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/poll.h>
49 #include <linux/sort.h>
51 #include <linux/seq_file.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>
58 #include <linux/file.h>
62 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys memory_cgrp_subsys __read_mostly;
70 EXPORT_SYMBOL(memory_cgrp_subsys);
72 #define MEM_CGROUP_RECLAIM_RETRIES 5
73 static struct mem_cgroup *root_mem_cgroup __read_mostly;
75 #ifdef CONFIG_MEMCG_SWAP
76 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
77 int do_swap_account __read_mostly;
79 /* for remember boot option*/
80 #ifdef CONFIG_MEMCG_SWAP_ENABLED
81 static int really_do_swap_account __initdata = 1;
83 static int really_do_swap_account __initdata;
87 #define do_swap_account 0
91 static const char * const mem_cgroup_stat_names[] = {
100 enum mem_cgroup_events_index {
101 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
102 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
103 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
104 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
105 MEM_CGROUP_EVENTS_NSTATS,
108 static const char * const mem_cgroup_events_names[] = {
115 static const char * const mem_cgroup_lru_names[] = {
124 * Per memcg event counter is incremented at every pagein/pageout. With THP,
125 * it will be incremated by the number of pages. This counter is used for
126 * for trigger some periodic events. This is straightforward and better
127 * than using jiffies etc. to handle periodic memcg event.
129 enum mem_cgroup_events_target {
130 MEM_CGROUP_TARGET_THRESH,
131 MEM_CGROUP_TARGET_SOFTLIMIT,
132 MEM_CGROUP_TARGET_NUMAINFO,
135 #define THRESHOLDS_EVENTS_TARGET 128
136 #define SOFTLIMIT_EVENTS_TARGET 1024
137 #define NUMAINFO_EVENTS_TARGET 1024
139 struct mem_cgroup_stat_cpu {
140 long count[MEM_CGROUP_STAT_NSTATS];
141 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
142 unsigned long nr_page_events;
143 unsigned long targets[MEM_CGROUP_NTARGETS];
146 struct mem_cgroup_reclaim_iter {
148 * last scanned hierarchy member. Valid only if last_dead_count
149 * matches memcg->dead_count of the hierarchy root group.
151 struct mem_cgroup *last_visited;
154 /* scan generation, increased every round-trip */
155 unsigned int generation;
159 * per-zone information in memory controller.
161 struct mem_cgroup_per_zone {
162 struct lruvec lruvec;
163 unsigned long lru_size[NR_LRU_LISTS];
165 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
167 struct rb_node tree_node; /* RB tree node */
168 unsigned long long usage_in_excess;/* Set to the value by which */
169 /* the soft limit is exceeded*/
171 struct mem_cgroup *memcg; /* Back pointer, we cannot */
172 /* use container_of */
175 struct mem_cgroup_per_node {
176 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
180 * Cgroups above their limits are maintained in a RB-Tree, independent of
181 * their hierarchy representation
184 struct mem_cgroup_tree_per_zone {
185 struct rb_root rb_root;
189 struct mem_cgroup_tree_per_node {
190 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
193 struct mem_cgroup_tree {
194 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
197 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
199 struct mem_cgroup_threshold {
200 struct eventfd_ctx *eventfd;
205 struct mem_cgroup_threshold_ary {
206 /* An array index points to threshold just below or equal to usage. */
207 int current_threshold;
208 /* Size of entries[] */
210 /* Array of thresholds */
211 struct mem_cgroup_threshold entries[0];
214 struct mem_cgroup_thresholds {
215 /* Primary thresholds array */
216 struct mem_cgroup_threshold_ary *primary;
218 * Spare threshold array.
219 * This is needed to make mem_cgroup_unregister_event() "never fail".
220 * It must be able to store at least primary->size - 1 entries.
222 struct mem_cgroup_threshold_ary *spare;
226 struct mem_cgroup_eventfd_list {
227 struct list_head list;
228 struct eventfd_ctx *eventfd;
232 * cgroup_event represents events which userspace want to receive.
234 struct mem_cgroup_event {
236 * memcg which the event belongs to.
238 struct mem_cgroup *memcg;
240 * eventfd to signal userspace about the event.
242 struct eventfd_ctx *eventfd;
244 * Each of these stored in a list by the cgroup.
246 struct list_head list;
248 * register_event() callback will be used to add new userspace
249 * waiter for changes related to this event. Use eventfd_signal()
250 * on eventfd to send notification to userspace.
252 int (*register_event)(struct mem_cgroup *memcg,
253 struct eventfd_ctx *eventfd, const char *args);
255 * unregister_event() callback will be called when userspace closes
256 * the eventfd or on cgroup removing. This callback must be set,
257 * if you want provide notification functionality.
259 void (*unregister_event)(struct mem_cgroup *memcg,
260 struct eventfd_ctx *eventfd);
262 * All fields below needed to unregister event when
263 * userspace closes eventfd.
266 wait_queue_head_t *wqh;
268 struct work_struct remove;
271 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
272 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
275 * The memory controller data structure. The memory controller controls both
276 * page cache and RSS per cgroup. We would eventually like to provide
277 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
278 * to help the administrator determine what knobs to tune.
280 * TODO: Add a water mark for the memory controller. Reclaim will begin when
281 * we hit the water mark. May be even add a low water mark, such that
282 * no reclaim occurs from a cgroup at it's low water mark, this is
283 * a feature that will be implemented much later in the future.
286 struct cgroup_subsys_state css;
288 * the counter to account for memory usage
290 struct res_counter res;
292 /* vmpressure notifications */
293 struct vmpressure vmpressure;
296 * the counter to account for mem+swap usage.
298 struct res_counter memsw;
301 * the counter to account for kernel memory usage.
303 struct res_counter kmem;
305 * Should the accounting and control be hierarchical, per subtree?
308 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
312 atomic_t oom_wakeups;
315 /* OOM-Killer disable */
316 int oom_kill_disable;
318 /* set when res.limit == memsw.limit */
319 bool memsw_is_minimum;
321 /* protect arrays of thresholds */
322 struct mutex thresholds_lock;
324 /* thresholds for memory usage. RCU-protected */
325 struct mem_cgroup_thresholds thresholds;
327 /* thresholds for mem+swap usage. RCU-protected */
328 struct mem_cgroup_thresholds memsw_thresholds;
330 /* For oom notifier event fd */
331 struct list_head oom_notify;
334 * Should we move charges of a task when a task is moved into this
335 * mem_cgroup ? And what type of charges should we move ?
337 unsigned long move_charge_at_immigrate;
339 * set > 0 if pages under this cgroup are moving to other cgroup.
341 atomic_t moving_account;
342 /* taken only while moving_account > 0 */
343 spinlock_t move_lock;
347 struct mem_cgroup_stat_cpu __percpu *stat;
349 * used when a cpu is offlined or other synchronizations
350 * See mem_cgroup_read_stat().
352 struct mem_cgroup_stat_cpu nocpu_base;
353 spinlock_t pcp_counter_lock;
356 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
357 struct cg_proto tcp_mem;
359 #if defined(CONFIG_MEMCG_KMEM)
360 /* analogous to slab_common's slab_caches list, but per-memcg;
361 * protected by memcg_slab_mutex */
362 struct list_head memcg_slab_caches;
363 /* Index in the kmem_cache->memcg_params->memcg_caches array */
367 int last_scanned_node;
369 nodemask_t scan_nodes;
370 atomic_t numainfo_events;
371 atomic_t numainfo_updating;
374 /* List of events which userspace want to receive */
375 struct list_head event_list;
376 spinlock_t event_list_lock;
378 struct mem_cgroup_per_node *nodeinfo[0];
379 /* WARNING: nodeinfo must be the last member here */
382 /* internal only representation about the status of kmem accounting. */
384 KMEM_ACCOUNTED_ACTIVE, /* accounted by this cgroup itself */
385 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
388 #ifdef CONFIG_MEMCG_KMEM
389 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
391 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
394 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
396 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
399 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
402 * Our caller must use css_get() first, because memcg_uncharge_kmem()
403 * will call css_put() if it sees the memcg is dead.
406 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
407 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
410 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
412 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
413 &memcg->kmem_account_flags);
417 /* Stuffs for move charges at task migration. */
419 * Types of charges to be moved. "move_charge_at_immitgrate" and
420 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
423 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
424 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
428 /* "mc" and its members are protected by cgroup_mutex */
429 static struct move_charge_struct {
430 spinlock_t lock; /* for from, to */
431 struct mem_cgroup *from;
432 struct mem_cgroup *to;
433 unsigned long immigrate_flags;
434 unsigned long precharge;
435 unsigned long moved_charge;
436 unsigned long moved_swap;
437 struct task_struct *moving_task; /* a task moving charges */
438 wait_queue_head_t waitq; /* a waitq for other context */
440 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
441 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
444 static bool move_anon(void)
446 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
449 static bool move_file(void)
451 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
455 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
456 * limit reclaim to prevent infinite loops, if they ever occur.
458 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
459 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
462 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
463 MEM_CGROUP_CHARGE_TYPE_ANON,
464 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
465 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
469 /* for encoding cft->private value on file */
477 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
478 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
479 #define MEMFILE_ATTR(val) ((val) & 0xffff)
480 /* Used for OOM nofiier */
481 #define OOM_CONTROL (0)
484 * Reclaim flags for mem_cgroup_hierarchical_reclaim
486 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
487 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
488 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
489 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
492 * The memcg_create_mutex will be held whenever a new cgroup is created.
493 * As a consequence, any change that needs to protect against new child cgroups
494 * appearing has to hold it as well.
496 static DEFINE_MUTEX(memcg_create_mutex);
498 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
500 return s ? container_of(s, struct mem_cgroup, css) : NULL;
503 /* Some nice accessors for the vmpressure. */
504 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
507 memcg = root_mem_cgroup;
508 return &memcg->vmpressure;
511 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
513 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
516 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
518 return (memcg == root_mem_cgroup);
522 * We restrict the id in the range of [1, 65535], so it can fit into
525 #define MEM_CGROUP_ID_MAX USHRT_MAX
527 static inline unsigned short mem_cgroup_id(struct mem_cgroup *memcg)
529 return memcg->css.id;
532 static inline struct mem_cgroup *mem_cgroup_from_id(unsigned short id)
534 struct cgroup_subsys_state *css;
536 css = css_from_id(id, &memory_cgrp_subsys);
537 return mem_cgroup_from_css(css);
540 /* Writing them here to avoid exposing memcg's inner layout */
541 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
543 void sock_update_memcg(struct sock *sk)
545 if (mem_cgroup_sockets_enabled) {
546 struct mem_cgroup *memcg;
547 struct cg_proto *cg_proto;
549 BUG_ON(!sk->sk_prot->proto_cgroup);
551 /* Socket cloning can throw us here with sk_cgrp already
552 * filled. It won't however, necessarily happen from
553 * process context. So the test for root memcg given
554 * the current task's memcg won't help us in this case.
556 * Respecting the original socket's memcg is a better
557 * decision in this case.
560 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
561 css_get(&sk->sk_cgrp->memcg->css);
566 memcg = mem_cgroup_from_task(current);
567 cg_proto = sk->sk_prot->proto_cgroup(memcg);
568 if (!mem_cgroup_is_root(memcg) &&
569 memcg_proto_active(cg_proto) &&
570 css_tryget_online(&memcg->css)) {
571 sk->sk_cgrp = cg_proto;
576 EXPORT_SYMBOL(sock_update_memcg);
578 void sock_release_memcg(struct sock *sk)
580 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
581 struct mem_cgroup *memcg;
582 WARN_ON(!sk->sk_cgrp->memcg);
583 memcg = sk->sk_cgrp->memcg;
584 css_put(&sk->sk_cgrp->memcg->css);
588 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
590 if (!memcg || mem_cgroup_is_root(memcg))
593 return &memcg->tcp_mem;
595 EXPORT_SYMBOL(tcp_proto_cgroup);
597 static void disarm_sock_keys(struct mem_cgroup *memcg)
599 if (!memcg_proto_activated(&memcg->tcp_mem))
601 static_key_slow_dec(&memcg_socket_limit_enabled);
604 static void disarm_sock_keys(struct mem_cgroup *memcg)
609 #ifdef CONFIG_MEMCG_KMEM
611 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
612 * The main reason for not using cgroup id for this:
613 * this works better in sparse environments, where we have a lot of memcgs,
614 * but only a few kmem-limited. Or also, if we have, for instance, 200
615 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
616 * 200 entry array for that.
618 * The current size of the caches array is stored in
619 * memcg_limited_groups_array_size. It will double each time we have to
622 static DEFINE_IDA(kmem_limited_groups);
623 int memcg_limited_groups_array_size;
626 * MIN_SIZE is different than 1, because we would like to avoid going through
627 * the alloc/free process all the time. In a small machine, 4 kmem-limited
628 * cgroups is a reasonable guess. In the future, it could be a parameter or
629 * tunable, but that is strictly not necessary.
631 * MAX_SIZE should be as large as the number of cgrp_ids. Ideally, we could get
632 * this constant directly from cgroup, but it is understandable that this is
633 * better kept as an internal representation in cgroup.c. In any case, the
634 * cgrp_id space is not getting any smaller, and we don't have to necessarily
635 * increase ours as well if it increases.
637 #define MEMCG_CACHES_MIN_SIZE 4
638 #define MEMCG_CACHES_MAX_SIZE MEM_CGROUP_ID_MAX
641 * A lot of the calls to the cache allocation functions are expected to be
642 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
643 * conditional to this static branch, we'll have to allow modules that does
644 * kmem_cache_alloc and the such to see this symbol as well
646 struct static_key memcg_kmem_enabled_key;
647 EXPORT_SYMBOL(memcg_kmem_enabled_key);
649 static void disarm_kmem_keys(struct mem_cgroup *memcg)
651 if (memcg_kmem_is_active(memcg)) {
652 static_key_slow_dec(&memcg_kmem_enabled_key);
653 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
656 * This check can't live in kmem destruction function,
657 * since the charges will outlive the cgroup
659 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
662 static void disarm_kmem_keys(struct mem_cgroup *memcg)
665 #endif /* CONFIG_MEMCG_KMEM */
667 static void disarm_static_keys(struct mem_cgroup *memcg)
669 disarm_sock_keys(memcg);
670 disarm_kmem_keys(memcg);
673 static void drain_all_stock_async(struct mem_cgroup *memcg);
675 static struct mem_cgroup_per_zone *
676 mem_cgroup_zone_zoneinfo(struct mem_cgroup *memcg, struct zone *zone)
678 int nid = zone_to_nid(zone);
679 int zid = zone_idx(zone);
681 return &memcg->nodeinfo[nid]->zoneinfo[zid];
684 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
689 static struct mem_cgroup_per_zone *
690 mem_cgroup_page_zoneinfo(struct mem_cgroup *memcg, struct page *page)
692 int nid = page_to_nid(page);
693 int zid = page_zonenum(page);
695 return &memcg->nodeinfo[nid]->zoneinfo[zid];
698 static struct mem_cgroup_tree_per_zone *
699 soft_limit_tree_node_zone(int nid, int zid)
701 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
704 static struct mem_cgroup_tree_per_zone *
705 soft_limit_tree_from_page(struct page *page)
707 int nid = page_to_nid(page);
708 int zid = page_zonenum(page);
710 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
713 static void __mem_cgroup_insert_exceeded(struct mem_cgroup_per_zone *mz,
714 struct mem_cgroup_tree_per_zone *mctz,
715 unsigned long long new_usage_in_excess)
717 struct rb_node **p = &mctz->rb_root.rb_node;
718 struct rb_node *parent = NULL;
719 struct mem_cgroup_per_zone *mz_node;
724 mz->usage_in_excess = new_usage_in_excess;
725 if (!mz->usage_in_excess)
729 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
731 if (mz->usage_in_excess < mz_node->usage_in_excess)
734 * We can't avoid mem cgroups that are over their soft
735 * limit by the same amount
737 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
740 rb_link_node(&mz->tree_node, parent, p);
741 rb_insert_color(&mz->tree_node, &mctz->rb_root);
745 static void __mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
746 struct mem_cgroup_tree_per_zone *mctz)
750 rb_erase(&mz->tree_node, &mctz->rb_root);
754 static void mem_cgroup_remove_exceeded(struct mem_cgroup_per_zone *mz,
755 struct mem_cgroup_tree_per_zone *mctz)
757 spin_lock(&mctz->lock);
758 __mem_cgroup_remove_exceeded(mz, mctz);
759 spin_unlock(&mctz->lock);
763 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
765 unsigned long long excess;
766 struct mem_cgroup_per_zone *mz;
767 struct mem_cgroup_tree_per_zone *mctz;
769 mctz = soft_limit_tree_from_page(page);
771 * Necessary to update all ancestors when hierarchy is used.
772 * because their event counter is not touched.
774 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
775 mz = mem_cgroup_page_zoneinfo(memcg, page);
776 excess = res_counter_soft_limit_excess(&memcg->res);
778 * We have to update the tree if mz is on RB-tree or
779 * mem is over its softlimit.
781 if (excess || mz->on_tree) {
782 spin_lock(&mctz->lock);
783 /* if on-tree, remove it */
785 __mem_cgroup_remove_exceeded(mz, mctz);
787 * Insert again. mz->usage_in_excess will be updated.
788 * If excess is 0, no tree ops.
790 __mem_cgroup_insert_exceeded(mz, mctz, excess);
791 spin_unlock(&mctz->lock);
796 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
798 struct mem_cgroup_tree_per_zone *mctz;
799 struct mem_cgroup_per_zone *mz;
803 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
804 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
805 mctz = soft_limit_tree_node_zone(nid, zid);
806 mem_cgroup_remove_exceeded(mz, mctz);
811 static struct mem_cgroup_per_zone *
812 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
814 struct rb_node *rightmost = NULL;
815 struct mem_cgroup_per_zone *mz;
819 rightmost = rb_last(&mctz->rb_root);
821 goto done; /* Nothing to reclaim from */
823 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
825 * Remove the node now but someone else can add it back,
826 * we will to add it back at the end of reclaim to its correct
827 * position in the tree.
829 __mem_cgroup_remove_exceeded(mz, mctz);
830 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
831 !css_tryget_online(&mz->memcg->css))
837 static struct mem_cgroup_per_zone *
838 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
840 struct mem_cgroup_per_zone *mz;
842 spin_lock(&mctz->lock);
843 mz = __mem_cgroup_largest_soft_limit_node(mctz);
844 spin_unlock(&mctz->lock);
849 * Implementation Note: reading percpu statistics for memcg.
851 * Both of vmstat[] and percpu_counter has threshold and do periodic
852 * synchronization to implement "quick" read. There are trade-off between
853 * reading cost and precision of value. Then, we may have a chance to implement
854 * a periodic synchronizion of counter in memcg's counter.
856 * But this _read() function is used for user interface now. The user accounts
857 * memory usage by memory cgroup and he _always_ requires exact value because
858 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
859 * have to visit all online cpus and make sum. So, for now, unnecessary
860 * synchronization is not implemented. (just implemented for cpu hotplug)
862 * If there are kernel internal actions which can make use of some not-exact
863 * value, and reading all cpu value can be performance bottleneck in some
864 * common workload, threashold and synchonization as vmstat[] should be
867 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
868 enum mem_cgroup_stat_index idx)
874 for_each_online_cpu(cpu)
875 val += per_cpu(memcg->stat->count[idx], cpu);
876 #ifdef CONFIG_HOTPLUG_CPU
877 spin_lock(&memcg->pcp_counter_lock);
878 val += memcg->nocpu_base.count[idx];
879 spin_unlock(&memcg->pcp_counter_lock);
885 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
888 int val = (charge) ? 1 : -1;
889 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
892 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
893 enum mem_cgroup_events_index idx)
895 unsigned long val = 0;
899 for_each_online_cpu(cpu)
900 val += per_cpu(memcg->stat->events[idx], cpu);
901 #ifdef CONFIG_HOTPLUG_CPU
902 spin_lock(&memcg->pcp_counter_lock);
903 val += memcg->nocpu_base.events[idx];
904 spin_unlock(&memcg->pcp_counter_lock);
910 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
912 bool anon, int nr_pages)
915 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
916 * counted as CACHE even if it's on ANON LRU.
