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>
57 #include <linux/lockdep.h>
61 #include <net/tcp_memcontrol.h>
63 #include <asm/uaccess.h>
65 #include <trace/events/vmscan.h>
67 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
68 EXPORT_SYMBOL(mem_cgroup_subsys);
70 #define MEM_CGROUP_RECLAIM_RETRIES 5
71 static struct mem_cgroup *root_mem_cgroup __read_mostly;
73 #ifdef CONFIG_MEMCG_SWAP
74 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
75 int do_swap_account __read_mostly;
77 /* for remember boot option*/
78 #ifdef CONFIG_MEMCG_SWAP_ENABLED
79 static int really_do_swap_account __initdata = 1;
81 static int really_do_swap_account __initdata = 0;
85 #define do_swap_account 0
89 static const char * const mem_cgroup_stat_names[] = {
98 enum mem_cgroup_events_index {
99 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
100 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
101 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
102 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
103 MEM_CGROUP_EVENTS_NSTATS,
106 static const char * const mem_cgroup_events_names[] = {
113 static const char * const mem_cgroup_lru_names[] = {
122 * Per memcg event counter is incremented at every pagein/pageout. With THP,
123 * it will be incremated by the number of pages. This counter is used for
124 * for trigger some periodic events. This is straightforward and better
125 * than using jiffies etc. to handle periodic memcg event.
127 enum mem_cgroup_events_target {
128 MEM_CGROUP_TARGET_THRESH,
129 MEM_CGROUP_TARGET_SOFTLIMIT,
130 MEM_CGROUP_TARGET_NUMAINFO,
133 #define THRESHOLDS_EVENTS_TARGET 128
134 #define SOFTLIMIT_EVENTS_TARGET 1024
135 #define NUMAINFO_EVENTS_TARGET 1024
137 struct mem_cgroup_stat_cpu {
138 long count[MEM_CGROUP_STAT_NSTATS];
139 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
140 unsigned long nr_page_events;
141 unsigned long targets[MEM_CGROUP_NTARGETS];
144 struct mem_cgroup_reclaim_iter {
146 * last scanned hierarchy member. Valid only if last_dead_count
147 * matches memcg->dead_count of the hierarchy root group.
149 struct mem_cgroup *last_visited;
150 unsigned long last_dead_count;
152 /* scan generation, increased every round-trip */
153 unsigned int generation;
157 * per-zone information in memory controller.
159 struct mem_cgroup_per_zone {
160 struct lruvec lruvec;
161 unsigned long lru_size[NR_LRU_LISTS];
163 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
165 struct rb_node tree_node; /* RB tree node */
166 unsigned long long usage_in_excess;/* Set to the value by which */
167 /* the soft limit is exceeded*/
169 struct mem_cgroup *memcg; /* Back pointer, we cannot */
170 /* use container_of */
173 struct mem_cgroup_per_node {
174 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
178 * Cgroups above their limits are maintained in a RB-Tree, independent of
179 * their hierarchy representation
182 struct mem_cgroup_tree_per_zone {
183 struct rb_root rb_root;
187 struct mem_cgroup_tree_per_node {
188 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
191 struct mem_cgroup_tree {
192 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
195 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
197 struct mem_cgroup_threshold {
198 struct eventfd_ctx *eventfd;
203 struct mem_cgroup_threshold_ary {
204 /* An array index points to threshold just below or equal to usage. */
205 int current_threshold;
206 /* Size of entries[] */
208 /* Array of thresholds */
209 struct mem_cgroup_threshold entries[0];
212 struct mem_cgroup_thresholds {
213 /* Primary thresholds array */
214 struct mem_cgroup_threshold_ary *primary;
216 * Spare threshold array.
217 * This is needed to make mem_cgroup_unregister_event() "never fail".
218 * It must be able to store at least primary->size - 1 entries.
220 struct mem_cgroup_threshold_ary *spare;
224 struct mem_cgroup_eventfd_list {
225 struct list_head list;
226 struct eventfd_ctx *eventfd;
229 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
230 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
233 * The memory controller data structure. The memory controller controls both
234 * page cache and RSS per cgroup. We would eventually like to provide
235 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
236 * to help the administrator determine what knobs to tune.
238 * TODO: Add a water mark for the memory controller. Reclaim will begin when
239 * we hit the water mark. May be even add a low water mark, such that
240 * no reclaim occurs from a cgroup at it's low water mark, this is
241 * a feature that will be implemented much later in the future.
244 struct cgroup_subsys_state css;
246 * the counter to account for memory usage
248 struct res_counter res;
250 /* vmpressure notifications */
251 struct vmpressure vmpressure;
254 * the counter to account for mem+swap usage.
256 struct res_counter memsw;
259 * the counter to account for kernel memory usage.
261 struct res_counter kmem;
263 * Should the accounting and control be hierarchical, per subtree?
266 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
270 atomic_t oom_wakeups;
273 /* OOM-Killer disable */
274 int oom_kill_disable;
276 /* set when res.limit == memsw.limit */
277 bool memsw_is_minimum;
279 /* protect arrays of thresholds */
280 struct mutex thresholds_lock;
282 /* thresholds for memory usage. RCU-protected */
283 struct mem_cgroup_thresholds thresholds;
285 /* thresholds for mem+swap usage. RCU-protected */
286 struct mem_cgroup_thresholds memsw_thresholds;
288 /* For oom notifier event fd */
289 struct list_head oom_notify;
292 * Should we move charges of a task when a task is moved into this
293 * mem_cgroup ? And what type of charges should we move ?
295 unsigned long move_charge_at_immigrate;
297 * set > 0 if pages under this cgroup are moving to other cgroup.
299 atomic_t moving_account;
300 /* taken only while moving_account > 0 */
301 spinlock_t move_lock;
305 struct mem_cgroup_stat_cpu __percpu *stat;
307 * used when a cpu is offlined or other synchronizations
308 * See mem_cgroup_read_stat().
310 struct mem_cgroup_stat_cpu nocpu_base;
311 spinlock_t pcp_counter_lock;
314 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
315 struct tcp_memcontrol tcp_mem;
317 #if defined(CONFIG_MEMCG_KMEM)
318 /* analogous to slab_common's slab_caches list. per-memcg */
319 struct list_head memcg_slab_caches;
320 /* Not a spinlock, we can take a lot of time walking the list */
321 struct mutex slab_caches_mutex;
322 /* Index in the kmem_cache->memcg_params->memcg_caches array */
326 int last_scanned_node;
328 nodemask_t scan_nodes;
329 atomic_t numainfo_events;
330 atomic_t numainfo_updating;
333 struct mem_cgroup_per_node *nodeinfo[0];
334 /* WARNING: nodeinfo must be the last member here */
337 static size_t memcg_size(void)
339 return sizeof(struct mem_cgroup) +
340 nr_node_ids * sizeof(struct mem_cgroup_per_node);
343 /* internal only representation about the status of kmem accounting. */
345 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
346 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
347 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
350 /* We account when limit is on, but only after call sites are patched */
351 #define KMEM_ACCOUNTED_MASK \
352 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
354 #ifdef CONFIG_MEMCG_KMEM
355 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
357 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
360 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
362 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
365 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
367 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
370 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
372 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
375 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
378 * Our caller must use css_get() first, because memcg_uncharge_kmem()
379 * will call css_put() if it sees the memcg is dead.
382 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
383 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
386 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
388 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
389 &memcg->kmem_account_flags);
393 /* Stuffs for move charges at task migration. */
395 * Types of charges to be moved. "move_charge_at_immitgrate" and
396 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
399 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
400 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
404 /* "mc" and its members are protected by cgroup_mutex */
405 static struct move_charge_struct {
406 spinlock_t lock; /* for from, to */
407 struct mem_cgroup *from;
408 struct mem_cgroup *to;
409 unsigned long immigrate_flags;
410 unsigned long precharge;
411 unsigned long moved_charge;
412 unsigned long moved_swap;
413 struct task_struct *moving_task; /* a task moving charges */
414 wait_queue_head_t waitq; /* a waitq for other context */
416 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
417 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
420 static bool move_anon(void)
422 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
425 static bool move_file(void)
427 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
431 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
432 * limit reclaim to prevent infinite loops, if they ever occur.
434 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
435 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
438 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
439 MEM_CGROUP_CHARGE_TYPE_ANON,
440 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
441 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
445 /* for encoding cft->private value on file */
453 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
454 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
455 #define MEMFILE_ATTR(val) ((val) & 0xffff)
456 /* Used for OOM nofiier */
457 #define OOM_CONTROL (0)
460 * Reclaim flags for mem_cgroup_hierarchical_reclaim
462 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
463 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
464 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
465 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
468 * The memcg_create_mutex will be held whenever a new cgroup is created.
469 * As a consequence, any change that needs to protect against new child cgroups
470 * appearing has to hold it as well.
472 static DEFINE_MUTEX(memcg_create_mutex);
474 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
476 return s ? container_of(s, struct mem_cgroup, css) : NULL;
479 /* Some nice accessors for the vmpressure. */
480 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
483 memcg = root_mem_cgroup;
484 return &memcg->vmpressure;
487 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
489 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
492 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
494 return &mem_cgroup_from_css(css)->vmpressure;
497 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
499 return (memcg == root_mem_cgroup);
502 /* Writing them here to avoid exposing memcg's inner layout */
503 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
505 void sock_update_memcg(struct sock *sk)
507 if (mem_cgroup_sockets_enabled) {
508 struct mem_cgroup *memcg;
509 struct cg_proto *cg_proto;
511 BUG_ON(!sk->sk_prot->proto_cgroup);
513 /* Socket cloning can throw us here with sk_cgrp already
514 * filled. It won't however, necessarily happen from
515 * process context. So the test for root memcg given
516 * the current task's memcg won't help us in this case.
518 * Respecting the original socket's memcg is a better
519 * decision in this case.
522 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
523 css_get(&sk->sk_cgrp->memcg->css);
528 memcg = mem_cgroup_from_task(current);
529 cg_proto = sk->sk_prot->proto_cgroup(memcg);
530 if (!mem_cgroup_is_root(memcg) &&
531 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
532 sk->sk_cgrp = cg_proto;
537 EXPORT_SYMBOL(sock_update_memcg);
539 void sock_release_memcg(struct sock *sk)
541 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
542 struct mem_cgroup *memcg;
543 WARN_ON(!sk->sk_cgrp->memcg);
544 memcg = sk->sk_cgrp->memcg;
545 css_put(&sk->sk_cgrp->memcg->css);
549 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
551 if (!memcg || mem_cgroup_is_root(memcg))
554 return &memcg->tcp_mem.cg_proto;
556 EXPORT_SYMBOL(tcp_proto_cgroup);
558 static void disarm_sock_keys(struct mem_cgroup *memcg)
560 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
562 static_key_slow_dec(&memcg_socket_limit_enabled);
565 static void disarm_sock_keys(struct mem_cgroup *memcg)
570 #ifdef CONFIG_MEMCG_KMEM
572 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
573 * There are two main reasons for not using the css_id for this:
574 * 1) this works better in sparse environments, where we have a lot of memcgs,
575 * but only a few kmem-limited. Or also, if we have, for instance, 200
576 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
577 * 200 entry array for that.
579 * 2) In order not to violate the cgroup API, we would like to do all memory
580 * allocation in ->create(). At that point, we haven't yet allocated the
581 * css_id. Having a separate index prevents us from messing with the cgroup
584 * The current size of the caches array is stored in
585 * memcg_limited_groups_array_size. It will double each time we have to
588 static DEFINE_IDA(kmem_limited_groups);
589 int memcg_limited_groups_array_size;
592 * MIN_SIZE is different than 1, because we would like to avoid going through
593 * the alloc/free process all the time. In a small machine, 4 kmem-limited
594 * cgroups is a reasonable guess. In the future, it could be a parameter or
595 * tunable, but that is strictly not necessary.
597 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
598 * this constant directly from cgroup, but it is understandable that this is
599 * better kept as an internal representation in cgroup.c. In any case, the
600 * css_id space is not getting any smaller, and we don't have to necessarily
601 * increase ours as well if it increases.
603 #define MEMCG_CACHES_MIN_SIZE 4
604 #define MEMCG_CACHES_MAX_SIZE 65535
607 * A lot of the calls to the cache allocation functions are expected to be
608 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
609 * conditional to this static branch, we'll have to allow modules that does
610 * kmem_cache_alloc and the such to see this symbol as well
612 struct static_key memcg_kmem_enabled_key;
613 EXPORT_SYMBOL(memcg_kmem_enabled_key);
615 static void disarm_kmem_keys(struct mem_cgroup *memcg)
617 if (memcg_kmem_is_active(memcg)) {
618 static_key_slow_dec(&memcg_kmem_enabled_key);
619 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
622 * This check can't live in kmem destruction function,
623 * since the charges will outlive the cgroup
625 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
628 static void disarm_kmem_keys(struct mem_cgroup *memcg)
631 #endif /* CONFIG_MEMCG_KMEM */
633 static void disarm_static_keys(struct mem_cgroup *memcg)
635 disarm_sock_keys(memcg);
636 disarm_kmem_keys(memcg);
639 static void drain_all_stock_async(struct mem_cgroup *memcg);
641 static struct mem_cgroup_per_zone *
642 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
644 VM_BUG_ON((unsigned)nid >= nr_node_ids);
645 return &memcg->nodeinfo[nid]->zoneinfo[zid];
648 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
653 static struct mem_cgroup_per_zone *
654 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
656 int nid = page_to_nid(page);
657 int zid = page_zonenum(page);
659 return mem_cgroup_zoneinfo(memcg, nid, zid);
662 static struct mem_cgroup_tree_per_zone *
663 soft_limit_tree_node_zone(int nid, int zid)
665 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
668 static struct mem_cgroup_tree_per_zone *
669 soft_limit_tree_from_page(struct page *page)
671 int nid = page_to_nid(page);
672 int zid = page_zonenum(page);
674 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
678 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
679 struct mem_cgroup_per_zone *mz,
680 struct mem_cgroup_tree_per_zone *mctz,
681 unsigned long long new_usage_in_excess)
683 struct rb_node **p = &mctz->rb_root.rb_node;
684 struct rb_node *parent = NULL;
685 struct mem_cgroup_per_zone *mz_node;
690 mz->usage_in_excess = new_usage_in_excess;
691 if (!mz->usage_in_excess)
695 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
697 if (mz->usage_in_excess < mz_node->usage_in_excess)
700 * We can't avoid mem cgroups that are over their soft
701 * limit by the same amount
703 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
706 rb_link_node(&mz->tree_node, parent, p);
707 rb_insert_color(&mz->tree_node, &mctz->rb_root);
712 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
713 struct mem_cgroup_per_zone *mz,
714 struct mem_cgroup_tree_per_zone *mctz)
718 rb_erase(&mz->tree_node, &mctz->rb_root);
723 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
724 struct mem_cgroup_per_zone *mz,
725 struct mem_cgroup_tree_per_zone *mctz)
727 spin_lock(&mctz->lock);
728 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
729 spin_unlock(&mctz->lock);
733 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
735 unsigned long long excess;
736 struct mem_cgroup_per_zone *mz;
737 struct mem_cgroup_tree_per_zone *mctz;
738 int nid = page_to_nid(page);
739 int zid = page_zonenum(page);
740 mctz = soft_limit_tree_from_page(page);
743 * Necessary to update all ancestors when hierarchy is used.
744 * because their event counter is not touched.
746 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
747 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
748 excess = res_counter_soft_limit_excess(&memcg->res);
750 * We have to update the tree if mz is on RB-tree or
751 * mem is over its softlimit.
753 if (excess || mz->on_tree) {
754 spin_lock(&mctz->lock);
755 /* if on-tree, remove it */
757 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
759 * Insert again. mz->usage_in_excess will be updated.
760 * If excess is 0, no tree ops.
762 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
763 spin_unlock(&mctz->lock);
768 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
771 struct mem_cgroup_per_zone *mz;
772 struct mem_cgroup_tree_per_zone *mctz;
774 for_each_node(node) {
775 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
776 mz = mem_cgroup_zoneinfo(memcg, node, zone);
777 mctz = soft_limit_tree_node_zone(node, zone);
778 mem_cgroup_remove_exceeded(memcg, mz, mctz);
783 static struct mem_cgroup_per_zone *
784 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
786 struct rb_node *rightmost = NULL;
787 struct mem_cgroup_per_zone *mz;
791 rightmost = rb_last(&mctz->rb_root);
793 goto done; /* Nothing to reclaim from */
795 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
797 * Remove the node now but someone else can add it back,
798 * we will to add it back at the end of reclaim to its correct
799 * position in the tree.
801 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
802 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
803 !css_tryget(&mz->memcg->css))
809 static struct mem_cgroup_per_zone *
810 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
812 struct mem_cgroup_per_zone *mz;
814 spin_lock(&mctz->lock);
815 mz = __mem_cgroup_largest_soft_limit_node(mctz);
816 spin_unlock(&mctz->lock);
821 * Implementation Note: reading percpu statistics for memcg.
823 * Both of vmstat[] and percpu_counter has threshold and do periodic
824 * synchronization to implement "quick" read. There are trade-off between
825 * reading cost and precision of value. Then, we may have a chance to implement
826 * a periodic synchronizion of counter in memcg's counter.
828 * But this _read() function is used for user interface now. The user accounts
829 * memory usage by memory cgroup and he _always_ requires exact value because
830 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
831 * have to visit all online cpus and make sum. So, for now, unnecessary
832 * synchronization is not implemented. (just implemented for cpu hotplug)
834 * If there are kernel internal actions which can make use of some not-exact
835 * value, and reading all cpu value can be performance bottleneck in some
836 * common workload, threashold and synchonization as vmstat[] should be
839 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
840 enum mem_cgroup_stat_index idx)
846 for_each_online_cpu(cpu)
847 val += per_cpu(memcg->stat->count[idx], cpu);
848 #ifdef CONFIG_HOTPLUG_CPU
849 spin_lock(&memcg->pcp_counter_lock);
850 val += memcg->nocpu_base.count[idx];
851 spin_unlock(&memcg->pcp_counter_lock);
857 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
860 int val = (charge) ? 1 : -1;
861 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
864 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
865 enum mem_cgroup_events_index idx)
867 unsigned long val = 0;
871 for_each_online_cpu(cpu)
872 val += per_cpu(memcg->stat->events[idx], cpu);
873 #ifdef CONFIG_HOTPLUG_CPU
874 spin_lock(&memcg->pcp_counter_lock);
875 val += memcg->nocpu_base.events[idx];
876 spin_unlock(&memcg->pcp_counter_lock);
882 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
884 bool anon, int nr_pages)
889 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
890 * counted as CACHE even if it's on ANON LRU.
