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/slab.h>
43 #include <linux/swap.h>
44 #include <linux/swapops.h>
45 #include <linux/spinlock.h>
46 #include <linux/eventfd.h>
47 #include <linux/sort.h>
49 #include <linux/seq_file.h>
50 #include <linux/vmalloc.h>
51 #include <linux/vmpressure.h>
52 #include <linux/mm_inline.h>
53 #include <linux/page_cgroup.h>
54 #include <linux/cpu.h>
55 #include <linux/oom.h>
59 #include <net/tcp_memcontrol.h>
61 #include <asm/uaccess.h>
63 #include <trace/events/vmscan.h>
65 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
66 EXPORT_SYMBOL(mem_cgroup_subsys);
68 #define MEM_CGROUP_RECLAIM_RETRIES 5
69 static struct mem_cgroup *root_mem_cgroup __read_mostly;
71 #ifdef CONFIG_MEMCG_SWAP
72 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
73 int do_swap_account __read_mostly;
75 /* for remember boot option*/
76 #ifdef CONFIG_MEMCG_SWAP_ENABLED
77 static int really_do_swap_account __initdata = 1;
79 static int really_do_swap_account __initdata = 0;
83 #define do_swap_account 0
87 static const char * const mem_cgroup_stat_names[] = {
96 enum mem_cgroup_events_index {
97 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
98 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
99 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
100 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
101 MEM_CGROUP_EVENTS_NSTATS,
104 static const char * const mem_cgroup_events_names[] = {
111 static const char * const mem_cgroup_lru_names[] = {
120 * Per memcg event counter is incremented at every pagein/pageout. With THP,
121 * it will be incremated by the number of pages. This counter is used for
122 * for trigger some periodic events. This is straightforward and better
123 * than using jiffies etc. to handle periodic memcg event.
125 enum mem_cgroup_events_target {
126 MEM_CGROUP_TARGET_THRESH,
127 MEM_CGROUP_TARGET_NUMAINFO,
130 #define THRESHOLDS_EVENTS_TARGET 128
131 #define SOFTLIMIT_EVENTS_TARGET 1024
132 #define NUMAINFO_EVENTS_TARGET 1024
134 struct mem_cgroup_stat_cpu {
135 long count[MEM_CGROUP_STAT_NSTATS];
136 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
137 unsigned long nr_page_events;
138 unsigned long targets[MEM_CGROUP_NTARGETS];
141 struct mem_cgroup_reclaim_iter {
143 * last scanned hierarchy member. Valid only if last_dead_count
144 * matches memcg->dead_count of the hierarchy root group.
146 struct mem_cgroup *last_visited;
147 unsigned long last_dead_count;
149 /* scan generation, increased every round-trip */
150 unsigned int generation;
154 * per-zone information in memory controller.
156 struct mem_cgroup_per_zone {
157 struct lruvec lruvec;
158 unsigned long lru_size[NR_LRU_LISTS];
160 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
162 struct mem_cgroup *memcg; /* Back pointer, we cannot */
163 /* use container_of */
166 struct mem_cgroup_per_node {
167 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
170 struct mem_cgroup_threshold {
171 struct eventfd_ctx *eventfd;
176 struct mem_cgroup_threshold_ary {
177 /* An array index points to threshold just below or equal to usage. */
178 int current_threshold;
179 /* Size of entries[] */
181 /* Array of thresholds */
182 struct mem_cgroup_threshold entries[0];
185 struct mem_cgroup_thresholds {
186 /* Primary thresholds array */
187 struct mem_cgroup_threshold_ary *primary;
189 * Spare threshold array.
190 * This is needed to make mem_cgroup_unregister_event() "never fail".
191 * It must be able to store at least primary->size - 1 entries.
193 struct mem_cgroup_threshold_ary *spare;
197 struct mem_cgroup_eventfd_list {
198 struct list_head list;
199 struct eventfd_ctx *eventfd;
202 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
203 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
206 * The memory controller data structure. The memory controller controls both
207 * page cache and RSS per cgroup. We would eventually like to provide
208 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
209 * to help the administrator determine what knobs to tune.
211 * TODO: Add a water mark for the memory controller. Reclaim will begin when
212 * we hit the water mark. May be even add a low water mark, such that
213 * no reclaim occurs from a cgroup at it's low water mark, this is
214 * a feature that will be implemented much later in the future.
217 struct cgroup_subsys_state css;
219 * the counter to account for memory usage
221 struct res_counter res;
223 /* vmpressure notifications */
224 struct vmpressure vmpressure;
227 * the counter to account for mem+swap usage.
229 struct res_counter memsw;
232 * the counter to account for kernel memory usage.
234 struct res_counter kmem;
236 * Should the accounting and control be hierarchical, per subtree?
239 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
243 atomic_t oom_wakeups;
246 /* OOM-Killer disable */
247 int oom_kill_disable;
249 /* set when res.limit == memsw.limit */
250 bool memsw_is_minimum;
252 /* protect arrays of thresholds */
253 struct mutex thresholds_lock;
255 /* thresholds for memory usage. RCU-protected */
256 struct mem_cgroup_thresholds thresholds;
258 /* thresholds for mem+swap usage. RCU-protected */
259 struct mem_cgroup_thresholds memsw_thresholds;
261 /* For oom notifier event fd */
262 struct list_head oom_notify;
265 * Should we move charges of a task when a task is moved into this
266 * mem_cgroup ? And what type of charges should we move ?
268 unsigned long move_charge_at_immigrate;
270 * set > 0 if pages under this cgroup are moving to other cgroup.
272 atomic_t moving_account;
273 /* taken only while moving_account > 0 */
274 spinlock_t move_lock;
278 struct mem_cgroup_stat_cpu __percpu *stat;
280 * used when a cpu is offlined or other synchronizations
281 * See mem_cgroup_read_stat().
283 struct mem_cgroup_stat_cpu nocpu_base;
284 spinlock_t pcp_counter_lock;
287 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
288 struct tcp_memcontrol tcp_mem;
290 #if defined(CONFIG_MEMCG_KMEM)
291 /* analogous to slab_common's slab_caches list. per-memcg */
292 struct list_head memcg_slab_caches;
293 /* Not a spinlock, we can take a lot of time walking the list */
294 struct mutex slab_caches_mutex;
295 /* Index in the kmem_cache->memcg_params->memcg_caches array */
299 int last_scanned_node;
301 nodemask_t scan_nodes;
302 atomic_t numainfo_events;
303 atomic_t numainfo_updating;
306 struct mem_cgroup_per_node *nodeinfo[0];
307 /* WARNING: nodeinfo must be the last member here */
310 static size_t memcg_size(void)
312 return sizeof(struct mem_cgroup) +
313 nr_node_ids * sizeof(struct mem_cgroup_per_node);
316 /* internal only representation about the status of kmem accounting. */
318 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
319 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
320 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
323 /* We account when limit is on, but only after call sites are patched */
324 #define KMEM_ACCOUNTED_MASK \
325 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
327 #ifdef CONFIG_MEMCG_KMEM
328 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
330 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
333 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
335 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
338 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
340 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
343 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
345 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
348 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
351 * Our caller must use css_get() first, because memcg_uncharge_kmem()
352 * will call css_put() if it sees the memcg is dead.
355 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
356 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
359 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
361 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
362 &memcg->kmem_account_flags);
366 /* Stuffs for move charges at task migration. */
368 * Types of charges to be moved. "move_charge_at_immitgrate" and
369 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
372 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
373 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
377 /* "mc" and its members are protected by cgroup_mutex */
378 static struct move_charge_struct {
379 spinlock_t lock; /* for from, to */
380 struct mem_cgroup *from;
381 struct mem_cgroup *to;
382 unsigned long immigrate_flags;
383 unsigned long precharge;
384 unsigned long moved_charge;
385 unsigned long moved_swap;
386 struct task_struct *moving_task; /* a task moving charges */
387 wait_queue_head_t waitq; /* a waitq for other context */
389 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
390 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
393 static bool move_anon(void)
395 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
398 static bool move_file(void)
400 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
404 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
405 * limit reclaim to prevent infinite loops, if they ever occur.
407 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
410 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
411 MEM_CGROUP_CHARGE_TYPE_ANON,
412 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
413 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
417 /* for encoding cft->private value on file */
425 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
426 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
427 #define MEMFILE_ATTR(val) ((val) & 0xffff)
428 /* Used for OOM nofiier */
429 #define OOM_CONTROL (0)
432 * Reclaim flags for mem_cgroup_hierarchical_reclaim
434 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
435 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
436 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
437 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
440 * The memcg_create_mutex will be held whenever a new cgroup is created.
441 * As a consequence, any change that needs to protect against new child cgroups
442 * appearing has to hold it as well.
444 static DEFINE_MUTEX(memcg_create_mutex);
446 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
448 return s ? container_of(s, struct mem_cgroup, css) : NULL;
451 /* Some nice accessors for the vmpressure. */
452 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
455 memcg = root_mem_cgroup;
456 return &memcg->vmpressure;
459 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
461 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
464 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
466 return &mem_cgroup_from_css(css)->vmpressure;
469 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
471 return (memcg == root_mem_cgroup);
474 /* Writing them here to avoid exposing memcg's inner layout */
475 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
477 void sock_update_memcg(struct sock *sk)
479 if (mem_cgroup_sockets_enabled) {
480 struct mem_cgroup *memcg;
481 struct cg_proto *cg_proto;
483 BUG_ON(!sk->sk_prot->proto_cgroup);
485 /* Socket cloning can throw us here with sk_cgrp already
486 * filled. It won't however, necessarily happen from
487 * process context. So the test for root memcg given
488 * the current task's memcg won't help us in this case.
490 * Respecting the original socket's memcg is a better
491 * decision in this case.
494 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
495 css_get(&sk->sk_cgrp->memcg->css);
500 memcg = mem_cgroup_from_task(current);
501 cg_proto = sk->sk_prot->proto_cgroup(memcg);
502 if (!mem_cgroup_is_root(memcg) &&
503 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
504 sk->sk_cgrp = cg_proto;
509 EXPORT_SYMBOL(sock_update_memcg);
511 void sock_release_memcg(struct sock *sk)
513 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
514 struct mem_cgroup *memcg;
515 WARN_ON(!sk->sk_cgrp->memcg);
516 memcg = sk->sk_cgrp->memcg;
517 css_put(&sk->sk_cgrp->memcg->css);
521 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
523 if (!memcg || mem_cgroup_is_root(memcg))
526 return &memcg->tcp_mem.cg_proto;
528 EXPORT_SYMBOL(tcp_proto_cgroup);
530 static void disarm_sock_keys(struct mem_cgroup *memcg)
532 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
534 static_key_slow_dec(&memcg_socket_limit_enabled);
537 static void disarm_sock_keys(struct mem_cgroup *memcg)
542 #ifdef CONFIG_MEMCG_KMEM
544 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
545 * There are two main reasons for not using the css_id for this:
546 * 1) this works better in sparse environments, where we have a lot of memcgs,
547 * but only a few kmem-limited. Or also, if we have, for instance, 200
548 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
549 * 200 entry array for that.
551 * 2) In order not to violate the cgroup API, we would like to do all memory
552 * allocation in ->create(). At that point, we haven't yet allocated the
553 * css_id. Having a separate index prevents us from messing with the cgroup
556 * The current size of the caches array is stored in
557 * memcg_limited_groups_array_size. It will double each time we have to
560 static DEFINE_IDA(kmem_limited_groups);
561 int memcg_limited_groups_array_size;
564 * MIN_SIZE is different than 1, because we would like to avoid going through
565 * the alloc/free process all the time. In a small machine, 4 kmem-limited
566 * cgroups is a reasonable guess. In the future, it could be a parameter or
567 * tunable, but that is strictly not necessary.
569 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
570 * this constant directly from cgroup, but it is understandable that this is
571 * better kept as an internal representation in cgroup.c. In any case, the
572 * css_id space is not getting any smaller, and we don't have to necessarily
573 * increase ours as well if it increases.
575 #define MEMCG_CACHES_MIN_SIZE 4
576 #define MEMCG_CACHES_MAX_SIZE 65535
579 * A lot of the calls to the cache allocation functions are expected to be
580 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
581 * conditional to this static branch, we'll have to allow modules that does
582 * kmem_cache_alloc and the such to see this symbol as well
584 struct static_key memcg_kmem_enabled_key;
585 EXPORT_SYMBOL(memcg_kmem_enabled_key);
587 static void disarm_kmem_keys(struct mem_cgroup *memcg)
589 if (memcg_kmem_is_active(memcg)) {
590 static_key_slow_dec(&memcg_kmem_enabled_key);
591 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
594 * This check can't live in kmem destruction function,
595 * since the charges will outlive the cgroup
597 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
600 static void disarm_kmem_keys(struct mem_cgroup *memcg)
603 #endif /* CONFIG_MEMCG_KMEM */
605 static void disarm_static_keys(struct mem_cgroup *memcg)
607 disarm_sock_keys(memcg);
608 disarm_kmem_keys(memcg);
611 static void drain_all_stock_async(struct mem_cgroup *memcg);
613 static struct mem_cgroup_per_zone *
614 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
616 VM_BUG_ON((unsigned)nid >= nr_node_ids);
617 return &memcg->nodeinfo[nid]->zoneinfo[zid];
620 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
625 static struct mem_cgroup_per_zone *
626 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
628 int nid = page_to_nid(page);
629 int zid = page_zonenum(page);
631 return mem_cgroup_zoneinfo(memcg, nid, zid);
635 * Implementation Note: reading percpu statistics for memcg.
637 * Both of vmstat[] and percpu_counter has threshold and do periodic
638 * synchronization to implement "quick" read. There are trade-off between
639 * reading cost and precision of value. Then, we may have a chance to implement
640 * a periodic synchronizion of counter in memcg's counter.
642 * But this _read() function is used for user interface now. The user accounts
643 * memory usage by memory cgroup and he _always_ requires exact value because
644 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
645 * have to visit all online cpus and make sum. So, for now, unnecessary
646 * synchronization is not implemented. (just implemented for cpu hotplug)
648 * If there are kernel internal actions which can make use of some not-exact
649 * value, and reading all cpu value can be performance bottleneck in some
650 * common workload, threashold and synchonization as vmstat[] should be
653 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
654 enum mem_cgroup_stat_index idx)
660 for_each_online_cpu(cpu)
661 val += per_cpu(memcg->stat->count[idx], cpu);
662 #ifdef CONFIG_HOTPLUG_CPU
663 spin_lock(&memcg->pcp_counter_lock);
664 val += memcg->nocpu_base.count[idx];
665 spin_unlock(&memcg->pcp_counter_lock);
671 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
674 int val = (charge) ? 1 : -1;
675 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
678 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
679 enum mem_cgroup_events_index idx)
681 unsigned long val = 0;
684 for_each_online_cpu(cpu)
685 val += per_cpu(memcg->stat->events[idx], cpu);
686 #ifdef CONFIG_HOTPLUG_CPU
687 spin_lock(&memcg->pcp_counter_lock);
688 val += memcg->nocpu_base.events[idx];
689 spin_unlock(&memcg->pcp_counter_lock);
694 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
696 bool anon, int nr_pages)
701 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
702 * counted as CACHE even if it's on ANON LRU.
705 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
708 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
711 if (PageTransHuge(page))
712 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
715 /* pagein of a big page is an event. So, ignore page size */
717 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
719 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
720 nr_pages = -nr_pages; /* for event */
723 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
729 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
731 struct mem_cgroup_per_zone *mz;
733 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
734 return mz->lru_size[lru];
738 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
739 unsigned int lru_mask)
741 struct mem_cgroup_per_zone *mz;
743 unsigned long ret = 0;
745 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
748 if (BIT(lru) & lru_mask)
749 ret += mz->lru_size[lru];
755 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
756 int nid, unsigned int lru_mask)
761 for (zid = 0; zid < MAX_NR_ZONES; zid++)
762 total += mem_cgroup_zone_nr_lru_pages(memcg,
768 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
769 unsigned int lru_mask)
774 for_each_node_state(nid, N_MEMORY)
775 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
779 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
780 enum mem_cgroup_events_target target)
782 unsigned long val, next;
784 val = __this_cpu_read(memcg->stat->nr_page_events);
785 next = __this_cpu_read(memcg->stat->targets[target]);
786 /* from time_after() in jiffies.h */
787 if ((long)next - (long)val < 0) {
789 case MEM_CGROUP_TARGET_THRESH:
790 next = val + THRESHOLDS_EVENTS_TARGET;
792 case MEM_CGROUP_TARGET_NUMAINFO:
793 next = val + NUMAINFO_EVENTS_TARGET;
798 __this_cpu_write(memcg->stat->targets[target], next);
805 * Check events in order.
808 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
811 /* threshold event is triggered in finer grain than soft limit */
812 if (unlikely(mem_cgroup_event_ratelimit(memcg,
813 MEM_CGROUP_TARGET_THRESH))) {
814 bool do_numainfo __maybe_unused;
817 do_numainfo = mem_cgroup_event_ratelimit(memcg,
818 MEM_CGROUP_TARGET_NUMAINFO);
822 mem_cgroup_threshold(memcg);
824 if (unlikely(do_numainfo))
825 atomic_inc(&memcg->numainfo_events);
831 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
834 * mm_update_next_owner() may clear mm->owner to NULL
835 * if it races with swapoff, page migration, etc.
836 * So this can be called with p == NULL.
841 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
844 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
846 struct mem_cgroup *memcg = NULL;
851 * Because we have no locks, mm->owner's may be being moved to other
852 * cgroup. We use css_tryget() here even if this looks
853 * pessimistic (rather than adding locks here).
857 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
858 if (unlikely(!memcg))
860 } while (!css_tryget(&memcg->css));
866 * Returns a next (in a pre-order walk) alive memcg (with elevated css
867 * ref. count) or NULL if the whole root's subtree has been visited.
869 * helper function to be used by mem_cgroup_iter
871 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
872 struct mem_cgroup *last_visited)
874 struct cgroup_subsys_state *prev_css, *next_css;
876 prev_css = last_visited ? &last_visited->css : NULL;
878 next_css = css_next_descendant_pre(prev_css, &root->css);
881 * Even if we found a group we have to make sure it is
882 * alive. css && !memcg means that the groups should be
883 * skipped and we should continue the tree walk.
884 * last_visited css is safe to use because it is
885 * protected by css_get and the tree walk is rcu safe.
888 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
890 if (css_tryget(&mem->css))
901 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
904 * When a group in the hierarchy below root is destroyed, the
905 * hierarchy iterator can no longer be trusted since it might
906 * have pointed to the destroyed group. Invalidate it.