919 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
922 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
925 if (PageTransHuge(page))
926 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
929 /* pagein of a big page is an event. So, ignore page size */
931 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
933 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
934 nr_pages = -nr_pages; /* for event */
937 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
940 unsigned long mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
942 struct mem_cgroup_per_zone *mz;
944 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
945 return mz->lru_size[lru];
948 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
950 unsigned int lru_mask)
952 unsigned long nr = 0;
955 VM_BUG_ON((unsigned)nid >= nr_node_ids);
957 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
958 struct mem_cgroup_per_zone *mz;
962 if (!(BIT(lru) & lru_mask))
964 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
965 nr += mz->lru_size[lru];
971 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
972 unsigned int lru_mask)
974 unsigned long nr = 0;
977 for_each_node_state(nid, N_MEMORY)
978 nr += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
982 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
983 enum mem_cgroup_events_target target)
985 unsigned long val, next;
987 val = __this_cpu_read(memcg->stat->nr_page_events);
988 next = __this_cpu_read(memcg->stat->targets[target]);
989 /* from time_after() in jiffies.h */
990 if ((long)next - (long)val < 0) {
992 case MEM_CGROUP_TARGET_THRESH:
993 next = val + THRESHOLDS_EVENTS_TARGET;
995 case MEM_CGROUP_TARGET_SOFTLIMIT:
996 next = val + SOFTLIMIT_EVENTS_TARGET;
998 case MEM_CGROUP_TARGET_NUMAINFO:
999 next = val + NUMAINFO_EVENTS_TARGET;
1004 __this_cpu_write(memcg->stat->targets[target], next);
1011 * Check events in order.
1014 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1017 /* threshold event is triggered in finer grain than soft limit */
1018 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1019 MEM_CGROUP_TARGET_THRESH))) {
1021 bool do_numainfo __maybe_unused;
1023 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1024 MEM_CGROUP_TARGET_SOFTLIMIT);
1025 #if MAX_NUMNODES > 1
1026 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1027 MEM_CGROUP_TARGET_NUMAINFO);
1031 mem_cgroup_threshold(memcg);
1032 if (unlikely(do_softlimit))
1033 mem_cgroup_update_tree(memcg, page);
1034 #if MAX_NUMNODES > 1
1035 if (unlikely(do_numainfo))
1036 atomic_inc(&memcg->numainfo_events);
1042 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1045 * mm_update_next_owner() may clear mm->owner to NULL
1046 * if it races with swapoff, page migration, etc.
1047 * So this can be called with p == NULL.
1052 return mem_cgroup_from_css(task_css(p, memory_cgrp_id));
1055 static struct mem_cgroup *get_mem_cgroup_from_mm(struct mm_struct *mm)
1057 struct mem_cgroup *memcg = NULL;
1062 * Page cache insertions can happen withou an
1063 * actual mm context, e.g. during disk probing
1064 * on boot, loopback IO, acct() writes etc.
1067 memcg = root_mem_cgroup;
1069 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1070 if (unlikely(!memcg))
1071 memcg = root_mem_cgroup;
1073 } while (!css_tryget_online(&memcg->css));
1079 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1080 * ref. count) or NULL if the whole root's subtree has been visited.
1082 * helper function to be used by mem_cgroup_iter
1084 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1085 struct mem_cgroup *last_visited)
1087 struct cgroup_subsys_state *prev_css, *next_css;
1089 prev_css = last_visited ? &last_visited->css : NULL;
1091 next_css = css_next_descendant_pre(prev_css, &root->css);
1094 * Even if we found a group we have to make sure it is
1095 * alive. css && !memcg means that the groups should be
1096 * skipped and we should continue the tree walk.
1097 * last_visited css is safe to use because it is
1098 * protected by css_get and the tree walk is rcu safe.
1100 * We do not take a reference on the root of the tree walk
1101 * because we might race with the root removal when it would
1102 * be the only node in the iterated hierarchy and mem_cgroup_iter
1103 * would end up in an endless loop because it expects that at
1104 * least one valid node will be returned. Root cannot disappear
1105 * because caller of the iterator should hold it already so
1106 * skipping css reference should be safe.
1109 if ((next_css == &root->css) ||
1110 ((next_css->flags & CSS_ONLINE) &&
1111 css_tryget_online(next_css)))
1112 return mem_cgroup_from_css(next_css);
1114 prev_css = next_css;
1121 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1124 * When a group in the hierarchy below root is destroyed, the
1125 * hierarchy iterator can no longer be trusted since it might
1126 * have pointed to the destroyed group. Invalidate it.
1128 atomic_inc(&root->dead_count);
1131 static struct mem_cgroup *
1132 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1133 struct mem_cgroup *root,
1136 struct mem_cgroup *position = NULL;
1138 * A cgroup destruction happens in two stages: offlining and
1139 * release. They are separated by a RCU grace period.
1141 * If the iterator is valid, we may still race with an
1142 * offlining. The RCU lock ensures the object won't be
1143 * released, tryget will fail if we lost the race.
1145 *sequence = atomic_read(&root->dead_count);
1146 if (iter->last_dead_count == *sequence) {
1148 position = iter->last_visited;
1151 * We cannot take a reference to root because we might race
1152 * with root removal and returning NULL would end up in
1153 * an endless loop on the iterator user level when root
1154 * would be returned all the time.
1156 if (position && position != root &&
1157 !css_tryget_online(&position->css))
1163 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1164 struct mem_cgroup *last_visited,
1165 struct mem_cgroup *new_position,
1166 struct mem_cgroup *root,
1169 /* root reference counting symmetric to mem_cgroup_iter_load */
1170 if (last_visited && last_visited != root)
1171 css_put(&last_visited->css);
1173 * We store the sequence count from the time @last_visited was
1174 * loaded successfully instead of rereading it here so that we
1175 * don't lose destruction events in between. We could have
1176 * raced with the destruction of @new_position after all.
1178 iter->last_visited = new_position;
1180 iter->last_dead_count = sequence;
1184 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1185 * @root: hierarchy root
1186 * @prev: previously returned memcg, NULL on first invocation
1187 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1189 * Returns references to children of the hierarchy below @root, or
1190 * @root itself, or %NULL after a full round-trip.
1192 * Caller must pass the return value in @prev on subsequent
1193 * invocations for reference counting, or use mem_cgroup_iter_break()
1194 * to cancel a hierarchy walk before the round-trip is complete.
1196 * Reclaimers can specify a zone and a priority level in @reclaim to
1197 * divide up the memcgs in the hierarchy among all concurrent
1198 * reclaimers operating on the same zone and priority.
1200 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1201 struct mem_cgroup *prev,
1202 struct mem_cgroup_reclaim_cookie *reclaim)
1204 struct mem_cgroup *memcg = NULL;
1205 struct mem_cgroup *last_visited = NULL;
1207 if (mem_cgroup_disabled())
1211 root = root_mem_cgroup;
1213 if (prev && !reclaim)
1214 last_visited = prev;
1216 if (!root->use_hierarchy && root != root_mem_cgroup) {
1224 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1225 int uninitialized_var(seq);
1228 struct mem_cgroup_per_zone *mz;
1230 mz = mem_cgroup_zone_zoneinfo(root, reclaim->zone);
1231 iter = &mz->reclaim_iter[reclaim->priority];
1232 if (prev && reclaim->generation != iter->generation) {
1233 iter->last_visited = NULL;
1237 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1240 memcg = __mem_cgroup_iter_next(root, last_visited);
1243 mem_cgroup_iter_update(iter, last_visited, memcg, root,
1248 else if (!prev && memcg)
1249 reclaim->generation = iter->generation;
1258 if (prev && prev != root)
1259 css_put(&prev->css);
1265 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1266 * @root: hierarchy root
1267 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1269 void mem_cgroup_iter_break(struct mem_cgroup *root,
1270 struct mem_cgroup *prev)
1273 root = root_mem_cgroup;
1274 if (prev && prev != root)
1275 css_put(&prev->css);
1279 * Iteration constructs for visiting all cgroups (under a tree). If
1280 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1281 * be used for reference counting.
1283 #define for_each_mem_cgroup_tree(iter, root) \
1284 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1286 iter = mem_cgroup_iter(root, iter, NULL))
1288 #define for_each_mem_cgroup(iter) \
1289 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1291 iter = mem_cgroup_iter(NULL, iter, NULL))
1293 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1295 struct mem_cgroup *memcg;
1298 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1299 if (unlikely(!memcg))
1304 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1307 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1315 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1318 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1319 * @zone: zone of the wanted lruvec
1320 * @memcg: memcg of the wanted lruvec
1322 * Returns the lru list vector holding pages for the given @zone and
1323 * @mem. This can be the global zone lruvec, if the memory controller
1326 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1327 struct mem_cgroup *memcg)
1329 struct mem_cgroup_per_zone *mz;
1330 struct lruvec *lruvec;
1332 if (mem_cgroup_disabled()) {
1333 lruvec = &zone->lruvec;
1337 mz = mem_cgroup_zone_zoneinfo(memcg, zone);
1338 lruvec = &mz->lruvec;
1341 * Since a node can be onlined after the mem_cgroup was created,
1342 * we have to be prepared to initialize lruvec->zone here;
1343 * and if offlined then reonlined, we need to reinitialize it.
1345 if (unlikely(lruvec->zone != zone))
1346 lruvec->zone = zone;
1351 * Following LRU functions are allowed to be used without PCG_LOCK.
1352 * Operations are called by routine of global LRU independently from memcg.
1353 * What we have to take care of here is validness of pc->mem_cgroup.
1355 * Changes to pc->mem_cgroup happens when
1358 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1359 * It is added to LRU before charge.
1360 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1361 * When moving account, the page is not on LRU. It's isolated.
1365 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1367 * @zone: zone of the page
1369 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1371 struct mem_cgroup_per_zone *mz;
1372 struct mem_cgroup *memcg;
1373 struct page_cgroup *pc;
1374 struct lruvec *lruvec;
1376 if (mem_cgroup_disabled()) {
1377 lruvec = &zone->lruvec;
1381 pc = lookup_page_cgroup(page);
1382 memcg = pc->mem_cgroup;
1385 * Surreptitiously switch any uncharged offlist page to root:
1386 * an uncharged page off lru does nothing to secure
1387 * its former mem_cgroup from sudden removal.
1389 * Our caller holds lru_lock, and PageCgroupUsed is updated
1390 * under page_cgroup lock: between them, they make all uses
1391 * of pc->mem_cgroup safe.
1393 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1394 pc->mem_cgroup = memcg = root_mem_cgroup;
1396 mz = mem_cgroup_page_zoneinfo(memcg, page);
1397 lruvec = &mz->lruvec;
1400 * Since a node can be onlined after the mem_cgroup was created,
1401 * we have to be prepared to initialize lruvec->zone here;
1402 * and if offlined then reonlined, we need to reinitialize it.
1404 if (unlikely(lruvec->zone != zone))
1405 lruvec->zone = zone;
1410 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1411 * @lruvec: mem_cgroup per zone lru vector
1412 * @lru: index of lru list the page is sitting on
1413 * @nr_pages: positive when adding or negative when removing
1415 * This function must be called when a page is added to or removed from an
1418 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1421 struct mem_cgroup_per_zone *mz;
1422 unsigned long *lru_size;
1424 if (mem_cgroup_disabled())
1427 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1428 lru_size = mz->lru_size + lru;
1429 *lru_size += nr_pages;
1430 VM_BUG_ON((long)(*lru_size) < 0);
1434 * Checks whether given mem is same or in the root_mem_cgroup's
1437 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1438 struct mem_cgroup *memcg)
1440 if (root_memcg == memcg)
1442 if (!root_memcg->use_hierarchy || !memcg)
1444 return cgroup_is_descendant(memcg->css.cgroup, root_memcg->css.cgroup);
1447 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1448 struct mem_cgroup *memcg)
1453 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1458 bool task_in_mem_cgroup(struct task_struct *task,
1459 const struct mem_cgroup *memcg)
1461 struct mem_cgroup *curr = NULL;
1462 struct task_struct *p;
1465 p = find_lock_task_mm(task);
1467 curr = get_mem_cgroup_from_mm(p->mm);
1471 * All threads may have already detached their mm's, but the oom
1472 * killer still needs to detect if they have already been oom
1473 * killed to prevent needlessly killing additional tasks.
1476 curr = mem_cgroup_from_task(task);
1478 css_get(&curr->css);
1482 * We should check use_hierarchy of "memcg" not "curr". Because checking
1483 * use_hierarchy of "curr" here make this function true if hierarchy is
1484 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1485 * hierarchy(even if use_hierarchy is disabled in "memcg").
1487 ret = mem_cgroup_same_or_subtree(memcg, curr);
1488 css_put(&curr->css);
1492 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1494 unsigned long inactive_ratio;
1495 unsigned long inactive;
1496 unsigned long active;
1499 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1500 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1502 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1504 inactive_ratio = int_sqrt(10 * gb);
1508 return inactive * inactive_ratio < active;
1511 #define mem_cgroup_from_res_counter(counter, member) \
1512 container_of(counter, struct mem_cgroup, member)
1515 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1516 * @memcg: the memory cgroup
1518 * Returns the maximum amount of memory @mem can be charged with, in
1521 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1523 unsigned long long margin;
1525 margin = res_counter_margin(&memcg->res);
1526 if (do_swap_account)
1527 margin = min(margin, res_counter_margin(&memcg->memsw));
1528 return margin >> PAGE_SHIFT;
1531 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1534 if (mem_cgroup_disabled() || !memcg->css.parent)
1535 return vm_swappiness;
1537 return memcg->swappiness;
1541 * memcg->moving_account is used for checking possibility that some thread is
1542 * calling move_account(). When a thread on CPU-A starts moving pages under
1543 * a memcg, other threads should check memcg->moving_account under
1544 * rcu_read_lock(), like this:
1548 * memcg->moving_account+1 if (memcg->mocing_account)
1550 * synchronize_rcu() update something.
1555 /* for quick checking without looking up memcg */
1556 atomic_t memcg_moving __read_mostly;
1558 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1560 atomic_inc(&memcg_moving);
1561 atomic_inc(&memcg->moving_account);
1565 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1568 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1569 * We check NULL in callee rather than caller.
1572 atomic_dec(&memcg_moving);
1573 atomic_dec(&memcg->moving_account);
1578 * A routine for checking "mem" is under move_account() or not.
1580 * Checking a cgroup is mc.from or mc.to or under hierarchy of
1581 * moving cgroups. This is for waiting at high-memory pressure
1584 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1586 struct mem_cgroup *from;
1587 struct mem_cgroup *to;
1590 * Unlike task_move routines, we access mc.to, mc.from not under
1591 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1593 spin_lock(&mc.lock);
1599 ret = mem_cgroup_same_or_subtree(memcg, from)
1600 || mem_cgroup_same_or_subtree(memcg, to);
1602 spin_unlock(&mc.lock);
1606 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1608 if (mc.moving_task && current != mc.moving_task) {
1609 if (mem_cgroup_under_move(memcg)) {
1611 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1612 /* moving charge context might have finished. */
1615 finish_wait(&mc.waitq, &wait);
1623 * Take this lock when
1624 * - a code tries to modify page's memcg while it's USED.
1625 * - a code tries to modify page state accounting in a memcg.
1627 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1628 unsigned long *flags)
1630 spin_lock_irqsave(&memcg->move_lock, *flags);
1633 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1634 unsigned long *flags)
1636 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1639 #define K(x) ((x) << (PAGE_SHIFT-10))
1641 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1642 * @memcg: The memory cgroup that went over limit
1643 * @p: Task that is going to be killed
1645 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1648 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1650 /* oom_info_lock ensures that parallel ooms do not interleave */
1651 static DEFINE_MUTEX(oom_info_lock);
1652 struct mem_cgroup *iter;
1658 mutex_lock(&oom_info_lock);
1661 pr_info("Task in ");
1662 pr_cont_cgroup_path(task_cgroup(p, memory_cgrp_id));
1663 pr_info(" killed as a result of limit of ");
1664 pr_cont_cgroup_path(memcg->css.cgroup);
1669 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1670 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1671 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1672 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1673 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1674 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1675 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1676 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1677 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1678 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1679 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1680 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1682 for_each_mem_cgroup_tree(iter, memcg) {
1683 pr_info("Memory cgroup stats for ");
1684 pr_cont_cgroup_path(iter->css.cgroup);
1687 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1688 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1690 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1691 K(mem_cgroup_read_stat(iter, i)));
1694 for (i = 0; i < NR_LRU_LISTS; i++)
1695 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1696 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1700 mutex_unlock(&oom_info_lock);
1704 * This function returns the number of memcg under hierarchy tree. Returns
1705 * 1(self count) if no children.
1707 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1710 struct mem_cgroup *iter;
1712 for_each_mem_cgroup_tree(iter, memcg)
1718 * Return the memory (and swap, if configured) limit for a memcg.
1720 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1724 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1727 * Do not consider swap space if we cannot swap due to swappiness
1729 if (mem_cgroup_swappiness(memcg)) {
1732 limit += total_swap_pages << PAGE_SHIFT;
1733 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1736 * If memsw is finite and limits the amount of swap space
1737 * available to this memcg, return that limit.
1739 limit = min(limit, memsw);
1745 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1748 struct mem_cgroup *iter;
1749 unsigned long chosen_points = 0;
1750 unsigned long totalpages;
1751 unsigned int points = 0;
1752 struct task_struct *chosen = NULL;
1755 * If current has a pending SIGKILL or is exiting, then automatically
1756 * select it. The goal is to allow it to allocate so that it may
1757 * quickly exit and free its memory.
1759 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1760 set_thread_flag(TIF_MEMDIE);
1764 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1765 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1766 for_each_mem_cgroup_tree(iter, memcg) {
1767 struct css_task_iter it;
1768 struct task_struct *task;
1770 css_task_iter_start(&iter->css, &it);
1771 while ((task = css_task_iter_next(&it))) {
1772 switch (oom_scan_process_thread(task, totalpages, NULL,
1774 case OOM_SCAN_SELECT:
1776 put_task_struct(chosen);
1778 chosen_points = ULONG_MAX;
1779 get_task_struct(chosen);
1781 case OOM_SCAN_CONTINUE:
1783 case OOM_SCAN_ABORT:
1784 css_task_iter_end(&it);
1785 mem_cgroup_iter_break(memcg, iter);
1787 put_task_struct(chosen);
1792 points = oom_badness(task, memcg, NULL, totalpages);
1793 if (!points || points < chosen_points)
1795 /* Prefer thread group leaders for display purposes */
1796 if (points == chosen_points &&
1797 thread_group_leader(chosen))
1801 put_task_struct(chosen);
1803 chosen_points = points;
1804 get_task_struct(chosen);
1806 css_task_iter_end(&it);
1811 points = chosen_points * 1000 / totalpages;
1812 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1813 NULL, "Memory cgroup out of memory");
1816 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1818 unsigned long flags)
1820 unsigned long total = 0;
1821 bool noswap = false;
1824 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1826 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1829 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1831 drain_all_stock_async(memcg);
1832 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1834 * Allow limit shrinkers, which are triggered directly
1835 * by userspace, to catch signals and stop reclaim
1836 * after minimal progress, regardless of the margin.
