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-limited 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 notified 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.
867 * We track children even outside the hierarchy for the root
868 * cgroup because tree walk starting at root should visit
869 * all cgroups and we want to prevent from pointless tree
870 * walk if no children is below the limit.
872 while (delta && (parent = parent_mem_cgroup(parent)))
873 atomic_add(delta, &parent->children_in_excess);
874 if (memcg != root_mem_cgroup && !root_mem_cgroup->use_hierarchy)
875 atomic_add(delta, &root_mem_cgroup->children_in_excess);
876 spin_unlock(&memcg->soft_lock);
880 * Check events in order.
883 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
886 /* threshold event is triggered in finer grain than soft limit */
887 if (unlikely(mem_cgroup_event_ratelimit(memcg,
888 MEM_CGROUP_TARGET_THRESH))) {
890 bool do_numainfo __maybe_unused;
892 do_softlimit = mem_cgroup_event_ratelimit(memcg,
893 MEM_CGROUP_TARGET_SOFTLIMIT);
895 do_numainfo = mem_cgroup_event_ratelimit(memcg,
896 MEM_CGROUP_TARGET_NUMAINFO);
900 mem_cgroup_threshold(memcg);
901 if (unlikely(do_softlimit))
902 mem_cgroup_update_soft_limit(memcg);
904 if (unlikely(do_numainfo))
905 atomic_inc(&memcg->numainfo_events);
911 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
914 * mm_update_next_owner() may clear mm->owner to NULL
915 * if it races with swapoff, page migration, etc.
916 * So this can be called with p == NULL.
921 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
924 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
926 struct mem_cgroup *memcg = NULL;
931 * Because we have no locks, mm->owner's may be being moved to other
932 * cgroup. We use css_tryget() here even if this looks
933 * pessimistic (rather than adding locks here).
937 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
938 if (unlikely(!memcg))
940 } while (!css_tryget(&memcg->css));
945 static enum mem_cgroup_filter_t
946 mem_cgroup_filter(struct mem_cgroup *memcg, struct mem_cgroup *root,
947 mem_cgroup_iter_filter cond)
951 return cond(memcg, root);
955 * Returns a next (in a pre-order walk) alive memcg (with elevated css
956 * ref. count) or NULL if the whole root's subtree has been visited.
958 * helper function to be used by mem_cgroup_iter
960 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
961 struct mem_cgroup *last_visited, mem_cgroup_iter_filter cond)
963 struct cgroup_subsys_state *prev_css, *next_css;
965 prev_css = last_visited ? &last_visited->css : NULL;
967 next_css = css_next_descendant_pre(prev_css, &root->css);
970 * Even if we found a group we have to make sure it is
971 * alive. css && !memcg means that the groups should be
972 * skipped and we should continue the tree walk.
973 * last_visited css is safe to use because it is
974 * protected by css_get and the tree walk is rcu safe.
977 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
979 switch (mem_cgroup_filter(mem, root, cond)) {
987 * css_rightmost_descendant is not an optimal way to
988 * skip through a subtree (especially for imbalanced
989 * trees leaning to right) but that's what we have right
990 * now. More effective solution would be traversing
991 * right-up for first non-NULL without calling
992 * css_next_descendant_pre afterwards.
994 prev_css = css_rightmost_descendant(next_css);
997 if (css_tryget(&mem->css))
1000 prev_css = next_css;
1010 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1013 * When a group in the hierarchy below root is destroyed, the
1014 * hierarchy iterator can no longer be trusted since it might
1015 * have pointed to the destroyed group. Invalidate it.
1017 atomic_inc(&root->dead_count);
1020 static struct mem_cgroup *
1021 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1022 struct mem_cgroup *root,
1025 struct mem_cgroup *position = NULL;
1027 * A cgroup destruction happens in two stages: offlining and
1028 * release. They are separated by a RCU grace period.
1030 * If the iterator is valid, we may still race with an
1031 * offlining. The RCU lock ensures the object won't be
1032 * released, tryget will fail if we lost the race.
1034 *sequence = atomic_read(&root->dead_count);
1035 if (iter->last_dead_count == *sequence) {
1037 position = iter->last_visited;
1038 if (position && !css_tryget(&position->css))
1044 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1045 struct mem_cgroup *last_visited,
1046 struct mem_cgroup *new_position,
1050 css_put(&last_visited->css);
1052 * We store the sequence count from the time @last_visited was
1053 * loaded successfully instead of rereading it here so that we
1054 * don't lose destruction events in between. We could have
1055 * raced with the destruction of @new_position after all.
1057 iter->last_visited = new_position;
1059 iter->last_dead_count = sequence;
1063 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1064 * @root: hierarchy root
1065 * @prev: previously returned memcg, NULL on first invocation
1066 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1067 * @cond: filter for visited nodes, NULL for no filter
1069 * Returns references to children of the hierarchy below @root, or
1070 * @root itself, or %NULL after a full round-trip.
1072 * Caller must pass the return value in @prev on subsequent
1073 * invocations for reference counting, or use mem_cgroup_iter_break()
1074 * to cancel a hierarchy walk before the round-trip is complete.
1076 * Reclaimers can specify a zone and a priority level in @reclaim to
1077 * divide up the memcgs in the hierarchy among all concurrent
1078 * reclaimers operating on the same zone and priority.
1080 struct mem_cgroup *mem_cgroup_iter_cond(struct mem_cgroup *root,
1081 struct mem_cgroup *prev,
1082 struct mem_cgroup_reclaim_cookie *reclaim,
1083 mem_cgroup_iter_filter cond)
1085 struct mem_cgroup *memcg = NULL;
1086 struct mem_cgroup *last_visited = NULL;
1088 if (mem_cgroup_disabled()) {
1089 /* first call must return non-NULL, second return NULL */
1090 return (struct mem_cgroup *)(unsigned long)!prev;
1094 root = root_mem_cgroup;
1096 if (prev && !reclaim)
1097 last_visited = prev;
1099 if (!root->use_hierarchy && root != root_mem_cgroup) {
1102 if (mem_cgroup_filter(root, root, cond) == VISIT)
1109 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1110 int uninitialized_var(seq);
1113 int nid = zone_to_nid(reclaim->zone);
1114 int zid = zone_idx(reclaim->zone);
1115 struct mem_cgroup_per_zone *mz;
1117 mz = mem_cgroup_zoneinfo(root, nid, zid);
1118 iter = &mz->reclaim_iter[reclaim->priority];
1119 if (prev && reclaim->generation != iter->generation) {
1120 iter->last_visited = NULL;
1124 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1127 memcg = __mem_cgroup_iter_next(root, last_visited, cond);
1130 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1134 else if (!prev && memcg)
1135 reclaim->generation = iter->generation;
1139 * We have finished the whole tree walk or no group has been
1140 * visited because filter told us to skip the root node.
1142 if (!memcg && (prev || (cond && !last_visited)))
1148 if (prev && prev != root)
1149 css_put(&prev->css);
1155 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1156 * @root: hierarchy root
1157 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1159 void mem_cgroup_iter_break(struct mem_cgroup *root,
1160 struct mem_cgroup *prev)
1163 root = root_mem_cgroup;
1164 if (prev && prev != root)
1165 css_put(&prev->css);
1169 * Iteration constructs for visiting all cgroups (under a tree). If
1170 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1171 * be used for reference counting.
1173 #define for_each_mem_cgroup_tree(iter, root) \
1174 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1176 iter = mem_cgroup_iter(root, iter, NULL))
1178 #define for_each_mem_cgroup(iter) \
1179 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1181 iter = mem_cgroup_iter(NULL, iter, NULL))
1183 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1185 struct mem_cgroup *memcg;
1188 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1189 if (unlikely(!memcg))
1194 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1197 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1205 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1208 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1209 * @zone: zone of the wanted lruvec
1210 * @memcg: memcg of the wanted lruvec
1212 * Returns the lru list vector holding pages for the given @zone and
1213 * @mem. This can be the global zone lruvec, if the memory controller
1216 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1217 struct mem_cgroup *memcg)
1219 struct mem_cgroup_per_zone *mz;
1220 struct lruvec *lruvec;
1222 if (mem_cgroup_disabled()) {
1223 lruvec = &zone->lruvec;
1227 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1228 lruvec = &mz->lruvec;
1231 * Since a node can be onlined after the mem_cgroup was created,
1232 * we have to be prepared to initialize lruvec->zone here;
1233 * and if offlined then reonlined, we need to reinitialize it.
1235 if (unlikely(lruvec->zone != zone))
1236 lruvec->zone = zone;
1241 * Following LRU functions are allowed to be used without PCG_LOCK.
1242 * Operations are called by routine of global LRU independently from memcg.
1243 * What we have to take care of here is validness of pc->mem_cgroup.
1245 * Changes to pc->mem_cgroup happens when
1248 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1249 * It is added to LRU before charge.
1250 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1251 * When moving account, the page is not on LRU. It's isolated.
1255 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1257 * @zone: zone of the page
1259 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1261 struct mem_cgroup_per_zone *mz;
1262 struct mem_cgroup *memcg;
1263 struct page_cgroup *pc;
1264 struct lruvec *lruvec;
1266 if (mem_cgroup_disabled()) {
1267 lruvec = &zone->lruvec;
1271 pc = lookup_page_cgroup(page);
1272 memcg = pc->mem_cgroup;
1275 * Surreptitiously switch any uncharged offlist page to root:
1276 * an uncharged page off lru does nothing to secure
1277 * its former mem_cgroup from sudden removal.
1279 * Our caller holds lru_lock, and PageCgroupUsed is updated
1280 * under page_cgroup lock: between them, they make all uses
1281 * of pc->mem_cgroup safe.
1283 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1284 pc->mem_cgroup = memcg = root_mem_cgroup;
1286 mz = page_cgroup_zoneinfo(memcg, page);
1287 lruvec = &mz->lruvec;
1290 * Since a node can be onlined after the mem_cgroup was created,
1291 * we have to be prepared to initialize lruvec->zone here;
1292 * and if offlined then reonlined, we need to reinitialize it.
1294 if (unlikely(lruvec->zone != zone))
1295 lruvec->zone = zone;
1300 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1301 * @lruvec: mem_cgroup per zone lru vector
1302 * @lru: index of lru list the page is sitting on
1303 * @nr_pages: positive when adding or negative when removing
1305 * This function must be called when a page is added to or removed from an
1308 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1311 struct mem_cgroup_per_zone *mz;
1312 unsigned long *lru_size;
1314 if (mem_cgroup_disabled())
1317 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1318 lru_size = mz->lru_size + lru;
1319 *lru_size += nr_pages;
1320 VM_BUG_ON((long)(*lru_size) < 0);
1324 * Checks whether given mem is same or in the root_mem_cgroup's
1327 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1328 struct mem_cgroup *memcg)
1330 if (root_memcg == memcg)
1332 if (!root_memcg->use_hierarchy || !memcg)
1334 return css_is_ancestor(&memcg->css, &root_memcg->css);
1337 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1338 struct mem_cgroup *memcg)
1343 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1348 bool task_in_mem_cgroup(struct task_struct *task,
1349 const struct mem_cgroup *memcg)
1351 struct mem_cgroup *curr = NULL;
1352 struct task_struct *p;
1355 p = find_lock_task_mm(task);
1357 curr = try_get_mem_cgroup_from_mm(p->mm);
1361 * All threads may have already detached their mm's, but the oom
1362 * killer still needs to detect if they have already been oom
1363 * killed to prevent needlessly killing additional tasks.
1366 curr = mem_cgroup_from_task(task);
1368 css_get(&curr->css);
1374 * We should check use_hierarchy of "memcg" not "curr". Because checking
1375 * use_hierarchy of "curr" here make this function true if hierarchy is
1376 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1377 * hierarchy(even if use_hierarchy is disabled in "memcg").
1379 ret = mem_cgroup_same_or_subtree(memcg, curr);
1380 css_put(&curr->css);
1384 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1386 unsigned long inactive_ratio;
1387 unsigned long inactive;
1388 unsigned long active;
1391 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1392 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1394 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1396 inactive_ratio = int_sqrt(10 * gb);
1400 return inactive * inactive_ratio < active;
1403 #define mem_cgroup_from_res_counter(counter, member) \
1404 container_of(counter, struct mem_cgroup, member)
1407 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1408 * @memcg: the memory cgroup
1410 * Returns the maximum amount of memory @mem can be charged with, in
1413 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1415 unsigned long long margin;
1417 margin = res_counter_margin(&memcg->res);
1418 if (do_swap_account)
1419 margin = min(margin, res_counter_margin(&memcg->memsw));
1420 return margin >> PAGE_SHIFT;
1423 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1426 if (!css_parent(&memcg->css))
1427 return vm_swappiness;
1429 return memcg->swappiness;
1433 * memcg->moving_account is used for checking possibility that some thread is
1434 * calling move_account(). When a thread on CPU-A starts moving pages under
1435 * a memcg, other threads should check memcg->moving_account under
1436 * rcu_read_lock(), like this:
1440 * memcg->moving_account+1 if (memcg->mocing_account)
1442 * synchronize_rcu() update something.
1447 /* for quick checking without looking up memcg */
1448 atomic_t memcg_moving __read_mostly;
1450 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1452 atomic_inc(&memcg_moving);
1453 atomic_inc(&memcg->moving_account);
1457 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1460 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1461 * We check NULL in callee rather than caller.
1464 atomic_dec(&memcg_moving);
1465 atomic_dec(&memcg->moving_account);
1470 * 2 routines for checking "mem" is under move_account() or not.
1472 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1473 * is used for avoiding races in accounting. If true,
1474 * pc->mem_cgroup may be overwritten.
1476 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1477 * under hierarchy of moving cgroups. This is for
1478 * waiting at hith-memory prressure caused by "move".
1481 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1483 VM_BUG_ON(!rcu_read_lock_held());
1484 return atomic_read(&memcg->moving_account) > 0;
1487 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1489 struct mem_cgroup *from;
1490 struct mem_cgroup *to;
1493 * Unlike task_move routines, we access mc.to, mc.from not under
1494 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1496 spin_lock(&mc.lock);
1502 ret = mem_cgroup_same_or_subtree(memcg, from)
1503 || mem_cgroup_same_or_subtree(memcg, to);
1505 spin_unlock(&mc.lock);
1509 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1511 if (mc.moving_task && current != mc.moving_task) {
1512 if (mem_cgroup_under_move(memcg)) {
1514 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1515 /* moving charge context might have finished. */
1518 finish_wait(&mc.waitq, &wait);
1526 * Take this lock when
1527 * - a code tries to modify page's memcg while it's USED.
1528 * - a code tries to modify page state accounting in a memcg.
1529 * see mem_cgroup_stolen(), too.
1531 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1532 unsigned long *flags)
1534 spin_lock_irqsave(&memcg->move_lock, *flags);
1537 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1538 unsigned long *flags)
1540 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1543 #define K(x) ((x) << (PAGE_SHIFT-10))
1545 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1546 * @memcg: The memory cgroup that went over limit
1547 * @p: Task that is going to be killed
1549 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1552 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1554 struct cgroup *task_cgrp;
1555 struct cgroup *mem_cgrp;
1557 * Need a buffer in BSS, can't rely on allocations. The code relies
1558 * on the assumption that OOM is serialized for memory controller.
1559 * If this assumption is broken, revisit this code.
1561 static char memcg_name[PATH_MAX];
1563 struct mem_cgroup *iter;
1571 mem_cgrp = memcg->css.cgroup;
1572 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1574 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1577 * Unfortunately, we are unable to convert to a useful name
1578 * But we'll still print out the usage information
1585 pr_info("Task in %s killed", memcg_name);
1588 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1596 * Continues from above, so we don't need an KERN_ level
1598 pr_cont(" as a result of limit of %s\n", memcg_name);
1601 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1602 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1603 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1604 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1605 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1606 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1607 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1608 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1609 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1610 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1611 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1612 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1614 for_each_mem_cgroup_tree(iter, memcg) {
1615 pr_info("Memory cgroup stats");
1618 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1620 pr_cont(" for %s", memcg_name);
1624 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1625 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1627 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1628 K(mem_cgroup_read_stat(iter, i)));
1631 for (i = 0; i < NR_LRU_LISTS; i++)
1632 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1633 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1640 * This function returns the number of memcg under hierarchy tree. Returns
1641 * 1(self count) if no children.
1643 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1646 struct mem_cgroup *iter;
1648 for_each_mem_cgroup_tree(iter, memcg)
1654 * Return the memory (and swap, if configured) limit for a memcg.
