1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * Kernel Memory Controller
14 * Copyright (C) 2012 Parallels Inc. and Google Inc.
15 * Authors: Glauber Costa and Suleiman Souhlal
17 * This program is free software; you can redistribute it and/or modify
18 * it under the terms of the GNU General Public License as published by
19 * the Free Software Foundation; either version 2 of the License, or
20 * (at your option) any later version.
22 * This program is distributed in the hope that it will be useful,
23 * but WITHOUT ANY WARRANTY; without even the implied warranty of
24 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
25 * GNU General Public License for more details.
28 #include <linux/res_counter.h>
29 #include <linux/memcontrol.h>
30 #include <linux/cgroup.h>
32 #include <linux/hugetlb.h>
33 #include <linux/pagemap.h>
34 #include <linux/smp.h>
35 #include <linux/page-flags.h>
36 #include <linux/backing-dev.h>
37 #include <linux/bit_spinlock.h>
38 #include <linux/rcupdate.h>
39 #include <linux/limits.h>
40 #include <linux/export.h>
41 #include <linux/mutex.h>
42 #include <linux/rbtree.h>
43 #include <linux/slab.h>
44 #include <linux/swap.h>
45 #include <linux/swapops.h>
46 #include <linux/spinlock.h>
47 #include <linux/eventfd.h>
48 #include <linux/sort.h>
50 #include <linux/seq_file.h>
51 #include <linux/vmalloc.h>
52 #include <linux/vmpressure.h>
53 #include <linux/mm_inline.h>
54 #include <linux/page_cgroup.h>
55 #include <linux/cpu.h>
56 #include <linux/oom.h>
60 #include <net/tcp_memcontrol.h>
62 #include <asm/uaccess.h>
64 #include <trace/events/vmscan.h>
66 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
67 EXPORT_SYMBOL(mem_cgroup_subsys);
69 #define MEM_CGROUP_RECLAIM_RETRIES 5
70 static struct mem_cgroup *root_mem_cgroup __read_mostly;
72 #ifdef CONFIG_MEMCG_SWAP
73 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
74 int do_swap_account __read_mostly;
76 /* for remember boot option*/
77 #ifdef CONFIG_MEMCG_SWAP_ENABLED
78 static int really_do_swap_account __initdata = 1;
80 static int really_do_swap_account __initdata = 0;
84 #define do_swap_account 0
88 static const char * const mem_cgroup_stat_names[] = {
97 enum mem_cgroup_events_index {
98 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
99 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
100 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
101 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
102 MEM_CGROUP_EVENTS_NSTATS,
105 static const char * const mem_cgroup_events_names[] = {
112 static const char * const mem_cgroup_lru_names[] = {
121 * Per memcg event counter is incremented at every pagein/pageout. With THP,
122 * it will be incremated by the number of pages. This counter is used for
123 * for trigger some periodic events. This is straightforward and better
124 * than using jiffies etc. to handle periodic memcg event.
126 enum mem_cgroup_events_target {
127 MEM_CGROUP_TARGET_THRESH,
128 MEM_CGROUP_TARGET_SOFTLIMIT,
129 MEM_CGROUP_TARGET_NUMAINFO,
132 #define THRESHOLDS_EVENTS_TARGET 128
133 #define SOFTLIMIT_EVENTS_TARGET 1024
134 #define NUMAINFO_EVENTS_TARGET 1024
136 struct mem_cgroup_stat_cpu {
137 long count[MEM_CGROUP_STAT_NSTATS];
138 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
139 unsigned long nr_page_events;
140 unsigned long targets[MEM_CGROUP_NTARGETS];
143 struct mem_cgroup_reclaim_iter {
145 * last scanned hierarchy member. Valid only if last_dead_count
146 * matches memcg->dead_count of the hierarchy root group.
148 struct mem_cgroup *last_visited;
149 unsigned long last_dead_count;
151 /* scan generation, increased every round-trip */
152 unsigned int generation;
156 * per-zone information in memory controller.
158 struct mem_cgroup_per_zone {
159 struct lruvec lruvec;
160 unsigned long lru_size[NR_LRU_LISTS];
162 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
164 struct rb_node tree_node; /* RB tree node */
165 unsigned long long usage_in_excess;/* Set to the value by which */
166 /* the soft limit is exceeded*/
168 struct mem_cgroup *memcg; /* Back pointer, we cannot */
169 /* use container_of */
172 struct mem_cgroup_per_node {
173 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
177 * Cgroups above their limits are maintained in a RB-Tree, independent of
178 * their hierarchy representation
181 struct mem_cgroup_tree_per_zone {
182 struct rb_root rb_root;
186 struct mem_cgroup_tree_per_node {
187 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
190 struct mem_cgroup_tree {
191 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
194 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
196 struct mem_cgroup_threshold {
197 struct eventfd_ctx *eventfd;
202 struct mem_cgroup_threshold_ary {
203 /* An array index points to threshold just below or equal to usage. */
204 int current_threshold;
205 /* Size of entries[] */
207 /* Array of thresholds */
208 struct mem_cgroup_threshold entries[0];
211 struct mem_cgroup_thresholds {
212 /* Primary thresholds array */
213 struct mem_cgroup_threshold_ary *primary;
215 * Spare threshold array.
216 * This is needed to make mem_cgroup_unregister_event() "never fail".
217 * It must be able to store at least primary->size - 1 entries.
219 struct mem_cgroup_threshold_ary *spare;
223 struct mem_cgroup_eventfd_list {
224 struct list_head list;
225 struct eventfd_ctx *eventfd;
228 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
229 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
232 * The memory controller data structure. The memory controller controls both
233 * page cache and RSS per cgroup. We would eventually like to provide
234 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
235 * to help the administrator determine what knobs to tune.
237 * TODO: Add a water mark for the memory controller. Reclaim will begin when
238 * we hit the water mark. May be even add a low water mark, such that
239 * no reclaim occurs from a cgroup at it's low water mark, this is
240 * a feature that will be implemented much later in the future.
243 struct cgroup_subsys_state css;
245 * the counter to account for memory usage
247 struct res_counter res;
249 /* vmpressure notifications */
250 struct vmpressure vmpressure;
253 * the counter to account for mem+swap usage.
255 struct res_counter memsw;
258 * the counter to account for kernel memory usage.
260 struct res_counter kmem;
262 * Should the accounting and control be hierarchical, per subtree?
265 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
269 atomic_t oom_wakeups;
272 /* OOM-Killer disable */
273 int oom_kill_disable;
275 /* set when res.limit == memsw.limit */
276 bool memsw_is_minimum;
278 /* protect arrays of thresholds */
279 struct mutex thresholds_lock;
281 /* thresholds for memory usage. RCU-protected */
282 struct mem_cgroup_thresholds thresholds;
284 /* thresholds for mem+swap usage. RCU-protected */
285 struct mem_cgroup_thresholds memsw_thresholds;
287 /* For oom notifier event fd */
288 struct list_head oom_notify;
291 * Should we move charges of a task when a task is moved into this
292 * mem_cgroup ? And what type of charges should we move ?
294 unsigned long move_charge_at_immigrate;
296 * set > 0 if pages under this cgroup are moving to other cgroup.
298 atomic_t moving_account;
299 /* taken only while moving_account > 0 */
300 spinlock_t move_lock;
304 struct mem_cgroup_stat_cpu __percpu *stat;
306 * used when a cpu is offlined or other synchronizations
307 * See mem_cgroup_read_stat().
309 struct mem_cgroup_stat_cpu nocpu_base;
310 spinlock_t pcp_counter_lock;
313 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
314 struct tcp_memcontrol tcp_mem;
316 #if defined(CONFIG_MEMCG_KMEM)
317 /* analogous to slab_common's slab_caches list. per-memcg */
318 struct list_head memcg_slab_caches;
319 /* Not a spinlock, we can take a lot of time walking the list */
320 struct mutex slab_caches_mutex;
321 /* Index in the kmem_cache->memcg_params->memcg_caches array */
325 int last_scanned_node;
327 nodemask_t scan_nodes;
328 atomic_t numainfo_events;
329 atomic_t numainfo_updating;
332 struct mem_cgroup_per_node *nodeinfo[0];
333 /* WARNING: nodeinfo must be the last member here */
336 static size_t memcg_size(void)
338 return sizeof(struct mem_cgroup) +
339 nr_node_ids * sizeof(struct mem_cgroup_per_node);
342 /* internal only representation about the status of kmem accounting. */
344 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
345 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
346 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
349 /* We account when limit is on, but only after call sites are patched */
350 #define KMEM_ACCOUNTED_MASK \
351 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
353 #ifdef CONFIG_MEMCG_KMEM
354 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
356 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
359 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
361 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
364 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
366 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
369 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
371 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
374 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
377 * Our caller must use css_get() first, because memcg_uncharge_kmem()
378 * will call css_put() if it sees the memcg is dead.
381 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
382 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
385 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
387 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
388 &memcg->kmem_account_flags);
392 /* Stuffs for move charges at task migration. */
394 * Types of charges to be moved. "move_charge_at_immitgrate" and
395 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
398 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
399 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
403 /* "mc" and its members are protected by cgroup_mutex */
404 static struct move_charge_struct {
405 spinlock_t lock; /* for from, to */
406 struct mem_cgroup *from;
407 struct mem_cgroup *to;
408 unsigned long immigrate_flags;
409 unsigned long precharge;
410 unsigned long moved_charge;
411 unsigned long moved_swap;
412 struct task_struct *moving_task; /* a task moving charges */
413 wait_queue_head_t waitq; /* a waitq for other context */
415 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
416 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
419 static bool move_anon(void)
421 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
424 static bool move_file(void)
426 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
430 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
431 * limit reclaim to prevent infinite loops, if they ever occur.
433 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
434 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
437 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
438 MEM_CGROUP_CHARGE_TYPE_ANON,
439 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
440 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
444 /* for encoding cft->private value on file */
452 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
453 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
454 #define MEMFILE_ATTR(val) ((val) & 0xffff)
455 /* Used for OOM nofiier */
456 #define OOM_CONTROL (0)
459 * Reclaim flags for mem_cgroup_hierarchical_reclaim
461 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
462 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
463 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
464 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
467 * The memcg_create_mutex will be held whenever a new cgroup is created.
468 * As a consequence, any change that needs to protect against new child cgroups
469 * appearing has to hold it as well.
471 static DEFINE_MUTEX(memcg_create_mutex);
473 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
475 return s ? container_of(s, struct mem_cgroup, css) : NULL;
478 /* Some nice accessors for the vmpressure. */
479 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
482 memcg = root_mem_cgroup;
483 return &memcg->vmpressure;
486 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
488 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
491 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
493 return &mem_cgroup_from_css(css)->vmpressure;
496 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
498 return (memcg == root_mem_cgroup);
501 /* Writing them here to avoid exposing memcg's inner layout */
502 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
504 void sock_update_memcg(struct sock *sk)
506 if (mem_cgroup_sockets_enabled) {
507 struct mem_cgroup *memcg;
508 struct cg_proto *cg_proto;
510 BUG_ON(!sk->sk_prot->proto_cgroup);
512 /* Socket cloning can throw us here with sk_cgrp already
513 * filled. It won't however, necessarily happen from
514 * process context. So the test for root memcg given
515 * the current task's memcg won't help us in this case.
517 * Respecting the original socket's memcg is a better
518 * decision in this case.
521 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
522 css_get(&sk->sk_cgrp->memcg->css);
527 memcg = mem_cgroup_from_task(current);
528 cg_proto = sk->sk_prot->proto_cgroup(memcg);
529 if (!mem_cgroup_is_root(memcg) &&
530 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
531 sk->sk_cgrp = cg_proto;
536 EXPORT_SYMBOL(sock_update_memcg);
538 void sock_release_memcg(struct sock *sk)
540 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
541 struct mem_cgroup *memcg;
542 WARN_ON(!sk->sk_cgrp->memcg);
543 memcg = sk->sk_cgrp->memcg;
544 css_put(&sk->sk_cgrp->memcg->css);
548 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
550 if (!memcg || mem_cgroup_is_root(memcg))
553 return &memcg->tcp_mem.cg_proto;
555 EXPORT_SYMBOL(tcp_proto_cgroup);
557 static void disarm_sock_keys(struct mem_cgroup *memcg)
559 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
561 static_key_slow_dec(&memcg_socket_limit_enabled);
564 static void disarm_sock_keys(struct mem_cgroup *memcg)
569 #ifdef CONFIG_MEMCG_KMEM
571 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
572 * There are two main reasons for not using the css_id for this:
573 * 1) this works better in sparse environments, where we have a lot of memcgs,
574 * but only a few kmem-limited. Or also, if we have, for instance, 200
575 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
576 * 200 entry array for that.
578 * 2) In order not to violate the cgroup API, we would like to do all memory
579 * allocation in ->create(). At that point, we haven't yet allocated the
580 * css_id. Having a separate index prevents us from messing with the cgroup
583 * The current size of the caches array is stored in
584 * memcg_limited_groups_array_size. It will double each time we have to
587 static DEFINE_IDA(kmem_limited_groups);
588 int memcg_limited_groups_array_size;
591 * MIN_SIZE is different than 1, because we would like to avoid going through
592 * the alloc/free process all the time. In a small machine, 4 kmem-limited
593 * cgroups is a reasonable guess. In the future, it could be a parameter or
594 * tunable, but that is strictly not necessary.
596 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
597 * this constant directly from cgroup, but it is understandable that this is
598 * better kept as an internal representation in cgroup.c. In any case, the
599 * css_id space is not getting any smaller, and we don't have to necessarily
600 * increase ours as well if it increases.
602 #define MEMCG_CACHES_MIN_SIZE 4
603 #define MEMCG_CACHES_MAX_SIZE 65535
606 * A lot of the calls to the cache allocation functions are expected to be
607 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
608 * conditional to this static branch, we'll have to allow modules that does
609 * kmem_cache_alloc and the such to see this symbol as well
611 struct static_key memcg_kmem_enabled_key;
612 EXPORT_SYMBOL(memcg_kmem_enabled_key);
614 static void disarm_kmem_keys(struct mem_cgroup *memcg)
616 if (memcg_kmem_is_active(memcg)) {
617 static_key_slow_dec(&memcg_kmem_enabled_key);
618 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
621 * This check can't live in kmem destruction function,
622 * since the charges will outlive the cgroup
624 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
627 static void disarm_kmem_keys(struct mem_cgroup *memcg)
630 #endif /* CONFIG_MEMCG_KMEM */
632 static void disarm_static_keys(struct mem_cgroup *memcg)
634 disarm_sock_keys(memcg);
635 disarm_kmem_keys(memcg);
638 static void drain_all_stock_async(struct mem_cgroup *memcg);
640 static struct mem_cgroup_per_zone *
641 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
643 VM_BUG_ON((unsigned)nid >= nr_node_ids);
644 return &memcg->nodeinfo[nid]->zoneinfo[zid];
647 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
652 static struct mem_cgroup_per_zone *
653 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
655 int nid = page_to_nid(page);
656 int zid = page_zonenum(page);
658 return mem_cgroup_zoneinfo(memcg, nid, zid);
661 static struct mem_cgroup_tree_per_zone *
662 soft_limit_tree_node_zone(int nid, int zid)
664 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
667 static struct mem_cgroup_tree_per_zone *
668 soft_limit_tree_from_page(struct page *page)
670 int nid = page_to_nid(page);
671 int zid = page_zonenum(page);
673 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
677 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
678 struct mem_cgroup_per_zone *mz,
679 struct mem_cgroup_tree_per_zone *mctz,
680 unsigned long long new_usage_in_excess)
682 struct rb_node **p = &mctz->rb_root.rb_node;
683 struct rb_node *parent = NULL;
684 struct mem_cgroup_per_zone *mz_node;
689 mz->usage_in_excess = new_usage_in_excess;
690 if (!mz->usage_in_excess)
694 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
696 if (mz->usage_in_excess < mz_node->usage_in_excess)
699 * We can't avoid mem cgroups that are over their soft
700 * limit by the same amount
702 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
705 rb_link_node(&mz->tree_node, parent, p);
706 rb_insert_color(&mz->tree_node, &mctz->rb_root);
711 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
712 struct mem_cgroup_per_zone *mz,
713 struct mem_cgroup_tree_per_zone *mctz)
717 rb_erase(&mz->tree_node, &mctz->rb_root);
722 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
723 struct mem_cgroup_per_zone *mz,
724 struct mem_cgroup_tree_per_zone *mctz)
726 spin_lock(&mctz->lock);
727 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
728 spin_unlock(&mctz->lock);
732 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
734 unsigned long long excess;
735 struct mem_cgroup_per_zone *mz;
736 struct mem_cgroup_tree_per_zone *mctz;
737 int nid = page_to_nid(page);
738 int zid = page_zonenum(page);
739 mctz = soft_limit_tree_from_page(page);
742 * Necessary to update all ancestors when hierarchy is used.
743 * because their event counter is not touched.
745 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
746 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
747 excess = res_counter_soft_limit_excess(&memcg->res);
749 * We have to update the tree if mz is on RB-tree or
750 * mem is over its softlimit.
752 if (excess || mz->on_tree) {
753 spin_lock(&mctz->lock);
754 /* if on-tree, remove it */
756 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
758 * Insert again. mz->usage_in_excess will be updated.
759 * If excess is 0, no tree ops.
761 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
762 spin_unlock(&mctz->lock);
767 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
770 struct mem_cgroup_per_zone *mz;
771 struct mem_cgroup_tree_per_zone *mctz;
773 for_each_node(node) {
774 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
775 mz = mem_cgroup_zoneinfo(memcg, node, zone);
776 mctz = soft_limit_tree_node_zone(node, zone);
777 mem_cgroup_remove_exceeded(memcg, mz, mctz);
782 static struct mem_cgroup_per_zone *
783 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
785 struct rb_node *rightmost = NULL;
786 struct mem_cgroup_per_zone *mz;
790 rightmost = rb_last(&mctz->rb_root);
792 goto done; /* Nothing to reclaim from */
794 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
796 * Remove the node now but someone else can add it back,
797 * we will to add it back at the end of reclaim to its correct
798 * position in the tree.
800 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
801 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
802 !css_tryget(&mz->memcg->css))
808 static struct mem_cgroup_per_zone *
809 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
811 struct mem_cgroup_per_zone *mz;
813 spin_lock(&mctz->lock);
814 mz = __mem_cgroup_largest_soft_limit_node(mctz);
815 spin_unlock(&mctz->lock);
820 * Implementation Note: reading percpu statistics for memcg.
822 * Both of vmstat[] and percpu_counter has threshold and do periodic
823 * synchronization to implement "quick" read. There are trade-off between
824 * reading cost and precision of value. Then, we may have a chance to implement
825 * a periodic synchronizion of counter in memcg's counter.
827 * But this _read() function is used for user interface now. The user accounts
828 * memory usage by memory cgroup and he _always_ requires exact value because
829 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
830 * have to visit all online cpus and make sum. So, for now, unnecessary
831 * synchronization is not implemented. (just implemented for cpu hotplug)
833 * If there are kernel internal actions which can make use of some not-exact
834 * value, and reading all cpu value can be performance bottleneck in some
835 * common workload, threashold and synchonization as vmstat[] should be
838 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
839 enum mem_cgroup_stat_index idx)
845 for_each_online_cpu(cpu)
846 val += per_cpu(memcg->stat->count[idx], cpu);
847 #ifdef CONFIG_HOTPLUG_CPU
848 spin_lock(&memcg->pcp_counter_lock);
849 val += memcg->nocpu_base.count[idx];
850 spin_unlock(&memcg->pcp_counter_lock);
856 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
859 int val = (charge) ? 1 : -1;
860 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
863 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
864 enum mem_cgroup_events_index idx)
866 unsigned long val = 0;
870 for_each_online_cpu(cpu)
871 val += per_cpu(memcg->stat->events[idx], cpu);
872 #ifdef CONFIG_HOTPLUG_CPU
873 spin_lock(&memcg->pcp_counter_lock);
874 val += memcg->nocpu_base.events[idx];
875 spin_unlock(&memcg->pcp_counter_lock);
881 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
883 bool anon, int nr_pages)
888 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
889 * counted as CACHE even if it's on ANON LRU.