1838 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1840 if (mem_cgroup_margin(memcg))
1843 * If nothing was reclaimed after two attempts, there
1844 * may be no reclaimable pages in this hierarchy.
1853 * test_mem_cgroup_node_reclaimable
1854 * @memcg: the target memcg
1855 * @nid: the node ID to be checked.
1856 * @noswap : specify true here if the user wants flle only information.
1858 * This function returns whether the specified memcg contains any
1859 * reclaimable pages on a node. Returns true if there are any reclaimable
1860 * pages in the node.
1862 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1863 int nid, bool noswap)
1865 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1867 if (noswap || !total_swap_pages)
1869 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1874 #if MAX_NUMNODES > 1
1877 * Always updating the nodemask is not very good - even if we have an empty
1878 * list or the wrong list here, we can start from some node and traverse all
1879 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1882 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1886 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1887 * pagein/pageout changes since the last update.
1889 if (!atomic_read(&memcg->numainfo_events))
1891 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1894 /* make a nodemask where this memcg uses memory from */
1895 memcg->scan_nodes = node_states[N_MEMORY];
1897 for_each_node_mask(nid, node_states[N_MEMORY]) {
1899 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1900 node_clear(nid, memcg->scan_nodes);
1903 atomic_set(&memcg->numainfo_events, 0);
1904 atomic_set(&memcg->numainfo_updating, 0);
1908 * Selecting a node where we start reclaim from. Because what we need is just
1909 * reducing usage counter, start from anywhere is O,K. Considering
1910 * memory reclaim from current node, there are pros. and cons.
1912 * Freeing memory from current node means freeing memory from a node which
1913 * we'll use or we've used. So, it may make LRU bad. And if several threads
1914 * hit limits, it will see a contention on a node. But freeing from remote
1915 * node means more costs for memory reclaim because of memory latency.
1917 * Now, we use round-robin. Better algorithm is welcomed.
1919 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1923 mem_cgroup_may_update_nodemask(memcg);
1924 node = memcg->last_scanned_node;
1926 node = next_node(node, memcg->scan_nodes);
1927 if (node == MAX_NUMNODES)
1928 node = first_node(memcg->scan_nodes);
1930 * We call this when we hit limit, not when pages are added to LRU.
1931 * No LRU may hold pages because all pages are UNEVICTABLE or
1932 * memcg is too small and all pages are not on LRU. In that case,
1933 * we use curret node.
1935 if (unlikely(node == MAX_NUMNODES))
1936 node = numa_node_id();
1938 memcg->last_scanned_node = node;
1943 * Check all nodes whether it contains reclaimable pages or not.
1944 * For quick scan, we make use of scan_nodes. This will allow us to skip
1945 * unused nodes. But scan_nodes is lazily updated and may not cotain
1946 * enough new information. We need to do double check.
1948 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1953 * quick check...making use of scan_node.
1954 * We can skip unused nodes.
1956 if (!nodes_empty(memcg->scan_nodes)) {
1957 for (nid = first_node(memcg->scan_nodes);
1959 nid = next_node(nid, memcg->scan_nodes)) {
1961 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1966 * Check rest of nodes.
1968 for_each_node_state(nid, N_MEMORY) {
1969 if (node_isset(nid, memcg->scan_nodes))
1971 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1978 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1983 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1985 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1989 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1992 unsigned long *total_scanned)
1994 struct mem_cgroup *victim = NULL;
1997 unsigned long excess;
1998 unsigned long nr_scanned;
1999 struct mem_cgroup_reclaim_cookie reclaim = {
2004 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2007 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2012 * If we have not been able to reclaim
2013 * anything, it might because there are
2014 * no reclaimable pages under this hierarchy
2019 * We want to do more targeted reclaim.
2020 * excess >> 2 is not to excessive so as to
2021 * reclaim too much, nor too less that we keep
2022 * coming back to reclaim from this cgroup
2024 if (total >= (excess >> 2) ||
2025 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2030 if (!mem_cgroup_reclaimable(victim, false))
2032 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2034 *total_scanned += nr_scanned;
2035 if (!res_counter_soft_limit_excess(&root_memcg->res))
2038 mem_cgroup_iter_break(root_memcg, victim);
2042 #ifdef CONFIG_LOCKDEP
2043 static struct lockdep_map memcg_oom_lock_dep_map = {
2044 .name = "memcg_oom_lock",
2048 static DEFINE_SPINLOCK(memcg_oom_lock);
2051 * Check OOM-Killer is already running under our hierarchy.
2052 * If someone is running, return false.
2054 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2056 struct mem_cgroup *iter, *failed = NULL;
2058 spin_lock(&memcg_oom_lock);
2060 for_each_mem_cgroup_tree(iter, memcg) {
2061 if (iter->oom_lock) {
2063 * this subtree of our hierarchy is already locked
2064 * so we cannot give a lock.
2067 mem_cgroup_iter_break(memcg, iter);
2070 iter->oom_lock = true;
2075 * OK, we failed to lock the whole subtree so we have
2076 * to clean up what we set up to the failing subtree
2078 for_each_mem_cgroup_tree(iter, memcg) {
2079 if (iter == failed) {
2080 mem_cgroup_iter_break(memcg, iter);
2083 iter->oom_lock = false;
2086 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2088 spin_unlock(&memcg_oom_lock);
2093 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2095 struct mem_cgroup *iter;
2097 spin_lock(&memcg_oom_lock);
2098 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2099 for_each_mem_cgroup_tree(iter, memcg)
2100 iter->oom_lock = false;
2101 spin_unlock(&memcg_oom_lock);
2104 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2106 struct mem_cgroup *iter;
2108 for_each_mem_cgroup_tree(iter, memcg)
2109 atomic_inc(&iter->under_oom);
2112 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2114 struct mem_cgroup *iter;
2117 * When a new child is created while the hierarchy is under oom,
2118 * mem_cgroup_oom_lock() may not be called. We have to use
2119 * atomic_add_unless() here.
2121 for_each_mem_cgroup_tree(iter, memcg)
2122 atomic_add_unless(&iter->under_oom, -1, 0);
2125 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2127 struct oom_wait_info {
2128 struct mem_cgroup *memcg;
2132 static int memcg_oom_wake_function(wait_queue_t *wait,
2133 unsigned mode, int sync, void *arg)
2135 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2136 struct mem_cgroup *oom_wait_memcg;
2137 struct oom_wait_info *oom_wait_info;
2139 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2140 oom_wait_memcg = oom_wait_info->memcg;
2143 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2144 * Then we can use css_is_ancestor without taking care of RCU.
2146 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2147 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2149 return autoremove_wake_function(wait, mode, sync, arg);
2152 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2154 atomic_inc(&memcg->oom_wakeups);
2155 /* for filtering, pass "memcg" as argument. */
2156 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2159 static void memcg_oom_recover(struct mem_cgroup *memcg)
2161 if (memcg && atomic_read(&memcg->under_oom))
2162 memcg_wakeup_oom(memcg);
2165 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2167 if (!current->memcg_oom.may_oom)
2170 * We are in the middle of the charge context here, so we
2171 * don't want to block when potentially sitting on a callstack
2172 * that holds all kinds of filesystem and mm locks.
2174 * Also, the caller may handle a failed allocation gracefully
2175 * (like optional page cache readahead) and so an OOM killer
2176 * invocation might not even be necessary.
2178 * That's why we don't do anything here except remember the
2179 * OOM context and then deal with it at the end of the page
2180 * fault when the stack is unwound, the locks are released,
2181 * and when we know whether the fault was overall successful.
2183 css_get(&memcg->css);
2184 current->memcg_oom.memcg = memcg;
2185 current->memcg_oom.gfp_mask = mask;
2186 current->memcg_oom.order = order;
2190 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2191 * @handle: actually kill/wait or just clean up the OOM state
2193 * This has to be called at the end of a page fault if the memcg OOM
2194 * handler was enabled.
2196 * Memcg supports userspace OOM handling where failed allocations must
2197 * sleep on a waitqueue until the userspace task resolves the
2198 * situation. Sleeping directly in the charge context with all kinds
2199 * of locks held is not a good idea, instead we remember an OOM state
2200 * in the task and mem_cgroup_oom_synchronize() has to be called at
2201 * the end of the page fault to complete the OOM handling.
2203 * Returns %true if an ongoing memcg OOM situation was detected and
2204 * completed, %false otherwise.
2206 bool mem_cgroup_oom_synchronize(bool handle)
2208 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2209 struct oom_wait_info owait;
2212 /* OOM is global, do not handle */
2219 owait.memcg = memcg;
2220 owait.wait.flags = 0;
2221 owait.wait.func = memcg_oom_wake_function;
2222 owait.wait.private = current;
2223 INIT_LIST_HEAD(&owait.wait.task_list);
2225 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2226 mem_cgroup_mark_under_oom(memcg);
2228 locked = mem_cgroup_oom_trylock(memcg);
2231 mem_cgroup_oom_notify(memcg);
2233 if (locked && !memcg->oom_kill_disable) {
2234 mem_cgroup_unmark_under_oom(memcg);
2235 finish_wait(&memcg_oom_waitq, &owait.wait);
2236 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2237 current->memcg_oom.order);
2240 mem_cgroup_unmark_under_oom(memcg);
2241 finish_wait(&memcg_oom_waitq, &owait.wait);
2245 mem_cgroup_oom_unlock(memcg);
2247 * There is no guarantee that an OOM-lock contender
2248 * sees the wakeups triggered by the OOM kill
2249 * uncharges. Wake any sleepers explicitely.
2251 memcg_oom_recover(memcg);
2254 current->memcg_oom.memcg = NULL;
2255 css_put(&memcg->css);
2260 * Used to update mapped file or writeback or other statistics.
2262 * Notes: Race condition
2264 * We usually use lock_page_cgroup() for accessing page_cgroup member but
2265 * it tends to be costly. But considering some conditions, we doesn't need
2266 * to do so _always_.
2268 * Considering "charge", lock_page_cgroup() is not required because all
2269 * file-stat operations happen after a page is attached to radix-tree. There
2270 * are no race with "charge".
2272 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2273 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2274 * if there are race with "uncharge". Statistics itself is properly handled
2277 * Considering "move", this is an only case we see a race. To make the race
2278 * small, we check memcg->moving_account and detect there are possibility
2279 * of race or not. If there is, we take a lock.
2282 void __mem_cgroup_begin_update_page_stat(struct page *page,
2283 bool *locked, unsigned long *flags)
2285 struct mem_cgroup *memcg;
2286 struct page_cgroup *pc;
2288 pc = lookup_page_cgroup(page);
2290 memcg = pc->mem_cgroup;
2291 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2294 * If this memory cgroup is not under account moving, we don't
2295 * need to take move_lock_mem_cgroup(). Because we already hold
2296 * rcu_read_lock(), any calls to move_account will be delayed until
2297 * rcu_read_unlock().
2299 VM_BUG_ON(!rcu_read_lock_held());
2300 if (atomic_read(&memcg->moving_account) <= 0)
2303 move_lock_mem_cgroup(memcg, flags);
2304 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2305 move_unlock_mem_cgroup(memcg, flags);
2311 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2313 struct page_cgroup *pc = lookup_page_cgroup(page);
2316 * It's guaranteed that pc->mem_cgroup never changes while
2317 * lock is held because a routine modifies pc->mem_cgroup
2318 * should take move_lock_mem_cgroup().
2320 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2323 void mem_cgroup_update_page_stat(struct page *page,
2324 enum mem_cgroup_stat_index idx, int val)
2326 struct mem_cgroup *memcg;
2327 struct page_cgroup *pc = lookup_page_cgroup(page);
2328 unsigned long uninitialized_var(flags);
2330 if (mem_cgroup_disabled())
2333 VM_BUG_ON(!rcu_read_lock_held());
2334 memcg = pc->mem_cgroup;
2335 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2338 this_cpu_add(memcg->stat->count[idx], val);
2342 * size of first charge trial. "32" comes from vmscan.c's magic value.
2343 * TODO: maybe necessary to use big numbers in big irons.
2345 #define CHARGE_BATCH 32U
2346 struct memcg_stock_pcp {
2347 struct mem_cgroup *cached; /* this never be root cgroup */
2348 unsigned int nr_pages;
2349 struct work_struct work;
2350 unsigned long flags;
2351 #define FLUSHING_CACHED_CHARGE 0
2353 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2354 static DEFINE_MUTEX(percpu_charge_mutex);
2357 * consume_stock: Try to consume stocked charge on this cpu.
2358 * @memcg: memcg to consume from.
2359 * @nr_pages: how many pages to charge.
2361 * The charges will only happen if @memcg matches the current cpu's memcg
2362 * stock, and at least @nr_pages are available in that stock. Failure to
2363 * service an allocation will refill the stock.
2365 * returns true if successful, false otherwise.
2367 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2369 struct memcg_stock_pcp *stock;
2372 if (nr_pages > CHARGE_BATCH)
2375 stock = &get_cpu_var(memcg_stock);
2376 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2377 stock->nr_pages -= nr_pages;
2378 else /* need to call res_counter_charge */
2380 put_cpu_var(memcg_stock);
2385 * Returns stocks cached in percpu to res_counter and reset cached information.
2387 static void drain_stock(struct memcg_stock_pcp *stock)
2389 struct mem_cgroup *old = stock->cached;
2391 if (stock->nr_pages) {
2392 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2394 res_counter_uncharge(&old->res, bytes);
2395 if (do_swap_account)
2396 res_counter_uncharge(&old->memsw, bytes);
2397 stock->nr_pages = 0;
2399 stock->cached = NULL;
2403 * This must be called under preempt disabled or must be called by
2404 * a thread which is pinned to local cpu.
2406 static void drain_local_stock(struct work_struct *dummy)
2408 struct memcg_stock_pcp *stock = this_cpu_ptr(&memcg_stock);
2410 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2413 static void __init memcg_stock_init(void)
2417 for_each_possible_cpu(cpu) {
2418 struct memcg_stock_pcp *stock =
2419 &per_cpu(memcg_stock, cpu);
2420 INIT_WORK(&stock->work, drain_local_stock);
2425 * Cache charges(val) which is from res_counter, to local per_cpu area.
2426 * This will be consumed by consume_stock() function, later.
2428 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2430 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2432 if (stock->cached != memcg) { /* reset if necessary */
2434 stock->cached = memcg;
2436 stock->nr_pages += nr_pages;
2437 put_cpu_var(memcg_stock);
2441 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2442 * of the hierarchy under it. sync flag says whether we should block
2443 * until the work is done.
2445 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2449 /* Notify other cpus that system-wide "drain" is running */
2452 for_each_online_cpu(cpu) {
2453 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2454 struct mem_cgroup *memcg;
2456 memcg = stock->cached;
2457 if (!memcg || !stock->nr_pages)
2459 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2461 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2463 drain_local_stock(&stock->work);
2465 schedule_work_on(cpu, &stock->work);
2473 for_each_online_cpu(cpu) {
2474 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2475 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2476 flush_work(&stock->work);
2483 * Tries to drain stocked charges in other cpus. This function is asynchronous
2484 * and just put a work per cpu for draining localy on each cpu. Caller can
2485 * expects some charges will be back to res_counter later but cannot wait for
2488 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2491 * If someone calls draining, avoid adding more kworker runs.
2493 if (!mutex_trylock(&percpu_charge_mutex))
2495 drain_all_stock(root_memcg, false);
2496 mutex_unlock(&percpu_charge_mutex);
2499 /* This is a synchronous drain interface. */
2500 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2502 /* called when force_empty is called */
2503 mutex_lock(&percpu_charge_mutex);
2504 drain_all_stock(root_memcg, true);
2505 mutex_unlock(&percpu_charge_mutex);
2509 * This function drains percpu counter value from DEAD cpu and
2510 * move it to local cpu. Note that this function can be preempted.
2512 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2516 spin_lock(&memcg->pcp_counter_lock);
2517 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2518 long x = per_cpu(memcg->stat->count[i], cpu);
2520 per_cpu(memcg->stat->count[i], cpu) = 0;
2521 memcg->nocpu_base.count[i] += x;
2523 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2524 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2526 per_cpu(memcg->stat->events[i], cpu) = 0;
2527 memcg->nocpu_base.events[i] += x;
2529 spin_unlock(&memcg->pcp_counter_lock);
2532 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2533 unsigned long action,
2536 int cpu = (unsigned long)hcpu;
2537 struct memcg_stock_pcp *stock;
2538 struct mem_cgroup *iter;
2540 if (action == CPU_ONLINE)
2543 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2546 for_each_mem_cgroup(iter)
2547 mem_cgroup_drain_pcp_counter(iter, cpu);
2549 stock = &per_cpu(memcg_stock, cpu);
2555 * mem_cgroup_try_charge - try charging a memcg
2556 * @memcg: memcg to charge
2557 * @nr_pages: number of pages to charge
2558 * @oom: trigger OOM if reclaim fails
2560 * Returns 0 if @memcg was charged successfully, -EINTR if the charge
2561 * was bypassed to root_mem_cgroup, and -ENOMEM if the charge failed.
2563 static int mem_cgroup_try_charge(struct mem_cgroup *memcg,
2565 unsigned int nr_pages,
2568 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2569 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2570 struct mem_cgroup *mem_over_limit;
2571 struct res_counter *fail_res;
2572 unsigned long nr_reclaimed;
2573 unsigned long flags = 0;
2574 unsigned long long size;
2576 if (mem_cgroup_is_root(memcg))
2579 if (consume_stock(memcg, nr_pages))
2582 size = batch * PAGE_SIZE;
2583 if (!res_counter_charge(&memcg->res, size, &fail_res)) {
2584 if (!do_swap_account)
2586 if (!res_counter_charge(&memcg->memsw, size, &fail_res))
2588 res_counter_uncharge(&memcg->res, size);
2589 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2590 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2592 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2594 if (batch > nr_pages) {
2600 * Unlike in global OOM situations, memcg is not in a physical
2601 * memory shortage. Allow dying and OOM-killed tasks to
2602 * bypass the last charges so that they can exit quickly and
2603 * free their memory.
2605 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2606 fatal_signal_pending(current) ||
2607 current->flags & PF_EXITING))
2610 if (unlikely(task_in_memcg_oom(current)))
2613 if (!(gfp_mask & __GFP_WAIT))
2616 nr_reclaimed = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2618 if (mem_cgroup_margin(mem_over_limit) >= batch)
2621 if (gfp_mask & __GFP_NORETRY)
2624 * Even though the limit is exceeded at this point, reclaim
2625 * may have been able to free some pages. Retry the charge
2626 * before killing the task.
2628 * Only for regular pages, though: huge pages are rather
2629 * unlikely to succeed so close to the limit, and we fall back
2630 * to regular pages anyway in case of failure.
2632 if (nr_reclaimed && batch <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2635 * At task move, charge accounts can be doubly counted. So, it's
2636 * better to wait until the end of task_move if something is going on.
2638 if (mem_cgroup_wait_acct_move(mem_over_limit))
2644 if (gfp_mask & __GFP_NOFAIL)
2647 if (fatal_signal_pending(current))
2653 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(batch));
2655 if (!(gfp_mask & __GFP_NOFAIL))
2661 if (batch > nr_pages)
2662 refill_stock(memcg, batch - nr_pages);
2668 * mem_cgroup_try_charge_mm - try charging a mm
2669 * @mm: mm_struct to charge
2670 * @nr_pages: number of pages to charge
2671 * @oom: trigger OOM if reclaim fails
2673 * Returns the charged mem_cgroup associated with the given mm_struct or
2674 * NULL the charge failed.
2676 static struct mem_cgroup *mem_cgroup_try_charge_mm(struct mm_struct *mm,
2678 unsigned int nr_pages,
2682 struct mem_cgroup *memcg;
2685 memcg = get_mem_cgroup_from_mm(mm);
2686 ret = mem_cgroup_try_charge(memcg, gfp_mask, nr_pages, oom);
2687 css_put(&memcg->css);
2689 memcg = root_mem_cgroup;
2697 * Somemtimes we have to undo a charge we got by try_charge().