1656 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1660 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1663 * Do not consider swap space if we cannot swap due to swappiness
1665 if (mem_cgroup_swappiness(memcg)) {
1668 limit += total_swap_pages << PAGE_SHIFT;
1669 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1672 * If memsw is finite and limits the amount of swap space
1673 * available to this memcg, return that limit.
1675 limit = min(limit, memsw);
1681 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1684 struct mem_cgroup *iter;
1685 unsigned long chosen_points = 0;
1686 unsigned long totalpages;
1687 unsigned int points = 0;
1688 struct task_struct *chosen = NULL;
1691 * If current has a pending SIGKILL or is exiting, then automatically
1692 * select it. The goal is to allow it to allocate so that it may
1693 * quickly exit and free its memory.
1695 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1696 set_thread_flag(TIF_MEMDIE);
1700 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1701 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1702 for_each_mem_cgroup_tree(iter, memcg) {
1703 struct css_task_iter it;
1704 struct task_struct *task;
1706 css_task_iter_start(&iter->css, &it);
1707 while ((task = css_task_iter_next(&it))) {
1708 switch (oom_scan_process_thread(task, totalpages, NULL,
1710 case OOM_SCAN_SELECT:
1712 put_task_struct(chosen);
1714 chosen_points = ULONG_MAX;
1715 get_task_struct(chosen);
1717 case OOM_SCAN_CONTINUE:
1719 case OOM_SCAN_ABORT:
1720 css_task_iter_end(&it);
1721 mem_cgroup_iter_break(memcg, iter);
1723 put_task_struct(chosen);
1728 points = oom_badness(task, memcg, NULL, totalpages);
1729 if (points > chosen_points) {
1731 put_task_struct(chosen);
1733 chosen_points = points;
1734 get_task_struct(chosen);
1737 css_task_iter_end(&it);
1742 points = chosen_points * 1000 / totalpages;
1743 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1744 NULL, "Memory cgroup out of memory");
1747 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1749 unsigned long flags)
1751 unsigned long total = 0;
1752 bool noswap = false;
1755 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1757 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1760 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1762 drain_all_stock_async(memcg);
1763 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1765 * Allow limit shrinkers, which are triggered directly
1766 * by userspace, to catch signals and stop reclaim
1767 * after minimal progress, regardless of the margin.
1769 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1771 if (mem_cgroup_margin(memcg))
1774 * If nothing was reclaimed after two attempts, there
1775 * may be no reclaimable pages in this hierarchy.
1783 #if MAX_NUMNODES > 1
1785 * test_mem_cgroup_node_reclaimable
1786 * @memcg: the target memcg
1787 * @nid: the node ID to be checked.
1788 * @noswap : specify true here if the user wants flle only information.
1790 * This function returns whether the specified memcg contains any
1791 * reclaimable pages on a node. Returns true if there are any reclaimable
1792 * pages in the node.
1794 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1795 int nid, bool noswap)
1797 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1799 if (noswap || !total_swap_pages)
1801 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1808 * Always updating the nodemask is not very good - even if we have an empty
1809 * list or the wrong list here, we can start from some node and traverse all
1810 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1813 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1817 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1818 * pagein/pageout changes since the last update.
1820 if (!atomic_read(&memcg->numainfo_events))
1822 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1825 /* make a nodemask where this memcg uses memory from */
1826 memcg->scan_nodes = node_states[N_MEMORY];
1828 for_each_node_mask(nid, node_states[N_MEMORY]) {
1830 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1831 node_clear(nid, memcg->scan_nodes);
1834 atomic_set(&memcg->numainfo_events, 0);
1835 atomic_set(&memcg->numainfo_updating, 0);
1839 * Selecting a node where we start reclaim from. Because what we need is just
1840 * reducing usage counter, start from anywhere is O,K. Considering
1841 * memory reclaim from current node, there are pros. and cons.
1843 * Freeing memory from current node means freeing memory from a node which
1844 * we'll use or we've used. So, it may make LRU bad. And if several threads
1845 * hit limits, it will see a contention on a node. But freeing from remote
1846 * node means more costs for memory reclaim because of memory latency.
1848 * Now, we use round-robin. Better algorithm is welcomed.
1850 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1854 mem_cgroup_may_update_nodemask(memcg);
1855 node = memcg->last_scanned_node;
1857 node = next_node(node, memcg->scan_nodes);
1858 if (node == MAX_NUMNODES)
1859 node = first_node(memcg->scan_nodes);
1861 * We call this when we hit limit, not when pages are added to LRU.
1862 * No LRU may hold pages because all pages are UNEVICTABLE or
1863 * memcg is too small and all pages are not on LRU. In that case,
1864 * we use curret node.
1866 if (unlikely(node == MAX_NUMNODES))
1867 node = numa_node_id();
1869 memcg->last_scanned_node = node;
1874 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1882 * A group is eligible for the soft limit reclaim under the given root
1884 * a) it is over its soft limit
1885 * b) any parent up the hierarchy is over its soft limit
1887 * If the given group doesn't have any children over the limit then it
1888 * doesn't make any sense to iterate its subtree.
1890 enum mem_cgroup_filter_t
1891 mem_cgroup_soft_reclaim_eligible(struct mem_cgroup *memcg,
1892 struct mem_cgroup *root)
1894 struct mem_cgroup *parent;
1897 memcg = root_mem_cgroup;
1900 if (res_counter_soft_limit_excess(&memcg->res))
1904 * If any parent up to the root in the hierarchy is over its soft limit
1905 * then we have to obey and reclaim from this group as well.
1907 while ((parent = parent_mem_cgroup(parent))) {
1908 if (res_counter_soft_limit_excess(&parent->res))
1914 if (!atomic_read(&memcg->children_in_excess))
1920 * Check OOM-Killer is already running under our hierarchy.
1921 * If someone is running, return false.
1922 * Has to be called with memcg_oom_lock
1924 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1926 struct mem_cgroup *iter, *failed = NULL;
1928 for_each_mem_cgroup_tree(iter, memcg) {
1929 if (iter->oom_lock) {
1931 * this subtree of our hierarchy is already locked
1932 * so we cannot give a lock.
1935 mem_cgroup_iter_break(memcg, iter);
1938 iter->oom_lock = true;
1945 * OK, we failed to lock the whole subtree so we have to clean up
1946 * what we set up to the failing subtree
1948 for_each_mem_cgroup_tree(iter, memcg) {
1949 if (iter == failed) {
1950 mem_cgroup_iter_break(memcg, iter);
1953 iter->oom_lock = false;
1959 * Has to be called with memcg_oom_lock
1961 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1963 struct mem_cgroup *iter;
1965 for_each_mem_cgroup_tree(iter, memcg)
1966 iter->oom_lock = false;
1970 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1972 struct mem_cgroup *iter;
1974 for_each_mem_cgroup_tree(iter, memcg)
1975 atomic_inc(&iter->under_oom);
1978 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1980 struct mem_cgroup *iter;
1983 * When a new child is created while the hierarchy is under oom,
1984 * mem_cgroup_oom_lock() may not be called. We have to use
1985 * atomic_add_unless() here.
1987 for_each_mem_cgroup_tree(iter, memcg)
1988 atomic_add_unless(&iter->under_oom, -1, 0);
1991 static DEFINE_SPINLOCK(memcg_oom_lock);
1992 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1994 struct oom_wait_info {
1995 struct mem_cgroup *memcg;
1999 static int memcg_oom_wake_function(wait_queue_t *wait,
2000 unsigned mode, int sync, void *arg)
2002 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2003 struct mem_cgroup *oom_wait_memcg;
2004 struct oom_wait_info *oom_wait_info;
2006 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2007 oom_wait_memcg = oom_wait_info->memcg;
2010 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2011 * Then we can use css_is_ancestor without taking care of RCU.
2013 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2014 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2016 return autoremove_wake_function(wait, mode, sync, arg);
2019 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2021 /* for filtering, pass "memcg" as argument. */
2022 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2025 static void memcg_oom_recover(struct mem_cgroup *memcg)
2027 if (memcg && atomic_read(&memcg->under_oom))
2028 memcg_wakeup_oom(memcg);
2032 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2034 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2037 struct oom_wait_info owait;
2038 bool locked, need_to_kill;
2040 owait.memcg = memcg;
2041 owait.wait.flags = 0;
2042 owait.wait.func = memcg_oom_wake_function;
2043 owait.wait.private = current;
2044 INIT_LIST_HEAD(&owait.wait.task_list);
2045 need_to_kill = true;
2046 mem_cgroup_mark_under_oom(memcg);
2048 /* At first, try to OOM lock hierarchy under memcg.*/
2049 spin_lock(&memcg_oom_lock);
2050 locked = mem_cgroup_oom_lock(memcg);
2052 * Even if signal_pending(), we can't quit charge() loop without
2053 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2054 * under OOM is always welcomed, use TASK_KILLABLE here.
2056 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2057 if (!locked || memcg->oom_kill_disable)
2058 need_to_kill = false;
2060 mem_cgroup_oom_notify(memcg);
2061 spin_unlock(&memcg_oom_lock);
2064 finish_wait(&memcg_oom_waitq, &owait.wait);
2065 mem_cgroup_out_of_memory(memcg, mask, order);
2068 finish_wait(&memcg_oom_waitq, &owait.wait);
2070 spin_lock(&memcg_oom_lock);
2072 mem_cgroup_oom_unlock(memcg);
2073 memcg_wakeup_oom(memcg);
2074 spin_unlock(&memcg_oom_lock);
2076 mem_cgroup_unmark_under_oom(memcg);
2078 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2080 /* Give chance to dying process */
2081 schedule_timeout_uninterruptible(1);
2086 * Currently used to update mapped file statistics, but the routine can be
2087 * generalized to update other statistics as well.
2089 * Notes: Race condition
2091 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2092 * it tends to be costly. But considering some conditions, we doesn't need
2093 * to do so _always_.
2095 * Considering "charge", lock_page_cgroup() is not required because all
2096 * file-stat operations happen after a page is attached to radix-tree. There
2097 * are no race with "charge".
2099 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2100 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2101 * if there are race with "uncharge". Statistics itself is properly handled
2104 * Considering "move", this is an only case we see a race. To make the race
2105 * small, we check mm->moving_account and detect there are possibility of race
2106 * If there is, we take a lock.
2109 void __mem_cgroup_begin_update_page_stat(struct page *page,
2110 bool *locked, unsigned long *flags)
2112 struct mem_cgroup *memcg;
2113 struct page_cgroup *pc;
2115 pc = lookup_page_cgroup(page);
2117 memcg = pc->mem_cgroup;
2118 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2121 * If this memory cgroup is not under account moving, we don't
2122 * need to take move_lock_mem_cgroup(). Because we already hold
2123 * rcu_read_lock(), any calls to move_account will be delayed until
2124 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2126 if (!mem_cgroup_stolen(memcg))
2129 move_lock_mem_cgroup(memcg, flags);
2130 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2131 move_unlock_mem_cgroup(memcg, flags);
2137 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2139 struct page_cgroup *pc = lookup_page_cgroup(page);
2142 * It's guaranteed that pc->mem_cgroup never changes while
2143 * lock is held because a routine modifies pc->mem_cgroup
2144 * should take move_lock_mem_cgroup().
2146 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2149 void mem_cgroup_update_page_stat(struct page *page,
2150 enum mem_cgroup_page_stat_item idx, int val)
2152 struct mem_cgroup *memcg;
2153 struct page_cgroup *pc = lookup_page_cgroup(page);
2154 unsigned long uninitialized_var(flags);
2156 if (mem_cgroup_disabled())
2159 memcg = pc->mem_cgroup;
2160 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2164 case MEMCG_NR_FILE_MAPPED:
2165 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2171 this_cpu_add(memcg->stat->count[idx], val);
2175 * size of first charge trial. "32" comes from vmscan.c's magic value.
2176 * TODO: maybe necessary to use big numbers in big irons.
2178 #define CHARGE_BATCH 32U
2179 struct memcg_stock_pcp {
2180 struct mem_cgroup *cached; /* this never be root cgroup */
2181 unsigned int nr_pages;
2182 struct work_struct work;
2183 unsigned long flags;
2184 #define FLUSHING_CACHED_CHARGE 0
2186 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2187 static DEFINE_MUTEX(percpu_charge_mutex);
2190 * consume_stock: Try to consume stocked charge on this cpu.
2191 * @memcg: memcg to consume from.
2192 * @nr_pages: how many pages to charge.
2194 * The charges will only happen if @memcg matches the current cpu's memcg
2195 * stock, and at least @nr_pages are available in that stock. Failure to
2196 * service an allocation will refill the stock.
2198 * returns true if successful, false otherwise.
2200 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2202 struct memcg_stock_pcp *stock;
2205 if (nr_pages > CHARGE_BATCH)
2208 stock = &get_cpu_var(memcg_stock);
2209 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2210 stock->nr_pages -= nr_pages;
2211 else /* need to call res_counter_charge */
2213 put_cpu_var(memcg_stock);
2218 * Returns stocks cached in percpu to res_counter and reset cached information.
2220 static void drain_stock(struct memcg_stock_pcp *stock)
2222 struct mem_cgroup *old = stock->cached;
2224 if (stock->nr_pages) {
2225 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2227 res_counter_uncharge(&old->res, bytes);
2228 if (do_swap_account)
2229 res_counter_uncharge(&old->memsw, bytes);
2230 stock->nr_pages = 0;
2232 stock->cached = NULL;
2236 * This must be called under preempt disabled or must be called by
2237 * a thread which is pinned to local cpu.
2239 static void drain_local_stock(struct work_struct *dummy)
2241 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2243 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2246 static void __init memcg_stock_init(void)
2250 for_each_possible_cpu(cpu) {
2251 struct memcg_stock_pcp *stock =
2252 &per_cpu(memcg_stock, cpu);
2253 INIT_WORK(&stock->work, drain_local_stock);
2258 * Cache charges(val) which is from res_counter, to local per_cpu area.
2259 * This will be consumed by consume_stock() function, later.
2261 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2263 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2265 if (stock->cached != memcg) { /* reset if necessary */
2267 stock->cached = memcg;
2269 stock->nr_pages += nr_pages;
2270 put_cpu_var(memcg_stock);
2274 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2275 * of the hierarchy under it. sync flag says whether we should block
2276 * until the work is done.
2278 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2282 /* Notify other cpus that system-wide "drain" is running */
2285 for_each_online_cpu(cpu) {
2286 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2287 struct mem_cgroup *memcg;
2289 memcg = stock->cached;
2290 if (!memcg || !stock->nr_pages)
2292 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2294 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2296 drain_local_stock(&stock->work);
2298 schedule_work_on(cpu, &stock->work);
2306 for_each_online_cpu(cpu) {
2307 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2308 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2309 flush_work(&stock->work);
2316 * Tries to drain stocked charges in other cpus. This function is asynchronous
2317 * and just put a work per cpu for draining localy on each cpu. Caller can
2318 * expects some charges will be back to res_counter later but cannot wait for
2321 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2324 * If someone calls draining, avoid adding more kworker runs.
2326 if (!mutex_trylock(&percpu_charge_mutex))
2328 drain_all_stock(root_memcg, false);
2329 mutex_unlock(&percpu_charge_mutex);
2332 /* This is a synchronous drain interface. */
2333 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2335 /* called when force_empty is called */
2336 mutex_lock(&percpu_charge_mutex);
2337 drain_all_stock(root_memcg, true);
2338 mutex_unlock(&percpu_charge_mutex);
2342 * This function drains percpu counter value from DEAD cpu and
2343 * move it to local cpu. Note that this function can be preempted.
2345 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2349 spin_lock(&memcg->pcp_counter_lock);
2350 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2351 long x = per_cpu(memcg->stat->count[i], cpu);
2353 per_cpu(memcg->stat->count[i], cpu) = 0;
2354 memcg->nocpu_base.count[i] += x;
2356 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2357 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2359 per_cpu(memcg->stat->events[i], cpu) = 0;
2360 memcg->nocpu_base.events[i] += x;
2362 spin_unlock(&memcg->pcp_counter_lock);
2365 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2366 unsigned long action,
2369 int cpu = (unsigned long)hcpu;
2370 struct memcg_stock_pcp *stock;
2371 struct mem_cgroup *iter;
2373 if (action == CPU_ONLINE)
2376 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2379 for_each_mem_cgroup(iter)
2380 mem_cgroup_drain_pcp_counter(iter, cpu);
2382 stock = &per_cpu(memcg_stock, cpu);
2388 /* See __mem_cgroup_try_charge() for details */
2390 CHARGE_OK, /* success */
2391 CHARGE_RETRY, /* need to retry but retry is not bad */
2392 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2393 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2394 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2397 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2398 unsigned int nr_pages, unsigned int min_pages,
2401 unsigned long csize = nr_pages * PAGE_SIZE;
2402 struct mem_cgroup *mem_over_limit;
2403 struct res_counter *fail_res;
2404 unsigned long flags = 0;
2407 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2410 if (!do_swap_account)
2412 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2416 res_counter_uncharge(&memcg->res, csize);
2417 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2418 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2420 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2422 * Never reclaim on behalf of optional batching, retry with a
2423 * single page instead.
