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
88 * Statistics for memory cgroup.
90 enum mem_cgroup_stat_index {
92 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
94 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
95 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
96 MEM_CGROUP_STAT_RSS_HUGE, /* # of pages charged as anon huge */
97 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
98 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
99 MEM_CGROUP_STAT_NSTATS,
102 static const char * const mem_cgroup_stat_names[] = {
110 enum mem_cgroup_events_index {
111 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
112 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
113 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
114 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
115 MEM_CGROUP_EVENTS_NSTATS,
118 static const char * const mem_cgroup_events_names[] = {
125 static const char * const mem_cgroup_lru_names[] = {
134 * Per memcg event counter is incremented at every pagein/pageout. With THP,
135 * it will be incremated by the number of pages. This counter is used for
136 * for trigger some periodic events. This is straightforward and better
137 * than using jiffies etc. to handle periodic memcg event.
139 enum mem_cgroup_events_target {
140 MEM_CGROUP_TARGET_THRESH,
141 MEM_CGROUP_TARGET_SOFTLIMIT,
142 MEM_CGROUP_TARGET_NUMAINFO,
145 #define THRESHOLDS_EVENTS_TARGET 128
146 #define SOFTLIMIT_EVENTS_TARGET 1024
147 #define NUMAINFO_EVENTS_TARGET 1024
149 struct mem_cgroup_stat_cpu {
150 long count[MEM_CGROUP_STAT_NSTATS];
151 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
152 unsigned long nr_page_events;
153 unsigned long targets[MEM_CGROUP_NTARGETS];
156 struct mem_cgroup_reclaim_iter {
158 * last scanned hierarchy member. Valid only if last_dead_count
159 * matches memcg->dead_count of the hierarchy root group.
161 struct mem_cgroup *last_visited;
162 unsigned long last_dead_count;
164 /* scan generation, increased every round-trip */
165 unsigned int generation;
169 * per-zone information in memory controller.
171 struct mem_cgroup_per_zone {
172 struct lruvec lruvec;
173 unsigned long lru_size[NR_LRU_LISTS];
175 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
177 struct mem_cgroup *memcg; /* Back pointer, we cannot */
178 /* use container_of */
181 struct mem_cgroup_per_node {
182 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
185 struct mem_cgroup_threshold {
186 struct eventfd_ctx *eventfd;
191 struct mem_cgroup_threshold_ary {
192 /* An array index points to threshold just below or equal to usage. */
193 int current_threshold;
194 /* Size of entries[] */
196 /* Array of thresholds */
197 struct mem_cgroup_threshold entries[0];
200 struct mem_cgroup_thresholds {
201 /* Primary thresholds array */
202 struct mem_cgroup_threshold_ary *primary;
204 * Spare threshold array.
205 * This is needed to make mem_cgroup_unregister_event() "never fail".
206 * It must be able to store at least primary->size - 1 entries.
208 struct mem_cgroup_threshold_ary *spare;
212 struct mem_cgroup_eventfd_list {
213 struct list_head list;
214 struct eventfd_ctx *eventfd;
217 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
218 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
221 * The memory controller data structure. The memory controller controls both
222 * page cache and RSS per cgroup. We would eventually like to provide
223 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
224 * to help the administrator determine what knobs to tune.
226 * TODO: Add a water mark for the memory controller. Reclaim will begin when
227 * we hit the water mark. May be even add a low water mark, such that
228 * no reclaim occurs from a cgroup at it's low water mark, this is
229 * a feature that will be implemented much later in the future.
232 struct cgroup_subsys_state css;
234 * the counter to account for memory usage
236 struct res_counter res;
238 /* vmpressure notifications */
239 struct vmpressure vmpressure;
242 * the counter to account for mem+swap usage.
244 struct res_counter memsw;
247 * the counter to account for kernel memory usage.
249 struct res_counter kmem;
251 * Should the accounting and control be hierarchical, per subtree?
254 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
260 /* OOM-Killer disable */
261 int oom_kill_disable;
263 /* set when res.limit == memsw.limit */
264 bool memsw_is_minimum;
266 /* protect arrays of thresholds */
267 struct mutex thresholds_lock;
269 /* thresholds for memory usage. RCU-protected */
270 struct mem_cgroup_thresholds thresholds;
272 /* thresholds for mem+swap usage. RCU-protected */
273 struct mem_cgroup_thresholds memsw_thresholds;
275 /* For oom notifier event fd */
276 struct list_head oom_notify;
279 * Should we move charges of a task when a task is moved into this
280 * mem_cgroup ? And what type of charges should we move ?
282 unsigned long move_charge_at_immigrate;
284 * set > 0 if pages under this cgroup are moving to other cgroup.
286 atomic_t moving_account;
287 /* taken only while moving_account > 0 */
288 spinlock_t move_lock;
292 struct mem_cgroup_stat_cpu __percpu *stat;
294 * used when a cpu is offlined or other synchronizations
295 * See mem_cgroup_read_stat().
297 struct mem_cgroup_stat_cpu nocpu_base;
298 spinlock_t pcp_counter_lock;
301 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
302 struct tcp_memcontrol tcp_mem;
304 #if defined(CONFIG_MEMCG_KMEM)
305 /* analogous to slab_common's slab_caches list. per-memcg */
306 struct list_head memcg_slab_caches;
307 /* Not a spinlock, we can take a lot of time walking the list */
308 struct mutex slab_caches_mutex;
309 /* Index in the kmem_cache->memcg_params->memcg_caches array */
313 int last_scanned_node;
315 nodemask_t scan_nodes;
316 atomic_t numainfo_events;
317 atomic_t numainfo_updating;
320 * Protects soft_contributed transitions.
321 * See mem_cgroup_update_soft_limit
323 spinlock_t soft_lock;
326 * If true then this group has increased parents' children_in_excess
327 * when it got over the soft limit.
328 * When a group falls bellow the soft limit, parents' children_in_excess
329 * is decreased and soft_contributed changed to false.
331 bool soft_contributed;
333 /* Number of children that are in soft limit excess */
334 atomic_t children_in_excess;
336 struct mem_cgroup_per_node *nodeinfo[0];
337 /* WARNING: nodeinfo must be the last member here */
340 static size_t memcg_size(void)
342 return sizeof(struct mem_cgroup) +
343 nr_node_ids * sizeof(struct mem_cgroup_per_node);
346 /* internal only representation about the status of kmem accounting. */
348 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
349 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
350 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
353 /* We account when limit is on, but only after call sites are patched */
354 #define KMEM_ACCOUNTED_MASK \
355 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
357 #ifdef CONFIG_MEMCG_KMEM
358 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
360 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
363 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
365 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
368 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
370 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
373 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
375 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
378 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
381 * Our caller must use css_get() first, because memcg_uncharge_kmem()
382 * will call css_put() if it sees the memcg is dead.
385 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
386 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
389 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
391 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
392 &memcg->kmem_account_flags);
396 /* Stuffs for move charges at task migration. */
398 * Types of charges to be moved. "move_charge_at_immitgrate" and
399 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
402 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
403 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
407 /* "mc" and its members are protected by cgroup_mutex */
408 static struct move_charge_struct {
409 spinlock_t lock; /* for from, to */
410 struct mem_cgroup *from;
411 struct mem_cgroup *to;
412 unsigned long immigrate_flags;
413 unsigned long precharge;
414 unsigned long moved_charge;
415 unsigned long moved_swap;
416 struct task_struct *moving_task; /* a task moving charges */
417 wait_queue_head_t waitq; /* a waitq for other context */
419 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
420 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
423 static bool move_anon(void)
425 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
428 static bool move_file(void)
430 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
434 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
435 * limit reclaim to prevent infinite loops, if they ever occur.
437 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
440 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
441 MEM_CGROUP_CHARGE_TYPE_ANON,
442 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
443 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
447 /* for encoding cft->private value on file */
455 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
456 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
457 #define MEMFILE_ATTR(val) ((val) & 0xffff)
458 /* Used for OOM nofiier */
459 #define OOM_CONTROL (0)
462 * Reclaim flags for mem_cgroup_hierarchical_reclaim
464 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
465 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
466 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
467 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
470 * The memcg_create_mutex will be held whenever a new cgroup is created.
471 * As a consequence, any change that needs to protect against new child cgroups
472 * appearing has to hold it as well.
474 static DEFINE_MUTEX(memcg_create_mutex);
476 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
478 return s ? container_of(s, struct mem_cgroup, css) : NULL;
481 /* Some nice accessors for the vmpressure. */
482 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
485 memcg = root_mem_cgroup;
486 return &memcg->vmpressure;
489 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
491 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
494 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
496 return &mem_cgroup_from_css(css)->vmpressure;
499 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
501 return (memcg == root_mem_cgroup);
504 /* Writing them here to avoid exposing memcg's inner layout */
505 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
507 void sock_update_memcg(struct sock *sk)
509 if (mem_cgroup_sockets_enabled) {
510 struct mem_cgroup *memcg;
511 struct cg_proto *cg_proto;
513 BUG_ON(!sk->sk_prot->proto_cgroup);
515 /* Socket cloning can throw us here with sk_cgrp already
516 * filled. It won't however, necessarily happen from
517 * process context. So the test for root memcg given
518 * the current task's memcg won't help us in this case.
520 * Respecting the original socket's memcg is a better
521 * decision in this case.
524 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
525 css_get(&sk->sk_cgrp->memcg->css);
530 memcg = mem_cgroup_from_task(current);
531 cg_proto = sk->sk_prot->proto_cgroup(memcg);
532 if (!mem_cgroup_is_root(memcg) &&
533 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
534 sk->sk_cgrp = cg_proto;
539 EXPORT_SYMBOL(sock_update_memcg);
541 void sock_release_memcg(struct sock *sk)
543 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
544 struct mem_cgroup *memcg;
545 WARN_ON(!sk->sk_cgrp->memcg);
546 memcg = sk->sk_cgrp->memcg;
547 css_put(&sk->sk_cgrp->memcg->css);
551 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
553 if (!memcg || mem_cgroup_is_root(memcg))
556 return &memcg->tcp_mem.cg_proto;
558 EXPORT_SYMBOL(tcp_proto_cgroup);
560 static void disarm_sock_keys(struct mem_cgroup *memcg)
562 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
564 static_key_slow_dec(&memcg_socket_limit_enabled);
567 static void disarm_sock_keys(struct mem_cgroup *memcg)
572 #ifdef CONFIG_MEMCG_KMEM
574 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
575 * There are two main reasons for not using the css_id for this:
576 * 1) this works better in sparse environments, where we have a lot of memcgs,
577 * but only a few kmem-limited. Or also, if we have, for instance, 200
578 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
579 * 200 entry array for that.
581 * 2) In order not to violate the cgroup API, we would like to do all memory
582 * allocation in ->create(). At that point, we haven't yet allocated the
583 * css_id. Having a separate index prevents us from messing with the cgroup
586 * The current size of the caches array is stored in
587 * memcg_limited_groups_array_size. It will double each time we have to
590 static DEFINE_IDA(kmem_limited_groups);
591 int memcg_limited_groups_array_size;
594 * MIN_SIZE is different than 1, because we would like to avoid going through
595 * the alloc/free process all the time. In a small machine, 4 kmem-limited
596 * cgroups is a reasonable guess. In the future, it could be a parameter or
597 * tunable, but that is strictly not necessary.
599 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
600 * this constant directly from cgroup, but it is understandable that this is
601 * better kept as an internal representation in cgroup.c. In any case, the
602 * css_id space is not getting any smaller, and we don't have to necessarily
603 * increase ours as well if it increases.
605 #define MEMCG_CACHES_MIN_SIZE 4
606 #define MEMCG_CACHES_MAX_SIZE 65535
609 * A lot of the calls to the cache allocation functions are expected to be
610 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
611 * conditional to this static branch, we'll have to allow modules that does
612 * kmem_cache_alloc and the such to see this symbol as well
614 struct static_key memcg_kmem_enabled_key;
615 EXPORT_SYMBOL(memcg_kmem_enabled_key);
617 static void disarm_kmem_keys(struct mem_cgroup *memcg)
619 if (memcg_kmem_is_active(memcg)) {
620 static_key_slow_dec(&memcg_kmem_enabled_key);
621 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
624 * This check can't live in kmem destruction function,
625 * since the charges will outlive the cgroup
627 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
630 static void disarm_kmem_keys(struct mem_cgroup *memcg)
633 #endif /* CONFIG_MEMCG_KMEM */
635 static void disarm_static_keys(struct mem_cgroup *memcg)
637 disarm_sock_keys(memcg);
638 disarm_kmem_keys(memcg);
641 static void drain_all_stock_async(struct mem_cgroup *memcg);
643 static struct mem_cgroup_per_zone *
644 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
646 VM_BUG_ON((unsigned)nid >= nr_node_ids);
647 return &memcg->nodeinfo[nid]->zoneinfo[zid];
650 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
655 static struct mem_cgroup_per_zone *
656 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
658 int nid = page_to_nid(page);
659 int zid = page_zonenum(page);
661 return mem_cgroup_zoneinfo(memcg, nid, zid);
665 * Implementation Note: reading percpu statistics for memcg.
667 * Both of vmstat[] and percpu_counter has threshold and do periodic
668 * synchronization to implement "quick" read. There are trade-off between
669 * reading cost and precision of value. Then, we may have a chance to implement
670 * a periodic synchronizion of counter in memcg's counter.
672 * But this _read() function is used for user interface now. The user accounts
673 * memory usage by memory cgroup and he _always_ requires exact value because
674 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
675 * have to visit all online cpus and make sum. So, for now, unnecessary
676 * synchronization is not implemented. (just implemented for cpu hotplug)
678 * If there are kernel internal actions which can make use of some not-exact
679 * value, and reading all cpu value can be performance bottleneck in some
680 * common workload, threashold and synchonization as vmstat[] should be
683 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
684 enum mem_cgroup_stat_index idx)
690 for_each_online_cpu(cpu)
691 val += per_cpu(memcg->stat->count[idx], cpu);
692 #ifdef CONFIG_HOTPLUG_CPU
693 spin_lock(&memcg->pcp_counter_lock);
694 val += memcg->nocpu_base.count[idx];
695 spin_unlock(&memcg->pcp_counter_lock);
701 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
704 int val = (charge) ? 1 : -1;
705 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
708 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
709 enum mem_cgroup_events_index idx)
711 unsigned long val = 0;
714 for_each_online_cpu(cpu)
715 val += per_cpu(memcg->stat->events[idx], cpu);
716 #ifdef CONFIG_HOTPLUG_CPU
717 spin_lock(&memcg->pcp_counter_lock);
718 val += memcg->nocpu_base.events[idx];
719 spin_unlock(&memcg->pcp_counter_lock);
724 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
726 bool anon, int nr_pages)
731 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
732 * counted as CACHE even if it's on ANON LRU.
735 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
738 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
741 if (PageTransHuge(page))
742 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
745 /* pagein of a big page is an event. So, ignore page size */
747 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
749 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
750 nr_pages = -nr_pages; /* for event */
753 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
759 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
761 struct mem_cgroup_per_zone *mz;
763 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
764 return mz->lru_size[lru];
768 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
769 unsigned int lru_mask)
771 struct mem_cgroup_per_zone *mz;
773 unsigned long ret = 0;
775 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
778 if (BIT(lru) & lru_mask)
779 ret += mz->lru_size[lru];
785 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
786 int nid, unsigned int lru_mask)
791 for (zid = 0; zid < MAX_NR_ZONES; zid++)
792 total += mem_cgroup_zone_nr_lru_pages(memcg,
798 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
799 unsigned int lru_mask)
804 for_each_node_state(nid, N_MEMORY)
805 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
809 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
810 enum mem_cgroup_events_target target)
812 unsigned long val, next;
814 val = __this_cpu_read(memcg->stat->nr_page_events);
815 next = __this_cpu_read(memcg->stat->targets[target]);
816 /* from time_after() in jiffies.h */
817 if ((long)next - (long)val < 0) {
819 case MEM_CGROUP_TARGET_THRESH:
820 next = val + THRESHOLDS_EVENTS_TARGET;
822 case MEM_CGROUP_TARGET_SOFTLIMIT:
823 next = val + SOFTLIMIT_EVENTS_TARGET;
825 case MEM_CGROUP_TARGET_NUMAINFO:
826 next = val + NUMAINFO_EVENTS_TARGET;
831 __this_cpu_write(memcg->stat->targets[target], next);
838 * Called from rate-limitted memcg_check_events when enough
839 * MEM_CGROUP_TARGET_SOFTLIMIT events are accumulated and it makes sure
840 * that all the parents up the hierarchy will be noticed that this group
841 * is in excess or that it is not in excess anymore. mmecg->soft_contributed
842 * makes the transition a single action whenever the state flips from one to
845 static void mem_cgroup_update_soft_limit(struct mem_cgroup *memcg)
847 unsigned long long excess = res_counter_soft_limit_excess(&memcg->res);
848 struct mem_cgroup *parent = memcg;
851 spin_lock(&memcg->soft_lock);
853 if (!memcg->soft_contributed) {
855 memcg->soft_contributed = true;
858 if (memcg->soft_contributed) {
860 memcg->soft_contributed = false;
865 * Necessary to update all ancestors when hierarchy is used
866 * because their event counter is not touched.
868 while (delta && (parent = parent_mem_cgroup(parent)))
869 atomic_add(delta, &parent->children_in_excess);
870 spin_unlock(&memcg->soft_lock);
874 * Check events in order.
877 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
880 /* threshold event is triggered in finer grain than soft limit */
881 if (unlikely(mem_cgroup_event_ratelimit(memcg,
882 MEM_CGROUP_TARGET_THRESH))) {
884 bool do_numainfo __maybe_unused;
886 do_softlimit = mem_cgroup_event_ratelimit(memcg,
887 MEM_CGROUP_TARGET_SOFTLIMIT);
889 do_numainfo = mem_cgroup_event_ratelimit(memcg,
890 MEM_CGROUP_TARGET_NUMAINFO);
894 mem_cgroup_threshold(memcg);
895 if (unlikely(do_softlimit))
896 mem_cgroup_update_soft_limit(memcg);
898 if (unlikely(do_numainfo))
899 atomic_inc(&memcg->numainfo_events);
905 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
908 * mm_update_next_owner() may clear mm->owner to NULL
909 * if it races with swapoff, page migration, etc.
910 * So this can be called with p == NULL.
915 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
918 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
920 struct mem_cgroup *memcg = NULL;
925 * Because we have no locks, mm->owner's may be being moved to other
926 * cgroup. We use css_tryget() here even if this looks
927 * pessimistic (rather than adding locks here).
