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/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/sort.h>
50 #include <linux/seq_file.h>
51 #include <linux/vmalloc.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
60 #include <net/tcp_memcontrol.h>
62 #include <asm/uaccess.h>
64 #include <trace/events/vmscan.h>
66 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
67 EXPORT_SYMBOL(mem_cgroup_subsys);
69 #define MEM_CGROUP_RECLAIM_RETRIES 5
70 static struct mem_cgroup *root_mem_cgroup __read_mostly;
72 #ifdef CONFIG_MEMCG_SWAP
73 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
74 int do_swap_account __read_mostly;
76 /* for remember boot option*/
77 #ifdef CONFIG_MEMCG_SWAP_ENABLED
78 static int really_do_swap_account __initdata = 1;
80 static int really_do_swap_account __initdata = 0;
84 #define do_swap_account 0
89 * Statistics for memory cgroup.
91 enum mem_cgroup_stat_index {
93 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
95 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
96 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
97 MEM_CGROUP_STAT_RSS_HUGE, /* # of pages charged as anon huge */
98 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
99 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
100 MEM_CGROUP_STAT_NSTATS,
103 static const char * const mem_cgroup_stat_names[] = {
111 enum mem_cgroup_events_index {
112 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
113 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
114 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
115 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
116 MEM_CGROUP_EVENTS_NSTATS,
119 static const char * const mem_cgroup_events_names[] = {
126 static const char * const mem_cgroup_lru_names[] = {
135 * Per memcg event counter is incremented at every pagein/pageout. With THP,
136 * it will be incremated by the number of pages. This counter is used for
137 * for trigger some periodic events. This is straightforward and better
138 * than using jiffies etc. to handle periodic memcg event.
140 enum mem_cgroup_events_target {
141 MEM_CGROUP_TARGET_THRESH,
142 MEM_CGROUP_TARGET_SOFTLIMIT,
143 MEM_CGROUP_TARGET_NUMAINFO,
146 #define THRESHOLDS_EVENTS_TARGET 128
147 #define SOFTLIMIT_EVENTS_TARGET 1024
148 #define NUMAINFO_EVENTS_TARGET 1024
150 struct mem_cgroup_stat_cpu {
151 long count[MEM_CGROUP_STAT_NSTATS];
152 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
153 unsigned long nr_page_events;
154 unsigned long targets[MEM_CGROUP_NTARGETS];
157 struct mem_cgroup_reclaim_iter {
159 * last scanned hierarchy member. Valid only if last_dead_count
160 * matches memcg->dead_count of the hierarchy root group.
162 struct mem_cgroup *last_visited;
163 unsigned long last_dead_count;
165 /* scan generation, increased every round-trip */
166 unsigned int generation;
170 * per-zone information in memory controller.
172 struct mem_cgroup_per_zone {
173 struct lruvec lruvec;
174 unsigned long lru_size[NR_LRU_LISTS];
176 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
178 struct rb_node tree_node; /* RB tree node */
179 unsigned long long usage_in_excess;/* Set to the value by which */
180 /* the soft limit is exceeded*/
182 struct mem_cgroup *memcg; /* Back pointer, we cannot */
183 /* use container_of */
186 struct mem_cgroup_per_node {
187 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
191 * Cgroups above their limits are maintained in a RB-Tree, independent of
192 * their hierarchy representation
195 struct mem_cgroup_tree_per_zone {
196 struct rb_root rb_root;
200 struct mem_cgroup_tree_per_node {
201 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
204 struct mem_cgroup_tree {
205 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
208 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
210 struct mem_cgroup_threshold {
211 struct eventfd_ctx *eventfd;
216 struct mem_cgroup_threshold_ary {
217 /* An array index points to threshold just below or equal to usage. */
218 int current_threshold;
219 /* Size of entries[] */
221 /* Array of thresholds */
222 struct mem_cgroup_threshold entries[0];
225 struct mem_cgroup_thresholds {
226 /* Primary thresholds array */
227 struct mem_cgroup_threshold_ary *primary;
229 * Spare threshold array.
230 * This is needed to make mem_cgroup_unregister_event() "never fail".
231 * It must be able to store at least primary->size - 1 entries.
233 struct mem_cgroup_threshold_ary *spare;
237 struct mem_cgroup_eventfd_list {
238 struct list_head list;
239 struct eventfd_ctx *eventfd;
242 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
243 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
246 * The memory controller data structure. The memory controller controls both
247 * page cache and RSS per cgroup. We would eventually like to provide
248 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
249 * to help the administrator determine what knobs to tune.
251 * TODO: Add a water mark for the memory controller. Reclaim will begin when
252 * we hit the water mark. May be even add a low water mark, such that
253 * no reclaim occurs from a cgroup at it's low water mark, this is
254 * a feature that will be implemented much later in the future.
257 struct cgroup_subsys_state css;
259 * the counter to account for memory usage
261 struct res_counter res;
263 /* vmpressure notifications */
264 struct vmpressure vmpressure;
267 * the counter to account for mem+swap usage.
269 struct res_counter memsw;
272 * the counter to account for kernel memory usage.
274 struct res_counter kmem;
276 * Should the accounting and control be hierarchical, per subtree?
279 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
285 /* OOM-Killer disable */
286 int oom_kill_disable;
288 /* set when res.limit == memsw.limit */
289 bool memsw_is_minimum;
291 /* protect arrays of thresholds */
292 struct mutex thresholds_lock;
294 /* thresholds for memory usage. RCU-protected */
295 struct mem_cgroup_thresholds thresholds;
297 /* thresholds for mem+swap usage. RCU-protected */
298 struct mem_cgroup_thresholds memsw_thresholds;
300 /* For oom notifier event fd */
301 struct list_head oom_notify;
304 * Should we move charges of a task when a task is moved into this
305 * mem_cgroup ? And what type of charges should we move ?
307 unsigned long move_charge_at_immigrate;
309 * set > 0 if pages under this cgroup are moving to other cgroup.
311 atomic_t moving_account;
312 /* taken only while moving_account > 0 */
313 spinlock_t move_lock;
317 struct mem_cgroup_stat_cpu __percpu *stat;
319 * used when a cpu is offlined or other synchronizations
320 * See mem_cgroup_read_stat().
322 struct mem_cgroup_stat_cpu nocpu_base;
323 spinlock_t pcp_counter_lock;
326 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
327 struct tcp_memcontrol tcp_mem;
329 #if defined(CONFIG_MEMCG_KMEM)
330 /* analogous to slab_common's slab_caches list. per-memcg */
331 struct list_head memcg_slab_caches;
332 /* Not a spinlock, we can take a lot of time walking the list */
333 struct mutex slab_caches_mutex;
334 /* Index in the kmem_cache->memcg_params->memcg_caches array */
338 int last_scanned_node;
340 nodemask_t scan_nodes;
341 atomic_t numainfo_events;
342 atomic_t numainfo_updating;
345 struct mem_cgroup_per_node *nodeinfo[0];
346 /* WARNING: nodeinfo must be the last member here */
349 static size_t memcg_size(void)
351 return sizeof(struct mem_cgroup) +
352 nr_node_ids * sizeof(struct mem_cgroup_per_node);
355 /* internal only representation about the status of kmem accounting. */
357 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
358 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
359 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
362 /* We account when limit is on, but only after call sites are patched */
363 #define KMEM_ACCOUNTED_MASK \
364 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
366 #ifdef CONFIG_MEMCG_KMEM
367 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
369 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
372 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
374 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
377 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
379 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
382 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
384 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
387 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
390 * Our caller must use css_get() first, because memcg_uncharge_kmem()
391 * will call css_put() if it sees the memcg is dead.
394 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
395 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
398 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
400 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
401 &memcg->kmem_account_flags);
405 /* Stuffs for move charges at task migration. */
407 * Types of charges to be moved. "move_charge_at_immitgrate" and
408 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
411 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
412 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
416 /* "mc" and its members are protected by cgroup_mutex */
417 static struct move_charge_struct {
418 spinlock_t lock; /* for from, to */
419 struct mem_cgroup *from;
420 struct mem_cgroup *to;
421 unsigned long immigrate_flags;
422 unsigned long precharge;
423 unsigned long moved_charge;
424 unsigned long moved_swap;
425 struct task_struct *moving_task; /* a task moving charges */
426 wait_queue_head_t waitq; /* a waitq for other context */
428 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
429 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
432 static bool move_anon(void)
434 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
437 static bool move_file(void)
439 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
443 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
444 * limit reclaim to prevent infinite loops, if they ever occur.
446 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
447 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
450 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
451 MEM_CGROUP_CHARGE_TYPE_ANON,
452 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
453 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
457 /* for encoding cft->private value on file */
465 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
466 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
467 #define MEMFILE_ATTR(val) ((val) & 0xffff)
468 /* Used for OOM nofiier */
469 #define OOM_CONTROL (0)
472 * Reclaim flags for mem_cgroup_hierarchical_reclaim
474 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
475 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
476 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
477 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
480 * The memcg_create_mutex will be held whenever a new cgroup is created.
481 * As a consequence, any change that needs to protect against new child cgroups
482 * appearing has to hold it as well.
484 static DEFINE_MUTEX(memcg_create_mutex);
487 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
489 return s ? container_of(s, struct mem_cgroup, css) : NULL;
492 /* Some nice accessors for the vmpressure. */
493 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
496 memcg = root_mem_cgroup;
497 return &memcg->vmpressure;
500 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
502 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
505 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
507 return &mem_cgroup_from_css(css)->vmpressure;
510 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
512 return (memcg == root_mem_cgroup);
515 /* Writing them here to avoid exposing memcg's inner layout */
516 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
518 void sock_update_memcg(struct sock *sk)
520 if (mem_cgroup_sockets_enabled) {
521 struct mem_cgroup *memcg;
522 struct cg_proto *cg_proto;
524 BUG_ON(!sk->sk_prot->proto_cgroup);
526 /* Socket cloning can throw us here with sk_cgrp already
527 * filled. It won't however, necessarily happen from
528 * process context. So the test for root memcg given
529 * the current task's memcg won't help us in this case.
531 * Respecting the original socket's memcg is a better
532 * decision in this case.
535 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
536 css_get(&sk->sk_cgrp->memcg->css);
541 memcg = mem_cgroup_from_task(current);
542 cg_proto = sk->sk_prot->proto_cgroup(memcg);
543 if (!mem_cgroup_is_root(memcg) &&
544 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
545 sk->sk_cgrp = cg_proto;
550 EXPORT_SYMBOL(sock_update_memcg);
552 void sock_release_memcg(struct sock *sk)
554 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
555 struct mem_cgroup *memcg;
556 WARN_ON(!sk->sk_cgrp->memcg);
557 memcg = sk->sk_cgrp->memcg;
558 css_put(&sk->sk_cgrp->memcg->css);
562 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
564 if (!memcg || mem_cgroup_is_root(memcg))
567 return &memcg->tcp_mem.cg_proto;
569 EXPORT_SYMBOL(tcp_proto_cgroup);
571 static void disarm_sock_keys(struct mem_cgroup *memcg)
573 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
575 static_key_slow_dec(&memcg_socket_limit_enabled);
578 static void disarm_sock_keys(struct mem_cgroup *memcg)
583 #ifdef CONFIG_MEMCG_KMEM
585 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
586 * There are two main reasons for not using the css_id for this:
587 * 1) this works better in sparse environments, where we have a lot of memcgs,
588 * but only a few kmem-limited. Or also, if we have, for instance, 200
589 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
590 * 200 entry array for that.
592 * 2) In order not to violate the cgroup API, we would like to do all memory
593 * allocation in ->create(). At that point, we haven't yet allocated the
594 * css_id. Having a separate index prevents us from messing with the cgroup
597 * The current size of the caches array is stored in
598 * memcg_limited_groups_array_size. It will double each time we have to
601 static DEFINE_IDA(kmem_limited_groups);
602 int memcg_limited_groups_array_size;
605 * MIN_SIZE is different than 1, because we would like to avoid going through
606 * the alloc/free process all the time. In a small machine, 4 kmem-limited
607 * cgroups is a reasonable guess. In the future, it could be a parameter or
608 * tunable, but that is strictly not necessary.
610 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
611 * this constant directly from cgroup, but it is understandable that this is
612 * better kept as an internal representation in cgroup.c. In any case, the
613 * css_id space is not getting any smaller, and we don't have to necessarily
614 * increase ours as well if it increases.
616 #define MEMCG_CACHES_MIN_SIZE 4
617 #define MEMCG_CACHES_MAX_SIZE 65535
620 * A lot of the calls to the cache allocation functions are expected to be
621 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
622 * conditional to this static branch, we'll have to allow modules that does
623 * kmem_cache_alloc and the such to see this symbol as well
625 struct static_key memcg_kmem_enabled_key;
626 EXPORT_SYMBOL(memcg_kmem_enabled_key);
628 static void disarm_kmem_keys(struct mem_cgroup *memcg)
630 if (memcg_kmem_is_active(memcg)) {
631 static_key_slow_dec(&memcg_kmem_enabled_key);
632 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
635 * This check can't live in kmem destruction function,
636 * since the charges will outlive the cgroup
638 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
641 static void disarm_kmem_keys(struct mem_cgroup *memcg)
644 #endif /* CONFIG_MEMCG_KMEM */
646 static void disarm_static_keys(struct mem_cgroup *memcg)
648 disarm_sock_keys(memcg);
649 disarm_kmem_keys(memcg);
652 static void drain_all_stock_async(struct mem_cgroup *memcg);
654 static struct mem_cgroup_per_zone *
655 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
657 VM_BUG_ON((unsigned)nid >= nr_node_ids);
658 return &memcg->nodeinfo[nid]->zoneinfo[zid];
661 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
666 static struct mem_cgroup_per_zone *
667 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
669 int nid = page_to_nid(page);
670 int zid = page_zonenum(page);
672 return mem_cgroup_zoneinfo(memcg, nid, zid);
675 static struct mem_cgroup_tree_per_zone *
676 soft_limit_tree_node_zone(int nid, int zid)
678 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
681 static struct mem_cgroup_tree_per_zone *
682 soft_limit_tree_from_page(struct page *page)
684 int nid = page_to_nid(page);
685 int zid = page_zonenum(page);
687 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
691 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
692 struct mem_cgroup_per_zone *mz,
693 struct mem_cgroup_tree_per_zone *mctz,
694 unsigned long long new_usage_in_excess)
696 struct rb_node **p = &mctz->rb_root.rb_node;
697 struct rb_node *parent = NULL;
698 struct mem_cgroup_per_zone *mz_node;
703 mz->usage_in_excess = new_usage_in_excess;
704 if (!mz->usage_in_excess)
708 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
710 if (mz->usage_in_excess < mz_node->usage_in_excess)
713 * We can't avoid mem cgroups that are over their soft
714 * limit by the same amount
716 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
719 rb_link_node(&mz->tree_node, parent, p);
720 rb_insert_color(&mz->tree_node, &mctz->rb_root);
725 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
726 struct mem_cgroup_per_zone *mz,
727 struct mem_cgroup_tree_per_zone *mctz)
731 rb_erase(&mz->tree_node, &mctz->rb_root);
736 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
737 struct mem_cgroup_per_zone *mz,
738 struct mem_cgroup_tree_per_zone *mctz)
740 spin_lock(&mctz->lock);
741 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
742 spin_unlock(&mctz->lock);
746 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
748 unsigned long long excess;
749 struct mem_cgroup_per_zone *mz;
750 struct mem_cgroup_tree_per_zone *mctz;
751 int nid = page_to_nid(page);
752 int zid = page_zonenum(page);
753 mctz = soft_limit_tree_from_page(page);
756 * Necessary to update all ancestors when hierarchy is used.
757 * because their event counter is not touched.
759 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
760 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
761 excess = res_counter_soft_limit_excess(&memcg->res);
763 * We have to update the tree if mz is on RB-tree or
764 * mem is over its softlimit.
766 if (excess || mz->on_tree) {
767 spin_lock(&mctz->lock);
768 /* if on-tree, remove it */
770 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
772 * Insert again. mz->usage_in_excess will be updated.
773 * If excess is 0, no tree ops.
775 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
776 spin_unlock(&mctz->lock);
781 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
784 struct mem_cgroup_per_zone *mz;
785 struct mem_cgroup_tree_per_zone *mctz;
787 for_each_node(node) {
788 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
789 mz = mem_cgroup_zoneinfo(memcg, node, zone);
790 mctz = soft_limit_tree_node_zone(node, zone);
791 mem_cgroup_remove_exceeded(memcg, mz, mctz);
796 static struct mem_cgroup_per_zone *
797 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
799 struct rb_node *rightmost = NULL;
800 struct mem_cgroup_per_zone *mz;
804 rightmost = rb_last(&mctz->rb_root);
806 goto done; /* Nothing to reclaim from */
808 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
810 * Remove the node now but someone else can add it back,
811 * we will to add it back at the end of reclaim to its correct
812 * position in the tree.
814 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
815 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
816 !css_tryget(&mz->memcg->css))
822 static struct mem_cgroup_per_zone *
823 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
825 struct mem_cgroup_per_zone *mz;
827 spin_lock(&mctz->lock);
828 mz = __mem_cgroup_largest_soft_limit_node(mctz);
829 spin_unlock(&mctz->lock);
834 * Implementation Note: reading percpu statistics for memcg.
836 * Both of vmstat[] and percpu_counter has threshold and do periodic
837 * synchronization to implement "quick" read. There are trade-off between
838 * reading cost and precision of value. Then, we may have a chance to implement
839 * a periodic synchronizion of counter in memcg's counter.
841 * But this _read() function is used for user interface now. The user accounts
842 * memory usage by memory cgroup and he _always_ requires exact value because
843 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
844 * have to visit all online cpus and make sum. So, for now, unnecessary
845 * synchronization is not implemented. (just implemented for cpu hotplug)
847 * If there are kernel internal actions which can make use of some not-exact
848 * value, and reading all cpu value can be performance bottleneck in some
849 * common workload, threashold and synchonization as vmstat[] should be
852 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
853 enum mem_cgroup_stat_index idx)
859 for_each_online_cpu(cpu)
860 val += per_cpu(memcg->stat->count[idx], cpu);
861 #ifdef CONFIG_HOTPLUG_CPU
862 spin_lock(&memcg->pcp_counter_lock);
863 val += memcg->nocpu_base.count[idx];
864 spin_unlock(&memcg->pcp_counter_lock);
870 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
873 int val = (charge) ? 1 : -1;
874 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
877 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
878 enum mem_cgroup_events_index idx)
880 unsigned long val = 0;
883 for_each_online_cpu(cpu)
884 val += per_cpu(memcg->stat->events[idx], cpu);
885 #ifdef CONFIG_HOTPLUG_CPU
886 spin_lock(&memcg->pcp_counter_lock);
887 val += memcg->nocpu_base.events[idx];
888 spin_unlock(&memcg->pcp_counter_lock);
893 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
895 bool anon, int nr_pages)
900 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
901 * counted as CACHE even if it's on ANON LRU.
