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
88 static const char * const mem_cgroup_stat_names[] = {
97 enum mem_cgroup_events_index {
98 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
99 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
100 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
101 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
102 MEM_CGROUP_EVENTS_NSTATS,
105 static const char * const mem_cgroup_events_names[] = {
112 static const char * const mem_cgroup_lru_names[] = {
121 * Per memcg event counter is incremented at every pagein/pageout. With THP,
122 * it will be incremated by the number of pages. This counter is used for
123 * for trigger some periodic events. This is straightforward and better
124 * than using jiffies etc. to handle periodic memcg event.
126 enum mem_cgroup_events_target {
127 MEM_CGROUP_TARGET_THRESH,
128 MEM_CGROUP_TARGET_SOFTLIMIT,
129 MEM_CGROUP_TARGET_NUMAINFO,
132 #define THRESHOLDS_EVENTS_TARGET 128
133 #define SOFTLIMIT_EVENTS_TARGET 1024
134 #define NUMAINFO_EVENTS_TARGET 1024
136 struct mem_cgroup_stat_cpu {
137 long count[MEM_CGROUP_STAT_NSTATS];
138 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
139 unsigned long nr_page_events;
140 unsigned long targets[MEM_CGROUP_NTARGETS];
143 struct mem_cgroup_reclaim_iter {
145 * last scanned hierarchy member. Valid only if last_dead_count
146 * matches memcg->dead_count of the hierarchy root group.
148 struct mem_cgroup *last_visited;
149 unsigned long last_dead_count;
151 /* scan generation, increased every round-trip */
152 unsigned int generation;
156 * per-zone information in memory controller.
158 struct mem_cgroup_per_zone {
159 struct lruvec lruvec;
160 unsigned long lru_size[NR_LRU_LISTS];
162 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
164 struct rb_node tree_node; /* RB tree node */
165 unsigned long long usage_in_excess;/* Set to the value by which */
166 /* the soft limit is exceeded*/
168 struct mem_cgroup *memcg; /* Back pointer, we cannot */
169 /* use container_of */
172 struct mem_cgroup_per_node {
173 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
177 * Cgroups above their limits are maintained in a RB-Tree, independent of
178 * their hierarchy representation
181 struct mem_cgroup_tree_per_zone {
182 struct rb_root rb_root;
186 struct mem_cgroup_tree_per_node {
187 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
190 struct mem_cgroup_tree {
191 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
194 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
196 struct mem_cgroup_threshold {
197 struct eventfd_ctx *eventfd;
202 struct mem_cgroup_threshold_ary {
203 /* An array index points to threshold just below or equal to usage. */
204 int current_threshold;
205 /* Size of entries[] */
207 /* Array of thresholds */
208 struct mem_cgroup_threshold entries[0];
211 struct mem_cgroup_thresholds {
212 /* Primary thresholds array */
213 struct mem_cgroup_threshold_ary *primary;
215 * Spare threshold array.
216 * This is needed to make mem_cgroup_unregister_event() "never fail".
217 * It must be able to store at least primary->size - 1 entries.
219 struct mem_cgroup_threshold_ary *spare;
223 struct mem_cgroup_eventfd_list {
224 struct list_head list;
225 struct eventfd_ctx *eventfd;
228 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
229 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
232 * The memory controller data structure. The memory controller controls both
233 * page cache and RSS per cgroup. We would eventually like to provide
234 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
235 * to help the administrator determine what knobs to tune.
237 * TODO: Add a water mark for the memory controller. Reclaim will begin when
238 * we hit the water mark. May be even add a low water mark, such that
239 * no reclaim occurs from a cgroup at it's low water mark, this is
240 * a feature that will be implemented much later in the future.
243 struct cgroup_subsys_state css;
245 * the counter to account for memory usage
247 struct res_counter res;
249 /* vmpressure notifications */
250 struct vmpressure vmpressure;
253 * the counter to account for mem+swap usage.
255 struct res_counter memsw;
258 * the counter to account for kernel memory usage.
260 struct res_counter kmem;
262 * Should the accounting and control be hierarchical, per subtree?
265 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
269 atomic_t oom_wakeups;
272 /* OOM-Killer disable */
273 int oom_kill_disable;
275 /* set when res.limit == memsw.limit */
276 bool memsw_is_minimum;
278 /* protect arrays of thresholds */
279 struct mutex thresholds_lock;
281 /* thresholds for memory usage. RCU-protected */
282 struct mem_cgroup_thresholds thresholds;
284 /* thresholds for mem+swap usage. RCU-protected */
285 struct mem_cgroup_thresholds memsw_thresholds;
287 /* For oom notifier event fd */
288 struct list_head oom_notify;
291 * Should we move charges of a task when a task is moved into this
292 * mem_cgroup ? And what type of charges should we move ?
294 unsigned long move_charge_at_immigrate;
296 * set > 0 if pages under this cgroup are moving to other cgroup.
298 atomic_t moving_account;
299 /* taken only while moving_account > 0 */
300 spinlock_t move_lock;
304 struct mem_cgroup_stat_cpu __percpu *stat;
306 * used when a cpu is offlined or other synchronizations
307 * See mem_cgroup_read_stat().
309 struct mem_cgroup_stat_cpu nocpu_base;
310 spinlock_t pcp_counter_lock;
313 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
314 struct cg_proto tcp_mem;
316 #if defined(CONFIG_MEMCG_KMEM)
317 /* analogous to slab_common's slab_caches list. per-memcg */
318 struct list_head memcg_slab_caches;
319 /* Not a spinlock, we can take a lot of time walking the list */
320 struct mutex slab_caches_mutex;
321 /* Index in the kmem_cache->memcg_params->memcg_caches array */
325 int last_scanned_node;
327 nodemask_t scan_nodes;
328 atomic_t numainfo_events;
329 atomic_t numainfo_updating;
332 struct mem_cgroup_per_node *nodeinfo[0];
333 /* WARNING: nodeinfo must be the last member here */
336 static size_t memcg_size(void)
338 return sizeof(struct mem_cgroup) +
339 nr_node_ids * sizeof(struct mem_cgroup_per_node);
342 /* internal only representation about the status of kmem accounting. */
344 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
345 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
346 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
349 /* We account when limit is on, but only after call sites are patched */
350 #define KMEM_ACCOUNTED_MASK \
351 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
353 #ifdef CONFIG_MEMCG_KMEM
354 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
356 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
359 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
361 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
364 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
366 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
369 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
371 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
374 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
377 * Our caller must use css_get() first, because memcg_uncharge_kmem()
378 * will call css_put() if it sees the memcg is dead.
381 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
382 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
385 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
387 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
388 &memcg->kmem_account_flags);
392 /* Stuffs for move charges at task migration. */
394 * Types of charges to be moved. "move_charge_at_immitgrate" and
395 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
398 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
399 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
403 /* "mc" and its members are protected by cgroup_mutex */
404 static struct move_charge_struct {
405 spinlock_t lock; /* for from, to */
406 struct mem_cgroup *from;
407 struct mem_cgroup *to;
408 unsigned long immigrate_flags;
409 unsigned long precharge;
410 unsigned long moved_charge;
411 unsigned long moved_swap;
412 struct task_struct *moving_task; /* a task moving charges */
413 wait_queue_head_t waitq; /* a waitq for other context */
415 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
416 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
419 static bool move_anon(void)
421 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
424 static bool move_file(void)
426 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
430 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
431 * limit reclaim to prevent infinite loops, if they ever occur.
433 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
434 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
437 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
438 MEM_CGROUP_CHARGE_TYPE_ANON,
439 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
440 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
444 /* for encoding cft->private value on file */
452 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
453 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
454 #define MEMFILE_ATTR(val) ((val) & 0xffff)
455 /* Used for OOM nofiier */
456 #define OOM_CONTROL (0)
459 * Reclaim flags for mem_cgroup_hierarchical_reclaim
461 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
462 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
463 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
464 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
467 * The memcg_create_mutex will be held whenever a new cgroup is created.
468 * As a consequence, any change that needs to protect against new child cgroups
469 * appearing has to hold it as well.
471 static DEFINE_MUTEX(memcg_create_mutex);
473 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
475 return s ? container_of(s, struct mem_cgroup, css) : NULL;
478 /* Some nice accessors for the vmpressure. */
479 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
482 memcg = root_mem_cgroup;
483 return &memcg->vmpressure;
486 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
488 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
491 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
493 return &mem_cgroup_from_css(css)->vmpressure;
496 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
498 return (memcg == root_mem_cgroup);
502 * We restrict the id in the range of [1, 65535], so it can fit into
505 #define MEM_CGROUP_ID_MAX USHRT_MAX
507 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
510 * The ID of the root cgroup is 0, but memcg treat 0 as an
511 * invalid ID, so we return (cgroup_id + 1).
513 return memcg->css.cgroup->id + 1;
516 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
518 struct cgroup_subsys_state *css;
520 css = css_from_id(id - 1, &mem_cgroup_subsys);
521 return mem_cgroup_from_css(css);
524 /* Writing them here to avoid exposing memcg's inner layout */
525 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
527 void sock_update_memcg(struct sock *sk)
529 if (mem_cgroup_sockets_enabled) {
530 struct mem_cgroup *memcg;
531 struct cg_proto *cg_proto;
533 BUG_ON(!sk->sk_prot->proto_cgroup);
535 /* Socket cloning can throw us here with sk_cgrp already
536 * filled. It won't however, necessarily happen from
537 * process context. So the test for root memcg given
538 * the current task's memcg won't help us in this case.
540 * Respecting the original socket's memcg is a better
541 * decision in this case.
544 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
545 css_get(&sk->sk_cgrp->memcg->css);
550 memcg = mem_cgroup_from_task(current);
551 cg_proto = sk->sk_prot->proto_cgroup(memcg);
552 if (!mem_cgroup_is_root(memcg) &&
553 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
554 sk->sk_cgrp = cg_proto;
559 EXPORT_SYMBOL(sock_update_memcg);
561 void sock_release_memcg(struct sock *sk)
563 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
564 struct mem_cgroup *memcg;
565 WARN_ON(!sk->sk_cgrp->memcg);
566 memcg = sk->sk_cgrp->memcg;
567 css_put(&sk->sk_cgrp->memcg->css);
571 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
573 if (!memcg || mem_cgroup_is_root(memcg))
576 return &memcg->tcp_mem;
578 EXPORT_SYMBOL(tcp_proto_cgroup);
580 static void disarm_sock_keys(struct mem_cgroup *memcg)
582 if (!memcg_proto_activated(&memcg->tcp_mem))
584 static_key_slow_dec(&memcg_socket_limit_enabled);
587 static void disarm_sock_keys(struct mem_cgroup *memcg)
592 #ifdef CONFIG_MEMCG_KMEM
594 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
595 * The main reason for not using cgroup id for this:
596 * this works better in sparse environments, where we have a lot of memcgs,
597 * but only a few kmem-limited. Or also, if we have, for instance, 200
598 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
599 * 200 entry array for that.
601 * The current size of the caches array is stored in
602 * memcg_limited_groups_array_size. It will double each time we have to
605 static DEFINE_IDA(kmem_limited_groups);
606 int memcg_limited_groups_array_size;
609 * MIN_SIZE is different than 1, because we would like to avoid going through
610 * the alloc/free process all the time. In a small machine, 4 kmem-limited
611 * cgroups is a reasonable guess. In the future, it could be a parameter or
612 * tunable, but that is strictly not necessary.
614 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
615 * this constant directly from cgroup, but it is understandable that this is
616 * better kept as an internal representation in cgroup.c. In any case, the
617 * cgrp_id space is not getting any smaller, and we don't have to necessarily
618 * increase ours as well if it increases.
620 #define MEMCG_CACHES_MIN_SIZE 4
621 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
624 * A lot of the calls to the cache allocation functions are expected to be
625 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
626 * conditional to this static branch, we'll have to allow modules that does
627 * kmem_cache_alloc and the such to see this symbol as well
629 struct static_key memcg_kmem_enabled_key;
630 EXPORT_SYMBOL(memcg_kmem_enabled_key);
632 static void disarm_kmem_keys(struct mem_cgroup *memcg)
634 if (memcg_kmem_is_active(memcg)) {
635 static_key_slow_dec(&memcg_kmem_enabled_key);
636 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
639 * This check can't live in kmem destruction function,
640 * since the charges will outlive the cgroup
642 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
645 static void disarm_kmem_keys(struct mem_cgroup *memcg)
648 #endif /* CONFIG_MEMCG_KMEM */
650 static void disarm_static_keys(struct mem_cgroup *memcg)
652 disarm_sock_keys(memcg);
653 disarm_kmem_keys(memcg);
656 static void drain_all_stock_async(struct mem_cgroup *memcg);
658 static struct mem_cgroup_per_zone *
659 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
661 VM_BUG_ON((unsigned)nid >= nr_node_ids);
662 return &memcg->nodeinfo[nid]->zoneinfo[zid];
665 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
670 static struct mem_cgroup_per_zone *
671 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
673 int nid = page_to_nid(page);
674 int zid = page_zonenum(page);
676 return mem_cgroup_zoneinfo(memcg, nid, zid);
679 static struct mem_cgroup_tree_per_zone *
680 soft_limit_tree_node_zone(int nid, int zid)
682 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
685 static struct mem_cgroup_tree_per_zone *
686 soft_limit_tree_from_page(struct page *page)
688 int nid = page_to_nid(page);
689 int zid = page_zonenum(page);
691 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
695 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
696 struct mem_cgroup_per_zone *mz,
697 struct mem_cgroup_tree_per_zone *mctz,
698 unsigned long long new_usage_in_excess)
700 struct rb_node **p = &mctz->rb_root.rb_node;
701 struct rb_node *parent = NULL;
702 struct mem_cgroup_per_zone *mz_node;
707 mz->usage_in_excess = new_usage_in_excess;
708 if (!mz->usage_in_excess)
712 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
714 if (mz->usage_in_excess < mz_node->usage_in_excess)
717 * We can't avoid mem cgroups that are over their soft
718 * limit by the same amount
720 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
723 rb_link_node(&mz->tree_node, parent, p);
724 rb_insert_color(&mz->tree_node, &mctz->rb_root);
729 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
730 struct mem_cgroup_per_zone *mz,
731 struct mem_cgroup_tree_per_zone *mctz)
735 rb_erase(&mz->tree_node, &mctz->rb_root);
740 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
741 struct mem_cgroup_per_zone *mz,
742 struct mem_cgroup_tree_per_zone *mctz)
744 spin_lock(&mctz->lock);
745 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
746 spin_unlock(&mctz->lock);
750 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
752 unsigned long long excess;
753 struct mem_cgroup_per_zone *mz;
754 struct mem_cgroup_tree_per_zone *mctz;
755 int nid = page_to_nid(page);
756 int zid = page_zonenum(page);
757 mctz = soft_limit_tree_from_page(page);
760 * Necessary to update all ancestors when hierarchy is used.
761 * because their event counter is not touched.
763 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
764 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
765 excess = res_counter_soft_limit_excess(&memcg->res);
767 * We have to update the tree if mz is on RB-tree or
768 * mem is over its softlimit.
770 if (excess || mz->on_tree) {
771 spin_lock(&mctz->lock);
772 /* if on-tree, remove it */
774 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
776 * Insert again. mz->usage_in_excess will be updated.
777 * If excess is 0, no tree ops.
779 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
780 spin_unlock(&mctz->lock);
785 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
788 struct mem_cgroup_per_zone *mz;
789 struct mem_cgroup_tree_per_zone *mctz;
791 for_each_node(node) {
792 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
793 mz = mem_cgroup_zoneinfo(memcg, node, zone);
794 mctz = soft_limit_tree_node_zone(node, zone);
795 mem_cgroup_remove_exceeded(memcg, mz, mctz);
800 static struct mem_cgroup_per_zone *
801 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
803 struct rb_node *rightmost = NULL;
804 struct mem_cgroup_per_zone *mz;
808 rightmost = rb_last(&mctz->rb_root);
810 goto done; /* Nothing to reclaim from */
812 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
814 * Remove the node now but someone else can add it back,
815 * we will to add it back at the end of reclaim to its correct
816 * position in the tree.
818 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
819 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
820 !css_tryget(&mz->memcg->css))
826 static struct mem_cgroup_per_zone *
827 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
829 struct mem_cgroup_per_zone *mz;
831 spin_lock(&mctz->lock);
832 mz = __mem_cgroup_largest_soft_limit_node(mctz);
833 spin_unlock(&mctz->lock);
838 * Implementation Note: reading percpu statistics for memcg.
840 * Both of vmstat[] and percpu_counter has threshold and do periodic
841 * synchronization to implement "quick" read. There are trade-off between
842 * reading cost and precision of value. Then, we may have a chance to implement
843 * a periodic synchronizion of counter in memcg's counter.
845 * But this _read() function is used for user interface now. The user accounts
846 * memory usage by memory cgroup and he _always_ requires exact value because
847 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
848 * have to visit all online cpus and make sum. So, for now, unnecessary
849 * synchronization is not implemented. (just implemented for cpu hotplug)
851 * If there are kernel internal actions which can make use of some not-exact
852 * value, and reading all cpu value can be performance bottleneck in some
853 * common workload, threashold and synchonization as vmstat[] should be
856 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
857 enum mem_cgroup_stat_index idx)
863 for_each_online_cpu(cpu)
864 val += per_cpu(memcg->stat->count[idx], cpu);
865 #ifdef CONFIG_HOTPLUG_CPU
866 spin_lock(&memcg->pcp_counter_lock);
867 val += memcg->nocpu_base.count[idx];
868 spin_unlock(&memcg->pcp_counter_lock);
874 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
877 int val = (charge) ? 1 : -1;
878 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
881 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
882 enum mem_cgroup_events_index idx)
884 unsigned long val = 0;
888 for_each_online_cpu(cpu)
889 val += per_cpu(memcg->stat->events[idx], cpu);
890 #ifdef CONFIG_HOTPLUG_CPU
891 spin_lock(&memcg->pcp_counter_lock);
892 val += memcg->nocpu_base.events[idx];
893 spin_unlock(&memcg->pcp_counter_lock);
899 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
901 bool anon, int nr_pages)
906 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
907 * counted as CACHE even if it's on ANON LRU.
910 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
913 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
916 if (PageTransHuge(page))
917 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
920 /* pagein of a big page is an event. So, ignore page size */
922 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
924 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
925 nr_pages = -nr_pages; /* for event */
928 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
934 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
936 struct mem_cgroup_per_zone *mz;
938 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
939 return mz->lru_size[lru];
943 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
944 unsigned int lru_mask)
946 struct mem_cgroup_per_zone *mz;
948 unsigned long ret = 0;
950 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
953 if (BIT(lru) & lru_mask)
954 ret += mz->lru_size[lru];
960 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
961 int nid, unsigned int lru_mask)
966 for (zid = 0; zid < MAX_NR_ZONES; zid++)
967 total += mem_cgroup_zone_nr_lru_pages(memcg,
973 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
974 unsigned int lru_mask)
979 for_each_node_state(nid, N_MEMORY)
980 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
984 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
985 enum mem_cgroup_events_target target)
987 unsigned long val, next;
989 val = __this_cpu_read(memcg->stat->nr_page_events);
990 next = __this_cpu_read(memcg->stat->targets[target]);
991 /* from time_after() in jiffies.h */
992 if ((long)next - (long)val < 0) {
994 case MEM_CGROUP_TARGET_THRESH:
995 next = val + THRESHOLDS_EVENTS_TARGET;
997 case MEM_CGROUP_TARGET_SOFTLIMIT:
998 next = val + SOFTLIMIT_EVENTS_TARGET;
1000 case MEM_CGROUP_TARGET_NUMAINFO:
1001 next = val + NUMAINFO_EVENTS_TARGET;
1006 __this_cpu_write(memcg->stat->targets[target], next);
1013 * Check events in order.
