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_oom_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 * Unlike in global OOM situations, memcg is not in a physical
2580 * memory shortage. Allow dying and OOM-killed tasks to
2581 * bypass the last charges so that they can exit quickly and
2582 * free their memory.
2584 if (unlikely(test_thread_flag(TIF_MEMDIE) ||
2585 fatal_signal_pending(current) ||
2586 current->flags & PF_EXITING))
2589 if (unlikely(task_in_memcg_oom(current)))
2592 if (gfp_mask & __GFP_NOFAIL)
2595 if (consume_stock(memcg, nr_pages))
2598 size = batch * PAGE_SIZE;
2599 if (!res_counter_charge(&memcg->res, size, &fail_res)) {
2600 if (!do_swap_account)
2602 if (!res_counter_charge(&memcg->memsw, size, &fail_res))
2604 res_counter_uncharge(&memcg->res, size);
2605 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2606 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2608 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2610 if (batch > nr_pages) {
2615 if (!(gfp_mask & __GFP_WAIT))
2618 if (gfp_mask & __GFP_NORETRY)
2621 nr_reclaimed = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2623 if (mem_cgroup_margin(mem_over_limit) >= batch)
2626 * Even though the limit is exceeded at this point, reclaim
2627 * may have been able to free some pages. Retry the charge
2628 * before killing the task.
2630 * Only for regular pages, though: huge pages are rather
2631 * unlikely to succeed so close to the limit, and we fall back
2632 * to regular pages anyway in case of failure.
2634 if (nr_reclaimed && batch <= (1 << PAGE_ALLOC_COSTLY_ORDER))
2637 * At task move, charge accounts can be doubly counted. So, it's
2638 * better to wait until the end of task_move if something is going on.
2640 if (mem_cgroup_wait_acct_move(mem_over_limit))
2643 if (fatal_signal_pending(current))
2649 if (nr_oom_retries--)
2652 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(batch));
2654 if (!(gfp_mask & __GFP_NOFAIL))
2660 if (batch > nr_pages)
2661 refill_stock(memcg, batch - nr_pages);
2667 * mem_cgroup_try_charge_mm - try charging a mm
2668 * @mm: mm_struct to charge
2669 * @nr_pages: number of pages to charge
2670 * @oom: trigger OOM if reclaim fails
2672 * Returns the charged mem_cgroup associated with the given mm_struct or
2673 * NULL the charge failed.
2675 static struct mem_cgroup *mem_cgroup_try_charge_mm(struct mm_struct *mm,
2677 unsigned int nr_pages,
2681 struct mem_cgroup *memcg;
2684 memcg = get_mem_cgroup_from_mm(mm);
2685 ret = mem_cgroup_try_charge(memcg, gfp_mask, nr_pages, oom);
2686 css_put(&memcg->css);
2688 memcg = root_mem_cgroup;
2696 * Somemtimes we have to undo a charge we got by try_charge().
2697 * This function is for that and do uncharge, put css's refcnt.
2698 * gotten by try_charge().
2700 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2701 unsigned int nr_pages)
2703 if (!mem_cgroup_is_root(memcg)) {
2704 unsigned long bytes = nr_pages * PAGE_SIZE;
2706 res_counter_uncharge(&memcg->res, bytes);
2707 if (do_swap_account)
2708 res_counter_uncharge(&memcg->memsw, bytes);
2713 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2714 * This is useful when moving usage to parent cgroup.
2716 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2717 unsigned int nr_pages)
2719 unsigned long bytes = nr_pages * PAGE_SIZE;
2721 if (mem_cgroup_is_root(memcg))
2724 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2725 if (do_swap_account)
2726 res_counter_uncharge_until(&memcg->memsw,
2727 memcg->memsw.parent, bytes);
2731 * A helper function to get mem_cgroup from ID. must be called under
2732 * rcu_read_lock(). The caller is responsible for calling
2733 * css_tryget_online() if the mem_cgroup is used for charging. (dropping
2734 * refcnt from swap can be called against removed memcg.)
2736 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2738 /* ID 0 is unused ID */
2741 return mem_cgroup_from_id(id);
2744 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2746 struct mem_cgroup *memcg = NULL;
2747 struct page_cgroup *pc;
2751 VM_BUG_ON_PAGE(!PageLocked(page), page);
2753 pc = lookup_page_cgroup(page);
2754 lock_page_cgroup(pc);
2755 if (PageCgroupUsed(pc)) {
2756 memcg = pc->mem_cgroup;
2757 if (memcg && !css_tryget_online(&memcg->css))
2759 } else if (PageSwapCache(page)) {
2760 ent.val = page_private(page);
2761 id = lookup_swap_cgroup_id(ent);
2763 memcg = mem_cgroup_lookup(id);
2764 if (memcg && !css_tryget_online(&memcg->css))
2768 unlock_page_cgroup(pc);
2772 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2774 unsigned int nr_pages,
2775 enum charge_type ctype,
2778 struct page_cgroup *pc = lookup_page_cgroup(page);
2779 struct zone *uninitialized_var(zone);
2780 struct lruvec *lruvec;
2781 bool was_on_lru = false;
2784 lock_page_cgroup(pc);
2785 VM_BUG_ON_PAGE(PageCgroupUsed(pc), page);
2787 * we don't need page_cgroup_lock about tail pages, becase they are not
2788 * accessed by any other context at this point.
2792 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2793 * may already be on some other mem_cgroup's LRU. Take care of it.
2796 zone = page_zone(page);
2797 spin_lock_irq(&zone->lru_lock);
2798 if (PageLRU(page)) {
2799 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2801 del_page_from_lru_list(page, lruvec, page_lru(page));
2806 pc->mem_cgroup = memcg;
2808 * We access a page_cgroup asynchronously without lock_page_cgroup().
2809 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2810 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2811 * before USED bit, we need memory barrier here.
2812 * See mem_cgroup_add_lru_list(), etc.
2815 SetPageCgroupUsed(pc);
2819 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2820 VM_BUG_ON_PAGE(PageLRU(page), page);
2822 add_page_to_lru_list(page, lruvec, page_lru(page));
2824 spin_unlock_irq(&zone->lru_lock);
2827 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2832 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2833 unlock_page_cgroup(pc);
2836 * "charge_statistics" updated event counter. Then, check it.
2837 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2838 * if they exceeds softlimit.
2840 memcg_check_events(memcg, page);
2843 static DEFINE_MUTEX(set_limit_mutex);
2845 #ifdef CONFIG_MEMCG_KMEM
2847 * The memcg_slab_mutex is held whenever a per memcg kmem cache is created or
2848 * destroyed. It protects memcg_caches arrays and memcg_slab_caches lists.
2850 static DEFINE_MUTEX(memcg_slab_mutex);
2852 static DEFINE_MUTEX(activate_kmem_mutex);
2854 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2856 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2857 memcg_kmem_is_active(memcg);
2861 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2862 * in the memcg_cache_params struct.
2864 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2866 struct kmem_cache *cachep;
2868 VM_BUG_ON(p->is_root_cache);
2869 cachep = p->root_cache;
2870 return cache_from_memcg_idx(cachep, memcg_cache_id(p->memcg));
2873 #ifdef CONFIG_SLABINFO
2874 static int mem_cgroup_slabinfo_read(struct seq_file *m, void *v)
2876 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
2877 struct memcg_cache_params *params;
2879 if (!memcg_can_account_kmem(memcg))
2882 print_slabinfo_header(m);
2884 mutex_lock(&memcg_slab_mutex);
2885 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2886 cache_show(memcg_params_to_cache(params), m);
2887 mutex_unlock(&memcg_slab_mutex);
2893 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2895 struct res_counter *fail_res;
2898 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2902 ret = mem_cgroup_try_charge(memcg, gfp, size >> PAGE_SHIFT,
2903 oom_gfp_allowed(gfp));
2904 if (ret == -EINTR) {
2906 * mem_cgroup_try_charge() chosed to bypass to root due to
2907 * OOM kill or fatal signal. Since our only options are to
2908 * either fail the allocation or charge it to this cgroup, do
2909 * it as a temporary condition. But we can't fail. From a
2910 * kmem/slab perspective, the cache has already been selected,
2911 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2914 * This condition will only trigger if the task entered
2915 * memcg_charge_kmem in a sane state, but was OOM-killed during
2916 * mem_cgroup_try_charge() above. Tasks that were already
2917 * dying when the allocation triggers should have been already
2918 * directed to the root cgroup in memcontrol.h
2920 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2921 if (do_swap_account)
2922 res_counter_charge_nofail(&memcg->memsw, size,
2926 res_counter_uncharge(&memcg->kmem, size);
2931 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2933 res_counter_uncharge(&memcg->res, size);
2934 if (do_swap_account)
2935 res_counter_uncharge(&memcg->memsw, size);
2938 if (res_counter_uncharge(&memcg->kmem, size))
2942 * Releases a reference taken in kmem_cgroup_css_offline in case
2943 * this last uncharge is racing with the offlining code or it is
2944 * outliving the memcg existence.
2946 * The memory barrier imposed by test&clear is paired with the
2947 * explicit one in memcg_kmem_mark_dead().
2949 if (memcg_kmem_test_and_clear_dead(memcg))
2950 css_put(&memcg->css);
2954 * helper for acessing a memcg's index. It will be used as an index in the
2955 * child cache array in kmem_cache, and also to derive its name. This function
2956 * will return -1 when this is not a kmem-limited memcg.
2958 int memcg_cache_id(struct mem_cgroup *memcg)
2960 return memcg ? memcg->kmemcg_id : -1;
2963 static size_t memcg_caches_array_size(int num_groups)
2966 if (num_groups <= 0)
2969 size = 2 * num_groups;
2970 if (size < MEMCG_CACHES_MIN_SIZE)
2971 size = MEMCG_CACHES_MIN_SIZE;
2972 else if (size > MEMCG_CACHES_MAX_SIZE)
2973 size = MEMCG_CACHES_MAX_SIZE;
2979 * We should update the current array size iff all caches updates succeed. This
2980 * can only be done from the slab side. The slab mutex needs to be held when
2983 void memcg_update_array_size(int num)
2985 if (num > memcg_limited_groups_array_size)
2986 memcg_limited_groups_array_size = memcg_caches_array_size(num);
2989 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
2991 struct memcg_cache_params *cur_params = s->memcg_params;
2993 VM_BUG_ON(!is_root_cache(s));
2995 if (num_groups > memcg_limited_groups_array_size) {
2997 struct memcg_cache_params *new_params;
2998 ssize_t size = memcg_caches_array_size(num_groups);
3000 size *= sizeof(void *);
3001 size += offsetof(struct memcg_cache_params, memcg_caches);
3003 new_params = kzalloc(size, GFP_KERNEL);
3007 new_params->is_root_cache = true;
3010 * There is the chance it will be bigger than
3011 * memcg_limited_groups_array_size, if we failed an allocation
3012 * in a cache, in which case all caches updated before it, will
3013 * have a bigger array.
3015 * But if that is the case, the data after
3016 * memcg_limited_groups_array_size is certainly unused
3018 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3019 if (!cur_params->memcg_caches[i])
3021 new_params->memcg_caches[i] =
3022 cur_params->memcg_caches[i];
3026 * Ideally, we would wait until all caches succeed, and only
3027 * then free the old one. But this is not worth the extra
3028 * pointer per-cache we'd have to have for this.
3030 * It is not a big deal if some caches are left with a size
3031 * bigger than the others. And all updates will reset this
3034 rcu_assign_pointer(s->memcg_params, new_params);
3036 kfree_rcu(cur_params, rcu_head);
3041 int memcg_alloc_cache_params(struct mem_cgroup *memcg, struct kmem_cache *s,
3042 struct kmem_cache *root_cache)
3046 if (!memcg_kmem_enabled())
3050 size = offsetof(struct memcg_cache_params, memcg_caches);
3051 size += memcg_limited_groups_array_size * sizeof(void *);
3053 size = sizeof(struct memcg_cache_params);
3055 s->memcg_params = kzalloc(size, GFP_KERNEL);
3056 if (!s->memcg_params)
3060 s->memcg_params->memcg = memcg;
3061 s->memcg_params->root_cache = root_cache;
3062 css_get(&memcg->css);
3064 s->memcg_params->is_root_cache = true;
3069 void memcg_free_cache_params(struct kmem_cache *s)
3071 if (!s->memcg_params)
3073 if (!s->memcg_params->is_root_cache)
3074 css_put(&s->memcg_params->memcg->css);
3075 kfree(s->memcg_params);
3078 static void memcg_register_cache(struct mem_cgroup *memcg,
3079 struct kmem_cache *root_cache)
3081 static char memcg_name_buf[NAME_MAX + 1]; /* protected by
3083 struct kmem_cache *cachep;
3086 lockdep_assert_held(&memcg_slab_mutex);
3088 id = memcg_cache_id(memcg);
3091 * Since per-memcg caches are created asynchronously on first
3092 * allocation (see memcg_kmem_get_cache()), several threads can try to
3093 * create the same cache, but only one of them may succeed.
