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 if
1764 * a) it is over its soft limit
1765 * b) any parent up the hierarchy is over its soft limit
1767 bool mem_cgroup_soft_reclaim_eligible(struct mem_cgroup *memcg)
1769 struct mem_cgroup *parent = memcg;
1771 if (res_counter_soft_limit_excess(&memcg->res))
1775 * If any parent up the hierarchy is over its soft limit then we
1776 * have to obey and reclaim from this group as well.
1778 while ((parent = parent_mem_cgroup(parent))) {
1779 if (res_counter_soft_limit_excess(&parent->res))
1786 static DEFINE_SPINLOCK(memcg_oom_lock);
1789 * Check OOM-Killer is already running under our hierarchy.
1790 * If someone is running, return false.
1792 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1794 struct mem_cgroup *iter, *failed = NULL;
1796 spin_lock(&memcg_oom_lock);
1798 for_each_mem_cgroup_tree(iter, memcg) {
1799 if (iter->oom_lock) {
1801 * this subtree of our hierarchy is already locked
1802 * so we cannot give a lock.
1805 mem_cgroup_iter_break(memcg, iter);
1808 iter->oom_lock = true;
1813 * OK, we failed to lock the whole subtree so we have
1814 * to clean up what we set up to the failing subtree
1816 for_each_mem_cgroup_tree(iter, memcg) {
1817 if (iter == failed) {
1818 mem_cgroup_iter_break(memcg, iter);
1821 iter->oom_lock = false;
1825 spin_unlock(&memcg_oom_lock);
1830 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1832 struct mem_cgroup *iter;
1834 spin_lock(&memcg_oom_lock);
1835 for_each_mem_cgroup_tree(iter, memcg)
1836 iter->oom_lock = false;
1837 spin_unlock(&memcg_oom_lock);
1840 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1842 struct mem_cgroup *iter;
1844 for_each_mem_cgroup_tree(iter, memcg)
1845 atomic_inc(&iter->under_oom);
1848 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1850 struct mem_cgroup *iter;
1853 * When a new child is created while the hierarchy is under oom,
1854 * mem_cgroup_oom_lock() may not be called. We have to use
1855 * atomic_add_unless() here.
1857 for_each_mem_cgroup_tree(iter, memcg)
1858 atomic_add_unless(&iter->under_oom, -1, 0);
1861 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1863 struct oom_wait_info {
1864 struct mem_cgroup *memcg;
1868 static int memcg_oom_wake_function(wait_queue_t *wait,
1869 unsigned mode, int sync, void *arg)
1871 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1872 struct mem_cgroup *oom_wait_memcg;
1873 struct oom_wait_info *oom_wait_info;
1875 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1876 oom_wait_memcg = oom_wait_info->memcg;
1879 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1880 * Then we can use css_is_ancestor without taking care of RCU.
1882 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1883 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1885 return autoremove_wake_function(wait, mode, sync, arg);
1888 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1890 atomic_inc(&memcg->oom_wakeups);
1891 /* for filtering, pass "memcg" as argument. */
1892 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1895 static void memcg_oom_recover(struct mem_cgroup *memcg)
1897 if (memcg && atomic_read(&memcg->under_oom))
1898 memcg_wakeup_oom(memcg);
1902 * try to call OOM killer
1904 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
1909 if (!current->memcg_oom.may_oom)
1912 current->memcg_oom.in_memcg_oom = 1;
1915 * As with any blocking lock, a contender needs to start
1916 * listening for wakeups before attempting the trylock,
1917 * otherwise it can miss the wakeup from the unlock and sleep
1918 * indefinitely. This is just open-coded because our locking
1919 * is so particular to memcg hierarchies.
1921 wakeups = atomic_read(&memcg->oom_wakeups);
1922 mem_cgroup_mark_under_oom(memcg);
1924 locked = mem_cgroup_oom_trylock(memcg);
1927 mem_cgroup_oom_notify(memcg);
1929 if (locked && !memcg->oom_kill_disable) {
1930 mem_cgroup_unmark_under_oom(memcg);
1931 mem_cgroup_out_of_memory(memcg, mask, order);
1932 mem_cgroup_oom_unlock(memcg);
1934 * There is no guarantee that an OOM-lock contender
1935 * sees the wakeups triggered by the OOM kill
1936 * uncharges. Wake any sleepers explicitely.
1938 memcg_oom_recover(memcg);
1941 * A system call can just return -ENOMEM, but if this
1942 * is a page fault and somebody else is handling the
1943 * OOM already, we need to sleep on the OOM waitqueue
1944 * for this memcg until the situation is resolved.
1945 * Which can take some time because it might be
1946 * handled by a userspace task.
1948 * However, this is the charge context, which means
1949 * that we may sit on a large call stack and hold
1950 * various filesystem locks, the mmap_sem etc. and we
1951 * don't want the OOM handler to deadlock on them
1952 * while we sit here and wait. Store the current OOM
1953 * context in the task_struct, then return -ENOMEM.
1954 * At the end of the page fault handler, with the
1955 * stack unwound, pagefault_out_of_memory() will check
1956 * back with us by calling
1957 * mem_cgroup_oom_synchronize(), possibly putting the
1960 current->memcg_oom.oom_locked = locked;
1961 current->memcg_oom.wakeups = wakeups;
1962 css_get(&memcg->css);
1963 current->memcg_oom.wait_on_memcg = memcg;
1968 * mem_cgroup_oom_synchronize - complete memcg OOM handling
1970 * This has to be called at the end of a page fault if the the memcg
1971 * OOM handler was enabled and the fault is returning %VM_FAULT_OOM.
1973 * Memcg supports userspace OOM handling, so failed allocations must
1974 * sleep on a waitqueue until the userspace task resolves the
1975 * situation. Sleeping directly in the charge context with all kinds
1976 * of locks held is not a good idea, instead we remember an OOM state
1977 * in the task and mem_cgroup_oom_synchronize() has to be called at
1978 * the end of the page fault to put the task to sleep and clean up the
1981 * Returns %true if an ongoing memcg OOM situation was detected and
1982 * finalized, %false otherwise.
1984 bool mem_cgroup_oom_synchronize(void)
1986 struct oom_wait_info owait;
1987 struct mem_cgroup *memcg;
1989 /* OOM is global, do not handle */
1990 if (!current->memcg_oom.in_memcg_oom)
1994 * We invoked the OOM killer but there is a chance that a kill
1995 * did not free up any charges. Everybody else might already
1996 * be sleeping, so restart the fault and keep the rampage
1997 * going until some charges are released.
1999 memcg = current->memcg_oom.wait_on_memcg;
2003 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2006 owait.memcg = memcg;
2007 owait.wait.flags = 0;
2008 owait.wait.func = memcg_oom_wake_function;
2009 owait.wait.private = current;
2010 INIT_LIST_HEAD(&owait.wait.task_list);
2012 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2013 /* Only sleep if we didn't miss any wakeups since OOM */
2014 if (atomic_read(&memcg->oom_wakeups) == current->memcg_oom.wakeups)
2016 finish_wait(&memcg_oom_waitq, &owait.wait);
2018 mem_cgroup_unmark_under_oom(memcg);
2019 if (current->memcg_oom.oom_locked) {
2020 mem_cgroup_oom_unlock(memcg);
2022 * There is no guarantee that an OOM-lock contender
2023 * sees the wakeups triggered by the OOM kill
2024 * uncharges. Wake any sleepers explicitely.
2026 memcg_oom_recover(memcg);
2028 css_put(&memcg->css);
2029 current->memcg_oom.wait_on_memcg = NULL;
2031 current->memcg_oom.in_memcg_oom = 0;
2036 * Currently used to update mapped file statistics, but the routine can be
2037 * generalized to update other statistics as well.
2039 * Notes: Race condition
2041 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2042 * it tends to be costly. But considering some conditions, we doesn't need
2043 * to do so _always_.
2045 * Considering "charge", lock_page_cgroup() is not required because all
2046 * file-stat operations happen after a page is attached to radix-tree. There
2047 * are no race with "charge".
2049 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2050 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2051 * if there are race with "uncharge". Statistics itself is properly handled
2054 * Considering "move", this is an only case we see a race. To make the race
2055 * small, we check mm->moving_account and detect there are possibility of race
2056 * If there is, we take a lock.
2059 void __mem_cgroup_begin_update_page_stat(struct page *page,
2060 bool *locked, unsigned long *flags)
2062 struct mem_cgroup *memcg;
2063 struct page_cgroup *pc;
2065 pc = lookup_page_cgroup(page);
2067 memcg = pc->mem_cgroup;
2068 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2071 * If this memory cgroup is not under account moving, we don't
2072 * need to take move_lock_mem_cgroup(). Because we already hold
2073 * rcu_read_lock(), any calls to move_account will be delayed until
2074 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2076 if (!mem_cgroup_stolen(memcg))
2079 move_lock_mem_cgroup(memcg, flags);
2080 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2081 move_unlock_mem_cgroup(memcg, flags);
2087 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2089 struct page_cgroup *pc = lookup_page_cgroup(page);
2092 * It's guaranteed that pc->mem_cgroup never changes while
2093 * lock is held because a routine modifies pc->mem_cgroup
2094 * should take move_lock_mem_cgroup().
2096 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2099 void mem_cgroup_update_page_stat(struct page *page,
2100 enum mem_cgroup_stat_index idx, int val)
2102 struct mem_cgroup *memcg;
2103 struct page_cgroup *pc = lookup_page_cgroup(page);
2104 unsigned long uninitialized_var(flags);
2106 if (mem_cgroup_disabled())
2109 VM_BUG_ON(!rcu_read_lock_held());
2110 memcg = pc->mem_cgroup;
2111 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2114 this_cpu_add(memcg->stat->count[idx], val);
2118 * size of first charge trial. "32" comes from vmscan.c's magic value.
2119 * TODO: maybe necessary to use big numbers in big irons.
2121 #define CHARGE_BATCH 32U
2122 struct memcg_stock_pcp {
2123 struct mem_cgroup *cached; /* this never be root cgroup */
2124 unsigned int nr_pages;
2125 struct work_struct work;
2126 unsigned long flags;
2127 #define FLUSHING_CACHED_CHARGE 0
2129 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2130 static DEFINE_MUTEX(percpu_charge_mutex);
2133 * consume_stock: Try to consume stocked charge on this cpu.
2134 * @memcg: memcg to consume from.
2135 * @nr_pages: how many pages to charge.
2137 * The charges will only happen if @memcg matches the current cpu's memcg
2138 * stock, and at least @nr_pages are available in that stock. Failure to
2139 * service an allocation will refill the stock.
2141 * returns true if successful, false otherwise.
2143 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2145 struct memcg_stock_pcp *stock;
2148 if (nr_pages > CHARGE_BATCH)
2151 stock = &get_cpu_var(memcg_stock);
2152 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2153 stock->nr_pages -= nr_pages;
2154 else /* need to call res_counter_charge */
2156 put_cpu_var(memcg_stock);
2161 * Returns stocks cached in percpu to res_counter and reset cached information.
2163 static void drain_stock(struct memcg_stock_pcp *stock)
2165 struct mem_cgroup *old = stock->cached;
2167 if (stock->nr_pages) {
2168 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2170 res_counter_uncharge(&old->res, bytes);
2171 if (do_swap_account)
2172 res_counter_uncharge(&old->memsw, bytes);
2173 stock->nr_pages = 0;
2175 stock->cached = NULL;
2179 * This must be called under preempt disabled or must be called by
2180 * a thread which is pinned to local cpu.
2182 static void drain_local_stock(struct work_struct *dummy)
2184 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2186 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2189 static void __init memcg_stock_init(void)
2193 for_each_possible_cpu(cpu) {
2194 struct memcg_stock_pcp *stock =
2195 &per_cpu(memcg_stock, cpu);
2196 INIT_WORK(&stock->work, drain_local_stock);
2201 * Cache charges(val) which is from res_counter, to local per_cpu area.
2202 * This will be consumed by consume_stock() function, later.
2204 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2206 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2208 if (stock->cached != memcg) { /* reset if necessary */
2210 stock->cached = memcg;
2212 stock->nr_pages += nr_pages;
2213 put_cpu_var(memcg_stock);
2217 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2218 * of the hierarchy under it. sync flag says whether we should block
2219 * until the work is done.
2221 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2225 /* Notify other cpus that system-wide "drain" is running */
2228 for_each_online_cpu(cpu) {
2229 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2230 struct mem_cgroup *memcg;
2232 memcg = stock->cached;
2233 if (!memcg || !stock->nr_pages)
2235 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2237 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2239 drain_local_stock(&stock->work);
2241 schedule_work_on(cpu, &stock->work);
2249 for_each_online_cpu(cpu) {
2250 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2251 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2252 flush_work(&stock->work);
2259 * Tries to drain stocked charges in other cpus. This function is asynchronous
2260 * and just put a work per cpu for draining localy on each cpu. Caller can
2261 * expects some charges will be back to res_counter later but cannot wait for
2264 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2267 * If someone calls draining, avoid adding more kworker runs.
2269 if (!mutex_trylock(&percpu_charge_mutex))
2271 drain_all_stock(root_memcg, false);
2272 mutex_unlock(&percpu_charge_mutex);
2275 /* This is a synchronous drain interface. */
2276 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2278 /* called when force_empty is called */
2279 mutex_lock(&percpu_charge_mutex);
2280 drain_all_stock(root_memcg, true);
2281 mutex_unlock(&percpu_charge_mutex);
2285 * This function drains percpu counter value from DEAD cpu and
2286 * move it to local cpu. Note that this function can be preempted.