892 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
895 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
898 if (PageTransHuge(page))
899 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
902 /* pagein of a big page is an event. So, ignore page size */
904 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
906 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
907 nr_pages = -nr_pages; /* for event */
910 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
916 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
918 struct mem_cgroup_per_zone *mz;
920 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
921 return mz->lru_size[lru];
925 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
926 unsigned int lru_mask)
928 struct mem_cgroup_per_zone *mz;
930 unsigned long ret = 0;
932 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
935 if (BIT(lru) & lru_mask)
936 ret += mz->lru_size[lru];
942 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
943 int nid, unsigned int lru_mask)
948 for (zid = 0; zid < MAX_NR_ZONES; zid++)
949 total += mem_cgroup_zone_nr_lru_pages(memcg,
955 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
956 unsigned int lru_mask)
961 for_each_node_state(nid, N_MEMORY)
962 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
966 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
967 enum mem_cgroup_events_target target)
969 unsigned long val, next;
971 val = __this_cpu_read(memcg->stat->nr_page_events);
972 next = __this_cpu_read(memcg->stat->targets[target]);
973 /* from time_after() in jiffies.h */
974 if ((long)next - (long)val < 0) {
976 case MEM_CGROUP_TARGET_THRESH:
977 next = val + THRESHOLDS_EVENTS_TARGET;
979 case MEM_CGROUP_TARGET_SOFTLIMIT:
980 next = val + SOFTLIMIT_EVENTS_TARGET;
982 case MEM_CGROUP_TARGET_NUMAINFO:
983 next = val + NUMAINFO_EVENTS_TARGET;
988 __this_cpu_write(memcg->stat->targets[target], next);
995 * Check events in order.
998 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1001 /* threshold event is triggered in finer grain than soft limit */
1002 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1003 MEM_CGROUP_TARGET_THRESH))) {
1005 bool do_numainfo __maybe_unused;
1007 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1008 MEM_CGROUP_TARGET_SOFTLIMIT);
1009 #if MAX_NUMNODES > 1
1010 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1011 MEM_CGROUP_TARGET_NUMAINFO);
1015 mem_cgroup_threshold(memcg);
1016 if (unlikely(do_softlimit))
1017 mem_cgroup_update_tree(memcg, page);
1018 #if MAX_NUMNODES > 1
1019 if (unlikely(do_numainfo))
1020 atomic_inc(&memcg->numainfo_events);
1026 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1029 * mm_update_next_owner() may clear mm->owner to NULL
1030 * if it races with swapoff, page migration, etc.
1031 * So this can be called with p == NULL.
1036 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1039 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1041 struct mem_cgroup *memcg = NULL;
1046 * Because we have no locks, mm->owner's may be being moved to other
1047 * cgroup. We use css_tryget() here even if this looks
1048 * pessimistic (rather than adding locks here).
1052 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1053 if (unlikely(!memcg))
1055 } while (!css_tryget(&memcg->css));
1061 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1062 * ref. count) or NULL if the whole root's subtree has been visited.
1064 * helper function to be used by mem_cgroup_iter
1066 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1067 struct mem_cgroup *last_visited)
1069 struct cgroup_subsys_state *prev_css, *next_css;
1071 prev_css = last_visited ? &last_visited->css : NULL;
1073 next_css = css_next_descendant_pre(prev_css, &root->css);
1076 * Even if we found a group we have to make sure it is
1077 * alive. css && !memcg means that the groups should be
1078 * skipped and we should continue the tree walk.
1079 * last_visited css is safe to use because it is
1080 * protected by css_get and the tree walk is rcu safe.
1083 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
1085 if (css_tryget(&mem->css))
1088 prev_css = next_css;
1096 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1099 * When a group in the hierarchy below root is destroyed, the
1100 * hierarchy iterator can no longer be trusted since it might
1101 * have pointed to the destroyed group. Invalidate it.
1103 atomic_inc(&root->dead_count);
1106 static struct mem_cgroup *
1107 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1108 struct mem_cgroup *root,
1111 struct mem_cgroup *position = NULL;
1113 * A cgroup destruction happens in two stages: offlining and
1114 * release. They are separated by a RCU grace period.
1116 * If the iterator is valid, we may still race with an
1117 * offlining. The RCU lock ensures the object won't be
1118 * released, tryget will fail if we lost the race.
1120 *sequence = atomic_read(&root->dead_count);
1121 if (iter->last_dead_count == *sequence) {
1123 position = iter->last_visited;
1124 if (position && !css_tryget(&position->css))
1130 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1131 struct mem_cgroup *last_visited,
1132 struct mem_cgroup *new_position,
1136 css_put(&last_visited->css);
1138 * We store the sequence count from the time @last_visited was
1139 * loaded successfully instead of rereading it here so that we
1140 * don't lose destruction events in between. We could have
1141 * raced with the destruction of @new_position after all.
1143 iter->last_visited = new_position;
1145 iter->last_dead_count = sequence;
1149 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1150 * @root: hierarchy root
1151 * @prev: previously returned memcg, NULL on first invocation
1152 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1154 * Returns references to children of the hierarchy below @root, or
1155 * @root itself, or %NULL after a full round-trip.
1157 * Caller must pass the return value in @prev on subsequent
1158 * invocations for reference counting, or use mem_cgroup_iter_break()
1159 * to cancel a hierarchy walk before the round-trip is complete.
1161 * Reclaimers can specify a zone and a priority level in @reclaim to
1162 * divide up the memcgs in the hierarchy among all concurrent
1163 * reclaimers operating on the same zone and priority.
1165 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1166 struct mem_cgroup *prev,
1167 struct mem_cgroup_reclaim_cookie *reclaim)
1169 struct mem_cgroup *memcg = NULL;
1170 struct mem_cgroup *last_visited = NULL;
1172 if (mem_cgroup_disabled())
1176 root = root_mem_cgroup;
1178 if (prev && !reclaim)
1179 last_visited = prev;
1181 if (!root->use_hierarchy && root != root_mem_cgroup) {
1189 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1190 int uninitialized_var(seq);
1193 int nid = zone_to_nid(reclaim->zone);
1194 int zid = zone_idx(reclaim->zone);
1195 struct mem_cgroup_per_zone *mz;
1197 mz = mem_cgroup_zoneinfo(root, nid, zid);
1198 iter = &mz->reclaim_iter[reclaim->priority];
1199 if (prev && reclaim->generation != iter->generation) {
1200 iter->last_visited = NULL;
1204 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1207 memcg = __mem_cgroup_iter_next(root, last_visited);
1210 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1214 else if (!prev && memcg)
1215 reclaim->generation = iter->generation;
1224 if (prev && prev != root)
1225 css_put(&prev->css);
1231 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1232 * @root: hierarchy root
1233 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1235 void mem_cgroup_iter_break(struct mem_cgroup *root,
1236 struct mem_cgroup *prev)
1239 root = root_mem_cgroup;
1240 if (prev && prev != root)
1241 css_put(&prev->css);
1245 * Iteration constructs for visiting all cgroups (under a tree). If
1246 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1247 * be used for reference counting.
1249 #define for_each_mem_cgroup_tree(iter, root) \
1250 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1252 iter = mem_cgroup_iter(root, iter, NULL))
1254 #define for_each_mem_cgroup(iter) \
1255 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1257 iter = mem_cgroup_iter(NULL, iter, NULL))
1259 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1261 struct mem_cgroup *memcg;
1264 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1265 if (unlikely(!memcg))
1270 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1273 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1281 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1284 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1285 * @zone: zone of the wanted lruvec
1286 * @memcg: memcg of the wanted lruvec
1288 * Returns the lru list vector holding pages for the given @zone and
1289 * @mem. This can be the global zone lruvec, if the memory controller
1292 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1293 struct mem_cgroup *memcg)
1295 struct mem_cgroup_per_zone *mz;
1296 struct lruvec *lruvec;
1298 if (mem_cgroup_disabled()) {
1299 lruvec = &zone->lruvec;
1303 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1304 lruvec = &mz->lruvec;
1307 * Since a node can be onlined after the mem_cgroup was created,
1308 * we have to be prepared to initialize lruvec->zone here;
1309 * and if offlined then reonlined, we need to reinitialize it.
1311 if (unlikely(lruvec->zone != zone))
1312 lruvec->zone = zone;
1317 * Following LRU functions are allowed to be used without PCG_LOCK.
1318 * Operations are called by routine of global LRU independently from memcg.
1319 * What we have to take care of here is validness of pc->mem_cgroup.
1321 * Changes to pc->mem_cgroup happens when
1324 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1325 * It is added to LRU before charge.
1326 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1327 * When moving account, the page is not on LRU. It's isolated.
1331 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1333 * @zone: zone of the page
1335 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1337 struct mem_cgroup_per_zone *mz;
1338 struct mem_cgroup *memcg;
1339 struct page_cgroup *pc;
1340 struct lruvec *lruvec;
1342 if (mem_cgroup_disabled()) {
1343 lruvec = &zone->lruvec;
1347 pc = lookup_page_cgroup(page);
1348 memcg = pc->mem_cgroup;
1351 * Surreptitiously switch any uncharged offlist page to root:
1352 * an uncharged page off lru does nothing to secure
1353 * its former mem_cgroup from sudden removal.
1355 * Our caller holds lru_lock, and PageCgroupUsed is updated
1356 * under page_cgroup lock: between them, they make all uses
1357 * of pc->mem_cgroup safe.
1359 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1360 pc->mem_cgroup = memcg = root_mem_cgroup;
1362 mz = page_cgroup_zoneinfo(memcg, page);
1363 lruvec = &mz->lruvec;
1366 * Since a node can be onlined after the mem_cgroup was created,
1367 * we have to be prepared to initialize lruvec->zone here;
1368 * and if offlined then reonlined, we need to reinitialize it.
1370 if (unlikely(lruvec->zone != zone))
1371 lruvec->zone = zone;
1376 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1377 * @lruvec: mem_cgroup per zone lru vector
1378 * @lru: index of lru list the page is sitting on
1379 * @nr_pages: positive when adding or negative when removing
1381 * This function must be called when a page is added to or removed from an
1384 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1387 struct mem_cgroup_per_zone *mz;
1388 unsigned long *lru_size;
1390 if (mem_cgroup_disabled())
1393 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1394 lru_size = mz->lru_size + lru;
1395 *lru_size += nr_pages;
1396 VM_BUG_ON((long)(*lru_size) < 0);
1400 * Checks whether given mem is same or in the root_mem_cgroup's
1403 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1404 struct mem_cgroup *memcg)
1406 if (root_memcg == memcg)
1408 if (!root_memcg->use_hierarchy || !memcg)
1410 return css_is_ancestor(&memcg->css, &root_memcg->css);
1413 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1414 struct mem_cgroup *memcg)
1419 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1424 bool task_in_mem_cgroup(struct task_struct *task,
1425 const struct mem_cgroup *memcg)
1427 struct mem_cgroup *curr = NULL;
1428 struct task_struct *p;
1431 p = find_lock_task_mm(task);
1433 curr = try_get_mem_cgroup_from_mm(p->mm);
1437 * All threads may have already detached their mm's, but the oom
1438 * killer still needs to detect if they have already been oom
1439 * killed to prevent needlessly killing additional tasks.
1442 curr = mem_cgroup_from_task(task);
1444 css_get(&curr->css);
1450 * We should check use_hierarchy of "memcg" not "curr". Because checking
1451 * use_hierarchy of "curr" here make this function true if hierarchy is
1452 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1453 * hierarchy(even if use_hierarchy is disabled in "memcg").
1455 ret = mem_cgroup_same_or_subtree(memcg, curr);
1456 css_put(&curr->css);
1460 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1462 unsigned long inactive_ratio;
1463 unsigned long inactive;
1464 unsigned long active;
1467 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1468 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1470 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1472 inactive_ratio = int_sqrt(10 * gb);
1476 return inactive * inactive_ratio < active;
1479 #define mem_cgroup_from_res_counter(counter, member) \
1480 container_of(counter, struct mem_cgroup, member)
1483 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1484 * @memcg: the memory cgroup
1486 * Returns the maximum amount of memory @mem can be charged with, in
1489 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1491 unsigned long long margin;
1493 margin = res_counter_margin(&memcg->res);
1494 if (do_swap_account)
1495 margin = min(margin, res_counter_margin(&memcg->memsw));
1496 return margin >> PAGE_SHIFT;
1499 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1502 if (!css_parent(&memcg->css))
1503 return vm_swappiness;
1505 return memcg->swappiness;
1509 * memcg->moving_account is used for checking possibility that some thread is
1510 * calling move_account(). When a thread on CPU-A starts moving pages under
1511 * a memcg, other threads should check memcg->moving_account under
1512 * rcu_read_lock(), like this:
1516 * memcg->moving_account+1 if (memcg->mocing_account)
1518 * synchronize_rcu() update something.
1523 /* for quick checking without looking up memcg */
1524 atomic_t memcg_moving __read_mostly;
1526 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1528 atomic_inc(&memcg_moving);
1529 atomic_inc(&memcg->moving_account);
1533 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1536 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1537 * We check NULL in callee rather than caller.
1540 atomic_dec(&memcg_moving);
1541 atomic_dec(&memcg->moving_account);
1546 * 2 routines for checking "mem" is under move_account() or not.
1548 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1549 * is used for avoiding races in accounting. If true,
1550 * pc->mem_cgroup may be overwritten.
1552 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1553 * under hierarchy of moving cgroups. This is for
1554 * waiting at hith-memory prressure caused by "move".
1557 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1559 VM_BUG_ON(!rcu_read_lock_held());
1560 return atomic_read(&memcg->moving_account) > 0;
1563 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1565 struct mem_cgroup *from;
1566 struct mem_cgroup *to;
1569 * Unlike task_move routines, we access mc.to, mc.from not under
1570 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1572 spin_lock(&mc.lock);
1578 ret = mem_cgroup_same_or_subtree(memcg, from)
1579 || mem_cgroup_same_or_subtree(memcg, to);
1581 spin_unlock(&mc.lock);
1585 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1587 if (mc.moving_task && current != mc.moving_task) {
1588 if (mem_cgroup_under_move(memcg)) {
1590 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1591 /* moving charge context might have finished. */
1594 finish_wait(&mc.waitq, &wait);
1602 * Take this lock when
1603 * - a code tries to modify page's memcg while it's USED.
1604 * - a code tries to modify page state accounting in a memcg.
1605 * see mem_cgroup_stolen(), too.
1607 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1608 unsigned long *flags)
1610 spin_lock_irqsave(&memcg->move_lock, *flags);
1613 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1614 unsigned long *flags)
1616 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1619 #define K(x) ((x) << (PAGE_SHIFT-10))
1621 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1622 * @memcg: The memory cgroup that went over limit
1623 * @p: Task that is going to be killed
1625 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1628 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1630 struct cgroup *task_cgrp;
1631 struct cgroup *mem_cgrp;
1633 * Need a buffer in BSS, can't rely on allocations. The code relies
1634 * on the assumption that OOM is serialized for memory controller.
1635 * If this assumption is broken, revisit this code.
1637 static char memcg_name[PATH_MAX];
1639 struct mem_cgroup *iter;
1647 mem_cgrp = memcg->css.cgroup;
1648 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1650 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1653 * Unfortunately, we are unable to convert to a useful name
1654 * But we'll still print out the usage information
1661 pr_info("Task in %s killed", memcg_name);
1664 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1672 * Continues from above, so we don't need an KERN_ level
1674 pr_cont(" as a result of limit of %s\n", memcg_name);
1677 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1678 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1679 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1680 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1681 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1682 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1683 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1684 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1685 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1686 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1687 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1688 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1690 for_each_mem_cgroup_tree(iter, memcg) {
1691 pr_info("Memory cgroup stats");
1694 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1696 pr_cont(" for %s", memcg_name);
1700 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1701 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1703 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1704 K(mem_cgroup_read_stat(iter, i)));
1707 for (i = 0; i < NR_LRU_LISTS; i++)
1708 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1709 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1716 * This function returns the number of memcg under hierarchy tree. Returns
1717 * 1(self count) if no children.
1719 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1722 struct mem_cgroup *iter;
1724 for_each_mem_cgroup_tree(iter, memcg)
1730 * Return the memory (and swap, if configured) limit for a memcg.
1732 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1736 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1739 * Do not consider swap space if we cannot swap due to swappiness
1741 if (mem_cgroup_swappiness(memcg)) {
1744 limit += total_swap_pages << PAGE_SHIFT;
1745 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1748 * If memsw is finite and limits the amount of swap space
1749 * available to this memcg, return that limit.
1751 limit = min(limit, memsw);
1757 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1760 struct mem_cgroup *iter;
1761 unsigned long chosen_points = 0;
1762 unsigned long totalpages;
1763 unsigned int points = 0;
1764 struct task_struct *chosen = NULL;
1767 * If current has a pending SIGKILL or is exiting, then automatically
1768 * select it. The goal is to allow it to allocate so that it may
1769 * quickly exit and free its memory.
1771 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1772 set_thread_flag(TIF_MEMDIE);
1776 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1777 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1778 for_each_mem_cgroup_tree(iter, memcg) {
1779 struct css_task_iter it;
1780 struct task_struct *task;
1782 css_task_iter_start(&iter->css, &it);
1783 while ((task = css_task_iter_next(&it))) {
1784 switch (oom_scan_process_thread(task, totalpages, NULL,
1786 case OOM_SCAN_SELECT:
1788 put_task_struct(chosen);
1790 chosen_points = ULONG_MAX;
1791 get_task_struct(chosen);
1793 case OOM_SCAN_CONTINUE:
1795 case OOM_SCAN_ABORT:
1796 css_task_iter_end(&it);
1797 mem_cgroup_iter_break(memcg, iter);
1799 put_task_struct(chosen);
1804 points = oom_badness(task, memcg, NULL, totalpages);
1805 if (points > chosen_points) {
1807 put_task_struct(chosen);
1809 chosen_points = points;
1810 get_task_struct(chosen);
1813 css_task_iter_end(&it);
1818 points = chosen_points * 1000 / totalpages;
1819 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1820 NULL, "Memory cgroup out of memory");
1823 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1825 unsigned long flags)
1827 unsigned long total = 0;
1828 bool noswap = false;
1831 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1833 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1836 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1838 drain_all_stock_async(memcg);
1839 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1841 * Allow limit shrinkers, which are triggered directly
1842 * by userspace, to catch signals and stop reclaim
1843 * after minimal progress, regardless of the margin.
1845 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1847 if (mem_cgroup_margin(memcg))
1850 * If nothing was reclaimed after two attempts, there
1851 * may be no reclaimable pages in this hierarchy.
1860 * test_mem_cgroup_node_reclaimable
1861 * @memcg: the target memcg
1862 * @nid: the node ID to be checked.
1863 * @noswap : specify true here if the user wants flle only information.
1865 * This function returns whether the specified memcg contains any
1866 * reclaimable pages on a node. Returns true if there are any reclaimable
1867 * pages in the node.