3095 if (cache_from_memcg_idx(root_cache, id))
3098 cgroup_name(memcg->css.cgroup, memcg_name_buf, NAME_MAX + 1);
3099 cachep = memcg_create_kmem_cache(memcg, root_cache, memcg_name_buf);
3101 * If we could not create a memcg cache, do not complain, because
3102 * that's not critical at all as we can always proceed with the root
3108 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3111 * Since readers won't lock (see cache_from_memcg_idx()), we need a
3112 * barrier here to ensure nobody will see the kmem_cache partially
3117 BUG_ON(root_cache->memcg_params->memcg_caches[id]);
3118 root_cache->memcg_params->memcg_caches[id] = cachep;
3121 static void memcg_unregister_cache(struct kmem_cache *cachep)
3123 struct kmem_cache *root_cache;
3124 struct mem_cgroup *memcg;
3127 lockdep_assert_held(&memcg_slab_mutex);
3129 BUG_ON(is_root_cache(cachep));
3131 root_cache = cachep->memcg_params->root_cache;
3132 memcg = cachep->memcg_params->memcg;
3133 id = memcg_cache_id(memcg);
3135 BUG_ON(root_cache->memcg_params->memcg_caches[id] != cachep);
3136 root_cache->memcg_params->memcg_caches[id] = NULL;
3138 list_del(&cachep->memcg_params->list);
3140 kmem_cache_destroy(cachep);
3144 * During the creation a new cache, we need to disable our accounting mechanism
3145 * altogether. This is true even if we are not creating, but rather just
3146 * enqueing new caches to be created.
3148 * This is because that process will trigger allocations; some visible, like
3149 * explicit kmallocs to auxiliary data structures, name strings and internal
3150 * cache structures; some well concealed, like INIT_WORK() that can allocate
3151 * objects during debug.
3153 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3154 * to it. This may not be a bounded recursion: since the first cache creation
3155 * failed to complete (waiting on the allocation), we'll just try to create the
3156 * cache again, failing at the same point.
3158 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3159 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3160 * inside the following two functions.
3162 static inline void memcg_stop_kmem_account(void)
3164 VM_BUG_ON(!current->mm);
3165 current->memcg_kmem_skip_account++;
3168 static inline void memcg_resume_kmem_account(void)
3170 VM_BUG_ON(!current->mm);
3171 current->memcg_kmem_skip_account--;
3174 int __memcg_cleanup_cache_params(struct kmem_cache *s)
3176 struct kmem_cache *c;
3179 mutex_lock(&memcg_slab_mutex);
3180 for_each_memcg_cache_index(i) {
3181 c = cache_from_memcg_idx(s, i);
3185 memcg_unregister_cache(c);
3187 if (cache_from_memcg_idx(s, i))
3190 mutex_unlock(&memcg_slab_mutex);
3194 static void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3196 struct kmem_cache *cachep;
3197 struct memcg_cache_params *params, *tmp;
3199 if (!memcg_kmem_is_active(memcg))
3202 mutex_lock(&memcg_slab_mutex);
3203 list_for_each_entry_safe(params, tmp, &memcg->memcg_slab_caches, list) {
3204 cachep = memcg_params_to_cache(params);
3205 kmem_cache_shrink(cachep);
3206 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3207 memcg_unregister_cache(cachep);
3209 mutex_unlock(&memcg_slab_mutex);
3212 struct memcg_register_cache_work {
3213 struct mem_cgroup *memcg;
3214 struct kmem_cache *cachep;
3215 struct work_struct work;
3218 static void memcg_register_cache_func(struct work_struct *w)
3220 struct memcg_register_cache_work *cw =
3221 container_of(w, struct memcg_register_cache_work, work);
3222 struct mem_cgroup *memcg = cw->memcg;
3223 struct kmem_cache *cachep = cw->cachep;
3225 mutex_lock(&memcg_slab_mutex);
3226 memcg_register_cache(memcg, cachep);
3227 mutex_unlock(&memcg_slab_mutex);
3229 css_put(&memcg->css);
3234 * Enqueue the creation of a per-memcg kmem_cache.
3236 static void __memcg_schedule_register_cache(struct mem_cgroup *memcg,
3237 struct kmem_cache *cachep)
3239 struct memcg_register_cache_work *cw;
3241 cw = kmalloc(sizeof(*cw), GFP_NOWAIT);
3243 css_put(&memcg->css);
3248 cw->cachep = cachep;
3250 INIT_WORK(&cw->work, memcg_register_cache_func);
3251 schedule_work(&cw->work);
3254 static void memcg_schedule_register_cache(struct mem_cgroup *memcg,
3255 struct kmem_cache *cachep)
3258 * We need to stop accounting when we kmalloc, because if the
3259 * corresponding kmalloc cache is not yet created, the first allocation
3260 * in __memcg_schedule_register_cache will recurse.
3262 * However, it is better to enclose the whole function. Depending on
3263 * the debugging options enabled, INIT_WORK(), for instance, can
3264 * trigger an allocation. This too, will make us recurse. Because at
3265 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3266 * the safest choice is to do it like this, wrapping the whole function.
3268 memcg_stop_kmem_account();
3269 __memcg_schedule_register_cache(memcg, cachep);
3270 memcg_resume_kmem_account();
3273 int __memcg_charge_slab(struct kmem_cache *cachep, gfp_t gfp, int order)
3277 res = memcg_charge_kmem(cachep->memcg_params->memcg, gfp,
3278 PAGE_SIZE << order);
3280 atomic_add(1 << order, &cachep->memcg_params->nr_pages);
3284 void __memcg_uncharge_slab(struct kmem_cache *cachep, int order)
3286 memcg_uncharge_kmem(cachep->memcg_params->memcg, PAGE_SIZE << order);
3287 atomic_sub(1 << order, &cachep->memcg_params->nr_pages);
3291 * Return the kmem_cache we're supposed to use for a slab allocation.
3292 * We try to use the current memcg's version of the cache.
3294 * If the cache does not exist yet, if we are the first user of it,
3295 * we either create it immediately, if possible, or create it asynchronously
3297 * In the latter case, we will let the current allocation go through with
3298 * the original cache.
3300 * Can't be called in interrupt context or from kernel threads.
3301 * This function needs to be called with rcu_read_lock() held.
3303 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3306 struct mem_cgroup *memcg;
3307 struct kmem_cache *memcg_cachep;
3309 VM_BUG_ON(!cachep->memcg_params);
3310 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3312 if (!current->mm || current->memcg_kmem_skip_account)
3316 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3318 if (!memcg_can_account_kmem(memcg))
3321 memcg_cachep = cache_from_memcg_idx(cachep, memcg_cache_id(memcg));
3322 if (likely(memcg_cachep)) {
3323 cachep = memcg_cachep;
3327 /* The corresponding put will be done in the workqueue. */
3328 if (!css_tryget_online(&memcg->css))
3333 * If we are in a safe context (can wait, and not in interrupt
3334 * context), we could be be predictable and return right away.
3335 * This would guarantee that the allocation being performed
3336 * already belongs in the new cache.
3338 * However, there are some clashes that can arrive from locking.
3339 * For instance, because we acquire the slab_mutex while doing
3340 * memcg_create_kmem_cache, this means no further allocation
3341 * could happen with the slab_mutex held. So it's better to
3344 memcg_schedule_register_cache(memcg, cachep);
3352 * We need to verify if the allocation against current->mm->owner's memcg is
3353 * possible for the given order. But the page is not allocated yet, so we'll
3354 * need a further commit step to do the final arrangements.
3356 * It is possible for the task to switch cgroups in this mean time, so at
3357 * commit time, we can't rely on task conversion any longer. We'll then use
3358 * the handle argument to return to the caller which cgroup we should commit
3359 * against. We could also return the memcg directly and avoid the pointer
3360 * passing, but a boolean return value gives better semantics considering
3361 * the compiled-out case as well.
3363 * Returning true means the allocation is possible.
3366 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3368 struct mem_cgroup *memcg;
3374 * Disabling accounting is only relevant for some specific memcg
3375 * internal allocations. Therefore we would initially not have such
3376 * check here, since direct calls to the page allocator that are
3377 * accounted to kmemcg (alloc_kmem_pages and friends) only happen
3378 * outside memcg core. We are mostly concerned with cache allocations,
3379 * and by having this test at memcg_kmem_get_cache, we are already able
3380 * to relay the allocation to the root cache and bypass the memcg cache
3383 * There is one exception, though: the SLUB allocator does not create
3384 * large order caches, but rather service large kmallocs directly from
3385 * the page allocator. Therefore, the following sequence when backed by
3386 * the SLUB allocator:
3388 * memcg_stop_kmem_account();
3389 * kmalloc(<large_number>)
3390 * memcg_resume_kmem_account();
3392 * would effectively ignore the fact that we should skip accounting,
3393 * since it will drive us directly to this function without passing
3394 * through the cache selector memcg_kmem_get_cache. Such large
3395 * allocations are extremely rare but can happen, for instance, for the
3396 * cache arrays. We bring this test here.
3398 if (!current->mm || current->memcg_kmem_skip_account)
3401 memcg = get_mem_cgroup_from_mm(current->mm);
3403 if (!memcg_can_account_kmem(memcg)) {
3404 css_put(&memcg->css);
3408 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3412 css_put(&memcg->css);
3416 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3419 struct page_cgroup *pc;
3421 VM_BUG_ON(mem_cgroup_is_root(memcg));
3423 /* The page allocation failed. Revert */
3425 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3429 pc = lookup_page_cgroup(page);
3430 lock_page_cgroup(pc);
3431 pc->mem_cgroup = memcg;
3432 SetPageCgroupUsed(pc);
3433 unlock_page_cgroup(pc);
3436 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3438 struct mem_cgroup *memcg = NULL;
3439 struct page_cgroup *pc;
3442 pc = lookup_page_cgroup(page);
3444 * Fast unlocked return. Theoretically might have changed, have to
3445 * check again after locking.
3447 if (!PageCgroupUsed(pc))
3450 lock_page_cgroup(pc);
3451 if (PageCgroupUsed(pc)) {
3452 memcg = pc->mem_cgroup;
3453 ClearPageCgroupUsed(pc);
3455 unlock_page_cgroup(pc);
3458 * We trust that only if there is a memcg associated with the page, it
3459 * is a valid allocation
3464 VM_BUG_ON_PAGE(mem_cgroup_is_root(memcg), page);
3465 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3468 static inline void memcg_unregister_all_caches(struct mem_cgroup *memcg)
3471 #endif /* CONFIG_MEMCG_KMEM */
3473 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3475 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3477 * Because tail pages are not marked as "used", set it. We're under
3478 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3479 * charge/uncharge will be never happen and move_account() is done under
3480 * compound_lock(), so we don't have to take care of races.
3482 void mem_cgroup_split_huge_fixup(struct page *head)
3484 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3485 struct page_cgroup *pc;
3486 struct mem_cgroup *memcg;
3489 if (mem_cgroup_disabled())
3492 memcg = head_pc->mem_cgroup;
3493 for (i = 1; i < HPAGE_PMD_NR; i++) {
3495 pc->mem_cgroup = memcg;
3496 smp_wmb();/* see __commit_charge() */
3497 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3499 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3502 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3505 * mem_cgroup_move_account - move account of the page
3507 * @nr_pages: number of regular pages (>1 for huge pages)
3508 * @pc: page_cgroup of the page.
3509 * @from: mem_cgroup which the page is moved from.
3510 * @to: mem_cgroup which the page is moved to. @from != @to.
3512 * The caller must confirm following.
3513 * - page is not on LRU (isolate_page() is useful.)
3514 * - compound_lock is held when nr_pages > 1
3516 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3519 static int mem_cgroup_move_account(struct page *page,
3520 unsigned int nr_pages,
3521 struct page_cgroup *pc,
3522 struct mem_cgroup *from,
3523 struct mem_cgroup *to)
3525 unsigned long flags;
3527 bool anon = PageAnon(page);
3529 VM_BUG_ON(from == to);
3530 VM_BUG_ON_PAGE(PageLRU(page), page);
3532 * The page is isolated from LRU. So, collapse function
3533 * will not handle this page. But page splitting can happen.
3534 * Do this check under compound_page_lock(). The caller should
3538 if (nr_pages > 1 && !PageTransHuge(page))
3541 lock_page_cgroup(pc);
3544 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3547 move_lock_mem_cgroup(from, &flags);
3549 if (!anon && page_mapped(page)) {
3550 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3552 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED],
3556 if (PageWriteback(page)) {
3557 __this_cpu_sub(from->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3559 __this_cpu_add(to->stat->count[MEM_CGROUP_STAT_WRITEBACK],
3563 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3565 /* caller should have done css_get */
3566 pc->mem_cgroup = to;
3567 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3568 move_unlock_mem_cgroup(from, &flags);
3571 unlock_page_cgroup(pc);
3575 memcg_check_events(to, page);
3576 memcg_check_events(from, page);
3582 * mem_cgroup_move_parent - moves page to the parent group
3583 * @page: the page to move
3584 * @pc: page_cgroup of the page
3585 * @child: page's cgroup
3587 * move charges to its parent or the root cgroup if the group has no
3588 * parent (aka use_hierarchy==0).
3589 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3590 * mem_cgroup_move_account fails) the failure is always temporary and
3591 * it signals a race with a page removal/uncharge or migration. In the
3592 * first case the page is on the way out and it will vanish from the LRU
3593 * on the next attempt and the call should be retried later.
3594 * Isolation from the LRU fails only if page has been isolated from
3595 * the LRU since we looked at it and that usually means either global
3596 * reclaim or migration going on. The page will either get back to the
3598 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3599 * (!PageCgroupUsed) or moved to a different group. The page will
3600 * disappear in the next attempt.
