1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/slab.h>
43 #include <linux/swap.h>
44 #include <linux/swapops.h>
45 #include <linux/spinlock.h>
46 #include <linux/eventfd.h>
47 #include <linux/sort.h>
49 #include <linux/seq_file.h>
50 #include <linux/vmalloc.h>
51 #include <linux/vmpressure.h>
52 #include <linux/mm_inline.h>
53 #include <linux/page_cgroup.h>
54 #include <linux/cpu.h>
55 #include <linux/oom.h>
59 #include <net/tcp_memcontrol.h>
61 #include <asm/uaccess.h>
63 #include <trace/events/vmscan.h>
65 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
66 EXPORT_SYMBOL(mem_cgroup_subsys);
68 #define MEM_CGROUP_RECLAIM_RETRIES 5
69 static struct mem_cgroup *root_mem_cgroup __read_mostly;
71 #ifdef CONFIG_MEMCG_SWAP
72 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
73 int do_swap_account __read_mostly;
75 /* for remember boot option*/
76 #ifdef CONFIG_MEMCG_SWAP_ENABLED
77 static int really_do_swap_account __initdata = 1;
79 static int really_do_swap_account __initdata = 0;
83 #define do_swap_account 0
87 static const char * const mem_cgroup_stat_names[] = {
95 enum mem_cgroup_events_index {
96 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
97 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
98 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
99 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
100 MEM_CGROUP_EVENTS_NSTATS,
103 static const char * const mem_cgroup_events_names[] = {
110 static const char * const mem_cgroup_lru_names[] = {
119 * Per memcg event counter is incremented at every pagein/pageout. With THP,
120 * it will be incremated by the number of pages. This counter is used for
121 * for trigger some periodic events. This is straightforward and better
122 * than using jiffies etc. to handle periodic memcg event.
124 enum mem_cgroup_events_target {
125 MEM_CGROUP_TARGET_THRESH,
126 MEM_CGROUP_TARGET_SOFTLIMIT,
127 MEM_CGROUP_TARGET_NUMAINFO,
130 #define THRESHOLDS_EVENTS_TARGET 128
131 #define SOFTLIMIT_EVENTS_TARGET 1024
132 #define NUMAINFO_EVENTS_TARGET 1024
134 struct mem_cgroup_stat_cpu {
135 long count[MEM_CGROUP_STAT_NSTATS];
136 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
137 unsigned long nr_page_events;
138 unsigned long targets[MEM_CGROUP_NTARGETS];
141 struct mem_cgroup_reclaim_iter {
143 * last scanned hierarchy member. Valid only if last_dead_count
144 * matches memcg->dead_count of the hierarchy root group.
146 struct mem_cgroup *last_visited;
147 unsigned long last_dead_count;
149 /* scan generation, increased every round-trip */
150 unsigned int generation;
154 * per-zone information in memory controller.
156 struct mem_cgroup_per_zone {
157 struct lruvec lruvec;
158 unsigned long lru_size[NR_LRU_LISTS];
160 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
162 struct mem_cgroup *memcg; /* Back pointer, we cannot */
163 /* use container_of */
166 struct mem_cgroup_per_node {
167 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
170 struct mem_cgroup_threshold {
171 struct eventfd_ctx *eventfd;
176 struct mem_cgroup_threshold_ary {
177 /* An array index points to threshold just below or equal to usage. */
178 int current_threshold;
179 /* Size of entries[] */
181 /* Array of thresholds */
182 struct mem_cgroup_threshold entries[0];
185 struct mem_cgroup_thresholds {
186 /* Primary thresholds array */
187 struct mem_cgroup_threshold_ary *primary;
189 * Spare threshold array.
190 * This is needed to make mem_cgroup_unregister_event() "never fail".
191 * It must be able to store at least primary->size - 1 entries.
193 struct mem_cgroup_threshold_ary *spare;
197 struct mem_cgroup_eventfd_list {
198 struct list_head list;
199 struct eventfd_ctx *eventfd;
202 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
203 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
206 * The memory controller data structure. The memory controller controls both
207 * page cache and RSS per cgroup. We would eventually like to provide
208 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
209 * to help the administrator determine what knobs to tune.
211 * TODO: Add a water mark for the memory controller. Reclaim will begin when
212 * we hit the water mark. May be even add a low water mark, such that
213 * no reclaim occurs from a cgroup at it's low water mark, this is
214 * a feature that will be implemented much later in the future.
217 struct cgroup_subsys_state css;
219 * the counter to account for memory usage
221 struct res_counter res;
223 /* vmpressure notifications */
224 struct vmpressure vmpressure;
227 * the counter to account for mem+swap usage.
229 struct res_counter memsw;
232 * the counter to account for kernel memory usage.
234 struct res_counter kmem;
236 * Should the accounting and control be hierarchical, per subtree?
239 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
243 atomic_t oom_wakeups;
246 /* OOM-Killer disable */
247 int oom_kill_disable;
249 /* set when res.limit == memsw.limit */
250 bool memsw_is_minimum;
252 /* protect arrays of thresholds */
253 struct mutex thresholds_lock;
255 /* thresholds for memory usage. RCU-protected */
256 struct mem_cgroup_thresholds thresholds;
258 /* thresholds for mem+swap usage. RCU-protected */
259 struct mem_cgroup_thresholds memsw_thresholds;
261 /* For oom notifier event fd */
262 struct list_head oom_notify;
265 * Should we move charges of a task when a task is moved into this
266 * mem_cgroup ? And what type of charges should we move ?
268 unsigned long move_charge_at_immigrate;
270 * set > 0 if pages under this cgroup are moving to other cgroup.
272 atomic_t moving_account;
273 /* taken only while moving_account > 0 */
274 spinlock_t move_lock;
278 struct mem_cgroup_stat_cpu __percpu *stat;
280 * used when a cpu is offlined or other synchronizations
281 * See mem_cgroup_read_stat().
283 struct mem_cgroup_stat_cpu nocpu_base;
284 spinlock_t pcp_counter_lock;
287 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
288 struct tcp_memcontrol tcp_mem;
290 #if defined(CONFIG_MEMCG_KMEM)
291 /* analogous to slab_common's slab_caches list. per-memcg */
292 struct list_head memcg_slab_caches;
293 /* Not a spinlock, we can take a lot of time walking the list */
294 struct mutex slab_caches_mutex;
295 /* Index in the kmem_cache->memcg_params->memcg_caches array */
299 int last_scanned_node;
301 nodemask_t scan_nodes;
302 atomic_t numainfo_events;
303 atomic_t numainfo_updating;
306 * Protects soft_contributed transitions.
307 * See mem_cgroup_update_soft_limit
309 spinlock_t soft_lock;
312 * If true then this group has increased parents' children_in_excess
313 * when it got over the soft limit.
314 * When a group falls bellow the soft limit, parents' children_in_excess
315 * is decreased and soft_contributed changed to false.
317 bool soft_contributed;
319 /* Number of children that are in soft limit excess */
320 atomic_t children_in_excess;
322 struct mem_cgroup_per_node *nodeinfo[0];
323 /* WARNING: nodeinfo must be the last member here */
326 static size_t memcg_size(void)
328 return sizeof(struct mem_cgroup) +
329 nr_node_ids * sizeof(struct mem_cgroup_per_node);
332 /* internal only representation about the status of kmem accounting. */
334 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
335 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
336 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
339 /* We account when limit is on, but only after call sites are patched */
340 #define KMEM_ACCOUNTED_MASK \
341 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
343 #ifdef CONFIG_MEMCG_KMEM
344 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
346 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
349 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
351 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
354 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
356 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
359 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
361 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
364 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
367 * Our caller must use css_get() first, because memcg_uncharge_kmem()
368 * will call css_put() if it sees the memcg is dead.
371 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
372 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
375 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
377 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
378 &memcg->kmem_account_flags);
382 /* Stuffs for move charges at task migration. */
384 * Types of charges to be moved. "move_charge_at_immitgrate" and
385 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
388 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
389 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
393 /* "mc" and its members are protected by cgroup_mutex */
394 static struct move_charge_struct {
395 spinlock_t lock; /* for from, to */
396 struct mem_cgroup *from;
397 struct mem_cgroup *to;
398 unsigned long immigrate_flags;
399 unsigned long precharge;
400 unsigned long moved_charge;
401 unsigned long moved_swap;
402 struct task_struct *moving_task; /* a task moving charges */
403 wait_queue_head_t waitq; /* a waitq for other context */
405 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
406 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
409 static bool move_anon(void)
411 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
414 static bool move_file(void)
416 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
420 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
421 * limit reclaim to prevent infinite loops, if they ever occur.
423 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
426 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
427 MEM_CGROUP_CHARGE_TYPE_ANON,
428 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
429 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
433 /* for encoding cft->private value on file */
441 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
442 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
443 #define MEMFILE_ATTR(val) ((val) & 0xffff)
444 /* Used for OOM nofiier */
445 #define OOM_CONTROL (0)
448 * Reclaim flags for mem_cgroup_hierarchical_reclaim
450 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
451 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
452 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
453 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
456 * The memcg_create_mutex will be held whenever a new cgroup is created.
457 * As a consequence, any change that needs to protect against new child cgroups
458 * appearing has to hold it as well.
460 static DEFINE_MUTEX(memcg_create_mutex);
462 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
464 return s ? container_of(s, struct mem_cgroup, css) : NULL;
467 /* Some nice accessors for the vmpressure. */
468 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
471 memcg = root_mem_cgroup;
472 return &memcg->vmpressure;
475 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
477 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
480 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
482 return &mem_cgroup_from_css(css)->vmpressure;
485 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
487 return (memcg == root_mem_cgroup);
490 /* Writing them here to avoid exposing memcg's inner layout */
491 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
493 void sock_update_memcg(struct sock *sk)
495 if (mem_cgroup_sockets_enabled) {
496 struct mem_cgroup *memcg;
497 struct cg_proto *cg_proto;
499 BUG_ON(!sk->sk_prot->proto_cgroup);
501 /* Socket cloning can throw us here with sk_cgrp already
502 * filled. It won't however, necessarily happen from
503 * process context. So the test for root memcg given
504 * the current task's memcg won't help us in this case.
506 * Respecting the original socket's memcg is a better
507 * decision in this case.
510 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
511 css_get(&sk->sk_cgrp->memcg->css);
516 memcg = mem_cgroup_from_task(current);
517 cg_proto = sk->sk_prot->proto_cgroup(memcg);
518 if (!mem_cgroup_is_root(memcg) &&
519 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
520 sk->sk_cgrp = cg_proto;
525 EXPORT_SYMBOL(sock_update_memcg);
527 void sock_release_memcg(struct sock *sk)
529 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
530 struct mem_cgroup *memcg;
531 WARN_ON(!sk->sk_cgrp->memcg);
532 memcg = sk->sk_cgrp->memcg;
533 css_put(&sk->sk_cgrp->memcg->css);
537 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
539 if (!memcg || mem_cgroup_is_root(memcg))
542 return &memcg->tcp_mem.cg_proto;
544 EXPORT_SYMBOL(tcp_proto_cgroup);
546 static void disarm_sock_keys(struct mem_cgroup *memcg)
548 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
550 static_key_slow_dec(&memcg_socket_limit_enabled);
553 static void disarm_sock_keys(struct mem_cgroup *memcg)
558 #ifdef CONFIG_MEMCG_KMEM
560 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
561 * There are two main reasons for not using the css_id for this:
562 * 1) this works better in sparse environments, where we have a lot of memcgs,
563 * but only a few kmem-limited. Or also, if we have, for instance, 200
564 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
565 * 200 entry array for that.
567 * 2) In order not to violate the cgroup API, we would like to do all memory
568 * allocation in ->create(). At that point, we haven't yet allocated the
569 * css_id. Having a separate index prevents us from messing with the cgroup
572 * The current size of the caches array is stored in
573 * memcg_limited_groups_array_size. It will double each time we have to
576 static DEFINE_IDA(kmem_limited_groups);
577 int memcg_limited_groups_array_size;
580 * MIN_SIZE is different than 1, because we would like to avoid going through
581 * the alloc/free process all the time. In a small machine, 4 kmem-limited
582 * cgroups is a reasonable guess. In the future, it could be a parameter or
583 * tunable, but that is strictly not necessary.
585 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
586 * this constant directly from cgroup, but it is understandable that this is
587 * better kept as an internal representation in cgroup.c. In any case, the
588 * css_id space is not getting any smaller, and we don't have to necessarily
589 * increase ours as well if it increases.
591 #define MEMCG_CACHES_MIN_SIZE 4
592 #define MEMCG_CACHES_MAX_SIZE 65535
595 * A lot of the calls to the cache allocation functions are expected to be
596 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
597 * conditional to this static branch, we'll have to allow modules that does
598 * kmem_cache_alloc and the such to see this symbol as well
600 struct static_key memcg_kmem_enabled_key;
601 EXPORT_SYMBOL(memcg_kmem_enabled_key);
603 static void disarm_kmem_keys(struct mem_cgroup *memcg)
605 if (memcg_kmem_is_active(memcg)) {
606 static_key_slow_dec(&memcg_kmem_enabled_key);
607 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
610 * This check can't live in kmem destruction function,
611 * since the charges will outlive the cgroup
613 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
616 static void disarm_kmem_keys(struct mem_cgroup *memcg)
619 #endif /* CONFIG_MEMCG_KMEM */
621 static void disarm_static_keys(struct mem_cgroup *memcg)
623 disarm_sock_keys(memcg);
624 disarm_kmem_keys(memcg);
627 static void drain_all_stock_async(struct mem_cgroup *memcg);
629 static struct mem_cgroup_per_zone *
630 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
632 VM_BUG_ON((unsigned)nid >= nr_node_ids);
633 return &memcg->nodeinfo[nid]->zoneinfo[zid];
636 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
641 static struct mem_cgroup_per_zone *
642 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
644 int nid = page_to_nid(page);
645 int zid = page_zonenum(page);
647 return mem_cgroup_zoneinfo(memcg, nid, zid);
651 * Implementation Note: reading percpu statistics for memcg.
653 * Both of vmstat[] and percpu_counter has threshold and do periodic
654 * synchronization to implement "quick" read. There are trade-off between
655 * reading cost and precision of value. Then, we may have a chance to implement
656 * a periodic synchronizion of counter in memcg's counter.
658 * But this _read() function is used for user interface now. The user accounts
659 * memory usage by memory cgroup and he _always_ requires exact value because
660 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
661 * have to visit all online cpus and make sum. So, for now, unnecessary
662 * synchronization is not implemented. (just implemented for cpu hotplug)
664 * If there are kernel internal actions which can make use of some not-exact
665 * value, and reading all cpu value can be performance bottleneck in some
666 * common workload, threashold and synchonization as vmstat[] should be
669 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
670 enum mem_cgroup_stat_index idx)
676 for_each_online_cpu(cpu)
677 val += per_cpu(memcg->stat->count[idx], cpu);
678 #ifdef CONFIG_HOTPLUG_CPU
679 spin_lock(&memcg->pcp_counter_lock);
680 val += memcg->nocpu_base.count[idx];
681 spin_unlock(&memcg->pcp_counter_lock);
687 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
690 int val = (charge) ? 1 : -1;
691 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
694 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
695 enum mem_cgroup_events_index idx)
697 unsigned long val = 0;
700 for_each_online_cpu(cpu)
701 val += per_cpu(memcg->stat->events[idx], cpu);
702 #ifdef CONFIG_HOTPLUG_CPU
703 spin_lock(&memcg->pcp_counter_lock);
704 val += memcg->nocpu_base.events[idx];
705 spin_unlock(&memcg->pcp_counter_lock);
710 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
712 bool anon, int nr_pages)
717 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
718 * counted as CACHE even if it's on ANON LRU.
721 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
724 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
727 if (PageTransHuge(page))
728 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
731 /* pagein of a big page is an event. So, ignore page size */
733 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
735 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
736 nr_pages = -nr_pages; /* for event */
739 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
745 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
747 struct mem_cgroup_per_zone *mz;
749 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
750 return mz->lru_size[lru];
754 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
755 unsigned int lru_mask)
757 struct mem_cgroup_per_zone *mz;
759 unsigned long ret = 0;
761 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
764 if (BIT(lru) & lru_mask)
765 ret += mz->lru_size[lru];
771 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
772 int nid, unsigned int lru_mask)
777 for (zid = 0; zid < MAX_NR_ZONES; zid++)
778 total += mem_cgroup_zone_nr_lru_pages(memcg,
784 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
785 unsigned int lru_mask)
790 for_each_node_state(nid, N_MEMORY)
791 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
795 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
796 enum mem_cgroup_events_target target)
798 unsigned long val, next;
800 val = __this_cpu_read(memcg->stat->nr_page_events);
801 next = __this_cpu_read(memcg->stat->targets[target]);
802 /* from time_after() in jiffies.h */
803 if ((long)next - (long)val < 0) {
805 case MEM_CGROUP_TARGET_THRESH:
806 next = val + THRESHOLDS_EVENTS_TARGET;
808 case MEM_CGROUP_TARGET_SOFTLIMIT:
809 next = val + SOFTLIMIT_EVENTS_TARGET;
811 case MEM_CGROUP_TARGET_NUMAINFO:
812 next = val + NUMAINFO_EVENTS_TARGET;
817 __this_cpu_write(memcg->stat->targets[target], next);
824 * Called from rate-limited memcg_check_events when enough
825 * MEM_CGROUP_TARGET_SOFTLIMIT events are accumulated and it makes sure
826 * that all the parents up the hierarchy will be notified that this group
827 * is in excess or that it is not in excess anymore. mmecg->soft_contributed
828 * makes the transition a single action whenever the state flips from one to
831 static void mem_cgroup_update_soft_limit(struct mem_cgroup *memcg)
833 unsigned long long excess = res_counter_soft_limit_excess(&memcg->res);
834 struct mem_cgroup *parent = memcg;
837 spin_lock(&memcg->soft_lock);
839 if (!memcg->soft_contributed) {
841 memcg->soft_contributed = true;
844 if (memcg->soft_contributed) {
846 memcg->soft_contributed = false;
851 * Necessary to update all ancestors when hierarchy is used
852 * because their event counter is not touched.
853 * We track children even outside the hierarchy for the root
854 * cgroup because tree walk starting at root should visit
855 * all cgroups and we want to prevent from pointless tree
856 * walk if no children is below the limit.
858 while (delta && (parent = parent_mem_cgroup(parent)))
859 atomic_add(delta, &parent->children_in_excess);
860 if (memcg != root_mem_cgroup && !root_mem_cgroup->use_hierarchy)
861 atomic_add(delta, &root_mem_cgroup->children_in_excess);
862 spin_unlock(&memcg->soft_lock);
866 * Check events in order.
