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;
869 for_each_online_cpu(cpu)
870 val += per_cpu(memcg->stat->events[idx], cpu);
871 #ifdef CONFIG_HOTPLUG_CPU
872 spin_lock(&memcg->pcp_counter_lock);
873 val += memcg->nocpu_base.events[idx];
874 spin_unlock(&memcg->pcp_counter_lock);
879 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
881 bool anon, int nr_pages)
886 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
887 * counted as CACHE even if it's on ANON LRU.
890 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
893 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
896 if (PageTransHuge(page))
897 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
900 /* pagein of a big page is an event. So, ignore page size */
902 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
904 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
905 nr_pages = -nr_pages; /* for event */
908 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
914 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
916 struct mem_cgroup_per_zone *mz;
918 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
919 return mz->lru_size[lru];
923 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
924 unsigned int lru_mask)
926 struct mem_cgroup_per_zone *mz;
928 unsigned long ret = 0;
930 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
933 if (BIT(lru) & lru_mask)
934 ret += mz->lru_size[lru];
940 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
941 int nid, unsigned int lru_mask)
946 for (zid = 0; zid < MAX_NR_ZONES; zid++)
947 total += mem_cgroup_zone_nr_lru_pages(memcg,
953 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
954 unsigned int lru_mask)
959 for_each_node_state(nid, N_MEMORY)
960 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
964 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
965 enum mem_cgroup_events_target target)
967 unsigned long val, next;
969 val = __this_cpu_read(memcg->stat->nr_page_events);
970 next = __this_cpu_read(memcg->stat->targets[target]);
971 /* from time_after() in jiffies.h */
972 if ((long)next - (long)val < 0) {
974 case MEM_CGROUP_TARGET_THRESH:
975 next = val + THRESHOLDS_EVENTS_TARGET;
977 case MEM_CGROUP_TARGET_SOFTLIMIT:
978 next = val + SOFTLIMIT_EVENTS_TARGET;
980 case MEM_CGROUP_TARGET_NUMAINFO:
981 next = val + NUMAINFO_EVENTS_TARGET;
986 __this_cpu_write(memcg->stat->targets[target], next);
993 * Check events in order.
996 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
999 /* threshold event is triggered in finer grain than soft limit */
1000 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1001 MEM_CGROUP_TARGET_THRESH))) {
1003 bool do_numainfo __maybe_unused;
1005 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1006 MEM_CGROUP_TARGET_SOFTLIMIT);
1007 #if MAX_NUMNODES > 1
1008 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1009 MEM_CGROUP_TARGET_NUMAINFO);
1013 mem_cgroup_threshold(memcg);
1014 if (unlikely(do_softlimit))
1015 mem_cgroup_update_tree(memcg, page);
1016 #if MAX_NUMNODES > 1
1017 if (unlikely(do_numainfo))
1018 atomic_inc(&memcg->numainfo_events);
1024 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1027 * mm_update_next_owner() may clear mm->owner to NULL
1028 * if it races with swapoff, page migration, etc.
1029 * So this can be called with p == NULL.
1034 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1037 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1039 struct mem_cgroup *memcg = NULL;
1044 * Because we have no locks, mm->owner's may be being moved to other
1045 * cgroup. We use css_tryget() here even if this looks
1046 * pessimistic (rather than adding locks here).
1050 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1051 if (unlikely(!memcg))
1053 } while (!css_tryget(&memcg->css));
1059 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1060 * ref. count) or NULL if the whole root's subtree has been visited.
1062 * helper function to be used by mem_cgroup_iter
1064 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1065 struct mem_cgroup *last_visited)
1067 struct cgroup_subsys_state *prev_css, *next_css;
1069 prev_css = last_visited ? &last_visited->css : NULL;
1071 next_css = css_next_descendant_pre(prev_css, &root->css);
1074 * Even if we found a group we have to make sure it is
1075 * alive. css && !memcg means that the groups should be
1076 * skipped and we should continue the tree walk.
1077 * last_visited css is safe to use because it is
1078 * protected by css_get and the tree walk is rcu safe.
1081 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
1083 if (css_tryget(&mem->css))
1086 prev_css = next_css;
1094 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1097 * When a group in the hierarchy below root is destroyed, the
1098 * hierarchy iterator can no longer be trusted since it might
1099 * have pointed to the destroyed group. Invalidate it.
1101 atomic_inc(&root->dead_count);
1104 static struct mem_cgroup *
1105 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1106 struct mem_cgroup *root,
1109 struct mem_cgroup *position = NULL;
1111 * A cgroup destruction happens in two stages: offlining and
1112 * release. They are separated by a RCU grace period.
1114 * If the iterator is valid, we may still race with an
1115 * offlining. The RCU lock ensures the object won't be
1116 * released, tryget will fail if we lost the race.
1118 *sequence = atomic_read(&root->dead_count);
1119 if (iter->last_dead_count == *sequence) {
1121 position = iter->last_visited;
1122 if (position && !css_tryget(&position->css))
1128 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1129 struct mem_cgroup *last_visited,
1130 struct mem_cgroup *new_position,
1134 css_put(&last_visited->css);
1136 * We store the sequence count from the time @last_visited was
1137 * loaded successfully instead of rereading it here so that we
1138 * don't lose destruction events in between. We could have
1139 * raced with the destruction of @new_position after all.
1141 iter->last_visited = new_position;
1143 iter->last_dead_count = sequence;
1147 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1148 * @root: hierarchy root
1149 * @prev: previously returned memcg, NULL on first invocation
1150 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1152 * Returns references to children of the hierarchy below @root, or
1153 * @root itself, or %NULL after a full round-trip.
1155 * Caller must pass the return value in @prev on subsequent
1156 * invocations for reference counting, or use mem_cgroup_iter_break()
1157 * to cancel a hierarchy walk before the round-trip is complete.
1159 * Reclaimers can specify a zone and a priority level in @reclaim to
1160 * divide up the memcgs in the hierarchy among all concurrent
1161 * reclaimers operating on the same zone and priority.
1163 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1164 struct mem_cgroup *prev,
1165 struct mem_cgroup_reclaim_cookie *reclaim)
1167 struct mem_cgroup *memcg = NULL;
1168 struct mem_cgroup *last_visited = NULL;
1170 if (mem_cgroup_disabled())
1174 root = root_mem_cgroup;
1176 if (prev && !reclaim)
1177 last_visited = prev;
1179 if (!root->use_hierarchy && root != root_mem_cgroup) {
1187 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1188 int uninitialized_var(seq);
1191 int nid = zone_to_nid(reclaim->zone);
1192 int zid = zone_idx(reclaim->zone);
1193 struct mem_cgroup_per_zone *mz;
1195 mz = mem_cgroup_zoneinfo(root, nid, zid);
1196 iter = &mz->reclaim_iter[reclaim->priority];
1197 if (prev && reclaim->generation != iter->generation) {
1198 iter->last_visited = NULL;
1202 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1205 memcg = __mem_cgroup_iter_next(root, last_visited);
1208 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1212 else if (!prev && memcg)
1213 reclaim->generation = iter->generation;
1222 if (prev && prev != root)
1223 css_put(&prev->css);
1229 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1230 * @root: hierarchy root
1231 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1233 void mem_cgroup_iter_break(struct mem_cgroup *root,
1234 struct mem_cgroup *prev)
1237 root = root_mem_cgroup;
1238 if (prev && prev != root)
1239 css_put(&prev->css);
1243 * Iteration constructs for visiting all cgroups (under a tree). If
1244 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1245 * be used for reference counting.