1869 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1870 int nid, bool noswap)
1872 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1874 if (noswap || !total_swap_pages)
1876 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1881 #if MAX_NUMNODES > 1
1884 * Always updating the nodemask is not very good - even if we have an empty
1885 * list or the wrong list here, we can start from some node and traverse all
1886 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1889 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1893 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1894 * pagein/pageout changes since the last update.
1896 if (!atomic_read(&memcg->numainfo_events))
1898 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1901 /* make a nodemask where this memcg uses memory from */
1902 memcg->scan_nodes = node_states[N_MEMORY];
1904 for_each_node_mask(nid, node_states[N_MEMORY]) {
1906 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1907 node_clear(nid, memcg->scan_nodes);
1910 atomic_set(&memcg->numainfo_events, 0);
1911 atomic_set(&memcg->numainfo_updating, 0);
1915 * Selecting a node where we start reclaim from. Because what we need is just
1916 * reducing usage counter, start from anywhere is O,K. Considering
1917 * memory reclaim from current node, there are pros. and cons.
1919 * Freeing memory from current node means freeing memory from a node which
1920 * we'll use or we've used. So, it may make LRU bad. And if several threads
1921 * hit limits, it will see a contention on a node. But freeing from remote
1922 * node means more costs for memory reclaim because of memory latency.
1924 * Now, we use round-robin. Better algorithm is welcomed.
1926 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1930 mem_cgroup_may_update_nodemask(memcg);
1931 node = memcg->last_scanned_node;
1933 node = next_node(node, memcg->scan_nodes);
1934 if (node == MAX_NUMNODES)
1935 node = first_node(memcg->scan_nodes);
1937 * We call this when we hit limit, not when pages are added to LRU.
1938 * No LRU may hold pages because all pages are UNEVICTABLE or
1939 * memcg is too small and all pages are not on LRU. In that case,
1940 * we use curret node.
1942 if (unlikely(node == MAX_NUMNODES))
1943 node = numa_node_id();
1945 memcg->last_scanned_node = node;
1950 * Check all nodes whether it contains reclaimable pages or not.
1951 * For quick scan, we make use of scan_nodes. This will allow us to skip
1952 * unused nodes. But scan_nodes is lazily updated and may not cotain
1953 * enough new information. We need to do double check.
1955 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1960 * quick check...making use of scan_node.
1961 * We can skip unused nodes.
1963 if (!nodes_empty(memcg->scan_nodes)) {
1964 for (nid = first_node(memcg->scan_nodes);
1966 nid = next_node(nid, memcg->scan_nodes)) {
1968 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1973 * Check rest of nodes.
1975 for_each_node_state(nid, N_MEMORY) {
1976 if (node_isset(nid, memcg->scan_nodes))
1978 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1985 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1990 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1992 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1996 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1999 unsigned long *total_scanned)
2001 struct mem_cgroup *victim = NULL;
2004 unsigned long excess;
2005 unsigned long nr_scanned;
2006 struct mem_cgroup_reclaim_cookie reclaim = {
2011 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2014 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2019 * If we have not been able to reclaim
2020 * anything, it might because there are
2021 * no reclaimable pages under this hierarchy
2026 * We want to do more targeted reclaim.
2027 * excess >> 2 is not to excessive so as to
2028 * reclaim too much, nor too less that we keep
2029 * coming back to reclaim from this cgroup
2031 if (total >= (excess >> 2) ||
2032 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2037 if (!mem_cgroup_reclaimable(victim, false))
2039 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2041 *total_scanned += nr_scanned;
2042 if (!res_counter_soft_limit_excess(&root_memcg->res))
2045 mem_cgroup_iter_break(root_memcg, victim);
2049 static DEFINE_SPINLOCK(memcg_oom_lock);
2052 * Check OOM-Killer is already running under our hierarchy.
2053 * If someone is running, return false.
2055 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2057 struct mem_cgroup *iter, *failed = NULL;
2059 spin_lock(&memcg_oom_lock);
2061 for_each_mem_cgroup_tree(iter, memcg) {
2062 if (iter->oom_lock) {
2064 * this subtree of our hierarchy is already locked
2065 * so we cannot give a lock.
2068 mem_cgroup_iter_break(memcg, iter);
2071 iter->oom_lock = true;
2076 * OK, we failed to lock the whole subtree so we have
2077 * to clean up what we set up to the failing subtree
2079 for_each_mem_cgroup_tree(iter, memcg) {
2080 if (iter == failed) {
2081 mem_cgroup_iter_break(memcg, iter);
2084 iter->oom_lock = false;
2088 spin_unlock(&memcg_oom_lock);
2093 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2095 struct mem_cgroup *iter;
2097 spin_lock(&memcg_oom_lock);
2098 for_each_mem_cgroup_tree(iter, memcg)
2099 iter->oom_lock = false;
2100 spin_unlock(&memcg_oom_lock);
2103 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2105 struct mem_cgroup *iter;
2107 for_each_mem_cgroup_tree(iter, memcg)
2108 atomic_inc(&iter->under_oom);
2111 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2113 struct mem_cgroup *iter;
2116 * When a new child is created while the hierarchy is under oom,
2117 * mem_cgroup_oom_lock() may not be called. We have to use
2118 * atomic_add_unless() here.
2120 for_each_mem_cgroup_tree(iter, memcg)
2121 atomic_add_unless(&iter->under_oom, -1, 0);
2124 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2126 struct oom_wait_info {
2127 struct mem_cgroup *memcg;
2131 static int memcg_oom_wake_function(wait_queue_t *wait,
2132 unsigned mode, int sync, void *arg)
2134 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2135 struct mem_cgroup *oom_wait_memcg;
2136 struct oom_wait_info *oom_wait_info;
2138 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2139 oom_wait_memcg = oom_wait_info->memcg;
2142 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2143 * Then we can use css_is_ancestor without taking care of RCU.
2145 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2146 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2148 return autoremove_wake_function(wait, mode, sync, arg);
2151 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2153 atomic_inc(&memcg->oom_wakeups);
2154 /* for filtering, pass "memcg" as argument. */
2155 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2158 static void memcg_oom_recover(struct mem_cgroup *memcg)
2160 if (memcg && atomic_read(&memcg->under_oom))
2161 memcg_wakeup_oom(memcg);
2165 * try to call OOM killer
2167 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2172 if (!current->memcg_oom.may_oom)
2175 current->memcg_oom.in_memcg_oom = 1;
2178 * As with any blocking lock, a contender needs to start
2179 * listening for wakeups before attempting the trylock,
2180 * otherwise it can miss the wakeup from the unlock and sleep
2181 * indefinitely. This is just open-coded because our locking
2182 * is so particular to memcg hierarchies.
2184 wakeups = atomic_read(&memcg->oom_wakeups);
2185 mem_cgroup_mark_under_oom(memcg);
2187 locked = mem_cgroup_oom_trylock(memcg);
2190 mem_cgroup_oom_notify(memcg);
2192 if (locked && !memcg->oom_kill_disable) {
2193 mem_cgroup_unmark_under_oom(memcg);
2194 mem_cgroup_out_of_memory(memcg, mask, order);
2195 mem_cgroup_oom_unlock(memcg);
2197 * There is no guarantee that an OOM-lock contender
2198 * sees the wakeups triggered by the OOM kill
2199 * uncharges. Wake any sleepers explicitely.
2201 memcg_oom_recover(memcg);
2204 * A system call can just return -ENOMEM, but if this
2205 * is a page fault and somebody else is handling the
2206 * OOM already, we need to sleep on the OOM waitqueue
2207 * for this memcg until the situation is resolved.
2208 * Which can take some time because it might be
2209 * handled by a userspace task.
2211 * However, this is the charge context, which means
2212 * that we may sit on a large call stack and hold
2213 * various filesystem locks, the mmap_sem etc. and we
2214 * don't want the OOM handler to deadlock on them
2215 * while we sit here and wait. Store the current OOM
2216 * context in the task_struct, then return -ENOMEM.
2217 * At the end of the page fault handler, with the
2218 * stack unwound, pagefault_out_of_memory() will check
2219 * back with us by calling
2220 * mem_cgroup_oom_synchronize(), possibly putting the
2223 current->memcg_oom.oom_locked = locked;
2224 current->memcg_oom.wakeups = wakeups;
2225 css_get(&memcg->css);
2226 current->memcg_oom.wait_on_memcg = memcg;
2231 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2233 * This has to be called at the end of a page fault if the the memcg
2234 * OOM handler was enabled and the fault is returning %VM_FAULT_OOM.
2236 * Memcg supports userspace OOM handling, so 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 put the task to sleep and clean up the
2244 * Returns %true if an ongoing memcg OOM situation was detected and
2245 * finalized, %false otherwise.
2247 bool mem_cgroup_oom_synchronize(void)
2249 struct oom_wait_info owait;
2250 struct mem_cgroup *memcg;
2252 /* OOM is global, do not handle */
2253 if (!current->memcg_oom.in_memcg_oom)
2257 * We invoked the OOM killer but there is a chance that a kill
2258 * did not free up any charges. Everybody else might already
2259 * be sleeping, so restart the fault and keep the rampage
2260 * going until some charges are released.
2262 memcg = current->memcg_oom.wait_on_memcg;
2266 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2269 owait.memcg = memcg;
2270 owait.wait.flags = 0;
2271 owait.wait.func = memcg_oom_wake_function;
2272 owait.wait.private = current;
2273 INIT_LIST_HEAD(&owait.wait.task_list);
2275 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2276 /* Only sleep if we didn't miss any wakeups since OOM */
2277 if (atomic_read(&memcg->oom_wakeups) == current->memcg_oom.wakeups)
2279 finish_wait(&memcg_oom_waitq, &owait.wait);
2281 mem_cgroup_unmark_under_oom(memcg);
2282 if (current->memcg_oom.oom_locked) {
2283 mem_cgroup_oom_unlock(memcg);
2285 * There is no guarantee that an OOM-lock contender
2286 * sees the wakeups triggered by the OOM kill
2287 * uncharges. Wake any sleepers explicitely.
2289 memcg_oom_recover(memcg);
2291 css_put(&memcg->css);
2292 current->memcg_oom.wait_on_memcg = NULL;
2294 current->memcg_oom.in_memcg_oom = 0;
2299 * Currently used to update mapped file statistics, but the routine can be
2300 * generalized to update other statistics as well.
2302 * Notes: Race condition
2304 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2305 * it tends to be costly. But considering some conditions, we doesn't need
2306 * to do so _always_.
2308 * Considering "charge", lock_page_cgroup() is not required because all
2309 * file-stat operations happen after a page is attached to radix-tree. There
2310 * are no race with "charge".
2312 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2313 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2314 * if there are race with "uncharge". Statistics itself is properly handled
2317 * Considering "move", this is an only case we see a race. To make the race
2318 * small, we check mm->moving_account and detect there are possibility of race
2319 * If there is, we take a lock.
2322 void __mem_cgroup_begin_update_page_stat(struct page *page,
2323 bool *locked, unsigned long *flags)
2325 struct mem_cgroup *memcg;
2326 struct page_cgroup *pc;
2328 pc = lookup_page_cgroup(page);
2330 memcg = pc->mem_cgroup;
2331 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2334 * If this memory cgroup is not under account moving, we don't
2335 * need to take move_lock_mem_cgroup(). Because we already hold
2336 * rcu_read_lock(), any calls to move_account will be delayed until
2337 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2339 if (!mem_cgroup_stolen(memcg))
2342 move_lock_mem_cgroup(memcg, flags);
2343 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2344 move_unlock_mem_cgroup(memcg, flags);
2350 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2352 struct page_cgroup *pc = lookup_page_cgroup(page);
2355 * It's guaranteed that pc->mem_cgroup never changes while
2356 * lock is held because a routine modifies pc->mem_cgroup
2357 * should take move_lock_mem_cgroup().
2359 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2362 void mem_cgroup_update_page_stat(struct page *page,
2363 enum mem_cgroup_stat_index idx, int val)
2365 struct mem_cgroup *memcg;
2366 struct page_cgroup *pc = lookup_page_cgroup(page);
2367 unsigned long uninitialized_var(flags);
2369 if (mem_cgroup_disabled())
2372 VM_BUG_ON(!rcu_read_lock_held());
2373 memcg = pc->mem_cgroup;
2374 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2377 this_cpu_add(memcg->stat->count[idx], val);
2381 * size of first charge trial. "32" comes from vmscan.c's magic value.
2382 * TODO: maybe necessary to use big numbers in big irons.
2384 #define CHARGE_BATCH 32U
2385 struct memcg_stock_pcp {
2386 struct mem_cgroup *cached; /* this never be root cgroup */
2387 unsigned int nr_pages;
2388 struct work_struct work;
2389 unsigned long flags;
2390 #define FLUSHING_CACHED_CHARGE 0
2392 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2393 static DEFINE_MUTEX(percpu_charge_mutex);
2396 * consume_stock: Try to consume stocked charge on this cpu.
2397 * @memcg: memcg to consume from.
2398 * @nr_pages: how many pages to charge.
2400 * The charges will only happen if @memcg matches the current cpu's memcg
2401 * stock, and at least @nr_pages are available in that stock. Failure to
2402 * service an allocation will refill the stock.
2404 * returns true if successful, false otherwise.
2406 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2408 struct memcg_stock_pcp *stock;
2411 if (nr_pages > CHARGE_BATCH)
2414 stock = &get_cpu_var(memcg_stock);
2415 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2416 stock->nr_pages -= nr_pages;
2417 else /* need to call res_counter_charge */
2419 put_cpu_var(memcg_stock);
2424 * Returns stocks cached in percpu to res_counter and reset cached information.
2426 static void drain_stock(struct memcg_stock_pcp *stock)
2428 struct mem_cgroup *old = stock->cached;
2430 if (stock->nr_pages) {
2431 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2433 res_counter_uncharge(&old->res, bytes);
2434 if (do_swap_account)
2435 res_counter_uncharge(&old->memsw, bytes);
2436 stock->nr_pages = 0;
2438 stock->cached = NULL;
2442 * This must be called under preempt disabled or must be called by
2443 * a thread which is pinned to local cpu.
2445 static void drain_local_stock(struct work_struct *dummy)
2447 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2449 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2452 static void __init memcg_stock_init(void)
2456 for_each_possible_cpu(cpu) {
2457 struct memcg_stock_pcp *stock =
2458 &per_cpu(memcg_stock, cpu);
2459 INIT_WORK(&stock->work, drain_local_stock);
2464 * Cache charges(val) which is from res_counter, to local per_cpu area.
2465 * This will be consumed by consume_stock() function, later.
2467 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2469 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2471 if (stock->cached != memcg) { /* reset if necessary */
2473 stock->cached = memcg;
2475 stock->nr_pages += nr_pages;
2476 put_cpu_var(memcg_stock);
2480 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2481 * of the hierarchy under it. sync flag says whether we should block
2482 * until the work is done.
2484 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2488 /* Notify other cpus that system-wide "drain" is running */
2491 for_each_online_cpu(cpu) {
2492 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2493 struct mem_cgroup *memcg;
2495 memcg = stock->cached;
2496 if (!memcg || !stock->nr_pages)
2498 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2500 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2502 drain_local_stock(&stock->work);
2504 schedule_work_on(cpu, &stock->work);
2512 for_each_online_cpu(cpu) {
2513 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2514 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2515 flush_work(&stock->work);
2522 * Tries to drain stocked charges in other cpus. This function is asynchronous
2523 * and just put a work per cpu for draining localy on each cpu. Caller can
2524 * expects some charges will be back to res_counter later but cannot wait for
2527 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2530 * If someone calls draining, avoid adding more kworker runs.
2532 if (!mutex_trylock(&percpu_charge_mutex))
2534 drain_all_stock(root_memcg, false);
2535 mutex_unlock(&percpu_charge_mutex);
2538 /* This is a synchronous drain interface. */
2539 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2541 /* called when force_empty is called */
2542 mutex_lock(&percpu_charge_mutex);
2543 drain_all_stock(root_memcg, true);
2544 mutex_unlock(&percpu_charge_mutex);
2548 * This function drains percpu counter value from DEAD cpu and
2549 * move it to local cpu. Note that this function can be preempted.
2551 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2555 spin_lock(&memcg->pcp_counter_lock);
2556 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2557 long x = per_cpu(memcg->stat->count[i], cpu);
2559 per_cpu(memcg->stat->count[i], cpu) = 0;
2560 memcg->nocpu_base.count[i] += x;
2562 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2563 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2565 per_cpu(memcg->stat->events[i], cpu) = 0;
2566 memcg->nocpu_base.events[i] += x;
2568 spin_unlock(&memcg->pcp_counter_lock);
2571 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2572 unsigned long action,
2575 int cpu = (unsigned long)hcpu;
2576 struct memcg_stock_pcp *stock;
2577 struct mem_cgroup *iter;
2579 if (action == CPU_ONLINE)
2582 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2585 for_each_mem_cgroup(iter)
2586 mem_cgroup_drain_pcp_counter(iter, cpu);
2588 stock = &per_cpu(memcg_stock, cpu);
2594 /* See __mem_cgroup_try_charge() for details */
2596 CHARGE_OK, /* success */
2597 CHARGE_RETRY, /* need to retry but retry is not bad */
2598 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2599 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2602 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2603 unsigned int nr_pages, unsigned int min_pages,
2606 unsigned long csize = nr_pages * PAGE_SIZE;
2607 struct mem_cgroup *mem_over_limit;
2608 struct res_counter *fail_res;
2609 unsigned long flags = 0;
2612 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2615 if (!do_swap_account)
2617 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2621 res_counter_uncharge(&memcg->res, csize);
2622 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2623 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2625 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2627 * Never reclaim on behalf of optional batching, retry with a
2628 * single page instead.
2630 if (nr_pages > min_pages)
2631 return CHARGE_RETRY;
2633 if (!(gfp_mask & __GFP_WAIT))
2634 return CHARGE_WOULDBLOCK;
2636 if (gfp_mask & __GFP_NORETRY)
2637 return CHARGE_NOMEM;
2639 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2640 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2641 return CHARGE_RETRY;
2643 * Even though the limit is exceeded at this point, reclaim
2644 * may have been able to free some pages. Retry the charge
2645 * before killing the task.
2647 * Only for regular pages, though: huge pages are rather
2648 * unlikely to succeed so close to the limit, and we fall back
2649 * to regular pages anyway in case of failure.
2651 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2652 return CHARGE_RETRY;
2655 * At task move, charge accounts can be doubly counted. So, it's
2656 * better to wait until the end of task_move if something is going on.
2658 if (mem_cgroup_wait_acct_move(mem_over_limit))
2659 return CHARGE_RETRY;
2662 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2664 return CHARGE_NOMEM;
2668 * __mem_cgroup_try_charge() does
2669 * 1. detect memcg to be charged against from passed *mm and *ptr,
2670 * 2. update res_counter
2671 * 3. call memory reclaim if necessary.
2673 * In some special case, if the task is fatal, fatal_signal_pending() or
2674 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2675 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2676 * as possible without any hazards. 2: all pages should have a valid
2677 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2678 * pointer, that is treated as a charge to root_mem_cgroup.
2680 * So __mem_cgroup_try_charge() will return
2681 * 0 ... on success, filling *ptr with a valid memcg pointer.
2682 * -ENOMEM ... charge failure because of resource limits.
2683 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2685 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2686 * the oom-killer can be invoked.
2688 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2690 unsigned int nr_pages,
2691 struct mem_cgroup **ptr,
2694 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2695 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2696 struct mem_cgroup *memcg = NULL;
2700 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2701 * in system level. So, allow to go ahead dying process in addition to
2704 if (unlikely(test_thread_flag(TIF_MEMDIE)
2705 || fatal_signal_pending(current)))
2709 * We always charge the cgroup the mm_struct belongs to.