869 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
872 /* threshold event is triggered in finer grain than soft limit */
873 if (unlikely(mem_cgroup_event_ratelimit(memcg,
874 MEM_CGROUP_TARGET_THRESH))) {
876 bool do_numainfo __maybe_unused;
878 do_softlimit = mem_cgroup_event_ratelimit(memcg,
879 MEM_CGROUP_TARGET_SOFTLIMIT);
881 do_numainfo = mem_cgroup_event_ratelimit(memcg,
882 MEM_CGROUP_TARGET_NUMAINFO);
886 mem_cgroup_threshold(memcg);
887 if (unlikely(do_softlimit))
888 mem_cgroup_update_soft_limit(memcg);
890 if (unlikely(do_numainfo))
891 atomic_inc(&memcg->numainfo_events);
897 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
900 * mm_update_next_owner() may clear mm->owner to NULL
901 * if it races with swapoff, page migration, etc.
902 * So this can be called with p == NULL.
907 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
910 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
912 struct mem_cgroup *memcg = NULL;
917 * Because we have no locks, mm->owner's may be being moved to other
918 * cgroup. We use css_tryget() here even if this looks
919 * pessimistic (rather than adding locks here).
923 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
924 if (unlikely(!memcg))
926 } while (!css_tryget(&memcg->css));
931 static enum mem_cgroup_filter_t
932 mem_cgroup_filter(struct mem_cgroup *memcg, struct mem_cgroup *root,
933 mem_cgroup_iter_filter cond)
937 return cond(memcg, root);
941 * Returns a next (in a pre-order walk) alive memcg (with elevated css
942 * ref. count) or NULL if the whole root's subtree has been visited.
944 * helper function to be used by mem_cgroup_iter
946 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
947 struct mem_cgroup *last_visited, mem_cgroup_iter_filter cond)
949 struct cgroup_subsys_state *prev_css, *next_css;
951 prev_css = last_visited ? &last_visited->css : NULL;
953 next_css = css_next_descendant_pre(prev_css, &root->css);
956 * Even if we found a group we have to make sure it is
957 * alive. css && !memcg means that the groups should be
958 * skipped and we should continue the tree walk.
959 * last_visited css is safe to use because it is
960 * protected by css_get and the tree walk is rcu safe.
963 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
965 switch (mem_cgroup_filter(mem, root, cond)) {
973 * css_rightmost_descendant is not an optimal way to
974 * skip through a subtree (especially for imbalanced
975 * trees leaning to right) but that's what we have right
976 * now. More effective solution would be traversing
977 * right-up for first non-NULL without calling
978 * css_next_descendant_pre afterwards.
980 prev_css = css_rightmost_descendant(next_css);
983 if (css_tryget(&mem->css))
996 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
999 * When a group in the hierarchy below root is destroyed, the
1000 * hierarchy iterator can no longer be trusted since it might
1001 * have pointed to the destroyed group. Invalidate it.
1003 atomic_inc(&root->dead_count);
1006 static struct mem_cgroup *
1007 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1008 struct mem_cgroup *root,
1011 struct mem_cgroup *position = NULL;
1013 * A cgroup destruction happens in two stages: offlining and
1014 * release. They are separated by a RCU grace period.
1016 * If the iterator is valid, we may still race with an
1017 * offlining. The RCU lock ensures the object won't be
1018 * released, tryget will fail if we lost the race.
1020 *sequence = atomic_read(&root->dead_count);
1021 if (iter->last_dead_count == *sequence) {
1023 position = iter->last_visited;
1024 if (position && !css_tryget(&position->css))
1030 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1031 struct mem_cgroup *last_visited,
1032 struct mem_cgroup *new_position,
1036 css_put(&last_visited->css);
1038 * We store the sequence count from the time @last_visited was
1039 * loaded successfully instead of rereading it here so that we
1040 * don't lose destruction events in between. We could have
1041 * raced with the destruction of @new_position after all.
1043 iter->last_visited = new_position;
1045 iter->last_dead_count = sequence;
1049 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1050 * @root: hierarchy root
1051 * @prev: previously returned memcg, NULL on first invocation
1052 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1053 * @cond: filter for visited nodes, NULL for no filter
1055 * Returns references to children of the hierarchy below @root, or
1056 * @root itself, or %NULL after a full round-trip.
1058 * Caller must pass the return value in @prev on subsequent
1059 * invocations for reference counting, or use mem_cgroup_iter_break()
1060 * to cancel a hierarchy walk before the round-trip is complete.
1062 * Reclaimers can specify a zone and a priority level in @reclaim to
1063 * divide up the memcgs in the hierarchy among all concurrent
1064 * reclaimers operating on the same zone and priority.
1066 struct mem_cgroup *mem_cgroup_iter_cond(struct mem_cgroup *root,
1067 struct mem_cgroup *prev,
1068 struct mem_cgroup_reclaim_cookie *reclaim,
1069 mem_cgroup_iter_filter cond)
1071 struct mem_cgroup *memcg = NULL;
1072 struct mem_cgroup *last_visited = NULL;
1074 if (mem_cgroup_disabled()) {
1075 /* first call must return non-NULL, second return NULL */
1076 return (struct mem_cgroup *)(unsigned long)!prev;
1080 root = root_mem_cgroup;
1082 if (prev && !reclaim)
1083 last_visited = prev;
1085 if (!root->use_hierarchy && root != root_mem_cgroup) {
1088 if (mem_cgroup_filter(root, root, cond) == VISIT)
1095 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1096 int uninitialized_var(seq);
1099 int nid = zone_to_nid(reclaim->zone);
1100 int zid = zone_idx(reclaim->zone);
1101 struct mem_cgroup_per_zone *mz;
1103 mz = mem_cgroup_zoneinfo(root, nid, zid);
1104 iter = &mz->reclaim_iter[reclaim->priority];
1105 if (prev && reclaim->generation != iter->generation) {
1106 iter->last_visited = NULL;
1110 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1113 memcg = __mem_cgroup_iter_next(root, last_visited, cond);
1116 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1120 else if (!prev && memcg)
1121 reclaim->generation = iter->generation;
1125 * We have finished the whole tree walk or no group has been
1126 * visited because filter told us to skip the root node.
1128 if (!memcg && (prev || (cond && !last_visited)))
1134 if (prev && prev != root)
1135 css_put(&prev->css);
1141 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1142 * @root: hierarchy root
1143 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1145 void mem_cgroup_iter_break(struct mem_cgroup *root,
1146 struct mem_cgroup *prev)
1149 root = root_mem_cgroup;
1150 if (prev && prev != root)
1151 css_put(&prev->css);
1155 * Iteration constructs for visiting all cgroups (under a tree). If
1156 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1157 * be used for reference counting.
1159 #define for_each_mem_cgroup_tree(iter, root) \
1160 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1162 iter = mem_cgroup_iter(root, iter, NULL))
1164 #define for_each_mem_cgroup(iter) \
1165 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1167 iter = mem_cgroup_iter(NULL, iter, NULL))
1169 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1171 struct mem_cgroup *memcg;
1174 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1175 if (unlikely(!memcg))
1180 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1183 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1191 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1194 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1195 * @zone: zone of the wanted lruvec
1196 * @memcg: memcg of the wanted lruvec
1198 * Returns the lru list vector holding pages for the given @zone and
1199 * @mem. This can be the global zone lruvec, if the memory controller
1202 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1203 struct mem_cgroup *memcg)
1205 struct mem_cgroup_per_zone *mz;
1206 struct lruvec *lruvec;
1208 if (mem_cgroup_disabled()) {
1209 lruvec = &zone->lruvec;
1213 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1214 lruvec = &mz->lruvec;
1217 * Since a node can be onlined after the mem_cgroup was created,
1218 * we have to be prepared to initialize lruvec->zone here;
1219 * and if offlined then reonlined, we need to reinitialize it.
1221 if (unlikely(lruvec->zone != zone))
1222 lruvec->zone = zone;
1227 * Following LRU functions are allowed to be used without PCG_LOCK.
1228 * Operations are called by routine of global LRU independently from memcg.
1229 * What we have to take care of here is validness of pc->mem_cgroup.
1231 * Changes to pc->mem_cgroup happens when
1234 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1235 * It is added to LRU before charge.
1236 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1237 * When moving account, the page is not on LRU. It's isolated.
1241 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1243 * @zone: zone of the page
1245 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1247 struct mem_cgroup_per_zone *mz;
1248 struct mem_cgroup *memcg;
1249 struct page_cgroup *pc;
1250 struct lruvec *lruvec;
1252 if (mem_cgroup_disabled()) {
1253 lruvec = &zone->lruvec;
1257 pc = lookup_page_cgroup(page);
1258 memcg = pc->mem_cgroup;
1261 * Surreptitiously switch any uncharged offlist page to root:
1262 * an uncharged page off lru does nothing to secure
1263 * its former mem_cgroup from sudden removal.
1265 * Our caller holds lru_lock, and PageCgroupUsed is updated
1266 * under page_cgroup lock: between them, they make all uses
1267 * of pc->mem_cgroup safe.
1269 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1270 pc->mem_cgroup = memcg = root_mem_cgroup;
1272 mz = page_cgroup_zoneinfo(memcg, page);
1273 lruvec = &mz->lruvec;
1276 * Since a node can be onlined after the mem_cgroup was created,
1277 * we have to be prepared to initialize lruvec->zone here;
1278 * and if offlined then reonlined, we need to reinitialize it.
1280 if (unlikely(lruvec->zone != zone))
1281 lruvec->zone = zone;
1286 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1287 * @lruvec: mem_cgroup per zone lru vector
1288 * @lru: index of lru list the page is sitting on
1289 * @nr_pages: positive when adding or negative when removing
1291 * This function must be called when a page is added to or removed from an
1294 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1297 struct mem_cgroup_per_zone *mz;
1298 unsigned long *lru_size;
1300 if (mem_cgroup_disabled())
1303 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1304 lru_size = mz->lru_size + lru;
1305 *lru_size += nr_pages;
1306 VM_BUG_ON((long)(*lru_size) < 0);
1310 * Checks whether given mem is same or in the root_mem_cgroup's
1313 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1314 struct mem_cgroup *memcg)
1316 if (root_memcg == memcg)
1318 if (!root_memcg->use_hierarchy || !memcg)
1320 return css_is_ancestor(&memcg->css, &root_memcg->css);
1323 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1324 struct mem_cgroup *memcg)
1329 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1334 bool task_in_mem_cgroup(struct task_struct *task,
1335 const struct mem_cgroup *memcg)
1337 struct mem_cgroup *curr = NULL;
1338 struct task_struct *p;
1341 p = find_lock_task_mm(task);
1343 curr = try_get_mem_cgroup_from_mm(p->mm);
1347 * All threads may have already detached their mm's, but the oom
1348 * killer still needs to detect if they have already been oom
1349 * killed to prevent needlessly killing additional tasks.
1352 curr = mem_cgroup_from_task(task);
1354 css_get(&curr->css);
1360 * We should check use_hierarchy of "memcg" not "curr". Because checking
1361 * use_hierarchy of "curr" here make this function true if hierarchy is
1362 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1363 * hierarchy(even if use_hierarchy is disabled in "memcg").
1365 ret = mem_cgroup_same_or_subtree(memcg, curr);
1366 css_put(&curr->css);
1370 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1372 unsigned long inactive_ratio;
1373 unsigned long inactive;
1374 unsigned long active;
1377 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1378 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1380 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1382 inactive_ratio = int_sqrt(10 * gb);
1386 return inactive * inactive_ratio < active;
1389 #define mem_cgroup_from_res_counter(counter, member) \
1390 container_of(counter, struct mem_cgroup, member)
1393 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1394 * @memcg: the memory cgroup
1396 * Returns the maximum amount of memory @mem can be charged with, in
1399 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1401 unsigned long long margin;
1403 margin = res_counter_margin(&memcg->res);
1404 if (do_swap_account)
1405 margin = min(margin, res_counter_margin(&memcg->memsw));
1406 return margin >> PAGE_SHIFT;
1409 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1412 if (!css_parent(&memcg->css))
1413 return vm_swappiness;
1415 return memcg->swappiness;
1419 * memcg->moving_account is used for checking possibility that some thread is
1420 * calling move_account(). When a thread on CPU-A starts moving pages under
1421 * a memcg, other threads should check memcg->moving_account under
1422 * rcu_read_lock(), like this:
1426 * memcg->moving_account+1 if (memcg->mocing_account)
1428 * synchronize_rcu() update something.
1433 /* for quick checking without looking up memcg */
1434 atomic_t memcg_moving __read_mostly;
1436 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1438 atomic_inc(&memcg_moving);
1439 atomic_inc(&memcg->moving_account);
1443 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1446 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1447 * We check NULL in callee rather than caller.
1450 atomic_dec(&memcg_moving);
1451 atomic_dec(&memcg->moving_account);
1456 * 2 routines for checking "mem" is under move_account() or not.
1458 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1459 * is used for avoiding races in accounting. If true,
1460 * pc->mem_cgroup may be overwritten.
1462 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1463 * under hierarchy of moving cgroups. This is for
1464 * waiting at hith-memory prressure caused by "move".
1467 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1469 VM_BUG_ON(!rcu_read_lock_held());
1470 return atomic_read(&memcg->moving_account) > 0;
1473 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1475 struct mem_cgroup *from;
1476 struct mem_cgroup *to;
1479 * Unlike task_move routines, we access mc.to, mc.from not under
1480 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1482 spin_lock(&mc.lock);
1488 ret = mem_cgroup_same_or_subtree(memcg, from)
1489 || mem_cgroup_same_or_subtree(memcg, to);
1491 spin_unlock(&mc.lock);
1495 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1497 if (mc.moving_task && current != mc.moving_task) {
1498 if (mem_cgroup_under_move(memcg)) {
1500 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1501 /* moving charge context might have finished. */
1504 finish_wait(&mc.waitq, &wait);
1512 * Take this lock when
1513 * - a code tries to modify page's memcg while it's USED.
1514 * - a code tries to modify page state accounting in a memcg.
1515 * see mem_cgroup_stolen(), too.
1517 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1518 unsigned long *flags)
1520 spin_lock_irqsave(&memcg->move_lock, *flags);
1523 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1524 unsigned long *flags)
1526 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1529 #define K(x) ((x) << (PAGE_SHIFT-10))
1531 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1532 * @memcg: The memory cgroup that went over limit
1533 * @p: Task that is going to be killed
1535 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1538 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1540 struct cgroup *task_cgrp;
1541 struct cgroup *mem_cgrp;
1543 * Need a buffer in BSS, can't rely on allocations. The code relies
1544 * on the assumption that OOM is serialized for memory controller.
1545 * If this assumption is broken, revisit this code.
1547 static char memcg_name[PATH_MAX];
1549 struct mem_cgroup *iter;
1557 mem_cgrp = memcg->css.cgroup;
1558 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1560 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1563 * Unfortunately, we are unable to convert to a useful name
1564 * But we'll still print out the usage information
1571 pr_info("Task in %s killed", memcg_name);
1574 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1582 * Continues from above, so we don't need an KERN_ level
1584 pr_cont(" as a result of limit of %s\n", memcg_name);
1587 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1588 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1589 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1590 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1591 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1592 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1593 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1594 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1595 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1596 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1597 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1598 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1600 for_each_mem_cgroup_tree(iter, memcg) {
1601 pr_info("Memory cgroup stats");
1604 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1606 pr_cont(" for %s", memcg_name);
1610 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1611 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1613 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1614 K(mem_cgroup_read_stat(iter, i)));
1617 for (i = 0; i < NR_LRU_LISTS; i++)
1618 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1619 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1626 * This function returns the number of memcg under hierarchy tree. Returns
1627 * 1(self count) if no children.
1629 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1632 struct mem_cgroup *iter;
1634 for_each_mem_cgroup_tree(iter, memcg)
1640 * Return the memory (and swap, if configured) limit for a memcg.
1642 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1646 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1649 * Do not consider swap space if we cannot swap due to swappiness
1651 if (mem_cgroup_swappiness(memcg)) {
1654 limit += total_swap_pages << PAGE_SHIFT;
1655 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1658 * If memsw is finite and limits the amount of swap space
1659 * available to this memcg, return that limit.
1661 limit = min(limit, memsw);
1667 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1670 struct mem_cgroup *iter;
1671 unsigned long chosen_points = 0;
1672 unsigned long totalpages;
1673 unsigned int points = 0;
1674 struct task_struct *chosen = NULL;
1677 * If current has a pending SIGKILL or is exiting, then automatically
1678 * select it. The goal is to allow it to allocate so that it may
1679 * quickly exit and free its memory.
1681 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1682 set_thread_flag(TIF_MEMDIE);
1686 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1687 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1688 for_each_mem_cgroup_tree(iter, memcg) {
1689 struct css_task_iter it;
1690 struct task_struct *task;
1692 css_task_iter_start(&iter->css, &it);
1693 while ((task = css_task_iter_next(&it))) {
1694 switch (oom_scan_process_thread(task, totalpages, NULL,
1696 case OOM_SCAN_SELECT:
1698 put_task_struct(chosen);
1700 chosen_points = ULONG_MAX;
1701 get_task_struct(chosen);
1703 case OOM_SCAN_CONTINUE:
1705 case OOM_SCAN_ABORT:
1706 css_task_iter_end(&it);
1707 mem_cgroup_iter_break(memcg, iter);
1709 put_task_struct(chosen);
1714 points = oom_badness(task, memcg, NULL, totalpages);
1715 if (points > chosen_points) {
1717 put_task_struct(chosen);
1719 chosen_points = points;
1720 get_task_struct(chosen);
1723 css_task_iter_end(&it);
1728 points = chosen_points * 1000 / totalpages;
1729 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1730 NULL, "Memory cgroup out of memory");
1733 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1735 unsigned long flags)
1737 unsigned long total = 0;
1738 bool noswap = false;
1741 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1743 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1746 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1748 drain_all_stock_async(memcg);
1749 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1751 * Allow limit shrinkers, which are triggered directly
1752 * by userspace, to catch signals and stop reclaim
1753 * after minimal progress, regardless of the margin.
1755 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1757 if (mem_cgroup_margin(memcg))
1760 * If nothing was reclaimed after two attempts, there
1761 * may be no reclaimable pages in this hierarchy.
1769 #if MAX_NUMNODES > 1
1771 * test_mem_cgroup_node_reclaimable
1772 * @memcg: the target memcg
1773 * @nid: the node ID to be checked.
1774 * @noswap : specify true here if the user wants flle only information.
1776 * This function returns whether the specified memcg contains any
1777 * reclaimable pages on a node. Returns true if there are any reclaimable
1778 * pages in the node.
1780 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1781 int nid, bool noswap)
1783 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1785 if (noswap || !total_swap_pages)
1787 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1794 * Always updating the nodemask is not very good - even if we have an empty
1795 * list or the wrong list here, we can start from some node and traverse all
1796 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1799 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1803 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1804 * pagein/pageout changes since the last update.