1247 #define for_each_mem_cgroup_tree(iter, root) \
1248 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1250 iter = mem_cgroup_iter(root, iter, NULL))
1252 #define for_each_mem_cgroup(iter) \
1253 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1255 iter = mem_cgroup_iter(NULL, iter, NULL))
1257 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1259 struct mem_cgroup *memcg;
1262 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1263 if (unlikely(!memcg))
1268 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1271 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1279 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1282 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1283 * @zone: zone of the wanted lruvec
1284 * @memcg: memcg of the wanted lruvec
1286 * Returns the lru list vector holding pages for the given @zone and
1287 * @mem. This can be the global zone lruvec, if the memory controller
1290 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1291 struct mem_cgroup *memcg)
1293 struct mem_cgroup_per_zone *mz;
1294 struct lruvec *lruvec;
1296 if (mem_cgroup_disabled()) {
1297 lruvec = &zone->lruvec;
1301 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1302 lruvec = &mz->lruvec;
1305 * Since a node can be onlined after the mem_cgroup was created,
1306 * we have to be prepared to initialize lruvec->zone here;
1307 * and if offlined then reonlined, we need to reinitialize it.
1309 if (unlikely(lruvec->zone != zone))
1310 lruvec->zone = zone;
1315 * Following LRU functions are allowed to be used without PCG_LOCK.
1316 * Operations are called by routine of global LRU independently from memcg.
1317 * What we have to take care of here is validness of pc->mem_cgroup.
1319 * Changes to pc->mem_cgroup happens when
1322 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1323 * It is added to LRU before charge.
1324 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1325 * When moving account, the page is not on LRU. It's isolated.
1329 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1331 * @zone: zone of the page
1333 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1335 struct mem_cgroup_per_zone *mz;
1336 struct mem_cgroup *memcg;
1337 struct page_cgroup *pc;
1338 struct lruvec *lruvec;
1340 if (mem_cgroup_disabled()) {
1341 lruvec = &zone->lruvec;
1345 pc = lookup_page_cgroup(page);
1346 memcg = pc->mem_cgroup;
1349 * Surreptitiously switch any uncharged offlist page to root:
1350 * an uncharged page off lru does nothing to secure
1351 * its former mem_cgroup from sudden removal.
1353 * Our caller holds lru_lock, and PageCgroupUsed is updated
1354 * under page_cgroup lock: between them, they make all uses
1355 * of pc->mem_cgroup safe.
1357 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1358 pc->mem_cgroup = memcg = root_mem_cgroup;
1360 mz = page_cgroup_zoneinfo(memcg, page);
1361 lruvec = &mz->lruvec;
1364 * Since a node can be onlined after the mem_cgroup was created,
1365 * we have to be prepared to initialize lruvec->zone here;
1366 * and if offlined then reonlined, we need to reinitialize it.
1368 if (unlikely(lruvec->zone != zone))
1369 lruvec->zone = zone;
1374 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1375 * @lruvec: mem_cgroup per zone lru vector
1376 * @lru: index of lru list the page is sitting on
1377 * @nr_pages: positive when adding or negative when removing
1379 * This function must be called when a page is added to or removed from an
1382 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1385 struct mem_cgroup_per_zone *mz;
1386 unsigned long *lru_size;
1388 if (mem_cgroup_disabled())
1391 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1392 lru_size = mz->lru_size + lru;
1393 *lru_size += nr_pages;
1394 VM_BUG_ON((long)(*lru_size) < 0);
1398 * Checks whether given mem is same or in the root_mem_cgroup's
1401 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1402 struct mem_cgroup *memcg)
1404 if (root_memcg == memcg)
1406 if (!root_memcg->use_hierarchy || !memcg)
1408 return css_is_ancestor(&memcg->css, &root_memcg->css);
1411 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1412 struct mem_cgroup *memcg)
1417 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1422 bool task_in_mem_cgroup(struct task_struct *task,
1423 const struct mem_cgroup *memcg)
1425 struct mem_cgroup *curr = NULL;
1426 struct task_struct *p;
1429 p = find_lock_task_mm(task);
1431 curr = try_get_mem_cgroup_from_mm(p->mm);
1435 * All threads may have already detached their mm's, but the oom
1436 * killer still needs to detect if they have already been oom
1437 * killed to prevent needlessly killing additional tasks.
1440 curr = mem_cgroup_from_task(task);
1442 css_get(&curr->css);
1448 * We should check use_hierarchy of "memcg" not "curr". Because checking
1449 * use_hierarchy of "curr" here make this function true if hierarchy is
1450 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1451 * hierarchy(even if use_hierarchy is disabled in "memcg").
1453 ret = mem_cgroup_same_or_subtree(memcg, curr);
1454 css_put(&curr->css);
1458 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1460 unsigned long inactive_ratio;
1461 unsigned long inactive;
1462 unsigned long active;
1465 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1466 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1468 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1470 inactive_ratio = int_sqrt(10 * gb);
1474 return inactive * inactive_ratio < active;
1477 #define mem_cgroup_from_res_counter(counter, member) \
1478 container_of(counter, struct mem_cgroup, member)
1481 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1482 * @memcg: the memory cgroup
1484 * Returns the maximum amount of memory @mem can be charged with, in
1487 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1489 unsigned long long margin;
1491 margin = res_counter_margin(&memcg->res);
1492 if (do_swap_account)
1493 margin = min(margin, res_counter_margin(&memcg->memsw));
1494 return margin >> PAGE_SHIFT;
1497 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1500 if (!css_parent(&memcg->css))
1501 return vm_swappiness;
1503 return memcg->swappiness;
1507 * memcg->moving_account is used for checking possibility that some thread is
1508 * calling move_account(). When a thread on CPU-A starts moving pages under
1509 * a memcg, other threads should check memcg->moving_account under
1510 * rcu_read_lock(), like this:
1514 * memcg->moving_account+1 if (memcg->mocing_account)
1516 * synchronize_rcu() update something.
