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_SOFTLIMIT,
128 MEM_CGROUP_TARGET_NUMAINFO,
131 #define THRESHOLDS_EVENTS_TARGET 128
132 #define SOFTLIMIT_EVENTS_TARGET 1024
133 #define NUMAINFO_EVENTS_TARGET 1024
135 struct mem_cgroup_stat_cpu {
136 long count[MEM_CGROUP_STAT_NSTATS];
137 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
138 unsigned long nr_page_events;
139 unsigned long targets[MEM_CGROUP_NTARGETS];
142 struct mem_cgroup_reclaim_iter {
144 * last scanned hierarchy member. Valid only if last_dead_count
145 * matches memcg->dead_count of the hierarchy root group.
147 struct mem_cgroup *last_visited;
148 unsigned long last_dead_count;
150 /* scan generation, increased every round-trip */
151 unsigned int generation;
155 * per-zone information in memory controller.
157 struct mem_cgroup_per_zone {
158 struct lruvec lruvec;
159 unsigned long lru_size[NR_LRU_LISTS];
161 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
163 struct mem_cgroup *memcg; /* Back pointer, we cannot */
164 /* use container_of */
167 struct mem_cgroup_per_node {
168 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
171 struct mem_cgroup_threshold {
172 struct eventfd_ctx *eventfd;
177 struct mem_cgroup_threshold_ary {
178 /* An array index points to threshold just below or equal to usage. */
179 int current_threshold;
180 /* Size of entries[] */
182 /* Array of thresholds */
183 struct mem_cgroup_threshold entries[0];
186 struct mem_cgroup_thresholds {
187 /* Primary thresholds array */
188 struct mem_cgroup_threshold_ary *primary;
190 * Spare threshold array.
191 * This is needed to make mem_cgroup_unregister_event() "never fail".
192 * It must be able to store at least primary->size - 1 entries.
194 struct mem_cgroup_threshold_ary *spare;
198 struct mem_cgroup_eventfd_list {
199 struct list_head list;
200 struct eventfd_ctx *eventfd;
203 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
204 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
207 * The memory controller data structure. The memory controller controls both
208 * page cache and RSS per cgroup. We would eventually like to provide
209 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
210 * to help the administrator determine what knobs to tune.
212 * TODO: Add a water mark for the memory controller. Reclaim will begin when
213 * we hit the water mark. May be even add a low water mark, such that
214 * no reclaim occurs from a cgroup at it's low water mark, this is
215 * a feature that will be implemented much later in the future.
218 struct cgroup_subsys_state css;
220 * the counter to account for memory usage
222 struct res_counter res;
224 /* vmpressure notifications */
225 struct vmpressure vmpressure;
228 * the counter to account for mem+swap usage.
230 struct res_counter memsw;
233 * the counter to account for kernel memory usage.
235 struct res_counter kmem;
237 * Should the accounting and control be hierarchical, per subtree?
240 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
244 atomic_t oom_wakeups;
247 /* OOM-Killer disable */
248 int oom_kill_disable;
250 /* set when res.limit == memsw.limit */
251 bool memsw_is_minimum;
253 /* protect arrays of thresholds */
254 struct mutex thresholds_lock;
256 /* thresholds for memory usage. RCU-protected */
257 struct mem_cgroup_thresholds thresholds;
259 /* thresholds for mem+swap usage. RCU-protected */
260 struct mem_cgroup_thresholds memsw_thresholds;
262 /* For oom notifier event fd */
263 struct list_head oom_notify;
266 * Should we move charges of a task when a task is moved into this
267 * mem_cgroup ? And what type of charges should we move ?
269 unsigned long move_charge_at_immigrate;
271 * set > 0 if pages under this cgroup are moving to other cgroup.
273 atomic_t moving_account;
274 /* taken only while moving_account > 0 */
275 spinlock_t move_lock;
279 struct mem_cgroup_stat_cpu __percpu *stat;
281 * used when a cpu is offlined or other synchronizations
282 * See mem_cgroup_read_stat().
284 struct mem_cgroup_stat_cpu nocpu_base;
285 spinlock_t pcp_counter_lock;
288 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
289 struct tcp_memcontrol tcp_mem;
291 #if defined(CONFIG_MEMCG_KMEM)
292 /* analogous to slab_common's slab_caches list. per-memcg */
293 struct list_head memcg_slab_caches;
294 /* Not a spinlock, we can take a lot of time walking the list */
295 struct mutex slab_caches_mutex;
296 /* Index in the kmem_cache->memcg_params->memcg_caches array */
300 int last_scanned_node;
302 nodemask_t scan_nodes;
303 atomic_t numainfo_events;
304 atomic_t numainfo_updating;
307 * Protects soft_contributed transitions.
308 * See mem_cgroup_update_soft_limit
310 spinlock_t soft_lock;
313 * If true then this group has increased parents' children_in_excess
314 * when it got over the soft limit.
315 * When a group falls bellow the soft limit, parents' children_in_excess
316 * is decreased and soft_contributed changed to false.
318 bool soft_contributed;
320 /* Number of children that are in soft limit excess */
321 atomic_t children_in_excess;
323 struct mem_cgroup_per_node *nodeinfo[0];
324 /* WARNING: nodeinfo must be the last member here */
327 static size_t memcg_size(void)
329 return sizeof(struct mem_cgroup) +
330 nr_node_ids * sizeof(struct mem_cgroup_per_node);
333 /* internal only representation about the status of kmem accounting. */
335 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
336 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
337 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
340 /* We account when limit is on, but only after call sites are patched */
341 #define KMEM_ACCOUNTED_MASK \
342 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
344 #ifdef CONFIG_MEMCG_KMEM
345 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
347 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
350 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
352 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
355 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
357 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
360 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
362 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
365 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
368 * Our caller must use css_get() first, because memcg_uncharge_kmem()
369 * will call css_put() if it sees the memcg is dead.
372 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
373 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
376 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
378 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
379 &memcg->kmem_account_flags);
383 /* Stuffs for move charges at task migration. */
385 * Types of charges to be moved. "move_charge_at_immitgrate" and
386 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
389 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
390 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
394 /* "mc" and its members are protected by cgroup_mutex */
395 static struct move_charge_struct {
396 spinlock_t lock; /* for from, to */
397 struct mem_cgroup *from;
398 struct mem_cgroup *to;
399 unsigned long immigrate_flags;
400 unsigned long precharge;
401 unsigned long moved_charge;
402 unsigned long moved_swap;
403 struct task_struct *moving_task; /* a task moving charges */
404 wait_queue_head_t waitq; /* a waitq for other context */
406 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
407 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
410 static bool move_anon(void)
412 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
415 static bool move_file(void)
417 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
421 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
422 * limit reclaim to prevent infinite loops, if they ever occur.
424 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
427 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
428 MEM_CGROUP_CHARGE_TYPE_ANON,
429 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
430 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
434 /* for encoding cft->private value on file */
442 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
443 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
444 #define MEMFILE_ATTR(val) ((val) & 0xffff)
445 /* Used for OOM nofiier */
446 #define OOM_CONTROL (0)
449 * Reclaim flags for mem_cgroup_hierarchical_reclaim
451 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
452 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
453 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
454 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
457 * The memcg_create_mutex will be held whenever a new cgroup is created.
458 * As a consequence, any change that needs to protect against new child cgroups
459 * appearing has to hold it as well.
461 static DEFINE_MUTEX(memcg_create_mutex);
463 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
465 return s ? container_of(s, struct mem_cgroup, css) : NULL;
468 /* Some nice accessors for the vmpressure. */
469 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
472 memcg = root_mem_cgroup;
473 return &memcg->vmpressure;
476 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
478 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
481 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
483 return &mem_cgroup_from_css(css)->vmpressure;
486 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
488 return (memcg == root_mem_cgroup);
491 /* Writing them here to avoid exposing memcg's inner layout */
492 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
494 void sock_update_memcg(struct sock *sk)
496 if (mem_cgroup_sockets_enabled) {
497 struct mem_cgroup *memcg;
498 struct cg_proto *cg_proto;
500 BUG_ON(!sk->sk_prot->proto_cgroup);
502 /* Socket cloning can throw us here with sk_cgrp already
503 * filled. It won't however, necessarily happen from
504 * process context. So the test for root memcg given
505 * the current task's memcg won't help us in this case.
507 * Respecting the original socket's memcg is a better
508 * decision in this case.
511 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
512 css_get(&sk->sk_cgrp->memcg->css);
517 memcg = mem_cgroup_from_task(current);
518 cg_proto = sk->sk_prot->proto_cgroup(memcg);
519 if (!mem_cgroup_is_root(memcg) &&
520 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
521 sk->sk_cgrp = cg_proto;
526 EXPORT_SYMBOL(sock_update_memcg);
528 void sock_release_memcg(struct sock *sk)
530 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
531 struct mem_cgroup *memcg;
532 WARN_ON(!sk->sk_cgrp->memcg);
533 memcg = sk->sk_cgrp->memcg;
534 css_put(&sk->sk_cgrp->memcg->css);
538 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
540 if (!memcg || mem_cgroup_is_root(memcg))
543 return &memcg->tcp_mem.cg_proto;
545 EXPORT_SYMBOL(tcp_proto_cgroup);
547 static void disarm_sock_keys(struct mem_cgroup *memcg)
549 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
551 static_key_slow_dec(&memcg_socket_limit_enabled);
554 static void disarm_sock_keys(struct mem_cgroup *memcg)
559 #ifdef CONFIG_MEMCG_KMEM
561 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
562 * There are two main reasons for not using the css_id for this:
563 * 1) this works better in sparse environments, where we have a lot of memcgs,
564 * but only a few kmem-limited. Or also, if we have, for instance, 200
565 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
566 * 200 entry array for that.
568 * 2) In order not to violate the cgroup API, we would like to do all memory
569 * allocation in ->create(). At that point, we haven't yet allocated the
570 * css_id. Having a separate index prevents us from messing with the cgroup
573 * The current size of the caches array is stored in
574 * memcg_limited_groups_array_size. It will double each time we have to
577 static DEFINE_IDA(kmem_limited_groups);
578 int memcg_limited_groups_array_size;
581 * MIN_SIZE is different than 1, because we would like to avoid going through
582 * the alloc/free process all the time. In a small machine, 4 kmem-limited
583 * cgroups is a reasonable guess. In the future, it could be a parameter or
584 * tunable, but that is strictly not necessary.
586 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
587 * this constant directly from cgroup, but it is understandable that this is
588 * better kept as an internal representation in cgroup.c. In any case, the
589 * css_id space is not getting any smaller, and we don't have to necessarily
590 * increase ours as well if it increases.
592 #define MEMCG_CACHES_MIN_SIZE 4
593 #define MEMCG_CACHES_MAX_SIZE 65535
596 * A lot of the calls to the cache allocation functions are expected to be
597 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
598 * conditional to this static branch, we'll have to allow modules that does
599 * kmem_cache_alloc and the such to see this symbol as well
601 struct static_key memcg_kmem_enabled_key;
602 EXPORT_SYMBOL(memcg_kmem_enabled_key);
604 static void disarm_kmem_keys(struct mem_cgroup *memcg)
606 if (memcg_kmem_is_active(memcg)) {
607 static_key_slow_dec(&memcg_kmem_enabled_key);
608 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
611 * This check can't live in kmem destruction function,
612 * since the charges will outlive the cgroup
614 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
617 static void disarm_kmem_keys(struct mem_cgroup *memcg)
620 #endif /* CONFIG_MEMCG_KMEM */
622 static void disarm_static_keys(struct mem_cgroup *memcg)
624 disarm_sock_keys(memcg);
625 disarm_kmem_keys(memcg);
628 static void drain_all_stock_async(struct mem_cgroup *memcg);
630 static struct mem_cgroup_per_zone *
631 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
633 VM_BUG_ON((unsigned)nid >= nr_node_ids);
634 return &memcg->nodeinfo[nid]->zoneinfo[zid];
637 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
642 static struct mem_cgroup_per_zone *
643 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
645 int nid = page_to_nid(page);
646 int zid = page_zonenum(page);
648 return mem_cgroup_zoneinfo(memcg, nid, zid);
652 * Implementation Note: reading percpu statistics for memcg.
654 * Both of vmstat[] and percpu_counter has threshold and do periodic
655 * synchronization to implement "quick" read. There are trade-off between
656 * reading cost and precision of value. Then, we may have a chance to implement
657 * a periodic synchronizion of counter in memcg's counter.
659 * But this _read() function is used for user interface now. The user accounts
660 * memory usage by memory cgroup and he _always_ requires exact value because
661 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
662 * have to visit all online cpus and make sum. So, for now, unnecessary
663 * synchronization is not implemented. (just implemented for cpu hotplug)
665 * If there are kernel internal actions which can make use of some not-exact
666 * value, and reading all cpu value can be performance bottleneck in some
667 * common workload, threashold and synchonization as vmstat[] should be
670 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
671 enum mem_cgroup_stat_index idx)
677 for_each_online_cpu(cpu)
678 val += per_cpu(memcg->stat->count[idx], cpu);
679 #ifdef CONFIG_HOTPLUG_CPU
680 spin_lock(&memcg->pcp_counter_lock);
681 val += memcg->nocpu_base.count[idx];
682 spin_unlock(&memcg->pcp_counter_lock);
688 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
691 int val = (charge) ? 1 : -1;
692 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
695 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
696 enum mem_cgroup_events_index idx)
698 unsigned long val = 0;
701 for_each_online_cpu(cpu)
702 val += per_cpu(memcg->stat->events[idx], cpu);
703 #ifdef CONFIG_HOTPLUG_CPU
704 spin_lock(&memcg->pcp_counter_lock);
705 val += memcg->nocpu_base.events[idx];
706 spin_unlock(&memcg->pcp_counter_lock);
711 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
713 bool anon, int nr_pages)
718 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
719 * counted as CACHE even if it's on ANON LRU.
722 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
725 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
728 if (PageTransHuge(page))
729 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
732 /* pagein of a big page is an event. So, ignore page size */
734 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
736 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
737 nr_pages = -nr_pages; /* for event */
740 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
746 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
748 struct mem_cgroup_per_zone *mz;
750 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
751 return mz->lru_size[lru];
755 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
756 unsigned int lru_mask)
758 struct mem_cgroup_per_zone *mz;
760 unsigned long ret = 0;
762 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
765 if (BIT(lru) & lru_mask)
766 ret += mz->lru_size[lru];
772 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
773 int nid, unsigned int lru_mask)
778 for (zid = 0; zid < MAX_NR_ZONES; zid++)
779 total += mem_cgroup_zone_nr_lru_pages(memcg,
785 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
786 unsigned int lru_mask)
791 for_each_node_state(nid, N_MEMORY)
792 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
796 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
797 enum mem_cgroup_events_target target)
799 unsigned long val, next;
801 val = __this_cpu_read(memcg->stat->nr_page_events);
802 next = __this_cpu_read(memcg->stat->targets[target]);
803 /* from time_after() in jiffies.h */
804 if ((long)next - (long)val < 0) {
806 case MEM_CGROUP_TARGET_THRESH:
807 next = val + THRESHOLDS_EVENTS_TARGET;
809 case MEM_CGROUP_TARGET_SOFTLIMIT:
810 next = val + SOFTLIMIT_EVENTS_TARGET;
812 case MEM_CGROUP_TARGET_NUMAINFO:
813 next = val + NUMAINFO_EVENTS_TARGET;
818 __this_cpu_write(memcg->stat->targets[target], next);
825 * Called from rate-limited memcg_check_events when enough
826 * MEM_CGROUP_TARGET_SOFTLIMIT events are accumulated and it makes sure
827 * that all the parents up the hierarchy will be notified that this group
828 * is in excess or that it is not in excess anymore. mmecg->soft_contributed
829 * makes the transition a single action whenever the state flips from one to
832 static void mem_cgroup_update_soft_limit(struct mem_cgroup *memcg)
834 unsigned long long excess = res_counter_soft_limit_excess(&memcg->res);
835 struct mem_cgroup *parent = memcg;
838 spin_lock(&memcg->soft_lock);
840 if (!memcg->soft_contributed) {
842 memcg->soft_contributed = true;
845 if (memcg->soft_contributed) {
847 memcg->soft_contributed = false;
852 * Necessary to update all ancestors when hierarchy is used
853 * because their event counter is not touched.
855 while (delta && (parent = parent_mem_cgroup(parent)))
856 atomic_add(delta, &parent->children_in_excess);
857 spin_unlock(&memcg->soft_lock);
861 * Check events in order.
864 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
867 /* threshold event is triggered in finer grain than soft limit */
868 if (unlikely(mem_cgroup_event_ratelimit(memcg,
869 MEM_CGROUP_TARGET_THRESH))) {
871 bool do_numainfo __maybe_unused;
873 do_softlimit = mem_cgroup_event_ratelimit(memcg,
874 MEM_CGROUP_TARGET_SOFTLIMIT);
876 do_numainfo = mem_cgroup_event_ratelimit(memcg,
877 MEM_CGROUP_TARGET_NUMAINFO);
881 mem_cgroup_threshold(memcg);
882 if (unlikely(do_softlimit))
883 mem_cgroup_update_soft_limit(memcg);
885 if (unlikely(do_numainfo))
886 atomic_inc(&memcg->numainfo_events);
892 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
895 * mm_update_next_owner() may clear mm->owner to NULL
896 * if it races with swapoff, page migration, etc.
897 * So this can be called with p == NULL.
902 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
905 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
907 struct mem_cgroup *memcg = NULL;
912 * Because we have no locks, mm->owner's may be being moved to other
913 * cgroup. We use css_tryget() here even if this looks
914 * pessimistic (rather than adding locks here).
918 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
919 if (unlikely(!memcg))
921 } while (!css_tryget(&memcg->css));
926 static enum mem_cgroup_filter_t
927 mem_cgroup_filter(struct mem_cgroup *memcg, struct mem_cgroup *root,
928 mem_cgroup_iter_filter cond)
932 return cond(memcg, root);
936 * Returns a next (in a pre-order walk) alive memcg (with elevated css
937 * ref. count) or NULL if the whole root's subtree has been visited.
939 * helper function to be used by mem_cgroup_iter
941 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
942 struct mem_cgroup *last_visited, mem_cgroup_iter_filter cond)
944 struct cgroup_subsys_state *prev_css, *next_css;
946 prev_css = last_visited ? &last_visited->css : NULL;
948 next_css = css_next_descendant_pre(prev_css, &root->css);
951 * Even if we found a group we have to make sure it is
952 * alive. css && !memcg means that the groups should be
953 * skipped and we should continue the tree walk.
954 * last_visited css is safe to use because it is
955 * protected by css_get and the tree walk is rcu safe.