2288 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2292 spin_lock(&memcg->pcp_counter_lock);
2293 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2294 long x = per_cpu(memcg->stat->count[i], cpu);
2296 per_cpu(memcg->stat->count[i], cpu) = 0;
2297 memcg->nocpu_base.count[i] += x;
2299 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2300 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2302 per_cpu(memcg->stat->events[i], cpu) = 0;
2303 memcg->nocpu_base.events[i] += x;
2305 spin_unlock(&memcg->pcp_counter_lock);
2308 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2309 unsigned long action,
2312 int cpu = (unsigned long)hcpu;
2313 struct memcg_stock_pcp *stock;
2314 struct mem_cgroup *iter;
2316 if (action == CPU_ONLINE)
2319 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2322 for_each_mem_cgroup(iter)
2323 mem_cgroup_drain_pcp_counter(iter, cpu);
2325 stock = &per_cpu(memcg_stock, cpu);
2331 /* See __mem_cgroup_try_charge() for details */
2333 CHARGE_OK, /* success */
2334 CHARGE_RETRY, /* need to retry but retry is not bad */
2335 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2336 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2339 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2340 unsigned int nr_pages, unsigned int min_pages,
2343 unsigned long csize = nr_pages * PAGE_SIZE;
2344 struct mem_cgroup *mem_over_limit;
2345 struct res_counter *fail_res;
2346 unsigned long flags = 0;
2349 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2352 if (!do_swap_account)
2354 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2358 res_counter_uncharge(&memcg->res, csize);
2359 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2360 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2362 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2364 * Never reclaim on behalf of optional batching, retry with a
2365 * single page instead.
2367 if (nr_pages > min_pages)
2368 return CHARGE_RETRY;
2370 if (!(gfp_mask & __GFP_WAIT))
2371 return CHARGE_WOULDBLOCK;
2373 if (gfp_mask & __GFP_NORETRY)
2374 return CHARGE_NOMEM;
2376 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2377 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2378 return CHARGE_RETRY;
2380 * Even though the limit is exceeded at this point, reclaim
2381 * may have been able to free some pages. Retry the charge
2382 * before killing the task.
2384 * Only for regular pages, though: huge pages are rather
2385 * unlikely to succeed so close to the limit, and we fall back
2386 * to regular pages anyway in case of failure.
2388 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2389 return CHARGE_RETRY;
2392 * At task move, charge accounts can be doubly counted. So, it's
2393 * better to wait until the end of task_move if something is going on.
2395 if (mem_cgroup_wait_acct_move(mem_over_limit))
2396 return CHARGE_RETRY;
2399 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2401 return CHARGE_NOMEM;
2405 * __mem_cgroup_try_charge() does
2406 * 1. detect memcg to be charged against from passed *mm and *ptr,
2407 * 2. update res_counter
2408 * 3. call memory reclaim if necessary.
2410 * In some special case, if the task is fatal, fatal_signal_pending() or
2411 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2412 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2413 * as possible without any hazards. 2: all pages should have a valid
2414 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2415 * pointer, that is treated as a charge to root_mem_cgroup.
2417 * So __mem_cgroup_try_charge() will return
2418 * 0 ... on success, filling *ptr with a valid memcg pointer.
2419 * -ENOMEM ... charge failure because of resource limits.
2420 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2422 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2423 * the oom-killer can be invoked.
2425 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2427 unsigned int nr_pages,
2428 struct mem_cgroup **ptr,
2431 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2432 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2433 struct mem_cgroup *memcg = NULL;
2437 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2438 * in system level. So, allow to go ahead dying process in addition to
2441 if (unlikely(test_thread_flag(TIF_MEMDIE)
2442 || fatal_signal_pending(current)))
2446 * We always charge the cgroup the mm_struct belongs to.
2447 * The mm_struct's mem_cgroup changes on task migration if the
2448 * thread group leader migrates. It's possible that mm is not
2449 * set, if so charge the root memcg (happens for pagecache usage).
2452 *ptr = root_mem_cgroup;
2454 if (*ptr) { /* css should be a valid one */
2456 if (mem_cgroup_is_root(memcg))
2458 if (consume_stock(memcg, nr_pages))
2460 css_get(&memcg->css);
2462 struct task_struct *p;
2465 p = rcu_dereference(mm->owner);
2467 * Because we don't have task_lock(), "p" can exit.
2468 * In that case, "memcg" can point to root or p can be NULL with
2469 * race with swapoff. Then, we have small risk of mis-accouning.
2470 * But such kind of mis-account by race always happens because
2471 * we don't have cgroup_mutex(). It's overkill and we allo that
2473 * (*) swapoff at el will charge against mm-struct not against
2474 * task-struct. So, mm->owner can be NULL.
2476 memcg = mem_cgroup_from_task(p);
2478 memcg = root_mem_cgroup;
2479 if (mem_cgroup_is_root(memcg)) {
2483 if (consume_stock(memcg, nr_pages)) {
2485 * It seems dagerous to access memcg without css_get().
2486 * But considering how consume_stok works, it's not
2487 * necessary. If consume_stock success, some charges
2488 * from this memcg are cached on this cpu. So, we
2489 * don't need to call css_get()/css_tryget() before
2490 * calling consume_stock().
2495 /* after here, we may be blocked. we need to get refcnt */
2496 if (!css_tryget(&memcg->css)) {
2504 bool invoke_oom = oom && !nr_oom_retries;
2506 /* If killed, bypass charge */
2507 if (fatal_signal_pending(current)) {
2508 css_put(&memcg->css);
2512 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2513 nr_pages, invoke_oom);
2517 case CHARGE_RETRY: /* not in OOM situation but retry */
2519 css_put(&memcg->css);
2522 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2523 css_put(&memcg->css);
2525 case CHARGE_NOMEM: /* OOM routine works */
2526 if (!oom || invoke_oom) {
2527 css_put(&memcg->css);
2533 } while (ret != CHARGE_OK);
2535 if (batch > nr_pages)
2536 refill_stock(memcg, batch - nr_pages);
2537 css_put(&memcg->css);
2545 *ptr = root_mem_cgroup;
2550 * Somemtimes we have to undo a charge we got by try_charge().
2551 * This function is for that and do uncharge, put css's refcnt.
2552 * gotten by try_charge().
2554 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2555 unsigned int nr_pages)
2557 if (!mem_cgroup_is_root(memcg)) {
2558 unsigned long bytes = nr_pages * PAGE_SIZE;
2560 res_counter_uncharge(&memcg->res, bytes);
2561 if (do_swap_account)
2562 res_counter_uncharge(&memcg->memsw, bytes);
2567 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2568 * This is useful when moving usage to parent cgroup.
2570 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2571 unsigned int nr_pages)
2573 unsigned long bytes = nr_pages * PAGE_SIZE;
2575 if (mem_cgroup_is_root(memcg))
2578 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2579 if (do_swap_account)
2580 res_counter_uncharge_until(&memcg->memsw,
2581 memcg->memsw.parent, bytes);
2585 * A helper function to get mem_cgroup from ID. must be called under
2586 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2587 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2588 * called against removed memcg.)
2590 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2592 struct cgroup_subsys_state *css;
2594 /* ID 0 is unused ID */
2597 css = css_lookup(&mem_cgroup_subsys, id);
2600 return mem_cgroup_from_css(css);
2603 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2605 struct mem_cgroup *memcg = NULL;
2606 struct page_cgroup *pc;
2610 VM_BUG_ON(!PageLocked(page));
2612 pc = lookup_page_cgroup(page);
2613 lock_page_cgroup(pc);
2614 if (PageCgroupUsed(pc)) {
2615 memcg = pc->mem_cgroup;
2616 if (memcg && !css_tryget(&memcg->css))
2618 } else if (PageSwapCache(page)) {
2619 ent.val = page_private(page);
2620 id = lookup_swap_cgroup_id(ent);
2622 memcg = mem_cgroup_lookup(id);
2623 if (memcg && !css_tryget(&memcg->css))
2627 unlock_page_cgroup(pc);
2631 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2633 unsigned int nr_pages,
2634 enum charge_type ctype,
2637 struct page_cgroup *pc = lookup_page_cgroup(page);
2638 struct zone *uninitialized_var(zone);
2639 struct lruvec *lruvec;
2640 bool was_on_lru = false;
2643 lock_page_cgroup(pc);
2644 VM_BUG_ON(PageCgroupUsed(pc));
2646 * we don't need page_cgroup_lock about tail pages, becase they are not
2647 * accessed by any other context at this point.
2651 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2652 * may already be on some other mem_cgroup's LRU. Take care of it.
2655 zone = page_zone(page);
2656 spin_lock_irq(&zone->lru_lock);
2657 if (PageLRU(page)) {
2658 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2660 del_page_from_lru_list(page, lruvec, page_lru(page));
2665 pc->mem_cgroup = memcg;
2667 * We access a page_cgroup asynchronously without lock_page_cgroup().
2668 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2669 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2670 * before USED bit, we need memory barrier here.
2671 * See mem_cgroup_add_lru_list(), etc.
2674 SetPageCgroupUsed(pc);
2678 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2679 VM_BUG_ON(PageLRU(page));
2681 add_page_to_lru_list(page, lruvec, page_lru(page));
2683 spin_unlock_irq(&zone->lru_lock);
2686 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2691 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2692 unlock_page_cgroup(pc);
2695 * "charge_statistics" updated event counter.
2697 memcg_check_events(memcg, page);
2700 static DEFINE_MUTEX(set_limit_mutex);
2702 #ifdef CONFIG_MEMCG_KMEM
2703 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2705 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2706 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2710 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2711 * in the memcg_cache_params struct.
2713 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2715 struct kmem_cache *cachep;
2717 VM_BUG_ON(p->is_root_cache);
2718 cachep = p->root_cache;
2719 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2722 #ifdef CONFIG_SLABINFO
2723 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2724 struct cftype *cft, struct seq_file *m)
2726 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2727 struct memcg_cache_params *params;
2729 if (!memcg_can_account_kmem(memcg))
2732 print_slabinfo_header(m);
2734 mutex_lock(&memcg->slab_caches_mutex);
2735 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2736 cache_show(memcg_params_to_cache(params), m);
2737 mutex_unlock(&memcg->slab_caches_mutex);
2743 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2745 struct res_counter *fail_res;
2746 struct mem_cgroup *_memcg;
2750 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2755 * Conditions under which we can wait for the oom_killer. Those are
2756 * the same conditions tested by the core page allocator
2758 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2761 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2764 if (ret == -EINTR) {
2766 * __mem_cgroup_try_charge() chosed to bypass to root due to
2767 * OOM kill or fatal signal. Since our only options are to
2768 * either fail the allocation or charge it to this cgroup, do
2769 * it as a temporary condition. But we can't fail. From a
2770 * kmem/slab perspective, the cache has already been selected,
2771 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2774 * This condition will only trigger if the task entered
2775 * memcg_charge_kmem in a sane state, but was OOM-killed during
2776 * __mem_cgroup_try_charge() above. Tasks that were already
2777 * dying when the allocation triggers should have been already
2778 * directed to the root cgroup in memcontrol.h
2780 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2781 if (do_swap_account)
2782 res_counter_charge_nofail(&memcg->memsw, size,
2786 res_counter_uncharge(&memcg->kmem, size);
2791 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2793 res_counter_uncharge(&memcg->res, size);
2794 if (do_swap_account)
2795 res_counter_uncharge(&memcg->memsw, size);
2798 if (res_counter_uncharge(&memcg->kmem, size))
2802 * Releases a reference taken in kmem_cgroup_css_offline in case
2803 * this last uncharge is racing with the offlining code or it is
2804 * outliving the memcg existence.
2806 * The memory barrier imposed by test&clear is paired with the
2807 * explicit one in memcg_kmem_mark_dead().
2809 if (memcg_kmem_test_and_clear_dead(memcg))
2810 css_put(&memcg->css);
2813 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2818 mutex_lock(&memcg->slab_caches_mutex);
2819 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2820 mutex_unlock(&memcg->slab_caches_mutex);
2824 * helper for acessing a memcg's index. It will be used as an index in the
2825 * child cache array in kmem_cache, and also to derive its name. This function
2826 * will return -1 when this is not a kmem-limited memcg.
2828 int memcg_cache_id(struct mem_cgroup *memcg)
2830 return memcg ? memcg->kmemcg_id : -1;
2834 * This ends up being protected by the set_limit mutex, during normal
2835 * operation, because that is its main call site.
2837 * But when we create a new cache, we can call this as well if its parent
2838 * is kmem-limited. That will have to hold set_limit_mutex as well.
2840 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2844 num = ida_simple_get(&kmem_limited_groups,
2845 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2849 * After this point, kmem_accounted (that we test atomically in
2850 * the beginning of this conditional), is no longer 0. This
2851 * guarantees only one process will set the following boolean
2852 * to true. We don't need test_and_set because we're protected
2853 * by the set_limit_mutex anyway.
2855 memcg_kmem_set_activated(memcg);
2857 ret = memcg_update_all_caches(num+1);
2859 ida_simple_remove(&kmem_limited_groups, num);
2860 memcg_kmem_clear_activated(memcg);
2864 memcg->kmemcg_id = num;
2865 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2866 mutex_init(&memcg->slab_caches_mutex);
2870 static size_t memcg_caches_array_size(int num_groups)
2873 if (num_groups <= 0)
2876 size = 2 * num_groups;
2877 if (size < MEMCG_CACHES_MIN_SIZE)
2878 size = MEMCG_CACHES_MIN_SIZE;
2879 else if (size > MEMCG_CACHES_MAX_SIZE)
2880 size = MEMCG_CACHES_MAX_SIZE;
2886 * We should update the current array size iff all caches updates succeed. This
2887 * can only be done from the slab side. The slab mutex needs to be held when
2890 void memcg_update_array_size(int num)
2892 if (num > memcg_limited_groups_array_size)
2893 memcg_limited_groups_array_size = memcg_caches_array_size(num);
2896 static void kmem_cache_destroy_work_func(struct work_struct *w);
2898 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
2900 struct memcg_cache_params *cur_params = s->memcg_params;
2902 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
2904 if (num_groups > memcg_limited_groups_array_size) {
2906 ssize_t size = memcg_caches_array_size(num_groups);
2908 size *= sizeof(void *);
2909 size += offsetof(struct memcg_cache_params, memcg_caches);
2911 s->memcg_params = kzalloc(size, GFP_KERNEL);
2912 if (!s->memcg_params) {
2913 s->memcg_params = cur_params;
2917 s->memcg_params->is_root_cache = true;
2920 * There is the chance it will be bigger than
2921 * memcg_limited_groups_array_size, if we failed an allocation
2922 * in a cache, in which case all caches updated before it, will
2923 * have a bigger array.
