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 VM_BUG_ON(!rcu_read_lock_held());
2229 memcg = pc->mem_cgroup;
2230 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2233 this_cpu_add(memcg->stat->count[idx], val);
2237 * size of first charge trial. "32" comes from vmscan.c's magic value.
2238 * TODO: maybe necessary to use big numbers in big irons.
2240 #define CHARGE_BATCH 32U
2241 struct memcg_stock_pcp {
2242 struct mem_cgroup *cached; /* this never be root cgroup */
2243 unsigned int nr_pages;
2244 struct work_struct work;
2245 unsigned long flags;
2246 #define FLUSHING_CACHED_CHARGE 0
2248 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2249 static DEFINE_MUTEX(percpu_charge_mutex);
2252 * consume_stock: Try to consume stocked charge on this cpu.
2253 * @memcg: memcg to consume from.
2254 * @nr_pages: how many pages to charge.
2256 * The charges will only happen if @memcg matches the current cpu's memcg
2257 * stock, and at least @nr_pages are available in that stock. Failure to
2258 * service an allocation will refill the stock.
2260 * returns true if successful, false otherwise.
2262 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2264 struct memcg_stock_pcp *stock;
2267 if (nr_pages > CHARGE_BATCH)
2270 stock = &get_cpu_var(memcg_stock);
2271 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2272 stock->nr_pages -= nr_pages;
2273 else /* need to call res_counter_charge */
2275 put_cpu_var(memcg_stock);
2280 * Returns stocks cached in percpu to res_counter and reset cached information.
2282 static void drain_stock(struct memcg_stock_pcp *stock)
2284 struct mem_cgroup *old = stock->cached;
2286 if (stock->nr_pages) {
2287 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2289 res_counter_uncharge(&old->res, bytes);
2290 if (do_swap_account)
2291 res_counter_uncharge(&old->memsw, bytes);
2292 stock->nr_pages = 0;
2294 stock->cached = NULL;
2298 * This must be called under preempt disabled or must be called by
2299 * a thread which is pinned to local cpu.
2301 static void drain_local_stock(struct work_struct *dummy)
2303 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2305 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2308 static void __init memcg_stock_init(void)
2312 for_each_possible_cpu(cpu) {
2313 struct memcg_stock_pcp *stock =
2314 &per_cpu(memcg_stock, cpu);
2315 INIT_WORK(&stock->work, drain_local_stock);
2320 * Cache charges(val) which is from res_counter, to local per_cpu area.
2321 * This will be consumed by consume_stock() function, later.
2323 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2325 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2327 if (stock->cached != memcg) { /* reset if necessary */
2329 stock->cached = memcg;
2331 stock->nr_pages += nr_pages;
2332 put_cpu_var(memcg_stock);
2336 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2337 * of the hierarchy under it. sync flag says whether we should block
2338 * until the work is done.
2340 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2344 /* Notify other cpus that system-wide "drain" is running */
2347 for_each_online_cpu(cpu) {
2348 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2349 struct mem_cgroup *memcg;
2351 memcg = stock->cached;
2352 if (!memcg || !stock->nr_pages)
2354 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2356 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2358 drain_local_stock(&stock->work);
2360 schedule_work_on(cpu, &stock->work);
2368 for_each_online_cpu(cpu) {
2369 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2370 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2371 flush_work(&stock->work);
2378 * Tries to drain stocked charges in other cpus. This function is asynchronous
2379 * and just put a work per cpu for draining localy on each cpu. Caller can
2380 * expects some charges will be back to res_counter later but cannot wait for
2383 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2386 * If someone calls draining, avoid adding more kworker runs.
2388 if (!mutex_trylock(&percpu_charge_mutex))
2390 drain_all_stock(root_memcg, false);
2391 mutex_unlock(&percpu_charge_mutex);
2394 /* This is a synchronous drain interface. */
2395 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2397 /* called when force_empty is called */
2398 mutex_lock(&percpu_charge_mutex);
2399 drain_all_stock(root_memcg, true);
2400 mutex_unlock(&percpu_charge_mutex);
2404 * This function drains percpu counter value from DEAD cpu and
2405 * move it to local cpu. Note that this function can be preempted.
2407 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2411 spin_lock(&memcg->pcp_counter_lock);
2412 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2413 long x = per_cpu(memcg->stat->count[i], cpu);
2415 per_cpu(memcg->stat->count[i], cpu) = 0;
2416 memcg->nocpu_base.count[i] += x;
2418 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2419 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2421 per_cpu(memcg->stat->events[i], cpu) = 0;
2422 memcg->nocpu_base.events[i] += x;
2424 spin_unlock(&memcg->pcp_counter_lock);
2427 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2428 unsigned long action,
2431 int cpu = (unsigned long)hcpu;
2432 struct memcg_stock_pcp *stock;
2433 struct mem_cgroup *iter;
2435 if (action == CPU_ONLINE)
2438 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2441 for_each_mem_cgroup(iter)
2442 mem_cgroup_drain_pcp_counter(iter, cpu);
2444 stock = &per_cpu(memcg_stock, cpu);
2450 /* See __mem_cgroup_try_charge() for details */
2452 CHARGE_OK, /* success */
2453 CHARGE_RETRY, /* need to retry but retry is not bad */
2454 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2455 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2458 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2459 unsigned int nr_pages, unsigned int min_pages,
2462 unsigned long csize = nr_pages * PAGE_SIZE;
2463 struct mem_cgroup *mem_over_limit;
2464 struct res_counter *fail_res;
2465 unsigned long flags = 0;
2468 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2471 if (!do_swap_account)
2473 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2477 res_counter_uncharge(&memcg->res, csize);
2478 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2479 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2481 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2483 * Never reclaim on behalf of optional batching, retry with a
2484 * single page instead.
2486 if (nr_pages > min_pages)
2487 return CHARGE_RETRY;
2489 if (!(gfp_mask & __GFP_WAIT))
2490 return CHARGE_WOULDBLOCK;
2492 if (gfp_mask & __GFP_NORETRY)
2493 return CHARGE_NOMEM;
2495 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2496 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2497 return CHARGE_RETRY;
2499 * Even though the limit is exceeded at this point, reclaim
2500 * may have been able to free some pages. Retry the charge
2501 * before killing the task.
2503 * Only for regular pages, though: huge pages are rather
2504 * unlikely to succeed so close to the limit, and we fall back
2505 * to regular pages anyway in case of failure.
2507 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2508 return CHARGE_RETRY;
2511 * At task move, charge accounts can be doubly counted. So, it's
2512 * better to wait until the end of task_move if something is going on.
2514 if (mem_cgroup_wait_acct_move(mem_over_limit))
2515 return CHARGE_RETRY;
2518 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2520 return CHARGE_NOMEM;
2524 * __mem_cgroup_try_charge() does
2525 * 1. detect memcg to be charged against from passed *mm and *ptr,
2526 * 2. update res_counter
2527 * 3. call memory reclaim if necessary.
2529 * In some special case, if the task is fatal, fatal_signal_pending() or
2530 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2531 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2532 * as possible without any hazards. 2: all pages should have a valid
2533 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2534 * pointer, that is treated as a charge to root_mem_cgroup.
2536 * So __mem_cgroup_try_charge() will return
2537 * 0 ... on success, filling *ptr with a valid memcg pointer.
2538 * -ENOMEM ... charge failure because of resource limits.
2539 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2541 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2542 * the oom-killer can be invoked.
2544 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2546 unsigned int nr_pages,
2547 struct mem_cgroup **ptr,
2550 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2551 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2552 struct mem_cgroup *memcg = NULL;
2556 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2557 * in system level. So, allow to go ahead dying process in addition to
2560 if (unlikely(test_thread_flag(TIF_MEMDIE)
2561 || fatal_signal_pending(current)))
2565 * We always charge the cgroup the mm_struct belongs to.
2566 * The mm_struct's mem_cgroup changes on task migration if the
2567 * thread group leader migrates. It's possible that mm is not
2568 * set, if so charge the root memcg (happens for pagecache usage).
2571 *ptr = root_mem_cgroup;
2573 if (*ptr) { /* css should be a valid one */
2575 if (mem_cgroup_is_root(memcg))
2577 if (consume_stock(memcg, nr_pages))
2579 css_get(&memcg->css);
2581 struct task_struct *p;
2584 p = rcu_dereference(mm->owner);
2586 * Because we don't have task_lock(), "p" can exit.
2587 * In that case, "memcg" can point to root or p can be NULL with
2588 * race with swapoff. Then, we have small risk of mis-accouning.
2589 * But such kind of mis-account by race always happens because
2590 * we don't have cgroup_mutex(). It's overkill and we allo that
2592 * (*) swapoff at el will charge against mm-struct not against
2593 * task-struct. So, mm->owner can be NULL.
2595 memcg = mem_cgroup_from_task(p);
2597 memcg = root_mem_cgroup;
2598 if (mem_cgroup_is_root(memcg)) {
2602 if (consume_stock(memcg, nr_pages)) {
2604 * It seems dagerous to access memcg without css_get().
2605 * But considering how consume_stok works, it's not
2606 * necessary. If consume_stock success, some charges
2607 * from this memcg are cached on this cpu. So, we
2608 * don't need to call css_get()/css_tryget() before
2609 * calling consume_stock().
2614 /* after here, we may be blocked. we need to get refcnt */
2615 if (!css_tryget(&memcg->css)) {
2623 bool invoke_oom = oom && !nr_oom_retries;
2625 /* If killed, bypass charge */
2626 if (fatal_signal_pending(current)) {
2627 css_put(&memcg->css);
2631 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2632 nr_pages, invoke_oom);
2636 case CHARGE_RETRY: /* not in OOM situation but retry */
2638 css_put(&memcg->css);
2641 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2642 css_put(&memcg->css);
2644 case CHARGE_NOMEM: /* OOM routine works */
2645 if (!oom || invoke_oom) {
2646 css_put(&memcg->css);
2652 } while (ret != CHARGE_OK);
2654 if (batch > nr_pages)
2655 refill_stock(memcg, batch - nr_pages);
2656 css_put(&memcg->css);
2664 *ptr = root_mem_cgroup;
2669 * Somemtimes we have to undo a charge we got by try_charge().
2670 * This function is for that and do uncharge, put css's refcnt.
2671 * gotten by try_charge().
2673 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2674 unsigned int nr_pages)
2676 if (!mem_cgroup_is_root(memcg)) {
2677 unsigned long bytes = nr_pages * PAGE_SIZE;
2679 res_counter_uncharge(&memcg->res, bytes);
2680 if (do_swap_account)
2681 res_counter_uncharge(&memcg->memsw, bytes);
2686 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2687 * This is useful when moving usage to parent cgroup.
2689 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2690 unsigned int nr_pages)
2692 unsigned long bytes = nr_pages * PAGE_SIZE;
2694 if (mem_cgroup_is_root(memcg))
2697 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2698 if (do_swap_account)
2699 res_counter_uncharge_until(&memcg->memsw,
2700 memcg->memsw.parent, bytes);
2704 * A helper function to get mem_cgroup from ID. must be called under
2705 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2706 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2707 * called against removed memcg.)
2709 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2711 struct cgroup_subsys_state *css;
2713 /* ID 0 is unused ID */
2716 css = css_lookup(&mem_cgroup_subsys, id);
2719 return mem_cgroup_from_css(css);
2722 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2724 struct mem_cgroup *memcg = NULL;
2725 struct page_cgroup *pc;
2729 VM_BUG_ON(!PageLocked(page));
2731 pc = lookup_page_cgroup(page);
2732 lock_page_cgroup(pc);
2733 if (PageCgroupUsed(pc)) {
2734 memcg = pc->mem_cgroup;
2735 if (memcg && !css_tryget(&memcg->css))
2737 } else if (PageSwapCache(page)) {
2738 ent.val = page_private(page);
2739 id = lookup_swap_cgroup_id(ent);
2741 memcg = mem_cgroup_lookup(id);
2742 if (memcg && !css_tryget(&memcg->css))
2746 unlock_page_cgroup(pc);
2750 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2752 unsigned int nr_pages,
2753 enum charge_type ctype,
2756 struct page_cgroup *pc = lookup_page_cgroup(page);
2757 struct zone *uninitialized_var(zone);
2758 struct lruvec *lruvec;
2759 bool was_on_lru = false;
2762 lock_page_cgroup(pc);
2763 VM_BUG_ON(PageCgroupUsed(pc));
2765 * we don't need page_cgroup_lock about tail pages, becase they are not
2766 * accessed by any other context at this point.
2770 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2771 * may already be on some other mem_cgroup's LRU. Take care of it.
2774 zone = page_zone(page);
2775 spin_lock_irq(&zone->lru_lock);
2776 if (PageLRU(page)) {
2777 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2779 del_page_from_lru_list(page, lruvec, page_lru(page));
2784 pc->mem_cgroup = memcg;
2786 * We access a page_cgroup asynchronously without lock_page_cgroup().
2787 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2788 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2789 * before USED bit, we need memory barrier here.
