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/vmalloc.h>
53 #include <linux/vmpressure.h>
54 #include <linux/mm_inline.h>
55 #include <linux/page_cgroup.h>
56 #include <linux/cpu.h>
57 #include <linux/oom.h>
58 #include <linux/lockdep.h>
59 #include <linux/file.h>
63 #include <net/tcp_memcontrol.h>
65 #include <asm/uaccess.h>
67 #include <trace/events/vmscan.h>
69 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
70 EXPORT_SYMBOL(mem_cgroup_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 = 0;
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;
152 unsigned long last_dead_count;
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 cgroup_event {
236 * css which the event belongs to.
238 struct cgroup_subsys_state *css;
240 * Control file which the event associated.
244 * eventfd to signal userspace about the event.
246 struct eventfd_ctx *eventfd;
248 * Each of these stored in a list by the cgroup.
250 struct list_head list;
252 * All fields below needed to unregister event when
253 * userspace closes eventfd.
256 wait_queue_head_t *wqh;
258 struct work_struct remove;
261 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
262 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
265 * The memory controller data structure. The memory controller controls both
266 * page cache and RSS per cgroup. We would eventually like to provide
267 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
268 * to help the administrator determine what knobs to tune.
270 * TODO: Add a water mark for the memory controller. Reclaim will begin when
271 * we hit the water mark. May be even add a low water mark, such that
272 * no reclaim occurs from a cgroup at it's low water mark, this is
273 * a feature that will be implemented much later in the future.
276 struct cgroup_subsys_state css;
278 * the counter to account for memory usage
280 struct res_counter res;
282 /* vmpressure notifications */
283 struct vmpressure vmpressure;
286 * the counter to account for mem+swap usage.
288 struct res_counter memsw;
291 * the counter to account for kernel memory usage.
293 struct res_counter kmem;
295 * Should the accounting and control be hierarchical, per subtree?
298 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
302 atomic_t oom_wakeups;
305 /* OOM-Killer disable */
306 int oom_kill_disable;
308 /* set when res.limit == memsw.limit */
309 bool memsw_is_minimum;
311 /* protect arrays of thresholds */
312 struct mutex thresholds_lock;
314 /* thresholds for memory usage. RCU-protected */
315 struct mem_cgroup_thresholds thresholds;
317 /* thresholds for mem+swap usage. RCU-protected */
318 struct mem_cgroup_thresholds memsw_thresholds;
320 /* For oom notifier event fd */
321 struct list_head oom_notify;
324 * Should we move charges of a task when a task is moved into this
325 * mem_cgroup ? And what type of charges should we move ?
327 unsigned long move_charge_at_immigrate;
329 * set > 0 if pages under this cgroup are moving to other cgroup.
331 atomic_t moving_account;
332 /* taken only while moving_account > 0 */
333 spinlock_t move_lock;
337 struct mem_cgroup_stat_cpu __percpu *stat;
339 * used when a cpu is offlined or other synchronizations
340 * See mem_cgroup_read_stat().
342 struct mem_cgroup_stat_cpu nocpu_base;
343 spinlock_t pcp_counter_lock;
346 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
347 struct tcp_memcontrol tcp_mem;
349 #if defined(CONFIG_MEMCG_KMEM)
350 /* analogous to slab_common's slab_caches list. per-memcg */
351 struct list_head memcg_slab_caches;
352 /* Not a spinlock, we can take a lot of time walking the list */
353 struct mutex slab_caches_mutex;
354 /* Index in the kmem_cache->memcg_params->memcg_caches array */
358 int last_scanned_node;
360 nodemask_t scan_nodes;
361 atomic_t numainfo_events;
362 atomic_t numainfo_updating;
365 struct mem_cgroup_per_node *nodeinfo[0];
366 /* WARNING: nodeinfo must be the last member here */
369 static size_t memcg_size(void)
371 return sizeof(struct mem_cgroup) +
372 nr_node_ids * sizeof(struct mem_cgroup_per_node);
375 /* internal only representation about the status of kmem accounting. */
377 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
378 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
379 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
382 /* We account when limit is on, but only after call sites are patched */
383 #define KMEM_ACCOUNTED_MASK \
384 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
386 #ifdef CONFIG_MEMCG_KMEM
387 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
389 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
392 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
394 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
397 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
399 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
402 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
404 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
407 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
410 * Our caller must use css_get() first, because memcg_uncharge_kmem()
411 * will call css_put() if it sees the memcg is dead.
414 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
415 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
418 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
420 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
421 &memcg->kmem_account_flags);
425 /* Stuffs for move charges at task migration. */
427 * Types of charges to be moved. "move_charge_at_immitgrate" and
428 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
431 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
432 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
436 /* "mc" and its members are protected by cgroup_mutex */
437 static struct move_charge_struct {
438 spinlock_t lock; /* for from, to */
439 struct mem_cgroup *from;
440 struct mem_cgroup *to;
441 unsigned long immigrate_flags;
442 unsigned long precharge;
443 unsigned long moved_charge;
444 unsigned long moved_swap;
445 struct task_struct *moving_task; /* a task moving charges */
446 wait_queue_head_t waitq; /* a waitq for other context */
448 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
449 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
452 static bool move_anon(void)
454 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
457 static bool move_file(void)
459 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
463 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
464 * limit reclaim to prevent infinite loops, if they ever occur.
466 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
467 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
470 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
471 MEM_CGROUP_CHARGE_TYPE_ANON,
472 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
473 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
477 /* for encoding cft->private value on file */
485 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
486 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
487 #define MEMFILE_ATTR(val) ((val) & 0xffff)
488 /* Used for OOM nofiier */
489 #define OOM_CONTROL (0)
492 * Reclaim flags for mem_cgroup_hierarchical_reclaim
494 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
495 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
496 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
497 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
500 * The memcg_create_mutex will be held whenever a new cgroup is created.
501 * As a consequence, any change that needs to protect against new child cgroups
502 * appearing has to hold it as well.
504 static DEFINE_MUTEX(memcg_create_mutex);
506 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
508 return s ? container_of(s, struct mem_cgroup, css) : NULL;
511 /* Some nice accessors for the vmpressure. */
512 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
515 memcg = root_mem_cgroup;
516 return &memcg->vmpressure;
519 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
521 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
524 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
526 return &mem_cgroup_from_css(css)->vmpressure;
529 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
531 return (memcg == root_mem_cgroup);
534 /* Writing them here to avoid exposing memcg's inner layout */
535 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
537 void sock_update_memcg(struct sock *sk)
539 if (mem_cgroup_sockets_enabled) {
540 struct mem_cgroup *memcg;
541 struct cg_proto *cg_proto;
543 BUG_ON(!sk->sk_prot->proto_cgroup);
545 /* Socket cloning can throw us here with sk_cgrp already
546 * filled. It won't however, necessarily happen from
547 * process context. So the test for root memcg given
548 * the current task's memcg won't help us in this case.
550 * Respecting the original socket's memcg is a better
551 * decision in this case.
554 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
555 css_get(&sk->sk_cgrp->memcg->css);
560 memcg = mem_cgroup_from_task(current);
561 cg_proto = sk->sk_prot->proto_cgroup(memcg);
562 if (!mem_cgroup_is_root(memcg) &&
563 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
564 sk->sk_cgrp = cg_proto;
569 EXPORT_SYMBOL(sock_update_memcg);
571 void sock_release_memcg(struct sock *sk)
573 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
574 struct mem_cgroup *memcg;
575 WARN_ON(!sk->sk_cgrp->memcg);
576 memcg = sk->sk_cgrp->memcg;
577 css_put(&sk->sk_cgrp->memcg->css);
581 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
583 if (!memcg || mem_cgroup_is_root(memcg))
586 return &memcg->tcp_mem.cg_proto;
588 EXPORT_SYMBOL(tcp_proto_cgroup);
590 static void disarm_sock_keys(struct mem_cgroup *memcg)
592 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
594 static_key_slow_dec(&memcg_socket_limit_enabled);
597 static void disarm_sock_keys(struct mem_cgroup *memcg)
602 #ifdef CONFIG_MEMCG_KMEM
604 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
605 * There are two main reasons for not using the css_id for this:
606 * 1) this works better in sparse environments, where we have a lot of memcgs,
607 * but only a few kmem-limited. Or also, if we have, for instance, 200
608 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
609 * 200 entry array for that.
611 * 2) In order not to violate the cgroup API, we would like to do all memory
612 * allocation in ->create(). At that point, we haven't yet allocated the
613 * css_id. Having a separate index prevents us from messing with the cgroup
616 * The current size of the caches array is stored in
617 * memcg_limited_groups_array_size. It will double each time we have to
620 static DEFINE_IDA(kmem_limited_groups);
621 int memcg_limited_groups_array_size;
624 * MIN_SIZE is different than 1, because we would like to avoid going through
625 * the alloc/free process all the time. In a small machine, 4 kmem-limited
626 * cgroups is a reasonable guess. In the future, it could be a parameter or
627 * tunable, but that is strictly not necessary.
629 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
630 * this constant directly from cgroup, but it is understandable that this is
631 * better kept as an internal representation in cgroup.c. In any case, the
632 * css_id space is not getting any smaller, and we don't have to necessarily
633 * increase ours as well if it increases.
635 #define MEMCG_CACHES_MIN_SIZE 4
636 #define MEMCG_CACHES_MAX_SIZE 65535
639 * A lot of the calls to the cache allocation functions are expected to be
640 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
641 * conditional to this static branch, we'll have to allow modules that does
642 * kmem_cache_alloc and the such to see this symbol as well
644 struct static_key memcg_kmem_enabled_key;
645 EXPORT_SYMBOL(memcg_kmem_enabled_key);
647 static void disarm_kmem_keys(struct mem_cgroup *memcg)
649 if (memcg_kmem_is_active(memcg)) {
650 static_key_slow_dec(&memcg_kmem_enabled_key);
651 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
654 * This check can't live in kmem destruction function,
655 * since the charges will outlive the cgroup
657 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
660 static void disarm_kmem_keys(struct mem_cgroup *memcg)
663 #endif /* CONFIG_MEMCG_KMEM */
665 static void disarm_static_keys(struct mem_cgroup *memcg)
667 disarm_sock_keys(memcg);
668 disarm_kmem_keys(memcg);
671 static void drain_all_stock_async(struct mem_cgroup *memcg);
673 static struct mem_cgroup_per_zone *
674 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
676 VM_BUG_ON((unsigned)nid >= nr_node_ids);
677 return &memcg->nodeinfo[nid]->zoneinfo[zid];
680 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
685 static struct mem_cgroup_per_zone *
686 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
688 int nid = page_to_nid(page);
689 int zid = page_zonenum(page);
691 return mem_cgroup_zoneinfo(memcg, nid, zid);
694 static struct mem_cgroup_tree_per_zone *
695 soft_limit_tree_node_zone(int nid, int zid)
697 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
700 static struct mem_cgroup_tree_per_zone *
701 soft_limit_tree_from_page(struct page *page)
703 int nid = page_to_nid(page);
704 int zid = page_zonenum(page);
706 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
710 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
711 struct mem_cgroup_per_zone *mz,
712 struct mem_cgroup_tree_per_zone *mctz,
713 unsigned long long new_usage_in_excess)
715 struct rb_node **p = &mctz->rb_root.rb_node;
716 struct rb_node *parent = NULL;
717 struct mem_cgroup_per_zone *mz_node;
722 mz->usage_in_excess = new_usage_in_excess;
723 if (!mz->usage_in_excess)
727 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
729 if (mz->usage_in_excess < mz_node->usage_in_excess)
732 * We can't avoid mem cgroups that are over their soft
733 * limit by the same amount
735 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
738 rb_link_node(&mz->tree_node, parent, p);
739 rb_insert_color(&mz->tree_node, &mctz->rb_root);
744 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
745 struct mem_cgroup_per_zone *mz,
746 struct mem_cgroup_tree_per_zone *mctz)
750 rb_erase(&mz->tree_node, &mctz->rb_root);
755 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
756 struct mem_cgroup_per_zone *mz,
757 struct mem_cgroup_tree_per_zone *mctz)
759 spin_lock(&mctz->lock);
760 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
761 spin_unlock(&mctz->lock);
765 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
767 unsigned long long excess;
768 struct mem_cgroup_per_zone *mz;
769 struct mem_cgroup_tree_per_zone *mctz;
770 int nid = page_to_nid(page);
771 int zid = page_zonenum(page);
772 mctz = soft_limit_tree_from_page(page);
775 * Necessary to update all ancestors when hierarchy is used.
776 * because their event counter is not touched.
778 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
779 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
780 excess = res_counter_soft_limit_excess(&memcg->res);
782 * We have to update the tree if mz is on RB-tree or
783 * mem is over its softlimit.
785 if (excess || mz->on_tree) {
786 spin_lock(&mctz->lock);
787 /* if on-tree, remove it */
789 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
791 * Insert again. mz->usage_in_excess will be updated.
792 * If excess is 0, no tree ops.
794 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
795 spin_unlock(&mctz->lock);
800 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
803 struct mem_cgroup_per_zone *mz;
804 struct mem_cgroup_tree_per_zone *mctz;
806 for_each_node(node) {
807 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
808 mz = mem_cgroup_zoneinfo(memcg, node, zone);
809 mctz = soft_limit_tree_node_zone(node, zone);
810 mem_cgroup_remove_exceeded(memcg, mz, mctz);
815 static struct mem_cgroup_per_zone *
816 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
818 struct rb_node *rightmost = NULL;
819 struct mem_cgroup_per_zone *mz;
823 rightmost = rb_last(&mctz->rb_root);
825 goto done; /* Nothing to reclaim from */
827 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
829 * Remove the node now but someone else can add it back,
830 * we will to add it back at the end of reclaim to its correct
831 * position in the tree.
833 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
834 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
835 !css_tryget(&mz->memcg->css))
841 static struct mem_cgroup_per_zone *
842 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
844 struct mem_cgroup_per_zone *mz;
846 spin_lock(&mctz->lock);
847 mz = __mem_cgroup_largest_soft_limit_node(mctz);
848 spin_unlock(&mctz->lock);
853 * Implementation Note: reading percpu statistics for memcg.
855 * Both of vmstat[] and percpu_counter has threshold and do periodic
856 * synchronization to implement "quick" read. There are trade-off between
857 * reading cost and precision of value. Then, we may have a chance to implement
858 * a periodic synchronizion of counter in memcg's counter.
860 * But this _read() function is used for user interface now. The user accounts
861 * memory usage by memory cgroup and he _always_ requires exact value because
862 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
863 * have to visit all online cpus and make sum. So, for now, unnecessary
864 * synchronization is not implemented. (just implemented for cpu hotplug)
866 * If there are kernel internal actions which can make use of some not-exact
867 * value, and reading all cpu value can be performance bottleneck in some
868 * common workload, threashold and synchonization as vmstat[] should be
871 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
872 enum mem_cgroup_stat_index idx)
878 for_each_online_cpu(cpu)
879 val += per_cpu(memcg->stat->count[idx], cpu);
880 #ifdef CONFIG_HOTPLUG_CPU
881 spin_lock(&memcg->pcp_counter_lock);
882 val += memcg->nocpu_base.count[idx];
883 spin_unlock(&memcg->pcp_counter_lock);
889 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
892 int val = (charge) ? 1 : -1;
893 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
896 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
897 enum mem_cgroup_events_index idx)
899 unsigned long val = 0;
903 for_each_online_cpu(cpu)
904 val += per_cpu(memcg->stat->events[idx], cpu);
905 #ifdef CONFIG_HOTPLUG_CPU
906 spin_lock(&memcg->pcp_counter_lock);
907 val += memcg->nocpu_base.events[idx];
908 spin_unlock(&memcg->pcp_counter_lock);
914 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
916 bool anon, int nr_pages)
921 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
922 * counted as CACHE even if it's on ANON LRU.
925 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
928 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
931 if (PageTransHuge(page))
932 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
935 /* pagein of a big page is an event. So, ignore page size */
937 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
939 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
940 nr_pages = -nr_pages; /* for event */
943 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
949 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
951 struct mem_cgroup_per_zone *mz;
953 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
954 return mz->lru_size[lru];
958 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
959 unsigned int lru_mask)
961 struct mem_cgroup_per_zone *mz;
963 unsigned long ret = 0;
965 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
968 if (BIT(lru) & lru_mask)
969 ret += mz->lru_size[lru];
975 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
976 int nid, unsigned int lru_mask)
981 for (zid = 0; zid < MAX_NR_ZONES; zid++)
982 total += mem_cgroup_zone_nr_lru_pages(memcg,
988 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
989 unsigned int lru_mask)
994 for_each_node_state(nid, N_MEMORY)
995 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
999 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
1000 enum mem_cgroup_events_target target)
1002 unsigned long val, next;
1004 val = __this_cpu_read(memcg->stat->nr_page_events);
1005 next = __this_cpu_read(memcg->stat->targets[target]);
1006 /* from time_after() in jiffies.h */
1007 if ((long)next - (long)val < 0) {
1009 case MEM_CGROUP_TARGET_THRESH:
1010 next = val + THRESHOLDS_EVENTS_TARGET;
1012 case MEM_CGROUP_TARGET_SOFTLIMIT:
1013 next = val + SOFTLIMIT_EVENTS_TARGET;
1015 case MEM_CGROUP_TARGET_NUMAINFO:
1016 next = val + NUMAINFO_EVENTS_TARGET;
1021 __this_cpu_write(memcg->stat->targets[target], next);
1028 * Check events in order.
1031 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1034 /* threshold event is triggered in finer grain than soft limit */
1035 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1036 MEM_CGROUP_TARGET_THRESH))) {
1038 bool do_numainfo __maybe_unused;
1040 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1041 MEM_CGROUP_TARGET_SOFTLIMIT);
1042 #if MAX_NUMNODES > 1
1043 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1044 MEM_CGROUP_TARGET_NUMAINFO);
1048 mem_cgroup_threshold(memcg);
1049 if (unlikely(do_softlimit))
1050 mem_cgroup_update_tree(memcg, page);
1051 #if MAX_NUMNODES > 1
1052 if (unlikely(do_numainfo))
1053 atomic_inc(&memcg->numainfo_events);
1059 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1062 * mm_update_next_owner() may clear mm->owner to NULL
1063 * if it races with swapoff, page migration, etc.
1064 * So this can be called with p == NULL.
1069 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1072 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1074 struct mem_cgroup *memcg = NULL;
1079 * Because we have no locks, mm->owner's may be being moved to other
1080 * cgroup. We use css_tryget() here even if this looks
1081 * pessimistic (rather than adding locks here).
1085 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1086 if (unlikely(!memcg))
1088 } while (!css_tryget(&memcg->css));
1094 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1095 * ref. count) or NULL if the whole root's subtree has been visited.
1097 * helper function to be used by mem_cgroup_iter
1099 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1100 struct mem_cgroup *last_visited)
1102 struct cgroup_subsys_state *prev_css, *next_css;
1104 prev_css = last_visited ? &last_visited->css : NULL;
1106 next_css = css_next_descendant_pre(prev_css, &root->css);
1109 * Even if we found a group we have to make sure it is
1110 * alive. css && !memcg means that the groups should be
1111 * skipped and we should continue the tree walk.
1112 * last_visited css is safe to use because it is
1113 * protected by css_get and the tree walk is rcu safe.
1116 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
1118 if (css_tryget(&mem->css))
1121 prev_css = next_css;
1129 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1132 * When a group in the hierarchy below root is destroyed, the
1133 * hierarchy iterator can no longer be trusted since it might
1134 * have pointed to the destroyed group. Invalidate it.