2925 * But if that is the case, the data after
2926 * memcg_limited_groups_array_size is certainly unused
2928 for (i = 0; i < memcg_limited_groups_array_size; i++) {
2929 if (!cur_params->memcg_caches[i])
2931 s->memcg_params->memcg_caches[i] =
2932 cur_params->memcg_caches[i];
2936 * Ideally, we would wait until all caches succeed, and only
2937 * then free the old one. But this is not worth the extra
2938 * pointer per-cache we'd have to have for this.
2940 * It is not a big deal if some caches are left with a size
2941 * bigger than the others. And all updates will reset this
2949 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
2950 struct kmem_cache *root_cache)
2954 if (!memcg_kmem_enabled())
2958 size = offsetof(struct memcg_cache_params, memcg_caches);
2959 size += memcg_limited_groups_array_size * sizeof(void *);
2961 size = sizeof(struct memcg_cache_params);
2963 s->memcg_params = kzalloc(size, GFP_KERNEL);
2964 if (!s->memcg_params)
2968 s->memcg_params->memcg = memcg;
2969 s->memcg_params->root_cache = root_cache;
2970 INIT_WORK(&s->memcg_params->destroy,
2971 kmem_cache_destroy_work_func);
2973 s->memcg_params->is_root_cache = true;
2978 void memcg_release_cache(struct kmem_cache *s)
2980 struct kmem_cache *root;
2981 struct mem_cgroup *memcg;
2985 * This happens, for instance, when a root cache goes away before we
2988 if (!s->memcg_params)
2991 if (s->memcg_params->is_root_cache)
2994 memcg = s->memcg_params->memcg;
2995 id = memcg_cache_id(memcg);
2997 root = s->memcg_params->root_cache;
2998 root->memcg_params->memcg_caches[id] = NULL;
3000 mutex_lock(&memcg->slab_caches_mutex);
3001 list_del(&s->memcg_params->list);
3002 mutex_unlock(&memcg->slab_caches_mutex);
3004 css_put(&memcg->css);
3006 kfree(s->memcg_params);
3010 * During the creation a new cache, we need to disable our accounting mechanism
3011 * altogether. This is true even if we are not creating, but rather just
3012 * enqueing new caches to be created.
3014 * This is because that process will trigger allocations; some visible, like
3015 * explicit kmallocs to auxiliary data structures, name strings and internal
3016 * cache structures; some well concealed, like INIT_WORK() that can allocate
3017 * objects during debug.
3019 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3020 * to it. This may not be a bounded recursion: since the first cache creation
3021 * failed to complete (waiting on the allocation), we'll just try to create the
3022 * cache again, failing at the same point.
3024 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3025 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3026 * inside the following two functions.
3028 static inline void memcg_stop_kmem_account(void)
3030 VM_BUG_ON(!current->mm);
3031 current->memcg_kmem_skip_account++;
3034 static inline void memcg_resume_kmem_account(void)
3036 VM_BUG_ON(!current->mm);
3037 current->memcg_kmem_skip_account--;
3040 static void kmem_cache_destroy_work_func(struct work_struct *w)
3042 struct kmem_cache *cachep;
3043 struct memcg_cache_params *p;
3045 p = container_of(w, struct memcg_cache_params, destroy);
3047 cachep = memcg_params_to_cache(p);
3050 * If we get down to 0 after shrink, we could delete right away.
3051 * However, memcg_release_pages() already puts us back in the workqueue
3052 * in that case. If we proceed deleting, we'll get a dangling
3053 * reference, and removing the object from the workqueue in that case
3054 * is unnecessary complication. We are not a fast path.
3056 * Note that this case is fundamentally different from racing with
3057 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3058 * kmem_cache_shrink, not only we would be reinserting a dead cache
3059 * into the queue, but doing so from inside the worker racing to
3062 * So if we aren't down to zero, we'll just schedule a worker and try
3065 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3066 kmem_cache_shrink(cachep);
3067 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3070 kmem_cache_destroy(cachep);
3073 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3075 if (!cachep->memcg_params->dead)
3079 * There are many ways in which we can get here.
3081 * We can get to a memory-pressure situation while the delayed work is
3082 * still pending to run. The vmscan shrinkers can then release all
3083 * cache memory and get us to destruction. If this is the case, we'll
3084 * be executed twice, which is a bug (the second time will execute over
3085 * bogus data). In this case, cancelling the work should be fine.
3087 * But we can also get here from the worker itself, if
3088 * kmem_cache_shrink is enough to shake all the remaining objects and
3089 * get the page count to 0. In this case, we'll deadlock if we try to
3090 * cancel the work (the worker runs with an internal lock held, which
3091 * is the same lock we would hold for cancel_work_sync().)
3093 * Since we can't possibly know who got us here, just refrain from
3094 * running if there is already work pending
3096 if (work_pending(&cachep->memcg_params->destroy))
3099 * We have to defer the actual destroying to a workqueue, because
3100 * we might currently be in a context that cannot sleep.
3102 schedule_work(&cachep->memcg_params->destroy);
3106 * This lock protects updaters, not readers. We want readers to be as fast as
3107 * they can, and they will either see NULL or a valid cache value. Our model
3108 * allow them to see NULL, in which case the root memcg will be selected.
3110 * We need this lock because multiple allocations to the same cache from a non
3111 * will span more than one worker. Only one of them can create the cache.
3113 static DEFINE_MUTEX(memcg_cache_mutex);
3116 * Called with memcg_cache_mutex held
3118 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3119 struct kmem_cache *s)
3121 struct kmem_cache *new;
3122 static char *tmp_name = NULL;
3124 lockdep_assert_held(&memcg_cache_mutex);
3127 * kmem_cache_create_memcg duplicates the given name and
3128 * cgroup_name for this name requires RCU context.
3129 * This static temporary buffer is used to prevent from
3130 * pointless shortliving allocation.
3133 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3139 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3140 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3143 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3144 (s->flags & ~SLAB_PANIC), s->ctor, s);
3147 new->allocflags |= __GFP_KMEMCG;
3152 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3153 struct kmem_cache *cachep)
3155 struct kmem_cache *new_cachep;
3158 BUG_ON(!memcg_can_account_kmem(memcg));
3160 idx = memcg_cache_id(memcg);
3162 mutex_lock(&memcg_cache_mutex);
3163 new_cachep = cachep->memcg_params->memcg_caches[idx];
3165 css_put(&memcg->css);
3169 new_cachep = kmem_cache_dup(memcg, cachep);
3170 if (new_cachep == NULL) {
3171 new_cachep = cachep;
3172 css_put(&memcg->css);
3176 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3178 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3180 * the readers won't lock, make sure everybody sees the updated value,
3181 * so they won't put stuff in the queue again for no reason
3185 mutex_unlock(&memcg_cache_mutex);
3189 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3191 struct kmem_cache *c;
3194 if (!s->memcg_params)
3196 if (!s->memcg_params->is_root_cache)
3200 * If the cache is being destroyed, we trust that there is no one else
3201 * requesting objects from it. Even if there are, the sanity checks in
3202 * kmem_cache_destroy should caught this ill-case.
3204 * Still, we don't want anyone else freeing memcg_caches under our
3205 * noses, which can happen if a new memcg comes to life. As usual,
3206 * we'll take the set_limit_mutex to protect ourselves against this.
3208 mutex_lock(&set_limit_mutex);
3209 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3210 c = s->memcg_params->memcg_caches[i];
3215 * We will now manually delete the caches, so to avoid races
3216 * we need to cancel all pending destruction workers and
3217 * proceed with destruction ourselves.
3219 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3220 * and that could spawn the workers again: it is likely that
3221 * the cache still have active pages until this very moment.
3222 * This would lead us back to mem_cgroup_destroy_cache.
3224 * But that will not execute at all if the "dead" flag is not
3225 * set, so flip it down to guarantee we are in control.
3227 c->memcg_params->dead = false;
3228 cancel_work_sync(&c->memcg_params->destroy);
3229 kmem_cache_destroy(c);
3231 mutex_unlock(&set_limit_mutex);
3234 struct create_work {
3235 struct mem_cgroup *memcg;
3236 struct kmem_cache *cachep;
3237 struct work_struct work;
3240 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3242 struct kmem_cache *cachep;
3243 struct memcg_cache_params *params;
3245 if (!memcg_kmem_is_active(memcg))
3248 mutex_lock(&memcg->slab_caches_mutex);
3249 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3250 cachep = memcg_params_to_cache(params);
3251 cachep->memcg_params->dead = true;
3252 schedule_work(&cachep->memcg_params->destroy);
3254 mutex_unlock(&memcg->slab_caches_mutex);
3257 static void memcg_create_cache_work_func(struct work_struct *w)
3259 struct create_work *cw;
3261 cw = container_of(w, struct create_work, work);
3262 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3267 * Enqueue the creation of a per-memcg kmem_cache.
3269 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3270 struct kmem_cache *cachep)
3272 struct create_work *cw;
3274 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3276 css_put(&memcg->css);
3281 cw->cachep = cachep;
3283 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3284 schedule_work(&cw->work);
3287 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3288 struct kmem_cache *cachep)
3291 * We need to stop accounting when we kmalloc, because if the
3292 * corresponding kmalloc cache is not yet created, the first allocation
3293 * in __memcg_create_cache_enqueue will recurse.
3295 * However, it is better to enclose the whole function. Depending on
3296 * the debugging options enabled, INIT_WORK(), for instance, can
3297 * trigger an allocation. This too, will make us recurse. Because at
3298 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3299 * the safest choice is to do it like this, wrapping the whole function.
3301 memcg_stop_kmem_account();
3302 __memcg_create_cache_enqueue(memcg, cachep);
3303 memcg_resume_kmem_account();
3306 * Return the kmem_cache we're supposed to use for a slab allocation.
3307 * We try to use the current memcg's version of the cache.
3309 * If the cache does not exist yet, if we are the first user of it,
3310 * we either create it immediately, if possible, or create it asynchronously
3312 * In the latter case, we will let the current allocation go through with
3313 * the original cache.
3315 * Can't be called in interrupt context or from kernel threads.
3316 * This function needs to be called with rcu_read_lock() held.
3318 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3321 struct mem_cgroup *memcg;
3324 VM_BUG_ON(!cachep->memcg_params);
3325 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3327 if (!current->mm || current->memcg_kmem_skip_account)
3331 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3333 if (!memcg_can_account_kmem(memcg))
3336 idx = memcg_cache_id(memcg);
3339 * barrier to mare sure we're always seeing the up to date value. The
3340 * code updating memcg_caches will issue a write barrier to match this.
3342 read_barrier_depends();
3343 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3344 cachep = cachep->memcg_params->memcg_caches[idx];
3348 /* The corresponding put will be done in the workqueue. */
3349 if (!css_tryget(&memcg->css))
3354 * If we are in a safe context (can wait, and not in interrupt
3355 * context), we could be be predictable and return right away.
3356 * This would guarantee that the allocation being performed
3357 * already belongs in the new cache.
3359 * However, there are some clashes that can arrive from locking.
3360 * For instance, because we acquire the slab_mutex while doing
3361 * kmem_cache_dup, this means no further allocation could happen
3362 * with the slab_mutex held.
3364 * Also, because cache creation issue get_online_cpus(), this
3365 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3366 * that ends up reversed during cpu hotplug. (cpuset allocates
3367 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3368 * better to defer everything.
3370 memcg_create_cache_enqueue(memcg, cachep);
3376 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3379 * We need to verify if the allocation against current->mm->owner's memcg is
3380 * possible for the given order. But the page is not allocated yet, so we'll
3381 * need a further commit step to do the final arrangements.
3383 * It is possible for the task to switch cgroups in this mean time, so at
3384 * commit time, we can't rely on task conversion any longer. We'll then use
3385 * the handle argument to return to the caller which cgroup we should commit
3386 * against. We could also return the memcg directly and avoid the pointer
3387 * passing, but a boolean return value gives better semantics considering
3388 * the compiled-out case as well.
3390 * Returning true means the allocation is possible.
3393 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3395 struct mem_cgroup *memcg;
3401 * Disabling accounting is only relevant for some specific memcg
3402 * internal allocations. Therefore we would initially not have such
3403 * check here, since direct calls to the page allocator that are marked
3404 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3405 * concerned with cache allocations, and by having this test at
3406 * memcg_kmem_get_cache, we are already able to relay the allocation to
3407 * the root cache and bypass the memcg cache altogether.
3409 * There is one exception, though: the SLUB allocator does not create
3410 * large order caches, but rather service large kmallocs directly from
3411 * the page allocator. Therefore, the following sequence when backed by
3412 * the SLUB allocator:
3414 * memcg_stop_kmem_account();
3415 * kmalloc(<large_number>)
3416 * memcg_resume_kmem_account();
3418 * would effectively ignore the fact that we should skip accounting,
3419 * since it will drive us directly to this function without passing
3420 * through the cache selector memcg_kmem_get_cache. Such large
3421 * allocations are extremely rare but can happen, for instance, for the
3422 * cache arrays. We bring this test here.
3424 if (!current->mm || current->memcg_kmem_skip_account)
3427 memcg = try_get_mem_cgroup_from_mm(current->mm);
3430 * very rare case described in mem_cgroup_from_task. Unfortunately there
3431 * isn't much we can do without complicating this too much, and it would
3432 * be gfp-dependent anyway. Just let it go
3434 if (unlikely(!memcg))
3437 if (!memcg_can_account_kmem(memcg)) {
3438 css_put(&memcg->css);
3442 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3446 css_put(&memcg->css);
3450 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3453 struct page_cgroup *pc;
3455 VM_BUG_ON(mem_cgroup_is_root(memcg));
3457 /* The page allocation failed. Revert */
3459 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3463 pc = lookup_page_cgroup(page);
3464 lock_page_cgroup(pc);
3465 pc->mem_cgroup = memcg;
3466 SetPageCgroupUsed(pc);
3467 unlock_page_cgroup(pc);
3470 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3472 struct mem_cgroup *memcg = NULL;
3473 struct page_cgroup *pc;
3476 pc = lookup_page_cgroup(page);
3478 * Fast unlocked return. Theoretically might have changed, have to
3479 * check again after locking.
