be checked
6681 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6684 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6685 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6686 * move charge. if @target is not NULL, the page is stored in target->page
6687 * with extra refcnt got(Callers should handle it).
6688 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6689 * target for charge migration. if @target is not NULL, the entry is stored
6692 * Called with pte lock held.
6699 enum mc_target_type {
6705 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6706 unsigned long addr, pte_t ptent)
6708 struct page *page = vm_normal_page(vma, addr, ptent);
6710 if (!page || !page_mapped(page))
6712 if (PageAnon(page)) {
6713 /* we don't move shared anon */
6716 } else if (!move_file())
6717 /* we ignore mapcount for file pages */
6719 if (!get_page_unless_zero(page))
6726 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6727 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6729 struct page *page = NULL;
6730 swp_entry_t ent = pte_to_swp_entry(ptent);
6732 if (!move_anon() || non_swap_entry(ent))
6735 * Because lookup_swap_cache() updates some statistics counter,
6736 * we call find_get_page() with swapper_space directly.
6738 page = find_get_page(swap_address_space(ent), ent.val);
6739 if (do_swap_account)
6740 entry->val = ent.val;
6745 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6746 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6752 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6753 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6755 struct page *page = NULL;
6756 struct address_space *mapping;
6759 if (!vma->vm_file) /* anonymous vma */
6764 mapping = vma->vm_file->f_mapping;
6765 if (pte_none(ptent))
6766 pgoff = linear_page_index(vma, addr);
6767 else /* pte_file(ptent) is true */
6768 pgoff = pte_to_pgoff(ptent);
6770 /* page is moved even if it's not RSS of this task(page-faulted). */
6771 page = find_get_page(mapping, pgoff);
6774 /* shmem/tmpfs may report page out on swap: account for that too. */
6775 if (radix_tree_exceptional_entry(page)) {
6776 swp_entry_t swap = radix_to_swp_entry(page);
6777 if (do_swap_account)
6779 page = find_get_page(swap_address_space(swap), swap.val);
6785 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6786 unsigned long addr, pte_t ptent, union mc_target *target)
6788 struct page *page = NULL;
6789 struct page_cgroup *pc;
6790 enum mc_target_type ret = MC_TARGET_NONE;
6791 swp_entry_t ent = { .val = 0 };
6793 if (pte_present(ptent))
6794 page = mc_handle_present_pte(vma, addr, ptent);
6795 else if (is_swap_pte(ptent))
6796 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6797 else if (pte_none(ptent) || pte_file(ptent))
6798 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6800 if (!page && !ent.val)
6803 pc = lookup_page_cgroup(page);
6805 * Do only loose check w/o page_cgroup lock.
6806 * mem_cgroup_move_account() checks the pc is valid or not under
6809 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6810 ret = MC_TARGET_PAGE;
6812 target->page = page;
6814 if (!ret || !target)
6817 /* There is a swap entry and a page doesn't exist or isn't charged */
6818 if (ent.val && !ret &&
6819 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6820 ret = MC_TARGET_SWAP;
6827 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6829 * We don't consider swapping or file mapped pages because THP does not
6830 * support them for now.
6831 * Caller should make sure that pmd_trans_huge(pmd) is true.
6833 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6834 unsigned long addr, pmd_t pmd, union mc_target *target)
6836 struct page *page = NULL;
6837 struct page_cgroup *pc;
6838 enum mc_target_type ret = MC_TARGET_NONE;
6840 page = pmd_page(pmd);
6841 VM_BUG_ON(!page || !PageHead(page));
6844 pc = lookup_page_cgroup(page);
6845 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6846 ret = MC_TARGET_PAGE;
6849 target->page = page;
6855 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6856 unsigned long addr, pmd_t pmd, union mc_target *target)
6858 return MC_TARGET_NONE;
6862 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6863 unsigned long addr, unsigned long end,
6864 struct mm_walk *walk)
6866 struct vm_area_struct *vma = walk->private;
6870 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6871 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6872 mc.precharge += HPAGE_PMD_NR;
6873 spin_unlock(&vma->vm_mm->page_table_lock);
6877 if (pmd_trans_unstable(pmd))
6879 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6880 for (; addr != end; pte++, addr += PAGE_SIZE)
6881 if (get_mctgt_type(vma, addr, *pte, NULL))
6882 mc.precharge++; /* increment precharge temporarily */
6883 pte_unmap_unlock(pte - 1, ptl);
6889 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6891 unsigned long precharge;
6892 struct vm_area_struct *vma;
6894 down_read(&mm->mmap_sem);
6895 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6896 struct mm_walk mem_cgroup_count_precharge_walk = {
6897 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6901 if (is_vm_hugetlb_page(vma))
6903 walk_page_range(vma->vm_start, vma->vm_end,
6904 &mem_cgroup_count_precharge_walk);
6906 up_read(&mm->mmap_sem);
6908 precharge = mc.precharge;
6914 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6916 unsigned long precharge = mem_cgroup_count_precharge(mm);
6918 VM_BUG_ON(mc.moving_task);
6919 mc.moving_task = current;
6920 return mem_cgroup_do_precharge(precharge);
6923 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6924 static void __mem_cgroup_clear_mc(void)
6926 struct mem_cgroup *from = mc.from;
6927 struct mem_cgroup *to = mc.to;
6930 /* we must uncharge all the leftover precharges from mc.to */
6932 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6936 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6937 * we must uncharge here.