931 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
932 if (unlikely(!memcg))
934 } while (!css_tryget(&memcg->css));
939 static enum mem_cgroup_filter_t
940 mem_cgroup_filter(struct mem_cgroup *memcg, struct mem_cgroup *root,
941 mem_cgroup_iter_filter cond)
945 return cond(memcg, root);
949 * Returns a next (in a pre-order walk) alive memcg (with elevated css
950 * ref. count) or NULL if the whole root's subtree has been visited.
952 * helper function to be used by mem_cgroup_iter
954 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
955 struct mem_cgroup *last_visited, mem_cgroup_iter_filter cond)
957 struct cgroup_subsys_state *prev_css, *next_css;
959 prev_css = last_visited ? &last_visited->css : NULL;
961 next_css = css_next_descendant_pre(prev_css, &root->css);
964 * Even if we found a group we have to make sure it is
965 * alive. css && !memcg means that the groups should be
966 * skipped and we should continue the tree walk.
967 * last_visited css is safe to use because it is
968 * protected by css_get and the tree walk is rcu safe.
971 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
973 switch (mem_cgroup_filter(mem, root, cond)) {
981 * css_rightmost_descendant is not an optimal way to
982 * skip through a subtree (especially for imbalanced
983 * trees leaning to right) but that's what we have right
984 * now. More effective solution would be traversing
985 * right-up for first non-NULL without calling
986 * css_next_descendant_pre afterwards.
988 prev_css = css_rightmost_descendant(next_css);
991 if (css_tryget(&mem->css))
1004 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1007 * When a group in the hierarchy below root is destroyed, the
1008 * hierarchy iterator can no longer be trusted since it might
1009 * have pointed to the destroyed group. Invalidate it.
1011 atomic_inc(&root->dead_count);
1014 static struct mem_cgroup *
1015 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1016 struct mem_cgroup *root,
1019 struct mem_cgroup *position = NULL;
1021 * A cgroup destruction happens in two stages: offlining and
1022 * release. They are separated by a RCU grace period.
1024 * If the iterator is valid, we may still race with an
1025 * offlining. The RCU lock ensures the object won't be
1026 * released, tryget will fail if we lost the race.
1028 *sequence = atomic_read(&root->dead_count);
1029 if (iter->last_dead_count == *sequence) {
1031 position = iter->last_visited;
1032 if (position && !css_tryget(&position->css))
1038 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1039 struct mem_cgroup *last_visited,
1040 struct mem_cgroup *new_position,
1044 css_put(&last_visited->css);
1046 * We store the sequence count from the time @last_visited was
1047 * loaded successfully instead of rereading it here so that we
1048 * don't lose destruction events in between. We could have
1049 * raced with the destruction of @new_position after all.
1051 iter->last_visited = new_position;
1053 iter->last_dead_count = sequence;
1057 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1058 * @root: hierarchy root
1059 * @prev: previously returned memcg, NULL on first invocation
1060 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1061 * @cond: filter for visited nodes, NULL for no filter
1063 * Returns references to children of the hierarchy below @root, or
1064 * @root itself, or %NULL after a full round-trip.
1066 * Caller must pass the return value in @prev on subsequent
1067 * invocations for reference counting, or use mem_cgroup_iter_break()
1068 * to cancel a hierarchy walk before the round-trip is complete.
1070 * Reclaimers can specify a zone and a priority level in @reclaim to
1071 * divide up the memcgs in the hierarchy among all concurrent
1072 * reclaimers operating on the same zone and priority.
1074 struct mem_cgroup *mem_cgroup_iter_cond(struct mem_cgroup *root,
1075 struct mem_cgroup *prev,
1076 struct mem_cgroup_reclaim_cookie *reclaim,
1077 mem_cgroup_iter_filter cond)
1079 struct mem_cgroup *memcg = NULL;
1080 struct mem_cgroup *last_visited = NULL;
1082 if (mem_cgroup_disabled()) {
1083 /* first call must return non-NULL, second return NULL */
1084 return (struct mem_cgroup *)(unsigned long)!prev;
1088 root = root_mem_cgroup;
1090 if (prev && !reclaim)
1091 last_visited = prev;
1093 if (!root->use_hierarchy && root != root_mem_cgroup) {
1096 if (mem_cgroup_filter(root, root, cond) == VISIT)
1103 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1104 int uninitialized_var(seq);
1107 int nid = zone_to_nid(reclaim->zone);
1108 int zid = zone_idx(reclaim->zone);
1109 struct mem_cgroup_per_zone *mz;
1111 mz = mem_cgroup_zoneinfo(root, nid, zid);
1112 iter = &mz->reclaim_iter[reclaim->priority];
1113 if (prev && reclaim->generation != iter->generation) {
1114 iter->last_visited = NULL;
1118 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1121 memcg = __mem_cgroup_iter_next(root, last_visited, cond);
1124 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1128 else if (!prev && memcg)
1129 reclaim->generation = iter->generation;
1133 * We have finished the whole tree walk or no group has been
1134 * visited because filter told us to skip the root node.
1136 if (!memcg && (prev || (cond && !last_visited)))
1142 if (prev && prev != root)
1143 css_put(&prev->css);
1149 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1150 * @root: hierarchy root
1151 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1153 void mem_cgroup_iter_break(struct mem_cgroup *root,
1154 struct mem_cgroup *prev)
1157 root = root_mem_cgroup;
1158 if (prev && prev != root)
1159 css_put(&prev->css);
1163 * Iteration constructs for visiting all cgroups (under a tree). If
1164 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1165 * be used for reference counting.
1167 #define for_each_mem_cgroup_tree(iter, root) \
1168 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1170 iter = mem_cgroup_iter(root, iter, NULL))
1172 #define for_each_mem_cgroup(iter) \
1173 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1175 iter = mem_cgroup_iter(NULL, iter, NULL))
1177 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1179 struct mem_cgroup *memcg;
1182 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1183 if (unlikely(!memcg))
1188 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1191 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1199 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1202 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1203 * @zone: zone of the wanted lruvec
1204 * @memcg: memcg of the wanted lruvec
1206 * Returns the lru list vector holding pages for the given @zone and
1207 * @mem. This can be the global zone lruvec, if the memory controller
1210 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1211 struct mem_cgroup *memcg)
1213 struct mem_cgroup_per_zone *mz;
1214 struct lruvec *lruvec;
1216 if (mem_cgroup_disabled()) {
1217 lruvec = &zone->lruvec;
1221 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1222 lruvec = &mz->lruvec;
1225 * Since a node can be onlined after the mem_cgroup was created,
1226 * we have to be prepared to initialize lruvec->zone here;
1227 * and if offlined then reonlined, we need to reinitialize it.
1229 if (unlikely(lruvec->zone != zone))
1230 lruvec->zone = zone;
1235 * Following LRU functions are allowed to be used without PCG_LOCK.
1236 * Operations are called by routine of global LRU independently from memcg.
1237 * What we have to take care of here is validness of pc->mem_cgroup.
1239 * Changes to pc->mem_cgroup happens when
1242 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1243 * It is added to LRU before charge.
1244 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1245 * When moving account, the page is not on LRU. It's isolated.
1249 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1251 * @zone: zone of the page
1253 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1255 struct mem_cgroup_per_zone *mz;
1256 struct mem_cgroup *memcg;
1257 struct page_cgroup *pc;
1258 struct lruvec *lruvec;
1260 if (mem_cgroup_disabled()) {
1261 lruvec = &zone->lruvec;
1265 pc = lookup_page_cgroup(page);
1266 memcg = pc->mem_cgroup;
1269 * Surreptitiously switch any uncharged offlist page to root:
1270 * an uncharged page off lru does nothing to secure
1271 * its former mem_cgroup from sudden removal.
1273 * Our caller holds lru_lock, and PageCgroupUsed is updated
1274 * under page_cgroup lock: between them, they make all uses
1275 * of pc->mem_cgroup safe.
1277 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1278 pc->mem_cgroup = memcg = root_mem_cgroup;
1280 mz = page_cgroup_zoneinfo(memcg, page);
1281 lruvec = &mz->lruvec;
1284 * Since a node can be onlined after the mem_cgroup was created,
1285 * we have to be prepared to initialize lruvec->zone here;
1286 * and if offlined then reonlined, we need to reinitialize it.
1288 if (unlikely(lruvec->zone != zone))
1289 lruvec->zone = zone;
1294 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1295 * @lruvec: mem_cgroup per zone lru vector
1296 * @lru: index of lru list the page is sitting on
1297 * @nr_pages: positive when adding or negative when removing
1299 * This function must be called when a page is added to or removed from an
1302 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1305 struct mem_cgroup_per_zone *mz;
1306 unsigned long *lru_size;
1308 if (mem_cgroup_disabled())
1311 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1312 lru_size = mz->lru_size + lru;
1313 *lru_size += nr_pages;
1314 VM_BUG_ON((long)(*lru_size) < 0);
1318 * Checks whether given mem is same or in the root_mem_cgroup's
1321 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1322 struct mem_cgroup *memcg)
1324 if (root_memcg == memcg)
1326 if (!root_memcg->use_hierarchy || !memcg)
1328 return css_is_ancestor(&memcg->css, &root_memcg->css);
1331 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1332 struct mem_cgroup *memcg)
1337 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1342 bool task_in_mem_cgroup(struct task_struct *task,
1343 const struct mem_cgroup *memcg)
1345 struct mem_cgroup *curr = NULL;
1346 struct task_struct *p;
1349 p = find_lock_task_mm(task);
1351 curr = try_get_mem_cgroup_from_mm(p->mm);
1355 * All threads may have already detached their mm's, but the oom
1356 * killer still needs to detect if they have already been oom
1357 * killed to prevent needlessly killing additional tasks.
1360 curr = mem_cgroup_from_task(task);
1362 css_get(&curr->css);
1368 * We should check use_hierarchy of "memcg" not "curr". Because checking
1369 * use_hierarchy of "curr" here make this function true if hierarchy is
1370 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1371 * hierarchy(even if use_hierarchy is disabled in "memcg").
1373 ret = mem_cgroup_same_or_subtree(memcg, curr);
1374 css_put(&curr->css);
1378 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1380 unsigned long inactive_ratio;
1381 unsigned long inactive;
1382 unsigned long active;
1385 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1386 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1388 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1390 inactive_ratio = int_sqrt(10 * gb);
1394 return inactive * inactive_ratio < active;
1397 #define mem_cgroup_from_res_counter(counter, member) \
1398 container_of(counter, struct mem_cgroup, member)
1401 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1402 * @memcg: the memory cgroup
1404 * Returns the maximum amount of memory @mem can be charged with, in
1407 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1409 unsigned long long margin;
1411 margin = res_counter_margin(&memcg->res);
1412 if (do_swap_account)
1413 margin = min(margin, res_counter_margin(&memcg->memsw));
1414 return margin >> PAGE_SHIFT;
1417 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1420 if (!css_parent(&memcg->css))
1421 return vm_swappiness;
1423 return memcg->swappiness;
1427 * memcg->moving_account is used for checking possibility that some thread is
1428 * calling move_account(). When a thread on CPU-A starts moving pages under
1429 * a memcg, other threads should check memcg->moving_account under
1430 * rcu_read_lock(), like this:
1434 * memcg->moving_account+1 if (memcg->mocing_account)
1436 * synchronize_rcu() update something.
1441 /* for quick checking without looking up memcg */
1442 atomic_t memcg_moving __read_mostly;
1444 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1446 atomic_inc(&memcg_moving);
1447 atomic_inc(&memcg->moving_account);
1451 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1454 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1455 * We check NULL in callee rather than caller.
1458 atomic_dec(&memcg_moving);
1459 atomic_dec(&memcg->moving_account);
1464 * 2 routines for checking "mem" is under move_account() or not.
1466 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1467 * is used for avoiding races in accounting. If true,
1468 * pc->mem_cgroup may be overwritten.
1470 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1471 * under hierarchy of moving cgroups. This is for
1472 * waiting at hith-memory prressure caused by "move".
1475 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1477 VM_BUG_ON(!rcu_read_lock_held());
1478 return atomic_read(&memcg->moving_account) > 0;
1481 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1483 struct mem_cgroup *from;
1484 struct mem_cgroup *to;
1487 * Unlike task_move routines, we access mc.to, mc.from not under
1488 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1490 spin_lock(&mc.lock);
1496 ret = mem_cgroup_same_or_subtree(memcg, from)
1497 || mem_cgroup_same_or_subtree(memcg, to);
1499 spin_unlock(&mc.lock);
1503 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1505 if (mc.moving_task && current != mc.moving_task) {
1506 if (mem_cgroup_under_move(memcg)) {
1508 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1509 /* moving charge context might have finished. */
1512 finish_wait(&mc.waitq, &wait);
1520 * Take this lock when
1521 * - a code tries to modify page's memcg while it's USED.
1522 * - a code tries to modify page state accounting in a memcg.
1523 * see mem_cgroup_stolen(), too.
1525 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1526 unsigned long *flags)
1528 spin_lock_irqsave(&memcg->move_lock, *flags);
1531 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1532 unsigned long *flags)
1534 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1537 #define K(x) ((x) << (PAGE_SHIFT-10))
1539 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1540 * @memcg: The memory cgroup that went over limit
1541 * @p: Task that is going to be killed
1543 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1546 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1548 struct cgroup *task_cgrp;
1549 struct cgroup *mem_cgrp;
1551 * Need a buffer in BSS, can't rely on allocations. The code relies
1552 * on the assumption that OOM is serialized for memory controller.
1553 * If this assumption is broken, revisit this code.
1555 static char memcg_name[PATH_MAX];
1557 struct mem_cgroup *iter;
1565 mem_cgrp = memcg->css.cgroup;
1566 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1568 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1571 * Unfortunately, we are unable to convert to a useful name
1572 * But we'll still print out the usage information
1579 pr_info("Task in %s killed", memcg_name);
1582 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1590 * Continues from above, so we don't need an KERN_ level
1592 pr_cont(" as a result of limit of %s\n", memcg_name);
1595 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1596 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1597 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1598 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1599 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1600 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1601 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1602 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1603 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1604 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1605 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1606 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1608 for_each_mem_cgroup_tree(iter, memcg) {
1609 pr_info("Memory cgroup stats");
1612 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1614 pr_cont(" for %s", memcg_name);
1618 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1619 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1621 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1622 K(mem_cgroup_read_stat(iter, i)));
1625 for (i = 0; i < NR_LRU_LISTS; i++)
1626 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1627 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1634 * This function returns the number of memcg under hierarchy tree. Returns
1635 * 1(self count) if no children.
1637 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1640 struct mem_cgroup *iter;
1642 for_each_mem_cgroup_tree(iter, memcg)
1648 * Return the memory (and swap, if configured) limit for a memcg.
1650 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1654 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1657 * Do not consider swap space if we cannot swap due to swappiness
1659 if (mem_cgroup_swappiness(memcg)) {
1662 limit += total_swap_pages << PAGE_SHIFT;
1663 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1666 * If memsw is finite and limits the amount of swap space
1667 * available to this memcg, return that limit.
1669 limit = min(limit, memsw);
1675 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1678 struct mem_cgroup *iter;
1679 unsigned long chosen_points = 0;
1680 unsigned long totalpages;
1681 unsigned int points = 0;
1682 struct task_struct *chosen = NULL;
1685 * If current has a pending SIGKILL or is exiting, then automatically
1686 * select it. The goal is to allow it to allocate so that it may
1687 * quickly exit and free its memory.
1689 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1690 set_thread_flag(TIF_MEMDIE);
1694 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1695 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1696 for_each_mem_cgroup_tree(iter, memcg) {
1697 struct css_task_iter it;
1698 struct task_struct *task;
1700 css_task_iter_start(&iter->css, &it);
1701 while ((task = css_task_iter_next(&it))) {
1702 switch (oom_scan_process_thread(task, totalpages, NULL,
1704 case OOM_SCAN_SELECT:
1706 put_task_struct(chosen);
1708 chosen_points = ULONG_MAX;
1709 get_task_struct(chosen);
1711 case OOM_SCAN_CONTINUE:
1713 case OOM_SCAN_ABORT:
1714 css_task_iter_end(&it);
1715 mem_cgroup_iter_break(memcg, iter);
1717 put_task_struct(chosen);
1722 points = oom_badness(task, memcg, NULL, totalpages);
1723 if (points > chosen_points) {
1725 put_task_struct(chosen);
1727 chosen_points = points;
1728 get_task_struct(chosen);
1731 css_task_iter_end(&it);
1736 points = chosen_points * 1000 / totalpages;
1737 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1738 NULL, "Memory cgroup out of memory");
1741 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1743 unsigned long flags)
1745 unsigned long total = 0;
1746 bool noswap = false;
1749 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1751 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1754 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1756 drain_all_stock_async(memcg);
1757 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1759 * Allow limit shrinkers, which are triggered directly
1760 * by userspace, to catch signals and stop reclaim
1761 * after minimal progress, regardless of the margin.
1763 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1765 if (mem_cgroup_margin(memcg))
1768 * If nothing was reclaimed after two attempts, there
1769 * may be no reclaimable pages in this hierarchy.
1777 #if MAX_NUMNODES > 1
1779 * test_mem_cgroup_node_reclaimable
1780 * @memcg: the target memcg
1781 * @nid: the node ID to be checked.
1782 * @noswap : specify true here if the user wants flle only information.
1784 * This function returns whether the specified memcg contains any
1785 * reclaimable pages on a node. Returns true if there are any reclaimable
1786 * pages in the node.
1788 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1789 int nid, bool noswap)
1791 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1793 if (noswap || !total_swap_pages)
1795 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1802 * Always updating the nodemask is not very good - even if we have an empty
1803 * list or the wrong list here, we can start from some node and traverse all
1804 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1807 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1811 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1812 * pagein/pageout changes since the last update.
1814 if (!atomic_read(&memcg->numainfo_events))
1816 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1819 /* make a nodemask where this memcg uses memory from */
1820 memcg->scan_nodes = node_states[N_MEMORY];
1822 for_each_node_mask(nid, node_states[N_MEMORY]) {
1824 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1825 node_clear(nid, memcg->scan_nodes);
1828 atomic_set(&memcg->numainfo_events, 0);
1829 atomic_set(&memcg->numainfo_updating, 0);
1833 * Selecting a node where we start reclaim from. Because what we need is just
1834 * reducing usage counter, start from anywhere is O,K. Considering
1835 * memory reclaim from current node, there are pros. and cons.