3481 if (!PageCgroupUsed(pc))
3484 lock_page_cgroup(pc);
3485 if (PageCgroupUsed(pc)) {
3486 memcg = pc->mem_cgroup;
3487 ClearPageCgroupUsed(pc);
3489 unlock_page_cgroup(pc);
3492 * We trust that only if there is a memcg associated with the page, it
3493 * is a valid allocation
3498 VM_BUG_ON(mem_cgroup_is_root(memcg));
3499 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3502 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3505 #endif /* CONFIG_MEMCG_KMEM */
3507 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3509 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3511 * Because tail pages are not marked as "used", set it. We're under
3512 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3513 * charge/uncharge will be never happen and move_account() is done under
3514 * compound_lock(), so we don't have to take care of races.
3516 void mem_cgroup_split_huge_fixup(struct page *head)
3518 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3519 struct page_cgroup *pc;
3520 struct mem_cgroup *memcg;
3523 if (mem_cgroup_disabled())
3526 memcg = head_pc->mem_cgroup;
3527 for (i = 1; i < HPAGE_PMD_NR; i++) {
3529 pc->mem_cgroup = memcg;
3530 smp_wmb();/* see __commit_charge() */
3531 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3533 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3536 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3539 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3540 struct mem_cgroup *to,
3541 unsigned int nr_pages,
3542 enum mem_cgroup_stat_index idx)
3544 /* Update stat data for mem_cgroup */
3546 WARN_ON_ONCE(from->stat->count[idx] < nr_pages);
3547 __this_cpu_add(from->stat->count[idx], -nr_pages);
3548 __this_cpu_add(to->stat->count[idx], nr_pages);
3553 * mem_cgroup_move_account - move account of the page
3555 * @nr_pages: number of regular pages (>1 for huge pages)
3556 * @pc: page_cgroup of the page.
3557 * @from: mem_cgroup which the page is moved from.
3558 * @to: mem_cgroup which the page is moved to. @from != @to.
3560 * The caller must confirm following.
3561 * - page is not on LRU (isolate_page() is useful.)
3562 * - compound_lock is held when nr_pages > 1
3564 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3567 static int mem_cgroup_move_account(struct page *page,
3568 unsigned int nr_pages,
3569 struct page_cgroup *pc,
3570 struct mem_cgroup *from,
3571 struct mem_cgroup *to)
3573 unsigned long flags;
3575 bool anon = PageAnon(page);
3577 VM_BUG_ON(from == to);
3578 VM_BUG_ON(PageLRU(page));
3580 * The page is isolated from LRU. So, collapse function
3581 * will not handle this page. But page splitting can happen.
3582 * Do this check under compound_page_lock(). The caller should
3586 if (nr_pages > 1 && !PageTransHuge(page))
3589 lock_page_cgroup(pc);
3592 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3595 move_lock_mem_cgroup(from, &flags);
3597 if (!anon && page_mapped(page))
3598 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3599 MEM_CGROUP_STAT_FILE_MAPPED);
3601 if (PageWriteback(page))
3602 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3603 MEM_CGROUP_STAT_WRITEBACK);
3605 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3607 /* caller should have done css_get */
3608 pc->mem_cgroup = to;
3609 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3610 move_unlock_mem_cgroup(from, &flags);
3613 unlock_page_cgroup(pc);
3617 memcg_check_events(to, page);
3618 memcg_check_events(from, page);
3624 * mem_cgroup_move_parent - moves page to the parent group
3625 * @page: the page to move
3626 * @pc: page_cgroup of the page
3627 * @child: page's cgroup
3629 * move charges to its parent or the root cgroup if the group has no
3630 * parent (aka use_hierarchy==0).
3631 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3632 * mem_cgroup_move_account fails) the failure is always temporary and
3633 * it signals a race with a page removal/uncharge or migration. In the
3634 * first case the page is on the way out and it will vanish from the LRU
3635 * on the next attempt and the call should be retried later.
3636 * Isolation from the LRU fails only if page has been isolated from
3637 * the LRU since we looked at it and that usually means either global
3638 * reclaim or migration going on. The page will either get back to the
3640 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3641 * (!PageCgroupUsed) or moved to a different group. The page will
3642 * disappear in the next attempt.
3644 static int mem_cgroup_move_parent(struct page *page,
3645 struct page_cgroup *pc,
3646 struct mem_cgroup *child)
3648 struct mem_cgroup *parent;
3649 unsigned int nr_pages;
3650 unsigned long uninitialized_var(flags);
3653 VM_BUG_ON(mem_cgroup_is_root(child));
3656 if (!get_page_unless_zero(page))
3658 if (isolate_lru_page(page))
3661 nr_pages = hpage_nr_pages(page);
3663 parent = parent_mem_cgroup(child);
3665 * If no parent, move charges to root cgroup.
3668 parent = root_mem_cgroup;
3671 VM_BUG_ON(!PageTransHuge(page));
3672 flags = compound_lock_irqsave(page);
3675 ret = mem_cgroup_move_account(page, nr_pages,
3678 __mem_cgroup_cancel_local_charge(child, nr_pages);
3681 compound_unlock_irqrestore(page, flags);
3682 putback_lru_page(page);
3690 * Charge the memory controller for page usage.
3692 * 0 if the charge was successful
3693 * < 0 if the cgroup is over its limit
3695 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3696 gfp_t gfp_mask, enum charge_type ctype)
3698 struct mem_cgroup *memcg = NULL;
3699 unsigned int nr_pages = 1;
3703 if (PageTransHuge(page)) {
3704 nr_pages <<= compound_order(page);
3705 VM_BUG_ON(!PageTransHuge(page));
3707 * Never OOM-kill a process for a huge page. The
3708 * fault handler will fall back to regular pages.
3713 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3716 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3720 int mem_cgroup_newpage_charge(struct page *page,
3721 struct mm_struct *mm, gfp_t gfp_mask)
3723 if (mem_cgroup_disabled())
3725 VM_BUG_ON(page_mapped(page));
3726 VM_BUG_ON(page->mapping && !PageAnon(page));
3728 return mem_cgroup_charge_common(page, mm, gfp_mask,
3729 MEM_CGROUP_CHARGE_TYPE_ANON);
3733 * While swap-in, try_charge -> commit or cancel, the page is locked.
3734 * And when try_charge() successfully returns, one refcnt to memcg without
3735 * struct page_cgroup is acquired. This refcnt will be consumed by
3736 * "commit()" or removed by "cancel()"
3738 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3741 struct mem_cgroup **memcgp)
3743 struct mem_cgroup *memcg;
3744 struct page_cgroup *pc;
3747 pc = lookup_page_cgroup(page);
3749 * Every swap fault against a single page tries to charge the
3750 * page, bail as early as possible. shmem_unuse() encounters
3751 * already charged pages, too. The USED bit is protected by
3752 * the page lock, which serializes swap cache removal, which
3753 * in turn serializes uncharging.
3755 if (PageCgroupUsed(pc))
3757 if (!do_swap_account)
3759 memcg = try_get_mem_cgroup_from_page(page);
3763 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3764 css_put(&memcg->css);
3769 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3775 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3776 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3779 if (mem_cgroup_disabled())
3782 * A racing thread's fault, or swapoff, may have already
3783 * updated the pte, and even removed page from swap cache: in
3784 * those cases unuse_pte()'s pte_same() test will fail; but
3785 * there's also a KSM case which does need to charge the page.
3787 if (!PageSwapCache(page)) {
3790 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3795 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3798 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3800 if (mem_cgroup_disabled())
3804 __mem_cgroup_cancel_charge(memcg, 1);
3808 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3809 enum charge_type ctype)
3811 if (mem_cgroup_disabled())
3816 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3818 * Now swap is on-memory. This means this page may be
3819 * counted both as mem and swap....double count.
3820 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3821 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3822 * may call delete_from_swap_cache() before reach here.
3824 if (do_swap_account && PageSwapCache(page)) {
3825 swp_entry_t ent = {.val = page_private(page)};
3826 mem_cgroup_uncharge_swap(ent);
3830 void mem_cgroup_commit_charge_swapin(struct page *page,
3831 struct mem_cgroup *memcg)
3833 __mem_cgroup_commit_charge_swapin(page, memcg,
3834 MEM_CGROUP_CHARGE_TYPE_ANON);
3837 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3840 struct mem_cgroup *memcg = NULL;
3841 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3844 if (mem_cgroup_disabled())
3846 if (PageCompound(page))
3849 if (!PageSwapCache(page))
3850 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3851 else { /* page is swapcache/shmem */
3852 ret = __mem_cgroup_try_charge_swapin(mm, page,
3855 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3860 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3861 unsigned int nr_pages,
3862 const enum charge_type ctype)
3864 struct memcg_batch_info *batch = NULL;
3865 bool uncharge_memsw = true;
3867 /* If swapout, usage of swap doesn't decrease */
3868 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3869 uncharge_memsw = false;
3871 batch = ¤t->memcg_batch;
3873 * In usual, we do css_get() when we remember memcg pointer.
3874 * But in this case, we keep res->usage until end of a series of
3875 * uncharges. Then, it's ok to ignore memcg's refcnt.
3878 batch->memcg = memcg;
3880 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3881 * In those cases, all pages freed continuously can be expected to be in
3882 * the same cgroup and we have chance to coalesce uncharges.
3883 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3884 * because we want to do uncharge as soon as possible.
3887 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3888 goto direct_uncharge;
3891 goto direct_uncharge;
3894 * In typical case, batch->memcg == mem. This means we can
3895 * merge a series of uncharges to an uncharge of res_counter.
3896 * If not, we uncharge res_counter ony by one.
3898 if (batch->memcg != memcg)
3899 goto direct_uncharge;
3900 /* remember freed charge and uncharge it later */
3903 batch->memsw_nr_pages++;
3906 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3908 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3909 if (unlikely(batch->memcg != memcg))
3910 memcg_oom_recover(memcg);
3914 * uncharge if !page_mapped(page)
3916 static struct mem_cgroup *
3917 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3920 struct mem_cgroup *memcg = NULL;
3921 unsigned int nr_pages = 1;
3922 struct page_cgroup *pc;
3925 if (mem_cgroup_disabled())
3928 if (PageTransHuge(page)) {
3929 nr_pages <<= compound_order(page);
3930 VM_BUG_ON(!PageTransHuge(page));
3933 * Check if our page_cgroup is valid
3935 pc = lookup_page_cgroup(page);
3936 if (unlikely(!PageCgroupUsed(pc)))
3939 lock_page_cgroup(pc);
3941 memcg = pc->mem_cgroup;
3943 if (!PageCgroupUsed(pc))
3946 anon = PageAnon(page);
3949 case MEM_CGROUP_CHARGE_TYPE_ANON:
3951 * Generally PageAnon tells if it's the anon statistics to be
3952 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3953 * used before page reached the stage of being marked PageAnon.
3957 case MEM_CGROUP_CHARGE_TYPE_DROP:
3958 /* See mem_cgroup_prepare_migration() */
3959 if (page_mapped(page))
3962 * Pages under migration may not be uncharged. But
3963 * end_migration() /must/ be the one uncharging the
3964 * unused post-migration page and so it has to call
3965 * here with the migration bit still set. See the
3966 * res_counter handling below.
3968 if (!end_migration && PageCgroupMigration(pc))
3971 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3972 if (!PageAnon(page)) { /* Shared memory */
3973 if (page->mapping && !page_is_file_cache(page))
3975 } else if (page_mapped(page)) /* Anon */
3982 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
3984 ClearPageCgroupUsed(pc);
3986 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3987 * freed from LRU. This is safe because uncharged page is expected not
3988 * to be reused (freed soon). Exception is SwapCache, it's handled by
3989 * special functions.
3992 unlock_page_cgroup(pc);
3994 * even after unlock, we have memcg->res.usage here and this memcg
3995 * will never be freed, so it's safe to call css_get().
3997 memcg_check_events(memcg, page);
3998 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3999 mem_cgroup_swap_statistics(memcg, true);
4000 css_get(&memcg->css);
4003 * Migration does not charge the res_counter for the
4004 * replacement page, so leave it alone when phasing out the
4005 * page that is unused after the migration.
4007 if (!end_migration && !mem_cgroup_is_root(memcg))
4008 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4013 unlock_page_cgroup(pc);
4017 void mem_cgroup_uncharge_page(struct page *page)
4020 if (page_mapped(page))
4022 VM_BUG_ON(page->mapping && !PageAnon(page));
4024 * If the page is in swap cache, uncharge should be deferred
4025 * to the swap path, which also properly accounts swap usage
4026 * and handles memcg lifetime.
4028 * Note that this check is not stable and reclaim may add the
4029 * page to swap cache at any time after this. However, if the
4030 * page is not in swap cache by the time page->mapcount hits
4031 * 0, there won't be any page table references to the swap
4032 * slot, and reclaim will free it and not actually write the
4035 if (PageSwapCache(page))
4037 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4040 void mem_cgroup_uncharge_cache_page(struct page *page)
4042 VM_BUG_ON(page_mapped(page));
4043 VM_BUG_ON(page->mapping);
4044 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4048 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4049 * In that cases, pages are freed continuously and we can expect pages
4050 * are in the same memcg. All these calls itself limits the number of
4051 * pages freed at once, then uncharge_start/end() is called properly.
4052 * This may be called prural(2) times in a context,
4055 void mem_cgroup_uncharge_start(void)
4057 current->memcg_batch.do_batch++;
4058 /* We can do nest. */
4059 if (current->memcg_batch.do_batch == 1) {
4060 current->memcg_batch.memcg = NULL;
4061 current->memcg_batch.nr_pages = 0;
4062 current->memcg_batch.memsw_nr_pages = 0;
4066 void mem_cgroup_uncharge_end(void)
4068 struct memcg_batch_info *batch = ¤t->memcg_batch;
4070 if (!batch->do_batch)
4074 if (batch->do_batch) /* If stacked, do nothing. */
4080 * This "batch->memcg" is valid without any css_get/put etc...
4081 * bacause we hide charges behind us.