6939 if (mc.moved_charge) {
6940 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6941 mc.moved_charge = 0;
6943 /* we must fixup refcnts and charges */
6944 if (mc.moved_swap) {
6945 /* uncharge swap account from the old cgroup */
6946 if (!mem_cgroup_is_root(mc.from))
6947 res_counter_uncharge(&mc.from->memsw,
6948 PAGE_SIZE * mc.moved_swap);
6950 for (i = 0; i < mc.moved_swap; i++)
6951 css_put(&mc.from->css);
6953 if (!mem_cgroup_is_root(mc.to)) {
6955 * we charged both to->res and to->memsw, so we should
6958 res_counter_uncharge(&mc.to->res,
6959 PAGE_SIZE * mc.moved_swap);
6961 /* we've already done css_get(mc.to) */
6964 memcg_oom_recover(from);
6965 memcg_oom_recover(to);
6966 wake_up_all(&mc.waitq);
6969 static void mem_cgroup_clear_mc(void)
6971 struct mem_cgroup *from = mc.from;
6974 * we must clear moving_task before waking up waiters at the end of
6977 mc.moving_task = NULL;
6978 __mem_cgroup_clear_mc();
6979 spin_lock(&mc.lock);
6982 spin_unlock(&mc.lock);
6983 mem_cgroup_end_move(from);
6986 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6987 struct cgroup_taskset *tset)
6989 struct task_struct *p = cgroup_taskset_first(tset);
6991 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6992 unsigned long move_charge_at_immigrate;
6995 * We are now commited to this value whatever it is. Changes in this
6996 * tunable will only affect upcoming migrations, not the current one.
6997 * So we need to save it, and keep it going.
6999 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
7000 if (move_charge_at_immigrate) {
7001 struct mm_struct *mm;
7002 struct mem_cgroup *from = mem_cgroup_from_task(p);
7004 VM_BUG_ON(from == memcg);
7006 mm = get_task_mm(p);
7009 /* We move charges only when we move a owner of the mm */
7010 if (mm->owner == p) {
7013 VM_BUG_ON(mc.precharge);
7014 VM_BUG_ON(mc.moved_charge);
7015 VM_BUG_ON(mc.moved_swap);
7016 mem_cgroup_start_move(from);
7017 spin_lock(&mc.lock);
7020 mc.immigrate_flags = move_charge_at_immigrate;
7021 spin_unlock(&mc.lock);
7022 /* We set mc.moving_task later */
7024 ret = mem_cgroup_precharge_mc(mm);
7026 mem_cgroup_clear_mc();
7033 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7034 struct cgroup_taskset *tset)
7036 mem_cgroup_clear_mc();
7039 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
7040 unsigned long addr, unsigned long end,
7041 struct mm_walk *walk)
7044 struct vm_area_struct *vma = walk->private;
7047 enum mc_target_type target_type;
7048 union mc_target target;
7050 struct page_cgroup *pc;
7053 * We don't take compound_lock() here but no race with splitting thp
7055 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7056 * under splitting, which means there's no concurrent thp split,
7057 * - if another thread runs into split_huge_page() just after we
7058 * entered this if-block, the thread must wait for page table lock
7059 * to be unlocked in __split_huge_page_splitting(), where the main
7060 * part of thp split is not executed yet.