1837 * Freeing memory from current node means freeing memory from a node which
1838 * we'll use or we've used. So, it may make LRU bad. And if several threads
1839 * hit limits, it will see a contention on a node. But freeing from remote
1840 * node means more costs for memory reclaim because of memory latency.
1842 * Now, we use round-robin. Better algorithm is welcomed.
1844 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1848 mem_cgroup_may_update_nodemask(memcg);
1849 node = memcg->last_scanned_node;
1851 node = next_node(node, memcg->scan_nodes);
1852 if (node == MAX_NUMNODES)
1853 node = first_node(memcg->scan_nodes);
1855 * We call this when we hit limit, not when pages are added to LRU.
1856 * No LRU may hold pages because all pages are UNEVICTABLE or
1857 * memcg is too small and all pages are not on LRU. In that case,
1858 * we use curret node.
1860 if (unlikely(node == MAX_NUMNODES))
1861 node = numa_node_id();
1863 memcg->last_scanned_node = node;
1868 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1876 * A group is eligible for the soft limit reclaim under the given root
1878 * a) it is over its soft limit
1879 * b) any parent up the hierarchy is over its soft limit
1881 * If the given group doesn't have any children over the limit then it
1882 * doesn't make any sense to iterate its subtree.
1884 enum mem_cgroup_filter_t
1885 mem_cgroup_soft_reclaim_eligible(struct mem_cgroup *memcg,
1886 struct mem_cgroup *root)
1888 struct mem_cgroup *parent;
1891 memcg = root_mem_cgroup;
1894 if (res_counter_soft_limit_excess(&memcg->res))
1898 * If any parent up to the root in the hierarchy is over its soft limit
1899 * then we have to obey and reclaim from this group as well.
1901 while((parent = parent_mem_cgroup(parent))) {
1902 if (res_counter_soft_limit_excess(&parent->res))
1908 if (!atomic_read(&memcg->children_in_excess))
1914 * Check OOM-Killer is already running under our hierarchy.
1915 * If someone is running, return false.
1916 * Has to be called with memcg_oom_lock
1918 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1920 struct mem_cgroup *iter, *failed = NULL;
1922 for_each_mem_cgroup_tree(iter, memcg) {
1923 if (iter->oom_lock) {
1925 * this subtree of our hierarchy is already locked
1926 * so we cannot give a lock.
1929 mem_cgroup_iter_break(memcg, iter);
1932 iter->oom_lock = true;
1939 * OK, we failed to lock the whole subtree so we have to clean up
1940 * what we set up to the failing subtree
1942 for_each_mem_cgroup_tree(iter, memcg) {
1943 if (iter == failed) {
1944 mem_cgroup_iter_break(memcg, iter);
1947 iter->oom_lock = false;
1953 * Has to be called with memcg_oom_lock
1955 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1957 struct mem_cgroup *iter;
1959 for_each_mem_cgroup_tree(iter, memcg)
1960 iter->oom_lock = false;
1964 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1966 struct mem_cgroup *iter;
1968 for_each_mem_cgroup_tree(iter, memcg)
1969 atomic_inc(&iter->under_oom);
1972 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1974 struct mem_cgroup *iter;
1977 * When a new child is created while the hierarchy is under oom,
1978 * mem_cgroup_oom_lock() may not be called. We have to use
1979 * atomic_add_unless() here.
1981 for_each_mem_cgroup_tree(iter, memcg)
1982 atomic_add_unless(&iter->under_oom, -1, 0);
1985 static DEFINE_SPINLOCK(memcg_oom_lock);
1986 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1988 struct oom_wait_info {
1989 struct mem_cgroup *memcg;
1993 static int memcg_oom_wake_function(wait_queue_t *wait,
1994 unsigned mode, int sync, void *arg)
1996 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1997 struct mem_cgroup *oom_wait_memcg;
1998 struct oom_wait_info *oom_wait_info;
2000 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2001 oom_wait_memcg = oom_wait_info->memcg;
2004 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2005 * Then we can use css_is_ancestor without taking care of RCU.
2007 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2008 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2010 return autoremove_wake_function(wait, mode, sync, arg);
2013 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2015 /* for filtering, pass "memcg" as argument. */
2016 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2019 static void memcg_oom_recover(struct mem_cgroup *memcg)
2021 if (memcg && atomic_read(&memcg->under_oom))
2022 memcg_wakeup_oom(memcg);
2026 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2028 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2031 struct oom_wait_info owait;
2032 bool locked, need_to_kill;
2034 owait.memcg = memcg;
2035 owait.wait.flags = 0;
2036 owait.wait.func = memcg_oom_wake_function;
2037 owait.wait.private = current;
2038 INIT_LIST_HEAD(&owait.wait.task_list);
2039 need_to_kill = true;
2040 mem_cgroup_mark_under_oom(memcg);
2042 /* At first, try to OOM lock hierarchy under memcg.*/
2043 spin_lock(&memcg_oom_lock);
2044 locked = mem_cgroup_oom_lock(memcg);
2046 * Even if signal_pending(), we can't quit charge() loop without
2047 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2048 * under OOM is always welcomed, use TASK_KILLABLE here.
2050 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2051 if (!locked || memcg->oom_kill_disable)
2052 need_to_kill = false;
2054 mem_cgroup_oom_notify(memcg);
2055 spin_unlock(&memcg_oom_lock);
2058 finish_wait(&memcg_oom_waitq, &owait.wait);
2059 mem_cgroup_out_of_memory(memcg, mask, order);
2062 finish_wait(&memcg_oom_waitq, &owait.wait);
2064 spin_lock(&memcg_oom_lock);
2066 mem_cgroup_oom_unlock(memcg);
2067 memcg_wakeup_oom(memcg);
2068 spin_unlock(&memcg_oom_lock);
2070 mem_cgroup_unmark_under_oom(memcg);
2072 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2074 /* Give chance to dying process */
2075 schedule_timeout_uninterruptible(1);
2080 * Currently used to update mapped file statistics, but the routine can be
2081 * generalized to update other statistics as well.
2083 * Notes: Race condition
2085 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2086 * it tends to be costly. But considering some conditions, we doesn't need
2087 * to do so _always_.
2089 * Considering "charge", lock_page_cgroup() is not required because all
2090 * file-stat operations happen after a page is attached to radix-tree. There
2091 * are no race with "charge".
2093 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2094 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2095 * if there are race with "uncharge". Statistics itself is properly handled
2098 * Considering "move", this is an only case we see a race. To make the race
2099 * small, we check mm->moving_account and detect there are possibility of race
2100 * If there is, we take a lock.
2103 void __mem_cgroup_begin_update_page_stat(struct page *page,
2104 bool *locked, unsigned long *flags)
2106 struct mem_cgroup *memcg;
2107 struct page_cgroup *pc;
2109 pc = lookup_page_cgroup(page);
2111 memcg = pc->mem_cgroup;
2112 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2115 * If this memory cgroup is not under account moving, we don't
2116 * need to take move_lock_mem_cgroup(). Because we already hold
2117 * rcu_read_lock(), any calls to move_account will be delayed until
2118 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2120 if (!mem_cgroup_stolen(memcg))
2123 move_lock_mem_cgroup(memcg, flags);
2124 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2125 move_unlock_mem_cgroup(memcg, flags);
2131 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2133 struct page_cgroup *pc = lookup_page_cgroup(page);
2136 * It's guaranteed that pc->mem_cgroup never changes while
2137 * lock is held because a routine modifies pc->mem_cgroup
2138 * should take move_lock_mem_cgroup().
2140 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2143 void mem_cgroup_update_page_stat(struct page *page,
2144 enum mem_cgroup_page_stat_item idx, int val)
2146 struct mem_cgroup *memcg;
2147 struct page_cgroup *pc = lookup_page_cgroup(page);
2148 unsigned long uninitialized_var(flags);
2150 if (mem_cgroup_disabled())
2153 memcg = pc->mem_cgroup;
2154 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2158 case MEMCG_NR_FILE_MAPPED:
2159 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2165 this_cpu_add(memcg->stat->count[idx], val);
2169 * size of first charge trial. "32" comes from vmscan.c's magic value.
2170 * TODO: maybe necessary to use big numbers in big irons.
2172 #define CHARGE_BATCH 32U
2173 struct memcg_stock_pcp {
2174 struct mem_cgroup *cached; /* this never be root cgroup */
2175 unsigned int nr_pages;
2176 struct work_struct work;
2177 unsigned long flags;
2178 #define FLUSHING_CACHED_CHARGE 0
2180 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2181 static DEFINE_MUTEX(percpu_charge_mutex);
2184 * consume_stock: Try to consume stocked charge on this cpu.
2185 * @memcg: memcg to consume from.
2186 * @nr_pages: how many pages to charge.
2188 * The charges will only happen if @memcg matches the current cpu's memcg
2189 * stock, and at least @nr_pages are available in that stock. Failure to
2190 * service an allocation will refill the stock.
2192 * returns true if successful, false otherwise.
2194 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2196 struct memcg_stock_pcp *stock;
2199 if (nr_pages > CHARGE_BATCH)
2202 stock = &get_cpu_var(memcg_stock);
2203 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2204 stock->nr_pages -= nr_pages;
2205 else /* need to call res_counter_charge */
2207 put_cpu_var(memcg_stock);
2212 * Returns stocks cached in percpu to res_counter and reset cached information.
2214 static void drain_stock(struct memcg_stock_pcp *stock)
2216 struct mem_cgroup *old = stock->cached;
2218 if (stock->nr_pages) {
2219 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2221 res_counter_uncharge(&old->res, bytes);
2222 if (do_swap_account)
2223 res_counter_uncharge(&old->memsw, bytes);
2224 stock->nr_pages = 0;
2226 stock->cached = NULL;
2230 * This must be called under preempt disabled or must be called by
2231 * a thread which is pinned to local cpu.
2233 static void drain_local_stock(struct work_struct *dummy)
2235 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2237 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2240 static void __init memcg_stock_init(void)
2244 for_each_possible_cpu(cpu) {
2245 struct memcg_stock_pcp *stock =
2246 &per_cpu(memcg_stock, cpu);
2247 INIT_WORK(&stock->work, drain_local_stock);
2252 * Cache charges(val) which is from res_counter, to local per_cpu area.
2253 * This will be consumed by consume_stock() function, later.
2255 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2257 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2259 if (stock->cached != memcg) { /* reset if necessary */
2261 stock->cached = memcg;
2263 stock->nr_pages += nr_pages;
2264 put_cpu_var(memcg_stock);
2268 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2269 * of the hierarchy under it. sync flag says whether we should block
2270 * until the work is done.
2272 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2276 /* Notify other cpus that system-wide "drain" is running */
2279 for_each_online_cpu(cpu) {
2280 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2281 struct mem_cgroup *memcg;
2283 memcg = stock->cached;
2284 if (!memcg || !stock->nr_pages)
2286 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2288 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2290 drain_local_stock(&stock->work);
2292 schedule_work_on(cpu, &stock->work);
2300 for_each_online_cpu(cpu) {
2301 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2302 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2303 flush_work(&stock->work);
2310 * Tries to drain stocked charges in other cpus. This function is asynchronous
2311 * and just put a work per cpu for draining localy on each cpu. Caller can
2312 * expects some charges will be back to res_counter later but cannot wait for
2315 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2318 * If someone calls draining, avoid adding more kworker runs.
2320 if (!mutex_trylock(&percpu_charge_mutex))
2322 drain_all_stock(root_memcg, false);
2323 mutex_unlock(&percpu_charge_mutex);
2326 /* This is a synchronous drain interface. */
2327 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2329 /* called when force_empty is called */
2330 mutex_lock(&percpu_charge_mutex);
2331 drain_all_stock(root_memcg, true);
2332 mutex_unlock(&percpu_charge_mutex);
2336 * This function drains percpu counter value from DEAD cpu and
2337 * move it to local cpu. Note that this function can be preempted.
2339 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2343 spin_lock(&memcg->pcp_counter_lock);
2344 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2345 long x = per_cpu(memcg->stat->count[i], cpu);
2347 per_cpu(memcg->stat->count[i], cpu) = 0;
2348 memcg->nocpu_base.count[i] += x;
2350 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2351 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2353 per_cpu(memcg->stat->events[i], cpu) = 0;
2354 memcg->nocpu_base.events[i] += x;
2356 spin_unlock(&memcg->pcp_counter_lock);
2359 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2360 unsigned long action,
2363 int cpu = (unsigned long)hcpu;
2364 struct memcg_stock_pcp *stock;
2365 struct mem_cgroup *iter;
2367 if (action == CPU_ONLINE)
2370 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2373 for_each_mem_cgroup(iter)
2374 mem_cgroup_drain_pcp_counter(iter, cpu);
2376 stock = &per_cpu(memcg_stock, cpu);
2382 /* See __mem_cgroup_try_charge() for details */
2384 CHARGE_OK, /* success */
2385 CHARGE_RETRY, /* need to retry but retry is not bad */
2386 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2387 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2388 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2391 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2392 unsigned int nr_pages, unsigned int min_pages,
2395 unsigned long csize = nr_pages * PAGE_SIZE;
2396 struct mem_cgroup *mem_over_limit;
2397 struct res_counter *fail_res;
2398 unsigned long flags = 0;
2401 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2404 if (!do_swap_account)
2406 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2410 res_counter_uncharge(&memcg->res, csize);
2411 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2412 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2414 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2416 * Never reclaim on behalf of optional batching, retry with a
2417 * single page instead.
2419 if (nr_pages > min_pages)
2420 return CHARGE_RETRY;
2422 if (!(gfp_mask & __GFP_WAIT))
2423 return CHARGE_WOULDBLOCK;
2425 if (gfp_mask & __GFP_NORETRY)
2426 return CHARGE_NOMEM;
2428 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2429 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2430 return CHARGE_RETRY;
2432 * Even though the limit is exceeded at this point, reclaim
2433 * may have been able to free some pages. Retry the charge
2434 * before killing the task.
2436 * Only for regular pages, though: huge pages are rather
2437 * unlikely to succeed so close to the limit, and we fall back
2438 * to regular pages anyway in case of failure.
2440 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2441 return CHARGE_RETRY;
2444 * At task move, charge accounts can be doubly counted. So, it's
2445 * better to wait until the end of task_move if something is going on.
2447 if (mem_cgroup_wait_acct_move(mem_over_limit))
2448 return CHARGE_RETRY;
2450 /* If we don't need to call oom-killer at el, return immediately */
2452 return CHARGE_NOMEM;
2454 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2455 return CHARGE_OOM_DIE;
2457 return CHARGE_RETRY;
2461 * __mem_cgroup_try_charge() does
2462 * 1. detect memcg to be charged against from passed *mm and *ptr,
2463 * 2. update res_counter
2464 * 3. call memory reclaim if necessary.
2466 * In some special case, if the task is fatal, fatal_signal_pending() or
2467 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2468 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2469 * as possible without any hazards. 2: all pages should have a valid
2470 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2471 * pointer, that is treated as a charge to root_mem_cgroup.
2473 * So __mem_cgroup_try_charge() will return
2474 * 0 ... on success, filling *ptr with a valid memcg pointer.
2475 * -ENOMEM ... charge failure because of resource limits.
2476 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2478 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2479 * the oom-killer can be invoked.
2481 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2483 unsigned int nr_pages,
2484 struct mem_cgroup **ptr,
2487 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2488 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2489 struct mem_cgroup *memcg = NULL;
2493 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2494 * in system level. So, allow to go ahead dying process in addition to
2497 if (unlikely(test_thread_flag(TIF_MEMDIE)
2498 || fatal_signal_pending(current)))
2502 * We always charge the cgroup the mm_struct belongs to.
2503 * The mm_struct's mem_cgroup changes on task migration if the
2504 * thread group leader migrates. It's possible that mm is not
2505 * set, if so charge the root memcg (happens for pagecache usage).
2508 *ptr = root_mem_cgroup;
2510 if (*ptr) { /* css should be a valid one */
2512 if (mem_cgroup_is_root(memcg))
2514 if (consume_stock(memcg, nr_pages))
2516 css_get(&memcg->css);
2518 struct task_struct *p;
2521 p = rcu_dereference(mm->owner);
2523 * Because we don't have task_lock(), "p" can exit.
2524 * In that case, "memcg" can point to root or p can be NULL with
2525 * race with swapoff. Then, we have small risk of mis-accouning.
2526 * But such kind of mis-account by race always happens because
2527 * we don't have cgroup_mutex(). It's overkill and we allo that
2529 * (*) swapoff at el will charge against mm-struct not against
2530 * task-struct. So, mm->owner can be NULL.
2532 memcg = mem_cgroup_from_task(p);
2534 memcg = root_mem_cgroup;
2535 if (mem_cgroup_is_root(memcg)) {
2539 if (consume_stock(memcg, nr_pages)) {
2541 * It seems dagerous to access memcg without css_get().
2542 * But considering how consume_stok works, it's not
2543 * necessary. If consume_stock success, some charges
2544 * from this memcg are cached on this cpu. So, we
2545 * don't need to call css_get()/css_tryget() before
2546 * calling consume_stock().
2551 /* after here, we may be blocked. we need to get refcnt */
2552 if (!css_tryget(&memcg->css)) {
2562 /* If killed, bypass charge */
2563 if (fatal_signal_pending(current)) {
2564 css_put(&memcg->css);
2569 if (oom && !nr_oom_retries) {
2571 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2574 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2579 case CHARGE_RETRY: /* not in OOM situation but retry */
2581 css_put(&memcg->css);
2584 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2585 css_put(&memcg->css);
2587 case CHARGE_NOMEM: /* OOM routine works */
2589 css_put(&memcg->css);
2592 /* If oom, we never return -ENOMEM */
2595 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2596 css_put(&memcg->css);
2599 } while (ret != CHARGE_OK);
2601 if (batch > nr_pages)
2602 refill_stock(memcg, batch - nr_pages);
2603 css_put(&memcg->css);
2611 *ptr = root_mem_cgroup;
2616 * Somemtimes we have to undo a charge we got by try_charge().
2617 * This function is for that and do uncharge, put css's refcnt.
2618 * gotten by try_charge().
2620 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2621 unsigned int nr_pages)
2623 if (!mem_cgroup_is_root(memcg)) {
2624 unsigned long bytes = nr_pages * PAGE_SIZE;
2626 res_counter_uncharge(&memcg->res, bytes);
2627 if (do_swap_account)
2628 res_counter_uncharge(&memcg->memsw, bytes);
2633 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2634 * This is useful when moving usage to parent cgroup.