4083 if (batch->nr_pages)
4084 res_counter_uncharge(&batch->memcg->res,
4085 batch->nr_pages * PAGE_SIZE);
4086 if (batch->memsw_nr_pages)
4087 res_counter_uncharge(&batch->memcg->memsw,
4088 batch->memsw_nr_pages * PAGE_SIZE);
4089 memcg_oom_recover(batch->memcg);
4090 /* forget this pointer (for sanity check) */
4091 batch->memcg = NULL;
4096 * called after __delete_from_swap_cache() and drop "page" account.
4097 * memcg information is recorded to swap_cgroup of "ent"
4100 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4102 struct mem_cgroup *memcg;
4103 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4105 if (!swapout) /* this was a swap cache but the swap is unused ! */
4106 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4108 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4111 * record memcg information, if swapout && memcg != NULL,
4112 * css_get() was called in uncharge().
4114 if (do_swap_account && swapout && memcg)
4115 swap_cgroup_record(ent, css_id(&memcg->css));
4119 #ifdef CONFIG_MEMCG_SWAP
4121 * called from swap_entry_free(). remove record in swap_cgroup and
4122 * uncharge "memsw" account.
4124 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4126 struct mem_cgroup *memcg;
4129 if (!do_swap_account)
4132 id = swap_cgroup_record(ent, 0);
4134 memcg = mem_cgroup_lookup(id);
4137 * We uncharge this because swap is freed.
4138 * This memcg can be obsolete one. We avoid calling css_tryget
4140 if (!mem_cgroup_is_root(memcg))
4141 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4142 mem_cgroup_swap_statistics(memcg, false);
4143 css_put(&memcg->css);
4149 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4150 * @entry: swap entry to be moved
4151 * @from: mem_cgroup which the entry is moved from
4152 * @to: mem_cgroup which the entry is moved to
4154 * It succeeds only when the swap_cgroup's record for this entry is the same
4155 * as the mem_cgroup's id of @from.
4157 * Returns 0 on success, -EINVAL on failure.
4159 * The caller must have charged to @to, IOW, called res_counter_charge() about
4160 * both res and memsw, and called css_get().
4162 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4163 struct mem_cgroup *from, struct mem_cgroup *to)
4165 unsigned short old_id, new_id;
4167 old_id = css_id(&from->css);
4168 new_id = css_id(&to->css);
4170 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4171 mem_cgroup_swap_statistics(from, false);
4172 mem_cgroup_swap_statistics(to, true);
4174 * This function is only called from task migration context now.
4175 * It postpones res_counter and refcount handling till the end
4176 * of task migration(mem_cgroup_clear_mc()) for performance
4177 * improvement. But we cannot postpone css_get(to) because if
4178 * the process that has been moved to @to does swap-in, the
4179 * refcount of @to might be decreased to 0.
4181 * We are in attach() phase, so the cgroup is guaranteed to be
4182 * alive, so we can just call css_get().
4190 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4191 struct mem_cgroup *from, struct mem_cgroup *to)
4198 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4201 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4202 struct mem_cgroup **memcgp)
4204 struct mem_cgroup *memcg = NULL;
4205 unsigned int nr_pages = 1;
4206 struct page_cgroup *pc;
4207 enum charge_type ctype;
4211 if (mem_cgroup_disabled())
4214 if (PageTransHuge(page))
4215 nr_pages <<= compound_order(page);
4217 pc = lookup_page_cgroup(page);
4218 lock_page_cgroup(pc);
4219 if (PageCgroupUsed(pc)) {
4220 memcg = pc->mem_cgroup;
4221 css_get(&memcg->css);
4223 * At migrating an anonymous page, its mapcount goes down
4224 * to 0 and uncharge() will be called. But, even if it's fully
4225 * unmapped, migration may fail and this page has to be
4226 * charged again. We set MIGRATION flag here and delay uncharge
4227 * until end_migration() is called
4229 * Corner Case Thinking
4231 * When the old page was mapped as Anon and it's unmap-and-freed
4232 * while migration was ongoing.
4233 * If unmap finds the old page, uncharge() of it will be delayed
4234 * until end_migration(). If unmap finds a new page, it's
4235 * uncharged when it make mapcount to be 1->0. If unmap code
4236 * finds swap_migration_entry, the new page will not be mapped
4237 * and end_migration() will find it(mapcount==0).
4240 * When the old page was mapped but migraion fails, the kernel
4241 * remaps it. A charge for it is kept by MIGRATION flag even
4242 * if mapcount goes down to 0. We can do remap successfully
4243 * without charging it again.
4246 * The "old" page is under lock_page() until the end of
4247 * migration, so, the old page itself will not be swapped-out.
4248 * If the new page is swapped out before end_migraton, our
4249 * hook to usual swap-out path will catch the event.
4252 SetPageCgroupMigration(pc);
4254 unlock_page_cgroup(pc);
4256 * If the page is not charged at this point,
4264 * We charge new page before it's used/mapped. So, even if unlock_page()
4265 * is called before end_migration, we can catch all events on this new
4266 * page. In the case new page is migrated but not remapped, new page's
4267 * mapcount will be finally 0 and we call uncharge in end_migration().
4270 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4272 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4274 * The page is committed to the memcg, but it's not actually
4275 * charged to the res_counter since we plan on replacing the
4276 * old one and only one page is going to be left afterwards.
4278 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4281 /* remove redundant charge if migration failed*/
4282 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4283 struct page *oldpage, struct page *newpage, bool migration_ok)
4285 struct page *used, *unused;
4286 struct page_cgroup *pc;
4292 if (!migration_ok) {
4299 anon = PageAnon(used);
4300 __mem_cgroup_uncharge_common(unused,
4301 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4302 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4304 css_put(&memcg->css);
4306 * We disallowed uncharge of pages under migration because mapcount
4307 * of the page goes down to zero, temporarly.
4308 * Clear the flag and check the page should be charged.
4310 pc = lookup_page_cgroup(oldpage);
4311 lock_page_cgroup(pc);
4312 ClearPageCgroupMigration(pc);
4313 unlock_page_cgroup(pc);
4316 * If a page is a file cache, radix-tree replacement is very atomic
4317 * and we can skip this check. When it was an Anon page, its mapcount
4318 * goes down to 0. But because we added MIGRATION flage, it's not
4319 * uncharged yet. There are several case but page->mapcount check
4320 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4321 * check. (see prepare_charge() also)
4324 mem_cgroup_uncharge_page(used);
4328 * At replace page cache, newpage is not under any memcg but it's on
4329 * LRU. So, this function doesn't touch res_counter but handles LRU
4330 * in correct way. Both pages are locked so we cannot race with uncharge.
4332 void mem_cgroup_replace_page_cache(struct page *oldpage,
4333 struct page *newpage)
4335 struct mem_cgroup *memcg = NULL;
4336 struct page_cgroup *pc;
4337 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4339 if (mem_cgroup_disabled())
4342 pc = lookup_page_cgroup(oldpage);
4343 /* fix accounting on old pages */
4344 lock_page_cgroup(pc);
4345 if (PageCgroupUsed(pc)) {
4346 memcg = pc->mem_cgroup;
4347 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4348 ClearPageCgroupUsed(pc);
4350 unlock_page_cgroup(pc);
4353 * When called from shmem_replace_page(), in some cases the
4354 * oldpage has already been charged, and in some cases not.
4359 * Even if newpage->mapping was NULL before starting replacement,
4360 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4361 * LRU while we overwrite pc->mem_cgroup.
4363 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4366 #ifdef CONFIG_DEBUG_VM
4367 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4369 struct page_cgroup *pc;
4371 pc = lookup_page_cgroup(page);
4373 * Can be NULL while feeding pages into the page allocator for
4374 * the first time, i.e. during boot or memory hotplug;
4375 * or when mem_cgroup_disabled().
4377 if (likely(pc) && PageCgroupUsed(pc))
4382 bool mem_cgroup_bad_page_check(struct page *page)
4384 if (mem_cgroup_disabled())
4387 return lookup_page_cgroup_used(page) != NULL;
4390 void mem_cgroup_print_bad_page(struct page *page)
4392 struct page_cgroup *pc;
4394 pc = lookup_page_cgroup_used(page);
4396 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4397 pc, pc->flags, pc->mem_cgroup);
4402 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4403 unsigned long long val)
4406 u64 memswlimit, memlimit;
4408 int children = mem_cgroup_count_children(memcg);
4409 u64 curusage, oldusage;
4413 * For keeping hierarchical_reclaim simple, how long we should retry
4414 * is depends on callers. We set our retry-count to be function
4415 * of # of children which we should visit in this loop.
4417 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4419 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4422 while (retry_count) {
4423 if (signal_pending(current)) {
4428 * Rather than hide all in some function, I do this in
4429 * open coded manner. You see what this really does.
4430 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4432 mutex_lock(&set_limit_mutex);
4433 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4434 if (memswlimit < val) {
4436 mutex_unlock(&set_limit_mutex);
4440 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4444 ret = res_counter_set_limit(&memcg->res, val);
4446 if (memswlimit == val)
4447 memcg->memsw_is_minimum = true;
4449 memcg->memsw_is_minimum = false;
4451 mutex_unlock(&set_limit_mutex);
4456 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4457 MEM_CGROUP_RECLAIM_SHRINK);
4458 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4459 /* Usage is reduced ? */
4460 if (curusage >= oldusage)
4463 oldusage = curusage;
4465 if (!ret && enlarge)
4466 memcg_oom_recover(memcg);
4471 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4472 unsigned long long val)
4475 u64 memlimit, memswlimit, oldusage, curusage;
4476 int children = mem_cgroup_count_children(memcg);
4480 /* see mem_cgroup_resize_res_limit */
4481 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4482 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4483 while (retry_count) {
4484 if (signal_pending(current)) {
4489 * Rather than hide all in some function, I do this in
4490 * open coded manner. You see what this really does.
4491 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4493 mutex_lock(&set_limit_mutex);
4494 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4495 if (memlimit > val) {
4497 mutex_unlock(&set_limit_mutex);
4500 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4501 if (memswlimit < val)
4503 ret = res_counter_set_limit(&memcg->memsw, val);
4505 if (memlimit == val)
4506 memcg->memsw_is_minimum = true;
4508 memcg->memsw_is_minimum = false;
4510 mutex_unlock(&set_limit_mutex);
4515 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4516 MEM_CGROUP_RECLAIM_NOSWAP |
4517 MEM_CGROUP_RECLAIM_SHRINK);
4518 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4519 /* Usage is reduced ? */
4520 if (curusage >= oldusage)
4523 oldusage = curusage;
4525 if (!ret && enlarge)
4526 memcg_oom_recover(memcg);
4531 * mem_cgroup_force_empty_list - clears LRU of a group
4532 * @memcg: group to clear
4535 * @lru: lru to to clear
4537 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4538 * reclaim the pages page themselves - pages are moved to the parent (or root)
4541 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4542 int node, int zid, enum lru_list lru)
4544 struct lruvec *lruvec;
4545 unsigned long flags;
4546 struct list_head *list;
4550 zone = &NODE_DATA(node)->node_zones[zid];
4551 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4552 list = &lruvec->lists[lru];
4556 struct page_cgroup *pc;
4559 spin_lock_irqsave(&zone->lru_lock, flags);
4560 if (list_empty(list)) {
4561 spin_unlock_irqrestore(&zone->lru_lock, flags);
4564 page = list_entry(list->prev, struct page, lru);
4566 list_move(&page->lru, list);
4568 spin_unlock_irqrestore(&zone->lru_lock, flags);
4571 spin_unlock_irqrestore(&zone->lru_lock, flags);
4573 pc = lookup_page_cgroup(page);
4575 if (mem_cgroup_move_parent(page, pc, memcg)) {
4576 /* found lock contention or "pc" is obsolete. */
4581 } while (!list_empty(list));
4585 * make mem_cgroup's charge to be 0 if there is no task by moving
4586 * all the charges and pages to the parent.
4587 * This enables deleting this mem_cgroup.
4589 * Caller is responsible for holding css reference on the memcg.
4591 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4597 /* This is for making all *used* pages to be on LRU. */
4598 lru_add_drain_all();
4599 drain_all_stock_sync(memcg);
4600 mem_cgroup_start_move(memcg);
4601 for_each_node_state(node, N_MEMORY) {
4602 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4605 mem_cgroup_force_empty_list(memcg,
4610 mem_cgroup_end_move(memcg);
4611 memcg_oom_recover(memcg);
4615 * Kernel memory may not necessarily be trackable to a specific
4616 * process. So they are not migrated, and therefore we can't
4617 * expect their value to drop to 0 here.
4618 * Having res filled up with kmem only is enough.
4620 * This is a safety check because mem_cgroup_force_empty_list
4621 * could have raced with mem_cgroup_replace_page_cache callers
4622 * so the lru seemed empty but the page could have been added
4623 * right after the check. RES_USAGE should be safe as we always
4624 * charge before adding to the LRU.
4626 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4627 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4628 } while (usage > 0);
4632 * This mainly exists for tests during the setting of set of use_hierarchy.
4633 * Since this is the very setting we are changing, the current hierarchy value
4636 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4638 struct cgroup_subsys_state *pos;
4640 /* bounce at first found */
4641 css_for_each_child(pos, &memcg->css)
4647 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4648 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4649 * from mem_cgroup_count_children(), in the sense that we don't really care how
4650 * many children we have; we only need to know if we have any. It also counts
4651 * any memcg without hierarchy as infertile.
4653 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4655 return memcg->use_hierarchy && __memcg_has_children(memcg);
4659 * Reclaims as many pages from the given memcg as possible and moves
4660 * the rest to the parent.
4662 * Caller is responsible for holding css reference for memcg.
4664 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4666 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4667 struct cgroup *cgrp = memcg->css.cgroup;
4669 /* returns EBUSY if there is a task or if we come here twice. */
4670 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4673 /* we call try-to-free pages for make this cgroup empty */
4674 lru_add_drain_all();
4675 /* try to free all pages in this cgroup */
4676 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4679 if (signal_pending(current))
4682 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4686 /* maybe some writeback is necessary */
4687 congestion_wait(BLK_RW_ASYNC, HZ/10);
4692 mem_cgroup_reparent_charges(memcg);
4697 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4700 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4702 if (mem_cgroup_is_root(memcg))
4704 return mem_cgroup_force_empty(memcg);
4707 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4710 return mem_cgroup_from_css(css)->use_hierarchy;
4713 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4714 struct cftype *cft, u64 val)
4717 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4718 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4720 mutex_lock(&memcg_create_mutex);
4722 if (memcg->use_hierarchy == val)
4726 * If parent's use_hierarchy is set, we can't make any modifications
4727 * in the child subtrees. If it is unset, then the change can
4728 * occur, provided the current cgroup has no children.