893 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
896 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
899 if (PageTransHuge(page))
900 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
903 /* pagein of a big page is an event. So, ignore page size */
905 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
907 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
908 nr_pages = -nr_pages; /* for event */
911 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
917 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
919 struct mem_cgroup_per_zone *mz;
921 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
922 return mz->lru_size[lru];
926 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
927 unsigned int lru_mask)
929 struct mem_cgroup_per_zone *mz;
931 unsigned long ret = 0;
933 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
936 if (BIT(lru) & lru_mask)
937 ret += mz->lru_size[lru];
943 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
944 int nid, unsigned int lru_mask)
949 for (zid = 0; zid < MAX_NR_ZONES; zid++)
950 total += mem_cgroup_zone_nr_lru_pages(memcg,
956 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
957 unsigned int lru_mask)
962 for_each_node_state(nid, N_MEMORY)
963 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
967 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
968 enum mem_cgroup_events_target target)
970 unsigned long val, next;
972 val = __this_cpu_read(memcg->stat->nr_page_events);
973 next = __this_cpu_read(memcg->stat->targets[target]);
974 /* from time_after() in jiffies.h */
975 if ((long)next - (long)val < 0) {
977 case MEM_CGROUP_TARGET_THRESH:
978 next = val + THRESHOLDS_EVENTS_TARGET;
980 case MEM_CGROUP_TARGET_SOFTLIMIT:
981 next = val + SOFTLIMIT_EVENTS_TARGET;
983 case MEM_CGROUP_TARGET_NUMAINFO:
984 next = val + NUMAINFO_EVENTS_TARGET;
989 __this_cpu_write(memcg->stat->targets[target], next);
996 * Check events in order.
999 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1002 /* threshold event is triggered in finer grain than soft limit */
1003 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1004 MEM_CGROUP_TARGET_THRESH))) {
1006 bool do_numainfo __maybe_unused;
1008 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1009 MEM_CGROUP_TARGET_SOFTLIMIT);
1010 #if MAX_NUMNODES > 1
1011 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1012 MEM_CGROUP_TARGET_NUMAINFO);
1016 mem_cgroup_threshold(memcg);
1017 if (unlikely(do_softlimit))
1018 mem_cgroup_update_tree(memcg, page);
1019 #if MAX_NUMNODES > 1
1020 if (unlikely(do_numainfo))
1021 atomic_inc(&memcg->numainfo_events);
1027 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1030 * mm_update_next_owner() may clear mm->owner to NULL
1031 * if it races with swapoff, page migration, etc.
1032 * So this can be called with p == NULL.
1037 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1040 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1042 struct mem_cgroup *memcg = NULL;
1047 * Because we have no locks, mm->owner's may be being moved to other
1048 * cgroup. We use css_tryget() here even if this looks
1049 * pessimistic (rather than adding locks here).
1053 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1054 if (unlikely(!memcg))
1056 } while (!css_tryget(&memcg->css));
1062 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1063 * ref. count) or NULL if the whole root's subtree has been visited.
1065 * helper function to be used by mem_cgroup_iter
1067 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1068 struct mem_cgroup *last_visited)
1070 struct cgroup_subsys_state *prev_css, *next_css;
1072 prev_css = last_visited ? &last_visited->css : NULL;
1074 next_css = css_next_descendant_pre(prev_css, &root->css);
1077 * Even if we found a group we have to make sure it is
1078 * alive. css && !memcg means that the groups should be
1079 * skipped and we should continue the tree walk.
1080 * last_visited css is safe to use because it is
1081 * protected by css_get and the tree walk is rcu safe.
1084 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
1086 if (css_tryget(&mem->css))
1089 prev_css = next_css;
1097 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1100 * When a group in the hierarchy below root is destroyed, the
1101 * hierarchy iterator can no longer be trusted since it might
1102 * have pointed to the destroyed group. Invalidate it.
1104 atomic_inc(&root->dead_count);
1107 static struct mem_cgroup *
1108 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1109 struct mem_cgroup *root,
1112 struct mem_cgroup *position = NULL;
1114 * A cgroup destruction happens in two stages: offlining and
1115 * release. They are separated by a RCU grace period.
1117 * If the iterator is valid, we may still race with an
1118 * offlining. The RCU lock ensures the object won't be
1119 * released, tryget will fail if we lost the race.
1121 *sequence = atomic_read(&root->dead_count);
1122 if (iter->last_dead_count == *sequence) {
1124 position = iter->last_visited;
1125 if (position && !css_tryget(&position->css))
1131 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1132 struct mem_cgroup *last_visited,
1133 struct mem_cgroup *new_position,
1137 css_put(&last_visited->css);
1139 * We store the sequence count from the time @last_visited was
1140 * loaded successfully instead of rereading it here so that we
1141 * don't lose destruction events in between. We could have
1142 * raced with the destruction of @new_position after all.
1144 iter->last_visited = new_position;
1146 iter->last_dead_count = sequence;
1150 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1151 * @root: hierarchy root
1152 * @prev: previously returned memcg, NULL on first invocation
1153 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1155 * Returns references to children of the hierarchy below @root, or
1156 * @root itself, or %NULL after a full round-trip.
1158 * Caller must pass the return value in @prev on subsequent
1159 * invocations for reference counting, or use mem_cgroup_iter_break()
1160 * to cancel a hierarchy walk before the round-trip is complete.
1162 * Reclaimers can specify a zone and a priority level in @reclaim to
1163 * divide up the memcgs in the hierarchy among all concurrent
1164 * reclaimers operating on the same zone and priority.
1166 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1167 struct mem_cgroup *prev,
1168 struct mem_cgroup_reclaim_cookie *reclaim)
1170 struct mem_cgroup *memcg = NULL;
1171 struct mem_cgroup *last_visited = NULL;
1173 if (mem_cgroup_disabled())
1177 root = root_mem_cgroup;
1179 if (prev && !reclaim)
1180 last_visited = prev;
1182 if (!root->use_hierarchy && root != root_mem_cgroup) {
1190 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1191 int uninitialized_var(seq);
1194 int nid = zone_to_nid(reclaim->zone);
1195 int zid = zone_idx(reclaim->zone);
1196 struct mem_cgroup_per_zone *mz;
1198 mz = mem_cgroup_zoneinfo(root, nid, zid);
1199 iter = &mz->reclaim_iter[reclaim->priority];
1200 if (prev && reclaim->generation != iter->generation) {
1201 iter->last_visited = NULL;
1205 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1208 memcg = __mem_cgroup_iter_next(root, last_visited);
1211 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1215 else if (!prev && memcg)
1216 reclaim->generation = iter->generation;
1225 if (prev && prev != root)
1226 css_put(&prev->css);
1232 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1233 * @root: hierarchy root
1234 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1236 void mem_cgroup_iter_break(struct mem_cgroup *root,
1237 struct mem_cgroup *prev)
1240 root = root_mem_cgroup;
1241 if (prev && prev != root)
1242 css_put(&prev->css);
1246 * Iteration constructs for visiting all cgroups (under a tree). If
1247 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1248 * be used for reference counting.
1250 #define for_each_mem_cgroup_tree(iter, root) \
1251 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1253 iter = mem_cgroup_iter(root, iter, NULL))
1255 #define for_each_mem_cgroup(iter) \
1256 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1258 iter = mem_cgroup_iter(NULL, iter, NULL))
1260 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1262 struct mem_cgroup *memcg;
1265 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1266 if (unlikely(!memcg))
1271 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1274 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1282 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1285 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1286 * @zone: zone of the wanted lruvec
1287 * @memcg: memcg of the wanted lruvec
1289 * Returns the lru list vector holding pages for the given @zone and
1290 * @mem. This can be the global zone lruvec, if the memory controller
1293 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1294 struct mem_cgroup *memcg)
1296 struct mem_cgroup_per_zone *mz;
1297 struct lruvec *lruvec;
1299 if (mem_cgroup_disabled()) {
1300 lruvec = &zone->lruvec;
1304 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1305 lruvec = &mz->lruvec;
1308 * Since a node can be onlined after the mem_cgroup was created,
1309 * we have to be prepared to initialize lruvec->zone here;
1310 * and if offlined then reonlined, we need to reinitialize it.
1312 if (unlikely(lruvec->zone != zone))
1313 lruvec->zone = zone;
1318 * Following LRU functions are allowed to be used without PCG_LOCK.
1319 * Operations are called by routine of global LRU independently from memcg.
1320 * What we have to take care of here is validness of pc->mem_cgroup.
1322 * Changes to pc->mem_cgroup happens when
1325 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1326 * It is added to LRU before charge.
1327 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1328 * When moving account, the page is not on LRU. It's isolated.
1332 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1334 * @zone: zone of the page
1336 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1338 struct mem_cgroup_per_zone *mz;
1339 struct mem_cgroup *memcg;
1340 struct page_cgroup *pc;
1341 struct lruvec *lruvec;
1343 if (mem_cgroup_disabled()) {
1344 lruvec = &zone->lruvec;
1348 pc = lookup_page_cgroup(page);
1349 memcg = pc->mem_cgroup;
1352 * Surreptitiously switch any uncharged offlist page to root:
1353 * an uncharged page off lru does nothing to secure
1354 * its former mem_cgroup from sudden removal.
1356 * Our caller holds lru_lock, and PageCgroupUsed is updated
1357 * under page_cgroup lock: between them, they make all uses
1358 * of pc->mem_cgroup safe.
1360 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1361 pc->mem_cgroup = memcg = root_mem_cgroup;
1363 mz = page_cgroup_zoneinfo(memcg, page);
1364 lruvec = &mz->lruvec;
1367 * Since a node can be onlined after the mem_cgroup was created,
1368 * we have to be prepared to initialize lruvec->zone here;
1369 * and if offlined then reonlined, we need to reinitialize it.
1371 if (unlikely(lruvec->zone != zone))
1372 lruvec->zone = zone;
1377 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1378 * @lruvec: mem_cgroup per zone lru vector
1379 * @lru: index of lru list the page is sitting on
1380 * @nr_pages: positive when adding or negative when removing
1382 * This function must be called when a page is added to or removed from an
1385 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1388 struct mem_cgroup_per_zone *mz;
1389 unsigned long *lru_size;
1391 if (mem_cgroup_disabled())
1394 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1395 lru_size = mz->lru_size + lru;
1396 *lru_size += nr_pages;
1397 VM_BUG_ON((long)(*lru_size) < 0);
1401 * Checks whether given mem is same or in the root_mem_cgroup's
1404 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1405 struct mem_cgroup *memcg)
1407 if (root_memcg == memcg)
1409 if (!root_memcg->use_hierarchy || !memcg)
1411 return css_is_ancestor(&memcg->css, &root_memcg->css);
1414 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1415 struct mem_cgroup *memcg)
1420 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1425 bool task_in_mem_cgroup(struct task_struct *task,
1426 const struct mem_cgroup *memcg)
1428 struct mem_cgroup *curr = NULL;
1429 struct task_struct *p;
1432 p = find_lock_task_mm(task);
1434 curr = try_get_mem_cgroup_from_mm(p->mm);
1438 * All threads may have already detached their mm's, but the oom
1439 * killer still needs to detect if they have already been oom
1440 * killed to prevent needlessly killing additional tasks.
1443 curr = mem_cgroup_from_task(task);
1445 css_get(&curr->css);
1451 * We should check use_hierarchy of "memcg" not "curr". Because checking
1452 * use_hierarchy of "curr" here make this function true if hierarchy is
1453 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1454 * hierarchy(even if use_hierarchy is disabled in "memcg").
1456 ret = mem_cgroup_same_or_subtree(memcg, curr);
1457 css_put(&curr->css);
1461 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1463 unsigned long inactive_ratio;
1464 unsigned long inactive;
1465 unsigned long active;
1468 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1469 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1471 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1473 inactive_ratio = int_sqrt(10 * gb);
1477 return inactive * inactive_ratio < active;
1480 #define mem_cgroup_from_res_counter(counter, member) \
1481 container_of(counter, struct mem_cgroup, member)
1484 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1485 * @memcg: the memory cgroup
1487 * Returns the maximum amount of memory @mem can be charged with, in
1490 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1492 unsigned long long margin;
1494 margin = res_counter_margin(&memcg->res);
1495 if (do_swap_account)
1496 margin = min(margin, res_counter_margin(&memcg->memsw));
1497 return margin >> PAGE_SHIFT;
1500 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1503 if (!css_parent(&memcg->css))
1504 return vm_swappiness;
1506 return memcg->swappiness;
1510 * memcg->moving_account is used for checking possibility that some thread is
1511 * calling move_account(). When a thread on CPU-A starts moving pages under
1512 * a memcg, other threads should check memcg->moving_account under
1513 * rcu_read_lock(), like this:
1517 * memcg->moving_account+1 if (memcg->mocing_account)
1519 * synchronize_rcu() update something.
1524 /* for quick checking without looking up memcg */
1525 atomic_t memcg_moving __read_mostly;
1527 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1529 atomic_inc(&memcg_moving);
1530 atomic_inc(&memcg->moving_account);
1534 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1537 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1538 * We check NULL in callee rather than caller.
1541 atomic_dec(&memcg_moving);
1542 atomic_dec(&memcg->moving_account);
1547 * 2 routines for checking "mem" is under move_account() or not.
1549 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1550 * is used for avoiding races in accounting. If true,
1551 * pc->mem_cgroup may be overwritten.
1553 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1554 * under hierarchy of moving cgroups. This is for
1555 * waiting at hith-memory prressure caused by "move".
1558 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1560 VM_BUG_ON(!rcu_read_lock_held());
1561 return atomic_read(&memcg->moving_account) > 0;
1564 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1566 struct mem_cgroup *from;
1567 struct mem_cgroup *to;
1570 * Unlike task_move routines, we access mc.to, mc.from not under
1571 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1573 spin_lock(&mc.lock);
1579 ret = mem_cgroup_same_or_subtree(memcg, from)
1580 || mem_cgroup_same_or_subtree(memcg, to);
1582 spin_unlock(&mc.lock);
1586 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1588 if (mc.moving_task && current != mc.moving_task) {
1589 if (mem_cgroup_under_move(memcg)) {
1591 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1592 /* moving charge context might have finished. */
1595 finish_wait(&mc.waitq, &wait);
1603 * Take this lock when
1604 * - a code tries to modify page's memcg while it's USED.
1605 * - a code tries to modify page state accounting in a memcg.
1606 * see mem_cgroup_stolen(), too.
1608 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1609 unsigned long *flags)
1611 spin_lock_irqsave(&memcg->move_lock, *flags);
1614 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1615 unsigned long *flags)
1617 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1620 #define K(x) ((x) << (PAGE_SHIFT-10))
1622 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1623 * @memcg: The memory cgroup that went over limit
1624 * @p: Task that is going to be killed
1626 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1629 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1631 struct cgroup *task_cgrp;
1632 struct cgroup *mem_cgrp;
1634 * Need a buffer in BSS, can't rely on allocations. The code relies
1635 * on the assumption that OOM is serialized for memory controller.
1636 * If this assumption is broken, revisit this code.
1638 static char memcg_name[PATH_MAX];
1640 struct mem_cgroup *iter;
1648 mem_cgrp = memcg->css.cgroup;
1649 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1651 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1654 * Unfortunately, we are unable to convert to a useful name
1655 * But we'll still print out the usage information
1662 pr_info("Task in %s killed", memcg_name);
1665 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1673 * Continues from above, so we don't need an KERN_ level
1675 pr_cont(" as a result of limit of %s\n", memcg_name);
1678 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1679 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1680 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1681 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1682 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1683 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1684 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1685 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1686 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1687 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1688 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1689 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1691 for_each_mem_cgroup_tree(iter, memcg) {
1692 pr_info("Memory cgroup stats");
1695 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1697 pr_cont(" for %s", memcg_name);
1701 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1702 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1704 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1705 K(mem_cgroup_read_stat(iter, i)));
1708 for (i = 0; i < NR_LRU_LISTS; i++)
1709 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1710 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1717 * This function returns the number of memcg under hierarchy tree. Returns
1718 * 1(self count) if no children.
1720 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1723 struct mem_cgroup *iter;
1725 for_each_mem_cgroup_tree(iter, memcg)
1731 * Return the memory (and swap, if configured) limit for a memcg.
1733 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1737 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1740 * Do not consider swap space if we cannot swap due to swappiness
1742 if (mem_cgroup_swappiness(memcg)) {
1745 limit += total_swap_pages << PAGE_SHIFT;
1746 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1749 * If memsw is finite and limits the amount of swap space
1750 * available to this memcg, return that limit.
1752 limit = min(limit, memsw);
1758 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1761 struct mem_cgroup *iter;
1762 unsigned long chosen_points = 0;
1763 unsigned long totalpages;
1764 unsigned int points = 0;
1765 struct task_struct *chosen = NULL;
1768 * If current has a pending SIGKILL or is exiting, then automatically
1769 * select it. The goal is to allow it to allocate so that it may
1770 * quickly exit and free its memory.
1772 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1773 set_thread_flag(TIF_MEMDIE);
1777 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1778 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1779 for_each_mem_cgroup_tree(iter, memcg) {
1780 struct css_task_iter it;
1781 struct task_struct *task;
1783 css_task_iter_start(&iter->css, &it);
1784 while ((task = css_task_iter_next(&it))) {
1785 switch (oom_scan_process_thread(task, totalpages, NULL,
1787 case OOM_SCAN_SELECT:
1789 put_task_struct(chosen);
1791 chosen_points = ULONG_MAX;
1792 get_task_struct(chosen);
1794 case OOM_SCAN_CONTINUE:
1796 case OOM_SCAN_ABORT:
1797 css_task_iter_end(&it);
1798 mem_cgroup_iter_break(memcg, iter);
1800 put_task_struct(chosen);
1805 points = oom_badness(task, memcg, NULL, totalpages);
1806 if (points > chosen_points) {
1808 put_task_struct(chosen);
1810 chosen_points = points;
1811 get_task_struct(chosen);
1814 css_task_iter_end(&it);
1819 points = chosen_points * 1000 / totalpages;
1820 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1821 NULL, "Memory cgroup out of memory");
1824 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1826 unsigned long flags)
1828 unsigned long total = 0;
1829 bool noswap = false;
1832 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1834 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1837 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1839 drain_all_stock_async(memcg);
1840 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1842 * Allow limit shrinkers, which are triggered directly
1843 * by userspace, to catch signals and stop reclaim
1844 * after minimal progress, regardless of the margin.
1846 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1848 if (mem_cgroup_margin(memcg))
1851 * If nothing was reclaimed after two attempts, there
1852 * may be no reclaimable pages in this hierarchy.
1861 * test_mem_cgroup_node_reclaimable
1862 * @memcg: the target memcg
1863 * @nid: the node ID to be checked.
1864 * @noswap : specify true here if the user wants flle only information.
1866 * This function returns whether the specified memcg contains any
1867 * reclaimable pages on a node. Returns true if there are any reclaimable
1868 * pages in the node.
1870 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1871 int nid, bool noswap)
1873 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1875 if (noswap || !total_swap_pages)
1877 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1882 #if MAX_NUMNODES > 1
1885 * Always updating the nodemask is not very good - even if we have an empty
1886 * list or the wrong list here, we can start from some node and traverse all
1887 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1890 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1894 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1895 * pagein/pageout changes since the last update.
1897 if (!atomic_read(&memcg->numainfo_events))
1899 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1902 /* make a nodemask where this memcg uses memory from */
1903 memcg->scan_nodes = node_states[N_MEMORY];
1905 for_each_node_mask(nid, node_states[N_MEMORY]) {
1907 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1908 node_clear(nid, memcg->scan_nodes);
1911 atomic_set(&memcg->numainfo_events, 0);
1912 atomic_set(&memcg->numainfo_updating, 0);
1916 * Selecting a node where we start reclaim from. Because what we need is just
1917 * reducing usage counter, start from anywhere is O,K. Considering
1918 * memory reclaim from current node, there are pros. and cons.