2790 * See mem_cgroup_add_lru_list(), etc.
2793 SetPageCgroupUsed(pc);
2797 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2798 VM_BUG_ON(PageLRU(page));
2800 add_page_to_lru_list(page, lruvec, page_lru(page));
2802 spin_unlock_irq(&zone->lru_lock);
2805 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2810 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2811 unlock_page_cgroup(pc);
2814 * "charge_statistics" updated event counter.
2816 memcg_check_events(memcg, page);
2819 static DEFINE_MUTEX(set_limit_mutex);
2821 #ifdef CONFIG_MEMCG_KMEM
2822 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2824 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2825 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2829 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2830 * in the memcg_cache_params struct.
2832 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2834 struct kmem_cache *cachep;
2836 VM_BUG_ON(p->is_root_cache);
2837 cachep = p->root_cache;
2838 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2841 #ifdef CONFIG_SLABINFO
2842 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2843 struct cftype *cft, struct seq_file *m)
2845 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2846 struct memcg_cache_params *params;
2848 if (!memcg_can_account_kmem(memcg))
2851 print_slabinfo_header(m);
2853 mutex_lock(&memcg->slab_caches_mutex);
2854 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2855 cache_show(memcg_params_to_cache(params), m);
2856 mutex_unlock(&memcg->slab_caches_mutex);
2862 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2864 struct res_counter *fail_res;
2865 struct mem_cgroup *_memcg;
2869 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2874 * Conditions under which we can wait for the oom_killer. Those are
2875 * the same conditions tested by the core page allocator
2877 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2880 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2883 if (ret == -EINTR) {
2885 * __mem_cgroup_try_charge() chosed to bypass to root due to
2886 * OOM kill or fatal signal. Since our only options are to
2887 * either fail the allocation or charge it to this cgroup, do
2888 * it as a temporary condition. But we can't fail. From a
2889 * kmem/slab perspective, the cache has already been selected,
2890 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2893 * This condition will only trigger if the task entered
2894 * memcg_charge_kmem in a sane state, but was OOM-killed during
2895 * __mem_cgroup_try_charge() above. Tasks that were already
2896 * dying when the allocation triggers should have been already
2897 * directed to the root cgroup in memcontrol.h
2899 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2900 if (do_swap_account)
2901 res_counter_charge_nofail(&memcg->memsw, size,
2905 res_counter_uncharge(&memcg->kmem, size);
2910 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2912 res_counter_uncharge(&memcg->res, size);
2913 if (do_swap_account)
2914 res_counter_uncharge(&memcg->memsw, size);
2917 if (res_counter_uncharge(&memcg->kmem, size))
2921 * Releases a reference taken in kmem_cgroup_css_offline in case
2922 * this last uncharge is racing with the offlining code or it is
2923 * outliving the memcg existence.
2925 * The memory barrier imposed by test&clear is paired with the
2926 * explicit one in memcg_kmem_mark_dead().
2928 if (memcg_kmem_test_and_clear_dead(memcg))
2929 css_put(&memcg->css);
2932 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
2937 mutex_lock(&memcg->slab_caches_mutex);
2938 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
2939 mutex_unlock(&memcg->slab_caches_mutex);
2943 * helper for acessing a memcg's index. It will be used as an index in the
2944 * child cache array in kmem_cache, and also to derive its name. This function
2945 * will return -1 when this is not a kmem-limited memcg.
2947 int memcg_cache_id(struct mem_cgroup *memcg)
2949 return memcg ? memcg->kmemcg_id : -1;
2953 * This ends up being protected by the set_limit mutex, during normal
2954 * operation, because that is its main call site.
2956 * But when we create a new cache, we can call this as well if its parent
2957 * is kmem-limited. That will have to hold set_limit_mutex as well.
2959 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
2963 num = ida_simple_get(&kmem_limited_groups,
2964 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
2968 * After this point, kmem_accounted (that we test atomically in
2969 * the beginning of this conditional), is no longer 0. This
2970 * guarantees only one process will set the following boolean
2971 * to true. We don't need test_and_set because we're protected
2972 * by the set_limit_mutex anyway.
2974 memcg_kmem_set_activated(memcg);
2976 ret = memcg_update_all_caches(num+1);
2978 ida_simple_remove(&kmem_limited_groups, num);
2979 memcg_kmem_clear_activated(memcg);
2983 memcg->kmemcg_id = num;
2984 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
2985 mutex_init(&memcg->slab_caches_mutex);
2989 static size_t memcg_caches_array_size(int num_groups)
2992 if (num_groups <= 0)
2995 size = 2 * num_groups;
2996 if (size < MEMCG_CACHES_MIN_SIZE)
2997 size = MEMCG_CACHES_MIN_SIZE;
2998 else if (size > MEMCG_CACHES_MAX_SIZE)
2999 size = MEMCG_CACHES_MAX_SIZE;
3005 * We should update the current array size iff all caches updates succeed. This
3006 * can only be done from the slab side. The slab mutex needs to be held when
3009 void memcg_update_array_size(int num)
3011 if (num > memcg_limited_groups_array_size)
3012 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3015 static void kmem_cache_destroy_work_func(struct work_struct *w);
3017 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3019 struct memcg_cache_params *cur_params = s->memcg_params;
3021 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3023 if (num_groups > memcg_limited_groups_array_size) {
3025 ssize_t size = memcg_caches_array_size(num_groups);
3027 size *= sizeof(void *);
3028 size += offsetof(struct memcg_cache_params, memcg_caches);
3030 s->memcg_params = kzalloc(size, GFP_KERNEL);
3031 if (!s->memcg_params) {
3032 s->memcg_params = cur_params;
3036 s->memcg_params->is_root_cache = true;
3039 * There is the chance it will be bigger than
3040 * memcg_limited_groups_array_size, if we failed an allocation
3041 * in a cache, in which case all caches updated before it, will
3042 * have a bigger array.
3044 * But if that is the case, the data after
3045 * memcg_limited_groups_array_size is certainly unused
3047 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3048 if (!cur_params->memcg_caches[i])
3050 s->memcg_params->memcg_caches[i] =
3051 cur_params->memcg_caches[i];
3055 * Ideally, we would wait until all caches succeed, and only
3056 * then free the old one. But this is not worth the extra
3057 * pointer per-cache we'd have to have for this.
3059 * It is not a big deal if some caches are left with a size
3060 * bigger than the others. And all updates will reset this
3068 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3069 struct kmem_cache *root_cache)
3073 if (!memcg_kmem_enabled())
3077 size = offsetof(struct memcg_cache_params, memcg_caches);
3078 size += memcg_limited_groups_array_size * sizeof(void *);
3080 size = sizeof(struct memcg_cache_params);
3082 s->memcg_params = kzalloc(size, GFP_KERNEL);
3083 if (!s->memcg_params)
3087 s->memcg_params->memcg = memcg;
3088 s->memcg_params->root_cache = root_cache;
3089 INIT_WORK(&s->memcg_params->destroy,
3090 kmem_cache_destroy_work_func);
3092 s->memcg_params->is_root_cache = true;
3097 void memcg_release_cache(struct kmem_cache *s)
3099 struct kmem_cache *root;
3100 struct mem_cgroup *memcg;
3104 * This happens, for instance, when a root cache goes away before we
3107 if (!s->memcg_params)
3110 if (s->memcg_params->is_root_cache)
3113 memcg = s->memcg_params->memcg;
3114 id = memcg_cache_id(memcg);
3116 root = s->memcg_params->root_cache;
3117 root->memcg_params->memcg_caches[id] = NULL;
3119 mutex_lock(&memcg->slab_caches_mutex);
3120 list_del(&s->memcg_params->list);
3121 mutex_unlock(&memcg->slab_caches_mutex);
3123 css_put(&memcg->css);
3125 kfree(s->memcg_params);
3129 * During the creation a new cache, we need to disable our accounting mechanism
3130 * altogether. This is true even if we are not creating, but rather just
3131 * enqueing new caches to be created.
3133 * This is because that process will trigger allocations; some visible, like
3134 * explicit kmallocs to auxiliary data structures, name strings and internal
3135 * cache structures; some well concealed, like INIT_WORK() that can allocate
3136 * objects during debug.
3138 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3139 * to it. This may not be a bounded recursion: since the first cache creation
3140 * failed to complete (waiting on the allocation), we'll just try to create the
3141 * cache again, failing at the same point.
3143 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3144 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3145 * inside the following two functions.
3147 static inline void memcg_stop_kmem_account(void)
3149 VM_BUG_ON(!current->mm);
3150 current->memcg_kmem_skip_account++;
3153 static inline void memcg_resume_kmem_account(void)
3155 VM_BUG_ON(!current->mm);
3156 current->memcg_kmem_skip_account--;
3159 static void kmem_cache_destroy_work_func(struct work_struct *w)
3161 struct kmem_cache *cachep;
3162 struct memcg_cache_params *p;
3164 p = container_of(w, struct memcg_cache_params, destroy);
3166 cachep = memcg_params_to_cache(p);
3169 * If we get down to 0 after shrink, we could delete right away.
3170 * However, memcg_release_pages() already puts us back in the workqueue
3171 * in that case. If we proceed deleting, we'll get a dangling
3172 * reference, and removing the object from the workqueue in that case
3173 * is unnecessary complication. We are not a fast path.
3175 * Note that this case is fundamentally different from racing with
3176 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3177 * kmem_cache_shrink, not only we would be reinserting a dead cache
3178 * into the queue, but doing so from inside the worker racing to
3181 * So if we aren't down to zero, we'll just schedule a worker and try
3184 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3185 kmem_cache_shrink(cachep);
3186 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3189 kmem_cache_destroy(cachep);
3192 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3194 if (!cachep->memcg_params->dead)
3198 * There are many ways in which we can get here.
3200 * We can get to a memory-pressure situation while the delayed work is
3201 * still pending to run. The vmscan shrinkers can then release all
3202 * cache memory and get us to destruction. If this is the case, we'll
3203 * be executed twice, which is a bug (the second time will execute over
3204 * bogus data). In this case, cancelling the work should be fine.
3206 * But we can also get here from the worker itself, if
3207 * kmem_cache_shrink is enough to shake all the remaining objects and
3208 * get the page count to 0. In this case, we'll deadlock if we try to
3209 * cancel the work (the worker runs with an internal lock held, which
3210 * is the same lock we would hold for cancel_work_sync().)
3212 * Since we can't possibly know who got us here, just refrain from
3213 * running if there is already work pending
3215 if (work_pending(&cachep->memcg_params->destroy))
3218 * We have to defer the actual destroying to a workqueue, because
3219 * we might currently be in a context that cannot sleep.
3221 schedule_work(&cachep->memcg_params->destroy);
3225 * This lock protects updaters, not readers. We want readers to be as fast as
3226 * they can, and they will either see NULL or a valid cache value. Our model
3227 * allow them to see NULL, in which case the root memcg will be selected.
3229 * We need this lock because multiple allocations to the same cache from a non
3230 * will span more than one worker. Only one of them can create the cache.
3232 static DEFINE_MUTEX(memcg_cache_mutex);
3235 * Called with memcg_cache_mutex held
3237 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3238 struct kmem_cache *s)
3240 struct kmem_cache *new;
3241 static char *tmp_name = NULL;
3243 lockdep_assert_held(&memcg_cache_mutex);
3246 * kmem_cache_create_memcg duplicates the given name and
3247 * cgroup_name for this name requires RCU context.
3248 * This static temporary buffer is used to prevent from
3249 * pointless shortliving allocation.
3252 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3258 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3259 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3262 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3263 (s->flags & ~SLAB_PANIC), s->ctor, s);
3266 new->allocflags |= __GFP_KMEMCG;
3271 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3272 struct kmem_cache *cachep)
3274 struct kmem_cache *new_cachep;
3277 BUG_ON(!memcg_can_account_kmem(memcg));
3279 idx = memcg_cache_id(memcg);
3281 mutex_lock(&memcg_cache_mutex);
3282 new_cachep = cachep->memcg_params->memcg_caches[idx];
3284 css_put(&memcg->css);
3288 new_cachep = kmem_cache_dup(memcg, cachep);
3289 if (new_cachep == NULL) {
3290 new_cachep = cachep;
3291 css_put(&memcg->css);
3295 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3297 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3299 * the readers won't lock, make sure everybody sees the updated value,
3300 * so they won't put stuff in the queue again for no reason
3304 mutex_unlock(&memcg_cache_mutex);
3308 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3310 struct kmem_cache *c;
3313 if (!s->memcg_params)
3315 if (!s->memcg_params->is_root_cache)
3319 * If the cache is being destroyed, we trust that there is no one else
3320 * requesting objects from it. Even if there are, the sanity checks in
3321 * kmem_cache_destroy should caught this ill-case.
3323 * Still, we don't want anyone else freeing memcg_caches under our
3324 * noses, which can happen if a new memcg comes to life. As usual,
3325 * we'll take the set_limit_mutex to protect ourselves against this.