2425 if (nr_pages > min_pages)
2426 return CHARGE_RETRY;
2428 if (!(gfp_mask & __GFP_WAIT))
2429 return CHARGE_WOULDBLOCK;
2431 if (gfp_mask & __GFP_NORETRY)
2432 return CHARGE_NOMEM;
2434 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2435 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2436 return CHARGE_RETRY;
2438 * Even though the limit is exceeded at this point, reclaim
2439 * may have been able to free some pages. Retry the charge
2440 * before killing the task.
2442 * Only for regular pages, though: huge pages are rather
2443 * unlikely to succeed so close to the limit, and we fall back
2444 * to regular pages anyway in case of failure.
2446 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2447 return CHARGE_RETRY;
2450 * At task move, charge accounts can be doubly counted. So, it's
2451 * better to wait until the end of task_move if something is going on.
2453 if (mem_cgroup_wait_acct_move(mem_over_limit))
2454 return CHARGE_RETRY;
2456 /* If we don't need to call oom-killer at el, return immediately */
2457 if (!oom_check || !current->memcg_oom.may_oom)
2458 return CHARGE_NOMEM;
2460 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2461 return CHARGE_OOM_DIE;
2463 return CHARGE_RETRY;
2467 * __mem_cgroup_try_charge() does
2468 * 1. detect memcg to be charged against from passed *mm and *ptr,
2469 * 2. update res_counter
2470 * 3. call memory reclaim if necessary.
2472 * In some special case, if the task is fatal, fatal_signal_pending() or
2473 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2474 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2475 * as possible without any hazards. 2: all pages should have a valid
2476 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2477 * pointer, that is treated as a charge to root_mem_cgroup.
2479 * So __mem_cgroup_try_charge() will return
2480 * 0 ... on success, filling *ptr with a valid memcg pointer.
2481 * -ENOMEM ... charge failure because of resource limits.
2482 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2484 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2485 * the oom-killer can be invoked.
2487 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2489 unsigned int nr_pages,
2490 struct mem_cgroup **ptr,
2493 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2494 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2495 struct mem_cgroup *memcg = NULL;
2499 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2500 * in system level. So, allow to go ahead dying process in addition to
2503 if (unlikely(test_thread_flag(TIF_MEMDIE)
2504 || fatal_signal_pending(current)))
2508 * We always charge the cgroup the mm_struct belongs to.
2509 * The mm_struct's mem_cgroup changes on task migration if the
2510 * thread group leader migrates. It's possible that mm is not
2511 * set, if so charge the root memcg (happens for pagecache usage).
2514 *ptr = root_mem_cgroup;
2516 if (*ptr) { /* css should be a valid one */
2518 if (mem_cgroup_is_root(memcg))
2520 if (consume_stock(memcg, nr_pages))
2522 css_get(&memcg->css);
2524 struct task_struct *p;
2527 p = rcu_dereference(mm->owner);
2529 * Because we don't have task_lock(), "p" can exit.
2530 * In that case, "memcg" can point to root or p can be NULL with
2531 * race with swapoff. Then, we have small risk of mis-accouning.
2532 * But such kind of mis-account by race always happens because
2533 * we don't have cgroup_mutex(). It's overkill and we allo that
2535 * (*) swapoff at el will charge against mm-struct not against
2536 * task-struct. So, mm->owner can be NULL.
2538 memcg = mem_cgroup_from_task(p);
2540 memcg = root_mem_cgroup;
2541 if (mem_cgroup_is_root(memcg)) {
2545 if (consume_stock(memcg, nr_pages)) {
2547 * It seems dagerous to access memcg without css_get().
2548 * But considering how consume_stok works, it's not
2549 * necessary. If consume_stock success, some charges
2550 * from this memcg are cached on this cpu. So, we
2551 * don't need to call css_get()/css_tryget() before
2552 * calling consume_stock().
2557 /* after here, we may be blocked. we need to get refcnt */
2558 if (!css_tryget(&memcg->css)) {
2568 /* If killed, bypass charge */
2569 if (fatal_signal_pending(current)) {
2570 css_put(&memcg->css);
2575 if (oom && !nr_oom_retries) {
2577 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2580 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2585 case CHARGE_RETRY: /* not in OOM situation but retry */
2587 css_put(&memcg->css);
2590 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2591 css_put(&memcg->css);
2593 case CHARGE_NOMEM: /* OOM routine works */
2595 css_put(&memcg->css);
2598 /* If oom, we never return -ENOMEM */
2601 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2602 css_put(&memcg->css);
2605 } while (ret != CHARGE_OK);
2607 if (batch > nr_pages)
2608 refill_stock(memcg, batch - nr_pages);
2609 css_put(&memcg->css);
2617 *ptr = root_mem_cgroup;
2622 * Somemtimes we have to undo a charge we got by try_charge().
2623 * This function is for that and do uncharge, put css's refcnt.
2624 * gotten by try_charge().
2626 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2627 unsigned int nr_pages)
2629 if (!mem_cgroup_is_root(memcg)) {
2630 unsigned long bytes = nr_pages * PAGE_SIZE;
2632 res_counter_uncharge(&memcg->res, bytes);
2633 if (do_swap_account)
2634 res_counter_uncharge(&memcg->memsw, bytes);
2639 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2640 * This is useful when moving usage to parent cgroup.
2642 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2643 unsigned int nr_pages)
2645 unsigned long bytes = nr_pages * PAGE_SIZE;
2647 if (mem_cgroup_is_root(memcg))
2650 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2651 if (do_swap_account)
2652 res_counter_uncharge_until(&memcg->memsw,
2653 memcg->memsw.parent, bytes);
2657 * A helper function to get mem_cgroup from ID. must be called under
2658 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2659 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2660 * called against removed memcg.)
2662 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2664 struct cgroup_subsys_state *css;
2666 /* ID 0 is unused ID */
2669 css = css_lookup(&mem_cgroup_subsys, id);
2672 return mem_cgroup_from_css(css);
2675 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2677 struct mem_cgroup *memcg = NULL;
2678 struct page_cgroup *pc;
2682 VM_BUG_ON(!PageLocked(page));
2684 pc = lookup_page_cgroup(page);
2685 lock_page_cgroup(pc);
2686 if (PageCgroupUsed(pc)) {
2687 memcg = pc->mem_cgroup;
2688 if (memcg && !css_tryget(&memcg->css))
2690 } else if (PageSwapCache(page)) {
2691 ent.val = page_private(page);
2692 id = lookup_swap_cgroup_id(ent);
2694 memcg = mem_cgroup_lookup(id);
2695 if (memcg && !css_tryget(&memcg->css))
2699 unlock_page_cgroup(pc);
2703 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2705 unsigned int nr_pages,
2706 enum charge_type ctype,
2709 struct page_cgroup *pc = lookup_page_cgroup(page);
2710 struct zone *uninitialized_var(zone);
2711 struct lruvec *lruvec;
2712 bool was_on_lru = false;
2715 lock_page_cgroup(pc);
2716 VM_BUG_ON(PageCgroupUsed(pc));
2718 * we don't need page_cgroup_lock about tail pages, becase they are not
2719 * accessed by any other context at this point.
2723 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2724 * may already be on some other mem_cgroup's LRU. Take care of it.
2727 zone = page_zone(page);
2728 spin_lock_irq(&zone->lru_lock);
2729 if (PageLRU(page)) {
2730 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2732 del_page_from_lru_list(page, lruvec, page_lru(page));
2737 pc->mem_cgroup = memcg;
2739 * We access a page_cgroup asynchronously without lock_page_cgroup().
2740 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2741 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2742 * before USED bit, we need memory barrier here.
2743 * See mem_cgroup_add_lru_list(), etc.
2746 SetPageCgroupUsed(pc);
2750 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2751 VM_BUG_ON(PageLRU(page));
2753 add_page_to_lru_list(page, lruvec, page_lru(page));
2755 spin_unlock_irq(&zone->lru_lock);
2758 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2763 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2764 unlock_page_cgroup(pc);
2767 * "charge_statistics" updated event counter.
2769 memcg_check_events(memcg, page);
2772 static DEFINE_MUTEX(set_limit_mutex);
2774 #ifdef CONFIG_MEMCG_KMEM
2775 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2777 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2778 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2782 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2783 * in the memcg_cache_params struct.
2785 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2787 struct kmem_cache *cachep;
2789 VM_BUG_ON(p->is_root_cache);
2790 cachep = p->root_cache;
2791 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2794 #ifdef CONFIG_SLABINFO
2795 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2796 struct cftype *cft, struct seq_file *m)
2798 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2799 struct memcg_cache_params *params;
2801 if (!memcg_can_account_kmem(memcg))
2804 print_slabinfo_header(m);
2806 mutex_lock(&memcg->slab_caches_mutex);
2807 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2808 cache_show(memcg_params_to_cache(params), m);
2809 mutex_unlock(&memcg->slab_caches_mutex);
2815 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2817 struct res_counter *fail_res;
2818 struct mem_cgroup *_memcg;
2822 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2827 * Conditions under which we can wait for the oom_killer. Those are
2828 * the same conditions tested by the core page allocator
2830 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2833 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2836 if (ret == -EINTR) {
2838 * __mem_cgroup_try_charge() chosed to bypass to root due to
2839 * OOM kill or fatal signal. Since our only options are to
2840 * either fail the allocation or charge it to this cgroup, do
2841 * it as a temporary condition. But we can't fail. From a
2842 * kmem/slab perspective, the cache has already been selected,
2843 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2846 * This condition will only trigger if the task entered
2847 * memcg_charge_kmem in a sane state, but was OOM-killed during
2848 * __mem_cgroup_try_charge() above. Tasks that were already
2849 * dying when the allocation triggers should have been already
2850 * directed to the root cgroup in memcontrol.h
2852 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2853 if (do_swap_account)
2854 res_counter_charge_nofail(&memcg->memsw, size,
2858 res_counter_uncharge(&memcg->kmem, size);
2863 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2865 res_counter_uncharge(&memcg->res, size);
2866 if (do_swap_account)
2867 res_counter_uncharge(&memcg->memsw, size);
2870 if (res_counter_uncharge(&memcg->kmem, size))
2874 * Releases a reference taken in kmem_cgroup_css_offline in case
2875 * this last uncharge is racing with the offlining code or it is
2876 * outliving the memcg existence.
2878 * The memory barrier imposed by test&clear is paired with the
2879 * explicit one in memcg_kmem_mark_dead().
2881 if (memcg_kmem_test_and_clear_dead(memcg))
2882 css_put(&memcg->css);
2885 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2890 mutex_lock(&memcg->slab_caches_mutex);
2891 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2892 mutex_unlock(&memcg->slab_caches_mutex);
2896 * helper for acessing a memcg's index. It will be used as an index in the
2897 * child cache array in kmem_cache, and also to derive its name. This function
2898 * will return -1 when this is not a kmem-limited memcg.
2900 int memcg_cache_id(struct mem_cgroup *memcg)
2902 return memcg ? memcg->kmemcg_id : -1;
2906 * This ends up being protected by the set_limit mutex, during normal
2907 * operation, because that is its main call site.
2909 * But when we create a new cache, we can call this as well if its parent
2910 * is kmem-limited. That will have to hold set_limit_mutex as well.
2912 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2916 num = ida_simple_get(&kmem_limited_groups,
2917 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2921 * After this point, kmem_accounted (that we test atomically in
2922 * the beginning of this conditional), is no longer 0. This
2923 * guarantees only one process will set the following boolean
2924 * to true. We don't need test_and_set because we're protected
2925 * by the set_limit_mutex anyway.
2927 memcg_kmem_set_activated(memcg);
2929 ret = memcg_update_all_caches(num+1);
2931 ida_simple_remove(&kmem_limited_groups, num);
2932 memcg_kmem_clear_activated(memcg);
2936 memcg->kmemcg_id = num;
2937 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2938 mutex_init(&memcg->slab_caches_mutex);
2942 static size_t memcg_caches_array_size(int num_groups)
2945 if (num_groups <= 0)
2948 size = 2 * num_groups;
2949 if (size < MEMCG_CACHES_MIN_SIZE)
2950 size = MEMCG_CACHES_MIN_SIZE;
2951 else if (size > MEMCG_CACHES_MAX_SIZE)
2952 size = MEMCG_CACHES_MAX_SIZE;
2958 * We should update the current array size iff all caches updates succeed. This
2959 * can only be done from the slab side. The slab mutex needs to be held when
2962 void memcg_update_array_size(int num)
2964 if (num > memcg_limited_groups_array_size)
2965 memcg_limited_groups_array_size = memcg_caches_array_size(num);
2968 static void kmem_cache_destroy_work_func(struct work_struct *w);
2970 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
2972 struct memcg_cache_params *cur_params = s->memcg_params;
2974 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
2976 if (num_groups > memcg_limited_groups_array_size) {
2978 ssize_t size = memcg_caches_array_size(num_groups);
2980 size *= sizeof(void *);
2981 size += offsetof(struct memcg_cache_params, memcg_caches);
2983 s->memcg_params = kzalloc(size, GFP_KERNEL);
2984 if (!s->memcg_params) {
2985 s->memcg_params = cur_params;
2989 s->memcg_params->is_root_cache = true;
2992 * There is the chance it will be bigger than
2993 * memcg_limited_groups_array_size, if we failed an allocation
2994 * in a cache, in which case all caches updated before it, will
2995 * have a bigger array.
2997 * But if that is the case, the data after
2998 * memcg_limited_groups_array_size is certainly unused
3000 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3001 if (!cur_params->memcg_caches[i])
3003 s->memcg_params->memcg_caches[i] =
3004 cur_params->memcg_caches[i];
3008 * Ideally, we would wait until all caches succeed, and only
3009 * then free the old one. But this is not worth the extra
3010 * pointer per-cache we'd have to have for this.
3012 * It is not a big deal if some caches are left with a size
3013 * bigger than the others. And all updates will reset this
3021 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3022 struct kmem_cache *root_cache)
3026 if (!memcg_kmem_enabled())
3030 size = offsetof(struct memcg_cache_params, memcg_caches);
3031 size += memcg_limited_groups_array_size * sizeof(void *);
3033 size = sizeof(struct memcg_cache_params);
3035 s->memcg_params = kzalloc(size, GFP_KERNEL);
3036 if (!s->memcg_params)
3040 s->memcg_params->memcg = memcg;
3041 s->memcg_params->root_cache = root_cache;
3042 INIT_WORK(&s->memcg_params->destroy,
3043 kmem_cache_destroy_work_func);
3045 s->memcg_params->is_root_cache = true;
3050 void memcg_release_cache(struct kmem_cache *s)
3052 struct kmem_cache *root;
3053 struct mem_cgroup *memcg;
3057 * This happens, for instance, when a root cache goes away before we
3060 if (!s->memcg_params)
3063 if (s->memcg_params->is_root_cache)
3066 memcg = s->memcg_params->memcg;
3067 id = memcg_cache_id(memcg);
3069 root = s->memcg_params->root_cache;
3070 root->memcg_params->memcg_caches[id] = NULL;
3072 mutex_lock(&memcg->slab_caches_mutex);
3073 list_del(&s->memcg_params->list);
3074 mutex_unlock(&memcg->slab_caches_mutex);
3076 css_put(&memcg->css);
3078 kfree(s->memcg_params);
3082 * During the creation a new cache, we need to disable our accounting mechanism
3083 * altogether. This is true even if we are not creating, but rather just
3084 * enqueing new caches to be created.
3086 * This is because that process will trigger allocations; some visible, like
3087 * explicit kmallocs to auxiliary data structures, name strings and internal
3088 * cache structures; some well concealed, like INIT_WORK() that can allocate
3089 * objects during debug.