1920 * Freeing memory from current node means freeing memory from a node which
1921 * we'll use or we've used. So, it may make LRU bad. And if several threads
1922 * hit limits, it will see a contention on a node. But freeing from remote
1923 * node means more costs for memory reclaim because of memory latency.
1925 * Now, we use round-robin. Better algorithm is welcomed.
1927 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1931 mem_cgroup_may_update_nodemask(memcg);
1932 node = memcg->last_scanned_node;
1934 node = next_node(node, memcg->scan_nodes);
1935 if (node == MAX_NUMNODES)
1936 node = first_node(memcg->scan_nodes);
1938 * We call this when we hit limit, not when pages are added to LRU.
1939 * No LRU may hold pages because all pages are UNEVICTABLE or
1940 * memcg is too small and all pages are not on LRU. In that case,
1941 * we use curret node.
1943 if (unlikely(node == MAX_NUMNODES))
1944 node = numa_node_id();
1946 memcg->last_scanned_node = node;
1951 * Check all nodes whether it contains reclaimable pages or not.
1952 * For quick scan, we make use of scan_nodes. This will allow us to skip
1953 * unused nodes. But scan_nodes is lazily updated and may not cotain
1954 * enough new information. We need to do double check.
1956 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1961 * quick check...making use of scan_node.
1962 * We can skip unused nodes.
1964 if (!nodes_empty(memcg->scan_nodes)) {
1965 for (nid = first_node(memcg->scan_nodes);
1967 nid = next_node(nid, memcg->scan_nodes)) {
1969 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1974 * Check rest of nodes.
1976 for_each_node_state(nid, N_MEMORY) {
1977 if (node_isset(nid, memcg->scan_nodes))
1979 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1986 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1991 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1993 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1997 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2000 unsigned long *total_scanned)
2002 struct mem_cgroup *victim = NULL;
2005 unsigned long excess;
2006 unsigned long nr_scanned;
2007 struct mem_cgroup_reclaim_cookie reclaim = {
2012 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2015 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2020 * If we have not been able to reclaim
2021 * anything, it might because there are
2022 * no reclaimable pages under this hierarchy
2027 * We want to do more targeted reclaim.
2028 * excess >> 2 is not to excessive so as to
2029 * reclaim too much, nor too less that we keep
2030 * coming back to reclaim from this cgroup
2032 if (total >= (excess >> 2) ||
2033 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2038 if (!mem_cgroup_reclaimable(victim, false))
2040 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2042 *total_scanned += nr_scanned;
2043 if (!res_counter_soft_limit_excess(&root_memcg->res))
2046 mem_cgroup_iter_break(root_memcg, victim);
2050 #ifdef CONFIG_LOCKDEP
2051 static struct lockdep_map memcg_oom_lock_dep_map = {
2052 .name = "memcg_oom_lock",
2056 static DEFINE_SPINLOCK(memcg_oom_lock);
2059 * Check OOM-Killer is already running under our hierarchy.
2060 * If someone is running, return false.
2062 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2064 struct mem_cgroup *iter, *failed = NULL;
2066 spin_lock(&memcg_oom_lock);
2068 for_each_mem_cgroup_tree(iter, memcg) {
2069 if (iter->oom_lock) {
2071 * this subtree of our hierarchy is already locked
2072 * so we cannot give a lock.
2075 mem_cgroup_iter_break(memcg, iter);
2078 iter->oom_lock = true;
2083 * OK, we failed to lock the whole subtree so we have
2084 * to clean up what we set up to the failing subtree
2086 for_each_mem_cgroup_tree(iter, memcg) {
2087 if (iter == failed) {
2088 mem_cgroup_iter_break(memcg, iter);
2091 iter->oom_lock = false;
2094 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2096 spin_unlock(&memcg_oom_lock);
2101 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2103 struct mem_cgroup *iter;
2105 spin_lock(&memcg_oom_lock);
2106 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2107 for_each_mem_cgroup_tree(iter, memcg)
2108 iter->oom_lock = false;
2109 spin_unlock(&memcg_oom_lock);
2112 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2114 struct mem_cgroup *iter;
2116 for_each_mem_cgroup_tree(iter, memcg)
2117 atomic_inc(&iter->under_oom);
2120 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2122 struct mem_cgroup *iter;
2125 * When a new child is created while the hierarchy is under oom,
2126 * mem_cgroup_oom_lock() may not be called. We have to use
2127 * atomic_add_unless() here.
2129 for_each_mem_cgroup_tree(iter, memcg)
2130 atomic_add_unless(&iter->under_oom, -1, 0);
2133 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2135 struct oom_wait_info {
2136 struct mem_cgroup *memcg;
2140 static int memcg_oom_wake_function(wait_queue_t *wait,
2141 unsigned mode, int sync, void *arg)
2143 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2144 struct mem_cgroup *oom_wait_memcg;
2145 struct oom_wait_info *oom_wait_info;
2147 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2148 oom_wait_memcg = oom_wait_info->memcg;
2151 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2152 * Then we can use css_is_ancestor without taking care of RCU.
2154 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2155 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2157 return autoremove_wake_function(wait, mode, sync, arg);
2160 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2162 atomic_inc(&memcg->oom_wakeups);
2163 /* for filtering, pass "memcg" as argument. */
2164 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2167 static void memcg_oom_recover(struct mem_cgroup *memcg)
2169 if (memcg && atomic_read(&memcg->under_oom))
2170 memcg_wakeup_oom(memcg);
2173 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2175 if (!current->memcg_oom.may_oom)
2178 * We are in the middle of the charge context here, so we
2179 * don't want to block when potentially sitting on a callstack
2180 * that holds all kinds of filesystem and mm locks.
2182 * Also, the caller may handle a failed allocation gracefully
2183 * (like optional page cache readahead) and so an OOM killer
2184 * invocation might not even be necessary.
2186 * That's why we don't do anything here except remember the
2187 * OOM context and then deal with it at the end of the page
2188 * fault when the stack is unwound, the locks are released,
2189 * and when we know whether the fault was overall successful.
2191 css_get(&memcg->css);
2192 current->memcg_oom.memcg = memcg;
2193 current->memcg_oom.gfp_mask = mask;
2194 current->memcg_oom.order = order;
2198 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2199 * @handle: actually kill/wait or just clean up the OOM state
2201 * This has to be called at the end of a page fault if the memcg OOM
2202 * handler was enabled.
2204 * Memcg supports userspace OOM handling where failed allocations must
2205 * sleep on a waitqueue until the userspace task resolves the
2206 * situation. Sleeping directly in the charge context with all kinds
2207 * of locks held is not a good idea, instead we remember an OOM state
2208 * in the task and mem_cgroup_oom_synchronize() has to be called at
2209 * the end of the page fault to complete the OOM handling.
2211 * Returns %true if an ongoing memcg OOM situation was detected and
2212 * completed, %false otherwise.
2214 bool mem_cgroup_oom_synchronize(bool handle)
2216 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2217 struct oom_wait_info owait;
2220 /* OOM is global, do not handle */
2227 owait.memcg = memcg;
2228 owait.wait.flags = 0;
2229 owait.wait.func = memcg_oom_wake_function;
2230 owait.wait.private = current;
2231 INIT_LIST_HEAD(&owait.wait.task_list);
2233 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2234 mem_cgroup_mark_under_oom(memcg);
2236 locked = mem_cgroup_oom_trylock(memcg);
2239 mem_cgroup_oom_notify(memcg);
2241 if (locked && !memcg->oom_kill_disable) {
2242 mem_cgroup_unmark_under_oom(memcg);
2243 finish_wait(&memcg_oom_waitq, &owait.wait);
2244 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2245 current->memcg_oom.order);
2248 mem_cgroup_unmark_under_oom(memcg);
2249 finish_wait(&memcg_oom_waitq, &owait.wait);
2253 mem_cgroup_oom_unlock(memcg);
2255 * There is no guarantee that an OOM-lock contender
2256 * sees the wakeups triggered by the OOM kill
2257 * uncharges. Wake any sleepers explicitely.
2259 memcg_oom_recover(memcg);
2262 current->memcg_oom.memcg = NULL;
2263 css_put(&memcg->css);
2268 * Currently used to update mapped file statistics, but the routine can be
2269 * generalized to update other statistics as well.
2271 * Notes: Race condition
2273 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2274 * it tends to be costly. But considering some conditions, we doesn't need
2275 * to do so _always_.
2277 * Considering "charge", lock_page_cgroup() is not required because all
2278 * file-stat operations happen after a page is attached to radix-tree. There
2279 * are no race with "charge".
2281 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2282 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2283 * if there are race with "uncharge". Statistics itself is properly handled
2286 * Considering "move", this is an only case we see a race. To make the race
2287 * small, we check mm->moving_account and detect there are possibility of race
2288 * If there is, we take a lock.
2291 void __mem_cgroup_begin_update_page_stat(struct page *page,
2292 bool *locked, unsigned long *flags)
2294 struct mem_cgroup *memcg;
2295 struct page_cgroup *pc;
2297 pc = lookup_page_cgroup(page);
2299 memcg = pc->mem_cgroup;
2300 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2303 * If this memory cgroup is not under account moving, we don't
2304 * need to take move_lock_mem_cgroup(). Because we already hold
2305 * rcu_read_lock(), any calls to move_account will be delayed until
2306 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2308 if (!mem_cgroup_stolen(memcg))
2311 move_lock_mem_cgroup(memcg, flags);
2312 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2313 move_unlock_mem_cgroup(memcg, flags);
2319 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2321 struct page_cgroup *pc = lookup_page_cgroup(page);
2324 * It's guaranteed that pc->mem_cgroup never changes while
2325 * lock is held because a routine modifies pc->mem_cgroup
2326 * should take move_lock_mem_cgroup().
2328 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2331 void mem_cgroup_update_page_stat(struct page *page,
2332 enum mem_cgroup_stat_index idx, int val)
2334 struct mem_cgroup *memcg;
2335 struct page_cgroup *pc = lookup_page_cgroup(page);
2336 unsigned long uninitialized_var(flags);
2338 if (mem_cgroup_disabled())
2341 VM_BUG_ON(!rcu_read_lock_held());
2342 memcg = pc->mem_cgroup;
2343 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2346 this_cpu_add(memcg->stat->count[idx], val);
2350 * size of first charge trial. "32" comes from vmscan.c's magic value.
2351 * TODO: maybe necessary to use big numbers in big irons.
2353 #define CHARGE_BATCH 32U
2354 struct memcg_stock_pcp {
2355 struct mem_cgroup *cached; /* this never be root cgroup */
2356 unsigned int nr_pages;
2357 struct work_struct work;
2358 unsigned long flags;
2359 #define FLUSHING_CACHED_CHARGE 0
2361 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2362 static DEFINE_MUTEX(percpu_charge_mutex);
2365 * consume_stock: Try to consume stocked charge on this cpu.
2366 * @memcg: memcg to consume from.
2367 * @nr_pages: how many pages to charge.
2369 * The charges will only happen if @memcg matches the current cpu's memcg
2370 * stock, and at least @nr_pages are available in that stock. Failure to
2371 * service an allocation will refill the stock.
2373 * returns true if successful, false otherwise.
2375 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2377 struct memcg_stock_pcp *stock;
2380 if (nr_pages > CHARGE_BATCH)
2383 stock = &get_cpu_var(memcg_stock);
2384 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2385 stock->nr_pages -= nr_pages;
2386 else /* need to call res_counter_charge */
2388 put_cpu_var(memcg_stock);
2393 * Returns stocks cached in percpu to res_counter and reset cached information.
2395 static void drain_stock(struct memcg_stock_pcp *stock)
2397 struct mem_cgroup *old = stock->cached;
2399 if (stock->nr_pages) {
2400 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2402 res_counter_uncharge(&old->res, bytes);
2403 if (do_swap_account)
2404 res_counter_uncharge(&old->memsw, bytes);
2405 stock->nr_pages = 0;
2407 stock->cached = NULL;
2411 * This must be called under preempt disabled or must be called by
2412 * a thread which is pinned to local cpu.
2414 static void drain_local_stock(struct work_struct *dummy)
2416 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2418 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2421 static void __init memcg_stock_init(void)
2425 for_each_possible_cpu(cpu) {
2426 struct memcg_stock_pcp *stock =
2427 &per_cpu(memcg_stock, cpu);
2428 INIT_WORK(&stock->work, drain_local_stock);
2433 * Cache charges(val) which is from res_counter, to local per_cpu area.
2434 * This will be consumed by consume_stock() function, later.
2436 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2438 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2440 if (stock->cached != memcg) { /* reset if necessary */
2442 stock->cached = memcg;
2444 stock->nr_pages += nr_pages;
2445 put_cpu_var(memcg_stock);
2449 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2450 * of the hierarchy under it. sync flag says whether we should block
2451 * until the work is done.
2453 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2457 /* Notify other cpus that system-wide "drain" is running */
2460 for_each_online_cpu(cpu) {
2461 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2462 struct mem_cgroup *memcg;
2464 memcg = stock->cached;
2465 if (!memcg || !stock->nr_pages)
2467 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2469 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2471 drain_local_stock(&stock->work);
2473 schedule_work_on(cpu, &stock->work);
2481 for_each_online_cpu(cpu) {
2482 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2483 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2484 flush_work(&stock->work);
2491 * Tries to drain stocked charges in other cpus. This function is asynchronous
2492 * and just put a work per cpu for draining localy on each cpu. Caller can
2493 * expects some charges will be back to res_counter later but cannot wait for
2496 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2499 * If someone calls draining, avoid adding more kworker runs.
2501 if (!mutex_trylock(&percpu_charge_mutex))
2503 drain_all_stock(root_memcg, false);
2504 mutex_unlock(&percpu_charge_mutex);
2507 /* This is a synchronous drain interface. */
2508 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2510 /* called when force_empty is called */
2511 mutex_lock(&percpu_charge_mutex);
2512 drain_all_stock(root_memcg, true);
2513 mutex_unlock(&percpu_charge_mutex);
2517 * This function drains percpu counter value from DEAD cpu and
2518 * move it to local cpu. Note that this function can be preempted.
2520 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2524 spin_lock(&memcg->pcp_counter_lock);
2525 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2526 long x = per_cpu(memcg->stat->count[i], cpu);
2528 per_cpu(memcg->stat->count[i], cpu) = 0;
2529 memcg->nocpu_base.count[i] += x;
2531 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2532 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2534 per_cpu(memcg->stat->events[i], cpu) = 0;
2535 memcg->nocpu_base.events[i] += x;
2537 spin_unlock(&memcg->pcp_counter_lock);
2540 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2541 unsigned long action,
2544 int cpu = (unsigned long)hcpu;
2545 struct memcg_stock_pcp *stock;
2546 struct mem_cgroup *iter;
2548 if (action == CPU_ONLINE)
2551 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2554 for_each_mem_cgroup(iter)
2555 mem_cgroup_drain_pcp_counter(iter, cpu);
2557 stock = &per_cpu(memcg_stock, cpu);
2563 /* See __mem_cgroup_try_charge() for details */
2565 CHARGE_OK, /* success */
2566 CHARGE_RETRY, /* need to retry but retry is not bad */
2567 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2568 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2571 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2572 unsigned int nr_pages, unsigned int min_pages,
2575 unsigned long csize = nr_pages * PAGE_SIZE;
2576 struct mem_cgroup *mem_over_limit;
2577 struct res_counter *fail_res;
2578 unsigned long flags = 0;
2581 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2584 if (!do_swap_account)
2586 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2590 res_counter_uncharge(&memcg->res, csize);
2591 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2592 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2594 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2596 * Never reclaim on behalf of optional batching, retry with a
2597 * single page instead.
2599 if (nr_pages > min_pages)
2600 return CHARGE_RETRY;
2602 if (!(gfp_mask & __GFP_WAIT))
2603 return CHARGE_WOULDBLOCK;
2605 if (gfp_mask & __GFP_NORETRY)
2606 return CHARGE_NOMEM;
2608 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2609 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2610 return CHARGE_RETRY;
2612 * Even though the limit is exceeded at this point, reclaim
2613 * may have been able to free some pages. Retry the charge
2614 * before killing the task.
2616 * Only for regular pages, though: huge pages are rather
2617 * unlikely to succeed so close to the limit, and we fall back
2618 * to regular pages anyway in case of failure.
2620 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2621 return CHARGE_RETRY;
2624 * At task move, charge accounts can be doubly counted. So, it's
2625 * better to wait until the end of task_move if something is going on.
2627 if (mem_cgroup_wait_acct_move(mem_over_limit))
2628 return CHARGE_RETRY;
2631 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2633 return CHARGE_NOMEM;
2637 * __mem_cgroup_try_charge() does
2638 * 1. detect memcg to be charged against from passed *mm and *ptr,
2639 * 2. update res_counter
2640 * 3. call memory reclaim if necessary.
2642 * In some special case, if the task is fatal, fatal_signal_pending() or
2643 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2644 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2645 * as possible without any hazards. 2: all pages should have a valid
2646 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2647 * pointer, that is treated as a charge to root_mem_cgroup.
2649 * So __mem_cgroup_try_charge() will return
2650 * 0 ... on success, filling *ptr with a valid memcg pointer.
2651 * -ENOMEM ... charge failure because of resource limits.
2652 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2654 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2655 * the oom-killer can be invoked.
2657 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2659 unsigned int nr_pages,
2660 struct mem_cgroup **ptr,
2663 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2664 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2665 struct mem_cgroup *memcg = NULL;
2669 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2670 * in system level. So, allow to go ahead dying process in addition to
2673 if (unlikely(test_thread_flag(TIF_MEMDIE)
2674 || fatal_signal_pending(current)))
2677 if (unlikely(task_in_memcg_oom(current)))
2681 * We always charge the cgroup the mm_struct belongs to.
2682 * The mm_struct's mem_cgroup changes on task migration if the
2683 * thread group leader migrates. It's possible that mm is not
2684 * set, if so charge the root memcg (happens for pagecache usage).
2687 *ptr = root_mem_cgroup;
2689 if (*ptr) { /* css should be a valid one */
2691 if (mem_cgroup_is_root(memcg))
2693 if (consume_stock(memcg, nr_pages))
2695 css_get(&memcg->css);
2697 struct task_struct *p;
2700 p = rcu_dereference(mm->owner);
2702 * Because we don't have task_lock(), "p" can exit.
2703 * In that case, "memcg" can point to root or p can be NULL with
2704 * race with swapoff. Then, we have small risk of mis-accouning.
2705 * But such kind of mis-account by race always happens because
2706 * we don't have cgroup_mutex(). It's overkill and we allo that
2708 * (*) swapoff at el will charge against mm-struct not against
2709 * task-struct. So, mm->owner can be NULL.
2711 memcg = mem_cgroup_from_task(p);
2713 memcg = root_mem_cgroup;
2714 if (mem_cgroup_is_root(memcg)) {
2718 if (consume_stock(memcg, nr_pages)) {
2720 * It seems dagerous to access memcg without css_get().
2721 * But considering how consume_stok works, it's not
2722 * necessary. If consume_stock success, some charges
2723 * from this memcg are cached on this cpu. So, we
2724 * don't need to call css_get()/css_tryget() before
2725 * calling consume_stock().