958 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
960 switch (mem_cgroup_filter(mem, root, cond)) {
968 * css_rightmost_descendant is not an optimal way to
969 * skip through a subtree (especially for imbalanced
970 * trees leaning to right) but that's what we have right
971 * now. More effective solution would be traversing
972 * right-up for first non-NULL without calling
973 * css_next_descendant_pre afterwards.
975 prev_css = css_rightmost_descendant(next_css);
978 if (css_tryget(&mem->css))
991 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
994 * When a group in the hierarchy below root is destroyed, the
995 * hierarchy iterator can no longer be trusted since it might
996 * have pointed to the destroyed group. Invalidate it.
998 atomic_inc(&root->dead_count);
1001 static struct mem_cgroup *
1002 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1003 struct mem_cgroup *root,
1006 struct mem_cgroup *position = NULL;
1008 * A cgroup destruction happens in two stages: offlining and
1009 * release. They are separated by a RCU grace period.
1011 * If the iterator is valid, we may still race with an
1012 * offlining. The RCU lock ensures the object won't be
1013 * released, tryget will fail if we lost the race.
1015 *sequence = atomic_read(&root->dead_count);
1016 if (iter->last_dead_count == *sequence) {
1018 position = iter->last_visited;
1019 if (position && !css_tryget(&position->css))
1025 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1026 struct mem_cgroup *last_visited,
1027 struct mem_cgroup *new_position,
1031 css_put(&last_visited->css);
1033 * We store the sequence count from the time @last_visited was
1034 * loaded successfully instead of rereading it here so that we
1035 * don't lose destruction events in between. We could have
1036 * raced with the destruction of @new_position after all.
1038 iter->last_visited = new_position;
1040 iter->last_dead_count = sequence;
1044 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1045 * @root: hierarchy root
1046 * @prev: previously returned memcg, NULL on first invocation
1047 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1048 * @cond: filter for visited nodes, NULL for no filter
1050 * Returns references to children of the hierarchy below @root, or
1051 * @root itself, or %NULL after a full round-trip.
1053 * Caller must pass the return value in @prev on subsequent
1054 * invocations for reference counting, or use mem_cgroup_iter_break()
1055 * to cancel a hierarchy walk before the round-trip is complete.
1057 * Reclaimers can specify a zone and a priority level in @reclaim to
1058 * divide up the memcgs in the hierarchy among all concurrent
1059 * reclaimers operating on the same zone and priority.
1061 struct mem_cgroup *mem_cgroup_iter_cond(struct mem_cgroup *root,
1062 struct mem_cgroup *prev,
1063 struct mem_cgroup_reclaim_cookie *reclaim,
1064 mem_cgroup_iter_filter cond)
1066 struct mem_cgroup *memcg = NULL;
1067 struct mem_cgroup *last_visited = NULL;
1069 if (mem_cgroup_disabled()) {
1070 /* first call must return non-NULL, second return NULL */
1071 return (struct mem_cgroup *)(unsigned long)!prev;
1075 root = root_mem_cgroup;
1077 if (prev && !reclaim)
1078 last_visited = prev;
1080 if (!root->use_hierarchy && root != root_mem_cgroup) {
1083 if (mem_cgroup_filter(root, root, cond) == VISIT)
1090 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1091 int uninitialized_var(seq);
1094 int nid = zone_to_nid(reclaim->zone);
1095 int zid = zone_idx(reclaim->zone);
1096 struct mem_cgroup_per_zone *mz;
1098 mz = mem_cgroup_zoneinfo(root, nid, zid);
1099 iter = &mz->reclaim_iter[reclaim->priority];
1100 if (prev && reclaim->generation != iter->generation) {
1101 iter->last_visited = NULL;
1105 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1108 memcg = __mem_cgroup_iter_next(root, last_visited, cond);
1111 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1115 else if (!prev && memcg)
1116 reclaim->generation = iter->generation;
1120 * We have finished the whole tree walk or no group has been
1121 * visited because filter told us to skip the root node.
1123 if (!memcg && (prev || (cond && !last_visited)))
1129 if (prev && prev != root)
1130 css_put(&prev->css);
1136 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1137 * @root: hierarchy root
1138 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1140 void mem_cgroup_iter_break(struct mem_cgroup *root,
1141 struct mem_cgroup *prev)
1144 root = root_mem_cgroup;
1145 if (prev && prev != root)
1146 css_put(&prev->css);
1150 * Iteration constructs for visiting all cgroups (under a tree). If
1151 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1152 * be used for reference counting.
1154 #define for_each_mem_cgroup_tree(iter, root) \
1155 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1157 iter = mem_cgroup_iter(root, iter, NULL))
1159 #define for_each_mem_cgroup(iter) \
1160 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1162 iter = mem_cgroup_iter(NULL, iter, NULL))
1164 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1166 struct mem_cgroup *memcg;
1169 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1170 if (unlikely(!memcg))
1175 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1178 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1186 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1189 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1190 * @zone: zone of the wanted lruvec
1191 * @memcg: memcg of the wanted lruvec
1193 * Returns the lru list vector holding pages for the given @zone and
1194 * @mem. This can be the global zone lruvec, if the memory controller
1197 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1198 struct mem_cgroup *memcg)
1200 struct mem_cgroup_per_zone *mz;
1201 struct lruvec *lruvec;
1203 if (mem_cgroup_disabled()) {
1204 lruvec = &zone->lruvec;
1208 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1209 lruvec = &mz->lruvec;
1212 * Since a node can be onlined after the mem_cgroup was created,
1213 * we have to be prepared to initialize lruvec->zone here;
1214 * and if offlined then reonlined, we need to reinitialize it.
1216 if (unlikely(lruvec->zone != zone))
1217 lruvec->zone = zone;
1222 * Following LRU functions are allowed to be used without PCG_LOCK.
1223 * Operations are called by routine of global LRU independently from memcg.
1224 * What we have to take care of here is validness of pc->mem_cgroup.
1226 * Changes to pc->mem_cgroup happens when
1229 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1230 * It is added to LRU before charge.
1231 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1232 * When moving account, the page is not on LRU. It's isolated.
1236 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1238 * @zone: zone of the page
1240 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1242 struct mem_cgroup_per_zone *mz;
1243 struct mem_cgroup *memcg;
1244 struct page_cgroup *pc;
1245 struct lruvec *lruvec;
1247 if (mem_cgroup_disabled()) {
1248 lruvec = &zone->lruvec;
1252 pc = lookup_page_cgroup(page);
1253 memcg = pc->mem_cgroup;
1256 * Surreptitiously switch any uncharged offlist page to root:
1257 * an uncharged page off lru does nothing to secure
1258 * its former mem_cgroup from sudden removal.
1260 * Our caller holds lru_lock, and PageCgroupUsed is updated
1261 * under page_cgroup lock: between them, they make all uses
1262 * of pc->mem_cgroup safe.
1264 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1265 pc->mem_cgroup = memcg = root_mem_cgroup;
1267 mz = page_cgroup_zoneinfo(memcg, page);
1268 lruvec = &mz->lruvec;
1271 * Since a node can be onlined after the mem_cgroup was created,
1272 * we have to be prepared to initialize lruvec->zone here;
1273 * and if offlined then reonlined, we need to reinitialize it.
1275 if (unlikely(lruvec->zone != zone))
1276 lruvec->zone = zone;
1281 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1282 * @lruvec: mem_cgroup per zone lru vector
1283 * @lru: index of lru list the page is sitting on
1284 * @nr_pages: positive when adding or negative when removing
1286 * This function must be called when a page is added to or removed from an
1289 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1292 struct mem_cgroup_per_zone *mz;
1293 unsigned long *lru_size;
1295 if (mem_cgroup_disabled())
1298 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1299 lru_size = mz->lru_size + lru;
1300 *lru_size += nr_pages;
1301 VM_BUG_ON((long)(*lru_size) < 0);
1305 * Checks whether given mem is same or in the root_mem_cgroup's
1308 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1309 struct mem_cgroup *memcg)
1311 if (root_memcg == memcg)
1313 if (!root_memcg->use_hierarchy || !memcg)
1315 return css_is_ancestor(&memcg->css, &root_memcg->css);
1318 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1319 struct mem_cgroup *memcg)
1324 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1329 bool task_in_mem_cgroup(struct task_struct *task,
1330 const struct mem_cgroup *memcg)
1332 struct mem_cgroup *curr = NULL;
1333 struct task_struct *p;
1336 p = find_lock_task_mm(task);
1338 curr = try_get_mem_cgroup_from_mm(p->mm);
1342 * All threads may have already detached their mm's, but the oom
1343 * killer still needs to detect if they have already been oom
1344 * killed to prevent needlessly killing additional tasks.
1347 curr = mem_cgroup_from_task(task);
1349 css_get(&curr->css);
1355 * We should check use_hierarchy of "memcg" not "curr". Because checking
1356 * use_hierarchy of "curr" here make this function true if hierarchy is
1357 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1358 * hierarchy(even if use_hierarchy is disabled in "memcg").
1360 ret = mem_cgroup_same_or_subtree(memcg, curr);
1361 css_put(&curr->css);
1365 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1367 unsigned long inactive_ratio;
1368 unsigned long inactive;
1369 unsigned long active;
1372 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1373 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1375 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1377 inactive_ratio = int_sqrt(10 * gb);
1381 return inactive * inactive_ratio < active;
1384 #define mem_cgroup_from_res_counter(counter, member) \
1385 container_of(counter, struct mem_cgroup, member)
1388 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1389 * @memcg: the memory cgroup
1391 * Returns the maximum amount of memory @mem can be charged with, in
1394 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1396 unsigned long long margin;
1398 margin = res_counter_margin(&memcg->res);
1399 if (do_swap_account)
1400 margin = min(margin, res_counter_margin(&memcg->memsw));
1401 return margin >> PAGE_SHIFT;
1404 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1407 if (!css_parent(&memcg->css))
1408 return vm_swappiness;
1410 return memcg->swappiness;
1414 * memcg->moving_account is used for checking possibility that some thread is
1415 * calling move_account(). When a thread on CPU-A starts moving pages under
1416 * a memcg, other threads should check memcg->moving_account under
1417 * rcu_read_lock(), like this:
1421 * memcg->moving_account+1 if (memcg->mocing_account)
1423 * synchronize_rcu() update something.
1428 /* for quick checking without looking up memcg */
1429 atomic_t memcg_moving __read_mostly;
1431 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1433 atomic_inc(&memcg_moving);
1434 atomic_inc(&memcg->moving_account);
1438 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1441 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1442 * We check NULL in callee rather than caller.
1445 atomic_dec(&memcg_moving);
1446 atomic_dec(&memcg->moving_account);
1451 * 2 routines for checking "mem" is under move_account() or not.
1453 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1454 * is used for avoiding races in accounting. If true,
1455 * pc->mem_cgroup may be overwritten.
1457 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1458 * under hierarchy of moving cgroups. This is for
1459 * waiting at hith-memory prressure caused by "move".
1462 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1464 VM_BUG_ON(!rcu_read_lock_held());
1465 return atomic_read(&memcg->moving_account) > 0;
1468 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1470 struct mem_cgroup *from;
1471 struct mem_cgroup *to;
1474 * Unlike task_move routines, we access mc.to, mc.from not under
1475 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1477 spin_lock(&mc.lock);
1483 ret = mem_cgroup_same_or_subtree(memcg, from)
1484 || mem_cgroup_same_or_subtree(memcg, to);
1486 spin_unlock(&mc.lock);
1490 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1492 if (mc.moving_task && current != mc.moving_task) {
1493 if (mem_cgroup_under_move(memcg)) {
1495 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1496 /* moving charge context might have finished. */
1499 finish_wait(&mc.waitq, &wait);
1507 * Take this lock when
1508 * - a code tries to modify page's memcg while it's USED.
1509 * - a code tries to modify page state accounting in a memcg.
1510 * see mem_cgroup_stolen(), too.
1512 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1513 unsigned long *flags)
1515 spin_lock_irqsave(&memcg->move_lock, *flags);
1518 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1519 unsigned long *flags)
1521 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1524 #define K(x) ((x) << (PAGE_SHIFT-10))
1526 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1527 * @memcg: The memory cgroup that went over limit
1528 * @p: Task that is going to be killed
1530 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1533 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1535 struct cgroup *task_cgrp;
1536 struct cgroup *mem_cgrp;
1538 * Need a buffer in BSS, can't rely on allocations. The code relies
1539 * on the assumption that OOM is serialized for memory controller.
1540 * If this assumption is broken, revisit this code.
1542 static char memcg_name[PATH_MAX];
1544 struct mem_cgroup *iter;
1552 mem_cgrp = memcg->css.cgroup;
1553 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1555 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1558 * Unfortunately, we are unable to convert to a useful name
1559 * But we'll still print out the usage information
1566 pr_info("Task in %s killed", memcg_name);
1569 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1577 * Continues from above, so we don't need an KERN_ level
1579 pr_cont(" as a result of limit of %s\n", memcg_name);
1582 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1583 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1584 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1585 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1586 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1587 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1588 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1589 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1590 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1591 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1592 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1593 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1595 for_each_mem_cgroup_tree(iter, memcg) {
1596 pr_info("Memory cgroup stats");
1599 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1601 pr_cont(" for %s", memcg_name);
1605 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1606 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1608 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1609 K(mem_cgroup_read_stat(iter, i)));
1612 for (i = 0; i < NR_LRU_LISTS; i++)
1613 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1614 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1621 * This function returns the number of memcg under hierarchy tree. Returns
1622 * 1(self count) if no children.
1624 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1627 struct mem_cgroup *iter;
1629 for_each_mem_cgroup_tree(iter, memcg)
1635 * Return the memory (and swap, if configured) limit for a memcg.
1637 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1641 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1644 * Do not consider swap space if we cannot swap due to swappiness
1646 if (mem_cgroup_swappiness(memcg)) {
1649 limit += total_swap_pages << PAGE_SHIFT;
1650 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1653 * If memsw is finite and limits the amount of swap space
1654 * available to this memcg, return that limit.
1656 limit = min(limit, memsw);
1662 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1665 struct mem_cgroup *iter;
1666 unsigned long chosen_points = 0;
1667 unsigned long totalpages;
1668 unsigned int points = 0;
1669 struct task_struct *chosen = NULL;
1672 * If current has a pending SIGKILL or is exiting, then automatically
1673 * select it. The goal is to allow it to allocate so that it may
1674 * quickly exit and free its memory.
1676 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1677 set_thread_flag(TIF_MEMDIE);
1681 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1682 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1683 for_each_mem_cgroup_tree(iter, memcg) {
1684 struct css_task_iter it;
1685 struct task_struct *task;
1687 css_task_iter_start(&iter->css, &it);
1688 while ((task = css_task_iter_next(&it))) {
1689 switch (oom_scan_process_thread(task, totalpages, NULL,
1691 case OOM_SCAN_SELECT:
1693 put_task_struct(chosen);
1695 chosen_points = ULONG_MAX;
1696 get_task_struct(chosen);
1698 case OOM_SCAN_CONTINUE:
1700 case OOM_SCAN_ABORT:
1701 css_task_iter_end(&it);
1702 mem_cgroup_iter_break(memcg, iter);
1704 put_task_struct(chosen);
1709 points = oom_badness(task, memcg, NULL, totalpages);
1710 if (points > chosen_points) {
1712 put_task_struct(chosen);
1714 chosen_points = points;
1715 get_task_struct(chosen);
1718 css_task_iter_end(&it);
1723 points = chosen_points * 1000 / totalpages;
1724 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1725 NULL, "Memory cgroup out of memory");
1728 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1730 unsigned long flags)
1732 unsigned long total = 0;
1733 bool noswap = false;
1736 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1738 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1741 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1743 drain_all_stock_async(memcg);
1744 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1746 * Allow limit shrinkers, which are triggered directly
1747 * by userspace, to catch signals and stop reclaim
1748 * after minimal progress, regardless of the margin.
1750 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1752 if (mem_cgroup_margin(memcg))
1755 * If nothing was reclaimed after two attempts, there
1756 * may be no reclaimable pages in this hierarchy.
1764 #if MAX_NUMNODES > 1
1766 * test_mem_cgroup_node_reclaimable
1767 * @memcg: the target memcg
1768 * @nid: the node ID to be checked.
1769 * @noswap : specify true here if the user wants flle only information.
1771 * This function returns whether the specified memcg contains any
1772 * reclaimable pages on a node. Returns true if there are any reclaimable
1773 * pages in the node.
1775 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1776 int nid, bool noswap)
1778 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1780 if (noswap || !total_swap_pages)
1782 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1789 * Always updating the nodemask is not very good - even if we have an empty
1790 * list or the wrong list here, we can start from some node and traverse all
1791 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1794 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1798 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1799 * pagein/pageout changes since the last update.
1801 if (!atomic_read(&memcg->numainfo_events))
1803 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1806 /* make a nodemask where this memcg uses memory from */
1807 memcg->scan_nodes = node_states[N_MEMORY];
1809 for_each_node_mask(nid, node_states[N_MEMORY]) {
1811 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1812 node_clear(nid, memcg->scan_nodes);
1815 atomic_set(&memcg->numainfo_events, 0);
1816 atomic_set(&memcg->numainfo_updating, 0);
1820 * Selecting a node where we start reclaim from. Because what we need is just
1821 * reducing usage counter, start from anywhere is O,K. Considering
1822 * memory reclaim from current node, there are pros. and cons.
1824 * Freeing memory from current node means freeing memory from a node which
1825 * we'll use or we've used. So, it may make LRU bad. And if several threads
1826 * hit limits, it will see a contention on a node. But freeing from remote
1827 * node means more costs for memory reclaim because of memory latency.
1829 * Now, we use round-robin. Better algorithm is welcomed.
1831 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1835 mem_cgroup_may_update_nodemask(memcg);
1836 node = memcg->last_scanned_node;
1838 node = next_node(node, memcg->scan_nodes);
1839 if (node == MAX_NUMNODES)
1840 node = first_node(memcg->scan_nodes);
1842 * We call this when we hit limit, not when pages are added to LRU.
1843 * No LRU may hold pages because all pages are UNEVICTABLE or
1844 * memcg is too small and all pages are not on LRU. In that case,
1845 * we use curret node.
1847 if (unlikely(node == MAX_NUMNODES))
1848 node = numa_node_id();
1850 memcg->last_scanned_node = node;
1855 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1863 * A group is eligible for the soft limit reclaim under the given root
1865 * a) it is over its soft limit
1866 * b) any parent up the hierarchy is over its soft limit
1868 * If the given group doesn't have any children over the limit then it
1869 * doesn't make any sense to iterate its subtree.