1016 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1019 /* threshold event is triggered in finer grain than soft limit */
1020 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1021 MEM_CGROUP_TARGET_THRESH))) {
1023 bool do_numainfo __maybe_unused;
1025 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1026 MEM_CGROUP_TARGET_SOFTLIMIT);
1027 #if MAX_NUMNODES > 1
1028 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1029 MEM_CGROUP_TARGET_NUMAINFO);
1033 mem_cgroup_threshold(memcg);
1034 if (unlikely(do_softlimit))
1035 mem_cgroup_update_tree(memcg, page);
1036 #if MAX_NUMNODES > 1
1037 if (unlikely(do_numainfo))
1038 atomic_inc(&memcg->numainfo_events);
1044 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1047 * mm_update_next_owner() may clear mm->owner to NULL
1048 * if it races with swapoff, page migration, etc.
1049 * So this can be called with p == NULL.
1054 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1057 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1059 struct mem_cgroup *memcg = NULL;
1064 * Because we have no locks, mm->owner's may be being moved to other
1065 * cgroup. We use css_tryget() here even if this looks
1066 * pessimistic (rather than adding locks here).
1070 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1071 if (unlikely(!memcg))
1073 } while (!css_tryget(&memcg->css));
1079 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1080 * ref. count) or NULL if the whole root's subtree has been visited.
1082 * helper function to be used by mem_cgroup_iter
1084 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1085 struct mem_cgroup *last_visited)
1087 struct cgroup_subsys_state *prev_css, *next_css;
1089 prev_css = last_visited ? &last_visited->css : NULL;
1091 next_css = css_next_descendant_pre(prev_css, &root->css);
1094 * Even if we found a group we have to make sure it is
1095 * alive. css && !memcg means that the groups should be
1096 * skipped and we should continue the tree walk.
1097 * last_visited css is safe to use because it is
1098 * protected by css_get and the tree walk is rcu safe.
1101 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
1103 if (css_tryget(&mem->css))
1106 prev_css = next_css;
1114 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1117 * When a group in the hierarchy below root is destroyed, the
1118 * hierarchy iterator can no longer be trusted since it might
1119 * have pointed to the destroyed group. Invalidate it.
1121 atomic_inc(&root->dead_count);
1124 static struct mem_cgroup *
1125 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1126 struct mem_cgroup *root,
1129 struct mem_cgroup *position = NULL;
1131 * A cgroup destruction happens in two stages: offlining and
1132 * release. They are separated by a RCU grace period.
1134 * If the iterator is valid, we may still race with an
1135 * offlining. The RCU lock ensures the object won't be
1136 * released, tryget will fail if we lost the race.
1138 *sequence = atomic_read(&root->dead_count);
1139 if (iter->last_dead_count == *sequence) {
1141 position = iter->last_visited;
1142 if (position && !css_tryget(&position->css))
1148 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1149 struct mem_cgroup *last_visited,
1150 struct mem_cgroup *new_position,
1154 css_put(&last_visited->css);
1156 * We store the sequence count from the time @last_visited was
1157 * loaded successfully instead of rereading it here so that we
1158 * don't lose destruction events in between. We could have
1159 * raced with the destruction of @new_position after all.
1161 iter->last_visited = new_position;
1163 iter->last_dead_count = sequence;
1167 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1168 * @root: hierarchy root
1169 * @prev: previously returned memcg, NULL on first invocation
1170 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1172 * Returns references to children of the hierarchy below @root, or
1173 * @root itself, or %NULL after a full round-trip.
1175 * Caller must pass the return value in @prev on subsequent
1176 * invocations for reference counting, or use mem_cgroup_iter_break()
1177 * to cancel a hierarchy walk before the round-trip is complete.
1179 * Reclaimers can specify a zone and a priority level in @reclaim to
1180 * divide up the memcgs in the hierarchy among all concurrent
1181 * reclaimers operating on the same zone and priority.
1183 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1184 struct mem_cgroup *prev,
1185 struct mem_cgroup_reclaim_cookie *reclaim)
1187 struct mem_cgroup *memcg = NULL;
1188 struct mem_cgroup *last_visited = NULL;
1190 if (mem_cgroup_disabled())
1194 root = root_mem_cgroup;
1196 if (prev && !reclaim)
1197 last_visited = prev;
1199 if (!root->use_hierarchy && root != root_mem_cgroup) {
1207 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1208 int uninitialized_var(seq);
1211 int nid = zone_to_nid(reclaim->zone);
1212 int zid = zone_idx(reclaim->zone);
1213 struct mem_cgroup_per_zone *mz;
1215 mz = mem_cgroup_zoneinfo(root, nid, zid);
1216 iter = &mz->reclaim_iter[reclaim->priority];
1217 if (prev && reclaim->generation != iter->generation) {
1218 iter->last_visited = NULL;
1222 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1225 memcg = __mem_cgroup_iter_next(root, last_visited);
1228 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1232 else if (!prev && memcg)
1233 reclaim->generation = iter->generation;
1242 if (prev && prev != root)
1243 css_put(&prev->css);
1249 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1250 * @root: hierarchy root
1251 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1253 void mem_cgroup_iter_break(struct mem_cgroup *root,
1254 struct mem_cgroup *prev)
1257 root = root_mem_cgroup;
1258 if (prev && prev != root)
1259 css_put(&prev->css);
1263 * Iteration constructs for visiting all cgroups (under a tree). If
1264 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1265 * be used for reference counting.
1267 #define for_each_mem_cgroup_tree(iter, root) \
1268 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1270 iter = mem_cgroup_iter(root, iter, NULL))
1272 #define for_each_mem_cgroup(iter) \
1273 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1275 iter = mem_cgroup_iter(NULL, iter, NULL))
1277 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1279 struct mem_cgroup *memcg;
1282 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1283 if (unlikely(!memcg))
1288 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1291 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1299 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1302 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1303 * @zone: zone of the wanted lruvec
1304 * @memcg: memcg of the wanted lruvec
1306 * Returns the lru list vector holding pages for the given @zone and
1307 * @mem. This can be the global zone lruvec, if the memory controller
1310 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1311 struct mem_cgroup *memcg)
1313 struct mem_cgroup_per_zone *mz;
1314 struct lruvec *lruvec;
1316 if (mem_cgroup_disabled()) {
1317 lruvec = &zone->lruvec;
1321 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1322 lruvec = &mz->lruvec;
1325 * Since a node can be onlined after the mem_cgroup was created,
1326 * we have to be prepared to initialize lruvec->zone here;
1327 * and if offlined then reonlined, we need to reinitialize it.
1329 if (unlikely(lruvec->zone != zone))
1330 lruvec->zone = zone;
1335 * Following LRU functions are allowed to be used without PCG_LOCK.
1336 * Operations are called by routine of global LRU independently from memcg.
1337 * What we have to take care of here is validness of pc->mem_cgroup.
1339 * Changes to pc->mem_cgroup happens when
1342 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1343 * It is added to LRU before charge.
1344 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1345 * When moving account, the page is not on LRU. It's isolated.
1349 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1351 * @zone: zone of the page
1353 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1355 struct mem_cgroup_per_zone *mz;
1356 struct mem_cgroup *memcg;
1357 struct page_cgroup *pc;
1358 struct lruvec *lruvec;
1360 if (mem_cgroup_disabled()) {
1361 lruvec = &zone->lruvec;
1365 pc = lookup_page_cgroup(page);
1366 memcg = pc->mem_cgroup;
1369 * Surreptitiously switch any uncharged offlist page to root:
1370 * an uncharged page off lru does nothing to secure
1371 * its former mem_cgroup from sudden removal.
1373 * Our caller holds lru_lock, and PageCgroupUsed is updated
1374 * under page_cgroup lock: between them, they make all uses
1375 * of pc->mem_cgroup safe.
1377 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1378 pc->mem_cgroup = memcg = root_mem_cgroup;
1380 mz = page_cgroup_zoneinfo(memcg, page);
1381 lruvec = &mz->lruvec;
1384 * Since a node can be onlined after the mem_cgroup was created,
1385 * we have to be prepared to initialize lruvec->zone here;
1386 * and if offlined then reonlined, we need to reinitialize it.
1388 if (unlikely(lruvec->zone != zone))
1389 lruvec->zone = zone;
1394 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1395 * @lruvec: mem_cgroup per zone lru vector
1396 * @lru: index of lru list the page is sitting on
1397 * @nr_pages: positive when adding or negative when removing
1399 * This function must be called when a page is added to or removed from an
1402 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1405 struct mem_cgroup_per_zone *mz;
1406 unsigned long *lru_size;
1408 if (mem_cgroup_disabled())
1411 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1412 lru_size = mz->lru_size + lru;
1413 *lru_size += nr_pages;
1414 VM_BUG_ON((long)(*lru_size) < 0);
1418 * Checks whether given mem is same or in the root_mem_cgroup's
1421 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1422 struct mem_cgroup *memcg)
1424 if (root_memcg == memcg)
1426 if (!root_memcg->use_hierarchy || !memcg)
1428 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1431 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1432 struct mem_cgroup *memcg)
1437 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1442 bool task_in_mem_cgroup(struct task_struct *task,
1443 const struct mem_cgroup *memcg)
1445 struct mem_cgroup *curr = NULL;
1446 struct task_struct *p;
1449 p = find_lock_task_mm(task);
1451 curr = try_get_mem_cgroup_from_mm(p->mm);
1455 * All threads may have already detached their mm's, but the oom
1456 * killer still needs to detect if they have already been oom
1457 * killed to prevent needlessly killing additional tasks.
1460 curr = mem_cgroup_from_task(task);
1462 css_get(&curr->css);
1468 * We should check use_hierarchy of "memcg" not "curr". Because checking
1469 * use_hierarchy of "curr" here make this function true if hierarchy is
1470 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1471 * hierarchy(even if use_hierarchy is disabled in "memcg").
1473 ret = mem_cgroup_same_or_subtree(memcg, curr);
1474 css_put(&curr->css);
1478 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1480 unsigned long inactive_ratio;
1481 unsigned long inactive;
1482 unsigned long active;
1485 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1486 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1488 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1490 inactive_ratio = int_sqrt(10 * gb);
1494 return inactive * inactive_ratio < active;
1497 #define mem_cgroup_from_res_counter(counter, member) \
1498 container_of(counter, struct mem_cgroup, member)
1501 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1502 * @memcg: the memory cgroup
1504 * Returns the maximum amount of memory @mem can be charged with, in
1507 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1509 unsigned long long margin;
1511 margin = res_counter_margin(&memcg->res);
1512 if (do_swap_account)
1513 margin = min(margin, res_counter_margin(&memcg->memsw));
1514 return margin >> PAGE_SHIFT;
1517 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1520 if (!css_parent(&memcg->css))
1521 return vm_swappiness;
1523 return memcg->swappiness;
1527 * memcg->moving_account is used for checking possibility that some thread is
1528 * calling move_account(). When a thread on CPU-A starts moving pages under
1529 * a memcg, other threads should check memcg->moving_account under
1530 * rcu_read_lock(), like this:
1534 * memcg->moving_account+1 if (memcg->mocing_account)
1536 * synchronize_rcu() update something.
1541 /* for quick checking without looking up memcg */
1542 atomic_t memcg_moving __read_mostly;
1544 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1546 atomic_inc(&memcg_moving);
1547 atomic_inc(&memcg->moving_account);
1551 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1554 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1555 * We check NULL in callee rather than caller.
1558 atomic_dec(&memcg_moving);
1559 atomic_dec(&memcg->moving_account);
1564 * 2 routines for checking "mem" is under move_account() or not.
1566 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1567 * is used for avoiding races in accounting. If true,
1568 * pc->mem_cgroup may be overwritten.
1570 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1571 * under hierarchy of moving cgroups. This is for
1572 * waiting at hith-memory prressure caused by "move".
1575 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1577 VM_BUG_ON(!rcu_read_lock_held());
1578 return atomic_read(&memcg->moving_account) > 0;
1581 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1583 struct mem_cgroup *from;
1584 struct mem_cgroup *to;
1587 * Unlike task_move routines, we access mc.to, mc.from not under
1588 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1590 spin_lock(&mc.lock);
1596 ret = mem_cgroup_same_or_subtree(memcg, from)
1597 || mem_cgroup_same_or_subtree(memcg, to);
1599 spin_unlock(&mc.lock);
1603 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1605 if (mc.moving_task && current != mc.moving_task) {
1606 if (mem_cgroup_under_move(memcg)) {
1608 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1609 /* moving charge context might have finished. */
1612 finish_wait(&mc.waitq, &wait);
1620 * Take this lock when
1621 * - a code tries to modify page's memcg while it's USED.
1622 * - a code tries to modify page state accounting in a memcg.
1623 * see mem_cgroup_stolen(), too.
1625 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1626 unsigned long *flags)
1628 spin_lock_irqsave(&memcg->move_lock, *flags);
1631 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1632 unsigned long *flags)
1634 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1637 #define K(x) ((x) << (PAGE_SHIFT-10))
1639 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1640 * @memcg: The memory cgroup that went over limit
1641 * @p: Task that is going to be killed
1643 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1646 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1648 struct cgroup *task_cgrp;
1649 struct cgroup *mem_cgrp;
1651 * Need a buffer in BSS, can't rely on allocations. The code relies
1652 * on the assumption that OOM is serialized for memory controller.
1653 * If this assumption is broken, revisit this code.
1655 static char memcg_name[PATH_MAX];
1657 struct mem_cgroup *iter;
1665 mem_cgrp = memcg->css.cgroup;
1666 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1668 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1671 * Unfortunately, we are unable to convert to a useful name
1672 * But we'll still print out the usage information
1679 pr_info("Task in %s killed", memcg_name);
1682 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1690 * Continues from above, so we don't need an KERN_ level
1692 pr_cont(" as a result of limit of %s\n", memcg_name);
1695 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1696 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1697 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1698 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1699 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1700 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1701 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1702 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1703 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1704 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1705 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1706 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1708 for_each_mem_cgroup_tree(iter, memcg) {
1709 pr_info("Memory cgroup stats");
1712 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1714 pr_cont(" for %s", memcg_name);
1718 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1719 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1721 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1722 K(mem_cgroup_read_stat(iter, i)));
1725 for (i = 0; i < NR_LRU_LISTS; i++)
1726 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1727 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1734 * This function returns the number of memcg under hierarchy tree. Returns
1735 * 1(self count) if no children.
1737 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1740 struct mem_cgroup *iter;
1742 for_each_mem_cgroup_tree(iter, memcg)
1748 * Return the memory (and swap, if configured) limit for a memcg.
1750 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1754 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1757 * Do not consider swap space if we cannot swap due to swappiness
1759 if (mem_cgroup_swappiness(memcg)) {
1762 limit += total_swap_pages << PAGE_SHIFT;
1763 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1766 * If memsw is finite and limits the amount of swap space
1767 * available to this memcg, return that limit.
1769 limit = min(limit, memsw);
1775 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1778 struct mem_cgroup *iter;
1779 unsigned long chosen_points = 0;
1780 unsigned long totalpages;
1781 unsigned int points = 0;
1782 struct task_struct *chosen = NULL;
1785 * If current has a pending SIGKILL or is exiting, then automatically
1786 * select it. The goal is to allow it to allocate so that it may
1787 * quickly exit and free its memory.
1789 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1790 set_thread_flag(TIF_MEMDIE);
1794 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1795 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1796 for_each_mem_cgroup_tree(iter, memcg) {
1797 struct css_task_iter it;
1798 struct task_struct *task;
1800 css_task_iter_start(&iter->css, &it);
1801 while ((task = css_task_iter_next(&it))) {
1802 switch (oom_scan_process_thread(task, totalpages, NULL,
1804 case OOM_SCAN_SELECT:
1806 put_task_struct(chosen);
1808 chosen_points = ULONG_MAX;
1809 get_task_struct(chosen);
1811 case OOM_SCAN_CONTINUE:
1813 case OOM_SCAN_ABORT:
1814 css_task_iter_end(&it);
1815 mem_cgroup_iter_break(memcg, iter);
1817 put_task_struct(chosen);
1822 points = oom_badness(task, memcg, NULL, totalpages);
1823 if (points > chosen_points) {
1825 put_task_struct(chosen);
1827 chosen_points = points;
1828 get_task_struct(chosen);
1831 css_task_iter_end(&it);
1836 points = chosen_points * 1000 / totalpages;
1837 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1838 NULL, "Memory cgroup out of memory");
1841 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1843 unsigned long flags)
1845 unsigned long total = 0;
1846 bool noswap = false;
1849 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1851 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1854 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1856 drain_all_stock_async(memcg);
1857 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1859 * Allow limit shrinkers, which are triggered directly
1860 * by userspace, to catch signals and stop reclaim
1861 * after minimal progress, regardless of the margin.
1863 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1865 if (mem_cgroup_margin(memcg))
1868 * If nothing was reclaimed after two attempts, there
1869 * may be no reclaimable pages in this hierarchy.
1878 * test_mem_cgroup_node_reclaimable
1879 * @memcg: the target memcg
1880 * @nid: the node ID to be checked.
1881 * @noswap : specify true here if the user wants flle only information.
1883 * This function returns whether the specified memcg contains any
1884 * reclaimable pages on a node. Returns true if there are any reclaimable
1885 * pages in the node.
1887 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1888 int nid, bool noswap)
1890 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1892 if (noswap || !total_swap_pages)
1894 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1899 #if MAX_NUMNODES > 1
1902 * Always updating the nodemask is not very good - even if we have an empty
1903 * list or the wrong list here, we can start from some node and traverse all
1904 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1907 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1911 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1912 * pagein/pageout changes since the last update.
1914 if (!atomic_read(&memcg->numainfo_events))
1916 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1919 /* make a nodemask where this memcg uses memory from */
1920 memcg->scan_nodes = node_states[N_MEMORY];
1922 for_each_node_mask(nid, node_states[N_MEMORY]) {
1924 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1925 node_clear(nid, memcg->scan_nodes);
1928 atomic_set(&memcg->numainfo_events, 0);
1929 atomic_set(&memcg->numainfo_updating, 0);
1933 * Selecting a node where we start reclaim from. Because what we need is just
1934 * reducing usage counter, start from anywhere is O,K. Considering
1935 * memory reclaim from current node, there are pros. and cons.
1937 * Freeing memory from current node means freeing memory from a node which
1938 * we'll use or we've used. So, it may make LRU bad. And if several threads
1939 * hit limits, it will see a contention on a node. But freeing from remote
1940 * node means more costs for memory reclaim because of memory latency.
1942 * Now, we use round-robin. Better algorithm is welcomed.