2730 /* after here, we may be blocked. we need to get refcnt */
2731 if (!css_tryget(&memcg->css)) {
2739 bool invoke_oom = oom && !nr_oom_retries;
2741 /* If killed, bypass charge */
2742 if (fatal_signal_pending(current)) {
2743 css_put(&memcg->css);
2747 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2748 nr_pages, invoke_oom);
2752 case CHARGE_RETRY: /* not in OOM situation but retry */
2754 css_put(&memcg->css);
2757 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2758 css_put(&memcg->css);
2760 case CHARGE_NOMEM: /* OOM routine works */
2761 if (!oom || invoke_oom) {
2762 css_put(&memcg->css);
2768 } while (ret != CHARGE_OK);
2770 if (batch > nr_pages)
2771 refill_stock(memcg, batch - nr_pages);
2772 css_put(&memcg->css);
2777 if (!(gfp_mask & __GFP_NOFAIL)) {
2782 *ptr = root_mem_cgroup;
2787 * Somemtimes we have to undo a charge we got by try_charge().
2788 * This function is for that and do uncharge, put css's refcnt.
2789 * gotten by try_charge().
2791 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2792 unsigned int nr_pages)
2794 if (!mem_cgroup_is_root(memcg)) {
2795 unsigned long bytes = nr_pages * PAGE_SIZE;
2797 res_counter_uncharge(&memcg->res, bytes);
2798 if (do_swap_account)
2799 res_counter_uncharge(&memcg->memsw, bytes);
2804 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2805 * This is useful when moving usage to parent cgroup.
2807 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2808 unsigned int nr_pages)
2810 unsigned long bytes = nr_pages * PAGE_SIZE;
2812 if (mem_cgroup_is_root(memcg))
2815 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2816 if (do_swap_account)
2817 res_counter_uncharge_until(&memcg->memsw,
2818 memcg->memsw.parent, bytes);
2822 * A helper function to get mem_cgroup from ID. must be called under
2823 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2824 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2825 * called against removed memcg.)
2827 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2829 struct cgroup_subsys_state *css;
2831 /* ID 0 is unused ID */
2834 css = css_lookup(&mem_cgroup_subsys, id);
2837 return mem_cgroup_from_css(css);
2840 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2842 struct mem_cgroup *memcg = NULL;
2843 struct page_cgroup *pc;
2847 VM_BUG_ON(!PageLocked(page));
2849 pc = lookup_page_cgroup(page);
2850 lock_page_cgroup(pc);
2851 if (PageCgroupUsed(pc)) {
2852 memcg = pc->mem_cgroup;
2853 if (memcg && !css_tryget(&memcg->css))
2855 } else if (PageSwapCache(page)) {
2856 ent.val = page_private(page);
2857 id = lookup_swap_cgroup_id(ent);
2859 memcg = mem_cgroup_lookup(id);
2860 if (memcg && !css_tryget(&memcg->css))
2864 unlock_page_cgroup(pc);
2868 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2870 unsigned int nr_pages,
2871 enum charge_type ctype,
2874 struct page_cgroup *pc = lookup_page_cgroup(page);
2875 struct zone *uninitialized_var(zone);
2876 struct lruvec *lruvec;
2877 bool was_on_lru = false;
2880 lock_page_cgroup(pc);
2881 VM_BUG_ON(PageCgroupUsed(pc));
2883 * we don't need page_cgroup_lock about tail pages, becase they are not
2884 * accessed by any other context at this point.
2888 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2889 * may already be on some other mem_cgroup's LRU. Take care of it.
2892 zone = page_zone(page);
2893 spin_lock_irq(&zone->lru_lock);
2894 if (PageLRU(page)) {
2895 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2897 del_page_from_lru_list(page, lruvec, page_lru(page));
2902 pc->mem_cgroup = memcg;
2904 * We access a page_cgroup asynchronously without lock_page_cgroup().
2905 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2906 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2907 * before USED bit, we need memory barrier here.
2908 * See mem_cgroup_add_lru_list(), etc.
2911 SetPageCgroupUsed(pc);
2915 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2916 VM_BUG_ON(PageLRU(page));
2918 add_page_to_lru_list(page, lruvec, page_lru(page));
2920 spin_unlock_irq(&zone->lru_lock);
2923 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2928 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2929 unlock_page_cgroup(pc);
2932 * "charge_statistics" updated event counter. Then, check it.
2933 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2934 * if they exceeds softlimit.
2936 memcg_check_events(memcg, page);
2939 static DEFINE_MUTEX(set_limit_mutex);
2941 #ifdef CONFIG_MEMCG_KMEM
2942 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2944 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2945 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2949 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2950 * in the memcg_cache_params struct.
2952 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2954 struct kmem_cache *cachep;
2956 VM_BUG_ON(p->is_root_cache);
2957 cachep = p->root_cache;
2958 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2961 #ifdef CONFIG_SLABINFO
2962 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2963 struct cftype *cft, struct seq_file *m)
2965 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2966 struct memcg_cache_params *params;
2968 if (!memcg_can_account_kmem(memcg))
2971 print_slabinfo_header(m);
2973 mutex_lock(&memcg->slab_caches_mutex);
2974 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2975 cache_show(memcg_params_to_cache(params), m);
2976 mutex_unlock(&memcg->slab_caches_mutex);
2982 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2984 struct res_counter *fail_res;
2985 struct mem_cgroup *_memcg;
2989 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2994 * Conditions under which we can wait for the oom_killer. Those are
2995 * the same conditions tested by the core page allocator
2997 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
3000 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3003 if (ret == -EINTR) {
3005 * __mem_cgroup_try_charge() chosed to bypass to root due to
3006 * OOM kill or fatal signal. Since our only options are to
3007 * either fail the allocation or charge it to this cgroup, do
3008 * it as a temporary condition. But we can't fail. From a
3009 * kmem/slab perspective, the cache has already been selected,
3010 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3013 * This condition will only trigger if the task entered
3014 * memcg_charge_kmem in a sane state, but was OOM-killed during
3015 * __mem_cgroup_try_charge() above. Tasks that were already
3016 * dying when the allocation triggers should have been already
3017 * directed to the root cgroup in memcontrol.h
3019 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3020 if (do_swap_account)
3021 res_counter_charge_nofail(&memcg->memsw, size,
3025 res_counter_uncharge(&memcg->kmem, size);
3030 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3032 res_counter_uncharge(&memcg->res, size);
3033 if (do_swap_account)
3034 res_counter_uncharge(&memcg->memsw, size);
3037 if (res_counter_uncharge(&memcg->kmem, size))
3041 * Releases a reference taken in kmem_cgroup_css_offline in case
3042 * this last uncharge is racing with the offlining code or it is
3043 * outliving the memcg existence.
3045 * The memory barrier imposed by test&clear is paired with the
3046 * explicit one in memcg_kmem_mark_dead().
3048 if (memcg_kmem_test_and_clear_dead(memcg))
3049 css_put(&memcg->css);
3052 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3057 mutex_lock(&memcg->slab_caches_mutex);
3058 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3059 mutex_unlock(&memcg->slab_caches_mutex);
3063 * helper for acessing a memcg's index. It will be used as an index in the
3064 * child cache array in kmem_cache, and also to derive its name. This function
3065 * will return -1 when this is not a kmem-limited memcg.
3067 int memcg_cache_id(struct mem_cgroup *memcg)
3069 return memcg ? memcg->kmemcg_id : -1;
3073 * This ends up being protected by the set_limit mutex, during normal
3074 * operation, because that is its main call site.
3076 * But when we create a new cache, we can call this as well if its parent
3077 * is kmem-limited. That will have to hold set_limit_mutex as well.
3079 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3083 num = ida_simple_get(&kmem_limited_groups,
3084 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3088 * After this point, kmem_accounted (that we test atomically in
3089 * the beginning of this conditional), is no longer 0. This
3090 * guarantees only one process will set the following boolean
3091 * to true. We don't need test_and_set because we're protected
3092 * by the set_limit_mutex anyway.
3094 memcg_kmem_set_activated(memcg);
3096 ret = memcg_update_all_caches(num+1);
3098 ida_simple_remove(&kmem_limited_groups, num);
3099 memcg_kmem_clear_activated(memcg);
3103 memcg->kmemcg_id = num;
3104 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3105 mutex_init(&memcg->slab_caches_mutex);
3109 static size_t memcg_caches_array_size(int num_groups)
3112 if (num_groups <= 0)
3115 size = 2 * num_groups;
3116 if (size < MEMCG_CACHES_MIN_SIZE)
3117 size = MEMCG_CACHES_MIN_SIZE;
3118 else if (size > MEMCG_CACHES_MAX_SIZE)
3119 size = MEMCG_CACHES_MAX_SIZE;
3125 * We should update the current array size iff all caches updates succeed. This
3126 * can only be done from the slab side. The slab mutex needs to be held when
3129 void memcg_update_array_size(int num)
3131 if (num > memcg_limited_groups_array_size)
3132 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3135 static void kmem_cache_destroy_work_func(struct work_struct *w);
3137 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3139 struct memcg_cache_params *cur_params = s->memcg_params;
3141 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3143 if (num_groups > memcg_limited_groups_array_size) {
3145 ssize_t size = memcg_caches_array_size(num_groups);
3147 size *= sizeof(void *);
3148 size += offsetof(struct memcg_cache_params, memcg_caches);
3150 s->memcg_params = kzalloc(size, GFP_KERNEL);
3151 if (!s->memcg_params) {
3152 s->memcg_params = cur_params;
3156 s->memcg_params->is_root_cache = true;
3159 * There is the chance it will be bigger than
3160 * memcg_limited_groups_array_size, if we failed an allocation
3161 * in a cache, in which case all caches updated before it, will
3162 * have a bigger array.
3164 * But if that is the case, the data after
3165 * memcg_limited_groups_array_size is certainly unused
3167 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3168 if (!cur_params->memcg_caches[i])
3170 s->memcg_params->memcg_caches[i] =
3171 cur_params->memcg_caches[i];
3175 * Ideally, we would wait until all caches succeed, and only
3176 * then free the old one. But this is not worth the extra
3177 * pointer per-cache we'd have to have for this.
3179 * It is not a big deal if some caches are left with a size
3180 * bigger than the others. And all updates will reset this
3188 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3189 struct kmem_cache *root_cache)
3193 if (!memcg_kmem_enabled())
3197 size = offsetof(struct memcg_cache_params, memcg_caches);
3198 size += memcg_limited_groups_array_size * sizeof(void *);
3200 size = sizeof(struct memcg_cache_params);
3202 s->memcg_params = kzalloc(size, GFP_KERNEL);
3203 if (!s->memcg_params)
3207 s->memcg_params->memcg = memcg;
3208 s->memcg_params->root_cache = root_cache;
3209 INIT_WORK(&s->memcg_params->destroy,
3210 kmem_cache_destroy_work_func);
3212 s->memcg_params->is_root_cache = true;
3217 void memcg_release_cache(struct kmem_cache *s)
3219 struct kmem_cache *root;
3220 struct mem_cgroup *memcg;
3224 * This happens, for instance, when a root cache goes away before we
3227 if (!s->memcg_params)
3230 if (s->memcg_params->is_root_cache)
3233 memcg = s->memcg_params->memcg;
3234 id = memcg_cache_id(memcg);
3236 root = s->memcg_params->root_cache;
3237 root->memcg_params->memcg_caches[id] = NULL;
3239 mutex_lock(&memcg->slab_caches_mutex);
3240 list_del(&s->memcg_params->list);
3241 mutex_unlock(&memcg->slab_caches_mutex);
3243 css_put(&memcg->css);
3245 kfree(s->memcg_params);
3249 * During the creation a new cache, we need to disable our accounting mechanism
3250 * altogether. This is true even if we are not creating, but rather just
3251 * enqueing new caches to be created.
3253 * This is because that process will trigger allocations; some visible, like
3254 * explicit kmallocs to auxiliary data structures, name strings and internal
3255 * cache structures; some well concealed, like INIT_WORK() that can allocate
3256 * objects during debug.
3258 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3259 * to it. This may not be a bounded recursion: since the first cache creation
3260 * failed to complete (waiting on the allocation), we'll just try to create the
3261 * cache again, failing at the same point.
3263 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3264 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3265 * inside the following two functions.
3267 static inline void memcg_stop_kmem_account(void)
3269 VM_BUG_ON(!current->mm);
3270 current->memcg_kmem_skip_account++;
3273 static inline void memcg_resume_kmem_account(void)
3275 VM_BUG_ON(!current->mm);
3276 current->memcg_kmem_skip_account--;
3279 static void kmem_cache_destroy_work_func(struct work_struct *w)
3281 struct kmem_cache *cachep;
3282 struct memcg_cache_params *p;
3284 p = container_of(w, struct memcg_cache_params, destroy);
3286 cachep = memcg_params_to_cache(p);
3289 * If we get down to 0 after shrink, we could delete right away.
3290 * However, memcg_release_pages() already puts us back in the workqueue
3291 * in that case. If we proceed deleting, we'll get a dangling
3292 * reference, and removing the object from the workqueue in that case
3293 * is unnecessary complication. We are not a fast path.
3295 * Note that this case is fundamentally different from racing with
3296 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3297 * kmem_cache_shrink, not only we would be reinserting a dead cache
3298 * into the queue, but doing so from inside the worker racing to
3301 * So if we aren't down to zero, we'll just schedule a worker and try
3304 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3305 kmem_cache_shrink(cachep);
3306 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3309 kmem_cache_destroy(cachep);
3312 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3314 if (!cachep->memcg_params->dead)
3318 * There are many ways in which we can get here.
3320 * We can get to a memory-pressure situation while the delayed work is
3321 * still pending to run. The vmscan shrinkers can then release all
3322 * cache memory and get us to destruction. If this is the case, we'll
3323 * be executed twice, which is a bug (the second time will execute over
3324 * bogus data). In this case, cancelling the work should be fine.
3326 * But we can also get here from the worker itself, if
3327 * kmem_cache_shrink is enough to shake all the remaining objects and
3328 * get the page count to 0. In this case, we'll deadlock if we try to
3329 * cancel the work (the worker runs with an internal lock held, which
3330 * is the same lock we would hold for cancel_work_sync().)
3332 * Since we can't possibly know who got us here, just refrain from
3333 * running if there is already work pending
3335 if (work_pending(&cachep->memcg_params->destroy))
3338 * We have to defer the actual destroying to a workqueue, because
3339 * we might currently be in a context that cannot sleep.
3341 schedule_work(&cachep->memcg_params->destroy);
3345 * This lock protects updaters, not readers. We want readers to be as fast as
3346 * they can, and they will either see NULL or a valid cache value. Our model
3347 * allow them to see NULL, in which case the root memcg will be selected.
3349 * We need this lock because multiple allocations to the same cache from a non
3350 * will span more than one worker. Only one of them can create the cache.
3352 static DEFINE_MUTEX(memcg_cache_mutex);
3355 * Called with memcg_cache_mutex held
3357 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3358 struct kmem_cache *s)
3360 struct kmem_cache *new;
3361 static char *tmp_name = NULL;
3363 lockdep_assert_held(&memcg_cache_mutex);
3366 * kmem_cache_create_memcg duplicates the given name and
3367 * cgroup_name for this name requires RCU context.
3368 * This static temporary buffer is used to prevent from
3369 * pointless shortliving allocation.
3372 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3378 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3379 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3382 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3383 (s->flags & ~SLAB_PANIC), s->ctor, s);
3386 new->allocflags |= __GFP_KMEMCG;
3391 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3392 struct kmem_cache *cachep)
3394 struct kmem_cache *new_cachep;
3397 BUG_ON(!memcg_can_account_kmem(memcg));
3399 idx = memcg_cache_id(memcg);
3401 mutex_lock(&memcg_cache_mutex);
3402 new_cachep = cachep->memcg_params->memcg_caches[idx];
3404 css_put(&memcg->css);
3408 new_cachep = kmem_cache_dup(memcg, cachep);
3409 if (new_cachep == NULL) {
3410 new_cachep = cachep;
3411 css_put(&memcg->css);
3415 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3417 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3419 * the readers won't lock, make sure everybody sees the updated value,
3420 * so they won't put stuff in the queue again for no reason
3424 mutex_unlock(&memcg_cache_mutex);
3428 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3430 struct kmem_cache *c;
3433 if (!s->memcg_params)
3435 if (!s->memcg_params->is_root_cache)
3439 * If the cache is being destroyed, we trust that there is no one else
3440 * requesting objects from it. Even if there are, the sanity checks in
3441 * kmem_cache_destroy should caught this ill-case.
3443 * Still, we don't want anyone else freeing memcg_caches under our
3444 * noses, which can happen if a new memcg comes to life. As usual,
3445 * we'll take the set_limit_mutex to protect ourselves against this.
3447 mutex_lock(&set_limit_mutex);
3448 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3449 c = s->memcg_params->memcg_caches[i];
3454 * We will now manually delete the caches, so to avoid races
3455 * we need to cancel all pending destruction workers and
3456 * proceed with destruction ourselves.
3458 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3459 * and that could spawn the workers again: it is likely that
3460 * the cache still have active pages until this very moment.
3461 * This would lead us back to mem_cgroup_destroy_cache.
3463 * But that will not execute at all if the "dead" flag is not
3464 * set, so flip it down to guarantee we are in control.
3466 c->memcg_params->dead = false;
3467 cancel_work_sync(&c->memcg_params->destroy);
3468 kmem_cache_destroy(c);
3470 mutex_unlock(&set_limit_mutex);
3473 struct create_work {
3474 struct mem_cgroup *memcg;
3475 struct kmem_cache *cachep;
3476 struct work_struct work;
3479 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3481 struct kmem_cache *cachep;
3482 struct memcg_cache_params *params;
3484 if (!memcg_kmem_is_active(memcg))
3487 mutex_lock(&memcg->slab_caches_mutex);
3488 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3489 cachep = memcg_params_to_cache(params);
3490 cachep->memcg_params->dead = true;
3491 schedule_work(&cachep->memcg_params->destroy);
3493 mutex_unlock(&memcg->slab_caches_mutex);
3496 static void memcg_create_cache_work_func(struct work_struct *w)
3498 struct create_work *cw;
3500 cw = container_of(w, struct create_work, work);
3501 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3506 * Enqueue the creation of a per-memcg kmem_cache.
3508 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3509 struct kmem_cache *cachep)
3511 struct create_work *cw;
3513 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3515 css_put(&memcg->css);
3520 cw->cachep = cachep;
3522 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3523 schedule_work(&cw->work);
3526 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3527 struct kmem_cache *cachep)
3530 * We need to stop accounting when we kmalloc, because if the
3531 * corresponding kmalloc cache is not yet created, the first allocation
3532 * in __memcg_create_cache_enqueue will recurse.
3534 * However, it is better to enclose the whole function. Depending on
3535 * the debugging options enabled, INIT_WORK(), for instance, can
3536 * trigger an allocation. This too, will make us recurse. Because at
3537 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3538 * the safest choice is to do it like this, wrapping the whole function.
3540 memcg_stop_kmem_account();
3541 __memcg_create_cache_enqueue(memcg, cachep);
3542 memcg_resume_kmem_account();
3545 * Return the kmem_cache we're supposed to use for a slab allocation.
3546 * We try to use the current memcg's version of the cache.