908 atomic_inc(&root->dead_count);
911 static struct mem_cgroup *
912 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
913 struct mem_cgroup *root,
916 struct mem_cgroup *position = NULL;
918 * A cgroup destruction happens in two stages: offlining and
919 * release. They are separated by a RCU grace period.
921 * If the iterator is valid, we may still race with an
922 * offlining. The RCU lock ensures the object won't be
923 * released, tryget will fail if we lost the race.
925 *sequence = atomic_read(&root->dead_count);
926 if (iter->last_dead_count == *sequence) {
928 position = iter->last_visited;
929 if (position && !css_tryget(&position->css))
935 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
936 struct mem_cgroup *last_visited,
937 struct mem_cgroup *new_position,
941 css_put(&last_visited->css);
943 * We store the sequence count from the time @last_visited was
944 * loaded successfully instead of rereading it here so that we
945 * don't lose destruction events in between. We could have
946 * raced with the destruction of @new_position after all.
948 iter->last_visited = new_position;
950 iter->last_dead_count = sequence;
954 * mem_cgroup_iter - iterate over memory cgroup hierarchy
955 * @root: hierarchy root
956 * @prev: previously returned memcg, NULL on first invocation
957 * @reclaim: cookie for shared reclaim walks, NULL for full walks
959 * Returns references to children of the hierarchy below @root, or
960 * @root itself, or %NULL after a full round-trip.
962 * Caller must pass the return value in @prev on subsequent
963 * invocations for reference counting, or use mem_cgroup_iter_break()
964 * to cancel a hierarchy walk before the round-trip is complete.
966 * Reclaimers can specify a zone and a priority level in @reclaim to
967 * divide up the memcgs in the hierarchy among all concurrent
968 * reclaimers operating on the same zone and priority.
970 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
971 struct mem_cgroup *prev,
972 struct mem_cgroup_reclaim_cookie *reclaim)
974 struct mem_cgroup *memcg = NULL;
975 struct mem_cgroup *last_visited = NULL;
977 if (mem_cgroup_disabled())
981 root = root_mem_cgroup;
983 if (prev && !reclaim)
986 if (!root->use_hierarchy && root != root_mem_cgroup) {
994 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
995 int uninitialized_var(seq);
998 int nid = zone_to_nid(reclaim->zone);
999 int zid = zone_idx(reclaim->zone);
1000 struct mem_cgroup_per_zone *mz;
1002 mz = mem_cgroup_zoneinfo(root, nid, zid);
1003 iter = &mz->reclaim_iter[reclaim->priority];
1004 if (prev && reclaim->generation != iter->generation) {
1005 iter->last_visited = NULL;
1009 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1012 memcg = __mem_cgroup_iter_next(root, last_visited);
1015 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1019 else if (!prev && memcg)
1020 reclaim->generation = iter->generation;
1029 if (prev && prev != root)
1030 css_put(&prev->css);
1036 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1037 * @root: hierarchy root
1038 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1040 void mem_cgroup_iter_break(struct mem_cgroup *root,
1041 struct mem_cgroup *prev)
1044 root = root_mem_cgroup;
1045 if (prev && prev != root)
1046 css_put(&prev->css);
1050 * Iteration constructs for visiting all cgroups (under a tree). If
1051 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1052 * be used for reference counting.
1054 #define for_each_mem_cgroup_tree(iter, root) \
1055 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1057 iter = mem_cgroup_iter(root, iter, NULL))
1059 #define for_each_mem_cgroup(iter) \
1060 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1062 iter = mem_cgroup_iter(NULL, iter, NULL))
1064 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1066 struct mem_cgroup *memcg;
1069 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1070 if (unlikely(!memcg))
1075 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1078 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1086 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1089 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1090 * @zone: zone of the wanted lruvec
1091 * @memcg: memcg of the wanted lruvec
1093 * Returns the lru list vector holding pages for the given @zone and
1094 * @mem. This can be the global zone lruvec, if the memory controller
1097 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1098 struct mem_cgroup *memcg)
1100 struct mem_cgroup_per_zone *mz;
1101 struct lruvec *lruvec;
1103 if (mem_cgroup_disabled()) {
1104 lruvec = &zone->lruvec;
1108 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1109 lruvec = &mz->lruvec;
1112 * Since a node can be onlined after the mem_cgroup was created,
1113 * we have to be prepared to initialize lruvec->zone here;
1114 * and if offlined then reonlined, we need to reinitialize it.
1116 if (unlikely(lruvec->zone != zone))
1117 lruvec->zone = zone;
1122 * Following LRU functions are allowed to be used without PCG_LOCK.
1123 * Operations are called by routine of global LRU independently from memcg.
1124 * What we have to take care of here is validness of pc->mem_cgroup.
1126 * Changes to pc->mem_cgroup happens when
1129 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1130 * It is added to LRU before charge.
1131 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1132 * When moving account, the page is not on LRU. It's isolated.
1136 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1138 * @zone: zone of the page
1140 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1142 struct mem_cgroup_per_zone *mz;
1143 struct mem_cgroup *memcg;
1144 struct page_cgroup *pc;
1145 struct lruvec *lruvec;
1147 if (mem_cgroup_disabled()) {
1148 lruvec = &zone->lruvec;
1152 pc = lookup_page_cgroup(page);
1153 memcg = pc->mem_cgroup;
1156 * Surreptitiously switch any uncharged offlist page to root:
1157 * an uncharged page off lru does nothing to secure
1158 * its former mem_cgroup from sudden removal.
1160 * Our caller holds lru_lock, and PageCgroupUsed is updated
1161 * under page_cgroup lock: between them, they make all uses
1162 * of pc->mem_cgroup safe.
1164 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1165 pc->mem_cgroup = memcg = root_mem_cgroup;
1167 mz = page_cgroup_zoneinfo(memcg, page);
1168 lruvec = &mz->lruvec;
1171 * Since a node can be onlined after the mem_cgroup was created,
1172 * we have to be prepared to initialize lruvec->zone here;
1173 * and if offlined then reonlined, we need to reinitialize it.
1175 if (unlikely(lruvec->zone != zone))
1176 lruvec->zone = zone;
1181 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1182 * @lruvec: mem_cgroup per zone lru vector
1183 * @lru: index of lru list the page is sitting on
1184 * @nr_pages: positive when adding or negative when removing
1186 * This function must be called when a page is added to or removed from an
1189 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1192 struct mem_cgroup_per_zone *mz;
1193 unsigned long *lru_size;
1195 if (mem_cgroup_disabled())
1198 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1199 lru_size = mz->lru_size + lru;
1200 *lru_size += nr_pages;
1201 VM_BUG_ON((long)(*lru_size) < 0);
1205 * Checks whether given mem is same or in the root_mem_cgroup's
1208 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1209 struct mem_cgroup *memcg)
1211 if (root_memcg == memcg)
1213 if (!root_memcg->use_hierarchy || !memcg)
1215 return css_is_ancestor(&memcg->css, &root_memcg->css);
1218 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1219 struct mem_cgroup *memcg)
1224 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1229 bool task_in_mem_cgroup(struct task_struct *task,
1230 const struct mem_cgroup *memcg)
1232 struct mem_cgroup *curr = NULL;
1233 struct task_struct *p;
1236 p = find_lock_task_mm(task);
1238 curr = try_get_mem_cgroup_from_mm(p->mm);
1242 * All threads may have already detached their mm's, but the oom
1243 * killer still needs to detect if they have already been oom
1244 * killed to prevent needlessly killing additional tasks.
1247 curr = mem_cgroup_from_task(task);
1249 css_get(&curr->css);
1255 * We should check use_hierarchy of "memcg" not "curr". Because checking
1256 * use_hierarchy of "curr" here make this function true if hierarchy is
1257 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1258 * hierarchy(even if use_hierarchy is disabled in "memcg").
1260 ret = mem_cgroup_same_or_subtree(memcg, curr);
1261 css_put(&curr->css);
1265 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1267 unsigned long inactive_ratio;
1268 unsigned long inactive;
1269 unsigned long active;
1272 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1273 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1275 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1277 inactive_ratio = int_sqrt(10 * gb);
1281 return inactive * inactive_ratio < active;
1284 #define mem_cgroup_from_res_counter(counter, member) \
1285 container_of(counter, struct mem_cgroup, member)
1288 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1289 * @memcg: the memory cgroup
1291 * Returns the maximum amount of memory @mem can be charged with, in
1294 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1296 unsigned long long margin;
1298 margin = res_counter_margin(&memcg->res);
1299 if (do_swap_account)
1300 margin = min(margin, res_counter_margin(&memcg->memsw));
1301 return margin >> PAGE_SHIFT;
1304 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1307 if (!css_parent(&memcg->css))
1308 return vm_swappiness;
1310 return memcg->swappiness;
1314 * memcg->moving_account is used for checking possibility that some thread is
1315 * calling move_account(). When a thread on CPU-A starts moving pages under
1316 * a memcg, other threads should check memcg->moving_account under
1317 * rcu_read_lock(), like this:
1321 * memcg->moving_account+1 if (memcg->mocing_account)
1323 * synchronize_rcu() update something.
1328 /* for quick checking without looking up memcg */
1329 atomic_t memcg_moving __read_mostly;
1331 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1333 atomic_inc(&memcg_moving);
1334 atomic_inc(&memcg->moving_account);
1338 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1341 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1342 * We check NULL in callee rather than caller.
1345 atomic_dec(&memcg_moving);
1346 atomic_dec(&memcg->moving_account);
1351 * 2 routines for checking "mem" is under move_account() or not.
1353 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1354 * is used for avoiding races in accounting. If true,
1355 * pc->mem_cgroup may be overwritten.
1357 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1358 * under hierarchy of moving cgroups. This is for
1359 * waiting at hith-memory prressure caused by "move".
1362 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1364 VM_BUG_ON(!rcu_read_lock_held());
1365 return atomic_read(&memcg->moving_account) > 0;
1368 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1370 struct mem_cgroup *from;
1371 struct mem_cgroup *to;
1374 * Unlike task_move routines, we access mc.to, mc.from not under
1375 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1377 spin_lock(&mc.lock);
1383 ret = mem_cgroup_same_or_subtree(memcg, from)
1384 || mem_cgroup_same_or_subtree(memcg, to);
1386 spin_unlock(&mc.lock);
1390 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1392 if (mc.moving_task && current != mc.moving_task) {
1393 if (mem_cgroup_under_move(memcg)) {
1395 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1396 /* moving charge context might have finished. */
1399 finish_wait(&mc.waitq, &wait);
1407 * Take this lock when
1408 * - a code tries to modify page's memcg while it's USED.
1409 * - a code tries to modify page state accounting in a memcg.
1410 * see mem_cgroup_stolen(), too.
1412 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1413 unsigned long *flags)
1415 spin_lock_irqsave(&memcg->move_lock, *flags);
1418 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1419 unsigned long *flags)
1421 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1424 #define K(x) ((x) << (PAGE_SHIFT-10))
1426 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1427 * @memcg: The memory cgroup that went over limit
1428 * @p: Task that is going to be killed
1430 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1433 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1435 struct cgroup *task_cgrp;
1436 struct cgroup *mem_cgrp;
1438 * Need a buffer in BSS, can't rely on allocations. The code relies
1439 * on the assumption that OOM is serialized for memory controller.
1440 * If this assumption is broken, revisit this code.
1442 static char memcg_name[PATH_MAX];
1444 struct mem_cgroup *iter;
1452 mem_cgrp = memcg->css.cgroup;
1453 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1455 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1458 * Unfortunately, we are unable to convert to a useful name
1459 * But we'll still print out the usage information
1466 pr_info("Task in %s killed", memcg_name);
1469 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1477 * Continues from above, so we don't need an KERN_ level
1479 pr_cont(" as a result of limit of %s\n", memcg_name);
1482 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1483 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1484 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1485 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1486 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1487 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1488 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1489 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1490 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1491 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1492 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1493 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1495 for_each_mem_cgroup_tree(iter, memcg) {
1496 pr_info("Memory cgroup stats");
1499 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1501 pr_cont(" for %s", memcg_name);
1505 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1506 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1508 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1509 K(mem_cgroup_read_stat(iter, i)));
1512 for (i = 0; i < NR_LRU_LISTS; i++)
1513 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1514 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1521 * This function returns the number of memcg under hierarchy tree. Returns
1522 * 1(self count) if no children.
1524 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1527 struct mem_cgroup *iter;
1529 for_each_mem_cgroup_tree(iter, memcg)
1535 * Return the memory (and swap, if configured) limit for a memcg.
1537 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1541 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1544 * Do not consider swap space if we cannot swap due to swappiness
1546 if (mem_cgroup_swappiness(memcg)) {
1549 limit += total_swap_pages << PAGE_SHIFT;
1550 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1553 * If memsw is finite and limits the amount of swap space
1554 * available to this memcg, return that limit.
1556 limit = min(limit, memsw);
1562 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1565 struct mem_cgroup *iter;
1566 unsigned long chosen_points = 0;
1567 unsigned long totalpages;
1568 unsigned int points = 0;
1569 struct task_struct *chosen = NULL;
1572 * If current has a pending SIGKILL or is exiting, then automatically
1573 * select it. The goal is to allow it to allocate so that it may
1574 * quickly exit and free its memory.
1576 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1577 set_thread_flag(TIF_MEMDIE);
1581 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1582 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1583 for_each_mem_cgroup_tree(iter, memcg) {
1584 struct css_task_iter it;
1585 struct task_struct *task;
1587 css_task_iter_start(&iter->css, &it);
1588 while ((task = css_task_iter_next(&it))) {
1589 switch (oom_scan_process_thread(task, totalpages, NULL,
1591 case OOM_SCAN_SELECT:
1593 put_task_struct(chosen);
1595 chosen_points = ULONG_MAX;
1596 get_task_struct(chosen);
1598 case OOM_SCAN_CONTINUE:
1600 case OOM_SCAN_ABORT:
1601 css_task_iter_end(&it);
1602 mem_cgroup_iter_break(memcg, iter);
1604 put_task_struct(chosen);
1609 points = oom_badness(task, memcg, NULL, totalpages);
1610 if (points > chosen_points) {
1612 put_task_struct(chosen);
1614 chosen_points = points;
1615 get_task_struct(chosen);
1618 css_task_iter_end(&it);
1623 points = chosen_points * 1000 / totalpages;
1624 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1625 NULL, "Memory cgroup out of memory");
1628 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1630 unsigned long flags)
1632 unsigned long total = 0;
1633 bool noswap = false;
1636 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1638 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1641 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1643 drain_all_stock_async(memcg);
1644 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1646 * Allow limit shrinkers, which are triggered directly
1647 * by userspace, to catch signals and stop reclaim
1648 * after minimal progress, regardless of the margin.
1650 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1652 if (mem_cgroup_margin(memcg))
1655 * If nothing was reclaimed after two attempts, there
1656 * may be no reclaimable pages in this hierarchy.
1664 #if MAX_NUMNODES > 1
1666 * test_mem_cgroup_node_reclaimable
1667 * @memcg: the target memcg
1668 * @nid: the node ID to be checked.
1669 * @noswap : specify true here if the user wants flle only information.
1671 * This function returns whether the specified memcg contains any
1672 * reclaimable pages on a node. Returns true if there are any reclaimable
1673 * pages in the node.
1675 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1676 int nid, bool noswap)
1678 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1680 if (noswap || !total_swap_pages)
1682 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1689 * Always updating the nodemask is not very good - even if we have an empty
1690 * list or the wrong list here, we can start from some node and traverse all
1691 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1694 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1698 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1699 * pagein/pageout changes since the last update.
1701 if (!atomic_read(&memcg->numainfo_events))
1703 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1706 /* make a nodemask where this memcg uses memory from */
1707 memcg->scan_nodes = node_states[N_MEMORY];
1709 for_each_node_mask(nid, node_states[N_MEMORY]) {
1711 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1712 node_clear(nid, memcg->scan_nodes);
1715 atomic_set(&memcg->numainfo_events, 0);
1716 atomic_set(&memcg->numainfo_updating, 0);
1720 * Selecting a node where we start reclaim from. Because what we need is just
1721 * reducing usage counter, start from anywhere is O,K. Considering
1722 * memory reclaim from current node, there are pros. and cons.
1724 * Freeing memory from current node means freeing memory from a node which
1725 * we'll use or we've used. So, it may make LRU bad. And if several threads
1726 * hit limits, it will see a contention on a node. But freeing from remote
1727 * node means more costs for memory reclaim because of memory latency.
1729 * Now, we use round-robin. Better algorithm is welcomed.
1731 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1735 mem_cgroup_may_update_nodemask(memcg);
1736 node = memcg->last_scanned_node;
1738 node = next_node(node, memcg->scan_nodes);
1739 if (node == MAX_NUMNODES)
1740 node = first_node(memcg->scan_nodes);
1742 * We call this when we hit limit, not when pages are added to LRU.
1743 * No LRU may hold pages because all pages are UNEVICTABLE or
1744 * memcg is too small and all pages are not on LRU. In that case,
1745 * we use curret node.
1747 if (unlikely(node == MAX_NUMNODES))
1748 node = numa_node_id();
1750 memcg->last_scanned_node = node;
1755 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1763 * A group is eligible for the soft limit reclaim under the given root
1765 * a) it is over its soft limit
1766 * b) any parent up the hierarchy is over its soft limit
1768 bool mem_cgroup_soft_reclaim_eligible(struct mem_cgroup *memcg,
1769 struct mem_cgroup *root)
1771 struct mem_cgroup *parent = memcg;
1773 if (res_counter_soft_limit_excess(&memcg->res))
1777 * If any parent up to the root in the hierarchy is over its soft limit
1778 * then we have to obey and reclaim from this group as well.
1780 while ((parent = parent_mem_cgroup(parent))) {
1781 if (res_counter_soft_limit_excess(&parent->res))
1790 static DEFINE_SPINLOCK(memcg_oom_lock);
1793 * Check OOM-Killer is already running under our hierarchy.
1794 * If someone is running, return false.
1796 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1798 struct mem_cgroup *iter, *failed = NULL;
1800 spin_lock(&memcg_oom_lock);
1802 for_each_mem_cgroup_tree(iter, memcg) {
1803 if (iter->oom_lock) {
1805 * this subtree of our hierarchy is already locked
1806 * so we cannot give a lock.