1871 enum mem_cgroup_filter_t
1872 mem_cgroup_soft_reclaim_eligible(struct mem_cgroup *memcg,
1873 struct mem_cgroup *root)
1875 struct mem_cgroup *parent;
1878 memcg = root_mem_cgroup;
1881 if (res_counter_soft_limit_excess(&memcg->res))
1885 * If any parent up to the root in the hierarchy is over its soft limit
1886 * then we have to obey and reclaim from this group as well.
1888 while ((parent = parent_mem_cgroup(parent))) {
1889 if (res_counter_soft_limit_excess(&parent->res))
1895 if (!atomic_read(&memcg->children_in_excess))
1900 static DEFINE_SPINLOCK(memcg_oom_lock);
1903 * Check OOM-Killer is already running under our hierarchy.
1904 * If someone is running, return false.
1906 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1908 struct mem_cgroup *iter, *failed = NULL;
1910 spin_lock(&memcg_oom_lock);
1912 for_each_mem_cgroup_tree(iter, memcg) {
1913 if (iter->oom_lock) {
1915 * this subtree of our hierarchy is already locked
1916 * so we cannot give a lock.
1919 mem_cgroup_iter_break(memcg, iter);
1922 iter->oom_lock = true;
1927 * OK, we failed to lock the whole subtree so we have
1928 * to clean up what we set up to the failing subtree
1930 for_each_mem_cgroup_tree(iter, memcg) {
1931 if (iter == failed) {
1932 mem_cgroup_iter_break(memcg, iter);
1935 iter->oom_lock = false;
1939 spin_unlock(&memcg_oom_lock);
1944 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1946 struct mem_cgroup *iter;
1948 spin_lock(&memcg_oom_lock);
1949 for_each_mem_cgroup_tree(iter, memcg)
1950 iter->oom_lock = false;
1951 spin_unlock(&memcg_oom_lock);
1954 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1956 struct mem_cgroup *iter;
1958 for_each_mem_cgroup_tree(iter, memcg)
1959 atomic_inc(&iter->under_oom);
1962 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1964 struct mem_cgroup *iter;
1967 * When a new child is created while the hierarchy is under oom,
1968 * mem_cgroup_oom_lock() may not be called. We have to use
1969 * atomic_add_unless() here.
1971 for_each_mem_cgroup_tree(iter, memcg)
1972 atomic_add_unless(&iter->under_oom, -1, 0);
1975 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1977 struct oom_wait_info {
1978 struct mem_cgroup *memcg;
1982 static int memcg_oom_wake_function(wait_queue_t *wait,
1983 unsigned mode, int sync, void *arg)
1985 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1986 struct mem_cgroup *oom_wait_memcg;
1987 struct oom_wait_info *oom_wait_info;
1989 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1990 oom_wait_memcg = oom_wait_info->memcg;
1993 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1994 * Then we can use css_is_ancestor without taking care of RCU.
1996 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1997 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1999 return autoremove_wake_function(wait, mode, sync, arg);
2002 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2004 atomic_inc(&memcg->oom_wakeups);
2005 /* for filtering, pass "memcg" as argument. */
2006 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2009 static void memcg_oom_recover(struct mem_cgroup *memcg)
2011 if (memcg && atomic_read(&memcg->under_oom))
2012 memcg_wakeup_oom(memcg);
2016 * try to call OOM killer
2018 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2023 if (!current->memcg_oom.may_oom)
2026 current->memcg_oom.in_memcg_oom = 1;
2029 * As with any blocking lock, a contender needs to start
2030 * listening for wakeups before attempting the trylock,
2031 * otherwise it can miss the wakeup from the unlock and sleep
2032 * indefinitely. This is just open-coded because our locking
2033 * is so particular to memcg hierarchies.
2035 wakeups = atomic_read(&memcg->oom_wakeups);
2036 mem_cgroup_mark_under_oom(memcg);
2038 locked = mem_cgroup_oom_trylock(memcg);
2041 mem_cgroup_oom_notify(memcg);
2043 if (locked && !memcg->oom_kill_disable) {
2044 mem_cgroup_unmark_under_oom(memcg);
2045 mem_cgroup_out_of_memory(memcg, mask, order);
2046 mem_cgroup_oom_unlock(memcg);
2048 * There is no guarantee that an OOM-lock contender
2049 * sees the wakeups triggered by the OOM kill
2050 * uncharges. Wake any sleepers explicitely.
2052 memcg_oom_recover(memcg);
2055 * A system call can just return -ENOMEM, but if this
2056 * is a page fault and somebody else is handling the
2057 * OOM already, we need to sleep on the OOM waitqueue
2058 * for this memcg until the situation is resolved.
2059 * Which can take some time because it might be
2060 * handled by a userspace task.
2062 * However, this is the charge context, which means
2063 * that we may sit on a large call stack and hold
2064 * various filesystem locks, the mmap_sem etc. and we
2065 * don't want the OOM handler to deadlock on them
2066 * while we sit here and wait. Store the current OOM
2067 * context in the task_struct, then return -ENOMEM.
2068 * At the end of the page fault handler, with the
2069 * stack unwound, pagefault_out_of_memory() will check
2070 * back with us by calling
2071 * mem_cgroup_oom_synchronize(), possibly putting the
2074 current->memcg_oom.oom_locked = locked;
2075 current->memcg_oom.wakeups = wakeups;
2076 css_get(&memcg->css);
2077 current->memcg_oom.wait_on_memcg = memcg;
2082 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2084 * This has to be called at the end of a page fault if the the memcg
2085 * OOM handler was enabled and the fault is returning %VM_FAULT_OOM.
2087 * Memcg supports userspace OOM handling, so failed allocations must
2088 * sleep on a waitqueue until the userspace task resolves the
2089 * situation. Sleeping directly in the charge context with all kinds
2090 * of locks held is not a good idea, instead we remember an OOM state
2091 * in the task and mem_cgroup_oom_synchronize() has to be called at
2092 * the end of the page fault to put the task to sleep and clean up the
2095 * Returns %true if an ongoing memcg OOM situation was detected and
2096 * finalized, %false otherwise.
2098 bool mem_cgroup_oom_synchronize(void)
2100 struct oom_wait_info owait;
2101 struct mem_cgroup *memcg;
2103 /* OOM is global, do not handle */
2104 if (!current->memcg_oom.in_memcg_oom)
2108 * We invoked the OOM killer but there is a chance that a kill
2109 * did not free up any charges. Everybody else might already
2110 * be sleeping, so restart the fault and keep the rampage
2111 * going until some charges are released.
2113 memcg = current->memcg_oom.wait_on_memcg;
2117 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2120 owait.memcg = memcg;
2121 owait.wait.flags = 0;
2122 owait.wait.func = memcg_oom_wake_function;
2123 owait.wait.private = current;
2124 INIT_LIST_HEAD(&owait.wait.task_list);
2126 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2127 /* Only sleep if we didn't miss any wakeups since OOM */
2128 if (atomic_read(&memcg->oom_wakeups) == current->memcg_oom.wakeups)
2130 finish_wait(&memcg_oom_waitq, &owait.wait);
2132 mem_cgroup_unmark_under_oom(memcg);
2133 if (current->memcg_oom.oom_locked) {
2134 mem_cgroup_oom_unlock(memcg);
2136 * There is no guarantee that an OOM-lock contender
2137 * sees the wakeups triggered by the OOM kill
2138 * uncharges. Wake any sleepers explicitely.
2140 memcg_oom_recover(memcg);
2142 css_put(&memcg->css);
2143 current->memcg_oom.wait_on_memcg = NULL;
2145 current->memcg_oom.in_memcg_oom = 0;
2150 * Currently used to update mapped file statistics, but the routine can be
2151 * generalized to update other statistics as well.
2153 * Notes: Race condition
2155 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2156 * it tends to be costly. But considering some conditions, we doesn't need
2157 * to do so _always_.
2159 * Considering "charge", lock_page_cgroup() is not required because all
2160 * file-stat operations happen after a page is attached to radix-tree. There
2161 * are no race with "charge".
2163 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2164 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2165 * if there are race with "uncharge". Statistics itself is properly handled
2168 * Considering "move", this is an only case we see a race. To make the race
2169 * small, we check mm->moving_account and detect there are possibility of race
2170 * If there is, we take a lock.
2173 void __mem_cgroup_begin_update_page_stat(struct page *page,
2174 bool *locked, unsigned long *flags)
2176 struct mem_cgroup *memcg;
2177 struct page_cgroup *pc;
2179 pc = lookup_page_cgroup(page);
2181 memcg = pc->mem_cgroup;
2182 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2185 * If this memory cgroup is not under account moving, we don't
2186 * need to take move_lock_mem_cgroup(). Because we already hold
2187 * rcu_read_lock(), any calls to move_account will be delayed until
2188 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2190 if (!mem_cgroup_stolen(memcg))
2193 move_lock_mem_cgroup(memcg, flags);
2194 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2195 move_unlock_mem_cgroup(memcg, flags);
2201 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2203 struct page_cgroup *pc = lookup_page_cgroup(page);
2206 * It's guaranteed that pc->mem_cgroup never changes while
2207 * lock is held because a routine modifies pc->mem_cgroup
2208 * should take move_lock_mem_cgroup().
2210 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2213 void mem_cgroup_update_page_stat(struct page *page,
2214 enum mem_cgroup_stat_index idx, int val)
2216 struct mem_cgroup *memcg;
2217 struct page_cgroup *pc = lookup_page_cgroup(page);
2218 unsigned long uninitialized_var(flags);
2220 if (mem_cgroup_disabled())
2223 VM_BUG_ON(!rcu_read_lock_held());
2224 memcg = pc->mem_cgroup;
2225 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2228 this_cpu_add(memcg->stat->count[idx], val);
2232 * size of first charge trial. "32" comes from vmscan.c's magic value.
2233 * TODO: maybe necessary to use big numbers in big irons.
2235 #define CHARGE_BATCH 32U
2236 struct memcg_stock_pcp {
2237 struct mem_cgroup *cached; /* this never be root cgroup */
2238 unsigned int nr_pages;
2239 struct work_struct work;
2240 unsigned long flags;
2241 #define FLUSHING_CACHED_CHARGE 0
2243 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2244 static DEFINE_MUTEX(percpu_charge_mutex);
2247 * consume_stock: Try to consume stocked charge on this cpu.
2248 * @memcg: memcg to consume from.
2249 * @nr_pages: how many pages to charge.
2251 * The charges will only happen if @memcg matches the current cpu's memcg
2252 * stock, and at least @nr_pages are available in that stock. Failure to
2253 * service an allocation will refill the stock.
2255 * returns true if successful, false otherwise.
2257 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2259 struct memcg_stock_pcp *stock;
2262 if (nr_pages > CHARGE_BATCH)
2265 stock = &get_cpu_var(memcg_stock);
2266 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2267 stock->nr_pages -= nr_pages;
2268 else /* need to call res_counter_charge */
2270 put_cpu_var(memcg_stock);
2275 * Returns stocks cached in percpu to res_counter and reset cached information.
2277 static void drain_stock(struct memcg_stock_pcp *stock)
2279 struct mem_cgroup *old = stock->cached;
2281 if (stock->nr_pages) {
2282 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2284 res_counter_uncharge(&old->res, bytes);
2285 if (do_swap_account)
2286 res_counter_uncharge(&old->memsw, bytes);
2287 stock->nr_pages = 0;
2289 stock->cached = NULL;
2293 * This must be called under preempt disabled or must be called by
2294 * a thread which is pinned to local cpu.
2296 static void drain_local_stock(struct work_struct *dummy)
2298 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2300 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2303 static void __init memcg_stock_init(void)
2307 for_each_possible_cpu(cpu) {
2308 struct memcg_stock_pcp *stock =
2309 &per_cpu(memcg_stock, cpu);
2310 INIT_WORK(&stock->work, drain_local_stock);
2315 * Cache charges(val) which is from res_counter, to local per_cpu area.
2316 * This will be consumed by consume_stock() function, later.
2318 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2320 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2322 if (stock->cached != memcg) { /* reset if necessary */
2324 stock->cached = memcg;
2326 stock->nr_pages += nr_pages;
2327 put_cpu_var(memcg_stock);
2331 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2332 * of the hierarchy under it. sync flag says whether we should block
2333 * until the work is done.
2335 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2339 /* Notify other cpus that system-wide "drain" is running */
2342 for_each_online_cpu(cpu) {
2343 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2344 struct mem_cgroup *memcg;
2346 memcg = stock->cached;
2347 if (!memcg || !stock->nr_pages)
2349 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2351 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2353 drain_local_stock(&stock->work);
2355 schedule_work_on(cpu, &stock->work);
2363 for_each_online_cpu(cpu) {
2364 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2365 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2366 flush_work(&stock->work);
2373 * Tries to drain stocked charges in other cpus. This function is asynchronous
2374 * and just put a work per cpu for draining localy on each cpu. Caller can
2375 * expects some charges will be back to res_counter later but cannot wait for
2378 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2381 * If someone calls draining, avoid adding more kworker runs.
2383 if (!mutex_trylock(&percpu_charge_mutex))
2385 drain_all_stock(root_memcg, false);
2386 mutex_unlock(&percpu_charge_mutex);
2389 /* This is a synchronous drain interface. */
2390 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2392 /* called when force_empty is called */
2393 mutex_lock(&percpu_charge_mutex);
2394 drain_all_stock(root_memcg, true);
2395 mutex_unlock(&percpu_charge_mutex);
2399 * This function drains percpu counter value from DEAD cpu and
2400 * move it to local cpu. Note that this function can be preempted.
2402 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2406 spin_lock(&memcg->pcp_counter_lock);
2407 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2408 long x = per_cpu(memcg->stat->count[i], cpu);
2410 per_cpu(memcg->stat->count[i], cpu) = 0;
2411 memcg->nocpu_base.count[i] += x;
2413 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2414 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2416 per_cpu(memcg->stat->events[i], cpu) = 0;
2417 memcg->nocpu_base.events[i] += x;
2419 spin_unlock(&memcg->pcp_counter_lock);
2422 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2423 unsigned long action,
2426 int cpu = (unsigned long)hcpu;
2427 struct memcg_stock_pcp *stock;
2428 struct mem_cgroup *iter;
2430 if (action == CPU_ONLINE)
2433 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2436 for_each_mem_cgroup(iter)
2437 mem_cgroup_drain_pcp_counter(iter, cpu);
2439 stock = &per_cpu(memcg_stock, cpu);
2445 /* See __mem_cgroup_try_charge() for details */
2447 CHARGE_OK, /* success */
2448 CHARGE_RETRY, /* need to retry but retry is not bad */
2449 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2450 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2453 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2454 unsigned int nr_pages, unsigned int min_pages,
2457 unsigned long csize = nr_pages * PAGE_SIZE;
2458 struct mem_cgroup *mem_over_limit;
2459 struct res_counter *fail_res;
2460 unsigned long flags = 0;
2463 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2466 if (!do_swap_account)
2468 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2472 res_counter_uncharge(&memcg->res, csize);
2473 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2474 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2476 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2478 * Never reclaim on behalf of optional batching, retry with a
2479 * single page instead.
2481 if (nr_pages > min_pages)
2482 return CHARGE_RETRY;
2484 if (!(gfp_mask & __GFP_WAIT))
2485 return CHARGE_WOULDBLOCK;
2487 if (gfp_mask & __GFP_NORETRY)
2488 return CHARGE_NOMEM;
2490 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2491 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2492 return CHARGE_RETRY;
2494 * Even though the limit is exceeded at this point, reclaim
2495 * may have been able to free some pages. Retry the charge
2496 * before killing the task.
2498 * Only for regular pages, though: huge pages are rather
2499 * unlikely to succeed so close to the limit, and we fall back
2500 * to regular pages anyway in case of failure.
2502 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2503 return CHARGE_RETRY;
2506 * At task move, charge accounts can be doubly counted. So, it's
2507 * better to wait until the end of task_move if something is going on.
2509 if (mem_cgroup_wait_acct_move(mem_over_limit))
2510 return CHARGE_RETRY;
2513 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2515 return CHARGE_NOMEM;
2519 * __mem_cgroup_try_charge() does
2520 * 1. detect memcg to be charged against from passed *mm and *ptr,
2521 * 2. update res_counter
2522 * 3. call memory reclaim if necessary.
2524 * In some special case, if the task is fatal, fatal_signal_pending() or
2525 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2526 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2527 * as possible without any hazards. 2: all pages should have a valid
2528 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2529 * pointer, that is treated as a charge to root_mem_cgroup.
2531 * So __mem_cgroup_try_charge() will return
2532 * 0 ... on success, filling *ptr with a valid memcg pointer.
2533 * -ENOMEM ... charge failure because of resource limits.
2534 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2536 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2537 * the oom-killer can be invoked.
2539 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2541 unsigned int nr_pages,
2542 struct mem_cgroup **ptr,
2545 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2546 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2547 struct mem_cgroup *memcg = NULL;
2551 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2552 * in system level. So, allow to go ahead dying process in addition to
2555 if (unlikely(test_thread_flag(TIF_MEMDIE)
2556 || fatal_signal_pending(current)))
2560 * We always charge the cgroup the mm_struct belongs to.
2561 * The mm_struct's mem_cgroup changes on task migration if the
2562 * thread group leader migrates. It's possible that mm is not
2563 * set, if so charge the root memcg (happens for pagecache usage).
2566 *ptr = root_mem_cgroup;
2568 if (*ptr) { /* css should be a valid one */
2570 if (mem_cgroup_is_root(memcg))
2572 if (consume_stock(memcg, nr_pages))
2574 css_get(&memcg->css);
2576 struct task_struct *p;
2579 p = rcu_dereference(mm->owner);
2581 * Because we don't have task_lock(), "p" can exit.
2582 * In that case, "memcg" can point to root or p can be NULL with
2583 * race with swapoff. Then, we have small risk of mis-accouning.
2584 * But such kind of mis-account by race always happens because
2585 * we don't have cgroup_mutex(). It's overkill and we allo that
2587 * (*) swapoff at el will charge against mm-struct not against
2588 * task-struct. So, mm->owner can be NULL.
2590 memcg = mem_cgroup_from_task(p);
2592 memcg = root_mem_cgroup;
2593 if (mem_cgroup_is_root(memcg)) {
2597 if (consume_stock(memcg, nr_pages)) {
2599 * It seems dagerous to access memcg without css_get().
2600 * But considering how consume_stok works, it's not
2601 * necessary. If consume_stock success, some charges
2602 * from this memcg are cached on this cpu. So, we
2603 * don't need to call css_get()/css_tryget() before
2604 * calling consume_stock().