3548 * If the cache does not exist yet, if we are the first user of it,
3549 * we either create it immediately, if possible, or create it asynchronously
3551 * In the latter case, we will let the current allocation go through with
3552 * the original cache.
3554 * Can't be called in interrupt context or from kernel threads.
3555 * This function needs to be called with rcu_read_lock() held.
3557 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3560 struct mem_cgroup *memcg;
3563 VM_BUG_ON(!cachep->memcg_params);
3564 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3566 if (!current->mm || current->memcg_kmem_skip_account)
3570 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3572 if (!memcg_can_account_kmem(memcg))
3575 idx = memcg_cache_id(memcg);
3578 * barrier to mare sure we're always seeing the up to date value. The
3579 * code updating memcg_caches will issue a write barrier to match this.
3581 read_barrier_depends();
3582 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3583 cachep = cachep->memcg_params->memcg_caches[idx];
3587 /* The corresponding put will be done in the workqueue. */
3588 if (!css_tryget(&memcg->css))
3593 * If we are in a safe context (can wait, and not in interrupt
3594 * context), we could be be predictable and return right away.
3595 * This would guarantee that the allocation being performed
3596 * already belongs in the new cache.
3598 * However, there are some clashes that can arrive from locking.
3599 * For instance, because we acquire the slab_mutex while doing
3600 * kmem_cache_dup, this means no further allocation could happen
3601 * with the slab_mutex held.
3603 * Also, because cache creation issue get_online_cpus(), this
3604 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3605 * that ends up reversed during cpu hotplug. (cpuset allocates
3606 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3607 * better to defer everything.
3609 memcg_create_cache_enqueue(memcg, cachep);
3615 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3618 * We need to verify if the allocation against current->mm->owner's memcg is
3619 * possible for the given order. But the page is not allocated yet, so we'll
3620 * need a further commit step to do the final arrangements.
3622 * It is possible for the task to switch cgroups in this mean time, so at
3623 * commit time, we can't rely on task conversion any longer. We'll then use
3624 * the handle argument to return to the caller which cgroup we should commit
3625 * against. We could also return the memcg directly and avoid the pointer
3626 * passing, but a boolean return value gives better semantics considering
3627 * the compiled-out case as well.
3629 * Returning true means the allocation is possible.
3632 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3634 struct mem_cgroup *memcg;
3640 * Disabling accounting is only relevant for some specific memcg
3641 * internal allocations. Therefore we would initially not have such
3642 * check here, since direct calls to the page allocator that are marked
3643 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3644 * concerned with cache allocations, and by having this test at
3645 * memcg_kmem_get_cache, we are already able to relay the allocation to
3646 * the root cache and bypass the memcg cache altogether.
3648 * There is one exception, though: the SLUB allocator does not create
3649 * large order caches, but rather service large kmallocs directly from
3650 * the page allocator. Therefore, the following sequence when backed by
3651 * the SLUB allocator:
3653 * memcg_stop_kmem_account();
3654 * kmalloc(<large_number>)
3655 * memcg_resume_kmem_account();
3657 * would effectively ignore the fact that we should skip accounting,
3658 * since it will drive us directly to this function without passing
3659 * through the cache selector memcg_kmem_get_cache. Such large
3660 * allocations are extremely rare but can happen, for instance, for the
3661 * cache arrays. We bring this test here.
3663 if (!current->mm || current->memcg_kmem_skip_account)
3666 memcg = try_get_mem_cgroup_from_mm(current->mm);
3669 * very rare case described in mem_cgroup_from_task. Unfortunately there
3670 * isn't much we can do without complicating this too much, and it would
3671 * be gfp-dependent anyway. Just let it go
3673 if (unlikely(!memcg))
3676 if (!memcg_can_account_kmem(memcg)) {
3677 css_put(&memcg->css);
3681 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3685 css_put(&memcg->css);
3689 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3692 struct page_cgroup *pc;
3694 VM_BUG_ON(mem_cgroup_is_root(memcg));
3696 /* The page allocation failed. Revert */
3698 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3702 pc = lookup_page_cgroup(page);
3703 lock_page_cgroup(pc);
3704 pc->mem_cgroup = memcg;
3705 SetPageCgroupUsed(pc);
3706 unlock_page_cgroup(pc);
3709 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3711 struct mem_cgroup *memcg = NULL;
3712 struct page_cgroup *pc;
3715 pc = lookup_page_cgroup(page);
3717 * Fast unlocked return. Theoretically might have changed, have to
3718 * check again after locking.
3720 if (!PageCgroupUsed(pc))
3723 lock_page_cgroup(pc);
3724 if (PageCgroupUsed(pc)) {
3725 memcg = pc->mem_cgroup;
3726 ClearPageCgroupUsed(pc);
3728 unlock_page_cgroup(pc);
3731 * We trust that only if there is a memcg associated with the page, it
3732 * is a valid allocation
3737 VM_BUG_ON(mem_cgroup_is_root(memcg));
3738 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3741 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3744 #endif /* CONFIG_MEMCG_KMEM */
3746 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3748 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3750 * Because tail pages are not marked as "used", set it. We're under
3751 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3752 * charge/uncharge will be never happen and move_account() is done under
3753 * compound_lock(), so we don't have to take care of races.
3755 void mem_cgroup_split_huge_fixup(struct page *head)
3757 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3758 struct page_cgroup *pc;
3759 struct mem_cgroup *memcg;
3762 if (mem_cgroup_disabled())
3765 memcg = head_pc->mem_cgroup;
3766 for (i = 1; i < HPAGE_PMD_NR; i++) {
3768 pc->mem_cgroup = memcg;
3769 smp_wmb();/* see __commit_charge() */
3770 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3772 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3775 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3778 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3779 struct mem_cgroup *to,
3780 unsigned int nr_pages,
3781 enum mem_cgroup_stat_index idx)
3783 /* Update stat data for mem_cgroup */
3785 __this_cpu_sub(from->stat->count[idx], nr_pages);
3786 __this_cpu_add(to->stat->count[idx], nr_pages);
3791 * mem_cgroup_move_account - move account of the page
3793 * @nr_pages: number of regular pages (>1 for huge pages)
3794 * @pc: page_cgroup of the page.
3795 * @from: mem_cgroup which the page is moved from.
3796 * @to: mem_cgroup which the page is moved to. @from != @to.
3798 * The caller must confirm following.
3799 * - page is not on LRU (isolate_page() is useful.)
3800 * - compound_lock is held when nr_pages > 1
3802 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3805 static int mem_cgroup_move_account(struct page *page,
3806 unsigned int nr_pages,
3807 struct page_cgroup *pc,
3808 struct mem_cgroup *from,
3809 struct mem_cgroup *to)
3811 unsigned long flags;
3813 bool anon = PageAnon(page);
3815 VM_BUG_ON(from == to);
3816 VM_BUG_ON(PageLRU(page));
3818 * The page is isolated from LRU. So, collapse function
3819 * will not handle this page. But page splitting can happen.
3820 * Do this check under compound_page_lock(). The caller should
3824 if (nr_pages > 1 && !PageTransHuge(page))
3827 lock_page_cgroup(pc);
3830 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3833 move_lock_mem_cgroup(from, &flags);
3835 if (!anon && page_mapped(page))
3836 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3837 MEM_CGROUP_STAT_FILE_MAPPED);
3839 if (PageWriteback(page))
3840 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3841 MEM_CGROUP_STAT_WRITEBACK);
3843 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3845 /* caller should have done css_get */
3846 pc->mem_cgroup = to;
3847 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3848 move_unlock_mem_cgroup(from, &flags);
3851 unlock_page_cgroup(pc);
3855 memcg_check_events(to, page);
3856 memcg_check_events(from, page);
3862 * mem_cgroup_move_parent - moves page to the parent group
3863 * @page: the page to move
3864 * @pc: page_cgroup of the page
3865 * @child: page's cgroup
3867 * move charges to its parent or the root cgroup if the group has no
3868 * parent (aka use_hierarchy==0).
3869 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3870 * mem_cgroup_move_account fails) the failure is always temporary and
3871 * it signals a race with a page removal/uncharge or migration. In the
3872 * first case the page is on the way out and it will vanish from the LRU
3873 * on the next attempt and the call should be retried later.
3874 * Isolation from the LRU fails only if page has been isolated from
3875 * the LRU since we looked at it and that usually means either global
3876 * reclaim or migration going on. The page will either get back to the
3878 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3879 * (!PageCgroupUsed) or moved to a different group. The page will
3880 * disappear in the next attempt.
3882 static int mem_cgroup_move_parent(struct page *page,
3883 struct page_cgroup *pc,
3884 struct mem_cgroup *child)
3886 struct mem_cgroup *parent;
3887 unsigned int nr_pages;
3888 unsigned long uninitialized_var(flags);
3891 VM_BUG_ON(mem_cgroup_is_root(child));
3894 if (!get_page_unless_zero(page))
3896 if (isolate_lru_page(page))
3899 nr_pages = hpage_nr_pages(page);
3901 parent = parent_mem_cgroup(child);
3903 * If no parent, move charges to root cgroup.
3906 parent = root_mem_cgroup;
3909 VM_BUG_ON(!PageTransHuge(page));
3910 flags = compound_lock_irqsave(page);
3913 ret = mem_cgroup_move_account(page, nr_pages,
3916 __mem_cgroup_cancel_local_charge(child, nr_pages);
3919 compound_unlock_irqrestore(page, flags);
3920 putback_lru_page(page);
3928 * Charge the memory controller for page usage.
3930 * 0 if the charge was successful
3931 * < 0 if the cgroup is over its limit
3933 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3934 gfp_t gfp_mask, enum charge_type ctype)
3936 struct mem_cgroup *memcg = NULL;
3937 unsigned int nr_pages = 1;
3941 if (PageTransHuge(page)) {
3942 nr_pages <<= compound_order(page);
3943 VM_BUG_ON(!PageTransHuge(page));
3945 * Never OOM-kill a process for a huge page. The
3946 * fault handler will fall back to regular pages.
3951 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3954 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3958 int mem_cgroup_newpage_charge(struct page *page,
3959 struct mm_struct *mm, gfp_t gfp_mask)
3961 if (mem_cgroup_disabled())
3963 VM_BUG_ON(page_mapped(page));
3964 VM_BUG_ON(page->mapping && !PageAnon(page));
3966 return mem_cgroup_charge_common(page, mm, gfp_mask,
3967 MEM_CGROUP_CHARGE_TYPE_ANON);
3971 * While swap-in, try_charge -> commit or cancel, the page is locked.
3972 * And when try_charge() successfully returns, one refcnt to memcg without
3973 * struct page_cgroup is acquired. This refcnt will be consumed by
3974 * "commit()" or removed by "cancel()"
3976 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3979 struct mem_cgroup **memcgp)
3981 struct mem_cgroup *memcg;
3982 struct page_cgroup *pc;
3985 pc = lookup_page_cgroup(page);
3987 * Every swap fault against a single page tries to charge the
3988 * page, bail as early as possible. shmem_unuse() encounters
3989 * already charged pages, too. The USED bit is protected by
3990 * the page lock, which serializes swap cache removal, which
3991 * in turn serializes uncharging.
3993 if (PageCgroupUsed(pc))
3995 if (!do_swap_account)
3997 memcg = try_get_mem_cgroup_from_page(page);
4001 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4002 css_put(&memcg->css);
4007 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4013 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4014 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4017 if (mem_cgroup_disabled())
4020 * A racing thread's fault, or swapoff, may have already
4021 * updated the pte, and even removed page from swap cache: in
4022 * those cases unuse_pte()'s pte_same() test will fail; but
4023 * there's also a KSM case which does need to charge the page.
4025 if (!PageSwapCache(page)) {
4028 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4033 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4036 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4038 if (mem_cgroup_disabled())
4042 __mem_cgroup_cancel_charge(memcg, 1);
4046 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4047 enum charge_type ctype)
4049 if (mem_cgroup_disabled())
4054 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4056 * Now swap is on-memory. This means this page may be
4057 * counted both as mem and swap....double count.
4058 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4059 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4060 * may call delete_from_swap_cache() before reach here.
4062 if (do_swap_account && PageSwapCache(page)) {
4063 swp_entry_t ent = {.val = page_private(page)};
4064 mem_cgroup_uncharge_swap(ent);
4068 void mem_cgroup_commit_charge_swapin(struct page *page,
4069 struct mem_cgroup *memcg)
4071 __mem_cgroup_commit_charge_swapin(page, memcg,
4072 MEM_CGROUP_CHARGE_TYPE_ANON);
4075 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4078 struct mem_cgroup *memcg = NULL;
4079 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4082 if (mem_cgroup_disabled())
4084 if (PageCompound(page))
4087 if (!PageSwapCache(page))
4088 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4089 else { /* page is swapcache/shmem */
4090 ret = __mem_cgroup_try_charge_swapin(mm, page,
4093 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4098 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4099 unsigned int nr_pages,
4100 const enum charge_type ctype)
4102 struct memcg_batch_info *batch = NULL;
4103 bool uncharge_memsw = true;
4105 /* If swapout, usage of swap doesn't decrease */
4106 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4107 uncharge_memsw = false;
4109 batch = ¤t->memcg_batch;
4111 * In usual, we do css_get() when we remember memcg pointer.
4112 * But in this case, we keep res->usage until end of a series of
4113 * uncharges. Then, it's ok to ignore memcg's refcnt.
4116 batch->memcg = memcg;
4118 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4119 * In those cases, all pages freed continuously can be expected to be in
4120 * the same cgroup and we have chance to coalesce uncharges.
4121 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4122 * because we want to do uncharge as soon as possible.
4125 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4126 goto direct_uncharge;
4129 goto direct_uncharge;
4132 * In typical case, batch->memcg == mem. This means we can
4133 * merge a series of uncharges to an uncharge of res_counter.
4134 * If not, we uncharge res_counter ony by one.
4136 if (batch->memcg != memcg)
4137 goto direct_uncharge;
4138 /* remember freed charge and uncharge it later */
4141 batch->memsw_nr_pages++;
4144 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4146 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4147 if (unlikely(batch->memcg != memcg))
4148 memcg_oom_recover(memcg);
4152 * uncharge if !page_mapped(page)
4154 static struct mem_cgroup *
4155 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4158 struct mem_cgroup *memcg = NULL;
4159 unsigned int nr_pages = 1;
4160 struct page_cgroup *pc;
4163 if (mem_cgroup_disabled())
4166 if (PageTransHuge(page)) {
4167 nr_pages <<= compound_order(page);
4168 VM_BUG_ON(!PageTransHuge(page));
4171 * Check if our page_cgroup is valid
4173 pc = lookup_page_cgroup(page);
4174 if (unlikely(!PageCgroupUsed(pc)))
4177 lock_page_cgroup(pc);
4179 memcg = pc->mem_cgroup;
4181 if (!PageCgroupUsed(pc))
4184 anon = PageAnon(page);
4187 case MEM_CGROUP_CHARGE_TYPE_ANON:
4189 * Generally PageAnon tells if it's the anon statistics to be
4190 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4191 * used before page reached the stage of being marked PageAnon.
4195 case MEM_CGROUP_CHARGE_TYPE_DROP:
4196 /* See mem_cgroup_prepare_migration() */
4197 if (page_mapped(page))
4200 * Pages under migration may not be uncharged. But
4201 * end_migration() /must/ be the one uncharging the
4202 * unused post-migration page and so it has to call
4203 * here with the migration bit still set. See the
4204 * res_counter handling below.
4206 if (!end_migration && PageCgroupMigration(pc))
4209 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4210 if (!PageAnon(page)) { /* Shared memory */
4211 if (page->mapping && !page_is_file_cache(page))
4213 } else if (page_mapped(page)) /* Anon */
4220 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4222 ClearPageCgroupUsed(pc);
4224 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4225 * freed from LRU. This is safe because uncharged page is expected not
4226 * to be reused (freed soon). Exception is SwapCache, it's handled by
4227 * special functions.
4230 unlock_page_cgroup(pc);
4232 * even after unlock, we have memcg->res.usage here and this memcg
4233 * will never be freed, so it's safe to call css_get().
4235 memcg_check_events(memcg, page);
4236 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4237 mem_cgroup_swap_statistics(memcg, true);
4238 css_get(&memcg->css);
4241 * Migration does not charge the res_counter for the
4242 * replacement page, so leave it alone when phasing out the
4243 * page that is unused after the migration.
4245 if (!end_migration && !mem_cgroup_is_root(memcg))
4246 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4251 unlock_page_cgroup(pc);
4255 void mem_cgroup_uncharge_page(struct page *page)
4258 if (page_mapped(page))
4260 VM_BUG_ON(page->mapping && !PageAnon(page));
4262 * If the page is in swap cache, uncharge should be deferred
4263 * to the swap path, which also properly accounts swap usage
4264 * and handles memcg lifetime.
4266 * Note that this check is not stable and reclaim may add the
4267 * page to swap cache at any time after this. However, if the
4268 * page is not in swap cache by the time page->mapcount hits
4269 * 0, there won't be any page table references to the swap
4270 * slot, and reclaim will free it and not actually write the
4273 if (PageSwapCache(page))
4275 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4278 void mem_cgroup_uncharge_cache_page(struct page *page)
4280 VM_BUG_ON(page_mapped(page));
4281 VM_BUG_ON(page->mapping);
4282 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4286 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4287 * In that cases, pages are freed continuously and we can expect pages
4288 * are in the same memcg. All these calls itself limits the number of
4289 * pages freed at once, then uncharge_start/end() is called properly.
4290 * This may be called prural(2) times in a context,
4293 void mem_cgroup_uncharge_start(void)
4295 current->memcg_batch.do_batch++;
4296 /* We can do nest. */
4297 if (current->memcg_batch.do_batch == 1) {
4298 current->memcg_batch.memcg = NULL;
4299 current->memcg_batch.nr_pages = 0;
4300 current->memcg_batch.memsw_nr_pages = 0;
4304 void mem_cgroup_uncharge_end(void)
4306 struct memcg_batch_info *batch = ¤t->memcg_batch;
4308 if (!batch->do_batch)
4312 if (batch->do_batch) /* If stacked, do nothing. */
4318 * This "batch->memcg" is valid without any css_get/put etc...
4319 * bacause we hide charges behind us.
4321 if (batch->nr_pages)
4322 res_counter_uncharge(&batch->memcg->res,
4323 batch->nr_pages * PAGE_SIZE);
4324 if (batch->memsw_nr_pages)
4325 res_counter_uncharge(&batch->memcg->memsw,
4326 batch->memsw_nr_pages * PAGE_SIZE);
4327 memcg_oom_recover(batch->memcg);
4328 /* forget this pointer (for sanity check) */
4329 batch->memcg = NULL;
4334 * called after __delete_from_swap_cache() and drop "page" account.
4335 * memcg information is recorded to swap_cgroup of "ent"
4338 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4340 struct mem_cgroup *memcg;
4341 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4343 if (!swapout) /* this was a swap cache but the swap is unused ! */
4344 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4346 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4349 * record memcg information, if swapout && memcg != NULL,
4350 * css_get() was called in uncharge().
4352 if (do_swap_account && swapout && memcg)
4353 swap_cgroup_record(ent, css_id(&memcg->css));
4357 #ifdef CONFIG_MEMCG_SWAP
4359 * called from swap_entry_free(). remove record in swap_cgroup and
4360 * uncharge "memsw" account.