1521 /* for quick checking without looking up memcg */
1522 atomic_t memcg_moving __read_mostly;
1524 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1526 atomic_inc(&memcg_moving);
1527 atomic_inc(&memcg->moving_account);
1531 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1534 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1535 * We check NULL in callee rather than caller.
1538 atomic_dec(&memcg_moving);
1539 atomic_dec(&memcg->moving_account);
1544 * 2 routines for checking "mem" is under move_account() or not.
1546 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1547 * is used for avoiding races in accounting. If true,
1548 * pc->mem_cgroup may be overwritten.
1550 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1551 * under hierarchy of moving cgroups. This is for
1552 * waiting at hith-memory prressure caused by "move".
1555 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1557 VM_BUG_ON(!rcu_read_lock_held());
1558 return atomic_read(&memcg->moving_account) > 0;
1561 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1563 struct mem_cgroup *from;
1564 struct mem_cgroup *to;
1567 * Unlike task_move routines, we access mc.to, mc.from not under
1568 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1570 spin_lock(&mc.lock);
1576 ret = mem_cgroup_same_or_subtree(memcg, from)
1577 || mem_cgroup_same_or_subtree(memcg, to);
1579 spin_unlock(&mc.lock);
1583 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1585 if (mc.moving_task && current != mc.moving_task) {
1586 if (mem_cgroup_under_move(memcg)) {
1588 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1589 /* moving charge context might have finished. */
1592 finish_wait(&mc.waitq, &wait);
1600 * Take this lock when
1601 * - a code tries to modify page's memcg while it's USED.
1602 * - a code tries to modify page state accounting in a memcg.
1603 * see mem_cgroup_stolen(), too.
1605 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1606 unsigned long *flags)
1608 spin_lock_irqsave(&memcg->move_lock, *flags);
1611 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1612 unsigned long *flags)
1614 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1617 #define K(x) ((x) << (PAGE_SHIFT-10))
1619 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1620 * @memcg: The memory cgroup that went over limit
1621 * @p: Task that is going to be killed
1623 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1626 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1628 struct cgroup *task_cgrp;
1629 struct cgroup *mem_cgrp;
1631 * Need a buffer in BSS, can't rely on allocations. The code relies
1632 * on the assumption that OOM is serialized for memory controller.
1633 * If this assumption is broken, revisit this code.
1635 static char memcg_name[PATH_MAX];
1637 struct mem_cgroup *iter;
1645 mem_cgrp = memcg->css.cgroup;
1646 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1648 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1651 * Unfortunately, we are unable to convert to a useful name
1652 * But we'll still print out the usage information
1659 pr_info("Task in %s killed", memcg_name);
1662 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1670 * Continues from above, so we don't need an KERN_ level
1672 pr_cont(" as a result of limit of %s\n", memcg_name);
1675 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1676 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1677 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1678 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1679 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1680 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1681 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1682 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1683 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1684 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1685 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1686 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1688 for_each_mem_cgroup_tree(iter, memcg) {
1689 pr_info("Memory cgroup stats");
1692 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1694 pr_cont(" for %s", memcg_name);
1698 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1699 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1701 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1702 K(mem_cgroup_read_stat(iter, i)));
1705 for (i = 0; i < NR_LRU_LISTS; i++)
1706 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1707 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1714 * This function returns the number of memcg under hierarchy tree. Returns
1715 * 1(self count) if no children.
1717 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1720 struct mem_cgroup *iter;
1722 for_each_mem_cgroup_tree(iter, memcg)
1728 * Return the memory (and swap, if configured) limit for a memcg.
1730 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1734 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1737 * Do not consider swap space if we cannot swap due to swappiness
1739 if (mem_cgroup_swappiness(memcg)) {
1742 limit += total_swap_pages << PAGE_SHIFT;
1743 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1746 * If memsw is finite and limits the amount of swap space
1747 * available to this memcg, return that limit.
1749 limit = min(limit, memsw);
1755 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1758 struct mem_cgroup *iter;
1759 unsigned long chosen_points = 0;
1760 unsigned long totalpages;
1761 unsigned int points = 0;
1762 struct task_struct *chosen = NULL;
1765 * If current has a pending SIGKILL or is exiting, then automatically
1766 * select it. The goal is to allow it to allocate so that it may
1767 * quickly exit and free its memory.
1769 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1770 set_thread_flag(TIF_MEMDIE);
1774 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1775 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1776 for_each_mem_cgroup_tree(iter, memcg) {
1777 struct css_task_iter it;
1778 struct task_struct *task;
1780 css_task_iter_start(&iter->css, &it);
1781 while ((task = css_task_iter_next(&it))) {
1782 switch (oom_scan_process_thread(task, totalpages, NULL,
1784 case OOM_SCAN_SELECT:
1786 put_task_struct(chosen);
1788 chosen_points = ULONG_MAX;
1789 get_task_struct(chosen);
1791 case OOM_SCAN_CONTINUE:
1793 case OOM_SCAN_ABORT:
1794 css_task_iter_end(&it);
1795 mem_cgroup_iter_break(memcg, iter);
1797 put_task_struct(chosen);
1802 points = oom_badness(task, memcg, NULL, totalpages);
1803 if (points > chosen_points) {
1805 put_task_struct(chosen);
1807 chosen_points = points;
1808 get_task_struct(chosen);
1811 css_task_iter_end(&it);
1816 points = chosen_points * 1000 / totalpages;
1817 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1818 NULL, "Memory cgroup out of memory");
1821 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1823 unsigned long flags)
1825 unsigned long total = 0;
1826 bool noswap = false;
1829 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1831 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1834 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1836 drain_all_stock_async(memcg);
1837 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1839 * Allow limit shrinkers, which are triggered directly
1840 * by userspace, to catch signals and stop reclaim
1841 * after minimal progress, regardless of the margin.
1843 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1845 if (mem_cgroup_margin(memcg))
1848 * If nothing was reclaimed after two attempts, there
1849 * may be no reclaimable pages in this hierarchy.
1858 * test_mem_cgroup_node_reclaimable
1859 * @memcg: the target memcg
1860 * @nid: the node ID to be checked.
1861 * @noswap : specify true here if the user wants flle only information.
1863 * This function returns whether the specified memcg contains any
1864 * reclaimable pages on a node. Returns true if there are any reclaimable
1865 * pages in the node.
1867 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1868 int nid, bool noswap)
1870 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1872 if (noswap || !total_swap_pages)
1874 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1879 #if MAX_NUMNODES > 1
1882 * Always updating the nodemask is not very good - even if we have an empty
1883 * list or the wrong list here, we can start from some node and traverse all
1884 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1887 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1891 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1892 * pagein/pageout changes since the last update.