904 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
907 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
910 if (PageTransHuge(page))
911 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
914 /* pagein of a big page is an event. So, ignore page size */
916 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
918 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
919 nr_pages = -nr_pages; /* for event */
922 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
928 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
930 struct mem_cgroup_per_zone *mz;
932 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
933 return mz->lru_size[lru];
937 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
938 unsigned int lru_mask)
940 struct mem_cgroup_per_zone *mz;
942 unsigned long ret = 0;
944 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
947 if (BIT(lru) & lru_mask)
948 ret += mz->lru_size[lru];
954 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
955 int nid, unsigned int lru_mask)
960 for (zid = 0; zid < MAX_NR_ZONES; zid++)
961 total += mem_cgroup_zone_nr_lru_pages(memcg,
967 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
968 unsigned int lru_mask)
973 for_each_node_state(nid, N_MEMORY)
974 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
978 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
979 enum mem_cgroup_events_target target)
981 unsigned long val, next;
983 val = __this_cpu_read(memcg->stat->nr_page_events);
984 next = __this_cpu_read(memcg->stat->targets[target]);
985 /* from time_after() in jiffies.h */
986 if ((long)next - (long)val < 0) {
988 case MEM_CGROUP_TARGET_THRESH:
989 next = val + THRESHOLDS_EVENTS_TARGET;
991 case MEM_CGROUP_TARGET_SOFTLIMIT:
992 next = val + SOFTLIMIT_EVENTS_TARGET;
994 case MEM_CGROUP_TARGET_NUMAINFO:
995 next = val + NUMAINFO_EVENTS_TARGET;
1000 __this_cpu_write(memcg->stat->targets[target], next);
1007 * Check events in order.
1010 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1013 /* threshold event is triggered in finer grain than soft limit */
1014 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1015 MEM_CGROUP_TARGET_THRESH))) {
1017 bool do_numainfo __maybe_unused;
1019 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1020 MEM_CGROUP_TARGET_SOFTLIMIT);
1021 #if MAX_NUMNODES > 1
1022 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1023 MEM_CGROUP_TARGET_NUMAINFO);
1027 mem_cgroup_threshold(memcg);
1028 if (unlikely(do_softlimit))
1029 mem_cgroup_update_tree(memcg, page);
1030 #if MAX_NUMNODES > 1
1031 if (unlikely(do_numainfo))
1032 atomic_inc(&memcg->numainfo_events);
1038 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
1040 return mem_cgroup_from_css(cgroup_css(cont, mem_cgroup_subsys_id));
1043 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1046 * mm_update_next_owner() may clear mm->owner to NULL
1047 * if it races with swapoff, page migration, etc.
1048 * So this can be called with p == NULL.
1053 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1056 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1058 struct mem_cgroup *memcg = NULL;
1063 * Because we have no locks, mm->owner's may be being moved to other
1064 * cgroup. We use css_tryget() here even if this looks
1065 * pessimistic (rather than adding locks here).
1069 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1070 if (unlikely(!memcg))
1072 } while (!css_tryget(&memcg->css));
1078 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1079 * ref. count) or NULL if the whole root's subtree has been visited.
1081 * helper function to be used by mem_cgroup_iter
1083 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1084 struct mem_cgroup *last_visited)
1086 struct cgroup *prev_cgroup, *next_cgroup;
1089 * Root is not visited by cgroup iterators so it needs an
1095 prev_cgroup = (last_visited == root) ? NULL
1096 : last_visited->css.cgroup;
1098 next_cgroup = cgroup_next_descendant_pre(
1099 prev_cgroup, root->css.cgroup);
1102 * Even if we found a group we have to make sure it is
1103 * alive. css && !memcg means that the groups should be
1104 * skipped and we should continue the tree walk.
1105 * last_visited css is safe to use because it is
1106 * protected by css_get and the tree walk is rcu safe.
1109 struct mem_cgroup *mem = mem_cgroup_from_cont(
1111 if (css_tryget(&mem->css))
1114 prev_cgroup = next_cgroup;
1122 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1125 * When a group in the hierarchy below root is destroyed, the
1126 * hierarchy iterator can no longer be trusted since it might
1127 * have pointed to the destroyed group. Invalidate it.
1129 atomic_inc(&root->dead_count);
1132 static struct mem_cgroup *
1133 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1134 struct mem_cgroup *root,
1137 struct mem_cgroup *position = NULL;
1139 * A cgroup destruction happens in two stages: offlining and
1140 * release. They are separated by a RCU grace period.
1142 * If the iterator is valid, we may still race with an
1143 * offlining. The RCU lock ensures the object won't be
1144 * released, tryget will fail if we lost the race.
1146 *sequence = atomic_read(&root->dead_count);
1147 if (iter->last_dead_count == *sequence) {
1149 position = iter->last_visited;
1150 if (position && !css_tryget(&position->css))
1156 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1157 struct mem_cgroup *last_visited,
1158 struct mem_cgroup *new_position,
1162 css_put(&last_visited->css);
1164 * We store the sequence count from the time @last_visited was
1165 * loaded successfully instead of rereading it here so that we
1166 * don't lose destruction events in between. We could have
1167 * raced with the destruction of @new_position after all.
1169 iter->last_visited = new_position;
1171 iter->last_dead_count = sequence;
1175 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1176 * @root: hierarchy root
1177 * @prev: previously returned memcg, NULL on first invocation
1178 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1180 * Returns references to children of the hierarchy below @root, or
1181 * @root itself, or %NULL after a full round-trip.
1183 * Caller must pass the return value in @prev on subsequent
1184 * invocations for reference counting, or use mem_cgroup_iter_break()
1185 * to cancel a hierarchy walk before the round-trip is complete.
1187 * Reclaimers can specify a zone and a priority level in @reclaim to
1188 * divide up the memcgs in the hierarchy among all concurrent
1189 * reclaimers operating on the same zone and priority.
1191 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1192 struct mem_cgroup *prev,
1193 struct mem_cgroup_reclaim_cookie *reclaim)
1195 struct mem_cgroup *memcg = NULL;
1196 struct mem_cgroup *last_visited = NULL;
1198 if (mem_cgroup_disabled())
1202 root = root_mem_cgroup;
1204 if (prev && !reclaim)
1205 last_visited = prev;
1207 if (!root->use_hierarchy && root != root_mem_cgroup) {
1215 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1216 int uninitialized_var(seq);
1219 int nid = zone_to_nid(reclaim->zone);
1220 int zid = zone_idx(reclaim->zone);
1221 struct mem_cgroup_per_zone *mz;
1223 mz = mem_cgroup_zoneinfo(root, nid, zid);
1224 iter = &mz->reclaim_iter[reclaim->priority];
1225 if (prev && reclaim->generation != iter->generation) {
1226 iter->last_visited = NULL;
1230 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1233 memcg = __mem_cgroup_iter_next(root, last_visited);
1236 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1240 else if (!prev && memcg)
1241 reclaim->generation = iter->generation;
1250 if (prev && prev != root)
1251 css_put(&prev->css);
1257 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1258 * @root: hierarchy root
1259 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1261 void mem_cgroup_iter_break(struct mem_cgroup *root,
1262 struct mem_cgroup *prev)
1265 root = root_mem_cgroup;
1266 if (prev && prev != root)
1267 css_put(&prev->css);
1271 * Iteration constructs for visiting all cgroups (under a tree). If
1272 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1273 * be used for reference counting.
1275 #define for_each_mem_cgroup_tree(iter, root) \
1276 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1278 iter = mem_cgroup_iter(root, iter, NULL))
1280 #define for_each_mem_cgroup(iter) \
1281 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1283 iter = mem_cgroup_iter(NULL, iter, NULL))
1285 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1287 struct mem_cgroup *memcg;
1290 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1291 if (unlikely(!memcg))
1296 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1299 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1307 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1310 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1311 * @zone: zone of the wanted lruvec
1312 * @memcg: memcg of the wanted lruvec
1314 * Returns the lru list vector holding pages for the given @zone and
1315 * @mem. This can be the global zone lruvec, if the memory controller
1318 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1319 struct mem_cgroup *memcg)
1321 struct mem_cgroup_per_zone *mz;
1322 struct lruvec *lruvec;
1324 if (mem_cgroup_disabled()) {
1325 lruvec = &zone->lruvec;
1329 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1330 lruvec = &mz->lruvec;
1333 * Since a node can be onlined after the mem_cgroup was created,
1334 * we have to be prepared to initialize lruvec->zone here;
1335 * and if offlined then reonlined, we need to reinitialize it.
1337 if (unlikely(lruvec->zone != zone))
1338 lruvec->zone = zone;
1343 * Following LRU functions are allowed to be used without PCG_LOCK.
1344 * Operations are called by routine of global LRU independently from memcg.
1345 * What we have to take care of here is validness of pc->mem_cgroup.
1347 * Changes to pc->mem_cgroup happens when
1350 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1351 * It is added to LRU before charge.
1352 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1353 * When moving account, the page is not on LRU. It's isolated.
1357 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1359 * @zone: zone of the page
1361 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1363 struct mem_cgroup_per_zone *mz;
1364 struct mem_cgroup *memcg;
1365 struct page_cgroup *pc;
1366 struct lruvec *lruvec;
1368 if (mem_cgroup_disabled()) {
1369 lruvec = &zone->lruvec;
1373 pc = lookup_page_cgroup(page);
1374 memcg = pc->mem_cgroup;
1377 * Surreptitiously switch any uncharged offlist page to root:
1378 * an uncharged page off lru does nothing to secure
1379 * its former mem_cgroup from sudden removal.
1381 * Our caller holds lru_lock, and PageCgroupUsed is updated
1382 * under page_cgroup lock: between them, they make all uses
1383 * of pc->mem_cgroup safe.
1385 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1386 pc->mem_cgroup = memcg = root_mem_cgroup;
1388 mz = page_cgroup_zoneinfo(memcg, page);
1389 lruvec = &mz->lruvec;
1392 * Since a node can be onlined after the mem_cgroup was created,
1393 * we have to be prepared to initialize lruvec->zone here;
1394 * and if offlined then reonlined, we need to reinitialize it.
1396 if (unlikely(lruvec->zone != zone))
1397 lruvec->zone = zone;
1402 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1403 * @lruvec: mem_cgroup per zone lru vector
1404 * @lru: index of lru list the page is sitting on
1405 * @nr_pages: positive when adding or negative when removing
1407 * This function must be called when a page is added to or removed from an
1410 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1413 struct mem_cgroup_per_zone *mz;
1414 unsigned long *lru_size;
1416 if (mem_cgroup_disabled())
1419 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1420 lru_size = mz->lru_size + lru;
1421 *lru_size += nr_pages;
1422 VM_BUG_ON((long)(*lru_size) < 0);
1426 * Checks whether given mem is same or in the root_mem_cgroup's
1429 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1430 struct mem_cgroup *memcg)
1432 if (root_memcg == memcg)
1434 if (!root_memcg->use_hierarchy || !memcg)
1436 return css_is_ancestor(&memcg->css, &root_memcg->css);
1439 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1440 struct mem_cgroup *memcg)
1445 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1450 bool task_in_mem_cgroup(struct task_struct *task,
1451 const struct mem_cgroup *memcg)
1453 struct mem_cgroup *curr = NULL;
1454 struct task_struct *p;
1457 p = find_lock_task_mm(task);
1459 curr = try_get_mem_cgroup_from_mm(p->mm);
1463 * All threads may have already detached their mm's, but the oom
1464 * killer still needs to detect if they have already been oom
1465 * killed to prevent needlessly killing additional tasks.
1468 curr = mem_cgroup_from_task(task);
1470 css_get(&curr->css);
1476 * We should check use_hierarchy of "memcg" not "curr". Because checking
1477 * use_hierarchy of "curr" here make this function true if hierarchy is
1478 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1479 * hierarchy(even if use_hierarchy is disabled in "memcg").
1481 ret = mem_cgroup_same_or_subtree(memcg, curr);
1482 css_put(&curr->css);
1486 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1488 unsigned long inactive_ratio;
1489 unsigned long inactive;
1490 unsigned long active;
1493 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1494 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1496 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1498 inactive_ratio = int_sqrt(10 * gb);
1502 return inactive * inactive_ratio < active;
1505 #define mem_cgroup_from_res_counter(counter, member) \
1506 container_of(counter, struct mem_cgroup, member)
1509 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1510 * @memcg: the memory cgroup
1512 * Returns the maximum amount of memory @mem can be charged with, in
1515 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1517 unsigned long long margin;
1519 margin = res_counter_margin(&memcg->res);
1520 if (do_swap_account)
1521 margin = min(margin, res_counter_margin(&memcg->memsw));
1522 return margin >> PAGE_SHIFT;
1525 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1527 struct cgroup *cgrp = memcg->css.cgroup;
1530 if (cgrp->parent == NULL)
1531 return vm_swappiness;
1533 return memcg->swappiness;
1537 * memcg->moving_account is used for checking possibility that some thread is
1538 * calling move_account(). When a thread on CPU-A starts moving pages under
1539 * a memcg, other threads should check memcg->moving_account under
1540 * rcu_read_lock(), like this:
1544 * memcg->moving_account+1 if (memcg->mocing_account)
1546 * synchronize_rcu() update something.
1551 /* for quick checking without looking up memcg */
1552 atomic_t memcg_moving __read_mostly;
1554 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1556 atomic_inc(&memcg_moving);
1557 atomic_inc(&memcg->moving_account);
1561 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1564 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1565 * We check NULL in callee rather than caller.
1568 atomic_dec(&memcg_moving);
1569 atomic_dec(&memcg->moving_account);
1574 * 2 routines for checking "mem" is under move_account() or not.
1576 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1577 * is used for avoiding races in accounting. If true,
1578 * pc->mem_cgroup may be overwritten.
1580 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1581 * under hierarchy of moving cgroups. This is for
1582 * waiting at hith-memory prressure caused by "move".
1585 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1587 VM_BUG_ON(!rcu_read_lock_held());
1588 return atomic_read(&memcg->moving_account) > 0;
1591 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1593 struct mem_cgroup *from;
1594 struct mem_cgroup *to;
1597 * Unlike task_move routines, we access mc.to, mc.from not under
1598 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1600 spin_lock(&mc.lock);
1606 ret = mem_cgroup_same_or_subtree(memcg, from)
1607 || mem_cgroup_same_or_subtree(memcg, to);
1609 spin_unlock(&mc.lock);
1613 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1615 if (mc.moving_task && current != mc.moving_task) {
1616 if (mem_cgroup_under_move(memcg)) {
1618 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1619 /* moving charge context might have finished. */
1622 finish_wait(&mc.waitq, &wait);
1630 * Take this lock when
1631 * - a code tries to modify page's memcg while it's USED.
1632 * - a code tries to modify page state accounting in a memcg.
1633 * see mem_cgroup_stolen(), too.
1635 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1636 unsigned long *flags)
1638 spin_lock_irqsave(&memcg->move_lock, *flags);
1641 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1642 unsigned long *flags)
1644 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1647 #define K(x) ((x) << (PAGE_SHIFT-10))
1649 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1650 * @memcg: The memory cgroup that went over limit
1651 * @p: Task that is going to be killed
1653 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1656 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1658 struct cgroup *task_cgrp;
1659 struct cgroup *mem_cgrp;
1661 * Need a buffer in BSS, can't rely on allocations. The code relies
1662 * on the assumption that OOM is serialized for memory controller.
1663 * If this assumption is broken, revisit this code.
1665 static char memcg_name[PATH_MAX];
1667 struct mem_cgroup *iter;
1675 mem_cgrp = memcg->css.cgroup;
1676 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1678 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1681 * Unfortunately, we are unable to convert to a useful name
1682 * But we'll still print out the usage information
1689 pr_info("Task in %s killed", memcg_name);
1692 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1700 * Continues from above, so we don't need an KERN_ level
1702 pr_cont(" as a result of limit of %s\n", memcg_name);
1705 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1706 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1707 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1708 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1709 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1710 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1711 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1712 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1713 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1714 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1715 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1716 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1718 for_each_mem_cgroup_tree(iter, memcg) {
1719 pr_info("Memory cgroup stats");
1722 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1724 pr_cont(" for %s", memcg_name);
1728 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1729 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1731 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1732 K(mem_cgroup_read_stat(iter, i)));
1735 for (i = 0; i < NR_LRU_LISTS; i++)
1736 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1737 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1744 * This function returns the number of memcg under hierarchy tree. Returns
1745 * 1(self count) if no children.
1747 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1750 struct mem_cgroup *iter;
1752 for_each_mem_cgroup_tree(iter, memcg)
1758 * Return the memory (and swap, if configured) limit for a memcg.
1760 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1764 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1767 * Do not consider swap space if we cannot swap due to swappiness
1769 if (mem_cgroup_swappiness(memcg)) {
1772 limit += total_swap_pages << PAGE_SHIFT;
1773 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1776 * If memsw is finite and limits the amount of swap space
1777 * available to this memcg, return that limit.
1779 limit = min(limit, memsw);
1785 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1788 struct mem_cgroup *iter;
1789 unsigned long chosen_points = 0;
1790 unsigned long totalpages;
1791 unsigned int points = 0;
1792 struct task_struct *chosen = NULL;
1795 * If current has a pending SIGKILL or is exiting, then automatically
1796 * select it. The goal is to allow it to allocate so that it may
1797 * quickly exit and free its memory.
1799 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1800 set_thread_flag(TIF_MEMDIE);
1804 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1805 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1806 for_each_mem_cgroup_tree(iter, memcg) {
1807 struct cgroup *cgroup = iter->css.cgroup;
1808 struct cgroup_iter it;
1809 struct task_struct *task;
1811 cgroup_iter_start(cgroup, &it);
1812 while ((task = cgroup_iter_next(cgroup, &it))) {
1813 switch (oom_scan_process_thread(task, totalpages, NULL,
1815 case OOM_SCAN_SELECT:
1817 put_task_struct(chosen);
1819 chosen_points = ULONG_MAX;
1820 get_task_struct(chosen);
1822 case OOM_SCAN_CONTINUE:
1824 case OOM_SCAN_ABORT:
1825 cgroup_iter_end(cgroup, &it);
1826 mem_cgroup_iter_break(memcg, iter);
1828 put_task_struct(chosen);
1833 points = oom_badness(task, memcg, NULL, totalpages);
1834 if (points > chosen_points) {
1836 put_task_struct(chosen);
1838 chosen_points = points;
1839 get_task_struct(chosen);
1842 cgroup_iter_end(cgroup, &it);
1847 points = chosen_points * 1000 / totalpages;
1848 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1849 NULL, "Memory cgroup out of memory");
1852 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1854 unsigned long flags)
1856 unsigned long total = 0;
1857 bool noswap = false;
1860 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1862 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1865 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1867 drain_all_stock_async(memcg);
1868 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1870 * Allow limit shrinkers, which are triggered directly
1871 * by userspace, to catch signals and stop reclaim
1872 * after minimal progress, regardless of the margin.
1874 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1876 if (mem_cgroup_margin(memcg))
1879 * If nothing was reclaimed after two attempts, there
1880 * may be no reclaimable pages in this hierarchy.
1889 * test_mem_cgroup_node_reclaimable
1890 * @memcg: the target memcg
1891 * @nid: the node ID to be checked.
1892 * @noswap : specify true here if the user wants flle only information.
1894 * This function returns whether the specified memcg contains any
1895 * reclaimable pages on a node. Returns true if there are any reclaimable
1896 * pages in the node.
1898 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1899 int nid, bool noswap)
1901 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1903 if (noswap || !total_swap_pages)
1905 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1910 #if MAX_NUMNODES > 1
1913 * Always updating the nodemask is not very good - even if we have an empty
1914 * list or the wrong list here, we can start from some node and traverse all
1915 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1918 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1922 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1923 * pagein/pageout changes since the last update.