1944 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1948 mem_cgroup_may_update_nodemask(memcg);
1949 node = memcg->last_scanned_node;
1951 node = next_node(node, memcg->scan_nodes);
1952 if (node == MAX_NUMNODES)
1953 node = first_node(memcg->scan_nodes);
1955 * We call this when we hit limit, not when pages are added to LRU.
1956 * No LRU may hold pages because all pages are UNEVICTABLE or
1957 * memcg is too small and all pages are not on LRU. In that case,
1958 * we use curret node.
1960 if (unlikely(node == MAX_NUMNODES))
1961 node = numa_node_id();
1963 memcg->last_scanned_node = node;
1968 * Check all nodes whether it contains reclaimable pages or not.
1969 * For quick scan, we make use of scan_nodes. This will allow us to skip
1970 * unused nodes. But scan_nodes is lazily updated and may not cotain
1971 * enough new information. We need to do double check.
1973 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1978 * quick check...making use of scan_node.
1979 * We can skip unused nodes.
1981 if (!nodes_empty(memcg->scan_nodes)) {
1982 for (nid = first_node(memcg->scan_nodes);
1984 nid = next_node(nid, memcg->scan_nodes)) {
1986 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1991 * Check rest of nodes.
1993 for_each_node_state(nid, N_MEMORY) {
1994 if (node_isset(nid, memcg->scan_nodes))
1996 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2003 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2008 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2010 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2014 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2017 unsigned long *total_scanned)
2019 struct mem_cgroup *victim = NULL;
2022 unsigned long excess;
2023 unsigned long nr_scanned;
2024 struct mem_cgroup_reclaim_cookie reclaim = {
2029 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2032 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2037 * If we have not been able to reclaim
2038 * anything, it might because there are
2039 * no reclaimable pages under this hierarchy
2044 * We want to do more targeted reclaim.
2045 * excess >> 2 is not to excessive so as to
2046 * reclaim too much, nor too less that we keep
2047 * coming back to reclaim from this cgroup
2049 if (total >= (excess >> 2) ||
2050 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2055 if (!mem_cgroup_reclaimable(victim, false))
2057 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2059 *total_scanned += nr_scanned;
2060 if (!res_counter_soft_limit_excess(&root_memcg->res))
2063 mem_cgroup_iter_break(root_memcg, victim);
2067 static DEFINE_SPINLOCK(memcg_oom_lock);
2070 * Check OOM-Killer is already running under our hierarchy.
2071 * If someone is running, return false.
2073 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2075 struct mem_cgroup *iter, *failed = NULL;
2077 spin_lock(&memcg_oom_lock);
2079 for_each_mem_cgroup_tree(iter, memcg) {
2080 if (iter->oom_lock) {
2082 * this subtree of our hierarchy is already locked
2083 * so we cannot give a lock.
2086 mem_cgroup_iter_break(memcg, iter);
2089 iter->oom_lock = true;
2094 * OK, we failed to lock the whole subtree so we have
2095 * to clean up what we set up to the failing subtree
2097 for_each_mem_cgroup_tree(iter, memcg) {
2098 if (iter == failed) {
2099 mem_cgroup_iter_break(memcg, iter);
2102 iter->oom_lock = false;
2106 spin_unlock(&memcg_oom_lock);
2111 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2113 struct mem_cgroup *iter;
2115 spin_lock(&memcg_oom_lock);
2116 for_each_mem_cgroup_tree(iter, memcg)
2117 iter->oom_lock = false;
2118 spin_unlock(&memcg_oom_lock);
2121 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2123 struct mem_cgroup *iter;
2125 for_each_mem_cgroup_tree(iter, memcg)
2126 atomic_inc(&iter->under_oom);
2129 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2131 struct mem_cgroup *iter;
2134 * When a new child is created while the hierarchy is under oom,
2135 * mem_cgroup_oom_lock() may not be called. We have to use
2136 * atomic_add_unless() here.
2138 for_each_mem_cgroup_tree(iter, memcg)
2139 atomic_add_unless(&iter->under_oom, -1, 0);
2142 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2144 struct oom_wait_info {
2145 struct mem_cgroup *memcg;
2149 static int memcg_oom_wake_function(wait_queue_t *wait,
2150 unsigned mode, int sync, void *arg)
2152 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2153 struct mem_cgroup *oom_wait_memcg;
2154 struct oom_wait_info *oom_wait_info;
2156 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2157 oom_wait_memcg = oom_wait_info->memcg;
2160 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2161 * Then we can use css_is_ancestor without taking care of RCU.
2163 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2164 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2166 return autoremove_wake_function(wait, mode, sync, arg);
2169 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2171 atomic_inc(&memcg->oom_wakeups);
2172 /* for filtering, pass "memcg" as argument. */
2173 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2176 static void memcg_oom_recover(struct mem_cgroup *memcg)
2178 if (memcg && atomic_read(&memcg->under_oom))
2179 memcg_wakeup_oom(memcg);
2182 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2184 if (!current->memcg_oom.may_oom)
2187 * We are in the middle of the charge context here, so we
2188 * don't want to block when potentially sitting on a callstack
2189 * that holds all kinds of filesystem and mm locks.
2191 * Also, the caller may handle a failed allocation gracefully
2192 * (like optional page cache readahead) and so an OOM killer
2193 * invocation might not even be necessary.
2195 * That's why we don't do anything here except remember the
2196 * OOM context and then deal with it at the end of the page
2197 * fault when the stack is unwound, the locks are released,
2198 * and when we know whether the fault was overall successful.
2200 css_get(&memcg->css);
2201 current->memcg_oom.memcg = memcg;
2202 current->memcg_oom.gfp_mask = mask;
2203 current->memcg_oom.order = order;
2207 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2208 * @handle: actually kill/wait or just clean up the OOM state
2210 * This has to be called at the end of a page fault if the memcg OOM
2211 * handler was enabled.
2213 * Memcg supports userspace OOM handling where failed allocations must
2214 * sleep on a waitqueue until the userspace task resolves the
2215 * situation. Sleeping directly in the charge context with all kinds
2216 * of locks held is not a good idea, instead we remember an OOM state
2217 * in the task and mem_cgroup_oom_synchronize() has to be called at
2218 * the end of the page fault to complete the OOM handling.
2220 * Returns %true if an ongoing memcg OOM situation was detected and
2221 * completed, %false otherwise.
2223 bool mem_cgroup_oom_synchronize(bool handle)
2225 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2226 struct oom_wait_info owait;
2229 /* OOM is global, do not handle */
2236 owait.memcg = memcg;
2237 owait.wait.flags = 0;
2238 owait.wait.func = memcg_oom_wake_function;
2239 owait.wait.private = current;
2240 INIT_LIST_HEAD(&owait.wait.task_list);
2242 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2243 mem_cgroup_mark_under_oom(memcg);
2245 locked = mem_cgroup_oom_trylock(memcg);
2248 mem_cgroup_oom_notify(memcg);
2250 if (locked && !memcg->oom_kill_disable) {
2251 mem_cgroup_unmark_under_oom(memcg);
2252 finish_wait(&memcg_oom_waitq, &owait.wait);
2253 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2254 current->memcg_oom.order);
2257 mem_cgroup_unmark_under_oom(memcg);
2258 finish_wait(&memcg_oom_waitq, &owait.wait);
2262 mem_cgroup_oom_unlock(memcg);
2264 * There is no guarantee that an OOM-lock contender
2265 * sees the wakeups triggered by the OOM kill
2266 * uncharges. Wake any sleepers explicitely.
2268 memcg_oom_recover(memcg);
2271 current->memcg_oom.memcg = NULL;
2272 css_put(&memcg->css);
2277 * Currently used to update mapped file statistics, but the routine can be
2278 * generalized to update other statistics as well.
2280 * Notes: Race condition
2282 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2283 * it tends to be costly. But considering some conditions, we doesn't need
2284 * to do so _always_.
2286 * Considering "charge", lock_page_cgroup() is not required because all
2287 * file-stat operations happen after a page is attached to radix-tree. There
2288 * are no race with "charge".
2290 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2291 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2292 * if there are race with "uncharge". Statistics itself is properly handled
2295 * Considering "move", this is an only case we see a race. To make the race
2296 * small, we check mm->moving_account and detect there are possibility of race
2297 * If there is, we take a lock.
2300 void __mem_cgroup_begin_update_page_stat(struct page *page,
2301 bool *locked, unsigned long *flags)
2303 struct mem_cgroup *memcg;
2304 struct page_cgroup *pc;
2306 pc = lookup_page_cgroup(page);
2308 memcg = pc->mem_cgroup;
2309 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2312 * If this memory cgroup is not under account moving, we don't
2313 * need to take move_lock_mem_cgroup(). Because we already hold
2314 * rcu_read_lock(), any calls to move_account will be delayed until
2315 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2317 if (!mem_cgroup_stolen(memcg))
2320 move_lock_mem_cgroup(memcg, flags);
2321 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2322 move_unlock_mem_cgroup(memcg, flags);
2328 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2330 struct page_cgroup *pc = lookup_page_cgroup(page);
2333 * It's guaranteed that pc->mem_cgroup never changes while
2334 * lock is held because a routine modifies pc->mem_cgroup
2335 * should take move_lock_mem_cgroup().
2337 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2340 void mem_cgroup_update_page_stat(struct page *page,
2341 enum mem_cgroup_stat_index idx, int val)
2343 struct mem_cgroup *memcg;
2344 struct page_cgroup *pc = lookup_page_cgroup(page);
2345 unsigned long uninitialized_var(flags);
2347 if (mem_cgroup_disabled())
2350 VM_BUG_ON(!rcu_read_lock_held());
2351 memcg = pc->mem_cgroup;
2352 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2355 this_cpu_add(memcg->stat->count[idx], val);
2359 * size of first charge trial. "32" comes from vmscan.c's magic value.
2360 * TODO: maybe necessary to use big numbers in big irons.
2362 #define CHARGE_BATCH 32U
2363 struct memcg_stock_pcp {
2364 struct mem_cgroup *cached; /* this never be root cgroup */
2365 unsigned int nr_pages;
2366 struct work_struct work;
2367 unsigned long flags;
2368 #define FLUSHING_CACHED_CHARGE 0
2370 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2371 static DEFINE_MUTEX(percpu_charge_mutex);
2374 * consume_stock: Try to consume stocked charge on this cpu.
2375 * @memcg: memcg to consume from.
2376 * @nr_pages: how many pages to charge.
2378 * The charges will only happen if @memcg matches the current cpu's memcg
2379 * stock, and at least @nr_pages are available in that stock. Failure to
2380 * service an allocation will refill the stock.
2382 * returns true if successful, false otherwise.
2384 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2386 struct memcg_stock_pcp *stock;
2389 if (nr_pages > CHARGE_BATCH)
2392 stock = &get_cpu_var(memcg_stock);
2393 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2394 stock->nr_pages -= nr_pages;
2395 else /* need to call res_counter_charge */
2397 put_cpu_var(memcg_stock);
2402 * Returns stocks cached in percpu to res_counter and reset cached information.
2404 static void drain_stock(struct memcg_stock_pcp *stock)
2406 struct mem_cgroup *old = stock->cached;
2408 if (stock->nr_pages) {
2409 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2411 res_counter_uncharge(&old->res, bytes);
2412 if (do_swap_account)
2413 res_counter_uncharge(&old->memsw, bytes);
2414 stock->nr_pages = 0;
2416 stock->cached = NULL;
2420 * This must be called under preempt disabled or must be called by
2421 * a thread which is pinned to local cpu.
2423 static void drain_local_stock(struct work_struct *dummy)
2425 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2427 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2430 static void __init memcg_stock_init(void)
2434 for_each_possible_cpu(cpu) {
2435 struct memcg_stock_pcp *stock =
2436 &per_cpu(memcg_stock, cpu);
2437 INIT_WORK(&stock->work, drain_local_stock);
2442 * Cache charges(val) which is from res_counter, to local per_cpu area.
2443 * This will be consumed by consume_stock() function, later.
2445 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2447 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2449 if (stock->cached != memcg) { /* reset if necessary */
2451 stock->cached = memcg;
2453 stock->nr_pages += nr_pages;
2454 put_cpu_var(memcg_stock);
2458 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2459 * of the hierarchy under it. sync flag says whether we should block
2460 * until the work is done.
2462 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2466 /* Notify other cpus that system-wide "drain" is running */
2469 for_each_online_cpu(cpu) {
2470 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2471 struct mem_cgroup *memcg;
2473 memcg = stock->cached;
2474 if (!memcg || !stock->nr_pages)
2476 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2478 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2480 drain_local_stock(&stock->work);
2482 schedule_work_on(cpu, &stock->work);
2490 for_each_online_cpu(cpu) {
2491 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2492 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2493 flush_work(&stock->work);
2500 * Tries to drain stocked charges in other cpus. This function is asynchronous
2501 * and just put a work per cpu for draining localy on each cpu. Caller can
2502 * expects some charges will be back to res_counter later but cannot wait for
2505 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2508 * If someone calls draining, avoid adding more kworker runs.
2510 if (!mutex_trylock(&percpu_charge_mutex))
2512 drain_all_stock(root_memcg, false);
2513 mutex_unlock(&percpu_charge_mutex);
2516 /* This is a synchronous drain interface. */
2517 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2519 /* called when force_empty is called */
2520 mutex_lock(&percpu_charge_mutex);
2521 drain_all_stock(root_memcg, true);
2522 mutex_unlock(&percpu_charge_mutex);
2526 * This function drains percpu counter value from DEAD cpu and
2527 * move it to local cpu. Note that this function can be preempted.
2529 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2533 spin_lock(&memcg->pcp_counter_lock);
2534 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2535 long x = per_cpu(memcg->stat->count[i], cpu);
2537 per_cpu(memcg->stat->count[i], cpu) = 0;
2538 memcg->nocpu_base.count[i] += x;
2540 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2541 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2543 per_cpu(memcg->stat->events[i], cpu) = 0;
2544 memcg->nocpu_base.events[i] += x;
2546 spin_unlock(&memcg->pcp_counter_lock);
2549 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2550 unsigned long action,
2553 int cpu = (unsigned long)hcpu;
2554 struct memcg_stock_pcp *stock;
2555 struct mem_cgroup *iter;
2557 if (action == CPU_ONLINE)
2560 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2563 for_each_mem_cgroup(iter)
2564 mem_cgroup_drain_pcp_counter(iter, cpu);
2566 stock = &per_cpu(memcg_stock, cpu);
2572 /* See __mem_cgroup_try_charge() for details */
2574 CHARGE_OK, /* success */
2575 CHARGE_RETRY, /* need to retry but retry is not bad */
2576 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2577 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2580 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2581 unsigned int nr_pages, unsigned int min_pages,
2584 unsigned long csize = nr_pages * PAGE_SIZE;
2585 struct mem_cgroup *mem_over_limit;
2586 struct res_counter *fail_res;
2587 unsigned long flags = 0;
2590 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2593 if (!do_swap_account)
2595 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2599 res_counter_uncharge(&memcg->res, csize);
2600 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2601 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2603 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2605 * Never reclaim on behalf of optional batching, retry with a
2606 * single page instead.
2608 if (nr_pages > min_pages)
2609 return CHARGE_RETRY;
2611 if (!(gfp_mask & __GFP_WAIT))
2612 return CHARGE_WOULDBLOCK;
2614 if (gfp_mask & __GFP_NORETRY)
2615 return CHARGE_NOMEM;
2617 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2618 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2619 return CHARGE_RETRY;
2621 * Even though the limit is exceeded at this point, reclaim
2622 * may have been able to free some pages. Retry the charge
2623 * before killing the task.
2625 * Only for regular pages, though: huge pages are rather
2626 * unlikely to succeed so close to the limit, and we fall back
2627 * to regular pages anyway in case of failure.
2629 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2630 return CHARGE_RETRY;
2633 * At task move, charge accounts can be doubly counted. So, it's
2634 * better to wait until the end of task_move if something is going on.
2636 if (mem_cgroup_wait_acct_move(mem_over_limit))
2637 return CHARGE_RETRY;
2640 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2642 return CHARGE_NOMEM;
2646 * __mem_cgroup_try_charge() does
2647 * 1. detect memcg to be charged against from passed *mm and *ptr,
2648 * 2. update res_counter
2649 * 3. call memory reclaim if necessary.
2651 * In some special case, if the task is fatal, fatal_signal_pending() or
2652 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2653 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2654 * as possible without any hazards. 2: all pages should have a valid
2655 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2656 * pointer, that is treated as a charge to root_mem_cgroup.
2658 * So __mem_cgroup_try_charge() will return
2659 * 0 ... on success, filling *ptr with a valid memcg pointer.
2660 * -ENOMEM ... charge failure because of resource limits.
2661 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2663 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2664 * the oom-killer can be invoked.
2666 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2668 unsigned int nr_pages,
2669 struct mem_cgroup **ptr,
2672 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2673 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2674 struct mem_cgroup *memcg = NULL;
2678 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2679 * in system level. So, allow to go ahead dying process in addition to
2682 if (unlikely(test_thread_flag(TIF_MEMDIE)
2683 || fatal_signal_pending(current)))
2686 if (unlikely(task_in_memcg_oom(current)))
2690 * We always charge the cgroup the mm_struct belongs to.
2691 * The mm_struct's mem_cgroup changes on task migration if the
2692 * thread group leader migrates. It's possible that mm is not
2693 * set, if so charge the root memcg (happens for pagecache usage).
2696 *ptr = root_mem_cgroup;
2698 if (*ptr) { /* css should be a valid one */
2700 if (mem_cgroup_is_root(memcg))
2702 if (consume_stock(memcg, nr_pages))
2704 css_get(&memcg->css);
2706 struct task_struct *p;
2709 p = rcu_dereference(mm->owner);
2711 * Because we don't have task_lock(), "p" can exit.
2712 * In that case, "memcg" can point to root or p can be NULL with
2713 * race with swapoff. Then, we have small risk of mis-accouning.
2714 * But such kind of mis-account by race always happens because
2715 * we don't have cgroup_mutex(). It's overkill and we allo that
2717 * (*) swapoff at el will charge against mm-struct not against
2718 * task-struct. So, mm->owner can be NULL.
2720 memcg = mem_cgroup_from_task(p);
2722 memcg = root_mem_cgroup;
2723 if (mem_cgroup_is_root(memcg)) {
2727 if (consume_stock(memcg, nr_pages)) {
2729 * It seems dagerous to access memcg without css_get().
2730 * But considering how consume_stok works, it's not
2731 * necessary. If consume_stock success, some charges
2732 * from this memcg are cached on this cpu. So, we
2733 * don't need to call css_get()/css_tryget() before
2734 * calling consume_stock().
2739 /* after here, we may be blocked. we need to get refcnt */
2740 if (!css_tryget(&memcg->css)) {
2748 bool invoke_oom = oom && !nr_oom_retries;
2750 /* If killed, bypass charge */
2751 if (fatal_signal_pending(current)) {
2752 css_put(&memcg->css);
2756 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2757 nr_pages, invoke_oom);
2761 case CHARGE_RETRY: /* not in OOM situation but retry */
2763 css_put(&memcg->css);
2766 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2767 css_put(&memcg->css);
2769 case CHARGE_NOMEM: /* OOM routine works */
2770 if (!oom || invoke_oom) {
2771 css_put(&memcg->css);
2777 } while (ret != CHARGE_OK);
2779 if (batch > nr_pages)
2780 refill_stock(memcg, batch - nr_pages);
2781 css_put(&memcg->css);
2787 if (gfp_mask & __GFP_NOFAIL)
2791 *ptr = root_mem_cgroup;
2796 * Somemtimes we have to undo a charge we got by try_charge().