3327 mutex_lock(&set_limit_mutex);
3328 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3329 c = s->memcg_params->memcg_caches[i];
3334 * We will now manually delete the caches, so to avoid races
3335 * we need to cancel all pending destruction workers and
3336 * proceed with destruction ourselves.
3338 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3339 * and that could spawn the workers again: it is likely that
3340 * the cache still have active pages until this very moment.
3341 * This would lead us back to mem_cgroup_destroy_cache.
3343 * But that will not execute at all if the "dead" flag is not
3344 * set, so flip it down to guarantee we are in control.
3346 c->memcg_params->dead = false;
3347 cancel_work_sync(&c->memcg_params->destroy);
3348 kmem_cache_destroy(c);
3350 mutex_unlock(&set_limit_mutex);
3353 struct create_work {
3354 struct mem_cgroup *memcg;
3355 struct kmem_cache *cachep;
3356 struct work_struct work;
3359 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3361 struct kmem_cache *cachep;
3362 struct memcg_cache_params *params;
3364 if (!memcg_kmem_is_active(memcg))
3367 mutex_lock(&memcg->slab_caches_mutex);
3368 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3369 cachep = memcg_params_to_cache(params);
3370 cachep->memcg_params->dead = true;
3371 schedule_work(&cachep->memcg_params->destroy);
3373 mutex_unlock(&memcg->slab_caches_mutex);
3376 static void memcg_create_cache_work_func(struct work_struct *w)
3378 struct create_work *cw;
3380 cw = container_of(w, struct create_work, work);
3381 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3386 * Enqueue the creation of a per-memcg kmem_cache.
3388 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3389 struct kmem_cache *cachep)
3391 struct create_work *cw;
3393 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3395 css_put(&memcg->css);
3400 cw->cachep = cachep;
3402 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3403 schedule_work(&cw->work);
3406 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3407 struct kmem_cache *cachep)
3410 * We need to stop accounting when we kmalloc, because if the
3411 * corresponding kmalloc cache is not yet created, the first allocation
3412 * in __memcg_create_cache_enqueue will recurse.
3414 * However, it is better to enclose the whole function. Depending on
3415 * the debugging options enabled, INIT_WORK(), for instance, can
3416 * trigger an allocation. This too, will make us recurse. Because at
3417 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3418 * the safest choice is to do it like this, wrapping the whole function.
3420 memcg_stop_kmem_account();
3421 __memcg_create_cache_enqueue(memcg, cachep);
3422 memcg_resume_kmem_account();
3425 * Return the kmem_cache we're supposed to use for a slab allocation.
3426 * We try to use the current memcg's version of the cache.
3428 * If the cache does not exist yet, if we are the first user of it,
3429 * we either create it immediately, if possible, or create it asynchronously
3431 * In the latter case, we will let the current allocation go through with
3432 * the original cache.
3434 * Can't be called in interrupt context or from kernel threads.
3435 * This function needs to be called with rcu_read_lock() held.
3437 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3440 struct mem_cgroup *memcg;
3443 VM_BUG_ON(!cachep->memcg_params);
3444 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3446 if (!current->mm || current->memcg_kmem_skip_account)
3450 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3452 if (!memcg_can_account_kmem(memcg))
3455 idx = memcg_cache_id(memcg);
3458 * barrier to mare sure we're always seeing the up to date value. The
3459 * code updating memcg_caches will issue a write barrier to match this.
3461 read_barrier_depends();
3462 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3463 cachep = cachep->memcg_params->memcg_caches[idx];
3467 /* The corresponding put will be done in the workqueue. */
3468 if (!css_tryget(&memcg->css))
3473 * If we are in a safe context (can wait, and not in interrupt
3474 * context), we could be be predictable and return right away.
3475 * This would guarantee that the allocation being performed
3476 * already belongs in the new cache.
3478 * However, there are some clashes that can arrive from locking.
3479 * For instance, because we acquire the slab_mutex while doing
3480 * kmem_cache_dup, this means no further allocation could happen
3481 * with the slab_mutex held.
3483 * Also, because cache creation issue get_online_cpus(), this
3484 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3485 * that ends up reversed during cpu hotplug. (cpuset allocates
3486 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3487 * better to defer everything.
3489 memcg_create_cache_enqueue(memcg, cachep);
3495 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3498 * We need to verify if the allocation against current->mm->owner's memcg is
3499 * possible for the given order. But the page is not allocated yet, so we'll
3500 * need a further commit step to do the final arrangements.
3502 * It is possible for the task to switch cgroups in this mean time, so at
3503 * commit time, we can't rely on task conversion any longer. We'll then use
3504 * the handle argument to return to the caller which cgroup we should commit
3505 * against. We could also return the memcg directly and avoid the pointer
3506 * passing, but a boolean return value gives better semantics considering
3507 * the compiled-out case as well.
3509 * Returning true means the allocation is possible.
3512 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3514 struct mem_cgroup *memcg;
3520 * Disabling accounting is only relevant for some specific memcg
3521 * internal allocations. Therefore we would initially not have such
3522 * check here, since direct calls to the page allocator that are marked
3523 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3524 * concerned with cache allocations, and by having this test at
3525 * memcg_kmem_get_cache, we are already able to relay the allocation to
3526 * the root cache and bypass the memcg cache altogether.
3528 * There is one exception, though: the SLUB allocator does not create
3529 * large order caches, but rather service large kmallocs directly from
3530 * the page allocator. Therefore, the following sequence when backed by
3531 * the SLUB allocator:
3533 * memcg_stop_kmem_account();
3534 * kmalloc(<large_number>)
3535 * memcg_resume_kmem_account();
3537 * would effectively ignore the fact that we should skip accounting,
3538 * since it will drive us directly to this function without passing
3539 * through the cache selector memcg_kmem_get_cache. Such large
3540 * allocations are extremely rare but can happen, for instance, for the
3541 * cache arrays. We bring this test here.
3543 if (!current->mm || current->memcg_kmem_skip_account)
3546 memcg = try_get_mem_cgroup_from_mm(current->mm);
3549 * very rare case described in mem_cgroup_from_task. Unfortunately there
3550 * isn't much we can do without complicating this too much, and it would
3551 * be gfp-dependent anyway. Just let it go
3553 if (unlikely(!memcg))
3556 if (!memcg_can_account_kmem(memcg)) {
3557 css_put(&memcg->css);
3561 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3565 css_put(&memcg->css);
3569 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3572 struct page_cgroup *pc;
3574 VM_BUG_ON(mem_cgroup_is_root(memcg));
3576 /* The page allocation failed. Revert */
3578 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3582 pc = lookup_page_cgroup(page);
3583 lock_page_cgroup(pc);
3584 pc->mem_cgroup = memcg;
3585 SetPageCgroupUsed(pc);
3586 unlock_page_cgroup(pc);
3589 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3591 struct mem_cgroup *memcg = NULL;
3592 struct page_cgroup *pc;
3595 pc = lookup_page_cgroup(page);
3597 * Fast unlocked return. Theoretically might have changed, have to
3598 * check again after locking.
3600 if (!PageCgroupUsed(pc))
3603 lock_page_cgroup(pc);
3604 if (PageCgroupUsed(pc)) {
3605 memcg = pc->mem_cgroup;
3606 ClearPageCgroupUsed(pc);
3608 unlock_page_cgroup(pc);
3611 * We trust that only if there is a memcg associated with the page, it
3612 * is a valid allocation
3617 VM_BUG_ON(mem_cgroup_is_root(memcg));
3618 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3621 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3624 #endif /* CONFIG_MEMCG_KMEM */
3626 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3628 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3630 * Because tail pages are not marked as "used", set it. We're under
3631 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3632 * charge/uncharge will be never happen and move_account() is done under
3633 * compound_lock(), so we don't have to take care of races.
3635 void mem_cgroup_split_huge_fixup(struct page *head)
3637 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3638 struct page_cgroup *pc;
3639 struct mem_cgroup *memcg;
3642 if (mem_cgroup_disabled())
3645 memcg = head_pc->mem_cgroup;
3646 for (i = 1; i < HPAGE_PMD_NR; i++) {
3648 pc->mem_cgroup = memcg;
3649 smp_wmb();/* see __commit_charge() */
3650 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3652 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3655 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3658 * mem_cgroup_move_account - move account of the page
3660 * @nr_pages: number of regular pages (>1 for huge pages)
3661 * @pc: page_cgroup of the page.
3662 * @from: mem_cgroup which the page is moved from.
3663 * @to: mem_cgroup which the page is moved to. @from != @to.
3665 * The caller must confirm following.
3666 * - page is not on LRU (isolate_page() is useful.)
3667 * - compound_lock is held when nr_pages > 1
3669 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3672 static int mem_cgroup_move_account(struct page *page,
3673 unsigned int nr_pages,
3674 struct page_cgroup *pc,
3675 struct mem_cgroup *from,
3676 struct mem_cgroup *to)
3678 unsigned long flags;
3680 bool anon = PageAnon(page);
3682 VM_BUG_ON(from == to);
3683 VM_BUG_ON(PageLRU(page));
3685 * The page is isolated from LRU. So, collapse function
3686 * will not handle this page. But page splitting can happen.
3687 * Do this check under compound_page_lock(). The caller should
3691 if (nr_pages > 1 && !PageTransHuge(page))
3694 lock_page_cgroup(pc);
3697 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3700 move_lock_mem_cgroup(from, &flags);
3702 if (!anon && page_mapped(page)) {
3703 /* Update mapped_file data for mem_cgroup */
3705 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3706 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3709 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3711 /* caller should have done css_get */
3712 pc->mem_cgroup = to;
3713 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3714 move_unlock_mem_cgroup(from, &flags);
3717 unlock_page_cgroup(pc);
3721 memcg_check_events(to, page);
3722 memcg_check_events(from, page);
3728 * mem_cgroup_move_parent - moves page to the parent group
3729 * @page: the page to move
3730 * @pc: page_cgroup of the page
3731 * @child: page's cgroup
3733 * move charges to its parent or the root cgroup if the group has no
3734 * parent (aka use_hierarchy==0).
3735 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3736 * mem_cgroup_move_account fails) the failure is always temporary and
3737 * it signals a race with a page removal/uncharge or migration. In the
3738 * first case the page is on the way out and it will vanish from the LRU
3739 * on the next attempt and the call should be retried later.
3740 * Isolation from the LRU fails only if page has been isolated from
3741 * the LRU since we looked at it and that usually means either global
3742 * reclaim or migration going on. The page will either get back to the
3744 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3745 * (!PageCgroupUsed) or moved to a different group. The page will
3746 * disappear in the next attempt.
3748 static int mem_cgroup_move_parent(struct page *page,
3749 struct page_cgroup *pc,
3750 struct mem_cgroup *child)
3752 struct mem_cgroup *parent;
3753 unsigned int nr_pages;
3754 unsigned long uninitialized_var(flags);
3757 VM_BUG_ON(mem_cgroup_is_root(child));
3760 if (!get_page_unless_zero(page))
3762 if (isolate_lru_page(page))
3765 nr_pages = hpage_nr_pages(page);
3767 parent = parent_mem_cgroup(child);
3769 * If no parent, move charges to root cgroup.
3772 parent = root_mem_cgroup;
3775 VM_BUG_ON(!PageTransHuge(page));
3776 flags = compound_lock_irqsave(page);
3779 ret = mem_cgroup_move_account(page, nr_pages,
3782 __mem_cgroup_cancel_local_charge(child, nr_pages);
3785 compound_unlock_irqrestore(page, flags);
3786 putback_lru_page(page);
3794 * Charge the memory controller for page usage.
3796 * 0 if the charge was successful
3797 * < 0 if the cgroup is over its limit
3799 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3800 gfp_t gfp_mask, enum charge_type ctype)
3802 struct mem_cgroup *memcg = NULL;
3803 unsigned int nr_pages = 1;
3807 if (PageTransHuge(page)) {
3808 nr_pages <<= compound_order(page);
3809 VM_BUG_ON(!PageTransHuge(page));
3811 * Never OOM-kill a process for a huge page. The
3812 * fault handler will fall back to regular pages.
3817 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3820 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3824 int mem_cgroup_newpage_charge(struct page *page,
3825 struct mm_struct *mm, gfp_t gfp_mask)
3827 if (mem_cgroup_disabled())
3829 VM_BUG_ON(page_mapped(page));
3830 VM_BUG_ON(page->mapping && !PageAnon(page));
3832 return mem_cgroup_charge_common(page, mm, gfp_mask,
3833 MEM_CGROUP_CHARGE_TYPE_ANON);
3837 * While swap-in, try_charge -> commit or cancel, the page is locked.
3838 * And when try_charge() successfully returns, one refcnt to memcg without
3839 * struct page_cgroup is acquired. This refcnt will be consumed by
3840 * "commit()" or removed by "cancel()"
3842 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3845 struct mem_cgroup **memcgp)
3847 struct mem_cgroup *memcg;
3848 struct page_cgroup *pc;
3851 pc = lookup_page_cgroup(page);
3853 * Every swap fault against a single page tries to charge the
3854 * page, bail as early as possible. shmem_unuse() encounters
3855 * already charged pages, too. The USED bit is protected by
3856 * the page lock, which serializes swap cache removal, which
3857 * in turn serializes uncharging.