2698 * This function is for that and do uncharge, put css's refcnt.
2699 * gotten by try_charge().
2701 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2702 unsigned int nr_pages)
2704 if (!mem_cgroup_is_root(memcg)) {
2705 unsigned long bytes = nr_pages * PAGE_SIZE;
2707 res_counter_uncharge(&memcg->res, bytes);
2708 if (do_swap_account)
2709 res_counter_uncharge(&memcg->memsw, bytes);
2714 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2715 * This is useful when moving usage to parent cgroup.
2717 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2718 unsigned int nr_pages)
2720 unsigned long bytes = nr_pages * PAGE_SIZE;
2722 if (mem_cgroup_is_root(memcg))
2725 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2726 if (do_swap_account)
2727 res_counter_uncharge_until(&memcg->memsw,
2728 memcg->memsw.parent, bytes);
2732 * A helper function to get mem_cgroup from ID. must be called under
2733 * rcu_read_lock(). The caller is responsible for calling
2734 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2735 * refcnt from swap can be called against removed memcg.)
2737 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2739 /* ID 0 is unused ID */
2742 return mem_cgroup_from_id(id);
2745 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2747 struct mem_cgroup *memcg = NULL;
2748 struct page_cgroup *pc;
2752 VM_BUG_ON_PAGE(!PageLocked(page), page);
2754 pc = lookup_page_cgroup(page);
2755 lock_page_cgroup(pc);
2756 if (PageCgroupUsed(pc)) {
2757 memcg = pc->mem_cgroup;
2758 if (memcg && !css_tryget_online(&memcg->css))
2760 } else if (PageSwapCache(page)) {
2761 ent.val = page_private(page);
2762 id = lookup_swap_cgroup_id(ent);
2764 memcg = mem_cgroup_lookup(id);
2765 if (memcg && !css_tryget_online(&memcg->css))
2769 unlock_page_cgroup(pc);
2773 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2775 unsigned int nr_pages,
2776 enum charge_type ctype,
2779 struct page_cgroup *pc = lookup_page_cgroup(page);
2780 struct zone *uninitialized_var(zone);
2781 struct lruvec *lruvec;
2782 bool was_on_lru = false;
2785 lock_page_cgroup(pc);
2786 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2788 * we don't need page_cgroup_lock about tail pages, becase they are not
2789 * accessed by any other context at this point.
2793 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2794 * may already be on some other mem_cgroup's LRU. Take care of it.
2797 zone = page_zone(page);
2798 spin_lock_irq(&zone->lru_lock);
2799 if (PageLRU(page)) {
2800 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2802 del_page_from_lru_list(page, lruvec, page_lru(page));
2807 pc->mem_cgroup = memcg;
2809 * We access a page_cgroup asynchronously without lock_page_cgroup().
2810 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2811 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2812 * before USED bit, we need memory barrier here.
2813 * See mem_cgroup_add_lru_list(), etc.
2816 SetPageCgroupUsed(pc);
2820 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2821 VM_BUG_ON_PAGE(PageLRU(page), page);
2823 add_page_to_lru_list(page, lruvec, page_lru(page));
2825 spin_unlock_irq(&zone->lru_lock);
2828 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2833 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2834 unlock_page_cgroup(pc);
2837 * "charge_statistics" updated event counter. Then, check it.
2838 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2839 * if they exceeds softlimit.
2841 memcg_check_events(memcg, page);
2844 static DEFINE_MUTEX(set_limit_mutex);
2846 #ifdef CONFIG_MEMCG_KMEM
2848 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2849 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2851 static DEFINE_MUTEX(memcg_slab_mutex);
2853 static DEFINE_MUTEX(activate_kmem_mutex);
2855 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2857 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2858 memcg_kmem_is_active(memcg);
2862 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2863 * in the memcg_cache_params struct.
2865 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2867 struct kmem_cache *cachep;
2869 VM_BUG_ON(p->is_root_cache);
2870 cachep = p->root_cache;
2871 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2874 #ifdef CONFIG_SLABINFO
2875 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2877 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2878 struct memcg_cache_params *params;
2880 if (!memcg_can_account_kmem(memcg))
2883 print_slabinfo_header(m);
2885 mutex_lock(&memcg_slab_mutex);
2886 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2887 cache_show(memcg_params_to_cache(params), m);
2888 mutex_unlock(&memcg_slab_mutex);
2894 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2896 struct res_counter *fail_res;
2899 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2903 ret = mem_cgroup_try_charge(memcg, gfp, size >> PAGE_SHIFT,
2904 oom_gfp_allowed(gfp));
2905 if (ret == -EINTR) {
2907 * mem_cgroup_try_charge() chosed to bypass to root due to
2908 * OOM kill or fatal signal. Since our only options are to
2909 * either fail the allocation or charge it to this cgroup, do
2910 * it as a temporary condition. But we can't fail. From a
2911 * kmem/slab perspective, the cache has already been selected,
2912 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2915 * This condition will only trigger if the task entered
2916 * memcg_charge_kmem in a sane state, but was OOM-killed during
2917 * mem_cgroup_try_charge() above. Tasks that were already
2918 * dying when the allocation triggers should have been already
2919 * directed to the root cgroup in memcontrol.h
2921 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2922 if (do_swap_account)
2923 res_counter_charge_nofail(&memcg->memsw, size,
2927 res_counter_uncharge(&memcg->kmem, size);
2932 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2934 res_counter_uncharge(&memcg->res, size);
2935 if (do_swap_account)
2936 res_counter_uncharge(&memcg->memsw, size);
2939 if (res_counter_uncharge(&memcg->kmem, size))
2943 * Releases a reference taken in kmem_cgroup_css_offline in case
2944 * this last uncharge is racing with the offlining code or it is
2945 * outliving the memcg existence.
2947 * The memory barrier imposed by test&clear is paired with the
2948 * explicit one in memcg_kmem_mark_dead().
2950 if (memcg_kmem_test_and_clear_dead(memcg))
2951 css_put(&memcg->css);
2955 * helper for acessing a memcg's index. It will be used as an index in the
2956 * child cache array in kmem_cache, and also to derive its name. This function
2957 * will return -1 when this is not a kmem-limited memcg.
2959 int memcg_cache_id(struct mem_cgroup *memcg)
2961 return memcg ? memcg->kmemcg_id : -1;
2964 static size_t memcg_caches_array_size(int num_groups)
2967 if (num_groups <= 0)
2970 size = 2 * num_groups;
2971 if (size < MEMCG_CACHES_MIN_SIZE)
2972 size = MEMCG_CACHES_MIN_SIZE;
2973 else if (size > MEMCG_CACHES_MAX_SIZE)
2974 size = MEMCG_CACHES_MAX_SIZE;
2980 * We should update the current array size iff all caches updates succeed. This
2981 * can only be done from the slab side. The slab mutex needs to be held when
2984 void memcg_update_array_size(int num)
2986 if (num > memcg_limited_groups_array_size)
2987 memcg_limited_groups_array_size = memcg_caches_array_size(num);
2990 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
2992 struct memcg_cache_params *cur_params = s->memcg_params;
2994 VM_BUG_ON(!is_root_cache(s));
2996 if (num_groups > memcg_limited_groups_array_size) {
2998 struct memcg_cache_params *new_params;
2999 ssize_t size = memcg_caches_array_size(num_groups);
3001 size *= sizeof(void *);
3002 size += offsetof(struct memcg_cache_params, memcg_caches);
3004 new_params = kzalloc(size, GFP_KERNEL);
3008 new_params->is_root_cache = true;
3011 * There is the chance it will be bigger than
3012 * memcg_limited_groups_array_size, if we failed an allocation
3013 * in a cache, in which case all caches updated before it, will
3014 * have a bigger array.
3016 * But if that is the case, the data after
3017 * memcg_limited_groups_array_size is certainly unused
3019 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3020 if (!cur_params->memcg_caches[i])
3022 new_params->memcg_caches[i] =
3023 cur_params->memcg_caches[i];
3027 * Ideally, we would wait until all caches succeed, and only
3028 * then free the old one. But this is not worth the extra
3029 * pointer per-cache we'd have to have for this.
3031 * It is not a big deal if some caches are left with a size
3032 * bigger than the others. And all updates will reset this
3035 rcu_assign_pointer(s->memcg_params, new_params);
3037 kfree_rcu(cur_params, rcu_head);
3042 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3043 struct kmem_cache *root_cache)
3047 if (!memcg_kmem_enabled())
3051 size = offsetof(struct memcg_cache_params, memcg_caches);
3052 size += memcg_limited_groups_array_size * sizeof(void *);
3054 size = sizeof(struct memcg_cache_params);
3056 s->memcg_params = kzalloc(size, GFP_KERNEL);
3057 if (!s->memcg_params)
3061 s->memcg_params->memcg = memcg;
3062 s->memcg_params->root_cache = root_cache;
3063 css_get(&memcg->css);
3065 s->memcg_params->is_root_cache = true;
3070 void memcg_free_cache_params(struct kmem_cache *s)
3072 if (!s->memcg_params)
3074 if (!s->memcg_params->is_root_cache)
3075 css_put(&s->memcg_params->memcg->css);
3076 kfree(s->memcg_params);
3079 static void memcg_register_cache(struct mem_cgroup *memcg,
3080 struct kmem_cache *root_cache)
3082 static char memcg_name_buf[NAME_MAX + 1]; /* protected by
3084 struct kmem_cache *cachep;
3087 lockdep_assert_held(&memcg_slab_mutex);
3089 id = memcg_cache_id(memcg);
3092 * Since per-memcg caches are created asynchronously on first
3093 * allocation (see memcg_kmem_get_cache()), several threads can try to
3094 * create the same cache, but only one of them may succeed.
3096 if (cache_from_memcg_idx(root_cache, id))
3099 cgroup_name(memcg->css.cgroup, memcg_name_buf, NAME_MAX + 1);
3100 cachep = memcg_create_kmem_cache(memcg, root_cache, memcg_name_buf);
3102 * If we could not create a memcg cache, do not complain, because
3103 * that's not critical at all as we can always proceed with the root
3109 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3112 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3113 * barrier here to ensure nobody will see the kmem_cache partially
3118 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
3119 root_cache->memcg_params->memcg_caches[id] = cachep;
3122 static void memcg_unregister_cache(struct kmem_cache *cachep)
3124 struct kmem_cache *root_cache;
3125 struct mem_cgroup *memcg;
3128 lockdep_assert_held(&memcg_slab_mutex);
3130 BUG_ON(is_root_cache(cachep));
3132 root_cache = cachep->memcg_params->root_cache;
3133 memcg = cachep->memcg_params->memcg;
3134 id = memcg_cache_id(memcg);
3136 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
3137 root_cache->memcg_params->memcg_caches[id] = NULL;
3139 list_del(&cachep->memcg_params->list);
3141 kmem_cache_destroy(cachep);
3145 * During the creation a new cache, we need to disable our accounting mechanism
3146 * altogether. This is true even if we are not creating, but rather just
3147 * enqueing new caches to be created.
3149 * This is because that process will trigger allocations; some visible, like
3150 * explicit kmallocs to auxiliary data structures, name strings and internal
3151 * cache structures; some well concealed, like INIT_WORK() that can allocate
3152 * objects during debug.
3154 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3155 * to it. This may not be a bounded recursion: since the first cache creation
3156 * failed to complete (waiting on the allocation), we'll just try to create the
3157 * cache again, failing at the same point.
3159 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3160 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3161 * inside the following two functions.
3163 static inline void memcg_stop_kmem_account(void)
3165 VM_BUG_ON(!current->mm);
3166 current->memcg_kmem_skip_account++;
3169 static inline void memcg_resume_kmem_account(void)
3171 VM_BUG_ON(!current->mm);
3172 current->memcg_kmem_skip_account--;
3175 int __memcg_cleanup_cache_params(struct kmem_cache *s)
3177 struct kmem_cache *c;
3180 mutex_lock(&memcg_slab_mutex);
3181 for_each_memcg_cache_index(i) {
3182 c = cache_from_memcg_idx(s, i);
3186 memcg_unregister_cache(c);
3188 if (cache_from_memcg_idx(s, i))
3191 mutex_unlock(&memcg_slab_mutex);
3195 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3197 struct kmem_cache *cachep;
3198 struct memcg_cache_params *params, *tmp;
3200 if (!memcg_kmem_is_active(memcg))
3203 mutex_lock(&memcg_slab_mutex);
3204 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
3205 cachep = memcg_params_to_cache(params);
3206 kmem_cache_shrink(cachep);
3207 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3208 memcg_unregister_cache(cachep);
3210 mutex_unlock(&memcg_slab_mutex);
3213 struct memcg_register_cache_work {
3214 struct mem_cgroup *memcg;
3215 struct kmem_cache *cachep;
3216 struct work_struct work;
3219 static void memcg_register_cache_func(struct work_struct *w)
3221 struct memcg_register_cache_work *cw =
3222 container_of(w, struct memcg_register_cache_work, work);
3223 struct mem_cgroup *memcg = cw->memcg;
3224 struct kmem_cache *cachep = cw->cachep;
3226 mutex_lock(&memcg_slab_mutex);
3227 memcg_register_cache(memcg, cachep);
3228 mutex_unlock(&memcg_slab_mutex);
3230 css_put(&memcg->css);
3235 * Enqueue the creation of a per-memcg kmem_cache.
3237 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
3238 struct kmem_cache *cachep)
3240 struct memcg_register_cache_work *cw;
3242 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
3244 css_put(&memcg->css);
3249 cw->cachep = cachep;
3251 INIT_WORK(&cw->work, memcg_register_cache_func);
3252 schedule_work(&cw->work);
3255 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
3256 struct kmem_cache *cachep)
3259 * We need to stop accounting when we kmalloc, because if the
3260 * corresponding kmalloc cache is not yet created, the first allocation
3261 * in __memcg_schedule_register_cache will recurse.
3263 * However, it is better to enclose the whole function. Depending on
3264 * the debugging options enabled, INIT_WORK(), for instance, can
3265 * trigger an allocation. This too, will make us recurse. Because at
3266 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3267 * the safest choice is to do it like this, wrapping the whole function.
3269 memcg_stop_kmem_account();
3270 __memcg_schedule_register_cache(memcg, cachep);
3271 memcg_resume_kmem_account();
3274 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
3278 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp,
3279 PAGE_SIZE << order);
3281 atomic_add(1 << order, &cachep->memcg_params->nr_pages);
3285 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
3287 memcg_uncharge_kmem(cachep->memcg_params->memcg, PAGE_SIZE << order);
3288 atomic_sub(1 << order, &cachep->memcg_params->nr_pages);
3292 * Return the kmem_cache we're supposed to use for a slab allocation.
3293 * We try to use the current memcg's version of the cache.
3295 * If the cache does not exist yet, if we are the first user of it,
3296 * we either create it immediately, if possible, or create it asynchronously
3298 * In the latter case, we will let the current allocation go through with
3299 * the original cache.
3301 * Can't be called in interrupt context or from kernel threads.
3302 * This function needs to be called with rcu_read_lock() held.
3304 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3307 struct mem_cgroup *memcg;
3308 struct kmem_cache *memcg_cachep;
3310 VM_BUG_ON(!cachep->memcg_params);
3311 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3313 if (!current->mm || current->memcg_kmem_skip_account)
3317 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3319 if (!memcg_can_account_kmem(memcg))
3322 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3323 if (likely(memcg_cachep)) {
3324 cachep = memcg_cachep;
3328 /* The corresponding put will be done in the workqueue. */
3329 if (!css_tryget_online(&memcg->css))
3334 * If we are in a safe context (can wait, and not in interrupt
3335 * context), we could be be predictable and return right away.
3336 * This would guarantee that the allocation being performed
3337 * already belongs in the new cache.
3339 * However, there are some clashes that can arrive from locking.
3340 * For instance, because we acquire the slab_mutex while doing
3341 * memcg_create_kmem_cache, this means no further allocation
3342 * could happen with the slab_mutex held. So it's better to
3345 memcg_schedule_register_cache(memcg, cachep);
3353 * We need to verify if the allocation against current->mm->owner's memcg is
3354 * possible for the given order. But the page is not allocated yet, so we'll
3355 * need a further commit step to do the final arrangements.
3357 * It is possible for the task to switch cgroups in this mean time, so at
3358 * commit time, we can't rely on task conversion any longer. We'll then use
3359 * the handle argument to return to the caller which cgroup we should commit
3360 * against. We could also return the memcg directly and avoid the pointer
3361 * passing, but a boolean return value gives better semantics considering
3362 * the compiled-out case as well.
3364 * Returning true means the allocation is possible.
3367 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3369 struct mem_cgroup *memcg;
3375 * Disabling accounting is only relevant for some specific memcg
3376 * internal allocations. Therefore we would initially not have such
3377 * check here, since direct calls to the page allocator that are
3378 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3379 * outside memcg core. We are mostly concerned with cache allocations,
3380 * and by having this test at memcg_kmem_get_cache, we are already able
3381 * to relay the allocation to the root cache and bypass the memcg cache
3384 * There is one exception, though: the SLUB allocator does not create
3385 * large order caches, but rather service large kmallocs directly from
3386 * the page allocator. Therefore, the following sequence when backed by
3387 * the SLUB allocator:
3389 * memcg_stop_kmem_account();
3390 * kmalloc(<large_number>)
3391 * memcg_resume_kmem_account();
3393 * would effectively ignore the fact that we should skip accounting,
3394 * since it will drive us directly to this function without passing
3395 * through the cache selector memcg_kmem_get_cache. Such large
3396 * allocations are extremely rare but can happen, for instance, for the
3397 * cache arrays. We bring this test here.
3399 if (!current->mm || current->memcg_kmem_skip_account)
3402 memcg = get_mem_cgroup_from_mm(current->mm);
3404 if (!memcg_can_account_kmem(memcg)) {
3405 css_put(&memcg->css);
3409 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3413 css_put(&memcg->css);
3417 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3420 struct page_cgroup *pc;
3422 VM_BUG_ON(mem_cgroup_is_root(memcg));
3424 /* The page allocation failed. Revert */
3426 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3430 pc = lookup_page_cgroup(page);
3431 lock_page_cgroup(pc);
3432 pc->mem_cgroup = memcg;
3433 SetPageCgroupUsed(pc);
3434 unlock_page_cgroup(pc);
3437 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3439 struct mem_cgroup *memcg = NULL;
3440 struct page_cgroup *pc;
3443 pc = lookup_page_cgroup(page);
3445 * Fast unlocked return. Theoretically might have changed, have to
3446 * check again after locking.
3448 if (!PageCgroupUsed(pc))
3451 lock_page_cgroup(pc);
3452 if (PageCgroupUsed(pc)) {
3453 memcg = pc->mem_cgroup;
3454 ClearPageCgroupUsed(pc);
3456 unlock_page_cgroup(pc);
3459 * We trust that only if there is a memcg associated with the page, it
3460 * is a valid allocation
3465 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3466 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3469 static inline void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3472 #endif /* CONFIG_MEMCG_KMEM */
3474 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3476 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3478 * Because tail pages are not marked as "used", set it. We're under
3479 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3480 * charge/uncharge will be never happen and move_account() is done under
3481 * compound_lock(), so we don't have to take care of races.
3483 void mem_cgroup_split_huge_fixup(struct page *head)
3485 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3486 struct page_cgroup *pc;
3487 struct mem_cgroup *memcg;
3490 if (mem_cgroup_disabled())
3493 memcg = head_pc->mem_cgroup;
3494 for (i = 1; i < HPAGE_PMD_NR; i++) {
3496 pc->mem_cgroup = memcg;
3497 smp_wmb();/* see __commit_charge() */
3498 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3500 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3503 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3506 * mem_cgroup_move_account - move account of the page
3508 * @nr_pages: number of regular pages (>1 for huge pages)
3509 * @pc: page_cgroup of the page.