2797 * This function is for that and do uncharge, put css's refcnt.
2798 * gotten by try_charge().
2800 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2801 unsigned int nr_pages)
2803 if (!mem_cgroup_is_root(memcg)) {
2804 unsigned long bytes = nr_pages * PAGE_SIZE;
2806 res_counter_uncharge(&memcg->res, bytes);
2807 if (do_swap_account)
2808 res_counter_uncharge(&memcg->memsw, bytes);
2813 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2814 * This is useful when moving usage to parent cgroup.
2816 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2817 unsigned int nr_pages)
2819 unsigned long bytes = nr_pages * PAGE_SIZE;
2821 if (mem_cgroup_is_root(memcg))
2824 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2825 if (do_swap_account)
2826 res_counter_uncharge_until(&memcg->memsw,
2827 memcg->memsw.parent, bytes);
2831 * A helper function to get mem_cgroup from ID. must be called under
2832 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2833 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2834 * called against removed memcg.)
2836 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2838 /* ID 0 is unused ID */
2841 return mem_cgroup_from_id(id);
2844 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2846 struct mem_cgroup *memcg = NULL;
2847 struct page_cgroup *pc;
2851 VM_BUG_ON(!PageLocked(page));
2853 pc = lookup_page_cgroup(page);
2854 lock_page_cgroup(pc);
2855 if (PageCgroupUsed(pc)) {
2856 memcg = pc->mem_cgroup;
2857 if (memcg && !css_tryget(&memcg->css))
2859 } else if (PageSwapCache(page)) {
2860 ent.val = page_private(page);
2861 id = lookup_swap_cgroup_id(ent);
2863 memcg = mem_cgroup_lookup(id);
2864 if (memcg && !css_tryget(&memcg->css))
2868 unlock_page_cgroup(pc);
2872 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2874 unsigned int nr_pages,
2875 enum charge_type ctype,
2878 struct page_cgroup *pc = lookup_page_cgroup(page);
2879 struct zone *uninitialized_var(zone);
2880 struct lruvec *lruvec;
2881 bool was_on_lru = false;
2884 lock_page_cgroup(pc);
2885 VM_BUG_ON(PageCgroupUsed(pc));
2887 * we don't need page_cgroup_lock about tail pages, becase they are not
2888 * accessed by any other context at this point.
2892 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2893 * may already be on some other mem_cgroup's LRU. Take care of it.
2896 zone = page_zone(page);
2897 spin_lock_irq(&zone->lru_lock);
2898 if (PageLRU(page)) {
2899 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2901 del_page_from_lru_list(page, lruvec, page_lru(page));
2906 pc->mem_cgroup = memcg;
2908 * We access a page_cgroup asynchronously without lock_page_cgroup().
2909 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2910 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2911 * before USED bit, we need memory barrier here.
2912 * See mem_cgroup_add_lru_list(), etc.
2915 SetPageCgroupUsed(pc);
2919 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2920 VM_BUG_ON(PageLRU(page));
2922 add_page_to_lru_list(page, lruvec, page_lru(page));
2924 spin_unlock_irq(&zone->lru_lock);
2927 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2932 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2933 unlock_page_cgroup(pc);
2936 * "charge_statistics" updated event counter. Then, check it.
2937 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2938 * if they exceeds softlimit.
2940 memcg_check_events(memcg, page);
2943 static DEFINE_MUTEX(set_limit_mutex);
2945 #ifdef CONFIG_MEMCG_KMEM
2946 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2948 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2949 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2953 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2954 * in the memcg_cache_params struct.
2956 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2958 struct kmem_cache *cachep;
2960 VM_BUG_ON(p->is_root_cache);
2961 cachep = p->root_cache;
2962 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2965 #ifdef CONFIG_SLABINFO
2966 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2967 struct cftype *cft, struct seq_file *m)
2969 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2970 struct memcg_cache_params *params;
2972 if (!memcg_can_account_kmem(memcg))
2975 print_slabinfo_header(m);
2977 mutex_lock(&memcg->slab_caches_mutex);
2978 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2979 cache_show(memcg_params_to_cache(params), m);
2980 mutex_unlock(&memcg->slab_caches_mutex);
2986 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2988 struct res_counter *fail_res;
2989 struct mem_cgroup *_memcg;
2993 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2998 * Conditions under which we can wait for the oom_killer. Those are
2999 * the same conditions tested by the core page allocator
3001 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
3004 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3007 if (ret == -EINTR) {
3009 * __mem_cgroup_try_charge() chosed to bypass to root due to
3010 * OOM kill or fatal signal. Since our only options are to
3011 * either fail the allocation or charge it to this cgroup, do
3012 * it as a temporary condition. But we can't fail. From a
3013 * kmem/slab perspective, the cache has already been selected,
3014 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3017 * This condition will only trigger if the task entered
3018 * memcg_charge_kmem in a sane state, but was OOM-killed during
3019 * __mem_cgroup_try_charge() above. Tasks that were already
3020 * dying when the allocation triggers should have been already
3021 * directed to the root cgroup in memcontrol.h
3023 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3024 if (do_swap_account)
3025 res_counter_charge_nofail(&memcg->memsw, size,
3029 res_counter_uncharge(&memcg->kmem, size);
3034 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3036 res_counter_uncharge(&memcg->res, size);
3037 if (do_swap_account)
3038 res_counter_uncharge(&memcg->memsw, size);
3041 if (res_counter_uncharge(&memcg->kmem, size))
3045 * Releases a reference taken in kmem_cgroup_css_offline in case
3046 * this last uncharge is racing with the offlining code or it is
3047 * outliving the memcg existence.
3049 * The memory barrier imposed by test&clear is paired with the
3050 * explicit one in memcg_kmem_mark_dead().
3052 if (memcg_kmem_test_and_clear_dead(memcg))
3053 css_put(&memcg->css);
3056 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3061 mutex_lock(&memcg->slab_caches_mutex);
3062 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3063 mutex_unlock(&memcg->slab_caches_mutex);
3067 * helper for acessing a memcg's index. It will be used as an index in the
3068 * child cache array in kmem_cache, and also to derive its name. This function
3069 * will return -1 when this is not a kmem-limited memcg.
3071 int memcg_cache_id(struct mem_cgroup *memcg)
3073 return memcg ? memcg->kmemcg_id : -1;
3077 * This ends up being protected by the set_limit mutex, during normal
3078 * operation, because that is its main call site.
3080 * But when we create a new cache, we can call this as well if its parent
3081 * is kmem-limited. That will have to hold set_limit_mutex as well.
3083 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3087 num = ida_simple_get(&kmem_limited_groups,
3088 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3092 * After this point, kmem_accounted (that we test atomically in
3093 * the beginning of this conditional), is no longer 0. This
3094 * guarantees only one process will set the following boolean
3095 * to true. We don't need test_and_set because we're protected
3096 * by the set_limit_mutex anyway.
3098 memcg_kmem_set_activated(memcg);
3100 ret = memcg_update_all_caches(num+1);
3102 ida_simple_remove(&kmem_limited_groups, num);
3103 memcg_kmem_clear_activated(memcg);
3107 memcg->kmemcg_id = num;
3108 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3109 mutex_init(&memcg->slab_caches_mutex);
3113 static size_t memcg_caches_array_size(int num_groups)
3116 if (num_groups <= 0)
3119 size = 2 * num_groups;
3120 if (size < MEMCG_CACHES_MIN_SIZE)
3121 size = MEMCG_CACHES_MIN_SIZE;
3122 else if (size > MEMCG_CACHES_MAX_SIZE)
3123 size = MEMCG_CACHES_MAX_SIZE;
3129 * We should update the current array size iff all caches updates succeed. This
3130 * can only be done from the slab side. The slab mutex needs to be held when
3133 void memcg_update_array_size(int num)
3135 if (num > memcg_limited_groups_array_size)
3136 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3139 static void kmem_cache_destroy_work_func(struct work_struct *w);
3141 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3143 struct memcg_cache_params *cur_params = s->memcg_params;
3145 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3147 if (num_groups > memcg_limited_groups_array_size) {
3149 ssize_t size = memcg_caches_array_size(num_groups);
3151 size *= sizeof(void *);
3152 size += offsetof(struct memcg_cache_params, memcg_caches);
3154 s->memcg_params = kzalloc(size, GFP_KERNEL);
3155 if (!s->memcg_params) {
3156 s->memcg_params = cur_params;
3160 s->memcg_params->is_root_cache = true;
3163 * There is the chance it will be bigger than
3164 * memcg_limited_groups_array_size, if we failed an allocation
3165 * in a cache, in which case all caches updated before it, will
3166 * have a bigger array.
3168 * But if that is the case, the data after
3169 * memcg_limited_groups_array_size is certainly unused
3171 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3172 if (!cur_params->memcg_caches[i])
3174 s->memcg_params->memcg_caches[i] =
3175 cur_params->memcg_caches[i];
3179 * Ideally, we would wait until all caches succeed, and only
3180 * then free the old one. But this is not worth the extra
3181 * pointer per-cache we'd have to have for this.
3183 * It is not a big deal if some caches are left with a size
3184 * bigger than the others. And all updates will reset this
3192 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3193 struct kmem_cache *root_cache)
3197 if (!memcg_kmem_enabled())
3201 size = offsetof(struct memcg_cache_params, memcg_caches);
3202 size += memcg_limited_groups_array_size * sizeof(void *);
3204 size = sizeof(struct memcg_cache_params);
3206 s->memcg_params = kzalloc(size, GFP_KERNEL);
3207 if (!s->memcg_params)
3211 s->memcg_params->memcg = memcg;
3212 s->memcg_params->root_cache = root_cache;
3213 INIT_WORK(&s->memcg_params->destroy,
3214 kmem_cache_destroy_work_func);
3216 s->memcg_params->is_root_cache = true;
3221 void memcg_release_cache(struct kmem_cache *s)
3223 struct kmem_cache *root;
3224 struct mem_cgroup *memcg;
3228 * This happens, for instance, when a root cache goes away before we
3231 if (!s->memcg_params)
3234 if (s->memcg_params->is_root_cache)
3237 memcg = s->memcg_params->memcg;
3238 id = memcg_cache_id(memcg);
3240 root = s->memcg_params->root_cache;
3241 root->memcg_params->memcg_caches[id] = NULL;
3243 mutex_lock(&memcg->slab_caches_mutex);
3244 list_del(&s->memcg_params->list);
3245 mutex_unlock(&memcg->slab_caches_mutex);
3247 css_put(&memcg->css);
3249 kfree(s->memcg_params);
3253 * During the creation a new cache, we need to disable our accounting mechanism
3254 * altogether. This is true even if we are not creating, but rather just
3255 * enqueing new caches to be created.
3257 * This is because that process will trigger allocations; some visible, like
3258 * explicit kmallocs to auxiliary data structures, name strings and internal
3259 * cache structures; some well concealed, like INIT_WORK() that can allocate
3260 * objects during debug.
3262 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3263 * to it. This may not be a bounded recursion: since the first cache creation
3264 * failed to complete (waiting on the allocation), we'll just try to create the
3265 * cache again, failing at the same point.
3267 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3268 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3269 * inside the following two functions.
3271 static inline void memcg_stop_kmem_account(void)
3273 VM_BUG_ON(!current->mm);
3274 current->memcg_kmem_skip_account++;
3277 static inline void memcg_resume_kmem_account(void)
3279 VM_BUG_ON(!current->mm);
3280 current->memcg_kmem_skip_account--;
3283 static void kmem_cache_destroy_work_func(struct work_struct *w)
3285 struct kmem_cache *cachep;
3286 struct memcg_cache_params *p;
3288 p = container_of(w, struct memcg_cache_params, destroy);
3290 cachep = memcg_params_to_cache(p);
3293 * If we get down to 0 after shrink, we could delete right away.
3294 * However, memcg_release_pages() already puts us back in the workqueue
3295 * in that case. If we proceed deleting, we'll get a dangling
3296 * reference, and removing the object from the workqueue in that case
3297 * is unnecessary complication. We are not a fast path.
3299 * Note that this case is fundamentally different from racing with
3300 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3301 * kmem_cache_shrink, not only we would be reinserting a dead cache
3302 * into the queue, but doing so from inside the worker racing to
3305 * So if we aren't down to zero, we'll just schedule a worker and try
3308 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3309 kmem_cache_shrink(cachep);
3310 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3313 kmem_cache_destroy(cachep);
3316 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3318 if (!cachep->memcg_params->dead)
3322 * There are many ways in which we can get here.
3324 * We can get to a memory-pressure situation while the delayed work is
3325 * still pending to run. The vmscan shrinkers can then release all
3326 * cache memory and get us to destruction. If this is the case, we'll
3327 * be executed twice, which is a bug (the second time will execute over
3328 * bogus data). In this case, cancelling the work should be fine.
3330 * But we can also get here from the worker itself, if
3331 * kmem_cache_shrink is enough to shake all the remaining objects and
3332 * get the page count to 0. In this case, we'll deadlock if we try to
3333 * cancel the work (the worker runs with an internal lock held, which
3334 * is the same lock we would hold for cancel_work_sync().)
3336 * Since we can't possibly know who got us here, just refrain from
3337 * running if there is already work pending
3339 if (work_pending(&cachep->memcg_params->destroy))
3342 * We have to defer the actual destroying to a workqueue, because
3343 * we might currently be in a context that cannot sleep.
3345 schedule_work(&cachep->memcg_params->destroy);
3349 * This lock protects updaters, not readers. We want readers to be as fast as
3350 * they can, and they will either see NULL or a valid cache value. Our model
3351 * allow them to see NULL, in which case the root memcg will be selected.
3353 * We need this lock because multiple allocations to the same cache from a non
3354 * will span more than one worker. Only one of them can create the cache.
3356 static DEFINE_MUTEX(memcg_cache_mutex);
3359 * Called with memcg_cache_mutex held
3361 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3362 struct kmem_cache *s)
3364 struct kmem_cache *new;
3365 static char *tmp_name = NULL;
3367 lockdep_assert_held(&memcg_cache_mutex);
3370 * kmem_cache_create_memcg duplicates the given name and
3371 * cgroup_name for this name requires RCU context.
3372 * This static temporary buffer is used to prevent from
3373 * pointless shortliving allocation.
3376 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3382 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3383 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3386 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3387 (s->flags & ~SLAB_PANIC), s->ctor, s);
3390 new->allocflags |= __GFP_KMEMCG;
3395 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3396 struct kmem_cache *cachep)
3398 struct kmem_cache *new_cachep;
3401 BUG_ON(!memcg_can_account_kmem(memcg));
3403 idx = memcg_cache_id(memcg);
3405 mutex_lock(&memcg_cache_mutex);
3406 new_cachep = cachep->memcg_params->memcg_caches[idx];
3408 css_put(&memcg->css);
3412 new_cachep = kmem_cache_dup(memcg, cachep);
3413 if (new_cachep == NULL) {
3414 new_cachep = cachep;
3415 css_put(&memcg->css);
3419 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3421 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3423 * the readers won't lock, make sure everybody sees the updated value,
3424 * so they won't put stuff in the queue again for no reason
3428 mutex_unlock(&memcg_cache_mutex);
3432 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3434 struct kmem_cache *c;
3437 if (!s->memcg_params)
3439 if (!s->memcg_params->is_root_cache)
3443 * If the cache is being destroyed, we trust that there is no one else
3444 * requesting objects from it. Even if there are, the sanity checks in
3445 * kmem_cache_destroy should caught this ill-case.
3447 * Still, we don't want anyone else freeing memcg_caches under our
3448 * noses, which can happen if a new memcg comes to life. As usual,
3449 * we'll take the set_limit_mutex to protect ourselves against this.
3451 mutex_lock(&set_limit_mutex);
3452 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3453 c = s->memcg_params->memcg_caches[i];
3458 * We will now manually delete the caches, so to avoid races
3459 * we need to cancel all pending destruction workers and
3460 * proceed with destruction ourselves.
3462 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3463 * and that could spawn the workers again: it is likely that
3464 * the cache still have active pages until this very moment.
3465 * This would lead us back to mem_cgroup_destroy_cache.
3467 * But that will not execute at all if the "dead" flag is not
3468 * set, so flip it down to guarantee we are in control.
3470 c->memcg_params->dead = false;
3471 cancel_work_sync(&c->memcg_params->destroy);
3472 kmem_cache_destroy(c);
3474 mutex_unlock(&set_limit_mutex);
3477 struct create_work {
3478 struct mem_cgroup *memcg;
3479 struct kmem_cache *cachep;
3480 struct work_struct work;
3483 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3485 struct kmem_cache *cachep;
3486 struct memcg_cache_params *params;
3488 if (!memcg_kmem_is_active(memcg))
3491 mutex_lock(&memcg->slab_caches_mutex);
3492 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3493 cachep = memcg_params_to_cache(params);
3494 cachep->memcg_params->dead = true;
3495 schedule_work(&cachep->memcg_params->destroy);
3497 mutex_unlock(&memcg->slab_caches_mutex);
3500 static void memcg_create_cache_work_func(struct work_struct *w)
3502 struct create_work *cw;
3504 cw = container_of(w, struct create_work, work);
3505 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3510 * Enqueue the creation of a per-memcg kmem_cache.
3512 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3513 struct kmem_cache *cachep)
3515 struct create_work *cw;
3517 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3519 css_put(&memcg->css);
3524 cw->cachep = cachep;
3526 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3527 schedule_work(&cw->work);
3530 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3531 struct kmem_cache *cachep)
3534 * We need to stop accounting when we kmalloc, because if the
3535 * corresponding kmalloc cache is not yet created, the first allocation
3536 * in __memcg_create_cache_enqueue will recurse.
3538 * However, it is better to enclose the whole function. Depending on
3539 * the debugging options enabled, INIT_WORK(), for instance, can
3540 * trigger an allocation. This too, will make us recurse. Because at
3541 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3542 * the safest choice is to do it like this, wrapping the whole function.
3544 memcg_stop_kmem_account();
3545 __memcg_create_cache_enqueue(memcg, cachep);
3546 memcg_resume_kmem_account();
3549 * Return the kmem_cache we're supposed to use for a slab allocation.
3550 * We try to use the current memcg's version of the cache.
3552 * If the cache does not exist yet, if we are the first user of it,
3553 * we either create it immediately, if possible, or create it asynchronously
3555 * In the latter case, we will let the current allocation go through with
3556 * the original cache.
3558 * Can't be called in interrupt context or from kernel threads.
3559 * This function needs to be called with rcu_read_lock() held.
3561 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3564 struct mem_cgroup *memcg;
3567 VM_BUG_ON(!cachep->memcg_params);
3568 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3570 if (!current->mm || current->memcg_kmem_skip_account)
3574 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3576 if (!memcg_can_account_kmem(memcg))
3579 idx = memcg_cache_id(memcg);
3582 * barrier to mare sure we're always seeing the up to date value. The
3583 * code updating memcg_caches will issue a write barrier to match this.