3859 if (PageCgroupUsed(pc))
3861 if (!do_swap_account)
3863 memcg = try_get_mem_cgroup_from_page(page);
3867 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3868 css_put(&memcg->css);
3873 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3879 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3880 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3883 if (mem_cgroup_disabled())
3886 * A racing thread's fault, or swapoff, may have already
3887 * updated the pte, and even removed page from swap cache: in
3888 * those cases unuse_pte()'s pte_same() test will fail; but
3889 * there's also a KSM case which does need to charge the page.
3891 if (!PageSwapCache(page)) {
3894 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3899 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3902 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3904 if (mem_cgroup_disabled())
3908 __mem_cgroup_cancel_charge(memcg, 1);
3912 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3913 enum charge_type ctype)
3915 if (mem_cgroup_disabled())
3920 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3922 * Now swap is on-memory. This means this page may be
3923 * counted both as mem and swap....double count.
3924 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3925 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3926 * may call delete_from_swap_cache() before reach here.
3928 if (do_swap_account && PageSwapCache(page)) {
3929 swp_entry_t ent = {.val = page_private(page)};
3930 mem_cgroup_uncharge_swap(ent);
3934 void mem_cgroup_commit_charge_swapin(struct page *page,
3935 struct mem_cgroup *memcg)
3937 __mem_cgroup_commit_charge_swapin(page, memcg,
3938 MEM_CGROUP_CHARGE_TYPE_ANON);
3941 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
3944 struct mem_cgroup *memcg = NULL;
3945 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3948 if (mem_cgroup_disabled())
3950 if (PageCompound(page))
3953 if (!PageSwapCache(page))
3954 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
3955 else { /* page is swapcache/shmem */
3956 ret = __mem_cgroup_try_charge_swapin(mm, page,
3959 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3964 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3965 unsigned int nr_pages,
3966 const enum charge_type ctype)
3968 struct memcg_batch_info *batch = NULL;
3969 bool uncharge_memsw = true;
3971 /* If swapout, usage of swap doesn't decrease */
3972 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3973 uncharge_memsw = false;
3975 batch = ¤t->memcg_batch;
3977 * In usual, we do css_get() when we remember memcg pointer.
3978 * But in this case, we keep res->usage until end of a series of
3979 * uncharges. Then, it's ok to ignore memcg's refcnt.
3982 batch->memcg = memcg;
3984 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3985 * In those cases, all pages freed continuously can be expected to be in
3986 * the same cgroup and we have chance to coalesce uncharges.
3987 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3988 * because we want to do uncharge as soon as possible.
3991 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3992 goto direct_uncharge;
3995 goto direct_uncharge;
3998 * In typical case, batch->memcg == mem. This means we can
3999 * merge a series of uncharges to an uncharge of res_counter.
4000 * If not, we uncharge res_counter ony by one.
4002 if (batch->memcg != memcg)
4003 goto direct_uncharge;
4004 /* remember freed charge and uncharge it later */
4007 batch->memsw_nr_pages++;
4010 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4012 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4013 if (unlikely(batch->memcg != memcg))
4014 memcg_oom_recover(memcg);
4018 * uncharge if !page_mapped(page)
4020 static struct mem_cgroup *
4021 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4024 struct mem_cgroup *memcg = NULL;
4025 unsigned int nr_pages = 1;
4026 struct page_cgroup *pc;
4029 if (mem_cgroup_disabled())
4032 if (PageTransHuge(page)) {
4033 nr_pages <<= compound_order(page);
4034 VM_BUG_ON(!PageTransHuge(page));
4037 * Check if our page_cgroup is valid
4039 pc = lookup_page_cgroup(page);
4040 if (unlikely(!PageCgroupUsed(pc)))
4043 lock_page_cgroup(pc);
4045 memcg = pc->mem_cgroup;
4047 if (!PageCgroupUsed(pc))
4050 anon = PageAnon(page);
4053 case MEM_CGROUP_CHARGE_TYPE_ANON:
4055 * Generally PageAnon tells if it's the anon statistics to be
4056 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4057 * used before page reached the stage of being marked PageAnon.
4061 case MEM_CGROUP_CHARGE_TYPE_DROP:
4062 /* See mem_cgroup_prepare_migration() */
4063 if (page_mapped(page))
4066 * Pages under migration may not be uncharged. But
4067 * end_migration() /must/ be the one uncharging the
4068 * unused post-migration page and so it has to call
4069 * here with the migration bit still set. See the
4070 * res_counter handling below.
4072 if (!end_migration && PageCgroupMigration(pc))
4075 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4076 if (!PageAnon(page)) { /* Shared memory */
4077 if (page->mapping && !page_is_file_cache(page))
4079 } else if (page_mapped(page)) /* Anon */
4086 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4088 ClearPageCgroupUsed(pc);
4090 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4091 * freed from LRU. This is safe because uncharged page is expected not
4092 * to be reused (freed soon). Exception is SwapCache, it's handled by
4093 * special functions.
4096 unlock_page_cgroup(pc);
4098 * even after unlock, we have memcg->res.usage here and this memcg
4099 * will never be freed, so it's safe to call css_get().
4101 memcg_check_events(memcg, page);
4102 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4103 mem_cgroup_swap_statistics(memcg, true);
4104 css_get(&memcg->css);
4107 * Migration does not charge the res_counter for the
4108 * replacement page, so leave it alone when phasing out the
4109 * page that is unused after the migration.
4111 if (!end_migration && !mem_cgroup_is_root(memcg))
4112 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4117 unlock_page_cgroup(pc);
4121 void mem_cgroup_uncharge_page(struct page *page)
4124 if (page_mapped(page))
4126 VM_BUG_ON(page->mapping && !PageAnon(page));
4128 * If the page is in swap cache, uncharge should be deferred
4129 * to the swap path, which also properly accounts swap usage
4130 * and handles memcg lifetime.
4132 * Note that this check is not stable and reclaim may add the
4133 * page to swap cache at any time after this. However, if the
4134 * page is not in swap cache by the time page->mapcount hits
4135 * 0, there won't be any page table references to the swap
4136 * slot, and reclaim will free it and not actually write the
4139 if (PageSwapCache(page))
4141 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4144 void mem_cgroup_uncharge_cache_page(struct page *page)
4146 VM_BUG_ON(page_mapped(page));
4147 VM_BUG_ON(page->mapping);
4148 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4152 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4153 * In that cases, pages are freed continuously and we can expect pages
4154 * are in the same memcg. All these calls itself limits the number of
4155 * pages freed at once, then uncharge_start/end() is called properly.
4156 * This may be called prural(2) times in a context,
4159 void mem_cgroup_uncharge_start(void)
4161 current->memcg_batch.do_batch++;
4162 /* We can do nest. */
4163 if (current->memcg_batch.do_batch == 1) {
4164 current->memcg_batch.memcg = NULL;
4165 current->memcg_batch.nr_pages = 0;
4166 current->memcg_batch.memsw_nr_pages = 0;
4170 void mem_cgroup_uncharge_end(void)
4172 struct memcg_batch_info *batch = ¤t->memcg_batch;
4174 if (!batch->do_batch)
4178 if (batch->do_batch) /* If stacked, do nothing. */
4184 * This "batch->memcg" is valid without any css_get/put etc...
4185 * bacause we hide charges behind us.
4187 if (batch->nr_pages)
4188 res_counter_uncharge(&batch->memcg->res,
4189 batch->nr_pages * PAGE_SIZE);
4190 if (batch->memsw_nr_pages)
4191 res_counter_uncharge(&batch->memcg->memsw,
4192 batch->memsw_nr_pages * PAGE_SIZE);
4193 memcg_oom_recover(batch->memcg);
4194 /* forget this pointer (for sanity check) */
4195 batch->memcg = NULL;
4200 * called after __delete_from_swap_cache() and drop "page" account.
4201 * memcg information is recorded to swap_cgroup of "ent"
4204 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4206 struct mem_cgroup *memcg;
4207 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4209 if (!swapout) /* this was a swap cache but the swap is unused ! */
4210 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4212 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4215 * record memcg information, if swapout && memcg != NULL,
4216 * css_get() was called in uncharge().
4218 if (do_swap_account && swapout && memcg)
4219 swap_cgroup_record(ent, css_id(&memcg->css));
4223 #ifdef CONFIG_MEMCG_SWAP
4225 * called from swap_entry_free(). remove record in swap_cgroup and
4226 * uncharge "memsw" account.
4228 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4230 struct mem_cgroup *memcg;
4233 if (!do_swap_account)
4236 id = swap_cgroup_record(ent, 0);
4238 memcg = mem_cgroup_lookup(id);
4241 * We uncharge this because swap is freed.
4242 * This memcg can be obsolete one. We avoid calling css_tryget
4244 if (!mem_cgroup_is_root(memcg))
4245 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4246 mem_cgroup_swap_statistics(memcg, false);
4247 css_put(&memcg->css);
4253 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4254 * @entry: swap entry to be moved
4255 * @from: mem_cgroup which the entry is moved from
4256 * @to: mem_cgroup which the entry is moved to
4258 * It succeeds only when the swap_cgroup's record for this entry is the same
4259 * as the mem_cgroup's id of @from.
4261 * Returns 0 on success, -EINVAL on failure.
4263 * The caller must have charged to @to, IOW, called res_counter_charge() about
4264 * both res and memsw, and called css_get().
4266 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4267 struct mem_cgroup *from, struct mem_cgroup *to)
4269 unsigned short old_id, new_id;
4271 old_id = css_id(&from->css);
4272 new_id = css_id(&to->css);
4274 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4275 mem_cgroup_swap_statistics(from, false);
4276 mem_cgroup_swap_statistics(to, true);
4278 * This function is only called from task migration context now.
4279 * It postpones res_counter and refcount handling till the end
4280 * of task migration(mem_cgroup_clear_mc()) for performance
4281 * improvement. But we cannot postpone css_get(to) because if
4282 * the process that has been moved to @to does swap-in, the
4283 * refcount of @to might be decreased to 0.
4285 * We are in attach() phase, so the cgroup is guaranteed to be
4286 * alive, so we can just call css_get().
4294 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4295 struct mem_cgroup *from, struct mem_cgroup *to)
4302 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4305 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4306 struct mem_cgroup **memcgp)
4308 struct mem_cgroup *memcg = NULL;
4309 unsigned int nr_pages = 1;
4310 struct page_cgroup *pc;
4311 enum charge_type ctype;
4315 if (mem_cgroup_disabled())
4318 if (PageTransHuge(page))
4319 nr_pages <<= compound_order(page);
4321 pc = lookup_page_cgroup(page);
4322 lock_page_cgroup(pc);
4323 if (PageCgroupUsed(pc)) {
4324 memcg = pc->mem_cgroup;
4325 css_get(&memcg->css);
4327 * At migrating an anonymous page, its mapcount goes down
4328 * to 0 and uncharge() will be called. But, even if it's fully
4329 * unmapped, migration may fail and this page has to be
4330 * charged again. We set MIGRATION flag here and delay uncharge
4331 * until end_migration() is called
4333 * Corner Case Thinking
4335 * When the old page was mapped as Anon and it's unmap-and-freed
4336 * while migration was ongoing.
4337 * If unmap finds the old page, uncharge() of it will be delayed
4338 * until end_migration(). If unmap finds a new page, it's
4339 * uncharged when it make mapcount to be 1->0. If unmap code
4340 * finds swap_migration_entry, the new page will not be mapped
4341 * and end_migration() will find it(mapcount==0).
4344 * When the old page was mapped but migraion fails, the kernel
4345 * remaps it. A charge for it is kept by MIGRATION flag even
4346 * if mapcount goes down to 0. We can do remap successfully
4347 * without charging it again.
4350 * The "old" page is under lock_page() until the end of
4351 * migration, so, the old page itself will not be swapped-out.
4352 * If the new page is swapped out before end_migraton, our
4353 * hook to usual swap-out path will catch the event.
4356 SetPageCgroupMigration(pc);
4358 unlock_page_cgroup(pc);
4360 * If the page is not charged at this point,
4368 * We charge new page before it's used/mapped. So, even if unlock_page()
4369 * is called before end_migration, we can catch all events on this new
4370 * page. In the case new page is migrated but not remapped, new page's
4371 * mapcount will be finally 0 and we call uncharge in end_migration().
4374 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4376 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4378 * The page is committed to the memcg, but it's not actually
4379 * charged to the res_counter since we plan on replacing the
4380 * old one and only one page is going to be left afterwards.