4730 * For the root cgroup, parent_mem is NULL, we allow value to be
4731 * set if there are no children.
4733 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4734 (val == 1 || val == 0)) {
4735 if (!__memcg_has_children(memcg))
4736 memcg->use_hierarchy = val;
4743 mutex_unlock(&memcg_create_mutex);
4749 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4750 enum mem_cgroup_stat_index idx)
4752 struct mem_cgroup *iter;
4755 /* Per-cpu values can be negative, use a signed accumulator */
4756 for_each_mem_cgroup_tree(iter, memcg)
4757 val += mem_cgroup_read_stat(iter, idx);
4759 if (val < 0) /* race ? */
4764 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4768 if (!mem_cgroup_is_root(memcg)) {
4770 return res_counter_read_u64(&memcg->res, RES_USAGE);
4772 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4776 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4777 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4779 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4780 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4783 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4785 return val << PAGE_SHIFT;
4788 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
4789 struct cftype *cft, struct file *file,
4790 char __user *buf, size_t nbytes, loff_t *ppos)
4792 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4798 type = MEMFILE_TYPE(cft->private);
4799 name = MEMFILE_ATTR(cft->private);
4803 if (name == RES_USAGE)
4804 val = mem_cgroup_usage(memcg, false);
4806 val = res_counter_read_u64(&memcg->res, name);
4809 if (name == RES_USAGE)
4810 val = mem_cgroup_usage(memcg, true);
4812 val = res_counter_read_u64(&memcg->memsw, name);
4815 val = res_counter_read_u64(&memcg->kmem, name);
4821 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4822 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4825 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
4828 #ifdef CONFIG_MEMCG_KMEM
4829 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4831 * For simplicity, we won't allow this to be disabled. It also can't
4832 * be changed if the cgroup has children already, or if tasks had
4835 * If tasks join before we set the limit, a person looking at
4836 * kmem.usage_in_bytes will have no way to determine when it took
4837 * place, which makes the value quite meaningless.
4839 * After it first became limited, changes in the value of the limit are
4840 * of course permitted.
4842 mutex_lock(&memcg_create_mutex);
4843 mutex_lock(&set_limit_mutex);
4844 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
4845 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
4849 ret = res_counter_set_limit(&memcg->kmem, val);
4852 ret = memcg_update_cache_sizes(memcg);
4854 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
4857 static_key_slow_inc(&memcg_kmem_enabled_key);
4859 * setting the active bit after the inc will guarantee no one
4860 * starts accounting before all call sites are patched
4862 memcg_kmem_set_active(memcg);
4864 ret = res_counter_set_limit(&memcg->kmem, val);
4866 mutex_unlock(&set_limit_mutex);
4867 mutex_unlock(&memcg_create_mutex);
4872 #ifdef CONFIG_MEMCG_KMEM
4873 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4876 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4880 memcg->kmem_account_flags = parent->kmem_account_flags;
4882 * When that happen, we need to disable the static branch only on those
4883 * memcgs that enabled it. To achieve this, we would be forced to
4884 * complicate the code by keeping track of which memcgs were the ones
4885 * that actually enabled limits, and which ones got it from its
4888 * It is a lot simpler just to do static_key_slow_inc() on every child
4889 * that is accounted.
4891 if (!memcg_kmem_is_active(memcg))
4895 * __mem_cgroup_free() will issue static_key_slow_dec() because this
4896 * memcg is active already. If the later initialization fails then the
4897 * cgroup core triggers the cleanup so we do not have to do it here.
4899 static_key_slow_inc(&memcg_kmem_enabled_key);
4901 mutex_lock(&set_limit_mutex);
4902 memcg_stop_kmem_account();
4903 ret = memcg_update_cache_sizes(memcg);
4904 memcg_resume_kmem_account();
4905 mutex_unlock(&set_limit_mutex);
4909 #endif /* CONFIG_MEMCG_KMEM */
4912 * The user of this function is...
4915 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
4918 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4921 unsigned long long val;
4924 type = MEMFILE_TYPE(cft->private);
4925 name = MEMFILE_ATTR(cft->private);
4929 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4933 /* This function does all necessary parse...reuse it */
4934 ret = res_counter_memparse_write_strategy(buffer, &val);
4938 ret = mem_cgroup_resize_limit(memcg, val);
4939 else if (type == _MEMSWAP)
4940 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4941 else if (type == _KMEM)
4942 ret = memcg_update_kmem_limit(css, val);
4946 case RES_SOFT_LIMIT:
4947 ret = res_counter_memparse_write_strategy(buffer, &val);
4951 * For memsw, soft limits are hard to implement in terms
4952 * of semantics, for now, we support soft limits for
4953 * control without swap
4956 ret = res_counter_set_soft_limit(&memcg->res, val);
4961 ret = -EINVAL; /* should be BUG() ? */
4967 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4968 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4970 unsigned long long min_limit, min_memsw_limit, tmp;
4972 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4973 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4974 if (!memcg->use_hierarchy)
4977 while (css_parent(&memcg->css)) {
4978 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4979 if (!memcg->use_hierarchy)
4981 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4982 min_limit = min(min_limit, tmp);
4983 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4984 min_memsw_limit = min(min_memsw_limit, tmp);
4987 *mem_limit = min_limit;
4988 *memsw_limit = min_memsw_limit;
4991 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
4993 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4997 type = MEMFILE_TYPE(event);
4998 name = MEMFILE_ATTR(event);
5003 res_counter_reset_max(&memcg->res);
5004 else if (type == _MEMSWAP)
5005 res_counter_reset_max(&memcg->memsw);
5006 else if (type == _KMEM)
5007 res_counter_reset_max(&memcg->kmem);
5013 res_counter_reset_failcnt(&memcg->res);
5014 else if (type == _MEMSWAP)
5015 res_counter_reset_failcnt(&memcg->memsw);
5016 else if (type == _KMEM)
5017 res_counter_reset_failcnt(&memcg->kmem);
5026 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5029 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5033 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5034 struct cftype *cft, u64 val)
5036 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5038 if (val >= (1 << NR_MOVE_TYPE))
5042 * No kind of locking is needed in here, because ->can_attach() will
5043 * check this value once in the beginning of the process, and then carry
5044 * on with stale data. This means that changes to this value will only
5045 * affect task migrations starting after the change.
5047 memcg->move_charge_at_immigrate = val;
5051 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5052 struct cftype *cft, u64 val)
5059 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5060 struct cftype *cft, struct seq_file *m)
5063 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5064 unsigned long node_nr;
5065 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5067 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5068 seq_printf(m, "total=%lu", total_nr);
5069 for_each_node_state(nid, N_MEMORY) {
5070 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5071 seq_printf(m, " N%d=%lu", nid, node_nr);
5075 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5076 seq_printf(m, "file=%lu", file_nr);
5077 for_each_node_state(nid, N_MEMORY) {
5078 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5080 seq_printf(m, " N%d=%lu", nid, node_nr);
5084 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5085 seq_printf(m, "anon=%lu", anon_nr);
5086 for_each_node_state(nid, N_MEMORY) {
5087 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5089 seq_printf(m, " N%d=%lu", nid, node_nr);
5093 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5094 seq_printf(m, "unevictable=%lu", unevictable_nr);
5095 for_each_node_state(nid, N_MEMORY) {
5096 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5097 BIT(LRU_UNEVICTABLE));
5098 seq_printf(m, " N%d=%lu", nid, node_nr);
5103 #endif /* CONFIG_NUMA */
5105 static inline void mem_cgroup_lru_names_not_uptodate(void)
5107 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5110 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5113 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5114 struct mem_cgroup *mi;
5117 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5118 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5120 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5121 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5124 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5125 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5126 mem_cgroup_read_events(memcg, i));
5128 for (i = 0; i < NR_LRU_LISTS; i++)
5129 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5130 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5132 /* Hierarchical information */
5134 unsigned long long limit, memsw_limit;
5135 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5136 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5137 if (do_swap_account)
5138 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5142 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5145 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5147 for_each_mem_cgroup_tree(mi, memcg)
5148 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5149 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5152 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5153 unsigned long long val = 0;
5155 for_each_mem_cgroup_tree(mi, memcg)
5156 val += mem_cgroup_read_events(mi, i);
5157 seq_printf(m, "total_%s %llu\n",
5158 mem_cgroup_events_names[i], val);
5161 for (i = 0; i < NR_LRU_LISTS; i++) {
5162 unsigned long long val = 0;
5164 for_each_mem_cgroup_tree(mi, memcg)
5165 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5166 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5169 #ifdef CONFIG_DEBUG_VM
5172 struct mem_cgroup_per_zone *mz;
5173 struct zone_reclaim_stat *rstat;
5174 unsigned long recent_rotated[2] = {0, 0};
5175 unsigned long recent_scanned[2] = {0, 0};
5177 for_each_online_node(nid)
5178 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5179 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5180 rstat = &mz->lruvec.reclaim_stat;
5182 recent_rotated[0] += rstat->recent_rotated[0];
5183 recent_rotated[1] += rstat->recent_rotated[1];
5184 recent_scanned[0] += rstat->recent_scanned[0];
5185 recent_scanned[1] += rstat->recent_scanned[1];
5187 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5188 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5189 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5190 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5197 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5200 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5202 return mem_cgroup_swappiness(memcg);
5205 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5206 struct cftype *cft, u64 val)
5208 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5209 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5211 if (val > 100 || !parent)
5214 mutex_lock(&memcg_create_mutex);
5216 /* If under hierarchy, only empty-root can set this value */
5217 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5218 mutex_unlock(&memcg_create_mutex);
5222 memcg->swappiness = val;
5224 mutex_unlock(&memcg_create_mutex);
5229 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5231 struct mem_cgroup_threshold_ary *t;
5237 t = rcu_dereference(memcg->thresholds.primary);
5239 t = rcu_dereference(memcg->memsw_thresholds.primary);
5244 usage = mem_cgroup_usage(memcg, swap);
5247 * current_threshold points to threshold just below or equal to usage.
5248 * If it's not true, a threshold was crossed after last
5249 * call of __mem_cgroup_threshold().
5251 i = t->current_threshold;
5254 * Iterate backward over array of thresholds starting from
5255 * current_threshold and check if a threshold is crossed.
5256 * If none of thresholds below usage is crossed, we read
5257 * only one element of the array here.
5259 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5260 eventfd_signal(t->entries[i].eventfd, 1);
5262 /* i = current_threshold + 1 */
5266 * Iterate forward over array of thresholds starting from
5267 * current_threshold+1 and check if a threshold is crossed.
5268 * If none of thresholds above usage is crossed, we read
5269 * only one element of the array here.