2636 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2637 unsigned int nr_pages)
2639 unsigned long bytes = nr_pages * PAGE_SIZE;
2641 if (mem_cgroup_is_root(memcg))
2644 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2645 if (do_swap_account)
2646 res_counter_uncharge_until(&memcg->memsw,
2647 memcg->memsw.parent, bytes);
2651 * A helper function to get mem_cgroup from ID. must be called under
2652 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2653 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2654 * called against removed memcg.)
2656 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2658 struct cgroup_subsys_state *css;
2660 /* ID 0 is unused ID */
2663 css = css_lookup(&mem_cgroup_subsys, id);
2666 return mem_cgroup_from_css(css);
2669 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2671 struct mem_cgroup *memcg = NULL;
2672 struct page_cgroup *pc;
2676 VM_BUG_ON(!PageLocked(page));
2678 pc = lookup_page_cgroup(page);
2679 lock_page_cgroup(pc);
2680 if (PageCgroupUsed(pc)) {
2681 memcg = pc->mem_cgroup;
2682 if (memcg && !css_tryget(&memcg->css))
2684 } else if (PageSwapCache(page)) {
2685 ent.val = page_private(page);
2686 id = lookup_swap_cgroup_id(ent);
2688 memcg = mem_cgroup_lookup(id);
2689 if (memcg && !css_tryget(&memcg->css))
2693 unlock_page_cgroup(pc);
2697 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2699 unsigned int nr_pages,
2700 enum charge_type ctype,
2703 struct page_cgroup *pc = lookup_page_cgroup(page);
2704 struct zone *uninitialized_var(zone);
2705 struct lruvec *lruvec;
2706 bool was_on_lru = false;
2709 lock_page_cgroup(pc);
2710 VM_BUG_ON(PageCgroupUsed(pc));
2712 * we don't need page_cgroup_lock about tail pages, becase they are not
2713 * accessed by any other context at this point.
2717 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2718 * may already be on some other mem_cgroup's LRU. Take care of it.
2721 zone = page_zone(page);
2722 spin_lock_irq(&zone->lru_lock);
2723 if (PageLRU(page)) {
2724 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2726 del_page_from_lru_list(page, lruvec, page_lru(page));
2731 pc->mem_cgroup = memcg;
2733 * We access a page_cgroup asynchronously without lock_page_cgroup().
2734 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2735 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2736 * before USED bit, we need memory barrier here.
2737 * See mem_cgroup_add_lru_list(), etc.
2740 SetPageCgroupUsed(pc);
2744 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2745 VM_BUG_ON(PageLRU(page));
2747 add_page_to_lru_list(page, lruvec, page_lru(page));
2749 spin_unlock_irq(&zone->lru_lock);
2752 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2757 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2758 unlock_page_cgroup(pc);
2761 * "charge_statistics" updated event counter.
2763 memcg_check_events(memcg, page);
2766 static DEFINE_MUTEX(set_limit_mutex);
2768 #ifdef CONFIG_MEMCG_KMEM
2769 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2771 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2772 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2776 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2777 * in the memcg_cache_params struct.
2779 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2781 struct kmem_cache *cachep;
2783 VM_BUG_ON(p->is_root_cache);
2784 cachep = p->root_cache;
2785 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2788 #ifdef CONFIG_SLABINFO
2789 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2790 struct cftype *cft, struct seq_file *m)
2792 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2793 struct memcg_cache_params *params;
2795 if (!memcg_can_account_kmem(memcg))
2798 print_slabinfo_header(m);
2800 mutex_lock(&memcg->slab_caches_mutex);
2801 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2802 cache_show(memcg_params_to_cache(params), m);
2803 mutex_unlock(&memcg->slab_caches_mutex);
2809 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2811 struct res_counter *fail_res;
2812 struct mem_cgroup *_memcg;
2816 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2821 * Conditions under which we can wait for the oom_killer. Those are
2822 * the same conditions tested by the core page allocator
2824 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2827 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2830 if (ret == -EINTR) {
2832 * __mem_cgroup_try_charge() chosed to bypass to root due to
2833 * OOM kill or fatal signal. Since our only options are to
2834 * either fail the allocation or charge it to this cgroup, do
2835 * it as a temporary condition. But we can't fail. From a
2836 * kmem/slab perspective, the cache has already been selected,
2837 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2840 * This condition will only trigger if the task entered
2841 * memcg_charge_kmem in a sane state, but was OOM-killed during
2842 * __mem_cgroup_try_charge() above. Tasks that were already
2843 * dying when the allocation triggers should have been already
2844 * directed to the root cgroup in memcontrol.h
2846 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2847 if (do_swap_account)
2848 res_counter_charge_nofail(&memcg->memsw, size,
2852 res_counter_uncharge(&memcg->kmem, size);
2857 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2859 res_counter_uncharge(&memcg->res, size);
2860 if (do_swap_account)
2861 res_counter_uncharge(&memcg->memsw, size);
2864 if (res_counter_uncharge(&memcg->kmem, size))
2868 * Releases a reference taken in kmem_cgroup_css_offline in case
2869 * this last uncharge is racing with the offlining code or it is
2870 * outliving the memcg existence.
2872 * The memory barrier imposed by test&clear is paired with the
2873 * explicit one in memcg_kmem_mark_dead().
2875 if (memcg_kmem_test_and_clear_dead(memcg))
2876 css_put(&memcg->css);
2879 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2884 mutex_lock(&memcg->slab_caches_mutex);
2885 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2886 mutex_unlock(&memcg->slab_caches_mutex);
2890 * helper for acessing a memcg's index. It will be used as an index in the
2891 * child cache array in kmem_cache, and also to derive its name. This function
2892 * will return -1 when this is not a kmem-limited memcg.
2894 int memcg_cache_id(struct mem_cgroup *memcg)
2896 return memcg ? memcg->kmemcg_id : -1;
2900 * This ends up being protected by the set_limit mutex, during normal
2901 * operation, because that is its main call site.
2903 * But when we create a new cache, we can call this as well if its parent
2904 * is kmem-limited. That will have to hold set_limit_mutex as well.
2906 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2910 num = ida_simple_get(&kmem_limited_groups,
2911 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2915 * After this point, kmem_accounted (that we test atomically in
2916 * the beginning of this conditional), is no longer 0. This
2917 * guarantees only one process will set the following boolean
2918 * to true. We don't need test_and_set because we're protected
2919 * by the set_limit_mutex anyway.
2921 memcg_kmem_set_activated(memcg);
2923 ret = memcg_update_all_caches(num+1);
2925 ida_simple_remove(&kmem_limited_groups, num);
2926 memcg_kmem_clear_activated(memcg);
2930 memcg->kmemcg_id = num;
2931 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2932 mutex_init(&memcg->slab_caches_mutex);
2936 static size_t memcg_caches_array_size(int num_groups)
2939 if (num_groups <= 0)
2942 size = 2 * num_groups;
2943 if (size < MEMCG_CACHES_MIN_SIZE)
2944 size = MEMCG_CACHES_MIN_SIZE;
2945 else if (size > MEMCG_CACHES_MAX_SIZE)
2946 size = MEMCG_CACHES_MAX_SIZE;
2952 * We should update the current array size iff all caches updates succeed. This
2953 * can only be done from the slab side. The slab mutex needs to be held when
2956 void memcg_update_array_size(int num)
2958 if (num > memcg_limited_groups_array_size)
2959 memcg_limited_groups_array_size = memcg_caches_array_size(num);
2962 static void kmem_cache_destroy_work_func(struct work_struct *w);
2964 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
2966 struct memcg_cache_params *cur_params = s->memcg_params;
2968 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
2970 if (num_groups > memcg_limited_groups_array_size) {
2972 ssize_t size = memcg_caches_array_size(num_groups);
2974 size *= sizeof(void *);
2975 size += offsetof(struct memcg_cache_params, memcg_caches);
2977 s->memcg_params = kzalloc(size, GFP_KERNEL);
2978 if (!s->memcg_params) {
2979 s->memcg_params = cur_params;
2983 s->memcg_params->is_root_cache = true;
2986 * There is the chance it will be bigger than
2987 * memcg_limited_groups_array_size, if we failed an allocation
2988 * in a cache, in which case all caches updated before it, will
2989 * have a bigger array.
2991 * But if that is the case, the data after
2992 * memcg_limited_groups_array_size is certainly unused
2994 for (i = 0; i < memcg_limited_groups_array_size; i++) {
2995 if (!cur_params->memcg_caches[i])
2997 s->memcg_params->memcg_caches[i] =
2998 cur_params->memcg_caches[i];
3002 * Ideally, we would wait until all caches succeed, and only
3003 * then free the old one. But this is not worth the extra
3004 * pointer per-cache we'd have to have for this.
3006 * It is not a big deal if some caches are left with a size
3007 * bigger than the others. And all updates will reset this
3015 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3016 struct kmem_cache *root_cache)
3020 if (!memcg_kmem_enabled())
3024 size = offsetof(struct memcg_cache_params, memcg_caches);
3025 size += memcg_limited_groups_array_size * sizeof(void *);
3027 size = sizeof(struct memcg_cache_params);
3029 s->memcg_params = kzalloc(size, GFP_KERNEL);
3030 if (!s->memcg_params)
3034 s->memcg_params->memcg = memcg;
3035 s->memcg_params->root_cache = root_cache;
3036 INIT_WORK(&s->memcg_params->destroy,
3037 kmem_cache_destroy_work_func);
3039 s->memcg_params->is_root_cache = true;
3044 void memcg_release_cache(struct kmem_cache *s)
3046 struct kmem_cache *root;
3047 struct mem_cgroup *memcg;
3051 * This happens, for instance, when a root cache goes away before we
3054 if (!s->memcg_params)
3057 if (s->memcg_params->is_root_cache)
3060 memcg = s->memcg_params->memcg;
3061 id = memcg_cache_id(memcg);
3063 root = s->memcg_params->root_cache;
3064 root->memcg_params->memcg_caches[id] = NULL;
3066 mutex_lock(&memcg->slab_caches_mutex);
3067 list_del(&s->memcg_params->list);
3068 mutex_unlock(&memcg->slab_caches_mutex);
3070 css_put(&memcg->css);
3072 kfree(s->memcg_params);
3076 * During the creation a new cache, we need to disable our accounting mechanism
3077 * altogether. This is true even if we are not creating, but rather just
3078 * enqueing new caches to be created.
3080 * This is because that process will trigger allocations; some visible, like
3081 * explicit kmallocs to auxiliary data structures, name strings and internal
3082 * cache structures; some well concealed, like INIT_WORK() that can allocate
3083 * objects during debug.
3085 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3086 * to it. This may not be a bounded recursion: since the first cache creation
3087 * failed to complete (waiting on the allocation), we'll just try to create the
3088 * cache again, failing at the same point.
3090 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3091 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3092 * inside the following two functions.
3094 static inline void memcg_stop_kmem_account(void)
3096 VM_BUG_ON(!current->mm);
3097 current->memcg_kmem_skip_account++;
3100 static inline void memcg_resume_kmem_account(void)
3102 VM_BUG_ON(!current->mm);
3103 current->memcg_kmem_skip_account--;
3106 static void kmem_cache_destroy_work_func(struct work_struct *w)
3108 struct kmem_cache *cachep;
3109 struct memcg_cache_params *p;
3111 p = container_of(w, struct memcg_cache_params, destroy);
3113 cachep = memcg_params_to_cache(p);
3116 * If we get down to 0 after shrink, we could delete right away.
3117 * However, memcg_release_pages() already puts us back in the workqueue
3118 * in that case. If we proceed deleting, we'll get a dangling
3119 * reference, and removing the object from the workqueue in that case
3120 * is unnecessary complication. We are not a fast path.
3122 * Note that this case is fundamentally different from racing with
3123 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3124 * kmem_cache_shrink, not only we would be reinserting a dead cache
3125 * into the queue, but doing so from inside the worker racing to
3128 * So if we aren't down to zero, we'll just schedule a worker and try
3131 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3132 kmem_cache_shrink(cachep);
3133 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3136 kmem_cache_destroy(cachep);
3139 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3141 if (!cachep->memcg_params->dead)
3145 * There are many ways in which we can get here.
3147 * We can get to a memory-pressure situation while the delayed work is
3148 * still pending to run. The vmscan shrinkers can then release all
3149 * cache memory and get us to destruction. If this is the case, we'll
3150 * be executed twice, which is a bug (the second time will execute over
3151 * bogus data). In this case, cancelling the work should be fine.
3153 * But we can also get here from the worker itself, if
3154 * kmem_cache_shrink is enough to shake all the remaining objects and
3155 * get the page count to 0. In this case, we'll deadlock if we try to
3156 * cancel the work (the worker runs with an internal lock held, which
3157 * is the same lock we would hold for cancel_work_sync().)
3159 * Since we can't possibly know who got us here, just refrain from
3160 * running if there is already work pending
3162 if (work_pending(&cachep->memcg_params->destroy))
3165 * We have to defer the actual destroying to a workqueue, because
3166 * we might currently be in a context that cannot sleep.
3168 schedule_work(&cachep->memcg_params->destroy);
3172 * This lock protects updaters, not readers. We want readers to be as fast as
3173 * they can, and they will either see NULL or a valid cache value. Our model
3174 * allow them to see NULL, in which case the root memcg will be selected.
3176 * We need this lock because multiple allocations to the same cache from a non
3177 * will span more than one worker. Only one of them can create the cache.
3179 static DEFINE_MUTEX(memcg_cache_mutex);
3182 * Called with memcg_cache_mutex held
3184 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3185 struct kmem_cache *s)
3187 struct kmem_cache *new;
3188 static char *tmp_name = NULL;
3190 lockdep_assert_held(&memcg_cache_mutex);
3193 * kmem_cache_create_memcg duplicates the given name and
3194 * cgroup_name for this name requires RCU context.
3195 * This static temporary buffer is used to prevent from
3196 * pointless shortliving allocation.
3199 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3205 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3206 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3209 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3210 (s->flags & ~SLAB_PANIC), s->ctor, s);
3213 new->allocflags |= __GFP_KMEMCG;
3218 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3219 struct kmem_cache *cachep)
3221 struct kmem_cache *new_cachep;
3224 BUG_ON(!memcg_can_account_kmem(memcg));
3226 idx = memcg_cache_id(memcg);
3228 mutex_lock(&memcg_cache_mutex);
3229 new_cachep = cachep->memcg_params->memcg_caches[idx];
3231 css_put(&memcg->css);
3235 new_cachep = kmem_cache_dup(memcg, cachep);
3236 if (new_cachep == NULL) {
3237 new_cachep = cachep;
3238 css_put(&memcg->css);
3242 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3244 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3246 * the readers won't lock, make sure everybody sees the updated value,
3247 * so they won't put stuff in the queue again for no reason
3251 mutex_unlock(&memcg_cache_mutex);
3255 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3257 struct kmem_cache *c;
3260 if (!s->memcg_params)
3262 if (!s->memcg_params->is_root_cache)
3266 * If the cache is being destroyed, we trust that there is no one else
3267 * requesting objects from it. Even if there are, the sanity checks in
3268 * kmem_cache_destroy should caught this ill-case.
3270 * Still, we don't want anyone else freeing memcg_caches under our
3271 * noses, which can happen if a new memcg comes to life. As usual,
3272 * we'll take the set_limit_mutex to protect ourselves against this.
3274 mutex_lock(&set_limit_mutex);
3275 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3276 c = s->memcg_params->memcg_caches[i];
3281 * We will now manually delete the caches, so to avoid races
3282 * we need to cancel all pending destruction workers and
3283 * proceed with destruction ourselves.
3285 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3286 * and that could spawn the workers again: it is likely that
3287 * the cache still have active pages until this very moment.
3288 * This would lead us back to mem_cgroup_destroy_cache.
3290 * But that will not execute at all if the "dead" flag is not
3291 * set, so flip it down to guarantee we are in control.
3293 c->memcg_params->dead = false;
3294 cancel_work_sync(&c->memcg_params->destroy);
3295 kmem_cache_destroy(c);
3297 mutex_unlock(&set_limit_mutex);
3300 struct create_work {
3301 struct mem_cgroup *memcg;
3302 struct kmem_cache *cachep;
3303 struct work_struct work;
3306 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3308 struct kmem_cache *cachep;
3309 struct memcg_cache_params *params;
3311 if (!memcg_kmem_is_active(memcg))
3314 mutex_lock(&memcg->slab_caches_mutex);
3315 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3316 cachep = memcg_params_to_cache(params);
3317 cachep->memcg_params->dead = true;
3318 schedule_work(&cachep->memcg_params->destroy);
3320 mutex_unlock(&memcg->slab_caches_mutex);
3323 static void memcg_create_cache_work_func(struct work_struct *w)
3325 struct create_work *cw;
3327 cw = container_of(w, struct create_work, work);
3328 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3333 * Enqueue the creation of a per-memcg kmem_cache.
3335 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3336 struct kmem_cache *cachep)
3338 struct create_work *cw;
3340 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3342 css_put(&memcg->css);
3347 cw->cachep = cachep;
3349 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3350 schedule_work(&cw->work);
3353 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3354 struct kmem_cache *cachep)
3357 * We need to stop accounting when we kmalloc, because if the
3358 * corresponding kmalloc cache is not yet created, the first allocation
3359 * in __memcg_create_cache_enqueue will recurse.
3361 * However, it is better to enclose the whole function. Depending on
3362 * the debugging options enabled, INIT_WORK(), for instance, can
3363 * trigger an allocation. This too, will make us recurse. Because at
3364 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3365 * the safest choice is to do it like this, wrapping the whole function.
3367 memcg_stop_kmem_account();
3368 __memcg_create_cache_enqueue(memcg, cachep);
3369 memcg_resume_kmem_account();
3372 * Return the kmem_cache we're supposed to use for a slab allocation.
3373 * We try to use the current memcg's version of the cache.
3375 * If the cache does not exist yet, if we are the first user of it,
3376 * we either create it immediately, if possible, or create it asynchronously
3378 * In the latter case, we will let the current allocation go through with
3379 * the original cache.
3381 * Can't be called in interrupt context or from kernel threads.
3382 * This function needs to be called with rcu_read_lock() held.
3384 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3387 struct mem_cgroup *memcg;
3390 VM_BUG_ON(!cachep->memcg_params);
3391 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3393 if (!current->mm || current->memcg_kmem_skip_account)
3397 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3399 if (!memcg_can_account_kmem(memcg))
3402 idx = memcg_cache_id(memcg);
3405 * barrier to mare sure we're always seeing the up to date value. The
3406 * code updating memcg_caches will issue a write barrier to match this.