1809 mem_cgroup_iter_break(memcg, iter);
1812 iter->oom_lock = true;
1817 * OK, we failed to lock the whole subtree so we have
1818 * to clean up what we set up to the failing subtree
1820 for_each_mem_cgroup_tree(iter, memcg) {
1821 if (iter == failed) {
1822 mem_cgroup_iter_break(memcg, iter);
1825 iter->oom_lock = false;
1829 spin_unlock(&memcg_oom_lock);
1834 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1836 struct mem_cgroup *iter;
1838 spin_lock(&memcg_oom_lock);
1839 for_each_mem_cgroup_tree(iter, memcg)
1840 iter->oom_lock = false;
1841 spin_unlock(&memcg_oom_lock);
1844 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1846 struct mem_cgroup *iter;
1848 for_each_mem_cgroup_tree(iter, memcg)
1849 atomic_inc(&iter->under_oom);
1852 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1854 struct mem_cgroup *iter;
1857 * When a new child is created while the hierarchy is under oom,
1858 * mem_cgroup_oom_lock() may not be called. We have to use
1859 * atomic_add_unless() here.
1861 for_each_mem_cgroup_tree(iter, memcg)
1862 atomic_add_unless(&iter->under_oom, -1, 0);
1865 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1867 struct oom_wait_info {
1868 struct mem_cgroup *memcg;
1872 static int memcg_oom_wake_function(wait_queue_t *wait,
1873 unsigned mode, int sync, void *arg)
1875 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1876 struct mem_cgroup *oom_wait_memcg;
1877 struct oom_wait_info *oom_wait_info;
1879 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1880 oom_wait_memcg = oom_wait_info->memcg;
1883 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1884 * Then we can use css_is_ancestor without taking care of RCU.
1886 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1887 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1889 return autoremove_wake_function(wait, mode, sync, arg);
1892 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1894 atomic_inc(&memcg->oom_wakeups);
1895 /* for filtering, pass "memcg" as argument. */
1896 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1899 static void memcg_oom_recover(struct mem_cgroup *memcg)
1901 if (memcg && atomic_read(&memcg->under_oom))
1902 memcg_wakeup_oom(memcg);
1906 * try to call OOM killer
1908 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1913 if (!current->memcg_oom.may_oom)
1916 current->memcg_oom.in_memcg_oom = 1;
1919 * As with any blocking lock, a contender needs to start
1920 * listening for wakeups before attempting the trylock,
1921 * otherwise it can miss the wakeup from the unlock and sleep
1922 * indefinitely. This is just open-coded because our locking
1923 * is so particular to memcg hierarchies.
1925 wakeups = atomic_read(&memcg->oom_wakeups);
1926 mem_cgroup_mark_under_oom(memcg);
1928 locked = mem_cgroup_oom_trylock(memcg);
1931 mem_cgroup_oom_notify(memcg);
1933 if (locked && !memcg->oom_kill_disable) {
1934 mem_cgroup_unmark_under_oom(memcg);
1935 mem_cgroup_out_of_memory(memcg, mask, order);
1936 mem_cgroup_oom_unlock(memcg);
1938 * There is no guarantee that an OOM-lock contender
1939 * sees the wakeups triggered by the OOM kill
1940 * uncharges. Wake any sleepers explicitely.
1942 memcg_oom_recover(memcg);
1945 * A system call can just return -ENOMEM, but if this
1946 * is a page fault and somebody else is handling the
1947 * OOM already, we need to sleep on the OOM waitqueue
1948 * for this memcg until the situation is resolved.
1949 * Which can take some time because it might be
1950 * handled by a userspace task.
1952 * However, this is the charge context, which means
1953 * that we may sit on a large call stack and hold
1954 * various filesystem locks, the mmap_sem etc. and we
1955 * don't want the OOM handler to deadlock on them
1956 * while we sit here and wait. Store the current OOM
1957 * context in the task_struct, then return -ENOMEM.
1958 * At the end of the page fault handler, with the
1959 * stack unwound, pagefault_out_of_memory() will check
1960 * back with us by calling
1961 * mem_cgroup_oom_synchronize(), possibly putting the
1964 current->memcg_oom.oom_locked = locked;
1965 current->memcg_oom.wakeups = wakeups;
1966 css_get(&memcg->css);
1967 current->memcg_oom.wait_on_memcg = memcg;
1972 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1974 * This has to be called at the end of a page fault if the the memcg
1975 * OOM handler was enabled and the fault is returning %VM_FAULT_OOM.
1977 * Memcg supports userspace OOM handling, so failed allocations must
1978 * sleep on a waitqueue until the userspace task resolves the
1979 * situation. Sleeping directly in the charge context with all kinds
1980 * of locks held is not a good idea, instead we remember an OOM state
1981 * in the task and mem_cgroup_oom_synchronize() has to be called at
1982 * the end of the page fault to put the task to sleep and clean up the
1985 * Returns %true if an ongoing memcg OOM situation was detected and
1986 * finalized, %false otherwise.
1988 bool mem_cgroup_oom_synchronize(void)
1990 struct oom_wait_info owait;
1991 struct mem_cgroup *memcg;
1993 /* OOM is global, do not handle */
1994 if (!current->memcg_oom.in_memcg_oom)
1998 * We invoked the OOM killer but there is a chance that a kill
1999 * did not free up any charges. Everybody else might already
2000 * be sleeping, so restart the fault and keep the rampage
2001 * going until some charges are released.
2003 memcg = current->memcg_oom.wait_on_memcg;
2007 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2010 owait.memcg = memcg;
2011 owait.wait.flags = 0;
2012 owait.wait.func = memcg_oom_wake_function;
2013 owait.wait.private = current;
2014 INIT_LIST_HEAD(&owait.wait.task_list);
2016 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2017 /* Only sleep if we didn't miss any wakeups since OOM */
2018 if (atomic_read(&memcg->oom_wakeups) == current->memcg_oom.wakeups)
2020 finish_wait(&memcg_oom_waitq, &owait.wait);
2022 mem_cgroup_unmark_under_oom(memcg);
2023 if (current->memcg_oom.oom_locked) {
2024 mem_cgroup_oom_unlock(memcg);
2026 * There is no guarantee that an OOM-lock contender
2027 * sees the wakeups triggered by the OOM kill
2028 * uncharges. Wake any sleepers explicitely.
2030 memcg_oom_recover(memcg);
2032 css_put(&memcg->css);
2033 current->memcg_oom.wait_on_memcg = NULL;
2035 current->memcg_oom.in_memcg_oom = 0;
2040 * Currently used to update mapped file statistics, but the routine can be
2041 * generalized to update other statistics as well.
2043 * Notes: Race condition
2045 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2046 * it tends to be costly. But considering some conditions, we doesn't need
2047 * to do so _always_.
2049 * Considering "charge", lock_page_cgroup() is not required because all
2050 * file-stat operations happen after a page is attached to radix-tree. There
2051 * are no race with "charge".
2053 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2054 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2055 * if there are race with "uncharge". Statistics itself is properly handled
2058 * Considering "move", this is an only case we see a race. To make the race
2059 * small, we check mm->moving_account and detect there are possibility of race
2060 * If there is, we take a lock.
2063 void __mem_cgroup_begin_update_page_stat(struct page *page,
2064 bool *locked, unsigned long *flags)
2066 struct mem_cgroup *memcg;
2067 struct page_cgroup *pc;
2069 pc = lookup_page_cgroup(page);
2071 memcg = pc->mem_cgroup;
2072 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2075 * If this memory cgroup is not under account moving, we don't
2076 * need to take move_lock_mem_cgroup(). Because we already hold
2077 * rcu_read_lock(), any calls to move_account will be delayed until
2078 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2080 if (!mem_cgroup_stolen(memcg))
2083 move_lock_mem_cgroup(memcg, flags);
2084 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2085 move_unlock_mem_cgroup(memcg, flags);
2091 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2093 struct page_cgroup *pc = lookup_page_cgroup(page);
2096 * It's guaranteed that pc->mem_cgroup never changes while
2097 * lock is held because a routine modifies pc->mem_cgroup
2098 * should take move_lock_mem_cgroup().
2100 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2103 void mem_cgroup_update_page_stat(struct page *page,
2104 enum mem_cgroup_stat_index idx, int val)
2106 struct mem_cgroup *memcg;
2107 struct page_cgroup *pc = lookup_page_cgroup(page);
2108 unsigned long uninitialized_var(flags);
2110 if (mem_cgroup_disabled())
2113 VM_BUG_ON(!rcu_read_lock_held());
2114 memcg = pc->mem_cgroup;
2115 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2118 this_cpu_add(memcg->stat->count[idx], val);
2122 * size of first charge trial. "32" comes from vmscan.c's magic value.
2123 * TODO: maybe necessary to use big numbers in big irons.
2125 #define CHARGE_BATCH 32U
2126 struct memcg_stock_pcp {
2127 struct mem_cgroup *cached; /* this never be root cgroup */
2128 unsigned int nr_pages;
2129 struct work_struct work;
2130 unsigned long flags;
2131 #define FLUSHING_CACHED_CHARGE 0
2133 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2134 static DEFINE_MUTEX(percpu_charge_mutex);
2137 * consume_stock: Try to consume stocked charge on this cpu.
2138 * @memcg: memcg to consume from.
2139 * @nr_pages: how many pages to charge.
2141 * The charges will only happen if @memcg matches the current cpu's memcg
2142 * stock, and at least @nr_pages are available in that stock. Failure to
2143 * service an allocation will refill the stock.
2145 * returns true if successful, false otherwise.
2147 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2149 struct memcg_stock_pcp *stock;
2152 if (nr_pages > CHARGE_BATCH)
2155 stock = &get_cpu_var(memcg_stock);
2156 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2157 stock->nr_pages -= nr_pages;
2158 else /* need to call res_counter_charge */
2160 put_cpu_var(memcg_stock);
2165 * Returns stocks cached in percpu to res_counter and reset cached information.
2167 static void drain_stock(struct memcg_stock_pcp *stock)
2169 struct mem_cgroup *old = stock->cached;
2171 if (stock->nr_pages) {
2172 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2174 res_counter_uncharge(&old->res, bytes);
2175 if (do_swap_account)
2176 res_counter_uncharge(&old->memsw, bytes);
2177 stock->nr_pages = 0;
2179 stock->cached = NULL;
2183 * This must be called under preempt disabled or must be called by
2184 * a thread which is pinned to local cpu.
2186 static void drain_local_stock(struct work_struct *dummy)
2188 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2190 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2193 static void __init memcg_stock_init(void)
2197 for_each_possible_cpu(cpu) {
2198 struct memcg_stock_pcp *stock =
2199 &per_cpu(memcg_stock, cpu);
2200 INIT_WORK(&stock->work, drain_local_stock);
2205 * Cache charges(val) which is from res_counter, to local per_cpu area.
2206 * This will be consumed by consume_stock() function, later.
2208 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2210 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2212 if (stock->cached != memcg) { /* reset if necessary */
2214 stock->cached = memcg;
2216 stock->nr_pages += nr_pages;
2217 put_cpu_var(memcg_stock);
2221 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2222 * of the hierarchy under it. sync flag says whether we should block
2223 * until the work is done.
2225 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2229 /* Notify other cpus that system-wide "drain" is running */
2232 for_each_online_cpu(cpu) {
2233 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2234 struct mem_cgroup *memcg;
2236 memcg = stock->cached;
2237 if (!memcg || !stock->nr_pages)
2239 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2241 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2243 drain_local_stock(&stock->work);
2245 schedule_work_on(cpu, &stock->work);
2253 for_each_online_cpu(cpu) {
2254 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2255 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2256 flush_work(&stock->work);
2263 * Tries to drain stocked charges in other cpus. This function is asynchronous
2264 * and just put a work per cpu for draining localy on each cpu. Caller can
2265 * expects some charges will be back to res_counter later but cannot wait for
2268 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2271 * If someone calls draining, avoid adding more kworker runs.
2273 if (!mutex_trylock(&percpu_charge_mutex))
2275 drain_all_stock(root_memcg, false);
2276 mutex_unlock(&percpu_charge_mutex);
2279 /* This is a synchronous drain interface. */
2280 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2282 /* called when force_empty is called */
2283 mutex_lock(&percpu_charge_mutex);
2284 drain_all_stock(root_memcg, true);
2285 mutex_unlock(&percpu_charge_mutex);
2289 * This function drains percpu counter value from DEAD cpu and
2290 * move it to local cpu. Note that this function can be preempted.
2292 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2296 spin_lock(&memcg->pcp_counter_lock);
2297 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2298 long x = per_cpu(memcg->stat->count[i], cpu);
2300 per_cpu(memcg->stat->count[i], cpu) = 0;
2301 memcg->nocpu_base.count[i] += x;
2303 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2304 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2306 per_cpu(memcg->stat->events[i], cpu) = 0;
2307 memcg->nocpu_base.events[i] += x;
2309 spin_unlock(&memcg->pcp_counter_lock);
2312 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2313 unsigned long action,
2316 int cpu = (unsigned long)hcpu;
2317 struct memcg_stock_pcp *stock;
2318 struct mem_cgroup *iter;
2320 if (action == CPU_ONLINE)
2323 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2326 for_each_mem_cgroup(iter)
2327 mem_cgroup_drain_pcp_counter(iter, cpu);
2329 stock = &per_cpu(memcg_stock, cpu);
2335 /* See __mem_cgroup_try_charge() for details */
2337 CHARGE_OK, /* success */
2338 CHARGE_RETRY, /* need to retry but retry is not bad */
2339 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2340 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2343 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2344 unsigned int nr_pages, unsigned int min_pages,
2347 unsigned long csize = nr_pages * PAGE_SIZE;
2348 struct mem_cgroup *mem_over_limit;
2349 struct res_counter *fail_res;
2350 unsigned long flags = 0;
2353 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2356 if (!do_swap_account)
2358 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2362 res_counter_uncharge(&memcg->res, csize);
2363 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2364 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2366 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2368 * Never reclaim on behalf of optional batching, retry with a
2369 * single page instead.
2371 if (nr_pages > min_pages)
2372 return CHARGE_RETRY;
2374 if (!(gfp_mask & __GFP_WAIT))
2375 return CHARGE_WOULDBLOCK;
2377 if (gfp_mask & __GFP_NORETRY)
2378 return CHARGE_NOMEM;
2380 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2381 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2382 return CHARGE_RETRY;
2384 * Even though the limit is exceeded at this point, reclaim
2385 * may have been able to free some pages. Retry the charge
2386 * before killing the task.
2388 * Only for regular pages, though: huge pages are rather
2389 * unlikely to succeed so close to the limit, and we fall back
2390 * to regular pages anyway in case of failure.
2392 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2393 return CHARGE_RETRY;
2396 * At task move, charge accounts can be doubly counted. So, it's
2397 * better to wait until the end of task_move if something is going on.
2399 if (mem_cgroup_wait_acct_move(mem_over_limit))
2400 return CHARGE_RETRY;
2403 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2405 return CHARGE_NOMEM;
2409 * __mem_cgroup_try_charge() does
2410 * 1. detect memcg to be charged against from passed *mm and *ptr,
2411 * 2. update res_counter
2412 * 3. call memory reclaim if necessary.
2414 * In some special case, if the task is fatal, fatal_signal_pending() or
2415 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2416 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2417 * as possible without any hazards. 2: all pages should have a valid
2418 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2419 * pointer, that is treated as a charge to root_mem_cgroup.
2421 * So __mem_cgroup_try_charge() will return
2422 * 0 ... on success, filling *ptr with a valid memcg pointer.
2423 * -ENOMEM ... charge failure because of resource limits.
2424 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2426 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2427 * the oom-killer can be invoked.
2429 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2431 unsigned int nr_pages,
2432 struct mem_cgroup **ptr,
2435 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2436 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2437 struct mem_cgroup *memcg = NULL;
2441 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2442 * in system level. So, allow to go ahead dying process in addition to
2445 if (unlikely(test_thread_flag(TIF_MEMDIE)
2446 || fatal_signal_pending(current)))
2450 * We always charge the cgroup the mm_struct belongs to.
2451 * The mm_struct's mem_cgroup changes on task migration if the
2452 * thread group leader migrates. It's possible that mm is not
2453 * set, if so charge the root memcg (happens for pagecache usage).
2456 *ptr = root_mem_cgroup;
2458 if (*ptr) { /* css should be a valid one */
2460 if (mem_cgroup_is_root(memcg))
2462 if (consume_stock(memcg, nr_pages))
2464 css_get(&memcg->css);
2466 struct task_struct *p;
2469 p = rcu_dereference(mm->owner);
2471 * Because we don't have task_lock(), "p" can exit.
2472 * In that case, "memcg" can point to root or p can be NULL with
2473 * race with swapoff. Then, we have small risk of mis-accouning.
2474 * But such kind of mis-account by race always happens because
2475 * we don't have cgroup_mutex(). It's overkill and we allo that
2477 * (*) swapoff at el will charge against mm-struct not against
2478 * task-struct. So, mm->owner can be NULL.
2480 memcg = mem_cgroup_from_task(p);
2482 memcg = root_mem_cgroup;
2483 if (mem_cgroup_is_root(memcg)) {
2487 if (consume_stock(memcg, nr_pages)) {
2489 * It seems dagerous to access memcg without css_get().
2490 * But considering how consume_stok works, it's not
2491 * necessary. If consume_stock success, some charges
2492 * from this memcg are cached on this cpu. So, we
2493 * don't need to call css_get()/css_tryget() before
2494 * calling consume_stock().
2499 /* after here, we may be blocked. we need to get refcnt */
2500 if (!css_tryget(&memcg->css)) {
2508 bool invoke_oom = oom && !nr_oom_retries;
2510 /* If killed, bypass charge */
2511 if (fatal_signal_pending(current)) {
2512 css_put(&memcg->css);
2516 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2517 nr_pages, invoke_oom);
2521 case CHARGE_RETRY: /* not in OOM situation but retry */
2523 css_put(&memcg->css);
2526 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2527 css_put(&memcg->css);
2529 case CHARGE_NOMEM: /* OOM routine works */
2530 if (!oom || invoke_oom) {
2531 css_put(&memcg->css);
2537 } while (ret != CHARGE_OK);
2539 if (batch > nr_pages)
2540 refill_stock(memcg, batch - nr_pages);
2541 css_put(&memcg->css);
2549 *ptr = root_mem_cgroup;
2554 * Somemtimes we have to undo a charge we got by try_charge().
2555 * This function is for that and do uncharge, put css's refcnt.
2556 * gotten by try_charge().
2558 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2559 unsigned int nr_pages)
2561 if (!mem_cgroup_is_root(memcg)) {
2562 unsigned long bytes = nr_pages * PAGE_SIZE;
2564 res_counter_uncharge(&memcg->res, bytes);
2565 if (do_swap_account)
2566 res_counter_uncharge(&memcg->memsw, bytes);
2571 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2572 * This is useful when moving usage to parent cgroup.