1925 if (!atomic_read(&memcg->numainfo_events))
1927 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1930 /* make a nodemask where this memcg uses memory from */
1931 memcg->scan_nodes = node_states[N_MEMORY];
1933 for_each_node_mask(nid, node_states[N_MEMORY]) {
1935 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1936 node_clear(nid, memcg->scan_nodes);
1939 atomic_set(&memcg->numainfo_events, 0);
1940 atomic_set(&memcg->numainfo_updating, 0);
1944 * Selecting a node where we start reclaim from. Because what we need is just
1945 * reducing usage counter, start from anywhere is O,K. Considering
1946 * memory reclaim from current node, there are pros. and cons.
1948 * Freeing memory from current node means freeing memory from a node which
1949 * we'll use or we've used. So, it may make LRU bad. And if several threads
1950 * hit limits, it will see a contention on a node. But freeing from remote
1951 * node means more costs for memory reclaim because of memory latency.
1953 * Now, we use round-robin. Better algorithm is welcomed.
1955 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1959 mem_cgroup_may_update_nodemask(memcg);
1960 node = memcg->last_scanned_node;
1962 node = next_node(node, memcg->scan_nodes);
1963 if (node == MAX_NUMNODES)
1964 node = first_node(memcg->scan_nodes);
1966 * We call this when we hit limit, not when pages are added to LRU.
1967 * No LRU may hold pages because all pages are UNEVICTABLE or
1968 * memcg is too small and all pages are not on LRU. In that case,
1969 * we use curret node.
1971 if (unlikely(node == MAX_NUMNODES))
1972 node = numa_node_id();
1974 memcg->last_scanned_node = node;
1979 * Check all nodes whether it contains reclaimable pages or not.
1980 * For quick scan, we make use of scan_nodes. This will allow us to skip
1981 * unused nodes. But scan_nodes is lazily updated and may not cotain
1982 * enough new information. We need to do double check.
1984 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1989 * quick check...making use of scan_node.
1990 * We can skip unused nodes.
1992 if (!nodes_empty(memcg->scan_nodes)) {
1993 for (nid = first_node(memcg->scan_nodes);
1995 nid = next_node(nid, memcg->scan_nodes)) {
1997 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2002 * Check rest of nodes.
2004 for_each_node_state(nid, N_MEMORY) {
2005 if (node_isset(nid, memcg->scan_nodes))
2007 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2014 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2019 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2021 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2025 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2028 unsigned long *total_scanned)
2030 struct mem_cgroup *victim = NULL;
2033 unsigned long excess;
2034 unsigned long nr_scanned;
2035 struct mem_cgroup_reclaim_cookie reclaim = {
2040 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2043 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2048 * If we have not been able to reclaim
2049 * anything, it might because there are
2050 * no reclaimable pages under this hierarchy
2055 * We want to do more targeted reclaim.
2056 * excess >> 2 is not to excessive so as to
2057 * reclaim too much, nor too less that we keep
2058 * coming back to reclaim from this cgroup
2060 if (total >= (excess >> 2) ||
2061 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2066 if (!mem_cgroup_reclaimable(victim, false))
2068 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2070 *total_scanned += nr_scanned;
2071 if (!res_counter_soft_limit_excess(&root_memcg->res))
2074 mem_cgroup_iter_break(root_memcg, victim);
2079 * Check OOM-Killer is already running under our hierarchy.
2080 * If someone is running, return false.
2081 * Has to be called with memcg_oom_lock
2083 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
2085 struct mem_cgroup *iter, *failed = NULL;
2087 for_each_mem_cgroup_tree(iter, memcg) {
2088 if (iter->oom_lock) {
2090 * this subtree of our hierarchy is already locked
2091 * so we cannot give a lock.
2094 mem_cgroup_iter_break(memcg, iter);
2097 iter->oom_lock = true;
2104 * OK, we failed to lock the whole subtree so we have to clean up
2105 * what we set up to the failing subtree
2107 for_each_mem_cgroup_tree(iter, memcg) {
2108 if (iter == failed) {
2109 mem_cgroup_iter_break(memcg, iter);
2112 iter->oom_lock = false;
2118 * Has to be called with memcg_oom_lock
2120 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2122 struct mem_cgroup *iter;
2124 for_each_mem_cgroup_tree(iter, memcg)
2125 iter->oom_lock = false;
2129 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2131 struct mem_cgroup *iter;
2133 for_each_mem_cgroup_tree(iter, memcg)
2134 atomic_inc(&iter->under_oom);
2137 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2139 struct mem_cgroup *iter;
2142 * When a new child is created while the hierarchy is under oom,
2143 * mem_cgroup_oom_lock() may not be called. We have to use
2144 * atomic_add_unless() here.
2146 for_each_mem_cgroup_tree(iter, memcg)
2147 atomic_add_unless(&iter->under_oom, -1, 0);
2150 static DEFINE_SPINLOCK(memcg_oom_lock);
2151 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2153 struct oom_wait_info {
2154 struct mem_cgroup *memcg;
2158 static int memcg_oom_wake_function(wait_queue_t *wait,
2159 unsigned mode, int sync, void *arg)
2161 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2162 struct mem_cgroup *oom_wait_memcg;
2163 struct oom_wait_info *oom_wait_info;
2165 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2166 oom_wait_memcg = oom_wait_info->memcg;
2169 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2170 * Then we can use css_is_ancestor without taking care of RCU.
2172 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2173 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2175 return autoremove_wake_function(wait, mode, sync, arg);
2178 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2180 /* for filtering, pass "memcg" as argument. */
2181 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2184 static void memcg_oom_recover(struct mem_cgroup *memcg)
2186 if (memcg && atomic_read(&memcg->under_oom))
2187 memcg_wakeup_oom(memcg);
2191 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2193 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2196 struct oom_wait_info owait;
2197 bool locked, need_to_kill;
2199 owait.memcg = memcg;
2200 owait.wait.flags = 0;
2201 owait.wait.func = memcg_oom_wake_function;
2202 owait.wait.private = current;
2203 INIT_LIST_HEAD(&owait.wait.task_list);
2204 need_to_kill = true;
2205 mem_cgroup_mark_under_oom(memcg);
2207 /* At first, try to OOM lock hierarchy under memcg.*/
2208 spin_lock(&memcg_oom_lock);
2209 locked = mem_cgroup_oom_lock(memcg);
2211 * Even if signal_pending(), we can't quit charge() loop without
2212 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2213 * under OOM is always welcomed, use TASK_KILLABLE here.
2215 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2216 if (!locked || memcg->oom_kill_disable)
2217 need_to_kill = false;
2219 mem_cgroup_oom_notify(memcg);
2220 spin_unlock(&memcg_oom_lock);
2223 finish_wait(&memcg_oom_waitq, &owait.wait);
2224 mem_cgroup_out_of_memory(memcg, mask, order);
2227 finish_wait(&memcg_oom_waitq, &owait.wait);
2229 spin_lock(&memcg_oom_lock);
2231 mem_cgroup_oom_unlock(memcg);
2232 memcg_wakeup_oom(memcg);
2233 spin_unlock(&memcg_oom_lock);
2235 mem_cgroup_unmark_under_oom(memcg);
2237 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2239 /* Give chance to dying process */
2240 schedule_timeout_uninterruptible(1);
2245 * Currently used to update mapped file statistics, but the routine can be
2246 * generalized to update other statistics as well.
2248 * Notes: Race condition
2250 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2251 * it tends to be costly. But considering some conditions, we doesn't need
2252 * to do so _always_.
2254 * Considering "charge", lock_page_cgroup() is not required because all
2255 * file-stat operations happen after a page is attached to radix-tree. There
2256 * are no race with "charge".
2258 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2259 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2260 * if there are race with "uncharge". Statistics itself is properly handled
2263 * Considering "move", this is an only case we see a race. To make the race
2264 * small, we check mm->moving_account and detect there are possibility of race
2265 * If there is, we take a lock.
2268 void __mem_cgroup_begin_update_page_stat(struct page *page,
2269 bool *locked, unsigned long *flags)
2271 struct mem_cgroup *memcg;
2272 struct page_cgroup *pc;
2274 pc = lookup_page_cgroup(page);
2276 memcg = pc->mem_cgroup;
2277 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2280 * If this memory cgroup is not under account moving, we don't
2281 * need to take move_lock_mem_cgroup(). Because we already hold
2282 * rcu_read_lock(), any calls to move_account will be delayed until
2283 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2285 if (!mem_cgroup_stolen(memcg))
2288 move_lock_mem_cgroup(memcg, flags);
2289 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2290 move_unlock_mem_cgroup(memcg, flags);
2296 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2298 struct page_cgroup *pc = lookup_page_cgroup(page);
2301 * It's guaranteed that pc->mem_cgroup never changes while
2302 * lock is held because a routine modifies pc->mem_cgroup
2303 * should take move_lock_mem_cgroup().
2305 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2308 void mem_cgroup_update_page_stat(struct page *page,
2309 enum mem_cgroup_page_stat_item idx, int val)
2311 struct mem_cgroup *memcg;
2312 struct page_cgroup *pc = lookup_page_cgroup(page);
2313 unsigned long uninitialized_var(flags);
2315 if (mem_cgroup_disabled())
2318 memcg = pc->mem_cgroup;
2319 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2323 case MEMCG_NR_FILE_MAPPED:
2324 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2330 this_cpu_add(memcg->stat->count[idx], val);
2334 * size of first charge trial. "32" comes from vmscan.c's magic value.
2335 * TODO: maybe necessary to use big numbers in big irons.
2337 #define CHARGE_BATCH 32U
2338 struct memcg_stock_pcp {
2339 struct mem_cgroup *cached; /* this never be root cgroup */
2340 unsigned int nr_pages;
2341 struct work_struct work;
2342 unsigned long flags;
2343 #define FLUSHING_CACHED_CHARGE 0
2345 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2346 static DEFINE_MUTEX(percpu_charge_mutex);
2349 * consume_stock: Try to consume stocked charge on this cpu.
2350 * @memcg: memcg to consume from.
2351 * @nr_pages: how many pages to charge.
2353 * The charges will only happen if @memcg matches the current cpu's memcg
2354 * stock, and at least @nr_pages are available in that stock. Failure to
2355 * service an allocation will refill the stock.
2357 * returns true if successful, false otherwise.
2359 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2361 struct memcg_stock_pcp *stock;
2364 if (nr_pages > CHARGE_BATCH)
2367 stock = &get_cpu_var(memcg_stock);
2368 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2369 stock->nr_pages -= nr_pages;
2370 else /* need to call res_counter_charge */
2372 put_cpu_var(memcg_stock);
2377 * Returns stocks cached in percpu to res_counter and reset cached information.
2379 static void drain_stock(struct memcg_stock_pcp *stock)
2381 struct mem_cgroup *old = stock->cached;
2383 if (stock->nr_pages) {
2384 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2386 res_counter_uncharge(&old->res, bytes);
2387 if (do_swap_account)
2388 res_counter_uncharge(&old->memsw, bytes);
2389 stock->nr_pages = 0;
2391 stock->cached = NULL;
2395 * This must be called under preempt disabled or must be called by
2396 * a thread which is pinned to local cpu.
2398 static void drain_local_stock(struct work_struct *dummy)
2400 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2402 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2405 static void __init memcg_stock_init(void)
2409 for_each_possible_cpu(cpu) {
2410 struct memcg_stock_pcp *stock =
2411 &per_cpu(memcg_stock, cpu);
2412 INIT_WORK(&stock->work, drain_local_stock);
2417 * Cache charges(val) which is from res_counter, to local per_cpu area.
2418 * This will be consumed by consume_stock() function, later.
2420 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2422 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2424 if (stock->cached != memcg) { /* reset if necessary */
2426 stock->cached = memcg;
2428 stock->nr_pages += nr_pages;
2429 put_cpu_var(memcg_stock);
2433 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2434 * of the hierarchy under it. sync flag says whether we should block
2435 * until the work is done.
2437 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2441 /* Notify other cpus that system-wide "drain" is running */
2444 for_each_online_cpu(cpu) {
2445 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2446 struct mem_cgroup *memcg;
2448 memcg = stock->cached;
2449 if (!memcg || !stock->nr_pages)
2451 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2453 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2455 drain_local_stock(&stock->work);
2457 schedule_work_on(cpu, &stock->work);
2465 for_each_online_cpu(cpu) {
2466 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2467 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2468 flush_work(&stock->work);
2475 * Tries to drain stocked charges in other cpus. This function is asynchronous
2476 * and just put a work per cpu for draining localy on each cpu. Caller can
2477 * expects some charges will be back to res_counter later but cannot wait for
2480 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2483 * If someone calls draining, avoid adding more kworker runs.
2485 if (!mutex_trylock(&percpu_charge_mutex))
2487 drain_all_stock(root_memcg, false);
2488 mutex_unlock(&percpu_charge_mutex);
2491 /* This is a synchronous drain interface. */
2492 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2494 /* called when force_empty is called */
2495 mutex_lock(&percpu_charge_mutex);
2496 drain_all_stock(root_memcg, true);
2497 mutex_unlock(&percpu_charge_mutex);
2501 * This function drains percpu counter value from DEAD cpu and
2502 * move it to local cpu. Note that this function can be preempted.
2504 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2508 spin_lock(&memcg->pcp_counter_lock);
2509 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2510 long x = per_cpu(memcg->stat->count[i], cpu);
2512 per_cpu(memcg->stat->count[i], cpu) = 0;
2513 memcg->nocpu_base.count[i] += x;
2515 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2516 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2518 per_cpu(memcg->stat->events[i], cpu) = 0;
2519 memcg->nocpu_base.events[i] += x;
2521 spin_unlock(&memcg->pcp_counter_lock);
2524 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2525 unsigned long action,
2528 int cpu = (unsigned long)hcpu;
2529 struct memcg_stock_pcp *stock;
2530 struct mem_cgroup *iter;
2532 if (action == CPU_ONLINE)
2535 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2538 for_each_mem_cgroup(iter)
2539 mem_cgroup_drain_pcp_counter(iter, cpu);
2541 stock = &per_cpu(memcg_stock, cpu);
2547 /* See __mem_cgroup_try_charge() for details */
2549 CHARGE_OK, /* success */
2550 CHARGE_RETRY, /* need to retry but retry is not bad */
2551 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2552 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2553 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2556 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2557 unsigned int nr_pages, unsigned int min_pages,
2560 unsigned long csize = nr_pages * PAGE_SIZE;
2561 struct mem_cgroup *mem_over_limit;
2562 struct res_counter *fail_res;
2563 unsigned long flags = 0;
2566 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2569 if (!do_swap_account)
2571 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2575 res_counter_uncharge(&memcg->res, csize);
2576 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2577 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2579 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2581 * Never reclaim on behalf of optional batching, retry with a
2582 * single page instead.
2584 if (nr_pages > min_pages)
2585 return CHARGE_RETRY;
2587 if (!(gfp_mask & __GFP_WAIT))
2588 return CHARGE_WOULDBLOCK;
2590 if (gfp_mask & __GFP_NORETRY)
2591 return CHARGE_NOMEM;
2593 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2594 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2595 return CHARGE_RETRY;
2597 * Even though the limit is exceeded at this point, reclaim
2598 * may have been able to free some pages. Retry the charge
2599 * before killing the task.
2601 * Only for regular pages, though: huge pages are rather
2602 * unlikely to succeed so close to the limit, and we fall back
2603 * to regular pages anyway in case of failure.
2605 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2606 return CHARGE_RETRY;
2609 * At task move, charge accounts can be doubly counted. So, it's
2610 * better to wait until the end of task_move if something is going on.
2612 if (mem_cgroup_wait_acct_move(mem_over_limit))
2613 return CHARGE_RETRY;
2615 /* If we don't need to call oom-killer at el, return immediately */
2617 return CHARGE_NOMEM;
2619 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2620 return CHARGE_OOM_DIE;
2622 return CHARGE_RETRY;
2626 * __mem_cgroup_try_charge() does
2627 * 1. detect memcg to be charged against from passed *mm and *ptr,
2628 * 2. update res_counter
2629 * 3. call memory reclaim if necessary.
2631 * In some special case, if the task is fatal, fatal_signal_pending() or
2632 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2633 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2634 * as possible without any hazards. 2: all pages should have a valid
2635 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2636 * pointer, that is treated as a charge to root_mem_cgroup.
2638 * So __mem_cgroup_try_charge() will return
2639 * 0 ... on success, filling *ptr with a valid memcg pointer.
2640 * -ENOMEM ... charge failure because of resource limits.
2641 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2643 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2644 * the oom-killer can be invoked.
2646 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2648 unsigned int nr_pages,
2649 struct mem_cgroup **ptr,
2652 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2653 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2654 struct mem_cgroup *memcg = NULL;
2658 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2659 * in system level. So, allow to go ahead dying process in addition to
2662 if (unlikely(test_thread_flag(TIF_MEMDIE)
2663 || fatal_signal_pending(current)))
2667 * We always charge the cgroup the mm_struct belongs to.
2668 * The mm_struct's mem_cgroup changes on task migration if the
2669 * thread group leader migrates. It's possible that mm is not
2670 * set, if so charge the root memcg (happens for pagecache usage).
2673 *ptr = root_mem_cgroup;
2675 if (*ptr) { /* css should be a valid one */
2677 if (mem_cgroup_is_root(memcg))
2679 if (consume_stock(memcg, nr_pages))
2681 css_get(&memcg->css);
2683 struct task_struct *p;
2686 p = rcu_dereference(mm->owner);
2688 * Because we don't have task_lock(), "p" can exit.
2689 * In that case, "memcg" can point to root or p can be NULL with
2690 * race with swapoff. Then, we have small risk of mis-accouning.
2691 * But such kind of mis-account by race always happens because
2692 * we don't have cgroup_mutex(). It's overkill and we allo that
2694 * (*) swapoff at el will charge against mm-struct not against
2695 * task-struct. So, mm->owner can be NULL.
2697 memcg = mem_cgroup_from_task(p);
2699 memcg = root_mem_cgroup;
2700 if (mem_cgroup_is_root(memcg)) {
2704 if (consume_stock(memcg, nr_pages)) {
2706 * It seems dagerous to access memcg without css_get().
2707 * But considering how consume_stok works, it's not
2708 * necessary. If consume_stock success, some charges
2709 * from this memcg are cached on this cpu. So, we
2710 * don't need to call css_get()/css_tryget() before
2711 * calling consume_stock().
2716 /* after here, we may be blocked. we need to get refcnt */
2717 if (!css_tryget(&memcg->css)) {
2727 /* If killed, bypass charge */
2728 if (fatal_signal_pending(current)) {
2729 css_put(&memcg->css);
2734 if (oom && !nr_oom_retries) {
2736 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2739 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2744 case CHARGE_RETRY: /* not in OOM situation but retry */
2746 css_put(&memcg->css);
2749 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2750 css_put(&memcg->css);
2752 case CHARGE_NOMEM: /* OOM routine works */
2754 css_put(&memcg->css);
2757 /* If oom, we never return -ENOMEM */
2760 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2761 css_put(&memcg->css);
2764 } while (ret != CHARGE_OK);
2766 if (batch > nr_pages)
2767 refill_stock(memcg, batch - nr_pages);
2768 css_put(&memcg->css);
2776 *ptr = root_mem_cgroup;
2781 * Somemtimes we have to undo a charge we got by try_charge().