3585 read_barrier_depends();
3586 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3587 cachep = cachep->memcg_params->memcg_caches[idx];
3591 /* The corresponding put will be done in the workqueue. */
3592 if (!css_tryget(&memcg->css))
3597 * If we are in a safe context (can wait, and not in interrupt
3598 * context), we could be be predictable and return right away.
3599 * This would guarantee that the allocation being performed
3600 * already belongs in the new cache.
3602 * However, there are some clashes that can arrive from locking.
3603 * For instance, because we acquire the slab_mutex while doing
3604 * kmem_cache_dup, this means no further allocation could happen
3605 * with the slab_mutex held.
3607 * Also, because cache creation issue get_online_cpus(), this
3608 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3609 * that ends up reversed during cpu hotplug. (cpuset allocates
3610 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3611 * better to defer everything.
3613 memcg_create_cache_enqueue(memcg, cachep);
3619 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3622 * We need to verify if the allocation against current->mm->owner's memcg is
3623 * possible for the given order. But the page is not allocated yet, so we'll
3624 * need a further commit step to do the final arrangements.
3626 * It is possible for the task to switch cgroups in this mean time, so at
3627 * commit time, we can't rely on task conversion any longer. We'll then use
3628 * the handle argument to return to the caller which cgroup we should commit
3629 * against. We could also return the memcg directly and avoid the pointer
3630 * passing, but a boolean return value gives better semantics considering
3631 * the compiled-out case as well.
3633 * Returning true means the allocation is possible.
3636 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3638 struct mem_cgroup *memcg;
3644 * Disabling accounting is only relevant for some specific memcg
3645 * internal allocations. Therefore we would initially not have such
3646 * check here, since direct calls to the page allocator that are marked
3647 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3648 * concerned with cache allocations, and by having this test at
3649 * memcg_kmem_get_cache, we are already able to relay the allocation to
3650 * the root cache and bypass the memcg cache altogether.
3652 * There is one exception, though: the SLUB allocator does not create
3653 * large order caches, but rather service large kmallocs directly from
3654 * the page allocator. Therefore, the following sequence when backed by
3655 * the SLUB allocator:
3657 * memcg_stop_kmem_account();
3658 * kmalloc(<large_number>)
3659 * memcg_resume_kmem_account();
3661 * would effectively ignore the fact that we should skip accounting,
3662 * since it will drive us directly to this function without passing
3663 * through the cache selector memcg_kmem_get_cache. Such large
3664 * allocations are extremely rare but can happen, for instance, for the
3665 * cache arrays. We bring this test here.
3667 if (!current->mm || current->memcg_kmem_skip_account)
3670 memcg = try_get_mem_cgroup_from_mm(current->mm);
3673 * very rare case described in mem_cgroup_from_task. Unfortunately there
3674 * isn't much we can do without complicating this too much, and it would
3675 * be gfp-dependent anyway. Just let it go
3677 if (unlikely(!memcg))
3680 if (!memcg_can_account_kmem(memcg)) {
3681 css_put(&memcg->css);
3685 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3689 css_put(&memcg->css);
3693 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3696 struct page_cgroup *pc;
3698 VM_BUG_ON(mem_cgroup_is_root(memcg));
3700 /* The page allocation failed. Revert */
3702 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3706 pc = lookup_page_cgroup(page);
3707 lock_page_cgroup(pc);
3708 pc->mem_cgroup = memcg;
3709 SetPageCgroupUsed(pc);
3710 unlock_page_cgroup(pc);
3713 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3715 struct mem_cgroup *memcg = NULL;
3716 struct page_cgroup *pc;
3719 pc = lookup_page_cgroup(page);
3721 * Fast unlocked return. Theoretically might have changed, have to
3722 * check again after locking.
3724 if (!PageCgroupUsed(pc))
3727 lock_page_cgroup(pc);
3728 if (PageCgroupUsed(pc)) {
3729 memcg = pc->mem_cgroup;
3730 ClearPageCgroupUsed(pc);
3732 unlock_page_cgroup(pc);
3735 * We trust that only if there is a memcg associated with the page, it
3736 * is a valid allocation
3741 VM_BUG_ON(mem_cgroup_is_root(memcg));
3742 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3745 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3748 #endif /* CONFIG_MEMCG_KMEM */
3750 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3752 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3754 * Because tail pages are not marked as "used", set it. We're under
3755 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3756 * charge/uncharge will be never happen and move_account() is done under
3757 * compound_lock(), so we don't have to take care of races.
3759 void mem_cgroup_split_huge_fixup(struct page *head)
3761 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3762 struct page_cgroup *pc;
3763 struct mem_cgroup *memcg;
3766 if (mem_cgroup_disabled())
3769 memcg = head_pc->mem_cgroup;
3770 for (i = 1; i < HPAGE_PMD_NR; i++) {
3772 pc->mem_cgroup = memcg;
3773 smp_wmb();/* see __commit_charge() */
3774 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3776 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3779 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3782 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3783 struct mem_cgroup *to,
3784 unsigned int nr_pages,
3785 enum mem_cgroup_stat_index idx)
3787 /* Update stat data for mem_cgroup */
3789 WARN_ON_ONCE(from->stat->count[idx] < nr_pages);
3790 __this_cpu_add(from->stat->count[idx], -nr_pages);
3791 __this_cpu_add(to->stat->count[idx], nr_pages);
3796 * mem_cgroup_move_account - move account of the page
3798 * @nr_pages: number of regular pages (>1 for huge pages)
3799 * @pc: page_cgroup of the page.
3800 * @from: mem_cgroup which the page is moved from.
3801 * @to: mem_cgroup which the page is moved to. @from != @to.
3803 * The caller must confirm following.
3804 * - page is not on LRU (isolate_page() is useful.)
3805 * - compound_lock is held when nr_pages > 1
3807 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3810 static int mem_cgroup_move_account(struct page *page,
3811 unsigned int nr_pages,
3812 struct page_cgroup *pc,
3813 struct mem_cgroup *from,
3814 struct mem_cgroup *to)
3816 unsigned long flags;
3818 bool anon = PageAnon(page);
3820 VM_BUG_ON(from == to);
3821 VM_BUG_ON(PageLRU(page));
3823 * The page is isolated from LRU. So, collapse function
3824 * will not handle this page. But page splitting can happen.
3825 * Do this check under compound_page_lock(). The caller should
3829 if (nr_pages > 1 && !PageTransHuge(page))
3832 lock_page_cgroup(pc);
3835 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3838 move_lock_mem_cgroup(from, &flags);
3840 if (!anon && page_mapped(page))
3841 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3842 MEM_CGROUP_STAT_FILE_MAPPED);
3844 if (PageWriteback(page))
3845 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3846 MEM_CGROUP_STAT_WRITEBACK);
3848 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3850 /* caller should have done css_get */
3851 pc->mem_cgroup = to;
3852 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3853 move_unlock_mem_cgroup(from, &flags);
3856 unlock_page_cgroup(pc);
3860 memcg_check_events(to, page);
3861 memcg_check_events(from, page);
3867 * mem_cgroup_move_parent - moves page to the parent group
3868 * @page: the page to move
3869 * @pc: page_cgroup of the page
3870 * @child: page's cgroup
3872 * move charges to its parent or the root cgroup if the group has no
3873 * parent (aka use_hierarchy==0).
3874 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3875 * mem_cgroup_move_account fails) the failure is always temporary and
3876 * it signals a race with a page removal/uncharge or migration. In the
3877 * first case the page is on the way out and it will vanish from the LRU
3878 * on the next attempt and the call should be retried later.
3879 * Isolation from the LRU fails only if page has been isolated from
3880 * the LRU since we looked at it and that usually means either global
3881 * reclaim or migration going on. The page will either get back to the
3883 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3884 * (!PageCgroupUsed) or moved to a different group. The page will
3885 * disappear in the next attempt.
3887 static int mem_cgroup_move_parent(struct page *page,
3888 struct page_cgroup *pc,
3889 struct mem_cgroup *child)
3891 struct mem_cgroup *parent;
3892 unsigned int nr_pages;
3893 unsigned long uninitialized_var(flags);
3896 VM_BUG_ON(mem_cgroup_is_root(child));
3899 if (!get_page_unless_zero(page))
3901 if (isolate_lru_page(page))
3904 nr_pages = hpage_nr_pages(page);
3906 parent = parent_mem_cgroup(child);
3908 * If no parent, move charges to root cgroup.
3911 parent = root_mem_cgroup;
3914 VM_BUG_ON(!PageTransHuge(page));
3915 flags = compound_lock_irqsave(page);
3918 ret = mem_cgroup_move_account(page, nr_pages,
3921 __mem_cgroup_cancel_local_charge(child, nr_pages);
3924 compound_unlock_irqrestore(page, flags);
3925 putback_lru_page(page);
3933 * Charge the memory controller for page usage.
3935 * 0 if the charge was successful
3936 * < 0 if the cgroup is over its limit
3938 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3939 gfp_t gfp_mask, enum charge_type ctype)
3941 struct mem_cgroup *memcg = NULL;
3942 unsigned int nr_pages = 1;
3946 if (PageTransHuge(page)) {
3947 nr_pages <<= compound_order(page);
3948 VM_BUG_ON(!PageTransHuge(page));
3950 * Never OOM-kill a process for a huge page. The
3951 * fault handler will fall back to regular pages.
3956 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3959 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3963 int mem_cgroup_newpage_charge(struct page *page,
3964 struct mm_struct *mm, gfp_t gfp_mask)
3966 if (mem_cgroup_disabled())
3968 VM_BUG_ON(page_mapped(page));
3969 VM_BUG_ON(page->mapping && !PageAnon(page));
3971 return mem_cgroup_charge_common(page, mm, gfp_mask,
3972 MEM_CGROUP_CHARGE_TYPE_ANON);
3976 * While swap-in, try_charge -> commit or cancel, the page is locked.
3977 * And when try_charge() successfully returns, one refcnt to memcg without
3978 * struct page_cgroup is acquired. This refcnt will be consumed by
3979 * "commit()" or removed by "cancel()"
3981 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3984 struct mem_cgroup **memcgp)
3986 struct mem_cgroup *memcg;
3987 struct page_cgroup *pc;
3990 pc = lookup_page_cgroup(page);
3992 * Every swap fault against a single page tries to charge the
3993 * page, bail as early as possible. shmem_unuse() encounters
3994 * already charged pages, too. The USED bit is protected by
3995 * the page lock, which serializes swap cache removal, which
3996 * in turn serializes uncharging.
3998 if (PageCgroupUsed(pc))
4000 if (!do_swap_account)
4002 memcg = try_get_mem_cgroup_from_page(page);
4006 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4007 css_put(&memcg->css);
4012 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4018 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4019 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4022 if (mem_cgroup_disabled())
4025 * A racing thread's fault, or swapoff, may have already
4026 * updated the pte, and even removed page from swap cache: in
4027 * those cases unuse_pte()'s pte_same() test will fail; but
4028 * there's also a KSM case which does need to charge the page.
4030 if (!PageSwapCache(page)) {
4033 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4038 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4041 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4043 if (mem_cgroup_disabled())
4047 __mem_cgroup_cancel_charge(memcg, 1);
4051 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4052 enum charge_type ctype)
4054 if (mem_cgroup_disabled())
4059 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4061 * Now swap is on-memory. This means this page may be
4062 * counted both as mem and swap....double count.
4063 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4064 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4065 * may call delete_from_swap_cache() before reach here.
4067 if (do_swap_account && PageSwapCache(page)) {
4068 swp_entry_t ent = {.val = page_private(page)};
4069 mem_cgroup_uncharge_swap(ent);
4073 void mem_cgroup_commit_charge_swapin(struct page *page,
4074 struct mem_cgroup *memcg)
4076 __mem_cgroup_commit_charge_swapin(page, memcg,
4077 MEM_CGROUP_CHARGE_TYPE_ANON);
4080 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4083 struct mem_cgroup *memcg = NULL;
4084 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4087 if (mem_cgroup_disabled())
4089 if (PageCompound(page))
4092 if (!PageSwapCache(page))
4093 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4094 else { /* page is swapcache/shmem */
4095 ret = __mem_cgroup_try_charge_swapin(mm, page,
4098 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4103 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4104 unsigned int nr_pages,
4105 const enum charge_type ctype)
4107 struct memcg_batch_info *batch = NULL;
4108 bool uncharge_memsw = true;
4110 /* If swapout, usage of swap doesn't decrease */
4111 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4112 uncharge_memsw = false;
4114 batch = ¤t->memcg_batch;
4116 * In usual, we do css_get() when we remember memcg pointer.
4117 * But in this case, we keep res->usage until end of a series of
4118 * uncharges. Then, it's ok to ignore memcg's refcnt.
4121 batch->memcg = memcg;
4123 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4124 * In those cases, all pages freed continuously can be expected to be in
4125 * the same cgroup and we have chance to coalesce uncharges.
4126 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4127 * because we want to do uncharge as soon as possible.
4130 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4131 goto direct_uncharge;
4134 goto direct_uncharge;
4137 * In typical case, batch->memcg == mem. This means we can
4138 * merge a series of uncharges to an uncharge of res_counter.
4139 * If not, we uncharge res_counter ony by one.
4141 if (batch->memcg != memcg)
4142 goto direct_uncharge;
4143 /* remember freed charge and uncharge it later */
4146 batch->memsw_nr_pages++;
4149 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4151 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4152 if (unlikely(batch->memcg != memcg))
4153 memcg_oom_recover(memcg);
4157 * uncharge if !page_mapped(page)
4159 static struct mem_cgroup *
4160 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4163 struct mem_cgroup *memcg = NULL;
4164 unsigned int nr_pages = 1;
4165 struct page_cgroup *pc;
4168 if (mem_cgroup_disabled())
4171 if (PageTransHuge(page)) {
4172 nr_pages <<= compound_order(page);
4173 VM_BUG_ON(!PageTransHuge(page));
4176 * Check if our page_cgroup is valid
4178 pc = lookup_page_cgroup(page);
4179 if (unlikely(!PageCgroupUsed(pc)))
4182 lock_page_cgroup(pc);
4184 memcg = pc->mem_cgroup;
4186 if (!PageCgroupUsed(pc))
4189 anon = PageAnon(page);
4192 case MEM_CGROUP_CHARGE_TYPE_ANON:
4194 * Generally PageAnon tells if it's the anon statistics to be
4195 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4196 * used before page reached the stage of being marked PageAnon.
4200 case MEM_CGROUP_CHARGE_TYPE_DROP:
4201 /* See mem_cgroup_prepare_migration() */
4202 if (page_mapped(page))
4205 * Pages under migration may not be uncharged. But
4206 * end_migration() /must/ be the one uncharging the
4207 * unused post-migration page and so it has to call
4208 * here with the migration bit still set. See the
4209 * res_counter handling below.
4211 if (!end_migration && PageCgroupMigration(pc))
4214 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4215 if (!PageAnon(page)) { /* Shared memory */
4216 if (page->mapping && !page_is_file_cache(page))
4218 } else if (page_mapped(page)) /* Anon */
4225 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4227 ClearPageCgroupUsed(pc);
4229 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4230 * freed from LRU. This is safe because uncharged page is expected not
4231 * to be reused (freed soon). Exception is SwapCache, it's handled by
4232 * special functions.
4235 unlock_page_cgroup(pc);
4237 * even after unlock, we have memcg->res.usage here and this memcg
4238 * will never be freed, so it's safe to call css_get().
4240 memcg_check_events(memcg, page);
4241 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4242 mem_cgroup_swap_statistics(memcg, true);
4243 css_get(&memcg->css);
4246 * Migration does not charge the res_counter for the
4247 * replacement page, so leave it alone when phasing out the
4248 * page that is unused after the migration.
4250 if (!end_migration && !mem_cgroup_is_root(memcg))
4251 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4256 unlock_page_cgroup(pc);
4260 void mem_cgroup_uncharge_page(struct page *page)
4263 if (page_mapped(page))
4265 VM_BUG_ON(page->mapping && !PageAnon(page));
4267 * If the page is in swap cache, uncharge should be deferred
4268 * to the swap path, which also properly accounts swap usage
4269 * and handles memcg lifetime.
4271 * Note that this check is not stable and reclaim may add the
4272 * page to swap cache at any time after this. However, if the
4273 * page is not in swap cache by the time page->mapcount hits
4274 * 0, there won't be any page table references to the swap
4275 * slot, and reclaim will free it and not actually write the
4278 if (PageSwapCache(page))
4280 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4283 void mem_cgroup_uncharge_cache_page(struct page *page)
4285 VM_BUG_ON(page_mapped(page));
4286 VM_BUG_ON(page->mapping);
4287 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4291 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4292 * In that cases, pages are freed continuously and we can expect pages
4293 * are in the same memcg. All these calls itself limits the number of
4294 * pages freed at once, then uncharge_start/end() is called properly.
4295 * This may be called prural(2) times in a context,
4298 void mem_cgroup_uncharge_start(void)
4300 current->memcg_batch.do_batch++;
4301 /* We can do nest. */
4302 if (current->memcg_batch.do_batch == 1) {
4303 current->memcg_batch.memcg = NULL;
4304 current->memcg_batch.nr_pages = 0;
4305 current->memcg_batch.memsw_nr_pages = 0;
4309 void mem_cgroup_uncharge_end(void)
4311 struct memcg_batch_info *batch = ¤t->memcg_batch;
4313 if (!batch->do_batch)
4317 if (batch->do_batch) /* If stacked, do nothing. */
4323 * This "batch->memcg" is valid without any css_get/put etc...
4324 * bacause we hide charges behind us.
4326 if (batch->nr_pages)
4327 res_counter_uncharge(&batch->memcg->res,
4328 batch->nr_pages * PAGE_SIZE);
4329 if (batch->memsw_nr_pages)
4330 res_counter_uncharge(&batch->memcg->memsw,
4331 batch->memsw_nr_pages * PAGE_SIZE);
4332 memcg_oom_recover(batch->memcg);
4333 /* forget this pointer (for sanity check) */
4334 batch->memcg = NULL;
4339 * called after __delete_from_swap_cache() and drop "page" account.
4340 * memcg information is recorded to swap_cgroup of "ent"
4343 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4345 struct mem_cgroup *memcg;
4346 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4348 if (!swapout) /* this was a swap cache but the swap is unused ! */
4349 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4351 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4354 * record memcg information, if swapout && memcg != NULL,
4355 * css_get() was called in uncharge().
4357 if (do_swap_account && swapout && memcg)
4358 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4362 #ifdef CONFIG_MEMCG_SWAP
4364 * called from swap_entry_free(). remove record in swap_cgroup and
4365 * uncharge "memsw" account.
4367 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4369 struct mem_cgroup *memcg;
4372 if (!do_swap_account)
4375 id = swap_cgroup_record(ent, 0);
4377 memcg = mem_cgroup_lookup(id);
4380 * We uncharge this because swap is freed.