2574 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2575 unsigned int nr_pages)
2577 unsigned long bytes = nr_pages * PAGE_SIZE;
2579 if (mem_cgroup_is_root(memcg))
2582 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2583 if (do_swap_account)
2584 res_counter_uncharge_until(&memcg->memsw,
2585 memcg->memsw.parent, bytes);
2589 * A helper function to get mem_cgroup from ID. must be called under
2590 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2591 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2592 * called against removed memcg.)
2594 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2596 struct cgroup_subsys_state *css;
2598 /* ID 0 is unused ID */
2601 css = css_lookup(&mem_cgroup_subsys, id);
2604 return mem_cgroup_from_css(css);
2607 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2609 struct mem_cgroup *memcg = NULL;
2610 struct page_cgroup *pc;
2614 VM_BUG_ON(!PageLocked(page));
2616 pc = lookup_page_cgroup(page);
2617 lock_page_cgroup(pc);
2618 if (PageCgroupUsed(pc)) {
2619 memcg = pc->mem_cgroup;
2620 if (memcg && !css_tryget(&memcg->css))
2622 } else if (PageSwapCache(page)) {
2623 ent.val = page_private(page);
2624 id = lookup_swap_cgroup_id(ent);
2626 memcg = mem_cgroup_lookup(id);
2627 if (memcg && !css_tryget(&memcg->css))
2631 unlock_page_cgroup(pc);
2635 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2637 unsigned int nr_pages,
2638 enum charge_type ctype,
2641 struct page_cgroup *pc = lookup_page_cgroup(page);
2642 struct zone *uninitialized_var(zone);
2643 struct lruvec *lruvec;
2644 bool was_on_lru = false;
2647 lock_page_cgroup(pc);
2648 VM_BUG_ON(PageCgroupUsed(pc));
2650 * we don't need page_cgroup_lock about tail pages, becase they are not
2651 * accessed by any other context at this point.
2655 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2656 * may already be on some other mem_cgroup's LRU. Take care of it.
2659 zone = page_zone(page);
2660 spin_lock_irq(&zone->lru_lock);
2661 if (PageLRU(page)) {
2662 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2664 del_page_from_lru_list(page, lruvec, page_lru(page));
2669 pc->mem_cgroup = memcg;
2671 * We access a page_cgroup asynchronously without lock_page_cgroup().
2672 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2673 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2674 * before USED bit, we need memory barrier here.
2675 * See mem_cgroup_add_lru_list(), etc.
2678 SetPageCgroupUsed(pc);
2682 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2683 VM_BUG_ON(PageLRU(page));
2685 add_page_to_lru_list(page, lruvec, page_lru(page));
2687 spin_unlock_irq(&zone->lru_lock);
2690 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2695 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2696 unlock_page_cgroup(pc);
2699 * "charge_statistics" updated event counter.
2701 memcg_check_events(memcg, page);
2704 static DEFINE_MUTEX(set_limit_mutex);
2706 #ifdef CONFIG_MEMCG_KMEM
2707 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2709 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2710 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2714 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2715 * in the memcg_cache_params struct.
2717 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2719 struct kmem_cache *cachep;
2721 VM_BUG_ON(p->is_root_cache);
2722 cachep = p->root_cache;
2723 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2726 #ifdef CONFIG_SLABINFO
2727 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2728 struct cftype *cft, struct seq_file *m)
2730 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2731 struct memcg_cache_params *params;
2733 if (!memcg_can_account_kmem(memcg))
2736 print_slabinfo_header(m);
2738 mutex_lock(&memcg->slab_caches_mutex);
2739 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2740 cache_show(memcg_params_to_cache(params), m);
2741 mutex_unlock(&memcg->slab_caches_mutex);
2747 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2749 struct res_counter *fail_res;
2750 struct mem_cgroup *_memcg;
2754 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2759 * Conditions under which we can wait for the oom_killer. Those are
2760 * the same conditions tested by the core page allocator
2762 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2765 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2768 if (ret == -EINTR) {
2770 * __mem_cgroup_try_charge() chosed to bypass to root due to
2771 * OOM kill or fatal signal. Since our only options are to
2772 * either fail the allocation or charge it to this cgroup, do
2773 * it as a temporary condition. But we can't fail. From a
2774 * kmem/slab perspective, the cache has already been selected,
2775 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2778 * This condition will only trigger if the task entered
2779 * memcg_charge_kmem in a sane state, but was OOM-killed during
2780 * __mem_cgroup_try_charge() above. Tasks that were already
2781 * dying when the allocation triggers should have been already
2782 * directed to the root cgroup in memcontrol.h
2784 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2785 if (do_swap_account)
2786 res_counter_charge_nofail(&memcg->memsw, size,
2790 res_counter_uncharge(&memcg->kmem, size);
2795 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2797 res_counter_uncharge(&memcg->res, size);
2798 if (do_swap_account)
2799 res_counter_uncharge(&memcg->memsw, size);
2802 if (res_counter_uncharge(&memcg->kmem, size))
2806 * Releases a reference taken in kmem_cgroup_css_offline in case
2807 * this last uncharge is racing with the offlining code or it is
2808 * outliving the memcg existence.
2810 * The memory barrier imposed by test&clear is paired with the
2811 * explicit one in memcg_kmem_mark_dead().
2813 if (memcg_kmem_test_and_clear_dead(memcg))
2814 css_put(&memcg->css);
2817 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2822 mutex_lock(&memcg->slab_caches_mutex);
2823 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2824 mutex_unlock(&memcg->slab_caches_mutex);
2828 * helper for acessing a memcg's index. It will be used as an index in the
2829 * child cache array in kmem_cache, and also to derive its name. This function
2830 * will return -1 when this is not a kmem-limited memcg.
2832 int memcg_cache_id(struct mem_cgroup *memcg)
2834 return memcg ? memcg->kmemcg_id : -1;
2838 * This ends up being protected by the set_limit mutex, during normal
2839 * operation, because that is its main call site.
2841 * But when we create a new cache, we can call this as well if its parent
2842 * is kmem-limited. That will have to hold set_limit_mutex as well.
2844 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2848 num = ida_simple_get(&kmem_limited_groups,
2849 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2853 * After this point, kmem_accounted (that we test atomically in
2854 * the beginning of this conditional), is no longer 0. This
2855 * guarantees only one process will set the following boolean
2856 * to true. We don't need test_and_set because we're protected
2857 * by the set_limit_mutex anyway.
2859 memcg_kmem_set_activated(memcg);
2861 ret = memcg_update_all_caches(num+1);
2863 ida_simple_remove(&kmem_limited_groups, num);
2864 memcg_kmem_clear_activated(memcg);
2868 memcg->kmemcg_id = num;
2869 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2870 mutex_init(&memcg->slab_caches_mutex);
2874 static size_t memcg_caches_array_size(int num_groups)
2877 if (num_groups <= 0)
2880 size = 2 * num_groups;
2881 if (size < MEMCG_CACHES_MIN_SIZE)
2882 size = MEMCG_CACHES_MIN_SIZE;
2883 else if (size > MEMCG_CACHES_MAX_SIZE)
2884 size = MEMCG_CACHES_MAX_SIZE;
2890 * We should update the current array size iff all caches updates succeed. This
2891 * can only be done from the slab side. The slab mutex needs to be held when
2894 void memcg_update_array_size(int num)
2896 if (num > memcg_limited_groups_array_size)
2897 memcg_limited_groups_array_size = memcg_caches_array_size(num);
2900 static void kmem_cache_destroy_work_func(struct work_struct *w);
2902 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
2904 struct memcg_cache_params *cur_params = s->memcg_params;
2906 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
2908 if (num_groups > memcg_limited_groups_array_size) {
2910 ssize_t size = memcg_caches_array_size(num_groups);
2912 size *= sizeof(void *);
2913 size += offsetof(struct memcg_cache_params, memcg_caches);
2915 s->memcg_params = kzalloc(size, GFP_KERNEL);
2916 if (!s->memcg_params) {
2917 s->memcg_params = cur_params;
2921 s->memcg_params->is_root_cache = true;
2924 * There is the chance it will be bigger than
2925 * memcg_limited_groups_array_size, if we failed an allocation
2926 * in a cache, in which case all caches updated before it, will
2927 * have a bigger array.
2929 * But if that is the case, the data after
2930 * memcg_limited_groups_array_size is certainly unused
2932 for (i = 0; i < memcg_limited_groups_array_size; i++) {
2933 if (!cur_params->memcg_caches[i])
2935 s->memcg_params->memcg_caches[i] =
2936 cur_params->memcg_caches[i];
2940 * Ideally, we would wait until all caches succeed, and only
2941 * then free the old one. But this is not worth the extra
2942 * pointer per-cache we'd have to have for this.
2944 * It is not a big deal if some caches are left with a size
2945 * bigger than the others. And all updates will reset this
2953 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
2954 struct kmem_cache *root_cache)
2958 if (!memcg_kmem_enabled())
2962 size = offsetof(struct memcg_cache_params, memcg_caches);
2963 size += memcg_limited_groups_array_size * sizeof(void *);
2965 size = sizeof(struct memcg_cache_params);
2967 s->memcg_params = kzalloc(size, GFP_KERNEL);
2968 if (!s->memcg_params)
2972 s->memcg_params->memcg = memcg;
2973 s->memcg_params->root_cache = root_cache;
2974 INIT_WORK(&s->memcg_params->destroy,
2975 kmem_cache_destroy_work_func);
2977 s->memcg_params->is_root_cache = true;
2982 void memcg_release_cache(struct kmem_cache *s)
2984 struct kmem_cache *root;
2985 struct mem_cgroup *memcg;
2989 * This happens, for instance, when a root cache goes away before we
2992 if (!s->memcg_params)
2995 if (s->memcg_params->is_root_cache)
2998 memcg = s->memcg_params->memcg;
2999 id = memcg_cache_id(memcg);
3001 root = s->memcg_params->root_cache;
3002 root->memcg_params->memcg_caches[id] = NULL;
3004 mutex_lock(&memcg->slab_caches_mutex);
3005 list_del(&s->memcg_params->list);
3006 mutex_unlock(&memcg->slab_caches_mutex);
3008 css_put(&memcg->css);
3010 kfree(s->memcg_params);
3014 * During the creation a new cache, we need to disable our accounting mechanism
3015 * altogether. This is true even if we are not creating, but rather just
3016 * enqueing new caches to be created.
3018 * This is because that process will trigger allocations; some visible, like
3019 * explicit kmallocs to auxiliary data structures, name strings and internal
3020 * cache structures; some well concealed, like INIT_WORK() that can allocate
3021 * objects during debug.
3023 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3024 * to it. This may not be a bounded recursion: since the first cache creation
3025 * failed to complete (waiting on the allocation), we'll just try to create the
3026 * cache again, failing at the same point.
3028 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3029 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3030 * inside the following two functions.
3032 static inline void memcg_stop_kmem_account(void)
3034 VM_BUG_ON(!current->mm);
3035 current->memcg_kmem_skip_account++;
3038 static inline void memcg_resume_kmem_account(void)
3040 VM_BUG_ON(!current->mm);
3041 current->memcg_kmem_skip_account--;
3044 static void kmem_cache_destroy_work_func(struct work_struct *w)
3046 struct kmem_cache *cachep;
3047 struct memcg_cache_params *p;
3049 p = container_of(w, struct memcg_cache_params, destroy);
3051 cachep = memcg_params_to_cache(p);
3054 * If we get down to 0 after shrink, we could delete right away.
3055 * However, memcg_release_pages() already puts us back in the workqueue
3056 * in that case. If we proceed deleting, we'll get a dangling
3057 * reference, and removing the object from the workqueue in that case
3058 * is unnecessary complication. We are not a fast path.
3060 * Note that this case is fundamentally different from racing with
3061 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3062 * kmem_cache_shrink, not only we would be reinserting a dead cache
3063 * into the queue, but doing so from inside the worker racing to
3066 * So if we aren't down to zero, we'll just schedule a worker and try
3069 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3070 kmem_cache_shrink(cachep);
3071 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3074 kmem_cache_destroy(cachep);
3077 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3079 if (!cachep->memcg_params->dead)
3083 * There are many ways in which we can get here.
3085 * We can get to a memory-pressure situation while the delayed work is
3086 * still pending to run. The vmscan shrinkers can then release all
3087 * cache memory and get us to destruction. If this is the case, we'll
3088 * be executed twice, which is a bug (the second time will execute over
3089 * bogus data). In this case, cancelling the work should be fine.
3091 * But we can also get here from the worker itself, if
3092 * kmem_cache_shrink is enough to shake all the remaining objects and
3093 * get the page count to 0. In this case, we'll deadlock if we try to
3094 * cancel the work (the worker runs with an internal lock held, which
3095 * is the same lock we would hold for cancel_work_sync().)
3097 * Since we can't possibly know who got us here, just refrain from
3098 * running if there is already work pending
3100 if (work_pending(&cachep->memcg_params->destroy))
3103 * We have to defer the actual destroying to a workqueue, because
3104 * we might currently be in a context that cannot sleep.
3106 schedule_work(&cachep->memcg_params->destroy);
3110 * This lock protects updaters, not readers. We want readers to be as fast as
3111 * they can, and they will either see NULL or a valid cache value. Our model
3112 * allow them to see NULL, in which case the root memcg will be selected.
3114 * We need this lock because multiple allocations to the same cache from a non
3115 * will span more than one worker. Only one of them can create the cache.
3117 static DEFINE_MUTEX(memcg_cache_mutex);
3120 * Called with memcg_cache_mutex held
3122 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3123 struct kmem_cache *s)
3125 struct kmem_cache *new;
3126 static char *tmp_name = NULL;
3128 lockdep_assert_held(&memcg_cache_mutex);
3131 * kmem_cache_create_memcg duplicates the given name and
3132 * cgroup_name for this name requires RCU context.
3133 * This static temporary buffer is used to prevent from
3134 * pointless shortliving allocation.
3137 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3143 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3144 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3147 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3148 (s->flags & ~SLAB_PANIC), s->ctor, s);
3151 new->allocflags |= __GFP_KMEMCG;
3156 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3157 struct kmem_cache *cachep)
3159 struct kmem_cache *new_cachep;
3162 BUG_ON(!memcg_can_account_kmem(memcg));
3164 idx = memcg_cache_id(memcg);
3166 mutex_lock(&memcg_cache_mutex);
3167 new_cachep = cachep->memcg_params->memcg_caches[idx];
3169 css_put(&memcg->css);
3173 new_cachep = kmem_cache_dup(memcg, cachep);
3174 if (new_cachep == NULL) {
3175 new_cachep = cachep;
3176 css_put(&memcg->css);
3180 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3182 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3184 * the readers won't lock, make sure everybody sees the updated value,
3185 * so they won't put stuff in the queue again for no reason
3189 mutex_unlock(&memcg_cache_mutex);
3193 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3195 struct kmem_cache *c;
3198 if (!s->memcg_params)
3200 if (!s->memcg_params->is_root_cache)
3204 * If the cache is being destroyed, we trust that there is no one else
3205 * requesting objects from it. Even if there are, the sanity checks in
3206 * kmem_cache_destroy should caught this ill-case.
3208 * Still, we don't want anyone else freeing memcg_caches under our
3209 * noses, which can happen if a new memcg comes to life. As usual,
3210 * we'll take the set_limit_mutex to protect ourselves against this.
3212 mutex_lock(&set_limit_mutex);
3213 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3214 c = s->memcg_params->memcg_caches[i];
3219 * We will now manually delete the caches, so to avoid races
3220 * we need to cancel all pending destruction workers and
3221 * proceed with destruction ourselves.
3223 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3224 * and that could spawn the workers again: it is likely that
3225 * the cache still have active pages until this very moment.
3226 * This would lead us back to mem_cgroup_destroy_cache.
3228 * But that will not execute at all if the "dead" flag is not
3229 * set, so flip it down to guarantee we are in control.
3231 c->memcg_params->dead = false;
3232 cancel_work_sync(&c->memcg_params->destroy);
3233 kmem_cache_destroy(c);
3235 mutex_unlock(&set_limit_mutex);
3238 struct create_work {
3239 struct mem_cgroup *memcg;
3240 struct kmem_cache *cachep;
3241 struct work_struct work;
3244 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3246 struct kmem_cache *cachep;
3247 struct memcg_cache_params *params;
3249 if (!memcg_kmem_is_active(memcg))
3252 mutex_lock(&memcg->slab_caches_mutex);
3253 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3254 cachep = memcg_params_to_cache(params);
3255 cachep->memcg_params->dead = true;
3256 schedule_work(&cachep->memcg_params->destroy);
3258 mutex_unlock(&memcg->slab_caches_mutex);
3261 static void memcg_create_cache_work_func(struct work_struct *w)
3263 struct create_work *cw;
3265 cw = container_of(w, struct create_work, work);
3266 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3271 * Enqueue the creation of a per-memcg kmem_cache.
3273 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3274 struct kmem_cache *cachep)
3276 struct create_work *cw;
3278 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3280 css_put(&memcg->css);
3285 cw->cachep = cachep;
3287 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3288 schedule_work(&cw->work);
3291 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3292 struct kmem_cache *cachep)
3295 * We need to stop accounting when we kmalloc, because if the
3296 * corresponding kmalloc cache is not yet created, the first allocation
3297 * in __memcg_create_cache_enqueue will recurse.
3299 * However, it is better to enclose the whole function. Depending on
3300 * the debugging options enabled, INIT_WORK(), for instance, can
3301 * trigger an allocation. This too, will make us recurse. Because at
3302 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3303 * the safest choice is to do it like this, wrapping the whole function.
3305 memcg_stop_kmem_account();
3306 __memcg_create_cache_enqueue(memcg, cachep);
3307 memcg_resume_kmem_account();
3310 * Return the kmem_cache we're supposed to use for a slab allocation.
3311 * We try to use the current memcg's version of the cache.
3313 * If the cache does not exist yet, if we are the first user of it,
3314 * we either create it immediately, if possible, or create it asynchronously
3316 * In the latter case, we will let the current allocation go through with
3317 * the original cache.
3319 * Can't be called in interrupt context or from kernel threads.
3320 * This function needs to be called with rcu_read_lock() held.
3322 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3325 struct mem_cgroup *memcg;
3328 VM_BUG_ON(!cachep->memcg_params);
3329 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3331 if (!current->mm || current->memcg_kmem_skip_account)
3335 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3337 if (!memcg_can_account_kmem(memcg))
3340 idx = memcg_cache_id(memcg);
3343 * barrier to mare sure we're always seeing the up to date value. The
3344 * code updating memcg_caches will issue a write barrier to match this.