1136 atomic_inc(&root->dead_count);
1139 static struct mem_cgroup *
1140 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1141 struct mem_cgroup *root,
1144 struct mem_cgroup *position = NULL;
1146 * A cgroup destruction happens in two stages: offlining and
1147 * release. They are separated by a RCU grace period.
1149 * If the iterator is valid, we may still race with an
1150 * offlining. The RCU lock ensures the object won't be
1151 * released, tryget will fail if we lost the race.
1153 *sequence = atomic_read(&root->dead_count);
1154 if (iter->last_dead_count == *sequence) {
1156 position = iter->last_visited;
1157 if (position && !css_tryget(&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,
1169 css_put(&last_visited->css);
1171 * We store the sequence count from the time @last_visited was
1172 * loaded successfully instead of rereading it here so that we
1173 * don't lose destruction events in between. We could have
1174 * raced with the destruction of @new_position after all.
1176 iter->last_visited = new_position;
1178 iter->last_dead_count = sequence;
1182 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1183 * @root: hierarchy root
1184 * @prev: previously returned memcg, NULL on first invocation
1185 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1187 * Returns references to children of the hierarchy below @root, or
1188 * @root itself, or %NULL after a full round-trip.
1190 * Caller must pass the return value in @prev on subsequent
1191 * invocations for reference counting, or use mem_cgroup_iter_break()
1192 * to cancel a hierarchy walk before the round-trip is complete.
1194 * Reclaimers can specify a zone and a priority level in @reclaim to
1195 * divide up the memcgs in the hierarchy among all concurrent
1196 * reclaimers operating on the same zone and priority.
1198 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1199 struct mem_cgroup *prev,
1200 struct mem_cgroup_reclaim_cookie *reclaim)
1202 struct mem_cgroup *memcg = NULL;
1203 struct mem_cgroup *last_visited = NULL;
1205 if (mem_cgroup_disabled())
1209 root = root_mem_cgroup;
1211 if (prev && !reclaim)
1212 last_visited = prev;
1214 if (!root->use_hierarchy && root != root_mem_cgroup) {
1222 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1223 int uninitialized_var(seq);
1226 int nid = zone_to_nid(reclaim->zone);
1227 int zid = zone_idx(reclaim->zone);
1228 struct mem_cgroup_per_zone *mz;
1230 mz = mem_cgroup_zoneinfo(root, nid, zid);
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, seq);
1247 else if (!prev && memcg)
1248 reclaim->generation = iter->generation;
1257 if (prev && prev != root)
1258 css_put(&prev->css);
1264 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1265 * @root: hierarchy root
1266 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1268 void mem_cgroup_iter_break(struct mem_cgroup *root,
1269 struct mem_cgroup *prev)
1272 root = root_mem_cgroup;
1273 if (prev && prev != root)
1274 css_put(&prev->css);
1278 * Iteration constructs for visiting all cgroups (under a tree). If
1279 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1280 * be used for reference counting.
1282 #define for_each_mem_cgroup_tree(iter, root) \
1283 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1285 iter = mem_cgroup_iter(root, iter, NULL))
1287 #define for_each_mem_cgroup(iter) \
1288 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1290 iter = mem_cgroup_iter(NULL, iter, NULL))
1292 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1294 struct mem_cgroup *memcg;
1297 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1298 if (unlikely(!memcg))
1303 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1306 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1314 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1317 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1318 * @zone: zone of the wanted lruvec
1319 * @memcg: memcg of the wanted lruvec
1321 * Returns the lru list vector holding pages for the given @zone and
1322 * @mem. This can be the global zone lruvec, if the memory controller
1325 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1326 struct mem_cgroup *memcg)
1328 struct mem_cgroup_per_zone *mz;
1329 struct lruvec *lruvec;
1331 if (mem_cgroup_disabled()) {
1332 lruvec = &zone->lruvec;
1336 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1337 lruvec = &mz->lruvec;
1340 * Since a node can be onlined after the mem_cgroup was created,
1341 * we have to be prepared to initialize lruvec->zone here;
1342 * and if offlined then reonlined, we need to reinitialize it.
1344 if (unlikely(lruvec->zone != zone))
1345 lruvec->zone = zone;
1350 * Following LRU functions are allowed to be used without PCG_LOCK.
1351 * Operations are called by routine of global LRU independently from memcg.
1352 * What we have to take care of here is validness of pc->mem_cgroup.
1354 * Changes to pc->mem_cgroup happens when
1357 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1358 * It is added to LRU before charge.
1359 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1360 * When moving account, the page is not on LRU. It's isolated.
1364 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1366 * @zone: zone of the page
1368 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1370 struct mem_cgroup_per_zone *mz;
1371 struct mem_cgroup *memcg;
1372 struct page_cgroup *pc;
1373 struct lruvec *lruvec;
1375 if (mem_cgroup_disabled()) {
1376 lruvec = &zone->lruvec;
1380 pc = lookup_page_cgroup(page);
1381 memcg = pc->mem_cgroup;
1384 * Surreptitiously switch any uncharged offlist page to root:
1385 * an uncharged page off lru does nothing to secure
1386 * its former mem_cgroup from sudden removal.
1388 * Our caller holds lru_lock, and PageCgroupUsed is updated
1389 * under page_cgroup lock: between them, they make all uses
1390 * of pc->mem_cgroup safe.
1392 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1393 pc->mem_cgroup = memcg = root_mem_cgroup;
1395 mz = page_cgroup_zoneinfo(memcg, page);
1396 lruvec = &mz->lruvec;
1399 * Since a node can be onlined after the mem_cgroup was created,
1400 * we have to be prepared to initialize lruvec->zone here;
1401 * and if offlined then reonlined, we need to reinitialize it.
1403 if (unlikely(lruvec->zone != zone))
1404 lruvec->zone = zone;
1409 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1410 * @lruvec: mem_cgroup per zone lru vector
1411 * @lru: index of lru list the page is sitting on
1412 * @nr_pages: positive when adding or negative when removing
1414 * This function must be called when a page is added to or removed from an
1417 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1420 struct mem_cgroup_per_zone *mz;
1421 unsigned long *lru_size;
1423 if (mem_cgroup_disabled())
1426 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1427 lru_size = mz->lru_size + lru;
1428 *lru_size += nr_pages;
1429 VM_BUG_ON((long)(*lru_size) < 0);
1433 * Checks whether given mem is same or in the root_mem_cgroup's
1436 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1437 struct mem_cgroup *memcg)
1439 if (root_memcg == memcg)
1441 if (!root_memcg->use_hierarchy || !memcg)
1443 return css_is_ancestor(&memcg->css, &root_memcg->css);
1446 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1447 struct mem_cgroup *memcg)
1452 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1457 bool task_in_mem_cgroup(struct task_struct *task,
1458 const struct mem_cgroup *memcg)
1460 struct mem_cgroup *curr = NULL;
1461 struct task_struct *p;
1464 p = find_lock_task_mm(task);
1466 curr = try_get_mem_cgroup_from_mm(p->mm);
1470 * All threads may have already detached their mm's, but the oom
1471 * killer still needs to detect if they have already been oom
1472 * killed to prevent needlessly killing additional tasks.
1475 curr = mem_cgroup_from_task(task);
1477 css_get(&curr->css);
1483 * We should check use_hierarchy of "memcg" not "curr". Because checking
1484 * use_hierarchy of "curr" here make this function true if hierarchy is
1485 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1486 * hierarchy(even if use_hierarchy is disabled in "memcg").
1488 ret = mem_cgroup_same_or_subtree(memcg, curr);
1489 css_put(&curr->css);
1493 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1495 unsigned long inactive_ratio;
1496 unsigned long inactive;
1497 unsigned long active;
1500 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1501 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1503 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1505 inactive_ratio = int_sqrt(10 * gb);
1509 return inactive * inactive_ratio < active;
1512 #define mem_cgroup_from_res_counter(counter, member) \
1513 container_of(counter, struct mem_cgroup, member)
1516 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1517 * @memcg: the memory cgroup
1519 * Returns the maximum amount of memory @mem can be charged with, in
1522 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1524 unsigned long long margin;
1526 margin = res_counter_margin(&memcg->res);
1527 if (do_swap_account)
1528 margin = min(margin, res_counter_margin(&memcg->memsw));
1529 return margin >> PAGE_SHIFT;
1532 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1535 if (!css_parent(&memcg->css))
1536 return vm_swappiness;
1538 return memcg->swappiness;
1542 * memcg->moving_account is used for checking possibility that some thread is
1543 * calling move_account(). When a thread on CPU-A starts moving pages under
1544 * a memcg, other threads should check memcg->moving_account under
1545 * rcu_read_lock(), like this:
1549 * memcg->moving_account+1 if (memcg->mocing_account)
1551 * synchronize_rcu() update something.
1556 /* for quick checking without looking up memcg */
1557 atomic_t memcg_moving __read_mostly;
1559 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1561 atomic_inc(&memcg_moving);
1562 atomic_inc(&memcg->moving_account);
1566 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1569 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1570 * We check NULL in callee rather than caller.
1573 atomic_dec(&memcg_moving);
1574 atomic_dec(&memcg->moving_account);
1579 * 2 routines for checking "mem" is under move_account() or not.
1581 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1582 * is used for avoiding races in accounting. If true,
1583 * pc->mem_cgroup may be overwritten.
1585 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1586 * under hierarchy of moving cgroups. This is for
1587 * waiting at hith-memory prressure caused by "move".
1590 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1592 VM_BUG_ON(!rcu_read_lock_held());
1593 return atomic_read(&memcg->moving_account) > 0;
1596 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1598 struct mem_cgroup *from;
1599 struct mem_cgroup *to;
1602 * Unlike task_move routines, we access mc.to, mc.from not under
1603 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1605 spin_lock(&mc.lock);
1611 ret = mem_cgroup_same_or_subtree(memcg, from)
1612 || mem_cgroup_same_or_subtree(memcg, to);
1614 spin_unlock(&mc.lock);
1618 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1620 if (mc.moving_task && current != mc.moving_task) {
1621 if (mem_cgroup_under_move(memcg)) {
1623 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1624 /* moving charge context might have finished. */
1627 finish_wait(&mc.waitq, &wait);
1635 * Take this lock when
1636 * - a code tries to modify page's memcg while it's USED.
1637 * - a code tries to modify page state accounting in a memcg.
1638 * see mem_cgroup_stolen(), too.
1640 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1641 unsigned long *flags)
1643 spin_lock_irqsave(&memcg->move_lock, *flags);
1646 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1647 unsigned long *flags)
1649 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1652 #define K(x) ((x) << (PAGE_SHIFT-10))
1654 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1655 * @memcg: The memory cgroup that went over limit
1656 * @p: Task that is going to be killed
1658 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1661 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1663 struct cgroup *task_cgrp;
1664 struct cgroup *mem_cgrp;
1666 * Need a buffer in BSS, can't rely on allocations. The code relies
1667 * on the assumption that OOM is serialized for memory controller.
1668 * If this assumption is broken, revisit this code.
1670 static char memcg_name[PATH_MAX];
1672 struct mem_cgroup *iter;
1680 mem_cgrp = memcg->css.cgroup;
1681 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1683 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1686 * Unfortunately, we are unable to convert to a useful name
1687 * But we'll still print out the usage information
1694 pr_info("Task in %s killed", memcg_name);
1697 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1705 * Continues from above, so we don't need an KERN_ level
1707 pr_cont(" as a result of limit of %s\n", memcg_name);
1710 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1711 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1712 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1713 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1714 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1715 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1716 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1717 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1718 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1719 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1720 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1721 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1723 for_each_mem_cgroup_tree(iter, memcg) {
1724 pr_info("Memory cgroup stats");
1727 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1729 pr_cont(" for %s", memcg_name);
1733 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1734 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1736 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1737 K(mem_cgroup_read_stat(iter, i)));
1740 for (i = 0; i < NR_LRU_LISTS; i++)
1741 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1742 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1749 * This function returns the number of memcg under hierarchy tree. Returns
1750 * 1(self count) if no children.
1752 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1755 struct mem_cgroup *iter;
1757 for_each_mem_cgroup_tree(iter, memcg)
1763 * Return the memory (and swap, if configured) limit for a memcg.
1765 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1769 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1772 * Do not consider swap space if we cannot swap due to swappiness
1774 if (mem_cgroup_swappiness(memcg)) {
1777 limit += total_swap_pages << PAGE_SHIFT;
1778 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1781 * If memsw is finite and limits the amount of swap space
1782 * available to this memcg, return that limit.
1784 limit = min(limit, memsw);
1790 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1793 struct mem_cgroup *iter;
1794 unsigned long chosen_points = 0;
1795 unsigned long totalpages;
1796 unsigned int points = 0;
1797 struct task_struct *chosen = NULL;
1800 * If current has a pending SIGKILL or is exiting, then automatically
1801 * select it. The goal is to allow it to allocate so that it may
1802 * quickly exit and free its memory.
1804 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1805 set_thread_flag(TIF_MEMDIE);
1809 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1810 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1811 for_each_mem_cgroup_tree(iter, memcg) {
1812 struct css_task_iter it;
1813 struct task_struct *task;
1815 css_task_iter_start(&iter->css, &it);
1816 while ((task = css_task_iter_next(&it))) {
1817 switch (oom_scan_process_thread(task, totalpages, NULL,
1819 case OOM_SCAN_SELECT:
1821 put_task_struct(chosen);
1823 chosen_points = ULONG_MAX;
1824 get_task_struct(chosen);
1826 case OOM_SCAN_CONTINUE:
1828 case OOM_SCAN_ABORT:
1829 css_task_iter_end(&it);
1830 mem_cgroup_iter_break(memcg, iter);
1832 put_task_struct(chosen);
1837 points = oom_badness(task, memcg, NULL, totalpages);
1838 if (points > chosen_points) {
1840 put_task_struct(chosen);
1842 chosen_points = points;
1843 get_task_struct(chosen);
1846 css_task_iter_end(&it);
1851 points = chosen_points * 1000 / totalpages;
1852 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1853 NULL, "Memory cgroup out of memory");
1856 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1858 unsigned long flags)
1860 unsigned long total = 0;
1861 bool noswap = false;
1864 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1866 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1869 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1871 drain_all_stock_async(memcg);
1872 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1874 * Allow limit shrinkers, which are triggered directly
1875 * by userspace, to catch signals and stop reclaim
1876 * after minimal progress, regardless of the margin.
1878 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1880 if (mem_cgroup_margin(memcg))
1883 * If nothing was reclaimed after two attempts, there
1884 * may be no reclaimable pages in this hierarchy.
1893 * test_mem_cgroup_node_reclaimable
1894 * @memcg: the target memcg
1895 * @nid: the node ID to be checked.
1896 * @noswap : specify true here if the user wants flle only information.
1898 * This function returns whether the specified memcg contains any
1899 * reclaimable pages on a node. Returns true if there are any reclaimable
1900 * pages in the node.
1902 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1903 int nid, bool noswap)
1905 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1907 if (noswap || !total_swap_pages)
1909 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1914 #if MAX_NUMNODES > 1
1917 * Always updating the nodemask is not very good - even if we have an empty
1918 * list or the wrong list here, we can start from some node and traverse all
1919 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1922 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1926 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1927 * pagein/pageout changes since the last update.
1929 if (!atomic_read(&memcg->numainfo_events))
1931 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1934 /* make a nodemask where this memcg uses memory from */
1935 memcg->scan_nodes = node_states[N_MEMORY];
1937 for_each_node_mask(nid, node_states[N_MEMORY]) {
1939 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1940 node_clear(nid, memcg->scan_nodes);
1943 atomic_set(&memcg->numainfo_events, 0);
1944 atomic_set(&memcg->numainfo_updating, 0);
1948 * Selecting a node where we start reclaim from. Because what we need is just
1949 * reducing usage counter, start from anywhere is O,K. Considering
1950 * memory reclaim from current node, there are pros. and cons.
1952 * Freeing memory from current node means freeing memory from a node which
1953 * we'll use or we've used. So, it may make LRU bad. And if several threads
1954 * hit limits, it will see a contention on a node. But freeing from remote
1955 * node means more costs for memory reclaim because of memory latency.
1957 * Now, we use round-robin. Better algorithm is welcomed.
1959 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1963 mem_cgroup_may_update_nodemask(memcg);
1964 node = memcg->last_scanned_node;
1966 node = next_node(node, memcg->scan_nodes);
1967 if (node == MAX_NUMNODES)
1968 node = first_node(memcg->scan_nodes);
1970 * We call this when we hit limit, not when pages are added to LRU.
1971 * No LRU may hold pages because all pages are UNEVICTABLE or
1972 * memcg is too small and all pages are not on LRU. In that case,
1973 * we use curret node.
1975 if (unlikely(node == MAX_NUMNODES))
1976 node = numa_node_id();
1978 memcg->last_scanned_node = node;
1983 * Check all nodes whether it contains reclaimable pages or not.
1984 * For quick scan, we make use of scan_nodes. This will allow us to skip
1985 * unused nodes. But scan_nodes is lazily updated and may not cotain
1986 * enough new information. We need to do double check.
1988 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1993 * quick check...making use of scan_node.
1994 * We can skip unused nodes.
1996 if (!nodes_empty(memcg->scan_nodes)) {
1997 for (nid = first_node(memcg->scan_nodes);
1999 nid = next_node(nid, memcg->scan_nodes)) {
2001 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2006 * Check rest of nodes.
2008 for_each_node_state(nid, N_MEMORY) {
2009 if (node_isset(nid, memcg->scan_nodes))
2011 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
2018 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2023 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2025 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2029 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2032 unsigned long *total_scanned)
2034 struct mem_cgroup *victim = NULL;
2037 unsigned long excess;
2038 unsigned long nr_scanned;
2039 struct mem_cgroup_reclaim_cookie reclaim = {
2044 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2047 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2052 * If we have not been able to reclaim
2053 * anything, it might because there are
2054 * no reclaimable pages under this hierarchy
2059 * We want to do more targeted reclaim.
2060 * excess >> 2 is not to excessive so as to
2061 * reclaim too much, nor too less that we keep
2062 * coming back to reclaim from this cgroup
2064 if (total >= (excess >> 2) ||
2065 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2070 if (!mem_cgroup_reclaimable(victim, false))
2072 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2074 *total_scanned += nr_scanned;
2075 if (!res_counter_soft_limit_excess(&root_memcg->res))
2078 mem_cgroup_iter_break(root_memcg, victim);
2082 #ifdef CONFIG_LOCKDEP
2083 static struct lockdep_map memcg_oom_lock_dep_map = {
2084 .name = "memcg_oom_lock",
2088 static DEFINE_SPINLOCK(memcg_oom_lock);
2091 * Check OOM-Killer is already running under our hierarchy.
2092 * If someone is running, return false.
2094 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2096 struct mem_cgroup *iter, *failed = NULL;
2098 spin_lock(&memcg_oom_lock);
2100 for_each_mem_cgroup_tree(iter, memcg) {
2101 if (iter->oom_lock) {
2103 * this subtree of our hierarchy is already locked
2104 * so we cannot give a lock.