2782 * This function is for that and do uncharge, put css's refcnt.
2783 * gotten by try_charge().
2785 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2786 unsigned int nr_pages)
2788 if (!mem_cgroup_is_root(memcg)) {
2789 unsigned long bytes = nr_pages * PAGE_SIZE;
2791 res_counter_uncharge(&memcg->res, bytes);
2792 if (do_swap_account)
2793 res_counter_uncharge(&memcg->memsw, bytes);
2798 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2799 * This is useful when moving usage to parent cgroup.
2801 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2802 unsigned int nr_pages)
2804 unsigned long bytes = nr_pages * PAGE_SIZE;
2806 if (mem_cgroup_is_root(memcg))
2809 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2810 if (do_swap_account)
2811 res_counter_uncharge_until(&memcg->memsw,
2812 memcg->memsw.parent, bytes);
2816 * A helper function to get mem_cgroup from ID. must be called under
2817 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2818 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2819 * called against removed memcg.)
2821 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2823 struct cgroup_subsys_state *css;
2825 /* ID 0 is unused ID */
2828 css = css_lookup(&mem_cgroup_subsys, id);
2831 return mem_cgroup_from_css(css);
2834 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2836 struct mem_cgroup *memcg = NULL;
2837 struct page_cgroup *pc;
2841 VM_BUG_ON(!PageLocked(page));
2843 pc = lookup_page_cgroup(page);
2844 lock_page_cgroup(pc);
2845 if (PageCgroupUsed(pc)) {
2846 memcg = pc->mem_cgroup;
2847 if (memcg && !css_tryget(&memcg->css))
2849 } else if (PageSwapCache(page)) {
2850 ent.val = page_private(page);
2851 id = lookup_swap_cgroup_id(ent);
2853 memcg = mem_cgroup_lookup(id);
2854 if (memcg && !css_tryget(&memcg->css))
2858 unlock_page_cgroup(pc);
2862 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2864 unsigned int nr_pages,
2865 enum charge_type ctype,
2868 struct page_cgroup *pc = lookup_page_cgroup(page);
2869 struct zone *uninitialized_var(zone);
2870 struct lruvec *lruvec;
2871 bool was_on_lru = false;
2874 lock_page_cgroup(pc);
2875 VM_BUG_ON(PageCgroupUsed(pc));
2877 * we don't need page_cgroup_lock about tail pages, becase they are not
2878 * accessed by any other context at this point.
2882 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2883 * may already be on some other mem_cgroup's LRU. Take care of it.
2886 zone = page_zone(page);
2887 spin_lock_irq(&zone->lru_lock);
2888 if (PageLRU(page)) {
2889 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2891 del_page_from_lru_list(page, lruvec, page_lru(page));
2896 pc->mem_cgroup = memcg;
2898 * We access a page_cgroup asynchronously without lock_page_cgroup().
2899 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2900 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2901 * before USED bit, we need memory barrier here.
2902 * See mem_cgroup_add_lru_list(), etc.
2905 SetPageCgroupUsed(pc);
2909 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2910 VM_BUG_ON(PageLRU(page));
2912 add_page_to_lru_list(page, lruvec, page_lru(page));
2914 spin_unlock_irq(&zone->lru_lock);
2917 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2922 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2923 unlock_page_cgroup(pc);
2926 * "charge_statistics" updated event counter. Then, check it.
2927 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2928 * if they exceeds softlimit.
2930 memcg_check_events(memcg, page);
2933 static DEFINE_MUTEX(set_limit_mutex);
2935 #ifdef CONFIG_MEMCG_KMEM
2936 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2938 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2939 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2943 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2944 * in the memcg_cache_params struct.
2946 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2948 struct kmem_cache *cachep;
2950 VM_BUG_ON(p->is_root_cache);
2951 cachep = p->root_cache;
2952 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2955 #ifdef CONFIG_SLABINFO
2956 static int mem_cgroup_slabinfo_read(struct cgroup *cont, struct cftype *cft,
2959 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
2960 struct memcg_cache_params *params;
2962 if (!memcg_can_account_kmem(memcg))
2965 print_slabinfo_header(m);
2967 mutex_lock(&memcg->slab_caches_mutex);
2968 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2969 cache_show(memcg_params_to_cache(params), m);
2970 mutex_unlock(&memcg->slab_caches_mutex);
2976 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2978 struct res_counter *fail_res;
2979 struct mem_cgroup *_memcg;
2983 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2988 * Conditions under which we can wait for the oom_killer. Those are
2989 * the same conditions tested by the core page allocator
2991 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2994 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2997 if (ret == -EINTR) {
2999 * __mem_cgroup_try_charge() chosed to bypass to root due to
3000 * OOM kill or fatal signal. Since our only options are to
3001 * either fail the allocation or charge it to this cgroup, do
3002 * it as a temporary condition. But we can't fail. From a
3003 * kmem/slab perspective, the cache has already been selected,
3004 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3007 * This condition will only trigger if the task entered
3008 * memcg_charge_kmem in a sane state, but was OOM-killed during
3009 * __mem_cgroup_try_charge() above. Tasks that were already
3010 * dying when the allocation triggers should have been already
3011 * directed to the root cgroup in memcontrol.h
3013 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3014 if (do_swap_account)
3015 res_counter_charge_nofail(&memcg->memsw, size,
3019 res_counter_uncharge(&memcg->kmem, size);
3024 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3026 res_counter_uncharge(&memcg->res, size);
3027 if (do_swap_account)
3028 res_counter_uncharge(&memcg->memsw, size);
3031 if (res_counter_uncharge(&memcg->kmem, size))
3035 * Releases a reference taken in kmem_cgroup_css_offline in case
3036 * this last uncharge is racing with the offlining code or it is
3037 * outliving the memcg existence.
3039 * The memory barrier imposed by test&clear is paired with the
3040 * explicit one in memcg_kmem_mark_dead().
3042 if (memcg_kmem_test_and_clear_dead(memcg))
3043 css_put(&memcg->css);
3046 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3051 mutex_lock(&memcg->slab_caches_mutex);
3052 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3053 mutex_unlock(&memcg->slab_caches_mutex);
3057 * helper for acessing a memcg's index. It will be used as an index in the
3058 * child cache array in kmem_cache, and also to derive its name. This function
3059 * will return -1 when this is not a kmem-limited memcg.
3061 int memcg_cache_id(struct mem_cgroup *memcg)
3063 return memcg ? memcg->kmemcg_id : -1;
3067 * This ends up being protected by the set_limit mutex, during normal
3068 * operation, because that is its main call site.
3070 * But when we create a new cache, we can call this as well if its parent
3071 * is kmem-limited. That will have to hold set_limit_mutex as well.
3073 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3077 num = ida_simple_get(&kmem_limited_groups,
3078 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3082 * After this point, kmem_accounted (that we test atomically in
3083 * the beginning of this conditional), is no longer 0. This
3084 * guarantees only one process will set the following boolean
3085 * to true. We don't need test_and_set because we're protected
3086 * by the set_limit_mutex anyway.
3088 memcg_kmem_set_activated(memcg);
3090 ret = memcg_update_all_caches(num+1);
3092 ida_simple_remove(&kmem_limited_groups, num);
3093 memcg_kmem_clear_activated(memcg);
3097 memcg->kmemcg_id = num;
3098 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3099 mutex_init(&memcg->slab_caches_mutex);
3103 static size_t memcg_caches_array_size(int num_groups)
3106 if (num_groups <= 0)
3109 size = 2 * num_groups;
3110 if (size < MEMCG_CACHES_MIN_SIZE)
3111 size = MEMCG_CACHES_MIN_SIZE;
3112 else if (size > MEMCG_CACHES_MAX_SIZE)
3113 size = MEMCG_CACHES_MAX_SIZE;
3119 * We should update the current array size iff all caches updates succeed. This
3120 * can only be done from the slab side. The slab mutex needs to be held when
3123 void memcg_update_array_size(int num)
3125 if (num > memcg_limited_groups_array_size)
3126 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3129 static void kmem_cache_destroy_work_func(struct work_struct *w);
3131 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3133 struct memcg_cache_params *cur_params = s->memcg_params;
3135 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3137 if (num_groups > memcg_limited_groups_array_size) {
3139 ssize_t size = memcg_caches_array_size(num_groups);
3141 size *= sizeof(void *);
3142 size += sizeof(struct memcg_cache_params);
3144 s->memcg_params = kzalloc(size, GFP_KERNEL);
3145 if (!s->memcg_params) {
3146 s->memcg_params = cur_params;
3150 s->memcg_params->is_root_cache = true;
3153 * There is the chance it will be bigger than
3154 * memcg_limited_groups_array_size, if we failed an allocation
3155 * in a cache, in which case all caches updated before it, will
3156 * have a bigger array.
3158 * But if that is the case, the data after
3159 * memcg_limited_groups_array_size is certainly unused
3161 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3162 if (!cur_params->memcg_caches[i])
3164 s->memcg_params->memcg_caches[i] =
3165 cur_params->memcg_caches[i];
3169 * Ideally, we would wait until all caches succeed, and only
3170 * then free the old one. But this is not worth the extra
3171 * pointer per-cache we'd have to have for this.
3173 * It is not a big deal if some caches are left with a size
3174 * bigger than the others. And all updates will reset this
3182 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3183 struct kmem_cache *root_cache)
3185 size_t size = sizeof(struct memcg_cache_params);
3187 if (!memcg_kmem_enabled())
3191 size += memcg_limited_groups_array_size * sizeof(void *);
3193 s->memcg_params = kzalloc(size, GFP_KERNEL);
3194 if (!s->memcg_params)
3197 INIT_WORK(&s->memcg_params->destroy,
3198 kmem_cache_destroy_work_func);
3200 s->memcg_params->memcg = memcg;
3201 s->memcg_params->root_cache = root_cache;
3203 s->memcg_params->is_root_cache = true;
3208 void memcg_release_cache(struct kmem_cache *s)
3210 struct kmem_cache *root;
3211 struct mem_cgroup *memcg;
3215 * This happens, for instance, when a root cache goes away before we
3218 if (!s->memcg_params)
3221 if (s->memcg_params->is_root_cache)
3224 memcg = s->memcg_params->memcg;
3225 id = memcg_cache_id(memcg);
3227 root = s->memcg_params->root_cache;
3228 root->memcg_params->memcg_caches[id] = NULL;
3230 mutex_lock(&memcg->slab_caches_mutex);
3231 list_del(&s->memcg_params->list);
3232 mutex_unlock(&memcg->slab_caches_mutex);
3234 css_put(&memcg->css);
3236 kfree(s->memcg_params);
3240 * During the creation a new cache, we need to disable our accounting mechanism
3241 * altogether. This is true even if we are not creating, but rather just
3242 * enqueing new caches to be created.
3244 * This is because that process will trigger allocations; some visible, like
3245 * explicit kmallocs to auxiliary data structures, name strings and internal
3246 * cache structures; some well concealed, like INIT_WORK() that can allocate
3247 * objects during debug.
3249 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3250 * to it. This may not be a bounded recursion: since the first cache creation
3251 * failed to complete (waiting on the allocation), we'll just try to create the
3252 * cache again, failing at the same point.
3254 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3255 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3256 * inside the following two functions.
3258 static inline void memcg_stop_kmem_account(void)
3260 VM_BUG_ON(!current->mm);
3261 current->memcg_kmem_skip_account++;
3264 static inline void memcg_resume_kmem_account(void)
3266 VM_BUG_ON(!current->mm);
3267 current->memcg_kmem_skip_account--;
3270 static void kmem_cache_destroy_work_func(struct work_struct *w)
3272 struct kmem_cache *cachep;
3273 struct memcg_cache_params *p;
3275 p = container_of(w, struct memcg_cache_params, destroy);
3277 cachep = memcg_params_to_cache(p);
3280 * If we get down to 0 after shrink, we could delete right away.
3281 * However, memcg_release_pages() already puts us back in the workqueue
3282 * in that case. If we proceed deleting, we'll get a dangling
3283 * reference, and removing the object from the workqueue in that case
3284 * is unnecessary complication. We are not a fast path.
3286 * Note that this case is fundamentally different from racing with
3287 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3288 * kmem_cache_shrink, not only we would be reinserting a dead cache
3289 * into the queue, but doing so from inside the worker racing to
3292 * So if we aren't down to zero, we'll just schedule a worker and try
3295 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3296 kmem_cache_shrink(cachep);
3297 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3300 kmem_cache_destroy(cachep);
3303 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3305 if (!cachep->memcg_params->dead)
3309 * There are many ways in which we can get here.
3311 * We can get to a memory-pressure situation while the delayed work is
3312 * still pending to run. The vmscan shrinkers can then release all
3313 * cache memory and get us to destruction. If this is the case, we'll
3314 * be executed twice, which is a bug (the second time will execute over
3315 * bogus data). In this case, cancelling the work should be fine.
3317 * But we can also get here from the worker itself, if
3318 * kmem_cache_shrink is enough to shake all the remaining objects and
3319 * get the page count to 0. In this case, we'll deadlock if we try to
3320 * cancel the work (the worker runs with an internal lock held, which
3321 * is the same lock we would hold for cancel_work_sync().)
3323 * Since we can't possibly know who got us here, just refrain from
3324 * running if there is already work pending
3326 if (work_pending(&cachep->memcg_params->destroy))
3329 * We have to defer the actual destroying to a workqueue, because
3330 * we might currently be in a context that cannot sleep.
3332 schedule_work(&cachep->memcg_params->destroy);
3336 * This lock protects updaters, not readers. We want readers to be as fast as
3337 * they can, and they will either see NULL or a valid cache value. Our model
3338 * allow them to see NULL, in which case the root memcg will be selected.
3340 * We need this lock because multiple allocations to the same cache from a non
3341 * will span more than one worker. Only one of them can create the cache.
3343 static DEFINE_MUTEX(memcg_cache_mutex);
3346 * Called with memcg_cache_mutex held
3348 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3349 struct kmem_cache *s)
3351 struct kmem_cache *new;
3352 static char *tmp_name = NULL;
3354 lockdep_assert_held(&memcg_cache_mutex);
3357 * kmem_cache_create_memcg duplicates the given name and
3358 * cgroup_name for this name requires RCU context.
3359 * This static temporary buffer is used to prevent from
3360 * pointless shortliving allocation.
3363 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3369 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3370 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3373 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3374 (s->flags & ~SLAB_PANIC), s->ctor, s);
3377 new->allocflags |= __GFP_KMEMCG;
3382 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3383 struct kmem_cache *cachep)
3385 struct kmem_cache *new_cachep;
3388 BUG_ON(!memcg_can_account_kmem(memcg));
3390 idx = memcg_cache_id(memcg);
3392 mutex_lock(&memcg_cache_mutex);
3393 new_cachep = cachep->memcg_params->memcg_caches[idx];
3395 css_put(&memcg->css);
3399 new_cachep = kmem_cache_dup(memcg, cachep);
3400 if (new_cachep == NULL) {
3401 new_cachep = cachep;
3402 css_put(&memcg->css);
3406 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3408 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3410 * the readers won't lock, make sure everybody sees the updated value,
3411 * so they won't put stuff in the queue again for no reason
3415 mutex_unlock(&memcg_cache_mutex);
3419 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3421 struct kmem_cache *c;
3424 if (!s->memcg_params)
3426 if (!s->memcg_params->is_root_cache)
3430 * If the cache is being destroyed, we trust that there is no one else
3431 * requesting objects from it. Even if there are, the sanity checks in
3432 * kmem_cache_destroy should caught this ill-case.
3434 * Still, we don't want anyone else freeing memcg_caches under our
3435 * noses, which can happen if a new memcg comes to life. As usual,
3436 * we'll take the set_limit_mutex to protect ourselves against this.
3438 mutex_lock(&set_limit_mutex);
3439 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3440 c = s->memcg_params->memcg_caches[i];
3445 * We will now manually delete the caches, so to avoid races
3446 * we need to cancel all pending destruction workers and
3447 * proceed with destruction ourselves.
3449 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3450 * and that could spawn the workers again: it is likely that
3451 * the cache still have active pages until this very moment.
3452 * This would lead us back to mem_cgroup_destroy_cache.
3454 * But that will not execute at all if the "dead" flag is not
3455 * set, so flip it down to guarantee we are in control.
3457 c->memcg_params->dead = false;
3458 cancel_work_sync(&c->memcg_params->destroy);
3459 kmem_cache_destroy(c);
3461 mutex_unlock(&set_limit_mutex);
3464 struct create_work {
3465 struct mem_cgroup *memcg;
3466 struct kmem_cache *cachep;
3467 struct work_struct work;
3470 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3472 struct kmem_cache *cachep;
3473 struct memcg_cache_params *params;
3475 if (!memcg_kmem_is_active(memcg))
3478 mutex_lock(&memcg->slab_caches_mutex);
3479 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3480 cachep = memcg_params_to_cache(params);
3481 cachep->memcg_params->dead = true;
3482 schedule_work(&cachep->memcg_params->destroy);
3484 mutex_unlock(&memcg->slab_caches_mutex);
3487 static void memcg_create_cache_work_func(struct work_struct *w)
3489 struct create_work *cw;
3491 cw = container_of(w, struct create_work, work);
3492 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3497 * Enqueue the creation of a per-memcg kmem_cache.
3499 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3500 struct kmem_cache *cachep)
3502 struct create_work *cw;
3504 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3506 css_put(&memcg->css);
3511 cw->cachep = cachep;
3513 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3514 schedule_work(&cw->work);
3517 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3518 struct kmem_cache *cachep)
3521 * We need to stop accounting when we kmalloc, because if the
3522 * corresponding kmalloc cache is not yet created, the first allocation
3523 * in __memcg_create_cache_enqueue will recurse.
3525 * However, it is better to enclose the whole function. Depending on
3526 * the debugging options enabled, INIT_WORK(), for instance, can
3527 * trigger an allocation. This too, will make us recurse. Because at
3528 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3529 * the safest choice is to do it like this, wrapping the whole function.
3531 memcg_stop_kmem_account();
3532 __memcg_create_cache_enqueue(memcg, cachep);
3533 memcg_resume_kmem_account();
3536 * Return the kmem_cache we're supposed to use for a slab allocation.
3537 * We try to use the current memcg's version of the cache.
3539 * If the cache does not exist yet, if we are the first user of it,
3540 * we either create it immediately, if possible, or create it asynchronously
3542 * In the latter case, we will let the current allocation go through with
3543 * the original cache.
3545 * Can't be called in interrupt context or from kernel threads.
3546 * This function needs to be called with rcu_read_lock() held.
3548 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3551 struct mem_cgroup *memcg;
3554 VM_BUG_ON(!cachep->memcg_params);
3555 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3557 if (!current->mm || current->memcg_kmem_skip_account)
3561 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3563 if (!memcg_can_account_kmem(memcg))
3566 idx = memcg_cache_id(memcg);
3569 * barrier to mare sure we're always seeing the up to date value. The
3570 * code updating memcg_caches will issue a write barrier to match this.