2710 * The mm_struct's mem_cgroup changes on task migration if the
2711 * thread group leader migrates. It's possible that mm is not
2712 * set, if so charge the root memcg (happens for pagecache usage).
2715 *ptr = root_mem_cgroup;
2717 if (*ptr) { /* css should be a valid one */
2719 if (mem_cgroup_is_root(memcg))
2721 if (consume_stock(memcg, nr_pages))
2723 css_get(&memcg->css);
2725 struct task_struct *p;
2728 p = rcu_dereference(mm->owner);
2730 * Because we don't have task_lock(), "p" can exit.
2731 * In that case, "memcg" can point to root or p can be NULL with
2732 * race with swapoff. Then, we have small risk of mis-accouning.
2733 * But such kind of mis-account by race always happens because
2734 * we don't have cgroup_mutex(). It's overkill and we allo that
2736 * (*) swapoff at el will charge against mm-struct not against
2737 * task-struct. So, mm->owner can be NULL.
2739 memcg = mem_cgroup_from_task(p);
2741 memcg = root_mem_cgroup;
2742 if (mem_cgroup_is_root(memcg)) {
2746 if (consume_stock(memcg, nr_pages)) {
2748 * It seems dagerous to access memcg without css_get().
2749 * But considering how consume_stok works, it's not
2750 * necessary. If consume_stock success, some charges
2751 * from this memcg are cached on this cpu. So, we
2752 * don't need to call css_get()/css_tryget() before
2753 * calling consume_stock().
2758 /* after here, we may be blocked. we need to get refcnt */
2759 if (!css_tryget(&memcg->css)) {
2767 bool invoke_oom = oom && !nr_oom_retries;
2769 /* If killed, bypass charge */
2770 if (fatal_signal_pending(current)) {
2771 css_put(&memcg->css);
2775 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2776 nr_pages, invoke_oom);
2780 case CHARGE_RETRY: /* not in OOM situation but retry */
2782 css_put(&memcg->css);
2785 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2786 css_put(&memcg->css);
2788 case CHARGE_NOMEM: /* OOM routine works */
2789 if (!oom || invoke_oom) {
2790 css_put(&memcg->css);
2796 } while (ret != CHARGE_OK);
2798 if (batch > nr_pages)
2799 refill_stock(memcg, batch - nr_pages);
2800 css_put(&memcg->css);
2808 *ptr = root_mem_cgroup;
2813 * Somemtimes we have to undo a charge we got by try_charge().
2814 * This function is for that and do uncharge, put css's refcnt.
2815 * gotten by try_charge().
2817 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2818 unsigned int nr_pages)
2820 if (!mem_cgroup_is_root(memcg)) {
2821 unsigned long bytes = nr_pages * PAGE_SIZE;
2823 res_counter_uncharge(&memcg->res, bytes);
2824 if (do_swap_account)
2825 res_counter_uncharge(&memcg->memsw, bytes);
2830 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2831 * This is useful when moving usage to parent cgroup.
2833 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2834 unsigned int nr_pages)
2836 unsigned long bytes = nr_pages * PAGE_SIZE;
2838 if (mem_cgroup_is_root(memcg))
2841 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2842 if (do_swap_account)
2843 res_counter_uncharge_until(&memcg->memsw,
2844 memcg->memsw.parent, bytes);
2848 * A helper function to get mem_cgroup from ID. must be called under
2849 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2850 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2851 * called against removed memcg.)
2853 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2855 struct cgroup_subsys_state *css;
2857 /* ID 0 is unused ID */
2860 css = css_lookup(&mem_cgroup_subsys, id);
2863 return mem_cgroup_from_css(css);
2866 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2868 struct mem_cgroup *memcg = NULL;
2869 struct page_cgroup *pc;
2873 VM_BUG_ON(!PageLocked(page));
2875 pc = lookup_page_cgroup(page);
2876 lock_page_cgroup(pc);
2877 if (PageCgroupUsed(pc)) {
2878 memcg = pc->mem_cgroup;
2879 if (memcg && !css_tryget(&memcg->css))
2881 } else if (PageSwapCache(page)) {
2882 ent.val = page_private(page);
2883 id = lookup_swap_cgroup_id(ent);
2885 memcg = mem_cgroup_lookup(id);
2886 if (memcg && !css_tryget(&memcg->css))
2890 unlock_page_cgroup(pc);
2894 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2896 unsigned int nr_pages,
2897 enum charge_type ctype,
2900 struct page_cgroup *pc = lookup_page_cgroup(page);
2901 struct zone *uninitialized_var(zone);
2902 struct lruvec *lruvec;
2903 bool was_on_lru = false;
2906 lock_page_cgroup(pc);
2907 VM_BUG_ON(PageCgroupUsed(pc));
2909 * we don't need page_cgroup_lock about tail pages, becase they are not
2910 * accessed by any other context at this point.
2914 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2915 * may already be on some other mem_cgroup's LRU. Take care of it.
2918 zone = page_zone(page);
2919 spin_lock_irq(&zone->lru_lock);
2920 if (PageLRU(page)) {
2921 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2923 del_page_from_lru_list(page, lruvec, page_lru(page));
2928 pc->mem_cgroup = memcg;
2930 * We access a page_cgroup asynchronously without lock_page_cgroup().
2931 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2932 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2933 * before USED bit, we need memory barrier here.
2934 * See mem_cgroup_add_lru_list(), etc.
2937 SetPageCgroupUsed(pc);
2941 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2942 VM_BUG_ON(PageLRU(page));
2944 add_page_to_lru_list(page, lruvec, page_lru(page));
2946 spin_unlock_irq(&zone->lru_lock);
2949 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2954 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2955 unlock_page_cgroup(pc);
2958 * "charge_statistics" updated event counter. Then, check it.
2959 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2960 * if they exceeds softlimit.
2962 memcg_check_events(memcg, page);
2965 static DEFINE_MUTEX(set_limit_mutex);
2967 #ifdef CONFIG_MEMCG_KMEM
2968 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2970 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2971 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2975 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2976 * in the memcg_cache_params struct.
2978 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2980 struct kmem_cache *cachep;
2982 VM_BUG_ON(p->is_root_cache);
2983 cachep = p->root_cache;
2984 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2987 #ifdef CONFIG_SLABINFO
2988 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2989 struct cftype *cft, struct seq_file *m)
2991 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2992 struct memcg_cache_params *params;
2994 if (!memcg_can_account_kmem(memcg))
2997 print_slabinfo_header(m);
2999 mutex_lock(&memcg->slab_caches_mutex);
3000 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
3001 cache_show(memcg_params_to_cache(params), m);
3002 mutex_unlock(&memcg->slab_caches_mutex);
3008 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
3010 struct res_counter *fail_res;
3011 struct mem_cgroup *_memcg;
3015 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
3020 * Conditions under which we can wait for the oom_killer. Those are
3021 * the same conditions tested by the core page allocator
3023 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
3026 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
3029 if (ret == -EINTR) {
3031 * __mem_cgroup_try_charge() chosed to bypass to root due to
3032 * OOM kill or fatal signal. Since our only options are to
3033 * either fail the allocation or charge it to this cgroup, do
3034 * it as a temporary condition. But we can't fail. From a
3035 * kmem/slab perspective, the cache has already been selected,
3036 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3039 * This condition will only trigger if the task entered
3040 * memcg_charge_kmem in a sane state, but was OOM-killed during
3041 * __mem_cgroup_try_charge() above. Tasks that were already
3042 * dying when the allocation triggers should have been already
3043 * directed to the root cgroup in memcontrol.h
3045 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3046 if (do_swap_account)
3047 res_counter_charge_nofail(&memcg->memsw, size,
3051 res_counter_uncharge(&memcg->kmem, size);
3056 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3058 res_counter_uncharge(&memcg->res, size);
3059 if (do_swap_account)
3060 res_counter_uncharge(&memcg->memsw, size);
3063 if (res_counter_uncharge(&memcg->kmem, size))
3067 * Releases a reference taken in kmem_cgroup_css_offline in case
3068 * this last uncharge is racing with the offlining code or it is
3069 * outliving the memcg existence.
3071 * The memory barrier imposed by test&clear is paired with the
3072 * explicit one in memcg_kmem_mark_dead().
3074 if (memcg_kmem_test_and_clear_dead(memcg))
3075 css_put(&memcg->css);
3078 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3083 mutex_lock(&memcg->slab_caches_mutex);
3084 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3085 mutex_unlock(&memcg->slab_caches_mutex);
3089 * helper for acessing a memcg's index. It will be used as an index in the
3090 * child cache array in kmem_cache, and also to derive its name. This function
3091 * will return -1 when this is not a kmem-limited memcg.
3093 int memcg_cache_id(struct mem_cgroup *memcg)
3095 return memcg ? memcg->kmemcg_id : -1;
3099 * This ends up being protected by the set_limit mutex, during normal
3100 * operation, because that is its main call site.
3102 * But when we create a new cache, we can call this as well if its parent
3103 * is kmem-limited. That will have to hold set_limit_mutex as well.
3105 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3109 num = ida_simple_get(&kmem_limited_groups,
3110 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3114 * After this point, kmem_accounted (that we test atomically in
3115 * the beginning of this conditional), is no longer 0. This
3116 * guarantees only one process will set the following boolean
3117 * to true. We don't need test_and_set because we're protected
3118 * by the set_limit_mutex anyway.
3120 memcg_kmem_set_activated(memcg);
3122 ret = memcg_update_all_caches(num+1);
3124 ida_simple_remove(&kmem_limited_groups, num);
3125 memcg_kmem_clear_activated(memcg);
3129 memcg->kmemcg_id = num;
3130 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3131 mutex_init(&memcg->slab_caches_mutex);
3135 static size_t memcg_caches_array_size(int num_groups)
3138 if (num_groups <= 0)
3141 size = 2 * num_groups;
3142 if (size < MEMCG_CACHES_MIN_SIZE)
3143 size = MEMCG_CACHES_MIN_SIZE;
3144 else if (size > MEMCG_CACHES_MAX_SIZE)
3145 size = MEMCG_CACHES_MAX_SIZE;
3151 * We should update the current array size iff all caches updates succeed. This
3152 * can only be done from the slab side. The slab mutex needs to be held when
3155 void memcg_update_array_size(int num)
3157 if (num > memcg_limited_groups_array_size)
3158 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3161 static void kmem_cache_destroy_work_func(struct work_struct *w);
3163 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3165 struct memcg_cache_params *cur_params = s->memcg_params;
3167 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3169 if (num_groups > memcg_limited_groups_array_size) {
3171 ssize_t size = memcg_caches_array_size(num_groups);
3173 size *= sizeof(void *);
3174 size += offsetof(struct memcg_cache_params, memcg_caches);
3176 s->memcg_params = kzalloc(size, GFP_KERNEL);
3177 if (!s->memcg_params) {
3178 s->memcg_params = cur_params;
3182 s->memcg_params->is_root_cache = true;
3185 * There is the chance it will be bigger than
3186 * memcg_limited_groups_array_size, if we failed an allocation
3187 * in a cache, in which case all caches updated before it, will
3188 * have a bigger array.
3190 * But if that is the case, the data after
3191 * memcg_limited_groups_array_size is certainly unused
3193 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3194 if (!cur_params->memcg_caches[i])
3196 s->memcg_params->memcg_caches[i] =
3197 cur_params->memcg_caches[i];
3201 * Ideally, we would wait until all caches succeed, and only
3202 * then free the old one. But this is not worth the extra
3203 * pointer per-cache we'd have to have for this.
3205 * It is not a big deal if some caches are left with a size
3206 * bigger than the others. And all updates will reset this
3214 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3215 struct kmem_cache *root_cache)
3219 if (!memcg_kmem_enabled())
3223 size = offsetof(struct memcg_cache_params, memcg_caches);
3224 size += memcg_limited_groups_array_size * sizeof(void *);
3226 size = sizeof(struct memcg_cache_params);
3228 s->memcg_params = kzalloc(size, GFP_KERNEL);
3229 if (!s->memcg_params)
3233 s->memcg_params->memcg = memcg;
3234 s->memcg_params->root_cache = root_cache;
3235 INIT_WORK(&s->memcg_params->destroy,
3236 kmem_cache_destroy_work_func);
3238 s->memcg_params->is_root_cache = true;
3243 void memcg_release_cache(struct kmem_cache *s)
3245 struct kmem_cache *root;
3246 struct mem_cgroup *memcg;
3250 * This happens, for instance, when a root cache goes away before we
3253 if (!s->memcg_params)
3256 if (s->memcg_params->is_root_cache)
3259 memcg = s->memcg_params->memcg;
3260 id = memcg_cache_id(memcg);
3262 root = s->memcg_params->root_cache;
3263 root->memcg_params->memcg_caches[id] = NULL;
3265 mutex_lock(&memcg->slab_caches_mutex);
3266 list_del(&s->memcg_params->list);
3267 mutex_unlock(&memcg->slab_caches_mutex);
3269 css_put(&memcg->css);
3271 kfree(s->memcg_params);
3275 * During the creation a new cache, we need to disable our accounting mechanism
3276 * altogether. This is true even if we are not creating, but rather just
3277 * enqueing new caches to be created.
3279 * This is because that process will trigger allocations; some visible, like
3280 * explicit kmallocs to auxiliary data structures, name strings and internal
3281 * cache structures; some well concealed, like INIT_WORK() that can allocate
3282 * objects during debug.
3284 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3285 * to it. This may not be a bounded recursion: since the first cache creation
3286 * failed to complete (waiting on the allocation), we'll just try to create the
3287 * cache again, failing at the same point.
3289 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3290 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3291 * inside the following two functions.
3293 static inline void memcg_stop_kmem_account(void)
3295 VM_BUG_ON(!current->mm);
3296 current->memcg_kmem_skip_account++;
3299 static inline void memcg_resume_kmem_account(void)
3301 VM_BUG_ON(!current->mm);
3302 current->memcg_kmem_skip_account--;
3305 static void kmem_cache_destroy_work_func(struct work_struct *w)
3307 struct kmem_cache *cachep;
3308 struct memcg_cache_params *p;
3310 p = container_of(w, struct memcg_cache_params, destroy);
3312 cachep = memcg_params_to_cache(p);
3315 * If we get down to 0 after shrink, we could delete right away.
3316 * However, memcg_release_pages() already puts us back in the workqueue
3317 * in that case. If we proceed deleting, we'll get a dangling
3318 * reference, and removing the object from the workqueue in that case
3319 * is unnecessary complication. We are not a fast path.
3321 * Note that this case is fundamentally different from racing with
3322 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3323 * kmem_cache_shrink, not only we would be reinserting a dead cache
3324 * into the queue, but doing so from inside the worker racing to
3327 * So if we aren't down to zero, we'll just schedule a worker and try
3330 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3331 kmem_cache_shrink(cachep);
3332 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3335 kmem_cache_destroy(cachep);
3338 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3340 if (!cachep->memcg_params->dead)
3344 * There are many ways in which we can get here.
3346 * We can get to a memory-pressure situation while the delayed work is
3347 * still pending to run. The vmscan shrinkers can then release all
3348 * cache memory and get us to destruction. If this is the case, we'll
3349 * be executed twice, which is a bug (the second time will execute over
3350 * bogus data). In this case, cancelling the work should be fine.
3352 * But we can also get here from the worker itself, if
3353 * kmem_cache_shrink is enough to shake all the remaining objects and
3354 * get the page count to 0. In this case, we'll deadlock if we try to
3355 * cancel the work (the worker runs with an internal lock held, which
3356 * is the same lock we would hold for cancel_work_sync().)
3358 * Since we can't possibly know who got us here, just refrain from
3359 * running if there is already work pending
3361 if (work_pending(&cachep->memcg_params->destroy))
3364 * We have to defer the actual destroying to a workqueue, because
3365 * we might currently be in a context that cannot sleep.
3367 schedule_work(&cachep->memcg_params->destroy);
3371 * This lock protects updaters, not readers. We want readers to be as fast as
3372 * they can, and they will either see NULL or a valid cache value. Our model
3373 * allow them to see NULL, in which case the root memcg will be selected.
3375 * We need this lock because multiple allocations to the same cache from a non
3376 * will span more than one worker. Only one of them can create the cache.
3378 static DEFINE_MUTEX(memcg_cache_mutex);
3381 * Called with memcg_cache_mutex held
3383 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3384 struct kmem_cache *s)
3386 struct kmem_cache *new;
3387 static char *tmp_name = NULL;
3389 lockdep_assert_held(&memcg_cache_mutex);
3392 * kmem_cache_create_memcg duplicates the given name and
3393 * cgroup_name for this name requires RCU context.
3394 * This static temporary buffer is used to prevent from
3395 * pointless shortliving allocation.
3398 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3404 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3405 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3408 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3409 (s->flags & ~SLAB_PANIC), s->ctor, s);
3412 new->allocflags |= __GFP_KMEMCG;
3417 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3418 struct kmem_cache *cachep)
3420 struct kmem_cache *new_cachep;
3423 BUG_ON(!memcg_can_account_kmem(memcg));
3425 idx = memcg_cache_id(memcg);
3427 mutex_lock(&memcg_cache_mutex);
3428 new_cachep = cachep->memcg_params->memcg_caches[idx];
3430 css_put(&memcg->css);
3434 new_cachep = kmem_cache_dup(memcg, cachep);
3435 if (new_cachep == NULL) {
3436 new_cachep = cachep;
3437 css_put(&memcg->css);
3441 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3443 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3445 * the readers won't lock, make sure everybody sees the updated value,
3446 * so they won't put stuff in the queue again for no reason
3450 mutex_unlock(&memcg_cache_mutex);
3454 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3456 struct kmem_cache *c;
3459 if (!s->memcg_params)
3461 if (!s->memcg_params->is_root_cache)
3465 * If the cache is being destroyed, we trust that there is no one else
3466 * requesting objects from it. Even if there are, the sanity checks in
3467 * kmem_cache_destroy should caught this ill-case.
3469 * Still, we don't want anyone else freeing memcg_caches under our
3470 * noses, which can happen if a new memcg comes to life. As usual,
3471 * we'll take the set_limit_mutex to protect ourselves against this.
3473 mutex_lock(&set_limit_mutex);
3474 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3475 c = s->memcg_params->memcg_caches[i];
3480 * We will now manually delete the caches, so to avoid races
3481 * we need to cancel all pending destruction workers and
3482 * proceed with destruction ourselves.
3484 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3485 * and that could spawn the workers again: it is likely that
3486 * the cache still have active pages until this very moment.
3487 * This would lead us back to mem_cgroup_destroy_cache.
3489 * But that will not execute at all if the "dead" flag is not
3490 * set, so flip it down to guarantee we are in control.
3492 c->memcg_params->dead = false;
3493 cancel_work_sync(&c->memcg_params->destroy);
3494 kmem_cache_destroy(c);
3496 mutex_unlock(&set_limit_mutex);
3499 struct create_work {
3500 struct mem_cgroup *memcg;
3501 struct kmem_cache *cachep;
3502 struct work_struct work;
3505 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3507 struct kmem_cache *cachep;
3508 struct memcg_cache_params *params;
3510 if (!memcg_kmem_is_active(memcg))
3513 mutex_lock(&memcg->slab_caches_mutex);
3514 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3515 cachep = memcg_params_to_cache(params);
3516 cachep->memcg_params->dead = true;
3517 schedule_work(&cachep->memcg_params->destroy);
3519 mutex_unlock(&memcg->slab_caches_mutex);
3522 static void memcg_create_cache_work_func(struct work_struct *w)
3524 struct create_work *cw;
3526 cw = container_of(w, struct create_work, work);
3527 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3532 * Enqueue the creation of a per-memcg kmem_cache.