4382 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4385 /* remove redundant charge if migration failed*/
4386 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4387 struct page *oldpage, struct page *newpage, bool migration_ok)
4389 struct page *used, *unused;
4390 struct page_cgroup *pc;
4396 if (!migration_ok) {
4403 anon = PageAnon(used);
4404 __mem_cgroup_uncharge_common(unused,
4405 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4406 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4408 css_put(&memcg->css);
4410 * We disallowed uncharge of pages under migration because mapcount
4411 * of the page goes down to zero, temporarly.
4412 * Clear the flag and check the page should be charged.
4414 pc = lookup_page_cgroup(oldpage);
4415 lock_page_cgroup(pc);
4416 ClearPageCgroupMigration(pc);
4417 unlock_page_cgroup(pc);
4420 * If a page is a file cache, radix-tree replacement is very atomic
4421 * and we can skip this check. When it was an Anon page, its mapcount
4422 * goes down to 0. But because we added MIGRATION flage, it's not
4423 * uncharged yet. There are several case but page->mapcount check
4424 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4425 * check. (see prepare_charge() also)
4428 mem_cgroup_uncharge_page(used);
4432 * At replace page cache, newpage is not under any memcg but it's on
4433 * LRU. So, this function doesn't touch res_counter but handles LRU
4434 * in correct way. Both pages are locked so we cannot race with uncharge.
4436 void mem_cgroup_replace_page_cache(struct page *oldpage,
4437 struct page *newpage)
4439 struct mem_cgroup *memcg = NULL;
4440 struct page_cgroup *pc;
4441 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4443 if (mem_cgroup_disabled())
4446 pc = lookup_page_cgroup(oldpage);
4447 /* fix accounting on old pages */
4448 lock_page_cgroup(pc);
4449 if (PageCgroupUsed(pc)) {
4450 memcg = pc->mem_cgroup;
4451 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4452 ClearPageCgroupUsed(pc);
4454 unlock_page_cgroup(pc);
4457 * When called from shmem_replace_page(), in some cases the
4458 * oldpage has already been charged, and in some cases not.
4463 * Even if newpage->mapping was NULL before starting replacement,
4464 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4465 * LRU while we overwrite pc->mem_cgroup.
4467 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4470 #ifdef CONFIG_DEBUG_VM
4471 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4473 struct page_cgroup *pc;
4475 pc = lookup_page_cgroup(page);
4477 * Can be NULL while feeding pages into the page allocator for
4478 * the first time, i.e. during boot or memory hotplug;
4479 * or when mem_cgroup_disabled().
4481 if (likely(pc) && PageCgroupUsed(pc))
4486 bool mem_cgroup_bad_page_check(struct page *page)
4488 if (mem_cgroup_disabled())
4491 return lookup_page_cgroup_used(page) != NULL;
4494 void mem_cgroup_print_bad_page(struct page *page)
4496 struct page_cgroup *pc;
4498 pc = lookup_page_cgroup_used(page);
4500 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4501 pc, pc->flags, pc->mem_cgroup);
4506 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4507 unsigned long long val)
4510 u64 memswlimit, memlimit;
4512 int children = mem_cgroup_count_children(memcg);
4513 u64 curusage, oldusage;
4517 * For keeping hierarchical_reclaim simple, how long we should retry
4518 * is depends on callers. We set our retry-count to be function
4519 * of # of children which we should visit in this loop.
4521 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4523 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4526 while (retry_count) {
4527 if (signal_pending(current)) {
4532 * Rather than hide all in some function, I do this in
4533 * open coded manner. You see what this really does.
4534 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4536 mutex_lock(&set_limit_mutex);
4537 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4538 if (memswlimit < val) {
4540 mutex_unlock(&set_limit_mutex);
4544 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4548 ret = res_counter_set_limit(&memcg->res, val);
4550 if (memswlimit == val)
4551 memcg->memsw_is_minimum = true;
4553 memcg->memsw_is_minimum = false;
4555 mutex_unlock(&set_limit_mutex);
4560 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4561 MEM_CGROUP_RECLAIM_SHRINK);
4562 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4563 /* Usage is reduced ? */
4564 if (curusage >= oldusage)
4567 oldusage = curusage;
4569 if (!ret && enlarge)
4570 memcg_oom_recover(memcg);
4575 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4576 unsigned long long val)
4579 u64 memlimit, memswlimit, oldusage, curusage;
4580 int children = mem_cgroup_count_children(memcg);
4584 /* see mem_cgroup_resize_res_limit */
4585 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4586 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4587 while (retry_count) {
4588 if (signal_pending(current)) {
4593 * Rather than hide all in some function, I do this in
4594 * open coded manner. You see what this really does.
4595 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4597 mutex_lock(&set_limit_mutex);
4598 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4599 if (memlimit > val) {
4601 mutex_unlock(&set_limit_mutex);
4604 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4605 if (memswlimit < val)
4607 ret = res_counter_set_limit(&memcg->memsw, val);
4609 if (memlimit == val)
4610 memcg->memsw_is_minimum = true;
4612 memcg->memsw_is_minimum = false;
4614 mutex_unlock(&set_limit_mutex);
4619 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4620 MEM_CGROUP_RECLAIM_NOSWAP |
4621 MEM_CGROUP_RECLAIM_SHRINK);
4622 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4623 /* Usage is reduced ? */
4624 if (curusage >= oldusage)
4627 oldusage = curusage;
4629 if (!ret && enlarge)
4630 memcg_oom_recover(memcg);
4635 * mem_cgroup_force_empty_list - clears LRU of a group
4636 * @memcg: group to clear
4639 * @lru: lru to to clear
4641 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4642 * reclaim the pages page themselves - pages are moved to the parent (or root)
4645 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4646 int node, int zid, enum lru_list lru)
4648 struct lruvec *lruvec;
4649 unsigned long flags;
4650 struct list_head *list;
4654 zone = &NODE_DATA(node)->node_zones[zid];
4655 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4656 list = &lruvec->lists[lru];
4660 struct page_cgroup *pc;
4663 spin_lock_irqsave(&zone->lru_lock, flags);
4664 if (list_empty(list)) {
4665 spin_unlock_irqrestore(&zone->lru_lock, flags);
4668 page = list_entry(list->prev, struct page, lru);
4670 list_move(&page->lru, list);
4672 spin_unlock_irqrestore(&zone->lru_lock, flags);
4675 spin_unlock_irqrestore(&zone->lru_lock, flags);
4677 pc = lookup_page_cgroup(page);
4679 if (mem_cgroup_move_parent(page, pc, memcg)) {
4680 /* found lock contention or "pc" is obsolete. */
4685 } while (!list_empty(list));
4689 * make mem_cgroup's charge to be 0 if there is no task by moving
4690 * all the charges and pages to the parent.
4691 * This enables deleting this mem_cgroup.
4693 * Caller is responsible for holding css reference on the memcg.
4695 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4701 /* This is for making all *used* pages to be on LRU. */
4702 lru_add_drain_all();
4703 drain_all_stock_sync(memcg);
4704 mem_cgroup_start_move(memcg);
4705 for_each_node_state(node, N_MEMORY) {
4706 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4709 mem_cgroup_force_empty_list(memcg,
4714 mem_cgroup_end_move(memcg);
4715 memcg_oom_recover(memcg);
4719 * Kernel memory may not necessarily be trackable to a specific
4720 * process. So they are not migrated, and therefore we can't
4721 * expect their value to drop to 0 here.
4722 * Having res filled up with kmem only is enough.
4724 * This is a safety check because mem_cgroup_force_empty_list
4725 * could have raced with mem_cgroup_replace_page_cache callers
4726 * so the lru seemed empty but the page could have been added
4727 * right after the check. RES_USAGE should be safe as we always
4728 * charge before adding to the LRU.
4730 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4731 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4732 } while (usage > 0);
4736 * This mainly exists for tests during the setting of set of use_hierarchy.
4737 * Since this is the very setting we are changing, the current hierarchy value
4740 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4742 struct cgroup_subsys_state *pos;
4744 /* bounce at first found */
4745 css_for_each_child(pos, &memcg->css)
4751 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4752 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4753 * from mem_cgroup_count_children(), in the sense that we don't really care how
4754 * many children we have; we only need to know if we have any. It also counts
4755 * any memcg without hierarchy as infertile.
4757 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4759 return memcg->use_hierarchy && __memcg_has_children(memcg);
4763 * Reclaims as many pages from the given memcg as possible and moves
4764 * the rest to the parent.
4766 * Caller is responsible for holding css reference for memcg.
4768 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4770 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4771 struct cgroup *cgrp = memcg->css.cgroup;
4773 /* returns EBUSY if there is a task or if we come here twice. */
4774 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4777 /* we call try-to-free pages for make this cgroup empty */
4778 lru_add_drain_all();
4779 /* try to free all pages in this cgroup */
4780 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4783 if (signal_pending(current))
4786 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4790 /* maybe some writeback is necessary */
4791 congestion_wait(BLK_RW_ASYNC, HZ/10);
4796 mem_cgroup_reparent_charges(memcg);
4801 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4804 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4806 if (mem_cgroup_is_root(memcg))
4808 return mem_cgroup_force_empty(memcg);
4811 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4814 return mem_cgroup_from_css(css)->use_hierarchy;
4817 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4818 struct cftype *cft, u64 val)
4821 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4822 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4824 mutex_lock(&memcg_create_mutex);
4826 if (memcg->use_hierarchy == val)
4830 * If parent's use_hierarchy is set, we can't make any modifications
4831 * in the child subtrees. If it is unset, then the change can
4832 * occur, provided the current cgroup has no children.
4834 * For the root cgroup, parent_mem is NULL, we allow value to be
4835 * set if there are no children.
4837 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4838 (val == 1 || val == 0)) {
4839 if (!__memcg_has_children(memcg))
4840 memcg->use_hierarchy = val;
4847 mutex_unlock(&memcg_create_mutex);
4853 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4854 enum mem_cgroup_stat_index idx)
4856 struct mem_cgroup *iter;
4859 /* Per-cpu values can be negative, use a signed accumulator */
4860 for_each_mem_cgroup_tree(iter, memcg)
4861 val += mem_cgroup_read_stat(iter, idx);
4863 if (val < 0) /* race ? */
4868 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4872 if (!mem_cgroup_is_root(memcg)) {
4874 return res_counter_read_u64(&memcg->res, RES_USAGE);
4876 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4880 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4881 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4883 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4884 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4887 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4889 return val << PAGE_SHIFT;
4892 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
4893 struct cftype *cft, struct file *file,
4894 char __user *buf, size_t nbytes, loff_t *ppos)
4896 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4902 type = MEMFILE_TYPE(cft->private);
4903 name = MEMFILE_ATTR(cft->private);
4907 if (name == RES_USAGE)
4908 val = mem_cgroup_usage(memcg, false);
4910 val = res_counter_read_u64(&memcg->res, name);
4913 if (name == RES_USAGE)
4914 val = mem_cgroup_usage(memcg, true);
4916 val = res_counter_read_u64(&memcg->memsw, name);
4919 val = res_counter_read_u64(&memcg->kmem, name);
4925 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4926 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4929 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
4932 #ifdef CONFIG_MEMCG_KMEM
4933 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4935 * For simplicity, we won't allow this to be disabled. It also can't
4936 * be changed if the cgroup has children already, or if tasks had
4939 * If tasks join before we set the limit, a person looking at
4940 * kmem.usage_in_bytes will have no way to determine when it took
4941 * place, which makes the value quite meaningless.
4943 * After it first became limited, changes in the value of the limit are
4944 * of course permitted.
4946 mutex_lock(&memcg_create_mutex);
4947 mutex_lock(&set_limit_mutex);
4948 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
4949 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
4953 ret = res_counter_set_limit(&memcg->kmem, val);
4956 ret = memcg_update_cache_sizes(memcg);
4958 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
4961 static_key_slow_inc(&memcg_kmem_enabled_key);
4963 * setting the active bit after the inc will guarantee no one
4964 * starts accounting before all call sites are patched
4966 memcg_kmem_set_active(memcg);
4968 ret = res_counter_set_limit(&memcg->kmem, val);
4970 mutex_unlock(&set_limit_mutex);
4971 mutex_unlock(&memcg_create_mutex);
4976 #ifdef CONFIG_MEMCG_KMEM
4977 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4980 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4984 memcg->kmem_account_flags = parent->kmem_account_flags;
4986 * When that happen, we need to disable the static branch only on those
4987 * memcgs that enabled it. To achieve this, we would be forced to
4988 * complicate the code by keeping track of which memcgs were the ones
4989 * that actually enabled limits, and which ones got it from its
4992 * It is a lot simpler just to do static_key_slow_inc() on every child
4993 * that is accounted.
4995 if (!memcg_kmem_is_active(memcg))
4999 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5000 * memcg is active already. If the later initialization fails then the
5001 * cgroup core triggers the cleanup so we do not have to do it here.
5003 static_key_slow_inc(&memcg_kmem_enabled_key);
5005 mutex_lock(&set_limit_mutex);
5006 memcg_stop_kmem_account();
5007 ret = memcg_update_cache_sizes(memcg);
5008 memcg_resume_kmem_account();
5009 mutex_unlock(&set_limit_mutex);
5013 #endif /* CONFIG_MEMCG_KMEM */
5016 * The user of this function is...