2609 /* after here, we may be blocked. we need to get refcnt */
2610 if (!css_tryget(&memcg->css)) {
2618 bool invoke_oom = oom && !nr_oom_retries;
2620 /* If killed, bypass charge */
2621 if (fatal_signal_pending(current)) {
2622 css_put(&memcg->css);
2626 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2627 nr_pages, invoke_oom);
2631 case CHARGE_RETRY: /* not in OOM situation but retry */
2633 css_put(&memcg->css);
2636 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2637 css_put(&memcg->css);
2639 case CHARGE_NOMEM: /* OOM routine works */
2640 if (!oom || invoke_oom) {
2641 css_put(&memcg->css);
2647 } while (ret != CHARGE_OK);
2649 if (batch > nr_pages)
2650 refill_stock(memcg, batch - nr_pages);
2651 css_put(&memcg->css);
2659 *ptr = root_mem_cgroup;
2664 * Somemtimes we have to undo a charge we got by try_charge().
2665 * This function is for that and do uncharge, put css's refcnt.
2666 * gotten by try_charge().
2668 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2669 unsigned int nr_pages)
2671 if (!mem_cgroup_is_root(memcg)) {
2672 unsigned long bytes = nr_pages * PAGE_SIZE;
2674 res_counter_uncharge(&memcg->res, bytes);
2675 if (do_swap_account)
2676 res_counter_uncharge(&memcg->memsw, bytes);
2681 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2682 * This is useful when moving usage to parent cgroup.
2684 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2685 unsigned int nr_pages)
2687 unsigned long bytes = nr_pages * PAGE_SIZE;
2689 if (mem_cgroup_is_root(memcg))
2692 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2693 if (do_swap_account)
2694 res_counter_uncharge_until(&memcg->memsw,
2695 memcg->memsw.parent, bytes);
2699 * A helper function to get mem_cgroup from ID. must be called under
2700 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2701 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2702 * called against removed memcg.)
2704 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2706 struct cgroup_subsys_state *css;
2708 /* ID 0 is unused ID */
2711 css = css_lookup(&mem_cgroup_subsys, id);
2714 return mem_cgroup_from_css(css);
2717 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2719 struct mem_cgroup *memcg = NULL;
2720 struct page_cgroup *pc;
2724 VM_BUG_ON(!PageLocked(page));
2726 pc = lookup_page_cgroup(page);
2727 lock_page_cgroup(pc);
2728 if (PageCgroupUsed(pc)) {
2729 memcg = pc->mem_cgroup;
2730 if (memcg && !css_tryget(&memcg->css))
2732 } else if (PageSwapCache(page)) {
2733 ent.val = page_private(page);
2734 id = lookup_swap_cgroup_id(ent);
2736 memcg = mem_cgroup_lookup(id);
2737 if (memcg && !css_tryget(&memcg->css))
2741 unlock_page_cgroup(pc);
2745 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2747 unsigned int nr_pages,
2748 enum charge_type ctype,
2751 struct page_cgroup *pc = lookup_page_cgroup(page);
2752 struct zone *uninitialized_var(zone);
2753 struct lruvec *lruvec;
2754 bool was_on_lru = false;
2757 lock_page_cgroup(pc);
2758 VM_BUG_ON(PageCgroupUsed(pc));
2760 * we don't need page_cgroup_lock about tail pages, becase they are not
2761 * accessed by any other context at this point.
2765 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2766 * may already be on some other mem_cgroup's LRU. Take care of it.
2769 zone = page_zone(page);
2770 spin_lock_irq(&zone->lru_lock);
2771 if (PageLRU(page)) {
2772 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2774 del_page_from_lru_list(page, lruvec, page_lru(page));
2779 pc->mem_cgroup = memcg;
2781 * We access a page_cgroup asynchronously without lock_page_cgroup().
2782 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2783 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2784 * before USED bit, we need memory barrier here.
2785 * See mem_cgroup_add_lru_list(), etc.
2788 SetPageCgroupUsed(pc);
2792 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2793 VM_BUG_ON(PageLRU(page));
2795 add_page_to_lru_list(page, lruvec, page_lru(page));
2797 spin_unlock_irq(&zone->lru_lock);
2800 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2805 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2806 unlock_page_cgroup(pc);
2809 * "charge_statistics" updated event counter.
2811 memcg_check_events(memcg, page);
2814 static DEFINE_MUTEX(set_limit_mutex);
2816 #ifdef CONFIG_MEMCG_KMEM
2817 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2819 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2820 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2824 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2825 * in the memcg_cache_params struct.
2827 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2829 struct kmem_cache *cachep;
2831 VM_BUG_ON(p->is_root_cache);
2832 cachep = p->root_cache;
2833 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2836 #ifdef CONFIG_SLABINFO
2837 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2838 struct cftype *cft, struct seq_file *m)
2840 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2841 struct memcg_cache_params *params;
2843 if (!memcg_can_account_kmem(memcg))
2846 print_slabinfo_header(m);
2848 mutex_lock(&memcg->slab_caches_mutex);
2849 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2850 cache_show(memcg_params_to_cache(params), m);
2851 mutex_unlock(&memcg->slab_caches_mutex);
2857 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2859 struct res_counter *fail_res;
2860 struct mem_cgroup *_memcg;
2864 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2869 * Conditions under which we can wait for the oom_killer. Those are
2870 * the same conditions tested by the core page allocator
2872 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2875 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2878 if (ret == -EINTR) {
2880 * __mem_cgroup_try_charge() chosed to bypass to root due to
2881 * OOM kill or fatal signal. Since our only options are to
2882 * either fail the allocation or charge it to this cgroup, do
2883 * it as a temporary condition. But we can't fail. From a
2884 * kmem/slab perspective, the cache has already been selected,
2885 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2888 * This condition will only trigger if the task entered
2889 * memcg_charge_kmem in a sane state, but was OOM-killed during
2890 * __mem_cgroup_try_charge() above. Tasks that were already
2891 * dying when the allocation triggers should have been already
2892 * directed to the root cgroup in memcontrol.h
2894 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2895 if (do_swap_account)
2896 res_counter_charge_nofail(&memcg->memsw, size,
2900 res_counter_uncharge(&memcg->kmem, size);
2905 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2907 res_counter_uncharge(&memcg->res, size);
2908 if (do_swap_account)
2909 res_counter_uncharge(&memcg->memsw, size);
2912 if (res_counter_uncharge(&memcg->kmem, size))
2916 * Releases a reference taken in kmem_cgroup_css_offline in case
2917 * this last uncharge is racing with the offlining code or it is
2918 * outliving the memcg existence.
2920 * The memory barrier imposed by test&clear is paired with the
2921 * explicit one in memcg_kmem_mark_dead().
2923 if (memcg_kmem_test_and_clear_dead(memcg))
2924 css_put(&memcg->css);
2927 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2932 mutex_lock(&memcg->slab_caches_mutex);
2933 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2934 mutex_unlock(&memcg->slab_caches_mutex);
2938 * helper for acessing a memcg's index. It will be used as an index in the
2939 * child cache array in kmem_cache, and also to derive its name. This function
2940 * will return -1 when this is not a kmem-limited memcg.
2942 int memcg_cache_id(struct mem_cgroup *memcg)
2944 return memcg ? memcg->kmemcg_id : -1;
2948 * This ends up being protected by the set_limit mutex, during normal
2949 * operation, because that is its main call site.
2951 * But when we create a new cache, we can call this as well if its parent
2952 * is kmem-limited. That will have to hold set_limit_mutex as well.
2954 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2958 num = ida_simple_get(&kmem_limited_groups,
2959 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2963 * After this point, kmem_accounted (that we test atomically in
2964 * the beginning of this conditional), is no longer 0. This
2965 * guarantees only one process will set the following boolean
2966 * to true. We don't need test_and_set because we're protected
2967 * by the set_limit_mutex anyway.
2969 memcg_kmem_set_activated(memcg);
2971 ret = memcg_update_all_caches(num+1);
2973 ida_simple_remove(&kmem_limited_groups, num);
2974 memcg_kmem_clear_activated(memcg);
2978 memcg->kmemcg_id = num;
2979 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2980 mutex_init(&memcg->slab_caches_mutex);
2984 static size_t memcg_caches_array_size(int num_groups)
2987 if (num_groups <= 0)
2990 size = 2 * num_groups;
2991 if (size < MEMCG_CACHES_MIN_SIZE)
2992 size = MEMCG_CACHES_MIN_SIZE;
2993 else if (size > MEMCG_CACHES_MAX_SIZE)
2994 size = MEMCG_CACHES_MAX_SIZE;
3000 * We should update the current array size iff all caches updates succeed. This
3001 * can only be done from the slab side. The slab mutex needs to be held when
3004 void memcg_update_array_size(int num)
3006 if (num > memcg_limited_groups_array_size)
3007 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3010 static void kmem_cache_destroy_work_func(struct work_struct *w);
3012 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3014 struct memcg_cache_params *cur_params = s->memcg_params;
3016 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3018 if (num_groups > memcg_limited_groups_array_size) {
3020 ssize_t size = memcg_caches_array_size(num_groups);
3022 size *= sizeof(void *);
3023 size += offsetof(struct memcg_cache_params, memcg_caches);
3025 s->memcg_params = kzalloc(size, GFP_KERNEL);
3026 if (!s->memcg_params) {
3027 s->memcg_params = cur_params;
3031 s->memcg_params->is_root_cache = true;
3034 * There is the chance it will be bigger than
3035 * memcg_limited_groups_array_size, if we failed an allocation
3036 * in a cache, in which case all caches updated before it, will
3037 * have a bigger array.
3039 * But if that is the case, the data after
3040 * memcg_limited_groups_array_size is certainly unused
3042 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3043 if (!cur_params->memcg_caches[i])
3045 s->memcg_params->memcg_caches[i] =
3046 cur_params->memcg_caches[i];
3050 * Ideally, we would wait until all caches succeed, and only
3051 * then free the old one. But this is not worth the extra
3052 * pointer per-cache we'd have to have for this.
3054 * It is not a big deal if some caches are left with a size
3055 * bigger than the others. And all updates will reset this
3063 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3064 struct kmem_cache *root_cache)
3068 if (!memcg_kmem_enabled())
3072 size = offsetof(struct memcg_cache_params, memcg_caches);
3073 size += memcg_limited_groups_array_size * sizeof(void *);
3075 size = sizeof(struct memcg_cache_params);
3077 s->memcg_params = kzalloc(size, GFP_KERNEL);
3078 if (!s->memcg_params)
3082 s->memcg_params->memcg = memcg;
3083 s->memcg_params->root_cache = root_cache;
3084 INIT_WORK(&s->memcg_params->destroy,
3085 kmem_cache_destroy_work_func);
3087 s->memcg_params->is_root_cache = true;
3092 void memcg_release_cache(struct kmem_cache *s)
3094 struct kmem_cache *root;
3095 struct mem_cgroup *memcg;
3099 * This happens, for instance, when a root cache goes away before we
3102 if (!s->memcg_params)
3105 if (s->memcg_params->is_root_cache)
3108 memcg = s->memcg_params->memcg;
3109 id = memcg_cache_id(memcg);
3111 root = s->memcg_params->root_cache;
3112 root->memcg_params->memcg_caches[id] = NULL;
3114 mutex_lock(&memcg->slab_caches_mutex);
3115 list_del(&s->memcg_params->list);
3116 mutex_unlock(&memcg->slab_caches_mutex);
3118 css_put(&memcg->css);
3120 kfree(s->memcg_params);
3124 * During the creation a new cache, we need to disable our accounting mechanism
3125 * altogether. This is true even if we are not creating, but rather just
3126 * enqueing new caches to be created.
3128 * This is because that process will trigger allocations; some visible, like
3129 * explicit kmallocs to auxiliary data structures, name strings and internal
3130 * cache structures; some well concealed, like INIT_WORK() that can allocate
3131 * objects during debug.
3133 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3134 * to it. This may not be a bounded recursion: since the first cache creation
3135 * failed to complete (waiting on the allocation), we'll just try to create the
3136 * cache again, failing at the same point.
3138 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3139 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3140 * inside the following two functions.
3142 static inline void memcg_stop_kmem_account(void)
3144 VM_BUG_ON(!current->mm);
3145 current->memcg_kmem_skip_account++;
3148 static inline void memcg_resume_kmem_account(void)
3150 VM_BUG_ON(!current->mm);
3151 current->memcg_kmem_skip_account--;
3154 static void kmem_cache_destroy_work_func(struct work_struct *w)
3156 struct kmem_cache *cachep;
3157 struct memcg_cache_params *p;
3159 p = container_of(w, struct memcg_cache_params, destroy);
3161 cachep = memcg_params_to_cache(p);
3164 * If we get down to 0 after shrink, we could delete right away.
3165 * However, memcg_release_pages() already puts us back in the workqueue
3166 * in that case. If we proceed deleting, we'll get a dangling
3167 * reference, and removing the object from the workqueue in that case
3168 * is unnecessary complication. We are not a fast path.
3170 * Note that this case is fundamentally different from racing with
3171 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3172 * kmem_cache_shrink, not only we would be reinserting a dead cache
3173 * into the queue, but doing so from inside the worker racing to
3176 * So if we aren't down to zero, we'll just schedule a worker and try
3179 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3180 kmem_cache_shrink(cachep);
3181 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3184 kmem_cache_destroy(cachep);
3187 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3189 if (!cachep->memcg_params->dead)
3193 * There are many ways in which we can get here.
3195 * We can get to a memory-pressure situation while the delayed work is
3196 * still pending to run. The vmscan shrinkers can then release all
3197 * cache memory and get us to destruction. If this is the case, we'll
3198 * be executed twice, which is a bug (the second time will execute over
3199 * bogus data). In this case, cancelling the work should be fine.
3201 * But we can also get here from the worker itself, if
3202 * kmem_cache_shrink is enough to shake all the remaining objects and
3203 * get the page count to 0. In this case, we'll deadlock if we try to
3204 * cancel the work (the worker runs with an internal lock held, which
3205 * is the same lock we would hold for cancel_work_sync().)
3207 * Since we can't possibly know who got us here, just refrain from
3208 * running if there is already work pending
3210 if (work_pending(&cachep->memcg_params->destroy))
3213 * We have to defer the actual destroying to a workqueue, because
3214 * we might currently be in a context that cannot sleep.
3216 schedule_work(&cachep->memcg_params->destroy);
3220 * This lock protects updaters, not readers. We want readers to be as fast as
3221 * they can, and they will either see NULL or a valid cache value. Our model
3222 * allow them to see NULL, in which case the root memcg will be selected.
3224 * We need this lock because multiple allocations to the same cache from a non
3225 * will span more than one worker. Only one of them can create the cache.
3227 static DEFINE_MUTEX(memcg_cache_mutex);
3230 * Called with memcg_cache_mutex held
3232 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3233 struct kmem_cache *s)
3235 struct kmem_cache *new;
3236 static char *tmp_name = NULL;
3238 lockdep_assert_held(&memcg_cache_mutex);
3241 * kmem_cache_create_memcg duplicates the given name and
3242 * cgroup_name for this name requires RCU context.
3243 * This static temporary buffer is used to prevent from
3244 * pointless shortliving allocation.
3247 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3253 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3254 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3257 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3258 (s->flags & ~SLAB_PANIC), s->ctor, s);
3261 new->allocflags |= __GFP_KMEMCG;
3266 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3267 struct kmem_cache *cachep)
3269 struct kmem_cache *new_cachep;
3272 BUG_ON(!memcg_can_account_kmem(memcg));
3274 idx = memcg_cache_id(memcg);
3276 mutex_lock(&memcg_cache_mutex);
3277 new_cachep = cachep->memcg_params->memcg_caches[idx];
3279 css_put(&memcg->css);
3283 new_cachep = kmem_cache_dup(memcg, cachep);
3284 if (new_cachep == NULL) {
3285 new_cachep = cachep;
3286 css_put(&memcg->css);
3290 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3292 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3294 * the readers won't lock, make sure everybody sees the updated value,
3295 * so they won't put stuff in the queue again for no reason
3299 mutex_unlock(&memcg_cache_mutex);
3303 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3305 struct kmem_cache *c;
3308 if (!s->memcg_params)
3310 if (!s->memcg_params->is_root_cache)
3314 * If the cache is being destroyed, we trust that there is no one else
3315 * requesting objects from it. Even if there are, the sanity checks in
3316 * kmem_cache_destroy should caught this ill-case.
3318 * Still, we don't want anyone else freeing memcg_caches under our
3319 * noses, which can happen if a new memcg comes to life. As usual,
3320 * we'll take the set_limit_mutex to protect ourselves against this.
3322 mutex_lock(&set_limit_mutex);
3323 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3324 c = s->memcg_params->memcg_caches[i];
3329 * We will now manually delete the caches, so to avoid races
3330 * we need to cancel all pending destruction workers and
3331 * proceed with destruction ourselves.
3333 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3334 * and that could spawn the workers again: it is likely that
3335 * the cache still have active pages until this very moment.
3336 * This would lead us back to mem_cgroup_destroy_cache.
3338 * But that will not execute at all if the "dead" flag is not
3339 * set, so flip it down to guarantee we are in control.
3341 c->memcg_params->dead = false;
3342 cancel_work_sync(&c->memcg_params->destroy);
3343 kmem_cache_destroy(c);
3345 mutex_unlock(&set_limit_mutex);
3348 struct create_work {
3349 struct mem_cgroup *memcg;
3350 struct kmem_cache *cachep;
3351 struct work_struct work;
3354 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3356 struct kmem_cache *cachep;
3357 struct memcg_cache_params *params;
3359 if (!memcg_kmem_is_active(memcg))
3362 mutex_lock(&memcg->slab_caches_mutex);
3363 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3364 cachep = memcg_params_to_cache(params);
3365 cachep->memcg_params->dead = true;
3366 schedule_work(&cachep->memcg_params->destroy);
3368 mutex_unlock(&memcg->slab_caches_mutex);
3371 static void memcg_create_cache_work_func(struct work_struct *w)
3373 struct create_work *cw;
3375 cw = container_of(w, struct create_work, work);
3376 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3381 * Enqueue the creation of a per-memcg kmem_cache.
3383 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3384 struct kmem_cache *cachep)
3386 struct create_work *cw;
3388 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3390 css_put(&memcg->css);
3395 cw->cachep = cachep;
3397 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3398 schedule_work(&cw->work);
3401 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3402 struct kmem_cache *cachep)
3405 * We need to stop accounting when we kmalloc, because if the
3406 * corresponding kmalloc cache is not yet created, the first allocation
3407 * in __memcg_create_cache_enqueue will recurse.
3409 * However, it is better to enclose the whole function. Depending on
3410 * the debugging options enabled, INIT_WORK(), for instance, can
3411 * trigger an allocation. This too, will make us recurse. Because at
3412 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3413 * the safest choice is to do it like this, wrapping the whole function.
3415 memcg_stop_kmem_account();
3416 __memcg_create_cache_enqueue(memcg, cachep);
3417 memcg_resume_kmem_account();
3420 * Return the kmem_cache we're supposed to use for a slab allocation.