1806 if (!atomic_read(&memcg->numainfo_events))
1808 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1811 /* make a nodemask where this memcg uses memory from */
1812 memcg->scan_nodes = node_states[N_MEMORY];
1814 for_each_node_mask(nid, node_states[N_MEMORY]) {
1816 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1817 node_clear(nid, memcg->scan_nodes);
1820 atomic_set(&memcg->numainfo_events, 0);
1821 atomic_set(&memcg->numainfo_updating, 0);
1825 * Selecting a node where we start reclaim from. Because what we need is just
1826 * reducing usage counter, start from anywhere is O,K. Considering
1827 * memory reclaim from current node, there are pros. and cons.
1829 * Freeing memory from current node means freeing memory from a node which
1830 * we'll use or we've used. So, it may make LRU bad. And if several threads
1831 * hit limits, it will see a contention on a node. But freeing from remote
1832 * node means more costs for memory reclaim because of memory latency.
1834 * Now, we use round-robin. Better algorithm is welcomed.
1836 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1840 mem_cgroup_may_update_nodemask(memcg);
1841 node = memcg->last_scanned_node;
1843 node = next_node(node, memcg->scan_nodes);
1844 if (node == MAX_NUMNODES)
1845 node = first_node(memcg->scan_nodes);
1847 * We call this when we hit limit, not when pages are added to LRU.
1848 * No LRU may hold pages because all pages are UNEVICTABLE or
1849 * memcg is too small and all pages are not on LRU. In that case,
1850 * we use curret node.
1852 if (unlikely(node == MAX_NUMNODES))
1853 node = numa_node_id();
1855 memcg->last_scanned_node = node;
1860 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1868 * A group is eligible for the soft limit reclaim under the given root
1870 * a) it is over its soft limit
1871 * b) any parent up the hierarchy is over its soft limit
1873 * If the given group doesn't have any children over the limit then it
1874 * doesn't make any sense to iterate its subtree.
1876 enum mem_cgroup_filter_t
1877 mem_cgroup_soft_reclaim_eligible(struct mem_cgroup *memcg,
1878 struct mem_cgroup *root)
1880 struct mem_cgroup *parent;
1883 memcg = root_mem_cgroup;
1886 if (res_counter_soft_limit_excess(&memcg->res))
1890 * If any parent up to the root in the hierarchy is over its soft limit
1891 * then we have to obey and reclaim from this group as well.
1893 while ((parent = parent_mem_cgroup(parent))) {
1894 if (res_counter_soft_limit_excess(&parent->res))
1900 if (!atomic_read(&memcg->children_in_excess))
1905 static DEFINE_SPINLOCK(memcg_oom_lock);
1908 * Check OOM-Killer is already running under our hierarchy.
1909 * If someone is running, return false.
1911 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
1913 struct mem_cgroup *iter, *failed = NULL;
1915 spin_lock(&memcg_oom_lock);
1917 for_each_mem_cgroup_tree(iter, memcg) {
1918 if (iter->oom_lock) {
1920 * this subtree of our hierarchy is already locked
1921 * so we cannot give a lock.
1924 mem_cgroup_iter_break(memcg, iter);
1927 iter->oom_lock = true;
1932 * OK, we failed to lock the whole subtree so we have
1933 * to clean up what we set up to the failing subtree
1935 for_each_mem_cgroup_tree(iter, memcg) {
1936 if (iter == failed) {
1937 mem_cgroup_iter_break(memcg, iter);
1940 iter->oom_lock = false;
1944 spin_unlock(&memcg_oom_lock);
1949 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1951 struct mem_cgroup *iter;
1953 spin_lock(&memcg_oom_lock);
1954 for_each_mem_cgroup_tree(iter, memcg)
1955 iter->oom_lock = false;
1956 spin_unlock(&memcg_oom_lock);
1959 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1961 struct mem_cgroup *iter;
1963 for_each_mem_cgroup_tree(iter, memcg)
1964 atomic_inc(&iter->under_oom);
1967 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1969 struct mem_cgroup *iter;
1972 * When a new child is created while the hierarchy is under oom,
1973 * mem_cgroup_oom_lock() may not be called. We have to use
1974 * atomic_add_unless() here.
1976 for_each_mem_cgroup_tree(iter, memcg)
1977 atomic_add_unless(&iter->under_oom, -1, 0);
1980 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1982 struct oom_wait_info {
1983 struct mem_cgroup *memcg;
1987 static int memcg_oom_wake_function(wait_queue_t *wait,
1988 unsigned mode, int sync, void *arg)
1990 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1991 struct mem_cgroup *oom_wait_memcg;
1992 struct oom_wait_info *oom_wait_info;
1994 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1995 oom_wait_memcg = oom_wait_info->memcg;
1998 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1999 * Then we can use css_is_ancestor without taking care of RCU.
2001 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2002 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2004 return autoremove_wake_function(wait, mode, sync, arg);
2007 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2009 atomic_inc(&memcg->oom_wakeups);
2010 /* for filtering, pass "memcg" as argument. */
2011 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2014 static void memcg_oom_recover(struct mem_cgroup *memcg)
2016 if (memcg && atomic_read(&memcg->under_oom))
2017 memcg_wakeup_oom(memcg);
2021 * try to call OOM killer
2023 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2028 if (!current->memcg_oom.may_oom)
2031 current->memcg_oom.in_memcg_oom = 1;
2034 * As with any blocking lock, a contender needs to start
2035 * listening for wakeups before attempting the trylock,
2036 * otherwise it can miss the wakeup from the unlock and sleep
2037 * indefinitely. This is just open-coded because our locking
2038 * is so particular to memcg hierarchies.
2040 wakeups = atomic_read(&memcg->oom_wakeups);
2041 mem_cgroup_mark_under_oom(memcg);
2043 locked = mem_cgroup_oom_trylock(memcg);
2046 mem_cgroup_oom_notify(memcg);
2048 if (locked && !memcg->oom_kill_disable) {
2049 mem_cgroup_unmark_under_oom(memcg);
2050 mem_cgroup_out_of_memory(memcg, mask, order);
2051 mem_cgroup_oom_unlock(memcg);
2053 * There is no guarantee that an OOM-lock contender
2054 * sees the wakeups triggered by the OOM kill
2055 * uncharges. Wake any sleepers explicitely.
2057 memcg_oom_recover(memcg);
2060 * A system call can just return -ENOMEM, but if this
2061 * is a page fault and somebody else is handling the
2062 * OOM already, we need to sleep on the OOM waitqueue
2063 * for this memcg until the situation is resolved.
2064 * Which can take some time because it might be
2065 * handled by a userspace task.
2067 * However, this is the charge context, which means
2068 * that we may sit on a large call stack and hold
2069 * various filesystem locks, the mmap_sem etc. and we
2070 * don't want the OOM handler to deadlock on them
2071 * while we sit here and wait. Store the current OOM
2072 * context in the task_struct, then return -ENOMEM.
2073 * At the end of the page fault handler, with the
2074 * stack unwound, pagefault_out_of_memory() will check
2075 * back with us by calling
2076 * mem_cgroup_oom_synchronize(), possibly putting the
2079 current->memcg_oom.oom_locked = locked;
2080 current->memcg_oom.wakeups = wakeups;
2081 css_get(&memcg->css);
2082 current->memcg_oom.wait_on_memcg = memcg;
2087 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2089 * This has to be called at the end of a page fault if the the memcg
2090 * OOM handler was enabled and the fault is returning %VM_FAULT_OOM.
2092 * Memcg supports userspace OOM handling, so failed allocations must
2093 * sleep on a waitqueue until the userspace task resolves the
2094 * situation. Sleeping directly in the charge context with all kinds
2095 * of locks held is not a good idea, instead we remember an OOM state
2096 * in the task and mem_cgroup_oom_synchronize() has to be called at
2097 * the end of the page fault to put the task to sleep and clean up the
2100 * Returns %true if an ongoing memcg OOM situation was detected and
2101 * finalized, %false otherwise.
2103 bool mem_cgroup_oom_synchronize(void)
2105 struct oom_wait_info owait;
2106 struct mem_cgroup *memcg;
2108 /* OOM is global, do not handle */
2109 if (!current->memcg_oom.in_memcg_oom)
2113 * We invoked the OOM killer but there is a chance that a kill
2114 * did not free up any charges. Everybody else might already
2115 * be sleeping, so restart the fault and keep the rampage
2116 * going until some charges are released.
2118 memcg = current->memcg_oom.wait_on_memcg;
2122 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2125 owait.memcg = memcg;
2126 owait.wait.flags = 0;
2127 owait.wait.func = memcg_oom_wake_function;
2128 owait.wait.private = current;
2129 INIT_LIST_HEAD(&owait.wait.task_list);
2131 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2132 /* Only sleep if we didn't miss any wakeups since OOM */
2133 if (atomic_read(&memcg->oom_wakeups) == current->memcg_oom.wakeups)
2135 finish_wait(&memcg_oom_waitq, &owait.wait);
2137 mem_cgroup_unmark_under_oom(memcg);
2138 if (current->memcg_oom.oom_locked) {
2139 mem_cgroup_oom_unlock(memcg);
2141 * There is no guarantee that an OOM-lock contender
2142 * sees the wakeups triggered by the OOM kill
2143 * uncharges. Wake any sleepers explicitely.
2145 memcg_oom_recover(memcg);
2147 css_put(&memcg->css);
2148 current->memcg_oom.wait_on_memcg = NULL;
2150 current->memcg_oom.in_memcg_oom = 0;
2155 * Currently used to update mapped file statistics, but the routine can be
2156 * generalized to update other statistics as well.
2158 * Notes: Race condition
2160 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2161 * it tends to be costly. But considering some conditions, we doesn't need
2162 * to do so _always_.
2164 * Considering "charge", lock_page_cgroup() is not required because all
2165 * file-stat operations happen after a page is attached to radix-tree. There
2166 * are no race with "charge".
2168 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2169 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2170 * if there are race with "uncharge". Statistics itself is properly handled
2173 * Considering "move", this is an only case we see a race. To make the race
2174 * small, we check mm->moving_account and detect there are possibility of race
2175 * If there is, we take a lock.
2178 void __mem_cgroup_begin_update_page_stat(struct page *page,
2179 bool *locked, unsigned long *flags)
2181 struct mem_cgroup *memcg;
2182 struct page_cgroup *pc;
2184 pc = lookup_page_cgroup(page);
2186 memcg = pc->mem_cgroup;
2187 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2190 * If this memory cgroup is not under account moving, we don't
2191 * need to take move_lock_mem_cgroup(). Because we already hold
2192 * rcu_read_lock(), any calls to move_account will be delayed until
2193 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2195 if (!mem_cgroup_stolen(memcg))
2198 move_lock_mem_cgroup(memcg, flags);
2199 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2200 move_unlock_mem_cgroup(memcg, flags);
2206 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2208 struct page_cgroup *pc = lookup_page_cgroup(page);
2211 * It's guaranteed that pc->mem_cgroup never changes while
2212 * lock is held because a routine modifies pc->mem_cgroup
2213 * should take move_lock_mem_cgroup().
2215 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2218 void mem_cgroup_update_page_stat(struct page *page,
2219 enum mem_cgroup_stat_index idx, int val)
2221 struct mem_cgroup *memcg;
2222 struct page_cgroup *pc = lookup_page_cgroup(page);
2223 unsigned long uninitialized_var(flags);
2225 if (mem_cgroup_disabled())
2228 memcg = pc->mem_cgroup;
2229 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2232 this_cpu_add(memcg->stat->count[idx], val);
2236 * size of first charge trial. "32" comes from vmscan.c's magic value.
2237 * TODO: maybe necessary to use big numbers in big irons.
2239 #define CHARGE_BATCH 32U
2240 struct memcg_stock_pcp {
2241 struct mem_cgroup *cached; /* this never be root cgroup */
2242 unsigned int nr_pages;
2243 struct work_struct work;
2244 unsigned long flags;
2245 #define FLUSHING_CACHED_CHARGE 0
2247 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2248 static DEFINE_MUTEX(percpu_charge_mutex);
2251 * consume_stock: Try to consume stocked charge on this cpu.
2252 * @memcg: memcg to consume from.
2253 * @nr_pages: how many pages to charge.
2255 * The charges will only happen if @memcg matches the current cpu's memcg
2256 * stock, and at least @nr_pages are available in that stock. Failure to
2257 * service an allocation will refill the stock.
2259 * returns true if successful, false otherwise.
2261 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2263 struct memcg_stock_pcp *stock;
2266 if (nr_pages > CHARGE_BATCH)
2269 stock = &get_cpu_var(memcg_stock);
2270 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2271 stock->nr_pages -= nr_pages;
2272 else /* need to call res_counter_charge */
2274 put_cpu_var(memcg_stock);
2279 * Returns stocks cached in percpu to res_counter and reset cached information.
2281 static void drain_stock(struct memcg_stock_pcp *stock)
2283 struct mem_cgroup *old = stock->cached;
2285 if (stock->nr_pages) {
2286 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2288 res_counter_uncharge(&old->res, bytes);
2289 if (do_swap_account)
2290 res_counter_uncharge(&old->memsw, bytes);
2291 stock->nr_pages = 0;
2293 stock->cached = NULL;
2297 * This must be called under preempt disabled or must be called by
2298 * a thread which is pinned to local cpu.
2300 static void drain_local_stock(struct work_struct *dummy)
2302 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2304 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2307 static void __init memcg_stock_init(void)
2311 for_each_possible_cpu(cpu) {
2312 struct memcg_stock_pcp *stock =
2313 &per_cpu(memcg_stock, cpu);
2314 INIT_WORK(&stock->work, drain_local_stock);
2319 * Cache charges(val) which is from res_counter, to local per_cpu area.
2320 * This will be consumed by consume_stock() function, later.
2322 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2324 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2326 if (stock->cached != memcg) { /* reset if necessary */
2328 stock->cached = memcg;
2330 stock->nr_pages += nr_pages;
2331 put_cpu_var(memcg_stock);
2335 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2336 * of the hierarchy under it. sync flag says whether we should block
2337 * until the work is done.
2339 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2343 /* Notify other cpus that system-wide "drain" is running */
2346 for_each_online_cpu(cpu) {
2347 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2348 struct mem_cgroup *memcg;
2350 memcg = stock->cached;
2351 if (!memcg || !stock->nr_pages)
2353 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2355 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2357 drain_local_stock(&stock->work);
2359 schedule_work_on(cpu, &stock->work);
2367 for_each_online_cpu(cpu) {
2368 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2369 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2370 flush_work(&stock->work);
2377 * Tries to drain stocked charges in other cpus. This function is asynchronous
2378 * and just put a work per cpu for draining localy on each cpu. Caller can
2379 * expects some charges will be back to res_counter later but cannot wait for
2382 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2385 * If someone calls draining, avoid adding more kworker runs.
2387 if (!mutex_trylock(&percpu_charge_mutex))
2389 drain_all_stock(root_memcg, false);
2390 mutex_unlock(&percpu_charge_mutex);
2393 /* This is a synchronous drain interface. */
2394 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2396 /* called when force_empty is called */
2397 mutex_lock(&percpu_charge_mutex);
2398 drain_all_stock(root_memcg, true);
2399 mutex_unlock(&percpu_charge_mutex);
2403 * This function drains percpu counter value from DEAD cpu and
2404 * move it to local cpu. Note that this function can be preempted.
2406 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2410 spin_lock(&memcg->pcp_counter_lock);
2411 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2412 long x = per_cpu(memcg->stat->count[i], cpu);
2414 per_cpu(memcg->stat->count[i], cpu) = 0;
2415 memcg->nocpu_base.count[i] += x;
2417 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2418 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2420 per_cpu(memcg->stat->events[i], cpu) = 0;
2421 memcg->nocpu_base.events[i] += x;
2423 spin_unlock(&memcg->pcp_counter_lock);
2426 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2427 unsigned long action,
2430 int cpu = (unsigned long)hcpu;
2431 struct memcg_stock_pcp *stock;
2432 struct mem_cgroup *iter;
2434 if (action == CPU_ONLINE)
2437 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2440 for_each_mem_cgroup(iter)
2441 mem_cgroup_drain_pcp_counter(iter, cpu);
2443 stock = &per_cpu(memcg_stock, cpu);
2449 /* See __mem_cgroup_try_charge() for details */
2451 CHARGE_OK, /* success */
2452 CHARGE_RETRY, /* need to retry but retry is not bad */
2453 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2454 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2457 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2458 unsigned int nr_pages, unsigned int min_pages,
2461 unsigned long csize = nr_pages * PAGE_SIZE;
2462 struct mem_cgroup *mem_over_limit;
2463 struct res_counter *fail_res;
2464 unsigned long flags = 0;
2467 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2470 if (!do_swap_account)
2472 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2476 res_counter_uncharge(&memcg->res, csize);
2477 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2478 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2480 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2482 * Never reclaim on behalf of optional batching, retry with a
2483 * single page instead.
2485 if (nr_pages > min_pages)
2486 return CHARGE_RETRY;
2488 if (!(gfp_mask & __GFP_WAIT))
2489 return CHARGE_WOULDBLOCK;
2491 if (gfp_mask & __GFP_NORETRY)
2492 return CHARGE_NOMEM;
2494 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2495 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2496 return CHARGE_RETRY;
2498 * Even though the limit is exceeded at this point, reclaim
2499 * may have been able to free some pages. Retry the charge
2500 * before killing the task.
2502 * Only for regular pages, though: huge pages are rather
2503 * unlikely to succeed so close to the limit, and we fall back
2504 * to regular pages anyway in case of failure.
2506 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2507 return CHARGE_RETRY;
2510 * At task move, charge accounts can be doubly counted. So, it's
2511 * better to wait until the end of task_move if something is going on.
2513 if (mem_cgroup_wait_acct_move(mem_over_limit))
2514 return CHARGE_RETRY;
2517 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2519 return CHARGE_NOMEM;
2523 * __mem_cgroup_try_charge() does
2524 * 1. detect memcg to be charged against from passed *mm and *ptr,
2525 * 2. update res_counter
2526 * 3. call memory reclaim if necessary.
2528 * In some special case, if the task is fatal, fatal_signal_pending() or
2529 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2530 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2531 * as possible without any hazards. 2: all pages should have a valid
2532 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2533 * pointer, that is treated as a charge to root_mem_cgroup.
2535 * So __mem_cgroup_try_charge() will return
2536 * 0 ... on success, filling *ptr with a valid memcg pointer.
2537 * -ENOMEM ... charge failure because of resource limits.
2538 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2540 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2541 * the oom-killer can be invoked.
2543 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2545 unsigned int nr_pages,
2546 struct mem_cgroup **ptr,
2549 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2550 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2551 struct mem_cgroup *memcg = NULL;
2555 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2556 * in system level. So, allow to go ahead dying process in addition to
2559 if (unlikely(test_thread_flag(TIF_MEMDIE)
2560 || fatal_signal_pending(current)))
2564 * We always charge the cgroup the mm_struct belongs to.
2565 * The mm_struct's mem_cgroup changes on task migration if the
2566 * thread group leader migrates. It's possible that mm is not
2567 * set, if so charge the root memcg (happens for pagecache usage).