3572 read_barrier_depends();
3573 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3574 cachep = cachep->memcg_params->memcg_caches[idx];
3578 /* The corresponding put will be done in the workqueue. */
3579 if (!css_tryget(&memcg->css))
3584 * If we are in a safe context (can wait, and not in interrupt
3585 * context), we could be be predictable and return right away.
3586 * This would guarantee that the allocation being performed
3587 * already belongs in the new cache.
3589 * However, there are some clashes that can arrive from locking.
3590 * For instance, because we acquire the slab_mutex while doing
3591 * kmem_cache_dup, this means no further allocation could happen
3592 * with the slab_mutex held.
3594 * Also, because cache creation issue get_online_cpus(), this
3595 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3596 * that ends up reversed during cpu hotplug. (cpuset allocates
3597 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3598 * better to defer everything.
3600 memcg_create_cache_enqueue(memcg, cachep);
3606 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3609 * We need to verify if the allocation against current->mm->owner's memcg is
3610 * possible for the given order. But the page is not allocated yet, so we'll
3611 * need a further commit step to do the final arrangements.
3613 * It is possible for the task to switch cgroups in this mean time, so at
3614 * commit time, we can't rely on task conversion any longer. We'll then use
3615 * the handle argument to return to the caller which cgroup we should commit
3616 * against. We could also return the memcg directly and avoid the pointer
3617 * passing, but a boolean return value gives better semantics considering
3618 * the compiled-out case as well.
3620 * Returning true means the allocation is possible.
3623 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3625 struct mem_cgroup *memcg;
3631 * Disabling accounting is only relevant for some specific memcg
3632 * internal allocations. Therefore we would initially not have such
3633 * check here, since direct calls to the page allocator that are marked
3634 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3635 * concerned with cache allocations, and by having this test at
3636 * memcg_kmem_get_cache, we are already able to relay the allocation to
3637 * the root cache and bypass the memcg cache altogether.
3639 * There is one exception, though: the SLUB allocator does not create
3640 * large order caches, but rather service large kmallocs directly from
3641 * the page allocator. Therefore, the following sequence when backed by
3642 * the SLUB allocator:
3644 * memcg_stop_kmem_account();
3645 * kmalloc(<large_number>)
3646 * memcg_resume_kmem_account();
3648 * would effectively ignore the fact that we should skip accounting,
3649 * since it will drive us directly to this function without passing
3650 * through the cache selector memcg_kmem_get_cache. Such large
3651 * allocations are extremely rare but can happen, for instance, for the
3652 * cache arrays. We bring this test here.
3654 if (!current->mm || current->memcg_kmem_skip_account)
3657 memcg = try_get_mem_cgroup_from_mm(current->mm);
3660 * very rare case described in mem_cgroup_from_task. Unfortunately there
3661 * isn't much we can do without complicating this too much, and it would
3662 * be gfp-dependent anyway. Just let it go
3664 if (unlikely(!memcg))
3667 if (!memcg_can_account_kmem(memcg)) {
3668 css_put(&memcg->css);
3672 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3676 css_put(&memcg->css);
3680 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3683 struct page_cgroup *pc;
3685 VM_BUG_ON(mem_cgroup_is_root(memcg));
3687 /* The page allocation failed. Revert */
3689 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3693 pc = lookup_page_cgroup(page);
3694 lock_page_cgroup(pc);
3695 pc->mem_cgroup = memcg;
3696 SetPageCgroupUsed(pc);
3697 unlock_page_cgroup(pc);
3700 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3702 struct mem_cgroup *memcg = NULL;
3703 struct page_cgroup *pc;
3706 pc = lookup_page_cgroup(page);
3708 * Fast unlocked return. Theoretically might have changed, have to
3709 * check again after locking.
3711 if (!PageCgroupUsed(pc))
3714 lock_page_cgroup(pc);
3715 if (PageCgroupUsed(pc)) {
3716 memcg = pc->mem_cgroup;
3717 ClearPageCgroupUsed(pc);
3719 unlock_page_cgroup(pc);
3722 * We trust that only if there is a memcg associated with the page, it
3723 * is a valid allocation
3728 VM_BUG_ON(mem_cgroup_is_root(memcg));
3729 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3732 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3735 #endif /* CONFIG_MEMCG_KMEM */
3737 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3739 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3741 * Because tail pages are not marked as "used", set it. We're under
3742 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3743 * charge/uncharge will be never happen and move_account() is done under
3744 * compound_lock(), so we don't have to take care of races.
3746 void mem_cgroup_split_huge_fixup(struct page *head)
3748 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3749 struct page_cgroup *pc;
3750 struct mem_cgroup *memcg;
3753 if (mem_cgroup_disabled())
3756 memcg = head_pc->mem_cgroup;
3757 for (i = 1; i < HPAGE_PMD_NR; i++) {
3759 pc->mem_cgroup = memcg;
3760 smp_wmb();/* see __commit_charge() */
3761 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3763 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3766 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3769 * mem_cgroup_move_account - move account of the page
3771 * @nr_pages: number of regular pages (>1 for huge pages)
3772 * @pc: page_cgroup of the page.
3773 * @from: mem_cgroup which the page is moved from.
3774 * @to: mem_cgroup which the page is moved to. @from != @to.
3776 * The caller must confirm following.
3777 * - page is not on LRU (isolate_page() is useful.)
3778 * - compound_lock is held when nr_pages > 1
3780 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3783 static int mem_cgroup_move_account(struct page *page,
3784 unsigned int nr_pages,
3785 struct page_cgroup *pc,
3786 struct mem_cgroup *from,
3787 struct mem_cgroup *to)
3789 unsigned long flags;
3791 bool anon = PageAnon(page);
3793 VM_BUG_ON(from == to);
3794 VM_BUG_ON(PageLRU(page));
3796 * The page is isolated from LRU. So, collapse function
3797 * will not handle this page. But page splitting can happen.
3798 * Do this check under compound_page_lock(). The caller should
3802 if (nr_pages > 1 && !PageTransHuge(page))
3805 lock_page_cgroup(pc);
3808 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3811 move_lock_mem_cgroup(from, &flags);
3813 if (!anon && page_mapped(page)) {
3814 /* Update mapped_file data for mem_cgroup */
3816 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3817 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3820 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3822 /* caller should have done css_get */
3823 pc->mem_cgroup = to;
3824 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3825 move_unlock_mem_cgroup(from, &flags);
3828 unlock_page_cgroup(pc);
3832 memcg_check_events(to, page);
3833 memcg_check_events(from, page);
3839 * mem_cgroup_move_parent - moves page to the parent group
3840 * @page: the page to move
3841 * @pc: page_cgroup of the page
3842 * @child: page's cgroup
3844 * move charges to its parent or the root cgroup if the group has no
3845 * parent (aka use_hierarchy==0).
3846 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3847 * mem_cgroup_move_account fails) the failure is always temporary and
3848 * it signals a race with a page removal/uncharge or migration. In the
3849 * first case the page is on the way out and it will vanish from the LRU
3850 * on the next attempt and the call should be retried later.
3851 * Isolation from the LRU fails only if page has been isolated from
3852 * the LRU since we looked at it and that usually means either global
3853 * reclaim or migration going on. The page will either get back to the
3855 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3856 * (!PageCgroupUsed) or moved to a different group. The page will
3857 * disappear in the next attempt.
3859 static int mem_cgroup_move_parent(struct page *page,
3860 struct page_cgroup *pc,
3861 struct mem_cgroup *child)
3863 struct mem_cgroup *parent;
3864 unsigned int nr_pages;
3865 unsigned long uninitialized_var(flags);
3868 VM_BUG_ON(mem_cgroup_is_root(child));
3871 if (!get_page_unless_zero(page))
3873 if (isolate_lru_page(page))
3876 nr_pages = hpage_nr_pages(page);
3878 parent = parent_mem_cgroup(child);
3880 * If no parent, move charges to root cgroup.
3883 parent = root_mem_cgroup;
3886 VM_BUG_ON(!PageTransHuge(page));
3887 flags = compound_lock_irqsave(page);
3890 ret = mem_cgroup_move_account(page, nr_pages,
3893 __mem_cgroup_cancel_local_charge(child, nr_pages);
3896 compound_unlock_irqrestore(page, flags);
3897 putback_lru_page(page);
3905 * Charge the memory controller for page usage.
3907 * 0 if the charge was successful
3908 * < 0 if the cgroup is over its limit
3910 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3911 gfp_t gfp_mask, enum charge_type ctype)
3913 struct mem_cgroup *memcg = NULL;
3914 unsigned int nr_pages = 1;
3918 if (PageTransHuge(page)) {
3919 nr_pages <<= compound_order(page);
3920 VM_BUG_ON(!PageTransHuge(page));
3922 * Never OOM-kill a process for a huge page. The
3923 * fault handler will fall back to regular pages.
3928 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3931 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3935 int mem_cgroup_newpage_charge(struct page *page,
3936 struct mm_struct *mm, gfp_t gfp_mask)
3938 if (mem_cgroup_disabled())
3940 VM_BUG_ON(page_mapped(page));
3941 VM_BUG_ON(page->mapping && !PageAnon(page));
3943 return mem_cgroup_charge_common(page, mm, gfp_mask,
3944 MEM_CGROUP_CHARGE_TYPE_ANON);
3948 * While swap-in, try_charge -> commit or cancel, the page is locked.
3949 * And when try_charge() successfully returns, one refcnt to memcg without
3950 * struct page_cgroup is acquired. This refcnt will be consumed by
3951 * "commit()" or removed by "cancel()"
3953 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3956 struct mem_cgroup **memcgp)
3958 struct mem_cgroup *memcg;
3959 struct page_cgroup *pc;
3962 pc = lookup_page_cgroup(page);
3964 * Every swap fault against a single page tries to charge the
3965 * page, bail as early as possible. shmem_unuse() encounters
3966 * already charged pages, too. The USED bit is protected by
3967 * the page lock, which serializes swap cache removal, which
3968 * in turn serializes uncharging.
3970 if (PageCgroupUsed(pc))
3972 if (!do_swap_account)
3974 memcg = try_get_mem_cgroup_from_page(page);
3978 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3979 css_put(&memcg->css);
3984 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3990 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3991 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3994 if (mem_cgroup_disabled())
3997 * A racing thread's fault, or swapoff, may have already
3998 * updated the pte, and even removed page from swap cache: in
3999 * those cases unuse_pte()'s pte_same() test will fail; but
4000 * there's also a KSM case which does need to charge the page.
4002 if (!PageSwapCache(page)) {
4005 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4010 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4013 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4015 if (mem_cgroup_disabled())
4019 __mem_cgroup_cancel_charge(memcg, 1);
4023 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4024 enum charge_type ctype)
4026 if (mem_cgroup_disabled())
4031 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4033 * Now swap is on-memory. This means this page may be
4034 * counted both as mem and swap....double count.
4035 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4036 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4037 * may call delete_from_swap_cache() before reach here.
4039 if (do_swap_account && PageSwapCache(page)) {
4040 swp_entry_t ent = {.val = page_private(page)};
4041 mem_cgroup_uncharge_swap(ent);
4045 void mem_cgroup_commit_charge_swapin(struct page *page,
4046 struct mem_cgroup *memcg)
4048 __mem_cgroup_commit_charge_swapin(page, memcg,
4049 MEM_CGROUP_CHARGE_TYPE_ANON);
4052 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4055 struct mem_cgroup *memcg = NULL;
4056 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4059 if (mem_cgroup_disabled())
4061 if (PageCompound(page))
4064 if (!PageSwapCache(page))
4065 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4066 else { /* page is swapcache/shmem */
4067 ret = __mem_cgroup_try_charge_swapin(mm, page,
4070 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4075 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4076 unsigned int nr_pages,
4077 const enum charge_type ctype)
4079 struct memcg_batch_info *batch = NULL;
4080 bool uncharge_memsw = true;
4082 /* If swapout, usage of swap doesn't decrease */
4083 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4084 uncharge_memsw = false;
4086 batch = ¤t->memcg_batch;
4088 * In usual, we do css_get() when we remember memcg pointer.
4089 * But in this case, we keep res->usage until end of a series of
4090 * uncharges. Then, it's ok to ignore memcg's refcnt.
4093 batch->memcg = memcg;
4095 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4096 * In those cases, all pages freed continuously can be expected to be in
4097 * the same cgroup and we have chance to coalesce uncharges.
4098 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4099 * because we want to do uncharge as soon as possible.
4102 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4103 goto direct_uncharge;
4106 goto direct_uncharge;
4109 * In typical case, batch->memcg == mem. This means we can
4110 * merge a series of uncharges to an uncharge of res_counter.
4111 * If not, we uncharge res_counter ony by one.
4113 if (batch->memcg != memcg)
4114 goto direct_uncharge;
4115 /* remember freed charge and uncharge it later */
4118 batch->memsw_nr_pages++;
4121 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4123 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4124 if (unlikely(batch->memcg != memcg))
4125 memcg_oom_recover(memcg);
4129 * uncharge if !page_mapped(page)
4131 static struct mem_cgroup *
4132 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4135 struct mem_cgroup *memcg = NULL;
4136 unsigned int nr_pages = 1;
4137 struct page_cgroup *pc;
4140 if (mem_cgroup_disabled())
4143 if (PageTransHuge(page)) {
4144 nr_pages <<= compound_order(page);
4145 VM_BUG_ON(!PageTransHuge(page));
4148 * Check if our page_cgroup is valid
4150 pc = lookup_page_cgroup(page);
4151 if (unlikely(!PageCgroupUsed(pc)))
4154 lock_page_cgroup(pc);
4156 memcg = pc->mem_cgroup;
4158 if (!PageCgroupUsed(pc))
4161 anon = PageAnon(page);
4164 case MEM_CGROUP_CHARGE_TYPE_ANON:
4166 * Generally PageAnon tells if it's the anon statistics to be
4167 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4168 * used before page reached the stage of being marked PageAnon.
4172 case MEM_CGROUP_CHARGE_TYPE_DROP:
4173 /* See mem_cgroup_prepare_migration() */
4174 if (page_mapped(page))
4177 * Pages under migration may not be uncharged. But
4178 * end_migration() /must/ be the one uncharging the
4179 * unused post-migration page and so it has to call
4180 * here with the migration bit still set. See the
4181 * res_counter handling below.
4183 if (!end_migration && PageCgroupMigration(pc))
4186 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4187 if (!PageAnon(page)) { /* Shared memory */
4188 if (page->mapping && !page_is_file_cache(page))
4190 } else if (page_mapped(page)) /* Anon */
4197 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4199 ClearPageCgroupUsed(pc);
4201 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4202 * freed from LRU. This is safe because uncharged page is expected not
4203 * to be reused (freed soon). Exception is SwapCache, it's handled by
4204 * special functions.
4207 unlock_page_cgroup(pc);
4209 * even after unlock, we have memcg->res.usage here and this memcg
4210 * will never be freed, so it's safe to call css_get().
4212 memcg_check_events(memcg, page);
4213 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4214 mem_cgroup_swap_statistics(memcg, true);
4215 css_get(&memcg->css);
4218 * Migration does not charge the res_counter for the
4219 * replacement page, so leave it alone when phasing out the
4220 * page that is unused after the migration.
4222 if (!end_migration && !mem_cgroup_is_root(memcg))
4223 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4228 unlock_page_cgroup(pc);
4232 void mem_cgroup_uncharge_page(struct page *page)
4235 if (page_mapped(page))
4237 VM_BUG_ON(page->mapping && !PageAnon(page));
4239 * If the page is in swap cache, uncharge should be deferred
4240 * to the swap path, which also properly accounts swap usage
4241 * and handles memcg lifetime.
4243 * Note that this check is not stable and reclaim may add the
4244 * page to swap cache at any time after this. However, if the
4245 * page is not in swap cache by the time page->mapcount hits
4246 * 0, there won't be any page table references to the swap
4247 * slot, and reclaim will free it and not actually write the
4250 if (PageSwapCache(page))
4252 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4255 void mem_cgroup_uncharge_cache_page(struct page *page)
4257 VM_BUG_ON(page_mapped(page));
4258 VM_BUG_ON(page->mapping);
4259 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4263 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4264 * In that cases, pages are freed continuously and we can expect pages
4265 * are in the same memcg. All these calls itself limits the number of
4266 * pages freed at once, then uncharge_start/end() is called properly.
4267 * This may be called prural(2) times in a context,
4270 void mem_cgroup_uncharge_start(void)
4272 current->memcg_batch.do_batch++;
4273 /* We can do nest. */
4274 if (current->memcg_batch.do_batch == 1) {
4275 current->memcg_batch.memcg = NULL;
4276 current->memcg_batch.nr_pages = 0;
4277 current->memcg_batch.memsw_nr_pages = 0;
4281 void mem_cgroup_uncharge_end(void)
4283 struct memcg_batch_info *batch = ¤t->memcg_batch;
4285 if (!batch->do_batch)
4289 if (batch->do_batch) /* If stacked, do nothing. */
4295 * This "batch->memcg" is valid without any css_get/put etc...
4296 * bacause we hide charges behind us.
4298 if (batch->nr_pages)
4299 res_counter_uncharge(&batch->memcg->res,
4300 batch->nr_pages * PAGE_SIZE);
4301 if (batch->memsw_nr_pages)
4302 res_counter_uncharge(&batch->memcg->memsw,
4303 batch->memsw_nr_pages * PAGE_SIZE);
4304 memcg_oom_recover(batch->memcg);
4305 /* forget this pointer (for sanity check) */
4306 batch->memcg = NULL;
4311 * called after __delete_from_swap_cache() and drop "page" account.
4312 * memcg information is recorded to swap_cgroup of "ent"
4315 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4317 struct mem_cgroup *memcg;
4318 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4320 if (!swapout) /* this was a swap cache but the swap is unused ! */
4321 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4323 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4326 * record memcg information, if swapout && memcg != NULL,
4327 * css_get() was called in uncharge().
4329 if (do_swap_account && swapout && memcg)
4330 swap_cgroup_record(ent, css_id(&memcg->css));
4334 #ifdef CONFIG_MEMCG_SWAP
4336 * called from swap_entry_free(). remove record in swap_cgroup and
4337 * uncharge "memsw" account.
4339 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4341 struct mem_cgroup *memcg;
4344 if (!do_swap_account)
4347 id = swap_cgroup_record(ent, 0);
4349 memcg = mem_cgroup_lookup(id);
4352 * We uncharge this because swap is freed.
4353 * This memcg can be obsolete one. We avoid calling css_tryget
4355 if (!mem_cgroup_is_root(memcg))
4356 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4357 mem_cgroup_swap_statistics(memcg, false);
4358 css_put(&memcg->css);
4364 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4365 * @entry: swap entry to be moved
4366 * @from: mem_cgroup which the entry is moved from
4367 * @to: mem_cgroup which the entry is moved to
4369 * It succeeds only when the swap_cgroup's record for this entry is the same
4370 * as the mem_cgroup's id of @from.
4372 * Returns 0 on success, -EINVAL on failure.
4374 * The caller must have charged to @to, IOW, called res_counter_charge() about
4375 * both res and memsw, and called css_get().
4377 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4378 struct mem_cgroup *from, struct mem_cgroup *to)
4380 unsigned short old_id, new_id;
4382 old_id = css_id(&from->css);
4383 new_id = css_id(&to->css);
4385 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4386 mem_cgroup_swap_statistics(from, false);
4387 mem_cgroup_swap_statistics(to, true);
4389 * This function is only called from task migration context now.