3602 static int mem_cgroup_move_parent(struct page *page,
3603 struct page_cgroup *pc,
3604 struct mem_cgroup *child)
3606 struct mem_cgroup *parent;
3607 unsigned int nr_pages;
3608 unsigned long uninitialized_var(flags);
3611 VM_BUG_ON(mem_cgroup_is_root(child));
3614 if (!get_page_unless_zero(page))
3616 if (isolate_lru_page(page))
3619 nr_pages = hpage_nr_pages(page);
3621 parent = parent_mem_cgroup(child);
3623 * If no parent, move charges to root cgroup.
3626 parent = root_mem_cgroup;
3629 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3630 flags = compound_lock_irqsave(page);
3633 ret = mem_cgroup_move_account(page, nr_pages,
3636 __mem_cgroup_cancel_local_charge(child, nr_pages);
3639 compound_unlock_irqrestore(page, flags);
3640 putback_lru_page(page);
3647 int mem_cgroup_charge_anon(struct page *page,
3648 struct mm_struct *mm, gfp_t gfp_mask)
3650 unsigned int nr_pages = 1;
3651 struct mem_cgroup *memcg;
3654 if (mem_cgroup_disabled())
3657 VM_BUG_ON_PAGE(page_mapped(page), page);
3658 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3661 if (PageTransHuge(page)) {
3662 nr_pages <<= compound_order(page);
3663 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3665 * Never OOM-kill a process for a huge page. The
3666 * fault handler will fall back to regular pages.
3671 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, nr_pages, oom);
3674 __mem_cgroup_commit_charge(memcg, page, nr_pages,
3675 MEM_CGROUP_CHARGE_TYPE_ANON, false);
3680 * While swap-in, try_charge -> commit or cancel, the page is locked.
3681 * And when try_charge() successfully returns, one refcnt to memcg without
3682 * struct page_cgroup is acquired. This refcnt will be consumed by
3683 * "commit()" or removed by "cancel()"
3685 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3688 struct mem_cgroup **memcgp)
3690 struct mem_cgroup *memcg = NULL;
3691 struct page_cgroup *pc;
3694 pc = lookup_page_cgroup(page);
3696 * Every swap fault against a single page tries to charge the
3697 * page, bail as early as possible. shmem_unuse() encounters
3698 * already charged pages, too. The USED bit is protected by
3699 * the page lock, which serializes swap cache removal, which
3700 * in turn serializes uncharging.
3702 if (PageCgroupUsed(pc))
3704 if (do_swap_account)
3705 memcg = try_get_mem_cgroup_from_page(page);
3707 memcg = get_mem_cgroup_from_mm(mm);
3708 ret = mem_cgroup_try_charge(memcg, mask, 1, true);
3709 css_put(&memcg->css);
3711 memcg = root_mem_cgroup;
3719 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3720 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3722 if (mem_cgroup_disabled()) {
3727 * A racing thread's fault, or swapoff, may have already
3728 * updated the pte, and even removed page from swap cache: in
3729 * those cases unuse_pte()'s pte_same() test will fail; but
3730 * there's also a KSM case which does need to charge the page.
3732 if (!PageSwapCache(page)) {
3733 struct mem_cgroup *memcg;
3735 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3741 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3744 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3746 if (mem_cgroup_disabled())
3750 __mem_cgroup_cancel_charge(memcg, 1);
3754 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3755 enum charge_type ctype)
3757 if (mem_cgroup_disabled())
3762 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3764 * Now swap is on-memory. This means this page may be
3765 * counted both as mem and swap....double count.
3766 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3767 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3768 * may call delete_from_swap_cache() before reach here.
3770 if (do_swap_account && PageSwapCache(page)) {
3771 swp_entry_t ent = {.val = page_private(page)};
3772 mem_cgroup_uncharge_swap(ent);
3776 void mem_cgroup_commit_charge_swapin(struct page *page,
3777 struct mem_cgroup *memcg)
3779 __mem_cgroup_commit_charge_swapin(page, memcg,
3780 MEM_CGROUP_CHARGE_TYPE_ANON);
3783 int mem_cgroup_charge_file(struct page *page, struct mm_struct *mm,
3786 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3787 struct mem_cgroup *memcg;
3790 if (mem_cgroup_disabled())
3792 if (PageCompound(page))
3795 if (PageSwapCache(page)) { /* shmem */
3796 ret = __mem_cgroup_try_charge_swapin(mm, page,
3800 __mem_cgroup_commit_charge_swapin(page, memcg, type);
3804 memcg = mem_cgroup_try_charge_mm(mm, gfp_mask, 1, true);
3807 __mem_cgroup_commit_charge(memcg, page, 1, type, false);
3811 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3812 unsigned int nr_pages,
3813 const enum charge_type ctype)
3815 struct memcg_batch_info *batch = NULL;
3816 bool uncharge_memsw = true;
3818 /* If swapout, usage of swap doesn't decrease */
3819 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3820 uncharge_memsw = false;
3822 batch = ¤t->memcg_batch;
3824 * In usual, we do css_get() when we remember memcg pointer.
3825 * But in this case, we keep res->usage until end of a series of
3826 * uncharges. Then, it's ok to ignore memcg's refcnt.
3829 batch->memcg = memcg;
3831 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3832 * In those cases, all pages freed continuously can be expected to be in
3833 * the same cgroup and we have chance to coalesce uncharges.
3834 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3835 * because we want to do uncharge as soon as possible.
3838 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3839 goto direct_uncharge;
3842 goto direct_uncharge;
3845 * In typical case, batch->memcg == mem. This means we can
3846 * merge a series of uncharges to an uncharge of res_counter.
3847 * If not, we uncharge res_counter ony by one.
3849 if (batch->memcg != memcg)
3850 goto direct_uncharge;
3851 /* remember freed charge and uncharge it later */
3854 batch->memsw_nr_pages++;
3857 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3859 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3860 if (unlikely(batch->memcg != memcg))
3861 memcg_oom_recover(memcg);
3865 * uncharge if !page_mapped(page)
3867 static struct mem_cgroup *
3868 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3871 struct mem_cgroup *memcg = NULL;
3872 unsigned int nr_pages = 1;
3873 struct page_cgroup *pc;
3876 if (mem_cgroup_disabled())
3879 if (PageTransHuge(page)) {
3880 nr_pages <<= compound_order(page);
3881 VM_BUG_ON_PAGE(!PageTransHuge(page), page);
3884 * Check if our page_cgroup is valid
3886 pc = lookup_page_cgroup(page);
3887 if (unlikely(!PageCgroupUsed(pc)))
3890 lock_page_cgroup(pc);
3892 memcg = pc->mem_cgroup;
3894 if (!PageCgroupUsed(pc))
3897 anon = PageAnon(page);
3900 case MEM_CGROUP_CHARGE_TYPE_ANON:
3902 * Generally PageAnon tells if it's the anon statistics to be
3903 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3904 * used before page reached the stage of being marked PageAnon.
3908 case MEM_CGROUP_CHARGE_TYPE_DROP:
3909 /* See mem_cgroup_prepare_migration() */
3910 if (page_mapped(page))
3913 * Pages under migration may not be uncharged. But
3914 * end_migration() /must/ be the one uncharging the
3915 * unused post-migration page and so it has to call
3916 * here with the migration bit still set. See the
3917 * res_counter handling below.
3919 if (!end_migration && PageCgroupMigration(pc))
3922 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3923 if (!PageAnon(page)) { /* Shared memory */
3924 if (page->mapping && !page_is_file_cache(page))
3926 } else if (page_mapped(page)) /* Anon */
3933 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
3935 ClearPageCgroupUsed(pc);
3937 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3938 * freed from LRU. This is safe because uncharged page is expected not
3939 * to be reused (freed soon). Exception is SwapCache, it's handled by
3940 * special functions.
3943 unlock_page_cgroup(pc);
3945 * even after unlock, we have memcg->res.usage here and this memcg
3946 * will never be freed, so it's safe to call css_get().
3948 memcg_check_events(memcg, page);
3949 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3950 mem_cgroup_swap_statistics(memcg, true);
3951 css_get(&memcg->css);
3954 * Migration does not charge the res_counter for the
3955 * replacement page, so leave it alone when phasing out the
3956 * page that is unused after the migration.
3958 if (!end_migration && !mem_cgroup_is_root(memcg))
3959 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3964 unlock_page_cgroup(pc);
3968 void mem_cgroup_uncharge_page(struct page *page)
3971 if (page_mapped(page))
3973 VM_BUG_ON_PAGE(page->mapping && !PageAnon(page), page);
3975 * If the page is in swap cache, uncharge should be deferred
3976 * to the swap path, which also properly accounts swap usage
3977 * and handles memcg lifetime.
3979 * Note that this check is not stable and reclaim may add the
3980 * page to swap cache at any time after this. However, if the
3981 * page is not in swap cache by the time page->mapcount hits
3982 * 0, there won't be any page table references to the swap
3983 * slot, and reclaim will free it and not actually write the
3986 if (PageSwapCache(page))
3988 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
3991 void mem_cgroup_uncharge_cache_page(struct page *page)
3993 VM_BUG_ON_PAGE(page_mapped(page), page);
3994 VM_BUG_ON_PAGE(page->mapping, page);
3995 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
3999 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4000 * In that cases, pages are freed continuously and we can expect pages
4001 * are in the same memcg. All these calls itself limits the number of
4002 * pages freed at once, then uncharge_start/end() is called properly.
4003 * This may be called prural(2) times in a context,
4006 void mem_cgroup_uncharge_start(void)
4008 current->memcg_batch.do_batch++;
4009 /* We can do nest. */
4010 if (current->memcg_batch.do_batch == 1) {
4011 current->memcg_batch.memcg = NULL;
4012 current->memcg_batch.nr_pages = 0;
4013 current->memcg_batch.memsw_nr_pages = 0;
4017 void mem_cgroup_uncharge_end(void)
4019 struct memcg_batch_info *batch = ¤t->memcg_batch;
4021 if (!batch->do_batch)
4025 if (batch->do_batch) /* If stacked, do nothing. */
4031 * This "batch->memcg" is valid without any css_get/put etc...
4032 * bacause we hide charges behind us.
4034 if (batch->nr_pages)
4035 res_counter_uncharge(&batch->memcg->res,
4036 batch->nr_pages * PAGE_SIZE);
4037 if (batch->memsw_nr_pages)
4038 res_counter_uncharge(&batch->memcg->memsw,
4039 batch->memsw_nr_pages * PAGE_SIZE);
4040 memcg_oom_recover(batch->memcg);
4041 /* forget this pointer (for sanity check) */
4042 batch->memcg = NULL;
4047 * called after __delete_from_swap_cache() and drop "page" account.
4048 * memcg information is recorded to swap_cgroup of "ent"
4051 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4053 struct mem_cgroup *memcg;
4054 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4056 if (!swapout) /* this was a swap cache but the swap is unused ! */
4057 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4059 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4062 * record memcg information, if swapout && memcg != NULL,
4063 * css_get() was called in uncharge().
4065 if (do_swap_account && swapout && memcg)
4066 swap_cgroup_record(ent, mem_cgroup_id(memcg));
4070 #ifdef CONFIG_MEMCG_SWAP
4072 * called from swap_entry_free(). remove record in swap_cgroup and
4073 * uncharge "memsw" account.
4075 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4077 struct mem_cgroup *memcg;
4080 if (!do_swap_account)
4083 id = swap_cgroup_record(ent, 0);
4085 memcg = mem_cgroup_lookup(id);
4088 * We uncharge this because swap is freed. This memcg can
4089 * be obsolete one. We avoid calling css_tryget_online().
4091 if (!mem_cgroup_is_root(memcg))
4092 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4093 mem_cgroup_swap_statistics(memcg, false);
4094 css_put(&memcg->css);
4100 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4101 * @entry: swap entry to be moved
4102 * @from: mem_cgroup which the entry is moved from
4103 * @to: mem_cgroup which the entry is moved to
4105 * It succeeds only when the swap_cgroup's record for this entry is the same
4106 * as the mem_cgroup's id of @from.
4108 * Returns 0 on success, -EINVAL on failure.
4110 * The caller must have charged to @to, IOW, called res_counter_charge() about
4111 * both res and memsw, and called css_get().
4113 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4114 struct mem_cgroup *from, struct mem_cgroup *to)
4116 unsigned short old_id, new_id;
4118 old_id = mem_cgroup_id(from);
4119 new_id = mem_cgroup_id(to);
4121 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4122 mem_cgroup_swap_statistics(from, false);
4123 mem_cgroup_swap_statistics(to, true);
4125 * This function is only called from task migration context now.
4126 * It postpones res_counter and refcount handling till the end
4127 * of task migration(mem_cgroup_clear_mc()) for performance
4128 * improvement. But we cannot postpone css_get(to) because if
4129 * the process that has been moved to @to does swap-in, the
4130 * refcount of @to might be decreased to 0.
4132 * We are in attach() phase, so the cgroup is guaranteed to be
4133 * alive, so we can just call css_get().