4381 * This memcg can be obsolete one. We avoid calling css_tryget
4383 if (!mem_cgroup_is_root(memcg))
4384 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4385 mem_cgroup_swap_statistics(memcg, false);
4386 css_put(&memcg->css);
4392 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4393 * @entry: swap entry to be moved
4394 * @from: mem_cgroup which the entry is moved from
4395 * @to: mem_cgroup which the entry is moved to
4397 * It succeeds only when the swap_cgroup's record for this entry is the same
4398 * as the mem_cgroup's id of @from.
4400 * Returns 0 on success, -EINVAL on failure.
4402 * The caller must have charged to @to, IOW, called res_counter_charge() about
4403 * both res and memsw, and called css_get().
4405 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4406 struct mem_cgroup *from, struct mem_cgroup *to)
4408 unsigned short old_id, new_id;
4410 old_id = mem_cgroup_id(from);
4411 new_id = mem_cgroup_id(to);
4413 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4414 mem_cgroup_swap_statistics(from, false);
4415 mem_cgroup_swap_statistics(to, true);
4417 * This function is only called from task migration context now.
4418 * It postpones res_counter and refcount handling till the end
4419 * of task migration(mem_cgroup_clear_mc()) for performance
4420 * improvement. But we cannot postpone css_get(to) because if
4421 * the process that has been moved to @to does swap-in, the
4422 * refcount of @to might be decreased to 0.
4424 * We are in attach() phase, so the cgroup is guaranteed to be
4425 * alive, so we can just call css_get().
4433 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4434 struct mem_cgroup *from, struct mem_cgroup *to)
4441 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4444 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4445 struct mem_cgroup **memcgp)
4447 struct mem_cgroup *memcg = NULL;
4448 unsigned int nr_pages = 1;
4449 struct page_cgroup *pc;
4450 enum charge_type ctype;
4454 if (mem_cgroup_disabled())
4457 if (PageTransHuge(page))
4458 nr_pages <<= compound_order(page);
4460 pc = lookup_page_cgroup(page);
4461 lock_page_cgroup(pc);
4462 if (PageCgroupUsed(pc)) {
4463 memcg = pc->mem_cgroup;
4464 css_get(&memcg->css);
4466 * At migrating an anonymous page, its mapcount goes down
4467 * to 0 and uncharge() will be called. But, even if it's fully
4468 * unmapped, migration may fail and this page has to be
4469 * charged again. We set MIGRATION flag here and delay uncharge
4470 * until end_migration() is called
4472 * Corner Case Thinking
4474 * When the old page was mapped as Anon and it's unmap-and-freed
4475 * while migration was ongoing.
4476 * If unmap finds the old page, uncharge() of it will be delayed
4477 * until end_migration(). If unmap finds a new page, it's
4478 * uncharged when it make mapcount to be 1->0. If unmap code
4479 * finds swap_migration_entry, the new page will not be mapped
4480 * and end_migration() will find it(mapcount==0).
4483 * When the old page was mapped but migraion fails, the kernel
4484 * remaps it. A charge for it is kept by MIGRATION flag even
4485 * if mapcount goes down to 0. We can do remap successfully
4486 * without charging it again.
4489 * The "old" page is under lock_page() until the end of
4490 * migration, so, the old page itself will not be swapped-out.
4491 * If the new page is swapped out before end_migraton, our
4492 * hook to usual swap-out path will catch the event.
4495 SetPageCgroupMigration(pc);
4497 unlock_page_cgroup(pc);
4499 * If the page is not charged at this point,
4507 * We charge new page before it's used/mapped. So, even if unlock_page()
4508 * is called before end_migration, we can catch all events on this new
4509 * page. In the case new page is migrated but not remapped, new page's
4510 * mapcount will be finally 0 and we call uncharge in end_migration().
4513 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4515 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4517 * The page is committed to the memcg, but it's not actually
4518 * charged to the res_counter since we plan on replacing the
4519 * old one and only one page is going to be left afterwards.
4521 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4524 /* remove redundant charge if migration failed*/
4525 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4526 struct page *oldpage, struct page *newpage, bool migration_ok)
4528 struct page *used, *unused;
4529 struct page_cgroup *pc;
4535 if (!migration_ok) {
4542 anon = PageAnon(used);
4543 __mem_cgroup_uncharge_common(unused,
4544 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4545 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4547 css_put(&memcg->css);
4549 * We disallowed uncharge of pages under migration because mapcount
4550 * of the page goes down to zero, temporarly.
4551 * Clear the flag and check the page should be charged.
4553 pc = lookup_page_cgroup(oldpage);
4554 lock_page_cgroup(pc);
4555 ClearPageCgroupMigration(pc);
4556 unlock_page_cgroup(pc);
4559 * If a page is a file cache, radix-tree replacement is very atomic
4560 * and we can skip this check. When it was an Anon page, its mapcount
4561 * goes down to 0. But because we added MIGRATION flage, it's not
4562 * uncharged yet. There are several case but page->mapcount check
4563 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4564 * check. (see prepare_charge() also)
4567 mem_cgroup_uncharge_page(used);
4571 * At replace page cache, newpage is not under any memcg but it's on
4572 * LRU. So, this function doesn't touch res_counter but handles LRU
4573 * in correct way. Both pages are locked so we cannot race with uncharge.
4575 void mem_cgroup_replace_page_cache(struct page *oldpage,
4576 struct page *newpage)
4578 struct mem_cgroup *memcg = NULL;
4579 struct page_cgroup *pc;
4580 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4582 if (mem_cgroup_disabled())
4585 pc = lookup_page_cgroup(oldpage);
4586 /* fix accounting on old pages */
4587 lock_page_cgroup(pc);
4588 if (PageCgroupUsed(pc)) {
4589 memcg = pc->mem_cgroup;
4590 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4591 ClearPageCgroupUsed(pc);
4593 unlock_page_cgroup(pc);
4596 * When called from shmem_replace_page(), in some cases the
4597 * oldpage has already been charged, and in some cases not.
4602 * Even if newpage->mapping was NULL before starting replacement,
4603 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4604 * LRU while we overwrite pc->mem_cgroup.
4606 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4609 #ifdef CONFIG_DEBUG_VM
4610 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4612 struct page_cgroup *pc;
4614 pc = lookup_page_cgroup(page);
4616 * Can be NULL while feeding pages into the page allocator for
4617 * the first time, i.e. during boot or memory hotplug;
4618 * or when mem_cgroup_disabled().
4620 if (likely(pc) && PageCgroupUsed(pc))
4625 bool mem_cgroup_bad_page_check(struct page *page)
4627 if (mem_cgroup_disabled())
4630 return lookup_page_cgroup_used(page) != NULL;
4633 void mem_cgroup_print_bad_page(struct page *page)
4635 struct page_cgroup *pc;
4637 pc = lookup_page_cgroup_used(page);
4639 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4640 pc, pc->flags, pc->mem_cgroup);
4645 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4646 unsigned long long val)
4649 u64 memswlimit, memlimit;
4651 int children = mem_cgroup_count_children(memcg);
4652 u64 curusage, oldusage;
4656 * For keeping hierarchical_reclaim simple, how long we should retry
4657 * is depends on callers. We set our retry-count to be function
4658 * of # of children which we should visit in this loop.
4660 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4662 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4665 while (retry_count) {
4666 if (signal_pending(current)) {
4671 * Rather than hide all in some function, I do this in
4672 * open coded manner. You see what this really does.
4673 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4675 mutex_lock(&set_limit_mutex);
4676 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4677 if (memswlimit < val) {
4679 mutex_unlock(&set_limit_mutex);
4683 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4687 ret = res_counter_set_limit(&memcg->res, val);
4689 if (memswlimit == val)
4690 memcg->memsw_is_minimum = true;
4692 memcg->memsw_is_minimum = false;
4694 mutex_unlock(&set_limit_mutex);
4699 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4700 MEM_CGROUP_RECLAIM_SHRINK);
4701 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4702 /* Usage is reduced ? */
4703 if (curusage >= oldusage)
4706 oldusage = curusage;
4708 if (!ret && enlarge)
4709 memcg_oom_recover(memcg);
4714 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4715 unsigned long long val)
4718 u64 memlimit, memswlimit, oldusage, curusage;
4719 int children = mem_cgroup_count_children(memcg);
4723 /* see mem_cgroup_resize_res_limit */
4724 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4725 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4726 while (retry_count) {
4727 if (signal_pending(current)) {
4732 * Rather than hide all in some function, I do this in
4733 * open coded manner. You see what this really does.
4734 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4736 mutex_lock(&set_limit_mutex);
4737 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4738 if (memlimit > val) {
4740 mutex_unlock(&set_limit_mutex);
4743 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4744 if (memswlimit < val)
4746 ret = res_counter_set_limit(&memcg->memsw, val);
4748 if (memlimit == val)
4749 memcg->memsw_is_minimum = true;
4751 memcg->memsw_is_minimum = false;
4753 mutex_unlock(&set_limit_mutex);
4758 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4759 MEM_CGROUP_RECLAIM_NOSWAP |
4760 MEM_CGROUP_RECLAIM_SHRINK);
4761 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4762 /* Usage is reduced ? */
4763 if (curusage >= oldusage)
4766 oldusage = curusage;
4768 if (!ret && enlarge)
4769 memcg_oom_recover(memcg);
4773 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4775 unsigned long *total_scanned)
4777 unsigned long nr_reclaimed = 0;
4778 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4779 unsigned long reclaimed;
4781 struct mem_cgroup_tree_per_zone *mctz;
4782 unsigned long long excess;
4783 unsigned long nr_scanned;
4788 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4790 * This loop can run a while, specially if mem_cgroup's continuously
4791 * keep exceeding their soft limit and putting the system under
4798 mz = mem_cgroup_largest_soft_limit_node(mctz);
4803 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4804 gfp_mask, &nr_scanned);
4805 nr_reclaimed += reclaimed;
4806 *total_scanned += nr_scanned;
4807 spin_lock(&mctz->lock);
4810 * If we failed to reclaim anything from this memory cgroup
4811 * it is time to move on to the next cgroup
4817 * Loop until we find yet another one.
4819 * By the time we get the soft_limit lock
4820 * again, someone might have aded the
4821 * group back on the RB tree. Iterate to
4822 * make sure we get a different mem.
4823 * mem_cgroup_largest_soft_limit_node returns
4824 * NULL if no other cgroup is present on
4828 __mem_cgroup_largest_soft_limit_node(mctz);
4830 css_put(&next_mz->memcg->css);
4831 else /* next_mz == NULL or other memcg */
4835 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4836 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4838 * One school of thought says that we should not add
4839 * back the node to the tree if reclaim returns 0.
4840 * But our reclaim could return 0, simply because due
4841 * to priority we are exposing a smaller subset of
4842 * memory to reclaim from. Consider this as a longer
4845 /* If excess == 0, no tree ops */
4846 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4847 spin_unlock(&mctz->lock);
4848 css_put(&mz->memcg->css);
4851 * Could not reclaim anything and there are no more
4852 * mem cgroups to try or we seem to be looping without
4853 * reclaiming anything.
4855 if (!nr_reclaimed &&
4857 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4859 } while (!nr_reclaimed);
4861 css_put(&next_mz->memcg->css);
4862 return nr_reclaimed;
4866 * mem_cgroup_force_empty_list - clears LRU of a group
4867 * @memcg: group to clear
4870 * @lru: lru to to clear
4872 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4873 * reclaim the pages page themselves - pages are moved to the parent (or root)
4876 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4877 int node, int zid, enum lru_list lru)
4879 struct lruvec *lruvec;
4880 unsigned long flags;
4881 struct list_head *list;
4885 zone = &NODE_DATA(node)->node_zones[zid];
4886 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4887 list = &lruvec->lists[lru];
4891 struct page_cgroup *pc;
4894 spin_lock_irqsave(&zone->lru_lock, flags);
4895 if (list_empty(list)) {
4896 spin_unlock_irqrestore(&zone->lru_lock, flags);
4899 page = list_entry(list->prev, struct page, lru);
4901 list_move(&page->lru, list);
4903 spin_unlock_irqrestore(&zone->lru_lock, flags);
4906 spin_unlock_irqrestore(&zone->lru_lock, flags);
4908 pc = lookup_page_cgroup(page);
4910 if (mem_cgroup_move_parent(page, pc, memcg)) {
4911 /* found lock contention or "pc" is obsolete. */
4916 } while (!list_empty(list));
4920 * make mem_cgroup's charge to be 0 if there is no task by moving
4921 * all the charges and pages to the parent.
4922 * This enables deleting this mem_cgroup.
4924 * Caller is responsible for holding css reference on the memcg.
4926 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4932 /* This is for making all *used* pages to be on LRU. */
4933 lru_add_drain_all();
4934 drain_all_stock_sync(memcg);
4935 mem_cgroup_start_move(memcg);
4936 for_each_node_state(node, N_MEMORY) {
4937 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4940 mem_cgroup_force_empty_list(memcg,
4945 mem_cgroup_end_move(memcg);
4946 memcg_oom_recover(memcg);
4950 * Kernel memory may not necessarily be trackable to a specific
4951 * process. So they are not migrated, and therefore we can't
4952 * expect their value to drop to 0 here.
4953 * Having res filled up with kmem only is enough.
4955 * This is a safety check because mem_cgroup_force_empty_list
4956 * could have raced with mem_cgroup_replace_page_cache callers
4957 * so the lru seemed empty but the page could have been added
4958 * right after the check. RES_USAGE should be safe as we always
4959 * charge before adding to the LRU.
4961 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4962 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4963 } while (usage > 0);
4967 * This mainly exists for tests during the setting of set of use_hierarchy.
4968 * Since this is the very setting we are changing, the current hierarchy value
4971 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4973 struct cgroup_subsys_state *pos;
4975 /* bounce at first found */
4976 css_for_each_child(pos, &memcg->css)
4982 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4983 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4984 * from mem_cgroup_count_children(), in the sense that we don't really care how
4985 * many children we have; we only need to know if we have any. It also counts
4986 * any memcg without hierarchy as infertile.
4988 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4990 return memcg->use_hierarchy && __memcg_has_children(memcg);
4994 * Reclaims as many pages from the given memcg as possible and moves
4995 * the rest to the parent.
4997 * Caller is responsible for holding css reference for memcg.
4999 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
5001 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
5002 struct cgroup *cgrp = memcg->css.cgroup;
5004 /* returns EBUSY if there is a task or if we come here twice. */
5005 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
5008 /* we call try-to-free pages for make this cgroup empty */
5009 lru_add_drain_all();
5010 /* try to free all pages in this cgroup */
5011 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
5014 if (signal_pending(current))
5017 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5021 /* maybe some writeback is necessary */
5022 congestion_wait(BLK_RW_ASYNC, HZ/10);
5027 mem_cgroup_reparent_charges(memcg);
5032 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
5035 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5037 if (mem_cgroup_is_root(memcg))
5039 return mem_cgroup_force_empty(memcg);
5042 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5045 return mem_cgroup_from_css(css)->use_hierarchy;
5048 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5049 struct cftype *cft, u64 val)
5052 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5053 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5055 mutex_lock(&memcg_create_mutex);
5057 if (memcg->use_hierarchy == val)
5061 * If parent's use_hierarchy is set, we can't make any modifications
5062 * in the child subtrees. If it is unset, then the change can
5063 * occur, provided the current cgroup has no children.
5065 * For the root cgroup, parent_mem is NULL, we allow value to be
5066 * set if there are no children.
5068 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5069 (val == 1 || val == 0)) {
5070 if (!__memcg_has_children(memcg))
5071 memcg->use_hierarchy = val;
5078 mutex_unlock(&memcg_create_mutex);
5084 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5085 enum mem_cgroup_stat_index idx)
5087 struct mem_cgroup *iter;
5090 /* Per-cpu values can be negative, use a signed accumulator */
5091 for_each_mem_cgroup_tree(iter, memcg)
5092 val += mem_cgroup_read_stat(iter, idx);
5094 if (val < 0) /* race ? */
5099 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5103 if (!mem_cgroup_is_root(memcg)) {
5105 return res_counter_read_u64(&memcg->res, RES_USAGE);
5107 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5111 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5112 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5114 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5115 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5118 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5120 return val << PAGE_SHIFT;
5123 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
5124 struct cftype *cft, struct file *file,
5125 char __user *buf, size_t nbytes, loff_t *ppos)
5127 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5133 type = MEMFILE_TYPE(cft->private);
5134 name = MEMFILE_ATTR(cft->private);
5138 if (name == RES_USAGE)
5139 val = mem_cgroup_usage(memcg, false);
5141 val = res_counter_read_u64(&memcg->res, name);
5144 if (name == RES_USAGE)
5145 val = mem_cgroup_usage(memcg, true);
5147 val = res_counter_read_u64(&memcg->memsw, name);
5150 val = res_counter_read_u64(&memcg->kmem, name);
5156 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5157 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5160 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
5163 #ifdef CONFIG_MEMCG_KMEM
5164 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5166 * For simplicity, we won't allow this to be disabled. It also can't
5167 * be changed if the cgroup has children already, or if tasks had
5170 * If tasks join before we set the limit, a person looking at
5171 * kmem.usage_in_bytes will have no way to determine when it took
5172 * place, which makes the value quite meaningless.
5174 * After it first became limited, changes in the value of the limit are
5175 * of course permitted.
5177 mutex_lock(&memcg_create_mutex);
5178 mutex_lock(&set_limit_mutex);
5179 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
5180 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
5184 ret = res_counter_set_limit(&memcg->kmem, val);
5187 ret = memcg_update_cache_sizes(memcg);
5189 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
5192 static_key_slow_inc(&memcg_kmem_enabled_key);
5194 * setting the active bit after the inc will guarantee no one
5195 * starts accounting before all call sites are patched
5197 memcg_kmem_set_active(memcg);
5199 ret = res_counter_set_limit(&memcg->kmem, val);
5201 mutex_unlock(&set_limit_mutex);
5202 mutex_unlock(&memcg_create_mutex);
5207 #ifdef CONFIG_MEMCG_KMEM
5208 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5211 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5215 memcg->kmem_account_flags = parent->kmem_account_flags;
5217 * When that happen, we need to disable the static branch only on those
5218 * memcgs that enabled it. To achieve this, we would be forced to
5219 * complicate the code by keeping track of which memcgs were the ones
5220 * that actually enabled limits, and which ones got it from its
5223 * It is a lot simpler just to do static_key_slow_inc() on every child
5224 * that is accounted.
5226 if (!memcg_kmem_is_active(memcg))
5230 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5231 * memcg is active already. If the later initialization fails then the
5232 * cgroup core triggers the cleanup so we do not have to do it here.
5234 static_key_slow_inc(&memcg_kmem_enabled_key);
5236 mutex_lock(&set_limit_mutex);
5237 memcg_stop_kmem_account();
5238 ret = memcg_update_cache_sizes(memcg);
5239 memcg_resume_kmem_account();
5240 mutex_unlock(&set_limit_mutex);
5244 #endif /* CONFIG_MEMCG_KMEM */
5247 * The user of this function is...