5019 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5022 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5025 unsigned long long val;
5028 type = MEMFILE_TYPE(cft->private);
5029 name = MEMFILE_ATTR(cft->private);
5033 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5037 /* This function does all necessary parse...reuse it */
5038 ret = res_counter_memparse_write_strategy(buffer, &val);
5042 ret = mem_cgroup_resize_limit(memcg, val);
5043 else if (type == _MEMSWAP)
5044 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5045 else if (type == _KMEM)
5046 ret = memcg_update_kmem_limit(css, val);
5050 case RES_SOFT_LIMIT:
5051 ret = res_counter_memparse_write_strategy(buffer, &val);
5055 * For memsw, soft limits are hard to implement in terms
5056 * of semantics, for now, we support soft limits for
5057 * control without swap
5060 ret = res_counter_set_soft_limit(&memcg->res, val);
5065 ret = -EINVAL; /* should be BUG() ? */
5071 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5072 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5074 unsigned long long min_limit, min_memsw_limit, tmp;
5076 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5077 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5078 if (!memcg->use_hierarchy)
5081 while (css_parent(&memcg->css)) {
5082 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5083 if (!memcg->use_hierarchy)
5085 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5086 min_limit = min(min_limit, tmp);
5087 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5088 min_memsw_limit = min(min_memsw_limit, tmp);
5091 *mem_limit = min_limit;
5092 *memsw_limit = min_memsw_limit;
5095 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5097 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5101 type = MEMFILE_TYPE(event);
5102 name = MEMFILE_ATTR(event);
5107 res_counter_reset_max(&memcg->res);
5108 else if (type == _MEMSWAP)
5109 res_counter_reset_max(&memcg->memsw);
5110 else if (type == _KMEM)
5111 res_counter_reset_max(&memcg->kmem);
5117 res_counter_reset_failcnt(&memcg->res);
5118 else if (type == _MEMSWAP)
5119 res_counter_reset_failcnt(&memcg->memsw);
5120 else if (type == _KMEM)
5121 res_counter_reset_failcnt(&memcg->kmem);
5130 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5133 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5137 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5138 struct cftype *cft, u64 val)
5140 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5142 if (val >= (1 << NR_MOVE_TYPE))
5146 * No kind of locking is needed in here, because ->can_attach() will
5147 * check this value once in the beginning of the process, and then carry
5148 * on with stale data. This means that changes to this value will only
5149 * affect task migrations starting after the change.
5151 memcg->move_charge_at_immigrate = val;
5155 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5156 struct cftype *cft, u64 val)
5163 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5164 struct cftype *cft, struct seq_file *m)
5167 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5168 unsigned long node_nr;
5169 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5171 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5172 seq_printf(m, "total=%lu", total_nr);
5173 for_each_node_state(nid, N_MEMORY) {
5174 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5175 seq_printf(m, " N%d=%lu", nid, node_nr);
5179 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5180 seq_printf(m, "file=%lu", file_nr);
5181 for_each_node_state(nid, N_MEMORY) {
5182 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5184 seq_printf(m, " N%d=%lu", nid, node_nr);
5188 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5189 seq_printf(m, "anon=%lu", anon_nr);
5190 for_each_node_state(nid, N_MEMORY) {
5191 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5193 seq_printf(m, " N%d=%lu", nid, node_nr);
5197 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5198 seq_printf(m, "unevictable=%lu", unevictable_nr);
5199 for_each_node_state(nid, N_MEMORY) {
5200 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5201 BIT(LRU_UNEVICTABLE));
5202 seq_printf(m, " N%d=%lu", nid, node_nr);
5207 #endif /* CONFIG_NUMA */
5209 static inline void mem_cgroup_lru_names_not_uptodate(void)
5211 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5214 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5217 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5218 struct mem_cgroup *mi;
5221 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5222 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5224 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5225 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5228 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5229 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5230 mem_cgroup_read_events(memcg, i));
5232 for (i = 0; i < NR_LRU_LISTS; i++)
5233 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5234 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5236 /* Hierarchical information */
5238 unsigned long long limit, memsw_limit;
5239 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5240 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5241 if (do_swap_account)
5242 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5246 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5249 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5251 for_each_mem_cgroup_tree(mi, memcg)
5252 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5253 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5256 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5257 unsigned long long val = 0;
5259 for_each_mem_cgroup_tree(mi, memcg)
5260 val += mem_cgroup_read_events(mi, i);
5261 seq_printf(m, "total_%s %llu\n",
5262 mem_cgroup_events_names[i], val);
5265 for (i = 0; i < NR_LRU_LISTS; i++) {
5266 unsigned long long val = 0;
5268 for_each_mem_cgroup_tree(mi, memcg)
5269 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5270 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5273 #ifdef CONFIG_DEBUG_VM
5276 struct mem_cgroup_per_zone *mz;
5277 struct zone_reclaim_stat *rstat;
5278 unsigned long recent_rotated[2] = {0, 0};
5279 unsigned long recent_scanned[2] = {0, 0};
5281 for_each_online_node(nid)
5282 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5283 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5284 rstat = &mz->lruvec.reclaim_stat;
5286 recent_rotated[0] += rstat->recent_rotated[0];
5287 recent_rotated[1] += rstat->recent_rotated[1];
5288 recent_scanned[0] += rstat->recent_scanned[0];
5289 recent_scanned[1] += rstat->recent_scanned[1];
5291 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5292 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5293 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5294 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5301 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5304 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5306 return mem_cgroup_swappiness(memcg);
5309 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5310 struct cftype *cft, u64 val)
5312 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5313 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5315 if (val > 100 || !parent)
5318 mutex_lock(&memcg_create_mutex);
5320 /* If under hierarchy, only empty-root can set this value */
5321 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5322 mutex_unlock(&memcg_create_mutex);
5326 memcg->swappiness = val;
5328 mutex_unlock(&memcg_create_mutex);
5333 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5335 struct mem_cgroup_threshold_ary *t;
5341 t = rcu_dereference(memcg->thresholds.primary);
5343 t = rcu_dereference(memcg->memsw_thresholds.primary);
5348 usage = mem_cgroup_usage(memcg, swap);
5351 * current_threshold points to threshold just below or equal to usage.
5352 * If it's not true, a threshold was crossed after last
5353 * call of __mem_cgroup_threshold().
5355 i = t->current_threshold;
5358 * Iterate backward over array of thresholds starting from
5359 * current_threshold and check if a threshold is crossed.
5360 * If none of thresholds below usage is crossed, we read
5361 * only one element of the array here.
5363 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5364 eventfd_signal(t->entries[i].eventfd, 1);
5366 /* i = current_threshold + 1 */
5370 * Iterate forward over array of thresholds starting from
5371 * current_threshold+1 and check if a threshold is crossed.
5372 * If none of thresholds above usage is crossed, we read
5373 * only one element of the array here.
5375 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5376 eventfd_signal(t->entries[i].eventfd, 1);
5378 /* Update current_threshold */
5379 t->current_threshold = i - 1;
5384 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5387 __mem_cgroup_threshold(memcg, false);
5388 if (do_swap_account)
5389 __mem_cgroup_threshold(memcg, true);
5391 memcg = parent_mem_cgroup(memcg);
5395 static int compare_thresholds(const void *a, const void *b)
5397 const struct mem_cgroup_threshold *_a = a;
5398 const struct mem_cgroup_threshold *_b = b;
5400 if (_a->threshold > _b->threshold)
5403 if (_a->threshold < _b->threshold)
5409 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5411 struct mem_cgroup_eventfd_list *ev;
5413 list_for_each_entry(ev, &memcg->oom_notify, list)
5414 eventfd_signal(ev->eventfd, 1);
5418 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5420 struct mem_cgroup *iter;
5422 for_each_mem_cgroup_tree(iter, memcg)
5423 mem_cgroup_oom_notify_cb(iter);
5426 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5427 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5429 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5430 struct mem_cgroup_thresholds *thresholds;
5431 struct mem_cgroup_threshold_ary *new;
5432 enum res_type type = MEMFILE_TYPE(cft->private);
5433 u64 threshold, usage;
5436 ret = res_counter_memparse_write_strategy(args, &threshold);
5440 mutex_lock(&memcg->thresholds_lock);
5443 thresholds = &memcg->thresholds;
5444 else if (type == _MEMSWAP)
5445 thresholds = &memcg->memsw_thresholds;
5449 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5451 /* Check if a threshold crossed before adding a new one */
5452 if (thresholds->primary)
5453 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5455 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5457 /* Allocate memory for new array of thresholds */
5458 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5466 /* Copy thresholds (if any) to new array */
5467 if (thresholds->primary) {
5468 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5469 sizeof(struct mem_cgroup_threshold));
5472 /* Add new threshold */
5473 new->entries[size - 1].eventfd = eventfd;
5474 new->entries[size - 1].threshold = threshold;
5476 /* Sort thresholds. Registering of new threshold isn't time-critical */
5477 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5478 compare_thresholds, NULL);
5480 /* Find current threshold */
5481 new->current_threshold = -1;
5482 for (i = 0; i < size; i++) {
5483 if (new->entries[i].threshold <= usage) {
5485 * new->current_threshold will not be used until
5486 * rcu_assign_pointer(), so it's safe to increment
5489 ++new->current_threshold;
5494 /* Free old spare buffer and save old primary buffer as spare */
5495 kfree(thresholds->spare);
5496 thresholds->spare = thresholds->primary;
5498 rcu_assign_pointer(thresholds->primary, new);
5500 /* To be sure that nobody uses thresholds */
5504 mutex_unlock(&memcg->thresholds_lock);
5509 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5510 struct cftype *cft, struct eventfd_ctx *eventfd)
5512 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5513 struct mem_cgroup_thresholds *thresholds;
5514 struct mem_cgroup_threshold_ary *new;
5515 enum res_type type = MEMFILE_TYPE(cft->private);
5519 mutex_lock(&memcg->thresholds_lock);
5521 thresholds = &memcg->thresholds;
5522 else if (type == _MEMSWAP)
5523 thresholds = &memcg->memsw_thresholds;
5527 if (!thresholds->primary)
5530 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5532 /* Check if a threshold crossed before removing */
5533 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5535 /* Calculate new number of threshold */
5537 for (i = 0; i < thresholds->primary->size; i++) {
5538 if (thresholds->primary->entries[i].eventfd != eventfd)
5542 new = thresholds->spare;
5544 /* Set thresholds array to NULL if we don't have thresholds */
5553 /* Copy thresholds and find current threshold */
5554 new->current_threshold = -1;
5555 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5556 if (thresholds->primary->entries[i].eventfd == eventfd)
5559 new->entries[j] = thresholds->primary->entries[i];
5560 if (new->entries[j].threshold <= usage) {
5562 * new->current_threshold will not be used
5563 * until rcu_assign_pointer(), so it's safe to increment
5566 ++new->current_threshold;
5572 /* Swap primary and spare array */
5573 thresholds->spare = thresholds->primary;
5574 /* If all events are unregistered, free the spare array */
5576 kfree(thresholds->spare);
5577 thresholds->spare = NULL;
5580 rcu_assign_pointer(thresholds->primary, new);
5582 /* To be sure that nobody uses thresholds */
5585 mutex_unlock(&memcg->thresholds_lock);
5588 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5589 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5591 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5592 struct mem_cgroup_eventfd_list *event;
5593 enum res_type type = MEMFILE_TYPE(cft->private);
5595 BUG_ON(type != _OOM_TYPE);
5596 event = kmalloc(sizeof(*event), GFP_KERNEL);
5600 spin_lock(&memcg_oom_lock);
5602 event->eventfd = eventfd;
5603 list_add(&event->list, &memcg->oom_notify);
5605 /* already in OOM ? */
5606 if (atomic_read(&memcg->under_oom))
5607 eventfd_signal(eventfd, 1);
5608 spin_unlock(&memcg_oom_lock);
5613 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5614 struct cftype *cft, struct eventfd_ctx *eventfd)
5616 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5617 struct mem_cgroup_eventfd_list *ev, *tmp;
5618 enum res_type type = MEMFILE_TYPE(cft->private);
5620 BUG_ON(type != _OOM_TYPE);
5622 spin_lock(&memcg_oom_lock);
5624 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5625 if (ev->eventfd == eventfd) {
5626 list_del(&ev->list);
5631 spin_unlock(&memcg_oom_lock);
5634 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5635 struct cftype *cft, struct cgroup_map_cb *cb)
5637 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5639 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5641 if (atomic_read(&memcg->under_oom))
5642 cb->fill(cb, "under_oom", 1);
5644 cb->fill(cb, "under_oom", 0);
5648 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5649 struct cftype *cft, u64 val)
5651 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5652 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5654 /* cannot set to root cgroup and only 0 and 1 are allowed */
5655 if (!parent || !((val == 0) || (val == 1)))
5658 mutex_lock(&memcg_create_mutex);
5659 /* oom-kill-disable is a flag for subhierarchy. */
5660 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5661 mutex_unlock(&memcg_create_mutex);
5664 memcg->oom_kill_disable = val;
5666 memcg_oom_recover(memcg);
5667 mutex_unlock(&memcg_create_mutex);
5671 #ifdef CONFIG_MEMCG_KMEM
5672 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5676 memcg->kmemcg_id = -1;
5677 ret = memcg_propagate_kmem(memcg);
5681 return mem_cgroup_sockets_init(memcg, ss);
5684 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5686 mem_cgroup_sockets_destroy(memcg);
5689 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5691 if (!memcg_kmem_is_active(memcg))
5695 * kmem charges can outlive the cgroup. In the case of slab
5696 * pages, for instance, a page contain objects from various
5697 * processes. As we prevent from taking a reference for every
5698 * such allocation we have to be careful when doing uncharge
5699 * (see memcg_uncharge_kmem) and here during offlining.