4390 * It postpones res_counter and refcount handling till the end
4391 * of task migration(mem_cgroup_clear_mc()) for performance
4392 * improvement. But we cannot postpone css_get(to) because if
4393 * the process that has been moved to @to does swap-in, the
4394 * refcount of @to might be decreased to 0.
4396 * We are in attach() phase, so the cgroup is guaranteed to be
4397 * alive, so we can just call css_get().
4405 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4406 struct mem_cgroup *from, struct mem_cgroup *to)
4413 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4416 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4417 struct mem_cgroup **memcgp)
4419 struct mem_cgroup *memcg = NULL;
4420 unsigned int nr_pages = 1;
4421 struct page_cgroup *pc;
4422 enum charge_type ctype;
4426 if (mem_cgroup_disabled())
4429 if (PageTransHuge(page))
4430 nr_pages <<= compound_order(page);
4432 pc = lookup_page_cgroup(page);
4433 lock_page_cgroup(pc);
4434 if (PageCgroupUsed(pc)) {
4435 memcg = pc->mem_cgroup;
4436 css_get(&memcg->css);
4438 * At migrating an anonymous page, its mapcount goes down
4439 * to 0 and uncharge() will be called. But, even if it's fully
4440 * unmapped, migration may fail and this page has to be
4441 * charged again. We set MIGRATION flag here and delay uncharge
4442 * until end_migration() is called
4444 * Corner Case Thinking
4446 * When the old page was mapped as Anon and it's unmap-and-freed
4447 * while migration was ongoing.
4448 * If unmap finds the old page, uncharge() of it will be delayed
4449 * until end_migration(). If unmap finds a new page, it's
4450 * uncharged when it make mapcount to be 1->0. If unmap code
4451 * finds swap_migration_entry, the new page will not be mapped
4452 * and end_migration() will find it(mapcount==0).
4455 * When the old page was mapped but migraion fails, the kernel
4456 * remaps it. A charge for it is kept by MIGRATION flag even
4457 * if mapcount goes down to 0. We can do remap successfully
4458 * without charging it again.
4461 * The "old" page is under lock_page() until the end of
4462 * migration, so, the old page itself will not be swapped-out.
4463 * If the new page is swapped out before end_migraton, our
4464 * hook to usual swap-out path will catch the event.
4467 SetPageCgroupMigration(pc);
4469 unlock_page_cgroup(pc);
4471 * If the page is not charged at this point,
4479 * We charge new page before it's used/mapped. So, even if unlock_page()
4480 * is called before end_migration, we can catch all events on this new
4481 * page. In the case new page is migrated but not remapped, new page's
4482 * mapcount will be finally 0 and we call uncharge in end_migration().
4485 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4487 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4489 * The page is committed to the memcg, but it's not actually
4490 * charged to the res_counter since we plan on replacing the
4491 * old one and only one page is going to be left afterwards.
4493 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4496 /* remove redundant charge if migration failed*/
4497 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4498 struct page *oldpage, struct page *newpage, bool migration_ok)
4500 struct page *used, *unused;
4501 struct page_cgroup *pc;
4507 if (!migration_ok) {
4514 anon = PageAnon(used);
4515 __mem_cgroup_uncharge_common(unused,
4516 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4517 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4519 css_put(&memcg->css);
4521 * We disallowed uncharge of pages under migration because mapcount
4522 * of the page goes down to zero, temporarly.
4523 * Clear the flag and check the page should be charged.
4525 pc = lookup_page_cgroup(oldpage);
4526 lock_page_cgroup(pc);
4527 ClearPageCgroupMigration(pc);
4528 unlock_page_cgroup(pc);
4531 * If a page is a file cache, radix-tree replacement is very atomic
4532 * and we can skip this check. When it was an Anon page, its mapcount
4533 * goes down to 0. But because we added MIGRATION flage, it's not
4534 * uncharged yet. There are several case but page->mapcount check
4535 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4536 * check. (see prepare_charge() also)
4539 mem_cgroup_uncharge_page(used);
4543 * At replace page cache, newpage is not under any memcg but it's on
4544 * LRU. So, this function doesn't touch res_counter but handles LRU
4545 * in correct way. Both pages are locked so we cannot race with uncharge.
4547 void mem_cgroup_replace_page_cache(struct page *oldpage,
4548 struct page *newpage)
4550 struct mem_cgroup *memcg = NULL;
4551 struct page_cgroup *pc;
4552 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4554 if (mem_cgroup_disabled())
4557 pc = lookup_page_cgroup(oldpage);
4558 /* fix accounting on old pages */
4559 lock_page_cgroup(pc);
4560 if (PageCgroupUsed(pc)) {
4561 memcg = pc->mem_cgroup;
4562 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4563 ClearPageCgroupUsed(pc);
4565 unlock_page_cgroup(pc);
4568 * When called from shmem_replace_page(), in some cases the
4569 * oldpage has already been charged, and in some cases not.
4574 * Even if newpage->mapping was NULL before starting replacement,
4575 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4576 * LRU while we overwrite pc->mem_cgroup.
4578 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4581 #ifdef CONFIG_DEBUG_VM
4582 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4584 struct page_cgroup *pc;
4586 pc = lookup_page_cgroup(page);
4588 * Can be NULL while feeding pages into the page allocator for
4589 * the first time, i.e. during boot or memory hotplug;
4590 * or when mem_cgroup_disabled().
4592 if (likely(pc) && PageCgroupUsed(pc))
4597 bool mem_cgroup_bad_page_check(struct page *page)
4599 if (mem_cgroup_disabled())
4602 return lookup_page_cgroup_used(page) != NULL;
4605 void mem_cgroup_print_bad_page(struct page *page)
4607 struct page_cgroup *pc;
4609 pc = lookup_page_cgroup_used(page);
4611 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4612 pc, pc->flags, pc->mem_cgroup);
4617 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4618 unsigned long long val)
4621 u64 memswlimit, memlimit;
4623 int children = mem_cgroup_count_children(memcg);
4624 u64 curusage, oldusage;
4628 * For keeping hierarchical_reclaim simple, how long we should retry
4629 * is depends on callers. We set our retry-count to be function
4630 * of # of children which we should visit in this loop.
4632 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4634 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4637 while (retry_count) {
4638 if (signal_pending(current)) {
4643 * Rather than hide all in some function, I do this in
4644 * open coded manner. You see what this really does.
4645 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4647 mutex_lock(&set_limit_mutex);
4648 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4649 if (memswlimit < val) {
4651 mutex_unlock(&set_limit_mutex);
4655 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4659 ret = res_counter_set_limit(&memcg->res, val);
4661 if (memswlimit == val)
4662 memcg->memsw_is_minimum = true;
4664 memcg->memsw_is_minimum = false;
4666 mutex_unlock(&set_limit_mutex);
4671 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4672 MEM_CGROUP_RECLAIM_SHRINK);
4673 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4674 /* Usage is reduced ? */
4675 if (curusage >= oldusage)
4678 oldusage = curusage;
4680 if (!ret && enlarge)
4681 memcg_oom_recover(memcg);
4686 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4687 unsigned long long val)
4690 u64 memlimit, memswlimit, oldusage, curusage;
4691 int children = mem_cgroup_count_children(memcg);
4695 /* see mem_cgroup_resize_res_limit */
4696 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4697 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4698 while (retry_count) {
4699 if (signal_pending(current)) {
4704 * Rather than hide all in some function, I do this in
4705 * open coded manner. You see what this really does.
4706 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4708 mutex_lock(&set_limit_mutex);
4709 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4710 if (memlimit > val) {
4712 mutex_unlock(&set_limit_mutex);
4715 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4716 if (memswlimit < val)
4718 ret = res_counter_set_limit(&memcg->memsw, val);
4720 if (memlimit == val)
4721 memcg->memsw_is_minimum = true;
4723 memcg->memsw_is_minimum = false;
4725 mutex_unlock(&set_limit_mutex);
4730 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4731 MEM_CGROUP_RECLAIM_NOSWAP |
4732 MEM_CGROUP_RECLAIM_SHRINK);
4733 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4734 /* Usage is reduced ? */
4735 if (curusage >= oldusage)
4738 oldusage = curusage;
4740 if (!ret && enlarge)
4741 memcg_oom_recover(memcg);
4745 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4747 unsigned long *total_scanned)
4749 unsigned long nr_reclaimed = 0;
4750 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4751 unsigned long reclaimed;
4753 struct mem_cgroup_tree_per_zone *mctz;
4754 unsigned long long excess;
4755 unsigned long nr_scanned;
4760 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4762 * This loop can run a while, specially if mem_cgroup's continuously
4763 * keep exceeding their soft limit and putting the system under
4770 mz = mem_cgroup_largest_soft_limit_node(mctz);
4775 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4776 gfp_mask, &nr_scanned);
4777 nr_reclaimed += reclaimed;
4778 *total_scanned += nr_scanned;
4779 spin_lock(&mctz->lock);
4782 * If we failed to reclaim anything from this memory cgroup
4783 * it is time to move on to the next cgroup
4789 * Loop until we find yet another one.
4791 * By the time we get the soft_limit lock
4792 * again, someone might have aded the
4793 * group back on the RB tree. Iterate to
4794 * make sure we get a different mem.
4795 * mem_cgroup_largest_soft_limit_node returns
4796 * NULL if no other cgroup is present on
4800 __mem_cgroup_largest_soft_limit_node(mctz);
4802 css_put(&next_mz->memcg->css);
4803 else /* next_mz == NULL or other memcg */
4807 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4808 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4810 * One school of thought says that we should not add
4811 * back the node to the tree if reclaim returns 0.
4812 * But our reclaim could return 0, simply because due
4813 * to priority we are exposing a smaller subset of
4814 * memory to reclaim from. Consider this as a longer
4817 /* If excess == 0, no tree ops */
4818 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4819 spin_unlock(&mctz->lock);
4820 css_put(&mz->memcg->css);
4823 * Could not reclaim anything and there are no more
4824 * mem cgroups to try or we seem to be looping without
4825 * reclaiming anything.
4827 if (!nr_reclaimed &&
4829 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4831 } while (!nr_reclaimed);
4833 css_put(&next_mz->memcg->css);
4834 return nr_reclaimed;
4838 * mem_cgroup_force_empty_list - clears LRU of a group
4839 * @memcg: group to clear
4842 * @lru: lru to to clear
4844 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4845 * reclaim the pages page themselves - pages are moved to the parent (or root)
4848 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4849 int node, int zid, enum lru_list lru)
4851 struct lruvec *lruvec;
4852 unsigned long flags;
4853 struct list_head *list;
4857 zone = &NODE_DATA(node)->node_zones[zid];
4858 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4859 list = &lruvec->lists[lru];
4863 struct page_cgroup *pc;
4866 spin_lock_irqsave(&zone->lru_lock, flags);
4867 if (list_empty(list)) {
4868 spin_unlock_irqrestore(&zone->lru_lock, flags);
4871 page = list_entry(list->prev, struct page, lru);
4873 list_move(&page->lru, list);
4875 spin_unlock_irqrestore(&zone->lru_lock, flags);
4878 spin_unlock_irqrestore(&zone->lru_lock, flags);
4880 pc = lookup_page_cgroup(page);
4882 if (mem_cgroup_move_parent(page, pc, memcg)) {
4883 /* found lock contention or "pc" is obsolete. */
4888 } while (!list_empty(list));
4892 * make mem_cgroup's charge to be 0 if there is no task by moving
4893 * all the charges and pages to the parent.
4894 * This enables deleting this mem_cgroup.
4896 * Caller is responsible for holding css reference on the memcg.
4898 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4904 /* This is for making all *used* pages to be on LRU. */
4905 lru_add_drain_all();
4906 drain_all_stock_sync(memcg);
4907 mem_cgroup_start_move(memcg);
4908 for_each_node_state(node, N_MEMORY) {
4909 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4912 mem_cgroup_force_empty_list(memcg,
4917 mem_cgroup_end_move(memcg);
4918 memcg_oom_recover(memcg);
4922 * Kernel memory may not necessarily be trackable to a specific
4923 * process. So they are not migrated, and therefore we can't
4924 * expect their value to drop to 0 here.
4925 * Having res filled up with kmem only is enough.
4927 * This is a safety check because mem_cgroup_force_empty_list
4928 * could have raced with mem_cgroup_replace_page_cache callers
4929 * so the lru seemed empty but the page could have been added
4930 * right after the check. RES_USAGE should be safe as we always
4931 * charge before adding to the LRU.
4933 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4934 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4935 } while (usage > 0);
4939 * This mainly exists for tests during the setting of set of use_hierarchy.
4940 * Since this is the very setting we are changing, the current hierarchy value
4943 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4947 /* bounce at first found */
4948 cgroup_for_each_child(pos, memcg->css.cgroup)
4954 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4955 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4956 * from mem_cgroup_count_children(), in the sense that we don't really care how
4957 * many children we have; we only need to know if we have any. It also counts
4958 * any memcg without hierarchy as infertile.
4960 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4962 return memcg->use_hierarchy && __memcg_has_children(memcg);
4966 * Reclaims as many pages from the given memcg as possible and moves
4967 * the rest to the parent.
4969 * Caller is responsible for holding css reference for memcg.
4971 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4973 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4974 struct cgroup *cgrp = memcg->css.cgroup;
4976 /* returns EBUSY if there is a task or if we come here twice. */
4977 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4980 /* we call try-to-free pages for make this cgroup empty */
4981 lru_add_drain_all();
4982 /* try to free all pages in this cgroup */
4983 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4986 if (signal_pending(current))
4989 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4993 /* maybe some writeback is necessary */
4994 congestion_wait(BLK_RW_ASYNC, HZ/10);
4999 mem_cgroup_reparent_charges(memcg);
5004 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
5006 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5009 if (mem_cgroup_is_root(memcg))
5011 css_get(&memcg->css);
5012 ret = mem_cgroup_force_empty(memcg);
5013 css_put(&memcg->css);
5019 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
5021 return mem_cgroup_from_cont(cont)->use_hierarchy;
5024 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
5028 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5029 struct cgroup *parent = cont->parent;
5030 struct mem_cgroup *parent_memcg = NULL;
5033 parent_memcg = mem_cgroup_from_cont(parent);
5035 mutex_lock(&memcg_create_mutex);
5037 if (memcg->use_hierarchy == val)
5041 * If parent's use_hierarchy is set, we can't make any modifications
5042 * in the child subtrees. If it is unset, then the change can
5043 * occur, provided the current cgroup has no children.
5045 * For the root cgroup, parent_mem is NULL, we allow value to be
5046 * set if there are no children.
5048 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5049 (val == 1 || val == 0)) {
5050 if (!__memcg_has_children(memcg))
5051 memcg->use_hierarchy = val;
5058 mutex_unlock(&memcg_create_mutex);
5064 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5065 enum mem_cgroup_stat_index idx)
5067 struct mem_cgroup *iter;
5070 /* Per-cpu values can be negative, use a signed accumulator */
5071 for_each_mem_cgroup_tree(iter, memcg)
5072 val += mem_cgroup_read_stat(iter, idx);
5074 if (val < 0) /* race ? */
5079 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5083 if (!mem_cgroup_is_root(memcg)) {
5085 return res_counter_read_u64(&memcg->res, RES_USAGE);
5087 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5091 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5092 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5094 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5095 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5098 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5100 return val << PAGE_SHIFT;
5103 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
5104 struct file *file, char __user *buf,
5105 size_t nbytes, loff_t *ppos)
5107 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5113 type = MEMFILE_TYPE(cft->private);
5114 name = MEMFILE_ATTR(cft->private);
5118 if (name == RES_USAGE)
5119 val = mem_cgroup_usage(memcg, false);
5121 val = res_counter_read_u64(&memcg->res, name);
5124 if (name == RES_USAGE)
5125 val = mem_cgroup_usage(memcg, true);
5127 val = res_counter_read_u64(&memcg->memsw, name);
5130 val = res_counter_read_u64(&memcg->kmem, name);
5136 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5137 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5140 static int memcg_update_kmem_limit(struct cgroup *cont, u64 val)
5143 #ifdef CONFIG_MEMCG_KMEM
5144 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5146 * For simplicity, we won't allow this to be disabled. It also can't
5147 * be changed if the cgroup has children already, or if tasks had
5150 * If tasks join before we set the limit, a person looking at
5151 * kmem.usage_in_bytes will have no way to determine when it took
5152 * place, which makes the value quite meaningless.
5154 * After it first became limited, changes in the value of the limit are
5155 * of course permitted.
5157 mutex_lock(&memcg_create_mutex);
5158 mutex_lock(&set_limit_mutex);
5159 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
5160 if (cgroup_task_count(cont) || memcg_has_children(memcg)) {
5164 ret = res_counter_set_limit(&memcg->kmem, val);
5167 ret = memcg_update_cache_sizes(memcg);
5169 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5172 static_key_slow_inc(&memcg_kmem_enabled_key);
5174 * setting the active bit after the inc will guarantee no one
5175 * starts accounting before all call sites are patched
5177 memcg_kmem_set_active(memcg);
5179 ret = res_counter_set_limit(&memcg->kmem, val);
5181 mutex_unlock(&set_limit_mutex);
5182 mutex_unlock(&memcg_create_mutex);
5187 #ifdef CONFIG_MEMCG_KMEM
5188 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5191 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5195 memcg->kmem_account_flags = parent->kmem_account_flags;
5197 * When that happen, we need to disable the static branch only on those
5198 * memcgs that enabled it. To achieve this, we would be forced to
5199 * complicate the code by keeping track of which memcgs were the ones
5200 * that actually enabled limits, and which ones got it from its
5203 * It is a lot simpler just to do static_key_slow_inc() on every child
5204 * that is accounted.
5206 if (!memcg_kmem_is_active(memcg))
5210 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5211 * memcg is active already. If the later initialization fails then the
5212 * cgroup core triggers the cleanup so we do not have to do it here.
5214 static_key_slow_inc(&memcg_kmem_enabled_key);
5216 mutex_lock(&set_limit_mutex);
5217 memcg_stop_kmem_account();
5218 ret = memcg_update_cache_sizes(memcg);
5219 memcg_resume_kmem_account();
5220 mutex_unlock(&set_limit_mutex);
5224 #endif /* CONFIG_MEMCG_KMEM */
5227 * The user of this function is...