1894 if (!atomic_read(&memcg->numainfo_events))
1896 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1899 /* make a nodemask where this memcg uses memory from */
1900 memcg->scan_nodes = node_states[N_MEMORY];
1902 for_each_node_mask(nid, node_states[N_MEMORY]) {
1904 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1905 node_clear(nid, memcg->scan_nodes);
1908 atomic_set(&memcg->numainfo_events, 0);
1909 atomic_set(&memcg->numainfo_updating, 0);
1913 * Selecting a node where we start reclaim from. Because what we need is just
1914 * reducing usage counter, start from anywhere is O,K. Considering
1915 * memory reclaim from current node, there are pros. and cons.
1917 * Freeing memory from current node means freeing memory from a node which
1918 * we'll use or we've used. So, it may make LRU bad. And if several threads
1919 * hit limits, it will see a contention on a node. But freeing from remote
1920 * node means more costs for memory reclaim because of memory latency.
1922 * Now, we use round-robin. Better algorithm is welcomed.
1924 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1928 mem_cgroup_may_update_nodemask(memcg);
1929 node = memcg->last_scanned_node;
1931 node = next_node(node, memcg->scan_nodes);
1932 if (node == MAX_NUMNODES)
1933 node = first_node(memcg->scan_nodes);
1935 * We call this when we hit limit, not when pages are added to LRU.
1936 * No LRU may hold pages because all pages are UNEVICTABLE or
1937 * memcg is too small and all pages are not on LRU. In that case,
1938 * we use curret node.
1940 if (unlikely(node == MAX_NUMNODES))
1941 node = numa_node_id();
1943 memcg->last_scanned_node = node;
1948 * Check all nodes whether it contains reclaimable pages or not.
1949 * For quick scan, we make use of scan_nodes. This will allow us to skip
1950 * unused nodes. But scan_nodes is lazily updated and may not cotain
1951 * enough new information. We need to do double check.
1953 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1958 * quick check...making use of scan_node.
1959 * We can skip unused nodes.
1961 if (!nodes_empty(memcg->scan_nodes)) {
1962 for (nid = first_node(memcg->scan_nodes);
1964 nid = next_node(nid, memcg->scan_nodes)) {
1966 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1971 * Check rest of nodes.
1973 for_each_node_state(nid, N_MEMORY) {
1974 if (node_isset(nid, memcg->scan_nodes))
1976 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1983 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1988 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1990 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1995 * A group is eligible for the soft limit reclaim if
1996 * a) it is over its soft limit
1997 * b) any parent up the hierarchy is over its soft limit
1999 bool mem_cgroup_soft_reclaim_eligible(struct mem_cgroup *memcg)
2001 struct mem_cgroup *parent = memcg;
2003 if (res_counter_soft_limit_excess(&memcg->res))
2007 * If any parent up the hierarchy is over its soft limit then we
2008 * have to obey and reclaim from this group as well.
2010 while ((parent = parent_mem_cgroup(parent))) {
2011 if (res_counter_soft_limit_excess(&parent->res))
2018 static DEFINE_SPINLOCK(memcg_oom_lock);
2021 * Check OOM-Killer is already running under our hierarchy.
2022 * If someone is running, return false.
2024 static bool mem_cgroup_oom_trylock(struct mem_cgroup *memcg)
2026 struct mem_cgroup *iter, *failed = NULL;
2028 spin_lock(&memcg_oom_lock);
2030 for_each_mem_cgroup_tree(iter, memcg) {
2031 if (iter->oom_lock) {
2033 * this subtree of our hierarchy is already locked
2034 * so we cannot give a lock.
2037 mem_cgroup_iter_break(memcg, iter);
2040 iter->oom_lock = true;
2045 * OK, we failed to lock the whole subtree so we have
2046 * to clean up what we set up to the failing subtree
2048 for_each_mem_cgroup_tree(iter, memcg) {
2049 if (iter == failed) {
2050 mem_cgroup_iter_break(memcg, iter);
2053 iter->oom_lock = false;
2057 spin_unlock(&memcg_oom_lock);
2062 static void mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2064 struct mem_cgroup *iter;
2066 spin_lock(&memcg_oom_lock);
2067 for_each_mem_cgroup_tree(iter, memcg)
2068 iter->oom_lock = false;
2069 spin_unlock(&memcg_oom_lock);
2072 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2074 struct mem_cgroup *iter;
2076 for_each_mem_cgroup_tree(iter, memcg)
2077 atomic_inc(&iter->under_oom);
2080 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2082 struct mem_cgroup *iter;
2085 * When a new child is created while the hierarchy is under oom,
2086 * mem_cgroup_oom_lock() may not be called. We have to use
2087 * atomic_add_unless() here.
2089 for_each_mem_cgroup_tree(iter, memcg)
2090 atomic_add_unless(&iter->under_oom, -1, 0);
2093 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2095 struct oom_wait_info {
2096 struct mem_cgroup *memcg;
2100 static int memcg_oom_wake_function(wait_queue_t *wait,
2101 unsigned mode, int sync, void *arg)
2103 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2104 struct mem_cgroup *oom_wait_memcg;
2105 struct oom_wait_info *oom_wait_info;
2107 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2108 oom_wait_memcg = oom_wait_info->memcg;
2111 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2112 * Then we can use css_is_ancestor without taking care of RCU.
2114 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2115 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2117 return autoremove_wake_function(wait, mode, sync, arg);
2120 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2122 atomic_inc(&memcg->oom_wakeups);
2123 /* for filtering, pass "memcg" as argument. */
2124 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2127 static void memcg_oom_recover(struct mem_cgroup *memcg)
2129 if (memcg && atomic_read(&memcg->under_oom))
2130 memcg_wakeup_oom(memcg);
2134 * try to call OOM killer
2136 static void mem_cgroup_oom(struct mem_cgroup *memcg, gfp_t mask, int order)
2141 if (!current->memcg_oom.may_oom)
2144 current->memcg_oom.in_memcg_oom = 1;
2147 * As with any blocking lock, a contender needs to start
2148 * listening for wakeups before attempting the trylock,
2149 * otherwise it can miss the wakeup from the unlock and sleep
2150 * indefinitely. This is just open-coded because our locking
2151 * is so particular to memcg hierarchies.
2153 wakeups = atomic_read(&memcg->oom_wakeups);
2154 mem_cgroup_mark_under_oom(memcg);
2156 locked = mem_cgroup_oom_trylock(memcg);
2159 mem_cgroup_oom_notify(memcg);
2161 if (locked && !memcg->oom_kill_disable) {
2162 mem_cgroup_unmark_under_oom(memcg);
2163 mem_cgroup_out_of_memory(memcg, mask, order);
2164 mem_cgroup_oom_unlock(memcg);
2166 * There is no guarantee that an OOM-lock contender
2167 * sees the wakeups triggered by the OOM kill
2168 * uncharges. Wake any sleepers explicitely.