2107 mem_cgroup_iter_break(memcg, iter);
2110 iter->oom_lock = true;
2115 * OK, we failed to lock the whole subtree so we have
2116 * to clean up what we set up to the failing subtree
2118 for_each_mem_cgroup_tree(iter, memcg) {
2119 if (iter == failed) {
2120 mem_cgroup_iter_break(memcg, iter);
2123 iter->oom_lock = false;
2126 mutex_acquire(&memcg_oom_lock_dep_map, 0, 1, _RET_IP_);
2128 spin_unlock(&memcg_oom_lock);
2133 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2135 struct mem_cgroup *iter;
2137 spin_lock(&memcg_oom_lock);
2138 mutex_release(&memcg_oom_lock_dep_map, 1, _RET_IP_);
2139 for_each_mem_cgroup_tree(iter, memcg)
2140 iter->oom_lock = false;
2141 spin_unlock(&memcg_oom_lock);
2144 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2146 struct mem_cgroup *iter;
2148 for_each_mem_cgroup_tree(iter, memcg)
2149 atomic_inc(&iter->under_oom);
2152 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2154 struct mem_cgroup *iter;
2157 * When a new child is created while the hierarchy is under oom,
2158 * mem_cgroup_oom_lock() may not be called. We have to use
2159 * atomic_add_unless() here.
2161 for_each_mem_cgroup_tree(iter, memcg)
2162 atomic_add_unless(&iter->under_oom, -1, 0);
2165 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2167 struct oom_wait_info {
2168 struct mem_cgroup *memcg;
2172 static int memcg_oom_wake_function(wait_queue_t *wait,
2173 unsigned mode, int sync, void *arg)
2175 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2176 struct mem_cgroup *oom_wait_memcg;
2177 struct oom_wait_info *oom_wait_info;
2179 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2180 oom_wait_memcg = oom_wait_info->memcg;
2183 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2184 * Then we can use css_is_ancestor without taking care of RCU.
2186 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2187 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2189 return autoremove_wake_function(wait, mode, sync, arg);
2192 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2194 atomic_inc(&memcg->oom_wakeups);
2195 /* for filtering, pass "memcg" as argument. */
2196 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2199 static void memcg_oom_recover(struct mem_cgroup *memcg)
2201 if (memcg && atomic_read(&memcg->under_oom))
2202 memcg_wakeup_oom(memcg);
2205 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2207 if (!current->memcg_oom.may_oom)
2210 * We are in the middle of the charge context here, so we
2211 * don't want to block when potentially sitting on a callstack
2212 * that holds all kinds of filesystem and mm locks.
2214 * Also, the caller may handle a failed allocation gracefully
2215 * (like optional page cache readahead) and so an OOM killer
2216 * invocation might not even be necessary.
2218 * That's why we don't do anything here except remember the
2219 * OOM context and then deal with it at the end of the page
2220 * fault when the stack is unwound, the locks are released,
2221 * and when we know whether the fault was overall successful.
2223 css_get(&memcg->css);
2224 current->memcg_oom.memcg = memcg;
2225 current->memcg_oom.gfp_mask = mask;
2226 current->memcg_oom.order = order;
2230 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2231 * @handle: actually kill/wait or just clean up the OOM state
2233 * This has to be called at the end of a page fault if the memcg OOM
2234 * handler was enabled.
2236 * Memcg supports userspace OOM handling where failed allocations must
2237 * sleep on a waitqueue until the userspace task resolves the
2238 * situation. Sleeping directly in the charge context with all kinds
2239 * of locks held is not a good idea, instead we remember an OOM state
2240 * in the task and mem_cgroup_oom_synchronize() has to be called at
2241 * the end of the page fault to complete the OOM handling.
2243 * Returns %true if an ongoing memcg OOM situation was detected and
2244 * completed, %false otherwise.
2246 bool mem_cgroup_oom_synchronize(bool handle)
2248 struct mem_cgroup *memcg = current->memcg_oom.memcg;
2249 struct oom_wait_info owait;
2252 /* OOM is global, do not handle */
2259 owait.memcg = memcg;
2260 owait.wait.flags = 0;
2261 owait.wait.func = memcg_oom_wake_function;
2262 owait.wait.private = current;
2263 INIT_LIST_HEAD(&owait.wait.task_list);
2265 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2266 mem_cgroup_mark_under_oom(memcg);
2268 locked = mem_cgroup_oom_trylock(memcg);
2271 mem_cgroup_oom_notify(memcg);
2273 if (locked && !memcg->oom_kill_disable) {
2274 mem_cgroup_unmark_under_oom(memcg);
2275 finish_wait(&memcg_oom_waitq, &owait.wait);
2276 mem_cgroup_out_of_memory(memcg, current->memcg_oom.gfp_mask,
2277 current->memcg_oom.order);
2280 mem_cgroup_unmark_under_oom(memcg);
2281 finish_wait(&memcg_oom_waitq, &owait.wait);
2285 mem_cgroup_oom_unlock(memcg);
2287 * There is no guarantee that an OOM-lock contender
2288 * sees the wakeups triggered by the OOM kill
2289 * uncharges. Wake any sleepers explicitely.
2291 memcg_oom_recover(memcg);
2294 current->memcg_oom.memcg = NULL;
2295 css_put(&memcg->css);
2300 * Currently used to update mapped file statistics, but the routine can be
2301 * generalized to update other statistics as well.
2303 * Notes: Race condition
2305 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2306 * it tends to be costly. But considering some conditions, we doesn't need
2307 * to do so _always_.
2309 * Considering "charge", lock_page_cgroup() is not required because all
2310 * file-stat operations happen after a page is attached to radix-tree. There
2311 * are no race with "charge".
2313 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2314 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2315 * if there are race with "uncharge". Statistics itself is properly handled
2318 * Considering "move", this is an only case we see a race. To make the race
2319 * small, we check mm->moving_account and detect there are possibility of race
2320 * If there is, we take a lock.
2323 void __mem_cgroup_begin_update_page_stat(struct page *page,
2324 bool *locked, unsigned long *flags)
2326 struct mem_cgroup *memcg;
2327 struct page_cgroup *pc;
2329 pc = lookup_page_cgroup(page);
2331 memcg = pc->mem_cgroup;
2332 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2335 * If this memory cgroup is not under account moving, we don't
2336 * need to take move_lock_mem_cgroup(). Because we already hold
2337 * rcu_read_lock(), any calls to move_account will be delayed until
2338 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2340 if (!mem_cgroup_stolen(memcg))
2343 move_lock_mem_cgroup(memcg, flags);
2344 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2345 move_unlock_mem_cgroup(memcg, flags);
2351 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2353 struct page_cgroup *pc = lookup_page_cgroup(page);
2356 * It's guaranteed that pc->mem_cgroup never changes while
2357 * lock is held because a routine modifies pc->mem_cgroup
2358 * should take move_lock_mem_cgroup().
2360 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2363 void mem_cgroup_update_page_stat(struct page *page,
2364 enum mem_cgroup_stat_index idx, int val)
2366 struct mem_cgroup *memcg;
2367 struct page_cgroup *pc = lookup_page_cgroup(page);
2368 unsigned long uninitialized_var(flags);
2370 if (mem_cgroup_disabled())
2373 VM_BUG_ON(!rcu_read_lock_held());
2374 memcg = pc->mem_cgroup;
2375 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2378 this_cpu_add(memcg->stat->count[idx], val);
2382 * size of first charge trial. "32" comes from vmscan.c's magic value.
2383 * TODO: maybe necessary to use big numbers in big irons.
2385 #define CHARGE_BATCH 32U
2386 struct memcg_stock_pcp {
2387 struct mem_cgroup *cached; /* this never be root cgroup */
2388 unsigned int nr_pages;
2389 struct work_struct work;
2390 unsigned long flags;
2391 #define FLUSHING_CACHED_CHARGE 0
2393 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2394 static DEFINE_MUTEX(percpu_charge_mutex);
2397 * consume_stock: Try to consume stocked charge on this cpu.
2398 * @memcg: memcg to consume from.
2399 * @nr_pages: how many pages to charge.
2401 * The charges will only happen if @memcg matches the current cpu's memcg
2402 * stock, and at least @nr_pages are available in that stock. Failure to
2403 * service an allocation will refill the stock.
2405 * returns true if successful, false otherwise.
2407 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2409 struct memcg_stock_pcp *stock;
2412 if (nr_pages > CHARGE_BATCH)
2415 stock = &get_cpu_var(memcg_stock);
2416 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2417 stock->nr_pages -= nr_pages;
2418 else /* need to call res_counter_charge */
2420 put_cpu_var(memcg_stock);
2425 * Returns stocks cached in percpu to res_counter and reset cached information.
2427 static void drain_stock(struct memcg_stock_pcp *stock)
2429 struct mem_cgroup *old = stock->cached;
2431 if (stock->nr_pages) {
2432 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2434 res_counter_uncharge(&old->res, bytes);
2435 if (do_swap_account)
2436 res_counter_uncharge(&old->memsw, bytes);
2437 stock->nr_pages = 0;
2439 stock->cached = NULL;
2443 * This must be called under preempt disabled or must be called by
2444 * a thread which is pinned to local cpu.
2446 static void drain_local_stock(struct work_struct *dummy)
2448 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2450 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2453 static void __init memcg_stock_init(void)
2457 for_each_possible_cpu(cpu) {
2458 struct memcg_stock_pcp *stock =
2459 &per_cpu(memcg_stock, cpu);
2460 INIT_WORK(&stock->work, drain_local_stock);
2465 * Cache charges(val) which is from res_counter, to local per_cpu area.
2466 * This will be consumed by consume_stock() function, later.
2468 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2470 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2472 if (stock->cached != memcg) { /* reset if necessary */
2474 stock->cached = memcg;
2476 stock->nr_pages += nr_pages;
2477 put_cpu_var(memcg_stock);
2481 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2482 * of the hierarchy under it. sync flag says whether we should block
2483 * until the work is done.
2485 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2489 /* Notify other cpus that system-wide "drain" is running */
2492 for_each_online_cpu(cpu) {
2493 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2494 struct mem_cgroup *memcg;
2496 memcg = stock->cached;
2497 if (!memcg || !stock->nr_pages)
2499 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2501 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2503 drain_local_stock(&stock->work);
2505 schedule_work_on(cpu, &stock->work);
2513 for_each_online_cpu(cpu) {
2514 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2515 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2516 flush_work(&stock->work);
2523 * Tries to drain stocked charges in other cpus. This function is asynchronous
2524 * and just put a work per cpu for draining localy on each cpu. Caller can
2525 * expects some charges will be back to res_counter later but cannot wait for
2528 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2531 * If someone calls draining, avoid adding more kworker runs.
2533 if (!mutex_trylock(&percpu_charge_mutex))
2535 drain_all_stock(root_memcg, false);
2536 mutex_unlock(&percpu_charge_mutex);
2539 /* This is a synchronous drain interface. */
2540 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2542 /* called when force_empty is called */
2543 mutex_lock(&percpu_charge_mutex);
2544 drain_all_stock(root_memcg, true);
2545 mutex_unlock(&percpu_charge_mutex);
2549 * This function drains percpu counter value from DEAD cpu and
2550 * move it to local cpu. Note that this function can be preempted.
2552 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2556 spin_lock(&memcg->pcp_counter_lock);
2557 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2558 long x = per_cpu(memcg->stat->count[i], cpu);
2560 per_cpu(memcg->stat->count[i], cpu) = 0;
2561 memcg->nocpu_base.count[i] += x;
2563 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2564 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2566 per_cpu(memcg->stat->events[i], cpu) = 0;
2567 memcg->nocpu_base.events[i] += x;
2569 spin_unlock(&memcg->pcp_counter_lock);
2572 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2573 unsigned long action,
2576 int cpu = (unsigned long)hcpu;
2577 struct memcg_stock_pcp *stock;
2578 struct mem_cgroup *iter;
2580 if (action == CPU_ONLINE)
2583 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2586 for_each_mem_cgroup(iter)
2587 mem_cgroup_drain_pcp_counter(iter, cpu);
2589 stock = &per_cpu(memcg_stock, cpu);
2595 /* See __mem_cgroup_try_charge() for details */
2597 CHARGE_OK, /* success */
2598 CHARGE_RETRY, /* need to retry but retry is not bad */
2599 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2600 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2603 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2604 unsigned int nr_pages, unsigned int min_pages,
2607 unsigned long csize = nr_pages * PAGE_SIZE;
2608 struct mem_cgroup *mem_over_limit;
2609 struct res_counter *fail_res;
2610 unsigned long flags = 0;
2613 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2616 if (!do_swap_account)
2618 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2622 res_counter_uncharge(&memcg->res, csize);
2623 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2624 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2626 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2628 * Never reclaim on behalf of optional batching, retry with a
2629 * single page instead.
2631 if (nr_pages > min_pages)
2632 return CHARGE_RETRY;
2634 if (!(gfp_mask & __GFP_WAIT))
2635 return CHARGE_WOULDBLOCK;
2637 if (gfp_mask & __GFP_NORETRY)
2638 return CHARGE_NOMEM;
2640 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2641 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2642 return CHARGE_RETRY;
2644 * Even though the limit is exceeded at this point, reclaim
2645 * may have been able to free some pages. Retry the charge
2646 * before killing the task.
2648 * Only for regular pages, though: huge pages are rather
2649 * unlikely to succeed so close to the limit, and we fall back
2650 * to regular pages anyway in case of failure.
2652 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2653 return CHARGE_RETRY;
2656 * At task move, charge accounts can be doubly counted. So, it's
2657 * better to wait until the end of task_move if something is going on.
2659 if (mem_cgroup_wait_acct_move(mem_over_limit))
2660 return CHARGE_RETRY;
2663 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2665 return CHARGE_NOMEM;
2669 * __mem_cgroup_try_charge() does
2670 * 1. detect memcg to be charged against from passed *mm and *ptr,
2671 * 2. update res_counter
2672 * 3. call memory reclaim if necessary.
2674 * In some special case, if the task is fatal, fatal_signal_pending() or
2675 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2676 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2677 * as possible without any hazards. 2: all pages should have a valid
2678 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2679 * pointer, that is treated as a charge to root_mem_cgroup.
2681 * So __mem_cgroup_try_charge() will return
2682 * 0 ... on success, filling *ptr with a valid memcg pointer.
2683 * -ENOMEM ... charge failure because of resource limits.
2684 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2686 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2687 * the oom-killer can be invoked.
2689 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2691 unsigned int nr_pages,
2692 struct mem_cgroup **ptr,
2695 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2696 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2697 struct mem_cgroup *memcg = NULL;
2701 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2702 * in system level. So, allow to go ahead dying process in addition to
2705 if (unlikely(test_thread_flag(TIF_MEMDIE)
2706 || fatal_signal_pending(current)))
2709 if (unlikely(task_in_memcg_oom(current)))
2713 * We always charge the cgroup the mm_struct belongs to.
2714 * The mm_struct's mem_cgroup changes on task migration if the
2715 * thread group leader migrates. It's possible that mm is not
2716 * set, if so charge the root memcg (happens for pagecache usage).
2719 *ptr = root_mem_cgroup;
2721 if (*ptr) { /* css should be a valid one */
2723 if (mem_cgroup_is_root(memcg))
2725 if (consume_stock(memcg, nr_pages))
2727 css_get(&memcg->css);
2729 struct task_struct *p;
2732 p = rcu_dereference(mm->owner);
2734 * Because we don't have task_lock(), "p" can exit.
2735 * In that case, "memcg" can point to root or p can be NULL with
2736 * race with swapoff. Then, we have small risk of mis-accouning.
2737 * But such kind of mis-account by race always happens because
2738 * we don't have cgroup_mutex(). It's overkill and we allo that
2740 * (*) swapoff at el will charge against mm-struct not against
2741 * task-struct. So, mm->owner can be NULL.
2743 memcg = mem_cgroup_from_task(p);
2745 memcg = root_mem_cgroup;
2746 if (mem_cgroup_is_root(memcg)) {
2750 if (consume_stock(memcg, nr_pages)) {
2752 * It seems dagerous to access memcg without css_get().
2753 * But considering how consume_stok works, it's not
2754 * necessary. If consume_stock success, some charges
2755 * from this memcg are cached on this cpu. So, we
2756 * don't need to call css_get()/css_tryget() before
2757 * calling consume_stock().
2762 /* after here, we may be blocked. we need to get refcnt */
2763 if (!css_tryget(&memcg->css)) {
2771 bool invoke_oom = oom && !nr_oom_retries;
2773 /* If killed, bypass charge */
2774 if (fatal_signal_pending(current)) {
2775 css_put(&memcg->css);
2779 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2780 nr_pages, invoke_oom);
2784 case CHARGE_RETRY: /* not in OOM situation but retry */
2786 css_put(&memcg->css);
2789 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2790 css_put(&memcg->css);
2792 case CHARGE_NOMEM: /* OOM routine works */
2793 if (!oom || invoke_oom) {
2794 css_put(&memcg->css);
2800 } while (ret != CHARGE_OK);
2802 if (batch > nr_pages)
2803 refill_stock(memcg, batch - nr_pages);
2804 css_put(&memcg->css);
2809 if (!(gfp_mask & __GFP_NOFAIL)) {
2814 *ptr = root_mem_cgroup;
2819 * Somemtimes we have to undo a charge we got by try_charge().
2820 * This function is for that and do uncharge, put css's refcnt.
2821 * gotten by try_charge().
2823 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2824 unsigned int nr_pages)
2826 if (!mem_cgroup_is_root(memcg)) {
2827 unsigned long bytes = nr_pages * PAGE_SIZE;
2829 res_counter_uncharge(&memcg->res, bytes);
2830 if (do_swap_account)
2831 res_counter_uncharge(&memcg->memsw, bytes);
2836 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2837 * This is useful when moving usage to parent cgroup.
2839 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2840 unsigned int nr_pages)
2842 unsigned long bytes = nr_pages * PAGE_SIZE;
2844 if (mem_cgroup_is_root(memcg))
2847 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2848 if (do_swap_account)
2849 res_counter_uncharge_until(&memcg->memsw,
2850 memcg->memsw.parent, bytes);
2854 * A helper function to get mem_cgroup from ID. must be called under
2855 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2856 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2857 * called against removed memcg.)
2859 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2861 struct cgroup_subsys_state *css;
2863 /* ID 0 is unused ID */
2866 css = css_lookup(&mem_cgroup_subsys, id);
2869 return mem_cgroup_from_css(css);
2872 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2874 struct mem_cgroup *memcg = NULL;
2875 struct page_cgroup *pc;
2879 VM_BUG_ON(!PageLocked(page));
2881 pc = lookup_page_cgroup(page);
2882 lock_page_cgroup(pc);
2883 if (PageCgroupUsed(pc)) {
2884 memcg = pc->mem_cgroup;
2885 if (memcg && !css_tryget(&memcg->css))
2887 } else if (PageSwapCache(page)) {
2888 ent.val = page_private(page);
2889 id = lookup_swap_cgroup_id(ent);
2891 memcg = mem_cgroup_lookup(id);
2892 if (memcg && !css_tryget(&memcg->css))
2896 unlock_page_cgroup(pc);
2900 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2902 unsigned int nr_pages,
2903 enum charge_type ctype,
2906 struct page_cgroup *pc = lookup_page_cgroup(page);
2907 struct zone *uninitialized_var(zone);
2908 struct lruvec *lruvec;
2909 bool was_on_lru = false;
2912 lock_page_cgroup(pc);
2913 VM_BUG_ON(PageCgroupUsed(pc));
2915 * we don't need page_cgroup_lock about tail pages, becase they are not
2916 * accessed by any other context at this point.
2920 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2921 * may already be on some other mem_cgroup's LRU. Take care of it.
2924 zone = page_zone(page);
2925 spin_lock_irq(&zone->lru_lock);
2926 if (PageLRU(page)) {
2927 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2929 del_page_from_lru_list(page, lruvec, page_lru(page));
2934 pc->mem_cgroup = memcg;
2936 * We access a page_cgroup asynchronously without lock_page_cgroup().
2937 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2938 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2939 * before USED bit, we need memory barrier here.
2940 * See mem_cgroup_add_lru_list(), etc.