3510 * @from: mem_cgroup which the page is moved from.
3511 * @to: mem_cgroup which the page is moved to. @from != @to.
3513 * The caller must confirm following.
3514 * - page is not on LRU (isolate_page() is useful.)
3515 * - compound_lock is held when nr_pages > 1
3517 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3520 static int mem_cgroup_move_account(struct page *page,
3521 unsigned int nr_pages,
3522 struct page_cgroup *pc,
3523 struct mem_cgroup *from,
3524 struct mem_cgroup *to)
3526 unsigned long flags;
3528 bool anon = PageAnon(page);
3530 VM_BUG_ON(from == to);
3531 VM_BUG_ON_PAGE(PageLRU(page), page);
3533 * The page is isolated from LRU. So, collapse function
3534 * will not handle this page. But page splitting can happen.
3535 * Do this check under compound_page_lock(). The caller should
3539 if (nr_pages > 1 && !PageTransHuge(page))
3542 lock_page_cgroup(pc);
3545 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3548 move_lock_mem_cgroup(from, &flags);
3550 if (!anon && page_mapped(page)) {
3551 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3553 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3557 if (PageWriteback(page)) {
3558 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3560 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3564 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3566 /* caller should have done css_get */
3567 pc->mem_cgroup = to;
3568 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3569 move_unlock_mem_cgroup(from, &flags);
3572 unlock_page_cgroup(pc);
3576 memcg_check_events(to, page);
3577 memcg_check_events(from, page);
3583 * mem_cgroup_move_parent - moves page to the parent group
3584 * @page: the page to move
3585 * @pc: page_cgroup of the page
3586 * @child: page's cgroup
3588 * move charges to its parent or the root cgroup if the group has no
3589 * parent (aka use_hierarchy==0).
3590 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3591 * mem_cgroup_move_account fails) the failure is always temporary and
3592 * it signals a race with a page removal/uncharge or migration. In the
3593 * first case the page is on the way out and it will vanish from the LRU
3594 * on the next attempt and the call should be retried later.
3595 * Isolation from the LRU fails only if page has been isolated from
3596 * the LRU since we looked at it and that usually means either global
3597 * reclaim or migration going on. The page will either get back to the
3599 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3600 * (!PageCgroupUsed) or moved to a different group. The page will
3601 * disappear in the next attempt.
3603 static int mem_cgroup_move_parent(struct page *page,
3604 struct page_cgroup *pc,
3605 struct mem_cgroup *child)
3607 struct mem_cgroup *parent;
3608 unsigned int nr_pages;
3609 unsigned long uninitialized_var(flags);
3612 VM_BUG_ON(mem_cgroup_is_root(child));
3615 if (!get_page_unless_zero(page))
3617 if (isolate_lru_page(page))
3620 nr_pages = hpage_nr_pages(page);
3622 parent = parent_mem_cgroup(child);
3624 * If no parent, move charges to root cgroup.
3627 parent = root_mem_cgroup;
3630 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3631 flags = compound_lock_irqsave(page);
3634 ret = mem_cgroup_move_account(page, nr_pages,
3637 __mem_cgroup_cancel_local_charge(child, nr_pages);
3640 compound_unlock_irqrestore(page, flags);
3641 putback_lru_page(page);
3648 int mem_cgroup_charge_anon(struct page *page,
3649 struct mm_struct *mm, gfp_t gfp_mask)
3651 unsigned int nr_pages = 1;
3652 struct mem_cgroup *memcg;
3655 if (mem_cgroup_disabled())
3658 VM_BUG_ON_PAGE(page_mapped(page), page);
3659 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3662 if (PageTransHuge(page)) {
3663 nr_pages <<= compound_order(page);
3664 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3666 * Never OOM-kill a process for a huge page. The
3667 * fault handler will fall back to regular pages.
3672 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, nr_pages, oom);
3675 __mem_cgroup_commit_charge(memcg, page, nr_pages,
3676 MEM_CGROUP_CHARGE_TYPE_ANON, false);
3681 * While swap-in, try_charge -> commit or cancel, the page is locked.
3682 * And when try_charge() successfully returns, one refcnt to memcg without
3683 * struct page_cgroup is acquired. This refcnt will be consumed by
3684 * "commit()" or removed by "cancel()"
3686 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3689 struct mem_cgroup **memcgp)
3691 struct mem_cgroup *memcg = NULL;
3692 struct page_cgroup *pc;
3695 pc = lookup_page_cgroup(page);
3697 * Every swap fault against a single page tries to charge the
3698 * page, bail as early as possible. shmem_unuse() encounters
3699 * already charged pages, too. The USED bit is protected by
3700 * the page lock, which serializes swap cache removal, which
3701 * in turn serializes uncharging.
3703 if (PageCgroupUsed(pc))
3705 if (do_swap_account)
3706 memcg = try_get_mem_cgroup_from_page(page);
3708 memcg = get_mem_cgroup_from_mm(mm);
3709 ret = mem_cgroup_try_charge(memcg, mask, 1, true);
3710 css_put(&memcg->css);
3712 memcg = root_mem_cgroup;
3720 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3721 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3723 if (mem_cgroup_disabled()) {
3728 * A racing thread's fault, or swapoff, may have already
3729 * updated the pte, and even removed page from swap cache: in
3730 * those cases unuse_pte()'s pte_same() test will fail; but
3731 * there's also a KSM case which does need to charge the page.
3733 if (!PageSwapCache(page)) {
3734 struct mem_cgroup *memcg;
3736 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3742 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3745 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3747 if (mem_cgroup_disabled())
3751 __mem_cgroup_cancel_charge(memcg, 1);
3755 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3756 enum charge_type ctype)
3758 if (mem_cgroup_disabled())
3763 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3765 * Now swap is on-memory. This means this page may be
3766 * counted both as mem and swap....double count.
3767 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3768 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3769 * may call delete_from_swap_cache() before reach here.
3771 if (do_swap_account && PageSwapCache(page)) {
3772 swp_entry_t ent = {.val = page_private(page)};
3773 mem_cgroup_uncharge_swap(ent);
3777 void mem_cgroup_commit_charge_swapin(struct page *page,
3778 struct mem_cgroup *memcg)
3780 __mem_cgroup_commit_charge_swapin(page, memcg,
3781 MEM_CGROUP_CHARGE_TYPE_ANON);
3784 int mem_cgroup_charge_file(struct page *page, struct mm_struct *mm,
3787 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3788 struct mem_cgroup *memcg;
3791 if (mem_cgroup_disabled())
3793 if (PageCompound(page))
3796 if (PageSwapCache(page)) { /* shmem */
3797 ret = __mem_cgroup_try_charge_swapin(mm, page,
3801 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3805 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3808 __mem_cgroup_commit_charge(memcg, page, 1, type, false);
3812 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3813 unsigned int nr_pages,
3814 const enum charge_type ctype)
3816 struct memcg_batch_info *batch = NULL;
3817 bool uncharge_memsw = true;
3819 /* If swapout, usage of swap doesn't decrease */
3820 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3821 uncharge_memsw = false;
3823 batch = ¤t->memcg_batch;
3825 * In usual, we do css_get() when we remember memcg pointer.
3826 * But in this case, we keep res->usage until end of a series of
3827 * uncharges. Then, it's ok to ignore memcg's refcnt.
3830 batch->memcg = memcg;
3832 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3833 * In those cases, all pages freed continuously can be expected to be in
3834 * the same cgroup and we have chance to coalesce uncharges.
3835 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3836 * because we want to do uncharge as soon as possible.
3839 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3840 goto direct_uncharge;
3843 goto direct_uncharge;
3846 * In typical case, batch->memcg == mem. This means we can
3847 * merge a series of uncharges to an uncharge of res_counter.
3848 * If not, we uncharge res_counter ony by one.
3850 if (batch->memcg != memcg)
3851 goto direct_uncharge;
3852 /* remember freed charge and uncharge it later */
3855 batch->memsw_nr_pages++;
3858 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3860 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3861 if (unlikely(batch->memcg != memcg))
3862 memcg_oom_recover(memcg);
3866 * uncharge if !page_mapped(page)
3868 static struct mem_cgroup *
3869 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3872 struct mem_cgroup *memcg = NULL;
3873 unsigned int nr_pages = 1;
3874 struct page_cgroup *pc;
3877 if (mem_cgroup_disabled())
3880 if (PageTransHuge(page)) {
3881 nr_pages <<= compound_order(page);
3882 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3885 * Check if our page_cgroup is valid
3887 pc = lookup_page_cgroup(page);
3888 if (unlikely(!PageCgroupUsed(pc)))
3891 lock_page_cgroup(pc);
3893 memcg = pc->mem_cgroup;
3895 if (!PageCgroupUsed(pc))
3898 anon = PageAnon(page);
3901 case MEM_CGROUP_CHARGE_TYPE_ANON:
3903 * Generally PageAnon tells if it's the anon statistics to be
3904 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3905 * used before page reached the stage of being marked PageAnon.
3909 case MEM_CGROUP_CHARGE_TYPE_DROP:
3910 /* See mem_cgroup_prepare_migration() */
3911 if (page_mapped(page))
3914 * Pages under migration may not be uncharged. But
3915 * end_migration() /must/ be the one uncharging the
3916 * unused post-migration page and so it has to call
3917 * here with the migration bit still set. See the
3918 * res_counter handling below.
3920 if (!end_migration && PageCgroupMigration(pc))
3923 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3924 if (!PageAnon(page)) { /* Shared memory */
3925 if (page->mapping && !page_is_file_cache(page))
3927 } else if (page_mapped(page)) /* Anon */
3934 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
3936 ClearPageCgroupUsed(pc);
3938 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3939 * freed from LRU. This is safe because uncharged page is expected not
3940 * to be reused (freed soon). Exception is SwapCache, it's handled by
3941 * special functions.
3944 unlock_page_cgroup(pc);
3946 * even after unlock, we have memcg->res.usage here and this memcg
3947 * will never be freed, so it's safe to call css_get().
3949 memcg_check_events(memcg, page);
3950 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3951 mem_cgroup_swap_statistics(memcg, true);
3952 css_get(&memcg->css);
3955 * Migration does not charge the res_counter for the
3956 * replacement page, so leave it alone when phasing out the
3957 * page that is unused after the migration.
3959 if (!end_migration && !mem_cgroup_is_root(memcg))
3960 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3965 unlock_page_cgroup(pc);
3969 void mem_cgroup_uncharge_page(struct page *page)
3972 if (page_mapped(page))
3974 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3976 * If the page is in swap cache, uncharge should be deferred
3977 * to the swap path, which also properly accounts swap usage
3978 * and handles memcg lifetime.
3980 * Note that this check is not stable and reclaim may add the
3981 * page to swap cache at any time after this. However, if the
3982 * page is not in swap cache by the time page->mapcount hits
3983 * 0, there won't be any page table references to the swap
3984 * slot, and reclaim will free it and not actually write the
3987 if (PageSwapCache(page))
3989 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
3992 void mem_cgroup_uncharge_cache_page(struct page *page)
3994 VM_BUG_ON_PAGE(page_mapped(page), page);
3995 VM_BUG_ON_PAGE(page->mapping, page);
3996 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4000 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4001 * In that cases, pages are freed continuously and we can expect pages
4002 * are in the same memcg. All these calls itself limits the number of
4003 * pages freed at once, then uncharge_start/end() is called properly.
4004 * This may be called prural(2) times in a context,
4007 void mem_cgroup_uncharge_start(void)
4009 current->memcg_batch.do_batch++;
4010 /* We can do nest. */
4011 if (current->memcg_batch.do_batch == 1) {
4012 current->memcg_batch.memcg = NULL;
4013 current->memcg_batch.nr_pages = 0;
4014 current->memcg_batch.memsw_nr_pages = 0;
4018 void mem_cgroup_uncharge_end(void)
4020 struct memcg_batch_info *batch = ¤t->memcg_batch;
4022 if (!batch->do_batch)
4026 if (batch->do_batch) /* If stacked, do nothing. */
4032 * This "batch->memcg" is valid without any css_get/put etc...
4033 * bacause we hide charges behind us.
4035 if (batch->nr_pages)
4036 res_counter_uncharge(&batch->memcg->res,
4037 batch->nr_pages * PAGE_SIZE);
4038 if (batch->memsw_nr_pages)
4039 res_counter_uncharge(&batch->memcg->memsw,
4040 batch->memsw_nr_pages * PAGE_SIZE);
4041 memcg_oom_recover(batch->memcg);
4042 /* forget this pointer (for sanity check) */
4043 batch->memcg = NULL;
4048 * called after __delete_from_swap_cache() and drop "page" account.
4049 * memcg information is recorded to swap_cgroup of "ent"
4052 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4054 struct mem_cgroup *memcg;
4055 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4057 if (!swapout) /* this was a swap cache but the swap is unused ! */
4058 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4060 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4063 * record memcg information, if swapout && memcg != NULL,
4064 * css_get() was called in uncharge().
4066 if (do_swap_account && swapout && memcg)
4067 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4071 #ifdef CONFIG_MEMCG_SWAP
4073 * called from swap_entry_free(). remove record in swap_cgroup and
4074 * uncharge "memsw" account.
4076 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4078 struct mem_cgroup *memcg;
4081 if (!do_swap_account)
4084 id = swap_cgroup_record(ent, 0);
4086 memcg = mem_cgroup_lookup(id);
4089 * We uncharge this because swap is freed. This memcg can
4090 * be obsolete one. We avoid calling css_tryget_online().
4092 if (!mem_cgroup_is_root(memcg))
4093 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4094 mem_cgroup_swap_statistics(memcg, false);
4095 css_put(&memcg->css);
4101 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4102 * @entry: swap entry to be moved
4103 * @from: mem_cgroup which the entry is moved from
4104 * @to: mem_cgroup which the entry is moved to
4106 * It succeeds only when the swap_cgroup's record for this entry is the same
4107 * as the mem_cgroup's id of @from.
4109 * Returns 0 on success, -EINVAL on failure.
4111 * The caller must have charged to @to, IOW, called res_counter_charge() about
4112 * both res and memsw, and called css_get().
4114 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4115 struct mem_cgroup *from, struct mem_cgroup *to)
4117 unsigned short old_id, new_id;
4119 old_id = mem_cgroup_id(from);
4120 new_id = mem_cgroup_id(to);
4122 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4123 mem_cgroup_swap_statistics(from, false);
4124 mem_cgroup_swap_statistics(to, true);
4126 * This function is only called from task migration context now.
4127 * It postpones res_counter and refcount handling till the end
4128 * of task migration(mem_cgroup_clear_mc()) for performance
4129 * improvement. But we cannot postpone css_get(to) because if
4130 * the process that has been moved to @to does swap-in, the
4131 * refcount of @to might be decreased to 0.
4133 * We are in attach() phase, so the cgroup is guaranteed to be
4134 * alive, so we can just call css_get().
4142 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4143 struct mem_cgroup *from, struct mem_cgroup *to)
4150 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4153 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4154 struct mem_cgroup **memcgp)
4156 struct mem_cgroup *memcg = NULL;
4157 unsigned int nr_pages = 1;
4158 struct page_cgroup *pc;
4159 enum charge_type ctype;
4163 if (mem_cgroup_disabled())
4166 if (PageTransHuge(page))
4167 nr_pages <<= compound_order(page);
4169 pc = lookup_page_cgroup(page);
4170 lock_page_cgroup(pc);
4171 if (PageCgroupUsed(pc)) {
4172 memcg = pc->mem_cgroup;
4173 css_get(&memcg->css);
4175 * At migrating an anonymous page, its mapcount goes down
4176 * to 0 and uncharge() will be called. But, even if it's fully
4177 * unmapped, migration may fail and this page has to be
4178 * charged again. We set MIGRATION flag here and delay uncharge
4179 * until end_migration() is called
4181 * Corner Case Thinking
4183 * When the old page was mapped as Anon and it's unmap-and-freed
4184 * while migration was ongoing.
4185 * If unmap finds the old page, uncharge() of it will be delayed
4186 * until end_migration(). If unmap finds a new page, it's
4187 * uncharged when it make mapcount to be 1->0. If unmap code
4188 * finds swap_migration_entry, the new page will not be mapped
4189 * and end_migration() will find it(mapcount==0).
4192 * When the old page was mapped but migraion fails, the kernel
4193 * remaps it. A charge for it is kept by MIGRATION flag even
4194 * if mapcount goes down to 0. We can do remap successfully
4195 * without charging it again.
4198 * The "old" page is under lock_page() until the end of
4199 * migration, so, the old page itself will not be swapped-out.
4200 * If the new page is swapped out before end_migraton, our
4201 * hook to usual swap-out path will catch the event.
4204 SetPageCgroupMigration(pc);
4206 unlock_page_cgroup(pc);
4208 * If the page is not charged at this point,
4216 * We charge new page before it's used/mapped. So, even if unlock_page()
4217 * is called before end_migration, we can catch all events on this new
4218 * page. In the case new page is migrated but not remapped, new page's
4219 * mapcount will be finally 0 and we call uncharge in end_migration().
4222 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4224 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4226 * The page is committed to the memcg, but it's not actually
4227 * charged to the res_counter since we plan on replacing the
4228 * old one and only one page is going to be left afterwards.
4230 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4233 /* remove redundant charge if migration failed*/
4234 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4235 struct page *oldpage, struct page *newpage, bool migration_ok)
4237 struct page *used, *unused;
4238 struct page_cgroup *pc;
4244 if (!migration_ok) {
4251 anon = PageAnon(used);
4252 __mem_cgroup_uncharge_common(unused,
4253 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4254 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4256 css_put(&memcg->css);
4258 * We disallowed uncharge of pages under migration because mapcount
4259 * of the page goes down to zero, temporarly.
4260 * Clear the flag and check the page should be charged.
4262 pc = lookup_page_cgroup(oldpage);
4263 lock_page_cgroup(pc);
4264 ClearPageCgroupMigration(pc);
4265 unlock_page_cgroup(pc);
4268 * If a page is a file cache, radix-tree replacement is very atomic
4269 * and we can skip this check. When it was an Anon page, its mapcount
4270 * goes down to 0. But because we added MIGRATION flage, it's not
4271 * uncharged yet. There are several case but page->mapcount check
4272 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4273 * check. (see prepare_charge() also)
4276 mem_cgroup_uncharge_page(used);
4280 * At replace page cache, newpage is not under any memcg but it's on
4281 * LRU. So, this function doesn't touch res_counter but handles LRU
4282 * in correct way. Both pages are locked so we cannot race with uncharge.
4284 void mem_cgroup_replace_page_cache(struct page *oldpage,
4285 struct page *newpage)
4287 struct mem_cgroup *memcg = NULL;
4288 struct page_cgroup *pc;
4289 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4291 if (mem_cgroup_disabled())
4294 pc = lookup_page_cgroup(oldpage);
4295 /* fix accounting on old pages */
4296 lock_page_cgroup(pc);
4297 if (PageCgroupUsed(pc)) {
4298 memcg = pc->mem_cgroup;
4299 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4300 ClearPageCgroupUsed(pc);
4302 unlock_page_cgroup(pc);
4305 * When called from shmem_replace_page(), in some cases the
4306 * oldpage has already been charged, and in some cases not.
4311 * Even if newpage->mapping was NULL before starting replacement,
4312 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4313 * LRU while we overwrite pc->mem_cgroup.