3421 * We try to use the current memcg's version of the cache.
3423 * If the cache does not exist yet, if we are the first user of it,
3424 * we either create it immediately, if possible, or create it asynchronously
3426 * In the latter case, we will let the current allocation go through with
3427 * the original cache.
3429 * Can't be called in interrupt context or from kernel threads.
3430 * This function needs to be called with rcu_read_lock() held.
3432 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3435 struct mem_cgroup *memcg;
3438 VM_BUG_ON(!cachep->memcg_params);
3439 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3441 if (!current->mm || current->memcg_kmem_skip_account)
3445 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3447 if (!memcg_can_account_kmem(memcg))
3450 idx = memcg_cache_id(memcg);
3453 * barrier to mare sure we're always seeing the up to date value. The
3454 * code updating memcg_caches will issue a write barrier to match this.
3456 read_barrier_depends();
3457 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3458 cachep = cachep->memcg_params->memcg_caches[idx];
3462 /* The corresponding put will be done in the workqueue. */
3463 if (!css_tryget(&memcg->css))
3468 * If we are in a safe context (can wait, and not in interrupt
3469 * context), we could be be predictable and return right away.
3470 * This would guarantee that the allocation being performed
3471 * already belongs in the new cache.
3473 * However, there are some clashes that can arrive from locking.
3474 * For instance, because we acquire the slab_mutex while doing
3475 * kmem_cache_dup, this means no further allocation could happen
3476 * with the slab_mutex held.
3478 * Also, because cache creation issue get_online_cpus(), this
3479 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3480 * that ends up reversed during cpu hotplug. (cpuset allocates
3481 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3482 * better to defer everything.
3484 memcg_create_cache_enqueue(memcg, cachep);
3490 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3493 * We need to verify if the allocation against current->mm->owner's memcg is
3494 * possible for the given order. But the page is not allocated yet, so we'll
3495 * need a further commit step to do the final arrangements.
3497 * It is possible for the task to switch cgroups in this mean time, so at
3498 * commit time, we can't rely on task conversion any longer. We'll then use
3499 * the handle argument to return to the caller which cgroup we should commit
3500 * against. We could also return the memcg directly and avoid the pointer
3501 * passing, but a boolean return value gives better semantics considering
3502 * the compiled-out case as well.
3504 * Returning true means the allocation is possible.
3507 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3509 struct mem_cgroup *memcg;
3515 * Disabling accounting is only relevant for some specific memcg
3516 * internal allocations. Therefore we would initially not have such
3517 * check here, since direct calls to the page allocator that are marked
3518 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3519 * concerned with cache allocations, and by having this test at
3520 * memcg_kmem_get_cache, we are already able to relay the allocation to
3521 * the root cache and bypass the memcg cache altogether.
3523 * There is one exception, though: the SLUB allocator does not create
3524 * large order caches, but rather service large kmallocs directly from
3525 * the page allocator. Therefore, the following sequence when backed by
3526 * the SLUB allocator:
3528 * memcg_stop_kmem_account();
3529 * kmalloc(<large_number>)
3530 * memcg_resume_kmem_account();
3532 * would effectively ignore the fact that we should skip accounting,
3533 * since it will drive us directly to this function without passing
3534 * through the cache selector memcg_kmem_get_cache. Such large
3535 * allocations are extremely rare but can happen, for instance, for the
3536 * cache arrays. We bring this test here.
3538 if (!current->mm || current->memcg_kmem_skip_account)
3541 memcg = try_get_mem_cgroup_from_mm(current->mm);
3544 * very rare case described in mem_cgroup_from_task. Unfortunately there
3545 * isn't much we can do without complicating this too much, and it would
3546 * be gfp-dependent anyway. Just let it go
3548 if (unlikely(!memcg))
3551 if (!memcg_can_account_kmem(memcg)) {
3552 css_put(&memcg->css);
3556 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3560 css_put(&memcg->css);
3564 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3567 struct page_cgroup *pc;
3569 VM_BUG_ON(mem_cgroup_is_root(memcg));
3571 /* The page allocation failed. Revert */
3573 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3577 pc = lookup_page_cgroup(page);
3578 lock_page_cgroup(pc);
3579 pc->mem_cgroup = memcg;
3580 SetPageCgroupUsed(pc);
3581 unlock_page_cgroup(pc);
3584 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3586 struct mem_cgroup *memcg = NULL;
3587 struct page_cgroup *pc;
3590 pc = lookup_page_cgroup(page);
3592 * Fast unlocked return. Theoretically might have changed, have to
3593 * check again after locking.
3595 if (!PageCgroupUsed(pc))
3598 lock_page_cgroup(pc);
3599 if (PageCgroupUsed(pc)) {
3600 memcg = pc->mem_cgroup;
3601 ClearPageCgroupUsed(pc);
3603 unlock_page_cgroup(pc);
3606 * We trust that only if there is a memcg associated with the page, it
3607 * is a valid allocation
3612 VM_BUG_ON(mem_cgroup_is_root(memcg));
3613 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3616 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3619 #endif /* CONFIG_MEMCG_KMEM */
3621 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3623 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3625 * Because tail pages are not marked as "used", set it. We're under
3626 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3627 * charge/uncharge will be never happen and move_account() is done under
3628 * compound_lock(), so we don't have to take care of races.
3630 void mem_cgroup_split_huge_fixup(struct page *head)
3632 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3633 struct page_cgroup *pc;
3634 struct mem_cgroup *memcg;
3637 if (mem_cgroup_disabled())
3640 memcg = head_pc->mem_cgroup;
3641 for (i = 1; i < HPAGE_PMD_NR; i++) {
3643 pc->mem_cgroup = memcg;
3644 smp_wmb();/* see __commit_charge() */
3645 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3647 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3650 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3653 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3654 struct mem_cgroup *to,
3655 unsigned int nr_pages,
3656 enum mem_cgroup_stat_index idx)
3658 /* Update stat data for mem_cgroup */
3660 WARN_ON_ONCE(from->stat->count[idx] < nr_pages);
3661 __this_cpu_add(from->stat->count[idx], -nr_pages);
3662 __this_cpu_add(to->stat->count[idx], nr_pages);
3667 * mem_cgroup_move_account - move account of the page
3669 * @nr_pages: number of regular pages (>1 for huge pages)
3670 * @pc: page_cgroup of the page.
3671 * @from: mem_cgroup which the page is moved from.
3672 * @to: mem_cgroup which the page is moved to. @from != @to.
3674 * The caller must confirm following.
3675 * - page is not on LRU (isolate_page() is useful.)
3676 * - compound_lock is held when nr_pages > 1
3678 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3681 static int mem_cgroup_move_account(struct page *page,
3682 unsigned int nr_pages,
3683 struct page_cgroup *pc,
3684 struct mem_cgroup *from,
3685 struct mem_cgroup *to)
3687 unsigned long flags;
3689 bool anon = PageAnon(page);
3691 VM_BUG_ON(from == to);
3692 VM_BUG_ON(PageLRU(page));
3694 * The page is isolated from LRU. So, collapse function
3695 * will not handle this page. But page splitting can happen.
3696 * Do this check under compound_page_lock(). The caller should
3700 if (nr_pages > 1 && !PageTransHuge(page))
3703 lock_page_cgroup(pc);
3706 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3709 move_lock_mem_cgroup(from, &flags);
3711 if (!anon && page_mapped(page))
3712 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3713 MEM_CGROUP_STAT_FILE_MAPPED);
3715 if (PageWriteback(page))
3716 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3717 MEM_CGROUP_STAT_WRITEBACK);
3719 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3721 /* caller should have done css_get */
3722 pc->mem_cgroup = to;
3723 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3724 move_unlock_mem_cgroup(from, &flags);
3727 unlock_page_cgroup(pc);
3731 memcg_check_events(to, page);
3732 memcg_check_events(from, page);
3738 * mem_cgroup_move_parent - moves page to the parent group
3739 * @page: the page to move
3740 * @pc: page_cgroup of the page
3741 * @child: page's cgroup
3743 * move charges to its parent or the root cgroup if the group has no
3744 * parent (aka use_hierarchy==0).
3745 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3746 * mem_cgroup_move_account fails) the failure is always temporary and
3747 * it signals a race with a page removal/uncharge or migration. In the
3748 * first case the page is on the way out and it will vanish from the LRU
3749 * on the next attempt and the call should be retried later.
3750 * Isolation from the LRU fails only if page has been isolated from
3751 * the LRU since we looked at it and that usually means either global
3752 * reclaim or migration going on. The page will either get back to the
3754 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3755 * (!PageCgroupUsed) or moved to a different group. The page will
3756 * disappear in the next attempt.
3758 static int mem_cgroup_move_parent(struct page *page,
3759 struct page_cgroup *pc,
3760 struct mem_cgroup *child)
3762 struct mem_cgroup *parent;
3763 unsigned int nr_pages;
3764 unsigned long uninitialized_var(flags);
3767 VM_BUG_ON(mem_cgroup_is_root(child));
3770 if (!get_page_unless_zero(page))
3772 if (isolate_lru_page(page))
3775 nr_pages = hpage_nr_pages(page);
3777 parent = parent_mem_cgroup(child);
3779 * If no parent, move charges to root cgroup.
3782 parent = root_mem_cgroup;
3785 VM_BUG_ON(!PageTransHuge(page));
3786 flags = compound_lock_irqsave(page);
3789 ret = mem_cgroup_move_account(page, nr_pages,
3792 __mem_cgroup_cancel_local_charge(child, nr_pages);
3795 compound_unlock_irqrestore(page, flags);
3796 putback_lru_page(page);
3804 * Charge the memory controller for page usage.
3806 * 0 if the charge was successful
3807 * < 0 if the cgroup is over its limit
3809 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3810 gfp_t gfp_mask, enum charge_type ctype)
3812 struct mem_cgroup *memcg = NULL;
3813 unsigned int nr_pages = 1;
3817 if (PageTransHuge(page)) {
3818 nr_pages <<= compound_order(page);
3819 VM_BUG_ON(!PageTransHuge(page));
3821 * Never OOM-kill a process for a huge page. The
3822 * fault handler will fall back to regular pages.
3827 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3830 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3834 int mem_cgroup_newpage_charge(struct page *page,
3835 struct mm_struct *mm, gfp_t gfp_mask)
3837 if (mem_cgroup_disabled())
3839 VM_BUG_ON(page_mapped(page));
3840 VM_BUG_ON(page->mapping && !PageAnon(page));
3842 return mem_cgroup_charge_common(page, mm, gfp_mask,
3843 MEM_CGROUP_CHARGE_TYPE_ANON);
3847 * While swap-in, try_charge -> commit or cancel, the page is locked.
3848 * And when try_charge() successfully returns, one refcnt to memcg without
3849 * struct page_cgroup is acquired. This refcnt will be consumed by
3850 * "commit()" or removed by "cancel()"
3852 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3855 struct mem_cgroup **memcgp)
3857 struct mem_cgroup *memcg;
3858 struct page_cgroup *pc;
3861 pc = lookup_page_cgroup(page);
3863 * Every swap fault against a single page tries to charge the
3864 * page, bail as early as possible. shmem_unuse() encounters
3865 * already charged pages, too. The USED bit is protected by
3866 * the page lock, which serializes swap cache removal, which
3867 * in turn serializes uncharging.
3869 if (PageCgroupUsed(pc))
3871 if (!do_swap_account)
3873 memcg = try_get_mem_cgroup_from_page(page);
3877 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3878 css_put(&memcg->css);
3883 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3889 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3890 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3893 if (mem_cgroup_disabled())
3896 * A racing thread's fault, or swapoff, may have already
3897 * updated the pte, and even removed page from swap cache: in
3898 * those cases unuse_pte()'s pte_same() test will fail; but
3899 * there's also a KSM case which does need to charge the page.
3901 if (!PageSwapCache(page)) {
3904 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3909 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3912 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3914 if (mem_cgroup_disabled())
3918 __mem_cgroup_cancel_charge(memcg, 1);
3922 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3923 enum charge_type ctype)
3925 if (mem_cgroup_disabled())
3930 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3932 * Now swap is on-memory. This means this page may be
3933 * counted both as mem and swap....double count.
3934 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3935 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3936 * may call delete_from_swap_cache() before reach here.
3938 if (do_swap_account && PageSwapCache(page)) {
3939 swp_entry_t ent = {.val = page_private(page)};
3940 mem_cgroup_uncharge_swap(ent);
3944 void mem_cgroup_commit_charge_swapin(struct page *page,
3945 struct mem_cgroup *memcg)
3947 __mem_cgroup_commit_charge_swapin(page, memcg,
3948 MEM_CGROUP_CHARGE_TYPE_ANON);
3951 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3954 struct mem_cgroup *memcg = NULL;
3955 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3958 if (mem_cgroup_disabled())
3960 if (PageCompound(page))
3963 if (!PageSwapCache(page))
3964 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3965 else { /* page is swapcache/shmem */
3966 ret = __mem_cgroup_try_charge_swapin(mm, page,
3969 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3974 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3975 unsigned int nr_pages,
3976 const enum charge_type ctype)
3978 struct memcg_batch_info *batch = NULL;
3979 bool uncharge_memsw = true;
3981 /* If swapout, usage of swap doesn't decrease */
3982 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3983 uncharge_memsw = false;
3985 batch = ¤t->memcg_batch;
3987 * In usual, we do css_get() when we remember memcg pointer.
3988 * But in this case, we keep res->usage until end of a series of
3989 * uncharges. Then, it's ok to ignore memcg's refcnt.
3992 batch->memcg = memcg;
3994 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3995 * In those cases, all pages freed continuously can be expected to be in
3996 * the same cgroup and we have chance to coalesce uncharges.
3997 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3998 * because we want to do uncharge as soon as possible.
4001 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4002 goto direct_uncharge;
4005 goto direct_uncharge;
4008 * In typical case, batch->memcg == mem. This means we can
4009 * merge a series of uncharges to an uncharge of res_counter.
4010 * If not, we uncharge res_counter ony by one.
4012 if (batch->memcg != memcg)
4013 goto direct_uncharge;
4014 /* remember freed charge and uncharge it later */
4017 batch->memsw_nr_pages++;
4020 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4022 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4023 if (unlikely(batch->memcg != memcg))
4024 memcg_oom_recover(memcg);
4028 * uncharge if !page_mapped(page)
4030 static struct mem_cgroup *
4031 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4034 struct mem_cgroup *memcg = NULL;
4035 unsigned int nr_pages = 1;
4036 struct page_cgroup *pc;
4039 if (mem_cgroup_disabled())
4042 if (PageTransHuge(page)) {
4043 nr_pages <<= compound_order(page);
4044 VM_BUG_ON(!PageTransHuge(page));
4047 * Check if our page_cgroup is valid
4049 pc = lookup_page_cgroup(page);
4050 if (unlikely(!PageCgroupUsed(pc)))
4053 lock_page_cgroup(pc);
4055 memcg = pc->mem_cgroup;
4057 if (!PageCgroupUsed(pc))
4060 anon = PageAnon(page);
4063 case MEM_CGROUP_CHARGE_TYPE_ANON:
4065 * Generally PageAnon tells if it's the anon statistics to be
4066 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4067 * used before page reached the stage of being marked PageAnon.
4071 case MEM_CGROUP_CHARGE_TYPE_DROP:
4072 /* See mem_cgroup_prepare_migration() */
4073 if (page_mapped(page))
4076 * Pages under migration may not be uncharged. But
4077 * end_migration() /must/ be the one uncharging the
4078 * unused post-migration page and so it has to call
4079 * here with the migration bit still set. See the
4080 * res_counter handling below.
4082 if (!end_migration && PageCgroupMigration(pc))
4085 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4086 if (!PageAnon(page)) { /* Shared memory */
4087 if (page->mapping && !page_is_file_cache(page))
4089 } else if (page_mapped(page)) /* Anon */
4096 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4098 ClearPageCgroupUsed(pc);
4100 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4101 * freed from LRU. This is safe because uncharged page is expected not
4102 * to be reused (freed soon). Exception is SwapCache, it's handled by
4103 * special functions.
4106 unlock_page_cgroup(pc);
4108 * even after unlock, we have memcg->res.usage here and this memcg
4109 * will never be freed, so it's safe to call css_get().
4111 memcg_check_events(memcg, page);
4112 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4113 mem_cgroup_swap_statistics(memcg, true);
4114 css_get(&memcg->css);
4117 * Migration does not charge the res_counter for the
4118 * replacement page, so leave it alone when phasing out the
4119 * page that is unused after the migration.
4121 if (!end_migration && !mem_cgroup_is_root(memcg))
4122 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4127 unlock_page_cgroup(pc);
4131 void mem_cgroup_uncharge_page(struct page *page)
4134 if (page_mapped(page))
4136 VM_BUG_ON(page->mapping && !PageAnon(page));
4138 * If the page is in swap cache, uncharge should be deferred
4139 * to the swap path, which also properly accounts swap usage
4140 * and handles memcg lifetime.
4142 * Note that this check is not stable and reclaim may add the
4143 * page to swap cache at any time after this. However, if the
4144 * page is not in swap cache by the time page->mapcount hits
4145 * 0, there won't be any page table references to the swap
4146 * slot, and reclaim will free it and not actually write the
4149 if (PageSwapCache(page))
4151 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4154 void mem_cgroup_uncharge_cache_page(struct page *page)
4156 VM_BUG_ON(page_mapped(page));
4157 VM_BUG_ON(page->mapping);
4158 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4162 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4163 * In that cases, pages are freed continuously and we can expect pages
4164 * are in the same memcg. All these calls itself limits the number of
4165 * pages freed at once, then uncharge_start/end() is called properly.
4166 * This may be called prural(2) times in a context,
4169 void mem_cgroup_uncharge_start(void)
4171 current->memcg_batch.do_batch++;
4172 /* We can do nest. */
4173 if (current->memcg_batch.do_batch == 1) {
4174 current->memcg_batch.memcg = NULL;
4175 current->memcg_batch.nr_pages = 0;
4176 current->memcg_batch.memsw_nr_pages = 0;
4180 void mem_cgroup_uncharge_end(void)
4182 struct memcg_batch_info *batch = ¤t->memcg_batch;
4184 if (!batch->do_batch)
4188 if (batch->do_batch) /* If stacked, do nothing. */
4194 * This "batch->memcg" is valid without any css_get/put etc...
4195 * bacause we hide charges behind us.
4197 if (batch->nr_pages)
4198 res_counter_uncharge(&batch->memcg->res,
4199 batch->nr_pages * PAGE_SIZE);
4200 if (batch->memsw_nr_pages)
4201 res_counter_uncharge(&batch->memcg->memsw,
4202 batch->memsw_nr_pages * PAGE_SIZE);
4203 memcg_oom_recover(batch->memcg);
4204 /* forget this pointer (for sanity check) */
4205 batch->memcg = NULL;
4210 * called after __delete_from_swap_cache() and drop "page" account.