3091 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3092 * to it. This may not be a bounded recursion: since the first cache creation
3093 * failed to complete (waiting on the allocation), we'll just try to create the
3094 * cache again, failing at the same point.
3096 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3097 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3098 * inside the following two functions.
3100 static inline void memcg_stop_kmem_account(void)
3102 VM_BUG_ON(!current->mm);
3103 current->memcg_kmem_skip_account++;
3106 static inline void memcg_resume_kmem_account(void)
3108 VM_BUG_ON(!current->mm);
3109 current->memcg_kmem_skip_account--;
3112 static void kmem_cache_destroy_work_func(struct work_struct *w)
3114 struct kmem_cache *cachep;
3115 struct memcg_cache_params *p;
3117 p = container_of(w, struct memcg_cache_params, destroy);
3119 cachep = memcg_params_to_cache(p);
3122 * If we get down to 0 after shrink, we could delete right away.
3123 * However, memcg_release_pages() already puts us back in the workqueue
3124 * in that case. If we proceed deleting, we'll get a dangling
3125 * reference, and removing the object from the workqueue in that case
3126 * is unnecessary complication. We are not a fast path.
3128 * Note that this case is fundamentally different from racing with
3129 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3130 * kmem_cache_shrink, not only we would be reinserting a dead cache
3131 * into the queue, but doing so from inside the worker racing to
3134 * So if we aren't down to zero, we'll just schedule a worker and try
3137 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3138 kmem_cache_shrink(cachep);
3139 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3142 kmem_cache_destroy(cachep);
3145 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3147 if (!cachep->memcg_params->dead)
3151 * There are many ways in which we can get here.
3153 * We can get to a memory-pressure situation while the delayed work is
3154 * still pending to run. The vmscan shrinkers can then release all
3155 * cache memory and get us to destruction. If this is the case, we'll
3156 * be executed twice, which is a bug (the second time will execute over
3157 * bogus data). In this case, cancelling the work should be fine.
3159 * But we can also get here from the worker itself, if
3160 * kmem_cache_shrink is enough to shake all the remaining objects and
3161 * get the page count to 0. In this case, we'll deadlock if we try to
3162 * cancel the work (the worker runs with an internal lock held, which
3163 * is the same lock we would hold for cancel_work_sync().)
3165 * Since we can't possibly know who got us here, just refrain from
3166 * running if there is already work pending
3168 if (work_pending(&cachep->memcg_params->destroy))
3171 * We have to defer the actual destroying to a workqueue, because
3172 * we might currently be in a context that cannot sleep.
3174 schedule_work(&cachep->memcg_params->destroy);
3178 * This lock protects updaters, not readers. We want readers to be as fast as
3179 * they can, and they will either see NULL or a valid cache value. Our model
3180 * allow them to see NULL, in which case the root memcg will be selected.
3182 * We need this lock because multiple allocations to the same cache from a non
3183 * will span more than one worker. Only one of them can create the cache.
3185 static DEFINE_MUTEX(memcg_cache_mutex);
3188 * Called with memcg_cache_mutex held
3190 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3191 struct kmem_cache *s)
3193 struct kmem_cache *new;
3194 static char *tmp_name = NULL;
3196 lockdep_assert_held(&memcg_cache_mutex);
3199 * kmem_cache_create_memcg duplicates the given name and
3200 * cgroup_name for this name requires RCU context.
3201 * This static temporary buffer is used to prevent from
3202 * pointless shortliving allocation.
3205 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3211 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3212 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3215 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3216 (s->flags & ~SLAB_PANIC), s->ctor, s);
3219 new->allocflags |= __GFP_KMEMCG;
3224 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3225 struct kmem_cache *cachep)
3227 struct kmem_cache *new_cachep;
3230 BUG_ON(!memcg_can_account_kmem(memcg));
3232 idx = memcg_cache_id(memcg);
3234 mutex_lock(&memcg_cache_mutex);
3235 new_cachep = cachep->memcg_params->memcg_caches[idx];
3237 css_put(&memcg->css);
3241 new_cachep = kmem_cache_dup(memcg, cachep);
3242 if (new_cachep == NULL) {
3243 new_cachep = cachep;
3244 css_put(&memcg->css);
3248 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3250 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3252 * the readers won't lock, make sure everybody sees the updated value,
3253 * so they won't put stuff in the queue again for no reason
3257 mutex_unlock(&memcg_cache_mutex);
3261 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3263 struct kmem_cache *c;
3266 if (!s->memcg_params)
3268 if (!s->memcg_params->is_root_cache)
3272 * If the cache is being destroyed, we trust that there is no one else
3273 * requesting objects from it. Even if there are, the sanity checks in
3274 * kmem_cache_destroy should caught this ill-case.
3276 * Still, we don't want anyone else freeing memcg_caches under our
3277 * noses, which can happen if a new memcg comes to life. As usual,
3278 * we'll take the set_limit_mutex to protect ourselves against this.
3280 mutex_lock(&set_limit_mutex);
3281 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3282 c = s->memcg_params->memcg_caches[i];
3287 * We will now manually delete the caches, so to avoid races
3288 * we need to cancel all pending destruction workers and
3289 * proceed with destruction ourselves.
3291 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3292 * and that could spawn the workers again: it is likely that
3293 * the cache still have active pages until this very moment.
3294 * This would lead us back to mem_cgroup_destroy_cache.
3296 * But that will not execute at all if the "dead" flag is not
3297 * set, so flip it down to guarantee we are in control.
3299 c->memcg_params->dead = false;
3300 cancel_work_sync(&c->memcg_params->destroy);
3301 kmem_cache_destroy(c);
3303 mutex_unlock(&set_limit_mutex);
3306 struct create_work {
3307 struct mem_cgroup *memcg;
3308 struct kmem_cache *cachep;
3309 struct work_struct work;
3312 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3314 struct kmem_cache *cachep;
3315 struct memcg_cache_params *params;
3317 if (!memcg_kmem_is_active(memcg))
3320 mutex_lock(&memcg->slab_caches_mutex);
3321 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3322 cachep = memcg_params_to_cache(params);
3323 cachep->memcg_params->dead = true;
3324 schedule_work(&cachep->memcg_params->destroy);
3326 mutex_unlock(&memcg->slab_caches_mutex);
3329 static void memcg_create_cache_work_func(struct work_struct *w)
3331 struct create_work *cw;
3333 cw = container_of(w, struct create_work, work);
3334 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3339 * Enqueue the creation of a per-memcg kmem_cache.
3341 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3342 struct kmem_cache *cachep)
3344 struct create_work *cw;
3346 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3348 css_put(&memcg->css);
3353 cw->cachep = cachep;
3355 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3356 schedule_work(&cw->work);
3359 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3360 struct kmem_cache *cachep)
3363 * We need to stop accounting when we kmalloc, because if the
3364 * corresponding kmalloc cache is not yet created, the first allocation
3365 * in __memcg_create_cache_enqueue will recurse.
3367 * However, it is better to enclose the whole function. Depending on
3368 * the debugging options enabled, INIT_WORK(), for instance, can
3369 * trigger an allocation. This too, will make us recurse. Because at
3370 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3371 * the safest choice is to do it like this, wrapping the whole function.
3373 memcg_stop_kmem_account();
3374 __memcg_create_cache_enqueue(memcg, cachep);
3375 memcg_resume_kmem_account();
3378 * Return the kmem_cache we're supposed to use for a slab allocation.
3379 * We try to use the current memcg's version of the cache.
3381 * If the cache does not exist yet, if we are the first user of it,
3382 * we either create it immediately, if possible, or create it asynchronously
3384 * In the latter case, we will let the current allocation go through with
3385 * the original cache.
3387 * Can't be called in interrupt context or from kernel threads.
3388 * This function needs to be called with rcu_read_lock() held.
3390 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3393 struct mem_cgroup *memcg;
3396 VM_BUG_ON(!cachep->memcg_params);
3397 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3399 if (!current->mm || current->memcg_kmem_skip_account)
3403 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3405 if (!memcg_can_account_kmem(memcg))
3408 idx = memcg_cache_id(memcg);
3411 * barrier to mare sure we're always seeing the up to date value. The
3412 * code updating memcg_caches will issue a write barrier to match this.
3414 read_barrier_depends();
3415 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3416 cachep = cachep->memcg_params->memcg_caches[idx];
3420 /* The corresponding put will be done in the workqueue. */
3421 if (!css_tryget(&memcg->css))
3426 * If we are in a safe context (can wait, and not in interrupt
3427 * context), we could be be predictable and return right away.
3428 * This would guarantee that the allocation being performed
3429 * already belongs in the new cache.
3431 * However, there are some clashes that can arrive from locking.
3432 * For instance, because we acquire the slab_mutex while doing
3433 * kmem_cache_dup, this means no further allocation could happen
3434 * with the slab_mutex held.
3436 * Also, because cache creation issue get_online_cpus(), this
3437 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3438 * that ends up reversed during cpu hotplug. (cpuset allocates
3439 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3440 * better to defer everything.
3442 memcg_create_cache_enqueue(memcg, cachep);
3448 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3451 * We need to verify if the allocation against current->mm->owner's memcg is
3452 * possible for the given order. But the page is not allocated yet, so we'll
3453 * need a further commit step to do the final arrangements.
3455 * It is possible for the task to switch cgroups in this mean time, so at
3456 * commit time, we can't rely on task conversion any longer. We'll then use
3457 * the handle argument to return to the caller which cgroup we should commit
3458 * against. We could also return the memcg directly and avoid the pointer
3459 * passing, but a boolean return value gives better semantics considering
3460 * the compiled-out case as well.
3462 * Returning true means the allocation is possible.
3465 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3467 struct mem_cgroup *memcg;
3473 * Disabling accounting is only relevant for some specific memcg
3474 * internal allocations. Therefore we would initially not have such
3475 * check here, since direct calls to the page allocator that are marked
3476 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3477 * concerned with cache allocations, and by having this test at
3478 * memcg_kmem_get_cache, we are already able to relay the allocation to
3479 * the root cache and bypass the memcg cache altogether.
3481 * There is one exception, though: the SLUB allocator does not create
3482 * large order caches, but rather service large kmallocs directly from
3483 * the page allocator. Therefore, the following sequence when backed by
3484 * the SLUB allocator:
3486 * memcg_stop_kmem_account();
3487 * kmalloc(<large_number>)
3488 * memcg_resume_kmem_account();
3490 * would effectively ignore the fact that we should skip accounting,
3491 * since it will drive us directly to this function without passing
3492 * through the cache selector memcg_kmem_get_cache. Such large
3493 * allocations are extremely rare but can happen, for instance, for the
3494 * cache arrays. We bring this test here.
3496 if (!current->mm || current->memcg_kmem_skip_account)
3499 memcg = try_get_mem_cgroup_from_mm(current->mm);
3502 * very rare case described in mem_cgroup_from_task. Unfortunately there
3503 * isn't much we can do without complicating this too much, and it would
3504 * be gfp-dependent anyway. Just let it go
3506 if (unlikely(!memcg))
3509 if (!memcg_can_account_kmem(memcg)) {
3510 css_put(&memcg->css);
3514 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3518 css_put(&memcg->css);
3522 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3525 struct page_cgroup *pc;
3527 VM_BUG_ON(mem_cgroup_is_root(memcg));
3529 /* The page allocation failed. Revert */
3531 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3535 pc = lookup_page_cgroup(page);
3536 lock_page_cgroup(pc);
3537 pc->mem_cgroup = memcg;
3538 SetPageCgroupUsed(pc);
3539 unlock_page_cgroup(pc);
3542 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3544 struct mem_cgroup *memcg = NULL;
3545 struct page_cgroup *pc;
3548 pc = lookup_page_cgroup(page);
3550 * Fast unlocked return. Theoretically might have changed, have to
3551 * check again after locking.
3553 if (!PageCgroupUsed(pc))
3556 lock_page_cgroup(pc);
3557 if (PageCgroupUsed(pc)) {
3558 memcg = pc->mem_cgroup;
3559 ClearPageCgroupUsed(pc);
3561 unlock_page_cgroup(pc);
3564 * We trust that only if there is a memcg associated with the page, it
3565 * is a valid allocation
3570 VM_BUG_ON(mem_cgroup_is_root(memcg));
3571 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3574 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3577 #endif /* CONFIG_MEMCG_KMEM */
3579 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3581 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3583 * Because tail pages are not marked as "used", set it. We're under
3584 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3585 * charge/uncharge will be never happen and move_account() is done under
3586 * compound_lock(), so we don't have to take care of races.
3588 void mem_cgroup_split_huge_fixup(struct page *head)
3590 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3591 struct page_cgroup *pc;
3592 struct mem_cgroup *memcg;
3595 if (mem_cgroup_disabled())
3598 memcg = head_pc->mem_cgroup;
3599 for (i = 1; i < HPAGE_PMD_NR; i++) {
3601 pc->mem_cgroup = memcg;
3602 smp_wmb();/* see __commit_charge() */
3603 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3605 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3608 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3611 * mem_cgroup_move_account - move account of the page
3613 * @nr_pages: number of regular pages (>1 for huge pages)
3614 * @pc: page_cgroup of the page.
3615 * @from: mem_cgroup which the page is moved from.
3616 * @to: mem_cgroup which the page is moved to. @from != @to.
3618 * The caller must confirm following.
3619 * - page is not on LRU (isolate_page() is useful.)
3620 * - compound_lock is held when nr_pages > 1
3622 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3625 static int mem_cgroup_move_account(struct page *page,
3626 unsigned int nr_pages,
3627 struct page_cgroup *pc,
3628 struct mem_cgroup *from,
3629 struct mem_cgroup *to)
3631 unsigned long flags;
3633 bool anon = PageAnon(page);
3635 VM_BUG_ON(from == to);
3636 VM_BUG_ON(PageLRU(page));
3638 * The page is isolated from LRU. So, collapse function
3639 * will not handle this page. But page splitting can happen.
3640 * Do this check under compound_page_lock(). The caller should
3644 if (nr_pages > 1 && !PageTransHuge(page))
3647 lock_page_cgroup(pc);
3650 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3653 move_lock_mem_cgroup(from, &flags);
3655 if (!anon && page_mapped(page)) {
3656 /* Update mapped_file data for mem_cgroup */
3658 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3659 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3662 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3664 /* caller should have done css_get */
3665 pc->mem_cgroup = to;
3666 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3667 move_unlock_mem_cgroup(from, &flags);
3670 unlock_page_cgroup(pc);
3674 memcg_check_events(to, page);
3675 memcg_check_events(from, page);
3681 * mem_cgroup_move_parent - moves page to the parent group
3682 * @page: the page to move
3683 * @pc: page_cgroup of the page
3684 * @child: page's cgroup
3686 * move charges to its parent or the root cgroup if the group has no
3687 * parent (aka use_hierarchy==0).
3688 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3689 * mem_cgroup_move_account fails) the failure is always temporary and
3690 * it signals a race with a page removal/uncharge or migration. In the
3691 * first case the page is on the way out and it will vanish from the LRU
3692 * on the next attempt and the call should be retried later.
3693 * Isolation from the LRU fails only if page has been isolated from
3694 * the LRU since we looked at it and that usually means either global
3695 * reclaim or migration going on. The page will either get back to the
3697 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3698 * (!PageCgroupUsed) or moved to a different group. The page will
3699 * disappear in the next attempt.
3701 static int mem_cgroup_move_parent(struct page *page,
3702 struct page_cgroup *pc,
3703 struct mem_cgroup *child)
3705 struct mem_cgroup *parent;
3706 unsigned int nr_pages;
3707 unsigned long uninitialized_var(flags);
3710 VM_BUG_ON(mem_cgroup_is_root(child));
3713 if (!get_page_unless_zero(page))
3715 if (isolate_lru_page(page))
3718 nr_pages = hpage_nr_pages(page);
3720 parent = parent_mem_cgroup(child);
3722 * If no parent, move charges to root cgroup.
3725 parent = root_mem_cgroup;
3728 VM_BUG_ON(!PageTransHuge(page));
3729 flags = compound_lock_irqsave(page);
3732 ret = mem_cgroup_move_account(page, nr_pages,
3735 __mem_cgroup_cancel_local_charge(child, nr_pages);
3738 compound_unlock_irqrestore(page, flags);
3739 putback_lru_page(page);
3747 * Charge the memory controller for page usage.