4315 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4318 #ifdef CONFIG_DEBUG_VM
4319 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4321 struct page_cgroup *pc;
4323 pc = lookup_page_cgroup(page);
4325 * Can be NULL while feeding pages into the page allocator for
4326 * the first time, i.e. during boot or memory hotplug;
4327 * or when mem_cgroup_disabled().
4329 if (likely(pc) && PageCgroupUsed(pc))
4334 bool mem_cgroup_bad_page_check(struct page *page)
4336 if (mem_cgroup_disabled())
4339 return lookup_page_cgroup_used(page) != NULL;
4342 void mem_cgroup_print_bad_page(struct page *page)
4344 struct page_cgroup *pc;
4346 pc = lookup_page_cgroup_used(page);
4348 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4349 pc, pc->flags, pc->mem_cgroup);
4354 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4355 unsigned long long val)
4358 u64 memswlimit, memlimit;
4360 int children = mem_cgroup_count_children(memcg);
4361 u64 curusage, oldusage;
4365 * For keeping hierarchical_reclaim simple, how long we should retry
4366 * is depends on callers. We set our retry-count to be function
4367 * of # of children which we should visit in this loop.
4369 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4371 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4374 while (retry_count) {
4375 if (signal_pending(current)) {
4380 * Rather than hide all in some function, I do this in
4381 * open coded manner. You see what this really does.
4382 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4384 mutex_lock(&set_limit_mutex);
4385 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4386 if (memswlimit < val) {
4388 mutex_unlock(&set_limit_mutex);
4392 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4396 ret = res_counter_set_limit(&memcg->res, val);
4398 if (memswlimit == val)
4399 memcg->memsw_is_minimum = true;
4401 memcg->memsw_is_minimum = false;
4403 mutex_unlock(&set_limit_mutex);
4408 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4409 MEM_CGROUP_RECLAIM_SHRINK);
4410 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4411 /* Usage is reduced ? */
4412 if (curusage >= oldusage)
4415 oldusage = curusage;
4417 if (!ret && enlarge)
4418 memcg_oom_recover(memcg);
4423 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4424 unsigned long long val)
4427 u64 memlimit, memswlimit, oldusage, curusage;
4428 int children = mem_cgroup_count_children(memcg);
4432 /* see mem_cgroup_resize_res_limit */
4433 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4434 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4435 while (retry_count) {
4436 if (signal_pending(current)) {
4441 * Rather than hide all in some function, I do this in
4442 * open coded manner. You see what this really does.
4443 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4445 mutex_lock(&set_limit_mutex);
4446 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4447 if (memlimit > val) {
4449 mutex_unlock(&set_limit_mutex);
4452 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4453 if (memswlimit < val)
4455 ret = res_counter_set_limit(&memcg->memsw, val);
4457 if (memlimit == val)
4458 memcg->memsw_is_minimum = true;
4460 memcg->memsw_is_minimum = false;
4462 mutex_unlock(&set_limit_mutex);
4467 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4468 MEM_CGROUP_RECLAIM_NOSWAP |
4469 MEM_CGROUP_RECLAIM_SHRINK);
4470 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4471 /* Usage is reduced ? */
4472 if (curusage >= oldusage)
4475 oldusage = curusage;
4477 if (!ret && enlarge)
4478 memcg_oom_recover(memcg);
4482 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4484 unsigned long *total_scanned)
4486 unsigned long nr_reclaimed = 0;
4487 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4488 unsigned long reclaimed;
4490 struct mem_cgroup_tree_per_zone *mctz;
4491 unsigned long long excess;
4492 unsigned long nr_scanned;
4497 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4499 * This loop can run a while, specially if mem_cgroup's continuously
4500 * keep exceeding their soft limit and putting the system under
4507 mz = mem_cgroup_largest_soft_limit_node(mctz);
4512 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4513 gfp_mask, &nr_scanned);
4514 nr_reclaimed += reclaimed;
4515 *total_scanned += nr_scanned;
4516 spin_lock(&mctz->lock);
4519 * If we failed to reclaim anything from this memory cgroup
4520 * it is time to move on to the next cgroup
4526 * Loop until we find yet another one.
4528 * By the time we get the soft_limit lock
4529 * again, someone might have aded the
4530 * group back on the RB tree. Iterate to
4531 * make sure we get a different mem.
4532 * mem_cgroup_largest_soft_limit_node returns
4533 * NULL if no other cgroup is present on
4537 __mem_cgroup_largest_soft_limit_node(mctz);
4539 css_put(&next_mz->memcg->css);
4540 else /* next_mz == NULL or other memcg */
4544 __mem_cgroup_remove_exceeded(mz, mctz);
4545 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4547 * One school of thought says that we should not add
4548 * back the node to the tree if reclaim returns 0.
4549 * But our reclaim could return 0, simply because due
4550 * to priority we are exposing a smaller subset of
4551 * memory to reclaim from. Consider this as a longer
4554 /* If excess == 0, no tree ops */
4555 __mem_cgroup_insert_exceeded(mz, mctz, excess);
4556 spin_unlock(&mctz->lock);
4557 css_put(&mz->memcg->css);
4560 * Could not reclaim anything and there are no more
4561 * mem cgroups to try or we seem to be looping without
4562 * reclaiming anything.
4564 if (!nr_reclaimed &&
4566 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4568 } while (!nr_reclaimed);
4570 css_put(&next_mz->memcg->css);
4571 return nr_reclaimed;
4575 * mem_cgroup_force_empty_list - clears LRU of a group
4576 * @memcg: group to clear
4579 * @lru: lru to to clear
4581 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4582 * reclaim the pages page themselves - pages are moved to the parent (or root)
4585 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4586 int node, int zid, enum lru_list lru)
4588 struct lruvec *lruvec;
4589 unsigned long flags;
4590 struct list_head *list;
4594 zone = &NODE_DATA(node)->node_zones[zid];
4595 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4596 list = &lruvec->lists[lru];
4600 struct page_cgroup *pc;
4603 spin_lock_irqsave(&zone->lru_lock, flags);
4604 if (list_empty(list)) {
4605 spin_unlock_irqrestore(&zone->lru_lock, flags);
4608 page = list_entry(list->prev, struct page, lru);
4610 list_move(&page->lru, list);
4612 spin_unlock_irqrestore(&zone->lru_lock, flags);
4615 spin_unlock_irqrestore(&zone->lru_lock, flags);
4617 pc = lookup_page_cgroup(page);
4619 if (mem_cgroup_move_parent(page, pc, memcg)) {
4620 /* found lock contention or "pc" is obsolete. */
4625 } while (!list_empty(list));
4629 * make mem_cgroup's charge to be 0 if there is no task by moving
4630 * all the charges and pages to the parent.
4631 * This enables deleting this mem_cgroup.
4633 * Caller is responsible for holding css reference on the memcg.
4635 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4641 /* This is for making all *used* pages to be on LRU. */
4642 lru_add_drain_all();
4643 drain_all_stock_sync(memcg);
4644 mem_cgroup_start_move(memcg);
4645 for_each_node_state(node, N_MEMORY) {
4646 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4649 mem_cgroup_force_empty_list(memcg,
4654 mem_cgroup_end_move(memcg);
4655 memcg_oom_recover(memcg);
4659 * Kernel memory may not necessarily be trackable to a specific
4660 * process. So they are not migrated, and therefore we can't
4661 * expect their value to drop to 0 here.
4662 * Having res filled up with kmem only is enough.
4664 * This is a safety check because mem_cgroup_force_empty_list
4665 * could have raced with mem_cgroup_replace_page_cache callers
4666 * so the lru seemed empty but the page could have been added
4667 * right after the check. RES_USAGE should be safe as we always
4668 * charge before adding to the LRU.
4670 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4671 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4672 } while (usage > 0);
4676 * Test whether @memcg has children, dead or alive. Note that this
4677 * function doesn't care whether @memcg has use_hierarchy enabled and
4678 * returns %true if there are child csses according to the cgroup
4679 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
4681 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4686 * The lock does not prevent addition or deletion of children, but
4687 * it prevents a new child from being initialized based on this
4688 * parent in css_online(), so it's enough to decide whether
4689 * hierarchically inherited attributes can still be changed or not.
4691 lockdep_assert_held(&memcg_create_mutex);
4694 ret = css_next_child(NULL, &memcg->css);
4700 * Reclaims as many pages from the given memcg as possible and moves
4701 * the rest to the parent.
4703 * Caller is responsible for holding css reference for memcg.
4705 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4707 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4709 /* we call try-to-free pages for make this cgroup empty */
4710 lru_add_drain_all();
4711 /* try to free all pages in this cgroup */
4712 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4715 if (signal_pending(current))
4718 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4722 /* maybe some writeback is necessary */
4723 congestion_wait(BLK_RW_ASYNC, HZ/10);
4731 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
4732 char *buf, size_t nbytes,
4735 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4737 if (mem_cgroup_is_root(memcg))
4739 return mem_cgroup_force_empty(memcg) ?: nbytes;
4742 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4745 return mem_cgroup_from_css(css)->use_hierarchy;
4748 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4749 struct cftype *cft, u64 val)
4752 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4753 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
4755 mutex_lock(&memcg_create_mutex);
4757 if (memcg->use_hierarchy == val)
4761 * If parent's use_hierarchy is set, we can't make any modifications
4762 * in the child subtrees. If it is unset, then the change can
4763 * occur, provided the current cgroup has no children.
4765 * For the root cgroup, parent_mem is NULL, we allow value to be
4766 * set if there are no children.
4768 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4769 (val == 1 || val == 0)) {
4770 if (!memcg_has_children(memcg))
4771 memcg->use_hierarchy = val;
4778 mutex_unlock(&memcg_create_mutex);
4784 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4785 enum mem_cgroup_stat_index idx)
4787 struct mem_cgroup *iter;
4790 /* Per-cpu values can be negative, use a signed accumulator */
4791 for_each_mem_cgroup_tree(iter, memcg)
4792 val += mem_cgroup_read_stat(iter, idx);
4794 if (val < 0) /* race ? */
4799 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4803 if (!mem_cgroup_is_root(memcg)) {
4805 return res_counter_read_u64(&memcg->res, RES_USAGE);
4807 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4811 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4812 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4814 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4815 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4818 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4820 return val << PAGE_SHIFT;
4823 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
4826 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4831 type = MEMFILE_TYPE(cft->private);
4832 name = MEMFILE_ATTR(cft->private);
4836 if (name == RES_USAGE)
4837 val = mem_cgroup_usage(memcg, false);
4839 val = res_counter_read_u64(&memcg->res, name);
4842 if (name == RES_USAGE)
4843 val = mem_cgroup_usage(memcg, true);
4845 val = res_counter_read_u64(&memcg->memsw, name);
4848 val = res_counter_read_u64(&memcg->kmem, name);
4857 #ifdef CONFIG_MEMCG_KMEM
4858 /* should be called with activate_kmem_mutex held */
4859 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
4860 unsigned long long limit)
4865 if (memcg_kmem_is_active(memcg))
4869 * We are going to allocate memory for data shared by all memory
4870 * cgroups so let's stop accounting here.
4872 memcg_stop_kmem_account();
4875 * For simplicity, we won't allow this to be disabled. It also can't
4876 * be changed if the cgroup has children already, or if tasks had
4879 * If tasks join before we set the limit, a person looking at
4880 * kmem.usage_in_bytes will have no way to determine when it took
4881 * place, which makes the value quite meaningless.
4883 * After it first became limited, changes in the value of the limit are
4884 * of course permitted.
4886 mutex_lock(&memcg_create_mutex);
4887 if (cgroup_has_tasks(memcg->css.cgroup) ||
4888 (memcg->use_hierarchy && memcg_has_children(memcg)))
4890 mutex_unlock(&memcg_create_mutex);
4894 memcg_id = ida_simple_get(&kmem_limited_groups,
4895 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
4902 * Make sure we have enough space for this cgroup in each root cache's
4905 mutex_lock(&memcg_slab_mutex);
4906 err = memcg_update_all_caches(memcg_id + 1);
4907 mutex_unlock(&memcg_slab_mutex);
4911 memcg->kmemcg_id = memcg_id;
4912 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
4915 * We couldn't have accounted to this cgroup, because it hasn't got the
4916 * active bit set yet, so this should succeed.
4918 err = res_counter_set_limit(&memcg->kmem, limit);
4921 static_key_slow_inc(&memcg_kmem_enabled_key);
4923 * Setting the active bit after enabling static branching will
4924 * guarantee no one starts accounting before all call sites are
4927 memcg_kmem_set_active(memcg);
4929 memcg_resume_kmem_account();
4933 ida_simple_remove(&kmem_limited_groups, memcg_id);
4937 static int memcg_activate_kmem(struct mem_cgroup *memcg,
4938 unsigned long long limit)
4942 mutex_lock(&activate_kmem_mutex);
4943 ret = __memcg_activate_kmem(memcg, limit);
4944 mutex_unlock(&activate_kmem_mutex);
4948 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4949 unsigned long long val)
4953 if (!memcg_kmem_is_active(memcg))
4954 ret = memcg_activate_kmem(memcg, val);
4956 ret = res_counter_set_limit(&memcg->kmem, val);
4960 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4963 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4968 mutex_lock(&activate_kmem_mutex);
4970 * If the parent cgroup is not kmem-active now, it cannot be activated
4971 * after this point, because it has at least one child already.
4973 if (memcg_kmem_is_active(parent))
4974 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
4975 mutex_unlock(&activate_kmem_mutex);
4979 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4980 unsigned long long val)
4984 #endif /* CONFIG_MEMCG_KMEM */
4987 * The user of this function is...
4990 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
4991 char *buf, size_t nbytes, loff_t off)
4993 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4996 unsigned long long val;
4999 buf = strstrip(buf);
5000 type = MEMFILE_TYPE(of_cft(of)->private);
5001 name = MEMFILE_ATTR(of_cft(of)->private);
5005 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5009 /* This function does all necessary parse...reuse it */
5010 ret = res_counter_memparse_write_strategy(buf, &val);
5014 ret = mem_cgroup_resize_limit(memcg, val);
5015 else if (type == _MEMSWAP)
5016 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5017 else if (type == _KMEM)
5018 ret = memcg_update_kmem_limit(memcg, val);
5022 case RES_SOFT_LIMIT:
5023 ret = res_counter_memparse_write_strategy(buf, &val);
5027 * For memsw, soft limits are hard to implement in terms
5028 * of semantics, for now, we support soft limits for
5029 * control without swap
5032 ret = res_counter_set_soft_limit(&memcg->res, val);
5037 ret = -EINVAL; /* should be BUG() ? */
5040 return ret ?: nbytes;
5043 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5044 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5046 unsigned long long min_limit, min_memsw_limit, tmp;
5048 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5049 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5050 if (!memcg->use_hierarchy)
5053 while (memcg->css.parent) {
5054 memcg = mem_cgroup_from_css(memcg->css.parent);
5055 if (!memcg->use_hierarchy)
5057 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5058 min_limit = min(min_limit, tmp);
5059 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5060 min_memsw_limit = min(min_memsw_limit, tmp);
5063 *mem_limit = min_limit;
5064 *memsw_limit = min_memsw_limit;
5067 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
5068 size_t nbytes, loff_t off)
5070 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5074 type = MEMFILE_TYPE(of_cft(of)->private);
5075 name = MEMFILE_ATTR(of_cft(of)->private);
5080 res_counter_reset_max(&memcg->res);
5081 else if (type == _MEMSWAP)
5082 res_counter_reset_max(&memcg->memsw);
5083 else if (type == _KMEM)
5084 res_counter_reset_max(&memcg->kmem);
5090 res_counter_reset_failcnt(&memcg->res);
5091 else if (type == _MEMSWAP)
5092 res_counter_reset_failcnt(&memcg->memsw);
5093 else if (type == _KMEM)
5094 res_counter_reset_failcnt(&memcg->kmem);
5103 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5106 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5110 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5111 struct cftype *cft, u64 val)
5113 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5115 if (val >= (1 << NR_MOVE_TYPE))
5119 * No kind of locking is needed in here, because ->can_attach() will
5120 * check this value once in the beginning of the process, and then carry
5121 * on with stale data. This means that changes to this value will only
5122 * affect task migrations starting after the change.
5124 memcg->move_charge_at_immigrate = val;
5128 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5129 struct cftype *cft, u64 val)
5136 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5140 unsigned int lru_mask;
5143 static const struct numa_stat stats[] = {
5144 { "total", LRU_ALL },
5145 { "file", LRU_ALL_FILE },
5146 { "anon", LRU_ALL_ANON },
5147 { "unevictable", BIT(LRU_UNEVICTABLE) },
5149 const struct numa_stat *stat;
5152 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5154 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5155 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5156 seq_printf(m, "%s=%lu", stat->name, nr);
5157 for_each_node_state(nid, N_MEMORY) {
5158 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5160 seq_printf(m, " N%d=%lu", nid, nr);
5165 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5166 struct mem_cgroup *iter;
5169 for_each_mem_cgroup_tree(iter, memcg)
5170 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5171 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5172 for_each_node_state(nid, N_MEMORY) {
5174 for_each_mem_cgroup_tree(iter, memcg)
5175 nr += mem_cgroup_node_nr_lru_pages(
5176 iter, nid, stat->lru_mask);
5177 seq_printf(m, " N%d=%lu", nid, nr);
5184 #endif /* CONFIG_NUMA */
5186 static inline void mem_cgroup_lru_names_not_uptodate(void)
5188 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5191 static int memcg_stat_show(struct seq_file *m, void *v)
5193 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5194 struct mem_cgroup *mi;
5197 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5198 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5200 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5201 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5204 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5205 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5206 mem_cgroup_read_events(memcg, i));
5208 for (i = 0; i < NR_LRU_LISTS; i++)
5209 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5210 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5212 /* Hierarchical information */
5214 unsigned long long limit, memsw_limit;
5215 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5216 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5217 if (do_swap_account)
5218 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5222 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5225 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5227 for_each_mem_cgroup_tree(mi, memcg)
5228 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5229 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5232 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5233 unsigned long long val = 0;
5235 for_each_mem_cgroup_tree(mi, memcg)
5236 val += mem_cgroup_read_events(mi, i);
5237 seq_printf(m, "total_%s %llu\n",
5238 mem_cgroup_events_names[i], val);
5241 for (i = 0; i < NR_LRU_LISTS; i++) {
5242 unsigned long long val = 0;
5244 for_each_mem_cgroup_tree(mi, memcg)
5245 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5246 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5249 #ifdef CONFIG_DEBUG_VM
5252 struct mem_cgroup_per_zone *mz;
5253 struct zone_reclaim_stat *rstat;
5254 unsigned long recent_rotated[2] = {0, 0};
5255 unsigned long recent_scanned[2] = {0, 0};
5257 for_each_online_node(nid)
5258 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5259 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
5260 rstat = &mz->lruvec.reclaim_stat;
5262 recent_rotated[0] += rstat->recent_rotated[0];
5263 recent_rotated[1] += rstat->recent_rotated[1];
5264 recent_scanned[0] += rstat->recent_scanned[0];
5265 recent_scanned[1] += rstat->recent_scanned[1];
5267 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5268 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5269 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5270 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5277 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5280 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5282 return mem_cgroup_swappiness(memcg);
5285 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5286 struct cftype *cft, u64 val)
5288 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5294 memcg->swappiness = val;
5296 vm_swappiness = val;
5301 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5303 struct mem_cgroup_threshold_ary *t;
5309 t = rcu_dereference(memcg->thresholds.primary);
5311 t = rcu_dereference(memcg->memsw_thresholds.primary);
5316 usage = mem_cgroup_usage(memcg, swap);
5319 * current_threshold points to threshold just below or equal to usage.
5320 * If it's not true, a threshold was crossed after last
5321 * call of __mem_cgroup_threshold().
5323 i = t->current_threshold;
5326 * Iterate backward over array of thresholds starting from
5327 * current_threshold and check if a threshold is crossed.