3346 read_barrier_depends();
3347 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3348 cachep = cachep->memcg_params->memcg_caches[idx];
3352 /* The corresponding put will be done in the workqueue. */
3353 if (!css_tryget(&memcg->css))
3358 * If we are in a safe context (can wait, and not in interrupt
3359 * context), we could be be predictable and return right away.
3360 * This would guarantee that the allocation being performed
3361 * already belongs in the new cache.
3363 * However, there are some clashes that can arrive from locking.
3364 * For instance, because we acquire the slab_mutex while doing
3365 * kmem_cache_dup, this means no further allocation could happen
3366 * with the slab_mutex held.
3368 * Also, because cache creation issue get_online_cpus(), this
3369 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3370 * that ends up reversed during cpu hotplug. (cpuset allocates
3371 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3372 * better to defer everything.
3374 memcg_create_cache_enqueue(memcg, cachep);
3380 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3383 * We need to verify if the allocation against current->mm->owner's memcg is
3384 * possible for the given order. But the page is not allocated yet, so we'll
3385 * need a further commit step to do the final arrangements.
3387 * It is possible for the task to switch cgroups in this mean time, so at
3388 * commit time, we can't rely on task conversion any longer. We'll then use
3389 * the handle argument to return to the caller which cgroup we should commit
3390 * against. We could also return the memcg directly and avoid the pointer
3391 * passing, but a boolean return value gives better semantics considering
3392 * the compiled-out case as well.
3394 * Returning true means the allocation is possible.
3397 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3399 struct mem_cgroup *memcg;
3405 * Disabling accounting is only relevant for some specific memcg
3406 * internal allocations. Therefore we would initially not have such
3407 * check here, since direct calls to the page allocator that are marked
3408 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3409 * concerned with cache allocations, and by having this test at
3410 * memcg_kmem_get_cache, we are already able to relay the allocation to
3411 * the root cache and bypass the memcg cache altogether.
3413 * There is one exception, though: the SLUB allocator does not create
3414 * large order caches, but rather service large kmallocs directly from
3415 * the page allocator. Therefore, the following sequence when backed by
3416 * the SLUB allocator:
3418 * memcg_stop_kmem_account();
3419 * kmalloc(<large_number>)
3420 * memcg_resume_kmem_account();
3422 * would effectively ignore the fact that we should skip accounting,
3423 * since it will drive us directly to this function without passing
3424 * through the cache selector memcg_kmem_get_cache. Such large
3425 * allocations are extremely rare but can happen, for instance, for the
3426 * cache arrays. We bring this test here.
3428 if (!current->mm || current->memcg_kmem_skip_account)
3431 memcg = try_get_mem_cgroup_from_mm(current->mm);
3434 * very rare case described in mem_cgroup_from_task. Unfortunately there
3435 * isn't much we can do without complicating this too much, and it would
3436 * be gfp-dependent anyway. Just let it go
3438 if (unlikely(!memcg))
3441 if (!memcg_can_account_kmem(memcg)) {
3442 css_put(&memcg->css);
3446 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3450 css_put(&memcg->css);
3454 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3457 struct page_cgroup *pc;
3459 VM_BUG_ON(mem_cgroup_is_root(memcg));
3461 /* The page allocation failed. Revert */
3463 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3467 pc = lookup_page_cgroup(page);
3468 lock_page_cgroup(pc);
3469 pc->mem_cgroup = memcg;
3470 SetPageCgroupUsed(pc);
3471 unlock_page_cgroup(pc);
3474 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3476 struct mem_cgroup *memcg = NULL;
3477 struct page_cgroup *pc;
3480 pc = lookup_page_cgroup(page);
3482 * Fast unlocked return. Theoretically might have changed, have to
3483 * check again after locking.
3485 if (!PageCgroupUsed(pc))
3488 lock_page_cgroup(pc);
3489 if (PageCgroupUsed(pc)) {
3490 memcg = pc->mem_cgroup;
3491 ClearPageCgroupUsed(pc);
3493 unlock_page_cgroup(pc);
3496 * We trust that only if there is a memcg associated with the page, it
3497 * is a valid allocation
3502 VM_BUG_ON(mem_cgroup_is_root(memcg));
3503 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3506 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3509 #endif /* CONFIG_MEMCG_KMEM */
3511 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3513 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3515 * Because tail pages are not marked as "used", set it. We're under
3516 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3517 * charge/uncharge will be never happen and move_account() is done under
3518 * compound_lock(), so we don't have to take care of races.
3520 void mem_cgroup_split_huge_fixup(struct page *head)
3522 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3523 struct page_cgroup *pc;
3524 struct mem_cgroup *memcg;
3527 if (mem_cgroup_disabled())
3530 memcg = head_pc->mem_cgroup;
3531 for (i = 1; i < HPAGE_PMD_NR; i++) {
3533 pc->mem_cgroup = memcg;
3534 smp_wmb();/* see __commit_charge() */
3535 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3537 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3540 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3543 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3544 struct mem_cgroup *to,
3545 unsigned int nr_pages,
3546 enum mem_cgroup_stat_index idx)
3548 /* Update stat data for mem_cgroup */
3550 WARN_ON_ONCE(from->stat->count[idx] < nr_pages);
3551 __this_cpu_add(from->stat->count[idx], -nr_pages);
3552 __this_cpu_add(to->stat->count[idx], nr_pages);
3557 * mem_cgroup_move_account - move account of the page
3559 * @nr_pages: number of regular pages (>1 for huge pages)
3560 * @pc: page_cgroup of the page.
3561 * @from: mem_cgroup which the page is moved from.
3562 * @to: mem_cgroup which the page is moved to. @from != @to.
3564 * The caller must confirm following.
3565 * - page is not on LRU (isolate_page() is useful.)
3566 * - compound_lock is held when nr_pages > 1
3568 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3571 static int mem_cgroup_move_account(struct page *page,
3572 unsigned int nr_pages,
3573 struct page_cgroup *pc,
3574 struct mem_cgroup *from,
3575 struct mem_cgroup *to)
3577 unsigned long flags;
3579 bool anon = PageAnon(page);
3581 VM_BUG_ON(from == to);
3582 VM_BUG_ON(PageLRU(page));
3584 * The page is isolated from LRU. So, collapse function
3585 * will not handle this page. But page splitting can happen.
3586 * Do this check under compound_page_lock(). The caller should
3590 if (nr_pages > 1 && !PageTransHuge(page))
3593 lock_page_cgroup(pc);
3596 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3599 move_lock_mem_cgroup(from, &flags);
3601 if (!anon && page_mapped(page))
3602 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3603 MEM_CGROUP_STAT_FILE_MAPPED);
3605 if (PageWriteback(page))
3606 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3607 MEM_CGROUP_STAT_WRITEBACK);
3609 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3611 /* caller should have done css_get */
3612 pc->mem_cgroup = to;
3613 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3614 move_unlock_mem_cgroup(from, &flags);
3617 unlock_page_cgroup(pc);
3621 memcg_check_events(to, page);
3622 memcg_check_events(from, page);
3628 * mem_cgroup_move_parent - moves page to the parent group
3629 * @page: the page to move
3630 * @pc: page_cgroup of the page
3631 * @child: page's cgroup
3633 * move charges to its parent or the root cgroup if the group has no
3634 * parent (aka use_hierarchy==0).
3635 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3636 * mem_cgroup_move_account fails) the failure is always temporary and
3637 * it signals a race with a page removal/uncharge or migration. In the
3638 * first case the page is on the way out and it will vanish from the LRU
3639 * on the next attempt and the call should be retried later.
3640 * Isolation from the LRU fails only if page has been isolated from
3641 * the LRU since we looked at it and that usually means either global
3642 * reclaim or migration going on. The page will either get back to the
3644 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3645 * (!PageCgroupUsed) or moved to a different group. The page will
3646 * disappear in the next attempt.
3648 static int mem_cgroup_move_parent(struct page *page,
3649 struct page_cgroup *pc,
3650 struct mem_cgroup *child)
3652 struct mem_cgroup *parent;
3653 unsigned int nr_pages;
3654 unsigned long uninitialized_var(flags);
3657 VM_BUG_ON(mem_cgroup_is_root(child));
3660 if (!get_page_unless_zero(page))
3662 if (isolate_lru_page(page))
3665 nr_pages = hpage_nr_pages(page);
3667 parent = parent_mem_cgroup(child);
3669 * If no parent, move charges to root cgroup.
3672 parent = root_mem_cgroup;
3675 VM_BUG_ON(!PageTransHuge(page));
3676 flags = compound_lock_irqsave(page);
3679 ret = mem_cgroup_move_account(page, nr_pages,
3682 __mem_cgroup_cancel_local_charge(child, nr_pages);
3685 compound_unlock_irqrestore(page, flags);
3686 putback_lru_page(page);
3694 * Charge the memory controller for page usage.
3696 * 0 if the charge was successful
3697 * < 0 if the cgroup is over its limit
3699 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3700 gfp_t gfp_mask, enum charge_type ctype)
3702 struct mem_cgroup *memcg = NULL;
3703 unsigned int nr_pages = 1;
3707 if (PageTransHuge(page)) {
3708 nr_pages <<= compound_order(page);
3709 VM_BUG_ON(!PageTransHuge(page));
3711 * Never OOM-kill a process for a huge page. The
3712 * fault handler will fall back to regular pages.
3717 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3720 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3724 int mem_cgroup_newpage_charge(struct page *page,
3725 struct mm_struct *mm, gfp_t gfp_mask)
3727 if (mem_cgroup_disabled())
3729 VM_BUG_ON(page_mapped(page));
3730 VM_BUG_ON(page->mapping && !PageAnon(page));
3732 return mem_cgroup_charge_common(page, mm, gfp_mask,
3733 MEM_CGROUP_CHARGE_TYPE_ANON);
3737 * While swap-in, try_charge -> commit or cancel, the page is locked.
3738 * And when try_charge() successfully returns, one refcnt to memcg without
3739 * struct page_cgroup is acquired. This refcnt will be consumed by
3740 * "commit()" or removed by "cancel()"
3742 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3745 struct mem_cgroup **memcgp)
3747 struct mem_cgroup *memcg;
3748 struct page_cgroup *pc;
3751 pc = lookup_page_cgroup(page);
3753 * Every swap fault against a single page tries to charge the
3754 * page, bail as early as possible. shmem_unuse() encounters
3755 * already charged pages, too. The USED bit is protected by
3756 * the page lock, which serializes swap cache removal, which
3757 * in turn serializes uncharging.
3759 if (PageCgroupUsed(pc))
3761 if (!do_swap_account)
3763 memcg = try_get_mem_cgroup_from_page(page);
3767 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3768 css_put(&memcg->css);
3773 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3779 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3780 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3783 if (mem_cgroup_disabled())
3786 * A racing thread's fault, or swapoff, may have already
3787 * updated the pte, and even removed page from swap cache: in
3788 * those cases unuse_pte()'s pte_same() test will fail; but
3789 * there's also a KSM case which does need to charge the page.
3791 if (!PageSwapCache(page)) {
3794 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3799 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3802 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3804 if (mem_cgroup_disabled())
3808 __mem_cgroup_cancel_charge(memcg, 1);
3812 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3813 enum charge_type ctype)
3815 if (mem_cgroup_disabled())
3820 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3822 * Now swap is on-memory. This means this page may be
3823 * counted both as mem and swap....double count.
3824 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3825 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3826 * may call delete_from_swap_cache() before reach here.
3828 if (do_swap_account && PageSwapCache(page)) {
3829 swp_entry_t ent = {.val = page_private(page)};
3830 mem_cgroup_uncharge_swap(ent);
3834 void mem_cgroup_commit_charge_swapin(struct page *page,
3835 struct mem_cgroup *memcg)
3837 __mem_cgroup_commit_charge_swapin(page, memcg,
3838 MEM_CGROUP_CHARGE_TYPE_ANON);
3841 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3844 struct mem_cgroup *memcg = NULL;
3845 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3848 if (mem_cgroup_disabled())
3850 if (PageCompound(page))
3853 if (!PageSwapCache(page))
3854 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3855 else { /* page is swapcache/shmem */
3856 ret = __mem_cgroup_try_charge_swapin(mm, page,
3859 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3864 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3865 unsigned int nr_pages,
3866 const enum charge_type ctype)
3868 struct memcg_batch_info *batch = NULL;
3869 bool uncharge_memsw = true;
3871 /* If swapout, usage of swap doesn't decrease */
3872 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3873 uncharge_memsw = false;
3875 batch = ¤t->memcg_batch;
3877 * In usual, we do css_get() when we remember memcg pointer.
3878 * But in this case, we keep res->usage until end of a series of
3879 * uncharges. Then, it's ok to ignore memcg's refcnt.
3882 batch->memcg = memcg;
3884 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3885 * In those cases, all pages freed continuously can be expected to be in
3886 * the same cgroup and we have chance to coalesce uncharges.
3887 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3888 * because we want to do uncharge as soon as possible.
3891 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3892 goto direct_uncharge;
3895 goto direct_uncharge;
3898 * In typical case, batch->memcg == mem. This means we can
3899 * merge a series of uncharges to an uncharge of res_counter.
3900 * If not, we uncharge res_counter ony by one.
3902 if (batch->memcg != memcg)
3903 goto direct_uncharge;
3904 /* remember freed charge and uncharge it later */
3907 batch->memsw_nr_pages++;
3910 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3912 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3913 if (unlikely(batch->memcg != memcg))
3914 memcg_oom_recover(memcg);
3918 * uncharge if !page_mapped(page)
3920 static struct mem_cgroup *
3921 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3924 struct mem_cgroup *memcg = NULL;
3925 unsigned int nr_pages = 1;
3926 struct page_cgroup *pc;
3929 if (mem_cgroup_disabled())
3932 if (PageTransHuge(page)) {
3933 nr_pages <<= compound_order(page);
3934 VM_BUG_ON(!PageTransHuge(page));
3937 * Check if our page_cgroup is valid
3939 pc = lookup_page_cgroup(page);
3940 if (unlikely(!PageCgroupUsed(pc)))
3943 lock_page_cgroup(pc);
3945 memcg = pc->mem_cgroup;
3947 if (!PageCgroupUsed(pc))
3950 anon = PageAnon(page);
3953 case MEM_CGROUP_CHARGE_TYPE_ANON:
3955 * Generally PageAnon tells if it's the anon statistics to be
3956 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3957 * used before page reached the stage of being marked PageAnon.
3961 case MEM_CGROUP_CHARGE_TYPE_DROP:
3962 /* See mem_cgroup_prepare_migration() */
3963 if (page_mapped(page))
3966 * Pages under migration may not be uncharged. But
3967 * end_migration() /must/ be the one uncharging the
3968 * unused post-migration page and so it has to call
3969 * here with the migration bit still set. See the
3970 * res_counter handling below.
3972 if (!end_migration && PageCgroupMigration(pc))
3975 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3976 if (!PageAnon(page)) { /* Shared memory */
3977 if (page->mapping && !page_is_file_cache(page))
3979 } else if (page_mapped(page)) /* Anon */
3986 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
3988 ClearPageCgroupUsed(pc);
3990 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3991 * freed from LRU. This is safe because uncharged page is expected not
3992 * to be reused (freed soon). Exception is SwapCache, it's handled by
3993 * special functions.
3996 unlock_page_cgroup(pc);
3998 * even after unlock, we have memcg->res.usage here and this memcg
3999 * will never be freed, so it's safe to call css_get().
4001 memcg_check_events(memcg, page);
4002 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4003 mem_cgroup_swap_statistics(memcg, true);
4004 css_get(&memcg->css);
4007 * Migration does not charge the res_counter for the
4008 * replacement page, so leave it alone when phasing out the
4009 * page that is unused after the migration.
4011 if (!end_migration && !mem_cgroup_is_root(memcg))
4012 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4017 unlock_page_cgroup(pc);
4021 void mem_cgroup_uncharge_page(struct page *page)
4024 if (page_mapped(page))
4026 VM_BUG_ON(page->mapping && !PageAnon(page));
4028 * If the page is in swap cache, uncharge should be deferred
4029 * to the swap path, which also properly accounts swap usage
4030 * and handles memcg lifetime.
4032 * Note that this check is not stable and reclaim may add the
4033 * page to swap cache at any time after this. However, if the
4034 * page is not in swap cache by the time page->mapcount hits
4035 * 0, there won't be any page table references to the swap
4036 * slot, and reclaim will free it and not actually write the
4039 if (PageSwapCache(page))
4041 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4044 void mem_cgroup_uncharge_cache_page(struct page *page)
4046 VM_BUG_ON(page_mapped(page));
4047 VM_BUG_ON(page->mapping);
4048 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4052 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4053 * In that cases, pages are freed continuously and we can expect pages
4054 * are in the same memcg. All these calls itself limits the number of
4055 * pages freed at once, then uncharge_start/end() is called properly.
4056 * This may be called prural(2) times in a context,
4059 void mem_cgroup_uncharge_start(void)
4061 current->memcg_batch.do_batch++;
4062 /* We can do nest. */
4063 if (current->memcg_batch.do_batch == 1) {
4064 current->memcg_batch.memcg = NULL;
4065 current->memcg_batch.nr_pages = 0;
4066 current->memcg_batch.memsw_nr_pages = 0;
4070 void mem_cgroup_uncharge_end(void)
4072 struct memcg_batch_info *batch = ¤t->memcg_batch;
4074 if (!batch->do_batch)
4078 if (batch->do_batch) /* If stacked, do nothing. */
4084 * This "batch->memcg" is valid without any css_get/put etc...
4085 * bacause we hide charges behind us.
4087 if (batch->nr_pages)
4088 res_counter_uncharge(&batch->memcg->res,
4089 batch->nr_pages * PAGE_SIZE);
4090 if (batch->memsw_nr_pages)
4091 res_counter_uncharge(&batch->memcg->memsw,
4092 batch->memsw_nr_pages * PAGE_SIZE);
4093 memcg_oom_recover(batch->memcg);
4094 /* forget this pointer (for sanity check) */
4095 batch->memcg = NULL;
4100 * called after __delete_from_swap_cache() and drop "page" account.
4101 * memcg information is recorded to swap_cgroup of "ent"
4104 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4106 struct mem_cgroup *memcg;
4107 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4109 if (!swapout) /* this was a swap cache but the swap is unused ! */
4110 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4112 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4115 * record memcg information, if swapout && memcg != NULL,
4116 * css_get() was called in uncharge().