5271 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5272 eventfd_signal(t->entries[i].eventfd, 1);
5274 /* Update current_threshold */
5275 t->current_threshold = i - 1;
5280 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5283 __mem_cgroup_threshold(memcg, false);
5284 if (do_swap_account)
5285 __mem_cgroup_threshold(memcg, true);
5287 memcg = parent_mem_cgroup(memcg);
5291 static int compare_thresholds(const void *a, const void *b)
5293 const struct mem_cgroup_threshold *_a = a;
5294 const struct mem_cgroup_threshold *_b = b;
5296 if (_a->threshold > _b->threshold)
5299 if (_a->threshold < _b->threshold)
5305 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5307 struct mem_cgroup_eventfd_list *ev;
5309 list_for_each_entry(ev, &memcg->oom_notify, list)
5310 eventfd_signal(ev->eventfd, 1);
5314 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5316 struct mem_cgroup *iter;
5318 for_each_mem_cgroup_tree(iter, memcg)
5319 mem_cgroup_oom_notify_cb(iter);
5322 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5323 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5325 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5326 struct mem_cgroup_thresholds *thresholds;
5327 struct mem_cgroup_threshold_ary *new;
5328 enum res_type type = MEMFILE_TYPE(cft->private);
5329 u64 threshold, usage;
5332 ret = res_counter_memparse_write_strategy(args, &threshold);
5336 mutex_lock(&memcg->thresholds_lock);
5339 thresholds = &memcg->thresholds;
5340 else if (type == _MEMSWAP)
5341 thresholds = &memcg->memsw_thresholds;
5345 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5347 /* Check if a threshold crossed before adding a new one */
5348 if (thresholds->primary)
5349 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5351 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5353 /* Allocate memory for new array of thresholds */
5354 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5362 /* Copy thresholds (if any) to new array */
5363 if (thresholds->primary) {
5364 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5365 sizeof(struct mem_cgroup_threshold));
5368 /* Add new threshold */
5369 new->entries[size - 1].eventfd = eventfd;
5370 new->entries[size - 1].threshold = threshold;
5372 /* Sort thresholds. Registering of new threshold isn't time-critical */
5373 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5374 compare_thresholds, NULL);
5376 /* Find current threshold */
5377 new->current_threshold = -1;
5378 for (i = 0; i < size; i++) {
5379 if (new->entries[i].threshold <= usage) {
5381 * new->current_threshold will not be used until
5382 * rcu_assign_pointer(), so it's safe to increment
5385 ++new->current_threshold;
5390 /* Free old spare buffer and save old primary buffer as spare */
5391 kfree(thresholds->spare);
5392 thresholds->spare = thresholds->primary;
5394 rcu_assign_pointer(thresholds->primary, new);
5396 /* To be sure that nobody uses thresholds */
5400 mutex_unlock(&memcg->thresholds_lock);
5405 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5406 struct cftype *cft, struct eventfd_ctx *eventfd)
5408 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5409 struct mem_cgroup_thresholds *thresholds;
5410 struct mem_cgroup_threshold_ary *new;
5411 enum res_type type = MEMFILE_TYPE(cft->private);
5415 mutex_lock(&memcg->thresholds_lock);
5417 thresholds = &memcg->thresholds;
5418 else if (type == _MEMSWAP)
5419 thresholds = &memcg->memsw_thresholds;
5423 if (!thresholds->primary)
5426 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5428 /* Check if a threshold crossed before removing */
5429 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5431 /* Calculate new number of threshold */
5433 for (i = 0; i < thresholds->primary->size; i++) {
5434 if (thresholds->primary->entries[i].eventfd != eventfd)
5438 new = thresholds->spare;
5440 /* Set thresholds array to NULL if we don't have thresholds */
5449 /* Copy thresholds and find current threshold */
5450 new->current_threshold = -1;
5451 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5452 if (thresholds->primary->entries[i].eventfd == eventfd)
5455 new->entries[j] = thresholds->primary->entries[i];
5456 if (new->entries[j].threshold <= usage) {
5458 * new->current_threshold will not be used
5459 * until rcu_assign_pointer(), so it's safe to increment
5462 ++new->current_threshold;
5468 /* Swap primary and spare array */
5469 thresholds->spare = thresholds->primary;
5470 /* If all events are unregistered, free the spare array */
5472 kfree(thresholds->spare);
5473 thresholds->spare = NULL;
5476 rcu_assign_pointer(thresholds->primary, new);
5478 /* To be sure that nobody uses thresholds */
5481 mutex_unlock(&memcg->thresholds_lock);
5484 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5485 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5487 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5488 struct mem_cgroup_eventfd_list *event;
5489 enum res_type type = MEMFILE_TYPE(cft->private);
5491 BUG_ON(type != _OOM_TYPE);
5492 event = kmalloc(sizeof(*event), GFP_KERNEL);
5496 spin_lock(&memcg_oom_lock);
5498 event->eventfd = eventfd;
5499 list_add(&event->list, &memcg->oom_notify);
5501 /* already in OOM ? */
5502 if (atomic_read(&memcg->under_oom))
5503 eventfd_signal(eventfd, 1);
5504 spin_unlock(&memcg_oom_lock);
5509 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5510 struct cftype *cft, struct eventfd_ctx *eventfd)
5512 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5513 struct mem_cgroup_eventfd_list *ev, *tmp;
5514 enum res_type type = MEMFILE_TYPE(cft->private);
5516 BUG_ON(type != _OOM_TYPE);
5518 spin_lock(&memcg_oom_lock);
5520 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5521 if (ev->eventfd == eventfd) {
5522 list_del(&ev->list);
5527 spin_unlock(&memcg_oom_lock);
5530 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5531 struct cftype *cft, struct cgroup_map_cb *cb)
5533 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5535 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5537 if (atomic_read(&memcg->under_oom))
5538 cb->fill(cb, "under_oom", 1);
5540 cb->fill(cb, "under_oom", 0);
5544 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5545 struct cftype *cft, u64 val)
5547 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5548 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5550 /* cannot set to root cgroup and only 0 and 1 are allowed */
5551 if (!parent || !((val == 0) || (val == 1)))
5554 mutex_lock(&memcg_create_mutex);
5555 /* oom-kill-disable is a flag for subhierarchy. */
5556 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5557 mutex_unlock(&memcg_create_mutex);
5560 memcg->oom_kill_disable = val;
5562 memcg_oom_recover(memcg);
5563 mutex_unlock(&memcg_create_mutex);
5567 #ifdef CONFIG_MEMCG_KMEM
5568 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5572 memcg->kmemcg_id = -1;
5573 ret = memcg_propagate_kmem(memcg);
5577 return mem_cgroup_sockets_init(memcg, ss);
5580 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5582 mem_cgroup_sockets_destroy(memcg);
5585 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5587 if (!memcg_kmem_is_active(memcg))
5591 * kmem charges can outlive the cgroup. In the case of slab
5592 * pages, for instance, a page contain objects from various
5593 * processes. As we prevent from taking a reference for every
5594 * such allocation we have to be careful when doing uncharge
5595 * (see memcg_uncharge_kmem) and here during offlining.
5597 * The idea is that that only the _last_ uncharge which sees
5598 * the dead memcg will drop the last reference. An additional
5599 * reference is taken here before the group is marked dead
5600 * which is then paired with css_put during uncharge resp. here.
5602 * Although this might sound strange as this path is called from
5603 * css_offline() when the referencemight have dropped down to 0
5604 * and shouldn't be incremented anymore (css_tryget would fail)
5605 * we do not have other options because of the kmem allocations
5608 css_get(&memcg->css);
5610 memcg_kmem_mark_dead(memcg);
5612 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5615 if (memcg_kmem_test_and_clear_dead(memcg))
5616 css_put(&memcg->css);
5619 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5624 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5628 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5633 static struct cftype mem_cgroup_files[] = {
5635 .name = "usage_in_bytes",
5636 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5637 .read = mem_cgroup_read,
5638 .register_event = mem_cgroup_usage_register_event,
5639 .unregister_event = mem_cgroup_usage_unregister_event,
5642 .name = "max_usage_in_bytes",
5643 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5644 .trigger = mem_cgroup_reset,
5645 .read = mem_cgroup_read,
5648 .name = "limit_in_bytes",
5649 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5650 .write_string = mem_cgroup_write,
5651 .read = mem_cgroup_read,
5654 .name = "soft_limit_in_bytes",
5655 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5656 .write_string = mem_cgroup_write,
5657 .read = mem_cgroup_read,
5661 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5662 .trigger = mem_cgroup_reset,
5663 .read = mem_cgroup_read,
5667 .read_seq_string = memcg_stat_show,
5670 .name = "force_empty",
5671 .trigger = mem_cgroup_force_empty_write,
5674 .name = "use_hierarchy",
5675 .flags = CFTYPE_INSANE,
5676 .write_u64 = mem_cgroup_hierarchy_write,
5677 .read_u64 = mem_cgroup_hierarchy_read,
5680 .name = "swappiness",
5681 .read_u64 = mem_cgroup_swappiness_read,
5682 .write_u64 = mem_cgroup_swappiness_write,
5685 .name = "move_charge_at_immigrate",
5686 .read_u64 = mem_cgroup_move_charge_read,
5687 .write_u64 = mem_cgroup_move_charge_write,
5690 .name = "oom_control",
5691 .read_map = mem_cgroup_oom_control_read,
5692 .write_u64 = mem_cgroup_oom_control_write,
5693 .register_event = mem_cgroup_oom_register_event,
5694 .unregister_event = mem_cgroup_oom_unregister_event,
5695 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5698 .name = "pressure_level",
5699 .register_event = vmpressure_register_event,
5700 .unregister_event = vmpressure_unregister_event,
5704 .name = "numa_stat",
5705 .read_seq_string = memcg_numa_stat_show,
5708 #ifdef CONFIG_MEMCG_KMEM
5710 .name = "kmem.limit_in_bytes",
5711 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5712 .write_string = mem_cgroup_write,
5713 .read = mem_cgroup_read,
5716 .name = "kmem.usage_in_bytes",
5717 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5718 .read = mem_cgroup_read,
5721 .name = "kmem.failcnt",
5722 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5723 .trigger = mem_cgroup_reset,
5724 .read = mem_cgroup_read,
5727 .name = "kmem.max_usage_in_bytes",
5728 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5729 .trigger = mem_cgroup_reset,
5730 .read = mem_cgroup_read,
5732 #ifdef CONFIG_SLABINFO
5734 .name = "kmem.slabinfo",
5735 .read_seq_string = mem_cgroup_slabinfo_read,
5739 { }, /* terminate */
5742 #ifdef CONFIG_MEMCG_SWAP
5743 static struct cftype memsw_cgroup_files[] = {
5745 .name = "memsw.usage_in_bytes",
5746 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5747 .read = mem_cgroup_read,
5748 .register_event = mem_cgroup_usage_register_event,
5749 .unregister_event = mem_cgroup_usage_unregister_event,
5752 .name = "memsw.max_usage_in_bytes",
5753 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5754 .trigger = mem_cgroup_reset,
5755 .read = mem_cgroup_read,
5758 .name = "memsw.limit_in_bytes",
5759 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5760 .write_string = mem_cgroup_write,
5761 .read = mem_cgroup_read,
5764 .name = "memsw.failcnt",
5765 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5766 .trigger = mem_cgroup_reset,
5767 .read = mem_cgroup_read,
5769 { }, /* terminate */
5772 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5774 struct mem_cgroup_per_node *pn;
5775 struct mem_cgroup_per_zone *mz;
5776 int zone, tmp = node;
5778 * This routine is called against possible nodes.
5779 * But it's BUG to call kmalloc() against offline node.
5781 * TODO: this routine can waste much memory for nodes which will
5782 * never be onlined. It's better to use memory hotplug callback
5785 if (!node_state(node, N_NORMAL_MEMORY))
5787 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5791 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5792 mz = &pn->zoneinfo[zone];
5793 lruvec_init(&mz->lruvec);
5796 memcg->nodeinfo[node] = pn;
5800 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5802 kfree(memcg->nodeinfo[node]);
5805 static struct mem_cgroup *mem_cgroup_alloc(void)
5807 struct mem_cgroup *memcg;
5808 size_t size = memcg_size();
5810 /* Can be very big if nr_node_ids is very big */
5811 if (size < PAGE_SIZE)
5812 memcg = kzalloc(size, GFP_KERNEL);
5814 memcg = vzalloc(size);
5819 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5822 spin_lock_init(&memcg->pcp_counter_lock);
5826 if (size < PAGE_SIZE)
5834 * At destroying mem_cgroup, references from swap_cgroup can remain.
5835 * (scanning all at force_empty is too costly...)
5837 * Instead of clearing all references at force_empty, we remember
5838 * the number of reference from swap_cgroup and free mem_cgroup when
5839 * it goes down to 0.
5841 * Removal of cgroup itself succeeds regardless of refs from swap.
5844 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5847 size_t size = memcg_size();
5849 free_css_id(&mem_cgroup_subsys, &memcg->css);
5852 free_mem_cgroup_per_zone_info(memcg, node);
5854 free_percpu(memcg->stat);
5857 * We need to make sure that (at least for now), the jump label
5858 * destruction code runs outside of the cgroup lock. This is because
5859 * get_online_cpus(), which is called from the static_branch update,
5860 * can't be called inside the cgroup_lock. cpusets are the ones
5861 * enforcing this dependency, so if they ever change, we might as well.
5863 * schedule_work() will guarantee this happens. Be careful if you need
5864 * to move this code around, and make sure it is outside
5867 disarm_static_keys(memcg);
5868 if (size < PAGE_SIZE)
5875 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5877 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5879 if (!memcg->res.parent)
5881 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5883 EXPORT_SYMBOL(parent_mem_cgroup);
5885 static struct cgroup_subsys_state * __ref
5886 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5888 struct mem_cgroup *memcg;
5889 long error = -ENOMEM;
5892 memcg = mem_cgroup_alloc();
5894 return ERR_PTR(error);
5897 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5901 if (parent_css == NULL) {
5902 root_mem_cgroup = memcg;
5903 res_counter_init(&memcg->res, NULL);
5904 res_counter_init(&memcg->memsw, NULL);
5905 res_counter_init(&memcg->kmem, NULL);
5908 memcg->last_scanned_node = MAX_NUMNODES;
5909 INIT_LIST_HEAD(&memcg->oom_notify);
5910 memcg->move_charge_at_immigrate = 0;
5911 mutex_init(&memcg->thresholds_lock);
5912 spin_lock_init(&memcg->move_lock);
5913 vmpressure_init(&memcg->vmpressure);
5918 __mem_cgroup_free(memcg);
5919 return ERR_PTR(error);
5923 mem_cgroup_css_online(struct cgroup_subsys_state *css)
5925 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5926 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
5932 mutex_lock(&memcg_create_mutex);
5934 memcg->use_hierarchy = parent->use_hierarchy;
5935 memcg->oom_kill_disable = parent->oom_kill_disable;
5936 memcg->swappiness = mem_cgroup_swappiness(parent);
5938 if (parent->use_hierarchy) {
5939 res_counter_init(&memcg->res, &parent->res);
5940 res_counter_init(&memcg->memsw, &parent->memsw);
5941 res_counter_init(&memcg->kmem, &parent->kmem);
5944 * No need to take a reference to the parent because cgroup
5945 * core guarantees its existence.
5948 res_counter_init(&memcg->res, NULL);
5949 res_counter_init(&memcg->memsw, NULL);
5950 res_counter_init(&memcg->kmem, NULL);
5952 * Deeper hierachy with use_hierarchy == false doesn't make
5953 * much sense so let cgroup subsystem know about this
5954 * unfortunate state in our controller.
5956 if (parent != root_mem_cgroup)
5957 mem_cgroup_subsys.broken_hierarchy = true;
5960 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
5961 mutex_unlock(&memcg_create_mutex);
5966 * Announce all parents that a group from their hierarchy is gone.
5968 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
5970 struct mem_cgroup *parent = memcg;
5972 while ((parent = parent_mem_cgroup(parent)))
5973 mem_cgroup_iter_invalidate(parent);
5976 * if the root memcg is not hierarchical we have to check it
5979 if (!root_mem_cgroup->use_hierarchy)
5980 mem_cgroup_iter_invalidate(root_mem_cgroup);
5983 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
5985 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5987 kmem_cgroup_css_offline(memcg);
5989 mem_cgroup_invalidate_reclaim_iterators(memcg);
5990 mem_cgroup_reparent_charges(memcg);
5991 mem_cgroup_destroy_all_caches(memcg);
5992 vmpressure_cleanup(&memcg->vmpressure);
5995 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
5997 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5999 memcg_destroy_kmem(memcg);
6000 __mem_cgroup_free(memcg);
6004 /* Handlers for move charge at task migration. */
6005 #define PRECHARGE_COUNT_AT_ONCE 256
6006 static int mem_cgroup_do_precharge(unsigned long count)
6009 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6010 struct mem_cgroup *memcg = mc.to;
6012 if (mem_cgroup_is_root(memcg)) {
6013 mc.precharge += count;
6014 /* we don't need css_get for root */
6017 /* try to charge at once */
6019 struct res_counter *dummy;
6021 * "memcg" cannot be under rmdir() because we've already checked
6022 * by cgroup_lock_live_cgroup() that it is not removed and we
6023 * are still under the same cgroup_mutex. So we can postpone
6026 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6028 if (do_swap_account && res_counter_charge(&memcg->memsw,
6029 PAGE_SIZE * count, &dummy)) {
6030 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6033 mc.precharge += count;
6037 /* fall back to one by one charge */
6039 if (signal_pending(current)) {
6043 if (!batch_count--) {
6044 batch_count = PRECHARGE_COUNT_AT_ONCE;
6047 ret = __mem_cgroup_try_charge(NULL,
6048 GFP_KERNEL, 1, &memcg, false);
6050 /* mem_cgroup_clear_mc() will do uncharge later */
6058 * get_mctgt_type - get target type of moving charge
6059 * @vma: the vma the pte to be checked belongs
6060 * @addr: the address corresponding to the pte to be checked
6061 * @ptent: the pte to be checked
6062 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6065 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6066 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6067 * move charge. if @target is not NULL, the page is stored in target->page
6068 * with extra refcnt got(Callers should handle it).