2943 SetPageCgroupUsed(pc);
2947 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2948 VM_BUG_ON(PageLRU(page));
2950 add_page_to_lru_list(page, lruvec, page_lru(page));
2952 spin_unlock_irq(&zone->lru_lock);
2955 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2960 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2961 unlock_page_cgroup(pc);
2964 * "charge_statistics" updated event counter. Then, check it.
2965 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2966 * if they exceeds softlimit.
2968 memcg_check_events(memcg, page);
2971 static DEFINE_MUTEX(set_limit_mutex);
2973 #ifdef CONFIG_MEMCG_KMEM
2974 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2976 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2977 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2981 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2982 * in the memcg_cache_params struct.
2984 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2986 struct kmem_cache *cachep;
2988 VM_BUG_ON(p->is_root_cache);
2989 cachep = p->root_cache;
2990 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2993 #ifdef CONFIG_SLABINFO
2994 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2995 struct cftype *cft, struct seq_file *m)
2997 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2998 struct memcg_cache_params *params;
3000 if (!memcg_can_account_kmem(memcg))
3003 print_slabinfo_header(m);
3005 mutex_lock(&memcg->slab_caches_mutex);
3006 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
3007 cache_show(memcg_params_to_cache(params), m);
3008 mutex_unlock(&memcg->slab_caches_mutex);
3014 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3016 struct res_counter *fail_res;
3017 struct mem_cgroup *_memcg;
3021 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3026 * Conditions under which we can wait for the oom_killer. Those are
3027 * the same conditions tested by the core page allocator
3029 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
3032 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3035 if (ret == -EINTR) {
3037 * __mem_cgroup_try_charge() chosed to bypass to root due to
3038 * OOM kill or fatal signal. Since our only options are to
3039 * either fail the allocation or charge it to this cgroup, do
3040 * it as a temporary condition. But we can't fail. From a
3041 * kmem/slab perspective, the cache has already been selected,
3042 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3045 * This condition will only trigger if the task entered
3046 * memcg_charge_kmem in a sane state, but was OOM-killed during
3047 * __mem_cgroup_try_charge() above. Tasks that were already
3048 * dying when the allocation triggers should have been already
3049 * directed to the root cgroup in memcontrol.h
3051 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3052 if (do_swap_account)
3053 res_counter_charge_nofail(&memcg->memsw, size,
3057 res_counter_uncharge(&memcg->kmem, size);
3062 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3064 res_counter_uncharge(&memcg->res, size);
3065 if (do_swap_account)
3066 res_counter_uncharge(&memcg->memsw, size);
3069 if (res_counter_uncharge(&memcg->kmem, size))
3073 * Releases a reference taken in kmem_cgroup_css_offline in case
3074 * this last uncharge is racing with the offlining code or it is
3075 * outliving the memcg existence.
3077 * The memory barrier imposed by test&clear is paired with the
3078 * explicit one in memcg_kmem_mark_dead().
3080 if (memcg_kmem_test_and_clear_dead(memcg))
3081 css_put(&memcg->css);
3084 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3089 mutex_lock(&memcg->slab_caches_mutex);
3090 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3091 mutex_unlock(&memcg->slab_caches_mutex);
3095 * helper for acessing a memcg's index. It will be used as an index in the
3096 * child cache array in kmem_cache, and also to derive its name. This function
3097 * will return -1 when this is not a kmem-limited memcg.
3099 int memcg_cache_id(struct mem_cgroup *memcg)
3101 return memcg ? memcg->kmemcg_id : -1;
3105 * This ends up being protected by the set_limit mutex, during normal
3106 * operation, because that is its main call site.
3108 * But when we create a new cache, we can call this as well if its parent
3109 * is kmem-limited. That will have to hold set_limit_mutex as well.
3111 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3115 num = ida_simple_get(&kmem_limited_groups,
3116 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3120 * After this point, kmem_accounted (that we test atomically in
3121 * the beginning of this conditional), is no longer 0. This
3122 * guarantees only one process will set the following boolean
3123 * to true. We don't need test_and_set because we're protected
3124 * by the set_limit_mutex anyway.
3126 memcg_kmem_set_activated(memcg);
3128 ret = memcg_update_all_caches(num+1);
3130 ida_simple_remove(&kmem_limited_groups, num);
3131 memcg_kmem_clear_activated(memcg);
3135 memcg->kmemcg_id = num;
3136 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3137 mutex_init(&memcg->slab_caches_mutex);
3141 static size_t memcg_caches_array_size(int num_groups)
3144 if (num_groups <= 0)
3147 size = 2 * num_groups;
3148 if (size < MEMCG_CACHES_MIN_SIZE)
3149 size = MEMCG_CACHES_MIN_SIZE;
3150 else if (size > MEMCG_CACHES_MAX_SIZE)
3151 size = MEMCG_CACHES_MAX_SIZE;
3157 * We should update the current array size iff all caches updates succeed. This
3158 * can only be done from the slab side. The slab mutex needs to be held when
3161 void memcg_update_array_size(int num)
3163 if (num > memcg_limited_groups_array_size)
3164 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3167 static void kmem_cache_destroy_work_func(struct work_struct *w);
3169 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3171 struct memcg_cache_params *cur_params = s->memcg_params;
3173 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3175 if (num_groups > memcg_limited_groups_array_size) {
3177 ssize_t size = memcg_caches_array_size(num_groups);
3179 size *= sizeof(void *);
3180 size += offsetof(struct memcg_cache_params, memcg_caches);
3182 s->memcg_params = kzalloc(size, GFP_KERNEL);
3183 if (!s->memcg_params) {
3184 s->memcg_params = cur_params;
3188 s->memcg_params->is_root_cache = true;
3191 * There is the chance it will be bigger than
3192 * memcg_limited_groups_array_size, if we failed an allocation
3193 * in a cache, in which case all caches updated before it, will
3194 * have a bigger array.
3196 * But if that is the case, the data after
3197 * memcg_limited_groups_array_size is certainly unused
3199 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3200 if (!cur_params->memcg_caches[i])
3202 s->memcg_params->memcg_caches[i] =
3203 cur_params->memcg_caches[i];
3207 * Ideally, we would wait until all caches succeed, and only
3208 * then free the old one. But this is not worth the extra
3209 * pointer per-cache we'd have to have for this.
3211 * It is not a big deal if some caches are left with a size
3212 * bigger than the others. And all updates will reset this
3220 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3221 struct kmem_cache *root_cache)
3225 if (!memcg_kmem_enabled())
3229 size = offsetof(struct memcg_cache_params, memcg_caches);
3230 size += memcg_limited_groups_array_size * sizeof(void *);
3232 size = sizeof(struct memcg_cache_params);
3234 s->memcg_params = kzalloc(size, GFP_KERNEL);
3235 if (!s->memcg_params)
3239 s->memcg_params->memcg = memcg;
3240 s->memcg_params->root_cache = root_cache;
3241 INIT_WORK(&s->memcg_params->destroy,
3242 kmem_cache_destroy_work_func);
3244 s->memcg_params->is_root_cache = true;
3249 void memcg_release_cache(struct kmem_cache *s)
3251 struct kmem_cache *root;
3252 struct mem_cgroup *memcg;
3256 * This happens, for instance, when a root cache goes away before we
3259 if (!s->memcg_params)
3262 if (s->memcg_params->is_root_cache)
3265 memcg = s->memcg_params->memcg;
3266 id = memcg_cache_id(memcg);
3268 root = s->memcg_params->root_cache;
3269 root->memcg_params->memcg_caches[id] = NULL;
3271 mutex_lock(&memcg->slab_caches_mutex);
3272 list_del(&s->memcg_params->list);
3273 mutex_unlock(&memcg->slab_caches_mutex);
3275 css_put(&memcg->css);
3277 kfree(s->memcg_params);
3281 * During the creation a new cache, we need to disable our accounting mechanism
3282 * altogether. This is true even if we are not creating, but rather just
3283 * enqueing new caches to be created.
3285 * This is because that process will trigger allocations; some visible, like
3286 * explicit kmallocs to auxiliary data structures, name strings and internal
3287 * cache structures; some well concealed, like INIT_WORK() that can allocate
3288 * objects during debug.
3290 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3291 * to it. This may not be a bounded recursion: since the first cache creation
3292 * failed to complete (waiting on the allocation), we'll just try to create the
3293 * cache again, failing at the same point.
3295 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3296 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3297 * inside the following two functions.
3299 static inline void memcg_stop_kmem_account(void)
3301 VM_BUG_ON(!current->mm);
3302 current->memcg_kmem_skip_account++;
3305 static inline void memcg_resume_kmem_account(void)
3307 VM_BUG_ON(!current->mm);
3308 current->memcg_kmem_skip_account--;
3311 static void kmem_cache_destroy_work_func(struct work_struct *w)
3313 struct kmem_cache *cachep;
3314 struct memcg_cache_params *p;
3316 p = container_of(w, struct memcg_cache_params, destroy);
3318 cachep = memcg_params_to_cache(p);
3321 * If we get down to 0 after shrink, we could delete right away.
3322 * However, memcg_release_pages() already puts us back in the workqueue
3323 * in that case. If we proceed deleting, we'll get a dangling
3324 * reference, and removing the object from the workqueue in that case
3325 * is unnecessary complication. We are not a fast path.
3327 * Note that this case is fundamentally different from racing with
3328 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3329 * kmem_cache_shrink, not only we would be reinserting a dead cache
3330 * into the queue, but doing so from inside the worker racing to
3333 * So if we aren't down to zero, we'll just schedule a worker and try
3336 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3337 kmem_cache_shrink(cachep);
3338 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3341 kmem_cache_destroy(cachep);
3344 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3346 if (!cachep->memcg_params->dead)
3350 * There are many ways in which we can get here.
3352 * We can get to a memory-pressure situation while the delayed work is
3353 * still pending to run. The vmscan shrinkers can then release all
3354 * cache memory and get us to destruction. If this is the case, we'll
3355 * be executed twice, which is a bug (the second time will execute over
3356 * bogus data). In this case, cancelling the work should be fine.
3358 * But we can also get here from the worker itself, if
3359 * kmem_cache_shrink is enough to shake all the remaining objects and
3360 * get the page count to 0. In this case, we'll deadlock if we try to
3361 * cancel the work (the worker runs with an internal lock held, which
3362 * is the same lock we would hold for cancel_work_sync().)
3364 * Since we can't possibly know who got us here, just refrain from
3365 * running if there is already work pending
3367 if (work_pending(&cachep->memcg_params->destroy))
3370 * We have to defer the actual destroying to a workqueue, because
3371 * we might currently be in a context that cannot sleep.
3373 schedule_work(&cachep->memcg_params->destroy);
3377 * This lock protects updaters, not readers. We want readers to be as fast as
3378 * they can, and they will either see NULL or a valid cache value. Our model
3379 * allow them to see NULL, in which case the root memcg will be selected.
3381 * We need this lock because multiple allocations to the same cache from a non
3382 * will span more than one worker. Only one of them can create the cache.
3384 static DEFINE_MUTEX(memcg_cache_mutex);
3387 * Called with memcg_cache_mutex held
3389 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3390 struct kmem_cache *s)
3392 struct kmem_cache *new;
3393 static char *tmp_name = NULL;
3395 lockdep_assert_held(&memcg_cache_mutex);
3398 * kmem_cache_create_memcg duplicates the given name and
3399 * cgroup_name for this name requires RCU context.
3400 * This static temporary buffer is used to prevent from
3401 * pointless shortliving allocation.
3404 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3410 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3411 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3414 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3415 (s->flags & ~SLAB_PANIC), s->ctor, s);
3418 new->allocflags |= __GFP_KMEMCG;
3423 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3424 struct kmem_cache *cachep)
3426 struct kmem_cache *new_cachep;
3429 BUG_ON(!memcg_can_account_kmem(memcg));
3431 idx = memcg_cache_id(memcg);
3433 mutex_lock(&memcg_cache_mutex);
3434 new_cachep = cachep->memcg_params->memcg_caches[idx];
3436 css_put(&memcg->css);
3440 new_cachep = kmem_cache_dup(memcg, cachep);
3441 if (new_cachep == NULL) {
3442 new_cachep = cachep;
3443 css_put(&memcg->css);
3447 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3449 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3451 * the readers won't lock, make sure everybody sees the updated value,
3452 * so they won't put stuff in the queue again for no reason
3456 mutex_unlock(&memcg_cache_mutex);
3460 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3462 struct kmem_cache *c;
3465 if (!s->memcg_params)
3467 if (!s->memcg_params->is_root_cache)
3471 * If the cache is being destroyed, we trust that there is no one else
3472 * requesting objects from it. Even if there are, the sanity checks in
3473 * kmem_cache_destroy should caught this ill-case.
3475 * Still, we don't want anyone else freeing memcg_caches under our
3476 * noses, which can happen if a new memcg comes to life. As usual,
3477 * we'll take the set_limit_mutex to protect ourselves against this.
3479 mutex_lock(&set_limit_mutex);
3480 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3481 c = s->memcg_params->memcg_caches[i];
3486 * We will now manually delete the caches, so to avoid races
3487 * we need to cancel all pending destruction workers and
3488 * proceed with destruction ourselves.
3490 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3491 * and that could spawn the workers again: it is likely that
3492 * the cache still have active pages until this very moment.
3493 * This would lead us back to mem_cgroup_destroy_cache.
3495 * But that will not execute at all if the "dead" flag is not
3496 * set, so flip it down to guarantee we are in control.
3498 c->memcg_params->dead = false;
3499 cancel_work_sync(&c->memcg_params->destroy);
3500 kmem_cache_destroy(c);
3502 mutex_unlock(&set_limit_mutex);
3505 struct create_work {
3506 struct mem_cgroup *memcg;
3507 struct kmem_cache *cachep;
3508 struct work_struct work;
3511 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3513 struct kmem_cache *cachep;
3514 struct memcg_cache_params *params;
3516 if (!memcg_kmem_is_active(memcg))
3519 mutex_lock(&memcg->slab_caches_mutex);
3520 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3521 cachep = memcg_params_to_cache(params);
3522 cachep->memcg_params->dead = true;
3523 schedule_work(&cachep->memcg_params->destroy);
3525 mutex_unlock(&memcg->slab_caches_mutex);
3528 static void memcg_create_cache_work_func(struct work_struct *w)
3530 struct create_work *cw;
3532 cw = container_of(w, struct create_work, work);
3533 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3538 * Enqueue the creation of a per-memcg kmem_cache.
3540 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3541 struct kmem_cache *cachep)
3543 struct create_work *cw;
3545 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3547 css_put(&memcg->css);
3552 cw->cachep = cachep;
3554 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3555 schedule_work(&cw->work);
3558 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3559 struct kmem_cache *cachep)
3562 * We need to stop accounting when we kmalloc, because if the
3563 * corresponding kmalloc cache is not yet created, the first allocation
3564 * in __memcg_create_cache_enqueue will recurse.
3566 * However, it is better to enclose the whole function. Depending on
3567 * the debugging options enabled, INIT_WORK(), for instance, can
3568 * trigger an allocation. This too, will make us recurse. Because at
3569 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3570 * the safest choice is to do it like this, wrapping the whole function.
3572 memcg_stop_kmem_account();
3573 __memcg_create_cache_enqueue(memcg, cachep);
3574 memcg_resume_kmem_account();
3577 * Return the kmem_cache we're supposed to use for a slab allocation.
3578 * We try to use the current memcg's version of the cache.
3580 * If the cache does not exist yet, if we are the first user of it,
3581 * we either create it immediately, if possible, or create it asynchronously
3583 * In the latter case, we will let the current allocation go through with
3584 * the original cache.
3586 * Can't be called in interrupt context or from kernel threads.
3587 * This function needs to be called with rcu_read_lock() held.
3589 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3592 struct mem_cgroup *memcg;
3595 VM_BUG_ON(!cachep->memcg_params);
3596 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3598 if (!current->mm || current->memcg_kmem_skip_account)
3602 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3604 if (!memcg_can_account_kmem(memcg))
3607 idx = memcg_cache_id(memcg);
3610 * barrier to mare sure we're always seeing the up to date value. The
3611 * code updating memcg_caches will issue a write barrier to match this.
3613 read_barrier_depends();
3614 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3615 cachep = cachep->memcg_params->memcg_caches[idx];
3619 /* The corresponding put will be done in the workqueue. */
3620 if (!css_tryget(&memcg->css))
3625 * If we are in a safe context (can wait, and not in interrupt
3626 * context), we could be be predictable and return right away.
3627 * This would guarantee that the allocation being performed
3628 * already belongs in the new cache.
3630 * However, there are some clashes that can arrive from locking.
3631 * For instance, because we acquire the slab_mutex while doing
3632 * kmem_cache_dup, this means no further allocation could happen
3633 * with the slab_mutex held.
3635 * Also, because cache creation issue get_online_cpus(), this
3636 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3637 * that ends up reversed during cpu hotplug. (cpuset allocates
3638 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3639 * better to defer everything.
3641 memcg_create_cache_enqueue(memcg, cachep);
3647 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3650 * We need to verify if the allocation against current->mm->owner's memcg is
3651 * possible for the given order. But the page is not allocated yet, so we'll
3652 * need a further commit step to do the final arrangements.
3654 * It is possible for the task to switch cgroups in this mean time, so at
3655 * commit time, we can't rely on task conversion any longer. We'll then use
3656 * the handle argument to return to the caller which cgroup we should commit
3657 * against. We could also return the memcg directly and avoid the pointer
3658 * passing, but a boolean return value gives better semantics considering
3659 * the compiled-out case as well.
3661 * Returning true means the allocation is possible.
3664 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3666 struct mem_cgroup *memcg;
3672 * Disabling accounting is only relevant for some specific memcg
3673 * internal allocations. Therefore we would initially not have such
3674 * check here, since direct calls to the page allocator that are marked
3675 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3676 * concerned with cache allocations, and by having this test at
3677 * memcg_kmem_get_cache, we are already able to relay the allocation to
3678 * the root cache and bypass the memcg cache altogether.
3680 * There is one exception, though: the SLUB allocator does not create
3681 * large order caches, but rather service large kmallocs directly from
3682 * the page allocator. Therefore, the following sequence when backed by
3683 * the SLUB allocator:
3685 * memcg_stop_kmem_account();
3686 * kmalloc(<large_number>)
3687 * memcg_resume_kmem_account();
3689 * would effectively ignore the fact that we should skip accounting,
3690 * since it will drive us directly to this function without passing
3691 * through the cache selector memcg_kmem_get_cache. Such large
3692 * allocations are extremely rare but can happen, for instance, for the
3693 * cache arrays. We bring this test here.
3695 if (!current->mm || current->memcg_kmem_skip_account)
3698 memcg = try_get_mem_cgroup_from_mm(current->mm);
3701 * very rare case described in mem_cgroup_from_task. Unfortunately there
3702 * isn't much we can do without complicating this too much, and it would
3703 * be gfp-dependent anyway. Just let it go
3705 if (unlikely(!memcg))
3708 if (!memcg_can_account_kmem(memcg)) {
3709 css_put(&memcg->css);
3713 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3717 css_put(&memcg->css);
3721 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3724 struct page_cgroup *pc;
3726 VM_BUG_ON(mem_cgroup_is_root(memcg));
3728 /* The page allocation failed. Revert */
3730 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3734 pc = lookup_page_cgroup(page);
3735 lock_page_cgroup(pc);
3736 pc->mem_cgroup = memcg;
3737 SetPageCgroupUsed(pc);
3738 unlock_page_cgroup(pc);
3741 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3743 struct mem_cgroup *memcg = NULL;
3744 struct page_cgroup *pc;
3747 pc = lookup_page_cgroup(page);
3749 * Fast unlocked return. Theoretically might have changed, have to
3750 * check again after locking.