5328 * If none of thresholds below usage is crossed, we read
5329 * only one element of the array here.
5331 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5332 eventfd_signal(t->entries[i].eventfd, 1);
5334 /* i = current_threshold + 1 */
5338 * Iterate forward over array of thresholds starting from
5339 * current_threshold+1 and check if a threshold is crossed.
5340 * If none of thresholds above usage is crossed, we read
5341 * only one element of the array here.
5343 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5344 eventfd_signal(t->entries[i].eventfd, 1);
5346 /* Update current_threshold */
5347 t->current_threshold = i - 1;
5352 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5355 __mem_cgroup_threshold(memcg, false);
5356 if (do_swap_account)
5357 __mem_cgroup_threshold(memcg, true);
5359 memcg = parent_mem_cgroup(memcg);
5363 static int compare_thresholds(const void *a, const void *b)
5365 const struct mem_cgroup_threshold *_a = a;
5366 const struct mem_cgroup_threshold *_b = b;
5368 if (_a->threshold > _b->threshold)
5371 if (_a->threshold < _b->threshold)
5377 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5379 struct mem_cgroup_eventfd_list *ev;
5381 spin_lock(&memcg_oom_lock);
5383 list_for_each_entry(ev, &memcg->oom_notify, list)
5384 eventfd_signal(ev->eventfd, 1);
5386 spin_unlock(&memcg_oom_lock);
5390 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5392 struct mem_cgroup *iter;
5394 for_each_mem_cgroup_tree(iter, memcg)
5395 mem_cgroup_oom_notify_cb(iter);
5398 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5399 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5401 struct mem_cgroup_thresholds *thresholds;
5402 struct mem_cgroup_threshold_ary *new;
5403 u64 threshold, usage;
5406 ret = res_counter_memparse_write_strategy(args, &threshold);
5410 mutex_lock(&memcg->thresholds_lock);
5413 thresholds = &memcg->thresholds;
5414 else if (type == _MEMSWAP)
5415 thresholds = &memcg->memsw_thresholds;
5419 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5421 /* Check if a threshold crossed before adding a new one */
5422 if (thresholds->primary)
5423 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5425 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5427 /* Allocate memory for new array of thresholds */
5428 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5436 /* Copy thresholds (if any) to new array */
5437 if (thresholds->primary) {
5438 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5439 sizeof(struct mem_cgroup_threshold));
5442 /* Add new threshold */
5443 new->entries[size - 1].eventfd = eventfd;
5444 new->entries[size - 1].threshold = threshold;
5446 /* Sort thresholds. Registering of new threshold isn't time-critical */
5447 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5448 compare_thresholds, NULL);
5450 /* Find current threshold */
5451 new->current_threshold = -1;
5452 for (i = 0; i < size; i++) {
5453 if (new->entries[i].threshold <= usage) {
5455 * new->current_threshold will not be used until
5456 * rcu_assign_pointer(), so it's safe to increment
5459 ++new->current_threshold;
5464 /* Free old spare buffer and save old primary buffer as spare */
5465 kfree(thresholds->spare);
5466 thresholds->spare = thresholds->primary;
5468 rcu_assign_pointer(thresholds->primary, new);
5470 /* To be sure that nobody uses thresholds */
5474 mutex_unlock(&memcg->thresholds_lock);
5479 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5480 struct eventfd_ctx *eventfd, const char *args)
5482 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5485 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5486 struct eventfd_ctx *eventfd, const char *args)
5488 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5491 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5492 struct eventfd_ctx *eventfd, enum res_type type)
5494 struct mem_cgroup_thresholds *thresholds;
5495 struct mem_cgroup_threshold_ary *new;
5499 mutex_lock(&memcg->thresholds_lock);
5501 thresholds = &memcg->thresholds;
5502 else if (type == _MEMSWAP)
5503 thresholds = &memcg->memsw_thresholds;
5507 if (!thresholds->primary)
5510 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5512 /* Check if a threshold crossed before removing */
5513 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5515 /* Calculate new number of threshold */
5517 for (i = 0; i < thresholds->primary->size; i++) {
5518 if (thresholds->primary->entries[i].eventfd != eventfd)
5522 new = thresholds->spare;
5524 /* Set thresholds array to NULL if we don't have thresholds */
5533 /* Copy thresholds and find current threshold */
5534 new->current_threshold = -1;
5535 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5536 if (thresholds->primary->entries[i].eventfd == eventfd)
5539 new->entries[j] = thresholds->primary->entries[i];
5540 if (new->entries[j].threshold <= usage) {
5542 * new->current_threshold will not be used
5543 * until rcu_assign_pointer(), so it's safe to increment
5546 ++new->current_threshold;
5552 /* Swap primary and spare array */
5553 thresholds->spare = thresholds->primary;
5554 /* If all events are unregistered, free the spare array */
5556 kfree(thresholds->spare);
5557 thresholds->spare = NULL;
5560 rcu_assign_pointer(thresholds->primary, new);
5562 /* To be sure that nobody uses thresholds */
5565 mutex_unlock(&memcg->thresholds_lock);
5568 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5569 struct eventfd_ctx *eventfd)
5571 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5574 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5575 struct eventfd_ctx *eventfd)
5577 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5580 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5581 struct eventfd_ctx *eventfd, const char *args)
5583 struct mem_cgroup_eventfd_list *event;
5585 event = kmalloc(sizeof(*event), GFP_KERNEL);
5589 spin_lock(&memcg_oom_lock);
5591 event->eventfd = eventfd;
5592 list_add(&event->list, &memcg->oom_notify);
5594 /* already in OOM ? */
5595 if (atomic_read(&memcg->under_oom))
5596 eventfd_signal(eventfd, 1);
5597 spin_unlock(&memcg_oom_lock);
5602 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5603 struct eventfd_ctx *eventfd)
5605 struct mem_cgroup_eventfd_list *ev, *tmp;
5607 spin_lock(&memcg_oom_lock);
5609 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5610 if (ev->eventfd == eventfd) {
5611 list_del(&ev->list);
5616 spin_unlock(&memcg_oom_lock);
5619 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5621 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5623 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5624 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5628 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5629 struct cftype *cft, u64 val)
5631 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5633 /* cannot set to root cgroup and only 0 and 1 are allowed */
5634 if (!css->parent || !((val == 0) || (val == 1)))
5637 memcg->oom_kill_disable = val;
5639 memcg_oom_recover(memcg);
5644 #ifdef CONFIG_MEMCG_KMEM
5645 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5649 memcg->kmemcg_id = -1;
5650 ret = memcg_propagate_kmem(memcg);
5654 return mem_cgroup_sockets_init(memcg, ss);
5657 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5659 mem_cgroup_sockets_destroy(memcg);
5662 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5664 if (!memcg_kmem_is_active(memcg))
5668 * kmem charges can outlive the cgroup. In the case of slab
5669 * pages, for instance, a page contain objects from various
5670 * processes. As we prevent from taking a reference for every
5671 * such allocation we have to be careful when doing uncharge
5672 * (see memcg_uncharge_kmem) and here during offlining.
5674 * The idea is that that only the _last_ uncharge which sees
5675 * the dead memcg will drop the last reference. An additional
5676 * reference is taken here before the group is marked dead
5677 * which is then paired with css_put during uncharge resp. here.
5679 * Although this might sound strange as this path is called from
5680 * css_offline() when the referencemight have dropped down to 0 and
5681 * shouldn't be incremented anymore (css_tryget_online() would
5682 * fail) we do not have other options because of the kmem
5683 * allocations lifetime.
5685 css_get(&memcg->css);
5687 memcg_kmem_mark_dead(memcg);
5689 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5692 if (memcg_kmem_test_and_clear_dead(memcg))
5693 css_put(&memcg->css);
5696 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5701 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5705 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5711 * DO NOT USE IN NEW FILES.
5713 * "cgroup.event_control" implementation.
5715 * This is way over-engineered. It tries to support fully configurable
5716 * events for each user. Such level of flexibility is completely
5717 * unnecessary especially in the light of the planned unified hierarchy.
5719 * Please deprecate this and replace with something simpler if at all
5724 * Unregister event and free resources.
5726 * Gets called from workqueue.
5728 static void memcg_event_remove(struct work_struct *work)
5730 struct mem_cgroup_event *event =
5731 container_of(work, struct mem_cgroup_event, remove);
5732 struct mem_cgroup *memcg = event->memcg;
5734 remove_wait_queue(event->wqh, &event->wait);
5736 event->unregister_event(memcg, event->eventfd);
5738 /* Notify userspace the event is going away. */
5739 eventfd_signal(event->eventfd, 1);
5741 eventfd_ctx_put(event->eventfd);
5743 css_put(&memcg->css);
5747 * Gets called on POLLHUP on eventfd when user closes it.
5749 * Called with wqh->lock held and interrupts disabled.
5751 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
5752 int sync, void *key)
5754 struct mem_cgroup_event *event =
5755 container_of(wait, struct mem_cgroup_event, wait);
5756 struct mem_cgroup *memcg = event->memcg;
5757 unsigned long flags = (unsigned long)key;
5759 if (flags & POLLHUP) {
5761 * If the event has been detached at cgroup removal, we
5762 * can simply return knowing the other side will cleanup
5765 * We can't race against event freeing since the other
5766 * side will require wqh->lock via remove_wait_queue(),
5769 spin_lock(&memcg->event_list_lock);
5770 if (!list_empty(&event->list)) {
5771 list_del_init(&event->list);
5773 * We are in atomic context, but cgroup_event_remove()
5774 * may sleep, so we have to call it in workqueue.
5776 schedule_work(&event->remove);
5778 spin_unlock(&memcg->event_list_lock);
5784 static void memcg_event_ptable_queue_proc(struct file *file,
5785 wait_queue_head_t *wqh, poll_table *pt)
5787 struct mem_cgroup_event *event =
5788 container_of(pt, struct mem_cgroup_event, pt);
5791 add_wait_queue(wqh, &event->wait);
5795 * DO NOT USE IN NEW FILES.
5797 * Parse input and register new cgroup event handler.
5799 * Input must be in format '<event_fd> <control_fd> <args>'.
5800 * Interpretation of args is defined by control file implementation.
5802 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
5803 char *buf, size_t nbytes, loff_t off)
5805 struct cgroup_subsys_state *css = of_css(of);
5806 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5807 struct mem_cgroup_event *event;
5808 struct cgroup_subsys_state *cfile_css;
5809 unsigned int efd, cfd;
5816 buf = strstrip(buf);
5818 efd = simple_strtoul(buf, &endp, 10);
5823 cfd = simple_strtoul(buf, &endp, 10);
5824 if ((*endp != ' ') && (*endp != '\0'))
5828 event = kzalloc(sizeof(*event), GFP_KERNEL);
5832 event->memcg = memcg;
5833 INIT_LIST_HEAD(&event->list);
5834 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
5835 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
5836 INIT_WORK(&event->remove, memcg_event_remove);
5844 event->eventfd = eventfd_ctx_fileget(efile.file);
5845 if (IS_ERR(event->eventfd)) {
5846 ret = PTR_ERR(event->eventfd);
5853 goto out_put_eventfd;
5856 /* the process need read permission on control file */
5857 /* AV: shouldn't we check that it's been opened for read instead? */
5858 ret = inode_permission(file_inode(cfile.file), MAY_READ);
5863 * Determine the event callbacks and set them in @event. This used
5864 * to be done via struct cftype but cgroup core no longer knows
5865 * about these events. The following is crude but the whole thing
5866 * is for compatibility anyway.
5868 * DO NOT ADD NEW FILES.
5870 name = cfile.file->f_dentry->d_name.name;
5872 if (!strcmp(name, "memory.usage_in_bytes")) {
5873 event->register_event = mem_cgroup_usage_register_event;
5874 event->unregister_event = mem_cgroup_usage_unregister_event;
5875 } else if (!strcmp(name, "memory.oom_control")) {
5876 event->register_event = mem_cgroup_oom_register_event;
5877 event->unregister_event = mem_cgroup_oom_unregister_event;
5878 } else if (!strcmp(name, "memory.pressure_level")) {
5879 event->register_event = vmpressure_register_event;
5880 event->unregister_event = vmpressure_unregister_event;
5881 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
5882 event->register_event = memsw_cgroup_usage_register_event;
5883 event->unregister_event = memsw_cgroup_usage_unregister_event;
5890 * Verify @cfile should belong to @css. Also, remaining events are
5891 * automatically removed on cgroup destruction but the removal is
5892 * asynchronous, so take an extra ref on @css.
5894 cfile_css = css_tryget_online_from_dir(cfile.file->f_dentry->d_parent,
5895 &memory_cgrp_subsys);
5897 if (IS_ERR(cfile_css))
5899 if (cfile_css != css) {
5904 ret = event->register_event(memcg, event->eventfd, buf);
5908 efile.file->f_op->poll(efile.file, &event->pt);
5910 spin_lock(&memcg->event_list_lock);
5911 list_add(&event->list, &memcg->event_list);
5912 spin_unlock(&memcg->event_list_lock);
5924 eventfd_ctx_put(event->eventfd);
5933 static struct cftype mem_cgroup_files[] = {
5935 .name = "usage_in_bytes",
5936 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5937 .read_u64 = mem_cgroup_read_u64,
5940 .name = "max_usage_in_bytes",
5941 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5942 .write = mem_cgroup_reset,
5943 .read_u64 = mem_cgroup_read_u64,
5946 .name = "limit_in_bytes",
5947 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5948 .write = mem_cgroup_write,
5949 .read_u64 = mem_cgroup_read_u64,
5952 .name = "soft_limit_in_bytes",
5953 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5954 .write = mem_cgroup_write,
5955 .read_u64 = mem_cgroup_read_u64,
5959 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5960 .write = mem_cgroup_reset,
5961 .read_u64 = mem_cgroup_read_u64,
5965 .seq_show = memcg_stat_show,
5968 .name = "force_empty",
5969 .write = mem_cgroup_force_empty_write,
5972 .name = "use_hierarchy",
5973 .write_u64 = mem_cgroup_hierarchy_write,
5974 .read_u64 = mem_cgroup_hierarchy_read,
5977 .name = "cgroup.event_control", /* XXX: for compat */
5978 .write = memcg_write_event_control,
5979 .flags = CFTYPE_NO_PREFIX,
5983 .name = "swappiness",
5984 .read_u64 = mem_cgroup_swappiness_read,
5985 .write_u64 = mem_cgroup_swappiness_write,
5988 .name = "move_charge_at_immigrate",
5989 .read_u64 = mem_cgroup_move_charge_read,
5990 .write_u64 = mem_cgroup_move_charge_write,
5993 .name = "oom_control",
5994 .seq_show = mem_cgroup_oom_control_read,
5995 .write_u64 = mem_cgroup_oom_control_write,
5996 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5999 .name = "pressure_level",
6003 .name = "numa_stat",
6004 .seq_show = memcg_numa_stat_show,
6007 #ifdef CONFIG_MEMCG_KMEM
6009 .name = "kmem.limit_in_bytes",
6010 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6011 .write = mem_cgroup_write,
6012 .read_u64 = mem_cgroup_read_u64,
6015 .name = "kmem.usage_in_bytes",
6016 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6017 .read_u64 = mem_cgroup_read_u64,
6020 .name = "kmem.failcnt",
6021 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6022 .write = mem_cgroup_reset,
6023 .read_u64 = mem_cgroup_read_u64,
6026 .name = "kmem.max_usage_in_bytes",
6027 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6028 .write = mem_cgroup_reset,
6029 .read_u64 = mem_cgroup_read_u64,
6031 #ifdef CONFIG_SLABINFO
6033 .name = "kmem.slabinfo",
6034 .seq_show = mem_cgroup_slabinfo_read,
6038 { }, /* terminate */
6041 #ifdef CONFIG_MEMCG_SWAP
6042 static struct cftype memsw_cgroup_files[] = {
6044 .name = "memsw.usage_in_bytes",
6045 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6046 .read_u64 = mem_cgroup_read_u64,
6049 .name = "memsw.max_usage_in_bytes",
6050 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6051 .write = mem_cgroup_reset,
6052 .read_u64 = mem_cgroup_read_u64,
6055 .name = "memsw.limit_in_bytes",
6056 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6057 .write = mem_cgroup_write,
6058 .read_u64 = mem_cgroup_read_u64,
6061 .name = "memsw.failcnt",
6062 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6063 .write = mem_cgroup_reset,
6064 .read_u64 = mem_cgroup_read_u64,
6066 { }, /* terminate */
6069 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6071 struct mem_cgroup_per_node *pn;
6072 struct mem_cgroup_per_zone *mz;
6073 int zone, tmp = node;
6075 * This routine is called against possible nodes.
6076 * But it's BUG to call kmalloc() against offline node.
6078 * TODO: this routine can waste much memory for nodes which will
6079 * never be onlined. It's better to use memory hotplug callback
6082 if (!node_state(node, N_NORMAL_MEMORY))
6084 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6088 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6089 mz = &pn->zoneinfo[zone];
6090 lruvec_init(&mz->lruvec);
6091 mz->usage_in_excess = 0;
6092 mz->on_tree = false;
6095 memcg->nodeinfo[node] = pn;
6099 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6101 kfree(memcg->nodeinfo[node]);
6104 static struct mem_cgroup *mem_cgroup_alloc(void)
6106 struct mem_cgroup *memcg;
6109 size = sizeof(struct mem_cgroup);
6110 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6112 memcg = kzalloc(size, GFP_KERNEL);
6116 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6119 spin_lock_init(&memcg->pcp_counter_lock);
6128 * At destroying mem_cgroup, references from swap_cgroup can remain.
6129 * (scanning all at force_empty is too costly...)
6131 * Instead of clearing all references at force_empty, we remember
6132 * the number of reference from swap_cgroup and free mem_cgroup when
6133 * it goes down to 0.
6135 * Removal of cgroup itself succeeds regardless of refs from swap.
6138 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6142 mem_cgroup_remove_from_trees(memcg);
6145 free_mem_cgroup_per_zone_info(memcg, node);
6147 free_percpu(memcg->stat);
6150 * We need to make sure that (at least for now), the jump label
6151 * destruction code runs outside of the cgroup lock. This is because
6152 * get_online_cpus(), which is called from the static_branch update,
6153 * can't be called inside the cgroup_lock. cpusets are the ones
6154 * enforcing this dependency, so if they ever change, we might as well.
6156 * schedule_work() will guarantee this happens. Be careful if you need
6157 * to move this code around, and make sure it is outside
6160 disarm_static_keys(memcg);
6165 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6167 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6169 if (!memcg->res.parent)
6171 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6173 EXPORT_SYMBOL(parent_mem_cgroup);
6175 static void __init mem_cgroup_soft_limit_tree_init(void)
6177 struct mem_cgroup_tree_per_node *rtpn;
6178 struct mem_cgroup_tree_per_zone *rtpz;
6179 int tmp, node, zone;
6181 for_each_node(node) {
6183 if (!node_state(node, N_NORMAL_MEMORY))
6185 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6188 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6190 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6191 rtpz = &rtpn->rb_tree_per_zone[zone];
6192 rtpz->rb_root = RB_ROOT;
6193 spin_lock_init(&rtpz->lock);
6198 static struct cgroup_subsys_state * __ref
6199 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6201 struct mem_cgroup *memcg;
6202 long error = -ENOMEM;
6205 memcg = mem_cgroup_alloc();
6207 return ERR_PTR(error);
6210 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6214 if (parent_css == NULL) {
6215 root_mem_cgroup = memcg;
6216 res_counter_init(&memcg->res, NULL);
6217 res_counter_init(&memcg->memsw, NULL);
6218 res_counter_init(&memcg->kmem, NULL);
6221 memcg->last_scanned_node = MAX_NUMNODES;
6222 INIT_LIST_HEAD(&memcg->oom_notify);
6223 memcg->move_charge_at_immigrate = 0;
6224 mutex_init(&memcg->thresholds_lock);
6225 spin_lock_init(&memcg->move_lock);
6226 vmpressure_init(&memcg->vmpressure);
6227 INIT_LIST_HEAD(&memcg->event_list);
6228 spin_lock_init(&memcg->event_list_lock);
6233 __mem_cgroup_free(memcg);
6234 return ERR_PTR(error);
6238 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6240 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6241 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
6243 if (css->id > MEM_CGROUP_ID_MAX)
6249 mutex_lock(&memcg_create_mutex);
6251 memcg->use_hierarchy = parent->use_hierarchy;
6252 memcg->oom_kill_disable = parent->oom_kill_disable;
6253 memcg->swappiness = mem_cgroup_swappiness(parent);
6255 if (parent->use_hierarchy) {
6256 res_counter_init(&memcg->res, &parent->res);
6257 res_counter_init(&memcg->memsw, &parent->memsw);
6258 res_counter_init(&memcg->kmem, &parent->kmem);
6261 * No need to take a reference to the parent because cgroup
6262 * core guarantees its existence.