3749 * 0 if the charge was successful
3750 * < 0 if the cgroup is over its limit
3752 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3753 gfp_t gfp_mask, enum charge_type ctype)
3755 struct mem_cgroup *memcg = NULL;
3756 unsigned int nr_pages = 1;
3760 if (PageTransHuge(page)) {
3761 nr_pages <<= compound_order(page);
3762 VM_BUG_ON(!PageTransHuge(page));
3764 * Never OOM-kill a process for a huge page. The
3765 * fault handler will fall back to regular pages.
3770 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3773 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3777 int mem_cgroup_newpage_charge(struct page *page,
3778 struct mm_struct *mm, gfp_t gfp_mask)
3780 if (mem_cgroup_disabled())
3782 VM_BUG_ON(page_mapped(page));
3783 VM_BUG_ON(page->mapping && !PageAnon(page));
3785 return mem_cgroup_charge_common(page, mm, gfp_mask,
3786 MEM_CGROUP_CHARGE_TYPE_ANON);
3790 * While swap-in, try_charge -> commit or cancel, the page is locked.
3791 * And when try_charge() successfully returns, one refcnt to memcg without
3792 * struct page_cgroup is acquired. This refcnt will be consumed by
3793 * "commit()" or removed by "cancel()"
3795 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3798 struct mem_cgroup **memcgp)
3800 struct mem_cgroup *memcg;
3801 struct page_cgroup *pc;
3804 pc = lookup_page_cgroup(page);
3806 * Every swap fault against a single page tries to charge the
3807 * page, bail as early as possible. shmem_unuse() encounters
3808 * already charged pages, too. The USED bit is protected by
3809 * the page lock, which serializes swap cache removal, which
3810 * in turn serializes uncharging.
3812 if (PageCgroupUsed(pc))
3814 if (!do_swap_account)
3816 memcg = try_get_mem_cgroup_from_page(page);
3820 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3821 css_put(&memcg->css);
3826 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3832 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3833 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3836 if (mem_cgroup_disabled())
3839 * A racing thread's fault, or swapoff, may have already
3840 * updated the pte, and even removed page from swap cache: in
3841 * those cases unuse_pte()'s pte_same() test will fail; but
3842 * there's also a KSM case which does need to charge the page.
3844 if (!PageSwapCache(page)) {
3847 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3852 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3855 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3857 if (mem_cgroup_disabled())
3861 __mem_cgroup_cancel_charge(memcg, 1);
3865 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3866 enum charge_type ctype)
3868 if (mem_cgroup_disabled())
3873 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3875 * Now swap is on-memory. This means this page may be
3876 * counted both as mem and swap....double count.
3877 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3878 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3879 * may call delete_from_swap_cache() before reach here.
3881 if (do_swap_account && PageSwapCache(page)) {
3882 swp_entry_t ent = {.val = page_private(page)};
3883 mem_cgroup_uncharge_swap(ent);
3887 void mem_cgroup_commit_charge_swapin(struct page *page,
3888 struct mem_cgroup *memcg)
3890 __mem_cgroup_commit_charge_swapin(page, memcg,
3891 MEM_CGROUP_CHARGE_TYPE_ANON);
3894 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3897 struct mem_cgroup *memcg = NULL;
3898 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3901 if (mem_cgroup_disabled())
3903 if (PageCompound(page))
3906 if (!PageSwapCache(page))
3907 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3908 else { /* page is swapcache/shmem */
3909 ret = __mem_cgroup_try_charge_swapin(mm, page,
3912 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3917 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3918 unsigned int nr_pages,
3919 const enum charge_type ctype)
3921 struct memcg_batch_info *batch = NULL;
3922 bool uncharge_memsw = true;
3924 /* If swapout, usage of swap doesn't decrease */
3925 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3926 uncharge_memsw = false;
3928 batch = ¤t->memcg_batch;
3930 * In usual, we do css_get() when we remember memcg pointer.
3931 * But in this case, we keep res->usage until end of a series of
3932 * uncharges. Then, it's ok to ignore memcg's refcnt.
3935 batch->memcg = memcg;
3937 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3938 * In those cases, all pages freed continuously can be expected to be in
3939 * the same cgroup and we have chance to coalesce uncharges.
3940 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3941 * because we want to do uncharge as soon as possible.
3944 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3945 goto direct_uncharge;
3948 goto direct_uncharge;
3951 * In typical case, batch->memcg == mem. This means we can
3952 * merge a series of uncharges to an uncharge of res_counter.
3953 * If not, we uncharge res_counter ony by one.
3955 if (batch->memcg != memcg)
3956 goto direct_uncharge;
3957 /* remember freed charge and uncharge it later */
3960 batch->memsw_nr_pages++;
3963 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3965 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3966 if (unlikely(batch->memcg != memcg))
3967 memcg_oom_recover(memcg);
3971 * uncharge if !page_mapped(page)
3973 static struct mem_cgroup *
3974 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3977 struct mem_cgroup *memcg = NULL;
3978 unsigned int nr_pages = 1;
3979 struct page_cgroup *pc;
3982 if (mem_cgroup_disabled())
3985 if (PageTransHuge(page)) {
3986 nr_pages <<= compound_order(page);
3987 VM_BUG_ON(!PageTransHuge(page));
3990 * Check if our page_cgroup is valid
3992 pc = lookup_page_cgroup(page);
3993 if (unlikely(!PageCgroupUsed(pc)))
3996 lock_page_cgroup(pc);
3998 memcg = pc->mem_cgroup;
4000 if (!PageCgroupUsed(pc))
4003 anon = PageAnon(page);
4006 case MEM_CGROUP_CHARGE_TYPE_ANON:
4008 * Generally PageAnon tells if it's the anon statistics to be
4009 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4010 * used before page reached the stage of being marked PageAnon.
4014 case MEM_CGROUP_CHARGE_TYPE_DROP:
4015 /* See mem_cgroup_prepare_migration() */
4016 if (page_mapped(page))
4019 * Pages under migration may not be uncharged. But
4020 * end_migration() /must/ be the one uncharging the
4021 * unused post-migration page and so it has to call
4022 * here with the migration bit still set. See the
4023 * res_counter handling below.
4025 if (!end_migration && PageCgroupMigration(pc))
4028 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4029 if (!PageAnon(page)) { /* Shared memory */
4030 if (page->mapping && !page_is_file_cache(page))
4032 } else if (page_mapped(page)) /* Anon */
4039 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4041 ClearPageCgroupUsed(pc);
4043 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4044 * freed from LRU. This is safe because uncharged page is expected not
4045 * to be reused (freed soon). Exception is SwapCache, it's handled by
4046 * special functions.
4049 unlock_page_cgroup(pc);
4051 * even after unlock, we have memcg->res.usage here and this memcg
4052 * will never be freed, so it's safe to call css_get().
4054 memcg_check_events(memcg, page);
4055 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4056 mem_cgroup_swap_statistics(memcg, true);
4057 css_get(&memcg->css);
4060 * Migration does not charge the res_counter for the
4061 * replacement page, so leave it alone when phasing out the
4062 * page that is unused after the migration.
4064 if (!end_migration && !mem_cgroup_is_root(memcg))
4065 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4070 unlock_page_cgroup(pc);
4074 void mem_cgroup_uncharge_page(struct page *page)
4077 if (page_mapped(page))
4079 VM_BUG_ON(page->mapping && !PageAnon(page));
4081 * If the page is in swap cache, uncharge should be deferred
4082 * to the swap path, which also properly accounts swap usage
4083 * and handles memcg lifetime.
4085 * Note that this check is not stable and reclaim may add the
4086 * page to swap cache at any time after this. However, if the
4087 * page is not in swap cache by the time page->mapcount hits
4088 * 0, there won't be any page table references to the swap
4089 * slot, and reclaim will free it and not actually write the
4092 if (PageSwapCache(page))
4094 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4097 void mem_cgroup_uncharge_cache_page(struct page *page)
4099 VM_BUG_ON(page_mapped(page));
4100 VM_BUG_ON(page->mapping);
4101 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4105 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4106 * In that cases, pages are freed continuously and we can expect pages
4107 * are in the same memcg. All these calls itself limits the number of
4108 * pages freed at once, then uncharge_start/end() is called properly.
4109 * This may be called prural(2) times in a context,
4112 void mem_cgroup_uncharge_start(void)
4114 current->memcg_batch.do_batch++;
4115 /* We can do nest. */
4116 if (current->memcg_batch.do_batch == 1) {
4117 current->memcg_batch.memcg = NULL;
4118 current->memcg_batch.nr_pages = 0;
4119 current->memcg_batch.memsw_nr_pages = 0;
4123 void mem_cgroup_uncharge_end(void)
4125 struct memcg_batch_info *batch = ¤t->memcg_batch;
4127 if (!batch->do_batch)
4131 if (batch->do_batch) /* If stacked, do nothing. */
4137 * This "batch->memcg" is valid without any css_get/put etc...
4138 * bacause we hide charges behind us.
4140 if (batch->nr_pages)
4141 res_counter_uncharge(&batch->memcg->res,
4142 batch->nr_pages * PAGE_SIZE);
4143 if (batch->memsw_nr_pages)
4144 res_counter_uncharge(&batch->memcg->memsw,
4145 batch->memsw_nr_pages * PAGE_SIZE);
4146 memcg_oom_recover(batch->memcg);
4147 /* forget this pointer (for sanity check) */
4148 batch->memcg = NULL;
4153 * called after __delete_from_swap_cache() and drop "page" account.
4154 * memcg information is recorded to swap_cgroup of "ent"
4157 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4159 struct mem_cgroup *memcg;
4160 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4162 if (!swapout) /* this was a swap cache but the swap is unused ! */
4163 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4165 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4168 * record memcg information, if swapout && memcg != NULL,
4169 * css_get() was called in uncharge().
4171 if (do_swap_account && swapout && memcg)
4172 swap_cgroup_record(ent, css_id(&memcg->css));
4176 #ifdef CONFIG_MEMCG_SWAP
4178 * called from swap_entry_free(). remove record in swap_cgroup and
4179 * uncharge "memsw" account.
4181 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4183 struct mem_cgroup *memcg;
4186 if (!do_swap_account)
4189 id = swap_cgroup_record(ent, 0);
4191 memcg = mem_cgroup_lookup(id);
4194 * We uncharge this because swap is freed.
4195 * This memcg can be obsolete one. We avoid calling css_tryget
4197 if (!mem_cgroup_is_root(memcg))
4198 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4199 mem_cgroup_swap_statistics(memcg, false);
4200 css_put(&memcg->css);
4206 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4207 * @entry: swap entry to be moved
4208 * @from: mem_cgroup which the entry is moved from
4209 * @to: mem_cgroup which the entry is moved to
4211 * It succeeds only when the swap_cgroup's record for this entry is the same
4212 * as the mem_cgroup's id of @from.
4214 * Returns 0 on success, -EINVAL on failure.
4216 * The caller must have charged to @to, IOW, called res_counter_charge() about
4217 * both res and memsw, and called css_get().
4219 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4220 struct mem_cgroup *from, struct mem_cgroup *to)
4222 unsigned short old_id, new_id;
4224 old_id = css_id(&from->css);
4225 new_id = css_id(&to->css);
4227 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4228 mem_cgroup_swap_statistics(from, false);
4229 mem_cgroup_swap_statistics(to, true);
4231 * This function is only called from task migration context now.
4232 * It postpones res_counter and refcount handling till the end
4233 * of task migration(mem_cgroup_clear_mc()) for performance
4234 * improvement. But we cannot postpone css_get(to) because if
4235 * the process that has been moved to @to does swap-in, the
4236 * refcount of @to might be decreased to 0.
4238 * We are in attach() phase, so the cgroup is guaranteed to be
4239 * alive, so we can just call css_get().
4247 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4248 struct mem_cgroup *from, struct mem_cgroup *to)
4255 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4258 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4259 struct mem_cgroup **memcgp)
4261 struct mem_cgroup *memcg = NULL;
4262 unsigned int nr_pages = 1;
4263 struct page_cgroup *pc;
4264 enum charge_type ctype;
4268 if (mem_cgroup_disabled())
4271 if (PageTransHuge(page))
4272 nr_pages <<= compound_order(page);
4274 pc = lookup_page_cgroup(page);
4275 lock_page_cgroup(pc);
4276 if (PageCgroupUsed(pc)) {
4277 memcg = pc->mem_cgroup;
4278 css_get(&memcg->css);
4280 * At migrating an anonymous page, its mapcount goes down
4281 * to 0 and uncharge() will be called. But, even if it's fully
4282 * unmapped, migration may fail and this page has to be
4283 * charged again. We set MIGRATION flag here and delay uncharge
4284 * until end_migration() is called
4286 * Corner Case Thinking
4288 * When the old page was mapped as Anon and it's unmap-and-freed
4289 * while migration was ongoing.
4290 * If unmap finds the old page, uncharge() of it will be delayed
4291 * until end_migration(). If unmap finds a new page, it's
4292 * uncharged when it make mapcount to be 1->0. If unmap code
4293 * finds swap_migration_entry, the new page will not be mapped
4294 * and end_migration() will find it(mapcount==0).
4297 * When the old page was mapped but migraion fails, the kernel
4298 * remaps it. A charge for it is kept by MIGRATION flag even
4299 * if mapcount goes down to 0. We can do remap successfully
4300 * without charging it again.
4303 * The "old" page is under lock_page() until the end of
4304 * migration, so, the old page itself will not be swapped-out.
4305 * If the new page is swapped out before end_migraton, our
4306 * hook to usual swap-out path will catch the event.
4309 SetPageCgroupMigration(pc);
4311 unlock_page_cgroup(pc);
4313 * If the page is not charged at this point,
4321 * We charge new page before it's used/mapped. So, even if unlock_page()
4322 * is called before end_migration, we can catch all events on this new
4323 * page. In the case new page is migrated but not remapped, new page's
4324 * mapcount will be finally 0 and we call uncharge in end_migration().
4327 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4329 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4331 * The page is committed to the memcg, but it's not actually
4332 * charged to the res_counter since we plan on replacing the
4333 * old one and only one page is going to be left afterwards.
4335 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4338 /* remove redundant charge if migration failed*/
4339 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4340 struct page *oldpage, struct page *newpage, bool migration_ok)
4342 struct page *used, *unused;
4343 struct page_cgroup *pc;
4349 if (!migration_ok) {
4356 anon = PageAnon(used);
4357 __mem_cgroup_uncharge_common(unused,
4358 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4359 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4361 css_put(&memcg->css);
4363 * We disallowed uncharge of pages under migration because mapcount
4364 * of the page goes down to zero, temporarly.
4365 * Clear the flag and check the page should be charged.
4367 pc = lookup_page_cgroup(oldpage);
4368 lock_page_cgroup(pc);
4369 ClearPageCgroupMigration(pc);
4370 unlock_page_cgroup(pc);
4373 * If a page is a file cache, radix-tree replacement is very atomic
4374 * and we can skip this check. When it was an Anon page, its mapcount
4375 * goes down to 0. But because we added MIGRATION flage, it's not
4376 * uncharged yet. There are several case but page->mapcount check
4377 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4378 * check. (see prepare_charge() also)
4381 mem_cgroup_uncharge_page(used);
4385 * At replace page cache, newpage is not under any memcg but it's on
4386 * LRU. So, this function doesn't touch res_counter but handles LRU
4387 * in correct way. Both pages are locked so we cannot race with uncharge.
4389 void mem_cgroup_replace_page_cache(struct page *oldpage,
4390 struct page *newpage)
4392 struct mem_cgroup *memcg = NULL;
4393 struct page_cgroup *pc;
4394 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4396 if (mem_cgroup_disabled())
4399 pc = lookup_page_cgroup(oldpage);
4400 /* fix accounting on old pages */
4401 lock_page_cgroup(pc);
4402 if (PageCgroupUsed(pc)) {
4403 memcg = pc->mem_cgroup;
4404 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4405 ClearPageCgroupUsed(pc);
4407 unlock_page_cgroup(pc);
4410 * When called from shmem_replace_page(), in some cases the
4411 * oldpage has already been charged, and in some cases not.
4416 * Even if newpage->mapping was NULL before starting replacement,
4417 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4418 * LRU while we overwrite pc->mem_cgroup.
4420 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4423 #ifdef CONFIG_DEBUG_VM
4424 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4426 struct page_cgroup *pc;
4428 pc = lookup_page_cgroup(page);
4430 * Can be NULL while feeding pages into the page allocator for
4431 * the first time, i.e. during boot or memory hotplug;
4432 * or when mem_cgroup_disabled().