3752 if (!PageCgroupUsed(pc))
3755 lock_page_cgroup(pc);
3756 if (PageCgroupUsed(pc)) {
3757 memcg = pc->mem_cgroup;
3758 ClearPageCgroupUsed(pc);
3760 unlock_page_cgroup(pc);
3763 * We trust that only if there is a memcg associated with the page, it
3764 * is a valid allocation
3769 VM_BUG_ON(mem_cgroup_is_root(memcg));
3770 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3773 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3776 #endif /* CONFIG_MEMCG_KMEM */
3778 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3780 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3782 * Because tail pages are not marked as "used", set it. We're under
3783 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3784 * charge/uncharge will be never happen and move_account() is done under
3785 * compound_lock(), so we don't have to take care of races.
3787 void mem_cgroup_split_huge_fixup(struct page *head)
3789 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3790 struct page_cgroup *pc;
3791 struct mem_cgroup *memcg;
3794 if (mem_cgroup_disabled())
3797 memcg = head_pc->mem_cgroup;
3798 for (i = 1; i < HPAGE_PMD_NR; i++) {
3800 pc->mem_cgroup = memcg;
3801 smp_wmb();/* see __commit_charge() */
3802 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3804 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3807 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3810 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3811 struct mem_cgroup *to,
3812 unsigned int nr_pages,
3813 enum mem_cgroup_stat_index idx)
3815 /* Update stat data for mem_cgroup */
3817 __this_cpu_sub(from->stat->count[idx], nr_pages);
3818 __this_cpu_add(to->stat->count[idx], nr_pages);
3823 * mem_cgroup_move_account - move account of the page
3825 * @nr_pages: number of regular pages (>1 for huge pages)
3826 * @pc: page_cgroup of the page.
3827 * @from: mem_cgroup which the page is moved from.
3828 * @to: mem_cgroup which the page is moved to. @from != @to.
3830 * The caller must confirm following.
3831 * - page is not on LRU (isolate_page() is useful.)
3832 * - compound_lock is held when nr_pages > 1
3834 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3837 static int mem_cgroup_move_account(struct page *page,
3838 unsigned int nr_pages,
3839 struct page_cgroup *pc,
3840 struct mem_cgroup *from,
3841 struct mem_cgroup *to)
3843 unsigned long flags;
3845 bool anon = PageAnon(page);
3847 VM_BUG_ON(from == to);
3848 VM_BUG_ON(PageLRU(page));
3850 * The page is isolated from LRU. So, collapse function
3851 * will not handle this page. But page splitting can happen.
3852 * Do this check under compound_page_lock(). The caller should
3856 if (nr_pages > 1 && !PageTransHuge(page))
3859 lock_page_cgroup(pc);
3862 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3865 move_lock_mem_cgroup(from, &flags);
3867 if (!anon && page_mapped(page))
3868 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3869 MEM_CGROUP_STAT_FILE_MAPPED);
3871 if (PageWriteback(page))
3872 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3873 MEM_CGROUP_STAT_WRITEBACK);
3875 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3877 /* caller should have done css_get */
3878 pc->mem_cgroup = to;
3879 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3880 move_unlock_mem_cgroup(from, &flags);
3883 unlock_page_cgroup(pc);
3887 memcg_check_events(to, page);
3888 memcg_check_events(from, page);
3894 * mem_cgroup_move_parent - moves page to the parent group
3895 * @page: the page to move
3896 * @pc: page_cgroup of the page
3897 * @child: page's cgroup
3899 * move charges to its parent or the root cgroup if the group has no
3900 * parent (aka use_hierarchy==0).
3901 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3902 * mem_cgroup_move_account fails) the failure is always temporary and
3903 * it signals a race with a page removal/uncharge or migration. In the
3904 * first case the page is on the way out and it will vanish from the LRU
3905 * on the next attempt and the call should be retried later.
3906 * Isolation from the LRU fails only if page has been isolated from
3907 * the LRU since we looked at it and that usually means either global
3908 * reclaim or migration going on. The page will either get back to the
3910 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3911 * (!PageCgroupUsed) or moved to a different group. The page will
3912 * disappear in the next attempt.
3914 static int mem_cgroup_move_parent(struct page *page,
3915 struct page_cgroup *pc,
3916 struct mem_cgroup *child)
3918 struct mem_cgroup *parent;
3919 unsigned int nr_pages;
3920 unsigned long uninitialized_var(flags);
3923 VM_BUG_ON(mem_cgroup_is_root(child));
3926 if (!get_page_unless_zero(page))
3928 if (isolate_lru_page(page))
3931 nr_pages = hpage_nr_pages(page);
3933 parent = parent_mem_cgroup(child);
3935 * If no parent, move charges to root cgroup.
3938 parent = root_mem_cgroup;
3941 VM_BUG_ON(!PageTransHuge(page));
3942 flags = compound_lock_irqsave(page);
3945 ret = mem_cgroup_move_account(page, nr_pages,
3948 __mem_cgroup_cancel_local_charge(child, nr_pages);
3951 compound_unlock_irqrestore(page, flags);
3952 putback_lru_page(page);
3960 * Charge the memory controller for page usage.
3962 * 0 if the charge was successful
3963 * < 0 if the cgroup is over its limit
3965 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3966 gfp_t gfp_mask, enum charge_type ctype)
3968 struct mem_cgroup *memcg = NULL;
3969 unsigned int nr_pages = 1;
3973 if (PageTransHuge(page)) {
3974 nr_pages <<= compound_order(page);
3975 VM_BUG_ON(!PageTransHuge(page));
3977 * Never OOM-kill a process for a huge page. The
3978 * fault handler will fall back to regular pages.
3983 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3986 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3990 int mem_cgroup_newpage_charge(struct page *page,
3991 struct mm_struct *mm, gfp_t gfp_mask)
3993 if (mem_cgroup_disabled())
3995 VM_BUG_ON(page_mapped(page));
3996 VM_BUG_ON(page->mapping && !PageAnon(page));
3998 return mem_cgroup_charge_common(page, mm, gfp_mask,
3999 MEM_CGROUP_CHARGE_TYPE_ANON);
4003 * While swap-in, try_charge -> commit or cancel, the page is locked.
4004 * And when try_charge() successfully returns, one refcnt to memcg without
4005 * struct page_cgroup is acquired. This refcnt will be consumed by
4006 * "commit()" or removed by "cancel()"
4008 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
4011 struct mem_cgroup **memcgp)
4013 struct mem_cgroup *memcg;
4014 struct page_cgroup *pc;
4017 pc = lookup_page_cgroup(page);
4019 * Every swap fault against a single page tries to charge the
4020 * page, bail as early as possible. shmem_unuse() encounters
4021 * already charged pages, too. The USED bit is protected by
4022 * the page lock, which serializes swap cache removal, which
4023 * in turn serializes uncharging.
4025 if (PageCgroupUsed(pc))
4027 if (!do_swap_account)
4029 memcg = try_get_mem_cgroup_from_page(page);
4033 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4034 css_put(&memcg->css);
4039 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4045 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4046 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4049 if (mem_cgroup_disabled())
4052 * A racing thread's fault, or swapoff, may have already
4053 * updated the pte, and even removed page from swap cache: in
4054 * those cases unuse_pte()'s pte_same() test will fail; but
4055 * there's also a KSM case which does need to charge the page.
4057 if (!PageSwapCache(page)) {
4060 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4065 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4068 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4070 if (mem_cgroup_disabled())
4074 __mem_cgroup_cancel_charge(memcg, 1);
4078 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4079 enum charge_type ctype)
4081 if (mem_cgroup_disabled())
4086 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4088 * Now swap is on-memory. This means this page may be
4089 * counted both as mem and swap....double count.
4090 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4091 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4092 * may call delete_from_swap_cache() before reach here.
4094 if (do_swap_account && PageSwapCache(page)) {
4095 swp_entry_t ent = {.val = page_private(page)};
4096 mem_cgroup_uncharge_swap(ent);
4100 void mem_cgroup_commit_charge_swapin(struct page *page,
4101 struct mem_cgroup *memcg)
4103 __mem_cgroup_commit_charge_swapin(page, memcg,
4104 MEM_CGROUP_CHARGE_TYPE_ANON);
4107 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4110 struct mem_cgroup *memcg = NULL;
4111 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4114 if (mem_cgroup_disabled())
4116 if (PageCompound(page))
4119 if (!PageSwapCache(page))
4120 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4121 else { /* page is swapcache/shmem */
4122 ret = __mem_cgroup_try_charge_swapin(mm, page,
4125 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4130 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4131 unsigned int nr_pages,
4132 const enum charge_type ctype)
4134 struct memcg_batch_info *batch = NULL;
4135 bool uncharge_memsw = true;
4137 /* If swapout, usage of swap doesn't decrease */
4138 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4139 uncharge_memsw = false;
4141 batch = ¤t->memcg_batch;
4143 * In usual, we do css_get() when we remember memcg pointer.
4144 * But in this case, we keep res->usage until end of a series of
4145 * uncharges. Then, it's ok to ignore memcg's refcnt.
4148 batch->memcg = memcg;
4150 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4151 * In those cases, all pages freed continuously can be expected to be in
4152 * the same cgroup and we have chance to coalesce uncharges.
4153 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4154 * because we want to do uncharge as soon as possible.
4157 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4158 goto direct_uncharge;
4161 goto direct_uncharge;
4164 * In typical case, batch->memcg == mem. This means we can
4165 * merge a series of uncharges to an uncharge of res_counter.
4166 * If not, we uncharge res_counter ony by one.
4168 if (batch->memcg != memcg)
4169 goto direct_uncharge;
4170 /* remember freed charge and uncharge it later */
4173 batch->memsw_nr_pages++;
4176 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4178 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4179 if (unlikely(batch->memcg != memcg))
4180 memcg_oom_recover(memcg);
4184 * uncharge if !page_mapped(page)
4186 static struct mem_cgroup *
4187 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4190 struct mem_cgroup *memcg = NULL;
4191 unsigned int nr_pages = 1;
4192 struct page_cgroup *pc;
4195 if (mem_cgroup_disabled())
4198 if (PageTransHuge(page)) {
4199 nr_pages <<= compound_order(page);
4200 VM_BUG_ON(!PageTransHuge(page));
4203 * Check if our page_cgroup is valid
4205 pc = lookup_page_cgroup(page);
4206 if (unlikely(!PageCgroupUsed(pc)))
4209 lock_page_cgroup(pc);
4211 memcg = pc->mem_cgroup;
4213 if (!PageCgroupUsed(pc))
4216 anon = PageAnon(page);
4219 case MEM_CGROUP_CHARGE_TYPE_ANON:
4221 * Generally PageAnon tells if it's the anon statistics to be
4222 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4223 * used before page reached the stage of being marked PageAnon.
4227 case MEM_CGROUP_CHARGE_TYPE_DROP:
4228 /* See mem_cgroup_prepare_migration() */
4229 if (page_mapped(page))
4232 * Pages under migration may not be uncharged. But
4233 * end_migration() /must/ be the one uncharging the
4234 * unused post-migration page and so it has to call
4235 * here with the migration bit still set. See the
4236 * res_counter handling below.
4238 if (!end_migration && PageCgroupMigration(pc))
4241 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4242 if (!PageAnon(page)) { /* Shared memory */
4243 if (page->mapping && !page_is_file_cache(page))
4245 } else if (page_mapped(page)) /* Anon */
4252 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4254 ClearPageCgroupUsed(pc);
4256 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4257 * freed from LRU. This is safe because uncharged page is expected not
4258 * to be reused (freed soon). Exception is SwapCache, it's handled by
4259 * special functions.
4262 unlock_page_cgroup(pc);
4264 * even after unlock, we have memcg->res.usage here and this memcg
4265 * will never be freed, so it's safe to call css_get().
4267 memcg_check_events(memcg, page);
4268 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4269 mem_cgroup_swap_statistics(memcg, true);
4270 css_get(&memcg->css);
4273 * Migration does not charge the res_counter for the
4274 * replacement page, so leave it alone when phasing out the
4275 * page that is unused after the migration.
4277 if (!end_migration && !mem_cgroup_is_root(memcg))
4278 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4283 unlock_page_cgroup(pc);
4287 void mem_cgroup_uncharge_page(struct page *page)
4290 if (page_mapped(page))
4292 VM_BUG_ON(page->mapping && !PageAnon(page));
4294 * If the page is in swap cache, uncharge should be deferred
4295 * to the swap path, which also properly accounts swap usage
4296 * and handles memcg lifetime.
4298 * Note that this check is not stable and reclaim may add the
4299 * page to swap cache at any time after this. However, if the
4300 * page is not in swap cache by the time page->mapcount hits
4301 * 0, there won't be any page table references to the swap
4302 * slot, and reclaim will free it and not actually write the
4305 if (PageSwapCache(page))
4307 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4310 void mem_cgroup_uncharge_cache_page(struct page *page)
4312 VM_BUG_ON(page_mapped(page));
4313 VM_BUG_ON(page->mapping);
4314 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4318 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4319 * In that cases, pages are freed continuously and we can expect pages
4320 * are in the same memcg. All these calls itself limits the number of
4321 * pages freed at once, then uncharge_start/end() is called properly.
4322 * This may be called prural(2) times in a context,
4325 void mem_cgroup_uncharge_start(void)
4327 current->memcg_batch.do_batch++;
4328 /* We can do nest. */
4329 if (current->memcg_batch.do_batch == 1) {
4330 current->memcg_batch.memcg = NULL;
4331 current->memcg_batch.nr_pages = 0;
4332 current->memcg_batch.memsw_nr_pages = 0;
4336 void mem_cgroup_uncharge_end(void)
4338 struct memcg_batch_info *batch = ¤t->memcg_batch;
4340 if (!batch->do_batch)
4344 if (batch->do_batch) /* If stacked, do nothing. */
4350 * This "batch->memcg" is valid without any css_get/put etc...
4351 * bacause we hide charges behind us.
4353 if (batch->nr_pages)
4354 res_counter_uncharge(&batch->memcg->res,
4355 batch->nr_pages * PAGE_SIZE);
4356 if (batch->memsw_nr_pages)
4357 res_counter_uncharge(&batch->memcg->memsw,
4358 batch->memsw_nr_pages * PAGE_SIZE);
4359 memcg_oom_recover(batch->memcg);
4360 /* forget this pointer (for sanity check) */
4361 batch->memcg = NULL;
4366 * called after __delete_from_swap_cache() and drop "page" account.
4367 * memcg information is recorded to swap_cgroup of "ent"
4370 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4372 struct mem_cgroup *memcg;
4373 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4375 if (!swapout) /* this was a swap cache but the swap is unused ! */
4376 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4378 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4381 * record memcg information, if swapout && memcg != NULL,
4382 * css_get() was called in uncharge().
4384 if (do_swap_account && swapout && memcg)
4385 swap_cgroup_record(ent, css_id(&memcg->css));
4389 #ifdef CONFIG_MEMCG_SWAP
4391 * called from swap_entry_free(). remove record in swap_cgroup and
4392 * uncharge "memsw" account.
4394 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4396 struct mem_cgroup *memcg;
4399 if (!do_swap_account)
4402 id = swap_cgroup_record(ent, 0);
4404 memcg = mem_cgroup_lookup(id);
4407 * We uncharge this because swap is freed.
4408 * This memcg can be obsolete one. We avoid calling css_tryget
4410 if (!mem_cgroup_is_root(memcg))
4411 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4412 mem_cgroup_swap_statistics(memcg, false);
4413 css_put(&memcg->css);
4419 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4420 * @entry: swap entry to be moved
4421 * @from: mem_cgroup which the entry is moved from
4422 * @to: mem_cgroup which the entry is moved to
4424 * It succeeds only when the swap_cgroup's record for this entry is the same
4425 * as the mem_cgroup's id of @from.
4427 * Returns 0 on success, -EINVAL on failure.
4429 * The caller must have charged to @to, IOW, called res_counter_charge() about
4430 * both res and memsw, and called css_get().
4432 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4433 struct mem_cgroup *from, struct mem_cgroup *to)
4435 unsigned short old_id, new_id;
4437 old_id = css_id(&from->css);
4438 new_id = css_id(&to->css);
4440 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4441 mem_cgroup_swap_statistics(from, false);
4442 mem_cgroup_swap_statistics(to, true);
4444 * This function is only called from task migration context now.
4445 * It postpones res_counter and refcount handling till the end
4446 * of task migration(mem_cgroup_clear_mc()) for performance
4447 * improvement. But we cannot postpone css_get(to) because if
4448 * the process that has been moved to @to does swap-in, the
4449 * refcount of @to might be decreased to 0.
4451 * We are in attach() phase, so the cgroup is guaranteed to be
4452 * alive, so we can just call css_get().
4460 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4461 struct mem_cgroup *from, struct mem_cgroup *to)
4468 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4471 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4472 struct mem_cgroup **memcgp)
4474 struct mem_cgroup *memcg = NULL;
4475 unsigned int nr_pages = 1;
4476 struct page_cgroup *pc;
4477 enum charge_type ctype;
4481 if (mem_cgroup_disabled())
4484 if (PageTransHuge(page))
4485 nr_pages <<= compound_order(page);
4487 pc = lookup_page_cgroup(page);
4488 lock_page_cgroup(pc);
4489 if (PageCgroupUsed(pc)) {
4490 memcg = pc->mem_cgroup;
4491 css_get(&memcg->css);
4493 * At migrating an anonymous page, its mapcount goes down
4494 * to 0 and uncharge() will be called. But, even if it's fully
4495 * unmapped, migration may fail and this page has to be
4496 * charged again. We set MIGRATION flag here and delay uncharge
4497 * until end_migration() is called
4499 * Corner Case Thinking
4501 * When the old page was mapped as Anon and it's unmap-and-freed
4502 * while migration was ongoing.
4503 * If unmap finds the old page, uncharge() of it will be delayed
4504 * until end_migration(). If unmap finds a new page, it's
4505 * uncharged when it make mapcount to be 1->0. If unmap code
4506 * finds swap_migration_entry, the new page will not be mapped
4507 * and end_migration() will find it(mapcount==0).
4510 * When the old page was mapped but migraion fails, the kernel
4511 * remaps it. A charge for it is kept by MIGRATION flag even
4512 * if mapcount goes down to 0. We can do remap successfully
4513 * without charging it again.
4516 * The "old" page is under lock_page() until the end of
4517 * migration, so, the old page itself will not be swapped-out.
4518 * If the new page is swapped out before end_migraton, our
4519 * hook to usual swap-out path will catch the event.
4522 SetPageCgroupMigration(pc);
4524 unlock_page_cgroup(pc);
4526 * If the page is not charged at this point,
4534 * We charge new page before it's used/mapped. So, even if unlock_page()
4535 * is called before end_migration, we can catch all events on this new
4536 * page. In the case new page is migrated but not remapped, new page's
4537 * mapcount will be finally 0 and we call uncharge in end_migration().
4540 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4542 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4544 * The page is committed to the memcg, but it's not actually
4545 * charged to the res_counter since we plan on replacing the
4546 * old one and only one page is going to be left afterwards.
4548 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4551 /* remove redundant charge if migration failed*/
4552 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4553 struct page *oldpage, struct page *newpage, bool migration_ok)
4555 struct page *used, *unused;
4556 struct page_cgroup *pc;
4562 if (!migration_ok) {
4569 anon = PageAnon(used);
4570 __mem_cgroup_uncharge_common(unused,
4571 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4572 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4574 css_put(&memcg->css);
4576 * We disallowed uncharge of pages under migration because mapcount
4577 * of the page goes down to zero, temporarly.