2170 memcg_oom_recover(memcg);
2173 * A system call can just return -ENOMEM, but if this
2174 * is a page fault and somebody else is handling the
2175 * OOM already, we need to sleep on the OOM waitqueue
2176 * for this memcg until the situation is resolved.
2177 * Which can take some time because it might be
2178 * handled by a userspace task.
2180 * However, this is the charge context, which means
2181 * that we may sit on a large call stack and hold
2182 * various filesystem locks, the mmap_sem etc. and we
2183 * don't want the OOM handler to deadlock on them
2184 * while we sit here and wait. Store the current OOM
2185 * context in the task_struct, then return -ENOMEM.
2186 * At the end of the page fault handler, with the
2187 * stack unwound, pagefault_out_of_memory() will check
2188 * back with us by calling
2189 * mem_cgroup_oom_synchronize(), possibly putting the
2192 current->memcg_oom.oom_locked = locked;
2193 current->memcg_oom.wakeups = wakeups;
2194 css_get(&memcg->css);
2195 current->memcg_oom.wait_on_memcg = memcg;
2200 * mem_cgroup_oom_synchronize - complete memcg OOM handling
2202 * This has to be called at the end of a page fault if the the memcg
2203 * OOM handler was enabled and the fault is returning %VM_FAULT_OOM.
2205 * Memcg supports userspace OOM handling, so failed allocations must
2206 * sleep on a waitqueue until the userspace task resolves the
2207 * situation. Sleeping directly in the charge context with all kinds
2208 * of locks held is not a good idea, instead we remember an OOM state
2209 * in the task and mem_cgroup_oom_synchronize() has to be called at
2210 * the end of the page fault to put the task to sleep and clean up the
2213 * Returns %true if an ongoing memcg OOM situation was detected and
2214 * finalized, %false otherwise.
2216 bool mem_cgroup_oom_synchronize(void)
2218 struct oom_wait_info owait;
2219 struct mem_cgroup *memcg;
2221 /* OOM is global, do not handle */
2222 if (!current->memcg_oom.in_memcg_oom)
2226 * We invoked the OOM killer but there is a chance that a kill
2227 * did not free up any charges. Everybody else might already
2228 * be sleeping, so restart the fault and keep the rampage
2229 * going until some charges are released.
2231 memcg = current->memcg_oom.wait_on_memcg;
2235 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2238 owait.memcg = memcg;
2239 owait.wait.flags = 0;
2240 owait.wait.func = memcg_oom_wake_function;
2241 owait.wait.private = current;
2242 INIT_LIST_HEAD(&owait.wait.task_list);
2244 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2245 /* Only sleep if we didn't miss any wakeups since OOM */
2246 if (atomic_read(&memcg->oom_wakeups) == current->memcg_oom.wakeups)
2248 finish_wait(&memcg_oom_waitq, &owait.wait);
2250 mem_cgroup_unmark_under_oom(memcg);
2251 if (current->memcg_oom.oom_locked) {
2252 mem_cgroup_oom_unlock(memcg);
2254 * There is no guarantee that an OOM-lock contender
2255 * sees the wakeups triggered by the OOM kill
2256 * uncharges. Wake any sleepers explicitely.
2258 memcg_oom_recover(memcg);
2260 css_put(&memcg->css);
2261 current->memcg_oom.wait_on_memcg = NULL;
2263 current->memcg_oom.in_memcg_oom = 0;
2268 * Currently used to update mapped file statistics, but the routine can be
2269 * generalized to update other statistics as well.
2271 * Notes: Race condition
2273 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2274 * it tends to be costly. But considering some conditions, we doesn't need
2275 * to do so _always_.
2277 * Considering "charge", lock_page_cgroup() is not required because all
2278 * file-stat operations happen after a page is attached to radix-tree. There
2279 * are no race with "charge".
2281 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2282 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2283 * if there are race with "uncharge". Statistics itself is properly handled
2286 * Considering "move", this is an only case we see a race. To make the race
2287 * small, we check mm->moving_account and detect there are possibility of race
2288 * If there is, we take a lock.
2291 void __mem_cgroup_begin_update_page_stat(struct page *page,
2292 bool *locked, unsigned long *flags)
2294 struct mem_cgroup *memcg;
2295 struct page_cgroup *pc;
2297 pc = lookup_page_cgroup(page);
2299 memcg = pc->mem_cgroup;
2300 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2303 * If this memory cgroup is not under account moving, we don't
2304 * need to take move_lock_mem_cgroup(). Because we already hold
2305 * rcu_read_lock(), any calls to move_account will be delayed until
2306 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2308 if (!mem_cgroup_stolen(memcg))
2311 move_lock_mem_cgroup(memcg, flags);
2312 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2313 move_unlock_mem_cgroup(memcg, flags);
2319 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2321 struct page_cgroup *pc = lookup_page_cgroup(page);
2324 * It's guaranteed that pc->mem_cgroup never changes while
2325 * lock is held because a routine modifies pc->mem_cgroup
2326 * should take move_lock_mem_cgroup().
2328 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2331 void mem_cgroup_update_page_stat(struct page *page,
2332 enum mem_cgroup_stat_index idx, int val)
2334 struct mem_cgroup *memcg;
2335 struct page_cgroup *pc = lookup_page_cgroup(page);
2336 unsigned long uninitialized_var(flags);
2338 if (mem_cgroup_disabled())
2341 VM_BUG_ON(!rcu_read_lock_held());
2342 memcg = pc->mem_cgroup;
2343 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2346 this_cpu_add(memcg->stat->count[idx], val);
2350 * size of first charge trial. "32" comes from vmscan.c's magic value.
2351 * TODO: maybe necessary to use big numbers in big irons.
2353 #define CHARGE_BATCH 32U
2354 struct memcg_stock_pcp {
2355 struct mem_cgroup *cached; /* this never be root cgroup */
2356 unsigned int nr_pages;
2357 struct work_struct work;
2358 unsigned long flags;
2359 #define FLUSHING_CACHED_CHARGE 0
2361 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2362 static DEFINE_MUTEX(percpu_charge_mutex);
2365 * consume_stock: Try to consume stocked charge on this cpu.
2366 * @memcg: memcg to consume from.
2367 * @nr_pages: how many pages to charge.
2369 * The charges will only happen if @memcg matches the current cpu's memcg
2370 * stock, and at least @nr_pages are available in that stock. Failure to
2371 * service an allocation will refill the stock.
2373 * returns true if successful, false otherwise.
2375 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2377 struct memcg_stock_pcp *stock;
2380 if (nr_pages > CHARGE_BATCH)
2383 stock = &get_cpu_var(memcg_stock);
2384 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2385 stock->nr_pages -= nr_pages;
2386 else /* need to call res_counter_charge */
2388 put_cpu_var(memcg_stock);
2393 * Returns stocks cached in percpu to res_counter and reset cached information.