2570 *ptr = root_mem_cgroup;
2572 if (*ptr) { /* css should be a valid one */
2574 if (mem_cgroup_is_root(memcg))
2576 if (consume_stock(memcg, nr_pages))
2578 css_get(&memcg->css);
2580 struct task_struct *p;
2583 p = rcu_dereference(mm->owner);
2585 * Because we don't have task_lock(), "p" can exit.
2586 * In that case, "memcg" can point to root or p can be NULL with
2587 * race with swapoff. Then, we have small risk of mis-accouning.
2588 * But such kind of mis-account by race always happens because
2589 * we don't have cgroup_mutex(). It's overkill and we allo that
2591 * (*) swapoff at el will charge against mm-struct not against
2592 * task-struct. So, mm->owner can be NULL.
2594 memcg = mem_cgroup_from_task(p);
2596 memcg = root_mem_cgroup;
2597 if (mem_cgroup_is_root(memcg)) {
2601 if (consume_stock(memcg, nr_pages)) {
2603 * It seems dagerous to access memcg without css_get().
2604 * But considering how consume_stok works, it's not
2605 * necessary. If consume_stock success, some charges
2606 * from this memcg are cached on this cpu. So, we
2607 * don't need to call css_get()/css_tryget() before
2608 * calling consume_stock().
2613 /* after here, we may be blocked. we need to get refcnt */
2614 if (!css_tryget(&memcg->css)) {
2622 bool invoke_oom = oom && !nr_oom_retries;
2624 /* If killed, bypass charge */
2625 if (fatal_signal_pending(current)) {
2626 css_put(&memcg->css);
2630 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2631 nr_pages, invoke_oom);
2635 case CHARGE_RETRY: /* not in OOM situation but retry */
2637 css_put(&memcg->css);
2640 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2641 css_put(&memcg->css);
2643 case CHARGE_NOMEM: /* OOM routine works */
2644 if (!oom || invoke_oom) {
2645 css_put(&memcg->css);
2651 } while (ret != CHARGE_OK);
2653 if (batch > nr_pages)
2654 refill_stock(memcg, batch - nr_pages);
2655 css_put(&memcg->css);
2663 *ptr = root_mem_cgroup;
2668 * Somemtimes we have to undo a charge we got by try_charge().
2669 * This function is for that and do uncharge, put css's refcnt.
2670 * gotten by try_charge().
2672 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2673 unsigned int nr_pages)
2675 if (!mem_cgroup_is_root(memcg)) {
2676 unsigned long bytes = nr_pages * PAGE_SIZE;
2678 res_counter_uncharge(&memcg->res, bytes);
2679 if (do_swap_account)
2680 res_counter_uncharge(&memcg->memsw, bytes);
2685 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2686 * This is useful when moving usage to parent cgroup.
2688 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2689 unsigned int nr_pages)
2691 unsigned long bytes = nr_pages * PAGE_SIZE;
2693 if (mem_cgroup_is_root(memcg))
2696 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2697 if (do_swap_account)
2698 res_counter_uncharge_until(&memcg->memsw,
2699 memcg->memsw.parent, bytes);
2703 * A helper function to get mem_cgroup from ID. must be called under
2704 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2705 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2706 * called against removed memcg.)
2708 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2710 struct cgroup_subsys_state *css;
2712 /* ID 0 is unused ID */
2715 css = css_lookup(&mem_cgroup_subsys, id);
2718 return mem_cgroup_from_css(css);
2721 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2723 struct mem_cgroup *memcg = NULL;
2724 struct page_cgroup *pc;
2728 VM_BUG_ON(!PageLocked(page));
2730 pc = lookup_page_cgroup(page);
2731 lock_page_cgroup(pc);
2732 if (PageCgroupUsed(pc)) {
2733 memcg = pc->mem_cgroup;
2734 if (memcg && !css_tryget(&memcg->css))
2736 } else if (PageSwapCache(page)) {
2737 ent.val = page_private(page);
2738 id = lookup_swap_cgroup_id(ent);
2740 memcg = mem_cgroup_lookup(id);
2741 if (memcg && !css_tryget(&memcg->css))
2745 unlock_page_cgroup(pc);
2749 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2751 unsigned int nr_pages,
2752 enum charge_type ctype,
2755 struct page_cgroup *pc = lookup_page_cgroup(page);
2756 struct zone *uninitialized_var(zone);
2757 struct lruvec *lruvec;
2758 bool was_on_lru = false;
2761 lock_page_cgroup(pc);
2762 VM_BUG_ON(PageCgroupUsed(pc));
2764 * we don't need page_cgroup_lock about tail pages, becase they are not
2765 * accessed by any other context at this point.
2769 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2770 * may already be on some other mem_cgroup's LRU. Take care of it.
2773 zone = page_zone(page);
2774 spin_lock_irq(&zone->lru_lock);
2775 if (PageLRU(page)) {
2776 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2778 del_page_from_lru_list(page, lruvec, page_lru(page));
2783 pc->mem_cgroup = memcg;
2785 * We access a page_cgroup asynchronously without lock_page_cgroup().
2786 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2787 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2788 * before USED bit, we need memory barrier here.
2789 * See mem_cgroup_add_lru_list(), etc.
2792 SetPageCgroupUsed(pc);
2796 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2797 VM_BUG_ON(PageLRU(page));
2799 add_page_to_lru_list(page, lruvec, page_lru(page));
2801 spin_unlock_irq(&zone->lru_lock);
2804 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2809 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2810 unlock_page_cgroup(pc);
2813 * "charge_statistics" updated event counter.
2815 memcg_check_events(memcg, page);
2818 static DEFINE_MUTEX(set_limit_mutex);
2820 #ifdef CONFIG_MEMCG_KMEM
2821 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2823 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2824 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2828 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2829 * in the memcg_cache_params struct.
2831 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2833 struct kmem_cache *cachep;
2835 VM_BUG_ON(p->is_root_cache);
2836 cachep = p->root_cache;
2837 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2840 #ifdef CONFIG_SLABINFO
2841 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2842 struct cftype *cft, struct seq_file *m)
2844 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2845 struct memcg_cache_params *params;
2847 if (!memcg_can_account_kmem(memcg))
2850 print_slabinfo_header(m);
2852 mutex_lock(&memcg->slab_caches_mutex);
2853 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2854 cache_show(memcg_params_to_cache(params), m);
2855 mutex_unlock(&memcg->slab_caches_mutex);
2861 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2863 struct res_counter *fail_res;
2864 struct mem_cgroup *_memcg;
2868 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2873 * Conditions under which we can wait for the oom_killer. Those are
2874 * the same conditions tested by the core page allocator
2876 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2879 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2882 if (ret == -EINTR) {
2884 * __mem_cgroup_try_charge() chosed to bypass to root due to
2885 * OOM kill or fatal signal. Since our only options are to
2886 * either fail the allocation or charge it to this cgroup, do
2887 * it as a temporary condition. But we can't fail. From a
2888 * kmem/slab perspective, the cache has already been selected,
2889 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2892 * This condition will only trigger if the task entered
2893 * memcg_charge_kmem in a sane state, but was OOM-killed during
2894 * __mem_cgroup_try_charge() above. Tasks that were already
2895 * dying when the allocation triggers should have been already
2896 * directed to the root cgroup in memcontrol.h
2898 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2899 if (do_swap_account)
2900 res_counter_charge_nofail(&memcg->memsw, size,
2904 res_counter_uncharge(&memcg->kmem, size);
2909 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2911 res_counter_uncharge(&memcg->res, size);
2912 if (do_swap_account)
2913 res_counter_uncharge(&memcg->memsw, size);
2916 if (res_counter_uncharge(&memcg->kmem, size))
2920 * Releases a reference taken in kmem_cgroup_css_offline in case
2921 * this last uncharge is racing with the offlining code or it is
2922 * outliving the memcg existence.
2924 * The memory barrier imposed by test&clear is paired with the
2925 * explicit one in memcg_kmem_mark_dead().
2927 if (memcg_kmem_test_and_clear_dead(memcg))
2928 css_put(&memcg->css);
2931 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2936 mutex_lock(&memcg->slab_caches_mutex);
2937 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2938 mutex_unlock(&memcg->slab_caches_mutex);
2942 * helper for acessing a memcg's index. It will be used as an index in the
2943 * child cache array in kmem_cache, and also to derive its name. This function
2944 * will return -1 when this is not a kmem-limited memcg.
2946 int memcg_cache_id(struct mem_cgroup *memcg)
2948 return memcg ? memcg->kmemcg_id : -1;
2952 * This ends up being protected by the set_limit mutex, during normal
2953 * operation, because that is its main call site.
2955 * But when we create a new cache, we can call this as well if its parent
2956 * is kmem-limited. That will have to hold set_limit_mutex as well.
2958 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2962 num = ida_simple_get(&kmem_limited_groups,
2963 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2967 * After this point, kmem_accounted (that we test atomically in
2968 * the beginning of this conditional), is no longer 0. This
2969 * guarantees only one process will set the following boolean
2970 * to true. We don't need test_and_set because we're protected
2971 * by the set_limit_mutex anyway.
2973 memcg_kmem_set_activated(memcg);
2975 ret = memcg_update_all_caches(num+1);
2977 ida_simple_remove(&kmem_limited_groups, num);
2978 memcg_kmem_clear_activated(memcg);
2982 memcg->kmemcg_id = num;
2983 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2984 mutex_init(&memcg->slab_caches_mutex);
2988 static size_t memcg_caches_array_size(int num_groups)
2991 if (num_groups <= 0)
2994 size = 2 * num_groups;
2995 if (size < MEMCG_CACHES_MIN_SIZE)
2996 size = MEMCG_CACHES_MIN_SIZE;
2997 else if (size > MEMCG_CACHES_MAX_SIZE)
2998 size = MEMCG_CACHES_MAX_SIZE;
3004 * We should update the current array size iff all caches updates succeed. This
3005 * can only be done from the slab side. The slab mutex needs to be held when
3008 void memcg_update_array_size(int num)
3010 if (num > memcg_limited_groups_array_size)
3011 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3014 static void kmem_cache_destroy_work_func(struct work_struct *w);
3016 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3018 struct memcg_cache_params *cur_params = s->memcg_params;
3020 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3022 if (num_groups > memcg_limited_groups_array_size) {
3024 ssize_t size = memcg_caches_array_size(num_groups);
3026 size *= sizeof(void *);
3027 size += offsetof(struct memcg_cache_params, memcg_caches);
3029 s->memcg_params = kzalloc(size, GFP_KERNEL);
3030 if (!s->memcg_params) {
3031 s->memcg_params = cur_params;
3035 s->memcg_params->is_root_cache = true;
3038 * There is the chance it will be bigger than
3039 * memcg_limited_groups_array_size, if we failed an allocation
3040 * in a cache, in which case all caches updated before it, will
3041 * have a bigger array.
3043 * But if that is the case, the data after
3044 * memcg_limited_groups_array_size is certainly unused
3046 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3047 if (!cur_params->memcg_caches[i])
3049 s->memcg_params->memcg_caches[i] =
3050 cur_params->memcg_caches[i];
3054 * Ideally, we would wait until all caches succeed, and only
3055 * then free the old one. But this is not worth the extra
3056 * pointer per-cache we'd have to have for this.
3058 * It is not a big deal if some caches are left with a size
3059 * bigger than the others. And all updates will reset this
3067 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3068 struct kmem_cache *root_cache)
3072 if (!memcg_kmem_enabled())
3076 size = offsetof(struct memcg_cache_params, memcg_caches);
3077 size += memcg_limited_groups_array_size * sizeof(void *);
3079 size = sizeof(struct memcg_cache_params);
3081 s->memcg_params = kzalloc(size, GFP_KERNEL);
3082 if (!s->memcg_params)
3086 s->memcg_params->memcg = memcg;
3087 s->memcg_params->root_cache = root_cache;
3088 INIT_WORK(&s->memcg_params->destroy,
3089 kmem_cache_destroy_work_func);
3091 s->memcg_params->is_root_cache = true;
3096 void memcg_release_cache(struct kmem_cache *s)
3098 struct kmem_cache *root;
3099 struct mem_cgroup *memcg;
3103 * This happens, for instance, when a root cache goes away before we
3106 if (!s->memcg_params)
3109 if (s->memcg_params->is_root_cache)
3112 memcg = s->memcg_params->memcg;
3113 id = memcg_cache_id(memcg);
3115 root = s->memcg_params->root_cache;
3116 root->memcg_params->memcg_caches[id] = NULL;
3118 mutex_lock(&memcg->slab_caches_mutex);
3119 list_del(&s->memcg_params->list);
3120 mutex_unlock(&memcg->slab_caches_mutex);
3122 css_put(&memcg->css);
3124 kfree(s->memcg_params);
3128 * During the creation a new cache, we need to disable our accounting mechanism
3129 * altogether. This is true even if we are not creating, but rather just
3130 * enqueing new caches to be created.
3132 * This is because that process will trigger allocations; some visible, like
3133 * explicit kmallocs to auxiliary data structures, name strings and internal
3134 * cache structures; some well concealed, like INIT_WORK() that can allocate
3135 * objects during debug.
3137 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3138 * to it. This may not be a bounded recursion: since the first cache creation
3139 * failed to complete (waiting on the allocation), we'll just try to create the
3140 * cache again, failing at the same point.
3142 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3143 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3144 * inside the following two functions.
3146 static inline void memcg_stop_kmem_account(void)
3148 VM_BUG_ON(!current->mm);
3149 current->memcg_kmem_skip_account++;
3152 static inline void memcg_resume_kmem_account(void)
3154 VM_BUG_ON(!current->mm);
3155 current->memcg_kmem_skip_account--;
3158 static void kmem_cache_destroy_work_func(struct work_struct *w)
3160 struct kmem_cache *cachep;
3161 struct memcg_cache_params *p;
3163 p = container_of(w, struct memcg_cache_params, destroy);
3165 cachep = memcg_params_to_cache(p);
3168 * If we get down to 0 after shrink, we could delete right away.
3169 * However, memcg_release_pages() already puts us back in the workqueue
3170 * in that case. If we proceed deleting, we'll get a dangling
3171 * reference, and removing the object from the workqueue in that case
3172 * is unnecessary complication. We are not a fast path.
3174 * Note that this case is fundamentally different from racing with
3175 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3176 * kmem_cache_shrink, not only we would be reinserting a dead cache
3177 * into the queue, but doing so from inside the worker racing to
3180 * So if we aren't down to zero, we'll just schedule a worker and try
3183 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3184 kmem_cache_shrink(cachep);
3185 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3188 kmem_cache_destroy(cachep);
3191 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3193 if (!cachep->memcg_params->dead)
3197 * There are many ways in which we can get here.
3199 * We can get to a memory-pressure situation while the delayed work is
3200 * still pending to run. The vmscan shrinkers can then release all
3201 * cache memory and get us to destruction. If this is the case, we'll
3202 * be executed twice, which is a bug (the second time will execute over
3203 * bogus data). In this case, cancelling the work should be fine.
3205 * But we can also get here from the worker itself, if
3206 * kmem_cache_shrink is enough to shake all the remaining objects and
3207 * get the page count to 0. In this case, we'll deadlock if we try to
3208 * cancel the work (the worker runs with an internal lock held, which
3209 * is the same lock we would hold for cancel_work_sync().)
3211 * Since we can't possibly know who got us here, just refrain from
3212 * running if there is already work pending
3214 if (work_pending(&cachep->memcg_params->destroy))
3217 * We have to defer the actual destroying to a workqueue, because
3218 * we might currently be in a context that cannot sleep.
3220 schedule_work(&cachep->memcg_params->destroy);
3224 * This lock protects updaters, not readers. We want readers to be as fast as
3225 * they can, and they will either see NULL or a valid cache value. Our model
3226 * allow them to see NULL, in which case the root memcg will be selected.
3228 * We need this lock because multiple allocations to the same cache from a non
3229 * will span more than one worker. Only one of them can create the cache.
3231 static DEFINE_MUTEX(memcg_cache_mutex);
3234 * Called with memcg_cache_mutex held
3236 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3237 struct kmem_cache *s)
3239 struct kmem_cache *new;
3240 static char *tmp_name = NULL;
3242 lockdep_assert_held(&memcg_cache_mutex);
3245 * kmem_cache_create_memcg duplicates the given name and
3246 * cgroup_name for this name requires RCU context.
3247 * This static temporary buffer is used to prevent from
3248 * pointless shortliving allocation.
3251 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3257 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3258 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3261 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3262 (s->flags & ~SLAB_PANIC), s->ctor, s);
3265 new->allocflags |= __GFP_KMEMCG;
3270 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3271 struct kmem_cache *cachep)
3273 struct kmem_cache *new_cachep;
3276 BUG_ON(!memcg_can_account_kmem(memcg));
3278 idx = memcg_cache_id(memcg);
3280 mutex_lock(&memcg_cache_mutex);
3281 new_cachep = cachep->memcg_params->memcg_caches[idx];
3283 css_put(&memcg->css);
3287 new_cachep = kmem_cache_dup(memcg, cachep);
3288 if (new_cachep == NULL) {
3289 new_cachep = cachep;
3290 css_put(&memcg->css);
3294 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3296 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3298 * the readers won't lock, make sure everybody sees the updated value,
3299 * so they won't put stuff in the queue again for no reason
3303 mutex_unlock(&memcg_cache_mutex);
3307 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3309 struct kmem_cache *c;
3312 if (!s->memcg_params)
3314 if (!s->memcg_params->is_root_cache)
3318 * If the cache is being destroyed, we trust that there is no one else
3319 * requesting objects from it. Even if there are, the sanity checks in
3320 * kmem_cache_destroy should caught this ill-case.
3322 * Still, we don't want anyone else freeing memcg_caches under our
3323 * noses, which can happen if a new memcg comes to life. As usual,
3324 * we'll take the set_limit_mutex to protect ourselves against this.
3326 mutex_lock(&set_limit_mutex);
3327 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3328 c = s->memcg_params->memcg_caches[i];
3333 * We will now manually delete the caches, so to avoid races
3334 * we need to cancel all pending destruction workers and
3335 * proceed with destruction ourselves.
3337 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3338 * and that could spawn the workers again: it is likely that
3339 * the cache still have active pages until this very moment.
3340 * This would lead us back to mem_cgroup_destroy_cache.
3342 * But that will not execute at all if the "dead" flag is not
3343 * set, so flip it down to guarantee we are in control.
3345 c->memcg_params->dead = false;
3346 cancel_work_sync(&c->memcg_params->destroy);
3347 kmem_cache_destroy(c);
3349 mutex_unlock(&set_limit_mutex);
3352 struct create_work {
3353 struct mem_cgroup *memcg;
3354 struct kmem_cache *cachep;
3355 struct work_struct work;
3358 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3360 struct kmem_cache *cachep;
3361 struct memcg_cache_params *params;
3363 if (!memcg_kmem_is_active(memcg))
3366 mutex_lock(&memcg->slab_caches_mutex);
3367 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3368 cachep = memcg_params_to_cache(params);
3369 cachep->memcg_params->dead = true;
3370 schedule_work(&cachep->memcg_params->destroy);
3372 mutex_unlock(&memcg->slab_caches_mutex);
3375 static void memcg_create_cache_work_func(struct work_struct *w)
3377 struct create_work *cw;
3379 cw = container_of(w, struct create_work, work);
3380 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3385 * Enqueue the creation of a per-memcg kmem_cache.