3534 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3535 struct kmem_cache *cachep)
3537 struct create_work *cw;
3539 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3541 css_put(&memcg->css);
3546 cw->cachep = cachep;
3548 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3549 schedule_work(&cw->work);
3552 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3553 struct kmem_cache *cachep)
3556 * We need to stop accounting when we kmalloc, because if the
3557 * corresponding kmalloc cache is not yet created, the first allocation
3558 * in __memcg_create_cache_enqueue will recurse.
3560 * However, it is better to enclose the whole function. Depending on
3561 * the debugging options enabled, INIT_WORK(), for instance, can
3562 * trigger an allocation. This too, will make us recurse. Because at
3563 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3564 * the safest choice is to do it like this, wrapping the whole function.
3566 memcg_stop_kmem_account();
3567 __memcg_create_cache_enqueue(memcg, cachep);
3568 memcg_resume_kmem_account();
3571 * Return the kmem_cache we're supposed to use for a slab allocation.
3572 * We try to use the current memcg's version of the cache.
3574 * If the cache does not exist yet, if we are the first user of it,
3575 * we either create it immediately, if possible, or create it asynchronously
3577 * In the latter case, we will let the current allocation go through with
3578 * the original cache.
3580 * Can't be called in interrupt context or from kernel threads.
3581 * This function needs to be called with rcu_read_lock() held.
3583 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3586 struct mem_cgroup *memcg;
3589 VM_BUG_ON(!cachep->memcg_params);
3590 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3592 if (!current->mm || current->memcg_kmem_skip_account)
3596 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3598 if (!memcg_can_account_kmem(memcg))
3601 idx = memcg_cache_id(memcg);
3604 * barrier to mare sure we're always seeing the up to date value. The
3605 * code updating memcg_caches will issue a write barrier to match this.
3607 read_barrier_depends();
3608 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3609 cachep = cachep->memcg_params->memcg_caches[idx];
3613 /* The corresponding put will be done in the workqueue. */
3614 if (!css_tryget(&memcg->css))
3619 * If we are in a safe context (can wait, and not in interrupt
3620 * context), we could be be predictable and return right away.
3621 * This would guarantee that the allocation being performed
3622 * already belongs in the new cache.
3624 * However, there are some clashes that can arrive from locking.
3625 * For instance, because we acquire the slab_mutex while doing
3626 * kmem_cache_dup, this means no further allocation could happen
3627 * with the slab_mutex held.
3629 * Also, because cache creation issue get_online_cpus(), this
3630 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3631 * that ends up reversed during cpu hotplug. (cpuset allocates
3632 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3633 * better to defer everything.
3635 memcg_create_cache_enqueue(memcg, cachep);
3641 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3644 * We need to verify if the allocation against current->mm->owner's memcg is
3645 * possible for the given order. But the page is not allocated yet, so we'll
3646 * need a further commit step to do the final arrangements.
3648 * It is possible for the task to switch cgroups in this mean time, so at
3649 * commit time, we can't rely on task conversion any longer. We'll then use
3650 * the handle argument to return to the caller which cgroup we should commit
3651 * against. We could also return the memcg directly and avoid the pointer
3652 * passing, but a boolean return value gives better semantics considering
3653 * the compiled-out case as well.
3655 * Returning true means the allocation is possible.
3658 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3660 struct mem_cgroup *memcg;
3666 * Disabling accounting is only relevant for some specific memcg
3667 * internal allocations. Therefore we would initially not have such
3668 * check here, since direct calls to the page allocator that are marked
3669 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3670 * concerned with cache allocations, and by having this test at
3671 * memcg_kmem_get_cache, we are already able to relay the allocation to
3672 * the root cache and bypass the memcg cache altogether.
3674 * There is one exception, though: the SLUB allocator does not create
3675 * large order caches, but rather service large kmallocs directly from
3676 * the page allocator. Therefore, the following sequence when backed by
3677 * the SLUB allocator:
3679 * memcg_stop_kmem_account();
3680 * kmalloc(<large_number>)
3681 * memcg_resume_kmem_account();
3683 * would effectively ignore the fact that we should skip accounting,
3684 * since it will drive us directly to this function without passing
3685 * through the cache selector memcg_kmem_get_cache. Such large
3686 * allocations are extremely rare but can happen, for instance, for the
3687 * cache arrays. We bring this test here.
3689 if (!current->mm || current->memcg_kmem_skip_account)
3692 memcg = try_get_mem_cgroup_from_mm(current->mm);
3695 * very rare case described in mem_cgroup_from_task. Unfortunately there
3696 * isn't much we can do without complicating this too much, and it would
3697 * be gfp-dependent anyway. Just let it go
3699 if (unlikely(!memcg))
3702 if (!memcg_can_account_kmem(memcg)) {
3703 css_put(&memcg->css);
3707 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3711 css_put(&memcg->css);
3715 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3718 struct page_cgroup *pc;
3720 VM_BUG_ON(mem_cgroup_is_root(memcg));
3722 /* The page allocation failed. Revert */
3724 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3728 pc = lookup_page_cgroup(page);
3729 lock_page_cgroup(pc);
3730 pc->mem_cgroup = memcg;
3731 SetPageCgroupUsed(pc);
3732 unlock_page_cgroup(pc);
3735 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3737 struct mem_cgroup *memcg = NULL;
3738 struct page_cgroup *pc;
3741 pc = lookup_page_cgroup(page);
3743 * Fast unlocked return. Theoretically might have changed, have to
3744 * check again after locking.
3746 if (!PageCgroupUsed(pc))
3749 lock_page_cgroup(pc);
3750 if (PageCgroupUsed(pc)) {
3751 memcg = pc->mem_cgroup;
3752 ClearPageCgroupUsed(pc);
3754 unlock_page_cgroup(pc);
3757 * We trust that only if there is a memcg associated with the page, it
3758 * is a valid allocation
3763 VM_BUG_ON(mem_cgroup_is_root(memcg));
3764 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3767 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3770 #endif /* CONFIG_MEMCG_KMEM */
3772 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3774 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3776 * Because tail pages are not marked as "used", set it. We're under
3777 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3778 * charge/uncharge will be never happen and move_account() is done under
3779 * compound_lock(), so we don't have to take care of races.
3781 void mem_cgroup_split_huge_fixup(struct page *head)
3783 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3784 struct page_cgroup *pc;
3785 struct mem_cgroup *memcg;
3788 if (mem_cgroup_disabled())
3791 memcg = head_pc->mem_cgroup;
3792 for (i = 1; i < HPAGE_PMD_NR; i++) {
3794 pc->mem_cgroup = memcg;
3795 smp_wmb();/* see __commit_charge() */
3796 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3798 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3801 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3804 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3805 struct mem_cgroup *to,
3806 unsigned int nr_pages,
3807 enum mem_cgroup_stat_index idx)
3809 /* Update stat data for mem_cgroup */
3811 WARN_ON_ONCE(from->stat->count[idx] < nr_pages);
3812 __this_cpu_add(from->stat->count[idx], -nr_pages);
3813 __this_cpu_add(to->stat->count[idx], nr_pages);
3818 * mem_cgroup_move_account - move account of the page
3820 * @nr_pages: number of regular pages (>1 for huge pages)
3821 * @pc: page_cgroup of the page.
3822 * @from: mem_cgroup which the page is moved from.
3823 * @to: mem_cgroup which the page is moved to. @from != @to.
3825 * The caller must confirm following.
3826 * - page is not on LRU (isolate_page() is useful.)
3827 * - compound_lock is held when nr_pages > 1
3829 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3832 static int mem_cgroup_move_account(struct page *page,
3833 unsigned int nr_pages,
3834 struct page_cgroup *pc,
3835 struct mem_cgroup *from,
3836 struct mem_cgroup *to)
3838 unsigned long flags;
3840 bool anon = PageAnon(page);
3842 VM_BUG_ON(from == to);
3843 VM_BUG_ON(PageLRU(page));
3845 * The page is isolated from LRU. So, collapse function
3846 * will not handle this page. But page splitting can happen.
3847 * Do this check under compound_page_lock(). The caller should
3851 if (nr_pages > 1 && !PageTransHuge(page))
3854 lock_page_cgroup(pc);
3857 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3860 move_lock_mem_cgroup(from, &flags);
3862 if (!anon && page_mapped(page))
3863 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3864 MEM_CGROUP_STAT_FILE_MAPPED);
3866 if (PageWriteback(page))
3867 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3868 MEM_CGROUP_STAT_WRITEBACK);
3870 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3872 /* caller should have done css_get */
3873 pc->mem_cgroup = to;
3874 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3875 move_unlock_mem_cgroup(from, &flags);
3878 unlock_page_cgroup(pc);
3882 memcg_check_events(to, page);
3883 memcg_check_events(from, page);
3889 * mem_cgroup_move_parent - moves page to the parent group
3890 * @page: the page to move
3891 * @pc: page_cgroup of the page
3892 * @child: page's cgroup
3894 * move charges to its parent or the root cgroup if the group has no
3895 * parent (aka use_hierarchy==0).
3896 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3897 * mem_cgroup_move_account fails) the failure is always temporary and
3898 * it signals a race with a page removal/uncharge or migration. In the
3899 * first case the page is on the way out and it will vanish from the LRU
3900 * on the next attempt and the call should be retried later.
3901 * Isolation from the LRU fails only if page has been isolated from
3902 * the LRU since we looked at it and that usually means either global
3903 * reclaim or migration going on. The page will either get back to the
3905 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3906 * (!PageCgroupUsed) or moved to a different group. The page will
3907 * disappear in the next attempt.
3909 static int mem_cgroup_move_parent(struct page *page,
3910 struct page_cgroup *pc,
3911 struct mem_cgroup *child)
3913 struct mem_cgroup *parent;
3914 unsigned int nr_pages;
3915 unsigned long uninitialized_var(flags);
3918 VM_BUG_ON(mem_cgroup_is_root(child));
3921 if (!get_page_unless_zero(page))
3923 if (isolate_lru_page(page))
3926 nr_pages = hpage_nr_pages(page);
3928 parent = parent_mem_cgroup(child);
3930 * If no parent, move charges to root cgroup.
3933 parent = root_mem_cgroup;
3936 VM_BUG_ON(!PageTransHuge(page));
3937 flags = compound_lock_irqsave(page);
3940 ret = mem_cgroup_move_account(page, nr_pages,
3943 __mem_cgroup_cancel_local_charge(child, nr_pages);
3946 compound_unlock_irqrestore(page, flags);
3947 putback_lru_page(page);
3955 * Charge the memory controller for page usage.
3957 * 0 if the charge was successful
3958 * < 0 if the cgroup is over its limit
3960 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3961 gfp_t gfp_mask, enum charge_type ctype)
3963 struct mem_cgroup *memcg = NULL;
3964 unsigned int nr_pages = 1;
3968 if (PageTransHuge(page)) {
3969 nr_pages <<= compound_order(page);
3970 VM_BUG_ON(!PageTransHuge(page));
3972 * Never OOM-kill a process for a huge page. The
3973 * fault handler will fall back to regular pages.
3978 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3981 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3985 int mem_cgroup_newpage_charge(struct page *page,
3986 struct mm_struct *mm, gfp_t gfp_mask)
3988 if (mem_cgroup_disabled())
3990 VM_BUG_ON(page_mapped(page));
3991 VM_BUG_ON(page->mapping && !PageAnon(page));
3993 return mem_cgroup_charge_common(page, mm, gfp_mask,
3994 MEM_CGROUP_CHARGE_TYPE_ANON);
3998 * While swap-in, try_charge -> commit or cancel, the page is locked.
3999 * And when try_charge() successfully returns, one refcnt to memcg without
4000 * struct page_cgroup is acquired. This refcnt will be consumed by
4001 * "commit()" or removed by "cancel()"
4003 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
4006 struct mem_cgroup **memcgp)
4008 struct mem_cgroup *memcg;
4009 struct page_cgroup *pc;
4012 pc = lookup_page_cgroup(page);
4014 * Every swap fault against a single page tries to charge the
4015 * page, bail as early as possible. shmem_unuse() encounters
4016 * already charged pages, too. The USED bit is protected by
4017 * the page lock, which serializes swap cache removal, which
4018 * in turn serializes uncharging.
4020 if (PageCgroupUsed(pc))
4022 if (!do_swap_account)
4024 memcg = try_get_mem_cgroup_from_page(page);
4028 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
4029 css_put(&memcg->css);
4034 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4040 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4041 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4044 if (mem_cgroup_disabled())
4047 * A racing thread's fault, or swapoff, may have already
4048 * updated the pte, and even removed page from swap cache: in
4049 * those cases unuse_pte()'s pte_same() test will fail; but
4050 * there's also a KSM case which does need to charge the page.
4052 if (!PageSwapCache(page)) {
4055 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4060 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4063 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4065 if (mem_cgroup_disabled())
4069 __mem_cgroup_cancel_charge(memcg, 1);
4073 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4074 enum charge_type ctype)
4076 if (mem_cgroup_disabled())
4081 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4083 * Now swap is on-memory. This means this page may be
4084 * counted both as mem and swap....double count.
4085 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4086 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4087 * may call delete_from_swap_cache() before reach here.
4089 if (do_swap_account && PageSwapCache(page)) {
4090 swp_entry_t ent = {.val = page_private(page)};
4091 mem_cgroup_uncharge_swap(ent);
4095 void mem_cgroup_commit_charge_swapin(struct page *page,
4096 struct mem_cgroup *memcg)
4098 __mem_cgroup_commit_charge_swapin(page, memcg,
4099 MEM_CGROUP_CHARGE_TYPE_ANON);
4102 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4105 struct mem_cgroup *memcg = NULL;
4106 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4109 if (mem_cgroup_disabled())
4111 if (PageCompound(page))
4114 if (!PageSwapCache(page))
4115 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4116 else { /* page is swapcache/shmem */
4117 ret = __mem_cgroup_try_charge_swapin(mm, page,
4120 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4125 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4126 unsigned int nr_pages,
4127 const enum charge_type ctype)
4129 struct memcg_batch_info *batch = NULL;
4130 bool uncharge_memsw = true;
4132 /* If swapout, usage of swap doesn't decrease */
4133 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4134 uncharge_memsw = false;
4136 batch = ¤t->memcg_batch;
4138 * In usual, we do css_get() when we remember memcg pointer.
4139 * But in this case, we keep res->usage until end of a series of
4140 * uncharges. Then, it's ok to ignore memcg's refcnt.
4143 batch->memcg = memcg;
4145 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4146 * In those cases, all pages freed continuously can be expected to be in
4147 * the same cgroup and we have chance to coalesce uncharges.
4148 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4149 * because we want to do uncharge as soon as possible.
4152 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4153 goto direct_uncharge;
4156 goto direct_uncharge;
4159 * In typical case, batch->memcg == mem. This means we can
4160 * merge a series of uncharges to an uncharge of res_counter.
4161 * If not, we uncharge res_counter ony by one.
4163 if (batch->memcg != memcg)
4164 goto direct_uncharge;
4165 /* remember freed charge and uncharge it later */
4168 batch->memsw_nr_pages++;
4171 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4173 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4174 if (unlikely(batch->memcg != memcg))
4175 memcg_oom_recover(memcg);
4179 * uncharge if !page_mapped(page)
4181 static struct mem_cgroup *
4182 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4185 struct mem_cgroup *memcg = NULL;
4186 unsigned int nr_pages = 1;
4187 struct page_cgroup *pc;
4190 if (mem_cgroup_disabled())
4193 if (PageTransHuge(page)) {
4194 nr_pages <<= compound_order(page);
4195 VM_BUG_ON(!PageTransHuge(page));
4198 * Check if our page_cgroup is valid
4200 pc = lookup_page_cgroup(page);
4201 if (unlikely(!PageCgroupUsed(pc)))
4204 lock_page_cgroup(pc);
4206 memcg = pc->mem_cgroup;
4208 if (!PageCgroupUsed(pc))
4211 anon = PageAnon(page);
4214 case MEM_CGROUP_CHARGE_TYPE_ANON:
4216 * Generally PageAnon tells if it's the anon statistics to be
4217 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4218 * used before page reached the stage of being marked PageAnon.
4222 case MEM_CGROUP_CHARGE_TYPE_DROP:
4223 /* See mem_cgroup_prepare_migration() */
4224 if (page_mapped(page))
4227 * Pages under migration may not be uncharged. But
4228 * end_migration() /must/ be the one uncharging the
4229 * unused post-migration page and so it has to call
4230 * here with the migration bit still set. See the
4231 * res_counter handling below.
4233 if (!end_migration && PageCgroupMigration(pc))
4236 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4237 if (!PageAnon(page)) { /* Shared memory */
4238 if (page->mapping && !page_is_file_cache(page))
4240 } else if (page_mapped(page)) /* Anon */
4247 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4249 ClearPageCgroupUsed(pc);
4251 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4252 * freed from LRU. This is safe because uncharged page is expected not
4253 * to be reused (freed soon). Exception is SwapCache, it's handled by
4254 * special functions.
4257 unlock_page_cgroup(pc);
4259 * even after unlock, we have memcg->res.usage here and this memcg
4260 * will never be freed, so it's safe to call css_get().
4262 memcg_check_events(memcg, page);
4263 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4264 mem_cgroup_swap_statistics(memcg, true);
4265 css_get(&memcg->css);
4268 * Migration does not charge the res_counter for the
4269 * replacement page, so leave it alone when phasing out the
4270 * page that is unused after the migration.
4272 if (!end_migration && !mem_cgroup_is_root(memcg))
4273 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4278 unlock_page_cgroup(pc);
4282 void mem_cgroup_uncharge_page(struct page *page)
4285 if (page_mapped(page))
4287 VM_BUG_ON(page->mapping && !PageAnon(page));
4289 * If the page is in swap cache, uncharge should be deferred
4290 * to the swap path, which also properly accounts swap usage
4291 * and handles memcg lifetime.
4293 * Note that this check is not stable and reclaim may add the
4294 * page to swap cache at any time after this. However, if the
4295 * page is not in swap cache by the time page->mapcount hits
4296 * 0, there won't be any page table references to the swap
4297 * slot, and reclaim will free it and not actually write the
4300 if (PageSwapCache(page))
4302 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4305 void mem_cgroup_uncharge_cache_page(struct page *page)
4307 VM_BUG_ON(page_mapped(page));
4308 VM_BUG_ON(page->mapping);
4309 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4313 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4314 * In that cases, pages are freed continuously and we can expect pages
4315 * are in the same memcg. All these calls itself limits the number of
4316 * pages freed at once, then uncharge_start/end() is called properly.
4317 * This may be called prural(2) times in a context,
4320 void mem_cgroup_uncharge_start(void)
4322 current->memcg_batch.do_batch++;
4323 /* We can do nest. */
4324 if (current->memcg_batch.do_batch == 1) {
4325 current->memcg_batch.memcg = NULL;
4326 current->memcg_batch.nr_pages = 0;
4327 current->memcg_batch.memsw_nr_pages = 0;
4331 void mem_cgroup_uncharge_end(void)
4333 struct memcg_batch_info *batch = ¤t->memcg_batch;
4335 if (!batch->do_batch)
4339 if (batch->do_batch) /* If stacked, do nothing. */
4345 * This "batch->memcg" is valid without any css_get/put etc...
4346 * bacause we hide charges behind us.
4348 if (batch->nr_pages)
4349 res_counter_uncharge(&batch->memcg->res,
4350 batch->nr_pages * PAGE_SIZE);
4351 if (batch->memsw_nr_pages)
4352 res_counter_uncharge(&batch->memcg->memsw,
4353 batch->memsw_nr_pages * PAGE_SIZE);
4354 memcg_oom_recover(batch->memcg);
4355 /* forget this pointer (for sanity check) */
4356 batch->memcg = NULL;
4361 * called after __delete_from_swap_cache() and drop "page" account.