2395 static void drain_stock(struct memcg_stock_pcp *stock)
2397 struct mem_cgroup *old = stock->cached;
2399 if (stock->nr_pages) {
2400 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2402 res_counter_uncharge(&old->res, bytes);
2403 if (do_swap_account)
2404 res_counter_uncharge(&old->memsw, bytes);
2405 stock->nr_pages = 0;
2407 stock->cached = NULL;
2411 * This must be called under preempt disabled or must be called by
2412 * a thread which is pinned to local cpu.
2414 static void drain_local_stock(struct work_struct *dummy)
2416 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2418 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2421 static void __init memcg_stock_init(void)
2425 for_each_possible_cpu(cpu) {
2426 struct memcg_stock_pcp *stock =
2427 &per_cpu(memcg_stock, cpu);
2428 INIT_WORK(&stock->work, drain_local_stock);
2433 * Cache charges(val) which is from res_counter, to local per_cpu area.
2434 * This will be consumed by consume_stock() function, later.
2436 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2438 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2440 if (stock->cached != memcg) { /* reset if necessary */
2442 stock->cached = memcg;
2444 stock->nr_pages += nr_pages;
2445 put_cpu_var(memcg_stock);
2449 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2450 * of the hierarchy under it. sync flag says whether we should block
2451 * until the work is done.
2453 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2457 /* Notify other cpus that system-wide "drain" is running */
2460 for_each_online_cpu(cpu) {
2461 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2462 struct mem_cgroup *memcg;
2464 memcg = stock->cached;
2465 if (!memcg || !stock->nr_pages)
2467 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2469 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2471 drain_local_stock(&stock->work);
2473 schedule_work_on(cpu, &stock->work);
2481 for_each_online_cpu(cpu) {
2482 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2483 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2484 flush_work(&stock->work);
2491 * Tries to drain stocked charges in other cpus. This function is asynchronous
2492 * and just put a work per cpu for draining localy on each cpu. Caller can
2493 * expects some charges will be back to res_counter later but cannot wait for
2496 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2499 * If someone calls draining, avoid adding more kworker runs.
2501 if (!mutex_trylock(&percpu_charge_mutex))
2503 drain_all_stock(root_memcg, false);
2504 mutex_unlock(&percpu_charge_mutex);
2507 /* This is a synchronous drain interface. */
2508 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2510 /* called when force_empty is called */
2511 mutex_lock(&percpu_charge_mutex);
2512 drain_all_stock(root_memcg, true);
2513 mutex_unlock(&percpu_charge_mutex);
2517 * This function drains percpu counter value from DEAD cpu and
2518 * move it to local cpu. Note that this function can be preempted.
2520 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2524 spin_lock(&memcg->pcp_counter_lock);
2525 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2526 long x = per_cpu(memcg->stat->count[i], cpu);
2528 per_cpu(memcg->stat->count[i], cpu) = 0;
2529 memcg->nocpu_base.count[i] += x;
2531 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2532 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2534 per_cpu(memcg->stat->events[i], cpu) = 0;
2535 memcg->nocpu_base.events[i] += x;
2537 spin_unlock(&memcg->pcp_counter_lock);
2540 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2541 unsigned long action,
2544 int cpu = (unsigned long)hcpu;
2545 struct memcg_stock_pcp *stock;
2546 struct mem_cgroup *iter;
2548 if (action == CPU_ONLINE)
2551 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2554 for_each_mem_cgroup(iter)
2555 mem_cgroup_drain_pcp_counter(iter, cpu);
2557 stock = &per_cpu(memcg_stock, cpu);
2563 /* See __mem_cgroup_try_charge() for details */
2565 CHARGE_OK, /* success */
2566 CHARGE_RETRY, /* need to retry but retry is not bad */
2567 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2568 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2571 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2572 unsigned int nr_pages, unsigned int min_pages,
2575 unsigned long csize = nr_pages * PAGE_SIZE;
2576 struct mem_cgroup *mem_over_limit;
2577 struct res_counter *fail_res;
2578 unsigned long flags = 0;
2581 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2584 if (!do_swap_account)
2586 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2590 res_counter_uncharge(&memcg->res, csize);
2591 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2592 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2594 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2596 * Never reclaim on behalf of optional batching, retry with a
2597 * single page instead.
2599 if (nr_pages > min_pages)
2600 return CHARGE_RETRY;
2602 if (!(gfp_mask & __GFP_WAIT))
2603 return CHARGE_WOULDBLOCK;
2605 if (gfp_mask & __GFP_NORETRY)
2606 return CHARGE_NOMEM;
2608 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2609 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2610 return CHARGE_RETRY;
2612 * Even though the limit is exceeded at this point, reclaim
2613 * may have been able to free some pages. Retry the charge
2614 * before killing the task.
2616 * Only for regular pages, though: huge pages are rather
2617 * unlikely to succeed so close to the limit, and we fall back
2618 * to regular pages anyway in case of failure.
2620 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2621 return CHARGE_RETRY;
2624 * At task move, charge accounts can be doubly counted. So, it's
2625 * better to wait until the end of task_move if something is going on.
2627 if (mem_cgroup_wait_acct_move(mem_over_limit))
2628 return CHARGE_RETRY;
2631 mem_cgroup_oom(mem_over_limit, gfp_mask, get_order(csize));
2633 return CHARGE_NOMEM;
2637 * __mem_cgroup_try_charge() does
2638 * 1. detect memcg to be charged against from passed *mm and *ptr,
2639 * 2. update res_counter
2640 * 3. call memory reclaim if necessary.
2642 * In some special case, if the task is fatal, fatal_signal_pending() or
2643 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2644 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2645 * as possible without any hazards. 2: all pages should have a valid
2646 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2647 * pointer, that is treated as a charge to root_mem_cgroup.
2649 * So __mem_cgroup_try_charge() will return
2650 * 0 ... on success, filling *ptr with a valid memcg pointer.
2651 * -ENOMEM ... charge failure because of resource limits.
2652 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2654 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2655 * the oom-killer can be invoked.
2657 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2659 unsigned int nr_pages,
2660 struct mem_cgroup **ptr,
2663 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2664 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2665 struct mem_cgroup *memcg = NULL;
2669 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2670 * in system level. So, allow to go ahead dying process in addition to
2673 if (unlikely(test_thread_flag(TIF_MEMDIE)
2674 || fatal_signal_pending(current)))
2678 * We always charge the cgroup the mm_struct belongs to.
2679 * The mm_struct's mem_cgroup changes on task migration if the
2680 * thread group leader migrates. It's possible that mm is not
2681 * set, if so charge the root memcg (happens for pagecache usage).
2684 *ptr = root_mem_cgroup;
2686 if (*ptr) { /* css should be a valid one */
2688 if (mem_cgroup_is_root(memcg))
2690 if (consume_stock(memcg, nr_pages))
2692 css_get(&memcg->css);
2694 struct task_struct *p;
2697 p = rcu_dereference(mm->owner);
2699 * Because we don't have task_lock(), "p" can exit.