4578 * Clear the flag and check the page should be charged.
4580 pc = lookup_page_cgroup(oldpage);
4581 lock_page_cgroup(pc);
4582 ClearPageCgroupMigration(pc);
4583 unlock_page_cgroup(pc);
4586 * If a page is a file cache, radix-tree replacement is very atomic
4587 * and we can skip this check. When it was an Anon page, its mapcount
4588 * goes down to 0. But because we added MIGRATION flage, it's not
4589 * uncharged yet. There are several case but page->mapcount check
4590 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4591 * check. (see prepare_charge() also)
4594 mem_cgroup_uncharge_page(used);
4598 * At replace page cache, newpage is not under any memcg but it's on
4599 * LRU. So, this function doesn't touch res_counter but handles LRU
4600 * in correct way. Both pages are locked so we cannot race with uncharge.
4602 void mem_cgroup_replace_page_cache(struct page *oldpage,
4603 struct page *newpage)
4605 struct mem_cgroup *memcg = NULL;
4606 struct page_cgroup *pc;
4607 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4609 if (mem_cgroup_disabled())
4612 pc = lookup_page_cgroup(oldpage);
4613 /* fix accounting on old pages */
4614 lock_page_cgroup(pc);
4615 if (PageCgroupUsed(pc)) {
4616 memcg = pc->mem_cgroup;
4617 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4618 ClearPageCgroupUsed(pc);
4620 unlock_page_cgroup(pc);
4623 * When called from shmem_replace_page(), in some cases the
4624 * oldpage has already been charged, and in some cases not.
4629 * Even if newpage->mapping was NULL before starting replacement,
4630 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4631 * LRU while we overwrite pc->mem_cgroup.
4633 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4636 #ifdef CONFIG_DEBUG_VM
4637 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4639 struct page_cgroup *pc;
4641 pc = lookup_page_cgroup(page);
4643 * Can be NULL while feeding pages into the page allocator for
4644 * the first time, i.e. during boot or memory hotplug;
4645 * or when mem_cgroup_disabled().
4647 if (likely(pc) && PageCgroupUsed(pc))
4652 bool mem_cgroup_bad_page_check(struct page *page)
4654 if (mem_cgroup_disabled())
4657 return lookup_page_cgroup_used(page) != NULL;
4660 void mem_cgroup_print_bad_page(struct page *page)
4662 struct page_cgroup *pc;
4664 pc = lookup_page_cgroup_used(page);
4666 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4667 pc, pc->flags, pc->mem_cgroup);
4672 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4673 unsigned long long val)
4676 u64 memswlimit, memlimit;
4678 int children = mem_cgroup_count_children(memcg);
4679 u64 curusage, oldusage;
4683 * For keeping hierarchical_reclaim simple, how long we should retry
4684 * is depends on callers. We set our retry-count to be function
4685 * of # of children which we should visit in this loop.
4687 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4689 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4692 while (retry_count) {
4693 if (signal_pending(current)) {
4698 * Rather than hide all in some function, I do this in
4699 * open coded manner. You see what this really does.
4700 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4702 mutex_lock(&set_limit_mutex);
4703 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4704 if (memswlimit < val) {
4706 mutex_unlock(&set_limit_mutex);
4710 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4714 ret = res_counter_set_limit(&memcg->res, val);
4716 if (memswlimit == val)
4717 memcg->memsw_is_minimum = true;
4719 memcg->memsw_is_minimum = false;
4721 mutex_unlock(&set_limit_mutex);
4726 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4727 MEM_CGROUP_RECLAIM_SHRINK);
4728 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4729 /* Usage is reduced ? */
4730 if (curusage >= oldusage)
4733 oldusage = curusage;
4735 if (!ret && enlarge)
4736 memcg_oom_recover(memcg);
4741 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4742 unsigned long long val)
4745 u64 memlimit, memswlimit, oldusage, curusage;
4746 int children = mem_cgroup_count_children(memcg);
4750 /* see mem_cgroup_resize_res_limit */
4751 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4752 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4753 while (retry_count) {
4754 if (signal_pending(current)) {
4759 * Rather than hide all in some function, I do this in
4760 * open coded manner. You see what this really does.
4761 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4763 mutex_lock(&set_limit_mutex);
4764 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4765 if (memlimit > val) {
4767 mutex_unlock(&set_limit_mutex);
4770 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4771 if (memswlimit < val)
4773 ret = res_counter_set_limit(&memcg->memsw, val);
4775 if (memlimit == val)
4776 memcg->memsw_is_minimum = true;
4778 memcg->memsw_is_minimum = false;
4780 mutex_unlock(&set_limit_mutex);
4785 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4786 MEM_CGROUP_RECLAIM_NOSWAP |
4787 MEM_CGROUP_RECLAIM_SHRINK);
4788 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4789 /* Usage is reduced ? */
4790 if (curusage >= oldusage)
4793 oldusage = curusage;
4795 if (!ret && enlarge)
4796 memcg_oom_recover(memcg);
4800 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4802 unsigned long *total_scanned)
4804 unsigned long nr_reclaimed = 0;
4805 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4806 unsigned long reclaimed;
4808 struct mem_cgroup_tree_per_zone *mctz;
4809 unsigned long long excess;
4810 unsigned long nr_scanned;
4815 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4817 * This loop can run a while, specially if mem_cgroup's continuously
4818 * keep exceeding their soft limit and putting the system under
4825 mz = mem_cgroup_largest_soft_limit_node(mctz);
4830 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4831 gfp_mask, &nr_scanned);
4832 nr_reclaimed += reclaimed;
4833 *total_scanned += nr_scanned;
4834 spin_lock(&mctz->lock);
4837 * If we failed to reclaim anything from this memory cgroup
4838 * it is time to move on to the next cgroup
4844 * Loop until we find yet another one.
4846 * By the time we get the soft_limit lock
4847 * again, someone might have aded the
4848 * group back on the RB tree. Iterate to
4849 * make sure we get a different mem.
4850 * mem_cgroup_largest_soft_limit_node returns
4851 * NULL if no other cgroup is present on
4855 __mem_cgroup_largest_soft_limit_node(mctz);
4857 css_put(&next_mz->memcg->css);
4858 else /* next_mz == NULL or other memcg */
4862 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4863 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4865 * One school of thought says that we should not add
4866 * back the node to the tree if reclaim returns 0.
4867 * But our reclaim could return 0, simply because due
4868 * to priority we are exposing a smaller subset of
4869 * memory to reclaim from. Consider this as a longer
4872 /* If excess == 0, no tree ops */
4873 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4874 spin_unlock(&mctz->lock);
4875 css_put(&mz->memcg->css);
4878 * Could not reclaim anything and there are no more
4879 * mem cgroups to try or we seem to be looping without
4880 * reclaiming anything.
4882 if (!nr_reclaimed &&
4884 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4886 } while (!nr_reclaimed);
4888 css_put(&next_mz->memcg->css);
4889 return nr_reclaimed;
4893 * mem_cgroup_force_empty_list - clears LRU of a group
4894 * @memcg: group to clear
4897 * @lru: lru to to clear
4899 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4900 * reclaim the pages page themselves - pages are moved to the parent (or root)
4903 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4904 int node, int zid, enum lru_list lru)
4906 struct lruvec *lruvec;
4907 unsigned long flags;
4908 struct list_head *list;
4912 zone = &NODE_DATA(node)->node_zones[zid];
4913 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4914 list = &lruvec->lists[lru];
4918 struct page_cgroup *pc;
4921 spin_lock_irqsave(&zone->lru_lock, flags);
4922 if (list_empty(list)) {
4923 spin_unlock_irqrestore(&zone->lru_lock, flags);
4926 page = list_entry(list->prev, struct page, lru);
4928 list_move(&page->lru, list);
4930 spin_unlock_irqrestore(&zone->lru_lock, flags);
4933 spin_unlock_irqrestore(&zone->lru_lock, flags);
4935 pc = lookup_page_cgroup(page);
4937 if (mem_cgroup_move_parent(page, pc, memcg)) {
4938 /* found lock contention or "pc" is obsolete. */
4943 } while (!list_empty(list));
4947 * make mem_cgroup's charge to be 0 if there is no task by moving
4948 * all the charges and pages to the parent.
4949 * This enables deleting this mem_cgroup.
4951 * Caller is responsible for holding css reference on the memcg.
4953 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4959 /* This is for making all *used* pages to be on LRU. */
4960 lru_add_drain_all();
4961 drain_all_stock_sync(memcg);
4962 mem_cgroup_start_move(memcg);
4963 for_each_node_state(node, N_MEMORY) {
4964 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4967 mem_cgroup_force_empty_list(memcg,
4972 mem_cgroup_end_move(memcg);
4973 memcg_oom_recover(memcg);
4977 * Kernel memory may not necessarily be trackable to a specific
4978 * process. So they are not migrated, and therefore we can't
4979 * expect their value to drop to 0 here.
4980 * Having res filled up with kmem only is enough.
4982 * This is a safety check because mem_cgroup_force_empty_list
4983 * could have raced with mem_cgroup_replace_page_cache callers
4984 * so the lru seemed empty but the page could have been added
4985 * right after the check. RES_USAGE should be safe as we always
4986 * charge before adding to the LRU.
4988 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4989 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4990 } while (usage > 0);
4993 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4995 lockdep_assert_held(&memcg_create_mutex);
4997 * The lock does not prevent addition or deletion to the list
4998 * of children, but it prevents a new child from being
4999 * initialized based on this parent in css_online(), so it's
5000 * enough to decide whether hierarchically inherited
5001 * attributes can still be changed or not.
5003 return memcg->use_hierarchy &&
5004 !list_empty(&memcg->css.cgroup->children);
5008 * Reclaims as many pages from the given memcg as possible and moves
5009 * the rest to the parent.
5011 * Caller is responsible for holding css reference for memcg.
5013 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
5015 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
5016 struct cgroup *cgrp = memcg->css.cgroup;
5018 /* returns EBUSY if there is a task or if we come here twice. */
5019 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
5022 /* we call try-to-free pages for make this cgroup empty */
5023 lru_add_drain_all();
5024 /* try to free all pages in this cgroup */
5025 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
5028 if (signal_pending(current))
5031 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5035 /* maybe some writeback is necessary */
5036 congestion_wait(BLK_RW_ASYNC, HZ/10);
5041 mem_cgroup_reparent_charges(memcg);
5046 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
5049 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5051 if (mem_cgroup_is_root(memcg))
5053 return mem_cgroup_force_empty(memcg);
5056 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5059 return mem_cgroup_from_css(css)->use_hierarchy;
5062 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5063 struct cftype *cft, u64 val)
5066 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5067 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5069 mutex_lock(&memcg_create_mutex);
5071 if (memcg->use_hierarchy == val)
5075 * If parent's use_hierarchy is set, we can't make any modifications
5076 * in the child subtrees. If it is unset, then the change can
5077 * occur, provided the current cgroup has no children.
5079 * For the root cgroup, parent_mem is NULL, we allow value to be
5080 * set if there are no children.
5082 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5083 (val == 1 || val == 0)) {
5084 if (list_empty(&memcg->css.cgroup->children))
5085 memcg->use_hierarchy = val;
5092 mutex_unlock(&memcg_create_mutex);
5098 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5099 enum mem_cgroup_stat_index idx)
5101 struct mem_cgroup *iter;
5104 /* Per-cpu values can be negative, use a signed accumulator */
5105 for_each_mem_cgroup_tree(iter, memcg)
5106 val += mem_cgroup_read_stat(iter, idx);
5108 if (val < 0) /* race ? */
5113 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5117 if (!mem_cgroup_is_root(memcg)) {
5119 return res_counter_read_u64(&memcg->res, RES_USAGE);
5121 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5125 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5126 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5128 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5129 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5132 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5134 return val << PAGE_SHIFT;
5137 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
5138 struct cftype *cft, struct file *file,
5139 char __user *buf, size_t nbytes, loff_t *ppos)
5141 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5147 type = MEMFILE_TYPE(cft->private);
5148 name = MEMFILE_ATTR(cft->private);
5152 if (name == RES_USAGE)
5153 val = mem_cgroup_usage(memcg, false);
5155 val = res_counter_read_u64(&memcg->res, name);
5158 if (name == RES_USAGE)
5159 val = mem_cgroup_usage(memcg, true);
5161 val = res_counter_read_u64(&memcg->memsw, name);
5164 val = res_counter_read_u64(&memcg->kmem, name);
5170 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5171 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5174 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
5177 #ifdef CONFIG_MEMCG_KMEM
5178 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5180 * For simplicity, we won't allow this to be disabled. It also can't
5181 * be changed if the cgroup has children already, or if tasks had
5184 * If tasks join before we set the limit, a person looking at
5185 * kmem.usage_in_bytes will have no way to determine when it took
5186 * place, which makes the value quite meaningless.
5188 * After it first became limited, changes in the value of the limit are
5189 * of course permitted.
5191 mutex_lock(&memcg_create_mutex);
5192 mutex_lock(&set_limit_mutex);
5193 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
5194 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
5198 ret = res_counter_set_limit(&memcg->kmem, val);
5201 ret = memcg_update_cache_sizes(memcg);
5203 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
5206 static_key_slow_inc(&memcg_kmem_enabled_key);
5208 * setting the active bit after the inc will guarantee no one
5209 * starts accounting before all call sites are patched
5211 memcg_kmem_set_active(memcg);
5213 ret = res_counter_set_limit(&memcg->kmem, val);
5215 mutex_unlock(&set_limit_mutex);
5216 mutex_unlock(&memcg_create_mutex);
5221 #ifdef CONFIG_MEMCG_KMEM
5222 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5225 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5229 memcg->kmem_account_flags = parent->kmem_account_flags;
5231 * When that happen, we need to disable the static branch only on those
5232 * memcgs that enabled it. To achieve this, we would be forced to
5233 * complicate the code by keeping track of which memcgs were the ones
5234 * that actually enabled limits, and which ones got it from its
5237 * It is a lot simpler just to do static_key_slow_inc() on every child
5238 * that is accounted.
5240 if (!memcg_kmem_is_active(memcg))
5244 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5245 * memcg is active already. If the later initialization fails then the
5246 * cgroup core triggers the cleanup so we do not have to do it here.
5248 static_key_slow_inc(&memcg_kmem_enabled_key);
5250 mutex_lock(&set_limit_mutex);
5251 memcg_stop_kmem_account();
5252 ret = memcg_update_cache_sizes(memcg);
5253 memcg_resume_kmem_account();
5254 mutex_unlock(&set_limit_mutex);
5258 #endif /* CONFIG_MEMCG_KMEM */
5261 * The user of this function is...
5264 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5267 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5270 unsigned long long val;
5273 type = MEMFILE_TYPE(cft->private);
5274 name = MEMFILE_ATTR(cft->private);
5278 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5282 /* This function does all necessary parse...reuse it */
5283 ret = res_counter_memparse_write_strategy(buffer, &val);
5287 ret = mem_cgroup_resize_limit(memcg, val);
5288 else if (type == _MEMSWAP)
5289 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5290 else if (type == _KMEM)
5291 ret = memcg_update_kmem_limit(css, val);
5295 case RES_SOFT_LIMIT:
5296 ret = res_counter_memparse_write_strategy(buffer, &val);
5300 * For memsw, soft limits are hard to implement in terms
5301 * of semantics, for now, we support soft limits for
5302 * control without swap
5305 ret = res_counter_set_soft_limit(&memcg->res, val);
5310 ret = -EINVAL; /* should be BUG() ? */
5316 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5317 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5319 unsigned long long min_limit, min_memsw_limit, tmp;
5321 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5322 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5323 if (!memcg->use_hierarchy)
5326 while (css_parent(&memcg->css)) {
5327 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5328 if (!memcg->use_hierarchy)
5330 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5331 min_limit = min(min_limit, tmp);
5332 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5333 min_memsw_limit = min(min_memsw_limit, tmp);
5336 *mem_limit = min_limit;
5337 *memsw_limit = min_memsw_limit;
5340 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5342 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5346 type = MEMFILE_TYPE(event);
5347 name = MEMFILE_ATTR(event);
5352 res_counter_reset_max(&memcg->res);
5353 else if (type == _MEMSWAP)
5354 res_counter_reset_max(&memcg->memsw);
5355 else if (type == _KMEM)
5356 res_counter_reset_max(&memcg->kmem);
5362 res_counter_reset_failcnt(&memcg->res);
5363 else if (type == _MEMSWAP)
5364 res_counter_reset_failcnt(&memcg->memsw);
5365 else if (type == _KMEM)
5366 res_counter_reset_failcnt(&memcg->kmem);
5375 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5378 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5382 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5383 struct cftype *cft, u64 val)
5385 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5387 if (val >= (1 << NR_MOVE_TYPE))
5391 * No kind of locking is needed in here, because ->can_attach() will
5392 * check this value once in the beginning of the process, and then carry
5393 * on with stale data. This means that changes to this value will only
5394 * affect task migrations starting after the change.
5396 memcg->move_charge_at_immigrate = val;
5400 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5401 struct cftype *cft, u64 val)
5408 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5409 struct cftype *cft, struct seq_file *m)
5412 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5413 unsigned long node_nr;
5414 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5416 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5417 seq_printf(m, "total=%lu", total_nr);
5418 for_each_node_state(nid, N_MEMORY) {
5419 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5420 seq_printf(m, " N%d=%lu", nid, node_nr);
5424 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5425 seq_printf(m, "file=%lu", file_nr);
5426 for_each_node_state(nid, N_MEMORY) {
5427 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5429 seq_printf(m, " N%d=%lu", nid, node_nr);
5433 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5434 seq_printf(m, "anon=%lu", anon_nr);
5435 for_each_node_state(nid, N_MEMORY) {
5436 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5438 seq_printf(m, " N%d=%lu", nid, node_nr);
5442 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5443 seq_printf(m, "unevictable=%lu", unevictable_nr);
5444 for_each_node_state(nid, N_MEMORY) {
5445 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5446 BIT(LRU_UNEVICTABLE));
5447 seq_printf(m, " N%d=%lu", nid, node_nr);
5452 #endif /* CONFIG_NUMA */
5454 static inline void mem_cgroup_lru_names_not_uptodate(void)
5456 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5459 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5462 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5463 struct mem_cgroup *mi;
5466 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5467 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5469 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5470 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5473 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5474 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5475 mem_cgroup_read_events(memcg, i));
5477 for (i = 0; i < NR_LRU_LISTS; i++)
5478 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5479 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5481 /* Hierarchical information */
5483 unsigned long long limit, memsw_limit;
5484 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5485 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5486 if (do_swap_account)
5487 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5491 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5494 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5496 for_each_mem_cgroup_tree(mi, memcg)
5497 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5498 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5501 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5502 unsigned long long val = 0;
5504 for_each_mem_cgroup_tree(mi, memcg)
5505 val += mem_cgroup_read_events(mi, i);
5506 seq_printf(m, "total_%s %llu\n",
5507 mem_cgroup_events_names[i], val);
5510 for (i = 0; i < NR_LRU_LISTS; i++) {
5511 unsigned long long val = 0;
5513 for_each_mem_cgroup_tree(mi, memcg)
5514 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5515 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5518 #ifdef CONFIG_DEBUG_VM
5521 struct mem_cgroup_per_zone *mz;
5522 struct zone_reclaim_stat *rstat;
5523 unsigned long recent_rotated[2] = {0, 0};
5524 unsigned long recent_scanned[2] = {0, 0};
5526 for_each_online_node(nid)
5527 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5528 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5529 rstat = &mz->lruvec.reclaim_stat;
5531 recent_rotated[0] += rstat->recent_rotated[0];
5532 recent_rotated[1] += rstat->recent_rotated[1];
5533 recent_scanned[0] += rstat->recent_scanned[0];
5534 recent_scanned[1] += rstat->recent_scanned[1];
5536 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5537 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5538 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5539 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5546 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5549 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5551 return mem_cgroup_swappiness(memcg);
5554 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5555 struct cftype *cft, u64 val)
5557 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5558 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5560 if (val > 100 || !parent)
5563 mutex_lock(&memcg_create_mutex);
5565 /* If under hierarchy, only empty-root can set this value */
5566 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5567 mutex_unlock(&memcg_create_mutex);
5571 memcg->swappiness = val;
5573 mutex_unlock(&memcg_create_mutex);
5578 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5580 struct mem_cgroup_threshold_ary *t;
5586 t = rcu_dereference(memcg->thresholds.primary);
5588 t = rcu_dereference(memcg->memsw_thresholds.primary);
5593 usage = mem_cgroup_usage(memcg, swap);
5596 * current_threshold points to threshold just below or equal to usage.