4141 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4142 struct mem_cgroup *from, struct mem_cgroup *to)
4149 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4152 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4153 struct mem_cgroup **memcgp)
4155 struct mem_cgroup *memcg = NULL;
4156 unsigned int nr_pages = 1;
4157 struct page_cgroup *pc;
4158 enum charge_type ctype;
4162 if (mem_cgroup_disabled())
4165 if (PageTransHuge(page))
4166 nr_pages <<= compound_order(page);
4168 pc = lookup_page_cgroup(page);
4169 lock_page_cgroup(pc);
4170 if (PageCgroupUsed(pc)) {
4171 memcg = pc->mem_cgroup;
4172 css_get(&memcg->css);
4174 * At migrating an anonymous page, its mapcount goes down
4175 * to 0 and uncharge() will be called. But, even if it's fully
4176 * unmapped, migration may fail and this page has to be
4177 * charged again. We set MIGRATION flag here and delay uncharge
4178 * until end_migration() is called
4180 * Corner Case Thinking
4182 * When the old page was mapped as Anon and it's unmap-and-freed
4183 * while migration was ongoing.
4184 * If unmap finds the old page, uncharge() of it will be delayed
4185 * until end_migration(). If unmap finds a new page, it's
4186 * uncharged when it make mapcount to be 1->0. If unmap code
4187 * finds swap_migration_entry, the new page will not be mapped
4188 * and end_migration() will find it(mapcount==0).
4191 * When the old page was mapped but migraion fails, the kernel
4192 * remaps it. A charge for it is kept by MIGRATION flag even
4193 * if mapcount goes down to 0. We can do remap successfully
4194 * without charging it again.
4197 * The "old" page is under lock_page() until the end of
4198 * migration, so, the old page itself will not be swapped-out.
4199 * If the new page is swapped out before end_migraton, our
4200 * hook to usual swap-out path will catch the event.
4203 SetPageCgroupMigration(pc);
4205 unlock_page_cgroup(pc);
4207 * If the page is not charged at this point,
4215 * We charge new page before it's used/mapped. So, even if unlock_page()
4216 * is called before end_migration, we can catch all events on this new
4217 * page. In the case new page is migrated but not remapped, new page's
4218 * mapcount will be finally 0 and we call uncharge in end_migration().
4221 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4223 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4225 * The page is committed to the memcg, but it's not actually
4226 * charged to the res_counter since we plan on replacing the
4227 * old one and only one page is going to be left afterwards.
4229 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4232 /* remove redundant charge if migration failed*/
4233 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4234 struct page *oldpage, struct page *newpage, bool migration_ok)
4236 struct page *used, *unused;
4237 struct page_cgroup *pc;
4243 if (!migration_ok) {
4250 anon = PageAnon(used);
4251 __mem_cgroup_uncharge_common(unused,
4252 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4253 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4255 css_put(&memcg->css);
4257 * We disallowed uncharge of pages under migration because mapcount
4258 * of the page goes down to zero, temporarly.
4259 * Clear the flag and check the page should be charged.
4261 pc = lookup_page_cgroup(oldpage);
4262 lock_page_cgroup(pc);
4263 ClearPageCgroupMigration(pc);
4264 unlock_page_cgroup(pc);
4267 * If a page is a file cache, radix-tree replacement is very atomic
4268 * and we can skip this check. When it was an Anon page, its mapcount
4269 * goes down to 0. But because we added MIGRATION flage, it's not
4270 * uncharged yet. There are several case but page->mapcount check
4271 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4272 * check. (see prepare_charge() also)
4275 mem_cgroup_uncharge_page(used);
4279 * At replace page cache, newpage is not under any memcg but it's on
4280 * LRU. So, this function doesn't touch res_counter but handles LRU
4281 * in correct way. Both pages are locked so we cannot race with uncharge.
4283 void mem_cgroup_replace_page_cache(struct page *oldpage,
4284 struct page *newpage)
4286 struct mem_cgroup *memcg = NULL;
4287 struct page_cgroup *pc;
4288 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4290 if (mem_cgroup_disabled())
4293 pc = lookup_page_cgroup(oldpage);
4294 /* fix accounting on old pages */
4295 lock_page_cgroup(pc);
4296 if (PageCgroupUsed(pc)) {
4297 memcg = pc->mem_cgroup;
4298 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4299 ClearPageCgroupUsed(pc);
4301 unlock_page_cgroup(pc);
4304 * When called from shmem_replace_page(), in some cases the
4305 * oldpage has already been charged, and in some cases not.
4310 * Even if newpage->mapping was NULL before starting replacement,
4311 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4312 * LRU while we overwrite pc->mem_cgroup.
4314 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4317 #ifdef CONFIG_DEBUG_VM
4318 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4320 struct page_cgroup *pc;
4322 pc = lookup_page_cgroup(page);
4324 * Can be NULL while feeding pages into the page allocator for
4325 * the first time, i.e. during boot or memory hotplug;
4326 * or when mem_cgroup_disabled().
4328 if (likely(pc) && PageCgroupUsed(pc))
4333 bool mem_cgroup_bad_page_check(struct page *page)
4335 if (mem_cgroup_disabled())
4338 return lookup_page_cgroup_used(page) != NULL;
4341 void mem_cgroup_print_bad_page(struct page *page)
4343 struct page_cgroup *pc;
4345 pc = lookup_page_cgroup_used(page);
4347 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4348 pc, pc->flags, pc->mem_cgroup);
4353 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4354 unsigned long long val)
4357 u64 memswlimit, memlimit;
4359 int children = mem_cgroup_count_children(memcg);
4360 u64 curusage, oldusage;
4364 * For keeping hierarchical_reclaim simple, how long we should retry
4365 * is depends on callers. We set our retry-count to be function
4366 * of # of children which we should visit in this loop.
4368 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4370 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4373 while (retry_count) {
4374 if (signal_pending(current)) {
4379 * Rather than hide all in some function, I do this in
4380 * open coded manner. You see what this really does.
4381 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4383 mutex_lock(&set_limit_mutex);
4384 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4385 if (memswlimit < val) {
4387 mutex_unlock(&set_limit_mutex);
4391 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4395 ret = res_counter_set_limit(&memcg->res, val);
4397 if (memswlimit == val)
4398 memcg->memsw_is_minimum = true;
4400 memcg->memsw_is_minimum = false;
4402 mutex_unlock(&set_limit_mutex);
4407 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4408 MEM_CGROUP_RECLAIM_SHRINK);
4409 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4410 /* Usage is reduced ? */
4411 if (curusage >= oldusage)
4414 oldusage = curusage;
4416 if (!ret && enlarge)
4417 memcg_oom_recover(memcg);
4422 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4423 unsigned long long val)
4426 u64 memlimit, memswlimit, oldusage, curusage;
4427 int children = mem_cgroup_count_children(memcg);
4431 /* see mem_cgroup_resize_res_limit */
4432 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4433 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4434 while (retry_count) {
4435 if (signal_pending(current)) {
4440 * Rather than hide all in some function, I do this in
4441 * open coded manner. You see what this really does.
4442 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4444 mutex_lock(&set_limit_mutex);
4445 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4446 if (memlimit > val) {
4448 mutex_unlock(&set_limit_mutex);
4451 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4452 if (memswlimit < val)
4454 ret = res_counter_set_limit(&memcg->memsw, val);
4456 if (memlimit == val)
4457 memcg->memsw_is_minimum = true;
4459 memcg->memsw_is_minimum = false;
4461 mutex_unlock(&set_limit_mutex);
4466 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4467 MEM_CGROUP_RECLAIM_NOSWAP |
4468 MEM_CGROUP_RECLAIM_SHRINK);
4469 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4470 /* Usage is reduced ? */
4471 if (curusage >= oldusage)
4474 oldusage = curusage;
4476 if (!ret && enlarge)
4477 memcg_oom_recover(memcg);
4481 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4483 unsigned long *total_scanned)
4485 unsigned long nr_reclaimed = 0;
4486 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4487 unsigned long reclaimed;
4489 struct mem_cgroup_tree_per_zone *mctz;
4490 unsigned long long excess;
4491 unsigned long nr_scanned;
4496 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4498 * This loop can run a while, specially if mem_cgroup's continuously
4499 * keep exceeding their soft limit and putting the system under
4506 mz = mem_cgroup_largest_soft_limit_node(mctz);
4511 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4512 gfp_mask, &nr_scanned);
4513 nr_reclaimed += reclaimed;
4514 *total_scanned += nr_scanned;
4515 spin_lock(&mctz->lock);
4518 * If we failed to reclaim anything from this memory cgroup
4519 * it is time to move on to the next cgroup
4525 * Loop until we find yet another one.
4527 * By the time we get the soft_limit lock
4528 * again, someone might have aded the
4529 * group back on the RB tree. Iterate to
4530 * make sure we get a different mem.
4531 * mem_cgroup_largest_soft_limit_node returns
4532 * NULL if no other cgroup is present on
4536 __mem_cgroup_largest_soft_limit_node(mctz);
4538 css_put(&next_mz->memcg->css);
4539 else /* next_mz == NULL or other memcg */
4543 __mem_cgroup_remove_exceeded(mz, mctz);
4544 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4546 * One school of thought says that we should not add
4547 * back the node to the tree if reclaim returns 0.
4548 * But our reclaim could return 0, simply because due
4549 * to priority we are exposing a smaller subset of
4550 * memory to reclaim from. Consider this as a longer
4553 /* If excess == 0, no tree ops */
4554 __mem_cgroup_insert_exceeded(mz, mctz, excess);
4555 spin_unlock(&mctz->lock);
4556 css_put(&mz->memcg->css);
4559 * Could not reclaim anything and there are no more
4560 * mem cgroups to try or we seem to be looping without
4561 * reclaiming anything.
4563 if (!nr_reclaimed &&
4565 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4567 } while (!nr_reclaimed);
4569 css_put(&next_mz->memcg->css);
4570 return nr_reclaimed;
4574 * mem_cgroup_force_empty_list - clears LRU of a group
4575 * @memcg: group to clear
4578 * @lru: lru to to clear
4580 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4581 * reclaim the pages page themselves - pages are moved to the parent (or root)
4584 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4585 int node, int zid, enum lru_list lru)
4587 struct lruvec *lruvec;
4588 unsigned long flags;
4589 struct list_head *list;
4593 zone = &NODE_DATA(node)->node_zones[zid];
4594 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4595 list = &lruvec->lists[lru];
4599 struct page_cgroup *pc;
4602 spin_lock_irqsave(&zone->lru_lock, flags);
4603 if (list_empty(list)) {
4604 spin_unlock_irqrestore(&zone->lru_lock, flags);
4607 page = list_entry(list->prev, struct page, lru);
4609 list_move(&page->lru, list);
4611 spin_unlock_irqrestore(&zone->lru_lock, flags);
4614 spin_unlock_irqrestore(&zone->lru_lock, flags);
4616 pc = lookup_page_cgroup(page);
4618 if (mem_cgroup_move_parent(page, pc, memcg)) {
4619 /* found lock contention or "pc" is obsolete. */
4624 } while (!list_empty(list));
4628 * make mem_cgroup's charge to be 0 if there is no task by moving
4629 * all the charges and pages to the parent.
4630 * This enables deleting this mem_cgroup.
4632 * Caller is responsible for holding css reference on the memcg.
4634 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4640 /* This is for making all *used* pages to be on LRU. */
4641 lru_add_drain_all();
4642 drain_all_stock_sync(memcg);
4643 mem_cgroup_start_move(memcg);
4644 for_each_node_state(node, N_MEMORY) {
4645 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4648 mem_cgroup_force_empty_list(memcg,
4653 mem_cgroup_end_move(memcg);
4654 memcg_oom_recover(memcg);
4658 * Kernel memory may not necessarily be trackable to a specific
4659 * process. So they are not migrated, and therefore we can't
4660 * expect their value to drop to 0 here.
4661 * Having res filled up with kmem only is enough.
4663 * This is a safety check because mem_cgroup_force_empty_list
4664 * could have raced with mem_cgroup_replace_page_cache callers
4665 * so the lru seemed empty but the page could have been added
4666 * right after the check. RES_USAGE should be safe as we always
4667 * charge before adding to the LRU.
4669 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4670 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4671 } while (usage > 0);
4675 * Test whether @memcg has children, dead or alive. Note that this
4676 * function doesn't care whether @memcg has use_hierarchy enabled and
4677 * returns %true if there are child csses according to the cgroup
4678 * hierarchy. Testing use_hierarchy is the caller's responsiblity.
4680 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4685 * The lock does not prevent addition or deletion of children, but
4686 * it prevents a new child from being initialized based on this
4687 * parent in css_online(), so it's enough to decide whether
4688 * hierarchically inherited attributes can still be changed or not.
4690 lockdep_assert_held(&memcg_create_mutex);
4693 ret = css_next_child(NULL, &memcg->css);
4699 * Reclaims as many pages from the given memcg as possible and moves
4700 * the rest to the parent.
4702 * Caller is responsible for holding css reference for memcg.
4704 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4706 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4708 /* we call try-to-free pages for make this cgroup empty */
4709 lru_add_drain_all();
4710 /* try to free all pages in this cgroup */
4711 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4714 if (signal_pending(current))
4717 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4721 /* maybe some writeback is necessary */
4722 congestion_wait(BLK_RW_ASYNC, HZ/10);
4730 static ssize_t mem_cgroup_force_empty_write(struct kernfs_open_file *of,
4731 char *buf, size_t nbytes,
4734 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4736 if (mem_cgroup_is_root(memcg))
4738 return mem_cgroup_force_empty(memcg) ?: nbytes;
4741 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4744 return mem_cgroup_from_css(css)->use_hierarchy;
4747 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4748 struct cftype *cft, u64 val)
4751 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4752 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(memcg->css.parent);
4754 mutex_lock(&memcg_create_mutex);
4756 if (memcg->use_hierarchy == val)
4760 * If parent's use_hierarchy is set, we can't make any modifications
4761 * in the child subtrees. If it is unset, then the change can
4762 * occur, provided the current cgroup has no children.