2700 * In that case, "memcg" can point to root or p can be NULL with
2701 * race with swapoff. Then, we have small risk of mis-accouning.
2702 * But such kind of mis-account by race always happens because
2703 * we don't have cgroup_mutex(). It's overkill and we allo that
2705 * (*) swapoff at el will charge against mm-struct not against
2706 * task-struct. So, mm->owner can be NULL.
2708 memcg = mem_cgroup_from_task(p);
2710 memcg = root_mem_cgroup;
2711 if (mem_cgroup_is_root(memcg)) {
2715 if (consume_stock(memcg, nr_pages)) {
2717 * It seems dagerous to access memcg without css_get().
2718 * But considering how consume_stok works, it's not
2719 * necessary. If consume_stock success, some charges
2720 * from this memcg are cached on this cpu. So, we
2721 * don't need to call css_get()/css_tryget() before
2722 * calling consume_stock().
2727 /* after here, we may be blocked. we need to get refcnt */
2728 if (!css_tryget(&memcg->css)) {
2736 bool invoke_oom = oom && !nr_oom_retries;
2738 /* If killed, bypass charge */
2739 if (fatal_signal_pending(current)) {
2740 css_put(&memcg->css);
2744 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch,
2745 nr_pages, invoke_oom);
2749 case CHARGE_RETRY: /* not in OOM situation but retry */
2751 css_put(&memcg->css);
2754 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2755 css_put(&memcg->css);
2757 case CHARGE_NOMEM: /* OOM routine works */
2758 if (!oom || invoke_oom) {
2759 css_put(&memcg->css);
2765 } while (ret != CHARGE_OK);
2767 if (batch > nr_pages)
2768 refill_stock(memcg, batch - nr_pages);
2769 css_put(&memcg->css);
2777 *ptr = root_mem_cgroup;
2782 * Somemtimes we have to undo a charge we got by try_charge().
2783 * This function is for that and do uncharge, put css's refcnt.
2784 * gotten by try_charge().
2786 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2787 unsigned int nr_pages)
2789 if (!mem_cgroup_is_root(memcg)) {
2790 unsigned long bytes = nr_pages * PAGE_SIZE;
2792 res_counter_uncharge(&memcg->res, bytes);
2793 if (do_swap_account)
2794 res_counter_uncharge(&memcg->memsw, bytes);
2799 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2800 * This is useful when moving usage to parent cgroup.
2802 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2803 unsigned int nr_pages)
2805 unsigned long bytes = nr_pages * PAGE_SIZE;
2807 if (mem_cgroup_is_root(memcg))
2810 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2811 if (do_swap_account)
2812 res_counter_uncharge_until(&memcg->memsw,
2813 memcg->memsw.parent, bytes);
2817 * A helper function to get mem_cgroup from ID. must be called under
2818 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2819 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2820 * called against removed memcg.)
2822 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2824 struct cgroup_subsys_state *css;
2826 /* ID 0 is unused ID */
2829 css = css_lookup(&mem_cgroup_subsys, id);
2832 return mem_cgroup_from_css(css);
2835 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2837 struct mem_cgroup *memcg = NULL;
2838 struct page_cgroup *pc;
2842 VM_BUG_ON(!PageLocked(page));
2844 pc = lookup_page_cgroup(page);
2845 lock_page_cgroup(pc);
2846 if (PageCgroupUsed(pc)) {
2847 memcg = pc->mem_cgroup;
2848 if (memcg && !css_tryget(&memcg->css))
2850 } else if (PageSwapCache(page)) {
2851 ent.val = page_private(page);
2852 id = lookup_swap_cgroup_id(ent);
2854 memcg = mem_cgroup_lookup(id);
2855 if (memcg && !css_tryget(&memcg->css))
2859 unlock_page_cgroup(pc);
2863 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2865 unsigned int nr_pages,
2866 enum charge_type ctype,
2869 struct page_cgroup *pc = lookup_page_cgroup(page);
2870 struct zone *uninitialized_var(zone);
2871 struct lruvec *lruvec;
2872 bool was_on_lru = false;
2875 lock_page_cgroup(pc);
2876 VM_BUG_ON(PageCgroupUsed(pc));
2878 * we don't need page_cgroup_lock about tail pages, becase they are not
2879 * accessed by any other context at this point.
2883 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2884 * may already be on some other mem_cgroup's LRU. Take care of it.
2887 zone = page_zone(page);
2888 spin_lock_irq(&zone->lru_lock);
2889 if (PageLRU(page)) {
2890 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2892 del_page_from_lru_list(page, lruvec, page_lru(page));
2897 pc->mem_cgroup = memcg;
2899 * We access a page_cgroup asynchronously without lock_page_cgroup().
2900 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2901 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2902 * before USED bit, we need memory barrier here.
2903 * See mem_cgroup_add_lru_list(), etc.
2906 SetPageCgroupUsed(pc);
2910 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2911 VM_BUG_ON(PageLRU(page));
2913 add_page_to_lru_list(page, lruvec, page_lru(page));
2915 spin_unlock_irq(&zone->lru_lock);
2918 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2923 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2924 unlock_page_cgroup(pc);
2927 * "charge_statistics" updated event counter. Then, check it.
2928 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2929 * if they exceeds softlimit.
2931 memcg_check_events(memcg, page);
2934 static DEFINE_MUTEX(set_limit_mutex);
2936 #ifdef CONFIG_MEMCG_KMEM
2937 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2939 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2940 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2944 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2945 * in the memcg_cache_params struct.
2947 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2949 struct kmem_cache *cachep;
2951 VM_BUG_ON(p->is_root_cache);
2952 cachep = p->root_cache;
2953 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2956 #ifdef CONFIG_SLABINFO
2957 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2958 struct cftype *cft, struct seq_file *m)
2960 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2961 struct memcg_cache_params *params;
2963 if (!memcg_can_account_kmem(memcg))
2966 print_slabinfo_header(m);
2968 mutex_lock(&memcg->slab_caches_mutex);
2969 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2970 cache_show(memcg_params_to_cache(params), m);
2971 mutex_unlock(&memcg->slab_caches_mutex);
2977 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2979 struct res_counter *fail_res;
2980 struct mem_cgroup *_memcg;
2984 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2989 * Conditions under which we can wait for the oom_killer. Those are
2990 * the same conditions tested by the core page allocator
2992 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2995 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2998 if (ret == -EINTR) {
3000 * __mem_cgroup_try_charge() chosed to bypass to root due to
3001 * OOM kill or fatal signal. Since our only options are to
3002 * either fail the allocation or charge it to this cgroup, do
3003 * it as a temporary condition. But we can't fail. From a
3004 * kmem/slab perspective, the cache has already been selected,
3005 * by mem_cgroup_kmem_get_cache(), so it is too late to change
3008 * This condition will only trigger if the task entered
3009 * memcg_charge_kmem in a sane state, but was OOM-killed during
3010 * __mem_cgroup_try_charge() above. Tasks that were already
3011 * dying when the allocation triggers should have been already
3012 * directed to the root cgroup in memcontrol.h
3014 res_counter_charge_nofail(&memcg->res, size, &fail_res);
3015 if (do_swap_account)
3016 res_counter_charge_nofail(&memcg->memsw, size,
3020 res_counter_uncharge(&memcg->kmem, size);
3025 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3027 res_counter_uncharge(&memcg->res, size);
3028 if (do_swap_account)
3029 res_counter_uncharge(&memcg->memsw, size);
3032 if (res_counter_uncharge(&memcg->kmem, size))
3036 * Releases a reference taken in kmem_cgroup_css_offline in case
3037 * this last uncharge is racing with the offlining code or it is
3038 * outliving the memcg existence.