4118 if (do_swap_account && swapout && memcg)
4119 swap_cgroup_record(ent, css_id(&memcg->css));
4123 #ifdef CONFIG_MEMCG_SWAP
4125 * called from swap_entry_free(). remove record in swap_cgroup and
4126 * uncharge "memsw" account.
4128 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4130 struct mem_cgroup *memcg;
4133 if (!do_swap_account)
4136 id = swap_cgroup_record(ent, 0);
4138 memcg = mem_cgroup_lookup(id);
4141 * We uncharge this because swap is freed.
4142 * This memcg can be obsolete one. We avoid calling css_tryget
4144 if (!mem_cgroup_is_root(memcg))
4145 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4146 mem_cgroup_swap_statistics(memcg, false);
4147 css_put(&memcg->css);
4153 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4154 * @entry: swap entry to be moved
4155 * @from: mem_cgroup which the entry is moved from
4156 * @to: mem_cgroup which the entry is moved to
4158 * It succeeds only when the swap_cgroup's record for this entry is the same
4159 * as the mem_cgroup's id of @from.
4161 * Returns 0 on success, -EINVAL on failure.
4163 * The caller must have charged to @to, IOW, called res_counter_charge() about
4164 * both res and memsw, and called css_get().
4166 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4167 struct mem_cgroup *from, struct mem_cgroup *to)
4169 unsigned short old_id, new_id;
4171 old_id = css_id(&from->css);
4172 new_id = css_id(&to->css);
4174 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4175 mem_cgroup_swap_statistics(from, false);
4176 mem_cgroup_swap_statistics(to, true);
4178 * This function is only called from task migration context now.
4179 * It postpones res_counter and refcount handling till the end
4180 * of task migration(mem_cgroup_clear_mc()) for performance
4181 * improvement. But we cannot postpone css_get(to) because if
4182 * the process that has been moved to @to does swap-in, the
4183 * refcount of @to might be decreased to 0.
4185 * We are in attach() phase, so the cgroup is guaranteed to be
4186 * alive, so we can just call css_get().
4194 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4195 struct mem_cgroup *from, struct mem_cgroup *to)
4202 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4205 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4206 struct mem_cgroup **memcgp)
4208 struct mem_cgroup *memcg = NULL;
4209 unsigned int nr_pages = 1;
4210 struct page_cgroup *pc;
4211 enum charge_type ctype;
4215 if (mem_cgroup_disabled())
4218 if (PageTransHuge(page))
4219 nr_pages <<= compound_order(page);
4221 pc = lookup_page_cgroup(page);
4222 lock_page_cgroup(pc);
4223 if (PageCgroupUsed(pc)) {
4224 memcg = pc->mem_cgroup;
4225 css_get(&memcg->css);
4227 * At migrating an anonymous page, its mapcount goes down
4228 * to 0 and uncharge() will be called. But, even if it's fully
4229 * unmapped, migration may fail and this page has to be
4230 * charged again. We set MIGRATION flag here and delay uncharge
4231 * until end_migration() is called
4233 * Corner Case Thinking
4235 * When the old page was mapped as Anon and it's unmap-and-freed
4236 * while migration was ongoing.
4237 * If unmap finds the old page, uncharge() of it will be delayed
4238 * until end_migration(). If unmap finds a new page, it's
4239 * uncharged when it make mapcount to be 1->0. If unmap code
4240 * finds swap_migration_entry, the new page will not be mapped
4241 * and end_migration() will find it(mapcount==0).
4244 * When the old page was mapped but migraion fails, the kernel
4245 * remaps it. A charge for it is kept by MIGRATION flag even
4246 * if mapcount goes down to 0. We can do remap successfully
4247 * without charging it again.
4250 * The "old" page is under lock_page() until the end of
4251 * migration, so, the old page itself will not be swapped-out.
4252 * If the new page is swapped out before end_migraton, our
4253 * hook to usual swap-out path will catch the event.
4256 SetPageCgroupMigration(pc);
4258 unlock_page_cgroup(pc);
4260 * If the page is not charged at this point,
4268 * We charge new page before it's used/mapped. So, even if unlock_page()
4269 * is called before end_migration, we can catch all events on this new
4270 * page. In the case new page is migrated but not remapped, new page's
4271 * mapcount will be finally 0 and we call uncharge in end_migration().
4274 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4276 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4278 * The page is committed to the memcg, but it's not actually
4279 * charged to the res_counter since we plan on replacing the
4280 * old one and only one page is going to be left afterwards.
4282 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4285 /* remove redundant charge if migration failed*/
4286 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4287 struct page *oldpage, struct page *newpage, bool migration_ok)
4289 struct page *used, *unused;
4290 struct page_cgroup *pc;
4296 if (!migration_ok) {
4303 anon = PageAnon(used);
4304 __mem_cgroup_uncharge_common(unused,
4305 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4306 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4308 css_put(&memcg->css);
4310 * We disallowed uncharge of pages under migration because mapcount
4311 * of the page goes down to zero, temporarly.
4312 * Clear the flag and check the page should be charged.
4314 pc = lookup_page_cgroup(oldpage);
4315 lock_page_cgroup(pc);
4316 ClearPageCgroupMigration(pc);
4317 unlock_page_cgroup(pc);
4320 * If a page is a file cache, radix-tree replacement is very atomic
4321 * and we can skip this check. When it was an Anon page, its mapcount
4322 * goes down to 0. But because we added MIGRATION flage, it's not
4323 * uncharged yet. There are several case but page->mapcount check
4324 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4325 * check. (see prepare_charge() also)
4328 mem_cgroup_uncharge_page(used);
4332 * At replace page cache, newpage is not under any memcg but it's on
4333 * LRU. So, this function doesn't touch res_counter but handles LRU
4334 * in correct way. Both pages are locked so we cannot race with uncharge.
4336 void mem_cgroup_replace_page_cache(struct page *oldpage,
4337 struct page *newpage)
4339 struct mem_cgroup *memcg = NULL;
4340 struct page_cgroup *pc;
4341 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4343 if (mem_cgroup_disabled())
4346 pc = lookup_page_cgroup(oldpage);
4347 /* fix accounting on old pages */
4348 lock_page_cgroup(pc);
4349 if (PageCgroupUsed(pc)) {
4350 memcg = pc->mem_cgroup;
4351 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4352 ClearPageCgroupUsed(pc);
4354 unlock_page_cgroup(pc);
4357 * When called from shmem_replace_page(), in some cases the
4358 * oldpage has already been charged, and in some cases not.
4363 * Even if newpage->mapping was NULL before starting replacement,
4364 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4365 * LRU while we overwrite pc->mem_cgroup.
4367 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4370 #ifdef CONFIG_DEBUG_VM
4371 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4373 struct page_cgroup *pc;
4375 pc = lookup_page_cgroup(page);
4377 * Can be NULL while feeding pages into the page allocator for
4378 * the first time, i.e. during boot or memory hotplug;
4379 * or when mem_cgroup_disabled().
4381 if (likely(pc) && PageCgroupUsed(pc))
4386 bool mem_cgroup_bad_page_check(struct page *page)
4388 if (mem_cgroup_disabled())
4391 return lookup_page_cgroup_used(page) != NULL;
4394 void mem_cgroup_print_bad_page(struct page *page)
4396 struct page_cgroup *pc;
4398 pc = lookup_page_cgroup_used(page);
4400 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4401 pc, pc->flags, pc->mem_cgroup);
4406 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4407 unsigned long long val)
4410 u64 memswlimit, memlimit;
4412 int children = mem_cgroup_count_children(memcg);
4413 u64 curusage, oldusage;
4417 * For keeping hierarchical_reclaim simple, how long we should retry
4418 * is depends on callers. We set our retry-count to be function
4419 * of # of children which we should visit in this loop.
4421 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4423 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4426 while (retry_count) {
4427 if (signal_pending(current)) {
4432 * Rather than hide all in some function, I do this in
4433 * open coded manner. You see what this really does.
4434 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4436 mutex_lock(&set_limit_mutex);
4437 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4438 if (memswlimit < val) {
4440 mutex_unlock(&set_limit_mutex);
4444 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4448 ret = res_counter_set_limit(&memcg->res, val);
4450 if (memswlimit == val)
4451 memcg->memsw_is_minimum = true;
4453 memcg->memsw_is_minimum = false;
4455 mutex_unlock(&set_limit_mutex);
4460 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4461 MEM_CGROUP_RECLAIM_SHRINK);
4462 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4463 /* Usage is reduced ? */
4464 if (curusage >= oldusage)
4467 oldusage = curusage;
4469 if (!ret && enlarge)
4470 memcg_oom_recover(memcg);
4475 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4476 unsigned long long val)
4479 u64 memlimit, memswlimit, oldusage, curusage;
4480 int children = mem_cgroup_count_children(memcg);
4484 /* see mem_cgroup_resize_res_limit */
4485 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4486 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4487 while (retry_count) {
4488 if (signal_pending(current)) {
4493 * Rather than hide all in some function, I do this in
4494 * open coded manner. You see what this really does.
4495 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4497 mutex_lock(&set_limit_mutex);
4498 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4499 if (memlimit > val) {
4501 mutex_unlock(&set_limit_mutex);
4504 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4505 if (memswlimit < val)
4507 ret = res_counter_set_limit(&memcg->memsw, val);
4509 if (memlimit == val)
4510 memcg->memsw_is_minimum = true;
4512 memcg->memsw_is_minimum = false;
4514 mutex_unlock(&set_limit_mutex);
4519 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4520 MEM_CGROUP_RECLAIM_NOSWAP |
4521 MEM_CGROUP_RECLAIM_SHRINK);
4522 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4523 /* Usage is reduced ? */
4524 if (curusage >= oldusage)
4527 oldusage = curusage;
4529 if (!ret && enlarge)
4530 memcg_oom_recover(memcg);
4535 * mem_cgroup_force_empty_list - clears LRU of a group
4536 * @memcg: group to clear
4539 * @lru: lru to to clear
4541 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4542 * reclaim the pages page themselves - pages are moved to the parent (or root)
4545 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4546 int node, int zid, enum lru_list lru)
4548 struct lruvec *lruvec;
4549 unsigned long flags;
4550 struct list_head *list;
4554 zone = &NODE_DATA(node)->node_zones[zid];
4555 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4556 list = &lruvec->lists[lru];
4560 struct page_cgroup *pc;
4563 spin_lock_irqsave(&zone->lru_lock, flags);
4564 if (list_empty(list)) {
4565 spin_unlock_irqrestore(&zone->lru_lock, flags);
4568 page = list_entry(list->prev, struct page, lru);
4570 list_move(&page->lru, list);
4572 spin_unlock_irqrestore(&zone->lru_lock, flags);
4575 spin_unlock_irqrestore(&zone->lru_lock, flags);
4577 pc = lookup_page_cgroup(page);
4579 if (mem_cgroup_move_parent(page, pc, memcg)) {
4580 /* found lock contention or "pc" is obsolete. */
4585 } while (!list_empty(list));
4589 * make mem_cgroup's charge to be 0 if there is no task by moving
4590 * all the charges and pages to the parent.
4591 * This enables deleting this mem_cgroup.
4593 * Caller is responsible for holding css reference on the memcg.
4595 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4601 /* This is for making all *used* pages to be on LRU. */
4602 lru_add_drain_all();
4603 drain_all_stock_sync(memcg);
4604 mem_cgroup_start_move(memcg);
4605 for_each_node_state(node, N_MEMORY) {
4606 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4609 mem_cgroup_force_empty_list(memcg,
4614 mem_cgroup_end_move(memcg);
4615 memcg_oom_recover(memcg);
4619 * Kernel memory may not necessarily be trackable to a specific
4620 * process. So they are not migrated, and therefore we can't
4621 * expect their value to drop to 0 here.
4622 * Having res filled up with kmem only is enough.
4624 * This is a safety check because mem_cgroup_force_empty_list
4625 * could have raced with mem_cgroup_replace_page_cache callers
4626 * so the lru seemed empty but the page could have been added
4627 * right after the check. RES_USAGE should be safe as we always
4628 * charge before adding to the LRU.
4630 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4631 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4632 } while (usage > 0);
4636 * This mainly exists for tests during the setting of set of use_hierarchy.
4637 * Since this is the very setting we are changing, the current hierarchy value
4640 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4642 struct cgroup_subsys_state *pos;
4644 /* bounce at first found */
4645 css_for_each_child(pos, &memcg->css)
4651 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4652 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4653 * from mem_cgroup_count_children(), in the sense that we don't really care how
4654 * many children we have; we only need to know if we have any. It also counts
4655 * any memcg without hierarchy as infertile.
4657 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4659 return memcg->use_hierarchy && __memcg_has_children(memcg);
4663 * Reclaims as many pages from the given memcg as possible and moves
4664 * the rest to the parent.
4666 * Caller is responsible for holding css reference for memcg.
4668 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4670 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4671 struct cgroup *cgrp = memcg->css.cgroup;
4673 /* returns EBUSY if there is a task or if we come here twice. */
4674 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4677 /* we call try-to-free pages for make this cgroup empty */
4678 lru_add_drain_all();
4679 /* try to free all pages in this cgroup */
4680 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4683 if (signal_pending(current))
4686 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4690 /* maybe some writeback is necessary */
4691 congestion_wait(BLK_RW_ASYNC, HZ/10);
4696 mem_cgroup_reparent_charges(memcg);
4701 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4704 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4706 if (mem_cgroup_is_root(memcg))
4708 return mem_cgroup_force_empty(memcg);
4711 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4714 return mem_cgroup_from_css(css)->use_hierarchy;
4717 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4718 struct cftype *cft, u64 val)
4721 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4722 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4724 mutex_lock(&memcg_create_mutex);
4726 if (memcg->use_hierarchy == val)
4730 * If parent's use_hierarchy is set, we can't make any modifications
4731 * in the child subtrees. If it is unset, then the change can
4732 * occur, provided the current cgroup has no children.
4734 * For the root cgroup, parent_mem is NULL, we allow value to be
4735 * set if there are no children.
4737 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4738 (val == 1 || val == 0)) {
4739 if (!__memcg_has_children(memcg))
4740 memcg->use_hierarchy = val;
4747 mutex_unlock(&memcg_create_mutex);
4753 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4754 enum mem_cgroup_stat_index idx)
4756 struct mem_cgroup *iter;
4759 /* Per-cpu values can be negative, use a signed accumulator */
4760 for_each_mem_cgroup_tree(iter, memcg)
4761 val += mem_cgroup_read_stat(iter, idx);
4763 if (val < 0) /* race ? */
4768 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4772 if (!mem_cgroup_is_root(memcg)) {
4774 return res_counter_read_u64(&memcg->res, RES_USAGE);
4776 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4780 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4781 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4783 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4784 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4787 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4789 return val << PAGE_SHIFT;
4792 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
4793 struct cftype *cft, struct file *file,
4794 char __user *buf, size_t nbytes, loff_t *ppos)
4796 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4802 type = MEMFILE_TYPE(cft->private);
4803 name = MEMFILE_ATTR(cft->private);
4807 if (name == RES_USAGE)
4808 val = mem_cgroup_usage(memcg, false);
4810 val = res_counter_read_u64(&memcg->res, name);
4813 if (name == RES_USAGE)
4814 val = mem_cgroup_usage(memcg, true);
4816 val = res_counter_read_u64(&memcg->memsw, name);
4819 val = res_counter_read_u64(&memcg->kmem, name);
4825 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4826 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4829 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
4832 #ifdef CONFIG_MEMCG_KMEM
4833 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4835 * For simplicity, we won't allow this to be disabled. It also can't
4836 * be changed if the cgroup has children already, or if tasks had
4839 * If tasks join before we set the limit, a person looking at
4840 * kmem.usage_in_bytes will have no way to determine when it took
4841 * place, which makes the value quite meaningless.
4843 * After it first became limited, changes in the value of the limit are
4844 * of course permitted.
4846 mutex_lock(&memcg_create_mutex);
4847 mutex_lock(&set_limit_mutex);
4848 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
4849 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
4853 ret = res_counter_set_limit(&memcg->kmem, val);
4856 ret = memcg_update_cache_sizes(memcg);
4858 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
4861 static_key_slow_inc(&memcg_kmem_enabled_key);
4863 * setting the active bit after the inc will guarantee no one
4864 * starts accounting before all call sites are patched
4866 memcg_kmem_set_active(memcg);
4868 ret = res_counter_set_limit(&memcg->kmem, val);
4870 mutex_unlock(&set_limit_mutex);
4871 mutex_unlock(&memcg_create_mutex);
4876 #ifdef CONFIG_MEMCG_KMEM
4877 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4880 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4884 memcg->kmem_account_flags = parent->kmem_account_flags;
4886 * When that happen, we need to disable the static branch only on those
4887 * memcgs that enabled it. To achieve this, we would be forced to
4888 * complicate the code by keeping track of which memcgs were the ones
4889 * that actually enabled limits, and which ones got it from its
4892 * It is a lot simpler just to do static_key_slow_inc() on every child
4893 * that is accounted.
4895 if (!memcg_kmem_is_active(memcg))
4899 * __mem_cgroup_free() will issue static_key_slow_dec() because this
4900 * memcg is active already. If the later initialization fails then the
4901 * cgroup core triggers the cleanup so we do not have to do it here.
4903 static_key_slow_inc(&memcg_kmem_enabled_key);
4905 mutex_lock(&set_limit_mutex);
4906 memcg_stop_kmem_account();
4907 ret = memcg_update_cache_sizes(memcg);
4908 memcg_resume_kmem_account();
4909 mutex_unlock(&set_limit_mutex);
4913 #endif /* CONFIG_MEMCG_KMEM */
4916 * The user of this function is...