5701 * The idea is that that only the _last_ uncharge which sees
5702 * the dead memcg will drop the last reference. An additional
5703 * reference is taken here before the group is marked dead
5704 * which is then paired with css_put during uncharge resp. here.
5706 * Although this might sound strange as this path is called from
5707 * css_offline() when the referencemight have dropped down to 0
5708 * and shouldn't be incremented anymore (css_tryget would fail)
5709 * we do not have other options because of the kmem allocations
5712 css_get(&memcg->css);
5714 memcg_kmem_mark_dead(memcg);
5716 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5719 if (memcg_kmem_test_and_clear_dead(memcg))
5720 css_put(&memcg->css);
5723 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5728 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5732 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5737 static struct cftype mem_cgroup_files[] = {
5739 .name = "usage_in_bytes",
5740 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5741 .read = mem_cgroup_read,
5742 .register_event = mem_cgroup_usage_register_event,
5743 .unregister_event = mem_cgroup_usage_unregister_event,
5746 .name = "max_usage_in_bytes",
5747 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5748 .trigger = mem_cgroup_reset,
5749 .read = mem_cgroup_read,
5752 .name = "limit_in_bytes",
5753 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5754 .write_string = mem_cgroup_write,
5755 .read = mem_cgroup_read,
5758 .name = "soft_limit_in_bytes",
5759 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5760 .write_string = mem_cgroup_write,
5761 .read = mem_cgroup_read,
5765 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5766 .trigger = mem_cgroup_reset,
5767 .read = mem_cgroup_read,
5771 .read_seq_string = memcg_stat_show,
5774 .name = "force_empty",
5775 .trigger = mem_cgroup_force_empty_write,
5778 .name = "use_hierarchy",
5779 .flags = CFTYPE_INSANE,
5780 .write_u64 = mem_cgroup_hierarchy_write,
5781 .read_u64 = mem_cgroup_hierarchy_read,
5784 .name = "swappiness",
5785 .read_u64 = mem_cgroup_swappiness_read,
5786 .write_u64 = mem_cgroup_swappiness_write,
5789 .name = "move_charge_at_immigrate",
5790 .read_u64 = mem_cgroup_move_charge_read,
5791 .write_u64 = mem_cgroup_move_charge_write,
5794 .name = "oom_control",
5795 .read_map = mem_cgroup_oom_control_read,
5796 .write_u64 = mem_cgroup_oom_control_write,
5797 .register_event = mem_cgroup_oom_register_event,
5798 .unregister_event = mem_cgroup_oom_unregister_event,
5799 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5802 .name = "pressure_level",
5803 .register_event = vmpressure_register_event,
5804 .unregister_event = vmpressure_unregister_event,
5808 .name = "numa_stat",
5809 .read_seq_string = memcg_numa_stat_show,
5812 #ifdef CONFIG_MEMCG_KMEM
5814 .name = "kmem.limit_in_bytes",
5815 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5816 .write_string = mem_cgroup_write,
5817 .read = mem_cgroup_read,
5820 .name = "kmem.usage_in_bytes",
5821 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5822 .read = mem_cgroup_read,
5825 .name = "kmem.failcnt",
5826 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5827 .trigger = mem_cgroup_reset,
5828 .read = mem_cgroup_read,
5831 .name = "kmem.max_usage_in_bytes",
5832 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5833 .trigger = mem_cgroup_reset,
5834 .read = mem_cgroup_read,
5836 #ifdef CONFIG_SLABINFO
5838 .name = "kmem.slabinfo",
5839 .read_seq_string = mem_cgroup_slabinfo_read,
5843 { }, /* terminate */
5846 #ifdef CONFIG_MEMCG_SWAP
5847 static struct cftype memsw_cgroup_files[] = {
5849 .name = "memsw.usage_in_bytes",
5850 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5851 .read = mem_cgroup_read,
5852 .register_event = mem_cgroup_usage_register_event,
5853 .unregister_event = mem_cgroup_usage_unregister_event,
5856 .name = "memsw.max_usage_in_bytes",
5857 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5858 .trigger = mem_cgroup_reset,
5859 .read = mem_cgroup_read,
5862 .name = "memsw.limit_in_bytes",
5863 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5864 .write_string = mem_cgroup_write,
5865 .read = mem_cgroup_read,
5868 .name = "memsw.failcnt",
5869 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5870 .trigger = mem_cgroup_reset,
5871 .read = mem_cgroup_read,
5873 { }, /* terminate */
5876 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5878 struct mem_cgroup_per_node *pn;
5879 struct mem_cgroup_per_zone *mz;
5880 int zone, tmp = node;
5882 * This routine is called against possible nodes.
5883 * But it's BUG to call kmalloc() against offline node.
5885 * TODO: this routine can waste much memory for nodes which will
5886 * never be onlined. It's better to use memory hotplug callback
5889 if (!node_state(node, N_NORMAL_MEMORY))
5891 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5895 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5896 mz = &pn->zoneinfo[zone];
5897 lruvec_init(&mz->lruvec);
5900 memcg->nodeinfo[node] = pn;
5904 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5906 kfree(memcg->nodeinfo[node]);
5909 static struct mem_cgroup *mem_cgroup_alloc(void)
5911 struct mem_cgroup *memcg;
5912 size_t size = memcg_size();
5914 /* Can be very big if nr_node_ids is very big */
5915 if (size < PAGE_SIZE)
5916 memcg = kzalloc(size, GFP_KERNEL);
5918 memcg = vzalloc(size);
5923 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5926 spin_lock_init(&memcg->pcp_counter_lock);
5930 if (size < PAGE_SIZE)
5938 * At destroying mem_cgroup, references from swap_cgroup can remain.
5939 * (scanning all at force_empty is too costly...)
5941 * Instead of clearing all references at force_empty, we remember
5942 * the number of reference from swap_cgroup and free mem_cgroup when
5943 * it goes down to 0.
5945 * Removal of cgroup itself succeeds regardless of refs from swap.
5948 static void __mem_cgroup_free(struct mem_cgroup *memcg)
5951 size_t size = memcg_size();
5953 free_css_id(&mem_cgroup_subsys, &memcg->css);
5956 free_mem_cgroup_per_zone_info(memcg, node);
5958 free_percpu(memcg->stat);
5961 * We need to make sure that (at least for now), the jump label
5962 * destruction code runs outside of the cgroup lock. This is because
5963 * get_online_cpus(), which is called from the static_branch update,
5964 * can't be called inside the cgroup_lock. cpusets are the ones
5965 * enforcing this dependency, so if they ever change, we might as well.
5967 * schedule_work() will guarantee this happens. Be careful if you need
5968 * to move this code around, and make sure it is outside
5971 disarm_static_keys(memcg);
5972 if (size < PAGE_SIZE)
5979 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5981 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5983 if (!memcg->res.parent)
5985 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5987 EXPORT_SYMBOL(parent_mem_cgroup);
5989 static struct cgroup_subsys_state * __ref
5990 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
5992 struct mem_cgroup *memcg;
5993 long error = -ENOMEM;
5996 memcg = mem_cgroup_alloc();
5998 return ERR_PTR(error);
6001 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6005 if (parent_css == NULL) {
6006 root_mem_cgroup = memcg;
6007 res_counter_init(&memcg->res, NULL);
6008 res_counter_init(&memcg->memsw, NULL);
6009 res_counter_init(&memcg->kmem, NULL);
6012 memcg->last_scanned_node = MAX_NUMNODES;
6013 INIT_LIST_HEAD(&memcg->oom_notify);
6014 memcg->move_charge_at_immigrate = 0;
6015 mutex_init(&memcg->thresholds_lock);
6016 spin_lock_init(&memcg->move_lock);
6017 vmpressure_init(&memcg->vmpressure);
6018 spin_lock_init(&memcg->soft_lock);
6023 __mem_cgroup_free(memcg);
6024 return ERR_PTR(error);
6028 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6030 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6031 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6037 mutex_lock(&memcg_create_mutex);
6039 memcg->use_hierarchy = parent->use_hierarchy;
6040 memcg->oom_kill_disable = parent->oom_kill_disable;
6041 memcg->swappiness = mem_cgroup_swappiness(parent);
6043 if (parent->use_hierarchy) {
6044 res_counter_init(&memcg->res, &parent->res);
6045 res_counter_init(&memcg->memsw, &parent->memsw);
6046 res_counter_init(&memcg->kmem, &parent->kmem);
6049 * No need to take a reference to the parent because cgroup
6050 * core guarantees its existence.
6053 res_counter_init(&memcg->res, NULL);
6054 res_counter_init(&memcg->memsw, NULL);
6055 res_counter_init(&memcg->kmem, NULL);
6057 * Deeper hierachy with use_hierarchy == false doesn't make
6058 * much sense so let cgroup subsystem know about this
6059 * unfortunate state in our controller.
6061 if (parent != root_mem_cgroup)
6062 mem_cgroup_subsys.broken_hierarchy = true;
6065 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6066 mutex_unlock(&memcg_create_mutex);
6071 * Announce all parents that a group from their hierarchy is gone.
6073 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6075 struct mem_cgroup *parent = memcg;
6077 while ((parent = parent_mem_cgroup(parent)))
6078 mem_cgroup_iter_invalidate(parent);
6081 * if the root memcg is not hierarchical we have to check it
6084 if (!root_mem_cgroup->use_hierarchy)
6085 mem_cgroup_iter_invalidate(root_mem_cgroup);
6088 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6090 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6092 kmem_cgroup_css_offline(memcg);
6094 mem_cgroup_invalidate_reclaim_iterators(memcg);
6095 mem_cgroup_reparent_charges(memcg);
6096 if (memcg->soft_contributed) {
6097 while ((memcg = parent_mem_cgroup(memcg)))
6098 atomic_dec(&memcg->children_in_excess);
6100 if (memcg != root_mem_cgroup && !root_mem_cgroup->use_hierarchy)
6101 atomic_dec(&root_mem_cgroup->children_in_excess);
6103 mem_cgroup_destroy_all_caches(memcg);
6104 vmpressure_cleanup(&memcg->vmpressure);
6107 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6109 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6111 memcg_destroy_kmem(memcg);
6112 __mem_cgroup_free(memcg);
6116 /* Handlers for move charge at task migration. */
6117 #define PRECHARGE_COUNT_AT_ONCE 256
6118 static int mem_cgroup_do_precharge(unsigned long count)
6121 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6122 struct mem_cgroup *memcg = mc.to;
6124 if (mem_cgroup_is_root(memcg)) {
6125 mc.precharge += count;
6126 /* we don't need css_get for root */
6129 /* try to charge at once */
6131 struct res_counter *dummy;
6133 * "memcg" cannot be under rmdir() because we've already checked
6134 * by cgroup_lock_live_cgroup() that it is not removed and we
6135 * are still under the same cgroup_mutex. So we can postpone
6138 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6140 if (do_swap_account && res_counter_charge(&memcg->memsw,
6141 PAGE_SIZE * count, &dummy)) {
6142 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6145 mc.precharge += count;
6149 /* fall back to one by one charge */
6151 if (signal_pending(current)) {
6155 if (!batch_count--) {
6156 batch_count = PRECHARGE_COUNT_AT_ONCE;
6159 ret = __mem_cgroup_try_charge(NULL,
6160 GFP_KERNEL, 1, &memcg, false);
6162 /* mem_cgroup_clear_mc() will do uncharge later */
6170 * get_mctgt_type - get target type of moving charge
6171 * @vma: the vma the pte to be checked belongs
6172 * @addr: the address corresponding to the pte to be checked
6173 * @ptent: the pte to be checked
6174 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6177 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6178 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6179 * move charge. if @target is not NULL, the page is stored in target->page
6180 * with extra refcnt got(Callers should handle it).
6181 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6182 * target for charge migration. if @target is not NULL, the entry is stored
6185 * Called with pte lock held.