5250 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5253 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5256 unsigned long long val;
5259 type = MEMFILE_TYPE(cft->private);
5260 name = MEMFILE_ATTR(cft->private);
5264 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5268 /* This function does all necessary parse...reuse it */
5269 ret = res_counter_memparse_write_strategy(buffer, &val);
5273 ret = mem_cgroup_resize_limit(memcg, val);
5274 else if (type == _MEMSWAP)
5275 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5276 else if (type == _KMEM)
5277 ret = memcg_update_kmem_limit(css, val);
5281 case RES_SOFT_LIMIT:
5282 ret = res_counter_memparse_write_strategy(buffer, &val);
5286 * For memsw, soft limits are hard to implement in terms
5287 * of semantics, for now, we support soft limits for
5288 * control without swap
5291 ret = res_counter_set_soft_limit(&memcg->res, val);
5296 ret = -EINVAL; /* should be BUG() ? */
5302 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5303 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5305 unsigned long long min_limit, min_memsw_limit, tmp;
5307 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5308 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5309 if (!memcg->use_hierarchy)
5312 while (css_parent(&memcg->css)) {
5313 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5314 if (!memcg->use_hierarchy)
5316 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5317 min_limit = min(min_limit, tmp);
5318 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5319 min_memsw_limit = min(min_memsw_limit, tmp);
5322 *mem_limit = min_limit;
5323 *memsw_limit = min_memsw_limit;
5326 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5328 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5332 type = MEMFILE_TYPE(event);
5333 name = MEMFILE_ATTR(event);
5338 res_counter_reset_max(&memcg->res);
5339 else if (type == _MEMSWAP)
5340 res_counter_reset_max(&memcg->memsw);
5341 else if (type == _KMEM)
5342 res_counter_reset_max(&memcg->kmem);
5348 res_counter_reset_failcnt(&memcg->res);
5349 else if (type == _MEMSWAP)
5350 res_counter_reset_failcnt(&memcg->memsw);
5351 else if (type == _KMEM)
5352 res_counter_reset_failcnt(&memcg->kmem);
5361 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5364 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5368 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5369 struct cftype *cft, u64 val)
5371 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5373 if (val >= (1 << NR_MOVE_TYPE))
5377 * No kind of locking is needed in here, because ->can_attach() will
5378 * check this value once in the beginning of the process, and then carry
5379 * on with stale data. This means that changes to this value will only
5380 * affect task migrations starting after the change.
5382 memcg->move_charge_at_immigrate = val;
5386 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5387 struct cftype *cft, u64 val)
5394 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5395 struct cftype *cft, struct seq_file *m)
5398 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5399 unsigned long node_nr;
5400 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5402 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5403 seq_printf(m, "total=%lu", total_nr);
5404 for_each_node_state(nid, N_MEMORY) {
5405 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5406 seq_printf(m, " N%d=%lu", nid, node_nr);
5410 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5411 seq_printf(m, "file=%lu", file_nr);
5412 for_each_node_state(nid, N_MEMORY) {
5413 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5415 seq_printf(m, " N%d=%lu", nid, node_nr);
5419 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5420 seq_printf(m, "anon=%lu", anon_nr);
5421 for_each_node_state(nid, N_MEMORY) {
5422 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5424 seq_printf(m, " N%d=%lu", nid, node_nr);
5428 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5429 seq_printf(m, "unevictable=%lu", unevictable_nr);
5430 for_each_node_state(nid, N_MEMORY) {
5431 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5432 BIT(LRU_UNEVICTABLE));
5433 seq_printf(m, " N%d=%lu", nid, node_nr);
5438 #endif /* CONFIG_NUMA */
5440 static inline void mem_cgroup_lru_names_not_uptodate(void)
5442 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5445 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5448 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5449 struct mem_cgroup *mi;
5452 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5453 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5455 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5456 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5459 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5460 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5461 mem_cgroup_read_events(memcg, i));
5463 for (i = 0; i < NR_LRU_LISTS; i++)
5464 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5465 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5467 /* Hierarchical information */
5469 unsigned long long limit, memsw_limit;
5470 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5471 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5472 if (do_swap_account)
5473 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5477 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5480 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5482 for_each_mem_cgroup_tree(mi, memcg)
5483 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5484 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5487 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5488 unsigned long long val = 0;
5490 for_each_mem_cgroup_tree(mi, memcg)
5491 val += mem_cgroup_read_events(mi, i);
5492 seq_printf(m, "total_%s %llu\n",
5493 mem_cgroup_events_names[i], val);
5496 for (i = 0; i < NR_LRU_LISTS; i++) {
5497 unsigned long long val = 0;
5499 for_each_mem_cgroup_tree(mi, memcg)
5500 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5501 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5504 #ifdef CONFIG_DEBUG_VM
5507 struct mem_cgroup_per_zone *mz;
5508 struct zone_reclaim_stat *rstat;
5509 unsigned long recent_rotated[2] = {0, 0};
5510 unsigned long recent_scanned[2] = {0, 0};
5512 for_each_online_node(nid)
5513 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5514 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5515 rstat = &mz->lruvec.reclaim_stat;
5517 recent_rotated[0] += rstat->recent_rotated[0];
5518 recent_rotated[1] += rstat->recent_rotated[1];
5519 recent_scanned[0] += rstat->recent_scanned[0];
5520 recent_scanned[1] += rstat->recent_scanned[1];
5522 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5523 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5524 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5525 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5532 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5535 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5537 return mem_cgroup_swappiness(memcg);
5540 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5541 struct cftype *cft, u64 val)
5543 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5544 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5546 if (val > 100 || !parent)
5549 mutex_lock(&memcg_create_mutex);
5551 /* If under hierarchy, only empty-root can set this value */
5552 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5553 mutex_unlock(&memcg_create_mutex);
5557 memcg->swappiness = val;
5559 mutex_unlock(&memcg_create_mutex);
5564 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5566 struct mem_cgroup_threshold_ary *t;
5572 t = rcu_dereference(memcg->thresholds.primary);
5574 t = rcu_dereference(memcg->memsw_thresholds.primary);
5579 usage = mem_cgroup_usage(memcg, swap);
5582 * current_threshold points to threshold just below or equal to usage.
5583 * If it's not true, a threshold was crossed after last
5584 * call of __mem_cgroup_threshold().
5586 i = t->current_threshold;
5589 * Iterate backward over array of thresholds starting from
5590 * current_threshold and check if a threshold is crossed.
5591 * If none of thresholds below usage is crossed, we read
5592 * only one element of the array here.
5594 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5595 eventfd_signal(t->entries[i].eventfd, 1);
5597 /* i = current_threshold + 1 */
5601 * Iterate forward over array of thresholds starting from
5602 * current_threshold+1 and check if a threshold is crossed.
5603 * If none of thresholds above usage is crossed, we read
5604 * only one element of the array here.
5606 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5607 eventfd_signal(t->entries[i].eventfd, 1);
5609 /* Update current_threshold */
5610 t->current_threshold = i - 1;
5615 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5618 __mem_cgroup_threshold(memcg, false);
5619 if (do_swap_account)
5620 __mem_cgroup_threshold(memcg, true);
5622 memcg = parent_mem_cgroup(memcg);
5626 static int compare_thresholds(const void *a, const void *b)
5628 const struct mem_cgroup_threshold *_a = a;
5629 const struct mem_cgroup_threshold *_b = b;
5631 if (_a->threshold > _b->threshold)
5634 if (_a->threshold < _b->threshold)
5640 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5642 struct mem_cgroup_eventfd_list *ev;
5644 list_for_each_entry(ev, &memcg->oom_notify, list)
5645 eventfd_signal(ev->eventfd, 1);
5649 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5651 struct mem_cgroup *iter;
5653 for_each_mem_cgroup_tree(iter, memcg)
5654 mem_cgroup_oom_notify_cb(iter);
5657 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5658 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5660 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5661 struct mem_cgroup_thresholds *thresholds;
5662 struct mem_cgroup_threshold_ary *new;
5663 enum res_type type = MEMFILE_TYPE(cft->private);
5664 u64 threshold, usage;
5667 ret = res_counter_memparse_write_strategy(args, &threshold);
5671 mutex_lock(&memcg->thresholds_lock);
5674 thresholds = &memcg->thresholds;
5675 else if (type == _MEMSWAP)
5676 thresholds = &memcg->memsw_thresholds;
5680 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5682 /* Check if a threshold crossed before adding a new one */
5683 if (thresholds->primary)
5684 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5686 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5688 /* Allocate memory for new array of thresholds */
5689 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5697 /* Copy thresholds (if any) to new array */
5698 if (thresholds->primary) {
5699 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5700 sizeof(struct mem_cgroup_threshold));
5703 /* Add new threshold */
5704 new->entries[size - 1].eventfd = eventfd;
5705 new->entries[size - 1].threshold = threshold;
5707 /* Sort thresholds. Registering of new threshold isn't time-critical */
5708 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5709 compare_thresholds, NULL);
5711 /* Find current threshold */
5712 new->current_threshold = -1;
5713 for (i = 0; i < size; i++) {
5714 if (new->entries[i].threshold <= usage) {
5716 * new->current_threshold will not be used until
5717 * rcu_assign_pointer(), so it's safe to increment
5720 ++new->current_threshold;
5725 /* Free old spare buffer and save old primary buffer as spare */
5726 kfree(thresholds->spare);
5727 thresholds->spare = thresholds->primary;
5729 rcu_assign_pointer(thresholds->primary, new);
5731 /* To be sure that nobody uses thresholds */
5735 mutex_unlock(&memcg->thresholds_lock);
5740 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5741 struct cftype *cft, struct eventfd_ctx *eventfd)
5743 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5744 struct mem_cgroup_thresholds *thresholds;
5745 struct mem_cgroup_threshold_ary *new;
5746 enum res_type type = MEMFILE_TYPE(cft->private);
5750 mutex_lock(&memcg->thresholds_lock);
5752 thresholds = &memcg->thresholds;
5753 else if (type == _MEMSWAP)
5754 thresholds = &memcg->memsw_thresholds;
5758 if (!thresholds->primary)
5761 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5763 /* Check if a threshold crossed before removing */
5764 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5766 /* Calculate new number of threshold */
5768 for (i = 0; i < thresholds->primary->size; i++) {
5769 if (thresholds->primary->entries[i].eventfd != eventfd)
5773 new = thresholds->spare;
5775 /* Set thresholds array to NULL if we don't have thresholds */
5784 /* Copy thresholds and find current threshold */
5785 new->current_threshold = -1;
5786 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5787 if (thresholds->primary->entries[i].eventfd == eventfd)
5790 new->entries[j] = thresholds->primary->entries[i];
5791 if (new->entries[j].threshold <= usage) {
5793 * new->current_threshold will not be used
5794 * until rcu_assign_pointer(), so it's safe to increment
5797 ++new->current_threshold;
5803 /* Swap primary and spare array */
5804 thresholds->spare = thresholds->primary;
5805 /* If all events are unregistered, free the spare array */
5807 kfree(thresholds->spare);
5808 thresholds->spare = NULL;
5811 rcu_assign_pointer(thresholds->primary, new);
5813 /* To be sure that nobody uses thresholds */
5816 mutex_unlock(&memcg->thresholds_lock);
5819 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5820 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5822 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5823 struct mem_cgroup_eventfd_list *event;
5824 enum res_type type = MEMFILE_TYPE(cft->private);
5826 BUG_ON(type != _OOM_TYPE);
5827 event = kmalloc(sizeof(*event), GFP_KERNEL);
5831 spin_lock(&memcg_oom_lock);
5833 event->eventfd = eventfd;
5834 list_add(&event->list, &memcg->oom_notify);
5836 /* already in OOM ? */
5837 if (atomic_read(&memcg->under_oom))
5838 eventfd_signal(eventfd, 1);
5839 spin_unlock(&memcg_oom_lock);
5844 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5845 struct cftype *cft, struct eventfd_ctx *eventfd)
5847 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5848 struct mem_cgroup_eventfd_list *ev, *tmp;
5849 enum res_type type = MEMFILE_TYPE(cft->private);
5851 BUG_ON(type != _OOM_TYPE);
5853 spin_lock(&memcg_oom_lock);
5855 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5856 if (ev->eventfd == eventfd) {
5857 list_del(&ev->list);
5862 spin_unlock(&memcg_oom_lock);
5865 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5866 struct cftype *cft, struct cgroup_map_cb *cb)
5868 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5870 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5872 if (atomic_read(&memcg->under_oom))
5873 cb->fill(cb, "under_oom", 1);
5875 cb->fill(cb, "under_oom", 0);
5879 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5880 struct cftype *cft, u64 val)
5882 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5883 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5885 /* cannot set to root cgroup and only 0 and 1 are allowed */
5886 if (!parent || !((val == 0) || (val == 1)))
5889 mutex_lock(&memcg_create_mutex);
5890 /* oom-kill-disable is a flag for subhierarchy. */
5891 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5892 mutex_unlock(&memcg_create_mutex);
5895 memcg->oom_kill_disable = val;
5897 memcg_oom_recover(memcg);
5898 mutex_unlock(&memcg_create_mutex);
5902 #ifdef CONFIG_MEMCG_KMEM
5903 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5907 memcg->kmemcg_id = -1;
5908 ret = memcg_propagate_kmem(memcg);
5912 return mem_cgroup_sockets_init(memcg, ss);
5915 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5917 mem_cgroup_sockets_destroy(memcg);
5920 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5922 if (!memcg_kmem_is_active(memcg))
5926 * kmem charges can outlive the cgroup. In the case of slab
5927 * pages, for instance, a page contain objects from various
5928 * processes. As we prevent from taking a reference for every
5929 * such allocation we have to be careful when doing uncharge
5930 * (see memcg_uncharge_kmem) and here during offlining.
5932 * The idea is that that only the _last_ uncharge which sees
5933 * the dead memcg will drop the last reference. An additional
5934 * reference is taken here before the group is marked dead
5935 * which is then paired with css_put during uncharge resp. here.
5937 * Although this might sound strange as this path is called from
5938 * css_offline() when the referencemight have dropped down to 0
5939 * and shouldn't be incremented anymore (css_tryget would fail)
5940 * we do not have other options because of the kmem allocations
5943 css_get(&memcg->css);
5945 memcg_kmem_mark_dead(memcg);
5947 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5950 if (memcg_kmem_test_and_clear_dead(memcg))
5951 css_put(&memcg->css);
5954 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5959 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5963 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5968 static struct cftype mem_cgroup_files[] = {
5970 .name = "usage_in_bytes",
5971 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5972 .read = mem_cgroup_read,
5973 .register_event = mem_cgroup_usage_register_event,
5974 .unregister_event = mem_cgroup_usage_unregister_event,
5977 .name = "max_usage_in_bytes",
5978 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5979 .trigger = mem_cgroup_reset,
5980 .read = mem_cgroup_read,
5983 .name = "limit_in_bytes",
5984 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5985 .write_string = mem_cgroup_write,
5986 .read = mem_cgroup_read,
5989 .name = "soft_limit_in_bytes",
5990 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5991 .write_string = mem_cgroup_write,
5992 .read = mem_cgroup_read,
5996 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5997 .trigger = mem_cgroup_reset,
5998 .read = mem_cgroup_read,
6002 .read_seq_string = memcg_stat_show,
6005 .name = "force_empty",
6006 .trigger = mem_cgroup_force_empty_write,
6009 .name = "use_hierarchy",
6010 .flags = CFTYPE_INSANE,
6011 .write_u64 = mem_cgroup_hierarchy_write,
6012 .read_u64 = mem_cgroup_hierarchy_read,
6015 .name = "swappiness",
6016 .read_u64 = mem_cgroup_swappiness_read,
6017 .write_u64 = mem_cgroup_swappiness_write,
6020 .name = "move_charge_at_immigrate",
6021 .read_u64 = mem_cgroup_move_charge_read,
6022 .write_u64 = mem_cgroup_move_charge_write,
6025 .name = "oom_control",
6026 .read_map = mem_cgroup_oom_control_read,
6027 .write_u64 = mem_cgroup_oom_control_write,
6028 .register_event = mem_cgroup_oom_register_event,
6029 .unregister_event = mem_cgroup_oom_unregister_event,
6030 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6033 .name = "pressure_level",
6034 .register_event = vmpressure_register_event,
6035 .unregister_event = vmpressure_unregister_event,
6039 .name = "numa_stat",
6040 .read_seq_string = memcg_numa_stat_show,
6043 #ifdef CONFIG_MEMCG_KMEM
6045 .name = "kmem.limit_in_bytes",
6046 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6047 .write_string = mem_cgroup_write,
6048 .read = mem_cgroup_read,
6051 .name = "kmem.usage_in_bytes",
6052 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6053 .read = mem_cgroup_read,
6056 .name = "kmem.failcnt",
6057 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6058 .trigger = mem_cgroup_reset,
6059 .read = mem_cgroup_read,
6062 .name = "kmem.max_usage_in_bytes",
6063 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6064 .trigger = mem_cgroup_reset,
6065 .read = mem_cgroup_read,
6067 #ifdef CONFIG_SLABINFO
6069 .name = "kmem.slabinfo",
6070 .read_seq_string = mem_cgroup_slabinfo_read,
6074 { }, /* terminate */
6077 #ifdef CONFIG_MEMCG_SWAP
6078 static struct cftype memsw_cgroup_files[] = {
6080 .name = "memsw.usage_in_bytes",
6081 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6082 .read = mem_cgroup_read,
6083 .register_event = mem_cgroup_usage_register_event,
6084 .unregister_event = mem_cgroup_usage_unregister_event,
6087 .name = "memsw.max_usage_in_bytes",
6088 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6089 .trigger = mem_cgroup_reset,
6090 .read = mem_cgroup_read,
6093 .name = "memsw.limit_in_bytes",
6094 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6095 .write_string = mem_cgroup_write,
6096 .read = mem_cgroup_read,
6099 .name = "memsw.failcnt",
6100 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6101 .trigger = mem_cgroup_reset,
6102 .read = mem_cgroup_read,
6104 { }, /* terminate */
6107 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6109 struct mem_cgroup_per_node *pn;
6110 struct mem_cgroup_per_zone *mz;
6111 int zone, tmp = node;
6113 * This routine is called against possible nodes.
6114 * But it's BUG to call kmalloc() against offline node.
6116 * TODO: this routine can waste much memory for nodes which will
6117 * never be onlined. It's better to use memory hotplug callback
6120 if (!node_state(node, N_NORMAL_MEMORY))
6122 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6126 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6127 mz = &pn->zoneinfo[zone];
6128 lruvec_init(&mz->lruvec);
6129 mz->usage_in_excess = 0;
6130 mz->on_tree = false;
6133 memcg->nodeinfo[node] = pn;
6137 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6139 kfree(memcg->nodeinfo[node]);
6142 static struct mem_cgroup *mem_cgroup_alloc(void)
6144 struct mem_cgroup *memcg;
6145 size_t size = memcg_size();
6147 /* Can be very big if nr_node_ids is very big */
6148 if (size < PAGE_SIZE)
6149 memcg = kzalloc(size, GFP_KERNEL);
6151 memcg = vzalloc(size);
6156 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6159 spin_lock_init(&memcg->pcp_counter_lock);
6163 if (size < PAGE_SIZE)
6171 * At destroying mem_cgroup, references from swap_cgroup can remain.
6172 * (scanning all at force_empty is too costly...)
6174 * Instead of clearing all references at force_empty, we remember
6175 * the number of reference from swap_cgroup and free mem_cgroup when
6176 * it goes down to 0.