3408 read_barrier_depends();
3409 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3410 cachep = cachep->memcg_params->memcg_caches[idx];
3414 /* The corresponding put will be done in the workqueue. */
3415 if (!css_tryget(&memcg->css))
3420 * If we are in a safe context (can wait, and not in interrupt
3421 * context), we could be be predictable and return right away.
3422 * This would guarantee that the allocation being performed
3423 * already belongs in the new cache.
3425 * However, there are some clashes that can arrive from locking.
3426 * For instance, because we acquire the slab_mutex while doing
3427 * kmem_cache_dup, this means no further allocation could happen
3428 * with the slab_mutex held.
3430 * Also, because cache creation issue get_online_cpus(), this
3431 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3432 * that ends up reversed during cpu hotplug. (cpuset allocates
3433 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3434 * better to defer everything.
3436 memcg_create_cache_enqueue(memcg, cachep);
3442 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3445 * We need to verify if the allocation against current->mm->owner's memcg is
3446 * possible for the given order. But the page is not allocated yet, so we'll
3447 * need a further commit step to do the final arrangements.
3449 * It is possible for the task to switch cgroups in this mean time, so at
3450 * commit time, we can't rely on task conversion any longer. We'll then use
3451 * the handle argument to return to the caller which cgroup we should commit
3452 * against. We could also return the memcg directly and avoid the pointer
3453 * passing, but a boolean return value gives better semantics considering
3454 * the compiled-out case as well.
3456 * Returning true means the allocation is possible.
3459 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3461 struct mem_cgroup *memcg;
3467 * Disabling accounting is only relevant for some specific memcg
3468 * internal allocations. Therefore we would initially not have such
3469 * check here, since direct calls to the page allocator that are marked
3470 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3471 * concerned with cache allocations, and by having this test at
3472 * memcg_kmem_get_cache, we are already able to relay the allocation to
3473 * the root cache and bypass the memcg cache altogether.
3475 * There is one exception, though: the SLUB allocator does not create
3476 * large order caches, but rather service large kmallocs directly from
3477 * the page allocator. Therefore, the following sequence when backed by
3478 * the SLUB allocator:
3480 * memcg_stop_kmem_account();
3481 * kmalloc(<large_number>)
3482 * memcg_resume_kmem_account();
3484 * would effectively ignore the fact that we should skip accounting,
3485 * since it will drive us directly to this function without passing
3486 * through the cache selector memcg_kmem_get_cache. Such large
3487 * allocations are extremely rare but can happen, for instance, for the
3488 * cache arrays. We bring this test here.
3490 if (!current->mm || current->memcg_kmem_skip_account)
3493 memcg = try_get_mem_cgroup_from_mm(current->mm);
3496 * very rare case described in mem_cgroup_from_task. Unfortunately there
3497 * isn't much we can do without complicating this too much, and it would
3498 * be gfp-dependent anyway. Just let it go
3500 if (unlikely(!memcg))
3503 if (!memcg_can_account_kmem(memcg)) {
3504 css_put(&memcg->css);
3508 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3512 css_put(&memcg->css);
3516 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3519 struct page_cgroup *pc;
3521 VM_BUG_ON(mem_cgroup_is_root(memcg));
3523 /* The page allocation failed. Revert */
3525 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3529 pc = lookup_page_cgroup(page);
3530 lock_page_cgroup(pc);
3531 pc->mem_cgroup = memcg;
3532 SetPageCgroupUsed(pc);
3533 unlock_page_cgroup(pc);
3536 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3538 struct mem_cgroup *memcg = NULL;
3539 struct page_cgroup *pc;
3542 pc = lookup_page_cgroup(page);
3544 * Fast unlocked return. Theoretically might have changed, have to
3545 * check again after locking.
3547 if (!PageCgroupUsed(pc))
3550 lock_page_cgroup(pc);
3551 if (PageCgroupUsed(pc)) {
3552 memcg = pc->mem_cgroup;
3553 ClearPageCgroupUsed(pc);
3555 unlock_page_cgroup(pc);
3558 * We trust that only if there is a memcg associated with the page, it
3559 * is a valid allocation
3564 VM_BUG_ON(mem_cgroup_is_root(memcg));
3565 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3568 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3571 #endif /* CONFIG_MEMCG_KMEM */
3573 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3575 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3577 * Because tail pages are not marked as "used", set it. We're under
3578 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3579 * charge/uncharge will be never happen and move_account() is done under
3580 * compound_lock(), so we don't have to take care of races.
3582 void mem_cgroup_split_huge_fixup(struct page *head)
3584 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3585 struct page_cgroup *pc;
3586 struct mem_cgroup *memcg;
3589 if (mem_cgroup_disabled())
3592 memcg = head_pc->mem_cgroup;
3593 for (i = 1; i < HPAGE_PMD_NR; i++) {
3595 pc->mem_cgroup = memcg;
3596 smp_wmb();/* see __commit_charge() */
3597 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3599 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3602 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3605 * mem_cgroup_move_account - move account of the page
3607 * @nr_pages: number of regular pages (>1 for huge pages)
3608 * @pc: page_cgroup of the page.
3609 * @from: mem_cgroup which the page is moved from.
3610 * @to: mem_cgroup which the page is moved to. @from != @to.
3612 * The caller must confirm following.
3613 * - page is not on LRU (isolate_page() is useful.)
3614 * - compound_lock is held when nr_pages > 1
3616 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3619 static int mem_cgroup_move_account(struct page *page,
3620 unsigned int nr_pages,
3621 struct page_cgroup *pc,
3622 struct mem_cgroup *from,
3623 struct mem_cgroup *to)
3625 unsigned long flags;
3627 bool anon = PageAnon(page);
3629 VM_BUG_ON(from == to);
3630 VM_BUG_ON(PageLRU(page));
3632 * The page is isolated from LRU. So, collapse function
3633 * will not handle this page. But page splitting can happen.
3634 * Do this check under compound_page_lock(). The caller should
3638 if (nr_pages > 1 && !PageTransHuge(page))
3641 lock_page_cgroup(pc);
3644 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3647 move_lock_mem_cgroup(from, &flags);
3649 if (!anon && page_mapped(page)) {
3650 /* Update mapped_file data for mem_cgroup */
3652 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3653 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3656 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3658 /* caller should have done css_get */
3659 pc->mem_cgroup = to;
3660 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3661 move_unlock_mem_cgroup(from, &flags);
3664 unlock_page_cgroup(pc);
3668 memcg_check_events(to, page);
3669 memcg_check_events(from, page);
3675 * mem_cgroup_move_parent - moves page to the parent group
3676 * @page: the page to move
3677 * @pc: page_cgroup of the page
3678 * @child: page's cgroup
3680 * move charges to its parent or the root cgroup if the group has no
3681 * parent (aka use_hierarchy==0).
3682 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3683 * mem_cgroup_move_account fails) the failure is always temporary and
3684 * it signals a race with a page removal/uncharge or migration. In the
3685 * first case the page is on the way out and it will vanish from the LRU
3686 * on the next attempt and the call should be retried later.
3687 * Isolation from the LRU fails only if page has been isolated from
3688 * the LRU since we looked at it and that usually means either global
3689 * reclaim or migration going on. The page will either get back to the
3691 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3692 * (!PageCgroupUsed) or moved to a different group. The page will
3693 * disappear in the next attempt.
3695 static int mem_cgroup_move_parent(struct page *page,
3696 struct page_cgroup *pc,
3697 struct mem_cgroup *child)
3699 struct mem_cgroup *parent;
3700 unsigned int nr_pages;
3701 unsigned long uninitialized_var(flags);
3704 VM_BUG_ON(mem_cgroup_is_root(child));
3707 if (!get_page_unless_zero(page))
3709 if (isolate_lru_page(page))
3712 nr_pages = hpage_nr_pages(page);
3714 parent = parent_mem_cgroup(child);
3716 * If no parent, move charges to root cgroup.
3719 parent = root_mem_cgroup;
3722 VM_BUG_ON(!PageTransHuge(page));
3723 flags = compound_lock_irqsave(page);
3726 ret = mem_cgroup_move_account(page, nr_pages,
3729 __mem_cgroup_cancel_local_charge(child, nr_pages);
3732 compound_unlock_irqrestore(page, flags);
3733 putback_lru_page(page);
3741 * Charge the memory controller for page usage.
3743 * 0 if the charge was successful
3744 * < 0 if the cgroup is over its limit
3746 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3747 gfp_t gfp_mask, enum charge_type ctype)
3749 struct mem_cgroup *memcg = NULL;
3750 unsigned int nr_pages = 1;
3754 if (PageTransHuge(page)) {
3755 nr_pages <<= compound_order(page);
3756 VM_BUG_ON(!PageTransHuge(page));
3758 * Never OOM-kill a process for a huge page. The
3759 * fault handler will fall back to regular pages.
3764 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3767 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3771 int mem_cgroup_newpage_charge(struct page *page,
3772 struct mm_struct *mm, gfp_t gfp_mask)
3774 if (mem_cgroup_disabled())
3776 VM_BUG_ON(page_mapped(page));
3777 VM_BUG_ON(page->mapping && !PageAnon(page));
3779 return mem_cgroup_charge_common(page, mm, gfp_mask,
3780 MEM_CGROUP_CHARGE_TYPE_ANON);
3784 * While swap-in, try_charge -> commit or cancel, the page is locked.
3785 * And when try_charge() successfully returns, one refcnt to memcg without
3786 * struct page_cgroup is acquired. This refcnt will be consumed by
3787 * "commit()" or removed by "cancel()"
3789 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3792 struct mem_cgroup **memcgp)
3794 struct mem_cgroup *memcg;
3795 struct page_cgroup *pc;
3798 pc = lookup_page_cgroup(page);
3800 * Every swap fault against a single page tries to charge the
3801 * page, bail as early as possible. shmem_unuse() encounters
3802 * already charged pages, too. The USED bit is protected by
3803 * the page lock, which serializes swap cache removal, which
3804 * in turn serializes uncharging.
3806 if (PageCgroupUsed(pc))
3808 if (!do_swap_account)
3810 memcg = try_get_mem_cgroup_from_page(page);
3814 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3815 css_put(&memcg->css);
3820 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3826 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3827 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3830 if (mem_cgroup_disabled())
3833 * A racing thread's fault, or swapoff, may have already
3834 * updated the pte, and even removed page from swap cache: in
3835 * those cases unuse_pte()'s pte_same() test will fail; but
3836 * there's also a KSM case which does need to charge the page.
3838 if (!PageSwapCache(page)) {
3841 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3846 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3849 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3851 if (mem_cgroup_disabled())
3855 __mem_cgroup_cancel_charge(memcg, 1);
3859 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3860 enum charge_type ctype)
3862 if (mem_cgroup_disabled())
3867 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3869 * Now swap is on-memory. This means this page may be
3870 * counted both as mem and swap....double count.
3871 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3872 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3873 * may call delete_from_swap_cache() before reach here.
3875 if (do_swap_account && PageSwapCache(page)) {
3876 swp_entry_t ent = {.val = page_private(page)};
3877 mem_cgroup_uncharge_swap(ent);
3881 void mem_cgroup_commit_charge_swapin(struct page *page,
3882 struct mem_cgroup *memcg)
3884 __mem_cgroup_commit_charge_swapin(page, memcg,
3885 MEM_CGROUP_CHARGE_TYPE_ANON);
3888 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3891 struct mem_cgroup *memcg = NULL;
3892 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3895 if (mem_cgroup_disabled())
3897 if (PageCompound(page))
3900 if (!PageSwapCache(page))
3901 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3902 else { /* page is swapcache/shmem */
3903 ret = __mem_cgroup_try_charge_swapin(mm, page,
3906 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3911 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3912 unsigned int nr_pages,
3913 const enum charge_type ctype)
3915 struct memcg_batch_info *batch = NULL;
3916 bool uncharge_memsw = true;
3918 /* If swapout, usage of swap doesn't decrease */
3919 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3920 uncharge_memsw = false;
3922 batch = ¤t->memcg_batch;
3924 * In usual, we do css_get() when we remember memcg pointer.
3925 * But in this case, we keep res->usage until end of a series of
3926 * uncharges. Then, it's ok to ignore memcg's refcnt.
3929 batch->memcg = memcg;
3931 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3932 * In those cases, all pages freed continuously can be expected to be in
3933 * the same cgroup and we have chance to coalesce uncharges.
3934 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3935 * because we want to do uncharge as soon as possible.
3938 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3939 goto direct_uncharge;
3942 goto direct_uncharge;
3945 * In typical case, batch->memcg == mem. This means we can
3946 * merge a series of uncharges to an uncharge of res_counter.
3947 * If not, we uncharge res_counter ony by one.
3949 if (batch->memcg != memcg)
3950 goto direct_uncharge;
3951 /* remember freed charge and uncharge it later */
3954 batch->memsw_nr_pages++;
3957 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3959 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3960 if (unlikely(batch->memcg != memcg))
3961 memcg_oom_recover(memcg);
3965 * uncharge if !page_mapped(page)
3967 static struct mem_cgroup *
3968 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3971 struct mem_cgroup *memcg = NULL;
3972 unsigned int nr_pages = 1;
3973 struct page_cgroup *pc;
3976 if (mem_cgroup_disabled())
3979 if (PageTransHuge(page)) {
3980 nr_pages <<= compound_order(page);
3981 VM_BUG_ON(!PageTransHuge(page));
3984 * Check if our page_cgroup is valid
3986 pc = lookup_page_cgroup(page);
3987 if (unlikely(!PageCgroupUsed(pc)))
3990 lock_page_cgroup(pc);
3992 memcg = pc->mem_cgroup;
3994 if (!PageCgroupUsed(pc))
3997 anon = PageAnon(page);
4000 case MEM_CGROUP_CHARGE_TYPE_ANON:
4002 * Generally PageAnon tells if it's the anon statistics to be
4003 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4004 * used before page reached the stage of being marked PageAnon.
4008 case MEM_CGROUP_CHARGE_TYPE_DROP:
4009 /* See mem_cgroup_prepare_migration() */
4010 if (page_mapped(page))
4013 * Pages under migration may not be uncharged. But
4014 * end_migration() /must/ be the one uncharging the
4015 * unused post-migration page and so it has to call
4016 * here with the migration bit still set. See the
4017 * res_counter handling below.
4019 if (!end_migration && PageCgroupMigration(pc))
4022 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4023 if (!PageAnon(page)) { /* Shared memory */
4024 if (page->mapping && !page_is_file_cache(page))
4026 } else if (page_mapped(page)) /* Anon */
4033 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4035 ClearPageCgroupUsed(pc);
4037 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4038 * freed from LRU. This is safe because uncharged page is expected not
4039 * to be reused (freed soon). Exception is SwapCache, it's handled by
4040 * special functions.
4043 unlock_page_cgroup(pc);
4045 * even after unlock, we have memcg->res.usage here and this memcg
4046 * will never be freed, so it's safe to call css_get().
4048 memcg_check_events(memcg, page);
4049 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4050 mem_cgroup_swap_statistics(memcg, true);
4051 css_get(&memcg->css);
4054 * Migration does not charge the res_counter for the
4055 * replacement page, so leave it alone when phasing out the
4056 * page that is unused after the migration.
4058 if (!end_migration && !mem_cgroup_is_root(memcg))
4059 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4064 unlock_page_cgroup(pc);
4068 void mem_cgroup_uncharge_page(struct page *page)
4071 if (page_mapped(page))
4073 VM_BUG_ON(page->mapping && !PageAnon(page));
4075 * If the page is in swap cache, uncharge should be deferred
4076 * to the swap path, which also properly accounts swap usage
4077 * and handles memcg lifetime.
4079 * Note that this check is not stable and reclaim may add the
4080 * page to swap cache at any time after this. However, if the
4081 * page is not in swap cache by the time page->mapcount hits
4082 * 0, there won't be any page table references to the swap
4083 * slot, and reclaim will free it and not actually write the
4086 if (PageSwapCache(page))
4088 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4091 void mem_cgroup_uncharge_cache_page(struct page *page)
4093 VM_BUG_ON(page_mapped(page));
4094 VM_BUG_ON(page->mapping);
4095 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4099 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4100 * In that cases, pages are freed continuously and we can expect pages
4101 * are in the same memcg. All these calls itself limits the number of
4102 * pages freed at once, then uncharge_start/end() is called properly.
4103 * This may be called prural(2) times in a context,
4106 void mem_cgroup_uncharge_start(void)
4108 current->memcg_batch.do_batch++;
4109 /* We can do nest. */
4110 if (current->memcg_batch.do_batch == 1) {
4111 current->memcg_batch.memcg = NULL;
4112 current->memcg_batch.nr_pages = 0;
4113 current->memcg_batch.memsw_nr_pages = 0;
4117 void mem_cgroup_uncharge_end(void)
4119 struct memcg_batch_info *batch = ¤t->memcg_batch;
4121 if (!batch->do_batch)
4125 if (batch->do_batch) /* If stacked, do nothing. */
4131 * This "batch->memcg" is valid without any css_get/put etc...
4132 * bacause we hide charges behind us.
4134 if (batch->nr_pages)
4135 res_counter_uncharge(&batch->memcg->res,
4136 batch->nr_pages * PAGE_SIZE);
4137 if (batch->memsw_nr_pages)
4138 res_counter_uncharge(&batch->memcg->memsw,
4139 batch->memsw_nr_pages * PAGE_SIZE);
4140 memcg_oom_recover(batch->memcg);
4141 /* forget this pointer (for sanity check) */
4142 batch->memcg = NULL;
4147 * called after __delete_from_swap_cache() and drop "page" account.
4148 * memcg information is recorded to swap_cgroup of "ent"
4151 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4153 struct mem_cgroup *memcg;
4154 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4156 if (!swapout) /* this was a swap cache but the swap is unused ! */
4157 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4159 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4162 * record memcg information, if swapout && memcg != NULL,
4163 * css_get() was called in uncharge().
4165 if (do_swap_account && swapout && memcg)
4166 swap_cgroup_record(ent, css_id(&memcg->css));
4170 #ifdef CONFIG_MEMCG_SWAP
4172 * called from swap_entry_free(). remove record in swap_cgroup and
4173 * uncharge "memsw" account.
4175 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4177 struct mem_cgroup *memcg;
4180 if (!do_swap_account)
4183 id = swap_cgroup_record(ent, 0);
4185 memcg = mem_cgroup_lookup(id);
4188 * We uncharge this because swap is freed.