7062 if (pmd_trans_huge_lock(pmd, vma) == 1) {
7063 if (mc.precharge < HPAGE_PMD_NR) {
7064 spin_unlock(&vma->vm_mm->page_table_lock);
7067 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
7068 if (target_type == MC_TARGET_PAGE) {
7070 if (!isolate_lru_page(page)) {
7071 pc = lookup_page_cgroup(page);
7072 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
7073 pc, mc.from, mc.to)) {
7074 mc.precharge -= HPAGE_PMD_NR;
7075 mc.moved_charge += HPAGE_PMD_NR;
7077 putback_lru_page(page);
7081 spin_unlock(&vma->vm_mm->page_table_lock);
7085 if (pmd_trans_unstable(pmd))
7088 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7089 for (; addr != end; addr += PAGE_SIZE) {
7090 pte_t ptent = *(pte++);
7096 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7097 case MC_TARGET_PAGE:
7099 if (isolate_lru_page(page))
7101 pc = lookup_page_cgroup(page);
7102 if (!mem_cgroup_move_account(page, 1, pc,
7105 /* we uncharge from mc.from later. */
7108 putback_lru_page(page);
7109 put: /* get_mctgt_type() gets the page */
7112 case MC_TARGET_SWAP:
7114 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7116 /* we fixup refcnts and charges later. */
7124 pte_unmap_unlock(pte - 1, ptl);
7129 * We have consumed all precharges we got in can_attach().
7130 * We try charge one by one, but don't do any additional
7131 * charges to mc.to if we have failed in charge once in attach()
7134 ret = mem_cgroup_do_precharge(1);
7142 static void mem_cgroup_move_charge(struct mm_struct *mm)
7144 struct vm_area_struct *vma;
7146 lru_add_drain_all();
7148 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7150 * Someone who are holding the mmap_sem might be waiting in
7151 * waitq. So we cancel all extra charges, wake up all waiters,
7152 * and retry. Because we cancel precharges, we might not be able
7153 * to move enough charges, but moving charge is a best-effort
7154 * feature anyway, so it wouldn't be a big problem.
7156 __mem_cgroup_clear_mc();
7160 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7162 struct mm_walk mem_cgroup_move_charge_walk = {
7163 .pmd_entry = mem_cgroup_move_charge_pte_range,
7167 if (is_vm_hugetlb_page(vma))
7169 ret = walk_page_range(vma->vm_start, vma->vm_end,
7170 &mem_cgroup_move_charge_walk);
7173 * means we have consumed all precharges and failed in
7174 * doing additional charge. Just abandon here.
7178 up_read(&mm->mmap_sem);
7181 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7182 struct cgroup_taskset *tset)
7184 struct task_struct *p = cgroup_taskset_first(tset);
7185 struct mm_struct *mm = get_task_mm(p);
7189 mem_cgroup_move_charge(mm);
7193 mem_cgroup_clear_mc();
7195 #else /* !CONFIG_MMU */
7196 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7197 struct cgroup_taskset *tset)
7201 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7202 struct cgroup_taskset *tset)
7205 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7206 struct cgroup_taskset *tset)
7212 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7213 * to verify sane_behavior flag on each mount attempt.
7215 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7218 * use_hierarchy is forced with sane_behavior. cgroup core
7219 * guarantees that @root doesn't have any children, so turning it
7220 * on for the root memcg is enough.
7222 if (cgroup_sane_behavior(root_css->cgroup))
7223 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7226 struct cgroup_subsys mem_cgroup_subsys = {
7228 .subsys_id = mem_cgroup_subsys_id,
7229 .css_alloc = mem_cgroup_css_alloc,
7230 .css_online = mem_cgroup_css_online,
7231 .css_offline = mem_cgroup_css_offline,
7232 .css_free = mem_cgroup_css_free,
7233 .can_attach = mem_cgroup_can_attach,
7234 .cancel_attach = mem_cgroup_cancel_attach,
7235 .attach = mem_cgroup_move_task,
7236 .bind = mem_cgroup_bind,
7237 .base_cftypes = mem_cgroup_files,
7242 #ifdef CONFIG_MEMCG_SWAP
7243 static int __init enable_swap_account(char *s)
7245 if (!strcmp(s, "1"))
7246 really_do_swap_account = 1;
7247 else if (!strcmp(s, "0"))
7248 really_do_swap_account = 0;
7251 __setup("swapaccount=", enable_swap_account);
7253 static void __init memsw_file_init(void)
7255 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7258 static void __init enable_swap_cgroup(void)
7260 if (!mem_cgroup_disabled() && really_do_swap_account) {
7261 do_swap_account = 1;
7267 static void __init enable_swap_cgroup(void)
7273 * subsys_initcall() for memory controller.
7275 * Some parts like hotcpu_notifier() have to be initialized from this context
7276 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7277 * everything that doesn't depend on a specific mem_cgroup structure should
7278 * be initialized from here.
7280 static int __init mem_cgroup_init(void)
7282 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7283 enable_swap_cgroup();
7284 mem_cgroup_soft_limit_tree_init();
7288 subsys_initcall(mem_cgroup_init);