6178 * Removal of cgroup itself succeeds regardless of refs from swap.
6181 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6184 size_t size = memcg_size();
6186 mem_cgroup_remove_from_trees(memcg);
6189 free_mem_cgroup_per_zone_info(memcg, node);
6191 free_percpu(memcg->stat);
6194 * We need to make sure that (at least for now), the jump label
6195 * destruction code runs outside of the cgroup lock. This is because
6196 * get_online_cpus(), which is called from the static_branch update,
6197 * can't be called inside the cgroup_lock. cpusets are the ones
6198 * enforcing this dependency, so if they ever change, we might as well.
6200 * schedule_work() will guarantee this happens. Be careful if you need
6201 * to move this code around, and make sure it is outside
6204 disarm_static_keys(memcg);
6205 if (size < PAGE_SIZE)
6212 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6214 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6216 if (!memcg->res.parent)
6218 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6220 EXPORT_SYMBOL(parent_mem_cgroup);
6222 static void __init mem_cgroup_soft_limit_tree_init(void)
6224 struct mem_cgroup_tree_per_node *rtpn;
6225 struct mem_cgroup_tree_per_zone *rtpz;
6226 int tmp, node, zone;
6228 for_each_node(node) {
6230 if (!node_state(node, N_NORMAL_MEMORY))
6232 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6235 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6237 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6238 rtpz = &rtpn->rb_tree_per_zone[zone];
6239 rtpz->rb_root = RB_ROOT;
6240 spin_lock_init(&rtpz->lock);
6245 static struct cgroup_subsys_state * __ref
6246 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6248 struct mem_cgroup *memcg;
6249 long error = -ENOMEM;
6252 memcg = mem_cgroup_alloc();
6254 return ERR_PTR(error);
6257 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6261 if (parent_css == NULL) {
6262 root_mem_cgroup = memcg;
6263 res_counter_init(&memcg->res, NULL);
6264 res_counter_init(&memcg->memsw, NULL);
6265 res_counter_init(&memcg->kmem, NULL);
6268 memcg->last_scanned_node = MAX_NUMNODES;
6269 INIT_LIST_HEAD(&memcg->oom_notify);
6270 memcg->move_charge_at_immigrate = 0;
6271 mutex_init(&memcg->thresholds_lock);
6272 spin_lock_init(&memcg->move_lock);
6273 vmpressure_init(&memcg->vmpressure);
6278 __mem_cgroup_free(memcg);
6279 return ERR_PTR(error);
6283 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6285 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6286 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6289 if (css->cgroup->id > MEM_CGROUP_ID_MAX)
6295 mutex_lock(&memcg_create_mutex);
6297 memcg->use_hierarchy = parent->use_hierarchy;
6298 memcg->oom_kill_disable = parent->oom_kill_disable;
6299 memcg->swappiness = mem_cgroup_swappiness(parent);
6301 if (parent->use_hierarchy) {
6302 res_counter_init(&memcg->res, &parent->res);
6303 res_counter_init(&memcg->memsw, &parent->memsw);
6304 res_counter_init(&memcg->kmem, &parent->kmem);
6307 * No need to take a reference to the parent because cgroup
6308 * core guarantees its existence.
6311 res_counter_init(&memcg->res, NULL);
6312 res_counter_init(&memcg->memsw, NULL);
6313 res_counter_init(&memcg->kmem, NULL);
6315 * Deeper hierachy with use_hierarchy == false doesn't make
6316 * much sense so let cgroup subsystem know about this
6317 * unfortunate state in our controller.
6319 if (parent != root_mem_cgroup)
6320 mem_cgroup_subsys.broken_hierarchy = true;
6323 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6324 mutex_unlock(&memcg_create_mutex);
6329 * Announce all parents that a group from their hierarchy is gone.
6331 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6333 struct mem_cgroup *parent = memcg;
6335 while ((parent = parent_mem_cgroup(parent)))
6336 mem_cgroup_iter_invalidate(parent);
6339 * if the root memcg is not hierarchical we have to check it
6342 if (!root_mem_cgroup->use_hierarchy)
6343 mem_cgroup_iter_invalidate(root_mem_cgroup);
6346 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6348 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6350 kmem_cgroup_css_offline(memcg);
6352 mem_cgroup_invalidate_reclaim_iterators(memcg);
6353 mem_cgroup_reparent_charges(memcg);
6354 mem_cgroup_destroy_all_caches(memcg);
6355 vmpressure_cleanup(&memcg->vmpressure);
6358 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6360 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6362 memcg_destroy_kmem(memcg);
6363 __mem_cgroup_free(memcg);
6367 /* Handlers for move charge at task migration. */
6368 #define PRECHARGE_COUNT_AT_ONCE 256
6369 static int mem_cgroup_do_precharge(unsigned long count)
6372 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6373 struct mem_cgroup *memcg = mc.to;
6375 if (mem_cgroup_is_root(memcg)) {
6376 mc.precharge += count;
6377 /* we don't need css_get for root */
6380 /* try to charge at once */
6382 struct res_counter *dummy;
6384 * "memcg" cannot be under rmdir() because we've already checked
6385 * by cgroup_lock_live_cgroup() that it is not removed and we
6386 * are still under the same cgroup_mutex. So we can postpone
6389 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6391 if (do_swap_account && res_counter_charge(&memcg->memsw,
6392 PAGE_SIZE * count, &dummy)) {
6393 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6396 mc.precharge += count;
6400 /* fall back to one by one charge */
6402 if (signal_pending(current)) {
6406 if (!batch_count--) {
6407 batch_count = PRECHARGE_COUNT_AT_ONCE;
6410 ret = __mem_cgroup_try_charge(NULL,
6411 GFP_KERNEL, 1, &memcg, false);
6413 /* mem_cgroup_clear_mc() will do uncharge later */
6421 * get_mctgt_type - get target type of moving charge
6422 * @vma: the vma the pte to be checked belongs
6423 * @addr: the address corresponding to the pte to be checked
6424 * @ptent: the pte to be checked
6425 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6428 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6429 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6430 * move charge. if @target is not NULL, the page is stored in target->page
6431 * with extra refcnt got(Callers should handle it).
6432 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6433 * target for charge migration. if @target is not NULL, the entry is stored
6436 * Called with pte lock held.
6443 enum mc_target_type {
6449 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6450 unsigned long addr, pte_t ptent)
6452 struct page *page = vm_normal_page(vma, addr, ptent);
6454 if (!page || !page_mapped(page))
6456 if (PageAnon(page)) {
6457 /* we don't move shared anon */
6460 } else if (!move_file())
6461 /* we ignore mapcount for file pages */
6463 if (!get_page_unless_zero(page))
6470 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6471 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6473 struct page *page = NULL;
6474 swp_entry_t ent = pte_to_swp_entry(ptent);
6476 if (!move_anon() || non_swap_entry(ent))
6479 * Because lookup_swap_cache() updates some statistics counter,
6480 * we call find_get_page() with swapper_space directly.
6482 page = find_get_page(swap_address_space(ent), ent.val);
6483 if (do_swap_account)
6484 entry->val = ent.val;
6489 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6490 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6496 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6497 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6499 struct page *page = NULL;
6500 struct address_space *mapping;
6503 if (!vma->vm_file) /* anonymous vma */
6508 mapping = vma->vm_file->f_mapping;
6509 if (pte_none(ptent))
6510 pgoff = linear_page_index(vma, addr);
6511 else /* pte_file(ptent) is true */
6512 pgoff = pte_to_pgoff(ptent);
6514 /* page is moved even if it's not RSS of this task(page-faulted). */
6515 page = find_get_page(mapping, pgoff);
6518 /* shmem/tmpfs may report page out on swap: account for that too. */
6519 if (radix_tree_exceptional_entry(page)) {
6520 swp_entry_t swap = radix_to_swp_entry(page);
6521 if (do_swap_account)
6523 page = find_get_page(swap_address_space(swap), swap.val);
6529 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6530 unsigned long addr, pte_t ptent, union mc_target *target)
6532 struct page *page = NULL;
6533 struct page_cgroup *pc;
6534 enum mc_target_type ret = MC_TARGET_NONE;
6535 swp_entry_t ent = { .val = 0 };
6537 if (pte_present(ptent))
6538 page = mc_handle_present_pte(vma, addr, ptent);
6539 else if (is_swap_pte(ptent))
6540 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6541 else if (pte_none(ptent) || pte_file(ptent))
6542 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6544 if (!page && !ent.val)
6547 pc = lookup_page_cgroup(page);
6549 * Do only loose check w/o page_cgroup lock.
6550 * mem_cgroup_move_account() checks the pc is valid or not under
6553 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6554 ret = MC_TARGET_PAGE;
6556 target->page = page;
6558 if (!ret || !target)
6561 /* There is a swap entry and a page doesn't exist or isn't charged */
6562 if (ent.val && !ret &&
6563 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6564 ret = MC_TARGET_SWAP;
6571 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6573 * We don't consider swapping or file mapped pages because THP does not
6574 * support them for now.
6575 * Caller should make sure that pmd_trans_huge(pmd) is true.
6577 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6578 unsigned long addr, pmd_t pmd, union mc_target *target)
6580 struct page *page = NULL;
6581 struct page_cgroup *pc;
6582 enum mc_target_type ret = MC_TARGET_NONE;
6584 page = pmd_page(pmd);
6585 VM_BUG_ON(!page || !PageHead(page));
6588 pc = lookup_page_cgroup(page);
6589 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6590 ret = MC_TARGET_PAGE;
6593 target->page = page;
6599 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6600 unsigned long addr, pmd_t pmd, union mc_target *target)
6602 return MC_TARGET_NONE;
6606 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6607 unsigned long addr, unsigned long end,
6608 struct mm_walk *walk)
6610 struct vm_area_struct *vma = walk->private;
6614 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6615 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6616 mc.precharge += HPAGE_PMD_NR;
6617 spin_unlock(&vma->vm_mm->page_table_lock);
6621 if (pmd_trans_unstable(pmd))
6623 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6624 for (; addr != end; pte++, addr += PAGE_SIZE)
6625 if (get_mctgt_type(vma, addr, *pte, NULL))
6626 mc.precharge++; /* increment precharge temporarily */
6627 pte_unmap_unlock(pte - 1, ptl);
6633 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6635 unsigned long precharge;
6636 struct vm_area_struct *vma;
6638 down_read(&mm->mmap_sem);
6639 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6640 struct mm_walk mem_cgroup_count_precharge_walk = {
6641 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6645 if (is_vm_hugetlb_page(vma))
6647 walk_page_range(vma->vm_start, vma->vm_end,
6648 &mem_cgroup_count_precharge_walk);
6650 up_read(&mm->mmap_sem);
6652 precharge = mc.precharge;
6658 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6660 unsigned long precharge = mem_cgroup_count_precharge(mm);
6662 VM_BUG_ON(mc.moving_task);
6663 mc.moving_task = current;
6664 return mem_cgroup_do_precharge(precharge);
6667 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6668 static void __mem_cgroup_clear_mc(void)
6670 struct mem_cgroup *from = mc.from;
6671 struct mem_cgroup *to = mc.to;
6674 /* we must uncharge all the leftover precharges from mc.to */
6676 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6680 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6681 * we must uncharge here.
6683 if (mc.moved_charge) {
6684 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6685 mc.moved_charge = 0;
6687 /* we must fixup refcnts and charges */
6688 if (mc.moved_swap) {
6689 /* uncharge swap account from the old cgroup */
6690 if (!mem_cgroup_is_root(mc.from))
6691 res_counter_uncharge(&mc.from->memsw,
6692 PAGE_SIZE * mc.moved_swap);
6694 for (i = 0; i < mc.moved_swap; i++)
6695 css_put(&mc.from->css);
6697 if (!mem_cgroup_is_root(mc.to)) {
6699 * we charged both to->res and to->memsw, so we should
6702 res_counter_uncharge(&mc.to->res,
6703 PAGE_SIZE * mc.moved_swap);
6705 /* we've already done css_get(mc.to) */
6708 memcg_oom_recover(from);
6709 memcg_oom_recover(to);
6710 wake_up_all(&mc.waitq);
6713 static void mem_cgroup_clear_mc(void)
6715 struct mem_cgroup *from = mc.from;
6718 * we must clear moving_task before waking up waiters at the end of
6721 mc.moving_task = NULL;
6722 __mem_cgroup_clear_mc();
6723 spin_lock(&mc.lock);
6726 spin_unlock(&mc.lock);
6727 mem_cgroup_end_move(from);
6730 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6731 struct cgroup_taskset *tset)
6733 struct task_struct *p = cgroup_taskset_first(tset);
6735 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6736 unsigned long move_charge_at_immigrate;
6739 * We are now commited to this value whatever it is. Changes in this
6740 * tunable will only affect upcoming migrations, not the current one.
6741 * So we need to save it, and keep it going.
6743 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6744 if (move_charge_at_immigrate) {
6745 struct mm_struct *mm;
6746 struct mem_cgroup *from = mem_cgroup_from_task(p);
6748 VM_BUG_ON(from == memcg);
6750 mm = get_task_mm(p);
6753 /* We move charges only when we move a owner of the mm */
6754 if (mm->owner == p) {
6757 VM_BUG_ON(mc.precharge);
6758 VM_BUG_ON(mc.moved_charge);
6759 VM_BUG_ON(mc.moved_swap);
6760 mem_cgroup_start_move(from);
6761 spin_lock(&mc.lock);
6764 mc.immigrate_flags = move_charge_at_immigrate;
6765 spin_unlock(&mc.lock);
6766 /* We set mc.moving_task later */
6768 ret = mem_cgroup_precharge_mc(mm);
6770 mem_cgroup_clear_mc();
6777 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6778 struct cgroup_taskset *tset)
6780 mem_cgroup_clear_mc();
6783 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6784 unsigned long addr, unsigned long end,
6785 struct mm_walk *walk)
6788 struct vm_area_struct *vma = walk->private;
6791 enum mc_target_type target_type;
6792 union mc_target target;
6794 struct page_cgroup *pc;
6797 * We don't take compound_lock() here but no race with splitting thp
6799 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6800 * under splitting, which means there's no concurrent thp split,
6801 * - if another thread runs into split_huge_page() just after we
6802 * entered this if-block, the thread must wait for page table lock
6803 * to be unlocked in __split_huge_page_splitting(), where the main
6804 * part of thp split is not executed yet.
6806 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6807 if (mc.precharge < HPAGE_PMD_NR) {
6808 spin_unlock(&vma->vm_mm->page_table_lock);
6811 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6812 if (target_type == MC_TARGET_PAGE) {
6814 if (!isolate_lru_page(page)) {
6815 pc = lookup_page_cgroup(page);
6816 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6817 pc, mc.from, mc.to)) {
6818 mc.precharge -= HPAGE_PMD_NR;
6819 mc.moved_charge += HPAGE_PMD_NR;
6821 putback_lru_page(page);
6825 spin_unlock(&vma->vm_mm->page_table_lock);
6829 if (pmd_trans_unstable(pmd))
6832 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6833 for (; addr != end; addr += PAGE_SIZE) {
6834 pte_t ptent = *(pte++);
6840 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6841 case MC_TARGET_PAGE:
6843 if (isolate_lru_page(page))
6845 pc = lookup_page_cgroup(page);
6846 if (!mem_cgroup_move_account(page, 1, pc,
6849 /* we uncharge from mc.from later. */
6852 putback_lru_page(page);
6853 put: /* get_mctgt_type() gets the page */
6856 case MC_TARGET_SWAP:
6858 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6860 /* we fixup refcnts and charges later. */
6868 pte_unmap_unlock(pte - 1, ptl);
6873 * We have consumed all precharges we got in can_attach().
6874 * We try charge one by one, but don't do any additional
6875 * charges to mc.to if we have failed in charge once in attach()
6878 ret = mem_cgroup_do_precharge(1);
6886 static void mem_cgroup_move_charge(struct mm_struct *mm)
6888 struct vm_area_struct *vma;
6890 lru_add_drain_all();
6892 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6894 * Someone who are holding the mmap_sem might be waiting in
6895 * waitq. So we cancel all extra charges, wake up all waiters,
6896 * and retry. Because we cancel precharges, we might not be able
6897 * to move enough charges, but moving charge is a best-effort
6898 * feature anyway, so it wouldn't be a big problem.
6900 __mem_cgroup_clear_mc();
6904 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6906 struct mm_walk mem_cgroup_move_charge_walk = {
6907 .pmd_entry = mem_cgroup_move_charge_pte_range,
6911 if (is_vm_hugetlb_page(vma))
6913 ret = walk_page_range(vma->vm_start, vma->vm_end,
6914 &mem_cgroup_move_charge_walk);
6917 * means we have consumed all precharges and failed in
6918 * doing additional charge. Just abandon here.
6922 up_read(&mm->mmap_sem);
6925 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6926 struct cgroup_taskset *tset)
6928 struct task_struct *p = cgroup_taskset_first(tset);
6929 struct mm_struct *mm = get_task_mm(p);
6933 mem_cgroup_move_charge(mm);
6937 mem_cgroup_clear_mc();
6939 #else /* !CONFIG_MMU */
6940 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6941 struct cgroup_taskset *tset)
6945 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6946 struct cgroup_taskset *tset)
6949 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6950 struct cgroup_taskset *tset)
6956 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6957 * to verify sane_behavior flag on each mount attempt.
6959 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6962 * use_hierarchy is forced with sane_behavior. cgroup core
6963 * guarantees that @root doesn't have any children, so turning it
6964 * on for the root memcg is enough.
6966 if (cgroup_sane_behavior(root_css->cgroup))
6967 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6970 struct cgroup_subsys mem_cgroup_subsys = {
6972 .subsys_id = mem_cgroup_subsys_id,
6973 .css_alloc = mem_cgroup_css_alloc,
6974 .css_online = mem_cgroup_css_online,
6975 .css_offline = mem_cgroup_css_offline,
6976 .css_free = mem_cgroup_css_free,
6977 .can_attach = mem_cgroup_can_attach,
6978 .cancel_attach = mem_cgroup_cancel_attach,
6979 .attach = mem_cgroup_move_task,
6980 .bind = mem_cgroup_bind,
6981 .base_cftypes = mem_cgroup_files,
6985 #ifdef CONFIG_MEMCG_SWAP
6986 static int __init enable_swap_account(char *s)
6988 if (!strcmp(s, "1"))
6989 really_do_swap_account = 1;
6990 else if (!strcmp(s, "0"))
6991 really_do_swap_account = 0;
6994 __setup("swapaccount=", enable_swap_account);
6996 static void __init memsw_file_init(void)
6998 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7001 static void __init enable_swap_cgroup(void)
7003 if (!mem_cgroup_disabled() && really_do_swap_account) {
7004 do_swap_account = 1;
7010 static void __init enable_swap_cgroup(void)
7016 * subsys_initcall() for memory controller.
7018 * Some parts like hotcpu_notifier() have to be initialized from this context
7019 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7020 * everything that doesn't depend on a specific mem_cgroup structure should
7021 * be initialized from here.
7023 static int __init mem_cgroup_init(void)
7025 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7026 enable_swap_cgroup();
7027 mem_cgroup_soft_limit_tree_init();
7031 subsys_initcall(mem_cgroup_init);