4434 if (likely(pc) && PageCgroupUsed(pc))
4439 bool mem_cgroup_bad_page_check(struct page *page)
4441 if (mem_cgroup_disabled())
4444 return lookup_page_cgroup_used(page) != NULL;
4447 void mem_cgroup_print_bad_page(struct page *page)
4449 struct page_cgroup *pc;
4451 pc = lookup_page_cgroup_used(page);
4453 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4454 pc, pc->flags, pc->mem_cgroup);
4459 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4460 unsigned long long val)
4463 u64 memswlimit, memlimit;
4465 int children = mem_cgroup_count_children(memcg);
4466 u64 curusage, oldusage;
4470 * For keeping hierarchical_reclaim simple, how long we should retry
4471 * is depends on callers. We set our retry-count to be function
4472 * of # of children which we should visit in this loop.
4474 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4476 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4479 while (retry_count) {
4480 if (signal_pending(current)) {
4485 * Rather than hide all in some function, I do this in
4486 * open coded manner. You see what this really does.
4487 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4489 mutex_lock(&set_limit_mutex);
4490 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4491 if (memswlimit < val) {
4493 mutex_unlock(&set_limit_mutex);
4497 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4501 ret = res_counter_set_limit(&memcg->res, val);
4503 if (memswlimit == val)
4504 memcg->memsw_is_minimum = true;
4506 memcg->memsw_is_minimum = false;
4508 mutex_unlock(&set_limit_mutex);
4513 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4514 MEM_CGROUP_RECLAIM_SHRINK);
4515 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4516 /* Usage is reduced ? */
4517 if (curusage >= oldusage)
4520 oldusage = curusage;
4522 if (!ret && enlarge)
4523 memcg_oom_recover(memcg);
4528 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4529 unsigned long long val)
4532 u64 memlimit, memswlimit, oldusage, curusage;
4533 int children = mem_cgroup_count_children(memcg);
4537 /* see mem_cgroup_resize_res_limit */
4538 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4539 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4540 while (retry_count) {
4541 if (signal_pending(current)) {
4546 * Rather than hide all in some function, I do this in
4547 * open coded manner. You see what this really does.
4548 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4550 mutex_lock(&set_limit_mutex);
4551 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4552 if (memlimit > val) {
4554 mutex_unlock(&set_limit_mutex);
4557 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4558 if (memswlimit < val)
4560 ret = res_counter_set_limit(&memcg->memsw, val);
4562 if (memlimit == val)
4563 memcg->memsw_is_minimum = true;
4565 memcg->memsw_is_minimum = false;
4567 mutex_unlock(&set_limit_mutex);
4572 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4573 MEM_CGROUP_RECLAIM_NOSWAP |
4574 MEM_CGROUP_RECLAIM_SHRINK);
4575 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4576 /* Usage is reduced ? */
4577 if (curusage >= oldusage)
4580 oldusage = curusage;
4582 if (!ret && enlarge)
4583 memcg_oom_recover(memcg);
4588 * mem_cgroup_force_empty_list - clears LRU of a group
4589 * @memcg: group to clear
4592 * @lru: lru to to clear
4594 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4595 * reclaim the pages page themselves - pages are moved to the parent (or root)
4598 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4599 int node, int zid, enum lru_list lru)
4601 struct lruvec *lruvec;
4602 unsigned long flags;
4603 struct list_head *list;
4607 zone = &NODE_DATA(node)->node_zones[zid];
4608 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4609 list = &lruvec->lists[lru];
4613 struct page_cgroup *pc;
4616 spin_lock_irqsave(&zone->lru_lock, flags);
4617 if (list_empty(list)) {
4618 spin_unlock_irqrestore(&zone->lru_lock, flags);
4621 page = list_entry(list->prev, struct page, lru);
4623 list_move(&page->lru, list);
4625 spin_unlock_irqrestore(&zone->lru_lock, flags);
4628 spin_unlock_irqrestore(&zone->lru_lock, flags);
4630 pc = lookup_page_cgroup(page);
4632 if (mem_cgroup_move_parent(page, pc, memcg)) {
4633 /* found lock contention or "pc" is obsolete. */
4638 } while (!list_empty(list));
4642 * make mem_cgroup's charge to be 0 if there is no task by moving
4643 * all the charges and pages to the parent.
4644 * This enables deleting this mem_cgroup.
4646 * Caller is responsible for holding css reference on the memcg.
4648 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4654 /* This is for making all *used* pages to be on LRU. */
4655 lru_add_drain_all();
4656 drain_all_stock_sync(memcg);
4657 mem_cgroup_start_move(memcg);
4658 for_each_node_state(node, N_MEMORY) {
4659 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4662 mem_cgroup_force_empty_list(memcg,
4667 mem_cgroup_end_move(memcg);
4668 memcg_oom_recover(memcg);
4672 * Kernel memory may not necessarily be trackable to a specific
4673 * process. So they are not migrated, and therefore we can't
4674 * expect their value to drop to 0 here.
4675 * Having res filled up with kmem only is enough.
4677 * This is a safety check because mem_cgroup_force_empty_list
4678 * could have raced with mem_cgroup_replace_page_cache callers
4679 * so the lru seemed empty but the page could have been added
4680 * right after the check. RES_USAGE should be safe as we always
4681 * charge before adding to the LRU.
4683 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4684 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4685 } while (usage > 0);
4689 * This mainly exists for tests during the setting of set of use_hierarchy.
4690 * Since this is the very setting we are changing, the current hierarchy value
4693 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4695 struct cgroup_subsys_state *pos;
4697 /* bounce at first found */
4698 css_for_each_child(pos, &memcg->css)
4704 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4705 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4706 * from mem_cgroup_count_children(), in the sense that we don't really care how
4707 * many children we have; we only need to know if we have any. It also counts
4708 * any memcg without hierarchy as infertile.
4710 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4712 return memcg->use_hierarchy && __memcg_has_children(memcg);
4716 * Reclaims as many pages from the given memcg as possible and moves
4717 * the rest to the parent.
4719 * Caller is responsible for holding css reference for memcg.
4721 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4723 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4724 struct cgroup *cgrp = memcg->css.cgroup;
4726 /* returns EBUSY if there is a task or if we come here twice. */
4727 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4730 /* we call try-to-free pages for make this cgroup empty */
4731 lru_add_drain_all();
4732 /* try to free all pages in this cgroup */
4733 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4736 if (signal_pending(current))
4739 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4743 /* maybe some writeback is necessary */
4744 congestion_wait(BLK_RW_ASYNC, HZ/10);
4749 mem_cgroup_reparent_charges(memcg);
4754 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4757 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4759 if (mem_cgroup_is_root(memcg))
4761 return mem_cgroup_force_empty(memcg);
4764 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4767 return mem_cgroup_from_css(css)->use_hierarchy;
4770 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4771 struct cftype *cft, u64 val)
4774 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4775 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4777 mutex_lock(&memcg_create_mutex);
4779 if (memcg->use_hierarchy == val)
4783 * If parent's use_hierarchy is set, we can't make any modifications
4784 * in the child subtrees. If it is unset, then the change can
4785 * occur, provided the current cgroup has no children.
4787 * For the root cgroup, parent_mem is NULL, we allow value to be
4788 * set if there are no children.
4790 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4791 (val == 1 || val == 0)) {
4792 if (!__memcg_has_children(memcg))
4793 memcg->use_hierarchy = val;
4800 mutex_unlock(&memcg_create_mutex);
4806 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4807 enum mem_cgroup_stat_index idx)
4809 struct mem_cgroup *iter;
4812 /* Per-cpu values can be negative, use a signed accumulator */
4813 for_each_mem_cgroup_tree(iter, memcg)
4814 val += mem_cgroup_read_stat(iter, idx);
4816 if (val < 0) /* race ? */
4821 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4825 if (!mem_cgroup_is_root(memcg)) {
4827 return res_counter_read_u64(&memcg->res, RES_USAGE);
4829 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4833 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4834 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4836 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4837 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4840 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4842 return val << PAGE_SHIFT;
4845 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
4846 struct cftype *cft, struct file *file,
4847 char __user *buf, size_t nbytes, loff_t *ppos)
4849 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4855 type = MEMFILE_TYPE(cft->private);
4856 name = MEMFILE_ATTR(cft->private);
4860 if (name == RES_USAGE)
4861 val = mem_cgroup_usage(memcg, false);
4863 val = res_counter_read_u64(&memcg->res, name);
4866 if (name == RES_USAGE)
4867 val = mem_cgroup_usage(memcg, true);
4869 val = res_counter_read_u64(&memcg->memsw, name);
4872 val = res_counter_read_u64(&memcg->kmem, name);
4878 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4879 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4882 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
4885 #ifdef CONFIG_MEMCG_KMEM
4886 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4888 * For simplicity, we won't allow this to be disabled. It also can't
4889 * be changed if the cgroup has children already, or if tasks had
4892 * If tasks join before we set the limit, a person looking at
4893 * kmem.usage_in_bytes will have no way to determine when it took
4894 * place, which makes the value quite meaningless.
4896 * After it first became limited, changes in the value of the limit are
4897 * of course permitted.
4899 mutex_lock(&memcg_create_mutex);
4900 mutex_lock(&set_limit_mutex);
4901 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
4902 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
4906 ret = res_counter_set_limit(&memcg->kmem, val);
4909 ret = memcg_update_cache_sizes(memcg);
4911 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
4914 static_key_slow_inc(&memcg_kmem_enabled_key);
4916 * setting the active bit after the inc will guarantee no one
4917 * starts accounting before all call sites are patched
4919 memcg_kmem_set_active(memcg);
4921 ret = res_counter_set_limit(&memcg->kmem, val);
4923 mutex_unlock(&set_limit_mutex);
4924 mutex_unlock(&memcg_create_mutex);
4929 #ifdef CONFIG_MEMCG_KMEM
4930 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4933 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4937 memcg->kmem_account_flags = parent->kmem_account_flags;
4939 * When that happen, we need to disable the static branch only on those
4940 * memcgs that enabled it. To achieve this, we would be forced to
4941 * complicate the code by keeping track of which memcgs were the ones
4942 * that actually enabled limits, and which ones got it from its
4945 * It is a lot simpler just to do static_key_slow_inc() on every child
4946 * that is accounted.
4948 if (!memcg_kmem_is_active(memcg))
4952 * __mem_cgroup_free() will issue static_key_slow_dec() because this
4953 * memcg is active already. If the later initialization fails then the
4954 * cgroup core triggers the cleanup so we do not have to do it here.
4956 static_key_slow_inc(&memcg_kmem_enabled_key);
4958 mutex_lock(&set_limit_mutex);
4959 memcg_stop_kmem_account();
4960 ret = memcg_update_cache_sizes(memcg);
4961 memcg_resume_kmem_account();
4962 mutex_unlock(&set_limit_mutex);
4966 #endif /* CONFIG_MEMCG_KMEM */
4969 * The user of this function is...
4972 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
4975 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4978 unsigned long long val;
4981 type = MEMFILE_TYPE(cft->private);
4982 name = MEMFILE_ATTR(cft->private);
4986 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4990 /* This function does all necessary parse...reuse it */
4991 ret = res_counter_memparse_write_strategy(buffer, &val);
4995 ret = mem_cgroup_resize_limit(memcg, val);
4996 else if (type == _MEMSWAP)
4997 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4998 else if (type == _KMEM)
4999 ret = memcg_update_kmem_limit(css, val);
5003 case RES_SOFT_LIMIT:
5004 ret = res_counter_memparse_write_strategy(buffer, &val);
5008 * For memsw, soft limits are hard to implement in terms
5009 * of semantics, for now, we support soft limits for
5010 * control without swap
5013 ret = res_counter_set_soft_limit(&memcg->res, val);
5018 ret = -EINVAL; /* should be BUG() ? */
5024 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5025 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5027 unsigned long long min_limit, min_memsw_limit, tmp;
5029 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5030 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5031 if (!memcg->use_hierarchy)
5034 while (css_parent(&memcg->css)) {
5035 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5036 if (!memcg->use_hierarchy)
5038 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5039 min_limit = min(min_limit, tmp);
5040 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5041 min_memsw_limit = min(min_memsw_limit, tmp);
5044 *mem_limit = min_limit;
5045 *memsw_limit = min_memsw_limit;
5048 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5050 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5054 type = MEMFILE_TYPE(event);
5055 name = MEMFILE_ATTR(event);
5060 res_counter_reset_max(&memcg->res);
5061 else if (type == _MEMSWAP)
5062 res_counter_reset_max(&memcg->memsw);
5063 else if (type == _KMEM)
5064 res_counter_reset_max(&memcg->kmem);
5070 res_counter_reset_failcnt(&memcg->res);
5071 else if (type == _MEMSWAP)
5072 res_counter_reset_failcnt(&memcg->memsw);
5073 else if (type == _KMEM)
5074 res_counter_reset_failcnt(&memcg->kmem);
5083 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5086 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5090 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5091 struct cftype *cft, u64 val)
5093 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5095 if (val >= (1 << NR_MOVE_TYPE))
5099 * No kind of locking is needed in here, because ->can_attach() will
5100 * check this value once in the beginning of the process, and then carry
5101 * on with stale data. This means that changes to this value will only
5102 * affect task migrations starting after the change.
5104 memcg->move_charge_at_immigrate = val;
5108 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5109 struct cftype *cft, u64 val)
5116 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5117 struct cftype *cft, struct seq_file *m)
5120 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5121 unsigned long node_nr;
5122 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5124 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5125 seq_printf(m, "total=%lu", total_nr);
5126 for_each_node_state(nid, N_MEMORY) {
5127 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5128 seq_printf(m, " N%d=%lu", nid, node_nr);
5132 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5133 seq_printf(m, "file=%lu", file_nr);
5134 for_each_node_state(nid, N_MEMORY) {
5135 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5137 seq_printf(m, " N%d=%lu", nid, node_nr);
5141 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5142 seq_printf(m, "anon=%lu", anon_nr);
5143 for_each_node_state(nid, N_MEMORY) {
5144 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5146 seq_printf(m, " N%d=%lu", nid, node_nr);
5150 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5151 seq_printf(m, "unevictable=%lu", unevictable_nr);
5152 for_each_node_state(nid, N_MEMORY) {
5153 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5154 BIT(LRU_UNEVICTABLE));
5155 seq_printf(m, " N%d=%lu", nid, node_nr);
5160 #endif /* CONFIG_NUMA */
5162 static inline void mem_cgroup_lru_names_not_uptodate(void)
5164 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5167 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5170 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5171 struct mem_cgroup *mi;
5174 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5175 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5177 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5178 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5181 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5182 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5183 mem_cgroup_read_events(memcg, i));
5185 for (i = 0; i < NR_LRU_LISTS; i++)
5186 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5187 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5189 /* Hierarchical information */
5191 unsigned long long limit, memsw_limit;
5192 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5193 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5194 if (do_swap_account)
5195 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5199 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5202 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5204 for_each_mem_cgroup_tree(mi, memcg)
5205 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5206 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5209 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5210 unsigned long long val = 0;
5212 for_each_mem_cgroup_tree(mi, memcg)
5213 val += mem_cgroup_read_events(mi, i);
5214 seq_printf(m, "total_%s %llu\n",
5215 mem_cgroup_events_names[i], val);
5218 for (i = 0; i < NR_LRU_LISTS; i++) {
5219 unsigned long long val = 0;
5221 for_each_mem_cgroup_tree(mi, memcg)
5222 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5223 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5226 #ifdef CONFIG_DEBUG_VM
5229 struct mem_cgroup_per_zone *mz;
5230 struct zone_reclaim_stat *rstat;
5231 unsigned long recent_rotated[2] = {0, 0};
5232 unsigned long recent_scanned[2] = {0, 0};
5234 for_each_online_node(nid)
5235 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5236 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5237 rstat = &mz->lruvec.reclaim_stat;
5239 recent_rotated[0] += rstat->recent_rotated[0];
5240 recent_rotated[1] += rstat->recent_rotated[1];
5241 recent_scanned[0] += rstat->recent_scanned[0];
5242 recent_scanned[1] += rstat->recent_scanned[1];
5244 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5245 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5246 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5247 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5254 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5257 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5259 return mem_cgroup_swappiness(memcg);
5262 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5263 struct cftype *cft, u64 val)
5265 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5266 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5268 if (val > 100 || !parent)
5271 mutex_lock(&memcg_create_mutex);
5273 /* If under hierarchy, only empty-root can set this value */
5274 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5275 mutex_unlock(&memcg_create_mutex);
5279 memcg->swappiness = val;
5281 mutex_unlock(&memcg_create_mutex);
5286 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5288 struct mem_cgroup_threshold_ary *t;
5294 t = rcu_dereference(memcg->thresholds.primary);
5296 t = rcu_dereference(memcg->memsw_thresholds.primary);
5301 usage = mem_cgroup_usage(memcg, swap);
5304 * current_threshold points to threshold just below or equal to usage.