4189 * This memcg can be obsolete one. We avoid calling css_tryget
4191 if (!mem_cgroup_is_root(memcg))
4192 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4193 mem_cgroup_swap_statistics(memcg, false);
4194 css_put(&memcg->css);
4200 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4201 * @entry: swap entry to be moved
4202 * @from: mem_cgroup which the entry is moved from
4203 * @to: mem_cgroup which the entry is moved to
4205 * It succeeds only when the swap_cgroup's record for this entry is the same
4206 * as the mem_cgroup's id of @from.
4208 * Returns 0 on success, -EINVAL on failure.
4210 * The caller must have charged to @to, IOW, called res_counter_charge() about
4211 * both res and memsw, and called css_get().
4213 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4214 struct mem_cgroup *from, struct mem_cgroup *to)
4216 unsigned short old_id, new_id;
4218 old_id = css_id(&from->css);
4219 new_id = css_id(&to->css);
4221 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4222 mem_cgroup_swap_statistics(from, false);
4223 mem_cgroup_swap_statistics(to, true);
4225 * This function is only called from task migration context now.
4226 * It postpones res_counter and refcount handling till the end
4227 * of task migration(mem_cgroup_clear_mc()) for performance
4228 * improvement. But we cannot postpone css_get(to) because if
4229 * the process that has been moved to @to does swap-in, the
4230 * refcount of @to might be decreased to 0.
4232 * We are in attach() phase, so the cgroup is guaranteed to be
4233 * alive, so we can just call css_get().
4241 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4242 struct mem_cgroup *from, struct mem_cgroup *to)
4249 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4252 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4253 struct mem_cgroup **memcgp)
4255 struct mem_cgroup *memcg = NULL;
4256 unsigned int nr_pages = 1;
4257 struct page_cgroup *pc;
4258 enum charge_type ctype;
4262 if (mem_cgroup_disabled())
4265 if (PageTransHuge(page))
4266 nr_pages <<= compound_order(page);
4268 pc = lookup_page_cgroup(page);
4269 lock_page_cgroup(pc);
4270 if (PageCgroupUsed(pc)) {
4271 memcg = pc->mem_cgroup;
4272 css_get(&memcg->css);
4274 * At migrating an anonymous page, its mapcount goes down
4275 * to 0 and uncharge() will be called. But, even if it's fully
4276 * unmapped, migration may fail and this page has to be
4277 * charged again. We set MIGRATION flag here and delay uncharge
4278 * until end_migration() is called
4280 * Corner Case Thinking
4282 * When the old page was mapped as Anon and it's unmap-and-freed
4283 * while migration was ongoing.
4284 * If unmap finds the old page, uncharge() of it will be delayed
4285 * until end_migration(). If unmap finds a new page, it's
4286 * uncharged when it make mapcount to be 1->0. If unmap code
4287 * finds swap_migration_entry, the new page will not be mapped
4288 * and end_migration() will find it(mapcount==0).
4291 * When the old page was mapped but migraion fails, the kernel
4292 * remaps it. A charge for it is kept by MIGRATION flag even
4293 * if mapcount goes down to 0. We can do remap successfully
4294 * without charging it again.
4297 * The "old" page is under lock_page() until the end of
4298 * migration, so, the old page itself will not be swapped-out.
4299 * If the new page is swapped out before end_migraton, our
4300 * hook to usual swap-out path will catch the event.
4303 SetPageCgroupMigration(pc);
4305 unlock_page_cgroup(pc);
4307 * If the page is not charged at this point,
4315 * We charge new page before it's used/mapped. So, even if unlock_page()
4316 * is called before end_migration, we can catch all events on this new
4317 * page. In the case new page is migrated but not remapped, new page's
4318 * mapcount will be finally 0 and we call uncharge in end_migration().
4321 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4323 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4325 * The page is committed to the memcg, but it's not actually
4326 * charged to the res_counter since we plan on replacing the
4327 * old one and only one page is going to be left afterwards.
4329 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4332 /* remove redundant charge if migration failed*/
4333 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4334 struct page *oldpage, struct page *newpage, bool migration_ok)
4336 struct page *used, *unused;
4337 struct page_cgroup *pc;
4343 if (!migration_ok) {
4350 anon = PageAnon(used);
4351 __mem_cgroup_uncharge_common(unused,
4352 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4353 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4355 css_put(&memcg->css);
4357 * We disallowed uncharge of pages under migration because mapcount
4358 * of the page goes down to zero, temporarly.
4359 * Clear the flag and check the page should be charged.
4361 pc = lookup_page_cgroup(oldpage);
4362 lock_page_cgroup(pc);
4363 ClearPageCgroupMigration(pc);
4364 unlock_page_cgroup(pc);
4367 * If a page is a file cache, radix-tree replacement is very atomic
4368 * and we can skip this check. When it was an Anon page, its mapcount
4369 * goes down to 0. But because we added MIGRATION flage, it's not
4370 * uncharged yet. There are several case but page->mapcount check
4371 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4372 * check. (see prepare_charge() also)
4375 mem_cgroup_uncharge_page(used);
4379 * At replace page cache, newpage is not under any memcg but it's on
4380 * LRU. So, this function doesn't touch res_counter but handles LRU
4381 * in correct way. Both pages are locked so we cannot race with uncharge.
4383 void mem_cgroup_replace_page_cache(struct page *oldpage,
4384 struct page *newpage)
4386 struct mem_cgroup *memcg = NULL;
4387 struct page_cgroup *pc;
4388 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4390 if (mem_cgroup_disabled())
4393 pc = lookup_page_cgroup(oldpage);
4394 /* fix accounting on old pages */
4395 lock_page_cgroup(pc);
4396 if (PageCgroupUsed(pc)) {
4397 memcg = pc->mem_cgroup;
4398 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4399 ClearPageCgroupUsed(pc);
4401 unlock_page_cgroup(pc);
4404 * When called from shmem_replace_page(), in some cases the
4405 * oldpage has already been charged, and in some cases not.
4410 * Even if newpage->mapping was NULL before starting replacement,
4411 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4412 * LRU while we overwrite pc->mem_cgroup.
4414 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4417 #ifdef CONFIG_DEBUG_VM
4418 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4420 struct page_cgroup *pc;
4422 pc = lookup_page_cgroup(page);
4424 * Can be NULL while feeding pages into the page allocator for
4425 * the first time, i.e. during boot or memory hotplug;
4426 * or when mem_cgroup_disabled().
4428 if (likely(pc) && PageCgroupUsed(pc))
4433 bool mem_cgroup_bad_page_check(struct page *page)
4435 if (mem_cgroup_disabled())
4438 return lookup_page_cgroup_used(page) != NULL;
4441 void mem_cgroup_print_bad_page(struct page *page)
4443 struct page_cgroup *pc;
4445 pc = lookup_page_cgroup_used(page);
4447 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4448 pc, pc->flags, pc->mem_cgroup);
4453 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4454 unsigned long long val)
4457 u64 memswlimit, memlimit;
4459 int children = mem_cgroup_count_children(memcg);
4460 u64 curusage, oldusage;
4464 * For keeping hierarchical_reclaim simple, how long we should retry
4465 * is depends on callers. We set our retry-count to be function
4466 * of # of children which we should visit in this loop.
4468 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4470 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4473 while (retry_count) {
4474 if (signal_pending(current)) {
4479 * Rather than hide all in some function, I do this in
4480 * open coded manner. You see what this really does.
4481 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4483 mutex_lock(&set_limit_mutex);
4484 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4485 if (memswlimit < val) {
4487 mutex_unlock(&set_limit_mutex);
4491 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4495 ret = res_counter_set_limit(&memcg->res, val);
4497 if (memswlimit == val)
4498 memcg->memsw_is_minimum = true;
4500 memcg->memsw_is_minimum = false;
4502 mutex_unlock(&set_limit_mutex);
4507 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4508 MEM_CGROUP_RECLAIM_SHRINK);
4509 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4510 /* Usage is reduced ? */
4511 if (curusage >= oldusage)
4514 oldusage = curusage;
4516 if (!ret && enlarge)
4517 memcg_oom_recover(memcg);
4522 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4523 unsigned long long val)
4526 u64 memlimit, memswlimit, oldusage, curusage;
4527 int children = mem_cgroup_count_children(memcg);
4531 /* see mem_cgroup_resize_res_limit */
4532 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4533 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4534 while (retry_count) {
4535 if (signal_pending(current)) {
4540 * Rather than hide all in some function, I do this in
4541 * open coded manner. You see what this really does.
4542 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4544 mutex_lock(&set_limit_mutex);
4545 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4546 if (memlimit > val) {
4548 mutex_unlock(&set_limit_mutex);
4551 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4552 if (memswlimit < val)
4554 ret = res_counter_set_limit(&memcg->memsw, val);
4556 if (memlimit == val)
4557 memcg->memsw_is_minimum = true;
4559 memcg->memsw_is_minimum = false;
4561 mutex_unlock(&set_limit_mutex);
4566 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4567 MEM_CGROUP_RECLAIM_NOSWAP |
4568 MEM_CGROUP_RECLAIM_SHRINK);
4569 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4570 /* Usage is reduced ? */
4571 if (curusage >= oldusage)
4574 oldusage = curusage;
4576 if (!ret && enlarge)
4577 memcg_oom_recover(memcg);
4582 * mem_cgroup_force_empty_list - clears LRU of a group
4583 * @memcg: group to clear
4586 * @lru: lru to to clear
4588 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4589 * reclaim the pages page themselves - pages are moved to the parent (or root)
4592 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4593 int node, int zid, enum lru_list lru)
4595 struct lruvec *lruvec;
4596 unsigned long flags;
4597 struct list_head *list;
4601 zone = &NODE_DATA(node)->node_zones[zid];
4602 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4603 list = &lruvec->lists[lru];
4607 struct page_cgroup *pc;
4610 spin_lock_irqsave(&zone->lru_lock, flags);
4611 if (list_empty(list)) {
4612 spin_unlock_irqrestore(&zone->lru_lock, flags);
4615 page = list_entry(list->prev, struct page, lru);
4617 list_move(&page->lru, list);
4619 spin_unlock_irqrestore(&zone->lru_lock, flags);
4622 spin_unlock_irqrestore(&zone->lru_lock, flags);
4624 pc = lookup_page_cgroup(page);
4626 if (mem_cgroup_move_parent(page, pc, memcg)) {
4627 /* found lock contention or "pc" is obsolete. */
4632 } while (!list_empty(list));
4636 * make mem_cgroup's charge to be 0 if there is no task by moving
4637 * all the charges and pages to the parent.
4638 * This enables deleting this mem_cgroup.
4640 * Caller is responsible for holding css reference on the memcg.
4642 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4648 /* This is for making all *used* pages to be on LRU. */
4649 lru_add_drain_all();
4650 drain_all_stock_sync(memcg);
4651 mem_cgroup_start_move(memcg);
4652 for_each_node_state(node, N_MEMORY) {
4653 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4656 mem_cgroup_force_empty_list(memcg,
4661 mem_cgroup_end_move(memcg);
4662 memcg_oom_recover(memcg);
4666 * Kernel memory may not necessarily be trackable to a specific
4667 * process. So they are not migrated, and therefore we can't
4668 * expect their value to drop to 0 here.
4669 * Having res filled up with kmem only is enough.
4671 * This is a safety check because mem_cgroup_force_empty_list
4672 * could have raced with mem_cgroup_replace_page_cache callers
4673 * so the lru seemed empty but the page could have been added
4674 * right after the check. RES_USAGE should be safe as we always
4675 * charge before adding to the LRU.
4677 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4678 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4679 } while (usage > 0);
4683 * This mainly exists for tests during the setting of set of use_hierarchy.
4684 * Since this is the very setting we are changing, the current hierarchy value
4687 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4689 struct cgroup_subsys_state *pos;
4691 /* bounce at first found */
4692 css_for_each_child(pos, &memcg->css)
4698 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4699 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4700 * from mem_cgroup_count_children(), in the sense that we don't really care how
4701 * many children we have; we only need to know if we have any. It also counts
4702 * any memcg without hierarchy as infertile.
4704 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4706 return memcg->use_hierarchy && __memcg_has_children(memcg);
4710 * Reclaims as many pages from the given memcg as possible and moves
4711 * the rest to the parent.
4713 * Caller is responsible for holding css reference for memcg.
4715 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4717 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4718 struct cgroup *cgrp = memcg->css.cgroup;
4720 /* returns EBUSY if there is a task or if we come here twice. */
4721 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4724 /* we call try-to-free pages for make this cgroup empty */
4725 lru_add_drain_all();
4726 /* try to free all pages in this cgroup */
4727 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4730 if (signal_pending(current))
4733 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4737 /* maybe some writeback is necessary */
4738 congestion_wait(BLK_RW_ASYNC, HZ/10);
4743 mem_cgroup_reparent_charges(memcg);
4748 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4751 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4753 if (mem_cgroup_is_root(memcg))
4755 return mem_cgroup_force_empty(memcg);
4758 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4761 return mem_cgroup_from_css(css)->use_hierarchy;
4764 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4765 struct cftype *cft, u64 val)
4768 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4769 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4771 mutex_lock(&memcg_create_mutex);
4773 if (memcg->use_hierarchy == val)
4777 * If parent's use_hierarchy is set, we can't make any modifications
4778 * in the child subtrees. If it is unset, then the change can
4779 * occur, provided the current cgroup has no children.
4781 * For the root cgroup, parent_mem is NULL, we allow value to be
4782 * set if there are no children.
4784 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4785 (val == 1 || val == 0)) {
4786 if (!__memcg_has_children(memcg))
4787 memcg->use_hierarchy = val;
4794 mutex_unlock(&memcg_create_mutex);
4800 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4801 enum mem_cgroup_stat_index idx)
4803 struct mem_cgroup *iter;
4806 /* Per-cpu values can be negative, use a signed accumulator */
4807 for_each_mem_cgroup_tree(iter, memcg)
4808 val += mem_cgroup_read_stat(iter, idx);
4810 if (val < 0) /* race ? */
4815 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4819 if (!mem_cgroup_is_root(memcg)) {
4821 return res_counter_read_u64(&memcg->res, RES_USAGE);
4823 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4827 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4828 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4830 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4831 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4834 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4836 return val << PAGE_SHIFT;
4839 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
4840 struct cftype *cft, struct file *file,
4841 char __user *buf, size_t nbytes, loff_t *ppos)
4843 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4849 type = MEMFILE_TYPE(cft->private);
4850 name = MEMFILE_ATTR(cft->private);
4854 if (name == RES_USAGE)
4855 val = mem_cgroup_usage(memcg, false);
4857 val = res_counter_read_u64(&memcg->res, name);
4860 if (name == RES_USAGE)
4861 val = mem_cgroup_usage(memcg, true);
4863 val = res_counter_read_u64(&memcg->memsw, name);
4866 val = res_counter_read_u64(&memcg->kmem, name);
4872 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4873 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4876 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
4879 #ifdef CONFIG_MEMCG_KMEM
4880 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4882 * For simplicity, we won't allow this to be disabled. It also can't
4883 * be changed if the cgroup has children already, or if tasks had
4886 * If tasks join before we set the limit, a person looking at
4887 * kmem.usage_in_bytes will have no way to determine when it took
4888 * place, which makes the value quite meaningless.
4890 * After it first became limited, changes in the value of the limit are
4891 * of course permitted.
4893 mutex_lock(&memcg_create_mutex);
4894 mutex_lock(&set_limit_mutex);
4895 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
4896 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
4900 ret = res_counter_set_limit(&memcg->kmem, val);
4903 ret = memcg_update_cache_sizes(memcg);
4905 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
4908 static_key_slow_inc(&memcg_kmem_enabled_key);
4910 * setting the active bit after the inc will guarantee no one
4911 * starts accounting before all call sites are patched
4913 memcg_kmem_set_active(memcg);
4915 ret = res_counter_set_limit(&memcg->kmem, val);
4917 mutex_unlock(&set_limit_mutex);
4918 mutex_unlock(&memcg_create_mutex);
4923 #ifdef CONFIG_MEMCG_KMEM
4924 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4927 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4931 memcg->kmem_account_flags = parent->kmem_account_flags;
4933 * When that happen, we need to disable the static branch only on those
4934 * memcgs that enabled it. To achieve this, we would be forced to
4935 * complicate the code by keeping track of which memcgs were the ones
4936 * that actually enabled limits, and which ones got it from its
4939 * It is a lot simpler just to do static_key_slow_inc() on every child
4940 * that is accounted.
4942 if (!memcg_kmem_is_active(memcg))
4946 * __mem_cgroup_free() will issue static_key_slow_dec() because this
4947 * memcg is active already. If the later initialization fails then the
4948 * cgroup core triggers the cleanup so we do not have to do it here.
4950 static_key_slow_inc(&memcg_kmem_enabled_key);
4952 mutex_lock(&set_limit_mutex);
4953 memcg_stop_kmem_account();
4954 ret = memcg_update_cache_sizes(memcg);
4955 memcg_resume_kmem_account();
4956 mutex_unlock(&set_limit_mutex);
4960 #endif /* CONFIG_MEMCG_KMEM */
4963 * The user of this function is...
4966 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
4969 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4972 unsigned long long val;
4975 type = MEMFILE_TYPE(cft->private);
4976 name = MEMFILE_ATTR(cft->private);
4980 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4984 /* This function does all necessary parse...reuse it */
4985 ret = res_counter_memparse_write_strategy(buffer, &val);
4989 ret = mem_cgroup_resize_limit(memcg, val);
4990 else if (type == _MEMSWAP)
4991 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4992 else if (type == _KMEM)
4993 ret = memcg_update_kmem_limit(css, val);
4997 case RES_SOFT_LIMIT:
4998 ret = res_counter_memparse_write_strategy(buffer, &val);
5002 * For memsw, soft limits are hard to implement in terms
5003 * of semantics, for now, we support soft limits for
5004 * control without swap
5007 ret = res_counter_set_soft_limit(&memcg->res, val);
5012 ret = -EINVAL; /* should be BUG() ? */
5018 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5019 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5021 unsigned long long min_limit, min_memsw_limit, tmp;
5023 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5024 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5025 if (!memcg->use_hierarchy)
5028 while (css_parent(&memcg->css)) {
5029 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5030 if (!memcg->use_hierarchy)
5032 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5033 min_limit = min(min_limit, tmp);
5034 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5035 min_memsw_limit = min(min_memsw_limit, tmp);
5038 *mem_limit = min_limit;
5039 *memsw_limit = min_memsw_limit;
5042 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5044 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5048 type = MEMFILE_TYPE(event);
5049 name = MEMFILE_ATTR(event);
5054 res_counter_reset_max(&memcg->res);
5055 else if (type == _MEMSWAP)
5056 res_counter_reset_max(&memcg->memsw);
5057 else if (type == _KMEM)
5058 res_counter_reset_max(&memcg->kmem);
5064 res_counter_reset_failcnt(&memcg->res);
5065 else if (type == _MEMSWAP)
5066 res_counter_reset_failcnt(&memcg->memsw);
5067 else if (type == _KMEM)
5068 res_counter_reset_failcnt(&memcg->kmem);
5077 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5080 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5084 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5085 struct cftype *cft, u64 val)
5087 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5089 if (val >= (1 << NR_MOVE_TYPE))
5093 * No kind of locking is needed in here, because ->can_attach() will
5094 * check this value once in the beginning of the process, and then carry
5095 * on with stale data. This means that changes to this value will only
5096 * affect task migrations starting after the change.