4211 * memcg information is recorded to swap_cgroup of "ent"
4214 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4216 struct mem_cgroup *memcg;
4217 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4219 if (!swapout) /* this was a swap cache but the swap is unused ! */
4220 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4222 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4225 * record memcg information, if swapout && memcg != NULL,
4226 * css_get() was called in uncharge().
4228 if (do_swap_account && swapout && memcg)
4229 swap_cgroup_record(ent, css_id(&memcg->css));
4233 #ifdef CONFIG_MEMCG_SWAP
4235 * called from swap_entry_free(). remove record in swap_cgroup and
4236 * uncharge "memsw" account.
4238 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4240 struct mem_cgroup *memcg;
4243 if (!do_swap_account)
4246 id = swap_cgroup_record(ent, 0);
4248 memcg = mem_cgroup_lookup(id);
4251 * We uncharge this because swap is freed.
4252 * This memcg can be obsolete one. We avoid calling css_tryget
4254 if (!mem_cgroup_is_root(memcg))
4255 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4256 mem_cgroup_swap_statistics(memcg, false);
4257 css_put(&memcg->css);
4263 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4264 * @entry: swap entry to be moved
4265 * @from: mem_cgroup which the entry is moved from
4266 * @to: mem_cgroup which the entry is moved to
4268 * It succeeds only when the swap_cgroup's record for this entry is the same
4269 * as the mem_cgroup's id of @from.
4271 * Returns 0 on success, -EINVAL on failure.
4273 * The caller must have charged to @to, IOW, called res_counter_charge() about
4274 * both res and memsw, and called css_get().
4276 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4277 struct mem_cgroup *from, struct mem_cgroup *to)
4279 unsigned short old_id, new_id;
4281 old_id = css_id(&from->css);
4282 new_id = css_id(&to->css);
4284 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4285 mem_cgroup_swap_statistics(from, false);
4286 mem_cgroup_swap_statistics(to, true);
4288 * This function is only called from task migration context now.
4289 * It postpones res_counter and refcount handling till the end
4290 * of task migration(mem_cgroup_clear_mc()) for performance
4291 * improvement. But we cannot postpone css_get(to) because if
4292 * the process that has been moved to @to does swap-in, the
4293 * refcount of @to might be decreased to 0.
4295 * We are in attach() phase, so the cgroup is guaranteed to be
4296 * alive, so we can just call css_get().
4304 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4305 struct mem_cgroup *from, struct mem_cgroup *to)
4312 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4315 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4316 struct mem_cgroup **memcgp)
4318 struct mem_cgroup *memcg = NULL;
4319 unsigned int nr_pages = 1;
4320 struct page_cgroup *pc;
4321 enum charge_type ctype;
4325 if (mem_cgroup_disabled())
4328 if (PageTransHuge(page))
4329 nr_pages <<= compound_order(page);
4331 pc = lookup_page_cgroup(page);
4332 lock_page_cgroup(pc);
4333 if (PageCgroupUsed(pc)) {
4334 memcg = pc->mem_cgroup;
4335 css_get(&memcg->css);
4337 * At migrating an anonymous page, its mapcount goes down
4338 * to 0 and uncharge() will be called. But, even if it's fully
4339 * unmapped, migration may fail and this page has to be
4340 * charged again. We set MIGRATION flag here and delay uncharge
4341 * until end_migration() is called
4343 * Corner Case Thinking
4345 * When the old page was mapped as Anon and it's unmap-and-freed
4346 * while migration was ongoing.
4347 * If unmap finds the old page, uncharge() of it will be delayed
4348 * until end_migration(). If unmap finds a new page, it's
4349 * uncharged when it make mapcount to be 1->0. If unmap code
4350 * finds swap_migration_entry, the new page will not be mapped
4351 * and end_migration() will find it(mapcount==0).
4354 * When the old page was mapped but migraion fails, the kernel
4355 * remaps it. A charge for it is kept by MIGRATION flag even
4356 * if mapcount goes down to 0. We can do remap successfully
4357 * without charging it again.
4360 * The "old" page is under lock_page() until the end of
4361 * migration, so, the old page itself will not be swapped-out.
4362 * If the new page is swapped out before end_migraton, our
4363 * hook to usual swap-out path will catch the event.
4366 SetPageCgroupMigration(pc);
4368 unlock_page_cgroup(pc);
4370 * If the page is not charged at this point,
4378 * We charge new page before it's used/mapped. So, even if unlock_page()
4379 * is called before end_migration, we can catch all events on this new
4380 * page. In the case new page is migrated but not remapped, new page's
4381 * mapcount will be finally 0 and we call uncharge in end_migration().
4384 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4386 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4388 * The page is committed to the memcg, but it's not actually
4389 * charged to the res_counter since we plan on replacing the
4390 * old one and only one page is going to be left afterwards.
4392 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4395 /* remove redundant charge if migration failed*/
4396 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4397 struct page *oldpage, struct page *newpage, bool migration_ok)
4399 struct page *used, *unused;
4400 struct page_cgroup *pc;
4406 if (!migration_ok) {
4413 anon = PageAnon(used);
4414 __mem_cgroup_uncharge_common(unused,
4415 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4416 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4418 css_put(&memcg->css);
4420 * We disallowed uncharge of pages under migration because mapcount
4421 * of the page goes down to zero, temporarly.
4422 * Clear the flag and check the page should be charged.
4424 pc = lookup_page_cgroup(oldpage);
4425 lock_page_cgroup(pc);
4426 ClearPageCgroupMigration(pc);
4427 unlock_page_cgroup(pc);
4430 * If a page is a file cache, radix-tree replacement is very atomic
4431 * and we can skip this check. When it was an Anon page, its mapcount
4432 * goes down to 0. But because we added MIGRATION flage, it's not
4433 * uncharged yet. There are several case but page->mapcount check
4434 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4435 * check. (see prepare_charge() also)
4438 mem_cgroup_uncharge_page(used);
4442 * At replace page cache, newpage is not under any memcg but it's on
4443 * LRU. So, this function doesn't touch res_counter but handles LRU
4444 * in correct way. Both pages are locked so we cannot race with uncharge.
4446 void mem_cgroup_replace_page_cache(struct page *oldpage,
4447 struct page *newpage)
4449 struct mem_cgroup *memcg = NULL;
4450 struct page_cgroup *pc;
4451 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4453 if (mem_cgroup_disabled())
4456 pc = lookup_page_cgroup(oldpage);
4457 /* fix accounting on old pages */
4458 lock_page_cgroup(pc);
4459 if (PageCgroupUsed(pc)) {
4460 memcg = pc->mem_cgroup;
4461 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4462 ClearPageCgroupUsed(pc);
4464 unlock_page_cgroup(pc);
4467 * When called from shmem_replace_page(), in some cases the
4468 * oldpage has already been charged, and in some cases not.
4473 * Even if newpage->mapping was NULL before starting replacement,
4474 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4475 * LRU while we overwrite pc->mem_cgroup.
4477 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4480 #ifdef CONFIG_DEBUG_VM
4481 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4483 struct page_cgroup *pc;
4485 pc = lookup_page_cgroup(page);
4487 * Can be NULL while feeding pages into the page allocator for
4488 * the first time, i.e. during boot or memory hotplug;
4489 * or when mem_cgroup_disabled().
4491 if (likely(pc) && PageCgroupUsed(pc))
4496 bool mem_cgroup_bad_page_check(struct page *page)
4498 if (mem_cgroup_disabled())
4501 return lookup_page_cgroup_used(page) != NULL;
4504 void mem_cgroup_print_bad_page(struct page *page)
4506 struct page_cgroup *pc;
4508 pc = lookup_page_cgroup_used(page);
4510 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4511 pc, pc->flags, pc->mem_cgroup);
4516 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4517 unsigned long long val)
4520 u64 memswlimit, memlimit;
4522 int children = mem_cgroup_count_children(memcg);
4523 u64 curusage, oldusage;
4527 * For keeping hierarchical_reclaim simple, how long we should retry
4528 * is depends on callers. We set our retry-count to be function
4529 * of # of children which we should visit in this loop.
4531 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4533 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4536 while (retry_count) {
4537 if (signal_pending(current)) {
4542 * Rather than hide all in some function, I do this in
4543 * open coded manner. You see what this really does.
4544 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4546 mutex_lock(&set_limit_mutex);
4547 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4548 if (memswlimit < val) {
4550 mutex_unlock(&set_limit_mutex);
4554 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4558 ret = res_counter_set_limit(&memcg->res, val);
4560 if (memswlimit == val)
4561 memcg->memsw_is_minimum = true;
4563 memcg->memsw_is_minimum = false;
4565 mutex_unlock(&set_limit_mutex);
4570 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4571 MEM_CGROUP_RECLAIM_SHRINK);
4572 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4573 /* Usage is reduced ? */
4574 if (curusage >= oldusage)
4577 oldusage = curusage;
4579 if (!ret && enlarge)
4580 memcg_oom_recover(memcg);
4585 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4586 unsigned long long val)
4589 u64 memlimit, memswlimit, oldusage, curusage;
4590 int children = mem_cgroup_count_children(memcg);
4594 /* see mem_cgroup_resize_res_limit */
4595 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4596 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4597 while (retry_count) {
4598 if (signal_pending(current)) {
4603 * Rather than hide all in some function, I do this in
4604 * open coded manner. You see what this really does.
4605 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4607 mutex_lock(&set_limit_mutex);
4608 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4609 if (memlimit > val) {
4611 mutex_unlock(&set_limit_mutex);
4614 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4615 if (memswlimit < val)
4617 ret = res_counter_set_limit(&memcg->memsw, val);
4619 if (memlimit == val)
4620 memcg->memsw_is_minimum = true;
4622 memcg->memsw_is_minimum = false;
4624 mutex_unlock(&set_limit_mutex);
4629 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4630 MEM_CGROUP_RECLAIM_NOSWAP |
4631 MEM_CGROUP_RECLAIM_SHRINK);
4632 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4633 /* Usage is reduced ? */
4634 if (curusage >= oldusage)
4637 oldusage = curusage;
4639 if (!ret && enlarge)
4640 memcg_oom_recover(memcg);
4645 * mem_cgroup_force_empty_list - clears LRU of a group
4646 * @memcg: group to clear
4649 * @lru: lru to to clear
4651 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4652 * reclaim the pages page themselves - pages are moved to the parent (or root)
4655 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4656 int node, int zid, enum lru_list lru)
4658 struct lruvec *lruvec;
4659 unsigned long flags;
4660 struct list_head *list;
4664 zone = &NODE_DATA(node)->node_zones[zid];
4665 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4666 list = &lruvec->lists[lru];
4670 struct page_cgroup *pc;
4673 spin_lock_irqsave(&zone->lru_lock, flags);
4674 if (list_empty(list)) {
4675 spin_unlock_irqrestore(&zone->lru_lock, flags);
4678 page = list_entry(list->prev, struct page, lru);
4680 list_move(&page->lru, list);
4682 spin_unlock_irqrestore(&zone->lru_lock, flags);
4685 spin_unlock_irqrestore(&zone->lru_lock, flags);
4687 pc = lookup_page_cgroup(page);
4689 if (mem_cgroup_move_parent(page, pc, memcg)) {
4690 /* found lock contention or "pc" is obsolete. */
4695 } while (!list_empty(list));
4699 * make mem_cgroup's charge to be 0 if there is no task by moving
4700 * all the charges and pages to the parent.
4701 * This enables deleting this mem_cgroup.
4703 * Caller is responsible for holding css reference on the memcg.
4705 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4711 /* This is for making all *used* pages to be on LRU. */
4712 lru_add_drain_all();
4713 drain_all_stock_sync(memcg);
4714 mem_cgroup_start_move(memcg);
4715 for_each_node_state(node, N_MEMORY) {
4716 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4719 mem_cgroup_force_empty_list(memcg,
4724 mem_cgroup_end_move(memcg);
4725 memcg_oom_recover(memcg);
4729 * Kernel memory may not necessarily be trackable to a specific
4730 * process. So they are not migrated, and therefore we can't
4731 * expect their value to drop to 0 here.
4732 * Having res filled up with kmem only is enough.
4734 * This is a safety check because mem_cgroup_force_empty_list
4735 * could have raced with mem_cgroup_replace_page_cache callers
4736 * so the lru seemed empty but the page could have been added
4737 * right after the check. RES_USAGE should be safe as we always
4738 * charge before adding to the LRU.
4740 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4741 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4742 } while (usage > 0);
4746 * This mainly exists for tests during the setting of set of use_hierarchy.
4747 * Since this is the very setting we are changing, the current hierarchy value
4750 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4752 struct cgroup_subsys_state *pos;
4754 /* bounce at first found */
4755 css_for_each_child(pos, &memcg->css)
4761 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4762 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4763 * from mem_cgroup_count_children(), in the sense that we don't really care how
4764 * many children we have; we only need to know if we have any. It also counts
4765 * any memcg without hierarchy as infertile.
4767 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4769 return memcg->use_hierarchy && __memcg_has_children(memcg);
4773 * Reclaims as many pages from the given memcg as possible and moves
4774 * the rest to the parent.
4776 * Caller is responsible for holding css reference for memcg.
4778 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4780 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4781 struct cgroup *cgrp = memcg->css.cgroup;
4783 /* returns EBUSY if there is a task or if we come here twice. */
4784 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4787 /* we call try-to-free pages for make this cgroup empty */
4788 lru_add_drain_all();
4789 /* try to free all pages in this cgroup */
4790 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4793 if (signal_pending(current))
4796 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4800 /* maybe some writeback is necessary */
4801 congestion_wait(BLK_RW_ASYNC, HZ/10);
4806 mem_cgroup_reparent_charges(memcg);
4811 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4814 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4816 if (mem_cgroup_is_root(memcg))
4818 return mem_cgroup_force_empty(memcg);
4821 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4824 return mem_cgroup_from_css(css)->use_hierarchy;
4827 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4828 struct cftype *cft, u64 val)
4831 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4832 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4834 mutex_lock(&memcg_create_mutex);
4836 if (memcg->use_hierarchy == val)
4840 * If parent's use_hierarchy is set, we can't make any modifications
4841 * in the child subtrees. If it is unset, then the change can
4842 * occur, provided the current cgroup has no children.
4844 * For the root cgroup, parent_mem is NULL, we allow value to be
4845 * set if there are no children.
4847 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4848 (val == 1 || val == 0)) {
4849 if (!__memcg_has_children(memcg))
4850 memcg->use_hierarchy = val;
4857 mutex_unlock(&memcg_create_mutex);
4863 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4864 enum mem_cgroup_stat_index idx)
4866 struct mem_cgroup *iter;
4869 /* Per-cpu values can be negative, use a signed accumulator */
4870 for_each_mem_cgroup_tree(iter, memcg)
4871 val += mem_cgroup_read_stat(iter, idx);
4873 if (val < 0) /* race ? */
4878 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4882 if (!mem_cgroup_is_root(memcg)) {
4884 return res_counter_read_u64(&memcg->res, RES_USAGE);
4886 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4890 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4891 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4893 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4894 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4897 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4899 return val << PAGE_SHIFT;
4902 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
4903 struct cftype *cft, struct file *file,
4904 char __user *buf, size_t nbytes, loff_t *ppos)
4906 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4912 type = MEMFILE_TYPE(cft->private);
4913 name = MEMFILE_ATTR(cft->private);
4917 if (name == RES_USAGE)
4918 val = mem_cgroup_usage(memcg, false);
4920 val = res_counter_read_u64(&memcg->res, name);
4923 if (name == RES_USAGE)
4924 val = mem_cgroup_usage(memcg, true);
4926 val = res_counter_read_u64(&memcg->memsw, name);
4929 val = res_counter_read_u64(&memcg->kmem, name);
4935 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4936 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4939 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
4942 #ifdef CONFIG_MEMCG_KMEM
4943 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4945 * For simplicity, we won't allow this to be disabled. It also can't
4946 * be changed if the cgroup has children already, or if tasks had
4949 * If tasks join before we set the limit, a person looking at
4950 * kmem.usage_in_bytes will have no way to determine when it took
4951 * place, which makes the value quite meaningless.
4953 * After it first became limited, changes in the value of the limit are
4954 * of course permitted.
4956 mutex_lock(&memcg_create_mutex);
4957 mutex_lock(&set_limit_mutex);
4958 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
4959 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
4963 ret = res_counter_set_limit(&memcg->kmem, val);
4966 ret = memcg_update_cache_sizes(memcg);
4968 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
4971 static_key_slow_inc(&memcg_kmem_enabled_key);
4973 * setting the active bit after the inc will guarantee no one
4974 * starts accounting before all call sites are patched
4976 memcg_kmem_set_active(memcg);
4978 ret = res_counter_set_limit(&memcg->kmem, val);
4980 mutex_unlock(&set_limit_mutex);
4981 mutex_unlock(&memcg_create_mutex);
4986 #ifdef CONFIG_MEMCG_KMEM
4987 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4990 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4994 memcg->kmem_account_flags = parent->kmem_account_flags;
4996 * When that happen, we need to disable the static branch only on those
4997 * memcgs that enabled it. To achieve this, we would be forced to
4998 * complicate the code by keeping track of which memcgs were the ones
4999 * that actually enabled limits, and which ones got it from its
5002 * It is a lot simpler just to do static_key_slow_inc() on every child
5003 * that is accounted.
5005 if (!memcg_kmem_is_active(memcg))
5009 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5010 * memcg is active already. If the later initialization fails then the
5011 * cgroup core triggers the cleanup so we do not have to do it here.
5013 static_key_slow_inc(&memcg_kmem_enabled_key);
5015 mutex_lock(&set_limit_mutex);
5016 memcg_stop_kmem_account();
5017 ret = memcg_update_cache_sizes(memcg);
5018 memcg_resume_kmem_account();
5019 mutex_unlock(&set_limit_mutex);
5023 #endif /* CONFIG_MEMCG_KMEM */
5026 * The user of this function is...