3387 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3388 struct kmem_cache *cachep)
3390 struct create_work *cw;
3392 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3394 css_put(&memcg->css);
3399 cw->cachep = cachep;
3401 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3402 schedule_work(&cw->work);
3405 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3406 struct kmem_cache *cachep)
3409 * We need to stop accounting when we kmalloc, because if the
3410 * corresponding kmalloc cache is not yet created, the first allocation
3411 * in __memcg_create_cache_enqueue will recurse.
3413 * However, it is better to enclose the whole function. Depending on
3414 * the debugging options enabled, INIT_WORK(), for instance, can
3415 * trigger an allocation. This too, will make us recurse. Because at
3416 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3417 * the safest choice is to do it like this, wrapping the whole function.
3419 memcg_stop_kmem_account();
3420 __memcg_create_cache_enqueue(memcg, cachep);
3421 memcg_resume_kmem_account();
3424 * Return the kmem_cache we're supposed to use for a slab allocation.
3425 * We try to use the current memcg's version of the cache.
3427 * If the cache does not exist yet, if we are the first user of it,
3428 * we either create it immediately, if possible, or create it asynchronously
3430 * In the latter case, we will let the current allocation go through with
3431 * the original cache.
3433 * Can't be called in interrupt context or from kernel threads.
3434 * This function needs to be called with rcu_read_lock() held.
3436 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3439 struct mem_cgroup *memcg;
3442 VM_BUG_ON(!cachep->memcg_params);
3443 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3445 if (!current->mm || current->memcg_kmem_skip_account)
3449 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3451 if (!memcg_can_account_kmem(memcg))
3454 idx = memcg_cache_id(memcg);
3457 * barrier to mare sure we're always seeing the up to date value. The
3458 * code updating memcg_caches will issue a write barrier to match this.
3460 read_barrier_depends();
3461 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3462 cachep = cachep->memcg_params->memcg_caches[idx];
3466 /* The corresponding put will be done in the workqueue. */
3467 if (!css_tryget(&memcg->css))
3472 * If we are in a safe context (can wait, and not in interrupt
3473 * context), we could be be predictable and return right away.
3474 * This would guarantee that the allocation being performed
3475 * already belongs in the new cache.
3477 * However, there are some clashes that can arrive from locking.
3478 * For instance, because we acquire the slab_mutex while doing
3479 * kmem_cache_dup, this means no further allocation could happen
3480 * with the slab_mutex held.
3482 * Also, because cache creation issue get_online_cpus(), this
3483 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3484 * that ends up reversed during cpu hotplug. (cpuset allocates
3485 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3486 * better to defer everything.
3488 memcg_create_cache_enqueue(memcg, cachep);
3494 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3497 * We need to verify if the allocation against current->mm->owner's memcg is
3498 * possible for the given order. But the page is not allocated yet, so we'll
3499 * need a further commit step to do the final arrangements.
3501 * It is possible for the task to switch cgroups in this mean time, so at
3502 * commit time, we can't rely on task conversion any longer. We'll then use
3503 * the handle argument to return to the caller which cgroup we should commit
3504 * against. We could also return the memcg directly and avoid the pointer
3505 * passing, but a boolean return value gives better semantics considering
3506 * the compiled-out case as well.
3508 * Returning true means the allocation is possible.
3511 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3513 struct mem_cgroup *memcg;
3519 * Disabling accounting is only relevant for some specific memcg
3520 * internal allocations. Therefore we would initially not have such
3521 * check here, since direct calls to the page allocator that are marked
3522 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3523 * concerned with cache allocations, and by having this test at
3524 * memcg_kmem_get_cache, we are already able to relay the allocation to
3525 * the root cache and bypass the memcg cache altogether.
3527 * There is one exception, though: the SLUB allocator does not create
3528 * large order caches, but rather service large kmallocs directly from
3529 * the page allocator. Therefore, the following sequence when backed by
3530 * the SLUB allocator:
3532 * memcg_stop_kmem_account();
3533 * kmalloc(<large_number>)
3534 * memcg_resume_kmem_account();
3536 * would effectively ignore the fact that we should skip accounting,
3537 * since it will drive us directly to this function without passing
3538 * through the cache selector memcg_kmem_get_cache. Such large
3539 * allocations are extremely rare but can happen, for instance, for the
3540 * cache arrays. We bring this test here.
3542 if (!current->mm || current->memcg_kmem_skip_account)
3545 memcg = try_get_mem_cgroup_from_mm(current->mm);
3548 * very rare case described in mem_cgroup_from_task. Unfortunately there
3549 * isn't much we can do without complicating this too much, and it would
3550 * be gfp-dependent anyway. Just let it go
3552 if (unlikely(!memcg))
3555 if (!memcg_can_account_kmem(memcg)) {
3556 css_put(&memcg->css);
3560 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3564 css_put(&memcg->css);
3568 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3571 struct page_cgroup *pc;
3573 VM_BUG_ON(mem_cgroup_is_root(memcg));
3575 /* The page allocation failed. Revert */
3577 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3581 pc = lookup_page_cgroup(page);
3582 lock_page_cgroup(pc);
3583 pc->mem_cgroup = memcg;
3584 SetPageCgroupUsed(pc);
3585 unlock_page_cgroup(pc);
3588 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3590 struct mem_cgroup *memcg = NULL;
3591 struct page_cgroup *pc;
3594 pc = lookup_page_cgroup(page);
3596 * Fast unlocked return. Theoretically might have changed, have to
3597 * check again after locking.
3599 if (!PageCgroupUsed(pc))
3602 lock_page_cgroup(pc);
3603 if (PageCgroupUsed(pc)) {
3604 memcg = pc->mem_cgroup;
3605 ClearPageCgroupUsed(pc);
3607 unlock_page_cgroup(pc);
3610 * We trust that only if there is a memcg associated with the page, it
3611 * is a valid allocation
3616 VM_BUG_ON(mem_cgroup_is_root(memcg));
3617 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3620 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3623 #endif /* CONFIG_MEMCG_KMEM */
3625 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3627 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3629 * Because tail pages are not marked as "used", set it. We're under
3630 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3631 * charge/uncharge will be never happen and move_account() is done under
3632 * compound_lock(), so we don't have to take care of races.
3634 void mem_cgroup_split_huge_fixup(struct page *head)
3636 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3637 struct page_cgroup *pc;
3638 struct mem_cgroup *memcg;
3641 if (mem_cgroup_disabled())
3644 memcg = head_pc->mem_cgroup;
3645 for (i = 1; i < HPAGE_PMD_NR; i++) {
3647 pc->mem_cgroup = memcg;
3648 smp_wmb();/* see __commit_charge() */
3649 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3651 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3654 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3657 * mem_cgroup_move_account - move account of the page
3659 * @nr_pages: number of regular pages (>1 for huge pages)
3660 * @pc: page_cgroup of the page.
3661 * @from: mem_cgroup which the page is moved from.
3662 * @to: mem_cgroup which the page is moved to. @from != @to.
3664 * The caller must confirm following.
3665 * - page is not on LRU (isolate_page() is useful.)
3666 * - compound_lock is held when nr_pages > 1
3668 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3671 static int mem_cgroup_move_account(struct page *page,
3672 unsigned int nr_pages,
3673 struct page_cgroup *pc,
3674 struct mem_cgroup *from,
3675 struct mem_cgroup *to)
3677 unsigned long flags;
3679 bool anon = PageAnon(page);
3681 VM_BUG_ON(from == to);
3682 VM_BUG_ON(PageLRU(page));
3684 * The page is isolated from LRU. So, collapse function
3685 * will not handle this page. But page splitting can happen.
3686 * Do this check under compound_page_lock(). The caller should
3690 if (nr_pages > 1 && !PageTransHuge(page))
3693 lock_page_cgroup(pc);
3696 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3699 move_lock_mem_cgroup(from, &flags);
3701 if (!anon && page_mapped(page)) {
3702 /* Update mapped_file data for mem_cgroup */
3704 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3705 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3708 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3710 /* caller should have done css_get */
3711 pc->mem_cgroup = to;
3712 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3713 move_unlock_mem_cgroup(from, &flags);
3716 unlock_page_cgroup(pc);
3720 memcg_check_events(to, page);
3721 memcg_check_events(from, page);
3727 * mem_cgroup_move_parent - moves page to the parent group
3728 * @page: the page to move
3729 * @pc: page_cgroup of the page
3730 * @child: page's cgroup
3732 * move charges to its parent or the root cgroup if the group has no
3733 * parent (aka use_hierarchy==0).
3734 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3735 * mem_cgroup_move_account fails) the failure is always temporary and
3736 * it signals a race with a page removal/uncharge or migration. In the
3737 * first case the page is on the way out and it will vanish from the LRU
3738 * on the next attempt and the call should be retried later.
3739 * Isolation from the LRU fails only if page has been isolated from
3740 * the LRU since we looked at it and that usually means either global
3741 * reclaim or migration going on. The page will either get back to the
3743 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3744 * (!PageCgroupUsed) or moved to a different group. The page will
3745 * disappear in the next attempt.
3747 static int mem_cgroup_move_parent(struct page *page,
3748 struct page_cgroup *pc,
3749 struct mem_cgroup *child)
3751 struct mem_cgroup *parent;
3752 unsigned int nr_pages;
3753 unsigned long uninitialized_var(flags);
3756 VM_BUG_ON(mem_cgroup_is_root(child));
3759 if (!get_page_unless_zero(page))
3761 if (isolate_lru_page(page))
3764 nr_pages = hpage_nr_pages(page);
3766 parent = parent_mem_cgroup(child);
3768 * If no parent, move charges to root cgroup.
3771 parent = root_mem_cgroup;
3774 VM_BUG_ON(!PageTransHuge(page));
3775 flags = compound_lock_irqsave(page);
3778 ret = mem_cgroup_move_account(page, nr_pages,
3781 __mem_cgroup_cancel_local_charge(child, nr_pages);
3784 compound_unlock_irqrestore(page, flags);
3785 putback_lru_page(page);
3793 * Charge the memory controller for page usage.
3795 * 0 if the charge was successful
3796 * < 0 if the cgroup is over its limit
3798 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3799 gfp_t gfp_mask, enum charge_type ctype)
3801 struct mem_cgroup *memcg = NULL;
3802 unsigned int nr_pages = 1;
3806 if (PageTransHuge(page)) {
3807 nr_pages <<= compound_order(page);
3808 VM_BUG_ON(!PageTransHuge(page));
3810 * Never OOM-kill a process for a huge page. The
3811 * fault handler will fall back to regular pages.
3816 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3819 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3823 int mem_cgroup_newpage_charge(struct page *page,
3824 struct mm_struct *mm, gfp_t gfp_mask)
3826 if (mem_cgroup_disabled())
3828 VM_BUG_ON(page_mapped(page));
3829 VM_BUG_ON(page->mapping && !PageAnon(page));
3831 return mem_cgroup_charge_common(page, mm, gfp_mask,
3832 MEM_CGROUP_CHARGE_TYPE_ANON);
3836 * While swap-in, try_charge -> commit or cancel, the page is locked.
3837 * And when try_charge() successfully returns, one refcnt to memcg without
3838 * struct page_cgroup is acquired. This refcnt will be consumed by
3839 * "commit()" or removed by "cancel()"
3841 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3844 struct mem_cgroup **memcgp)
3846 struct mem_cgroup *memcg;
3847 struct page_cgroup *pc;
3850 pc = lookup_page_cgroup(page);
3852 * Every swap fault against a single page tries to charge the
3853 * page, bail as early as possible. shmem_unuse() encounters
3854 * already charged pages, too. The USED bit is protected by
3855 * the page lock, which serializes swap cache removal, which
3856 * in turn serializes uncharging.
3858 if (PageCgroupUsed(pc))
3860 if (!do_swap_account)
3862 memcg = try_get_mem_cgroup_from_page(page);
3866 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3867 css_put(&memcg->css);
3872 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3878 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3879 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3882 if (mem_cgroup_disabled())
3885 * A racing thread's fault, or swapoff, may have already
3886 * updated the pte, and even removed page from swap cache: in
3887 * those cases unuse_pte()'s pte_same() test will fail; but
3888 * there's also a KSM case which does need to charge the page.
3890 if (!PageSwapCache(page)) {
3893 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3898 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3901 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3903 if (mem_cgroup_disabled())
3907 __mem_cgroup_cancel_charge(memcg, 1);
3911 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3912 enum charge_type ctype)
3914 if (mem_cgroup_disabled())
3919 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3921 * Now swap is on-memory. This means this page may be
3922 * counted both as mem and swap....double count.
3923 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3924 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3925 * may call delete_from_swap_cache() before reach here.
3927 if (do_swap_account && PageSwapCache(page)) {
3928 swp_entry_t ent = {.val = page_private(page)};
3929 mem_cgroup_uncharge_swap(ent);
3933 void mem_cgroup_commit_charge_swapin(struct page *page,
3934 struct mem_cgroup *memcg)
3936 __mem_cgroup_commit_charge_swapin(page, memcg,
3937 MEM_CGROUP_CHARGE_TYPE_ANON);
3940 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3943 struct mem_cgroup *memcg = NULL;
3944 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3947 if (mem_cgroup_disabled())
3949 if (PageCompound(page))
3952 if (!PageSwapCache(page))
3953 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3954 else { /* page is swapcache/shmem */
3955 ret = __mem_cgroup_try_charge_swapin(mm, page,
3958 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3963 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3964 unsigned int nr_pages,
3965 const enum charge_type ctype)
3967 struct memcg_batch_info *batch = NULL;
3968 bool uncharge_memsw = true;
3970 /* If swapout, usage of swap doesn't decrease */
3971 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3972 uncharge_memsw = false;
3974 batch = ¤t->memcg_batch;
3976 * In usual, we do css_get() when we remember memcg pointer.
3977 * But in this case, we keep res->usage until end of a series of
3978 * uncharges. Then, it's ok to ignore memcg's refcnt.
3981 batch->memcg = memcg;
3983 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3984 * In those cases, all pages freed continuously can be expected to be in
3985 * the same cgroup and we have chance to coalesce uncharges.
3986 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3987 * because we want to do uncharge as soon as possible.
3990 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3991 goto direct_uncharge;
3994 goto direct_uncharge;
3997 * In typical case, batch->memcg == mem. This means we can
3998 * merge a series of uncharges to an uncharge of res_counter.
3999 * If not, we uncharge res_counter ony by one.
4001 if (batch->memcg != memcg)
4002 goto direct_uncharge;
4003 /* remember freed charge and uncharge it later */
4006 batch->memsw_nr_pages++;
4009 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4011 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4012 if (unlikely(batch->memcg != memcg))
4013 memcg_oom_recover(memcg);
4017 * uncharge if !page_mapped(page)
4019 static struct mem_cgroup *
4020 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4023 struct mem_cgroup *memcg = NULL;
4024 unsigned int nr_pages = 1;
4025 struct page_cgroup *pc;
4028 if (mem_cgroup_disabled())
4031 if (PageTransHuge(page)) {
4032 nr_pages <<= compound_order(page);
4033 VM_BUG_ON(!PageTransHuge(page));
4036 * Check if our page_cgroup is valid
4038 pc = lookup_page_cgroup(page);
4039 if (unlikely(!PageCgroupUsed(pc)))
4042 lock_page_cgroup(pc);
4044 memcg = pc->mem_cgroup;
4046 if (!PageCgroupUsed(pc))
4049 anon = PageAnon(page);
4052 case MEM_CGROUP_CHARGE_TYPE_ANON:
4054 * Generally PageAnon tells if it's the anon statistics to be
4055 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4056 * used before page reached the stage of being marked PageAnon.
4060 case MEM_CGROUP_CHARGE_TYPE_DROP:
4061 /* See mem_cgroup_prepare_migration() */
4062 if (page_mapped(page))
4065 * Pages under migration may not be uncharged. But
4066 * end_migration() /must/ be the one uncharging the
4067 * unused post-migration page and so it has to call
4068 * here with the migration bit still set. See the
4069 * res_counter handling below.
4071 if (!end_migration && PageCgroupMigration(pc))
4074 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4075 if (!PageAnon(page)) { /* Shared memory */
4076 if (page->mapping && !page_is_file_cache(page))
4078 } else if (page_mapped(page)) /* Anon */
4085 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4087 ClearPageCgroupUsed(pc);
4089 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4090 * freed from LRU. This is safe because uncharged page is expected not
4091 * to be reused (freed soon). Exception is SwapCache, it's handled by
4092 * special functions.
4095 unlock_page_cgroup(pc);
4097 * even after unlock, we have memcg->res.usage here and this memcg
4098 * will never be freed, so it's safe to call css_get().
4100 memcg_check_events(memcg, page);
4101 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4102 mem_cgroup_swap_statistics(memcg, true);
4103 css_get(&memcg->css);
4106 * Migration does not charge the res_counter for the
4107 * replacement page, so leave it alone when phasing out the
4108 * page that is unused after the migration.
4110 if (!end_migration && !mem_cgroup_is_root(memcg))
4111 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4116 unlock_page_cgroup(pc);
4120 void mem_cgroup_uncharge_page(struct page *page)
4123 if (page_mapped(page))
4125 VM_BUG_ON(page->mapping && !PageAnon(page));
4127 * If the page is in swap cache, uncharge should be deferred
4128 * to the swap path, which also properly accounts swap usage
4129 * and handles memcg lifetime.
4131 * Note that this check is not stable and reclaim may add the
4132 * page to swap cache at any time after this. However, if the
4133 * page is not in swap cache by the time page->mapcount hits
4134 * 0, there won't be any page table references to the swap
4135 * slot, and reclaim will free it and not actually write the
4138 if (PageSwapCache(page))
4140 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4143 void mem_cgroup_uncharge_cache_page(struct page *page)
4145 VM_BUG_ON(page_mapped(page));
4146 VM_BUG_ON(page->mapping);
4147 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4151 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4152 * In that cases, pages are freed continuously and we can expect pages
4153 * are in the same memcg. All these calls itself limits the number of
4154 * pages freed at once, then uncharge_start/end() is called properly.
4155 * This may be called prural(2) times in a context,
4158 void mem_cgroup_uncharge_start(void)
4160 current->memcg_batch.do_batch++;
4161 /* We can do nest. */
4162 if (current->memcg_batch.do_batch == 1) {
4163 current->memcg_batch.memcg = NULL;
4164 current->memcg_batch.nr_pages = 0;
4165 current->memcg_batch.memsw_nr_pages = 0;
4169 void mem_cgroup_uncharge_end(void)
4171 struct memcg_batch_info *batch = ¤t->memcg_batch;
4173 if (!batch->do_batch)
4177 if (batch->do_batch) /* If stacked, do nothing. */
4183 * This "batch->memcg" is valid without any css_get/put etc...