5230 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
5233 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5236 unsigned long long val;
5239 type = MEMFILE_TYPE(cft->private);
5240 name = MEMFILE_ATTR(cft->private);
5244 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5248 /* This function does all necessary parse...reuse it */
5249 ret = res_counter_memparse_write_strategy(buffer, &val);
5253 ret = mem_cgroup_resize_limit(memcg, val);
5254 else if (type == _MEMSWAP)
5255 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5256 else if (type == _KMEM)
5257 ret = memcg_update_kmem_limit(cont, val);
5261 case RES_SOFT_LIMIT:
5262 ret = res_counter_memparse_write_strategy(buffer, &val);
5266 * For memsw, soft limits are hard to implement in terms
5267 * of semantics, for now, we support soft limits for
5268 * control without swap
5271 ret = res_counter_set_soft_limit(&memcg->res, val);
5276 ret = -EINVAL; /* should be BUG() ? */
5282 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5283 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5285 struct cgroup *cgroup;
5286 unsigned long long min_limit, min_memsw_limit, tmp;
5288 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5289 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5290 cgroup = memcg->css.cgroup;
5291 if (!memcg->use_hierarchy)
5294 while (cgroup->parent) {
5295 cgroup = cgroup->parent;
5296 memcg = mem_cgroup_from_cont(cgroup);
5297 if (!memcg->use_hierarchy)
5299 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5300 min_limit = min(min_limit, tmp);
5301 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5302 min_memsw_limit = min(min_memsw_limit, tmp);
5305 *mem_limit = min_limit;
5306 *memsw_limit = min_memsw_limit;
5309 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
5311 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5315 type = MEMFILE_TYPE(event);
5316 name = MEMFILE_ATTR(event);
5321 res_counter_reset_max(&memcg->res);
5322 else if (type == _MEMSWAP)
5323 res_counter_reset_max(&memcg->memsw);
5324 else if (type == _KMEM)
5325 res_counter_reset_max(&memcg->kmem);
5331 res_counter_reset_failcnt(&memcg->res);
5332 else if (type == _MEMSWAP)
5333 res_counter_reset_failcnt(&memcg->memsw);
5334 else if (type == _KMEM)
5335 res_counter_reset_failcnt(&memcg->kmem);
5344 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
5347 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
5351 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5352 struct cftype *cft, u64 val)
5354 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5356 if (val >= (1 << NR_MOVE_TYPE))
5360 * No kind of locking is needed in here, because ->can_attach() will
5361 * check this value once in the beginning of the process, and then carry
5362 * on with stale data. This means that changes to this value will only
5363 * affect task migrations starting after the change.
5365 memcg->move_charge_at_immigrate = val;
5369 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
5370 struct cftype *cft, u64 val)
5377 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
5381 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5382 unsigned long node_nr;
5383 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5385 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5386 seq_printf(m, "total=%lu", total_nr);
5387 for_each_node_state(nid, N_MEMORY) {
5388 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5389 seq_printf(m, " N%d=%lu", nid, node_nr);
5393 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5394 seq_printf(m, "file=%lu", file_nr);
5395 for_each_node_state(nid, N_MEMORY) {
5396 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5398 seq_printf(m, " N%d=%lu", nid, node_nr);
5402 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5403 seq_printf(m, "anon=%lu", anon_nr);
5404 for_each_node_state(nid, N_MEMORY) {
5405 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5407 seq_printf(m, " N%d=%lu", nid, node_nr);
5411 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5412 seq_printf(m, "unevictable=%lu", unevictable_nr);
5413 for_each_node_state(nid, N_MEMORY) {
5414 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5415 BIT(LRU_UNEVICTABLE));
5416 seq_printf(m, " N%d=%lu", nid, node_nr);
5421 #endif /* CONFIG_NUMA */
5423 static inline void mem_cgroup_lru_names_not_uptodate(void)
5425 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5428 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
5431 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5432 struct mem_cgroup *mi;
5435 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5436 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5438 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5439 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5442 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5443 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5444 mem_cgroup_read_events(memcg, i));
5446 for (i = 0; i < NR_LRU_LISTS; i++)
5447 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5448 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5450 /* Hierarchical information */
5452 unsigned long long limit, memsw_limit;
5453 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5454 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5455 if (do_swap_account)
5456 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5460 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5463 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5465 for_each_mem_cgroup_tree(mi, memcg)
5466 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5467 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5470 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5471 unsigned long long val = 0;
5473 for_each_mem_cgroup_tree(mi, memcg)
5474 val += mem_cgroup_read_events(mi, i);
5475 seq_printf(m, "total_%s %llu\n",
5476 mem_cgroup_events_names[i], val);
5479 for (i = 0; i < NR_LRU_LISTS; i++) {
5480 unsigned long long val = 0;
5482 for_each_mem_cgroup_tree(mi, memcg)
5483 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5484 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5487 #ifdef CONFIG_DEBUG_VM
5490 struct mem_cgroup_per_zone *mz;
5491 struct zone_reclaim_stat *rstat;
5492 unsigned long recent_rotated[2] = {0, 0};
5493 unsigned long recent_scanned[2] = {0, 0};
5495 for_each_online_node(nid)
5496 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5497 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5498 rstat = &mz->lruvec.reclaim_stat;
5500 recent_rotated[0] += rstat->recent_rotated[0];
5501 recent_rotated[1] += rstat->recent_rotated[1];
5502 recent_scanned[0] += rstat->recent_scanned[0];
5503 recent_scanned[1] += rstat->recent_scanned[1];
5505 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5506 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5507 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5508 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5515 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
5517 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5519 return mem_cgroup_swappiness(memcg);
5522 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
5525 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5526 struct mem_cgroup *parent;
5531 if (cgrp->parent == NULL)
5534 parent = mem_cgroup_from_cont(cgrp->parent);
5536 mutex_lock(&memcg_create_mutex);
5538 /* If under hierarchy, only empty-root can set this value */
5539 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5540 mutex_unlock(&memcg_create_mutex);
5544 memcg->swappiness = val;
5546 mutex_unlock(&memcg_create_mutex);
5551 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5553 struct mem_cgroup_threshold_ary *t;
5559 t = rcu_dereference(memcg->thresholds.primary);
5561 t = rcu_dereference(memcg->memsw_thresholds.primary);
5566 usage = mem_cgroup_usage(memcg, swap);
5569 * current_threshold points to threshold just below or equal to usage.
5570 * If it's not true, a threshold was crossed after last
5571 * call of __mem_cgroup_threshold().
5573 i = t->current_threshold;
5576 * Iterate backward over array of thresholds starting from
5577 * current_threshold and check if a threshold is crossed.
5578 * If none of thresholds below usage is crossed, we read
5579 * only one element of the array here.
5581 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5582 eventfd_signal(t->entries[i].eventfd, 1);
5584 /* i = current_threshold + 1 */
5588 * Iterate forward over array of thresholds starting from
5589 * current_threshold+1 and check if a threshold is crossed.
5590 * If none of thresholds above usage is crossed, we read
5591 * only one element of the array here.
5593 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5594 eventfd_signal(t->entries[i].eventfd, 1);
5596 /* Update current_threshold */
5597 t->current_threshold = i - 1;
5602 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5605 __mem_cgroup_threshold(memcg, false);
5606 if (do_swap_account)
5607 __mem_cgroup_threshold(memcg, true);
5609 memcg = parent_mem_cgroup(memcg);
5613 static int compare_thresholds(const void *a, const void *b)
5615 const struct mem_cgroup_threshold *_a = a;
5616 const struct mem_cgroup_threshold *_b = b;
5618 return _a->threshold - _b->threshold;
5621 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5623 struct mem_cgroup_eventfd_list *ev;
5625 list_for_each_entry(ev, &memcg->oom_notify, list)
5626 eventfd_signal(ev->eventfd, 1);
5630 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5632 struct mem_cgroup *iter;
5634 for_each_mem_cgroup_tree(iter, memcg)
5635 mem_cgroup_oom_notify_cb(iter);
5638 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
5639 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5641 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5642 struct mem_cgroup_thresholds *thresholds;
5643 struct mem_cgroup_threshold_ary *new;
5644 enum res_type type = MEMFILE_TYPE(cft->private);
5645 u64 threshold, usage;
5648 ret = res_counter_memparse_write_strategy(args, &threshold);
5652 mutex_lock(&memcg->thresholds_lock);
5655 thresholds = &memcg->thresholds;
5656 else if (type == _MEMSWAP)
5657 thresholds = &memcg->memsw_thresholds;
5661 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5663 /* Check if a threshold crossed before adding a new one */
5664 if (thresholds->primary)
5665 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5667 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5669 /* Allocate memory for new array of thresholds */
5670 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5678 /* Copy thresholds (if any) to new array */
5679 if (thresholds->primary) {
5680 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5681 sizeof(struct mem_cgroup_threshold));
5684 /* Add new threshold */
5685 new->entries[size - 1].eventfd = eventfd;
5686 new->entries[size - 1].threshold = threshold;
5688 /* Sort thresholds. Registering of new threshold isn't time-critical */
5689 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5690 compare_thresholds, NULL);
5692 /* Find current threshold */
5693 new->current_threshold = -1;
5694 for (i = 0; i < size; i++) {
5695 if (new->entries[i].threshold <= usage) {
5697 * new->current_threshold will not be used until
5698 * rcu_assign_pointer(), so it's safe to increment
5701 ++new->current_threshold;
5706 /* Free old spare buffer and save old primary buffer as spare */
5707 kfree(thresholds->spare);
5708 thresholds->spare = thresholds->primary;
5710 rcu_assign_pointer(thresholds->primary, new);
5712 /* To be sure that nobody uses thresholds */
5716 mutex_unlock(&memcg->thresholds_lock);
5721 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
5722 struct cftype *cft, struct eventfd_ctx *eventfd)
5724 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5725 struct mem_cgroup_thresholds *thresholds;
5726 struct mem_cgroup_threshold_ary *new;
5727 enum res_type type = MEMFILE_TYPE(cft->private);
5731 mutex_lock(&memcg->thresholds_lock);
5733 thresholds = &memcg->thresholds;
5734 else if (type == _MEMSWAP)
5735 thresholds = &memcg->memsw_thresholds;
5739 if (!thresholds->primary)
5742 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5744 /* Check if a threshold crossed before removing */
5745 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5747 /* Calculate new number of threshold */
5749 for (i = 0; i < thresholds->primary->size; i++) {
5750 if (thresholds->primary->entries[i].eventfd != eventfd)
5754 new = thresholds->spare;
5756 /* Set thresholds array to NULL if we don't have thresholds */
5765 /* Copy thresholds and find current threshold */
5766 new->current_threshold = -1;
5767 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5768 if (thresholds->primary->entries[i].eventfd == eventfd)
5771 new->entries[j] = thresholds->primary->entries[i];
5772 if (new->entries[j].threshold <= usage) {
5774 * new->current_threshold will not be used
5775 * until rcu_assign_pointer(), so it's safe to increment
5778 ++new->current_threshold;
5784 /* Swap primary and spare array */
5785 thresholds->spare = thresholds->primary;
5786 /* If all events are unregistered, free the spare array */
5788 kfree(thresholds->spare);
5789 thresholds->spare = NULL;
5792 rcu_assign_pointer(thresholds->primary, new);
5794 /* To be sure that nobody uses thresholds */
5797 mutex_unlock(&memcg->thresholds_lock);
5800 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
5801 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5803 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5804 struct mem_cgroup_eventfd_list *event;
5805 enum res_type type = MEMFILE_TYPE(cft->private);
5807 BUG_ON(type != _OOM_TYPE);
5808 event = kmalloc(sizeof(*event), GFP_KERNEL);
5812 spin_lock(&memcg_oom_lock);
5814 event->eventfd = eventfd;
5815 list_add(&event->list, &memcg->oom_notify);
5817 /* already in OOM ? */
5818 if (atomic_read(&memcg->under_oom))
5819 eventfd_signal(eventfd, 1);
5820 spin_unlock(&memcg_oom_lock);
5825 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
5826 struct cftype *cft, struct eventfd_ctx *eventfd)
5828 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5829 struct mem_cgroup_eventfd_list *ev, *tmp;
5830 enum res_type type = MEMFILE_TYPE(cft->private);
5832 BUG_ON(type != _OOM_TYPE);
5834 spin_lock(&memcg_oom_lock);
5836 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5837 if (ev->eventfd == eventfd) {
5838 list_del(&ev->list);
5843 spin_unlock(&memcg_oom_lock);
5846 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
5847 struct cftype *cft, struct cgroup_map_cb *cb)
5849 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5851 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5853 if (atomic_read(&memcg->under_oom))
5854 cb->fill(cb, "under_oom", 1);
5856 cb->fill(cb, "under_oom", 0);
5860 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
5861 struct cftype *cft, u64 val)
5863 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
5864 struct mem_cgroup *parent;
5866 /* cannot set to root cgroup and only 0 and 1 are allowed */
5867 if (!cgrp->parent || !((val == 0) || (val == 1)))
5870 parent = mem_cgroup_from_cont(cgrp->parent);
5872 mutex_lock(&memcg_create_mutex);
5873 /* oom-kill-disable is a flag for subhierarchy. */
5874 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5875 mutex_unlock(&memcg_create_mutex);
5878 memcg->oom_kill_disable = val;
5880 memcg_oom_recover(memcg);
5881 mutex_unlock(&memcg_create_mutex);
5885 #ifdef CONFIG_MEMCG_KMEM
5886 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5890 memcg->kmemcg_id = -1;
5891 ret = memcg_propagate_kmem(memcg);
5895 return mem_cgroup_sockets_init(memcg, ss);
5898 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5900 mem_cgroup_sockets_destroy(memcg);
5903 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5905 if (!memcg_kmem_is_active(memcg))
5909 * kmem charges can outlive the cgroup. In the case of slab
5910 * pages, for instance, a page contain objects from various
5911 * processes. As we prevent from taking a reference for every
5912 * such allocation we have to be careful when doing uncharge
5913 * (see memcg_uncharge_kmem) and here during offlining.
5915 * The idea is that that only the _last_ uncharge which sees
5916 * the dead memcg will drop the last reference. An additional
5917 * reference is taken here before the group is marked dead
5918 * which is then paired with css_put during uncharge resp. here.
5920 * Although this might sound strange as this path is called from
5921 * css_offline() when the referencemight have dropped down to 0
5922 * and shouldn't be incremented anymore (css_tryget would fail)
5923 * we do not have other options because of the kmem allocations
5926 css_get(&memcg->css);
5928 memcg_kmem_mark_dead(memcg);
5930 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5933 if (memcg_kmem_test_and_clear_dead(memcg))
5934 css_put(&memcg->css);
5937 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5942 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5946 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5951 static struct cftype mem_cgroup_files[] = {
5953 .name = "usage_in_bytes",
5954 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5955 .read = mem_cgroup_read,
5956 .register_event = mem_cgroup_usage_register_event,
5957 .unregister_event = mem_cgroup_usage_unregister_event,
5960 .name = "max_usage_in_bytes",
5961 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5962 .trigger = mem_cgroup_reset,
5963 .read = mem_cgroup_read,
5966 .name = "limit_in_bytes",
5967 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5968 .write_string = mem_cgroup_write,
5969 .read = mem_cgroup_read,
5972 .name = "soft_limit_in_bytes",
5973 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5974 .write_string = mem_cgroup_write,
5975 .read = mem_cgroup_read,
5979 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5980 .trigger = mem_cgroup_reset,
5981 .read = mem_cgroup_read,
5985 .read_seq_string = memcg_stat_show,
5988 .name = "force_empty",
5989 .trigger = mem_cgroup_force_empty_write,
5992 .name = "use_hierarchy",
5993 .flags = CFTYPE_INSANE,
5994 .write_u64 = mem_cgroup_hierarchy_write,
5995 .read_u64 = mem_cgroup_hierarchy_read,
5998 .name = "swappiness",
5999 .read_u64 = mem_cgroup_swappiness_read,
6000 .write_u64 = mem_cgroup_swappiness_write,
6003 .name = "move_charge_at_immigrate",
6004 .read_u64 = mem_cgroup_move_charge_read,
6005 .write_u64 = mem_cgroup_move_charge_write,
6008 .name = "oom_control",
6009 .read_map = mem_cgroup_oom_control_read,
6010 .write_u64 = mem_cgroup_oom_control_write,
6011 .register_event = mem_cgroup_oom_register_event,
6012 .unregister_event = mem_cgroup_oom_unregister_event,
6013 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6016 .name = "pressure_level",
6017 .register_event = vmpressure_register_event,
6018 .unregister_event = vmpressure_unregister_event,
6022 .name = "numa_stat",
6023 .read_seq_string = memcg_numa_stat_show,
6026 #ifdef CONFIG_MEMCG_KMEM
6028 .name = "kmem.limit_in_bytes",
6029 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6030 .write_string = mem_cgroup_write,
6031 .read = mem_cgroup_read,
6034 .name = "kmem.usage_in_bytes",
6035 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6036 .read = mem_cgroup_read,
6039 .name = "kmem.failcnt",
6040 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6041 .trigger = mem_cgroup_reset,
6042 .read = mem_cgroup_read,
6045 .name = "kmem.max_usage_in_bytes",
6046 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6047 .trigger = mem_cgroup_reset,
6048 .read = mem_cgroup_read,
6050 #ifdef CONFIG_SLABINFO
6052 .name = "kmem.slabinfo",
6053 .read_seq_string = mem_cgroup_slabinfo_read,
6057 { }, /* terminate */
6060 #ifdef CONFIG_MEMCG_SWAP
6061 static struct cftype memsw_cgroup_files[] = {
6063 .name = "memsw.usage_in_bytes",
6064 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6065 .read = mem_cgroup_read,
6066 .register_event = mem_cgroup_usage_register_event,
6067 .unregister_event = mem_cgroup_usage_unregister_event,
6070 .name = "memsw.max_usage_in_bytes",
6071 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6072 .trigger = mem_cgroup_reset,
6073 .read = mem_cgroup_read,
6076 .name = "memsw.limit_in_bytes",
6077 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6078 .write_string = mem_cgroup_write,
6079 .read = mem_cgroup_read,
6082 .name = "memsw.failcnt",
6083 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6084 .trigger = mem_cgroup_reset,
6085 .read = mem_cgroup_read,
6087 { }, /* terminate */
6090 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6092 struct mem_cgroup_per_node *pn;
6093 struct mem_cgroup_per_zone *mz;
6094 int zone, tmp = node;
6096 * This routine is called against possible nodes.
6097 * But it's BUG to call kmalloc() against offline node.
6099 * TODO: this routine can waste much memory for nodes which will
6100 * never be onlined. It's better to use memory hotplug callback
6103 if (!node_state(node, N_NORMAL_MEMORY))
6105 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6109 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6110 mz = &pn->zoneinfo[zone];
6111 lruvec_init(&mz->lruvec);
6112 mz->usage_in_excess = 0;
6113 mz->on_tree = false;
6116 memcg->nodeinfo[node] = pn;
6120 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6122 kfree(memcg->nodeinfo[node]);
6125 static struct mem_cgroup *mem_cgroup_alloc(void)
6127 struct mem_cgroup *memcg;
6128 size_t size = memcg_size();
6130 /* Can be very big if nr_node_ids is very big */
6131 if (size < PAGE_SIZE)
6132 memcg = kzalloc(size, GFP_KERNEL);
6134 memcg = vzalloc(size);
6139 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6142 spin_lock_init(&memcg->pcp_counter_lock);
6146 if (size < PAGE_SIZE)
6154 * At destroying mem_cgroup, references from swap_cgroup can remain.
6155 * (scanning all at force_empty is too costly...)
6157 * Instead of clearing all references at force_empty, we remember
6158 * the number of reference from swap_cgroup and free mem_cgroup when
6159 * it goes down to 0.
6161 * Removal of cgroup itself succeeds regardless of refs from swap.
6164 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6167 size_t size = memcg_size();
6169 mem_cgroup_remove_from_trees(memcg);
6170 free_css_id(&mem_cgroup_subsys, &memcg->css);
6173 free_mem_cgroup_per_zone_info(memcg, node);
6175 free_percpu(memcg->stat);
6178 * We need to make sure that (at least for now), the jump label
6179 * destruction code runs outside of the cgroup lock. This is because
6180 * get_online_cpus(), which is called from the static_branch update,
6181 * can't be called inside the cgroup_lock. cpusets are the ones
6182 * enforcing this dependency, so if they ever change, we might as well.