4362 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4364 struct mem_cgroup *memcg;
4367 if (!do_swap_account)
4370 id = swap_cgroup_record(ent, 0);
4372 memcg = mem_cgroup_lookup(id);
4375 * We uncharge this because swap is freed.
4376 * This memcg can be obsolete one. We avoid calling css_tryget
4378 if (!mem_cgroup_is_root(memcg))
4379 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4380 mem_cgroup_swap_statistics(memcg, false);
4381 css_put(&memcg->css);
4387 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4388 * @entry: swap entry to be moved
4389 * @from: mem_cgroup which the entry is moved from
4390 * @to: mem_cgroup which the entry is moved to
4392 * It succeeds only when the swap_cgroup's record for this entry is the same
4393 * as the mem_cgroup's id of @from.
4395 * Returns 0 on success, -EINVAL on failure.
4397 * The caller must have charged to @to, IOW, called res_counter_charge() about
4398 * both res and memsw, and called css_get().
4400 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4401 struct mem_cgroup *from, struct mem_cgroup *to)
4403 unsigned short old_id, new_id;
4405 old_id = css_id(&from->css);
4406 new_id = css_id(&to->css);
4408 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4409 mem_cgroup_swap_statistics(from, false);
4410 mem_cgroup_swap_statistics(to, true);
4412 * This function is only called from task migration context now.
4413 * It postpones res_counter and refcount handling till the end
4414 * of task migration(mem_cgroup_clear_mc()) for performance
4415 * improvement. But we cannot postpone css_get(to) because if
4416 * the process that has been moved to @to does swap-in, the
4417 * refcount of @to might be decreased to 0.
4419 * We are in attach() phase, so the cgroup is guaranteed to be
4420 * alive, so we can just call css_get().
4428 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4429 struct mem_cgroup *from, struct mem_cgroup *to)
4436 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4439 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4440 struct mem_cgroup **memcgp)
4442 struct mem_cgroup *memcg = NULL;
4443 unsigned int nr_pages = 1;
4444 struct page_cgroup *pc;
4445 enum charge_type ctype;
4449 if (mem_cgroup_disabled())
4452 if (PageTransHuge(page))
4453 nr_pages <<= compound_order(page);
4455 pc = lookup_page_cgroup(page);
4456 lock_page_cgroup(pc);
4457 if (PageCgroupUsed(pc)) {
4458 memcg = pc->mem_cgroup;
4459 css_get(&memcg->css);
4461 * At migrating an anonymous page, its mapcount goes down
4462 * to 0 and uncharge() will be called. But, even if it's fully
4463 * unmapped, migration may fail and this page has to be
4464 * charged again. We set MIGRATION flag here and delay uncharge
4465 * until end_migration() is called
4467 * Corner Case Thinking
4469 * When the old page was mapped as Anon and it's unmap-and-freed
4470 * while migration was ongoing.
4471 * If unmap finds the old page, uncharge() of it will be delayed
4472 * until end_migration(). If unmap finds a new page, it's
4473 * uncharged when it make mapcount to be 1->0. If unmap code
4474 * finds swap_migration_entry, the new page will not be mapped
4475 * and end_migration() will find it(mapcount==0).
4478 * When the old page was mapped but migraion fails, the kernel
4479 * remaps it. A charge for it is kept by MIGRATION flag even
4480 * if mapcount goes down to 0. We can do remap successfully
4481 * without charging it again.
4484 * The "old" page is under lock_page() until the end of
4485 * migration, so, the old page itself will not be swapped-out.
4486 * If the new page is swapped out before end_migraton, our
4487 * hook to usual swap-out path will catch the event.
4490 SetPageCgroupMigration(pc);
4492 unlock_page_cgroup(pc);
4494 * If the page is not charged at this point,
4502 * We charge new page before it's used/mapped. So, even if unlock_page()
4503 * is called before end_migration, we can catch all events on this new
4504 * page. In the case new page is migrated but not remapped, new page's
4505 * mapcount will be finally 0 and we call uncharge in end_migration().
4508 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4510 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4512 * The page is committed to the memcg, but it's not actually
4513 * charged to the res_counter since we plan on replacing the
4514 * old one and only one page is going to be left afterwards.
4516 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4519 /* remove redundant charge if migration failed*/
4520 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4521 struct page *oldpage, struct page *newpage, bool migration_ok)
4523 struct page *used, *unused;
4524 struct page_cgroup *pc;
4530 if (!migration_ok) {
4537 anon = PageAnon(used);
4538 __mem_cgroup_uncharge_common(unused,
4539 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4540 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4542 css_put(&memcg->css);
4544 * We disallowed uncharge of pages under migration because mapcount
4545 * of the page goes down to zero, temporarly.
4546 * Clear the flag and check the page should be charged.
4548 pc = lookup_page_cgroup(oldpage);
4549 lock_page_cgroup(pc);
4550 ClearPageCgroupMigration(pc);
4551 unlock_page_cgroup(pc);
4554 * If a page is a file cache, radix-tree replacement is very atomic
4555 * and we can skip this check. When it was an Anon page, its mapcount
4556 * goes down to 0. But because we added MIGRATION flage, it's not
4557 * uncharged yet. There are several case but page->mapcount check
4558 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4559 * check. (see prepare_charge() also)
4562 mem_cgroup_uncharge_page(used);
4566 * At replace page cache, newpage is not under any memcg but it's on
4567 * LRU. So, this function doesn't touch res_counter but handles LRU
4568 * in correct way. Both pages are locked so we cannot race with uncharge.
4570 void mem_cgroup_replace_page_cache(struct page *oldpage,
4571 struct page *newpage)
4573 struct mem_cgroup *memcg = NULL;
4574 struct page_cgroup *pc;
4575 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4577 if (mem_cgroup_disabled())
4580 pc = lookup_page_cgroup(oldpage);
4581 /* fix accounting on old pages */
4582 lock_page_cgroup(pc);
4583 if (PageCgroupUsed(pc)) {
4584 memcg = pc->mem_cgroup;
4585 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4586 ClearPageCgroupUsed(pc);
4588 unlock_page_cgroup(pc);
4591 * When called from shmem_replace_page(), in some cases the
4592 * oldpage has already been charged, and in some cases not.
4597 * Even if newpage->mapping was NULL before starting replacement,
4598 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4599 * LRU while we overwrite pc->mem_cgroup.
4601 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4604 #ifdef CONFIG_DEBUG_VM
4605 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4607 struct page_cgroup *pc;
4609 pc = lookup_page_cgroup(page);
4611 * Can be NULL while feeding pages into the page allocator for
4612 * the first time, i.e. during boot or memory hotplug;
4613 * or when mem_cgroup_disabled().
4615 if (likely(pc) && PageCgroupUsed(pc))
4620 bool mem_cgroup_bad_page_check(struct page *page)
4622 if (mem_cgroup_disabled())
4625 return lookup_page_cgroup_used(page) != NULL;
4628 void mem_cgroup_print_bad_page(struct page *page)
4630 struct page_cgroup *pc;
4632 pc = lookup_page_cgroup_used(page);
4634 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4635 pc, pc->flags, pc->mem_cgroup);
4640 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4641 unsigned long long val)
4644 u64 memswlimit, memlimit;
4646 int children = mem_cgroup_count_children(memcg);
4647 u64 curusage, oldusage;
4651 * For keeping hierarchical_reclaim simple, how long we should retry
4652 * is depends on callers. We set our retry-count to be function
4653 * of # of children which we should visit in this loop.
4655 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4657 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4660 while (retry_count) {
4661 if (signal_pending(current)) {
4666 * Rather than hide all in some function, I do this in
4667 * open coded manner. You see what this really does.
4668 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4670 mutex_lock(&set_limit_mutex);
4671 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4672 if (memswlimit < val) {
4674 mutex_unlock(&set_limit_mutex);
4678 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4682 ret = res_counter_set_limit(&memcg->res, val);
4684 if (memswlimit == val)
4685 memcg->memsw_is_minimum = true;
4687 memcg->memsw_is_minimum = false;
4689 mutex_unlock(&set_limit_mutex);
4694 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4695 MEM_CGROUP_RECLAIM_SHRINK);
4696 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4697 /* Usage is reduced ? */
4698 if (curusage >= oldusage)
4701 oldusage = curusage;
4703 if (!ret && enlarge)
4704 memcg_oom_recover(memcg);
4709 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4710 unsigned long long val)
4713 u64 memlimit, memswlimit, oldusage, curusage;
4714 int children = mem_cgroup_count_children(memcg);
4718 /* see mem_cgroup_resize_res_limit */
4719 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4720 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4721 while (retry_count) {
4722 if (signal_pending(current)) {
4727 * Rather than hide all in some function, I do this in
4728 * open coded manner. You see what this really does.
4729 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4731 mutex_lock(&set_limit_mutex);
4732 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4733 if (memlimit > val) {
4735 mutex_unlock(&set_limit_mutex);
4738 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4739 if (memswlimit < val)
4741 ret = res_counter_set_limit(&memcg->memsw, val);
4743 if (memlimit == val)
4744 memcg->memsw_is_minimum = true;
4746 memcg->memsw_is_minimum = false;
4748 mutex_unlock(&set_limit_mutex);
4753 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4754 MEM_CGROUP_RECLAIM_NOSWAP |
4755 MEM_CGROUP_RECLAIM_SHRINK);
4756 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4757 /* Usage is reduced ? */
4758 if (curusage >= oldusage)
4761 oldusage = curusage;
4763 if (!ret && enlarge)
4764 memcg_oom_recover(memcg);
4768 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4770 unsigned long *total_scanned)
4772 unsigned long nr_reclaimed = 0;
4773 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4774 unsigned long reclaimed;
4776 struct mem_cgroup_tree_per_zone *mctz;
4777 unsigned long long excess;
4778 unsigned long nr_scanned;
4783 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4785 * This loop can run a while, specially if mem_cgroup's continuously
4786 * keep exceeding their soft limit and putting the system under
4793 mz = mem_cgroup_largest_soft_limit_node(mctz);
4798 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4799 gfp_mask, &nr_scanned);
4800 nr_reclaimed += reclaimed;
4801 *total_scanned += nr_scanned;
4802 spin_lock(&mctz->lock);
4805 * If we failed to reclaim anything from this memory cgroup
4806 * it is time to move on to the next cgroup
4812 * Loop until we find yet another one.
4814 * By the time we get the soft_limit lock
4815 * again, someone might have aded the
4816 * group back on the RB tree. Iterate to
4817 * make sure we get a different mem.
4818 * mem_cgroup_largest_soft_limit_node returns
4819 * NULL if no other cgroup is present on
4823 __mem_cgroup_largest_soft_limit_node(mctz);
4825 css_put(&next_mz->memcg->css);
4826 else /* next_mz == NULL or other memcg */
4830 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4831 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4833 * One school of thought says that we should not add
4834 * back the node to the tree if reclaim returns 0.
4835 * But our reclaim could return 0, simply because due
4836 * to priority we are exposing a smaller subset of
4837 * memory to reclaim from. Consider this as a longer
4840 /* If excess == 0, no tree ops */
4841 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4842 spin_unlock(&mctz->lock);
4843 css_put(&mz->memcg->css);
4846 * Could not reclaim anything and there are no more
4847 * mem cgroups to try or we seem to be looping without
4848 * reclaiming anything.
4850 if (!nr_reclaimed &&
4852 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4854 } while (!nr_reclaimed);
4856 css_put(&next_mz->memcg->css);
4857 return nr_reclaimed;
4861 * mem_cgroup_force_empty_list - clears LRU of a group
4862 * @memcg: group to clear
4865 * @lru: lru to to clear
4867 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4868 * reclaim the pages page themselves - pages are moved to the parent (or root)
4871 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4872 int node, int zid, enum lru_list lru)
4874 struct lruvec *lruvec;
4875 unsigned long flags;
4876 struct list_head *list;
4880 zone = &NODE_DATA(node)->node_zones[zid];
4881 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4882 list = &lruvec->lists[lru];
4886 struct page_cgroup *pc;
4889 spin_lock_irqsave(&zone->lru_lock, flags);
4890 if (list_empty(list)) {
4891 spin_unlock_irqrestore(&zone->lru_lock, flags);
4894 page = list_entry(list->prev, struct page, lru);
4896 list_move(&page->lru, list);
4898 spin_unlock_irqrestore(&zone->lru_lock, flags);
4901 spin_unlock_irqrestore(&zone->lru_lock, flags);
4903 pc = lookup_page_cgroup(page);
4905 if (mem_cgroup_move_parent(page, pc, memcg)) {
4906 /* found lock contention or "pc" is obsolete. */
4911 } while (!list_empty(list));
4915 * make mem_cgroup's charge to be 0 if there is no task by moving
4916 * all the charges and pages to the parent.
4917 * This enables deleting this mem_cgroup.
4919 * Caller is responsible for holding css reference on the memcg.
4921 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4927 /* This is for making all *used* pages to be on LRU. */
4928 lru_add_drain_all();
4929 drain_all_stock_sync(memcg);
4930 mem_cgroup_start_move(memcg);
4931 for_each_node_state(node, N_MEMORY) {
4932 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4935 mem_cgroup_force_empty_list(memcg,
4940 mem_cgroup_end_move(memcg);
4941 memcg_oom_recover(memcg);
4945 * Kernel memory may not necessarily be trackable to a specific
4946 * process. So they are not migrated, and therefore we can't
4947 * expect their value to drop to 0 here.
4948 * Having res filled up with kmem only is enough.
4950 * This is a safety check because mem_cgroup_force_empty_list
4951 * could have raced with mem_cgroup_replace_page_cache callers
4952 * so the lru seemed empty but the page could have been added
4953 * right after the check. RES_USAGE should be safe as we always
4954 * charge before adding to the LRU.
4956 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4957 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4958 } while (usage > 0);
4961 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4963 lockdep_assert_held(&memcg_create_mutex);
4965 * The lock does not prevent addition or deletion to the list
4966 * of children, but it prevents a new child from being
4967 * initialized based on this parent in css_online(), so it's
4968 * enough to decide whether hierarchically inherited
4969 * attributes can still be changed or not.
4971 return memcg->use_hierarchy &&
4972 !list_empty(&memcg->css.cgroup->children);
4976 * Reclaims as many pages from the given memcg as possible and moves
4977 * the rest to the parent.
4979 * Caller is responsible for holding css reference for memcg.
4981 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4983 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4984 struct cgroup *cgrp = memcg->css.cgroup;
4986 /* returns EBUSY if there is a task or if we come here twice. */
4987 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4990 /* we call try-to-free pages for make this cgroup empty */
4991 lru_add_drain_all();
4992 /* try to free all pages in this cgroup */
4993 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4996 if (signal_pending(current))
4999 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5003 /* maybe some writeback is necessary */
5004 congestion_wait(BLK_RW_ASYNC, HZ/10);
5009 mem_cgroup_reparent_charges(memcg);
5014 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
5017 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5019 if (mem_cgroup_is_root(memcg))
5021 return mem_cgroup_force_empty(memcg);
5024 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5027 return mem_cgroup_from_css(css)->use_hierarchy;
5030 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5031 struct cftype *cft, u64 val)
5034 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5035 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5037 mutex_lock(&memcg_create_mutex);
5039 if (memcg->use_hierarchy == val)
5043 * If parent's use_hierarchy is set, we can't make any modifications
5044 * in the child subtrees. If it is unset, then the change can
5045 * occur, provided the current cgroup has no children.
5047 * For the root cgroup, parent_mem is NULL, we allow value to be
5048 * set if there are no children.
5050 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5051 (val == 1 || val == 0)) {
5052 if (list_empty(&memcg->css.cgroup->children))
5053 memcg->use_hierarchy = val;
5060 mutex_unlock(&memcg_create_mutex);
5066 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5067 enum mem_cgroup_stat_index idx)
5069 struct mem_cgroup *iter;
5072 /* Per-cpu values can be negative, use a signed accumulator */
5073 for_each_mem_cgroup_tree(iter, memcg)
5074 val += mem_cgroup_read_stat(iter, idx);
5076 if (val < 0) /* race ? */
5081 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5085 if (!mem_cgroup_is_root(memcg)) {
5087 return res_counter_read_u64(&memcg->res, RES_USAGE);
5089 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5093 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5094 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5096 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5097 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5100 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5102 return val << PAGE_SHIFT;
5105 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
5106 struct cftype *cft, struct file *file,
5107 char __user *buf, size_t nbytes, loff_t *ppos)
5109 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5115 type = MEMFILE_TYPE(cft->private);
5116 name = MEMFILE_ATTR(cft->private);
5120 if (name == RES_USAGE)
5121 val = mem_cgroup_usage(memcg, false);
5123 val = res_counter_read_u64(&memcg->res, name);
5126 if (name == RES_USAGE)
5127 val = mem_cgroup_usage(memcg, true);
5129 val = res_counter_read_u64(&memcg->memsw, name);
5132 val = res_counter_read_u64(&memcg->kmem, name);
5138 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5139 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5142 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
5145 #ifdef CONFIG_MEMCG_KMEM
5146 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5148 * For simplicity, we won't allow this to be disabled. It also can't
5149 * be changed if the cgroup has children already, or if tasks had
5152 * If tasks join before we set the limit, a person looking at
5153 * kmem.usage_in_bytes will have no way to determine when it took
5154 * place, which makes the value quite meaningless.
5156 * After it first became limited, changes in the value of the limit are
5157 * of course permitted.
5159 mutex_lock(&memcg_create_mutex);
5160 mutex_lock(&set_limit_mutex);
5161 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
5162 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
5166 ret = res_counter_set_limit(&memcg->kmem, val);
5169 ret = memcg_update_cache_sizes(memcg);
5171 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
5174 static_key_slow_inc(&memcg_kmem_enabled_key);
5176 * setting the active bit after the inc will guarantee no one
5177 * starts accounting before all call sites are patched
5179 memcg_kmem_set_active(memcg);
5181 ret = res_counter_set_limit(&memcg->kmem, val);
5183 mutex_unlock(&set_limit_mutex);
5184 mutex_unlock(&memcg_create_mutex);
5189 #ifdef CONFIG_MEMCG_KMEM
5190 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5193 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5197 memcg->kmem_account_flags = parent->kmem_account_flags;
5199 * When that happen, we need to disable the static branch only on those
5200 * memcgs that enabled it. To achieve this, we would be forced to
5201 * complicate the code by keeping track of which memcgs were the ones
5202 * that actually enabled limits, and which ones got it from its
5205 * It is a lot simpler just to do static_key_slow_inc() on every child
5206 * that is accounted.
5208 if (!memcg_kmem_is_active(memcg))
5212 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5213 * memcg is active already. If the later initialization fails then the
5214 * cgroup core triggers the cleanup so we do not have to do it here.
5216 static_key_slow_inc(&memcg_kmem_enabled_key);
5218 mutex_lock(&set_limit_mutex);
5219 memcg_stop_kmem_account();
5220 ret = memcg_update_cache_sizes(memcg);
5221 memcg_resume_kmem_account();
5222 mutex_unlock(&set_limit_mutex);
5226 #endif /* CONFIG_MEMCG_KMEM */
5229 * The user of this function is...