4362 * memcg information is recorded to swap_cgroup of "ent"
4365 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4367 struct mem_cgroup *memcg;
4368 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4370 if (!swapout) /* this was a swap cache but the swap is unused ! */
4371 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4373 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4376 * record memcg information, if swapout && memcg != NULL,
4377 * css_get() was called in uncharge().
4379 if (do_swap_account && swapout && memcg)
4380 swap_cgroup_record(ent, css_id(&memcg->css));
4384 #ifdef CONFIG_MEMCG_SWAP
4386 * called from swap_entry_free(). remove record in swap_cgroup and
4387 * uncharge "memsw" account.
4389 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4391 struct mem_cgroup *memcg;
4394 if (!do_swap_account)
4397 id = swap_cgroup_record(ent, 0);
4399 memcg = mem_cgroup_lookup(id);
4402 * We uncharge this because swap is freed.
4403 * This memcg can be obsolete one. We avoid calling css_tryget
4405 if (!mem_cgroup_is_root(memcg))
4406 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4407 mem_cgroup_swap_statistics(memcg, false);
4408 css_put(&memcg->css);
4414 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4415 * @entry: swap entry to be moved
4416 * @from: mem_cgroup which the entry is moved from
4417 * @to: mem_cgroup which the entry is moved to
4419 * It succeeds only when the swap_cgroup's record for this entry is the same
4420 * as the mem_cgroup's id of @from.
4422 * Returns 0 on success, -EINVAL on failure.
4424 * The caller must have charged to @to, IOW, called res_counter_charge() about
4425 * both res and memsw, and called css_get().
4427 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4428 struct mem_cgroup *from, struct mem_cgroup *to)
4430 unsigned short old_id, new_id;
4432 old_id = css_id(&from->css);
4433 new_id = css_id(&to->css);
4435 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4436 mem_cgroup_swap_statistics(from, false);
4437 mem_cgroup_swap_statistics(to, true);
4439 * This function is only called from task migration context now.
4440 * It postpones res_counter and refcount handling till the end
4441 * of task migration(mem_cgroup_clear_mc()) for performance
4442 * improvement. But we cannot postpone css_get(to) because if
4443 * the process that has been moved to @to does swap-in, the
4444 * refcount of @to might be decreased to 0.
4446 * We are in attach() phase, so the cgroup is guaranteed to be
4447 * alive, so we can just call css_get().
4455 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4456 struct mem_cgroup *from, struct mem_cgroup *to)
4463 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4466 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4467 struct mem_cgroup **memcgp)
4469 struct mem_cgroup *memcg = NULL;
4470 unsigned int nr_pages = 1;
4471 struct page_cgroup *pc;
4472 enum charge_type ctype;
4476 if (mem_cgroup_disabled())
4479 if (PageTransHuge(page))
4480 nr_pages <<= compound_order(page);
4482 pc = lookup_page_cgroup(page);
4483 lock_page_cgroup(pc);
4484 if (PageCgroupUsed(pc)) {
4485 memcg = pc->mem_cgroup;
4486 css_get(&memcg->css);
4488 * At migrating an anonymous page, its mapcount goes down
4489 * to 0 and uncharge() will be called. But, even if it's fully
4490 * unmapped, migration may fail and this page has to be
4491 * charged again. We set MIGRATION flag here and delay uncharge
4492 * until end_migration() is called
4494 * Corner Case Thinking
4496 * When the old page was mapped as Anon and it's unmap-and-freed
4497 * while migration was ongoing.
4498 * If unmap finds the old page, uncharge() of it will be delayed
4499 * until end_migration(). If unmap finds a new page, it's
4500 * uncharged when it make mapcount to be 1->0. If unmap code
4501 * finds swap_migration_entry, the new page will not be mapped
4502 * and end_migration() will find it(mapcount==0).
4505 * When the old page was mapped but migraion fails, the kernel
4506 * remaps it. A charge for it is kept by MIGRATION flag even
4507 * if mapcount goes down to 0. We can do remap successfully
4508 * without charging it again.
4511 * The "old" page is under lock_page() until the end of
4512 * migration, so, the old page itself will not be swapped-out.
4513 * If the new page is swapped out before end_migraton, our
4514 * hook to usual swap-out path will catch the event.
4517 SetPageCgroupMigration(pc);
4519 unlock_page_cgroup(pc);
4521 * If the page is not charged at this point,
4529 * We charge new page before it's used/mapped. So, even if unlock_page()
4530 * is called before end_migration, we can catch all events on this new
4531 * page. In the case new page is migrated but not remapped, new page's
4532 * mapcount will be finally 0 and we call uncharge in end_migration().
4535 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4537 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4539 * The page is committed to the memcg, but it's not actually
4540 * charged to the res_counter since we plan on replacing the
4541 * old one and only one page is going to be left afterwards.
4543 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4546 /* remove redundant charge if migration failed*/
4547 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4548 struct page *oldpage, struct page *newpage, bool migration_ok)
4550 struct page *used, *unused;
4551 struct page_cgroup *pc;
4557 if (!migration_ok) {
4564 anon = PageAnon(used);
4565 __mem_cgroup_uncharge_common(unused,
4566 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4567 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4569 css_put(&memcg->css);
4571 * We disallowed uncharge of pages under migration because mapcount
4572 * of the page goes down to zero, temporarly.
4573 * Clear the flag and check the page should be charged.
4575 pc = lookup_page_cgroup(oldpage);
4576 lock_page_cgroup(pc);
4577 ClearPageCgroupMigration(pc);
4578 unlock_page_cgroup(pc);
4581 * If a page is a file cache, radix-tree replacement is very atomic
4582 * and we can skip this check. When it was an Anon page, its mapcount
4583 * goes down to 0. But because we added MIGRATION flage, it's not
4584 * uncharged yet. There are several case but page->mapcount check
4585 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4586 * check. (see prepare_charge() also)
4589 mem_cgroup_uncharge_page(used);
4593 * At replace page cache, newpage is not under any memcg but it's on
4594 * LRU. So, this function doesn't touch res_counter but handles LRU
4595 * in correct way. Both pages are locked so we cannot race with uncharge.
4597 void mem_cgroup_replace_page_cache(struct page *oldpage,
4598 struct page *newpage)
4600 struct mem_cgroup *memcg = NULL;
4601 struct page_cgroup *pc;
4602 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4604 if (mem_cgroup_disabled())
4607 pc = lookup_page_cgroup(oldpage);
4608 /* fix accounting on old pages */
4609 lock_page_cgroup(pc);
4610 if (PageCgroupUsed(pc)) {
4611 memcg = pc->mem_cgroup;
4612 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4613 ClearPageCgroupUsed(pc);
4615 unlock_page_cgroup(pc);
4618 * When called from shmem_replace_page(), in some cases the
4619 * oldpage has already been charged, and in some cases not.
4624 * Even if newpage->mapping was NULL before starting replacement,
4625 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4626 * LRU while we overwrite pc->mem_cgroup.
4628 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4631 #ifdef CONFIG_DEBUG_VM
4632 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4634 struct page_cgroup *pc;
4636 pc = lookup_page_cgroup(page);
4638 * Can be NULL while feeding pages into the page allocator for
4639 * the first time, i.e. during boot or memory hotplug;
4640 * or when mem_cgroup_disabled().
4642 if (likely(pc) && PageCgroupUsed(pc))
4647 bool mem_cgroup_bad_page_check(struct page *page)
4649 if (mem_cgroup_disabled())
4652 return lookup_page_cgroup_used(page) != NULL;
4655 void mem_cgroup_print_bad_page(struct page *page)
4657 struct page_cgroup *pc;
4659 pc = lookup_page_cgroup_used(page);
4661 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4662 pc, pc->flags, pc->mem_cgroup);
4667 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4668 unsigned long long val)
4671 u64 memswlimit, memlimit;
4673 int children = mem_cgroup_count_children(memcg);
4674 u64 curusage, oldusage;
4678 * For keeping hierarchical_reclaim simple, how long we should retry
4679 * is depends on callers. We set our retry-count to be function
4680 * of # of children which we should visit in this loop.
4682 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4684 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4687 while (retry_count) {
4688 if (signal_pending(current)) {
4693 * Rather than hide all in some function, I do this in
4694 * open coded manner. You see what this really does.
4695 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4697 mutex_lock(&set_limit_mutex);
4698 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4699 if (memswlimit < val) {
4701 mutex_unlock(&set_limit_mutex);
4705 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4709 ret = res_counter_set_limit(&memcg->res, val);
4711 if (memswlimit == val)
4712 memcg->memsw_is_minimum = true;
4714 memcg->memsw_is_minimum = false;
4716 mutex_unlock(&set_limit_mutex);
4721 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4722 MEM_CGROUP_RECLAIM_SHRINK);
4723 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4724 /* Usage is reduced ? */
4725 if (curusage >= oldusage)
4728 oldusage = curusage;
4730 if (!ret && enlarge)
4731 memcg_oom_recover(memcg);
4736 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4737 unsigned long long val)
4740 u64 memlimit, memswlimit, oldusage, curusage;
4741 int children = mem_cgroup_count_children(memcg);
4745 /* see mem_cgroup_resize_res_limit */
4746 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4747 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4748 while (retry_count) {
4749 if (signal_pending(current)) {
4754 * Rather than hide all in some function, I do this in
4755 * open coded manner. You see what this really does.
4756 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4758 mutex_lock(&set_limit_mutex);
4759 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4760 if (memlimit > val) {
4762 mutex_unlock(&set_limit_mutex);
4765 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4766 if (memswlimit < val)
4768 ret = res_counter_set_limit(&memcg->memsw, val);
4770 if (memlimit == val)
4771 memcg->memsw_is_minimum = true;
4773 memcg->memsw_is_minimum = false;
4775 mutex_unlock(&set_limit_mutex);
4780 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4781 MEM_CGROUP_RECLAIM_NOSWAP |
4782 MEM_CGROUP_RECLAIM_SHRINK);
4783 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4784 /* Usage is reduced ? */
4785 if (curusage >= oldusage)
4788 oldusage = curusage;
4790 if (!ret && enlarge)
4791 memcg_oom_recover(memcg);
4795 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4797 unsigned long *total_scanned)
4799 unsigned long nr_reclaimed = 0;
4800 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4801 unsigned long reclaimed;
4803 struct mem_cgroup_tree_per_zone *mctz;
4804 unsigned long long excess;
4805 unsigned long nr_scanned;
4810 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4812 * This loop can run a while, specially if mem_cgroup's continuously
4813 * keep exceeding their soft limit and putting the system under
4820 mz = mem_cgroup_largest_soft_limit_node(mctz);
4825 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4826 gfp_mask, &nr_scanned);
4827 nr_reclaimed += reclaimed;
4828 *total_scanned += nr_scanned;
4829 spin_lock(&mctz->lock);
4832 * If we failed to reclaim anything from this memory cgroup
4833 * it is time to move on to the next cgroup
4839 * Loop until we find yet another one.
4841 * By the time we get the soft_limit lock
4842 * again, someone might have aded the
4843 * group back on the RB tree. Iterate to
4844 * make sure we get a different mem.
4845 * mem_cgroup_largest_soft_limit_node returns
4846 * NULL if no other cgroup is present on
4850 __mem_cgroup_largest_soft_limit_node(mctz);
4852 css_put(&next_mz->memcg->css);
4853 else /* next_mz == NULL or other memcg */
4857 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4858 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4860 * One school of thought says that we should not add
4861 * back the node to the tree if reclaim returns 0.
4862 * But our reclaim could return 0, simply because due
4863 * to priority we are exposing a smaller subset of
4864 * memory to reclaim from. Consider this as a longer
4867 /* If excess == 0, no tree ops */
4868 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4869 spin_unlock(&mctz->lock);
4870 css_put(&mz->memcg->css);
4873 * Could not reclaim anything and there are no more
4874 * mem cgroups to try or we seem to be looping without
4875 * reclaiming anything.
4877 if (!nr_reclaimed &&
4879 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4881 } while (!nr_reclaimed);
4883 css_put(&next_mz->memcg->css);
4884 return nr_reclaimed;
4888 * mem_cgroup_force_empty_list - clears LRU of a group
4889 * @memcg: group to clear
4892 * @lru: lru to to clear
4894 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4895 * reclaim the pages page themselves - pages are moved to the parent (or root)
4898 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4899 int node, int zid, enum lru_list lru)
4901 struct lruvec *lruvec;
4902 unsigned long flags;
4903 struct list_head *list;
4907 zone = &NODE_DATA(node)->node_zones[zid];
4908 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4909 list = &lruvec->lists[lru];
4913 struct page_cgroup *pc;
4916 spin_lock_irqsave(&zone->lru_lock, flags);
4917 if (list_empty(list)) {
4918 spin_unlock_irqrestore(&zone->lru_lock, flags);
4921 page = list_entry(list->prev, struct page, lru);
4923 list_move(&page->lru, list);
4925 spin_unlock_irqrestore(&zone->lru_lock, flags);
4928 spin_unlock_irqrestore(&zone->lru_lock, flags);
4930 pc = lookup_page_cgroup(page);
4932 if (mem_cgroup_move_parent(page, pc, memcg)) {
4933 /* found lock contention or "pc" is obsolete. */
4938 } while (!list_empty(list));
4942 * make mem_cgroup's charge to be 0 if there is no task by moving
4943 * all the charges and pages to the parent.
4944 * This enables deleting this mem_cgroup.
4946 * Caller is responsible for holding css reference on the memcg.
4948 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4954 /* This is for making all *used* pages to be on LRU. */
4955 lru_add_drain_all();
4956 drain_all_stock_sync(memcg);
4957 mem_cgroup_start_move(memcg);
4958 for_each_node_state(node, N_MEMORY) {
4959 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4962 mem_cgroup_force_empty_list(memcg,
4967 mem_cgroup_end_move(memcg);
4968 memcg_oom_recover(memcg);
4972 * Kernel memory may not necessarily be trackable to a specific
4973 * process. So they are not migrated, and therefore we can't
4974 * expect their value to drop to 0 here.
4975 * Having res filled up with kmem only is enough.
4977 * This is a safety check because mem_cgroup_force_empty_list
4978 * could have raced with mem_cgroup_replace_page_cache callers
4979 * so the lru seemed empty but the page could have been added
4980 * right after the check. RES_USAGE should be safe as we always
4981 * charge before adding to the LRU.
4983 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4984 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4985 } while (usage > 0);
4989 * This mainly exists for tests during the setting of set of use_hierarchy.
4990 * Since this is the very setting we are changing, the current hierarchy value
4993 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4995 struct cgroup_subsys_state *pos;
4997 /* bounce at first found */
4998 css_for_each_child(pos, &memcg->css)
5004 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
5005 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
5006 * from mem_cgroup_count_children(), in the sense that we don't really care how
5007 * many children we have; we only need to know if we have any. It also counts
5008 * any memcg without hierarchy as infertile.
5010 static inline bool memcg_has_children(struct mem_cgroup *memcg)
5012 return memcg->use_hierarchy && __memcg_has_children(memcg);
5016 * Reclaims as many pages from the given memcg as possible and moves
5017 * the rest to the parent.
5019 * Caller is responsible for holding css reference for memcg.
5021 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
5023 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
5024 struct cgroup *cgrp = memcg->css.cgroup;
5026 /* returns EBUSY if there is a task or if we come here twice. */
5027 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
5030 /* we call try-to-free pages for make this cgroup empty */
5031 lru_add_drain_all();
5032 /* try to free all pages in this cgroup */
5033 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
5036 if (signal_pending(current))
5039 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
5043 /* maybe some writeback is necessary */
5044 congestion_wait(BLK_RW_ASYNC, HZ/10);
5049 mem_cgroup_reparent_charges(memcg);
5054 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
5057 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5059 if (mem_cgroup_is_root(memcg))
5061 return mem_cgroup_force_empty(memcg);
5064 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5067 return mem_cgroup_from_css(css)->use_hierarchy;
5070 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5071 struct cftype *cft, u64 val)
5074 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5075 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5077 mutex_lock(&memcg_create_mutex);
5079 if (memcg->use_hierarchy == val)
5083 * If parent's use_hierarchy is set, we can't make any modifications
5084 * in the child subtrees. If it is unset, then the change can
5085 * occur, provided the current cgroup has no children.
5087 * For the root cgroup, parent_mem is NULL, we allow value to be
5088 * set if there are no children.
5090 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5091 (val == 1 || val == 0)) {
5092 if (!__memcg_has_children(memcg))
5093 memcg->use_hierarchy = val;
5100 mutex_unlock(&memcg_create_mutex);
5106 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5107 enum mem_cgroup_stat_index idx)
5109 struct mem_cgroup *iter;
5112 /* Per-cpu values can be negative, use a signed accumulator */
5113 for_each_mem_cgroup_tree(iter, memcg)
5114 val += mem_cgroup_read_stat(iter, idx);
5116 if (val < 0) /* race ? */
5121 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5125 if (!mem_cgroup_is_root(memcg)) {
5127 return res_counter_read_u64(&memcg->res, RES_USAGE);
5129 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5133 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5134 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5136 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5137 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5140 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5142 return val << PAGE_SHIFT;
5145 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
5146 struct cftype *cft, struct file *file,
5147 char __user *buf, size_t nbytes, loff_t *ppos)
5149 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5155 type = MEMFILE_TYPE(cft->private);
5156 name = MEMFILE_ATTR(cft->private);
5160 if (name == RES_USAGE)
5161 val = mem_cgroup_usage(memcg, false);
5163 val = res_counter_read_u64(&memcg->res, name);
5166 if (name == RES_USAGE)
5167 val = mem_cgroup_usage(memcg, true);
5169 val = res_counter_read_u64(&memcg->memsw, name);
5172 val = res_counter_read_u64(&memcg->kmem, name);
5178 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5179 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5182 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
5185 #ifdef CONFIG_MEMCG_KMEM
5186 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5188 * For simplicity, we won't allow this to be disabled. It also can't
5189 * be changed if the cgroup has children already, or if tasks had
5192 * If tasks join before we set the limit, a person looking at
5193 * kmem.usage_in_bytes will have no way to determine when it took
5194 * place, which makes the value quite meaningless.
5196 * After it first became limited, changes in the value of the limit are
5197 * of course permitted.
5199 mutex_lock(&memcg_create_mutex);
5200 mutex_lock(&set_limit_mutex);
5201 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
5202 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
5206 ret = res_counter_set_limit(&memcg->kmem, val);
5209 ret = memcg_update_cache_sizes(memcg);
5211 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
5214 static_key_slow_inc(&memcg_kmem_enabled_key);
5216 * setting the active bit after the inc will guarantee no one
5217 * starts accounting before all call sites are patched
5219 memcg_kmem_set_active(memcg);
5221 ret = res_counter_set_limit(&memcg->kmem, val);
5223 mutex_unlock(&set_limit_mutex);
5224 mutex_unlock(&memcg_create_mutex);
5229 #ifdef CONFIG_MEMCG_KMEM
5230 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5233 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5237 memcg->kmem_account_flags = parent->kmem_account_flags;
5239 * When that happen, we need to disable the static branch only on those
5240 * memcgs that enabled it. To achieve this, we would be forced to
5241 * complicate the code by keeping track of which memcgs were the ones
5242 * that actually enabled limits, and which ones got it from its
5245 * It is a lot simpler just to do static_key_slow_inc() on every child
5246 * that is accounted.