5098 memcg->move_charge_at_immigrate = val;
5102 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5103 struct cftype *cft, u64 val)
5110 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5111 struct cftype *cft, struct seq_file *m)
5114 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5115 unsigned long node_nr;
5116 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5118 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5119 seq_printf(m, "total=%lu", total_nr);
5120 for_each_node_state(nid, N_MEMORY) {
5121 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5122 seq_printf(m, " N%d=%lu", nid, node_nr);
5126 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5127 seq_printf(m, "file=%lu", file_nr);
5128 for_each_node_state(nid, N_MEMORY) {
5129 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5131 seq_printf(m, " N%d=%lu", nid, node_nr);
5135 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5136 seq_printf(m, "anon=%lu", anon_nr);
5137 for_each_node_state(nid, N_MEMORY) {
5138 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5140 seq_printf(m, " N%d=%lu", nid, node_nr);
5144 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5145 seq_printf(m, "unevictable=%lu", unevictable_nr);
5146 for_each_node_state(nid, N_MEMORY) {
5147 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5148 BIT(LRU_UNEVICTABLE));
5149 seq_printf(m, " N%d=%lu", nid, node_nr);
5154 #endif /* CONFIG_NUMA */
5156 static inline void mem_cgroup_lru_names_not_uptodate(void)
5158 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5161 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5164 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5165 struct mem_cgroup *mi;
5168 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5169 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5171 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5172 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5175 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5176 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5177 mem_cgroup_read_events(memcg, i));
5179 for (i = 0; i < NR_LRU_LISTS; i++)
5180 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5181 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5183 /* Hierarchical information */
5185 unsigned long long limit, memsw_limit;
5186 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5187 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5188 if (do_swap_account)
5189 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5193 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5196 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5198 for_each_mem_cgroup_tree(mi, memcg)
5199 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5200 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5203 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5204 unsigned long long val = 0;
5206 for_each_mem_cgroup_tree(mi, memcg)
5207 val += mem_cgroup_read_events(mi, i);
5208 seq_printf(m, "total_%s %llu\n",
5209 mem_cgroup_events_names[i], val);
5212 for (i = 0; i < NR_LRU_LISTS; i++) {
5213 unsigned long long val = 0;
5215 for_each_mem_cgroup_tree(mi, memcg)
5216 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5217 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5220 #ifdef CONFIG_DEBUG_VM
5223 struct mem_cgroup_per_zone *mz;
5224 struct zone_reclaim_stat *rstat;
5225 unsigned long recent_rotated[2] = {0, 0};
5226 unsigned long recent_scanned[2] = {0, 0};
5228 for_each_online_node(nid)
5229 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5230 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5231 rstat = &mz->lruvec.reclaim_stat;
5233 recent_rotated[0] += rstat->recent_rotated[0];
5234 recent_rotated[1] += rstat->recent_rotated[1];
5235 recent_scanned[0] += rstat->recent_scanned[0];
5236 recent_scanned[1] += rstat->recent_scanned[1];
5238 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5239 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5240 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5241 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5248 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5251 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5253 return mem_cgroup_swappiness(memcg);
5256 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5257 struct cftype *cft, u64 val)
5259 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5260 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5262 if (val > 100 || !parent)
5265 mutex_lock(&memcg_create_mutex);
5267 /* If under hierarchy, only empty-root can set this value */
5268 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5269 mutex_unlock(&memcg_create_mutex);
5273 memcg->swappiness = val;
5275 mutex_unlock(&memcg_create_mutex);
5280 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5282 struct mem_cgroup_threshold_ary *t;
5288 t = rcu_dereference(memcg->thresholds.primary);
5290 t = rcu_dereference(memcg->memsw_thresholds.primary);
5295 usage = mem_cgroup_usage(memcg, swap);
5298 * current_threshold points to threshold just below or equal to usage.
5299 * If it's not true, a threshold was crossed after last
5300 * call of __mem_cgroup_threshold().
5302 i = t->current_threshold;
5305 * Iterate backward over array of thresholds starting from
5306 * current_threshold and check if a threshold is crossed.
5307 * If none of thresholds below usage is crossed, we read
5308 * only one element of the array here.
5310 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5311 eventfd_signal(t->entries[i].eventfd, 1);
5313 /* i = current_threshold + 1 */
5317 * Iterate forward over array of thresholds starting from
5318 * current_threshold+1 and check if a threshold is crossed.
5319 * If none of thresholds above usage is crossed, we read
5320 * only one element of the array here.
5322 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5323 eventfd_signal(t->entries[i].eventfd, 1);
5325 /* Update current_threshold */
5326 t->current_threshold = i - 1;
5331 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5334 __mem_cgroup_threshold(memcg, false);
5335 if (do_swap_account)
5336 __mem_cgroup_threshold(memcg, true);
5338 memcg = parent_mem_cgroup(memcg);
5342 static int compare_thresholds(const void *a, const void *b)
5344 const struct mem_cgroup_threshold *_a = a;
5345 const struct mem_cgroup_threshold *_b = b;
5347 if (_a->threshold > _b->threshold)
5350 if (_a->threshold < _b->threshold)
5356 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5358 struct mem_cgroup_eventfd_list *ev;
5360 list_for_each_entry(ev, &memcg->oom_notify, list)
5361 eventfd_signal(ev->eventfd, 1);
5365 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5367 struct mem_cgroup *iter;
5369 for_each_mem_cgroup_tree(iter, memcg)
5370 mem_cgroup_oom_notify_cb(iter);
5373 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5374 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5376 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5377 struct mem_cgroup_thresholds *thresholds;
5378 struct mem_cgroup_threshold_ary *new;
5379 enum res_type type = MEMFILE_TYPE(cft->private);
5380 u64 threshold, usage;
5383 ret = res_counter_memparse_write_strategy(args, &threshold);
5387 mutex_lock(&memcg->thresholds_lock);
5390 thresholds = &memcg->thresholds;
5391 else if (type == _MEMSWAP)
5392 thresholds = &memcg->memsw_thresholds;
5396 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5398 /* Check if a threshold crossed before adding a new one */
5399 if (thresholds->primary)
5400 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5402 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5404 /* Allocate memory for new array of thresholds */
5405 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5413 /* Copy thresholds (if any) to new array */
5414 if (thresholds->primary) {
5415 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5416 sizeof(struct mem_cgroup_threshold));
5419 /* Add new threshold */
5420 new->entries[size - 1].eventfd = eventfd;
5421 new->entries[size - 1].threshold = threshold;
5423 /* Sort thresholds. Registering of new threshold isn't time-critical */
5424 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5425 compare_thresholds, NULL);
5427 /* Find current threshold */
5428 new->current_threshold = -1;
5429 for (i = 0; i < size; i++) {
5430 if (new->entries[i].threshold <= usage) {
5432 * new->current_threshold will not be used until
5433 * rcu_assign_pointer(), so it's safe to increment
5436 ++new->current_threshold;
5441 /* Free old spare buffer and save old primary buffer as spare */
5442 kfree(thresholds->spare);
5443 thresholds->spare = thresholds->primary;
5445 rcu_assign_pointer(thresholds->primary, new);
5447 /* To be sure that nobody uses thresholds */
5451 mutex_unlock(&memcg->thresholds_lock);
5456 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5457 struct cftype *cft, struct eventfd_ctx *eventfd)
5459 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5460 struct mem_cgroup_thresholds *thresholds;
5461 struct mem_cgroup_threshold_ary *new;
5462 enum res_type type = MEMFILE_TYPE(cft->private);
5466 mutex_lock(&memcg->thresholds_lock);
5468 thresholds = &memcg->thresholds;
5469 else if (type == _MEMSWAP)
5470 thresholds = &memcg->memsw_thresholds;
5474 if (!thresholds->primary)
5477 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5479 /* Check if a threshold crossed before removing */
5480 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5482 /* Calculate new number of threshold */
5484 for (i = 0; i < thresholds->primary->size; i++) {
5485 if (thresholds->primary->entries[i].eventfd != eventfd)
5489 new = thresholds->spare;
5491 /* Set thresholds array to NULL if we don't have thresholds */
5500 /* Copy thresholds and find current threshold */
5501 new->current_threshold = -1;
5502 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5503 if (thresholds->primary->entries[i].eventfd == eventfd)
5506 new->entries[j] = thresholds->primary->entries[i];
5507 if (new->entries[j].threshold <= usage) {
5509 * new->current_threshold will not be used
5510 * until rcu_assign_pointer(), so it's safe to increment
5513 ++new->current_threshold;
5519 /* Swap primary and spare array */
5520 thresholds->spare = thresholds->primary;
5521 /* If all events are unregistered, free the spare array */
5523 kfree(thresholds->spare);
5524 thresholds->spare = NULL;
5527 rcu_assign_pointer(thresholds->primary, new);
5529 /* To be sure that nobody uses thresholds */
5532 mutex_unlock(&memcg->thresholds_lock);
5535 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5536 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5538 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5539 struct mem_cgroup_eventfd_list *event;
5540 enum res_type type = MEMFILE_TYPE(cft->private);
5542 BUG_ON(type != _OOM_TYPE);
5543 event = kmalloc(sizeof(*event), GFP_KERNEL);
5547 spin_lock(&memcg_oom_lock);
5549 event->eventfd = eventfd;
5550 list_add(&event->list, &memcg->oom_notify);
5552 /* already in OOM ? */
5553 if (atomic_read(&memcg->under_oom))
5554 eventfd_signal(eventfd, 1);
5555 spin_unlock(&memcg_oom_lock);
5560 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5561 struct cftype *cft, struct eventfd_ctx *eventfd)
5563 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5564 struct mem_cgroup_eventfd_list *ev, *tmp;
5565 enum res_type type = MEMFILE_TYPE(cft->private);
5567 BUG_ON(type != _OOM_TYPE);
5569 spin_lock(&memcg_oom_lock);
5571 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5572 if (ev->eventfd == eventfd) {
5573 list_del(&ev->list);
5578 spin_unlock(&memcg_oom_lock);
5581 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5582 struct cftype *cft, struct cgroup_map_cb *cb)
5584 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5586 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5588 if (atomic_read(&memcg->under_oom))
5589 cb->fill(cb, "under_oom", 1);
5591 cb->fill(cb, "under_oom", 0);
5595 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5596 struct cftype *cft, u64 val)
5598 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5599 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5601 /* cannot set to root cgroup and only 0 and 1 are allowed */
5602 if (!parent || !((val == 0) || (val == 1)))
5605 mutex_lock(&memcg_create_mutex);
5606 /* oom-kill-disable is a flag for subhierarchy. */
5607 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5608 mutex_unlock(&memcg_create_mutex);
5611 memcg->oom_kill_disable = val;
5613 memcg_oom_recover(memcg);
5614 mutex_unlock(&memcg_create_mutex);
5618 #ifdef CONFIG_MEMCG_KMEM
5619 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5623 memcg->kmemcg_id = -1;
5624 ret = memcg_propagate_kmem(memcg);
5628 return mem_cgroup_sockets_init(memcg, ss);
5631 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5633 mem_cgroup_sockets_destroy(memcg);
5636 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5638 if (!memcg_kmem_is_active(memcg))
5642 * kmem charges can outlive the cgroup. In the case of slab
5643 * pages, for instance, a page contain objects from various
5644 * processes. As we prevent from taking a reference for every
5645 * such allocation we have to be careful when doing uncharge
5646 * (see memcg_uncharge_kmem) and here during offlining.
5648 * The idea is that that only the _last_ uncharge which sees
5649 * the dead memcg will drop the last reference. An additional
5650 * reference is taken here before the group is marked dead
5651 * which is then paired with css_put during uncharge resp. here.
5653 * Although this might sound strange as this path is called from
5654 * css_offline() when the referencemight have dropped down to 0
5655 * and shouldn't be incremented anymore (css_tryget would fail)
5656 * we do not have other options because of the kmem allocations
5659 css_get(&memcg->css);
5661 memcg_kmem_mark_dead(memcg);
5663 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5666 if (memcg_kmem_test_and_clear_dead(memcg))
5667 css_put(&memcg->css);
5670 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5675 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5679 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5684 static struct cftype mem_cgroup_files[] = {
5686 .name = "usage_in_bytes",
5687 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5688 .read = mem_cgroup_read,
5689 .register_event = mem_cgroup_usage_register_event,
5690 .unregister_event = mem_cgroup_usage_unregister_event,
5693 .name = "max_usage_in_bytes",
5694 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5695 .trigger = mem_cgroup_reset,
5696 .read = mem_cgroup_read,
5699 .name = "limit_in_bytes",
5700 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5701 .write_string = mem_cgroup_write,
5702 .read = mem_cgroup_read,
5705 .name = "soft_limit_in_bytes",
5706 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5707 .write_string = mem_cgroup_write,
5708 .read = mem_cgroup_read,
5712 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5713 .trigger = mem_cgroup_reset,
5714 .read = mem_cgroup_read,
5718 .read_seq_string = memcg_stat_show,
5721 .name = "force_empty",
5722 .trigger = mem_cgroup_force_empty_write,
5725 .name = "use_hierarchy",
5726 .flags = CFTYPE_INSANE,
5727 .write_u64 = mem_cgroup_hierarchy_write,
5728 .read_u64 = mem_cgroup_hierarchy_read,
5731 .name = "swappiness",
5732 .read_u64 = mem_cgroup_swappiness_read,
5733 .write_u64 = mem_cgroup_swappiness_write,
5736 .name = "move_charge_at_immigrate",
5737 .read_u64 = mem_cgroup_move_charge_read,
5738 .write_u64 = mem_cgroup_move_charge_write,
5741 .name = "oom_control",
5742 .read_map = mem_cgroup_oom_control_read,
5743 .write_u64 = mem_cgroup_oom_control_write,
5744 .register_event = mem_cgroup_oom_register_event,
5745 .unregister_event = mem_cgroup_oom_unregister_event,
5746 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5749 .name = "pressure_level",
5750 .register_event = vmpressure_register_event,
5751 .unregister_event = vmpressure_unregister_event,
5755 .name = "numa_stat",
5756 .read_seq_string = memcg_numa_stat_show,
5759 #ifdef CONFIG_MEMCG_KMEM
5761 .name = "kmem.limit_in_bytes",
5762 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5763 .write_string = mem_cgroup_write,
5764 .read = mem_cgroup_read,
5767 .name = "kmem.usage_in_bytes",
5768 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5769 .read = mem_cgroup_read,
5772 .name = "kmem.failcnt",
5773 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5774 .trigger = mem_cgroup_reset,
5775 .read = mem_cgroup_read,
5778 .name = "kmem.max_usage_in_bytes",
5779 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5780 .trigger = mem_cgroup_reset,
5781 .read = mem_cgroup_read,
5783 #ifdef CONFIG_SLABINFO
5785 .name = "kmem.slabinfo",
5786 .read_seq_string = mem_cgroup_slabinfo_read,
5790 { }, /* terminate */
5793 #ifdef CONFIG_MEMCG_SWAP
5794 static struct cftype memsw_cgroup_files[] = {
5796 .name = "memsw.usage_in_bytes",
5797 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5798 .read = mem_cgroup_read,
5799 .register_event = mem_cgroup_usage_register_event,
5800 .unregister_event = mem_cgroup_usage_unregister_event,
5803 .name = "memsw.max_usage_in_bytes",
5804 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5805 .trigger = mem_cgroup_reset,
5806 .read = mem_cgroup_read,
5809 .name = "memsw.limit_in_bytes",
5810 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5811 .write_string = mem_cgroup_write,
5812 .read = mem_cgroup_read,
5815 .name = "memsw.failcnt",
5816 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5817 .trigger = mem_cgroup_reset,
5818 .read = mem_cgroup_read,
5820 { }, /* terminate */
5823 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5825 struct mem_cgroup_per_node *pn;
5826 struct mem_cgroup_per_zone *mz;
5827 int zone, tmp = node;
5829 * This routine is called against possible nodes.
5830 * But it's BUG to call kmalloc() against offline node.
5832 * TODO: this routine can waste much memory for nodes which will
5833 * never be onlined. It's better to use memory hotplug callback
5836 if (!node_state(node, N_NORMAL_MEMORY))
5838 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5842 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5843 mz = &pn->zoneinfo[zone];
5844 lruvec_init(&mz->lruvec);
5847 memcg->nodeinfo[node] = pn;
5851 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5853 kfree(memcg->nodeinfo[node]);
5856 static struct mem_cgroup *mem_cgroup_alloc(void)
5858 struct mem_cgroup *memcg;
5859 size_t size = memcg_size();
5861 /* Can be very big if nr_node_ids is very big */
5862 if (size < PAGE_SIZE)
5863 memcg = kzalloc(size, GFP_KERNEL);
5865 memcg = vzalloc(size);
5870 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5873 spin_lock_init(&memcg->pcp_counter_lock);
5877 if (size < PAGE_SIZE)
5885 * At destroying mem_cgroup, references from swap_cgroup can remain.
5886 * (scanning all at force_empty is too costly...)
5888 * Instead of clearing all references at force_empty, we remember
5889 * the number of reference from swap_cgroup and free mem_cgroup when
5890 * it goes down to 0.
5892 * Removal of cgroup itself succeeds regardless of refs from swap.
5895 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5898 size_t size = memcg_size();
5900 free_css_id(&mem_cgroup_subsys, &memcg->css);
5903 free_mem_cgroup_per_zone_info(memcg, node);
5905 free_percpu(memcg->stat);
5908 * We need to make sure that (at least for now), the jump label
5909 * destruction code runs outside of the cgroup lock. This is because
5910 * get_online_cpus(), which is called from the static_branch update,
5911 * can't be called inside the cgroup_lock. cpusets are the ones
5912 * enforcing this dependency, so if they ever change, we might as well.