5232 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5235 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5238 unsigned long long val;
5241 type = MEMFILE_TYPE(cft->private);
5242 name = MEMFILE_ATTR(cft->private);
5246 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5250 /* This function does all necessary parse...reuse it */
5251 ret = res_counter_memparse_write_strategy(buffer, &val);
5255 ret = mem_cgroup_resize_limit(memcg, val);
5256 else if (type == _MEMSWAP)
5257 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5258 else if (type == _KMEM)
5259 ret = memcg_update_kmem_limit(css, val);
5263 case RES_SOFT_LIMIT:
5264 ret = res_counter_memparse_write_strategy(buffer, &val);
5268 * For memsw, soft limits are hard to implement in terms
5269 * of semantics, for now, we support soft limits for
5270 * control without swap
5273 ret = res_counter_set_soft_limit(&memcg->res, val);
5278 ret = -EINVAL; /* should be BUG() ? */
5284 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5285 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5287 unsigned long long min_limit, min_memsw_limit, tmp;
5289 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5290 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5291 if (!memcg->use_hierarchy)
5294 while (css_parent(&memcg->css)) {
5295 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5296 if (!memcg->use_hierarchy)
5298 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5299 min_limit = min(min_limit, tmp);
5300 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5301 min_memsw_limit = min(min_memsw_limit, tmp);
5304 *mem_limit = min_limit;
5305 *memsw_limit = min_memsw_limit;
5308 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5310 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5314 type = MEMFILE_TYPE(event);
5315 name = MEMFILE_ATTR(event);
5320 res_counter_reset_max(&memcg->res);
5321 else if (type == _MEMSWAP)
5322 res_counter_reset_max(&memcg->memsw);
5323 else if (type == _KMEM)
5324 res_counter_reset_max(&memcg->kmem);
5330 res_counter_reset_failcnt(&memcg->res);
5331 else if (type == _MEMSWAP)
5332 res_counter_reset_failcnt(&memcg->memsw);
5333 else if (type == _KMEM)
5334 res_counter_reset_failcnt(&memcg->kmem);
5343 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5346 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5350 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5351 struct cftype *cft, u64 val)
5353 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5355 if (val >= (1 << NR_MOVE_TYPE))
5359 * No kind of locking is needed in here, because ->can_attach() will
5360 * check this value once in the beginning of the process, and then carry
5361 * on with stale data. This means that changes to this value will only
5362 * affect task migrations starting after the change.
5364 memcg->move_charge_at_immigrate = val;
5368 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5369 struct cftype *cft, u64 val)
5376 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5377 struct cftype *cft, struct seq_file *m)
5380 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5381 unsigned long node_nr;
5382 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5384 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5385 seq_printf(m, "total=%lu", total_nr);
5386 for_each_node_state(nid, N_MEMORY) {
5387 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5388 seq_printf(m, " N%d=%lu", nid, node_nr);
5392 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5393 seq_printf(m, "file=%lu", file_nr);
5394 for_each_node_state(nid, N_MEMORY) {
5395 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5397 seq_printf(m, " N%d=%lu", nid, node_nr);
5401 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5402 seq_printf(m, "anon=%lu", anon_nr);
5403 for_each_node_state(nid, N_MEMORY) {
5404 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5406 seq_printf(m, " N%d=%lu", nid, node_nr);
5410 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5411 seq_printf(m, "unevictable=%lu", unevictable_nr);
5412 for_each_node_state(nid, N_MEMORY) {
5413 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5414 BIT(LRU_UNEVICTABLE));
5415 seq_printf(m, " N%d=%lu", nid, node_nr);
5420 #endif /* CONFIG_NUMA */
5422 static inline void mem_cgroup_lru_names_not_uptodate(void)
5424 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5427 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5430 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5431 struct mem_cgroup *mi;
5434 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5435 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5437 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5438 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5441 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5442 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5443 mem_cgroup_read_events(memcg, i));
5445 for (i = 0; i < NR_LRU_LISTS; i++)
5446 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5447 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5449 /* Hierarchical information */
5451 unsigned long long limit, memsw_limit;
5452 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5453 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5454 if (do_swap_account)
5455 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5459 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5462 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5464 for_each_mem_cgroup_tree(mi, memcg)
5465 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5466 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5469 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5470 unsigned long long val = 0;
5472 for_each_mem_cgroup_tree(mi, memcg)
5473 val += mem_cgroup_read_events(mi, i);
5474 seq_printf(m, "total_%s %llu\n",
5475 mem_cgroup_events_names[i], val);
5478 for (i = 0; i < NR_LRU_LISTS; i++) {
5479 unsigned long long val = 0;
5481 for_each_mem_cgroup_tree(mi, memcg)
5482 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5483 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5486 #ifdef CONFIG_DEBUG_VM
5489 struct mem_cgroup_per_zone *mz;
5490 struct zone_reclaim_stat *rstat;
5491 unsigned long recent_rotated[2] = {0, 0};
5492 unsigned long recent_scanned[2] = {0, 0};
5494 for_each_online_node(nid)
5495 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5496 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5497 rstat = &mz->lruvec.reclaim_stat;
5499 recent_rotated[0] += rstat->recent_rotated[0];
5500 recent_rotated[1] += rstat->recent_rotated[1];
5501 recent_scanned[0] += rstat->recent_scanned[0];
5502 recent_scanned[1] += rstat->recent_scanned[1];
5504 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5505 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5506 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5507 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5514 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5517 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5519 return mem_cgroup_swappiness(memcg);
5522 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5523 struct cftype *cft, u64 val)
5525 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5526 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5528 if (val > 100 || !parent)
5531 mutex_lock(&memcg_create_mutex);
5533 /* If under hierarchy, only empty-root can set this value */
5534 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5535 mutex_unlock(&memcg_create_mutex);
5539 memcg->swappiness = val;
5541 mutex_unlock(&memcg_create_mutex);
5546 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5548 struct mem_cgroup_threshold_ary *t;
5554 t = rcu_dereference(memcg->thresholds.primary);
5556 t = rcu_dereference(memcg->memsw_thresholds.primary);
5561 usage = mem_cgroup_usage(memcg, swap);
5564 * current_threshold points to threshold just below or equal to usage.
5565 * If it's not true, a threshold was crossed after last
5566 * call of __mem_cgroup_threshold().
5568 i = t->current_threshold;
5571 * Iterate backward over array of thresholds starting from
5572 * current_threshold and check if a threshold is crossed.
5573 * If none of thresholds below usage is crossed, we read
5574 * only one element of the array here.
5576 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5577 eventfd_signal(t->entries[i].eventfd, 1);
5579 /* i = current_threshold + 1 */
5583 * Iterate forward over array of thresholds starting from
5584 * current_threshold+1 and check if a threshold is crossed.
5585 * If none of thresholds above usage is crossed, we read
5586 * only one element of the array here.
5588 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5589 eventfd_signal(t->entries[i].eventfd, 1);
5591 /* Update current_threshold */
5592 t->current_threshold = i - 1;
5597 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5600 __mem_cgroup_threshold(memcg, false);
5601 if (do_swap_account)
5602 __mem_cgroup_threshold(memcg, true);
5604 memcg = parent_mem_cgroup(memcg);
5608 static int compare_thresholds(const void *a, const void *b)
5610 const struct mem_cgroup_threshold *_a = a;
5611 const struct mem_cgroup_threshold *_b = b;
5613 if (_a->threshold > _b->threshold)
5616 if (_a->threshold < _b->threshold)
5622 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5624 struct mem_cgroup_eventfd_list *ev;
5626 list_for_each_entry(ev, &memcg->oom_notify, list)
5627 eventfd_signal(ev->eventfd, 1);
5631 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5633 struct mem_cgroup *iter;
5635 for_each_mem_cgroup_tree(iter, memcg)
5636 mem_cgroup_oom_notify_cb(iter);
5639 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5640 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5642 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5643 struct mem_cgroup_thresholds *thresholds;
5644 struct mem_cgroup_threshold_ary *new;
5645 enum res_type type = MEMFILE_TYPE(cft->private);
5646 u64 threshold, usage;
5649 ret = res_counter_memparse_write_strategy(args, &threshold);
5653 mutex_lock(&memcg->thresholds_lock);
5656 thresholds = &memcg->thresholds;
5657 else if (type == _MEMSWAP)
5658 thresholds = &memcg->memsw_thresholds;
5662 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5664 /* Check if a threshold crossed before adding a new one */
5665 if (thresholds->primary)
5666 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5668 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5670 /* Allocate memory for new array of thresholds */
5671 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5679 /* Copy thresholds (if any) to new array */
5680 if (thresholds->primary) {
5681 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5682 sizeof(struct mem_cgroup_threshold));
5685 /* Add new threshold */
5686 new->entries[size - 1].eventfd = eventfd;
5687 new->entries[size - 1].threshold = threshold;
5689 /* Sort thresholds. Registering of new threshold isn't time-critical */
5690 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5691 compare_thresholds, NULL);
5693 /* Find current threshold */
5694 new->current_threshold = -1;
5695 for (i = 0; i < size; i++) {
5696 if (new->entries[i].threshold <= usage) {
5698 * new->current_threshold will not be used until
5699 * rcu_assign_pointer(), so it's safe to increment
5702 ++new->current_threshold;
5707 /* Free old spare buffer and save old primary buffer as spare */
5708 kfree(thresholds->spare);
5709 thresholds->spare = thresholds->primary;
5711 rcu_assign_pointer(thresholds->primary, new);
5713 /* To be sure that nobody uses thresholds */
5717 mutex_unlock(&memcg->thresholds_lock);
5722 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5723 struct cftype *cft, struct eventfd_ctx *eventfd)
5725 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5726 struct mem_cgroup_thresholds *thresholds;
5727 struct mem_cgroup_threshold_ary *new;
5728 enum res_type type = MEMFILE_TYPE(cft->private);
5732 mutex_lock(&memcg->thresholds_lock);
5734 thresholds = &memcg->thresholds;
5735 else if (type == _MEMSWAP)
5736 thresholds = &memcg->memsw_thresholds;
5740 if (!thresholds->primary)
5743 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5745 /* Check if a threshold crossed before removing */
5746 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5748 /* Calculate new number of threshold */
5750 for (i = 0; i < thresholds->primary->size; i++) {
5751 if (thresholds->primary->entries[i].eventfd != eventfd)
5755 new = thresholds->spare;
5757 /* Set thresholds array to NULL if we don't have thresholds */
5766 /* Copy thresholds and find current threshold */
5767 new->current_threshold = -1;
5768 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5769 if (thresholds->primary->entries[i].eventfd == eventfd)
5772 new->entries[j] = thresholds->primary->entries[i];
5773 if (new->entries[j].threshold <= usage) {
5775 * new->current_threshold will not be used
5776 * until rcu_assign_pointer(), so it's safe to increment
5779 ++new->current_threshold;
5785 /* Swap primary and spare array */
5786 thresholds->spare = thresholds->primary;
5787 /* If all events are unregistered, free the spare array */
5789 kfree(thresholds->spare);
5790 thresholds->spare = NULL;
5793 rcu_assign_pointer(thresholds->primary, new);
5795 /* To be sure that nobody uses thresholds */
5798 mutex_unlock(&memcg->thresholds_lock);
5801 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5802 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5804 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5805 struct mem_cgroup_eventfd_list *event;
5806 enum res_type type = MEMFILE_TYPE(cft->private);
5808 BUG_ON(type != _OOM_TYPE);
5809 event = kmalloc(sizeof(*event), GFP_KERNEL);
5813 spin_lock(&memcg_oom_lock);
5815 event->eventfd = eventfd;
5816 list_add(&event->list, &memcg->oom_notify);
5818 /* already in OOM ? */
5819 if (atomic_read(&memcg->under_oom))
5820 eventfd_signal(eventfd, 1);
5821 spin_unlock(&memcg_oom_lock);
5826 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5827 struct cftype *cft, struct eventfd_ctx *eventfd)
5829 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5830 struct mem_cgroup_eventfd_list *ev, *tmp;
5831 enum res_type type = MEMFILE_TYPE(cft->private);
5833 BUG_ON(type != _OOM_TYPE);
5835 spin_lock(&memcg_oom_lock);
5837 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5838 if (ev->eventfd == eventfd) {
5839 list_del(&ev->list);
5844 spin_unlock(&memcg_oom_lock);
5847 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5848 struct cftype *cft, struct cgroup_map_cb *cb)
5850 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5852 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5854 if (atomic_read(&memcg->under_oom))
5855 cb->fill(cb, "under_oom", 1);
5857 cb->fill(cb, "under_oom", 0);
5861 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5862 struct cftype *cft, u64 val)
5864 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5865 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5867 /* cannot set to root cgroup and only 0 and 1 are allowed */
5868 if (!parent || !((val == 0) || (val == 1)))
5871 mutex_lock(&memcg_create_mutex);
5872 /* oom-kill-disable is a flag for subhierarchy. */
5873 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5874 mutex_unlock(&memcg_create_mutex);
5877 memcg->oom_kill_disable = val;
5879 memcg_oom_recover(memcg);
5880 mutex_unlock(&memcg_create_mutex);
5884 #ifdef CONFIG_MEMCG_KMEM
5885 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5889 memcg->kmemcg_id = -1;
5890 ret = memcg_propagate_kmem(memcg);
5894 return mem_cgroup_sockets_init(memcg, ss);
5897 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5899 mem_cgroup_sockets_destroy(memcg);
5902 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5904 if (!memcg_kmem_is_active(memcg))
5908 * kmem charges can outlive the cgroup. In the case of slab
5909 * pages, for instance, a page contain objects from various
5910 * processes. As we prevent from taking a reference for every
5911 * such allocation we have to be careful when doing uncharge
5912 * (see memcg_uncharge_kmem) and here during offlining.
5914 * The idea is that that only the _last_ uncharge which sees
5915 * the dead memcg will drop the last reference. An additional
5916 * reference is taken here before the group is marked dead
5917 * which is then paired with css_put during uncharge resp. here.
5919 * Although this might sound strange as this path is called from
5920 * css_offline() when the referencemight have dropped down to 0
5921 * and shouldn't be incremented anymore (css_tryget would fail)
5922 * we do not have other options because of the kmem allocations
5925 css_get(&memcg->css);
5927 memcg_kmem_mark_dead(memcg);
5929 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5932 if (memcg_kmem_test_and_clear_dead(memcg))
5933 css_put(&memcg->css);
5936 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5941 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5945 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5950 static struct cftype mem_cgroup_files[] = {
5952 .name = "usage_in_bytes",
5953 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5954 .read = mem_cgroup_read,
5955 .register_event = mem_cgroup_usage_register_event,
5956 .unregister_event = mem_cgroup_usage_unregister_event,
5959 .name = "max_usage_in_bytes",
5960 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5961 .trigger = mem_cgroup_reset,
5962 .read = mem_cgroup_read,
5965 .name = "limit_in_bytes",
5966 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5967 .write_string = mem_cgroup_write,
5968 .read = mem_cgroup_read,
5971 .name = "soft_limit_in_bytes",
5972 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5973 .write_string = mem_cgroup_write,
5974 .read = mem_cgroup_read,
5978 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5979 .trigger = mem_cgroup_reset,
5980 .read = mem_cgroup_read,
5984 .read_seq_string = memcg_stat_show,
5987 .name = "force_empty",
5988 .trigger = mem_cgroup_force_empty_write,
5991 .name = "use_hierarchy",
5992 .flags = CFTYPE_INSANE,
5993 .write_u64 = mem_cgroup_hierarchy_write,
5994 .read_u64 = mem_cgroup_hierarchy_read,
5997 .name = "swappiness",
5998 .read_u64 = mem_cgroup_swappiness_read,
5999 .write_u64 = mem_cgroup_swappiness_write,
6002 .name = "move_charge_at_immigrate",
6003 .read_u64 = mem_cgroup_move_charge_read,
6004 .write_u64 = mem_cgroup_move_charge_write,
6007 .name = "oom_control",
6008 .read_map = mem_cgroup_oom_control_read,
6009 .write_u64 = mem_cgroup_oom_control_write,
6010 .register_event = mem_cgroup_oom_register_event,
6011 .unregister_event = mem_cgroup_oom_unregister_event,
6012 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6015 .name = "pressure_level",
6016 .register_event = vmpressure_register_event,
6017 .unregister_event = vmpressure_unregister_event,
6021 .name = "numa_stat",
6022 .read_seq_string = memcg_numa_stat_show,
6025 #ifdef CONFIG_MEMCG_KMEM
6027 .name = "kmem.limit_in_bytes",
6028 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6029 .write_string = mem_cgroup_write,
6030 .read = mem_cgroup_read,
6033 .name = "kmem.usage_in_bytes",
6034 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6035 .read = mem_cgroup_read,
6038 .name = "kmem.failcnt",
6039 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6040 .trigger = mem_cgroup_reset,
6041 .read = mem_cgroup_read,
6044 .name = "kmem.max_usage_in_bytes",
6045 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6046 .trigger = mem_cgroup_reset,
6047 .read = mem_cgroup_read,
6049 #ifdef CONFIG_SLABINFO
6051 .name = "kmem.slabinfo",
6052 .read_seq_string = mem_cgroup_slabinfo_read,
6056 { }, /* terminate */
6059 #ifdef CONFIG_MEMCG_SWAP
6060 static struct cftype memsw_cgroup_files[] = {
6062 .name = "memsw.usage_in_bytes",
6063 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6064 .read = mem_cgroup_read,
6065 .register_event = mem_cgroup_usage_register_event,
6066 .unregister_event = mem_cgroup_usage_unregister_event,
6069 .name = "memsw.max_usage_in_bytes",
6070 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6071 .trigger = mem_cgroup_reset,
6072 .read = mem_cgroup_read,
6075 .name = "memsw.limit_in_bytes",
6076 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6077 .write_string = mem_cgroup_write,
6078 .read = mem_cgroup_read,
6081 .name = "memsw.failcnt",
6082 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6083 .trigger = mem_cgroup_reset,
6084 .read = mem_cgroup_read,
6086 { }, /* terminate */
6089 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6091 struct mem_cgroup_per_node *pn;
6092 struct mem_cgroup_per_zone *mz;
6093 int zone, tmp = node;
6095 * This routine is called against possible nodes.
6096 * But it's BUG to call kmalloc() against offline node.
6098 * TODO: this routine can waste much memory for nodes which will
6099 * never be onlined. It's better to use memory hotplug callback
6102 if (!node_state(node, N_NORMAL_MEMORY))
6104 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6108 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6109 mz = &pn->zoneinfo[zone];
6110 lruvec_init(&mz->lruvec);
6111 mz->usage_in_excess = 0;
6112 mz->on_tree = false;
6115 memcg->nodeinfo[node] = pn;
6119 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6121 kfree(memcg->nodeinfo[node]);
6124 static struct mem_cgroup *mem_cgroup_alloc(void)
6126 struct mem_cgroup *memcg;
6127 size_t size = memcg_size();
6129 /* Can be very big if nr_node_ids is very big */
6130 if (size < PAGE_SIZE)
6131 memcg = kzalloc(size, GFP_KERNEL);
6133 memcg = vzalloc(size);
6138 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6141 spin_lock_init(&memcg->pcp_counter_lock);
6145 if (size < PAGE_SIZE)
6153 * At destroying mem_cgroup, references from swap_cgroup can remain.
6154 * (scanning all at force_empty is too costly...)