4919 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
4922 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4925 unsigned long long val;
4928 type = MEMFILE_TYPE(cft->private);
4929 name = MEMFILE_ATTR(cft->private);
4933 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4937 /* This function does all necessary parse...reuse it */
4938 ret = res_counter_memparse_write_strategy(buffer, &val);
4942 ret = mem_cgroup_resize_limit(memcg, val);
4943 else if (type == _MEMSWAP)
4944 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4945 else if (type == _KMEM)
4946 ret = memcg_update_kmem_limit(css, val);
4950 case RES_SOFT_LIMIT:
4951 ret = res_counter_memparse_write_strategy(buffer, &val);
4955 * For memsw, soft limits are hard to implement in terms
4956 * of semantics, for now, we support soft limits for
4957 * control without swap
4960 ret = res_counter_set_soft_limit(&memcg->res, val);
4965 ret = -EINVAL; /* should be BUG() ? */
4971 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4972 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4974 unsigned long long min_limit, min_memsw_limit, tmp;
4976 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4977 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4978 if (!memcg->use_hierarchy)
4981 while (css_parent(&memcg->css)) {
4982 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4983 if (!memcg->use_hierarchy)
4985 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4986 min_limit = min(min_limit, tmp);
4987 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4988 min_memsw_limit = min(min_memsw_limit, tmp);
4991 *mem_limit = min_limit;
4992 *memsw_limit = min_memsw_limit;
4995 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
4997 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5001 type = MEMFILE_TYPE(event);
5002 name = MEMFILE_ATTR(event);
5007 res_counter_reset_max(&memcg->res);
5008 else if (type == _MEMSWAP)
5009 res_counter_reset_max(&memcg->memsw);
5010 else if (type == _KMEM)
5011 res_counter_reset_max(&memcg->kmem);
5017 res_counter_reset_failcnt(&memcg->res);
5018 else if (type == _MEMSWAP)
5019 res_counter_reset_failcnt(&memcg->memsw);
5020 else if (type == _KMEM)
5021 res_counter_reset_failcnt(&memcg->kmem);
5030 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5033 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5037 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5038 struct cftype *cft, u64 val)
5040 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5042 if (val >= (1 << NR_MOVE_TYPE))
5046 * No kind of locking is needed in here, because ->can_attach() will
5047 * check this value once in the beginning of the process, and then carry
5048 * on with stale data. This means that changes to this value will only
5049 * affect task migrations starting after the change.
5051 memcg->move_charge_at_immigrate = val;
5055 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5056 struct cftype *cft, u64 val)
5063 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5064 struct cftype *cft, struct seq_file *m)
5067 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5068 unsigned long node_nr;
5069 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5071 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5072 seq_printf(m, "total=%lu", total_nr);
5073 for_each_node_state(nid, N_MEMORY) {
5074 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5075 seq_printf(m, " N%d=%lu", nid, node_nr);
5079 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5080 seq_printf(m, "file=%lu", file_nr);
5081 for_each_node_state(nid, N_MEMORY) {
5082 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5084 seq_printf(m, " N%d=%lu", nid, node_nr);
5088 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5089 seq_printf(m, "anon=%lu", anon_nr);
5090 for_each_node_state(nid, N_MEMORY) {
5091 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5093 seq_printf(m, " N%d=%lu", nid, node_nr);
5097 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5098 seq_printf(m, "unevictable=%lu", unevictable_nr);
5099 for_each_node_state(nid, N_MEMORY) {
5100 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5101 BIT(LRU_UNEVICTABLE));
5102 seq_printf(m, " N%d=%lu", nid, node_nr);
5107 #endif /* CONFIG_NUMA */
5109 static inline void mem_cgroup_lru_names_not_uptodate(void)
5111 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5114 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5117 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5118 struct mem_cgroup *mi;
5121 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5122 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5124 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5125 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5128 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5129 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5130 mem_cgroup_read_events(memcg, i));
5132 for (i = 0; i < NR_LRU_LISTS; i++)
5133 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5134 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5136 /* Hierarchical information */
5138 unsigned long long limit, memsw_limit;
5139 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5140 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5141 if (do_swap_account)
5142 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5146 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5149 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5151 for_each_mem_cgroup_tree(mi, memcg)
5152 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5153 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5156 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5157 unsigned long long val = 0;
5159 for_each_mem_cgroup_tree(mi, memcg)
5160 val += mem_cgroup_read_events(mi, i);
5161 seq_printf(m, "total_%s %llu\n",
5162 mem_cgroup_events_names[i], val);
5165 for (i = 0; i < NR_LRU_LISTS; i++) {
5166 unsigned long long val = 0;
5168 for_each_mem_cgroup_tree(mi, memcg)
5169 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5170 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5173 #ifdef CONFIG_DEBUG_VM
5176 struct mem_cgroup_per_zone *mz;
5177 struct zone_reclaim_stat *rstat;
5178 unsigned long recent_rotated[2] = {0, 0};
5179 unsigned long recent_scanned[2] = {0, 0};
5181 for_each_online_node(nid)
5182 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5183 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5184 rstat = &mz->lruvec.reclaim_stat;
5186 recent_rotated[0] += rstat->recent_rotated[0];
5187 recent_rotated[1] += rstat->recent_rotated[1];
5188 recent_scanned[0] += rstat->recent_scanned[0];
5189 recent_scanned[1] += rstat->recent_scanned[1];
5191 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5192 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5193 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5194 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5201 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5204 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5206 return mem_cgroup_swappiness(memcg);
5209 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5210 struct cftype *cft, u64 val)
5212 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5213 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5215 if (val > 100 || !parent)
5218 mutex_lock(&memcg_create_mutex);
5220 /* If under hierarchy, only empty-root can set this value */
5221 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5222 mutex_unlock(&memcg_create_mutex);
5226 memcg->swappiness = val;
5228 mutex_unlock(&memcg_create_mutex);
5233 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5235 struct mem_cgroup_threshold_ary *t;
5241 t = rcu_dereference(memcg->thresholds.primary);
5243 t = rcu_dereference(memcg->memsw_thresholds.primary);
5248 usage = mem_cgroup_usage(memcg, swap);
5251 * current_threshold points to threshold just below or equal to usage.
5252 * If it's not true, a threshold was crossed after last
5253 * call of __mem_cgroup_threshold().
5255 i = t->current_threshold;
5258 * Iterate backward over array of thresholds starting from
5259 * current_threshold and check if a threshold is crossed.
5260 * If none of thresholds below usage is crossed, we read
5261 * only one element of the array here.
5263 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5264 eventfd_signal(t->entries[i].eventfd, 1);
5266 /* i = current_threshold + 1 */
5270 * Iterate forward over array of thresholds starting from
5271 * current_threshold+1 and check if a threshold is crossed.
5272 * If none of thresholds above usage is crossed, we read
5273 * only one element of the array here.
5275 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5276 eventfd_signal(t->entries[i].eventfd, 1);
5278 /* Update current_threshold */
5279 t->current_threshold = i - 1;
5284 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5287 __mem_cgroup_threshold(memcg, false);
5288 if (do_swap_account)
5289 __mem_cgroup_threshold(memcg, true);
5291 memcg = parent_mem_cgroup(memcg);
5295 static int compare_thresholds(const void *a, const void *b)
5297 const struct mem_cgroup_threshold *_a = a;
5298 const struct mem_cgroup_threshold *_b = b;
5300 if (_a->threshold > _b->threshold)
5303 if (_a->threshold < _b->threshold)
5309 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5311 struct mem_cgroup_eventfd_list *ev;
5313 list_for_each_entry(ev, &memcg->oom_notify, list)
5314 eventfd_signal(ev->eventfd, 1);
5318 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5320 struct mem_cgroup *iter;
5322 for_each_mem_cgroup_tree(iter, memcg)
5323 mem_cgroup_oom_notify_cb(iter);
5326 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5327 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5329 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5330 struct mem_cgroup_thresholds *thresholds;
5331 struct mem_cgroup_threshold_ary *new;
5332 enum res_type type = MEMFILE_TYPE(cft->private);
5333 u64 threshold, usage;
5336 ret = res_counter_memparse_write_strategy(args, &threshold);
5340 mutex_lock(&memcg->thresholds_lock);
5343 thresholds = &memcg->thresholds;
5344 else if (type == _MEMSWAP)
5345 thresholds = &memcg->memsw_thresholds;
5349 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5351 /* Check if a threshold crossed before adding a new one */
5352 if (thresholds->primary)
5353 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5355 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5357 /* Allocate memory for new array of thresholds */
5358 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5366 /* Copy thresholds (if any) to new array */
5367 if (thresholds->primary) {
5368 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5369 sizeof(struct mem_cgroup_threshold));
5372 /* Add new threshold */
5373 new->entries[size - 1].eventfd = eventfd;
5374 new->entries[size - 1].threshold = threshold;
5376 /* Sort thresholds. Registering of new threshold isn't time-critical */
5377 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5378 compare_thresholds, NULL);
5380 /* Find current threshold */
5381 new->current_threshold = -1;
5382 for (i = 0; i < size; i++) {
5383 if (new->entries[i].threshold <= usage) {
5385 * new->current_threshold will not be used until
5386 * rcu_assign_pointer(), so it's safe to increment
5389 ++new->current_threshold;
5394 /* Free old spare buffer and save old primary buffer as spare */
5395 kfree(thresholds->spare);
5396 thresholds->spare = thresholds->primary;
5398 rcu_assign_pointer(thresholds->primary, new);
5400 /* To be sure that nobody uses thresholds */
5404 mutex_unlock(&memcg->thresholds_lock);
5409 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5410 struct cftype *cft, struct eventfd_ctx *eventfd)
5412 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5413 struct mem_cgroup_thresholds *thresholds;
5414 struct mem_cgroup_threshold_ary *new;
5415 enum res_type type = MEMFILE_TYPE(cft->private);
5419 mutex_lock(&memcg->thresholds_lock);
5421 thresholds = &memcg->thresholds;
5422 else if (type == _MEMSWAP)
5423 thresholds = &memcg->memsw_thresholds;
5427 if (!thresholds->primary)
5430 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5432 /* Check if a threshold crossed before removing */
5433 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5435 /* Calculate new number of threshold */
5437 for (i = 0; i < thresholds->primary->size; i++) {
5438 if (thresholds->primary->entries[i].eventfd != eventfd)
5442 new = thresholds->spare;
5444 /* Set thresholds array to NULL if we don't have thresholds */
5453 /* Copy thresholds and find current threshold */
5454 new->current_threshold = -1;
5455 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5456 if (thresholds->primary->entries[i].eventfd == eventfd)
5459 new->entries[j] = thresholds->primary->entries[i];
5460 if (new->entries[j].threshold <= usage) {
5462 * new->current_threshold will not be used
5463 * until rcu_assign_pointer(), so it's safe to increment
5466 ++new->current_threshold;
5472 /* Swap primary and spare array */
5473 thresholds->spare = thresholds->primary;
5474 /* If all events are unregistered, free the spare array */
5476 kfree(thresholds->spare);
5477 thresholds->spare = NULL;
5480 rcu_assign_pointer(thresholds->primary, new);
5482 /* To be sure that nobody uses thresholds */
5485 mutex_unlock(&memcg->thresholds_lock);
5488 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5489 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5491 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5492 struct mem_cgroup_eventfd_list *event;
5493 enum res_type type = MEMFILE_TYPE(cft->private);
5495 BUG_ON(type != _OOM_TYPE);
5496 event = kmalloc(sizeof(*event), GFP_KERNEL);
5500 spin_lock(&memcg_oom_lock);
5502 event->eventfd = eventfd;
5503 list_add(&event->list, &memcg->oom_notify);
5505 /* already in OOM ? */
5506 if (atomic_read(&memcg->under_oom))
5507 eventfd_signal(eventfd, 1);
5508 spin_unlock(&memcg_oom_lock);
5513 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5514 struct cftype *cft, struct eventfd_ctx *eventfd)
5516 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5517 struct mem_cgroup_eventfd_list *ev, *tmp;
5518 enum res_type type = MEMFILE_TYPE(cft->private);
5520 BUG_ON(type != _OOM_TYPE);
5522 spin_lock(&memcg_oom_lock);
5524 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5525 if (ev->eventfd == eventfd) {
5526 list_del(&ev->list);
5531 spin_unlock(&memcg_oom_lock);
5534 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5535 struct cftype *cft, struct cgroup_map_cb *cb)
5537 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5539 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5541 if (atomic_read(&memcg->under_oom))
5542 cb->fill(cb, "under_oom", 1);
5544 cb->fill(cb, "under_oom", 0);
5548 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5549 struct cftype *cft, u64 val)
5551 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5552 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5554 /* cannot set to root cgroup and only 0 and 1 are allowed */
5555 if (!parent || !((val == 0) || (val == 1)))
5558 mutex_lock(&memcg_create_mutex);
5559 /* oom-kill-disable is a flag for subhierarchy. */
5560 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5561 mutex_unlock(&memcg_create_mutex);
5564 memcg->oom_kill_disable = val;
5566 memcg_oom_recover(memcg);
5567 mutex_unlock(&memcg_create_mutex);
5571 #ifdef CONFIG_MEMCG_KMEM
5572 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5576 memcg->kmemcg_id = -1;
5577 ret = memcg_propagate_kmem(memcg);
5581 return mem_cgroup_sockets_init(memcg, ss);
5584 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5586 mem_cgroup_sockets_destroy(memcg);
5589 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5591 if (!memcg_kmem_is_active(memcg))
5595 * kmem charges can outlive the cgroup. In the case of slab
5596 * pages, for instance, a page contain objects from various
5597 * processes. As we prevent from taking a reference for every
5598 * such allocation we have to be careful when doing uncharge
5599 * (see memcg_uncharge_kmem) and here during offlining.
5601 * The idea is that that only the _last_ uncharge which sees
5602 * the dead memcg will drop the last reference. An additional
5603 * reference is taken here before the group is marked dead
5604 * which is then paired with css_put during uncharge resp. here.
5606 * Although this might sound strange as this path is called from
5607 * css_offline() when the referencemight have dropped down to 0
5608 * and shouldn't be incremented anymore (css_tryget would fail)
5609 * we do not have other options because of the kmem allocations
5612 css_get(&memcg->css);
5614 memcg_kmem_mark_dead(memcg);
5616 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5619 if (memcg_kmem_test_and_clear_dead(memcg))
5620 css_put(&memcg->css);
5623 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5628 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5632 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5637 static struct cftype mem_cgroup_files[] = {
5639 .name = "usage_in_bytes",
5640 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5641 .read = mem_cgroup_read,
5642 .register_event = mem_cgroup_usage_register_event,
5643 .unregister_event = mem_cgroup_usage_unregister_event,
5646 .name = "max_usage_in_bytes",
5647 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5648 .trigger = mem_cgroup_reset,
5649 .read = mem_cgroup_read,
5652 .name = "limit_in_bytes",
5653 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5654 .write_string = mem_cgroup_write,
5655 .read = mem_cgroup_read,
5658 .name = "soft_limit_in_bytes",
5659 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5660 .write_string = mem_cgroup_write,
5661 .read = mem_cgroup_read,
5665 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5666 .trigger = mem_cgroup_reset,
5667 .read = mem_cgroup_read,
5671 .read_seq_string = memcg_stat_show,
5674 .name = "force_empty",
5675 .trigger = mem_cgroup_force_empty_write,
5678 .name = "use_hierarchy",
5679 .flags = CFTYPE_INSANE,
5680 .write_u64 = mem_cgroup_hierarchy_write,
5681 .read_u64 = mem_cgroup_hierarchy_read,
5684 .name = "swappiness",
5685 .read_u64 = mem_cgroup_swappiness_read,
5686 .write_u64 = mem_cgroup_swappiness_write,
5689 .name = "move_charge_at_immigrate",
5690 .read_u64 = mem_cgroup_move_charge_read,
5691 .write_u64 = mem_cgroup_move_charge_write,
5694 .name = "oom_control",
5695 .read_map = mem_cgroup_oom_control_read,
5696 .write_u64 = mem_cgroup_oom_control_write,
5697 .register_event = mem_cgroup_oom_register_event,
5698 .unregister_event = mem_cgroup_oom_unregister_event,
5699 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5702 .name = "pressure_level",
5703 .register_event = vmpressure_register_event,
5704 .unregister_event = vmpressure_unregister_event,
5708 .name = "numa_stat",
5709 .read_seq_string = memcg_numa_stat_show,
5712 #ifdef CONFIG_MEMCG_KMEM
5714 .name = "kmem.limit_in_bytes",
5715 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5716 .write_string = mem_cgroup_write,
5717 .read = mem_cgroup_read,
5720 .name = "kmem.usage_in_bytes",
5721 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5722 .read = mem_cgroup_read,
5725 .name = "kmem.failcnt",
5726 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5727 .trigger = mem_cgroup_reset,
5728 .read = mem_cgroup_read,
5731 .name = "kmem.max_usage_in_bytes",
5732 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5733 .trigger = mem_cgroup_reset,
5734 .read = mem_cgroup_read,
5736 #ifdef CONFIG_SLABINFO
5738 .name = "kmem.slabinfo",
5739 .read_seq_string = mem_cgroup_slabinfo_read,
5743 { }, /* terminate */
5746 #ifdef CONFIG_MEMCG_SWAP
5747 static struct cftype memsw_cgroup_files[] = {
5749 .name = "memsw.usage_in_bytes",
5750 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5751 .read = mem_cgroup_read,
5752 .register_event = mem_cgroup_usage_register_event,
5753 .unregister_event = mem_cgroup_usage_unregister_event,
5756 .name = "memsw.max_usage_in_bytes",
5757 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5758 .trigger = mem_cgroup_reset,
5759 .read = mem_cgroup_read,
5762 .name = "memsw.limit_in_bytes",
5763 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5764 .write_string = mem_cgroup_write,
5765 .read = mem_cgroup_read,
5768 .name = "memsw.failcnt",
5769 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5770 .trigger = mem_cgroup_reset,
5771 .read = mem_cgroup_read,
5773 { }, /* terminate */
5776 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5778 struct mem_cgroup_per_node *pn;
5779 struct mem_cgroup_per_zone *mz;
5780 int zone, tmp = node;
5782 * This routine is called against possible nodes.
5783 * But it's BUG to call kmalloc() against offline node.
5785 * TODO: this routine can waste much memory for nodes which will
5786 * never be onlined. It's better to use memory hotplug callback
5789 if (!node_state(node, N_NORMAL_MEMORY))
5791 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5795 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5796 mz = &pn->zoneinfo[zone];
5797 lruvec_init(&mz->lruvec);
5800 memcg->nodeinfo[node] = pn;
5804 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5806 kfree(memcg->nodeinfo[node]);
5809 static struct mem_cgroup *mem_cgroup_alloc(void)
5811 struct mem_cgroup *memcg;
5812 size_t size = memcg_size();
5814 /* Can be very big if nr_node_ids is very big */
5815 if (size < PAGE_SIZE)
5816 memcg = kzalloc(size, GFP_KERNEL);
5818 memcg = vzalloc(size);
5823 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5826 spin_lock_init(&memcg->pcp_counter_lock);
5830 if (size < PAGE_SIZE)
5838 * At destroying mem_cgroup, references from swap_cgroup can remain.
5839 * (scanning all at force_empty is too costly...)
5841 * Instead of clearing all references at force_empty, we remember
5842 * the number of reference from swap_cgroup and free mem_cgroup when
5843 * it goes down to 0.
5845 * Removal of cgroup itself succeeds regardless of refs from swap.