5029 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5032 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5035 unsigned long long val;
5038 type = MEMFILE_TYPE(cft->private);
5039 name = MEMFILE_ATTR(cft->private);
5043 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5047 /* This function does all necessary parse...reuse it */
5048 ret = res_counter_memparse_write_strategy(buffer, &val);
5052 ret = mem_cgroup_resize_limit(memcg, val);
5053 else if (type == _MEMSWAP)
5054 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5055 else if (type == _KMEM)
5056 ret = memcg_update_kmem_limit(css, val);
5060 case RES_SOFT_LIMIT:
5061 ret = res_counter_memparse_write_strategy(buffer, &val);
5065 * For memsw, soft limits are hard to implement in terms
5066 * of semantics, for now, we support soft limits for
5067 * control without swap
5070 ret = res_counter_set_soft_limit(&memcg->res, val);
5075 ret = -EINVAL; /* should be BUG() ? */
5081 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5082 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5084 unsigned long long min_limit, min_memsw_limit, tmp;
5086 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5087 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5088 if (!memcg->use_hierarchy)
5091 while (css_parent(&memcg->css)) {
5092 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5093 if (!memcg->use_hierarchy)
5095 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5096 min_limit = min(min_limit, tmp);
5097 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5098 min_memsw_limit = min(min_memsw_limit, tmp);
5101 *mem_limit = min_limit;
5102 *memsw_limit = min_memsw_limit;
5105 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5107 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5111 type = MEMFILE_TYPE(event);
5112 name = MEMFILE_ATTR(event);
5117 res_counter_reset_max(&memcg->res);
5118 else if (type == _MEMSWAP)
5119 res_counter_reset_max(&memcg->memsw);
5120 else if (type == _KMEM)
5121 res_counter_reset_max(&memcg->kmem);
5127 res_counter_reset_failcnt(&memcg->res);
5128 else if (type == _MEMSWAP)
5129 res_counter_reset_failcnt(&memcg->memsw);
5130 else if (type == _KMEM)
5131 res_counter_reset_failcnt(&memcg->kmem);
5140 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5143 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5147 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5148 struct cftype *cft, u64 val)
5150 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5152 if (val >= (1 << NR_MOVE_TYPE))
5156 * No kind of locking is needed in here, because ->can_attach() will
5157 * check this value once in the beginning of the process, and then carry
5158 * on with stale data. This means that changes to this value will only
5159 * affect task migrations starting after the change.
5161 memcg->move_charge_at_immigrate = val;
5165 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5166 struct cftype *cft, u64 val)
5173 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5174 struct cftype *cft, struct seq_file *m)
5177 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5178 unsigned long node_nr;
5179 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5181 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5182 seq_printf(m, "total=%lu", total_nr);
5183 for_each_node_state(nid, N_MEMORY) {
5184 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5185 seq_printf(m, " N%d=%lu", nid, node_nr);
5189 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5190 seq_printf(m, "file=%lu", file_nr);
5191 for_each_node_state(nid, N_MEMORY) {
5192 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5194 seq_printf(m, " N%d=%lu", nid, node_nr);
5198 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5199 seq_printf(m, "anon=%lu", anon_nr);
5200 for_each_node_state(nid, N_MEMORY) {
5201 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5203 seq_printf(m, " N%d=%lu", nid, node_nr);
5207 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5208 seq_printf(m, "unevictable=%lu", unevictable_nr);
5209 for_each_node_state(nid, N_MEMORY) {
5210 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5211 BIT(LRU_UNEVICTABLE));
5212 seq_printf(m, " N%d=%lu", nid, node_nr);
5217 #endif /* CONFIG_NUMA */
5219 static inline void mem_cgroup_lru_names_not_uptodate(void)
5221 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5224 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5227 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5228 struct mem_cgroup *mi;
5231 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5232 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5234 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5235 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5238 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5239 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5240 mem_cgroup_read_events(memcg, i));
5242 for (i = 0; i < NR_LRU_LISTS; i++)
5243 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5244 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5246 /* Hierarchical information */
5248 unsigned long long limit, memsw_limit;
5249 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5250 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5251 if (do_swap_account)
5252 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5256 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5259 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5261 for_each_mem_cgroup_tree(mi, memcg)
5262 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5263 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5266 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5267 unsigned long long val = 0;
5269 for_each_mem_cgroup_tree(mi, memcg)
5270 val += mem_cgroup_read_events(mi, i);
5271 seq_printf(m, "total_%s %llu\n",
5272 mem_cgroup_events_names[i], val);
5275 for (i = 0; i < NR_LRU_LISTS; i++) {
5276 unsigned long long val = 0;
5278 for_each_mem_cgroup_tree(mi, memcg)
5279 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5280 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5283 #ifdef CONFIG_DEBUG_VM
5286 struct mem_cgroup_per_zone *mz;
5287 struct zone_reclaim_stat *rstat;
5288 unsigned long recent_rotated[2] = {0, 0};
5289 unsigned long recent_scanned[2] = {0, 0};
5291 for_each_online_node(nid)
5292 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5293 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5294 rstat = &mz->lruvec.reclaim_stat;
5296 recent_rotated[0] += rstat->recent_rotated[0];
5297 recent_rotated[1] += rstat->recent_rotated[1];
5298 recent_scanned[0] += rstat->recent_scanned[0];
5299 recent_scanned[1] += rstat->recent_scanned[1];
5301 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5302 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5303 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5304 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5311 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5314 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5316 return mem_cgroup_swappiness(memcg);
5319 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5320 struct cftype *cft, u64 val)
5322 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5323 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5325 if (val > 100 || !parent)
5328 mutex_lock(&memcg_create_mutex);
5330 /* If under hierarchy, only empty-root can set this value */
5331 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5332 mutex_unlock(&memcg_create_mutex);
5336 memcg->swappiness = val;
5338 mutex_unlock(&memcg_create_mutex);
5343 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5345 struct mem_cgroup_threshold_ary *t;
5351 t = rcu_dereference(memcg->thresholds.primary);
5353 t = rcu_dereference(memcg->memsw_thresholds.primary);
5358 usage = mem_cgroup_usage(memcg, swap);
5361 * current_threshold points to threshold just below or equal to usage.
5362 * If it's not true, a threshold was crossed after last
5363 * call of __mem_cgroup_threshold().
5365 i = t->current_threshold;
5368 * Iterate backward over array of thresholds starting from
5369 * current_threshold and check if a threshold is crossed.
5370 * If none of thresholds below usage is crossed, we read
5371 * only one element of the array here.
5373 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5374 eventfd_signal(t->entries[i].eventfd, 1);
5376 /* i = current_threshold + 1 */
5380 * Iterate forward over array of thresholds starting from
5381 * current_threshold+1 and check if a threshold is crossed.
5382 * If none of thresholds above usage is crossed, we read
5383 * only one element of the array here.
5385 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5386 eventfd_signal(t->entries[i].eventfd, 1);
5388 /* Update current_threshold */
5389 t->current_threshold = i - 1;
5394 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5397 __mem_cgroup_threshold(memcg, false);
5398 if (do_swap_account)
5399 __mem_cgroup_threshold(memcg, true);
5401 memcg = parent_mem_cgroup(memcg);
5405 static int compare_thresholds(const void *a, const void *b)
5407 const struct mem_cgroup_threshold *_a = a;
5408 const struct mem_cgroup_threshold *_b = b;
5410 if (_a->threshold > _b->threshold)
5413 if (_a->threshold < _b->threshold)
5419 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5421 struct mem_cgroup_eventfd_list *ev;
5423 list_for_each_entry(ev, &memcg->oom_notify, list)
5424 eventfd_signal(ev->eventfd, 1);
5428 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5430 struct mem_cgroup *iter;
5432 for_each_mem_cgroup_tree(iter, memcg)
5433 mem_cgroup_oom_notify_cb(iter);
5436 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5437 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5439 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5440 struct mem_cgroup_thresholds *thresholds;
5441 struct mem_cgroup_threshold_ary *new;
5442 enum res_type type = MEMFILE_TYPE(cft->private);
5443 u64 threshold, usage;
5446 ret = res_counter_memparse_write_strategy(args, &threshold);
5450 mutex_lock(&memcg->thresholds_lock);
5453 thresholds = &memcg->thresholds;
5454 else if (type == _MEMSWAP)
5455 thresholds = &memcg->memsw_thresholds;
5459 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5461 /* Check if a threshold crossed before adding a new one */
5462 if (thresholds->primary)
5463 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5465 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5467 /* Allocate memory for new array of thresholds */
5468 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5476 /* Copy thresholds (if any) to new array */
5477 if (thresholds->primary) {
5478 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5479 sizeof(struct mem_cgroup_threshold));
5482 /* Add new threshold */
5483 new->entries[size - 1].eventfd = eventfd;
5484 new->entries[size - 1].threshold = threshold;
5486 /* Sort thresholds. Registering of new threshold isn't time-critical */
5487 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5488 compare_thresholds, NULL);
5490 /* Find current threshold */
5491 new->current_threshold = -1;
5492 for (i = 0; i < size; i++) {
5493 if (new->entries[i].threshold <= usage) {
5495 * new->current_threshold will not be used until
5496 * rcu_assign_pointer(), so it's safe to increment
5499 ++new->current_threshold;
5504 /* Free old spare buffer and save old primary buffer as spare */
5505 kfree(thresholds->spare);
5506 thresholds->spare = thresholds->primary;
5508 rcu_assign_pointer(thresholds->primary, new);
5510 /* To be sure that nobody uses thresholds */
5514 mutex_unlock(&memcg->thresholds_lock);
5519 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5520 struct cftype *cft, struct eventfd_ctx *eventfd)
5522 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5523 struct mem_cgroup_thresholds *thresholds;
5524 struct mem_cgroup_threshold_ary *new;
5525 enum res_type type = MEMFILE_TYPE(cft->private);
5529 mutex_lock(&memcg->thresholds_lock);
5531 thresholds = &memcg->thresholds;
5532 else if (type == _MEMSWAP)
5533 thresholds = &memcg->memsw_thresholds;
5537 if (!thresholds->primary)
5540 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5542 /* Check if a threshold crossed before removing */
5543 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5545 /* Calculate new number of threshold */
5547 for (i = 0; i < thresholds->primary->size; i++) {
5548 if (thresholds->primary->entries[i].eventfd != eventfd)
5552 new = thresholds->spare;
5554 /* Set thresholds array to NULL if we don't have thresholds */
5563 /* Copy thresholds and find current threshold */
5564 new->current_threshold = -1;
5565 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5566 if (thresholds->primary->entries[i].eventfd == eventfd)
5569 new->entries[j] = thresholds->primary->entries[i];
5570 if (new->entries[j].threshold <= usage) {
5572 * new->current_threshold will not be used
5573 * until rcu_assign_pointer(), so it's safe to increment
5576 ++new->current_threshold;
5582 /* Swap primary and spare array */
5583 thresholds->spare = thresholds->primary;
5584 /* If all events are unregistered, free the spare array */
5586 kfree(thresholds->spare);
5587 thresholds->spare = NULL;
5590 rcu_assign_pointer(thresholds->primary, new);
5592 /* To be sure that nobody uses thresholds */
5595 mutex_unlock(&memcg->thresholds_lock);
5598 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5599 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5601 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5602 struct mem_cgroup_eventfd_list *event;
5603 enum res_type type = MEMFILE_TYPE(cft->private);
5605 BUG_ON(type != _OOM_TYPE);
5606 event = kmalloc(sizeof(*event), GFP_KERNEL);
5610 spin_lock(&memcg_oom_lock);
5612 event->eventfd = eventfd;
5613 list_add(&event->list, &memcg->oom_notify);
5615 /* already in OOM ? */
5616 if (atomic_read(&memcg->under_oom))
5617 eventfd_signal(eventfd, 1);
5618 spin_unlock(&memcg_oom_lock);
5623 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5624 struct cftype *cft, struct eventfd_ctx *eventfd)
5626 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5627 struct mem_cgroup_eventfd_list *ev, *tmp;
5628 enum res_type type = MEMFILE_TYPE(cft->private);
5630 BUG_ON(type != _OOM_TYPE);
5632 spin_lock(&memcg_oom_lock);
5634 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5635 if (ev->eventfd == eventfd) {
5636 list_del(&ev->list);
5641 spin_unlock(&memcg_oom_lock);
5644 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5645 struct cftype *cft, struct cgroup_map_cb *cb)
5647 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5649 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5651 if (atomic_read(&memcg->under_oom))
5652 cb->fill(cb, "under_oom", 1);
5654 cb->fill(cb, "under_oom", 0);
5658 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5659 struct cftype *cft, u64 val)
5661 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5662 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5664 /* cannot set to root cgroup and only 0 and 1 are allowed */
5665 if (!parent || !((val == 0) || (val == 1)))
5668 mutex_lock(&memcg_create_mutex);
5669 /* oom-kill-disable is a flag for subhierarchy. */
5670 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5671 mutex_unlock(&memcg_create_mutex);
5674 memcg->oom_kill_disable = val;
5676 memcg_oom_recover(memcg);
5677 mutex_unlock(&memcg_create_mutex);
5681 #ifdef CONFIG_MEMCG_KMEM
5682 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5686 memcg->kmemcg_id = -1;
5687 ret = memcg_propagate_kmem(memcg);
5691 return mem_cgroup_sockets_init(memcg, ss);
5694 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5696 mem_cgroup_sockets_destroy(memcg);
5699 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5701 if (!memcg_kmem_is_active(memcg))
5705 * kmem charges can outlive the cgroup. In the case of slab
5706 * pages, for instance, a page contain objects from various
5707 * processes. As we prevent from taking a reference for every
5708 * such allocation we have to be careful when doing uncharge
5709 * (see memcg_uncharge_kmem) and here during offlining.
5711 * The idea is that that only the _last_ uncharge which sees
5712 * the dead memcg will drop the last reference. An additional
5713 * reference is taken here before the group is marked dead
5714 * which is then paired with css_put during uncharge resp. here.
5716 * Although this might sound strange as this path is called from
5717 * css_offline() when the referencemight have dropped down to 0
5718 * and shouldn't be incremented anymore (css_tryget would fail)
5719 * we do not have other options because of the kmem allocations
5722 css_get(&memcg->css);
5724 memcg_kmem_mark_dead(memcg);
5726 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5729 if (memcg_kmem_test_and_clear_dead(memcg))
5730 css_put(&memcg->css);
5733 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5738 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5742 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5747 static struct cftype mem_cgroup_files[] = {
5749 .name = "usage_in_bytes",
5750 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5751 .read = mem_cgroup_read,
5752 .register_event = mem_cgroup_usage_register_event,
5753 .unregister_event = mem_cgroup_usage_unregister_event,
5756 .name = "max_usage_in_bytes",
5757 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5758 .trigger = mem_cgroup_reset,
5759 .read = mem_cgroup_read,
5762 .name = "limit_in_bytes",
5763 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5764 .write_string = mem_cgroup_write,
5765 .read = mem_cgroup_read,
5768 .name = "soft_limit_in_bytes",
5769 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5770 .write_string = mem_cgroup_write,
5771 .read = mem_cgroup_read,
5775 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5776 .trigger = mem_cgroup_reset,
5777 .read = mem_cgroup_read,
5781 .read_seq_string = memcg_stat_show,
5784 .name = "force_empty",
5785 .trigger = mem_cgroup_force_empty_write,
5788 .name = "use_hierarchy",
5789 .flags = CFTYPE_INSANE,
5790 .write_u64 = mem_cgroup_hierarchy_write,
5791 .read_u64 = mem_cgroup_hierarchy_read,
5794 .name = "swappiness",
5795 .read_u64 = mem_cgroup_swappiness_read,
5796 .write_u64 = mem_cgroup_swappiness_write,
5799 .name = "move_charge_at_immigrate",
5800 .read_u64 = mem_cgroup_move_charge_read,
5801 .write_u64 = mem_cgroup_move_charge_write,
5804 .name = "oom_control",
5805 .read_map = mem_cgroup_oom_control_read,
5806 .write_u64 = mem_cgroup_oom_control_write,
5807 .register_event = mem_cgroup_oom_register_event,
5808 .unregister_event = mem_cgroup_oom_unregister_event,
5809 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5812 .name = "pressure_level",
5813 .register_event = vmpressure_register_event,
5814 .unregister_event = vmpressure_unregister_event,
5818 .name = "numa_stat",
5819 .read_seq_string = memcg_numa_stat_show,
5822 #ifdef CONFIG_MEMCG_KMEM
5824 .name = "kmem.limit_in_bytes",
5825 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5826 .write_string = mem_cgroup_write,
5827 .read = mem_cgroup_read,
5830 .name = "kmem.usage_in_bytes",
5831 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5832 .read = mem_cgroup_read,
5835 .name = "kmem.failcnt",
5836 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5837 .trigger = mem_cgroup_reset,
5838 .read = mem_cgroup_read,
5841 .name = "kmem.max_usage_in_bytes",
5842 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5843 .trigger = mem_cgroup_reset,
5844 .read = mem_cgroup_read,
5846 #ifdef CONFIG_SLABINFO
5848 .name = "kmem.slabinfo",
5849 .read_seq_string = mem_cgroup_slabinfo_read,
5853 { }, /* terminate */
5856 #ifdef CONFIG_MEMCG_SWAP
5857 static struct cftype memsw_cgroup_files[] = {
5859 .name = "memsw.usage_in_bytes",
5860 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5861 .read = mem_cgroup_read,
5862 .register_event = mem_cgroup_usage_register_event,
5863 .unregister_event = mem_cgroup_usage_unregister_event,
5866 .name = "memsw.max_usage_in_bytes",
5867 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5868 .trigger = mem_cgroup_reset,
5869 .read = mem_cgroup_read,
5872 .name = "memsw.limit_in_bytes",
5873 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5874 .write_string = mem_cgroup_write,
5875 .read = mem_cgroup_read,
5878 .name = "memsw.failcnt",
5879 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5880 .trigger = mem_cgroup_reset,
5881 .read = mem_cgroup_read,
5883 { }, /* terminate */
5886 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5888 struct mem_cgroup_per_node *pn;
5889 struct mem_cgroup_per_zone *mz;
5890 int zone, tmp = node;
5892 * This routine is called against possible nodes.
5893 * But it's BUG to call kmalloc() against offline node.
5895 * TODO: this routine can waste much memory for nodes which will
5896 * never be onlined. It's better to use memory hotplug callback
5899 if (!node_state(node, N_NORMAL_MEMORY))
5901 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5905 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5906 mz = &pn->zoneinfo[zone];
5907 lruvec_init(&mz->lruvec);
5910 memcg->nodeinfo[node] = pn;
5914 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5916 kfree(memcg->nodeinfo[node]);
5919 static struct mem_cgroup *mem_cgroup_alloc(void)
5921 struct mem_cgroup *memcg;
5922 size_t size = memcg_size();
5924 /* Can be very big if nr_node_ids is very big */
5925 if (size < PAGE_SIZE)
5926 memcg = kzalloc(size, GFP_KERNEL);
5928 memcg = vzalloc(size);
5933 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5936 spin_lock_init(&memcg->pcp_counter_lock);
5940 if (size < PAGE_SIZE)
5948 * At destroying mem_cgroup, references from swap_cgroup can remain.
5949 * (scanning all at force_empty is too costly...)
5951 * Instead of clearing all references at force_empty, we remember
5952 * the number of reference from swap_cgroup and free mem_cgroup when
5953 * it goes down to 0.