6156 * Instead of clearing all references at force_empty, we remember
6157 * the number of reference from swap_cgroup and free mem_cgroup when
6158 * it goes down to 0.
6160 * Removal of cgroup itself succeeds regardless of refs from swap.
6163 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6166 size_t size = memcg_size();
6168 mem_cgroup_remove_from_trees(memcg);
6169 free_css_id(&mem_cgroup_subsys, &memcg->css);
6172 free_mem_cgroup_per_zone_info(memcg, node);
6174 free_percpu(memcg->stat);
6177 * We need to make sure that (at least for now), the jump label
6178 * destruction code runs outside of the cgroup lock. This is because
6179 * get_online_cpus(), which is called from the static_branch update,
6180 * can't be called inside the cgroup_lock. cpusets are the ones
6181 * enforcing this dependency, so if they ever change, we might as well.
6183 * schedule_work() will guarantee this happens. Be careful if you need
6184 * to move this code around, and make sure it is outside
6187 disarm_static_keys(memcg);
6188 if (size < PAGE_SIZE)
6195 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6197 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6199 if (!memcg->res.parent)
6201 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6203 EXPORT_SYMBOL(parent_mem_cgroup);
6205 static void __init mem_cgroup_soft_limit_tree_init(void)
6207 struct mem_cgroup_tree_per_node *rtpn;
6208 struct mem_cgroup_tree_per_zone *rtpz;
6209 int tmp, node, zone;
6211 for_each_node(node) {
6213 if (!node_state(node, N_NORMAL_MEMORY))
6215 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6218 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6220 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6221 rtpz = &rtpn->rb_tree_per_zone[zone];
6222 rtpz->rb_root = RB_ROOT;
6223 spin_lock_init(&rtpz->lock);
6228 static struct cgroup_subsys_state * __ref
6229 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6231 struct mem_cgroup *memcg;
6232 long error = -ENOMEM;
6235 memcg = mem_cgroup_alloc();
6237 return ERR_PTR(error);
6240 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6244 if (parent_css == NULL) {
6245 root_mem_cgroup = memcg;
6246 res_counter_init(&memcg->res, NULL);
6247 res_counter_init(&memcg->memsw, NULL);
6248 res_counter_init(&memcg->kmem, NULL);
6251 memcg->last_scanned_node = MAX_NUMNODES;
6252 INIT_LIST_HEAD(&memcg->oom_notify);
6253 memcg->move_charge_at_immigrate = 0;
6254 mutex_init(&memcg->thresholds_lock);
6255 spin_lock_init(&memcg->move_lock);
6256 vmpressure_init(&memcg->vmpressure);
6261 __mem_cgroup_free(memcg);
6262 return ERR_PTR(error);
6266 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6268 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6269 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6275 mutex_lock(&memcg_create_mutex);
6277 memcg->use_hierarchy = parent->use_hierarchy;
6278 memcg->oom_kill_disable = parent->oom_kill_disable;
6279 memcg->swappiness = mem_cgroup_swappiness(parent);
6281 if (parent->use_hierarchy) {
6282 res_counter_init(&memcg->res, &parent->res);
6283 res_counter_init(&memcg->memsw, &parent->memsw);
6284 res_counter_init(&memcg->kmem, &parent->kmem);
6287 * No need to take a reference to the parent because cgroup
6288 * core guarantees its existence.
6291 res_counter_init(&memcg->res, NULL);
6292 res_counter_init(&memcg->memsw, NULL);
6293 res_counter_init(&memcg->kmem, NULL);
6295 * Deeper hierachy with use_hierarchy == false doesn't make
6296 * much sense so let cgroup subsystem know about this
6297 * unfortunate state in our controller.
6299 if (parent != root_mem_cgroup)
6300 mem_cgroup_subsys.broken_hierarchy = true;
6303 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6304 mutex_unlock(&memcg_create_mutex);
6309 * Announce all parents that a group from their hierarchy is gone.
6311 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6313 struct mem_cgroup *parent = memcg;
6315 while ((parent = parent_mem_cgroup(parent)))
6316 mem_cgroup_iter_invalidate(parent);
6319 * if the root memcg is not hierarchical we have to check it
6322 if (!root_mem_cgroup->use_hierarchy)
6323 mem_cgroup_iter_invalidate(root_mem_cgroup);
6326 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6328 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6330 kmem_cgroup_css_offline(memcg);
6332 mem_cgroup_invalidate_reclaim_iterators(memcg);
6333 mem_cgroup_reparent_charges(memcg);
6334 mem_cgroup_destroy_all_caches(memcg);
6335 vmpressure_cleanup(&memcg->vmpressure);
6338 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6340 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6342 memcg_destroy_kmem(memcg);
6343 __mem_cgroup_free(memcg);
6347 /* Handlers for move charge at task migration. */
6348 #define PRECHARGE_COUNT_AT_ONCE 256
6349 static int mem_cgroup_do_precharge(unsigned long count)
6352 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6353 struct mem_cgroup *memcg = mc.to;
6355 if (mem_cgroup_is_root(memcg)) {
6356 mc.precharge += count;
6357 /* we don't need css_get for root */
6360 /* try to charge at once */
6362 struct res_counter *dummy;
6364 * "memcg" cannot be under rmdir() because we've already checked
6365 * by cgroup_lock_live_cgroup() that it is not removed and we
6366 * are still under the same cgroup_mutex. So we can postpone
6369 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6371 if (do_swap_account && res_counter_charge(&memcg->memsw,
6372 PAGE_SIZE * count, &dummy)) {
6373 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6376 mc.precharge += count;
6380 /* fall back to one by one charge */
6382 if (signal_pending(current)) {
6386 if (!batch_count--) {
6387 batch_count = PRECHARGE_COUNT_AT_ONCE;
6390 ret = __mem_cgroup_try_charge(NULL,
6391 GFP_KERNEL, 1, &memcg, false);
6393 /* mem_cgroup_clear_mc() will do uncharge later */
6401 * get_mctgt_type - get target type of moving charge
6402 * @vma: the vma the pte to be checked belongs
6403 * @addr: the address corresponding to the pte to be checked
6404 * @ptent: the pte to be checked
6405 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6408 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6409 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6410 * move charge. if @target is not NULL, the page is stored in target->page
6411 * with extra refcnt got(Callers should handle it).
6412 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6413 * target for charge migration. if @target is not NULL, the entry is stored
6416 * Called with pte lock held.
6423 enum mc_target_type {
6429 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6430 unsigned long addr, pte_t ptent)
6432 struct page *page = vm_normal_page(vma, addr, ptent);
6434 if (!page || !page_mapped(page))
6436 if (PageAnon(page)) {
6437 /* we don't move shared anon */
6440 } else if (!move_file())
6441 /* we ignore mapcount for file pages */
6443 if (!get_page_unless_zero(page))
6450 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6451 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6453 struct page *page = NULL;
6454 swp_entry_t ent = pte_to_swp_entry(ptent);
6456 if (!move_anon() || non_swap_entry(ent))
6459 * Because lookup_swap_cache() updates some statistics counter,
6460 * we call find_get_page() with swapper_space directly.
6462 page = find_get_page(swap_address_space(ent), ent.val);
6463 if (do_swap_account)
6464 entry->val = ent.val;
6469 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6470 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6476 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6477 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6479 struct page *page = NULL;
6480 struct address_space *mapping;
6483 if (!vma->vm_file) /* anonymous vma */
6488 mapping = vma->vm_file->f_mapping;
6489 if (pte_none(ptent))
6490 pgoff = linear_page_index(vma, addr);
6491 else /* pte_file(ptent) is true */
6492 pgoff = pte_to_pgoff(ptent);
6494 /* page is moved even if it's not RSS of this task(page-faulted). */
6495 page = find_get_page(mapping, pgoff);
6498 /* shmem/tmpfs may report page out on swap: account for that too. */
6499 if (radix_tree_exceptional_entry(page)) {
6500 swp_entry_t swap = radix_to_swp_entry(page);
6501 if (do_swap_account)
6503 page = find_get_page(swap_address_space(swap), swap.val);
6509 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6510 unsigned long addr, pte_t ptent, union mc_target *target)
6512 struct page *page = NULL;
6513 struct page_cgroup *pc;
6514 enum mc_target_type ret = MC_TARGET_NONE;
6515 swp_entry_t ent = { .val = 0 };
6517 if (pte_present(ptent))
6518 page = mc_handle_present_pte(vma, addr, ptent);
6519 else if (is_swap_pte(ptent))
6520 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6521 else if (pte_none(ptent) || pte_file(ptent))
6522 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6524 if (!page && !ent.val)
6527 pc = lookup_page_cgroup(page);
6529 * Do only loose check w/o page_cgroup lock.
6530 * mem_cgroup_move_account() checks the pc is valid or not under
6533 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6534 ret = MC_TARGET_PAGE;
6536 target->page = page;
6538 if (!ret || !target)
6541 /* There is a swap entry and a page doesn't exist or isn't charged */
6542 if (ent.val && !ret &&
6543 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6544 ret = MC_TARGET_SWAP;
6551 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6553 * We don't consider swapping or file mapped pages because THP does not
6554 * support them for now.
6555 * Caller should make sure that pmd_trans_huge(pmd) is true.
6557 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6558 unsigned long addr, pmd_t pmd, union mc_target *target)
6560 struct page *page = NULL;
6561 struct page_cgroup *pc;
6562 enum mc_target_type ret = MC_TARGET_NONE;
6564 page = pmd_page(pmd);
6565 VM_BUG_ON(!page || !PageHead(page));
6568 pc = lookup_page_cgroup(page);
6569 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6570 ret = MC_TARGET_PAGE;
6573 target->page = page;
6579 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6580 unsigned long addr, pmd_t pmd, union mc_target *target)
6582 return MC_TARGET_NONE;
6586 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6587 unsigned long addr, unsigned long end,
6588 struct mm_walk *walk)
6590 struct vm_area_struct *vma = walk->private;
6594 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6595 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6596 mc.precharge += HPAGE_PMD_NR;
6597 spin_unlock(&vma->vm_mm->page_table_lock);
6601 if (pmd_trans_unstable(pmd))
6603 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6604 for (; addr != end; pte++, addr += PAGE_SIZE)
6605 if (get_mctgt_type(vma, addr, *pte, NULL))
6606 mc.precharge++; /* increment precharge temporarily */
6607 pte_unmap_unlock(pte - 1, ptl);
6613 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6615 unsigned long precharge;
6616 struct vm_area_struct *vma;
6618 down_read(&mm->mmap_sem);
6619 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6620 struct mm_walk mem_cgroup_count_precharge_walk = {
6621 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6625 if (is_vm_hugetlb_page(vma))
6627 walk_page_range(vma->vm_start, vma->vm_end,
6628 &mem_cgroup_count_precharge_walk);
6630 up_read(&mm->mmap_sem);
6632 precharge = mc.precharge;
6638 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6640 unsigned long precharge = mem_cgroup_count_precharge(mm);
6642 VM_BUG_ON(mc.moving_task);
6643 mc.moving_task = current;
6644 return mem_cgroup_do_precharge(precharge);
6647 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6648 static void __mem_cgroup_clear_mc(void)
6650 struct mem_cgroup *from = mc.from;
6651 struct mem_cgroup *to = mc.to;
6654 /* we must uncharge all the leftover precharges from mc.to */
6656 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6660 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6661 * we must uncharge here.
6663 if (mc.moved_charge) {
6664 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6665 mc.moved_charge = 0;
6667 /* we must fixup refcnts and charges */
6668 if (mc.moved_swap) {
6669 /* uncharge swap account from the old cgroup */
6670 if (!mem_cgroup_is_root(mc.from))
6671 res_counter_uncharge(&mc.from->memsw,
6672 PAGE_SIZE * mc.moved_swap);
6674 for (i = 0; i < mc.moved_swap; i++)
6675 css_put(&mc.from->css);
6677 if (!mem_cgroup_is_root(mc.to)) {
6679 * we charged both to->res and to->memsw, so we should
6682 res_counter_uncharge(&mc.to->res,
6683 PAGE_SIZE * mc.moved_swap);
6685 /* we've already done css_get(mc.to) */
6688 memcg_oom_recover(from);
6689 memcg_oom_recover(to);
6690 wake_up_all(&mc.waitq);
6693 static void mem_cgroup_clear_mc(void)
6695 struct mem_cgroup *from = mc.from;
6698 * we must clear moving_task before waking up waiters at the end of
6701 mc.moving_task = NULL;
6702 __mem_cgroup_clear_mc();
6703 spin_lock(&mc.lock);
6706 spin_unlock(&mc.lock);
6707 mem_cgroup_end_move(from);
6710 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6711 struct cgroup_taskset *tset)
6713 struct task_struct *p = cgroup_taskset_first(tset);
6715 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6716 unsigned long move_charge_at_immigrate;
6719 * We are now commited to this value whatever it is. Changes in this
6720 * tunable will only affect upcoming migrations, not the current one.
6721 * So we need to save it, and keep it going.
6723 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6724 if (move_charge_at_immigrate) {
6725 struct mm_struct *mm;
6726 struct mem_cgroup *from = mem_cgroup_from_task(p);
6728 VM_BUG_ON(from == memcg);
6730 mm = get_task_mm(p);
6733 /* We move charges only when we move a owner of the mm */
6734 if (mm->owner == p) {
6737 VM_BUG_ON(mc.precharge);
6738 VM_BUG_ON(mc.moved_charge);
6739 VM_BUG_ON(mc.moved_swap);
6740 mem_cgroup_start_move(from);
6741 spin_lock(&mc.lock);
6744 mc.immigrate_flags = move_charge_at_immigrate;
6745 spin_unlock(&mc.lock);
6746 /* We set mc.moving_task later */
6748 ret = mem_cgroup_precharge_mc(mm);
6750 mem_cgroup_clear_mc();
6757 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6758 struct cgroup_taskset *tset)
6760 mem_cgroup_clear_mc();
6763 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6764 unsigned long addr, unsigned long end,
6765 struct mm_walk *walk)
6768 struct vm_area_struct *vma = walk->private;
6771 enum mc_target_type target_type;
6772 union mc_target target;
6774 struct page_cgroup *pc;
6777 * We don't take compound_lock() here but no race with splitting thp
6779 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6780 * under splitting, which means there's no concurrent thp split,
6781 * - if another thread runs into split_huge_page() just after we
6782 * entered this if-block, the thread must wait for page table lock
6783 * to be unlocked in __split_huge_page_splitting(), where the main
6784 * part of thp split is not executed yet.
6786 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6787 if (mc.precharge < HPAGE_PMD_NR) {
6788 spin_unlock(&vma->vm_mm->page_table_lock);
6791 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6792 if (target_type == MC_TARGET_PAGE) {
6794 if (!isolate_lru_page(page)) {
6795 pc = lookup_page_cgroup(page);
6796 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6797 pc, mc.from, mc.to)) {
6798 mc.precharge -= HPAGE_PMD_NR;
6799 mc.moved_charge += HPAGE_PMD_NR;
6801 putback_lru_page(page);
6805 spin_unlock(&vma->vm_mm->page_table_lock);
6809 if (pmd_trans_unstable(pmd))
6812 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6813 for (; addr != end; addr += PAGE_SIZE) {
6814 pte_t ptent = *(pte++);
6820 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6821 case MC_TARGET_PAGE:
6823 if (isolate_lru_page(page))
6825 pc = lookup_page_cgroup(page);
6826 if (!mem_cgroup_move_account(page, 1, pc,
6829 /* we uncharge from mc.from later. */
6832 putback_lru_page(page);
6833 put: /* get_mctgt_type() gets the page */
6836 case MC_TARGET_SWAP:
6838 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6840 /* we fixup refcnts and charges later. */
6848 pte_unmap_unlock(pte - 1, ptl);
6853 * We have consumed all precharges we got in can_attach().
6854 * We try charge one by one, but don't do any additional
6855 * charges to mc.to if we have failed in charge once in attach()
6858 ret = mem_cgroup_do_precharge(1);
6866 static void mem_cgroup_move_charge(struct mm_struct *mm)
6868 struct vm_area_struct *vma;
6870 lru_add_drain_all();
6872 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6874 * Someone who are holding the mmap_sem might be waiting in
6875 * waitq. So we cancel all extra charges, wake up all waiters,
6876 * and retry. Because we cancel precharges, we might not be able
6877 * to move enough charges, but moving charge is a best-effort
6878 * feature anyway, so it wouldn't be a big problem.
6880 __mem_cgroup_clear_mc();
6884 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6886 struct mm_walk mem_cgroup_move_charge_walk = {
6887 .pmd_entry = mem_cgroup_move_charge_pte_range,
6891 if (is_vm_hugetlb_page(vma))
6893 ret = walk_page_range(vma->vm_start, vma->vm_end,
6894 &mem_cgroup_move_charge_walk);
6897 * means we have consumed all precharges and failed in
6898 * doing additional charge. Just abandon here.
6902 up_read(&mm->mmap_sem);
6905 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6906 struct cgroup_taskset *tset)
6908 struct task_struct *p = cgroup_taskset_first(tset);
6909 struct mm_struct *mm = get_task_mm(p);
6913 mem_cgroup_move_charge(mm);
6917 mem_cgroup_clear_mc();
6919 #else /* !CONFIG_MMU */
6920 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6921 struct cgroup_taskset *tset)
6925 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6926 struct cgroup_taskset *tset)
6929 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6930 struct cgroup_taskset *tset)
6936 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6937 * to verify sane_behavior flag on each mount attempt.
6939 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6942 * use_hierarchy is forced with sane_behavior. cgroup core
6943 * guarantees that @root doesn't have any children, so turning it
6944 * on for the root memcg is enough.
6946 if (cgroup_sane_behavior(root_css->cgroup))
6947 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6950 struct cgroup_subsys mem_cgroup_subsys = {
6952 .subsys_id = mem_cgroup_subsys_id,
6953 .css_alloc = mem_cgroup_css_alloc,
6954 .css_online = mem_cgroup_css_online,
6955 .css_offline = mem_cgroup_css_offline,
6956 .css_free = mem_cgroup_css_free,
6957 .can_attach = mem_cgroup_can_attach,
6958 .cancel_attach = mem_cgroup_cancel_attach,
6959 .attach = mem_cgroup_move_task,
6960 .bind = mem_cgroup_bind,
6961 .base_cftypes = mem_cgroup_files,
6966 #ifdef CONFIG_MEMCG_SWAP
6967 static int __init enable_swap_account(char *s)
6969 if (!strcmp(s, "1"))
6970 really_do_swap_account = 1;
6971 else if (!strcmp(s, "0"))
6972 really_do_swap_account = 0;
6975 __setup("swapaccount=", enable_swap_account);
6977 static void __init memsw_file_init(void)
6979 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6982 static void __init enable_swap_cgroup(void)
6984 if (!mem_cgroup_disabled() && really_do_swap_account) {
6985 do_swap_account = 1;
6991 static void __init enable_swap_cgroup(void)
6997 * subsys_initcall() for memory controller.
6999 * Some parts like hotcpu_notifier() have to be initialized from this context
7000 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7001 * everything that doesn't depend on a specific mem_cgroup structure should
7002 * be initialized from here.
7004 static int __init mem_cgroup_init(void)
7006 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7007 enable_swap_cgroup();
7008 mem_cgroup_soft_limit_tree_init();
7012 subsys_initcall(mem_cgroup_init);