6265 res_counter_init(&memcg->res, NULL);
6266 res_counter_init(&memcg->memsw, NULL);
6267 res_counter_init(&memcg->kmem, NULL);
6269 * Deeper hierachy with use_hierarchy == false doesn't make
6270 * much sense so let cgroup subsystem know about this
6271 * unfortunate state in our controller.
6273 if (parent != root_mem_cgroup)
6274 memory_cgrp_subsys.broken_hierarchy = true;
6276 mutex_unlock(&memcg_create_mutex);
6278 return memcg_init_kmem(memcg, &memory_cgrp_subsys);
6282 * Announce all parents that a group from their hierarchy is gone.
6284 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6286 struct mem_cgroup *parent = memcg;
6288 while ((parent = parent_mem_cgroup(parent)))
6289 mem_cgroup_iter_invalidate(parent);
6292 * if the root memcg is not hierarchical we have to check it
6295 if (!root_mem_cgroup->use_hierarchy)
6296 mem_cgroup_iter_invalidate(root_mem_cgroup);
6299 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6301 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6302 struct mem_cgroup_event *event, *tmp;
6303 struct cgroup_subsys_state *iter;
6306 * Unregister events and notify userspace.
6307 * Notify userspace about cgroup removing only after rmdir of cgroup
6308 * directory to avoid race between userspace and kernelspace.
6310 spin_lock(&memcg->event_list_lock);
6311 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6312 list_del_init(&event->list);
6313 schedule_work(&event->remove);
6315 spin_unlock(&memcg->event_list_lock);
6317 kmem_cgroup_css_offline(memcg);
6319 mem_cgroup_invalidate_reclaim_iterators(memcg);
6322 * This requires that offlining is serialized. Right now that is
6323 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6325 css_for_each_descendant_post(iter, css)
6326 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
6328 memcg_unregister_all_caches(memcg);
6329 vmpressure_cleanup(&memcg->vmpressure);
6332 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6334 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6336 * XXX: css_offline() would be where we should reparent all
6337 * memory to prepare the cgroup for destruction. However,
6338 * memcg does not do css_tryget_online() and res_counter charging
6339 * under the same RCU lock region, which means that charging
6340 * could race with offlining. Offlining only happens to
6341 * cgroups with no tasks in them but charges can show up
6342 * without any tasks from the swapin path when the target
6343 * memcg is looked up from the swapout record and not from the
6344 * current task as it usually is. A race like this can leak
6345 * charges and put pages with stale cgroup pointers into
6349 * lookup_swap_cgroup_id()
6351 * mem_cgroup_lookup()
6352 * css_tryget_online()
6354 * disable css_tryget_online()
6357 * reparent_charges()
6358 * res_counter_charge()
6361 * pc->mem_cgroup = dead memcg
6364 * The bulk of the charges are still moved in offline_css() to
6365 * avoid pinning a lot of pages in case a long-term reference
6366 * like a swapout record is deferring the css_free() to long
6367 * after offlining. But this makes sure we catch any charges
6368 * made after offlining:
6370 mem_cgroup_reparent_charges(memcg);
6372 memcg_destroy_kmem(memcg);
6373 __mem_cgroup_free(memcg);
6377 * mem_cgroup_css_reset - reset the states of a mem_cgroup
6378 * @css: the target css
6380 * Reset the states of the mem_cgroup associated with @css. This is
6381 * invoked when the userland requests disabling on the default hierarchy
6382 * but the memcg is pinned through dependency. The memcg should stop
6383 * applying policies and should revert to the vanilla state as it may be
6384 * made visible again.
6386 * The current implementation only resets the essential configurations.
6387 * This needs to be expanded to cover all the visible parts.
6389 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
6391 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6393 mem_cgroup_resize_limit(memcg, ULLONG_MAX);
6394 mem_cgroup_resize_memsw_limit(memcg, ULLONG_MAX);
6395 memcg_update_kmem_limit(memcg, ULLONG_MAX);
6396 res_counter_set_soft_limit(&memcg->res, ULLONG_MAX);
6400 /* Handlers for move charge at task migration. */
6401 #define PRECHARGE_COUNT_AT_ONCE 256
6402 static int mem_cgroup_do_precharge(unsigned long count)
6405 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6406 struct mem_cgroup *memcg = mc.to;
6408 if (mem_cgroup_is_root(memcg)) {
6409 mc.precharge += count;
6410 /* we don't need css_get for root */
6413 /* try to charge at once */
6415 struct res_counter *dummy;
6417 * "memcg" cannot be under rmdir() because we've already checked
6418 * by cgroup_lock_live_cgroup() that it is not removed and we
6419 * are still under the same cgroup_mutex. So we can postpone
6422 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6424 if (do_swap_account && res_counter_charge(&memcg->memsw,
6425 PAGE_SIZE * count, &dummy)) {
6426 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6429 mc.precharge += count;
6433 /* fall back to one by one charge */
6435 if (signal_pending(current)) {
6439 if (!batch_count--) {
6440 batch_count = PRECHARGE_COUNT_AT_ONCE;
6443 ret = mem_cgroup_try_charge(memcg, GFP_KERNEL, 1, false);
6445 /* mem_cgroup_clear_mc() will do uncharge later */
6453 * get_mctgt_type - get target type of moving charge
6454 * @vma: the vma the pte to be checked belongs
6455 * @addr: the address corresponding to the pte to be checked
6456 * @ptent: the pte to be checked
6457 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6460 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6461 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6462 * move charge. if @target is not NULL, the page is stored in target->page
6463 * with extra refcnt got(Callers should handle it).
6464 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6465 * target for charge migration. if @target is not NULL, the entry is stored
6468 * Called with pte lock held.
6475 enum mc_target_type {
6481 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6482 unsigned long addr, pte_t ptent)
6484 struct page *page = vm_normal_page(vma, addr, ptent);
6486 if (!page || !page_mapped(page))
6488 if (PageAnon(page)) {
6489 /* we don't move shared anon */
6492 } else if (!move_file())
6493 /* we ignore mapcount for file pages */
6495 if (!get_page_unless_zero(page))
6502 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6503 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6505 struct page *page = NULL;
6506 swp_entry_t ent = pte_to_swp_entry(ptent);
6508 if (!move_anon() || non_swap_entry(ent))
6511 * Because lookup_swap_cache() updates some statistics counter,
6512 * we call find_get_page() with swapper_space directly.
6514 page = find_get_page(swap_address_space(ent), ent.val);
6515 if (do_swap_account)
6516 entry->val = ent.val;
6521 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6522 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6528 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6529 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6531 struct page *page = NULL;
6532 struct address_space *mapping;
6535 if (!vma->vm_file) /* anonymous vma */
6540 mapping = vma->vm_file->f_mapping;
6541 if (pte_none(ptent))
6542 pgoff = linear_page_index(vma, addr);
6543 else /* pte_file(ptent) is true */
6544 pgoff = pte_to_pgoff(ptent);
6546 /* page is moved even if it's not RSS of this task(page-faulted). */
6548 /* shmem/tmpfs may report page out on swap: account for that too. */
6549 if (shmem_mapping(mapping)) {
6550 page = find_get_entry(mapping, pgoff);
6551 if (radix_tree_exceptional_entry(page)) {
6552 swp_entry_t swp = radix_to_swp_entry(page);
6553 if (do_swap_account)
6555 page = find_get_page(swap_address_space(swp), swp.val);
6558 page = find_get_page(mapping, pgoff);
6560 page = find_get_page(mapping, pgoff);
6565 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6566 unsigned long addr, pte_t ptent, union mc_target *target)
6568 struct page *page = NULL;
6569 struct page_cgroup *pc;
6570 enum mc_target_type ret = MC_TARGET_NONE;
6571 swp_entry_t ent = { .val = 0 };
6573 if (pte_present(ptent))
6574 page = mc_handle_present_pte(vma, addr, ptent);
6575 else if (is_swap_pte(ptent))
6576 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6577 else if (pte_none(ptent) || pte_file(ptent))
6578 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6580 if (!page && !ent.val)
6583 pc = lookup_page_cgroup(page);
6585 * Do only loose check w/o page_cgroup lock.
6586 * mem_cgroup_move_account() checks the pc is valid or not under
6589 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6590 ret = MC_TARGET_PAGE;
6592 target->page = page;
6594 if (!ret || !target)
6597 /* There is a swap entry and a page doesn't exist or isn't charged */
6598 if (ent.val && !ret &&
6599 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6600 ret = MC_TARGET_SWAP;
6607 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6609 * We don't consider swapping or file mapped pages because THP does not
6610 * support them for now.
6611 * Caller should make sure that pmd_trans_huge(pmd) is true.
6613 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6614 unsigned long addr, pmd_t pmd, union mc_target *target)
6616 struct page *page = NULL;
6617 struct page_cgroup *pc;
6618 enum mc_target_type ret = MC_TARGET_NONE;
6620 page = pmd_page(pmd);
6621 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6624 pc = lookup_page_cgroup(page);
6625 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6626 ret = MC_TARGET_PAGE;
6629 target->page = page;
6635 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6636 unsigned long addr, pmd_t pmd, union mc_target *target)
6638 return MC_TARGET_NONE;
6642 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6643 unsigned long addr, unsigned long end,
6644 struct mm_walk *walk)
6646 struct vm_area_struct *vma = walk->private;
6650 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6651 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6652 mc.precharge += HPAGE_PMD_NR;
6657 if (pmd_trans_unstable(pmd))
6659 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6660 for (; addr != end; pte++, addr += PAGE_SIZE)
6661 if (get_mctgt_type(vma, addr, *pte, NULL))
6662 mc.precharge++; /* increment precharge temporarily */
6663 pte_unmap_unlock(pte - 1, ptl);
6669 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6671 unsigned long precharge;
6672 struct vm_area_struct *vma;
6674 down_read(&mm->mmap_sem);
6675 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6676 struct mm_walk mem_cgroup_count_precharge_walk = {
6677 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6681 if (is_vm_hugetlb_page(vma))
6683 walk_page_range(vma->vm_start, vma->vm_end,
6684 &mem_cgroup_count_precharge_walk);
6686 up_read(&mm->mmap_sem);
6688 precharge = mc.precharge;
6694 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6696 unsigned long precharge = mem_cgroup_count_precharge(mm);
6698 VM_BUG_ON(mc.moving_task);
6699 mc.moving_task = current;
6700 return mem_cgroup_do_precharge(precharge);
6703 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6704 static void __mem_cgroup_clear_mc(void)
6706 struct mem_cgroup *from = mc.from;
6707 struct mem_cgroup *to = mc.to;
6710 /* we must uncharge all the leftover precharges from mc.to */
6712 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6716 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6717 * we must uncharge here.
6719 if (mc.moved_charge) {
6720 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6721 mc.moved_charge = 0;
6723 /* we must fixup refcnts and charges */
6724 if (mc.moved_swap) {
6725 /* uncharge swap account from the old cgroup */
6726 if (!mem_cgroup_is_root(mc.from))
6727 res_counter_uncharge(&mc.from->memsw,
6728 PAGE_SIZE * mc.moved_swap);
6730 for (i = 0; i < mc.moved_swap; i++)
6731 css_put(&mc.from->css);
6733 if (!mem_cgroup_is_root(mc.to)) {
6735 * we charged both to->res and to->memsw, so we should
6738 res_counter_uncharge(&mc.to->res,
6739 PAGE_SIZE * mc.moved_swap);
6741 /* we've already done css_get(mc.to) */
6744 memcg_oom_recover(from);
6745 memcg_oom_recover(to);
6746 wake_up_all(&mc.waitq);
6749 static void mem_cgroup_clear_mc(void)
6751 struct mem_cgroup *from = mc.from;
6754 * we must clear moving_task before waking up waiters at the end of
6757 mc.moving_task = NULL;
6758 __mem_cgroup_clear_mc();
6759 spin_lock(&mc.lock);
6762 spin_unlock(&mc.lock);
6763 mem_cgroup_end_move(from);
6766 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6767 struct cgroup_taskset *tset)
6769 struct task_struct *p = cgroup_taskset_first(tset);
6771 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6772 unsigned long move_charge_at_immigrate;
6775 * We are now commited to this value whatever it is. Changes in this
6776 * tunable will only affect upcoming migrations, not the current one.
6777 * So we need to save it, and keep it going.
6779 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6780 if (move_charge_at_immigrate) {
6781 struct mm_struct *mm;
6782 struct mem_cgroup *from = mem_cgroup_from_task(p);
6784 VM_BUG_ON(from == memcg);
6786 mm = get_task_mm(p);
6789 /* We move charges only when we move a owner of the mm */
6790 if (mm->owner == p) {
6793 VM_BUG_ON(mc.precharge);
6794 VM_BUG_ON(mc.moved_charge);
6795 VM_BUG_ON(mc.moved_swap);
6796 mem_cgroup_start_move(from);
6797 spin_lock(&mc.lock);
6800 mc.immigrate_flags = move_charge_at_immigrate;
6801 spin_unlock(&mc.lock);
6802 /* We set mc.moving_task later */
6804 ret = mem_cgroup_precharge_mc(mm);
6806 mem_cgroup_clear_mc();
6813 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6814 struct cgroup_taskset *tset)
6816 mem_cgroup_clear_mc();
6819 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6820 unsigned long addr, unsigned long end,
6821 struct mm_walk *walk)
6824 struct vm_area_struct *vma = walk->private;
6827 enum mc_target_type target_type;
6828 union mc_target target;
6830 struct page_cgroup *pc;
6833 * We don't take compound_lock() here but no race with splitting thp
6835 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6836 * under splitting, which means there's no concurrent thp split,
6837 * - if another thread runs into split_huge_page() just after we
6838 * entered this if-block, the thread must wait for page table lock
6839 * to be unlocked in __split_huge_page_splitting(), where the main
6840 * part of thp split is not executed yet.
6842 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6843 if (mc.precharge < HPAGE_PMD_NR) {
6847 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6848 if (target_type == MC_TARGET_PAGE) {
6850 if (!isolate_lru_page(page)) {
6851 pc = lookup_page_cgroup(page);
6852 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6853 pc, mc.from, mc.to)) {
6854 mc.precharge -= HPAGE_PMD_NR;
6855 mc.moved_charge += HPAGE_PMD_NR;
6857 putback_lru_page(page);
6865 if (pmd_trans_unstable(pmd))
6868 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6869 for (; addr != end; addr += PAGE_SIZE) {
6870 pte_t ptent = *(pte++);
6876 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6877 case MC_TARGET_PAGE:
6879 if (isolate_lru_page(page))
6881 pc = lookup_page_cgroup(page);
6882 if (!mem_cgroup_move_account(page, 1, pc,
6885 /* we uncharge from mc.from later. */
6888 putback_lru_page(page);
6889 put: /* get_mctgt_type() gets the page */
6892 case MC_TARGET_SWAP:
6894 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6896 /* we fixup refcnts and charges later. */
6904 pte_unmap_unlock(pte - 1, ptl);
6909 * We have consumed all precharges we got in can_attach().
6910 * We try charge one by one, but don't do any additional
6911 * charges to mc.to if we have failed in charge once in attach()
6914 ret = mem_cgroup_do_precharge(1);
6922 static void mem_cgroup_move_charge(struct mm_struct *mm)
6924 struct vm_area_struct *vma;
6926 lru_add_drain_all();
6928 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6930 * Someone who are holding the mmap_sem might be waiting in
6931 * waitq. So we cancel all extra charges, wake up all waiters,
6932 * and retry. Because we cancel precharges, we might not be able
6933 * to move enough charges, but moving charge is a best-effort
6934 * feature anyway, so it wouldn't be a big problem.
6936 __mem_cgroup_clear_mc();
6940 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6942 struct mm_walk mem_cgroup_move_charge_walk = {
6943 .pmd_entry = mem_cgroup_move_charge_pte_range,
6947 if (is_vm_hugetlb_page(vma))
6949 ret = walk_page_range(vma->vm_start, vma->vm_end,
6950 &mem_cgroup_move_charge_walk);
6953 * means we have consumed all precharges and failed in
6954 * doing additional charge. Just abandon here.
6958 up_read(&mm->mmap_sem);
6961 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6962 struct cgroup_taskset *tset)
6964 struct task_struct *p = cgroup_taskset_first(tset);
6965 struct mm_struct *mm = get_task_mm(p);
6969 mem_cgroup_move_charge(mm);
6973 mem_cgroup_clear_mc();
6975 #else /* !CONFIG_MMU */
6976 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6977 struct cgroup_taskset *tset)
6981 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6982 struct cgroup_taskset *tset)
6985 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6986 struct cgroup_taskset *tset)
6992 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6993 * to verify whether we're attached to the default hierarchy on each mount
6996 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6999 * use_hierarchy is forced on the default hierarchy. cgroup core
7000 * guarantees that @root doesn't have any children, so turning it
7001 * on for the root memcg is enough.
7003 if (cgroup_on_dfl(root_css->cgroup))
7004 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7007 struct cgroup_subsys memory_cgrp_subsys = {
7008 .css_alloc = mem_cgroup_css_alloc,
7009 .css_online = mem_cgroup_css_online,
7010 .css_offline = mem_cgroup_css_offline,
7011 .css_free = mem_cgroup_css_free,
7012 .css_reset = mem_cgroup_css_reset,
7013 .can_attach = mem_cgroup_can_attach,
7014 .cancel_attach = mem_cgroup_cancel_attach,
7015 .attach = mem_cgroup_move_task,
7016 .bind = mem_cgroup_bind,
7017 .legacy_cftypes = mem_cgroup_files,
7021 #ifdef CONFIG_MEMCG_SWAP
7022 static int __init enable_swap_account(char *s)
7024 if (!strcmp(s, "1"))
7025 really_do_swap_account = 1;
7026 else if (!strcmp(s, "0"))
7027 really_do_swap_account = 0;
7030 __setup("swapaccount=", enable_swap_account);
7032 static void __init memsw_file_init(void)
7034 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7035 memsw_cgroup_files));
7038 static void __init enable_swap_cgroup(void)
7040 if (!mem_cgroup_disabled() && really_do_swap_account) {
7041 do_swap_account = 1;
7047 static void __init enable_swap_cgroup(void)
7053 * subsys_initcall() for memory controller.
7055 * Some parts like hotcpu_notifier() have to be initialized from this context
7056 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7057 * everything that doesn't depend on a specific mem_cgroup structure should
7058 * be initialized from here.
7060 static int __init mem_cgroup_init(void)
7062 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7063 enable_swap_cgroup();
7064 mem_cgroup_soft_limit_tree_init();
7068 subsys_initcall(mem_cgroup_init);