5848 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5851 size_t size = memcg_size();
5853 free_css_id(&mem_cgroup_subsys, &memcg->css);
5856 free_mem_cgroup_per_zone_info(memcg, node);
5858 free_percpu(memcg->stat);
5861 * We need to make sure that (at least for now), the jump label
5862 * destruction code runs outside of the cgroup lock. This is because
5863 * get_online_cpus(), which is called from the static_branch update,
5864 * can't be called inside the cgroup_lock. cpusets are the ones
5865 * enforcing this dependency, so if they ever change, we might as well.
5867 * schedule_work() will guarantee this happens. Be careful if you need
5868 * to move this code around, and make sure it is outside
5871 disarm_static_keys(memcg);
5872 if (size < PAGE_SIZE)
5879 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5881 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5883 if (!memcg->res.parent)
5885 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5887 EXPORT_SYMBOL(parent_mem_cgroup);
5889 static struct cgroup_subsys_state * __ref
5890 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5892 struct mem_cgroup *memcg;
5893 long error = -ENOMEM;
5896 memcg = mem_cgroup_alloc();
5898 return ERR_PTR(error);
5901 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5905 if (parent_css == NULL) {
5906 root_mem_cgroup = memcg;
5907 res_counter_init(&memcg->res, NULL);
5908 res_counter_init(&memcg->memsw, NULL);
5909 res_counter_init(&memcg->kmem, NULL);
5912 memcg->last_scanned_node = MAX_NUMNODES;
5913 INIT_LIST_HEAD(&memcg->oom_notify);
5914 memcg->move_charge_at_immigrate = 0;
5915 mutex_init(&memcg->thresholds_lock);
5916 spin_lock_init(&memcg->move_lock);
5917 vmpressure_init(&memcg->vmpressure);
5922 __mem_cgroup_free(memcg);
5923 return ERR_PTR(error);
5927 mem_cgroup_css_online(struct cgroup_subsys_state *css)
5929 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5930 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
5936 mutex_lock(&memcg_create_mutex);
5938 memcg->use_hierarchy = parent->use_hierarchy;
5939 memcg->oom_kill_disable = parent->oom_kill_disable;
5940 memcg->swappiness = mem_cgroup_swappiness(parent);
5942 if (parent->use_hierarchy) {
5943 res_counter_init(&memcg->res, &parent->res);
5944 res_counter_init(&memcg->memsw, &parent->memsw);
5945 res_counter_init(&memcg->kmem, &parent->kmem);
5948 * No need to take a reference to the parent because cgroup
5949 * core guarantees its existence.
5952 res_counter_init(&memcg->res, NULL);
5953 res_counter_init(&memcg->memsw, NULL);
5954 res_counter_init(&memcg->kmem, NULL);
5956 * Deeper hierachy with use_hierarchy == false doesn't make
5957 * much sense so let cgroup subsystem know about this
5958 * unfortunate state in our controller.
5960 if (parent != root_mem_cgroup)
5961 mem_cgroup_subsys.broken_hierarchy = true;
5964 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
5965 mutex_unlock(&memcg_create_mutex);
5970 * Announce all parents that a group from their hierarchy is gone.
5972 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
5974 struct mem_cgroup *parent = memcg;
5976 while ((parent = parent_mem_cgroup(parent)))
5977 mem_cgroup_iter_invalidate(parent);
5980 * if the root memcg is not hierarchical we have to check it
5983 if (!root_mem_cgroup->use_hierarchy)
5984 mem_cgroup_iter_invalidate(root_mem_cgroup);
5987 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5989 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5991 kmem_cgroup_css_offline(memcg);
5993 mem_cgroup_invalidate_reclaim_iterators(memcg);
5994 mem_cgroup_reparent_charges(memcg);
5995 mem_cgroup_destroy_all_caches(memcg);
5996 vmpressure_cleanup(&memcg->vmpressure);
5999 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6001 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6003 memcg_destroy_kmem(memcg);
6004 __mem_cgroup_free(memcg);
6008 /* Handlers for move charge at task migration. */
6009 #define PRECHARGE_COUNT_AT_ONCE 256
6010 static int mem_cgroup_do_precharge(unsigned long count)
6013 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6014 struct mem_cgroup *memcg = mc.to;
6016 if (mem_cgroup_is_root(memcg)) {
6017 mc.precharge += count;
6018 /* we don't need css_get for root */
6021 /* try to charge at once */
6023 struct res_counter *dummy;
6025 * "memcg" cannot be under rmdir() because we've already checked
6026 * by cgroup_lock_live_cgroup() that it is not removed and we
6027 * are still under the same cgroup_mutex. So we can postpone
6030 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6032 if (do_swap_account && res_counter_charge(&memcg->memsw,
6033 PAGE_SIZE * count, &dummy)) {
6034 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6037 mc.precharge += count;
6041 /* fall back to one by one charge */
6043 if (signal_pending(current)) {
6047 if (!batch_count--) {
6048 batch_count = PRECHARGE_COUNT_AT_ONCE;
6051 ret = __mem_cgroup_try_charge(NULL,
6052 GFP_KERNEL, 1, &memcg, false);
6054 /* mem_cgroup_clear_mc() will do uncharge later */
6062 * get_mctgt_type - get target type of moving charge
6063 * @vma: the vma the pte to be checked belongs
6064 * @addr: the address corresponding to the pte to be checked
6065 * @ptent: the pte to be checked
6066 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6069 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6070 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6071 * move charge. if @target is not NULL, the page is stored in target->page
6072 * with extra refcnt got(Callers should handle it).
6073 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6074 * target for charge migration. if @target is not NULL, the entry is stored
6077 * Called with pte lock held.
6084 enum mc_target_type {
6090 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6091 unsigned long addr, pte_t ptent)
6093 struct page *page = vm_normal_page(vma, addr, ptent);
6095 if (!page || !page_mapped(page))
6097 if (PageAnon(page)) {
6098 /* we don't move shared anon */
6101 } else if (!move_file())
6102 /* we ignore mapcount for file pages */
6104 if (!get_page_unless_zero(page))
6111 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6112 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6114 struct page *page = NULL;
6115 swp_entry_t ent = pte_to_swp_entry(ptent);
6117 if (!move_anon() || non_swap_entry(ent))
6120 * Because lookup_swap_cache() updates some statistics counter,
6121 * we call find_get_page() with swapper_space directly.
6123 page = find_get_page(swap_address_space(ent), ent.val);
6124 if (do_swap_account)
6125 entry->val = ent.val;
6130 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6131 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6137 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6138 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6140 struct page *page = NULL;
6141 struct address_space *mapping;
6144 if (!vma->vm_file) /* anonymous vma */
6149 mapping = vma->vm_file->f_mapping;
6150 if (pte_none(ptent))
6151 pgoff = linear_page_index(vma, addr);
6152 else /* pte_file(ptent) is true */
6153 pgoff = pte_to_pgoff(ptent);
6155 /* page is moved even if it's not RSS of this task(page-faulted). */
6156 page = find_get_page(mapping, pgoff);
6159 /* shmem/tmpfs may report page out on swap: account for that too. */
6160 if (radix_tree_exceptional_entry(page)) {
6161 swp_entry_t swap = radix_to_swp_entry(page);
6162 if (do_swap_account)
6164 page = find_get_page(swap_address_space(swap), swap.val);
6170 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6171 unsigned long addr, pte_t ptent, union mc_target *target)
6173 struct page *page = NULL;
6174 struct page_cgroup *pc;
6175 enum mc_target_type ret = MC_TARGET_NONE;
6176 swp_entry_t ent = { .val = 0 };
6178 if (pte_present(ptent))
6179 page = mc_handle_present_pte(vma, addr, ptent);
6180 else if (is_swap_pte(ptent))
6181 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6182 else if (pte_none(ptent) || pte_file(ptent))
6183 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6185 if (!page && !ent.val)
6188 pc = lookup_page_cgroup(page);
6190 * Do only loose check w/o page_cgroup lock.
6191 * mem_cgroup_move_account() checks the pc is valid or not under
6194 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6195 ret = MC_TARGET_PAGE;
6197 target->page = page;
6199 if (!ret || !target)
6202 /* There is a swap entry and a page doesn't exist or isn't charged */
6203 if (ent.val && !ret &&
6204 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6205 ret = MC_TARGET_SWAP;
6212 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6214 * We don't consider swapping or file mapped pages because THP does not
6215 * support them for now.
6216 * Caller should make sure that pmd_trans_huge(pmd) is true.
6218 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6219 unsigned long addr, pmd_t pmd, union mc_target *target)
6221 struct page *page = NULL;
6222 struct page_cgroup *pc;
6223 enum mc_target_type ret = MC_TARGET_NONE;
6225 page = pmd_page(pmd);
6226 VM_BUG_ON(!page || !PageHead(page));
6229 pc = lookup_page_cgroup(page);
6230 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6231 ret = MC_TARGET_PAGE;
6234 target->page = page;
6240 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6241 unsigned long addr, pmd_t pmd, union mc_target *target)
6243 return MC_TARGET_NONE;
6247 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6248 unsigned long addr, unsigned long end,
6249 struct mm_walk *walk)
6251 struct vm_area_struct *vma = walk->private;
6255 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6256 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6257 mc.precharge += HPAGE_PMD_NR;
6258 spin_unlock(&vma->vm_mm->page_table_lock);
6262 if (pmd_trans_unstable(pmd))
6264 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6265 for (; addr != end; pte++, addr += PAGE_SIZE)
6266 if (get_mctgt_type(vma, addr, *pte, NULL))
6267 mc.precharge++; /* increment precharge temporarily */
6268 pte_unmap_unlock(pte - 1, ptl);
6274 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6276 unsigned long precharge;
6277 struct vm_area_struct *vma;
6279 down_read(&mm->mmap_sem);
6280 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6281 struct mm_walk mem_cgroup_count_precharge_walk = {
6282 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6286 if (is_vm_hugetlb_page(vma))
6288 walk_page_range(vma->vm_start, vma->vm_end,
6289 &mem_cgroup_count_precharge_walk);
6291 up_read(&mm->mmap_sem);
6293 precharge = mc.precharge;
6299 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6301 unsigned long precharge = mem_cgroup_count_precharge(mm);
6303 VM_BUG_ON(mc.moving_task);
6304 mc.moving_task = current;
6305 return mem_cgroup_do_precharge(precharge);
6308 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6309 static void __mem_cgroup_clear_mc(void)
6311 struct mem_cgroup *from = mc.from;
6312 struct mem_cgroup *to = mc.to;
6315 /* we must uncharge all the leftover precharges from mc.to */
6317 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6321 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6322 * we must uncharge here.
6324 if (mc.moved_charge) {
6325 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6326 mc.moved_charge = 0;
6328 /* we must fixup refcnts and charges */
6329 if (mc.moved_swap) {
6330 /* uncharge swap account from the old cgroup */
6331 if (!mem_cgroup_is_root(mc.from))
6332 res_counter_uncharge(&mc.from->memsw,
6333 PAGE_SIZE * mc.moved_swap);
6335 for (i = 0; i < mc.moved_swap; i++)
6336 css_put(&mc.from->css);
6338 if (!mem_cgroup_is_root(mc.to)) {
6340 * we charged both to->res and to->memsw, so we should
6343 res_counter_uncharge(&mc.to->res,
6344 PAGE_SIZE * mc.moved_swap);
6346 /* we've already done css_get(mc.to) */
6349 memcg_oom_recover(from);
6350 memcg_oom_recover(to);
6351 wake_up_all(&mc.waitq);
6354 static void mem_cgroup_clear_mc(void)
6356 struct mem_cgroup *from = mc.from;
6359 * we must clear moving_task before waking up waiters at the end of
6362 mc.moving_task = NULL;
6363 __mem_cgroup_clear_mc();
6364 spin_lock(&mc.lock);
6367 spin_unlock(&mc.lock);
6368 mem_cgroup_end_move(from);
6371 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6372 struct cgroup_taskset *tset)
6374 struct task_struct *p = cgroup_taskset_first(tset);
6376 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6377 unsigned long move_charge_at_immigrate;
6380 * We are now commited to this value whatever it is. Changes in this
6381 * tunable will only affect upcoming migrations, not the current one.
6382 * So we need to save it, and keep it going.
6384 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6385 if (move_charge_at_immigrate) {
6386 struct mm_struct *mm;
6387 struct mem_cgroup *from = mem_cgroup_from_task(p);
6389 VM_BUG_ON(from == memcg);
6391 mm = get_task_mm(p);
6394 /* We move charges only when we move a owner of the mm */
6395 if (mm->owner == p) {
6398 VM_BUG_ON(mc.precharge);
6399 VM_BUG_ON(mc.moved_charge);
6400 VM_BUG_ON(mc.moved_swap);
6401 mem_cgroup_start_move(from);
6402 spin_lock(&mc.lock);
6405 mc.immigrate_flags = move_charge_at_immigrate;
6406 spin_unlock(&mc.lock);
6407 /* We set mc.moving_task later */
6409 ret = mem_cgroup_precharge_mc(mm);
6411 mem_cgroup_clear_mc();
6418 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6419 struct cgroup_taskset *tset)
6421 mem_cgroup_clear_mc();
6424 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6425 unsigned long addr, unsigned long end,
6426 struct mm_walk *walk)
6429 struct vm_area_struct *vma = walk->private;
6432 enum mc_target_type target_type;
6433 union mc_target target;
6435 struct page_cgroup *pc;
6438 * We don't take compound_lock() here but no race with splitting thp
6440 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6441 * under splitting, which means there's no concurrent thp split,
6442 * - if another thread runs into split_huge_page() just after we
6443 * entered this if-block, the thread must wait for page table lock
6444 * to be unlocked in __split_huge_page_splitting(), where the main
6445 * part of thp split is not executed yet.
6447 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6448 if (mc.precharge < HPAGE_PMD_NR) {
6449 spin_unlock(&vma->vm_mm->page_table_lock);
6452 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6453 if (target_type == MC_TARGET_PAGE) {
6455 if (!isolate_lru_page(page)) {
6456 pc = lookup_page_cgroup(page);
6457 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6458 pc, mc.from, mc.to)) {
6459 mc.precharge -= HPAGE_PMD_NR;
6460 mc.moved_charge += HPAGE_PMD_NR;
6462 putback_lru_page(page);
6466 spin_unlock(&vma->vm_mm->page_table_lock);
6470 if (pmd_trans_unstable(pmd))
6473 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6474 for (; addr != end; addr += PAGE_SIZE) {
6475 pte_t ptent = *(pte++);
6481 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6482 case MC_TARGET_PAGE:
6484 if (isolate_lru_page(page))
6486 pc = lookup_page_cgroup(page);
6487 if (!mem_cgroup_move_account(page, 1, pc,
6490 /* we uncharge from mc.from later. */
6493 putback_lru_page(page);
6494 put: /* get_mctgt_type() gets the page */
6497 case MC_TARGET_SWAP:
6499 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6501 /* we fixup refcnts and charges later. */
6509 pte_unmap_unlock(pte - 1, ptl);
6514 * We have consumed all precharges we got in can_attach().
6515 * We try charge one by one, but don't do any additional
6516 * charges to mc.to if we have failed in charge once in attach()
6519 ret = mem_cgroup_do_precharge(1);
6527 static void mem_cgroup_move_charge(struct mm_struct *mm)
6529 struct vm_area_struct *vma;
6531 lru_add_drain_all();
6533 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6535 * Someone who are holding the mmap_sem might be waiting in
6536 * waitq. So we cancel all extra charges, wake up all waiters,
6537 * and retry. Because we cancel precharges, we might not be able
6538 * to move enough charges, but moving charge is a best-effort
6539 * feature anyway, so it wouldn't be a big problem.
6541 __mem_cgroup_clear_mc();
6545 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6547 struct mm_walk mem_cgroup_move_charge_walk = {
6548 .pmd_entry = mem_cgroup_move_charge_pte_range,
6552 if (is_vm_hugetlb_page(vma))
6554 ret = walk_page_range(vma->vm_start, vma->vm_end,
6555 &mem_cgroup_move_charge_walk);
6558 * means we have consumed all precharges and failed in
6559 * doing additional charge. Just abandon here.
6563 up_read(&mm->mmap_sem);
6566 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6567 struct cgroup_taskset *tset)
6569 struct task_struct *p = cgroup_taskset_first(tset);
6570 struct mm_struct *mm = get_task_mm(p);
6574 mem_cgroup_move_charge(mm);
6578 mem_cgroup_clear_mc();
6580 #else /* !CONFIG_MMU */
6581 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6582 struct cgroup_taskset *tset)
6586 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6587 struct cgroup_taskset *tset)
6590 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6591 struct cgroup_taskset *tset)
6597 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6598 * to verify sane_behavior flag on each mount attempt.
6600 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6603 * use_hierarchy is forced with sane_behavior. cgroup core
6604 * guarantees that @root doesn't have any children, so turning it
6605 * on for the root memcg is enough.
6607 if (cgroup_sane_behavior(root_css->cgroup))
6608 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6611 struct cgroup_subsys mem_cgroup_subsys = {
6613 .subsys_id = mem_cgroup_subsys_id,
6614 .css_alloc = mem_cgroup_css_alloc,
6615 .css_online = mem_cgroup_css_online,
6616 .css_offline = mem_cgroup_css_offline,
6617 .css_free = mem_cgroup_css_free,
6618 .can_attach = mem_cgroup_can_attach,
6619 .cancel_attach = mem_cgroup_cancel_attach,
6620 .attach = mem_cgroup_move_task,
6621 .bind = mem_cgroup_bind,
6622 .base_cftypes = mem_cgroup_files,
6627 #ifdef CONFIG_MEMCG_SWAP
6628 static int __init enable_swap_account(char *s)
6630 if (!strcmp(s, "1"))
6631 really_do_swap_account = 1;
6632 else if (!strcmp(s, "0"))
6633 really_do_swap_account = 0;
6636 __setup("swapaccount=", enable_swap_account);
6638 static void __init memsw_file_init(void)
6640 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6643 static void __init enable_swap_cgroup(void)
6645 if (!mem_cgroup_disabled() && really_do_swap_account) {
6646 do_swap_account = 1;
6652 static void __init enable_swap_cgroup(void)
6658 * subsys_initcall() for memory controller.
6660 * Some parts like hotcpu_notifier() have to be initialized from this context
6661 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6662 * everything that doesn't depend on a specific mem_cgroup structure should
6663 * be initialized from here.
6665 static int __init mem_cgroup_init(void)
6667 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6668 enable_swap_cgroup();
6672 subsys_initcall(mem_cgroup_init);