4764 * For the root cgroup, parent_mem is NULL, we allow value to be
4765 * set if there are no children.
4767 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4768 (val == 1 || val == 0)) {
4769 if (!memcg_has_children(memcg))
4770 memcg->use_hierarchy = val;
4777 mutex_unlock(&memcg_create_mutex);
4783 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4784 enum mem_cgroup_stat_index idx)
4786 struct mem_cgroup *iter;
4789 /* Per-cpu values can be negative, use a signed accumulator */
4790 for_each_mem_cgroup_tree(iter, memcg)
4791 val += mem_cgroup_read_stat(iter, idx);
4793 if (val < 0) /* race ? */
4798 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4802 if (!mem_cgroup_is_root(memcg)) {
4804 return res_counter_read_u64(&memcg->res, RES_USAGE);
4806 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4810 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4811 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4813 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4814 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4817 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4819 return val << PAGE_SHIFT;
4822 static u64 mem_cgroup_read_u64(struct cgroup_subsys_state *css,
4825 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4830 type = MEMFILE_TYPE(cft->private);
4831 name = MEMFILE_ATTR(cft->private);
4835 if (name == RES_USAGE)
4836 val = mem_cgroup_usage(memcg, false);
4838 val = res_counter_read_u64(&memcg->res, name);
4841 if (name == RES_USAGE)
4842 val = mem_cgroup_usage(memcg, true);
4844 val = res_counter_read_u64(&memcg->memsw, name);
4847 val = res_counter_read_u64(&memcg->kmem, name);
4856 #ifdef CONFIG_MEMCG_KMEM
4857 /* should be called with activate_kmem_mutex held */
4858 static int __memcg_activate_kmem(struct mem_cgroup *memcg,
4859 unsigned long long limit)
4864 if (memcg_kmem_is_active(memcg))
4868 * We are going to allocate memory for data shared by all memory
4869 * cgroups so let's stop accounting here.
4871 memcg_stop_kmem_account();
4874 * For simplicity, we won't allow this to be disabled. It also can't
4875 * be changed if the cgroup has children already, or if tasks had
4878 * If tasks join before we set the limit, a person looking at
4879 * kmem.usage_in_bytes will have no way to determine when it took
4880 * place, which makes the value quite meaningless.
4882 * After it first became limited, changes in the value of the limit are
4883 * of course permitted.
4885 mutex_lock(&memcg_create_mutex);
4886 if (cgroup_has_tasks(memcg->css.cgroup) ||
4887 (memcg->use_hierarchy && memcg_has_children(memcg)))
4889 mutex_unlock(&memcg_create_mutex);
4893 memcg_id = ida_simple_get(&kmem_limited_groups,
4894 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
4901 * Make sure we have enough space for this cgroup in each root cache's
4904 mutex_lock(&memcg_slab_mutex);
4905 err = memcg_update_all_caches(memcg_id + 1);
4906 mutex_unlock(&memcg_slab_mutex);
4910 memcg->kmemcg_id = memcg_id;
4911 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
4914 * We couldn't have accounted to this cgroup, because it hasn't got the
4915 * active bit set yet, so this should succeed.
4917 err = res_counter_set_limit(&memcg->kmem, limit);
4920 static_key_slow_inc(&memcg_kmem_enabled_key);
4922 * Setting the active bit after enabling static branching will
4923 * guarantee no one starts accounting before all call sites are
4926 memcg_kmem_set_active(memcg);
4928 memcg_resume_kmem_account();
4932 ida_simple_remove(&kmem_limited_groups, memcg_id);
4936 static int memcg_activate_kmem(struct mem_cgroup *memcg,
4937 unsigned long long limit)
4941 mutex_lock(&activate_kmem_mutex);
4942 ret = __memcg_activate_kmem(memcg, limit);
4943 mutex_unlock(&activate_kmem_mutex);
4947 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4948 unsigned long long val)
4952 if (!memcg_kmem_is_active(memcg))
4953 ret = memcg_activate_kmem(memcg, val);
4955 ret = res_counter_set_limit(&memcg->kmem, val);
4959 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
4962 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4967 mutex_lock(&activate_kmem_mutex);
4969 * If the parent cgroup is not kmem-active now, it cannot be activated
4970 * after this point, because it has at least one child already.
4972 if (memcg_kmem_is_active(parent))
4973 ret = __memcg_activate_kmem(memcg, RES_COUNTER_MAX);
4974 mutex_unlock(&activate_kmem_mutex);
4978 static int memcg_update_kmem_limit(struct mem_cgroup *memcg,
4979 unsigned long long val)
4983 #endif /* CONFIG_MEMCG_KMEM */
4986 * The user of this function is...
4989 static ssize_t mem_cgroup_write(struct kernfs_open_file *of,
4990 char *buf, size_t nbytes, loff_t off)
4992 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
4995 unsigned long long val;
4998 buf = strstrip(buf);
4999 type = MEMFILE_TYPE(of_cft(of)->private);
5000 name = MEMFILE_ATTR(of_cft(of)->private);
5004 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5008 /* This function does all necessary parse...reuse it */
5009 ret = res_counter_memparse_write_strategy(buf, &val);
5013 ret = mem_cgroup_resize_limit(memcg, val);
5014 else if (type == _MEMSWAP)
5015 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5016 else if (type == _KMEM)
5017 ret = memcg_update_kmem_limit(memcg, val);
5021 case RES_SOFT_LIMIT:
5022 ret = res_counter_memparse_write_strategy(buf, &val);
5026 * For memsw, soft limits are hard to implement in terms
5027 * of semantics, for now, we support soft limits for
5028 * control without swap
5031 ret = res_counter_set_soft_limit(&memcg->res, val);
5036 ret = -EINVAL; /* should be BUG() ? */
5039 return ret ?: nbytes;
5042 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5043 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5045 unsigned long long min_limit, min_memsw_limit, tmp;
5047 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5048 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5049 if (!memcg->use_hierarchy)
5052 while (memcg->css.parent) {
5053 memcg = mem_cgroup_from_css(memcg->css.parent);
5054 if (!memcg->use_hierarchy)
5056 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5057 min_limit = min(min_limit, tmp);
5058 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5059 min_memsw_limit = min(min_memsw_limit, tmp);
5062 *mem_limit = min_limit;
5063 *memsw_limit = min_memsw_limit;
5066 static ssize_t mem_cgroup_reset(struct kernfs_open_file *of, char *buf,
5067 size_t nbytes, loff_t off)
5069 struct mem_cgroup *memcg = mem_cgroup_from_css(of_css(of));
5073 type = MEMFILE_TYPE(of_cft(of)->private);
5074 name = MEMFILE_ATTR(of_cft(of)->private);
5079 res_counter_reset_max(&memcg->res);
5080 else if (type == _MEMSWAP)
5081 res_counter_reset_max(&memcg->memsw);
5082 else if (type == _KMEM)
5083 res_counter_reset_max(&memcg->kmem);
5089 res_counter_reset_failcnt(&memcg->res);
5090 else if (type == _MEMSWAP)
5091 res_counter_reset_failcnt(&memcg->memsw);
5092 else if (type == _KMEM)
5093 res_counter_reset_failcnt(&memcg->kmem);
5102 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5105 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5109 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5110 struct cftype *cft, u64 val)
5112 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5114 if (val >= (1 << NR_MOVE_TYPE))
5118 * No kind of locking is needed in here, because ->can_attach() will
5119 * check this value once in the beginning of the process, and then carry
5120 * on with stale data. This means that changes to this value will only
5121 * affect task migrations starting after the change.
5123 memcg->move_charge_at_immigrate = val;
5127 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5128 struct cftype *cft, u64 val)
5135 static int memcg_numa_stat_show(struct seq_file *m, void *v)
5139 unsigned int lru_mask;
5142 static const struct numa_stat stats[] = {
5143 { "total", LRU_ALL },
5144 { "file", LRU_ALL_FILE },
5145 { "anon", LRU_ALL_ANON },
5146 { "unevictable", BIT(LRU_UNEVICTABLE) },
5148 const struct numa_stat *stat;
5151 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5153 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5154 nr = mem_cgroup_nr_lru_pages(memcg, stat->lru_mask);
5155 seq_printf(m, "%s=%lu", stat->name, nr);
5156 for_each_node_state(nid, N_MEMORY) {
5157 nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5159 seq_printf(m, " N%d=%lu", nid, nr);
5164 for (stat = stats; stat < stats + ARRAY_SIZE(stats); stat++) {
5165 struct mem_cgroup *iter;
5168 for_each_mem_cgroup_tree(iter, memcg)
5169 nr += mem_cgroup_nr_lru_pages(iter, stat->lru_mask);
5170 seq_printf(m, "hierarchical_%s=%lu", stat->name, nr);
5171 for_each_node_state(nid, N_MEMORY) {
5173 for_each_mem_cgroup_tree(iter, memcg)
5174 nr += mem_cgroup_node_nr_lru_pages(
5175 iter, nid, stat->lru_mask);
5176 seq_printf(m, " N%d=%lu", nid, nr);
5183 #endif /* CONFIG_NUMA */
5185 static inline void mem_cgroup_lru_names_not_uptodate(void)
5187 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5190 static int memcg_stat_show(struct seq_file *m, void *v)
5192 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
5193 struct mem_cgroup *mi;
5196 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5197 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5199 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5200 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5203 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5204 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5205 mem_cgroup_read_events(memcg, i));
5207 for (i = 0; i < NR_LRU_LISTS; i++)
5208 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5209 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5211 /* Hierarchical information */
5213 unsigned long long limit, memsw_limit;
5214 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5215 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5216 if (do_swap_account)
5217 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5221 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5224 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5226 for_each_mem_cgroup_tree(mi, memcg)
5227 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5228 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5231 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5232 unsigned long long val = 0;
5234 for_each_mem_cgroup_tree(mi, memcg)
5235 val += mem_cgroup_read_events(mi, i);
5236 seq_printf(m, "total_%s %llu\n",
5237 mem_cgroup_events_names[i], val);
5240 for (i = 0; i < NR_LRU_LISTS; i++) {
5241 unsigned long long val = 0;
5243 for_each_mem_cgroup_tree(mi, memcg)
5244 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5245 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5248 #ifdef CONFIG_DEBUG_VM
5251 struct mem_cgroup_per_zone *mz;
5252 struct zone_reclaim_stat *rstat;
5253 unsigned long recent_rotated[2] = {0, 0};
5254 unsigned long recent_scanned[2] = {0, 0};
5256 for_each_online_node(nid)
5257 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5258 mz = &memcg->nodeinfo[nid]->zoneinfo[zid];
5259 rstat = &mz->lruvec.reclaim_stat;
5261 recent_rotated[0] += rstat->recent_rotated[0];
5262 recent_rotated[1] += rstat->recent_rotated[1];
5263 recent_scanned[0] += rstat->recent_scanned[0];
5264 recent_scanned[1] += rstat->recent_scanned[1];
5266 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5267 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5268 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5269 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5276 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5279 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5281 return mem_cgroup_swappiness(memcg);
5284 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5285 struct cftype *cft, u64 val)
5287 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5293 memcg->swappiness = val;
5295 vm_swappiness = val;
5300 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5302 struct mem_cgroup_threshold_ary *t;
5308 t = rcu_dereference(memcg->thresholds.primary);
5310 t = rcu_dereference(memcg->memsw_thresholds.primary);
5315 usage = mem_cgroup_usage(memcg, swap);
5318 * current_threshold points to threshold just below or equal to usage.
5319 * If it's not true, a threshold was crossed after last
5320 * call of __mem_cgroup_threshold().
5322 i = t->current_threshold;
5325 * Iterate backward over array of thresholds starting from
5326 * current_threshold and check if a threshold is crossed.
5327 * If none of thresholds below usage is crossed, we read
5328 * only one element of the array here.
5330 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5331 eventfd_signal(t->entries[i].eventfd, 1);
5333 /* i = current_threshold + 1 */
5337 * Iterate forward over array of thresholds starting from
5338 * current_threshold+1 and check if a threshold is crossed.
5339 * If none of thresholds above usage is crossed, we read
5340 * only one element of the array here.