4184 * bacause we hide charges behind us.
4186 if (batch->nr_pages)
4187 res_counter_uncharge(&batch->memcg->res,
4188 batch->nr_pages * PAGE_SIZE);
4189 if (batch->memsw_nr_pages)
4190 res_counter_uncharge(&batch->memcg->memsw,
4191 batch->memsw_nr_pages * PAGE_SIZE);
4192 memcg_oom_recover(batch->memcg);
4193 /* forget this pointer (for sanity check) */
4194 batch->memcg = NULL;
4199 * called after __delete_from_swap_cache() and drop "page" account.
4200 * memcg information is recorded to swap_cgroup of "ent"
4203 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4205 struct mem_cgroup *memcg;
4206 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4208 if (!swapout) /* this was a swap cache but the swap is unused ! */
4209 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4211 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4214 * record memcg information, if swapout && memcg != NULL,
4215 * css_get() was called in uncharge().
4217 if (do_swap_account && swapout && memcg)
4218 swap_cgroup_record(ent, css_id(&memcg->css));
4222 #ifdef CONFIG_MEMCG_SWAP
4224 * called from swap_entry_free(). remove record in swap_cgroup and
4225 * uncharge "memsw" account.
4227 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4229 struct mem_cgroup *memcg;
4232 if (!do_swap_account)
4235 id = swap_cgroup_record(ent, 0);
4237 memcg = mem_cgroup_lookup(id);
4240 * We uncharge this because swap is freed.
4241 * This memcg can be obsolete one. We avoid calling css_tryget
4243 if (!mem_cgroup_is_root(memcg))
4244 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4245 mem_cgroup_swap_statistics(memcg, false);
4246 css_put(&memcg->css);
4252 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4253 * @entry: swap entry to be moved
4254 * @from: mem_cgroup which the entry is moved from
4255 * @to: mem_cgroup which the entry is moved to
4257 * It succeeds only when the swap_cgroup's record for this entry is the same
4258 * as the mem_cgroup's id of @from.
4260 * Returns 0 on success, -EINVAL on failure.
4262 * The caller must have charged to @to, IOW, called res_counter_charge() about
4263 * both res and memsw, and called css_get().
4265 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4266 struct mem_cgroup *from, struct mem_cgroup *to)
4268 unsigned short old_id, new_id;
4270 old_id = css_id(&from->css);
4271 new_id = css_id(&to->css);
4273 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4274 mem_cgroup_swap_statistics(from, false);
4275 mem_cgroup_swap_statistics(to, true);
4277 * This function is only called from task migration context now.
4278 * It postpones res_counter and refcount handling till the end
4279 * of task migration(mem_cgroup_clear_mc()) for performance
4280 * improvement. But we cannot postpone css_get(to) because if
4281 * the process that has been moved to @to does swap-in, the
4282 * refcount of @to might be decreased to 0.
4284 * We are in attach() phase, so the cgroup is guaranteed to be
4285 * alive, so we can just call css_get().
4293 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4294 struct mem_cgroup *from, struct mem_cgroup *to)
4301 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4304 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4305 struct mem_cgroup **memcgp)
4307 struct mem_cgroup *memcg = NULL;
4308 unsigned int nr_pages = 1;
4309 struct page_cgroup *pc;
4310 enum charge_type ctype;
4314 if (mem_cgroup_disabled())
4317 if (PageTransHuge(page))
4318 nr_pages <<= compound_order(page);
4320 pc = lookup_page_cgroup(page);
4321 lock_page_cgroup(pc);
4322 if (PageCgroupUsed(pc)) {
4323 memcg = pc->mem_cgroup;
4324 css_get(&memcg->css);
4326 * At migrating an anonymous page, its mapcount goes down
4327 * to 0 and uncharge() will be called. But, even if it's fully
4328 * unmapped, migration may fail and this page has to be
4329 * charged again. We set MIGRATION flag here and delay uncharge
4330 * until end_migration() is called
4332 * Corner Case Thinking
4334 * When the old page was mapped as Anon and it's unmap-and-freed
4335 * while migration was ongoing.
4336 * If unmap finds the old page, uncharge() of it will be delayed
4337 * until end_migration(). If unmap finds a new page, it's
4338 * uncharged when it make mapcount to be 1->0. If unmap code
4339 * finds swap_migration_entry, the new page will not be mapped
4340 * and end_migration() will find it(mapcount==0).
4343 * When the old page was mapped but migraion fails, the kernel
4344 * remaps it. A charge for it is kept by MIGRATION flag even
4345 * if mapcount goes down to 0. We can do remap successfully
4346 * without charging it again.
4349 * The "old" page is under lock_page() until the end of
4350 * migration, so, the old page itself will not be swapped-out.
4351 * If the new page is swapped out before end_migraton, our
4352 * hook to usual swap-out path will catch the event.
4355 SetPageCgroupMigration(pc);
4357 unlock_page_cgroup(pc);
4359 * If the page is not charged at this point,
4367 * We charge new page before it's used/mapped. So, even if unlock_page()
4368 * is called before end_migration, we can catch all events on this new
4369 * page. In the case new page is migrated but not remapped, new page's
4370 * mapcount will be finally 0 and we call uncharge in end_migration().
4373 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4375 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4377 * The page is committed to the memcg, but it's not actually
4378 * charged to the res_counter since we plan on replacing the
4379 * old one and only one page is going to be left afterwards.
4381 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4384 /* remove redundant charge if migration failed*/
4385 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4386 struct page *oldpage, struct page *newpage, bool migration_ok)
4388 struct page *used, *unused;
4389 struct page_cgroup *pc;
4395 if (!migration_ok) {
4402 anon = PageAnon(used);
4403 __mem_cgroup_uncharge_common(unused,
4404 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4405 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4407 css_put(&memcg->css);
4409 * We disallowed uncharge of pages under migration because mapcount
4410 * of the page goes down to zero, temporarly.
4411 * Clear the flag and check the page should be charged.
4413 pc = lookup_page_cgroup(oldpage);
4414 lock_page_cgroup(pc);
4415 ClearPageCgroupMigration(pc);
4416 unlock_page_cgroup(pc);
4419 * If a page is a file cache, radix-tree replacement is very atomic
4420 * and we can skip this check. When it was an Anon page, its mapcount
4421 * goes down to 0. But because we added MIGRATION flage, it's not
4422 * uncharged yet. There are several case but page->mapcount check
4423 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4424 * check. (see prepare_charge() also)
4427 mem_cgroup_uncharge_page(used);
4431 * At replace page cache, newpage is not under any memcg but it's on
4432 * LRU. So, this function doesn't touch res_counter but handles LRU
4433 * in correct way. Both pages are locked so we cannot race with uncharge.
4435 void mem_cgroup_replace_page_cache(struct page *oldpage,
4436 struct page *newpage)
4438 struct mem_cgroup *memcg = NULL;
4439 struct page_cgroup *pc;
4440 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4442 if (mem_cgroup_disabled())
4445 pc = lookup_page_cgroup(oldpage);
4446 /* fix accounting on old pages */
4447 lock_page_cgroup(pc);
4448 if (PageCgroupUsed(pc)) {
4449 memcg = pc->mem_cgroup;
4450 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4451 ClearPageCgroupUsed(pc);
4453 unlock_page_cgroup(pc);
4456 * When called from shmem_replace_page(), in some cases the
4457 * oldpage has already been charged, and in some cases not.
4462 * Even if newpage->mapping was NULL before starting replacement,
4463 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4464 * LRU while we overwrite pc->mem_cgroup.
4466 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4469 #ifdef CONFIG_DEBUG_VM
4470 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4472 struct page_cgroup *pc;
4474 pc = lookup_page_cgroup(page);
4476 * Can be NULL while feeding pages into the page allocator for
4477 * the first time, i.e. during boot or memory hotplug;
4478 * or when mem_cgroup_disabled().
4480 if (likely(pc) && PageCgroupUsed(pc))
4485 bool mem_cgroup_bad_page_check(struct page *page)
4487 if (mem_cgroup_disabled())
4490 return lookup_page_cgroup_used(page) != NULL;
4493 void mem_cgroup_print_bad_page(struct page *page)
4495 struct page_cgroup *pc;
4497 pc = lookup_page_cgroup_used(page);
4499 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4500 pc, pc->flags, pc->mem_cgroup);
4505 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4506 unsigned long long val)
4509 u64 memswlimit, memlimit;
4511 int children = mem_cgroup_count_children(memcg);
4512 u64 curusage, oldusage;
4516 * For keeping hierarchical_reclaim simple, how long we should retry
4517 * is depends on callers. We set our retry-count to be function
4518 * of # of children which we should visit in this loop.
4520 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4522 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4525 while (retry_count) {
4526 if (signal_pending(current)) {
4531 * Rather than hide all in some function, I do this in
4532 * open coded manner. You see what this really does.
4533 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4535 mutex_lock(&set_limit_mutex);
4536 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4537 if (memswlimit < val) {
4539 mutex_unlock(&set_limit_mutex);
4543 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4547 ret = res_counter_set_limit(&memcg->res, val);
4549 if (memswlimit == val)
4550 memcg->memsw_is_minimum = true;
4552 memcg->memsw_is_minimum = false;
4554 mutex_unlock(&set_limit_mutex);
4559 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4560 MEM_CGROUP_RECLAIM_SHRINK);
4561 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4562 /* Usage is reduced ? */
4563 if (curusage >= oldusage)
4566 oldusage = curusage;
4568 if (!ret && enlarge)
4569 memcg_oom_recover(memcg);
4574 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4575 unsigned long long val)
4578 u64 memlimit, memswlimit, oldusage, curusage;
4579 int children = mem_cgroup_count_children(memcg);
4583 /* see mem_cgroup_resize_res_limit */
4584 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4585 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4586 while (retry_count) {
4587 if (signal_pending(current)) {
4592 * Rather than hide all in some function, I do this in
4593 * open coded manner. You see what this really does.
4594 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4596 mutex_lock(&set_limit_mutex);
4597 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4598 if (memlimit > val) {
4600 mutex_unlock(&set_limit_mutex);
4603 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4604 if (memswlimit < val)
4606 ret = res_counter_set_limit(&memcg->memsw, val);
4608 if (memlimit == val)
4609 memcg->memsw_is_minimum = true;
4611 memcg->memsw_is_minimum = false;
4613 mutex_unlock(&set_limit_mutex);
4618 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4619 MEM_CGROUP_RECLAIM_NOSWAP |
4620 MEM_CGROUP_RECLAIM_SHRINK);
4621 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4622 /* Usage is reduced ? */
4623 if (curusage >= oldusage)
4626 oldusage = curusage;
4628 if (!ret && enlarge)
4629 memcg_oom_recover(memcg);
4634 * mem_cgroup_force_empty_list - clears LRU of a group
4635 * @memcg: group to clear
4638 * @lru: lru to to clear
4640 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4641 * reclaim the pages page themselves - pages are moved to the parent (or root)
4644 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4645 int node, int zid, enum lru_list lru)
4647 struct lruvec *lruvec;
4648 unsigned long flags;
4649 struct list_head *list;
4653 zone = &NODE_DATA(node)->node_zones[zid];
4654 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4655 list = &lruvec->lists[lru];
4659 struct page_cgroup *pc;
4662 spin_lock_irqsave(&zone->lru_lock, flags);
4663 if (list_empty(list)) {
4664 spin_unlock_irqrestore(&zone->lru_lock, flags);
4667 page = list_entry(list->prev, struct page, lru);
4669 list_move(&page->lru, list);
4671 spin_unlock_irqrestore(&zone->lru_lock, flags);
4674 spin_unlock_irqrestore(&zone->lru_lock, flags);
4676 pc = lookup_page_cgroup(page);
4678 if (mem_cgroup_move_parent(page, pc, memcg)) {
4679 /* found lock contention or "pc" is obsolete. */
4684 } while (!list_empty(list));
4688 * make mem_cgroup's charge to be 0 if there is no task by moving
4689 * all the charges and pages to the parent.
4690 * This enables deleting this mem_cgroup.
4692 * Caller is responsible for holding css reference on the memcg.
4694 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4700 /* This is for making all *used* pages to be on LRU. */
4701 lru_add_drain_all();
4702 drain_all_stock_sync(memcg);
4703 mem_cgroup_start_move(memcg);
4704 for_each_node_state(node, N_MEMORY) {
4705 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4708 mem_cgroup_force_empty_list(memcg,
4713 mem_cgroup_end_move(memcg);
4714 memcg_oom_recover(memcg);
4718 * Kernel memory may not necessarily be trackable to a specific
4719 * process. So they are not migrated, and therefore we can't
4720 * expect their value to drop to 0 here.
4721 * Having res filled up with kmem only is enough.
4723 * This is a safety check because mem_cgroup_force_empty_list
4724 * could have raced with mem_cgroup_replace_page_cache callers
4725 * so the lru seemed empty but the page could have been added
4726 * right after the check. RES_USAGE should be safe as we always
4727 * charge before adding to the LRU.
4729 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4730 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4731 } while (usage > 0);
4735 * This mainly exists for tests during the setting of set of use_hierarchy.
4736 * Since this is the very setting we are changing, the current hierarchy value
4739 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4741 struct cgroup_subsys_state *pos;
4743 /* bounce at first found */
4744 css_for_each_child(pos, &memcg->css)
4750 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4751 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4752 * from mem_cgroup_count_children(), in the sense that we don't really care how
4753 * many children we have; we only need to know if we have any. It also counts
4754 * any memcg without hierarchy as infertile.
4756 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4758 return memcg->use_hierarchy && __memcg_has_children(memcg);
4762 * Reclaims as many pages from the given memcg as possible and moves
4763 * the rest to the parent.
4765 * Caller is responsible for holding css reference for memcg.
4767 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4769 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4770 struct cgroup *cgrp = memcg->css.cgroup;
4772 /* returns EBUSY if there is a task or if we come here twice. */
4773 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4776 /* we call try-to-free pages for make this cgroup empty */
4777 lru_add_drain_all();
4778 /* try to free all pages in this cgroup */
4779 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4782 if (signal_pending(current))
4785 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4789 /* maybe some writeback is necessary */
4790 congestion_wait(BLK_RW_ASYNC, HZ/10);
4795 mem_cgroup_reparent_charges(memcg);
4800 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4803 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4805 if (mem_cgroup_is_root(memcg))
4807 return mem_cgroup_force_empty(memcg);
4810 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4813 return mem_cgroup_from_css(css)->use_hierarchy;
4816 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4817 struct cftype *cft, u64 val)
4820 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4821 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4823 mutex_lock(&memcg_create_mutex);
4825 if (memcg->use_hierarchy == val)
4829 * If parent's use_hierarchy is set, we can't make any modifications
4830 * in the child subtrees. If it is unset, then the change can
4831 * occur, provided the current cgroup has no children.
4833 * For the root cgroup, parent_mem is NULL, we allow value to be
4834 * set if there are no children.
4836 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4837 (val == 1 || val == 0)) {
4838 if (!__memcg_has_children(memcg))
4839 memcg->use_hierarchy = val;
4846 mutex_unlock(&memcg_create_mutex);
4852 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4853 enum mem_cgroup_stat_index idx)
4855 struct mem_cgroup *iter;
4858 /* Per-cpu values can be negative, use a signed accumulator */
4859 for_each_mem_cgroup_tree(iter, memcg)
4860 val += mem_cgroup_read_stat(iter, idx);
4862 if (val < 0) /* race ? */
4867 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4871 if (!mem_cgroup_is_root(memcg)) {
4873 return res_counter_read_u64(&memcg->res, RES_USAGE);
4875 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4879 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4880 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4882 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4883 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4886 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4888 return val << PAGE_SHIFT;
4891 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
4892 struct cftype *cft, struct file *file,
4893 char __user *buf, size_t nbytes, loff_t *ppos)
4895 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4901 type = MEMFILE_TYPE(cft->private);
4902 name = MEMFILE_ATTR(cft->private);
4906 if (name == RES_USAGE)
4907 val = mem_cgroup_usage(memcg, false);
4909 val = res_counter_read_u64(&memcg->res, name);
4912 if (name == RES_USAGE)
4913 val = mem_cgroup_usage(memcg, true);
4915 val = res_counter_read_u64(&memcg->memsw, name);
4918 val = res_counter_read_u64(&memcg->kmem, name);
4924 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4925 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4928 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
4931 #ifdef CONFIG_MEMCG_KMEM
4932 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4934 * For simplicity, we won't allow this to be disabled. It also can't
4935 * be changed if the cgroup has children already, or if tasks had
4938 * If tasks join before we set the limit, a person looking at
4939 * kmem.usage_in_bytes will have no way to determine when it took
4940 * place, which makes the value quite meaningless.
4942 * After it first became limited, changes in the value of the limit are
4943 * of course permitted.
4945 mutex_lock(&memcg_create_mutex);
4946 mutex_lock(&set_limit_mutex);
4947 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
4948 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
4952 ret = res_counter_set_limit(&memcg->kmem, val);
4955 ret = memcg_update_cache_sizes(memcg);
4957 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
4960 static_key_slow_inc(&memcg_kmem_enabled_key);
4962 * setting the active bit after the inc will guarantee no one
4963 * starts accounting before all call sites are patched
4965 memcg_kmem_set_active(memcg);
4967 ret = res_counter_set_limit(&memcg->kmem, val);
4969 mutex_unlock(&set_limit_mutex);
4970 mutex_unlock(&memcg_create_mutex);
4975 #ifdef CONFIG_MEMCG_KMEM
4976 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4979 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4983 memcg->kmem_account_flags = parent->kmem_account_flags;
4985 * When that happen, we need to disable the static branch only on those
4986 * memcgs that enabled it. To achieve this, we would be forced to
4987 * complicate the code by keeping track of which memcgs were the ones
4988 * that actually enabled limits, and which ones got it from its
4991 * It is a lot simpler just to do static_key_slow_inc() on every child
4992 * that is accounted.
4994 if (!memcg_kmem_is_active(memcg))
4998 * __mem_cgroup_free() will issue static_key_slow_dec() because this
4999 * memcg is active already. If the later initialization fails then the
5000 * cgroup core triggers the cleanup so we do not have to do it here.
5002 static_key_slow_inc(&memcg_kmem_enabled_key);
5004 mutex_lock(&set_limit_mutex);
5005 memcg_stop_kmem_account();
5006 ret = memcg_update_cache_sizes(memcg);
5007 memcg_resume_kmem_account();
5008 mutex_unlock(&set_limit_mutex);
5012 #endif /* CONFIG_MEMCG_KMEM */
5015 * The user of this function is...