5305 * If it's not true, a threshold was crossed after last
5306 * call of __mem_cgroup_threshold().
5308 i = t->current_threshold;
5311 * Iterate backward over array of thresholds starting from
5312 * current_threshold and check if a threshold is crossed.
5313 * If none of thresholds below usage is crossed, we read
5314 * only one element of the array here.
5316 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5317 eventfd_signal(t->entries[i].eventfd, 1);
5319 /* i = current_threshold + 1 */
5323 * Iterate forward over array of thresholds starting from
5324 * current_threshold+1 and check if a threshold is crossed.
5325 * If none of thresholds above usage is crossed, we read
5326 * only one element of the array here.
5328 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5329 eventfd_signal(t->entries[i].eventfd, 1);
5331 /* Update current_threshold */
5332 t->current_threshold = i - 1;
5337 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5340 __mem_cgroup_threshold(memcg, false);
5341 if (do_swap_account)
5342 __mem_cgroup_threshold(memcg, true);
5344 memcg = parent_mem_cgroup(memcg);
5348 static int compare_thresholds(const void *a, const void *b)
5350 const struct mem_cgroup_threshold *_a = a;
5351 const struct mem_cgroup_threshold *_b = b;
5353 if (_a->threshold > _b->threshold)
5356 if (_a->threshold < _b->threshold)
5362 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5364 struct mem_cgroup_eventfd_list *ev;
5366 list_for_each_entry(ev, &memcg->oom_notify, list)
5367 eventfd_signal(ev->eventfd, 1);
5371 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5373 struct mem_cgroup *iter;
5375 for_each_mem_cgroup_tree(iter, memcg)
5376 mem_cgroup_oom_notify_cb(iter);
5379 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5380 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5382 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5383 struct mem_cgroup_thresholds *thresholds;
5384 struct mem_cgroup_threshold_ary *new;
5385 enum res_type type = MEMFILE_TYPE(cft->private);
5386 u64 threshold, usage;
5389 ret = res_counter_memparse_write_strategy(args, &threshold);
5393 mutex_lock(&memcg->thresholds_lock);
5396 thresholds = &memcg->thresholds;
5397 else if (type == _MEMSWAP)
5398 thresholds = &memcg->memsw_thresholds;
5402 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5404 /* Check if a threshold crossed before adding a new one */
5405 if (thresholds->primary)
5406 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5408 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5410 /* Allocate memory for new array of thresholds */
5411 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5419 /* Copy thresholds (if any) to new array */
5420 if (thresholds->primary) {
5421 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5422 sizeof(struct mem_cgroup_threshold));
5425 /* Add new threshold */
5426 new->entries[size - 1].eventfd = eventfd;
5427 new->entries[size - 1].threshold = threshold;
5429 /* Sort thresholds. Registering of new threshold isn't time-critical */
5430 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5431 compare_thresholds, NULL);
5433 /* Find current threshold */
5434 new->current_threshold = -1;
5435 for (i = 0; i < size; i++) {
5436 if (new->entries[i].threshold <= usage) {
5438 * new->current_threshold will not be used until
5439 * rcu_assign_pointer(), so it's safe to increment
5442 ++new->current_threshold;
5447 /* Free old spare buffer and save old primary buffer as spare */
5448 kfree(thresholds->spare);
5449 thresholds->spare = thresholds->primary;
5451 rcu_assign_pointer(thresholds->primary, new);
5453 /* To be sure that nobody uses thresholds */
5457 mutex_unlock(&memcg->thresholds_lock);
5462 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5463 struct cftype *cft, struct eventfd_ctx *eventfd)
5465 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5466 struct mem_cgroup_thresholds *thresholds;
5467 struct mem_cgroup_threshold_ary *new;
5468 enum res_type type = MEMFILE_TYPE(cft->private);
5472 mutex_lock(&memcg->thresholds_lock);
5474 thresholds = &memcg->thresholds;
5475 else if (type == _MEMSWAP)
5476 thresholds = &memcg->memsw_thresholds;
5480 if (!thresholds->primary)
5483 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5485 /* Check if a threshold crossed before removing */
5486 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5488 /* Calculate new number of threshold */
5490 for (i = 0; i < thresholds->primary->size; i++) {
5491 if (thresholds->primary->entries[i].eventfd != eventfd)
5495 new = thresholds->spare;
5497 /* Set thresholds array to NULL if we don't have thresholds */
5506 /* Copy thresholds and find current threshold */
5507 new->current_threshold = -1;
5508 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5509 if (thresholds->primary->entries[i].eventfd == eventfd)
5512 new->entries[j] = thresholds->primary->entries[i];
5513 if (new->entries[j].threshold <= usage) {
5515 * new->current_threshold will not be used
5516 * until rcu_assign_pointer(), so it's safe to increment
5519 ++new->current_threshold;
5525 /* Swap primary and spare array */
5526 thresholds->spare = thresholds->primary;
5527 /* If all events are unregistered, free the spare array */
5529 kfree(thresholds->spare);
5530 thresholds->spare = NULL;
5533 rcu_assign_pointer(thresholds->primary, new);
5535 /* To be sure that nobody uses thresholds */
5538 mutex_unlock(&memcg->thresholds_lock);
5541 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5542 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5544 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5545 struct mem_cgroup_eventfd_list *event;
5546 enum res_type type = MEMFILE_TYPE(cft->private);
5548 BUG_ON(type != _OOM_TYPE);
5549 event = kmalloc(sizeof(*event), GFP_KERNEL);
5553 spin_lock(&memcg_oom_lock);
5555 event->eventfd = eventfd;
5556 list_add(&event->list, &memcg->oom_notify);
5558 /* already in OOM ? */
5559 if (atomic_read(&memcg->under_oom))
5560 eventfd_signal(eventfd, 1);
5561 spin_unlock(&memcg_oom_lock);
5566 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5567 struct cftype *cft, struct eventfd_ctx *eventfd)
5569 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5570 struct mem_cgroup_eventfd_list *ev, *tmp;
5571 enum res_type type = MEMFILE_TYPE(cft->private);
5573 BUG_ON(type != _OOM_TYPE);
5575 spin_lock(&memcg_oom_lock);
5577 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5578 if (ev->eventfd == eventfd) {
5579 list_del(&ev->list);
5584 spin_unlock(&memcg_oom_lock);
5587 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5588 struct cftype *cft, struct cgroup_map_cb *cb)
5590 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5592 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5594 if (atomic_read(&memcg->under_oom))
5595 cb->fill(cb, "under_oom", 1);
5597 cb->fill(cb, "under_oom", 0);
5601 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5602 struct cftype *cft, u64 val)
5604 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5605 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5607 /* cannot set to root cgroup and only 0 and 1 are allowed */
5608 if (!parent || !((val == 0) || (val == 1)))
5611 mutex_lock(&memcg_create_mutex);
5612 /* oom-kill-disable is a flag for subhierarchy. */
5613 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5614 mutex_unlock(&memcg_create_mutex);
5617 memcg->oom_kill_disable = val;
5619 memcg_oom_recover(memcg);
5620 mutex_unlock(&memcg_create_mutex);
5624 #ifdef CONFIG_MEMCG_KMEM
5625 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5629 memcg->kmemcg_id = -1;
5630 ret = memcg_propagate_kmem(memcg);
5634 return mem_cgroup_sockets_init(memcg, ss);
5637 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5639 mem_cgroup_sockets_destroy(memcg);
5642 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5644 if (!memcg_kmem_is_active(memcg))
5648 * kmem charges can outlive the cgroup. In the case of slab
5649 * pages, for instance, a page contain objects from various
5650 * processes. As we prevent from taking a reference for every
5651 * such allocation we have to be careful when doing uncharge
5652 * (see memcg_uncharge_kmem) and here during offlining.
5654 * The idea is that that only the _last_ uncharge which sees
5655 * the dead memcg will drop the last reference. An additional
5656 * reference is taken here before the group is marked dead
5657 * which is then paired with css_put during uncharge resp. here.
5659 * Although this might sound strange as this path is called from
5660 * css_offline() when the referencemight have dropped down to 0
5661 * and shouldn't be incremented anymore (css_tryget would fail)
5662 * we do not have other options because of the kmem allocations
5665 css_get(&memcg->css);
5667 memcg_kmem_mark_dead(memcg);
5669 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5672 if (memcg_kmem_test_and_clear_dead(memcg))
5673 css_put(&memcg->css);
5676 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5681 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5685 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5690 static struct cftype mem_cgroup_files[] = {
5692 .name = "usage_in_bytes",
5693 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5694 .read = mem_cgroup_read,
5695 .register_event = mem_cgroup_usage_register_event,
5696 .unregister_event = mem_cgroup_usage_unregister_event,
5699 .name = "max_usage_in_bytes",
5700 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5701 .trigger = mem_cgroup_reset,
5702 .read = mem_cgroup_read,
5705 .name = "limit_in_bytes",
5706 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5707 .write_string = mem_cgroup_write,
5708 .read = mem_cgroup_read,
5711 .name = "soft_limit_in_bytes",
5712 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5713 .write_string = mem_cgroup_write,
5714 .read = mem_cgroup_read,
5718 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5719 .trigger = mem_cgroup_reset,
5720 .read = mem_cgroup_read,
5724 .read_seq_string = memcg_stat_show,
5727 .name = "force_empty",
5728 .trigger = mem_cgroup_force_empty_write,
5731 .name = "use_hierarchy",
5732 .flags = CFTYPE_INSANE,
5733 .write_u64 = mem_cgroup_hierarchy_write,
5734 .read_u64 = mem_cgroup_hierarchy_read,
5737 .name = "swappiness",
5738 .read_u64 = mem_cgroup_swappiness_read,
5739 .write_u64 = mem_cgroup_swappiness_write,
5742 .name = "move_charge_at_immigrate",
5743 .read_u64 = mem_cgroup_move_charge_read,
5744 .write_u64 = mem_cgroup_move_charge_write,
5747 .name = "oom_control",
5748 .read_map = mem_cgroup_oom_control_read,
5749 .write_u64 = mem_cgroup_oom_control_write,
5750 .register_event = mem_cgroup_oom_register_event,
5751 .unregister_event = mem_cgroup_oom_unregister_event,
5752 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5755 .name = "pressure_level",
5756 .register_event = vmpressure_register_event,
5757 .unregister_event = vmpressure_unregister_event,
5761 .name = "numa_stat",
5762 .read_seq_string = memcg_numa_stat_show,
5765 #ifdef CONFIG_MEMCG_KMEM
5767 .name = "kmem.limit_in_bytes",
5768 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5769 .write_string = mem_cgroup_write,
5770 .read = mem_cgroup_read,
5773 .name = "kmem.usage_in_bytes",
5774 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5775 .read = mem_cgroup_read,
5778 .name = "kmem.failcnt",
5779 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5780 .trigger = mem_cgroup_reset,
5781 .read = mem_cgroup_read,
5784 .name = "kmem.max_usage_in_bytes",
5785 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5786 .trigger = mem_cgroup_reset,
5787 .read = mem_cgroup_read,
5789 #ifdef CONFIG_SLABINFO
5791 .name = "kmem.slabinfo",
5792 .read_seq_string = mem_cgroup_slabinfo_read,
5796 { }, /* terminate */
5799 #ifdef CONFIG_MEMCG_SWAP
5800 static struct cftype memsw_cgroup_files[] = {
5802 .name = "memsw.usage_in_bytes",
5803 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5804 .read = mem_cgroup_read,
5805 .register_event = mem_cgroup_usage_register_event,
5806 .unregister_event = mem_cgroup_usage_unregister_event,
5809 .name = "memsw.max_usage_in_bytes",
5810 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5811 .trigger = mem_cgroup_reset,
5812 .read = mem_cgroup_read,
5815 .name = "memsw.limit_in_bytes",
5816 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5817 .write_string = mem_cgroup_write,
5818 .read = mem_cgroup_read,
5821 .name = "memsw.failcnt",
5822 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5823 .trigger = mem_cgroup_reset,
5824 .read = mem_cgroup_read,
5826 { }, /* terminate */
5829 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5831 struct mem_cgroup_per_node *pn;
5832 struct mem_cgroup_per_zone *mz;
5833 int zone, tmp = node;
5835 * This routine is called against possible nodes.
5836 * But it's BUG to call kmalloc() against offline node.
5838 * TODO: this routine can waste much memory for nodes which will
5839 * never be onlined. It's better to use memory hotplug callback
5842 if (!node_state(node, N_NORMAL_MEMORY))
5844 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5848 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5849 mz = &pn->zoneinfo[zone];
5850 lruvec_init(&mz->lruvec);
5853 memcg->nodeinfo[node] = pn;
5857 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5859 kfree(memcg->nodeinfo[node]);
5862 static struct mem_cgroup *mem_cgroup_alloc(void)
5864 struct mem_cgroup *memcg;
5865 size_t size = memcg_size();
5867 /* Can be very big if nr_node_ids is very big */
5868 if (size < PAGE_SIZE)
5869 memcg = kzalloc(size, GFP_KERNEL);
5871 memcg = vzalloc(size);
5876 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5879 spin_lock_init(&memcg->pcp_counter_lock);
5883 if (size < PAGE_SIZE)
5891 * At destroying mem_cgroup, references from swap_cgroup can remain.
5892 * (scanning all at force_empty is too costly...)
5894 * Instead of clearing all references at force_empty, we remember
5895 * the number of reference from swap_cgroup and free mem_cgroup when
5896 * it goes down to 0.
5898 * Removal of cgroup itself succeeds regardless of refs from swap.
5901 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5904 size_t size = memcg_size();
5906 free_css_id(&mem_cgroup_subsys, &memcg->css);
5909 free_mem_cgroup_per_zone_info(memcg, node);
5911 free_percpu(memcg->stat);
5914 * We need to make sure that (at least for now), the jump label
5915 * destruction code runs outside of the cgroup lock. This is because
5916 * get_online_cpus(), which is called from the static_branch update,
5917 * can't be called inside the cgroup_lock. cpusets are the ones
5918 * enforcing this dependency, so if they ever change, we might as well.
5920 * schedule_work() will guarantee this happens. Be careful if you need
5921 * to move this code around, and make sure it is outside
5924 disarm_static_keys(memcg);
5925 if (size < PAGE_SIZE)
5932 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5934 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5936 if (!memcg->res.parent)
5938 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5940 EXPORT_SYMBOL(parent_mem_cgroup);
5942 static struct cgroup_subsys_state * __ref
5943 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5945 struct mem_cgroup *memcg;
5946 long error = -ENOMEM;
5949 memcg = mem_cgroup_alloc();
5951 return ERR_PTR(error);
5954 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5958 if (parent_css == NULL) {
5959 root_mem_cgroup = memcg;
5960 res_counter_init(&memcg->res, NULL);
5961 res_counter_init(&memcg->memsw, NULL);
5962 res_counter_init(&memcg->kmem, NULL);
5965 memcg->last_scanned_node = MAX_NUMNODES;
5966 INIT_LIST_HEAD(&memcg->oom_notify);
5967 memcg->move_charge_at_immigrate = 0;
5968 mutex_init(&memcg->thresholds_lock);
5969 spin_lock_init(&memcg->move_lock);
5970 vmpressure_init(&memcg->vmpressure);
5971 spin_lock_init(&memcg->soft_lock);
5976 __mem_cgroup_free(memcg);
5977 return ERR_PTR(error);
5981 mem_cgroup_css_online(struct cgroup_subsys_state *css)
5983 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5984 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
5990 mutex_lock(&memcg_create_mutex);
5992 memcg->use_hierarchy = parent->use_hierarchy;
5993 memcg->oom_kill_disable = parent->oom_kill_disable;
5994 memcg->swappiness = mem_cgroup_swappiness(parent);
5996 if (parent->use_hierarchy) {
5997 res_counter_init(&memcg->res, &parent->res);
5998 res_counter_init(&memcg->memsw, &parent->memsw);
5999 res_counter_init(&memcg->kmem, &parent->kmem);
6002 * No need to take a reference to the parent because cgroup
6003 * core guarantees its existence.