6192 enum mc_target_type {
6198 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6199 unsigned long addr, pte_t ptent)
6201 struct page *page = vm_normal_page(vma, addr, ptent);
6203 if (!page || !page_mapped(page))
6205 if (PageAnon(page)) {
6206 /* we don't move shared anon */
6209 } else if (!move_file())
6210 /* we ignore mapcount for file pages */
6212 if (!get_page_unless_zero(page))
6219 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6220 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6222 struct page *page = NULL;
6223 swp_entry_t ent = pte_to_swp_entry(ptent);
6225 if (!move_anon() || non_swap_entry(ent))
6228 * Because lookup_swap_cache() updates some statistics counter,
6229 * we call find_get_page() with swapper_space directly.
6231 page = find_get_page(swap_address_space(ent), ent.val);
6232 if (do_swap_account)
6233 entry->val = ent.val;
6238 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6239 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6245 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6246 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6248 struct page *page = NULL;
6249 struct address_space *mapping;
6252 if (!vma->vm_file) /* anonymous vma */
6257 mapping = vma->vm_file->f_mapping;
6258 if (pte_none(ptent))
6259 pgoff = linear_page_index(vma, addr);
6260 else /* pte_file(ptent) is true */
6261 pgoff = pte_to_pgoff(ptent);
6263 /* page is moved even if it's not RSS of this task(page-faulted). */
6264 page = find_get_page(mapping, pgoff);
6267 /* shmem/tmpfs may report page out on swap: account for that too. */
6268 if (radix_tree_exceptional_entry(page)) {
6269 swp_entry_t swap = radix_to_swp_entry(page);
6270 if (do_swap_account)
6272 page = find_get_page(swap_address_space(swap), swap.val);
6278 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6279 unsigned long addr, pte_t ptent, union mc_target *target)
6281 struct page *page = NULL;
6282 struct page_cgroup *pc;
6283 enum mc_target_type ret = MC_TARGET_NONE;
6284 swp_entry_t ent = { .val = 0 };
6286 if (pte_present(ptent))
6287 page = mc_handle_present_pte(vma, addr, ptent);
6288 else if (is_swap_pte(ptent))
6289 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6290 else if (pte_none(ptent) || pte_file(ptent))
6291 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6293 if (!page && !ent.val)
6296 pc = lookup_page_cgroup(page);
6298 * Do only loose check w/o page_cgroup lock.
6299 * mem_cgroup_move_account() checks the pc is valid or not under
6302 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6303 ret = MC_TARGET_PAGE;
6305 target->page = page;
6307 if (!ret || !target)
6310 /* There is a swap entry and a page doesn't exist or isn't charged */
6311 if (ent.val && !ret &&
6312 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6313 ret = MC_TARGET_SWAP;
6320 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6322 * We don't consider swapping or file mapped pages because THP does not
6323 * support them for now.
6324 * Caller should make sure that pmd_trans_huge(pmd) is true.
6326 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6327 unsigned long addr, pmd_t pmd, union mc_target *target)
6329 struct page *page = NULL;
6330 struct page_cgroup *pc;
6331 enum mc_target_type ret = MC_TARGET_NONE;
6333 page = pmd_page(pmd);
6334 VM_BUG_ON(!page || !PageHead(page));
6337 pc = lookup_page_cgroup(page);
6338 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6339 ret = MC_TARGET_PAGE;
6342 target->page = page;
6348 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6349 unsigned long addr, pmd_t pmd, union mc_target *target)
6351 return MC_TARGET_NONE;
6355 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6356 unsigned long addr, unsigned long end,
6357 struct mm_walk *walk)
6359 struct vm_area_struct *vma = walk->private;
6363 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6364 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6365 mc.precharge += HPAGE_PMD_NR;
6366 spin_unlock(&vma->vm_mm->page_table_lock);
6370 if (pmd_trans_unstable(pmd))
6372 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6373 for (; addr != end; pte++, addr += PAGE_SIZE)
6374 if (get_mctgt_type(vma, addr, *pte, NULL))
6375 mc.precharge++; /* increment precharge temporarily */
6376 pte_unmap_unlock(pte - 1, ptl);
6382 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6384 unsigned long precharge;
6385 struct vm_area_struct *vma;
6387 down_read(&mm->mmap_sem);
6388 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6389 struct mm_walk mem_cgroup_count_precharge_walk = {
6390 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6394 if (is_vm_hugetlb_page(vma))
6396 walk_page_range(vma->vm_start, vma->vm_end,
6397 &mem_cgroup_count_precharge_walk);
6399 up_read(&mm->mmap_sem);
6401 precharge = mc.precharge;
6407 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6409 unsigned long precharge = mem_cgroup_count_precharge(mm);
6411 VM_BUG_ON(mc.moving_task);
6412 mc.moving_task = current;
6413 return mem_cgroup_do_precharge(precharge);
6416 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6417 static void __mem_cgroup_clear_mc(void)
6419 struct mem_cgroup *from = mc.from;
6420 struct mem_cgroup *to = mc.to;
6423 /* we must uncharge all the leftover precharges from mc.to */
6425 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6429 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6430 * we must uncharge here.
6432 if (mc.moved_charge) {
6433 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6434 mc.moved_charge = 0;
6436 /* we must fixup refcnts and charges */
6437 if (mc.moved_swap) {
6438 /* uncharge swap account from the old cgroup */
6439 if (!mem_cgroup_is_root(mc.from))
6440 res_counter_uncharge(&mc.from->memsw,
6441 PAGE_SIZE * mc.moved_swap);
6443 for (i = 0; i < mc.moved_swap; i++)
6444 css_put(&mc.from->css);
6446 if (!mem_cgroup_is_root(mc.to)) {
6448 * we charged both to->res and to->memsw, so we should
6451 res_counter_uncharge(&mc.to->res,
6452 PAGE_SIZE * mc.moved_swap);
6454 /* we've already done css_get(mc.to) */
6457 memcg_oom_recover(from);
6458 memcg_oom_recover(to);
6459 wake_up_all(&mc.waitq);
6462 static void mem_cgroup_clear_mc(void)
6464 struct mem_cgroup *from = mc.from;
6467 * we must clear moving_task before waking up waiters at the end of
6470 mc.moving_task = NULL;
6471 __mem_cgroup_clear_mc();
6472 spin_lock(&mc.lock);
6475 spin_unlock(&mc.lock);
6476 mem_cgroup_end_move(from);
6479 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6480 struct cgroup_taskset *tset)
6482 struct task_struct *p = cgroup_taskset_first(tset);
6484 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6485 unsigned long move_charge_at_immigrate;
6488 * We are now commited to this value whatever it is. Changes in this
6489 * tunable will only affect upcoming migrations, not the current one.
6490 * So we need to save it, and keep it going.
6492 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6493 if (move_charge_at_immigrate) {
6494 struct mm_struct *mm;
6495 struct mem_cgroup *from = mem_cgroup_from_task(p);
6497 VM_BUG_ON(from == memcg);
6499 mm = get_task_mm(p);
6502 /* We move charges only when we move a owner of the mm */
6503 if (mm->owner == p) {
6506 VM_BUG_ON(mc.precharge);
6507 VM_BUG_ON(mc.moved_charge);
6508 VM_BUG_ON(mc.moved_swap);
6509 mem_cgroup_start_move(from);
6510 spin_lock(&mc.lock);
6513 mc.immigrate_flags = move_charge_at_immigrate;
6514 spin_unlock(&mc.lock);
6515 /* We set mc.moving_task later */
6517 ret = mem_cgroup_precharge_mc(mm);
6519 mem_cgroup_clear_mc();
6526 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6527 struct cgroup_taskset *tset)
6529 mem_cgroup_clear_mc();
6532 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6533 unsigned long addr, unsigned long end,
6534 struct mm_walk *walk)
6537 struct vm_area_struct *vma = walk->private;
6540 enum mc_target_type target_type;
6541 union mc_target target;
6543 struct page_cgroup *pc;
6546 * We don't take compound_lock() here but no race with splitting thp
6548 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6549 * under splitting, which means there's no concurrent thp split,
6550 * - if another thread runs into split_huge_page() just after we
6551 * entered this if-block, the thread must wait for page table lock
6552 * to be unlocked in __split_huge_page_splitting(), where the main
6553 * part of thp split is not executed yet.
6555 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6556 if (mc.precharge < HPAGE_PMD_NR) {
6557 spin_unlock(&vma->vm_mm->page_table_lock);
6560 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6561 if (target_type == MC_TARGET_PAGE) {
6563 if (!isolate_lru_page(page)) {
6564 pc = lookup_page_cgroup(page);
6565 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6566 pc, mc.from, mc.to)) {
6567 mc.precharge -= HPAGE_PMD_NR;
6568 mc.moved_charge += HPAGE_PMD_NR;
6570 putback_lru_page(page);
6574 spin_unlock(&vma->vm_mm->page_table_lock);
6578 if (pmd_trans_unstable(pmd))
6581 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6582 for (; addr != end; addr += PAGE_SIZE) {
6583 pte_t ptent = *(pte++);
6589 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6590 case MC_TARGET_PAGE:
6592 if (isolate_lru_page(page))
6594 pc = lookup_page_cgroup(page);
6595 if (!mem_cgroup_move_account(page, 1, pc,
6598 /* we uncharge from mc.from later. */
6601 putback_lru_page(page);
6602 put: /* get_mctgt_type() gets the page */
6605 case MC_TARGET_SWAP:
6607 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6609 /* we fixup refcnts and charges later. */
6617 pte_unmap_unlock(pte - 1, ptl);
6622 * We have consumed all precharges we got in can_attach().
6623 * We try charge one by one, but don't do any additional
6624 * charges to mc.to if we have failed in charge once in attach()
6627 ret = mem_cgroup_do_precharge(1);
6635 static void mem_cgroup_move_charge(struct mm_struct *mm)
6637 struct vm_area_struct *vma;
6639 lru_add_drain_all();
6641 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6643 * Someone who are holding the mmap_sem might be waiting in
6644 * waitq. So we cancel all extra charges, wake up all waiters,
6645 * and retry. Because we cancel precharges, we might not be able
6646 * to move enough charges, but moving charge is a best-effort
6647 * feature anyway, so it wouldn't be a big problem.
6649 __mem_cgroup_clear_mc();
6653 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6655 struct mm_walk mem_cgroup_move_charge_walk = {
6656 .pmd_entry = mem_cgroup_move_charge_pte_range,
6660 if (is_vm_hugetlb_page(vma))
6662 ret = walk_page_range(vma->vm_start, vma->vm_end,
6663 &mem_cgroup_move_charge_walk);
6666 * means we have consumed all precharges and failed in
6667 * doing additional charge. Just abandon here.
6671 up_read(&mm->mmap_sem);
6674 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6675 struct cgroup_taskset *tset)
6677 struct task_struct *p = cgroup_taskset_first(tset);
6678 struct mm_struct *mm = get_task_mm(p);
6682 mem_cgroup_move_charge(mm);
6686 mem_cgroup_clear_mc();
6688 #else /* !CONFIG_MMU */
6689 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6690 struct cgroup_taskset *tset)
6694 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6695 struct cgroup_taskset *tset)
6698 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6699 struct cgroup_taskset *tset)
6705 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6706 * to verify sane_behavior flag on each mount attempt.
6708 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6711 * use_hierarchy is forced with sane_behavior. cgroup core
6712 * guarantees that @root doesn't have any children, so turning it
6713 * on for the root memcg is enough.
6715 if (cgroup_sane_behavior(root_css->cgroup))
6716 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6719 struct cgroup_subsys mem_cgroup_subsys = {
6721 .subsys_id = mem_cgroup_subsys_id,
6722 .css_alloc = mem_cgroup_css_alloc,
6723 .css_online = mem_cgroup_css_online,
6724 .css_offline = mem_cgroup_css_offline,
6725 .css_free = mem_cgroup_css_free,
6726 .can_attach = mem_cgroup_can_attach,
6727 .cancel_attach = mem_cgroup_cancel_attach,
6728 .attach = mem_cgroup_move_task,
6729 .bind = mem_cgroup_bind,
6730 .base_cftypes = mem_cgroup_files,
6735 #ifdef CONFIG_MEMCG_SWAP
6736 static int __init enable_swap_account(char *s)
6738 if (!strcmp(s, "1"))
6739 really_do_swap_account = 1;
6740 else if (!strcmp(s, "0"))
6741 really_do_swap_account = 0;
6744 __setup("swapaccount=", enable_swap_account);
6746 static void __init memsw_file_init(void)
6748 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6751 static void __init enable_swap_cgroup(void)
6753 if (!mem_cgroup_disabled() && really_do_swap_account) {
6754 do_swap_account = 1;
6760 static void __init enable_swap_cgroup(void)
6766 * subsys_initcall() for memory controller.
6768 * Some parts like hotcpu_notifier() have to be initialized from this context
6769 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6770 * everything that doesn't depend on a specific mem_cgroup structure should
6771 * be initialized from here.
6773 static int __init mem_cgroup_init(void)
6775 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6776 enable_swap_cgroup();
6780 subsys_initcall(mem_cgroup_init);