5018 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5021 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5024 unsigned long long val;
5027 type = MEMFILE_TYPE(cft->private);
5028 name = MEMFILE_ATTR(cft->private);
5032 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5036 /* This function does all necessary parse...reuse it */
5037 ret = res_counter_memparse_write_strategy(buffer, &val);
5041 ret = mem_cgroup_resize_limit(memcg, val);
5042 else if (type == _MEMSWAP)
5043 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5044 else if (type == _KMEM)
5045 ret = memcg_update_kmem_limit(css, val);
5049 case RES_SOFT_LIMIT:
5050 ret = res_counter_memparse_write_strategy(buffer, &val);
5054 * For memsw, soft limits are hard to implement in terms
5055 * of semantics, for now, we support soft limits for
5056 * control without swap
5059 ret = res_counter_set_soft_limit(&memcg->res, val);
5064 ret = -EINVAL; /* should be BUG() ? */
5070 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5071 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5073 unsigned long long min_limit, min_memsw_limit, tmp;
5075 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5076 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5077 if (!memcg->use_hierarchy)
5080 while (css_parent(&memcg->css)) {
5081 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5082 if (!memcg->use_hierarchy)
5084 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5085 min_limit = min(min_limit, tmp);
5086 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5087 min_memsw_limit = min(min_memsw_limit, tmp);
5090 *mem_limit = min_limit;
5091 *memsw_limit = min_memsw_limit;
5094 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5096 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5100 type = MEMFILE_TYPE(event);
5101 name = MEMFILE_ATTR(event);
5106 res_counter_reset_max(&memcg->res);
5107 else if (type == _MEMSWAP)
5108 res_counter_reset_max(&memcg->memsw);
5109 else if (type == _KMEM)
5110 res_counter_reset_max(&memcg->kmem);
5116 res_counter_reset_failcnt(&memcg->res);
5117 else if (type == _MEMSWAP)
5118 res_counter_reset_failcnt(&memcg->memsw);
5119 else if (type == _KMEM)
5120 res_counter_reset_failcnt(&memcg->kmem);
5129 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5132 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5136 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5137 struct cftype *cft, u64 val)
5139 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5141 if (val >= (1 << NR_MOVE_TYPE))
5145 * No kind of locking is needed in here, because ->can_attach() will
5146 * check this value once in the beginning of the process, and then carry
5147 * on with stale data. This means that changes to this value will only
5148 * affect task migrations starting after the change.
5150 memcg->move_charge_at_immigrate = val;
5154 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5155 struct cftype *cft, u64 val)
5162 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5163 struct cftype *cft, struct seq_file *m)
5166 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5167 unsigned long node_nr;
5168 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5170 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5171 seq_printf(m, "total=%lu", total_nr);
5172 for_each_node_state(nid, N_MEMORY) {
5173 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5174 seq_printf(m, " N%d=%lu", nid, node_nr);
5178 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5179 seq_printf(m, "file=%lu", file_nr);
5180 for_each_node_state(nid, N_MEMORY) {
5181 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5183 seq_printf(m, " N%d=%lu", nid, node_nr);
5187 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5188 seq_printf(m, "anon=%lu", anon_nr);
5189 for_each_node_state(nid, N_MEMORY) {
5190 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5192 seq_printf(m, " N%d=%lu", nid, node_nr);
5196 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5197 seq_printf(m, "unevictable=%lu", unevictable_nr);
5198 for_each_node_state(nid, N_MEMORY) {
5199 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5200 BIT(LRU_UNEVICTABLE));
5201 seq_printf(m, " N%d=%lu", nid, node_nr);
5206 #endif /* CONFIG_NUMA */
5208 static inline void mem_cgroup_lru_names_not_uptodate(void)
5210 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5213 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5216 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5217 struct mem_cgroup *mi;
5220 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5221 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5223 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5224 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5227 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5228 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5229 mem_cgroup_read_events(memcg, i));
5231 for (i = 0; i < NR_LRU_LISTS; i++)
5232 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5233 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5235 /* Hierarchical information */
5237 unsigned long long limit, memsw_limit;
5238 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5239 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5240 if (do_swap_account)
5241 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5245 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5248 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5250 for_each_mem_cgroup_tree(mi, memcg)
5251 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5252 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5255 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5256 unsigned long long val = 0;
5258 for_each_mem_cgroup_tree(mi, memcg)
5259 val += mem_cgroup_read_events(mi, i);
5260 seq_printf(m, "total_%s %llu\n",
5261 mem_cgroup_events_names[i], val);
5264 for (i = 0; i < NR_LRU_LISTS; i++) {
5265 unsigned long long val = 0;
5267 for_each_mem_cgroup_tree(mi, memcg)
5268 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5269 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5272 #ifdef CONFIG_DEBUG_VM
5275 struct mem_cgroup_per_zone *mz;
5276 struct zone_reclaim_stat *rstat;
5277 unsigned long recent_rotated[2] = {0, 0};
5278 unsigned long recent_scanned[2] = {0, 0};
5280 for_each_online_node(nid)
5281 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5282 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5283 rstat = &mz->lruvec.reclaim_stat;
5285 recent_rotated[0] += rstat->recent_rotated[0];
5286 recent_rotated[1] += rstat->recent_rotated[1];
5287 recent_scanned[0] += rstat->recent_scanned[0];
5288 recent_scanned[1] += rstat->recent_scanned[1];
5290 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5291 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5292 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5293 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5300 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5303 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5305 return mem_cgroup_swappiness(memcg);
5308 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5309 struct cftype *cft, u64 val)
5311 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5312 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5314 if (val > 100 || !parent)
5317 mutex_lock(&memcg_create_mutex);
5319 /* If under hierarchy, only empty-root can set this value */
5320 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5321 mutex_unlock(&memcg_create_mutex);
5325 memcg->swappiness = val;
5327 mutex_unlock(&memcg_create_mutex);
5332 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5334 struct mem_cgroup_threshold_ary *t;
5340 t = rcu_dereference(memcg->thresholds.primary);
5342 t = rcu_dereference(memcg->memsw_thresholds.primary);
5347 usage = mem_cgroup_usage(memcg, swap);
5350 * current_threshold points to threshold just below or equal to usage.
5351 * If it's not true, a threshold was crossed after last
5352 * call of __mem_cgroup_threshold().
5354 i = t->current_threshold;
5357 * Iterate backward over array of thresholds starting from
5358 * current_threshold and check if a threshold is crossed.
5359 * If none of thresholds below usage is crossed, we read
5360 * only one element of the array here.
5362 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5363 eventfd_signal(t->entries[i].eventfd, 1);
5365 /* i = current_threshold + 1 */
5369 * Iterate forward over array of thresholds starting from
5370 * current_threshold+1 and check if a threshold is crossed.
5371 * If none of thresholds above usage is crossed, we read
5372 * only one element of the array here.
5374 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5375 eventfd_signal(t->entries[i].eventfd, 1);
5377 /* Update current_threshold */
5378 t->current_threshold = i - 1;
5383 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5386 __mem_cgroup_threshold(memcg, false);
5387 if (do_swap_account)
5388 __mem_cgroup_threshold(memcg, true);
5390 memcg = parent_mem_cgroup(memcg);
5394 static int compare_thresholds(const void *a, const void *b)
5396 const struct mem_cgroup_threshold *_a = a;
5397 const struct mem_cgroup_threshold *_b = b;
5399 if (_a->threshold > _b->threshold)
5402 if (_a->threshold < _b->threshold)
5408 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5410 struct mem_cgroup_eventfd_list *ev;
5412 list_for_each_entry(ev, &memcg->oom_notify, list)
5413 eventfd_signal(ev->eventfd, 1);
5417 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5419 struct mem_cgroup *iter;
5421 for_each_mem_cgroup_tree(iter, memcg)
5422 mem_cgroup_oom_notify_cb(iter);
5425 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5426 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5428 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5429 struct mem_cgroup_thresholds *thresholds;
5430 struct mem_cgroup_threshold_ary *new;
5431 enum res_type type = MEMFILE_TYPE(cft->private);
5432 u64 threshold, usage;
5435 ret = res_counter_memparse_write_strategy(args, &threshold);
5439 mutex_lock(&memcg->thresholds_lock);
5442 thresholds = &memcg->thresholds;
5443 else if (type == _MEMSWAP)
5444 thresholds = &memcg->memsw_thresholds;
5448 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5450 /* Check if a threshold crossed before adding a new one */
5451 if (thresholds->primary)
5452 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5454 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5456 /* Allocate memory for new array of thresholds */
5457 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5465 /* Copy thresholds (if any) to new array */
5466 if (thresholds->primary) {
5467 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5468 sizeof(struct mem_cgroup_threshold));
5471 /* Add new threshold */
5472 new->entries[size - 1].eventfd = eventfd;
5473 new->entries[size - 1].threshold = threshold;
5475 /* Sort thresholds. Registering of new threshold isn't time-critical */
5476 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5477 compare_thresholds, NULL);
5479 /* Find current threshold */
5480 new->current_threshold = -1;
5481 for (i = 0; i < size; i++) {
5482 if (new->entries[i].threshold <= usage) {
5484 * new->current_threshold will not be used until
5485 * rcu_assign_pointer(), so it's safe to increment
5488 ++new->current_threshold;
5493 /* Free old spare buffer and save old primary buffer as spare */
5494 kfree(thresholds->spare);
5495 thresholds->spare = thresholds->primary;
5497 rcu_assign_pointer(thresholds->primary, new);
5499 /* To be sure that nobody uses thresholds */
5503 mutex_unlock(&memcg->thresholds_lock);
5508 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5509 struct cftype *cft, struct eventfd_ctx *eventfd)
5511 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5512 struct mem_cgroup_thresholds *thresholds;
5513 struct mem_cgroup_threshold_ary *new;
5514 enum res_type type = MEMFILE_TYPE(cft->private);
5518 mutex_lock(&memcg->thresholds_lock);
5520 thresholds = &memcg->thresholds;
5521 else if (type == _MEMSWAP)
5522 thresholds = &memcg->memsw_thresholds;
5526 if (!thresholds->primary)
5529 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5531 /* Check if a threshold crossed before removing */
5532 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5534 /* Calculate new number of threshold */
5536 for (i = 0; i < thresholds->primary->size; i++) {
5537 if (thresholds->primary->entries[i].eventfd != eventfd)
5541 new = thresholds->spare;
5543 /* Set thresholds array to NULL if we don't have thresholds */
5552 /* Copy thresholds and find current threshold */
5553 new->current_threshold = -1;
5554 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5555 if (thresholds->primary->entries[i].eventfd == eventfd)
5558 new->entries[j] = thresholds->primary->entries[i];
5559 if (new->entries[j].threshold <= usage) {
5561 * new->current_threshold will not be used
5562 * until rcu_assign_pointer(), so it's safe to increment
5565 ++new->current_threshold;
5571 /* Swap primary and spare array */
5572 thresholds->spare = thresholds->primary;
5573 /* If all events are unregistered, free the spare array */
5575 kfree(thresholds->spare);
5576 thresholds->spare = NULL;
5579 rcu_assign_pointer(thresholds->primary, new);
5581 /* To be sure that nobody uses thresholds */
5584 mutex_unlock(&memcg->thresholds_lock);
5587 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5588 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5590 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5591 struct mem_cgroup_eventfd_list *event;
5592 enum res_type type = MEMFILE_TYPE(cft->private);
5594 BUG_ON(type != _OOM_TYPE);
5595 event = kmalloc(sizeof(*event), GFP_KERNEL);
5599 spin_lock(&memcg_oom_lock);
5601 event->eventfd = eventfd;
5602 list_add(&event->list, &memcg->oom_notify);
5604 /* already in OOM ? */
5605 if (atomic_read(&memcg->under_oom))
5606 eventfd_signal(eventfd, 1);
5607 spin_unlock(&memcg_oom_lock);
5612 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5613 struct cftype *cft, struct eventfd_ctx *eventfd)
5615 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5616 struct mem_cgroup_eventfd_list *ev, *tmp;
5617 enum res_type type = MEMFILE_TYPE(cft->private);
5619 BUG_ON(type != _OOM_TYPE);
5621 spin_lock(&memcg_oom_lock);
5623 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5624 if (ev->eventfd == eventfd) {
5625 list_del(&ev->list);
5630 spin_unlock(&memcg_oom_lock);
5633 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5634 struct cftype *cft, struct cgroup_map_cb *cb)
5636 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5638 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5640 if (atomic_read(&memcg->under_oom))
5641 cb->fill(cb, "under_oom", 1);
5643 cb->fill(cb, "under_oom", 0);
5647 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5648 struct cftype *cft, u64 val)
5650 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5651 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5653 /* cannot set to root cgroup and only 0 and 1 are allowed */
5654 if (!parent || !((val == 0) || (val == 1)))
5657 mutex_lock(&memcg_create_mutex);
5658 /* oom-kill-disable is a flag for subhierarchy. */
5659 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5660 mutex_unlock(&memcg_create_mutex);
5663 memcg->oom_kill_disable = val;
5665 memcg_oom_recover(memcg);
5666 mutex_unlock(&memcg_create_mutex);
5670 #ifdef CONFIG_MEMCG_KMEM
5671 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5675 memcg->kmemcg_id = -1;
5676 ret = memcg_propagate_kmem(memcg);
5680 return mem_cgroup_sockets_init(memcg, ss);
5683 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5685 mem_cgroup_sockets_destroy(memcg);
5688 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5690 if (!memcg_kmem_is_active(memcg))
5694 * kmem charges can outlive the cgroup. In the case of slab
5695 * pages, for instance, a page contain objects from various
5696 * processes. As we prevent from taking a reference for every
5697 * such allocation we have to be careful when doing uncharge
5698 * (see memcg_uncharge_kmem) and here during offlining.
5700 * The idea is that that only the _last_ uncharge which sees
5701 * the dead memcg will drop the last reference. An additional
5702 * reference is taken here before the group is marked dead
5703 * which is then paired with css_put during uncharge resp. here.
5705 * Although this might sound strange as this path is called from
5706 * css_offline() when the referencemight have dropped down to 0
5707 * and shouldn't be incremented anymore (css_tryget would fail)
5708 * we do not have other options because of the kmem allocations
5711 css_get(&memcg->css);
5713 memcg_kmem_mark_dead(memcg);
5715 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5718 if (memcg_kmem_test_and_clear_dead(memcg))
5719 css_put(&memcg->css);
5722 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5727 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5731 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5736 static struct cftype mem_cgroup_files[] = {
5738 .name = "usage_in_bytes",
5739 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5740 .read = mem_cgroup_read,
5741 .register_event = mem_cgroup_usage_register_event,
5742 .unregister_event = mem_cgroup_usage_unregister_event,
5745 .name = "max_usage_in_bytes",
5746 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5747 .trigger = mem_cgroup_reset,
5748 .read = mem_cgroup_read,
5751 .name = "limit_in_bytes",
5752 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5753 .write_string = mem_cgroup_write,
5754 .read = mem_cgroup_read,
5757 .name = "soft_limit_in_bytes",
5758 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5759 .write_string = mem_cgroup_write,
5760 .read = mem_cgroup_read,
5764 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5765 .trigger = mem_cgroup_reset,
5766 .read = mem_cgroup_read,
5770 .read_seq_string = memcg_stat_show,
5773 .name = "force_empty",
5774 .trigger = mem_cgroup_force_empty_write,
5777 .name = "use_hierarchy",
5778 .flags = CFTYPE_INSANE,
5779 .write_u64 = mem_cgroup_hierarchy_write,
5780 .read_u64 = mem_cgroup_hierarchy_read,
5783 .name = "swappiness",
5784 .read_u64 = mem_cgroup_swappiness_read,
5785 .write_u64 = mem_cgroup_swappiness_write,
5788 .name = "move_charge_at_immigrate",
5789 .read_u64 = mem_cgroup_move_charge_read,
5790 .write_u64 = mem_cgroup_move_charge_write,
5793 .name = "oom_control",
5794 .read_map = mem_cgroup_oom_control_read,
5795 .write_u64 = mem_cgroup_oom_control_write,
5796 .register_event = mem_cgroup_oom_register_event,
5797 .unregister_event = mem_cgroup_oom_unregister_event,
5798 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5801 .name = "pressure_level",
5802 .register_event = vmpressure_register_event,
5803 .unregister_event = vmpressure_unregister_event,
5807 .name = "numa_stat",
5808 .read_seq_string = memcg_numa_stat_show,
5811 #ifdef CONFIG_MEMCG_KMEM
5813 .name = "kmem.limit_in_bytes",
5814 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5815 .write_string = mem_cgroup_write,
5816 .read = mem_cgroup_read,
5819 .name = "kmem.usage_in_bytes",
5820 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5821 .read = mem_cgroup_read,
5824 .name = "kmem.failcnt",
5825 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5826 .trigger = mem_cgroup_reset,
5827 .read = mem_cgroup_read,
5830 .name = "kmem.max_usage_in_bytes",
5831 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5832 .trigger = mem_cgroup_reset,
5833 .read = mem_cgroup_read,
5835 #ifdef CONFIG_SLABINFO
5837 .name = "kmem.slabinfo",
5838 .read_seq_string = mem_cgroup_slabinfo_read,
5842 { }, /* terminate */
5845 #ifdef CONFIG_MEMCG_SWAP
5846 static struct cftype memsw_cgroup_files[] = {
5848 .name = "memsw.usage_in_bytes",
5849 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5850 .read = mem_cgroup_read,
5851 .register_event = mem_cgroup_usage_register_event,
5852 .unregister_event = mem_cgroup_usage_unregister_event,
5855 .name = "memsw.max_usage_in_bytes",
5856 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5857 .trigger = mem_cgroup_reset,
5858 .read = mem_cgroup_read,
5861 .name = "memsw.limit_in_bytes",
5862 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5863 .write_string = mem_cgroup_write,
5864 .read = mem_cgroup_read,
5867 .name = "memsw.failcnt",
5868 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5869 .trigger = mem_cgroup_reset,
5870 .read = mem_cgroup_read,
5872 { }, /* terminate */
5875 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5877 struct mem_cgroup_per_node *pn;
5878 struct mem_cgroup_per_zone *mz;
5879 int zone, tmp = node;
5881 * This routine is called against possible nodes.
5882 * But it's BUG to call kmalloc() against offline node.
5884 * TODO: this routine can waste much memory for nodes which will
5885 * never be onlined. It's better to use memory hotplug callback
5888 if (!node_state(node, N_NORMAL_MEMORY))
5890 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5894 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5895 mz = &pn->zoneinfo[zone];
5896 lruvec_init(&mz->lruvec);
5899 memcg->nodeinfo[node] = pn;
5903 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5905 kfree(memcg->nodeinfo[node]);
5908 static struct mem_cgroup *mem_cgroup_alloc(void)
5910 struct mem_cgroup *memcg;
5911 size_t size = memcg_size();
5913 /* Can be very big if nr_node_ids is very big */
5914 if (size < PAGE_SIZE)
5915 memcg = kzalloc(size, GFP_KERNEL);
5917 memcg = vzalloc(size);
5922 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5925 spin_lock_init(&memcg->pcp_counter_lock);
5929 if (size < PAGE_SIZE)
5937 * At destroying mem_cgroup, references from swap_cgroup can remain.
5938 * (scanning all at force_empty is too costly...)
5940 * Instead of clearing all references at force_empty, we remember
5941 * the number of reference from swap_cgroup and free mem_cgroup when
5942 * it goes down to 0.