6006 res_counter_init(&memcg->res, NULL);
6007 res_counter_init(&memcg->memsw, NULL);
6008 res_counter_init(&memcg->kmem, NULL);
6010 * Deeper hierachy with use_hierarchy == false doesn't make
6011 * much sense so let cgroup subsystem know about this
6012 * unfortunate state in our controller.
6014 if (parent != root_mem_cgroup)
6015 mem_cgroup_subsys.broken_hierarchy = true;
6018 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6019 mutex_unlock(&memcg_create_mutex);
6024 * Announce all parents that a group from their hierarchy is gone.
6026 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6028 struct mem_cgroup *parent = memcg;
6030 while ((parent = parent_mem_cgroup(parent)))
6031 mem_cgroup_iter_invalidate(parent);
6034 * if the root memcg is not hierarchical we have to check it
6037 if (!root_mem_cgroup->use_hierarchy)
6038 mem_cgroup_iter_invalidate(root_mem_cgroup);
6041 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6043 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6045 kmem_cgroup_css_offline(memcg);
6047 mem_cgroup_invalidate_reclaim_iterators(memcg);
6048 mem_cgroup_reparent_charges(memcg);
6049 if (memcg->soft_contributed) {
6050 while ((memcg = parent_mem_cgroup(memcg)))
6051 atomic_dec(&memcg->children_in_excess);
6053 if (memcg != root_mem_cgroup && !root_mem_cgroup->use_hierarchy)
6054 atomic_dec(&root_mem_cgroup->children_in_excess);
6056 mem_cgroup_destroy_all_caches(memcg);
6057 vmpressure_cleanup(&memcg->vmpressure);
6060 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6062 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6064 memcg_destroy_kmem(memcg);
6065 __mem_cgroup_free(memcg);
6069 /* Handlers for move charge at task migration. */
6070 #define PRECHARGE_COUNT_AT_ONCE 256
6071 static int mem_cgroup_do_precharge(unsigned long count)
6074 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6075 struct mem_cgroup *memcg = mc.to;
6077 if (mem_cgroup_is_root(memcg)) {
6078 mc.precharge += count;
6079 /* we don't need css_get for root */
6082 /* try to charge at once */
6084 struct res_counter *dummy;
6086 * "memcg" cannot be under rmdir() because we've already checked
6087 * by cgroup_lock_live_cgroup() that it is not removed and we
6088 * are still under the same cgroup_mutex. So we can postpone
6091 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6093 if (do_swap_account && res_counter_charge(&memcg->memsw,
6094 PAGE_SIZE * count, &dummy)) {
6095 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6098 mc.precharge += count;
6102 /* fall back to one by one charge */
6104 if (signal_pending(current)) {
6108 if (!batch_count--) {
6109 batch_count = PRECHARGE_COUNT_AT_ONCE;
6112 ret = __mem_cgroup_try_charge(NULL,
6113 GFP_KERNEL, 1, &memcg, false);
6115 /* mem_cgroup_clear_mc() will do uncharge later */
6123 * get_mctgt_type - get target type of moving charge
6124 * @vma: the vma the pte to be checked belongs
6125 * @addr: the address corresponding to the pte to be checked
6126 * @ptent: the pte to be checked
6127 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6130 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6131 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6132 * move charge. if @target is not NULL, the page is stored in target->page
6133 * with extra refcnt got(Callers should handle it).
6134 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6135 * target for charge migration. if @target is not NULL, the entry is stored
6138 * Called with pte lock held.
6145 enum mc_target_type {
6151 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6152 unsigned long addr, pte_t ptent)
6154 struct page *page = vm_normal_page(vma, addr, ptent);
6156 if (!page || !page_mapped(page))
6158 if (PageAnon(page)) {
6159 /* we don't move shared anon */
6162 } else if (!move_file())
6163 /* we ignore mapcount for file pages */
6165 if (!get_page_unless_zero(page))
6172 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6173 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6175 struct page *page = NULL;
6176 swp_entry_t ent = pte_to_swp_entry(ptent);
6178 if (!move_anon() || non_swap_entry(ent))
6181 * Because lookup_swap_cache() updates some statistics counter,
6182 * we call find_get_page() with swapper_space directly.
6184 page = find_get_page(swap_address_space(ent), ent.val);
6185 if (do_swap_account)
6186 entry->val = ent.val;
6191 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6192 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6198 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6199 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6201 struct page *page = NULL;
6202 struct address_space *mapping;
6205 if (!vma->vm_file) /* anonymous vma */
6210 mapping = vma->vm_file->f_mapping;
6211 if (pte_none(ptent))
6212 pgoff = linear_page_index(vma, addr);
6213 else /* pte_file(ptent) is true */
6214 pgoff = pte_to_pgoff(ptent);
6216 /* page is moved even if it's not RSS of this task(page-faulted). */
6217 page = find_get_page(mapping, pgoff);
6220 /* shmem/tmpfs may report page out on swap: account for that too. */
6221 if (radix_tree_exceptional_entry(page)) {
6222 swp_entry_t swap = radix_to_swp_entry(page);
6223 if (do_swap_account)
6225 page = find_get_page(swap_address_space(swap), swap.val);
6231 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6232 unsigned long addr, pte_t ptent, union mc_target *target)
6234 struct page *page = NULL;
6235 struct page_cgroup *pc;
6236 enum mc_target_type ret = MC_TARGET_NONE;
6237 swp_entry_t ent = { .val = 0 };
6239 if (pte_present(ptent))
6240 page = mc_handle_present_pte(vma, addr, ptent);
6241 else if (is_swap_pte(ptent))
6242 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6243 else if (pte_none(ptent) || pte_file(ptent))
6244 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6246 if (!page && !ent.val)
6249 pc = lookup_page_cgroup(page);
6251 * Do only loose check w/o page_cgroup lock.
6252 * mem_cgroup_move_account() checks the pc is valid or not under
6255 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6256 ret = MC_TARGET_PAGE;
6258 target->page = page;
6260 if (!ret || !target)
6263 /* There is a swap entry and a page doesn't exist or isn't charged */
6264 if (ent.val && !ret &&
6265 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6266 ret = MC_TARGET_SWAP;
6273 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6275 * We don't consider swapping or file mapped pages because THP does not
6276 * support them for now.
6277 * Caller should make sure that pmd_trans_huge(pmd) is true.
6279 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6280 unsigned long addr, pmd_t pmd, union mc_target *target)
6282 struct page *page = NULL;
6283 struct page_cgroup *pc;
6284 enum mc_target_type ret = MC_TARGET_NONE;
6286 page = pmd_page(pmd);
6287 VM_BUG_ON(!page || !PageHead(page));
6290 pc = lookup_page_cgroup(page);
6291 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6292 ret = MC_TARGET_PAGE;
6295 target->page = page;
6301 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6302 unsigned long addr, pmd_t pmd, union mc_target *target)
6304 return MC_TARGET_NONE;
6308 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6309 unsigned long addr, unsigned long end,
6310 struct mm_walk *walk)
6312 struct vm_area_struct *vma = walk->private;
6316 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6317 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6318 mc.precharge += HPAGE_PMD_NR;
6319 spin_unlock(&vma->vm_mm->page_table_lock);
6323 if (pmd_trans_unstable(pmd))
6325 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6326 for (; addr != end; pte++, addr += PAGE_SIZE)
6327 if (get_mctgt_type(vma, addr, *pte, NULL))
6328 mc.precharge++; /* increment precharge temporarily */
6329 pte_unmap_unlock(pte - 1, ptl);
6335 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6337 unsigned long precharge;
6338 struct vm_area_struct *vma;
6340 down_read(&mm->mmap_sem);
6341 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6342 struct mm_walk mem_cgroup_count_precharge_walk = {
6343 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6347 if (is_vm_hugetlb_page(vma))
6349 walk_page_range(vma->vm_start, vma->vm_end,
6350 &mem_cgroup_count_precharge_walk);
6352 up_read(&mm->mmap_sem);
6354 precharge = mc.precharge;
6360 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6362 unsigned long precharge = mem_cgroup_count_precharge(mm);
6364 VM_BUG_ON(mc.moving_task);
6365 mc.moving_task = current;
6366 return mem_cgroup_do_precharge(precharge);
6369 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6370 static void __mem_cgroup_clear_mc(void)
6372 struct mem_cgroup *from = mc.from;
6373 struct mem_cgroup *to = mc.to;
6376 /* we must uncharge all the leftover precharges from mc.to */
6378 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6382 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6383 * we must uncharge here.
6385 if (mc.moved_charge) {
6386 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6387 mc.moved_charge = 0;
6389 /* we must fixup refcnts and charges */
6390 if (mc.moved_swap) {
6391 /* uncharge swap account from the old cgroup */
6392 if (!mem_cgroup_is_root(mc.from))
6393 res_counter_uncharge(&mc.from->memsw,
6394 PAGE_SIZE * mc.moved_swap);
6396 for (i = 0; i < mc.moved_swap; i++)
6397 css_put(&mc.from->css);
6399 if (!mem_cgroup_is_root(mc.to)) {
6401 * we charged both to->res and to->memsw, so we should
6404 res_counter_uncharge(&mc.to->res,
6405 PAGE_SIZE * mc.moved_swap);
6407 /* we've already done css_get(mc.to) */
6410 memcg_oom_recover(from);
6411 memcg_oom_recover(to);
6412 wake_up_all(&mc.waitq);
6415 static void mem_cgroup_clear_mc(void)
6417 struct mem_cgroup *from = mc.from;
6420 * we must clear moving_task before waking up waiters at the end of
6423 mc.moving_task = NULL;
6424 __mem_cgroup_clear_mc();
6425 spin_lock(&mc.lock);
6428 spin_unlock(&mc.lock);
6429 mem_cgroup_end_move(from);
6432 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6433 struct cgroup_taskset *tset)
6435 struct task_struct *p = cgroup_taskset_first(tset);
6437 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6438 unsigned long move_charge_at_immigrate;
6441 * We are now commited to this value whatever it is. Changes in this
6442 * tunable will only affect upcoming migrations, not the current one.
6443 * So we need to save it, and keep it going.
6445 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6446 if (move_charge_at_immigrate) {
6447 struct mm_struct *mm;
6448 struct mem_cgroup *from = mem_cgroup_from_task(p);
6450 VM_BUG_ON(from == memcg);
6452 mm = get_task_mm(p);
6455 /* We move charges only when we move a owner of the mm */
6456 if (mm->owner == p) {
6459 VM_BUG_ON(mc.precharge);
6460 VM_BUG_ON(mc.moved_charge);
6461 VM_BUG_ON(mc.moved_swap);
6462 mem_cgroup_start_move(from);
6463 spin_lock(&mc.lock);
6466 mc.immigrate_flags = move_charge_at_immigrate;
6467 spin_unlock(&mc.lock);
6468 /* We set mc.moving_task later */
6470 ret = mem_cgroup_precharge_mc(mm);
6472 mem_cgroup_clear_mc();
6479 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6480 struct cgroup_taskset *tset)
6482 mem_cgroup_clear_mc();
6485 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6486 unsigned long addr, unsigned long end,
6487 struct mm_walk *walk)
6490 struct vm_area_struct *vma = walk->private;
6493 enum mc_target_type target_type;
6494 union mc_target target;
6496 struct page_cgroup *pc;
6499 * We don't take compound_lock() here but no race with splitting thp
6501 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6502 * under splitting, which means there's no concurrent thp split,
6503 * - if another thread runs into split_huge_page() just after we
6504 * entered this if-block, the thread must wait for page table lock
6505 * to be unlocked in __split_huge_page_splitting(), where the main
6506 * part of thp split is not executed yet.
6508 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6509 if (mc.precharge < HPAGE_PMD_NR) {
6510 spin_unlock(&vma->vm_mm->page_table_lock);
6513 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6514 if (target_type == MC_TARGET_PAGE) {
6516 if (!isolate_lru_page(page)) {
6517 pc = lookup_page_cgroup(page);
6518 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6519 pc, mc.from, mc.to)) {
6520 mc.precharge -= HPAGE_PMD_NR;
6521 mc.moved_charge += HPAGE_PMD_NR;
6523 putback_lru_page(page);
6527 spin_unlock(&vma->vm_mm->page_table_lock);
6531 if (pmd_trans_unstable(pmd))
6534 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6535 for (; addr != end; addr += PAGE_SIZE) {
6536 pte_t ptent = *(pte++);
6542 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6543 case MC_TARGET_PAGE:
6545 if (isolate_lru_page(page))
6547 pc = lookup_page_cgroup(page);
6548 if (!mem_cgroup_move_account(page, 1, pc,
6551 /* we uncharge from mc.from later. */
6554 putback_lru_page(page);
6555 put: /* get_mctgt_type() gets the page */
6558 case MC_TARGET_SWAP:
6560 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6562 /* we fixup refcnts and charges later. */
6570 pte_unmap_unlock(pte - 1, ptl);
6575 * We have consumed all precharges we got in can_attach().
6576 * We try charge one by one, but don't do any additional
6577 * charges to mc.to if we have failed in charge once in attach()
6580 ret = mem_cgroup_do_precharge(1);
6588 static void mem_cgroup_move_charge(struct mm_struct *mm)
6590 struct vm_area_struct *vma;
6592 lru_add_drain_all();
6594 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6596 * Someone who are holding the mmap_sem might be waiting in
6597 * waitq. So we cancel all extra charges, wake up all waiters,
6598 * and retry. Because we cancel precharges, we might not be able
6599 * to move enough charges, but moving charge is a best-effort
6600 * feature anyway, so it wouldn't be a big problem.
6602 __mem_cgroup_clear_mc();
6606 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6608 struct mm_walk mem_cgroup_move_charge_walk = {
6609 .pmd_entry = mem_cgroup_move_charge_pte_range,
6613 if (is_vm_hugetlb_page(vma))
6615 ret = walk_page_range(vma->vm_start, vma->vm_end,
6616 &mem_cgroup_move_charge_walk);
6619 * means we have consumed all precharges and failed in
6620 * doing additional charge. Just abandon here.
6624 up_read(&mm->mmap_sem);
6627 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6628 struct cgroup_taskset *tset)
6630 struct task_struct *p = cgroup_taskset_first(tset);
6631 struct mm_struct *mm = get_task_mm(p);
6635 mem_cgroup_move_charge(mm);
6639 mem_cgroup_clear_mc();
6641 #else /* !CONFIG_MMU */
6642 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6643 struct cgroup_taskset *tset)
6647 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6648 struct cgroup_taskset *tset)
6651 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6652 struct cgroup_taskset *tset)
6658 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6659 * to verify sane_behavior flag on each mount attempt.
6661 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6664 * use_hierarchy is forced with sane_behavior. cgroup core
6665 * guarantees that @root doesn't have any children, so turning it
6666 * on for the root memcg is enough.
6668 if (cgroup_sane_behavior(root_css->cgroup))
6669 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6672 struct cgroup_subsys mem_cgroup_subsys = {
6674 .subsys_id = mem_cgroup_subsys_id,
6675 .css_alloc = mem_cgroup_css_alloc,
6676 .css_online = mem_cgroup_css_online,
6677 .css_offline = mem_cgroup_css_offline,
6678 .css_free = mem_cgroup_css_free,
6679 .can_attach = mem_cgroup_can_attach,
6680 .cancel_attach = mem_cgroup_cancel_attach,
6681 .attach = mem_cgroup_move_task,
6682 .bind = mem_cgroup_bind,
6683 .base_cftypes = mem_cgroup_files,
6688 #ifdef CONFIG_MEMCG_SWAP
6689 static int __init enable_swap_account(char *s)
6691 if (!strcmp(s, "1"))
6692 really_do_swap_account = 1;
6693 else if (!strcmp(s, "0"))
6694 really_do_swap_account = 0;
6697 __setup("swapaccount=", enable_swap_account);
6699 static void __init memsw_file_init(void)
6701 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6704 static void __init enable_swap_cgroup(void)
6706 if (!mem_cgroup_disabled() && really_do_swap_account) {
6707 do_swap_account = 1;
6713 static void __init enable_swap_cgroup(void)
6719 * subsys_initcall() for memory controller.
6721 * Some parts like hotcpu_notifier() have to be initialized from this context
6722 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6723 * everything that doesn't depend on a specific mem_cgroup structure should
6724 * be initialized from here.
6726 static int __init mem_cgroup_init(void)
6728 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6729 enable_swap_cgroup();
6733 subsys_initcall(mem_cgroup_init);