5955 * Removal of cgroup itself succeeds regardless of refs from swap.
5958 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5961 size_t size = memcg_size();
5963 free_css_id(&mem_cgroup_subsys, &memcg->css);
5966 free_mem_cgroup_per_zone_info(memcg, node);
5968 free_percpu(memcg->stat);
5971 * We need to make sure that (at least for now), the jump label
5972 * destruction code runs outside of the cgroup lock. This is because
5973 * get_online_cpus(), which is called from the static_branch update,
5974 * can't be called inside the cgroup_lock. cpusets are the ones
5975 * enforcing this dependency, so if they ever change, we might as well.
5977 * schedule_work() will guarantee this happens. Be careful if you need
5978 * to move this code around, and make sure it is outside
5981 disarm_static_keys(memcg);
5982 if (size < PAGE_SIZE)
5989 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5991 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5993 if (!memcg->res.parent)
5995 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5997 EXPORT_SYMBOL(parent_mem_cgroup);
5999 static struct cgroup_subsys_state * __ref
6000 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6002 struct mem_cgroup *memcg;
6003 long error = -ENOMEM;
6006 memcg = mem_cgroup_alloc();
6008 return ERR_PTR(error);
6011 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6015 if (parent_css == NULL) {
6016 root_mem_cgroup = memcg;
6017 res_counter_init(&memcg->res, NULL);
6018 res_counter_init(&memcg->memsw, NULL);
6019 res_counter_init(&memcg->kmem, NULL);
6022 memcg->last_scanned_node = MAX_NUMNODES;
6023 INIT_LIST_HEAD(&memcg->oom_notify);
6024 memcg->move_charge_at_immigrate = 0;
6025 mutex_init(&memcg->thresholds_lock);
6026 spin_lock_init(&memcg->move_lock);
6027 vmpressure_init(&memcg->vmpressure);
6028 spin_lock_init(&memcg->soft_lock);
6033 __mem_cgroup_free(memcg);
6034 return ERR_PTR(error);
6038 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6040 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6041 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6047 mutex_lock(&memcg_create_mutex);
6049 memcg->use_hierarchy = parent->use_hierarchy;
6050 memcg->oom_kill_disable = parent->oom_kill_disable;
6051 memcg->swappiness = mem_cgroup_swappiness(parent);
6053 if (parent->use_hierarchy) {
6054 res_counter_init(&memcg->res, &parent->res);
6055 res_counter_init(&memcg->memsw, &parent->memsw);
6056 res_counter_init(&memcg->kmem, &parent->kmem);
6059 * No need to take a reference to the parent because cgroup
6060 * core guarantees its existence.
6063 res_counter_init(&memcg->res, NULL);
6064 res_counter_init(&memcg->memsw, NULL);
6065 res_counter_init(&memcg->kmem, NULL);
6067 * Deeper hierachy with use_hierarchy == false doesn't make
6068 * much sense so let cgroup subsystem know about this
6069 * unfortunate state in our controller.
6071 if (parent != root_mem_cgroup)
6072 mem_cgroup_subsys.broken_hierarchy = true;
6075 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6076 mutex_unlock(&memcg_create_mutex);
6081 * Announce all parents that a group from their hierarchy is gone.
6083 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6085 struct mem_cgroup *parent = memcg;
6087 while ((parent = parent_mem_cgroup(parent)))
6088 mem_cgroup_iter_invalidate(parent);
6091 * if the root memcg is not hierarchical we have to check it
6094 if (!root_mem_cgroup->use_hierarchy)
6095 mem_cgroup_iter_invalidate(root_mem_cgroup);
6098 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6100 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6102 kmem_cgroup_css_offline(memcg);
6104 mem_cgroup_invalidate_reclaim_iterators(memcg);
6105 mem_cgroup_reparent_charges(memcg);
6106 if (memcg->soft_contributed) {
6107 while ((memcg = parent_mem_cgroup(memcg)))
6108 atomic_dec(&memcg->children_in_excess);
6110 mem_cgroup_destroy_all_caches(memcg);
6111 vmpressure_cleanup(&memcg->vmpressure);
6114 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6116 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6118 memcg_destroy_kmem(memcg);
6119 __mem_cgroup_free(memcg);
6123 /* Handlers for move charge at task migration. */
6124 #define PRECHARGE_COUNT_AT_ONCE 256
6125 static int mem_cgroup_do_precharge(unsigned long count)
6128 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6129 struct mem_cgroup *memcg = mc.to;
6131 if (mem_cgroup_is_root(memcg)) {
6132 mc.precharge += count;
6133 /* we don't need css_get for root */
6136 /* try to charge at once */
6138 struct res_counter *dummy;
6140 * "memcg" cannot be under rmdir() because we've already checked
6141 * by cgroup_lock_live_cgroup() that it is not removed and we
6142 * are still under the same cgroup_mutex. So we can postpone
6145 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6147 if (do_swap_account && res_counter_charge(&memcg->memsw,
6148 PAGE_SIZE * count, &dummy)) {
6149 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6152 mc.precharge += count;
6156 /* fall back to one by one charge */
6158 if (signal_pending(current)) {
6162 if (!batch_count--) {
6163 batch_count = PRECHARGE_COUNT_AT_ONCE;
6166 ret = __mem_cgroup_try_charge(NULL,
6167 GFP_KERNEL, 1, &memcg, false);
6169 /* mem_cgroup_clear_mc() will do uncharge later */
6177 * get_mctgt_type - get target type of moving charge
6178 * @vma: the vma the pte to be checked belongs
6179 * @addr: the address corresponding to the pte to be checked
6180 * @ptent: the pte to be checked
6181 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6184 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6185 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6186 * move charge. if @target is not NULL, the page is stored in target->page
6187 * with extra refcnt got(Callers should handle it).
6188 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6189 * target for charge migration. if @target is not NULL, the entry is stored
6192 * Called with pte lock held.
6199 enum mc_target_type {
6205 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6206 unsigned long addr, pte_t ptent)
6208 struct page *page = vm_normal_page(vma, addr, ptent);
6210 if (!page || !page_mapped(page))
6212 if (PageAnon(page)) {
6213 /* we don't move shared anon */
6216 } else if (!move_file())
6217 /* we ignore mapcount for file pages */
6219 if (!get_page_unless_zero(page))
6226 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6227 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6229 struct page *page = NULL;
6230 swp_entry_t ent = pte_to_swp_entry(ptent);
6232 if (!move_anon() || non_swap_entry(ent))
6235 * Because lookup_swap_cache() updates some statistics counter,
6236 * we call find_get_page() with swapper_space directly.
6238 page = find_get_page(swap_address_space(ent), ent.val);
6239 if (do_swap_account)
6240 entry->val = ent.val;
6245 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6246 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6252 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6253 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6255 struct page *page = NULL;
6256 struct address_space *mapping;
6259 if (!vma->vm_file) /* anonymous vma */
6264 mapping = vma->vm_file->f_mapping;
6265 if (pte_none(ptent))
6266 pgoff = linear_page_index(vma, addr);
6267 else /* pte_file(ptent) is true */
6268 pgoff = pte_to_pgoff(ptent);
6270 /* page is moved even if it's not RSS of this task(page-faulted). */
6271 page = find_get_page(mapping, pgoff);
6274 /* shmem/tmpfs may report page out on swap: account for that too. */
6275 if (radix_tree_exceptional_entry(page)) {
6276 swp_entry_t swap = radix_to_swp_entry(page);
6277 if (do_swap_account)
6279 page = find_get_page(swap_address_space(swap), swap.val);
6285 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6286 unsigned long addr, pte_t ptent, union mc_target *target)
6288 struct page *page = NULL;
6289 struct page_cgroup *pc;
6290 enum mc_target_type ret = MC_TARGET_NONE;
6291 swp_entry_t ent = { .val = 0 };
6293 if (pte_present(ptent))
6294 page = mc_handle_present_pte(vma, addr, ptent);
6295 else if (is_swap_pte(ptent))
6296 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6297 else if (pte_none(ptent) || pte_file(ptent))
6298 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6300 if (!page && !ent.val)
6303 pc = lookup_page_cgroup(page);
6305 * Do only loose check w/o page_cgroup lock.
6306 * mem_cgroup_move_account() checks the pc is valid or not under
6309 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6310 ret = MC_TARGET_PAGE;
6312 target->page = page;
6314 if (!ret || !target)
6317 /* There is a swap entry and a page doesn't exist or isn't charged */
6318 if (ent.val && !ret &&
6319 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6320 ret = MC_TARGET_SWAP;
6327 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6329 * We don't consider swapping or file mapped pages because THP does not
6330 * support them for now.
6331 * Caller should make sure that pmd_trans_huge(pmd) is true.
6333 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6334 unsigned long addr, pmd_t pmd, union mc_target *target)
6336 struct page *page = NULL;
6337 struct page_cgroup *pc;
6338 enum mc_target_type ret = MC_TARGET_NONE;
6340 page = pmd_page(pmd);
6341 VM_BUG_ON(!page || !PageHead(page));
6344 pc = lookup_page_cgroup(page);
6345 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6346 ret = MC_TARGET_PAGE;
6349 target->page = page;
6355 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6356 unsigned long addr, pmd_t pmd, union mc_target *target)
6358 return MC_TARGET_NONE;
6362 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6363 unsigned long addr, unsigned long end,
6364 struct mm_walk *walk)
6366 struct vm_area_struct *vma = walk->private;
6370 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6371 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6372 mc.precharge += HPAGE_PMD_NR;
6373 spin_unlock(&vma->vm_mm->page_table_lock);
6377 if (pmd_trans_unstable(pmd))
6379 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6380 for (; addr != end; pte++, addr += PAGE_SIZE)
6381 if (get_mctgt_type(vma, addr, *pte, NULL))
6382 mc.precharge++; /* increment precharge temporarily */
6383 pte_unmap_unlock(pte - 1, ptl);
6389 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6391 unsigned long precharge;
6392 struct vm_area_struct *vma;
6394 down_read(&mm->mmap_sem);
6395 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6396 struct mm_walk mem_cgroup_count_precharge_walk = {
6397 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6401 if (is_vm_hugetlb_page(vma))
6403 walk_page_range(vma->vm_start, vma->vm_end,
6404 &mem_cgroup_count_precharge_walk);
6406 up_read(&mm->mmap_sem);
6408 precharge = mc.precharge;
6414 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6416 unsigned long precharge = mem_cgroup_count_precharge(mm);
6418 VM_BUG_ON(mc.moving_task);
6419 mc.moving_task = current;
6420 return mem_cgroup_do_precharge(precharge);
6423 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6424 static void __mem_cgroup_clear_mc(void)
6426 struct mem_cgroup *from = mc.from;
6427 struct mem_cgroup *to = mc.to;
6430 /* we must uncharge all the leftover precharges from mc.to */
6432 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6436 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6437 * we must uncharge here.
6439 if (mc.moved_charge) {
6440 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6441 mc.moved_charge = 0;
6443 /* we must fixup refcnts and charges */
6444 if (mc.moved_swap) {
6445 /* uncharge swap account from the old cgroup */
6446 if (!mem_cgroup_is_root(mc.from))
6447 res_counter_uncharge(&mc.from->memsw,
6448 PAGE_SIZE * mc.moved_swap);
6450 for (i = 0; i < mc.moved_swap; i++)
6451 css_put(&mc.from->css);
6453 if (!mem_cgroup_is_root(mc.to)) {
6455 * we charged both to->res and to->memsw, so we should
6458 res_counter_uncharge(&mc.to->res,
6459 PAGE_SIZE * mc.moved_swap);
6461 /* we've already done css_get(mc.to) */
6464 memcg_oom_recover(from);
6465 memcg_oom_recover(to);
6466 wake_up_all(&mc.waitq);
6469 static void mem_cgroup_clear_mc(void)
6471 struct mem_cgroup *from = mc.from;
6474 * we must clear moving_task before waking up waiters at the end of
6477 mc.moving_task = NULL;
6478 __mem_cgroup_clear_mc();
6479 spin_lock(&mc.lock);
6482 spin_unlock(&mc.lock);
6483 mem_cgroup_end_move(from);
6486 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6487 struct cgroup_taskset *tset)
6489 struct task_struct *p = cgroup_taskset_first(tset);
6491 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6492 unsigned long move_charge_at_immigrate;
6495 * We are now commited to this value whatever it is. Changes in this
6496 * tunable will only affect upcoming migrations, not the current one.
6497 * So we need to save it, and keep it going.
6499 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6500 if (move_charge_at_immigrate) {
6501 struct mm_struct *mm;
6502 struct mem_cgroup *from = mem_cgroup_from_task(p);
6504 VM_BUG_ON(from == memcg);
6506 mm = get_task_mm(p);
6509 /* We move charges only when we move a owner of the mm */
6510 if (mm->owner == p) {
6513 VM_BUG_ON(mc.precharge);
6514 VM_BUG_ON(mc.moved_charge);
6515 VM_BUG_ON(mc.moved_swap);
6516 mem_cgroup_start_move(from);
6517 spin_lock(&mc.lock);
6520 mc.immigrate_flags = move_charge_at_immigrate;
6521 spin_unlock(&mc.lock);
6522 /* We set mc.moving_task later */
6524 ret = mem_cgroup_precharge_mc(mm);
6526 mem_cgroup_clear_mc();
6533 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6534 struct cgroup_taskset *tset)
6536 mem_cgroup_clear_mc();
6539 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6540 unsigned long addr, unsigned long end,
6541 struct mm_walk *walk)
6544 struct vm_area_struct *vma = walk->private;
6547 enum mc_target_type target_type;
6548 union mc_target target;
6550 struct page_cgroup *pc;
6553 * We don't take compound_lock() here but no race with splitting thp
6555 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6556 * under splitting, which means there's no concurrent thp split,
6557 * - if another thread runs into split_huge_page() just after we
6558 * entered this if-block, the thread must wait for page table lock
6559 * to be unlocked in __split_huge_page_splitting(), where the main
6560 * part of thp split is not executed yet.
6562 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6563 if (mc.precharge < HPAGE_PMD_NR) {
6564 spin_unlock(&vma->vm_mm->page_table_lock);
6567 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6568 if (target_type == MC_TARGET_PAGE) {
6570 if (!isolate_lru_page(page)) {
6571 pc = lookup_page_cgroup(page);
6572 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6573 pc, mc.from, mc.to)) {
6574 mc.precharge -= HPAGE_PMD_NR;
6575 mc.moved_charge += HPAGE_PMD_NR;
6577 putback_lru_page(page);
6581 spin_unlock(&vma->vm_mm->page_table_lock);
6585 if (pmd_trans_unstable(pmd))
6588 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6589 for (; addr != end; addr += PAGE_SIZE) {
6590 pte_t ptent = *(pte++);
6596 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6597 case MC_TARGET_PAGE:
6599 if (isolate_lru_page(page))
6601 pc = lookup_page_cgroup(page);
6602 if (!mem_cgroup_move_account(page, 1, pc,
6605 /* we uncharge from mc.from later. */
6608 putback_lru_page(page);
6609 put: /* get_mctgt_type() gets the page */
6612 case MC_TARGET_SWAP:
6614 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6616 /* we fixup refcnts and charges later. */
6624 pte_unmap_unlock(pte - 1, ptl);
6629 * We have consumed all precharges we got in can_attach().
6630 * We try charge one by one, but don't do any additional
6631 * charges to mc.to if we have failed in charge once in attach()
6634 ret = mem_cgroup_do_precharge(1);
6642 static void mem_cgroup_move_charge(struct mm_struct *mm)
6644 struct vm_area_struct *vma;
6646 lru_add_drain_all();
6648 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6650 * Someone who are holding the mmap_sem might be waiting in
6651 * waitq. So we cancel all extra charges, wake up all waiters,
6652 * and retry. Because we cancel precharges, we might not be able
6653 * to move enough charges, but moving charge is a best-effort
6654 * feature anyway, so it wouldn't be a big problem.
6656 __mem_cgroup_clear_mc();
6660 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6662 struct mm_walk mem_cgroup_move_charge_walk = {
6663 .pmd_entry = mem_cgroup_move_charge_pte_range,
6667 if (is_vm_hugetlb_page(vma))
6669 ret = walk_page_range(vma->vm_start, vma->vm_end,
6670 &mem_cgroup_move_charge_walk);
6673 * means we have consumed all precharges and failed in
6674 * doing additional charge. Just abandon here.
6678 up_read(&mm->mmap_sem);
6681 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6682 struct cgroup_taskset *tset)
6684 struct task_struct *p = cgroup_taskset_first(tset);
6685 struct mm_struct *mm = get_task_mm(p);
6689 mem_cgroup_move_charge(mm);
6693 mem_cgroup_clear_mc();
6695 #else /* !CONFIG_MMU */
6696 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6697 struct cgroup_taskset *tset)
6701 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6702 struct cgroup_taskset *tset)
6705 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6706 struct cgroup_taskset *tset)
6712 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6713 * to verify sane_behavior flag on each mount attempt.
6715 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6718 * use_hierarchy is forced with sane_behavior. cgroup core
6719 * guarantees that @root doesn't have any children, so turning it
6720 * on for the root memcg is enough.
6722 if (cgroup_sane_behavior(root_css->cgroup))
6723 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6726 struct cgroup_subsys mem_cgroup_subsys = {
6728 .subsys_id = mem_cgroup_subsys_id,
6729 .css_alloc = mem_cgroup_css_alloc,
6730 .css_online = mem_cgroup_css_online,
6731 .css_offline = mem_cgroup_css_offline,
6732 .css_free = mem_cgroup_css_free,
6733 .can_attach = mem_cgroup_can_attach,
6734 .cancel_attach = mem_cgroup_cancel_attach,
6735 .attach = mem_cgroup_move_task,
6736 .bind = mem_cgroup_bind,
6737 .base_cftypes = mem_cgroup_files,
6742 #ifdef CONFIG_MEMCG_SWAP
6743 static int __init enable_swap_account(char *s)
6745 if (!strcmp(s, "1"))
6746 really_do_swap_account = 1;
6747 else if (!strcmp(s, "0"))
6748 really_do_swap_account = 0;
6751 __setup("swapaccount=", enable_swap_account);
6753 static void __init memsw_file_init(void)
6755 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6758 static void __init enable_swap_cgroup(void)
6760 if (!mem_cgroup_disabled() && really_do_swap_account) {
6761 do_swap_account = 1;
6767 static void __init enable_swap_cgroup(void)
6773 * subsys_initcall() for memory controller.
6775 * Some parts like hotcpu_notifier() have to be initialized from this context
6776 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6777 * everything that doesn't depend on a specific mem_cgroup structure should
6778 * be initialized from here.
6780 static int __init mem_cgroup_init(void)
6782 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6783 enable_swap_cgroup();
6787 subsys_initcall(mem_cgroup_init);