5342 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5343 eventfd_signal(t->entries[i].eventfd, 1);
5345 /* Update current_threshold */
5346 t->current_threshold = i - 1;
5351 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5354 __mem_cgroup_threshold(memcg, false);
5355 if (do_swap_account)
5356 __mem_cgroup_threshold(memcg, true);
5358 memcg = parent_mem_cgroup(memcg);
5362 static int compare_thresholds(const void *a, const void *b)
5364 const struct mem_cgroup_threshold *_a = a;
5365 const struct mem_cgroup_threshold *_b = b;
5367 if (_a->threshold > _b->threshold)
5370 if (_a->threshold < _b->threshold)
5376 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5378 struct mem_cgroup_eventfd_list *ev;
5380 spin_lock(&memcg_oom_lock);
5382 list_for_each_entry(ev, &memcg->oom_notify, list)
5383 eventfd_signal(ev->eventfd, 1);
5385 spin_unlock(&memcg_oom_lock);
5389 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5391 struct mem_cgroup *iter;
5393 for_each_mem_cgroup_tree(iter, memcg)
5394 mem_cgroup_oom_notify_cb(iter);
5397 static int __mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5398 struct eventfd_ctx *eventfd, const char *args, enum res_type type)
5400 struct mem_cgroup_thresholds *thresholds;
5401 struct mem_cgroup_threshold_ary *new;
5402 u64 threshold, usage;
5405 ret = res_counter_memparse_write_strategy(args, &threshold);
5409 mutex_lock(&memcg->thresholds_lock);
5412 thresholds = &memcg->thresholds;
5413 else if (type == _MEMSWAP)
5414 thresholds = &memcg->memsw_thresholds;
5418 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5420 /* Check if a threshold crossed before adding a new one */
5421 if (thresholds->primary)
5422 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5424 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5426 /* Allocate memory for new array of thresholds */
5427 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5435 /* Copy thresholds (if any) to new array */
5436 if (thresholds->primary) {
5437 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5438 sizeof(struct mem_cgroup_threshold));
5441 /* Add new threshold */
5442 new->entries[size - 1].eventfd = eventfd;
5443 new->entries[size - 1].threshold = threshold;
5445 /* Sort thresholds. Registering of new threshold isn't time-critical */
5446 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5447 compare_thresholds, NULL);
5449 /* Find current threshold */
5450 new->current_threshold = -1;
5451 for (i = 0; i < size; i++) {
5452 if (new->entries[i].threshold <= usage) {
5454 * new->current_threshold will not be used until
5455 * rcu_assign_pointer(), so it's safe to increment
5458 ++new->current_threshold;
5463 /* Free old spare buffer and save old primary buffer as spare */
5464 kfree(thresholds->spare);
5465 thresholds->spare = thresholds->primary;
5467 rcu_assign_pointer(thresholds->primary, new);
5469 /* To be sure that nobody uses thresholds */
5473 mutex_unlock(&memcg->thresholds_lock);
5478 static int mem_cgroup_usage_register_event(struct mem_cgroup *memcg,
5479 struct eventfd_ctx *eventfd, const char *args)
5481 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEM);
5484 static int memsw_cgroup_usage_register_event(struct mem_cgroup *memcg,
5485 struct eventfd_ctx *eventfd, const char *args)
5487 return __mem_cgroup_usage_register_event(memcg, eventfd, args, _MEMSWAP);
5490 static void __mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5491 struct eventfd_ctx *eventfd, enum res_type type)
5493 struct mem_cgroup_thresholds *thresholds;
5494 struct mem_cgroup_threshold_ary *new;
5498 mutex_lock(&memcg->thresholds_lock);
5500 thresholds = &memcg->thresholds;
5501 else if (type == _MEMSWAP)
5502 thresholds = &memcg->memsw_thresholds;
5506 if (!thresholds->primary)
5509 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5511 /* Check if a threshold crossed before removing */
5512 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5514 /* Calculate new number of threshold */
5516 for (i = 0; i < thresholds->primary->size; i++) {
5517 if (thresholds->primary->entries[i].eventfd != eventfd)
5521 new = thresholds->spare;
5523 /* Set thresholds array to NULL if we don't have thresholds */
5532 /* Copy thresholds and find current threshold */
5533 new->current_threshold = -1;
5534 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5535 if (thresholds->primary->entries[i].eventfd == eventfd)
5538 new->entries[j] = thresholds->primary->entries[i];
5539 if (new->entries[j].threshold <= usage) {
5541 * new->current_threshold will not be used
5542 * until rcu_assign_pointer(), so it's safe to increment
5545 ++new->current_threshold;
5551 /* Swap primary and spare array */
5552 thresholds->spare = thresholds->primary;
5553 /* If all events are unregistered, free the spare array */
5555 kfree(thresholds->spare);
5556 thresholds->spare = NULL;
5559 rcu_assign_pointer(thresholds->primary, new);
5561 /* To be sure that nobody uses thresholds */
5564 mutex_unlock(&memcg->thresholds_lock);
5567 static void mem_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5568 struct eventfd_ctx *eventfd)
5570 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEM);
5573 static void memsw_cgroup_usage_unregister_event(struct mem_cgroup *memcg,
5574 struct eventfd_ctx *eventfd)
5576 return __mem_cgroup_usage_unregister_event(memcg, eventfd, _MEMSWAP);
5579 static int mem_cgroup_oom_register_event(struct mem_cgroup *memcg,
5580 struct eventfd_ctx *eventfd, const char *args)
5582 struct mem_cgroup_eventfd_list *event;
5584 event = kmalloc(sizeof(*event), GFP_KERNEL);
5588 spin_lock(&memcg_oom_lock);
5590 event->eventfd = eventfd;
5591 list_add(&event->list, &memcg->oom_notify);
5593 /* already in OOM ? */
5594 if (atomic_read(&memcg->under_oom))
5595 eventfd_signal(eventfd, 1);
5596 spin_unlock(&memcg_oom_lock);
5601 static void mem_cgroup_oom_unregister_event(struct mem_cgroup *memcg,
5602 struct eventfd_ctx *eventfd)
5604 struct mem_cgroup_eventfd_list *ev, *tmp;
5606 spin_lock(&memcg_oom_lock);
5608 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5609 if (ev->eventfd == eventfd) {
5610 list_del(&ev->list);
5615 spin_unlock(&memcg_oom_lock);
5618 static int mem_cgroup_oom_control_read(struct seq_file *sf, void *v)
5620 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(sf));
5622 seq_printf(sf, "oom_kill_disable %d\n", memcg->oom_kill_disable);
5623 seq_printf(sf, "under_oom %d\n", (bool)atomic_read(&memcg->under_oom));
5627 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5628 struct cftype *cft, u64 val)
5630 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5632 /* cannot set to root cgroup and only 0 and 1 are allowed */
5633 if (!css->parent || !((val == 0) || (val == 1)))
5636 memcg->oom_kill_disable = val;
5638 memcg_oom_recover(memcg);
5643 #ifdef CONFIG_MEMCG_KMEM
5644 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5648 memcg->kmemcg_id = -1;
5649 ret = memcg_propagate_kmem(memcg);
5653 return mem_cgroup_sockets_init(memcg, ss);
5656 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5658 mem_cgroup_sockets_destroy(memcg);
5661 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5663 if (!memcg_kmem_is_active(memcg))
5667 * kmem charges can outlive the cgroup. In the case of slab
5668 * pages, for instance, a page contain objects from various
5669 * processes. As we prevent from taking a reference for every
5670 * such allocation we have to be careful when doing uncharge
5671 * (see memcg_uncharge_kmem) and here during offlining.
5673 * The idea is that that only the _last_ uncharge which sees
5674 * the dead memcg will drop the last reference. An additional
5675 * reference is taken here before the group is marked dead
5676 * which is then paired with css_put during uncharge resp. here.
5678 * Although this might sound strange as this path is called from
5679 * css_offline() when the referencemight have dropped down to 0 and
5680 * shouldn't be incremented anymore (css_tryget_online() would
5681 * fail) we do not have other options because of the kmem
5682 * allocations lifetime.
5684 css_get(&memcg->css);
5686 memcg_kmem_mark_dead(memcg);
5688 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5691 if (memcg_kmem_test_and_clear_dead(memcg))
5692 css_put(&memcg->css);
5695 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5700 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5704 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5710 * DO NOT USE IN NEW FILES.
5712 * "cgroup.event_control" implementation.
5714 * This is way over-engineered. It tries to support fully configurable
5715 * events for each user. Such level of flexibility is completely
5716 * unnecessary especially in the light of the planned unified hierarchy.
5718 * Please deprecate this and replace with something simpler if at all
5723 * Unregister event and free resources.
5725 * Gets called from workqueue.
5727 static void memcg_event_remove(struct work_struct *work)
5729 struct mem_cgroup_event *event =
5730 container_of(work, struct mem_cgroup_event, remove);
5731 struct mem_cgroup *memcg = event->memcg;
5733 remove_wait_queue(event->wqh, &event->wait);
5735 event->unregister_event(memcg, event->eventfd);
5737 /* Notify userspace the event is going away. */
5738 eventfd_signal(event->eventfd, 1);
5740 eventfd_ctx_put(event->eventfd);
5742 css_put(&memcg->css);
5746 * Gets called on POLLHUP on eventfd when user closes it.
5748 * Called with wqh->lock held and interrupts disabled.
5750 static int memcg_event_wake(wait_queue_t *wait, unsigned mode,
5751 int sync, void *key)
5753 struct mem_cgroup_event *event =
5754 container_of(wait, struct mem_cgroup_event, wait);
5755 struct mem_cgroup *memcg = event->memcg;
5756 unsigned long flags = (unsigned long)key;
5758 if (flags & POLLHUP) {
5760 * If the event has been detached at cgroup removal, we
5761 * can simply return knowing the other side will cleanup
5764 * We can't race against event freeing since the other
5765 * side will require wqh->lock via remove_wait_queue(),
5768 spin_lock(&memcg->event_list_lock);
5769 if (!list_empty(&event->list)) {
5770 list_del_init(&event->list);
5772 * We are in atomic context, but cgroup_event_remove()
5773 * may sleep, so we have to call it in workqueue.
5775 schedule_work(&event->remove);
5777 spin_unlock(&memcg->event_list_lock);
5783 static void memcg_event_ptable_queue_proc(struct file *file,
5784 wait_queue_head_t *wqh, poll_table *pt)
5786 struct mem_cgroup_event *event =
5787 container_of(pt, struct mem_cgroup_event, pt);
5790 add_wait_queue(wqh, &event->wait);
5794 * DO NOT USE IN NEW FILES.
5796 * Parse input and register new cgroup event handler.
5798 * Input must be in format '<event_fd> <control_fd> <args>'.
5799 * Interpretation of args is defined by control file implementation.
5801 static ssize_t memcg_write_event_control(struct kernfs_open_file *of,
5802 char *buf, size_t nbytes, loff_t off)
5804 struct cgroup_subsys_state *css = of_css(of);
5805 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5806 struct mem_cgroup_event *event;
5807 struct cgroup_subsys_state *cfile_css;
5808 unsigned int efd, cfd;
5815 buf = strstrip(buf);
5817 efd = simple_strtoul(buf, &endp, 10);
5822 cfd = simple_strtoul(buf, &endp, 10);
5823 if ((*endp != ' ') && (*endp != '\0'))
5827 event = kzalloc(sizeof(*event), GFP_KERNEL);
5831 event->memcg = memcg;
5832 INIT_LIST_HEAD(&event->list);
5833 init_poll_funcptr(&event->pt, memcg_event_ptable_queue_proc);
5834 init_waitqueue_func_entry(&event->wait, memcg_event_wake);
5835 INIT_WORK(&event->remove, memcg_event_remove);
5843 event->eventfd = eventfd_ctx_fileget(efile.file);
5844 if (IS_ERR(event->eventfd)) {
5845 ret = PTR_ERR(event->eventfd);
5852 goto out_put_eventfd;
5855 /* the process need read permission on control file */
5856 /* AV: shouldn't we check that it's been opened for read instead? */
5857 ret = inode_permission(file_inode(cfile.file), MAY_READ);
5862 * Determine the event callbacks and set them in @event. This used
5863 * to be done via struct cftype but cgroup core no longer knows
5864 * about these events. The following is crude but the whole thing
5865 * is for compatibility anyway.
5867 * DO NOT ADD NEW FILES.
5869 name = cfile.file->f_dentry->d_name.name;
5871 if (!strcmp(name, "memory.usage_in_bytes")) {
5872 event->register_event = mem_cgroup_usage_register_event;
5873 event->unregister_event = mem_cgroup_usage_unregister_event;
5874 } else if (!strcmp(name, "memory.oom_control")) {
5875 event->register_event = mem_cgroup_oom_register_event;
5876 event->unregister_event = mem_cgroup_oom_unregister_event;
5877 } else if (!strcmp(name, "memory.pressure_level")) {
5878 event->register_event = vmpressure_register_event;
5879 event->unregister_event = vmpressure_unregister_event;
5880 } else if (!strcmp(name, "memory.memsw.usage_in_bytes")) {
5881 event->register_event = memsw_cgroup_usage_register_event;
5882 event->unregister_event = memsw_cgroup_usage_unregister_event;
5889 * Verify @cfile should belong to @css. Also, remaining events are
5890 * automatically removed on cgroup destruction but the removal is
5891 * asynchronous, so take an extra ref on @css.