6184 * schedule_work() will guarantee this happens. Be careful if you need
6185 * to move this code around, and make sure it is outside
6188 disarm_static_keys(memcg);
6189 if (size < PAGE_SIZE)
6196 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6198 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6200 if (!memcg->res.parent)
6202 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6204 EXPORT_SYMBOL(parent_mem_cgroup);
6206 static void __init mem_cgroup_soft_limit_tree_init(void)
6208 struct mem_cgroup_tree_per_node *rtpn;
6209 struct mem_cgroup_tree_per_zone *rtpz;
6210 int tmp, node, zone;
6212 for_each_node(node) {
6214 if (!node_state(node, N_NORMAL_MEMORY))
6216 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6219 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6221 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6222 rtpz = &rtpn->rb_tree_per_zone[zone];
6223 rtpz->rb_root = RB_ROOT;
6224 spin_lock_init(&rtpz->lock);
6229 static struct cgroup_subsys_state * __ref
6230 mem_cgroup_css_alloc(struct cgroup *cont)
6232 struct mem_cgroup *memcg;
6233 long error = -ENOMEM;
6236 memcg = mem_cgroup_alloc();
6238 return ERR_PTR(error);
6241 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6245 if (cont->parent == NULL) {
6246 root_mem_cgroup = memcg;
6247 res_counter_init(&memcg->res, NULL);
6248 res_counter_init(&memcg->memsw, NULL);
6249 res_counter_init(&memcg->kmem, NULL);
6252 memcg->last_scanned_node = MAX_NUMNODES;
6253 INIT_LIST_HEAD(&memcg->oom_notify);
6254 memcg->move_charge_at_immigrate = 0;
6255 mutex_init(&memcg->thresholds_lock);
6256 spin_lock_init(&memcg->move_lock);
6257 vmpressure_init(&memcg->vmpressure);
6262 __mem_cgroup_free(memcg);
6263 return ERR_PTR(error);
6267 mem_cgroup_css_online(struct cgroup *cont)
6269 struct mem_cgroup *memcg, *parent;
6275 mutex_lock(&memcg_create_mutex);
6276 memcg = mem_cgroup_from_cont(cont);
6277 parent = mem_cgroup_from_cont(cont->parent);
6279 memcg->use_hierarchy = parent->use_hierarchy;
6280 memcg->oom_kill_disable = parent->oom_kill_disable;
6281 memcg->swappiness = mem_cgroup_swappiness(parent);
6283 if (parent->use_hierarchy) {
6284 res_counter_init(&memcg->res, &parent->res);
6285 res_counter_init(&memcg->memsw, &parent->memsw);
6286 res_counter_init(&memcg->kmem, &parent->kmem);
6289 * No need to take a reference to the parent because cgroup
6290 * core guarantees its existence.
6293 res_counter_init(&memcg->res, NULL);
6294 res_counter_init(&memcg->memsw, NULL);
6295 res_counter_init(&memcg->kmem, NULL);
6297 * Deeper hierachy with use_hierarchy == false doesn't make
6298 * much sense so let cgroup subsystem know about this
6299 * unfortunate state in our controller.
6301 if (parent != root_mem_cgroup)
6302 mem_cgroup_subsys.broken_hierarchy = true;
6305 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6306 mutex_unlock(&memcg_create_mutex);
6311 * Announce all parents that a group from their hierarchy is gone.
6313 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6315 struct mem_cgroup *parent = memcg;
6317 while ((parent = parent_mem_cgroup(parent)))
6318 mem_cgroup_iter_invalidate(parent);
6321 * if the root memcg is not hierarchical we have to check it
6324 if (!root_mem_cgroup->use_hierarchy)
6325 mem_cgroup_iter_invalidate(root_mem_cgroup);
6328 static void mem_cgroup_css_offline(struct cgroup *cont)
6330 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6332 kmem_cgroup_css_offline(memcg);
6334 mem_cgroup_invalidate_reclaim_iterators(memcg);
6335 mem_cgroup_reparent_charges(memcg);
6336 mem_cgroup_destroy_all_caches(memcg);
6339 static void mem_cgroup_css_free(struct cgroup *cont)
6341 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
6343 memcg_destroy_kmem(memcg);
6344 __mem_cgroup_free(memcg);
6348 /* Handlers for move charge at task migration. */
6349 #define PRECHARGE_COUNT_AT_ONCE 256
6350 static int mem_cgroup_do_precharge(unsigned long count)
6353 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6354 struct mem_cgroup *memcg = mc.to;
6356 if (mem_cgroup_is_root(memcg)) {
6357 mc.precharge += count;
6358 /* we don't need css_get for root */
6361 /* try to charge at once */
6363 struct res_counter *dummy;
6365 * "memcg" cannot be under rmdir() because we've already checked
6366 * by cgroup_lock_live_cgroup() that it is not removed and we
6367 * are still under the same cgroup_mutex. So we can postpone
6370 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6372 if (do_swap_account && res_counter_charge(&memcg->memsw,
6373 PAGE_SIZE * count, &dummy)) {
6374 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6377 mc.precharge += count;
6381 /* fall back to one by one charge */
6383 if (signal_pending(current)) {
6387 if (!batch_count--) {
6388 batch_count = PRECHARGE_COUNT_AT_ONCE;
6391 ret = __mem_cgroup_try_charge(NULL,
6392 GFP_KERNEL, 1, &memcg, false);
6394 /* mem_cgroup_clear_mc() will do uncharge later */
6402 * get_mctgt_type - get target type of moving charge
6403 * @vma: the vma the pte to be checked belongs
6404 * @addr: the address corresponding to the pte to be checked
6405 * @ptent: the pte to be checked
6406 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6409 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6410 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6411 * move charge. if @target is not NULL, the page is stored in target->page
6412 * with extra refcnt got(Callers should handle it).
6413 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6414 * target for charge migration. if @target is not NULL, the entry is stored
6417 * Called with pte lock held.
6424 enum mc_target_type {
6430 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6431 unsigned long addr, pte_t ptent)
6433 struct page *page = vm_normal_page(vma, addr, ptent);
6435 if (!page || !page_mapped(page))
6437 if (PageAnon(page)) {
6438 /* we don't move shared anon */
6441 } else if (!move_file())
6442 /* we ignore mapcount for file pages */
6444 if (!get_page_unless_zero(page))
6451 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6452 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6454 struct page *page = NULL;
6455 swp_entry_t ent = pte_to_swp_entry(ptent);
6457 if (!move_anon() || non_swap_entry(ent))
6460 * Because lookup_swap_cache() updates some statistics counter,
6461 * we call find_get_page() with swapper_space directly.
6463 page = find_get_page(swap_address_space(ent), ent.val);
6464 if (do_swap_account)
6465 entry->val = ent.val;
6470 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6471 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6477 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6478 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6480 struct page *page = NULL;
6481 struct address_space *mapping;
6484 if (!vma->vm_file) /* anonymous vma */
6489 mapping = vma->vm_file->f_mapping;
6490 if (pte_none(ptent))
6491 pgoff = linear_page_index(vma, addr);
6492 else /* pte_file(ptent) is true */
6493 pgoff = pte_to_pgoff(ptent);
6495 /* page is moved even if it's not RSS of this task(page-faulted). */
6496 page = find_get_page(mapping, pgoff);
6499 /* shmem/tmpfs may report page out on swap: account for that too. */
6500 if (radix_tree_exceptional_entry(page)) {
6501 swp_entry_t swap = radix_to_swp_entry(page);
6502 if (do_swap_account)
6504 page = find_get_page(swap_address_space(swap), swap.val);
6510 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6511 unsigned long addr, pte_t ptent, union mc_target *target)
6513 struct page *page = NULL;
6514 struct page_cgroup *pc;
6515 enum mc_target_type ret = MC_TARGET_NONE;
6516 swp_entry_t ent = { .val = 0 };
6518 if (pte_present(ptent))
6519 page = mc_handle_present_pte(vma, addr, ptent);
6520 else if (is_swap_pte(ptent))
6521 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6522 else if (pte_none(ptent) || pte_file(ptent))
6523 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6525 if (!page && !ent.val)
6528 pc = lookup_page_cgroup(page);
6530 * Do only loose check w/o page_cgroup lock.
6531 * mem_cgroup_move_account() checks the pc is valid or not under
6534 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6535 ret = MC_TARGET_PAGE;
6537 target->page = page;
6539 if (!ret || !target)
6542 /* There is a swap entry and a page doesn't exist or isn't charged */
6543 if (ent.val && !ret &&
6544 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6545 ret = MC_TARGET_SWAP;
6552 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6554 * We don't consider swapping or file mapped pages because THP does not
6555 * support them for now.
6556 * Caller should make sure that pmd_trans_huge(pmd) is true.
6558 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6559 unsigned long addr, pmd_t pmd, union mc_target *target)
6561 struct page *page = NULL;
6562 struct page_cgroup *pc;
6563 enum mc_target_type ret = MC_TARGET_NONE;
6565 page = pmd_page(pmd);
6566 VM_BUG_ON(!page || !PageHead(page));
6569 pc = lookup_page_cgroup(page);
6570 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6571 ret = MC_TARGET_PAGE;
6574 target->page = page;
6580 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6581 unsigned long addr, pmd_t pmd, union mc_target *target)
6583 return MC_TARGET_NONE;
6587 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6588 unsigned long addr, unsigned long end,
6589 struct mm_walk *walk)
6591 struct vm_area_struct *vma = walk->private;
6595 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6596 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6597 mc.precharge += HPAGE_PMD_NR;
6598 spin_unlock(&vma->vm_mm->page_table_lock);
6602 if (pmd_trans_unstable(pmd))
6604 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6605 for (; addr != end; pte++, addr += PAGE_SIZE)
6606 if (get_mctgt_type(vma, addr, *pte, NULL))
6607 mc.precharge++; /* increment precharge temporarily */
6608 pte_unmap_unlock(pte - 1, ptl);
6614 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6616 unsigned long precharge;
6617 struct vm_area_struct *vma;
6619 down_read(&mm->mmap_sem);
6620 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6621 struct mm_walk mem_cgroup_count_precharge_walk = {
6622 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6626 if (is_vm_hugetlb_page(vma))
6628 walk_page_range(vma->vm_start, vma->vm_end,
6629 &mem_cgroup_count_precharge_walk);
6631 up_read(&mm->mmap_sem);
6633 precharge = mc.precharge;
6639 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6641 unsigned long precharge = mem_cgroup_count_precharge(mm);
6643 VM_BUG_ON(mc.moving_task);
6644 mc.moving_task = current;
6645 return mem_cgroup_do_precharge(precharge);
6648 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6649 static void __mem_cgroup_clear_mc(void)
6651 struct mem_cgroup *from = mc.from;
6652 struct mem_cgroup *to = mc.to;
6655 /* we must uncharge all the leftover precharges from mc.to */
6657 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6661 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6662 * we must uncharge here.
6664 if (mc.moved_charge) {
6665 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6666 mc.moved_charge = 0;
6668 /* we must fixup refcnts and charges */
6669 if (mc.moved_swap) {
6670 /* uncharge swap account from the old cgroup */
6671 if (!mem_cgroup_is_root(mc.from))
6672 res_counter_uncharge(&mc.from->memsw,
6673 PAGE_SIZE * mc.moved_swap);
6675 for (i = 0; i < mc.moved_swap; i++)
6676 css_put(&mc.from->css);
6678 if (!mem_cgroup_is_root(mc.to)) {
6680 * we charged both to->res and to->memsw, so we should
6683 res_counter_uncharge(&mc.to->res,
6684 PAGE_SIZE * mc.moved_swap);
6686 /* we've already done css_get(mc.to) */
6689 memcg_oom_recover(from);
6690 memcg_oom_recover(to);
6691 wake_up_all(&mc.waitq);
6694 static void mem_cgroup_clear_mc(void)
6696 struct mem_cgroup *from = mc.from;
6699 * we must clear moving_task before waking up waiters at the end of
6702 mc.moving_task = NULL;
6703 __mem_cgroup_clear_mc();
6704 spin_lock(&mc.lock);
6707 spin_unlock(&mc.lock);
6708 mem_cgroup_end_move(from);
6711 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6712 struct cgroup_taskset *tset)
6714 struct task_struct *p = cgroup_taskset_first(tset);
6716 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
6717 unsigned long move_charge_at_immigrate;
6720 * We are now commited to this value whatever it is. Changes in this
6721 * tunable will only affect upcoming migrations, not the current one.
6722 * So we need to save it, and keep it going.
6724 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6725 if (move_charge_at_immigrate) {
6726 struct mm_struct *mm;
6727 struct mem_cgroup *from = mem_cgroup_from_task(p);
6729 VM_BUG_ON(from == memcg);
6731 mm = get_task_mm(p);
6734 /* We move charges only when we move a owner of the mm */
6735 if (mm->owner == p) {
6738 VM_BUG_ON(mc.precharge);
6739 VM_BUG_ON(mc.moved_charge);
6740 VM_BUG_ON(mc.moved_swap);
6741 mem_cgroup_start_move(from);
6742 spin_lock(&mc.lock);
6745 mc.immigrate_flags = move_charge_at_immigrate;
6746 spin_unlock(&mc.lock);
6747 /* We set mc.moving_task later */
6749 ret = mem_cgroup_precharge_mc(mm);
6751 mem_cgroup_clear_mc();
6758 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6759 struct cgroup_taskset *tset)
6761 mem_cgroup_clear_mc();
6764 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6765 unsigned long addr, unsigned long end,
6766 struct mm_walk *walk)
6769 struct vm_area_struct *vma = walk->private;
6772 enum mc_target_type target_type;
6773 union mc_target target;
6775 struct page_cgroup *pc;
6778 * We don't take compound_lock() here but no race with splitting thp
6780 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6781 * under splitting, which means there's no concurrent thp split,
6782 * - if another thread runs into split_huge_page() just after we
6783 * entered this if-block, the thread must wait for page table lock
6784 * to be unlocked in __split_huge_page_splitting(), where the main
6785 * part of thp split is not executed yet.
6787 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6788 if (mc.precharge < HPAGE_PMD_NR) {
6789 spin_unlock(&vma->vm_mm->page_table_lock);
6792 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6793 if (target_type == MC_TARGET_PAGE) {
6795 if (!isolate_lru_page(page)) {
6796 pc = lookup_page_cgroup(page);
6797 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6798 pc, mc.from, mc.to)) {
6799 mc.precharge -= HPAGE_PMD_NR;
6800 mc.moved_charge += HPAGE_PMD_NR;
6802 putback_lru_page(page);
6806 spin_unlock(&vma->vm_mm->page_table_lock);
6810 if (pmd_trans_unstable(pmd))
6813 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6814 for (; addr != end; addr += PAGE_SIZE) {
6815 pte_t ptent = *(pte++);
6821 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6822 case MC_TARGET_PAGE:
6824 if (isolate_lru_page(page))
6826 pc = lookup_page_cgroup(page);
6827 if (!mem_cgroup_move_account(page, 1, pc,
6830 /* we uncharge from mc.from later. */
6833 putback_lru_page(page);
6834 put: /* get_mctgt_type() gets the page */
6837 case MC_TARGET_SWAP:
6839 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6841 /* we fixup refcnts and charges later. */
6849 pte_unmap_unlock(pte - 1, ptl);
6854 * We have consumed all precharges we got in can_attach().
6855 * We try charge one by one, but don't do any additional
6856 * charges to mc.to if we have failed in charge once in attach()
6859 ret = mem_cgroup_do_precharge(1);
6867 static void mem_cgroup_move_charge(struct mm_struct *mm)
6869 struct vm_area_struct *vma;
6871 lru_add_drain_all();
6873 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6875 * Someone who are holding the mmap_sem might be waiting in
6876 * waitq. So we cancel all extra charges, wake up all waiters,
6877 * and retry. Because we cancel precharges, we might not be able
6878 * to move enough charges, but moving charge is a best-effort
6879 * feature anyway, so it wouldn't be a big problem.
6881 __mem_cgroup_clear_mc();
6885 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6887 struct mm_walk mem_cgroup_move_charge_walk = {
6888 .pmd_entry = mem_cgroup_move_charge_pte_range,
6892 if (is_vm_hugetlb_page(vma))
6894 ret = walk_page_range(vma->vm_start, vma->vm_end,
6895 &mem_cgroup_move_charge_walk);
6898 * means we have consumed all precharges and failed in
6899 * doing additional charge. Just abandon here.
6903 up_read(&mm->mmap_sem);
6906 static void mem_cgroup_move_task(struct cgroup *cont,
6907 struct cgroup_taskset *tset)
6909 struct task_struct *p = cgroup_taskset_first(tset);
6910 struct mm_struct *mm = get_task_mm(p);
6914 mem_cgroup_move_charge(mm);
6918 mem_cgroup_clear_mc();
6920 #else /* !CONFIG_MMU */
6921 static int mem_cgroup_can_attach(struct cgroup *cgroup,
6922 struct cgroup_taskset *tset)
6926 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
6927 struct cgroup_taskset *tset)
6930 static void mem_cgroup_move_task(struct cgroup *cont,
6931 struct cgroup_taskset *tset)
6937 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6938 * to verify sane_behavior flag on each mount attempt.
6940 static void mem_cgroup_bind(struct cgroup *root)
6943 * use_hierarchy is forced with sane_behavior. cgroup core
6944 * guarantees that @root doesn't have any children, so turning it
6945 * on for the root memcg is enough.
6947 if (cgroup_sane_behavior(root))
6948 mem_cgroup_from_cont(root)->use_hierarchy = true;
6951 struct cgroup_subsys mem_cgroup_subsys = {
6953 .subsys_id = mem_cgroup_subsys_id,
6954 .css_alloc = mem_cgroup_css_alloc,
6955 .css_online = mem_cgroup_css_online,
6956 .css_offline = mem_cgroup_css_offline,
6957 .css_free = mem_cgroup_css_free,
6958 .can_attach = mem_cgroup_can_attach,
6959 .cancel_attach = mem_cgroup_cancel_attach,
6960 .attach = mem_cgroup_move_task,
6961 .bind = mem_cgroup_bind,
6962 .base_cftypes = mem_cgroup_files,
6967 #ifdef CONFIG_MEMCG_SWAP
6968 static int __init enable_swap_account(char *s)
6970 /* consider enabled if no parameter or 1 is given */
6971 if (!strcmp(s, "1"))
6972 really_do_swap_account = 1;
6973 else if (!strcmp(s, "0"))
6974 really_do_swap_account = 0;
6977 __setup("swapaccount=", enable_swap_account);
6979 static void __init memsw_file_init(void)
6981 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6984 static void __init enable_swap_cgroup(void)
6986 if (!mem_cgroup_disabled() && really_do_swap_account) {
6987 do_swap_account = 1;
6993 static void __init enable_swap_cgroup(void)
6999 * subsys_initcall() for memory controller.
7001 * Some parts like hotcpu_notifier() have to be initialized from this context
7002 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7003 * everything that doesn't depend on a specific mem_cgroup structure should
7004 * be initialized from here.
7006 static int __init mem_cgroup_init(void)
7008 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7009 enable_swap_cgroup();
7010 mem_cgroup_soft_limit_tree_init();
7014 subsys_initcall(mem_cgroup_init);