5248 if (!memcg_kmem_is_active(memcg))
5252 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5253 * memcg is active already. If the later initialization fails then the
5254 * cgroup core triggers the cleanup so we do not have to do it here.
5256 static_key_slow_inc(&memcg_kmem_enabled_key);
5258 mutex_lock(&set_limit_mutex);
5259 memcg_stop_kmem_account();
5260 ret = memcg_update_cache_sizes(memcg);
5261 memcg_resume_kmem_account();
5262 mutex_unlock(&set_limit_mutex);
5266 #endif /* CONFIG_MEMCG_KMEM */
5269 * The user of this function is...
5272 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5275 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5278 unsigned long long val;
5281 type = MEMFILE_TYPE(cft->private);
5282 name = MEMFILE_ATTR(cft->private);
5286 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5290 /* This function does all necessary parse...reuse it */
5291 ret = res_counter_memparse_write_strategy(buffer, &val);
5295 ret = mem_cgroup_resize_limit(memcg, val);
5296 else if (type == _MEMSWAP)
5297 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5298 else if (type == _KMEM)
5299 ret = memcg_update_kmem_limit(css, val);
5303 case RES_SOFT_LIMIT:
5304 ret = res_counter_memparse_write_strategy(buffer, &val);
5308 * For memsw, soft limits are hard to implement in terms
5309 * of semantics, for now, we support soft limits for
5310 * control without swap
5313 ret = res_counter_set_soft_limit(&memcg->res, val);
5318 ret = -EINVAL; /* should be BUG() ? */
5324 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5325 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5327 unsigned long long min_limit, min_memsw_limit, tmp;
5329 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5330 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5331 if (!memcg->use_hierarchy)
5334 while (css_parent(&memcg->css)) {
5335 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5336 if (!memcg->use_hierarchy)
5338 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5339 min_limit = min(min_limit, tmp);
5340 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5341 min_memsw_limit = min(min_memsw_limit, tmp);
5344 *mem_limit = min_limit;
5345 *memsw_limit = min_memsw_limit;
5348 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5350 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5354 type = MEMFILE_TYPE(event);
5355 name = MEMFILE_ATTR(event);
5360 res_counter_reset_max(&memcg->res);
5361 else if (type == _MEMSWAP)
5362 res_counter_reset_max(&memcg->memsw);
5363 else if (type == _KMEM)
5364 res_counter_reset_max(&memcg->kmem);
5370 res_counter_reset_failcnt(&memcg->res);
5371 else if (type == _MEMSWAP)
5372 res_counter_reset_failcnt(&memcg->memsw);
5373 else if (type == _KMEM)
5374 res_counter_reset_failcnt(&memcg->kmem);
5383 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5386 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5390 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5391 struct cftype *cft, u64 val)
5393 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5395 if (val >= (1 << NR_MOVE_TYPE))
5399 * No kind of locking is needed in here, because ->can_attach() will
5400 * check this value once in the beginning of the process, and then carry
5401 * on with stale data. This means that changes to this value will only
5402 * affect task migrations starting after the change.
5404 memcg->move_charge_at_immigrate = val;
5408 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5409 struct cftype *cft, u64 val)
5416 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5417 struct cftype *cft, struct seq_file *m)
5420 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5421 unsigned long node_nr;
5422 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5424 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5425 seq_printf(m, "total=%lu", total_nr);
5426 for_each_node_state(nid, N_MEMORY) {
5427 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5428 seq_printf(m, " N%d=%lu", nid, node_nr);
5432 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5433 seq_printf(m, "file=%lu", file_nr);
5434 for_each_node_state(nid, N_MEMORY) {
5435 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5437 seq_printf(m, " N%d=%lu", nid, node_nr);
5441 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5442 seq_printf(m, "anon=%lu", anon_nr);
5443 for_each_node_state(nid, N_MEMORY) {
5444 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5446 seq_printf(m, " N%d=%lu", nid, node_nr);
5450 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5451 seq_printf(m, "unevictable=%lu", unevictable_nr);
5452 for_each_node_state(nid, N_MEMORY) {
5453 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5454 BIT(LRU_UNEVICTABLE));
5455 seq_printf(m, " N%d=%lu", nid, node_nr);
5460 #endif /* CONFIG_NUMA */
5462 static inline void mem_cgroup_lru_names_not_uptodate(void)
5464 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5467 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5470 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5471 struct mem_cgroup *mi;
5474 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5475 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5477 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5478 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5481 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5482 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5483 mem_cgroup_read_events(memcg, i));
5485 for (i = 0; i < NR_LRU_LISTS; i++)
5486 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5487 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5489 /* Hierarchical information */
5491 unsigned long long limit, memsw_limit;
5492 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5493 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5494 if (do_swap_account)
5495 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5499 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5502 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5504 for_each_mem_cgroup_tree(mi, memcg)
5505 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5506 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5509 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5510 unsigned long long val = 0;
5512 for_each_mem_cgroup_tree(mi, memcg)
5513 val += mem_cgroup_read_events(mi, i);
5514 seq_printf(m, "total_%s %llu\n",
5515 mem_cgroup_events_names[i], val);
5518 for (i = 0; i < NR_LRU_LISTS; i++) {
5519 unsigned long long val = 0;
5521 for_each_mem_cgroup_tree(mi, memcg)
5522 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5523 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5526 #ifdef CONFIG_DEBUG_VM
5529 struct mem_cgroup_per_zone *mz;
5530 struct zone_reclaim_stat *rstat;
5531 unsigned long recent_rotated[2] = {0, 0};
5532 unsigned long recent_scanned[2] = {0, 0};
5534 for_each_online_node(nid)
5535 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5536 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5537 rstat = &mz->lruvec.reclaim_stat;
5539 recent_rotated[0] += rstat->recent_rotated[0];
5540 recent_rotated[1] += rstat->recent_rotated[1];
5541 recent_scanned[0] += rstat->recent_scanned[0];
5542 recent_scanned[1] += rstat->recent_scanned[1];
5544 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5545 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5546 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5547 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5554 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5557 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5559 return mem_cgroup_swappiness(memcg);
5562 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5563 struct cftype *cft, u64 val)
5565 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5566 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5568 if (val > 100 || !parent)
5571 mutex_lock(&memcg_create_mutex);
5573 /* If under hierarchy, only empty-root can set this value */
5574 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5575 mutex_unlock(&memcg_create_mutex);
5579 memcg->swappiness = val;
5581 mutex_unlock(&memcg_create_mutex);
5586 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5588 struct mem_cgroup_threshold_ary *t;
5594 t = rcu_dereference(memcg->thresholds.primary);
5596 t = rcu_dereference(memcg->memsw_thresholds.primary);
5601 usage = mem_cgroup_usage(memcg, swap);
5604 * current_threshold points to threshold just below or equal to usage.
5605 * If it's not true, a threshold was crossed after last
5606 * call of __mem_cgroup_threshold().
5608 i = t->current_threshold;
5611 * Iterate backward over array of thresholds starting from
5612 * current_threshold and check if a threshold is crossed.
5613 * If none of thresholds below usage is crossed, we read
5614 * only one element of the array here.
5616 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5617 eventfd_signal(t->entries[i].eventfd, 1);
5619 /* i = current_threshold + 1 */
5623 * Iterate forward over array of thresholds starting from
5624 * current_threshold+1 and check if a threshold is crossed.
5625 * If none of thresholds above usage is crossed, we read
5626 * only one element of the array here.
5628 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5629 eventfd_signal(t->entries[i].eventfd, 1);
5631 /* Update current_threshold */
5632 t->current_threshold = i - 1;
5637 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5640 __mem_cgroup_threshold(memcg, false);
5641 if (do_swap_account)
5642 __mem_cgroup_threshold(memcg, true);
5644 memcg = parent_mem_cgroup(memcg);
5648 static int compare_thresholds(const void *a, const void *b)
5650 const struct mem_cgroup_threshold *_a = a;
5651 const struct mem_cgroup_threshold *_b = b;
5653 if (_a->threshold > _b->threshold)
5656 if (_a->threshold < _b->threshold)
5662 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5664 struct mem_cgroup_eventfd_list *ev;
5666 list_for_each_entry(ev, &memcg->oom_notify, list)
5667 eventfd_signal(ev->eventfd, 1);
5671 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5673 struct mem_cgroup *iter;
5675 for_each_mem_cgroup_tree(iter, memcg)
5676 mem_cgroup_oom_notify_cb(iter);
5679 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5680 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5682 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5683 struct mem_cgroup_thresholds *thresholds;
5684 struct mem_cgroup_threshold_ary *new;
5685 enum res_type type = MEMFILE_TYPE(cft->private);
5686 u64 threshold, usage;
5689 ret = res_counter_memparse_write_strategy(args, &threshold);
5693 mutex_lock(&memcg->thresholds_lock);
5696 thresholds = &memcg->thresholds;
5697 else if (type == _MEMSWAP)
5698 thresholds = &memcg->memsw_thresholds;
5702 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5704 /* Check if a threshold crossed before adding a new one */
5705 if (thresholds->primary)
5706 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5708 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5710 /* Allocate memory for new array of thresholds */
5711 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5719 /* Copy thresholds (if any) to new array */
5720 if (thresholds->primary) {
5721 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5722 sizeof(struct mem_cgroup_threshold));
5725 /* Add new threshold */
5726 new->entries[size - 1].eventfd = eventfd;
5727 new->entries[size - 1].threshold = threshold;
5729 /* Sort thresholds. Registering of new threshold isn't time-critical */
5730 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5731 compare_thresholds, NULL);
5733 /* Find current threshold */
5734 new->current_threshold = -1;
5735 for (i = 0; i < size; i++) {
5736 if (new->entries[i].threshold <= usage) {
5738 * new->current_threshold will not be used until
5739 * rcu_assign_pointer(), so it's safe to increment
5742 ++new->current_threshold;
5747 /* Free old spare buffer and save old primary buffer as spare */
5748 kfree(thresholds->spare);
5749 thresholds->spare = thresholds->primary;
5751 rcu_assign_pointer(thresholds->primary, new);
5753 /* To be sure that nobody uses thresholds */
5757 mutex_unlock(&memcg->thresholds_lock);
5762 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5763 struct cftype *cft, struct eventfd_ctx *eventfd)
5765 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5766 struct mem_cgroup_thresholds *thresholds;
5767 struct mem_cgroup_threshold_ary *new;
5768 enum res_type type = MEMFILE_TYPE(cft->private);
5772 mutex_lock(&memcg->thresholds_lock);
5774 thresholds = &memcg->thresholds;
5775 else if (type == _MEMSWAP)
5776 thresholds = &memcg->memsw_thresholds;
5780 if (!thresholds->primary)
5783 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5785 /* Check if a threshold crossed before removing */
5786 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5788 /* Calculate new number of threshold */
5790 for (i = 0; i < thresholds->primary->size; i++) {
5791 if (thresholds->primary->entries[i].eventfd != eventfd)
5795 new = thresholds->spare;
5797 /* Set thresholds array to NULL if we don't have thresholds */
5806 /* Copy thresholds and find current threshold */
5807 new->current_threshold = -1;
5808 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5809 if (thresholds->primary->entries[i].eventfd == eventfd)
5812 new->entries[j] = thresholds->primary->entries[i];
5813 if (new->entries[j].threshold <= usage) {
5815 * new->current_threshold will not be used
5816 * until rcu_assign_pointer(), so it's safe to increment
5819 ++new->current_threshold;
5825 /* Swap primary and spare array */
5826 thresholds->spare = thresholds->primary;
5827 /* If all events are unregistered, free the spare array */
5829 kfree(thresholds->spare);
5830 thresholds->spare = NULL;
5833 rcu_assign_pointer(thresholds->primary, new);
5835 /* To be sure that nobody uses thresholds */
5838 mutex_unlock(&memcg->thresholds_lock);
5841 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5842 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5844 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5845 struct mem_cgroup_eventfd_list *event;
5846 enum res_type type = MEMFILE_TYPE(cft->private);
5848 BUG_ON(type != _OOM_TYPE);
5849 event = kmalloc(sizeof(*event), GFP_KERNEL);
5853 spin_lock(&memcg_oom_lock);
5855 event->eventfd = eventfd;
5856 list_add(&event->list, &memcg->oom_notify);
5858 /* already in OOM ? */
5859 if (atomic_read(&memcg->under_oom))
5860 eventfd_signal(eventfd, 1);
5861 spin_unlock(&memcg_oom_lock);
5866 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5867 struct cftype *cft, struct eventfd_ctx *eventfd)
5869 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5870 struct mem_cgroup_eventfd_list *ev, *tmp;
5871 enum res_type type = MEMFILE_TYPE(cft->private);
5873 BUG_ON(type != _OOM_TYPE);
5875 spin_lock(&memcg_oom_lock);
5877 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5878 if (ev->eventfd == eventfd) {
5879 list_del(&ev->list);
5884 spin_unlock(&memcg_oom_lock);
5887 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5888 struct cftype *cft, struct cgroup_map_cb *cb)
5890 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5892 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5894 if (atomic_read(&memcg->under_oom))
5895 cb->fill(cb, "under_oom", 1);
5897 cb->fill(cb, "under_oom", 0);
5901 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5902 struct cftype *cft, u64 val)
5904 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5905 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5907 /* cannot set to root cgroup and only 0 and 1 are allowed */
5908 if (!parent || !((val == 0) || (val == 1)))
5911 mutex_lock(&memcg_create_mutex);
5912 /* oom-kill-disable is a flag for subhierarchy. */
5913 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5914 mutex_unlock(&memcg_create_mutex);
5917 memcg->oom_kill_disable = val;
5919 memcg_oom_recover(memcg);
5920 mutex_unlock(&memcg_create_mutex);
5924 #ifdef CONFIG_MEMCG_KMEM
5925 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5929 memcg->kmemcg_id = -1;
5930 ret = memcg_propagate_kmem(memcg);
5934 return mem_cgroup_sockets_init(memcg, ss);
5937 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5939 mem_cgroup_sockets_destroy(memcg);
5942 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5944 if (!memcg_kmem_is_active(memcg))
5948 * kmem charges can outlive the cgroup. In the case of slab
5949 * pages, for instance, a page contain objects from various
5950 * processes. As we prevent from taking a reference for every
5951 * such allocation we have to be careful when doing uncharge
5952 * (see memcg_uncharge_kmem) and here during offlining.
5954 * The idea is that that only the _last_ uncharge which sees
5955 * the dead memcg will drop the last reference. An additional
5956 * reference is taken here before the group is marked dead
5957 * which is then paired with css_put during uncharge resp. here.
5959 * Although this might sound strange as this path is called from
5960 * css_offline() when the referencemight have dropped down to 0
5961 * and shouldn't be incremented anymore (css_tryget would fail)
5962 * we do not have other options because of the kmem allocations
5965 css_get(&memcg->css);
5967 memcg_kmem_mark_dead(memcg);
5969 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5972 if (memcg_kmem_test_and_clear_dead(memcg))
5973 css_put(&memcg->css);
5976 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5981 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5985 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5990 static struct cftype mem_cgroup_files[] = {
5992 .name = "usage_in_bytes",
5993 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5994 .read = mem_cgroup_read,
5995 .register_event = mem_cgroup_usage_register_event,
5996 .unregister_event = mem_cgroup_usage_unregister_event,
5999 .name = "max_usage_in_bytes",
6000 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
6001 .trigger = mem_cgroup_reset,
6002 .read = mem_cgroup_read,
6005 .name = "limit_in_bytes",
6006 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
6007 .write_string = mem_cgroup_write,
6008 .read = mem_cgroup_read,
6011 .name = "soft_limit_in_bytes",
6012 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
6013 .write_string = mem_cgroup_write,
6014 .read = mem_cgroup_read,
6018 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
6019 .trigger = mem_cgroup_reset,
6020 .read = mem_cgroup_read,
6024 .read_seq_string = memcg_stat_show,
6027 .name = "force_empty",
6028 .trigger = mem_cgroup_force_empty_write,
6031 .name = "use_hierarchy",
6032 .flags = CFTYPE_INSANE,
6033 .write_u64 = mem_cgroup_hierarchy_write,
6034 .read_u64 = mem_cgroup_hierarchy_read,
6037 .name = "swappiness",
6038 .read_u64 = mem_cgroup_swappiness_read,
6039 .write_u64 = mem_cgroup_swappiness_write,
6042 .name = "move_charge_at_immigrate",
6043 .read_u64 = mem_cgroup_move_charge_read,
6044 .write_u64 = mem_cgroup_move_charge_write,
6047 .name = "oom_control",
6048 .read_map = mem_cgroup_oom_control_read,
6049 .write_u64 = mem_cgroup_oom_control_write,
6050 .register_event = mem_cgroup_oom_register_event,
6051 .unregister_event = mem_cgroup_oom_unregister_event,
6052 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
6055 .name = "pressure_level",
6056 .register_event = vmpressure_register_event,
6057 .unregister_event = vmpressure_unregister_event,
6061 .name = "numa_stat",
6062 .read_seq_string = memcg_numa_stat_show,
6065 #ifdef CONFIG_MEMCG_KMEM
6067 .name = "kmem.limit_in_bytes",
6068 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6069 .write_string = mem_cgroup_write,
6070 .read = mem_cgroup_read,
6073 .name = "kmem.usage_in_bytes",
6074 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6075 .read = mem_cgroup_read,
6078 .name = "kmem.failcnt",
6079 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6080 .trigger = mem_cgroup_reset,
6081 .read = mem_cgroup_read,
6084 .name = "kmem.max_usage_in_bytes",
6085 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6086 .trigger = mem_cgroup_reset,
6087 .read = mem_cgroup_read,
6089 #ifdef CONFIG_SLABINFO
6091 .name = "kmem.slabinfo",
6092 .read_seq_string = mem_cgroup_slabinfo_read,
6096 { }, /* terminate */
6099 #ifdef CONFIG_MEMCG_SWAP
6100 static struct cftype memsw_cgroup_files[] = {
6102 .name = "memsw.usage_in_bytes",
6103 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6104 .read = mem_cgroup_read,
6105 .register_event = mem_cgroup_usage_register_event,
6106 .unregister_event = mem_cgroup_usage_unregister_event,
6109 .name = "memsw.max_usage_in_bytes",
6110 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6111 .trigger = mem_cgroup_reset,
6112 .read = mem_cgroup_read,
6115 .name = "memsw.limit_in_bytes",
6116 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6117 .write_string = mem_cgroup_write,
6118 .read = mem_cgroup_read,
6121 .name = "memsw.failcnt",
6122 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6123 .trigger = mem_cgroup_reset,
6124 .read = mem_cgroup_read,
6126 { }, /* terminate */
6129 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6131 struct mem_cgroup_per_node *pn;
6132 struct mem_cgroup_per_zone *mz;
6133 int zone, tmp = node;
6135 * This routine is called against possible nodes.
6136 * But it's BUG to call kmalloc() against offline node.
6138 * TODO: this routine can waste much memory for nodes which will
6139 * never be onlined. It's better to use memory hotplug callback
6142 if (!node_state(node, N_NORMAL_MEMORY))
6144 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6148 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6149 mz = &pn->zoneinfo[zone];
6150 lruvec_init(&mz->lruvec);
6151 mz->usage_in_excess = 0;
6152 mz->on_tree = false;
6155 memcg->nodeinfo[node] = pn;
6159 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6161 kfree(memcg->nodeinfo[node]);
6164 static struct mem_cgroup *mem_cgroup_alloc(void)
6166 struct mem_cgroup *memcg;
6167 size_t size = memcg_size();
6169 /* Can be very big if nr_node_ids is very big */
6170 if (size < PAGE_SIZE)
6171 memcg = kzalloc(size, GFP_KERNEL);
6173 memcg = vzalloc(size);
6178 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6181 spin_lock_init(&memcg->pcp_counter_lock);
6185 if (size < PAGE_SIZE)
6193 * At destroying mem_cgroup, references from swap_cgroup can remain.