5893 cfile_css = css_tryget_online_from_dir(cfile.file->f_dentry->d_parent,
5894 &memory_cgrp_subsys);
5896 if (IS_ERR(cfile_css))
5898 if (cfile_css != css) {
5903 ret = event->register_event(memcg, event->eventfd, buf);
5907 efile.file->f_op->poll(efile.file, &event->pt);
5909 spin_lock(&memcg->event_list_lock);
5910 list_add(&event->list, &memcg->event_list);
5911 spin_unlock(&memcg->event_list_lock);
5923 eventfd_ctx_put(event->eventfd);
5932 static struct cftype mem_cgroup_files[] = {
5934 .name = "usage_in_bytes",
5935 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5936 .read_u64 = mem_cgroup_read_u64,
5939 .name = "max_usage_in_bytes",
5940 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5941 .write = mem_cgroup_reset,
5942 .read_u64 = mem_cgroup_read_u64,
5945 .name = "limit_in_bytes",
5946 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5947 .write = mem_cgroup_write,
5948 .read_u64 = mem_cgroup_read_u64,
5951 .name = "soft_limit_in_bytes",
5952 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5953 .write = mem_cgroup_write,
5954 .read_u64 = mem_cgroup_read_u64,
5958 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5959 .write = mem_cgroup_reset,
5960 .read_u64 = mem_cgroup_read_u64,
5964 .seq_show = memcg_stat_show,
5967 .name = "force_empty",
5968 .write = mem_cgroup_force_empty_write,
5971 .name = "use_hierarchy",
5972 .write_u64 = mem_cgroup_hierarchy_write,
5973 .read_u64 = mem_cgroup_hierarchy_read,
5976 .name = "cgroup.event_control", /* XXX: for compat */
5977 .write = memcg_write_event_control,
5978 .flags = CFTYPE_NO_PREFIX,
5982 .name = "swappiness",
5983 .read_u64 = mem_cgroup_swappiness_read,
5984 .write_u64 = mem_cgroup_swappiness_write,
5987 .name = "move_charge_at_immigrate",
5988 .read_u64 = mem_cgroup_move_charge_read,
5989 .write_u64 = mem_cgroup_move_charge_write,
5992 .name = "oom_control",
5993 .seq_show = mem_cgroup_oom_control_read,
5994 .write_u64 = mem_cgroup_oom_control_write,
5995 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5998 .name = "pressure_level",
6002 .name = "numa_stat",
6003 .seq_show = memcg_numa_stat_show,
6006 #ifdef CONFIG_MEMCG_KMEM
6008 .name = "kmem.limit_in_bytes",
6009 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6010 .write = mem_cgroup_write,
6011 .read_u64 = mem_cgroup_read_u64,
6014 .name = "kmem.usage_in_bytes",
6015 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6016 .read_u64 = mem_cgroup_read_u64,
6019 .name = "kmem.failcnt",
6020 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6021 .write = mem_cgroup_reset,
6022 .read_u64 = mem_cgroup_read_u64,
6025 .name = "kmem.max_usage_in_bytes",
6026 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6027 .write = mem_cgroup_reset,
6028 .read_u64 = mem_cgroup_read_u64,
6030 #ifdef CONFIG_SLABINFO
6032 .name = "kmem.slabinfo",
6033 .seq_show = mem_cgroup_slabinfo_read,
6037 { }, /* terminate */
6040 #ifdef CONFIG_MEMCG_SWAP
6041 static struct cftype memsw_cgroup_files[] = {
6043 .name = "memsw.usage_in_bytes",
6044 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6045 .read_u64 = mem_cgroup_read_u64,
6048 .name = "memsw.max_usage_in_bytes",
6049 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6050 .write = mem_cgroup_reset,
6051 .read_u64 = mem_cgroup_read_u64,
6054 .name = "memsw.limit_in_bytes",
6055 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6056 .write = mem_cgroup_write,
6057 .read_u64 = mem_cgroup_read_u64,
6060 .name = "memsw.failcnt",
6061 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6062 .write = mem_cgroup_reset,
6063 .read_u64 = mem_cgroup_read_u64,
6065 { }, /* terminate */
6068 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6070 struct mem_cgroup_per_node *pn;
6071 struct mem_cgroup_per_zone *mz;
6072 int zone, tmp = node;
6074 * This routine is called against possible nodes.
6075 * But it's BUG to call kmalloc() against offline node.
6077 * TODO: this routine can waste much memory for nodes which will
6078 * never be onlined. It's better to use memory hotplug callback
6081 if (!node_state(node, N_NORMAL_MEMORY))
6083 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6087 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6088 mz = &pn->zoneinfo[zone];
6089 lruvec_init(&mz->lruvec);
6090 mz->usage_in_excess = 0;
6091 mz->on_tree = false;
6094 memcg->nodeinfo[node] = pn;
6098 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6100 kfree(memcg->nodeinfo[node]);
6103 static struct mem_cgroup *mem_cgroup_alloc(void)
6105 struct mem_cgroup *memcg;
6108 size = sizeof(struct mem_cgroup);
6109 size += nr_node_ids * sizeof(struct mem_cgroup_per_node *);
6111 memcg = kzalloc(size, GFP_KERNEL);
6115 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6118 spin_lock_init(&memcg->pcp_counter_lock);
6127 * At destroying mem_cgroup, references from swap_cgroup can remain.
6128 * (scanning all at force_empty is too costly...)
6130 * Instead of clearing all references at force_empty, we remember
6131 * the number of reference from swap_cgroup and free mem_cgroup when
6132 * it goes down to 0.
6134 * Removal of cgroup itself succeeds regardless of refs from swap.
6137 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6141 mem_cgroup_remove_from_trees(memcg);
6144 free_mem_cgroup_per_zone_info(memcg, node);
6146 free_percpu(memcg->stat);
6149 * We need to make sure that (at least for now), the jump label
6150 * destruction code runs outside of the cgroup lock. This is because
6151 * get_online_cpus(), which is called from the static_branch update,
6152 * can't be called inside the cgroup_lock. cpusets are the ones
6153 * enforcing this dependency, so if they ever change, we might as well.
6155 * schedule_work() will guarantee this happens. Be careful if you need
6156 * to move this code around, and make sure it is outside
6159 disarm_static_keys(memcg);
6164 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6166 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6168 if (!memcg->res.parent)
6170 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6172 EXPORT_SYMBOL(parent_mem_cgroup);
6174 static void __init mem_cgroup_soft_limit_tree_init(void)
6176 struct mem_cgroup_tree_per_node *rtpn;
6177 struct mem_cgroup_tree_per_zone *rtpz;
6178 int tmp, node, zone;
6180 for_each_node(node) {
6182 if (!node_state(node, N_NORMAL_MEMORY))
6184 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6187 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6189 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6190 rtpz = &rtpn->rb_tree_per_zone[zone];
6191 rtpz->rb_root = RB_ROOT;
6192 spin_lock_init(&rtpz->lock);
6197 static struct cgroup_subsys_state * __ref
6198 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6200 struct mem_cgroup *memcg;
6201 long error = -ENOMEM;
6204 memcg = mem_cgroup_alloc();
6206 return ERR_PTR(error);
6209 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6213 if (parent_css == NULL) {
6214 root_mem_cgroup = memcg;
6215 res_counter_init(&memcg->res, NULL);
6216 res_counter_init(&memcg->memsw, NULL);
6217 res_counter_init(&memcg->kmem, NULL);
6220 memcg->last_scanned_node = MAX_NUMNODES;
6221 INIT_LIST_HEAD(&memcg->oom_notify);
6222 memcg->move_charge_at_immigrate = 0;
6223 mutex_init(&memcg->thresholds_lock);
6224 spin_lock_init(&memcg->move_lock);
6225 vmpressure_init(&memcg->vmpressure);
6226 INIT_LIST_HEAD(&memcg->event_list);
6227 spin_lock_init(&memcg->event_list_lock);
6232 __mem_cgroup_free(memcg);
6233 return ERR_PTR(error);
6237 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6239 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6240 struct mem_cgroup *parent = mem_cgroup_from_css(css->parent);
6242 if (css->id > MEM_CGROUP_ID_MAX)
6248 mutex_lock(&memcg_create_mutex);
6250 memcg->use_hierarchy = parent->use_hierarchy;
6251 memcg->oom_kill_disable = parent->oom_kill_disable;
6252 memcg->swappiness = mem_cgroup_swappiness(parent);
6254 if (parent->use_hierarchy) {
6255 res_counter_init(&memcg->res, &parent->res);
6256 res_counter_init(&memcg->memsw, &parent->memsw);
6257 res_counter_init(&memcg->kmem, &parent->kmem);
6260 * No need to take a reference to the parent because cgroup
6261 * core guarantees its existence.
6264 res_counter_init(&memcg->res, NULL);
6265 res_counter_init(&memcg->memsw, NULL);
6266 res_counter_init(&memcg->kmem, NULL);
6268 * Deeper hierachy with use_hierarchy == false doesn't make
6269 * much sense so let cgroup subsystem know about this
6270 * unfortunate state in our controller.
6272 if (parent != root_mem_cgroup)
6273 memory_cgrp_subsys.broken_hierarchy = true;
6275 mutex_unlock(&memcg_create_mutex);
6277 return memcg_init_kmem(memcg, &memory_cgrp_subsys);
6281 * Announce all parents that a group from their hierarchy is gone.
6283 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6285 struct mem_cgroup *parent = memcg;
6287 while ((parent = parent_mem_cgroup(parent)))
6288 mem_cgroup_iter_invalidate(parent);
6291 * if the root memcg is not hierarchical we have to check it
6294 if (!root_mem_cgroup->use_hierarchy)
6295 mem_cgroup_iter_invalidate(root_mem_cgroup);
6298 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6300 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6301 struct mem_cgroup_event *event, *tmp;
6302 struct cgroup_subsys_state *iter;
6305 * Unregister events and notify userspace.
6306 * Notify userspace about cgroup removing only after rmdir of cgroup
6307 * directory to avoid race between userspace and kernelspace.
6309 spin_lock(&memcg->event_list_lock);
6310 list_for_each_entry_safe(event, tmp, &memcg->event_list, list) {
6311 list_del_init(&event->list);
6312 schedule_work(&event->remove);
6314 spin_unlock(&memcg->event_list_lock);
6316 kmem_cgroup_css_offline(memcg);
6318 mem_cgroup_invalidate_reclaim_iterators(memcg);
6321 * This requires that offlining is serialized. Right now that is
6322 * guaranteed because css_killed_work_fn() holds the cgroup_mutex.
6324 css_for_each_descendant_post(iter, css)
6325 mem_cgroup_reparent_charges(mem_cgroup_from_css(iter));
6327 memcg_unregister_all_caches(memcg);
6328 vmpressure_cleanup(&memcg->vmpressure);
6331 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6333 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6335 * XXX: css_offline() would be where we should reparent all
6336 * memory to prepare the cgroup for destruction. However,
6337 * memcg does not do css_tryget_online() and res_counter charging
6338 * under the same RCU lock region, which means that charging
6339 * could race with offlining. Offlining only happens to
6340 * cgroups with no tasks in them but charges can show up
6341 * without any tasks from the swapin path when the target
6342 * memcg is looked up from the swapout record and not from the
6343 * current task as it usually is. A race like this can leak
6344 * charges and put pages with stale cgroup pointers into
6348 * lookup_swap_cgroup_id()
6350 * mem_cgroup_lookup()
6351 * css_tryget_online()
6353 * disable css_tryget_online()
6356 * reparent_charges()
6357 * res_counter_charge()
6360 * pc->mem_cgroup = dead memcg
6363 * The bulk of the charges are still moved in offline_css() to
6364 * avoid pinning a lot of pages in case a long-term reference
6365 * like a swapout record is deferring the css_free() to long
6366 * after offlining. But this makes sure we catch any charges
6367 * made after offlining:
6369 mem_cgroup_reparent_charges(memcg);
6371 memcg_destroy_kmem(memcg);
6372 __mem_cgroup_free(memcg);
6376 * mem_cgroup_css_reset - reset the states of a mem_cgroup
6377 * @css: the target css
6379 * Reset the states of the mem_cgroup associated with @css. This is
6380 * invoked when the userland requests disabling on the default hierarchy
6381 * but the memcg is pinned through dependency. The memcg should stop
6382 * applying policies and should revert to the vanilla state as it may be
6383 * made visible again.
6385 * The current implementation only resets the essential configurations.
6386 * This needs to be expanded to cover all the visible parts.
6388 static void mem_cgroup_css_reset(struct cgroup_subsys_state *css)
6390 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6392 mem_cgroup_resize_limit(memcg, ULLONG_MAX);
6393 mem_cgroup_resize_memsw_limit(memcg, ULLONG_MAX);
6394 memcg_update_kmem_limit(memcg, ULLONG_MAX);
6395 res_counter_set_soft_limit(&memcg->res, ULLONG_MAX);
6399 /* Handlers for move charge at task migration. */
6400 #define PRECHARGE_COUNT_AT_ONCE 256
6401 static int mem_cgroup_do_precharge(unsigned long count)
6404 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6405 struct mem_cgroup *memcg = mc.to;
6407 if (mem_cgroup_is_root(memcg)) {
6408 mc.precharge += count;
6409 /* we don't need css_get for root */
6412 /* try to charge at once */
6414 struct res_counter *dummy;
6416 * "memcg" cannot be under rmdir() because we've already checked
6417 * by cgroup_lock_live_cgroup() that it is not removed and we
6418 * are still under the same cgroup_mutex. So we can postpone
6421 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6423 if (do_swap_account && res_counter_charge(&memcg->memsw,
6424 PAGE_SIZE * count, &dummy)) {
6425 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6428 mc.precharge += count;
6432 /* fall back to one by one charge */
6434 if (signal_pending(current)) {
6438 if (!batch_count--) {
6439 batch_count = PRECHARGE_COUNT_AT_ONCE;
6442 ret = mem_cgroup_try_charge(memcg, GFP_KERNEL, 1, false);
6444 /* mem_cgroup_clear_mc() will do uncharge later */
6452 * get_mctgt_type - get target type of moving charge
6453 * @vma: the vma the pte to be checked belongs
6454 * @addr: the address corresponding to the pte to be checked
6455 * @ptent: the pte to be checked
6456 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6459 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6460 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6461 * move charge. if @target is not NULL, the page is stored in target->page
6462 * with extra refcnt got(Callers should handle it).
6463 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6464 * target for charge migration. if @target is not NULL, the entry is stored
6467 * Called with pte lock held.