5944 * Removal of cgroup itself succeeds regardless of refs from swap.
5947 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5950 size_t size = memcg_size();
5952 free_css_id(&mem_cgroup_subsys, &memcg->css);
5955 free_mem_cgroup_per_zone_info(memcg, node);
5957 free_percpu(memcg->stat);
5960 * We need to make sure that (at least for now), the jump label
5961 * destruction code runs outside of the cgroup lock. This is because
5962 * get_online_cpus(), which is called from the static_branch update,
5963 * can't be called inside the cgroup_lock. cpusets are the ones
5964 * enforcing this dependency, so if they ever change, we might as well.
5966 * schedule_work() will guarantee this happens. Be careful if you need
5967 * to move this code around, and make sure it is outside
5970 disarm_static_keys(memcg);
5971 if (size < PAGE_SIZE)
5978 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5980 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5982 if (!memcg->res.parent)
5984 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5986 EXPORT_SYMBOL(parent_mem_cgroup);
5988 static struct cgroup_subsys_state * __ref
5989 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5991 struct mem_cgroup *memcg;
5992 long error = -ENOMEM;
5995 memcg = mem_cgroup_alloc();
5997 return ERR_PTR(error);
6000 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6004 if (parent_css == NULL) {
6005 root_mem_cgroup = memcg;
6006 res_counter_init(&memcg->res, NULL);
6007 res_counter_init(&memcg->memsw, NULL);
6008 res_counter_init(&memcg->kmem, NULL);
6011 memcg->last_scanned_node = MAX_NUMNODES;
6012 INIT_LIST_HEAD(&memcg->oom_notify);
6013 memcg->move_charge_at_immigrate = 0;
6014 mutex_init(&memcg->thresholds_lock);
6015 spin_lock_init(&memcg->move_lock);
6016 vmpressure_init(&memcg->vmpressure);
6017 spin_lock_init(&memcg->soft_lock);
6022 __mem_cgroup_free(memcg);
6023 return ERR_PTR(error);
6027 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6029 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6030 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6036 mutex_lock(&memcg_create_mutex);
6038 memcg->use_hierarchy = parent->use_hierarchy;
6039 memcg->oom_kill_disable = parent->oom_kill_disable;
6040 memcg->swappiness = mem_cgroup_swappiness(parent);
6042 if (parent->use_hierarchy) {
6043 res_counter_init(&memcg->res, &parent->res);
6044 res_counter_init(&memcg->memsw, &parent->memsw);
6045 res_counter_init(&memcg->kmem, &parent->kmem);
6048 * No need to take a reference to the parent because cgroup
6049 * core guarantees its existence.
6052 res_counter_init(&memcg->res, NULL);
6053 res_counter_init(&memcg->memsw, NULL);
6054 res_counter_init(&memcg->kmem, NULL);
6056 * Deeper hierachy with use_hierarchy == false doesn't make
6057 * much sense so let cgroup subsystem know about this
6058 * unfortunate state in our controller.
6060 if (parent != root_mem_cgroup)
6061 mem_cgroup_subsys.broken_hierarchy = true;
6064 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6065 mutex_unlock(&memcg_create_mutex);
6070 * Announce all parents that a group from their hierarchy is gone.
6072 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6074 struct mem_cgroup *parent = memcg;
6076 while ((parent = parent_mem_cgroup(parent)))
6077 mem_cgroup_iter_invalidate(parent);
6080 * if the root memcg is not hierarchical we have to check it
6083 if (!root_mem_cgroup->use_hierarchy)
6084 mem_cgroup_iter_invalidate(root_mem_cgroup);
6087 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6089 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6091 kmem_cgroup_css_offline(memcg);
6093 mem_cgroup_invalidate_reclaim_iterators(memcg);
6094 mem_cgroup_reparent_charges(memcg);
6095 if (memcg->soft_contributed) {
6096 while ((memcg = parent_mem_cgroup(memcg)))
6097 atomic_dec(&memcg->children_in_excess);
6099 if (memcg != root_mem_cgroup && !root_mem_cgroup->use_hierarchy)
6100 atomic_dec(&root_mem_cgroup->children_in_excess);
6102 mem_cgroup_destroy_all_caches(memcg);
6103 vmpressure_cleanup(&memcg->vmpressure);
6106 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6108 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6110 memcg_destroy_kmem(memcg);
6111 __mem_cgroup_free(memcg);
6115 /* Handlers for move charge at task migration. */
6116 #define PRECHARGE_COUNT_AT_ONCE 256
6117 static int mem_cgroup_do_precharge(unsigned long count)
6120 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6121 struct mem_cgroup *memcg = mc.to;
6123 if (mem_cgroup_is_root(memcg)) {
6124 mc.precharge += count;
6125 /* we don't need css_get for root */
6128 /* try to charge at once */
6130 struct res_counter *dummy;
6132 * "memcg" cannot be under rmdir() because we've already checked
6133 * by cgroup_lock_live_cgroup() that it is not removed and we
6134 * are still under the same cgroup_mutex. So we can postpone
6137 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6139 if (do_swap_account && res_counter_charge(&memcg->memsw,
6140 PAGE_SIZE * count, &dummy)) {
6141 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6144 mc.precharge += count;
6148 /* fall back to one by one charge */
6150 if (signal_pending(current)) {
6154 if (!batch_count--) {
6155 batch_count = PRECHARGE_COUNT_AT_ONCE;
6158 ret = __mem_cgroup_try_charge(NULL,
6159 GFP_KERNEL, 1, &memcg, false);
6161 /* mem_cgroup_clear_mc() will do uncharge later */
6169 * get_mctgt_type - get target type of moving charge
6170 * @vma: the vma the pte to be checked belongs
6171 * @addr: the address corresponding to the pte to be checked
6172 * @ptent: the pte to be checked
6173 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6176 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6177 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6178 * move charge. if @target is not NULL, the page is stored in target->page
6179 * with extra refcnt got(Callers should handle it).
6180 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6181 * target for charge migration. if @target is not NULL, the entry is stored
6184 * Called with pte lock held.
6191 enum mc_target_type {
6197 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6198 unsigned long addr, pte_t ptent)
6200 struct page *page = vm_normal_page(vma, addr, ptent);
6202 if (!page || !page_mapped(page))
6204 if (PageAnon(page)) {
6205 /* we don't move shared anon */
6208 } else if (!move_file())
6209 /* we ignore mapcount for file pages */
6211 if (!get_page_unless_zero(page))
6218 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6219 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6221 struct page *page = NULL;
6222 swp_entry_t ent = pte_to_swp_entry(ptent);
6224 if (!move_anon() || non_swap_entry(ent))
6227 * Because lookup_swap_cache() updates some statistics counter,
6228 * we call find_get_page() with swapper_space directly.
6230 page = find_get_page(swap_address_space(ent), ent.val);
6231 if (do_swap_account)
6232 entry->val = ent.val;
6237 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6238 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6244 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6245 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6247 struct page *page = NULL;
6248 struct address_space *mapping;
6251 if (!vma->vm_file) /* anonymous vma */
6256 mapping = vma->vm_file->f_mapping;
6257 if (pte_none(ptent))
6258 pgoff = linear_page_index(vma, addr);
6259 else /* pte_file(ptent) is true */
6260 pgoff = pte_to_pgoff(ptent);
6262 /* page is moved even if it's not RSS of this task(page-faulted). */
6263 page = find_get_page(mapping, pgoff);
6266 /* shmem/tmpfs may report page out on swap: account for that too. */
6267 if (radix_tree_exceptional_entry(page)) {
6268 swp_entry_t swap = radix_to_swp_entry(page);
6269 if (do_swap_account)
6271 page = find_get_page(swap_address_space(swap), swap.val);
6277 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6278 unsigned long addr, pte_t ptent, union mc_target *target)
6280 struct page *page = NULL;
6281 struct page_cgroup *pc;
6282 enum mc_target_type ret = MC_TARGET_NONE;
6283 swp_entry_t ent = { .val = 0 };
6285 if (pte_present(ptent))
6286 page = mc_handle_present_pte(vma, addr, ptent);
6287 else if (is_swap_pte(ptent))
6288 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6289 else if (pte_none(ptent) || pte_file(ptent))
6290 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6292 if (!page && !ent.val)
6295 pc = lookup_page_cgroup(page);
6297 * Do only loose check w/o page_cgroup lock.
6298 * mem_cgroup_move_account() checks the pc is valid or not under
6301 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6302 ret = MC_TARGET_PAGE;
6304 target->page = page;
6306 if (!ret || !target)
6309 /* There is a swap entry and a page doesn't exist or isn't charged */
6310 if (ent.val && !ret &&
6311 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6312 ret = MC_TARGET_SWAP;
6319 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6321 * We don't consider swapping or file mapped pages because THP does not
6322 * support them for now.
6323 * Caller should make sure that pmd_trans_huge(pmd) is true.
6325 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6326 unsigned long addr, pmd_t pmd, union mc_target *target)
6328 struct page *page = NULL;
6329 struct page_cgroup *pc;
6330 enum mc_target_type ret = MC_TARGET_NONE;
6332 page = pmd_page(pmd);
6333 VM_BUG_ON(!page || !PageHead(page));
6336 pc = lookup_page_cgroup(page);
6337 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6338 ret = MC_TARGET_PAGE;
6341 target->page = page;
6347 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6348 unsigned long addr, pmd_t pmd, union mc_target *target)
6350 return MC_TARGET_NONE;
6354 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6355 unsigned long addr, unsigned long end,
6356 struct mm_walk *walk)
6358 struct vm_area_struct *vma = walk->private;
6362 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6363 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6364 mc.precharge += HPAGE_PMD_NR;
6365 spin_unlock(&vma->vm_mm->page_table_lock);
6369 if (pmd_trans_unstable(pmd))
6371 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6372 for (; addr != end; pte++, addr += PAGE_SIZE)
6373 if (get_mctgt_type(vma, addr, *pte, NULL))
6374 mc.precharge++; /* increment precharge temporarily */
6375 pte_unmap_unlock(pte - 1, ptl);
6381 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6383 unsigned long precharge;
6384 struct vm_area_struct *vma;
6386 down_read(&mm->mmap_sem);
6387 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6388 struct mm_walk mem_cgroup_count_precharge_walk = {
6389 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6393 if (is_vm_hugetlb_page(vma))
6395 walk_page_range(vma->vm_start, vma->vm_end,
6396 &mem_cgroup_count_precharge_walk);
6398 up_read(&mm->mmap_sem);
6400 precharge = mc.precharge;
6406 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6408 unsigned long precharge = mem_cgroup_count_precharge(mm);
6410 VM_BUG_ON(mc.moving_task);
6411 mc.moving_task = current;
6412 return mem_cgroup_do_precharge(precharge);
6415 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6416 static void __mem_cgroup_clear_mc(void)
6418 struct mem_cgroup *from = mc.from;
6419 struct mem_cgroup *to = mc.to;
6422 /* we must uncharge all the leftover precharges from mc.to */
6424 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6428 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6429 * we must uncharge here.
6431 if (mc.moved_charge) {
6432 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6433 mc.moved_charge = 0;
6435 /* we must fixup refcnts and charges */
6436 if (mc.moved_swap) {
6437 /* uncharge swap account from the old cgroup */
6438 if (!mem_cgroup_is_root(mc.from))
6439 res_counter_uncharge(&mc.from->memsw,
6440 PAGE_SIZE * mc.moved_swap);
6442 for (i = 0; i < mc.moved_swap; i++)
6443 css_put(&mc.from->css);
6445 if (!mem_cgroup_is_root(mc.to)) {
6447 * we charged both to->res and to->memsw, so we should
6450 res_counter_uncharge(&mc.to->res,
6451 PAGE_SIZE * mc.moved_swap);
6453 /* we've already done css_get(mc.to) */
6456 memcg_oom_recover(from);
6457 memcg_oom_recover(to);
6458 wake_up_all(&mc.waitq);
6461 static void mem_cgroup_clear_mc(void)
6463 struct mem_cgroup *from = mc.from;
6466 * we must clear moving_task before waking up waiters at the end of
6469 mc.moving_task = NULL;
6470 __mem_cgroup_clear_mc();
6471 spin_lock(&mc.lock);
6474 spin_unlock(&mc.lock);
6475 mem_cgroup_end_move(from);
6478 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6479 struct cgroup_taskset *tset)
6481 struct task_struct *p = cgroup_taskset_first(tset);
6483 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6484 unsigned long move_charge_at_immigrate;
6487 * We are now commited to this value whatever it is. Changes in this
6488 * tunable will only affect upcoming migrations, not the current one.
6489 * So we need to save it, and keep it going.
6491 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6492 if (move_charge_at_immigrate) {
6493 struct mm_struct *mm;
6494 struct mem_cgroup *from = mem_cgroup_from_task(p);
6496 VM_BUG_ON(from == memcg);
6498 mm = get_task_mm(p);
6501 /* We move charges only when we move a owner of the mm */
6502 if (mm->owner == p) {
6505 VM_BUG_ON(mc.precharge);
6506 VM_BUG_ON(mc.moved_charge);
6507 VM_BUG_ON(mc.moved_swap);
6508 mem_cgroup_start_move(from);
6509 spin_lock(&mc.lock);
6512 mc.immigrate_flags = move_charge_at_immigrate;
6513 spin_unlock(&mc.lock);
6514 /* We set mc.moving_task later */
6516 ret = mem_cgroup_precharge_mc(mm);
6518 mem_cgroup_clear_mc();
6525 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6526 struct cgroup_taskset *tset)
6528 mem_cgroup_clear_mc();
6531 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6532 unsigned long addr, unsigned long end,
6533 struct mm_walk *walk)
6536 struct vm_area_struct *vma = walk->private;
6539 enum mc_target_type target_type;
6540 union mc_target target;
6542 struct page_cgroup *pc;
6545 * We don't take compound_lock() here but no race with splitting thp
6547 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6548 * under splitting, which means there's no concurrent thp split,
6549 * - if another thread runs into split_huge_page() just after we
6550 * entered this if-block, the thread must wait for page table lock
6551 * to be unlocked in __split_huge_page_splitting(), where the main
6552 * part of thp split is not executed yet.
6554 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6555 if (mc.precharge < HPAGE_PMD_NR) {
6556 spin_unlock(&vma->vm_mm->page_table_lock);
6559 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6560 if (target_type == MC_TARGET_PAGE) {
6562 if (!isolate_lru_page(page)) {
6563 pc = lookup_page_cgroup(page);
6564 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6565 pc, mc.from, mc.to)) {
6566 mc.precharge -= HPAGE_PMD_NR;
6567 mc.moved_charge += HPAGE_PMD_NR;
6569 putback_lru_page(page);
6573 spin_unlock(&vma->vm_mm->page_table_lock);
6577 if (pmd_trans_unstable(pmd))
6580 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6581 for (; addr != end; addr += PAGE_SIZE) {
6582 pte_t ptent = *(pte++);
6588 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6589 case MC_TARGET_PAGE:
6591 if (isolate_lru_page(page))
6593 pc = lookup_page_cgroup(page);
6594 if (!mem_cgroup_move_account(page, 1, pc,
6597 /* we uncharge from mc.from later. */
6600 putback_lru_page(page);
6601 put: /* get_mctgt_type() gets the page */
6604 case MC_TARGET_SWAP:
6606 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6608 /* we fixup refcnts and charges later. */
6616 pte_unmap_unlock(pte - 1, ptl);
6621 * We have consumed all precharges we got in can_attach().
6622 * We try charge one by one, but don't do any additional
6623 * charges to mc.to if we have failed in charge once in attach()
6626 ret = mem_cgroup_do_precharge(1);
6634 static void mem_cgroup_move_charge(struct mm_struct *mm)
6636 struct vm_area_struct *vma;
6638 lru_add_drain_all();
6640 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6642 * Someone who are holding the mmap_sem might be waiting in
6643 * waitq. So we cancel all extra charges, wake up all waiters,
6644 * and retry. Because we cancel precharges, we might not be able
6645 * to move enough charges, but moving charge is a best-effort
6646 * feature anyway, so it wouldn't be a big problem.
6648 __mem_cgroup_clear_mc();
6652 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6654 struct mm_walk mem_cgroup_move_charge_walk = {
6655 .pmd_entry = mem_cgroup_move_charge_pte_range,
6659 if (is_vm_hugetlb_page(vma))
6661 ret = walk_page_range(vma->vm_start, vma->vm_end,
6662 &mem_cgroup_move_charge_walk);
6665 * means we have consumed all precharges and failed in
6666 * doing additional charge. Just abandon here.
6670 up_read(&mm->mmap_sem);
6673 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6674 struct cgroup_taskset *tset)
6676 struct task_struct *p = cgroup_taskset_first(tset);
6677 struct mm_struct *mm = get_task_mm(p);
6681 mem_cgroup_move_charge(mm);
6685 mem_cgroup_clear_mc();
6687 #else /* !CONFIG_MMU */
6688 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6689 struct cgroup_taskset *tset)
6693 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6694 struct cgroup_taskset *tset)
6697 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6698 struct cgroup_taskset *tset)
6704 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6705 * to verify sane_behavior flag on each mount attempt.
6707 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6710 * use_hierarchy is forced with sane_behavior. cgroup core
6711 * guarantees that @root doesn't have any children, so turning it
6712 * on for the root memcg is enough.
6714 if (cgroup_sane_behavior(root_css->cgroup))
6715 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6718 struct cgroup_subsys mem_cgroup_subsys = {
6720 .subsys_id = mem_cgroup_subsys_id,
6721 .css_alloc = mem_cgroup_css_alloc,
6722 .css_online = mem_cgroup_css_online,
6723 .css_offline = mem_cgroup_css_offline,
6724 .css_free = mem_cgroup_css_free,
6725 .can_attach = mem_cgroup_can_attach,
6726 .cancel_attach = mem_cgroup_cancel_attach,
6727 .attach = mem_cgroup_move_task,
6728 .bind = mem_cgroup_bind,
6729 .base_cftypes = mem_cgroup_files,
6734 #ifdef CONFIG_MEMCG_SWAP
6735 static int __init enable_swap_account(char *s)
6737 if (!strcmp(s, "1"))
6738 really_do_swap_account = 1;
6739 else if (!strcmp(s, "0"))
6740 really_do_swap_account = 0;
6743 __setup("swapaccount=", enable_swap_account);
6745 static void __init memsw_file_init(void)
6747 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6750 static void __init enable_swap_cgroup(void)
6752 if (!mem_cgroup_disabled() && really_do_swap_account) {
6753 do_swap_account = 1;
6759 static void __init enable_swap_cgroup(void)
6765 * subsys_initcall() for memory controller.
6767 * Some parts like hotcpu_notifier() have to be initialized from this context
6768 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6769 * everything that doesn't depend on a specific mem_cgroup structure should
6770 * be initialized from here.
6772 static int __init mem_cgroup_init(void)
6774 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6775 enable_swap_cgroup();
6779 subsys_initcall(mem_cgroup_init);