5597 * If it's not true, a threshold was crossed after last
5598 * call of __mem_cgroup_threshold().
5600 i = t->current_threshold;
5603 * Iterate backward over array of thresholds starting from
5604 * current_threshold and check if a threshold is crossed.
5605 * If none of thresholds below usage is crossed, we read
5606 * only one element of the array here.
5608 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5609 eventfd_signal(t->entries[i].eventfd, 1);
5611 /* i = current_threshold + 1 */
5615 * Iterate forward over array of thresholds starting from
5616 * current_threshold+1 and check if a threshold is crossed.
5617 * If none of thresholds above usage is crossed, we read
5618 * only one element of the array here.
5620 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5621 eventfd_signal(t->entries[i].eventfd, 1);
5623 /* Update current_threshold */
5624 t->current_threshold = i - 1;
5629 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5632 __mem_cgroup_threshold(memcg, false);
5633 if (do_swap_account)
5634 __mem_cgroup_threshold(memcg, true);
5636 memcg = parent_mem_cgroup(memcg);
5640 static int compare_thresholds(const void *a, const void *b)
5642 const struct mem_cgroup_threshold *_a = a;
5643 const struct mem_cgroup_threshold *_b = b;
5645 if (_a->threshold > _b->threshold)
5648 if (_a->threshold < _b->threshold)
5654 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5656 struct mem_cgroup_eventfd_list *ev;
5658 list_for_each_entry(ev, &memcg->oom_notify, list)
5659 eventfd_signal(ev->eventfd, 1);
5663 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5665 struct mem_cgroup *iter;
5667 for_each_mem_cgroup_tree(iter, memcg)
5668 mem_cgroup_oom_notify_cb(iter);
5671 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5672 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5674 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5675 struct mem_cgroup_thresholds *thresholds;
5676 struct mem_cgroup_threshold_ary *new;
5677 enum res_type type = MEMFILE_TYPE(cft->private);
5678 u64 threshold, usage;
5681 ret = res_counter_memparse_write_strategy(args, &threshold);
5685 mutex_lock(&memcg->thresholds_lock);
5688 thresholds = &memcg->thresholds;
5689 else if (type == _MEMSWAP)
5690 thresholds = &memcg->memsw_thresholds;
5694 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5696 /* Check if a threshold crossed before adding a new one */
5697 if (thresholds->primary)
5698 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5700 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5702 /* Allocate memory for new array of thresholds */
5703 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5711 /* Copy thresholds (if any) to new array */
5712 if (thresholds->primary) {
5713 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5714 sizeof(struct mem_cgroup_threshold));
5717 /* Add new threshold */
5718 new->entries[size - 1].eventfd = eventfd;
5719 new->entries[size - 1].threshold = threshold;
5721 /* Sort thresholds. Registering of new threshold isn't time-critical */
5722 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5723 compare_thresholds, NULL);
5725 /* Find current threshold */
5726 new->current_threshold = -1;
5727 for (i = 0; i < size; i++) {
5728 if (new->entries[i].threshold <= usage) {
5730 * new->current_threshold will not be used until
5731 * rcu_assign_pointer(), so it's safe to increment
5734 ++new->current_threshold;
5739 /* Free old spare buffer and save old primary buffer as spare */
5740 kfree(thresholds->spare);
5741 thresholds->spare = thresholds->primary;
5743 rcu_assign_pointer(thresholds->primary, new);
5745 /* To be sure that nobody uses thresholds */
5749 mutex_unlock(&memcg->thresholds_lock);
5754 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5755 struct cftype *cft, struct eventfd_ctx *eventfd)
5757 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5758 struct mem_cgroup_thresholds *thresholds;
5759 struct mem_cgroup_threshold_ary *new;
5760 enum res_type type = MEMFILE_TYPE(cft->private);
5764 mutex_lock(&memcg->thresholds_lock);
5766 thresholds = &memcg->thresholds;
5767 else if (type == _MEMSWAP)
5768 thresholds = &memcg->memsw_thresholds;
5772 if (!thresholds->primary)
5775 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5777 /* Check if a threshold crossed before removing */
5778 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5780 /* Calculate new number of threshold */
5782 for (i = 0; i < thresholds->primary->size; i++) {
5783 if (thresholds->primary->entries[i].eventfd != eventfd)
5787 new = thresholds->spare;
5789 /* Set thresholds array to NULL if we don't have thresholds */
5798 /* Copy thresholds and find current threshold */
5799 new->current_threshold = -1;
5800 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5801 if (thresholds->primary->entries[i].eventfd == eventfd)
5804 new->entries[j] = thresholds->primary->entries[i];
5805 if (new->entries[j].threshold <= usage) {
5807 * new->current_threshold will not be used
5808 * until rcu_assign_pointer(), so it's safe to increment
5811 ++new->current_threshold;
5817 /* Swap primary and spare array */
5818 thresholds->spare = thresholds->primary;
5819 /* If all events are unregistered, free the spare array */
5821 kfree(thresholds->spare);
5822 thresholds->spare = NULL;
5825 rcu_assign_pointer(thresholds->primary, new);
5827 /* To be sure that nobody uses thresholds */
5830 mutex_unlock(&memcg->thresholds_lock);
5833 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5834 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5836 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5837 struct mem_cgroup_eventfd_list *event;
5838 enum res_type type = MEMFILE_TYPE(cft->private);
5840 BUG_ON(type != _OOM_TYPE);
5841 event = kmalloc(sizeof(*event), GFP_KERNEL);
5845 spin_lock(&memcg_oom_lock);
5847 event->eventfd = eventfd;
5848 list_add(&event->list, &memcg->oom_notify);
5850 /* already in OOM ? */
5851 if (atomic_read(&memcg->under_oom))
5852 eventfd_signal(eventfd, 1);
5853 spin_unlock(&memcg_oom_lock);
5858 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5859 struct cftype *cft, struct eventfd_ctx *eventfd)
5861 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5862 struct mem_cgroup_eventfd_list *ev, *tmp;
5863 enum res_type type = MEMFILE_TYPE(cft->private);
5865 BUG_ON(type != _OOM_TYPE);
5867 spin_lock(&memcg_oom_lock);
5869 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5870 if (ev->eventfd == eventfd) {
5871 list_del(&ev->list);
5876 spin_unlock(&memcg_oom_lock);
5879 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5880 struct cftype *cft, struct cgroup_map_cb *cb)
5882 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5884 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5886 if (atomic_read(&memcg->under_oom))
5887 cb->fill(cb, "under_oom", 1);
5889 cb->fill(cb, "under_oom", 0);
5893 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5894 struct cftype *cft, u64 val)
5896 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5897 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5899 /* cannot set to root cgroup and only 0 and 1 are allowed */
5900 if (!parent || !((val == 0) || (val == 1)))
5903 mutex_lock(&memcg_create_mutex);
5904 /* oom-kill-disable is a flag for subhierarchy. */
5905 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5906 mutex_unlock(&memcg_create_mutex);
5909 memcg->oom_kill_disable = val;
5911 memcg_oom_recover(memcg);
5912 mutex_unlock(&memcg_create_mutex);
5916 #ifdef CONFIG_MEMCG_KMEM
5917 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5921 memcg->kmemcg_id = -1;
5922 ret = memcg_propagate_kmem(memcg);
5926 return mem_cgroup_sockets_init(memcg, ss);
5929 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5931 mem_cgroup_sockets_destroy(memcg);
5934 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5936 if (!memcg_kmem_is_active(memcg))
5940 * kmem charges can outlive the cgroup. In the case of slab
5941 * pages, for instance, a page contain objects from various
5942 * processes. As we prevent from taking a reference for every
5943 * such allocation we have to be careful when doing uncharge
5944 * (see memcg_uncharge_kmem) and here during offlining.
5946 * The idea is that that only the _last_ uncharge which sees
5947 * the dead memcg will drop the last reference. An additional
5948 * reference is taken here before the group is marked dead
5949 * which is then paired with css_put during uncharge resp. here.
5951 * Although this might sound strange as this path is called from
5952 * css_offline() when the referencemight have dropped down to 0
5953 * and shouldn't be incremented anymore (css_tryget would fail)
5954 * we do not have other options because of the kmem allocations
5957 css_get(&memcg->css);
5959 memcg_kmem_mark_dead(memcg);
5961 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5964 if (memcg_kmem_test_and_clear_dead(memcg))
5965 css_put(&memcg->css);
5968 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5973 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5977 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5983 * Unregister event and free resources.
5985 * Gets called from workqueue.
5987 static void cgroup_event_remove(struct work_struct *work)
5989 struct cgroup_event *event = container_of(work, struct cgroup_event,
5991 struct cgroup_subsys_state *css = event->css;
5993 remove_wait_queue(event->wqh, &event->wait);
5995 event->cft->unregister_event(css, event->cft, event->eventfd);
5997 /* Notify userspace the event is going away. */
5998 eventfd_signal(event->eventfd, 1);
6000 eventfd_ctx_put(event->eventfd);
6006 * Gets called on POLLHUP on eventfd when user closes it.
6008 * Called with wqh->lock held and interrupts disabled.
6010 static int cgroup_event_wake(wait_queue_t *wait, unsigned mode,
6011 int sync, void *key)
6013 struct cgroup_event *event = container_of(wait,
6014 struct cgroup_event, wait);
6015 struct cgroup *cgrp = event->css->cgroup;
6016 unsigned long flags = (unsigned long)key;
6018 if (flags & POLLHUP) {
6020 * If the event has been detached at cgroup removal, we
6021 * can simply return knowing the other side will cleanup
6024 * We can't race against event freeing since the other
6025 * side will require wqh->lock via remove_wait_queue(),
6028 spin_lock(&cgrp->event_list_lock);
6029 if (!list_empty(&event->list)) {
6030 list_del_init(&event->list);
6032 * We are in atomic context, but cgroup_event_remove()
6033 * may sleep, so we have to call it in workqueue.
6035 schedule_work(&event->remove);
6037 spin_unlock(&cgrp->event_list_lock);
6043 static void cgroup_event_ptable_queue_proc(struct file *file,
6044 wait_queue_head_t *wqh, poll_table *pt)
6046 struct cgroup_event *event = container_of(pt,
6047 struct cgroup_event, pt);
6050 add_wait_queue(wqh, &event->wait);
6054 * Parse input and register new cgroup event handler.
6056 * Input must be in format '<event_fd> <control_fd> <args>'.
6057 * Interpretation of args is defined by control file implementation.
6059 static int cgroup_write_event_control(struct cgroup_subsys_state *dummy_css,
6060 struct cftype *cft, const char *buffer)
6062 struct cgroup *cgrp = dummy_css->cgroup;
6063 struct cgroup_event *event;
6064 struct cgroup_subsys_state *cfile_css;
6065 unsigned int efd, cfd;
6071 efd = simple_strtoul(buffer, &endp, 10);
6076 cfd = simple_strtoul(buffer, &endp, 10);
6077 if ((*endp != ' ') && (*endp != '\0'))
6081 event = kzalloc(sizeof(*event), GFP_KERNEL);
6085 INIT_LIST_HEAD(&event->list);
6086 init_poll_funcptr(&event->pt, cgroup_event_ptable_queue_proc);
6087 init_waitqueue_func_entry(&event->wait, cgroup_event_wake);
6088 INIT_WORK(&event->remove, cgroup_event_remove);
6096 event->eventfd = eventfd_ctx_fileget(efile.file);
6097 if (IS_ERR(event->eventfd)) {
6098 ret = PTR_ERR(event->eventfd);
6105 goto out_put_eventfd;
6108 /* the process need read permission on control file */
6109 /* AV: shouldn't we check that it's been opened for read instead? */
6110 ret = inode_permission(file_inode(cfile.file), MAY_READ);
6114 event->cft = __file_cft(cfile.file);
6115 if (IS_ERR(event->cft)) {
6116 ret = PTR_ERR(event->cft);
6120 if (!event->cft->ss) {
6126 * Determine the css of @cfile, verify it belongs to the same
6127 * cgroup as cgroup.event_control, and associate @event with it.
6128 * Remaining events are automatically removed on cgroup destruction
6129 * but the removal is asynchronous, so take an extra ref.
6134 event->css = cgroup_css(cgrp, event->cft->ss);
6135 cfile_css = css_from_dir(cfile.file->f_dentry->d_parent, event->cft->ss);
6136 if (event->css && event->css == cfile_css && css_tryget(event->css))
6143 if (!event->cft->register_event || !event->cft->unregister_event) {
6148 ret = event->cft->register_event(event->css, event->cft,
6149 event->eventfd, buffer);
6153 efile.file->f_op->poll(efile.file, &event->pt);
6155 spin_lock(&cgrp->event_list_lock);
6156 list_add(&event->list, &cgrp->event_list);
6157 spin_unlock(&cgrp->event_list_lock);
6165 css_put(event->css);
6169 eventfd_ctx_put(event->eventfd);
6178 static struct cftype mem_cgroup_files[] = {
6180 .name = "usage_in_bytes",
6181 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
6182 .read = mem_cgroup_read,
6183 .register_event = mem_cgroup_usage_register_event,
6184 .unregister_event = mem_cgroup_usage_unregister_event,
6187 .name = "max_usage_in_bytes",
6188 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6189 .trigger = mem_cgroup_reset,
6190 .read = mem_cgroup_read,
6193 .name = "limit_in_bytes",
6194 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6195 .write_string = mem_cgroup_write,
6196 .read = mem_cgroup_read,
6199 .name = "soft_limit_in_bytes",
6200 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6201 .write_string = mem_cgroup_write,
6202 .read = mem_cgroup_read,
6206 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6207 .trigger = mem_cgroup_reset,
6208 .read = mem_cgroup_read,
6212 .read_seq_string = memcg_stat_show,
6215 .name = "force_empty",
6216 .trigger = mem_cgroup_force_empty_write,
6219 .name = "use_hierarchy",
6220 .flags = CFTYPE_INSANE,
6221 .write_u64 = mem_cgroup_hierarchy_write,
6222 .read_u64 = mem_cgroup_hierarchy_read,
6225 .name = "cgroup.event_control",
6226 .write_string = cgroup_write_event_control,
6227 .flags = CFTYPE_NO_PREFIX,
6231 .name = "swappiness",
6232 .read_u64 = mem_cgroup_swappiness_read,
6233 .write_u64 = mem_cgroup_swappiness_write,
6236 .name = "move_charge_at_immigrate",
6237 .read_u64 = mem_cgroup_move_charge_read,
6238 .write_u64 = mem_cgroup_move_charge_write,
6241 .name = "oom_control",
6242 .read_map = mem_cgroup_oom_control_read,
6243 .write_u64 = mem_cgroup_oom_control_write,
6244 .register_event = mem_cgroup_oom_register_event,
6245 .unregister_event = mem_cgroup_oom_unregister_event,
6246 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6249 .name = "pressure_level",
6250 .register_event = vmpressure_register_event,
6251 .unregister_event = vmpressure_unregister_event,
6255 .name = "numa_stat",
6256 .read_seq_string = memcg_numa_stat_show,
6259 #ifdef CONFIG_MEMCG_KMEM
6261 .name = "kmem.limit_in_bytes",
6262 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6263 .write_string = mem_cgroup_write,
6264 .read = mem_cgroup_read,
6267 .name = "kmem.usage_in_bytes",
6268 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6269 .read = mem_cgroup_read,
6272 .name = "kmem.failcnt",
6273 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6274 .trigger = mem_cgroup_reset,
6275 .read = mem_cgroup_read,
6278 .name = "kmem.max_usage_in_bytes",
6279 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6280 .trigger = mem_cgroup_reset,
6281 .read = mem_cgroup_read,
6283 #ifdef CONFIG_SLABINFO
6285 .name = "kmem.slabinfo",
6286 .read_seq_string = mem_cgroup_slabinfo_read,
6290 { }, /* terminate */
6293 #ifdef CONFIG_MEMCG_SWAP
6294 static struct cftype memsw_cgroup_files[] = {
6296 .name = "memsw.usage_in_bytes",
6297 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6298 .read = mem_cgroup_read,
6299 .register_event = mem_cgroup_usage_register_event,
6300 .unregister_event = mem_cgroup_usage_unregister_event,
6303 .name = "memsw.max_usage_in_bytes",
6304 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6305 .trigger = mem_cgroup_reset,
6306 .read = mem_cgroup_read,
6309 .name = "memsw.limit_in_bytes",
6310 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6311 .write_string = mem_cgroup_write,
6312 .read = mem_cgroup_read,
6315 .name = "memsw.failcnt",
6316 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6317 .trigger = mem_cgroup_reset,
6318 .read = mem_cgroup_read,
6320 { }, /* terminate */
6323 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6325 struct mem_cgroup_per_node *pn;
6326 struct mem_cgroup_per_zone *mz;
6327 int zone, tmp = node;
6329 * This routine is called against possible nodes.
6330 * But it's BUG to call kmalloc() against offline node.
6332 * TODO: this routine can waste much memory for nodes which will
6333 * never be onlined. It's better to use memory hotplug callback
6336 if (!node_state(node, N_NORMAL_MEMORY))
6338 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6342 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6343 mz = &pn->zoneinfo[zone];
6344 lruvec_init(&mz->lruvec);
6345 mz->usage_in_excess = 0;
6346 mz->on_tree = false;
6349 memcg->nodeinfo[node] = pn;
6353 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6355 kfree(memcg->nodeinfo[node]);
6358 static struct mem_cgroup *mem_cgroup_alloc(void)
6360 struct mem_cgroup *memcg;
6361 size_t size = memcg_size();
6363 /* Can be very big if nr_node_ids is very big */
6364 if (size < PAGE_SIZE)
6365 memcg = kzalloc(size, GFP_KERNEL);
6367 memcg = vzalloc(size);
6372 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6375 spin_lock_init(&memcg->pcp_counter_lock);
6379 if (size < PAGE_SIZE)
6387 * At destroying mem_cgroup, references from swap_cgroup can remain.
6388 * (scanning all at force_empty is too costly...)
6390 * Instead of clearing all references at force_empty, we remember
6391 * the number of reference from swap_cgroup and free mem_cgroup when
6392 * it goes down to 0.
6394 * Removal of cgroup itself succeeds regardless of refs from swap.
6397 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6400 size_t size = memcg_size();
6402 mem_cgroup_remove_from_trees(memcg);
6403 free_css_id(&mem_cgroup_subsys, &memcg->css);
6406 free_mem_cgroup_per_zone_info(memcg, node);
6408 free_percpu(memcg->stat);
6411 * We need to make sure that (at least for now), the jump label
6412 * destruction code runs outside of the cgroup lock. This is because
6413 * get_online_cpus(), which is called from the static_branch update,
6414 * can't be called inside the cgroup_lock. cpusets are the ones
6415 * enforcing this dependency, so if they ever change, we might as well.