6069 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6070 * target for charge migration. if @target is not NULL, the entry is stored
6073 * Called with pte lock held.
6080 enum mc_target_type {
6086 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6087 unsigned long addr, pte_t ptent)
6089 struct page *page = vm_normal_page(vma, addr, ptent);
6091 if (!page || !page_mapped(page))
6093 if (PageAnon(page)) {
6094 /* we don't move shared anon */
6097 } else if (!move_file())
6098 /* we ignore mapcount for file pages */
6100 if (!get_page_unless_zero(page))
6107 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6108 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6110 struct page *page = NULL;
6111 swp_entry_t ent = pte_to_swp_entry(ptent);
6113 if (!move_anon() || non_swap_entry(ent))
6116 * Because lookup_swap_cache() updates some statistics counter,
6117 * we call find_get_page() with swapper_space directly.
6119 page = find_get_page(swap_address_space(ent), ent.val);
6120 if (do_swap_account)
6121 entry->val = ent.val;
6126 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6127 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6133 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6134 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6136 struct page *page = NULL;
6137 struct address_space *mapping;
6140 if (!vma->vm_file) /* anonymous vma */
6145 mapping = vma->vm_file->f_mapping;
6146 if (pte_none(ptent))
6147 pgoff = linear_page_index(vma, addr);
6148 else /* pte_file(ptent) is true */
6149 pgoff = pte_to_pgoff(ptent);
6151 /* page is moved even if it's not RSS of this task(page-faulted). */
6152 page = find_get_page(mapping, pgoff);
6155 /* shmem/tmpfs may report page out on swap: account for that too. */
6156 if (radix_tree_exceptional_entry(page)) {
6157 swp_entry_t swap = radix_to_swp_entry(page);
6158 if (do_swap_account)
6160 page = find_get_page(swap_address_space(swap), swap.val);
6166 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6167 unsigned long addr, pte_t ptent, union mc_target *target)
6169 struct page *page = NULL;
6170 struct page_cgroup *pc;
6171 enum mc_target_type ret = MC_TARGET_NONE;
6172 swp_entry_t ent = { .val = 0 };
6174 if (pte_present(ptent))
6175 page = mc_handle_present_pte(vma, addr, ptent);
6176 else if (is_swap_pte(ptent))
6177 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6178 else if (pte_none(ptent) || pte_file(ptent))
6179 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6181 if (!page && !ent.val)
6184 pc = lookup_page_cgroup(page);
6186 * Do only loose check w/o page_cgroup lock.
6187 * mem_cgroup_move_account() checks the pc is valid or not under
6190 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6191 ret = MC_TARGET_PAGE;
6193 target->page = page;
6195 if (!ret || !target)
6198 /* There is a swap entry and a page doesn't exist or isn't charged */
6199 if (ent.val && !ret &&
6200 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6201 ret = MC_TARGET_SWAP;
6208 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6210 * We don't consider swapping or file mapped pages because THP does not
6211 * support them for now.
6212 * Caller should make sure that pmd_trans_huge(pmd) is true.
6214 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6215 unsigned long addr, pmd_t pmd, union mc_target *target)
6217 struct page *page = NULL;
6218 struct page_cgroup *pc;
6219 enum mc_target_type ret = MC_TARGET_NONE;
6221 page = pmd_page(pmd);
6222 VM_BUG_ON(!page || !PageHead(page));
6225 pc = lookup_page_cgroup(page);
6226 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6227 ret = MC_TARGET_PAGE;
6230 target->page = page;
6236 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6237 unsigned long addr, pmd_t pmd, union mc_target *target)
6239 return MC_TARGET_NONE;
6243 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6244 unsigned long addr, unsigned long end,
6245 struct mm_walk *walk)
6247 struct vm_area_struct *vma = walk->private;
6251 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6252 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6253 mc.precharge += HPAGE_PMD_NR;
6254 spin_unlock(&vma->vm_mm->page_table_lock);
6258 if (pmd_trans_unstable(pmd))
6260 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6261 for (; addr != end; pte++, addr += PAGE_SIZE)
6262 if (get_mctgt_type(vma, addr, *pte, NULL))
6263 mc.precharge++; /* increment precharge temporarily */
6264 pte_unmap_unlock(pte - 1, ptl);
6270 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6272 unsigned long precharge;
6273 struct vm_area_struct *vma;
6275 down_read(&mm->mmap_sem);
6276 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6277 struct mm_walk mem_cgroup_count_precharge_walk = {
6278 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6282 if (is_vm_hugetlb_page(vma))
6284 walk_page_range(vma->vm_start, vma->vm_end,
6285 &mem_cgroup_count_precharge_walk);
6287 up_read(&mm->mmap_sem);
6289 precharge = mc.precharge;
6295 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6297 unsigned long precharge = mem_cgroup_count_precharge(mm);
6299 VM_BUG_ON(mc.moving_task);
6300 mc.moving_task = current;
6301 return mem_cgroup_do_precharge(precharge);
6304 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6305 static void __mem_cgroup_clear_mc(void)
6307 struct mem_cgroup *from = mc.from;
6308 struct mem_cgroup *to = mc.to;
6311 /* we must uncharge all the leftover precharges from mc.to */
6313 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6317 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6318 * we must uncharge here.
6320 if (mc.moved_charge) {
6321 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6322 mc.moved_charge = 0;
6324 /* we must fixup refcnts and charges */
6325 if (mc.moved_swap) {
6326 /* uncharge swap account from the old cgroup */
6327 if (!mem_cgroup_is_root(mc.from))
6328 res_counter_uncharge(&mc.from->memsw,
6329 PAGE_SIZE * mc.moved_swap);
6331 for (i = 0; i < mc.moved_swap; i++)
6332 css_put(&mc.from->css);
6334 if (!mem_cgroup_is_root(mc.to)) {
6336 * we charged both to->res and to->memsw, so we should
6339 res_counter_uncharge(&mc.to->res,
6340 PAGE_SIZE * mc.moved_swap);
6342 /* we've already done css_get(mc.to) */
6345 memcg_oom_recover(from);
6346 memcg_oom_recover(to);
6347 wake_up_all(&mc.waitq);
6350 static void mem_cgroup_clear_mc(void)
6352 struct mem_cgroup *from = mc.from;
6355 * we must clear moving_task before waking up waiters at the end of
6358 mc.moving_task = NULL;
6359 __mem_cgroup_clear_mc();
6360 spin_lock(&mc.lock);
6363 spin_unlock(&mc.lock);
6364 mem_cgroup_end_move(from);
6367 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6368 struct cgroup_taskset *tset)
6370 struct task_struct *p = cgroup_taskset_first(tset);
6372 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6373 unsigned long move_charge_at_immigrate;
6376 * We are now commited to this value whatever it is. Changes in this
6377 * tunable will only affect upcoming migrations, not the current one.
6378 * So we need to save it, and keep it going.
6380 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6381 if (move_charge_at_immigrate) {
6382 struct mm_struct *mm;
6383 struct mem_cgroup *from = mem_cgroup_from_task(p);
6385 VM_BUG_ON(from == memcg);
6387 mm = get_task_mm(p);
6390 /* We move charges only when we move a owner of the mm */
6391 if (mm->owner == p) {
6394 VM_BUG_ON(mc.precharge);
6395 VM_BUG_ON(mc.moved_charge);
6396 VM_BUG_ON(mc.moved_swap);
6397 mem_cgroup_start_move(from);
6398 spin_lock(&mc.lock);
6401 mc.immigrate_flags = move_charge_at_immigrate;
6402 spin_unlock(&mc.lock);
6403 /* We set mc.moving_task later */
6405 ret = mem_cgroup_precharge_mc(mm);
6407 mem_cgroup_clear_mc();
6414 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6415 struct cgroup_taskset *tset)
6417 mem_cgroup_clear_mc();
6420 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6421 unsigned long addr, unsigned long end,
6422 struct mm_walk *walk)
6425 struct vm_area_struct *vma = walk->private;
6428 enum mc_target_type target_type;
6429 union mc_target target;
6431 struct page_cgroup *pc;
6434 * We don't take compound_lock() here but no race with splitting thp
6436 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6437 * under splitting, which means there's no concurrent thp split,
6438 * - if another thread runs into split_huge_page() just after we
6439 * entered this if-block, the thread must wait for page table lock
6440 * to be unlocked in __split_huge_page_splitting(), where the main
6441 * part of thp split is not executed yet.
6443 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6444 if (mc.precharge < HPAGE_PMD_NR) {
6445 spin_unlock(&vma->vm_mm->page_table_lock);
6448 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6449 if (target_type == MC_TARGET_PAGE) {
6451 if (!isolate_lru_page(page)) {
6452 pc = lookup_page_cgroup(page);
6453 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6454 pc, mc.from, mc.to)) {
6455 mc.precharge -= HPAGE_PMD_NR;
6456 mc.moved_charge += HPAGE_PMD_NR;
6458 putback_lru_page(page);
6462 spin_unlock(&vma->vm_mm->page_table_lock);
6466 if (pmd_trans_unstable(pmd))
6469 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6470 for (; addr != end; addr += PAGE_SIZE) {
6471 pte_t ptent = *(pte++);
6477 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6478 case MC_TARGET_PAGE:
6480 if (isolate_lru_page(page))
6482 pc = lookup_page_cgroup(page);
6483 if (!mem_cgroup_move_account(page, 1, pc,
6486 /* we uncharge from mc.from later. */
6489 putback_lru_page(page);
6490 put: /* get_mctgt_type() gets the page */
6493 case MC_TARGET_SWAP:
6495 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6497 /* we fixup refcnts and charges later. */
6505 pte_unmap_unlock(pte - 1, ptl);
6510 * We have consumed all precharges we got in can_attach().
6511 * We try charge one by one, but don't do any additional
6512 * charges to mc.to if we have failed in charge once in attach()
6515 ret = mem_cgroup_do_precharge(1);
6523 static void mem_cgroup_move_charge(struct mm_struct *mm)
6525 struct vm_area_struct *vma;
6527 lru_add_drain_all();
6529 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6531 * Someone who are holding the mmap_sem might be waiting in
6532 * waitq. So we cancel all extra charges, wake up all waiters,
6533 * and retry. Because we cancel precharges, we might not be able
6534 * to move enough charges, but moving charge is a best-effort
6535 * feature anyway, so it wouldn't be a big problem.
6537 __mem_cgroup_clear_mc();
6541 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6543 struct mm_walk mem_cgroup_move_charge_walk = {
6544 .pmd_entry = mem_cgroup_move_charge_pte_range,
6548 if (is_vm_hugetlb_page(vma))
6550 ret = walk_page_range(vma->vm_start, vma->vm_end,
6551 &mem_cgroup_move_charge_walk);
6554 * means we have consumed all precharges and failed in
6555 * doing additional charge. Just abandon here.
6559 up_read(&mm->mmap_sem);
6562 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6563 struct cgroup_taskset *tset)
6565 struct task_struct *p = cgroup_taskset_first(tset);
6566 struct mm_struct *mm = get_task_mm(p);
6570 mem_cgroup_move_charge(mm);
6574 mem_cgroup_clear_mc();
6576 #else /* !CONFIG_MMU */
6577 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6578 struct cgroup_taskset *tset)
6582 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6583 struct cgroup_taskset *tset)
6586 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6587 struct cgroup_taskset *tset)
6593 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6594 * to verify sane_behavior flag on each mount attempt.
6596 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6599 * use_hierarchy is forced with sane_behavior. cgroup core
6600 * guarantees that @root doesn't have any children, so turning it
6601 * on for the root memcg is enough.
6603 if (cgroup_sane_behavior(root_css->cgroup))
6604 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6607 struct cgroup_subsys mem_cgroup_subsys = {
6609 .subsys_id = mem_cgroup_subsys_id,
6610 .css_alloc = mem_cgroup_css_alloc,
6611 .css_online = mem_cgroup_css_online,
6612 .css_offline = mem_cgroup_css_offline,
6613 .css_free = mem_cgroup_css_free,
6614 .can_attach = mem_cgroup_can_attach,
6615 .cancel_attach = mem_cgroup_cancel_attach,
6616 .attach = mem_cgroup_move_task,
6617 .bind = mem_cgroup_bind,
6618 .base_cftypes = mem_cgroup_files,
6623 #ifdef CONFIG_MEMCG_SWAP
6624 static int __init enable_swap_account(char *s)
6626 if (!strcmp(s, "1"))
6627 really_do_swap_account = 1;
6628 else if (!strcmp(s, "0"))
6629 really_do_swap_account = 0;
6632 __setup("swapaccount=", enable_swap_account);
6634 static void __init memsw_file_init(void)
6636 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6639 static void __init enable_swap_cgroup(void)
6641 if (!mem_cgroup_disabled() && really_do_swap_account) {
6642 do_swap_account = 1;
6648 static void __init enable_swap_cgroup(void)
6654 * subsys_initcall() for memory controller.
6656 * Some parts like hotcpu_notifier() have to be initialized from this context
6657 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6658 * everything that doesn't depend on a specific mem_cgroup structure should
6659 * be initialized from here.
6661 static int __init mem_cgroup_init(void)
6663 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6664 enable_swap_cgroup();
6668 subsys_initcall(mem_cgroup_init);