5914 * schedule_work() will guarantee this happens. Be careful if you need
5915 * to move this code around, and make sure it is outside
5918 disarm_static_keys(memcg);
5919 if (size < PAGE_SIZE)
5926 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5928 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5930 if (!memcg->res.parent)
5932 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5934 EXPORT_SYMBOL(parent_mem_cgroup);
5936 static struct cgroup_subsys_state * __ref
5937 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5939 struct mem_cgroup *memcg;
5940 long error = -ENOMEM;
5943 memcg = mem_cgroup_alloc();
5945 return ERR_PTR(error);
5948 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5952 if (parent_css == NULL) {
5953 root_mem_cgroup = memcg;
5954 res_counter_init(&memcg->res, NULL);
5955 res_counter_init(&memcg->memsw, NULL);
5956 res_counter_init(&memcg->kmem, NULL);
5959 memcg->last_scanned_node = MAX_NUMNODES;
5960 INIT_LIST_HEAD(&memcg->oom_notify);
5961 memcg->move_charge_at_immigrate = 0;
5962 mutex_init(&memcg->thresholds_lock);
5963 spin_lock_init(&memcg->move_lock);
5964 vmpressure_init(&memcg->vmpressure);
5965 spin_lock_init(&memcg->soft_lock);
5970 __mem_cgroup_free(memcg);
5971 return ERR_PTR(error);
5975 mem_cgroup_css_online(struct cgroup_subsys_state *css)
5977 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5978 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
5984 mutex_lock(&memcg_create_mutex);
5986 memcg->use_hierarchy = parent->use_hierarchy;
5987 memcg->oom_kill_disable = parent->oom_kill_disable;
5988 memcg->swappiness = mem_cgroup_swappiness(parent);
5990 if (parent->use_hierarchy) {
5991 res_counter_init(&memcg->res, &parent->res);
5992 res_counter_init(&memcg->memsw, &parent->memsw);
5993 res_counter_init(&memcg->kmem, &parent->kmem);
5996 * No need to take a reference to the parent because cgroup
5997 * core guarantees its existence.
6000 res_counter_init(&memcg->res, NULL);
6001 res_counter_init(&memcg->memsw, NULL);
6002 res_counter_init(&memcg->kmem, NULL);
6004 * Deeper hierachy with use_hierarchy == false doesn't make
6005 * much sense so let cgroup subsystem know about this
6006 * unfortunate state in our controller.
6008 if (parent != root_mem_cgroup)
6009 mem_cgroup_subsys.broken_hierarchy = true;
6012 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6013 mutex_unlock(&memcg_create_mutex);
6018 * Announce all parents that a group from their hierarchy is gone.
6020 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6022 struct mem_cgroup *parent = memcg;
6024 while ((parent = parent_mem_cgroup(parent)))
6025 mem_cgroup_iter_invalidate(parent);
6028 * if the root memcg is not hierarchical we have to check it
6031 if (!root_mem_cgroup->use_hierarchy)
6032 mem_cgroup_iter_invalidate(root_mem_cgroup);
6035 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6037 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6039 kmem_cgroup_css_offline(memcg);
6041 mem_cgroup_invalidate_reclaim_iterators(memcg);
6042 mem_cgroup_reparent_charges(memcg);
6043 if (memcg->soft_contributed) {
6044 while ((memcg = parent_mem_cgroup(memcg)))
6045 atomic_dec(&memcg->children_in_excess);
6047 mem_cgroup_destroy_all_caches(memcg);
6048 vmpressure_cleanup(&memcg->vmpressure);
6051 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6053 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6055 memcg_destroy_kmem(memcg);
6056 __mem_cgroup_free(memcg);
6060 /* Handlers for move charge at task migration. */
6061 #define PRECHARGE_COUNT_AT_ONCE 256
6062 static int mem_cgroup_do_precharge(unsigned long count)
6065 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6066 struct mem_cgroup *memcg = mc.to;
6068 if (mem_cgroup_is_root(memcg)) {
6069 mc.precharge += count;
6070 /* we don't need css_get for root */
6073 /* try to charge at once */
6075 struct res_counter *dummy;
6077 * "memcg" cannot be under rmdir() because we've already checked
6078 * by cgroup_lock_live_cgroup() that it is not removed and we
6079 * are still under the same cgroup_mutex. So we can postpone
6082 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6084 if (do_swap_account && res_counter_charge(&memcg->memsw,
6085 PAGE_SIZE * count, &dummy)) {
6086 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6089 mc.precharge += count;
6093 /* fall back to one by one charge */
6095 if (signal_pending(current)) {
6099 if (!batch_count--) {
6100 batch_count = PRECHARGE_COUNT_AT_ONCE;
6103 ret = __mem_cgroup_try_charge(NULL,
6104 GFP_KERNEL, 1, &memcg, false);
6106 /* mem_cgroup_clear_mc() will do uncharge later */
6114 * get_mctgt_type - get target type of moving charge
6115 * @vma: the vma the pte to be checked belongs
6116 * @addr: the address corresponding to the pte to be checked
6117 * @ptent: the pte to be checked
6118 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6121 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6122 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6123 * move charge. if @target is not NULL, the page is stored in target->page
6124 * with extra refcnt got(Callers should handle it).
6125 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6126 * target for charge migration. if @target is not NULL, the entry is stored
6129 * Called with pte lock held.
6136 enum mc_target_type {
6142 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6143 unsigned long addr, pte_t ptent)
6145 struct page *page = vm_normal_page(vma, addr, ptent);
6147 if (!page || !page_mapped(page))
6149 if (PageAnon(page)) {
6150 /* we don't move shared anon */
6153 } else if (!move_file())
6154 /* we ignore mapcount for file pages */
6156 if (!get_page_unless_zero(page))
6163 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6164 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6166 struct page *page = NULL;
6167 swp_entry_t ent = pte_to_swp_entry(ptent);
6169 if (!move_anon() || non_swap_entry(ent))
6172 * Because lookup_swap_cache() updates some statistics counter,
6173 * we call find_get_page() with swapper_space directly.
6175 page = find_get_page(swap_address_space(ent), ent.val);
6176 if (do_swap_account)
6177 entry->val = ent.val;
6182 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6183 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6189 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6190 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6192 struct page *page = NULL;
6193 struct address_space *mapping;
6196 if (!vma->vm_file) /* anonymous vma */
6201 mapping = vma->vm_file->f_mapping;
6202 if (pte_none(ptent))
6203 pgoff = linear_page_index(vma, addr);
6204 else /* pte_file(ptent) is true */
6205 pgoff = pte_to_pgoff(ptent);
6207 /* page is moved even if it's not RSS of this task(page-faulted). */
6208 page = find_get_page(mapping, pgoff);
6211 /* shmem/tmpfs may report page out on swap: account for that too. */
6212 if (radix_tree_exceptional_entry(page)) {
6213 swp_entry_t swap = radix_to_swp_entry(page);
6214 if (do_swap_account)
6216 page = find_get_page(swap_address_space(swap), swap.val);
6222 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6223 unsigned long addr, pte_t ptent, union mc_target *target)
6225 struct page *page = NULL;
6226 struct page_cgroup *pc;
6227 enum mc_target_type ret = MC_TARGET_NONE;
6228 swp_entry_t ent = { .val = 0 };
6230 if (pte_present(ptent))
6231 page = mc_handle_present_pte(vma, addr, ptent);
6232 else if (is_swap_pte(ptent))
6233 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6234 else if (pte_none(ptent) || pte_file(ptent))
6235 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6237 if (!page && !ent.val)
6240 pc = lookup_page_cgroup(page);
6242 * Do only loose check w/o page_cgroup lock.
6243 * mem_cgroup_move_account() checks the pc is valid or not under
6246 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6247 ret = MC_TARGET_PAGE;
6249 target->page = page;
6251 if (!ret || !target)
6254 /* There is a swap entry and a page doesn't exist or isn't charged */
6255 if (ent.val && !ret &&
6256 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6257 ret = MC_TARGET_SWAP;
6264 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6266 * We don't consider swapping or file mapped pages because THP does not
6267 * support them for now.
6268 * Caller should make sure that pmd_trans_huge(pmd) is true.
6270 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6271 unsigned long addr, pmd_t pmd, union mc_target *target)
6273 struct page *page = NULL;
6274 struct page_cgroup *pc;
6275 enum mc_target_type ret = MC_TARGET_NONE;
6277 page = pmd_page(pmd);
6278 VM_BUG_ON(!page || !PageHead(page));
6281 pc = lookup_page_cgroup(page);
6282 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6283 ret = MC_TARGET_PAGE;
6286 target->page = page;
6292 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6293 unsigned long addr, pmd_t pmd, union mc_target *target)
6295 return MC_TARGET_NONE;
6299 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6300 unsigned long addr, unsigned long end,
6301 struct mm_walk *walk)
6303 struct vm_area_struct *vma = walk->private;
6307 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6308 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6309 mc.precharge += HPAGE_PMD_NR;
6310 spin_unlock(&vma->vm_mm->page_table_lock);
6314 if (pmd_trans_unstable(pmd))
6316 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6317 for (; addr != end; pte++, addr += PAGE_SIZE)
6318 if (get_mctgt_type(vma, addr, *pte, NULL))
6319 mc.precharge++; /* increment precharge temporarily */
6320 pte_unmap_unlock(pte - 1, ptl);
6326 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6328 unsigned long precharge;
6329 struct vm_area_struct *vma;
6331 down_read(&mm->mmap_sem);
6332 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6333 struct mm_walk mem_cgroup_count_precharge_walk = {
6334 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6338 if (is_vm_hugetlb_page(vma))
6340 walk_page_range(vma->vm_start, vma->vm_end,
6341 &mem_cgroup_count_precharge_walk);
6343 up_read(&mm->mmap_sem);
6345 precharge = mc.precharge;
6351 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6353 unsigned long precharge = mem_cgroup_count_precharge(mm);
6355 VM_BUG_ON(mc.moving_task);
6356 mc.moving_task = current;
6357 return mem_cgroup_do_precharge(precharge);
6360 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6361 static void __mem_cgroup_clear_mc(void)
6363 struct mem_cgroup *from = mc.from;
6364 struct mem_cgroup *to = mc.to;
6367 /* we must uncharge all the leftover precharges from mc.to */
6369 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6373 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6374 * we must uncharge here.
6376 if (mc.moved_charge) {
6377 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6378 mc.moved_charge = 0;
6380 /* we must fixup refcnts and charges */
6381 if (mc.moved_swap) {
6382 /* uncharge swap account from the old cgroup */
6383 if (!mem_cgroup_is_root(mc.from))
6384 res_counter_uncharge(&mc.from->memsw,
6385 PAGE_SIZE * mc.moved_swap);
6387 for (i = 0; i < mc.moved_swap; i++)
6388 css_put(&mc.from->css);
6390 if (!mem_cgroup_is_root(mc.to)) {
6392 * we charged both to->res and to->memsw, so we should
6395 res_counter_uncharge(&mc.to->res,
6396 PAGE_SIZE * mc.moved_swap);
6398 /* we've already done css_get(mc.to) */
6401 memcg_oom_recover(from);
6402 memcg_oom_recover(to);
6403 wake_up_all(&mc.waitq);
6406 static void mem_cgroup_clear_mc(void)
6408 struct mem_cgroup *from = mc.from;
6411 * we must clear moving_task before waking up waiters at the end of
6414 mc.moving_task = NULL;
6415 __mem_cgroup_clear_mc();
6416 spin_lock(&mc.lock);
6419 spin_unlock(&mc.lock);
6420 mem_cgroup_end_move(from);
6423 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6424 struct cgroup_taskset *tset)
6426 struct task_struct *p = cgroup_taskset_first(tset);
6428 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6429 unsigned long move_charge_at_immigrate;
6432 * We are now commited to this value whatever it is. Changes in this
6433 * tunable will only affect upcoming migrations, not the current one.
6434 * So we need to save it, and keep it going.
6436 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6437 if (move_charge_at_immigrate) {
6438 struct mm_struct *mm;
6439 struct mem_cgroup *from = mem_cgroup_from_task(p);
6441 VM_BUG_ON(from == memcg);
6443 mm = get_task_mm(p);
6446 /* We move charges only when we move a owner of the mm */
6447 if (mm->owner == p) {
6450 VM_BUG_ON(mc.precharge);
6451 VM_BUG_ON(mc.moved_charge);
6452 VM_BUG_ON(mc.moved_swap);
6453 mem_cgroup_start_move(from);
6454 spin_lock(&mc.lock);
6457 mc.immigrate_flags = move_charge_at_immigrate;
6458 spin_unlock(&mc.lock);
6459 /* We set mc.moving_task later */
6461 ret = mem_cgroup_precharge_mc(mm);
6463 mem_cgroup_clear_mc();
6470 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6471 struct cgroup_taskset *tset)
6473 mem_cgroup_clear_mc();
6476 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6477 unsigned long addr, unsigned long end,
6478 struct mm_walk *walk)
6481 struct vm_area_struct *vma = walk->private;
6484 enum mc_target_type target_type;
6485 union mc_target target;
6487 struct page_cgroup *pc;
6490 * We don't take compound_lock() here but no race with splitting thp
6492 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6493 * under splitting, which means there's no concurrent thp split,
6494 * - if another thread runs into split_huge_page() just after we
6495 * entered this if-block, the thread must wait for page table lock
6496 * to be unlocked in __split_huge_page_splitting(), where the main
6497 * part of thp split is not executed yet.
6499 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6500 if (mc.precharge < HPAGE_PMD_NR) {
6501 spin_unlock(&vma->vm_mm->page_table_lock);
6504 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6505 if (target_type == MC_TARGET_PAGE) {
6507 if (!isolate_lru_page(page)) {
6508 pc = lookup_page_cgroup(page);
6509 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6510 pc, mc.from, mc.to)) {
6511 mc.precharge -= HPAGE_PMD_NR;
6512 mc.moved_charge += HPAGE_PMD_NR;
6514 putback_lru_page(page);
6518 spin_unlock(&vma->vm_mm->page_table_lock);
6522 if (pmd_trans_unstable(pmd))
6525 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6526 for (; addr != end; addr += PAGE_SIZE) {
6527 pte_t ptent = *(pte++);
6533 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6534 case MC_TARGET_PAGE:
6536 if (isolate_lru_page(page))
6538 pc = lookup_page_cgroup(page);
6539 if (!mem_cgroup_move_account(page, 1, pc,
6542 /* we uncharge from mc.from later. */
6545 putback_lru_page(page);
6546 put: /* get_mctgt_type() gets the page */
6549 case MC_TARGET_SWAP:
6551 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6553 /* we fixup refcnts and charges later. */
6561 pte_unmap_unlock(pte - 1, ptl);
6566 * We have consumed all precharges we got in can_attach().
6567 * We try charge one by one, but don't do any additional
6568 * charges to mc.to if we have failed in charge once in attach()
6571 ret = mem_cgroup_do_precharge(1);
6579 static void mem_cgroup_move_charge(struct mm_struct *mm)
6581 struct vm_area_struct *vma;
6583 lru_add_drain_all();
6585 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6587 * Someone who are holding the mmap_sem might be waiting in
6588 * waitq. So we cancel all extra charges, wake up all waiters,
6589 * and retry. Because we cancel precharges, we might not be able
6590 * to move enough charges, but moving charge is a best-effort
6591 * feature anyway, so it wouldn't be a big problem.
6593 __mem_cgroup_clear_mc();
6597 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6599 struct mm_walk mem_cgroup_move_charge_walk = {
6600 .pmd_entry = mem_cgroup_move_charge_pte_range,
6604 if (is_vm_hugetlb_page(vma))
6606 ret = walk_page_range(vma->vm_start, vma->vm_end,
6607 &mem_cgroup_move_charge_walk);
6610 * means we have consumed all precharges and failed in
6611 * doing additional charge. Just abandon here.
6615 up_read(&mm->mmap_sem);
6618 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6619 struct cgroup_taskset *tset)
6621 struct task_struct *p = cgroup_taskset_first(tset);
6622 struct mm_struct *mm = get_task_mm(p);
6626 mem_cgroup_move_charge(mm);
6630 mem_cgroup_clear_mc();
6632 #else /* !CONFIG_MMU */
6633 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6634 struct cgroup_taskset *tset)
6638 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6639 struct cgroup_taskset *tset)
6642 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6643 struct cgroup_taskset *tset)
6649 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6650 * to verify sane_behavior flag on each mount attempt.
6652 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6655 * use_hierarchy is forced with sane_behavior. cgroup core
6656 * guarantees that @root doesn't have any children, so turning it
6657 * on for the root memcg is enough.
6659 if (cgroup_sane_behavior(root_css->cgroup))
6660 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6663 struct cgroup_subsys mem_cgroup_subsys = {
6665 .subsys_id = mem_cgroup_subsys_id,
6666 .css_alloc = mem_cgroup_css_alloc,
6667 .css_online = mem_cgroup_css_online,
6668 .css_offline = mem_cgroup_css_offline,
6669 .css_free = mem_cgroup_css_free,
6670 .can_attach = mem_cgroup_can_attach,
6671 .cancel_attach = mem_cgroup_cancel_attach,
6672 .attach = mem_cgroup_move_task,
6673 .bind = mem_cgroup_bind,
6674 .base_cftypes = mem_cgroup_files,
6679 #ifdef CONFIG_MEMCG_SWAP
6680 static int __init enable_swap_account(char *s)
6682 if (!strcmp(s, "1"))
6683 really_do_swap_account = 1;
6684 else if (!strcmp(s, "0"))
6685 really_do_swap_account = 0;
6688 __setup("swapaccount=", enable_swap_account);
6690 static void __init memsw_file_init(void)
6692 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6695 static void __init enable_swap_cgroup(void)
6697 if (!mem_cgroup_disabled() && really_do_swap_account) {
6698 do_swap_account = 1;
6704 static void __init enable_swap_cgroup(void)
6710 * subsys_initcall() for memory controller.
6712 * Some parts like hotcpu_notifier() have to be initialized from this context
6713 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6714 * everything that doesn't depend on a specific mem_cgroup structure should
6715 * be initialized from here.
6717 static int __init mem_cgroup_init(void)
6719 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6720 enable_swap_cgroup();
6724 subsys_initcall(mem_cgroup_init);