6474 enum mc_target_type {
6480 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6481 unsigned long addr, pte_t ptent)
6483 struct page *page = vm_normal_page(vma, addr, ptent);
6485 if (!page || !page_mapped(page))
6487 if (PageAnon(page)) {
6488 /* we don't move shared anon */
6491 } else if (!move_file())
6492 /* we ignore mapcount for file pages */
6494 if (!get_page_unless_zero(page))
6501 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6502 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6504 struct page *page = NULL;
6505 swp_entry_t ent = pte_to_swp_entry(ptent);
6507 if (!move_anon() || non_swap_entry(ent))
6510 * Because lookup_swap_cache() updates some statistics counter,
6511 * we call find_get_page() with swapper_space directly.
6513 page = find_get_page(swap_address_space(ent), ent.val);
6514 if (do_swap_account)
6515 entry->val = ent.val;
6520 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6521 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6527 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6528 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6530 struct page *page = NULL;
6531 struct address_space *mapping;
6534 if (!vma->vm_file) /* anonymous vma */
6539 mapping = vma->vm_file->f_mapping;
6540 if (pte_none(ptent))
6541 pgoff = linear_page_index(vma, addr);
6542 else /* pte_file(ptent) is true */
6543 pgoff = pte_to_pgoff(ptent);
6545 /* page is moved even if it's not RSS of this task(page-faulted). */
6547 /* shmem/tmpfs may report page out on swap: account for that too. */
6548 if (shmem_mapping(mapping)) {
6549 page = find_get_entry(mapping, pgoff);
6550 if (radix_tree_exceptional_entry(page)) {
6551 swp_entry_t swp = radix_to_swp_entry(page);
6552 if (do_swap_account)
6554 page = find_get_page(swap_address_space(swp), swp.val);
6557 page = find_get_page(mapping, pgoff);
6559 page = find_get_page(mapping, pgoff);
6564 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6565 unsigned long addr, pte_t ptent, union mc_target *target)
6567 struct page *page = NULL;
6568 struct page_cgroup *pc;
6569 enum mc_target_type ret = MC_TARGET_NONE;
6570 swp_entry_t ent = { .val = 0 };
6572 if (pte_present(ptent))
6573 page = mc_handle_present_pte(vma, addr, ptent);
6574 else if (is_swap_pte(ptent))
6575 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6576 else if (pte_none(ptent) || pte_file(ptent))
6577 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6579 if (!page && !ent.val)
6582 pc = lookup_page_cgroup(page);
6584 * Do only loose check w/o page_cgroup lock.
6585 * mem_cgroup_move_account() checks the pc is valid or not under
6588 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6589 ret = MC_TARGET_PAGE;
6591 target->page = page;
6593 if (!ret || !target)
6596 /* There is a swap entry and a page doesn't exist or isn't charged */
6597 if (ent.val && !ret &&
6598 mem_cgroup_id(mc.from) == lookup_swap_cgroup_id(ent)) {
6599 ret = MC_TARGET_SWAP;
6606 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6608 * We don't consider swapping or file mapped pages because THP does not
6609 * support them for now.
6610 * Caller should make sure that pmd_trans_huge(pmd) is true.
6612 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6613 unsigned long addr, pmd_t pmd, union mc_target *target)
6615 struct page *page = NULL;
6616 struct page_cgroup *pc;
6617 enum mc_target_type ret = MC_TARGET_NONE;
6619 page = pmd_page(pmd);
6620 VM_BUG_ON_PAGE(!page || !PageHead(page), page);
6623 pc = lookup_page_cgroup(page);
6624 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6625 ret = MC_TARGET_PAGE;
6628 target->page = page;
6634 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6635 unsigned long addr, pmd_t pmd, union mc_target *target)
6637 return MC_TARGET_NONE;
6641 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6642 unsigned long addr, unsigned long end,
6643 struct mm_walk *walk)
6645 struct vm_area_struct *vma = walk->private;
6649 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6650 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6651 mc.precharge += HPAGE_PMD_NR;
6656 if (pmd_trans_unstable(pmd))
6658 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6659 for (; addr != end; pte++, addr += PAGE_SIZE)
6660 if (get_mctgt_type(vma, addr, *pte, NULL))
6661 mc.precharge++; /* increment precharge temporarily */
6662 pte_unmap_unlock(pte - 1, ptl);
6668 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6670 unsigned long precharge;
6671 struct vm_area_struct *vma;
6673 down_read(&mm->mmap_sem);
6674 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6675 struct mm_walk mem_cgroup_count_precharge_walk = {
6676 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6680 if (is_vm_hugetlb_page(vma))
6682 walk_page_range(vma->vm_start, vma->vm_end,
6683 &mem_cgroup_count_precharge_walk);
6685 up_read(&mm->mmap_sem);
6687 precharge = mc.precharge;
6693 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6695 unsigned long precharge = mem_cgroup_count_precharge(mm);
6697 VM_BUG_ON(mc.moving_task);
6698 mc.moving_task = current;
6699 return mem_cgroup_do_precharge(precharge);
6702 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6703 static void __mem_cgroup_clear_mc(void)
6705 struct mem_cgroup *from = mc.from;
6706 struct mem_cgroup *to = mc.to;
6709 /* we must uncharge all the leftover precharges from mc.to */
6711 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6715 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6716 * we must uncharge here.
6718 if (mc.moved_charge) {
6719 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6720 mc.moved_charge = 0;
6722 /* we must fixup refcnts and charges */
6723 if (mc.moved_swap) {
6724 /* uncharge swap account from the old cgroup */
6725 if (!mem_cgroup_is_root(mc.from))
6726 res_counter_uncharge(&mc.from->memsw,
6727 PAGE_SIZE * mc.moved_swap);
6729 for (i = 0; i < mc.moved_swap; i++)
6730 css_put(&mc.from->css);
6732 if (!mem_cgroup_is_root(mc.to)) {
6734 * we charged both to->res and to->memsw, so we should
6737 res_counter_uncharge(&mc.to->res,
6738 PAGE_SIZE * mc.moved_swap);
6740 /* we've already done css_get(mc.to) */
6743 memcg_oom_recover(from);
6744 memcg_oom_recover(to);
6745 wake_up_all(&mc.waitq);
6748 static void mem_cgroup_clear_mc(void)
6750 struct mem_cgroup *from = mc.from;
6753 * we must clear moving_task before waking up waiters at the end of
6756 mc.moving_task = NULL;
6757 __mem_cgroup_clear_mc();
6758 spin_lock(&mc.lock);
6761 spin_unlock(&mc.lock);
6762 mem_cgroup_end_move(from);
6765 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6766 struct cgroup_taskset *tset)
6768 struct task_struct *p = cgroup_taskset_first(tset);
6770 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6771 unsigned long move_charge_at_immigrate;
6774 * We are now commited to this value whatever it is. Changes in this
6775 * tunable will only affect upcoming migrations, not the current one.
6776 * So we need to save it, and keep it going.
6778 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6779 if (move_charge_at_immigrate) {
6780 struct mm_struct *mm;
6781 struct mem_cgroup *from = mem_cgroup_from_task(p);
6783 VM_BUG_ON(from == memcg);
6785 mm = get_task_mm(p);
6788 /* We move charges only when we move a owner of the mm */
6789 if (mm->owner == p) {
6792 VM_BUG_ON(mc.precharge);
6793 VM_BUG_ON(mc.moved_charge);
6794 VM_BUG_ON(mc.moved_swap);
6795 mem_cgroup_start_move(from);
6796 spin_lock(&mc.lock);
6799 mc.immigrate_flags = move_charge_at_immigrate;
6800 spin_unlock(&mc.lock);
6801 /* We set mc.moving_task later */
6803 ret = mem_cgroup_precharge_mc(mm);
6805 mem_cgroup_clear_mc();
6812 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6813 struct cgroup_taskset *tset)
6815 mem_cgroup_clear_mc();
6818 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6819 unsigned long addr, unsigned long end,
6820 struct mm_walk *walk)
6823 struct vm_area_struct *vma = walk->private;
6826 enum mc_target_type target_type;
6827 union mc_target target;
6829 struct page_cgroup *pc;
6832 * We don't take compound_lock() here but no race with splitting thp
6834 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6835 * under splitting, which means there's no concurrent thp split,
6836 * - if another thread runs into split_huge_page() just after we
6837 * entered this if-block, the thread must wait for page table lock
6838 * to be unlocked in __split_huge_page_splitting(), where the main
6839 * part of thp split is not executed yet.
6841 if (pmd_trans_huge_lock(pmd, vma, &ptl) == 1) {
6842 if (mc.precharge < HPAGE_PMD_NR) {
6846 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6847 if (target_type == MC_TARGET_PAGE) {
6849 if (!isolate_lru_page(page)) {
6850 pc = lookup_page_cgroup(page);
6851 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6852 pc, mc.from, mc.to)) {
6853 mc.precharge -= HPAGE_PMD_NR;
6854 mc.moved_charge += HPAGE_PMD_NR;
6856 putback_lru_page(page);
6864 if (pmd_trans_unstable(pmd))
6867 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6868 for (; addr != end; addr += PAGE_SIZE) {
6869 pte_t ptent = *(pte++);
6875 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6876 case MC_TARGET_PAGE:
6878 if (isolate_lru_page(page))
6880 pc = lookup_page_cgroup(page);
6881 if (!mem_cgroup_move_account(page, 1, pc,
6884 /* we uncharge from mc.from later. */
6887 putback_lru_page(page);
6888 put: /* get_mctgt_type() gets the page */
6891 case MC_TARGET_SWAP:
6893 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6895 /* we fixup refcnts and charges later. */
6903 pte_unmap_unlock(pte - 1, ptl);
6908 * We have consumed all precharges we got in can_attach().
6909 * We try charge one by one, but don't do any additional
6910 * charges to mc.to if we have failed in charge once in attach()
6913 ret = mem_cgroup_do_precharge(1);
6921 static void mem_cgroup_move_charge(struct mm_struct *mm)
6923 struct vm_area_struct *vma;
6925 lru_add_drain_all();
6927 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6929 * Someone who are holding the mmap_sem might be waiting in
6930 * waitq. So we cancel all extra charges, wake up all waiters,
6931 * and retry. Because we cancel precharges, we might not be able
6932 * to move enough charges, but moving charge is a best-effort
6933 * feature anyway, so it wouldn't be a big problem.
6935 __mem_cgroup_clear_mc();
6939 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6941 struct mm_walk mem_cgroup_move_charge_walk = {
6942 .pmd_entry = mem_cgroup_move_charge_pte_range,
6946 if (is_vm_hugetlb_page(vma))
6948 ret = walk_page_range(vma->vm_start, vma->vm_end,
6949 &mem_cgroup_move_charge_walk);
6952 * means we have consumed all precharges and failed in
6953 * doing additional charge. Just abandon here.
6957 up_read(&mm->mmap_sem);
6960 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6961 struct cgroup_taskset *tset)
6963 struct task_struct *p = cgroup_taskset_first(tset);
6964 struct mm_struct *mm = get_task_mm(p);
6968 mem_cgroup_move_charge(mm);
6972 mem_cgroup_clear_mc();
6974 #else /* !CONFIG_MMU */
6975 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6976 struct cgroup_taskset *tset)
6980 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6981 struct cgroup_taskset *tset)
6984 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6985 struct cgroup_taskset *tset)
6991 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6992 * to verify whether we're attached to the default hierarchy on each mount
6995 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6998 * use_hierarchy is forced on the default hierarchy. cgroup core
6999 * guarantees that @root doesn't have any children, so turning it
7000 * on for the root memcg is enough.
7002 if (cgroup_on_dfl(root_css->cgroup))
7003 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7006 struct cgroup_subsys memory_cgrp_subsys = {
7007 .css_alloc = mem_cgroup_css_alloc,
7008 .css_online = mem_cgroup_css_online,
7009 .css_offline = mem_cgroup_css_offline,
7010 .css_free = mem_cgroup_css_free,
7011 .css_reset = mem_cgroup_css_reset,
7012 .can_attach = mem_cgroup_can_attach,
7013 .cancel_attach = mem_cgroup_cancel_attach,
7014 .attach = mem_cgroup_move_task,
7015 .bind = mem_cgroup_bind,
7016 .legacy_cftypes = mem_cgroup_files,
7020 #ifdef CONFIG_MEMCG_SWAP
7021 static int __init enable_swap_account(char *s)
7023 if (!strcmp(s, "1"))
7024 really_do_swap_account = 1;
7025 else if (!strcmp(s, "0"))
7026 really_do_swap_account = 0;
7029 __setup("swapaccount=", enable_swap_account);
7031 static void __init memsw_file_init(void)
7033 WARN_ON(cgroup_add_legacy_cftypes(&memory_cgrp_subsys,
7034 memsw_cgroup_files));
7037 static void __init enable_swap_cgroup(void)
7039 if (!mem_cgroup_disabled() && really_do_swap_account) {
7040 do_swap_account = 1;
7046 static void __init enable_swap_cgroup(void)
7052 * subsys_initcall() for memory controller.
7054 * Some parts like hotcpu_notifier() have to be initialized from this context
7055 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7056 * everything that doesn't depend on a specific mem_cgroup structure should
7057 * be initialized from here.
7059 static int __init mem_cgroup_init(void)
7061 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7062 enable_swap_cgroup();
7063 mem_cgroup_soft_limit_tree_init();
7067 subsys_initcall(mem_cgroup_init);