6417 * schedule_work() will guarantee this happens. Be careful if you need
6418 * to move this code around, and make sure it is outside
6421 disarm_static_keys(memcg);
6422 if (size < PAGE_SIZE)
6429 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6431 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6433 if (!memcg->res.parent)
6435 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6437 EXPORT_SYMBOL(parent_mem_cgroup);
6439 static void __init mem_cgroup_soft_limit_tree_init(void)
6441 struct mem_cgroup_tree_per_node *rtpn;
6442 struct mem_cgroup_tree_per_zone *rtpz;
6443 int tmp, node, zone;
6445 for_each_node(node) {
6447 if (!node_state(node, N_NORMAL_MEMORY))
6449 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6452 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6454 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6455 rtpz = &rtpn->rb_tree_per_zone[zone];
6456 rtpz->rb_root = RB_ROOT;
6457 spin_lock_init(&rtpz->lock);
6462 static struct cgroup_subsys_state * __ref
6463 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6465 struct mem_cgroup *memcg;
6466 long error = -ENOMEM;
6469 memcg = mem_cgroup_alloc();
6471 return ERR_PTR(error);
6474 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6478 if (parent_css == NULL) {
6479 root_mem_cgroup = memcg;
6480 res_counter_init(&memcg->res, NULL);
6481 res_counter_init(&memcg->memsw, NULL);
6482 res_counter_init(&memcg->kmem, NULL);
6485 memcg->last_scanned_node = MAX_NUMNODES;
6486 INIT_LIST_HEAD(&memcg->oom_notify);
6487 memcg->move_charge_at_immigrate = 0;
6488 mutex_init(&memcg->thresholds_lock);
6489 spin_lock_init(&memcg->move_lock);
6490 vmpressure_init(&memcg->vmpressure);
6495 __mem_cgroup_free(memcg);
6496 return ERR_PTR(error);
6500 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6502 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6503 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6509 mutex_lock(&memcg_create_mutex);
6511 memcg->use_hierarchy = parent->use_hierarchy;
6512 memcg->oom_kill_disable = parent->oom_kill_disable;
6513 memcg->swappiness = mem_cgroup_swappiness(parent);
6515 if (parent->use_hierarchy) {
6516 res_counter_init(&memcg->res, &parent->res);
6517 res_counter_init(&memcg->memsw, &parent->memsw);
6518 res_counter_init(&memcg->kmem, &parent->kmem);
6521 * No need to take a reference to the parent because cgroup
6522 * core guarantees its existence.
6525 res_counter_init(&memcg->res, NULL);
6526 res_counter_init(&memcg->memsw, NULL);
6527 res_counter_init(&memcg->kmem, NULL);
6529 * Deeper hierachy with use_hierarchy == false doesn't make
6530 * much sense so let cgroup subsystem know about this
6531 * unfortunate state in our controller.
6533 if (parent != root_mem_cgroup)
6534 mem_cgroup_subsys.broken_hierarchy = true;
6537 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6538 mutex_unlock(&memcg_create_mutex);
6543 * Announce all parents that a group from their hierarchy is gone.
6545 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6547 struct mem_cgroup *parent = memcg;
6549 while ((parent = parent_mem_cgroup(parent)))
6550 mem_cgroup_iter_invalidate(parent);
6553 * if the root memcg is not hierarchical we have to check it
6556 if (!root_mem_cgroup->use_hierarchy)
6557 mem_cgroup_iter_invalidate(root_mem_cgroup);
6560 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6562 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6563 struct cgroup *cgrp = css->cgroup;
6564 struct cgroup_event *event, *tmp;
6567 * Unregister events and notify userspace.
6568 * Notify userspace about cgroup removing only after rmdir of cgroup
6569 * directory to avoid race between userspace and kernelspace.
6571 spin_lock(&cgrp->event_list_lock);
6572 list_for_each_entry_safe(event, tmp, &cgrp->event_list, list) {
6573 list_del_init(&event->list);
6574 schedule_work(&event->remove);
6576 spin_unlock(&cgrp->event_list_lock);
6578 kmem_cgroup_css_offline(memcg);
6580 mem_cgroup_invalidate_reclaim_iterators(memcg);
6581 mem_cgroup_reparent_charges(memcg);
6582 mem_cgroup_destroy_all_caches(memcg);
6583 vmpressure_cleanup(&memcg->vmpressure);
6586 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6588 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6590 memcg_destroy_kmem(memcg);
6591 __mem_cgroup_free(memcg);
6595 /* Handlers for move charge at task migration. */
6596 #define PRECHARGE_COUNT_AT_ONCE 256
6597 static int mem_cgroup_do_precharge(unsigned long count)
6600 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6601 struct mem_cgroup *memcg = mc.to;
6603 if (mem_cgroup_is_root(memcg)) {
6604 mc.precharge += count;
6605 /* we don't need css_get for root */
6608 /* try to charge at once */
6610 struct res_counter *dummy;
6612 * "memcg" cannot be under rmdir() because we've already checked
6613 * by cgroup_lock_live_cgroup() that it is not removed and we
6614 * are still under the same cgroup_mutex. So we can postpone
6617 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6619 if (do_swap_account && res_counter_charge(&memcg->memsw,
6620 PAGE_SIZE * count, &dummy)) {
6621 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6624 mc.precharge += count;
6628 /* fall back to one by one charge */
6630 if (signal_pending(current)) {
6634 if (!batch_count--) {
6635 batch_count = PRECHARGE_COUNT_AT_ONCE;
6638 ret = __mem_cgroup_try_charge(NULL,
6639 GFP_KERNEL, 1, &memcg, false);
6641 /* mem_cgroup_clear_mc() will do uncharge later */
6649 * get_mctgt_type - get target type of moving charge
6650 * @vma: the vma the pte to be checked belongs
6651 * @addr: the address corresponding to the pte to be checked
6652 * @ptent: the pte to be checked
6653 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6656 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6657 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6658 * move charge. if @target is not NULL, the page is stored in target->page
6659 * with extra refcnt got(Callers should handle it).
6660 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6661 * target for charge migration. if @target is not NULL, the entry is stored
6664 * Called with pte lock held.
6671 enum mc_target_type {
6677 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6678 unsigned long addr, pte_t ptent)
6680 struct page *page = vm_normal_page(vma, addr, ptent);
6682 if (!page || !page_mapped(page))
6684 if (PageAnon(page)) {
6685 /* we don't move shared anon */
6688 } else if (!move_file())
6689 /* we ignore mapcount for file pages */
6691 if (!get_page_unless_zero(page))
6698 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6699 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6701 struct page *page = NULL;
6702 swp_entry_t ent = pte_to_swp_entry(ptent);
6704 if (!move_anon() || non_swap_entry(ent))
6707 * Because lookup_swap_cache() updates some statistics counter,
6708 * we call find_get_page() with swapper_space directly.
6710 page = find_get_page(swap_address_space(ent), ent.val);
6711 if (do_swap_account)
6712 entry->val = ent.val;
6717 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6718 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6724 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6725 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6727 struct page *page = NULL;
6728 struct address_space *mapping;
6731 if (!vma->vm_file) /* anonymous vma */
6736 mapping = vma->vm_file->f_mapping;
6737 if (pte_none(ptent))
6738 pgoff = linear_page_index(vma, addr);
6739 else /* pte_file(ptent) is true */
6740 pgoff = pte_to_pgoff(ptent);
6742 /* page is moved even if it's not RSS of this task(page-faulted). */
6743 page = find_get_page(mapping, pgoff);
6746 /* shmem/tmpfs may report page out on swap: account for that too. */
6747 if (radix_tree_exceptional_entry(page)) {
6748 swp_entry_t swap = radix_to_swp_entry(page);
6749 if (do_swap_account)
6751 page = find_get_page(swap_address_space(swap), swap.val);
6757 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6758 unsigned long addr, pte_t ptent, union mc_target *target)
6760 struct page *page = NULL;
6761 struct page_cgroup *pc;
6762 enum mc_target_type ret = MC_TARGET_NONE;
6763 swp_entry_t ent = { .val = 0 };
6765 if (pte_present(ptent))
6766 page = mc_handle_present_pte(vma, addr, ptent);
6767 else if (is_swap_pte(ptent))
6768 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6769 else if (pte_none(ptent) || pte_file(ptent))
6770 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6772 if (!page && !ent.val)
6775 pc = lookup_page_cgroup(page);
6777 * Do only loose check w/o page_cgroup lock.
6778 * mem_cgroup_move_account() checks the pc is valid or not under
6781 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6782 ret = MC_TARGET_PAGE;
6784 target->page = page;
6786 if (!ret || !target)
6789 /* There is a swap entry and a page doesn't exist or isn't charged */
6790 if (ent.val && !ret &&
6791 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6792 ret = MC_TARGET_SWAP;
6799 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6801 * We don't consider swapping or file mapped pages because THP does not
6802 * support them for now.
6803 * Caller should make sure that pmd_trans_huge(pmd) is true.
6805 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6806 unsigned long addr, pmd_t pmd, union mc_target *target)
6808 struct page *page = NULL;
6809 struct page_cgroup *pc;
6810 enum mc_target_type ret = MC_TARGET_NONE;
6812 page = pmd_page(pmd);
6813 VM_BUG_ON(!page || !PageHead(page));
6816 pc = lookup_page_cgroup(page);
6817 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6818 ret = MC_TARGET_PAGE;
6821 target->page = page;
6827 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6828 unsigned long addr, pmd_t pmd, union mc_target *target)
6830 return MC_TARGET_NONE;
6834 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6835 unsigned long addr, unsigned long end,
6836 struct mm_walk *walk)
6838 struct vm_area_struct *vma = walk->private;
6842 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6843 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6844 mc.precharge += HPAGE_PMD_NR;
6845 spin_unlock(&vma->vm_mm->page_table_lock);
6849 if (pmd_trans_unstable(pmd))
6851 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6852 for (; addr != end; pte++, addr += PAGE_SIZE)
6853 if (get_mctgt_type(vma, addr, *pte, NULL))
6854 mc.precharge++; /* increment precharge temporarily */
6855 pte_unmap_unlock(pte - 1, ptl);
6861 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6863 unsigned long precharge;
6864 struct vm_area_struct *vma;
6866 down_read(&mm->mmap_sem);
6867 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6868 struct mm_walk mem_cgroup_count_precharge_walk = {
6869 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6873 if (is_vm_hugetlb_page(vma))
6875 walk_page_range(vma->vm_start, vma->vm_end,
6876 &mem_cgroup_count_precharge_walk);
6878 up_read(&mm->mmap_sem);
6880 precharge = mc.precharge;
6886 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6888 unsigned long precharge = mem_cgroup_count_precharge(mm);
6890 VM_BUG_ON(mc.moving_task);
6891 mc.moving_task = current;
6892 return mem_cgroup_do_precharge(precharge);
6895 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6896 static void __mem_cgroup_clear_mc(void)
6898 struct mem_cgroup *from = mc.from;
6899 struct mem_cgroup *to = mc.to;
6902 /* we must uncharge all the leftover precharges from mc.to */
6904 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6908 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6909 * we must uncharge here.
6911 if (mc.moved_charge) {
6912 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6913 mc.moved_charge = 0;
6915 /* we must fixup refcnts and charges */
6916 if (mc.moved_swap) {
6917 /* uncharge swap account from the old cgroup */
6918 if (!mem_cgroup_is_root(mc.from))
6919 res_counter_uncharge(&mc.from->memsw,
6920 PAGE_SIZE * mc.moved_swap);
6922 for (i = 0; i < mc.moved_swap; i++)
6923 css_put(&mc.from->css);
6925 if (!mem_cgroup_is_root(mc.to)) {
6927 * we charged both to->res and to->memsw, so we should
6930 res_counter_uncharge(&mc.to->res,
6931 PAGE_SIZE * mc.moved_swap);
6933 /* we've already done css_get(mc.to) */
6936 memcg_oom_recover(from);
6937 memcg_oom_recover(to);
6938 wake_up_all(&mc.waitq);
6941 static void mem_cgroup_clear_mc(void)
6943 struct mem_cgroup *from = mc.from;
6946 * we must clear moving_task before waking up waiters at the end of
6949 mc.moving_task = NULL;
6950 __mem_cgroup_clear_mc();
6951 spin_lock(&mc.lock);
6954 spin_unlock(&mc.lock);
6955 mem_cgroup_end_move(from);
6958 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6959 struct cgroup_taskset *tset)
6961 struct task_struct *p = cgroup_taskset_first(tset);
6963 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6964 unsigned long move_charge_at_immigrate;
6967 * We are now commited to this value whatever it is. Changes in this
6968 * tunable will only affect upcoming migrations, not the current one.
6969 * So we need to save it, and keep it going.
6971 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6972 if (move_charge_at_immigrate) {
6973 struct mm_struct *mm;
6974 struct mem_cgroup *from = mem_cgroup_from_task(p);
6976 VM_BUG_ON(from == memcg);
6978 mm = get_task_mm(p);
6981 /* We move charges only when we move a owner of the mm */
6982 if (mm->owner == p) {
6985 VM_BUG_ON(mc.precharge);
6986 VM_BUG_ON(mc.moved_charge);
6987 VM_BUG_ON(mc.moved_swap);
6988 mem_cgroup_start_move(from);
6989 spin_lock(&mc.lock);
6992 mc.immigrate_flags = move_charge_at_immigrate;
6993 spin_unlock(&mc.lock);
6994 /* We set mc.moving_task later */
6996 ret = mem_cgroup_precharge_mc(mm);
6998 mem_cgroup_clear_mc();
7005 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7006 struct cgroup_taskset *tset)
7008 mem_cgroup_clear_mc();
7011 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
7012 unsigned long addr, unsigned long end,
7013 struct mm_walk *walk)
7016 struct vm_area_struct *vma = walk->private;
7019 enum mc_target_type target_type;
7020 union mc_target target;
7022 struct page_cgroup *pc;
7025 * We don't take compound_lock() here but no race with splitting thp
7027 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
7028 * under splitting, which means there's no concurrent thp split,
7029 * - if another thread runs into split_huge_page() just after we
7030 * entered this if-block, the thread must wait for page table lock
7031 * to be unlocked in __split_huge_page_splitting(), where the main
7032 * part of thp split is not executed yet.
7034 if (pmd_trans_huge_lock(pmd, vma) == 1) {
7035 if (mc.precharge < HPAGE_PMD_NR) {
7036 spin_unlock(&vma->vm_mm->page_table_lock);
7039 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
7040 if (target_type == MC_TARGET_PAGE) {
7042 if (!isolate_lru_page(page)) {
7043 pc = lookup_page_cgroup(page);
7044 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
7045 pc, mc.from, mc.to)) {
7046 mc.precharge -= HPAGE_PMD_NR;
7047 mc.moved_charge += HPAGE_PMD_NR;
7049 putback_lru_page(page);
7053 spin_unlock(&vma->vm_mm->page_table_lock);
7057 if (pmd_trans_unstable(pmd))
7060 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
7061 for (; addr != end; addr += PAGE_SIZE) {
7062 pte_t ptent = *(pte++);
7068 switch (get_mctgt_type(vma, addr, ptent, &target)) {
7069 case MC_TARGET_PAGE:
7071 if (isolate_lru_page(page))
7073 pc = lookup_page_cgroup(page);
7074 if (!mem_cgroup_move_account(page, 1, pc,
7077 /* we uncharge from mc.from later. */
7080 putback_lru_page(page);
7081 put: /* get_mctgt_type() gets the page */
7084 case MC_TARGET_SWAP:
7086 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
7088 /* we fixup refcnts and charges later. */
7096 pte_unmap_unlock(pte - 1, ptl);
7101 * We have consumed all precharges we got in can_attach().
7102 * We try charge one by one, but don't do any additional
7103 * charges to mc.to if we have failed in charge once in attach()
7106 ret = mem_cgroup_do_precharge(1);
7114 static void mem_cgroup_move_charge(struct mm_struct *mm)
7116 struct vm_area_struct *vma;
7118 lru_add_drain_all();
7120 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
7122 * Someone who are holding the mmap_sem might be waiting in
7123 * waitq. So we cancel all extra charges, wake up all waiters,
7124 * and retry. Because we cancel precharges, we might not be able
7125 * to move enough charges, but moving charge is a best-effort
7126 * feature anyway, so it wouldn't be a big problem.
7128 __mem_cgroup_clear_mc();
7132 for (vma = mm->mmap; vma; vma = vma->vm_next) {
7134 struct mm_walk mem_cgroup_move_charge_walk = {
7135 .pmd_entry = mem_cgroup_move_charge_pte_range,
7139 if (is_vm_hugetlb_page(vma))
7141 ret = walk_page_range(vma->vm_start, vma->vm_end,
7142 &mem_cgroup_move_charge_walk);
7145 * means we have consumed all precharges and failed in
7146 * doing additional charge. Just abandon here.
7150 up_read(&mm->mmap_sem);
7153 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7154 struct cgroup_taskset *tset)
7156 struct task_struct *p = cgroup_taskset_first(tset);
7157 struct mm_struct *mm = get_task_mm(p);
7161 mem_cgroup_move_charge(mm);
7165 mem_cgroup_clear_mc();
7167 #else /* !CONFIG_MMU */
7168 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
7169 struct cgroup_taskset *tset)
7173 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
7174 struct cgroup_taskset *tset)
7177 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
7178 struct cgroup_taskset *tset)
7184 * Cgroup retains root cgroups across [un]mount cycles making it necessary
7185 * to verify sane_behavior flag on each mount attempt.
7187 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
7190 * use_hierarchy is forced with sane_behavior. cgroup core
7191 * guarantees that @root doesn't have any children, so turning it
7192 * on for the root memcg is enough.
7194 if (cgroup_sane_behavior(root_css->cgroup))
7195 mem_cgroup_from_css(root_css)->use_hierarchy = true;
7198 struct cgroup_subsys mem_cgroup_subsys = {
7200 .subsys_id = mem_cgroup_subsys_id,
7201 .css_alloc = mem_cgroup_css_alloc,
7202 .css_online = mem_cgroup_css_online,
7203 .css_offline = mem_cgroup_css_offline,
7204 .css_free = mem_cgroup_css_free,
7205 .can_attach = mem_cgroup_can_attach,
7206 .cancel_attach = mem_cgroup_cancel_attach,
7207 .attach = mem_cgroup_move_task,
7208 .bind = mem_cgroup_bind,
7209 .base_cftypes = mem_cgroup_files,
7214 #ifdef CONFIG_MEMCG_SWAP
7215 static int __init enable_swap_account(char *s)
7217 if (!strcmp(s, "1"))
7218 really_do_swap_account = 1;
7219 else if (!strcmp(s, "0"))
7220 really_do_swap_account = 0;
7223 __setup("swapaccount=", enable_swap_account);
7225 static void __init memsw_file_init(void)
7227 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7230 static void __init enable_swap_cgroup(void)
7232 if (!mem_cgroup_disabled() && really_do_swap_account) {
7233 do_swap_account = 1;
7239 static void __init enable_swap_cgroup(void)
7245 * subsys_initcall() for memory controller.
7247 * Some parts like hotcpu_notifier() have to be initialized from this context
7248 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7249 * everything that doesn't depend on a specific mem_cgroup structure should
7250 * be initialized from here.
7252 static int __init mem_cgroup_init(void)
7254 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7255 enable_swap_cgroup();
7256 mem_cgroup_soft_limit_tree_init();
7260 subsys_initcall(mem_cgroup_init);