6194 * (scanning all at force_empty is too costly...)
6196 * Instead of clearing all references at force_empty, we remember
6197 * the number of reference from swap_cgroup and free mem_cgroup when
6198 * it goes down to 0.
6200 * Removal of cgroup itself succeeds regardless of refs from swap.
6203 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6206 size_t size = memcg_size();
6208 mem_cgroup_remove_from_trees(memcg);
6209 free_css_id(&mem_cgroup_subsys, &memcg->css);
6212 free_mem_cgroup_per_zone_info(memcg, node);
6214 free_percpu(memcg->stat);
6217 * We need to make sure that (at least for now), the jump label
6218 * destruction code runs outside of the cgroup lock. This is because
6219 * get_online_cpus(), which is called from the static_branch update,
6220 * can't be called inside the cgroup_lock. cpusets are the ones
6221 * enforcing this dependency, so if they ever change, we might as well.
6223 * schedule_work() will guarantee this happens. Be careful if you need
6224 * to move this code around, and make sure it is outside
6227 disarm_static_keys(memcg);
6228 if (size < PAGE_SIZE)
6235 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6237 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6239 if (!memcg->res.parent)
6241 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6243 EXPORT_SYMBOL(parent_mem_cgroup);
6245 static void __init mem_cgroup_soft_limit_tree_init(void)
6247 struct mem_cgroup_tree_per_node *rtpn;
6248 struct mem_cgroup_tree_per_zone *rtpz;
6249 int tmp, node, zone;
6251 for_each_node(node) {
6253 if (!node_state(node, N_NORMAL_MEMORY))
6255 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6258 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6260 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6261 rtpz = &rtpn->rb_tree_per_zone[zone];
6262 rtpz->rb_root = RB_ROOT;
6263 spin_lock_init(&rtpz->lock);
6268 static struct cgroup_subsys_state * __ref
6269 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6271 struct mem_cgroup *memcg;
6272 long error = -ENOMEM;
6275 memcg = mem_cgroup_alloc();
6277 return ERR_PTR(error);
6280 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6284 if (parent_css == NULL) {
6285 root_mem_cgroup = memcg;
6286 res_counter_init(&memcg->res, NULL);
6287 res_counter_init(&memcg->memsw, NULL);
6288 res_counter_init(&memcg->kmem, NULL);
6291 memcg->last_scanned_node = MAX_NUMNODES;
6292 INIT_LIST_HEAD(&memcg->oom_notify);
6293 memcg->move_charge_at_immigrate = 0;
6294 mutex_init(&memcg->thresholds_lock);
6295 spin_lock_init(&memcg->move_lock);
6296 vmpressure_init(&memcg->vmpressure);
6301 __mem_cgroup_free(memcg);
6302 return ERR_PTR(error);
6306 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6308 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6309 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6315 mutex_lock(&memcg_create_mutex);
6317 memcg->use_hierarchy = parent->use_hierarchy;
6318 memcg->oom_kill_disable = parent->oom_kill_disable;
6319 memcg->swappiness = mem_cgroup_swappiness(parent);
6321 if (parent->use_hierarchy) {
6322 res_counter_init(&memcg->res, &parent->res);
6323 res_counter_init(&memcg->memsw, &parent->memsw);
6324 res_counter_init(&memcg->kmem, &parent->kmem);
6327 * No need to take a reference to the parent because cgroup
6328 * core guarantees its existence.
6331 res_counter_init(&memcg->res, NULL);
6332 res_counter_init(&memcg->memsw, NULL);
6333 res_counter_init(&memcg->kmem, NULL);
6335 * Deeper hierachy with use_hierarchy == false doesn't make
6336 * much sense so let cgroup subsystem know about this
6337 * unfortunate state in our controller.
6339 if (parent != root_mem_cgroup)
6340 mem_cgroup_subsys.broken_hierarchy = true;
6343 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6344 mutex_unlock(&memcg_create_mutex);
6349 * Announce all parents that a group from their hierarchy is gone.
6351 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6353 struct mem_cgroup *parent = memcg;
6355 while ((parent = parent_mem_cgroup(parent)))
6356 mem_cgroup_iter_invalidate(parent);
6359 * if the root memcg is not hierarchical we have to check it
6362 if (!root_mem_cgroup->use_hierarchy)
6363 mem_cgroup_iter_invalidate(root_mem_cgroup);
6366 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6368 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6370 kmem_cgroup_css_offline(memcg);
6372 mem_cgroup_invalidate_reclaim_iterators(memcg);
6373 mem_cgroup_reparent_charges(memcg);
6374 mem_cgroup_destroy_all_caches(memcg);
6375 vmpressure_cleanup(&memcg->vmpressure);
6378 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6380 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6382 memcg_destroy_kmem(memcg);
6383 __mem_cgroup_free(memcg);
6387 /* Handlers for move charge at task migration. */
6388 #define PRECHARGE_COUNT_AT_ONCE 256
6389 static int mem_cgroup_do_precharge(unsigned long count)
6392 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6393 struct mem_cgroup *memcg = mc.to;
6395 if (mem_cgroup_is_root(memcg)) {
6396 mc.precharge += count;
6397 /* we don't need css_get for root */
6400 /* try to charge at once */
6402 struct res_counter *dummy;
6404 * "memcg" cannot be under rmdir() because we've already checked
6405 * by cgroup_lock_live_cgroup() that it is not removed and we
6406 * are still under the same cgroup_mutex. So we can postpone
6409 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6411 if (do_swap_account && res_counter_charge(&memcg->memsw,
6412 PAGE_SIZE * count, &dummy)) {
6413 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6416 mc.precharge += count;
6420 /* fall back to one by one charge */
6422 if (signal_pending(current)) {
6426 if (!batch_count--) {
6427 batch_count = PRECHARGE_COUNT_AT_ONCE;
6430 ret = __mem_cgroup_try_charge(NULL,
6431 GFP_KERNEL, 1, &memcg, false);
6433 /* mem_cgroup_clear_mc() will do uncharge later */
6441 * get_mctgt_type - get target type of moving charge
6442 * @vma: the vma the pte to be checked belongs
6443 * @addr: the address corresponding to the pte to be checked
6444 * @ptent: the pte to be checked
6445 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6448 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6449 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6450 * move charge. if @target is not NULL, the page is stored in target->page
6451 * with extra refcnt got(Callers should handle it).
6452 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6453 * target for charge migration. if @target is not NULL, the entry is stored
6456 * Called with pte lock held.
6463 enum mc_target_type {
6469 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6470 unsigned long addr, pte_t ptent)
6472 struct page *page = vm_normal_page(vma, addr, ptent);
6474 if (!page || !page_mapped(page))
6476 if (PageAnon(page)) {
6477 /* we don't move shared anon */
6480 } else if (!move_file())
6481 /* we ignore mapcount for file pages */
6483 if (!get_page_unless_zero(page))
6490 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6491 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6493 struct page *page = NULL;
6494 swp_entry_t ent = pte_to_swp_entry(ptent);
6496 if (!move_anon() || non_swap_entry(ent))
6499 * Because lookup_swap_cache() updates some statistics counter,
6500 * we call find_get_page() with swapper_space directly.
6502 page = find_get_page(swap_address_space(ent), ent.val);
6503 if (do_swap_account)
6504 entry->val = ent.val;
6509 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6510 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6516 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6517 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6519 struct page *page = NULL;
6520 struct address_space *mapping;
6523 if (!vma->vm_file) /* anonymous vma */
6528 mapping = vma->vm_file->f_mapping;
6529 if (pte_none(ptent))
6530 pgoff = linear_page_index(vma, addr);
6531 else /* pte_file(ptent) is true */
6532 pgoff = pte_to_pgoff(ptent);
6534 /* page is moved even if it's not RSS of this task(page-faulted). */
6535 page = find_get_page(mapping, pgoff);
6538 /* shmem/tmpfs may report page out on swap: account for that too. */
6539 if (radix_tree_exceptional_entry(page)) {
6540 swp_entry_t swap = radix_to_swp_entry(page);
6541 if (do_swap_account)
6543 page = find_get_page(swap_address_space(swap), swap.val);
6549 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6550 unsigned long addr, pte_t ptent, union mc_target *target)
6552 struct page *page = NULL;
6553 struct page_cgroup *pc;
6554 enum mc_target_type ret = MC_TARGET_NONE;
6555 swp_entry_t ent = { .val = 0 };
6557 if (pte_present(ptent))
6558 page = mc_handle_present_pte(vma, addr, ptent);
6559 else if (is_swap_pte(ptent))
6560 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6561 else if (pte_none(ptent) || pte_file(ptent))
6562 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6564 if (!page && !ent.val)
6567 pc = lookup_page_cgroup(page);
6569 * Do only loose check w/o page_cgroup lock.
6570 * mem_cgroup_move_account() checks the pc is valid or not under
6573 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6574 ret = MC_TARGET_PAGE;
6576 target->page = page;
6578 if (!ret || !target)
6581 /* There is a swap entry and a page doesn't exist or isn't charged */
6582 if (ent.val && !ret &&
6583 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6584 ret = MC_TARGET_SWAP;
6591 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6593 * We don't consider swapping or file mapped pages because THP does not
6594 * support them for now.
6595 * Caller should make sure that pmd_trans_huge(pmd) is true.
6597 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6598 unsigned long addr, pmd_t pmd, union mc_target *target)
6600 struct page *page = NULL;
6601 struct page_cgroup *pc;
6602 enum mc_target_type ret = MC_TARGET_NONE;
6604 page = pmd_page(pmd);
6605 VM_BUG_ON(!page || !PageHead(page));
6608 pc = lookup_page_cgroup(page);
6609 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6610 ret = MC_TARGET_PAGE;
6613 target->page = page;
6619 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6620 unsigned long addr, pmd_t pmd, union mc_target *target)
6622 return MC_TARGET_NONE;
6626 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6627 unsigned long addr, unsigned long end,
6628 struct mm_walk *walk)
6630 struct vm_area_struct *vma = walk->private;
6634 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6635 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6636 mc.precharge += HPAGE_PMD_NR;
6637 spin_unlock(&vma->vm_mm->page_table_lock);
6641 if (pmd_trans_unstable(pmd))
6643 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6644 for (; addr != end; pte++, addr += PAGE_SIZE)
6645 if (get_mctgt_type(vma, addr, *pte, NULL))
6646 mc.precharge++; /* increment precharge temporarily */
6647 pte_unmap_unlock(pte - 1, ptl);
6653 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6655 unsigned long precharge;
6656 struct vm_area_struct *vma;
6658 down_read(&mm->mmap_sem);
6659 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6660 struct mm_walk mem_cgroup_count_precharge_walk = {
6661 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6665 if (is_vm_hugetlb_page(vma))
6667 walk_page_range(vma->vm_start, vma->vm_end,
6668 &mem_cgroup_count_precharge_walk);
6670 up_read(&mm->mmap_sem);
6672 precharge = mc.precharge;
6678 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6680 unsigned long precharge = mem_cgroup_count_precharge(mm);
6682 VM_BUG_ON(mc.moving_task);
6683 mc.moving_task = current;
6684 return mem_cgroup_do_precharge(precharge);
6687 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6688 static void __mem_cgroup_clear_mc(void)
6690 struct mem_cgroup *from = mc.from;
6691 struct mem_cgroup *to = mc.to;
6694 /* we must uncharge all the leftover precharges from mc.to */
6696 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6700 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6701 * we must uncharge here.
6703 if (mc.moved_charge) {
6704 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6705 mc.moved_charge = 0;
6707 /* we must fixup refcnts and charges */
6708 if (mc.moved_swap) {
6709 /* uncharge swap account from the old cgroup */
6710 if (!mem_cgroup_is_root(mc.from))
6711 res_counter_uncharge(&mc.from->memsw,
6712 PAGE_SIZE * mc.moved_swap);
6714 for (i = 0; i < mc.moved_swap; i++)
6715 css_put(&mc.from->css);
6717 if (!mem_cgroup_is_root(mc.to)) {
6719 * we charged both to->res and to->memsw, so we should
6722 res_counter_uncharge(&mc.to->res,
6723 PAGE_SIZE * mc.moved_swap);
6725 /* we've already done css_get(mc.to) */
6728 memcg_oom_recover(from);
6729 memcg_oom_recover(to);
6730 wake_up_all(&mc.waitq);
6733 static void mem_cgroup_clear_mc(void)
6735 struct mem_cgroup *from = mc.from;
6738 * we must clear moving_task before waking up waiters at the end of
6741 mc.moving_task = NULL;
6742 __mem_cgroup_clear_mc();
6743 spin_lock(&mc.lock);
6746 spin_unlock(&mc.lock);
6747 mem_cgroup_end_move(from);
6750 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6751 struct cgroup_taskset *tset)
6753 struct task_struct *p = cgroup_taskset_first(tset);
6755 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6756 unsigned long move_charge_at_immigrate;
6759 * We are now commited to this value whatever it is. Changes in this
6760 * tunable will only affect upcoming migrations, not the current one.
6761 * So we need to save it, and keep it going.
6763 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6764 if (move_charge_at_immigrate) {
6765 struct mm_struct *mm;
6766 struct mem_cgroup *from = mem_cgroup_from_task(p);
6768 VM_BUG_ON(from == memcg);
6770 mm = get_task_mm(p);
6773 /* We move charges only when we move a owner of the mm */
6774 if (mm->owner == p) {
6777 VM_BUG_ON(mc.precharge);
6778 VM_BUG_ON(mc.moved_charge);
6779 VM_BUG_ON(mc.moved_swap);
6780 mem_cgroup_start_move(from);
6781 spin_lock(&mc.lock);
6784 mc.immigrate_flags = move_charge_at_immigrate;
6785 spin_unlock(&mc.lock);
6786 /* We set mc.moving_task later */
6788 ret = mem_cgroup_precharge_mc(mm);
6790 mem_cgroup_clear_mc();
6797 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6798 struct cgroup_taskset *tset)
6800 mem_cgroup_clear_mc();
6803 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6804 unsigned long addr, unsigned long end,
6805 struct mm_walk *walk)
6808 struct vm_area_struct *vma = walk->private;
6811 enum mc_target_type target_type;
6812 union mc_target target;
6814 struct page_cgroup *pc;
6817 * We don't take compound_lock() here but no race with splitting thp
6819 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6820 * under splitting, which means there's no concurrent thp split,
6821 * - if another thread runs into split_huge_page() just after we
6822 * entered this if-block, the thread must wait for page table lock
6823 * to be unlocked in __split_huge_page_splitting(), where the main
6824 * part of thp split is not executed yet.
6826 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6827 if (mc.precharge < HPAGE_PMD_NR) {
6828 spin_unlock(&vma->vm_mm->page_table_lock);
6831 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6832 if (target_type == MC_TARGET_PAGE) {
6834 if (!isolate_lru_page(page)) {
6835 pc = lookup_page_cgroup(page);
6836 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6837 pc, mc.from, mc.to)) {
6838 mc.precharge -= HPAGE_PMD_NR;
6839 mc.moved_charge += HPAGE_PMD_NR;
6841 putback_lru_page(page);
6845 spin_unlock(&vma->vm_mm->page_table_lock);
6849 if (pmd_trans_unstable(pmd))
6852 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6853 for (; addr != end; addr += PAGE_SIZE) {
6854 pte_t ptent = *(pte++);
6860 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6861 case MC_TARGET_PAGE:
6863 if (isolate_lru_page(page))
6865 pc = lookup_page_cgroup(page);
6866 if (!mem_cgroup_move_account(page, 1, pc,
6869 /* we uncharge from mc.from later. */
6872 putback_lru_page(page);
6873 put: /* get_mctgt_type() gets the page */
6876 case MC_TARGET_SWAP:
6878 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6880 /* we fixup refcnts and charges later. */
6888 pte_unmap_unlock(pte - 1, ptl);
6893 * We have consumed all precharges we got in can_attach().
6894 * We try charge one by one, but don't do any additional
6895 * charges to mc.to if we have failed in charge once in attach()
6898 ret = mem_cgroup_do_precharge(1);
6906 static void mem_cgroup_move_charge(struct mm_struct *mm)
6908 struct vm_area_struct *vma;
6910 lru_add_drain_all();
6912 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6914 * Someone who are holding the mmap_sem might be waiting in
6915 * waitq. So we cancel all extra charges, wake up all waiters,
6916 * and retry. Because we cancel precharges, we might not be able
6917 * to move enough charges, but moving charge is a best-effort
6918 * feature anyway, so it wouldn't be a big problem.
6920 __mem_cgroup_clear_mc();
6924 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6926 struct mm_walk mem_cgroup_move_charge_walk = {
6927 .pmd_entry = mem_cgroup_move_charge_pte_range,
6931 if (is_vm_hugetlb_page(vma))
6933 ret = walk_page_range(vma->vm_start, vma->vm_end,
6934 &mem_cgroup_move_charge_walk);
6937 * means we have consumed all precharges and failed in
6938 * doing additional charge. Just abandon here.
6942 up_read(&mm->mmap_sem);
6945 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6946 struct cgroup_taskset *tset)
6948 struct task_struct *p = cgroup_taskset_first(tset);
6949 struct mm_struct *mm = get_task_mm(p);
6953 mem_cgroup_move_charge(mm);
6957 mem_cgroup_clear_mc();
6959 #else /* !CONFIG_MMU */
6960 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6961 struct cgroup_taskset *tset)
6965 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6966 struct cgroup_taskset *tset)
6969 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6970 struct cgroup_taskset *tset)
6976 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6977 * to verify sane_behavior flag on each mount attempt.
6979 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6982 * use_hierarchy is forced with sane_behavior. cgroup core
6983 * guarantees that @root doesn't have any children, so turning it
6984 * on for the root memcg is enough.
6986 if (cgroup_sane_behavior(root_css->cgroup))
6987 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6990 struct cgroup_subsys mem_cgroup_subsys = {
6992 .subsys_id = mem_cgroup_subsys_id,
6993 .css_alloc = mem_cgroup_css_alloc,
6994 .css_online = mem_cgroup_css_online,
6995 .css_offline = mem_cgroup_css_offline,
6996 .css_free = mem_cgroup_css_free,
6997 .can_attach = mem_cgroup_can_attach,
6998 .cancel_attach = mem_cgroup_cancel_attach,
6999 .attach = mem_cgroup_move_task,
7000 .bind = mem_cgroup_bind,
7001 .base_cftypes = mem_cgroup_files,
7006 #ifdef CONFIG_MEMCG_SWAP
7007 static int __init enable_swap_account(char *s)
7009 if (!strcmp(s, "1"))
7010 really_do_swap_account = 1;
7011 else if (!strcmp(s, "0"))
7012 really_do_swap_account = 0;
7015 __setup("swapaccount=", enable_swap_account);
7017 static void __init memsw_file_init(void)
7019 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
7022 static void __init enable_swap_cgroup(void)
7024 if (!mem_cgroup_disabled() && really_do_swap_account) {
7025 do_swap_account = 1;
7031 static void __init enable_swap_cgroup(void)
7037 * subsys_initcall() for memory controller.
7039 * Some parts like hotcpu_notifier() have to be initialized from this context
7040 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
7041 * everything that doesn't depend on a specific mem_cgroup structure should
7042 * be initialized from here.
7044 static int __init mem_cgroup_init(void)
7046 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
7047 enable_swap_cgroup();
7048 mem_cgroup_soft_limit_tree_init();
7052 subsys_initcall(mem_cgroup_init);