3040 * The memory barrier imposed by test&clear is paired with the
3041 * explicit one in memcg_kmem_mark_dead().
3043 if (memcg_kmem_test_and_clear_dead(memcg))
3044 css_put(&memcg->css);
3047 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3052 mutex_lock(&memcg->slab_caches_mutex);
3053 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3054 mutex_unlock(&memcg->slab_caches_mutex);
3058 * helper for acessing a memcg's index. It will be used as an index in the
3059 * child cache array in kmem_cache, and also to derive its name. This function
3060 * will return -1 when this is not a kmem-limited memcg.
3062 int memcg_cache_id(struct mem_cgroup *memcg)
3064 return memcg ? memcg->kmemcg_id : -1;
3068 * This ends up being protected by the set_limit mutex, during normal
3069 * operation, because that is its main call site.
3071 * But when we create a new cache, we can call this as well if its parent
3072 * is kmem-limited. That will have to hold set_limit_mutex as well.
3074 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3078 num = ida_simple_get(&kmem_limited_groups,
3079 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3083 * After this point, kmem_accounted (that we test atomically in
3084 * the beginning of this conditional), is no longer 0. This
3085 * guarantees only one process will set the following boolean
3086 * to true. We don't need test_and_set because we're protected
3087 * by the set_limit_mutex anyway.
3089 memcg_kmem_set_activated(memcg);
3091 ret = memcg_update_all_caches(num+1);
3093 ida_simple_remove(&kmem_limited_groups, num);
3094 memcg_kmem_clear_activated(memcg);
3098 memcg->kmemcg_id = num;
3099 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3100 mutex_init(&memcg->slab_caches_mutex);
3104 static size_t memcg_caches_array_size(int num_groups)
3107 if (num_groups <= 0)
3110 size = 2 * num_groups;
3111 if (size < MEMCG_CACHES_MIN_SIZE)
3112 size = MEMCG_CACHES_MIN_SIZE;
3113 else if (size > MEMCG_CACHES_MAX_SIZE)
3114 size = MEMCG_CACHES_MAX_SIZE;
3120 * We should update the current array size iff all caches updates succeed. This
3121 * can only be done from the slab side. The slab mutex needs to be held when
3124 void memcg_update_array_size(int num)
3126 if (num > memcg_limited_groups_array_size)
3127 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3130 static void kmem_cache_destroy_work_func(struct work_struct *w);
3132 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3134 struct memcg_cache_params *cur_params = s->memcg_params;
3136 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3138 if (num_groups > memcg_limited_groups_array_size) {
3140 ssize_t size = memcg_caches_array_size(num_groups);
3142 size *= sizeof(void *);
3143 size += offsetof(struct memcg_cache_params, memcg_caches);
3145 s->memcg_params = kzalloc(size, GFP_KERNEL);
3146 if (!s->memcg_params) {
3147 s->memcg_params = cur_params;
3151 s->memcg_params->is_root_cache = true;
3154 * There is the chance it will be bigger than
3155 * memcg_limited_groups_array_size, if we failed an allocation
3156 * in a cache, in which case all caches updated before it, will
3157 * have a bigger array.
3159 * But if that is the case, the data after
3160 * memcg_limited_groups_array_size is certainly unused
3162 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3163 if (!cur_params->memcg_caches[i])
3165 s->memcg_params->memcg_caches[i] =
3166 cur_params->memcg_caches[i];
3170 * Ideally, we would wait until all caches succeed, and only
3171 * then free the old one. But this is not worth the extra
3172 * pointer per-cache we'd have to have for this.
3174 * It is not a big deal if some caches are left with a size
3175 * bigger than the others. And all updates will reset this
3183 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3184 struct kmem_cache *root_cache)
3188 if (!memcg_kmem_enabled())
3192 size = offsetof(struct memcg_cache_params, memcg_caches);
3193 size += memcg_limited_groups_array_size * sizeof(void *);
3195 size = sizeof(struct memcg_cache_params);
3197 s->memcg_params = kzalloc(size, GFP_KERNEL);
3198 if (!s->memcg_params)
3202 s->memcg_params->memcg = memcg;
3203 s->memcg_params->root_cache = root_cache;
3204 INIT_WORK(&s->memcg_params->destroy,
3205 kmem_cache_destroy_work_func);
3207 s->memcg_params->is_root_cache = true;
3212 void memcg_release_cache(struct kmem_cache *s)
3214 struct kmem_cache *root;
3215 struct mem_cgroup *memcg;
3219 * This happens, for instance, when a root cache goes away before we
3222 if (!s->memcg_params)
3225 if (s->memcg_params->is_root_cache)
3228 memcg = s->memcg_params->memcg;
3229 id = memcg_cache_id(memcg);
3231 root = s->memcg_params->root_cache;
3232 root->memcg_params->memcg_caches[id] = NULL;
3234 mutex_lock(&memcg->slab_caches_mutex);
3235 list_del(&s->memcg_params->list);
3236 mutex_unlock(&memcg->slab_caches_mutex);
3238 css_put(&memcg->css);
3240 kfree(s->memcg_params);
3244 * During the creation a new cache, we need to disable our accounting mechanism
3245 * altogether. This is true even if we are not creating, but rather just
3246 * enqueing new caches to be created.
3248 * This is because that process will trigger allocations; some visible, like
3249 * explicit kmallocs to auxiliary data structures, name strings and internal
3250 * cache structures; some well concealed, like INIT_WORK() that can allocate
3251 * objects during debug.
3253 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3254 * to it. This may not be a bounded recursion: since the first cache creation
3255 * failed to complete (waiting on the allocation), we'll just try to create the
3256 * cache again, failing at the same point.
3258 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3259 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3260 * inside the following two functions.
3262 static inline void memcg_stop_kmem_account(void)
3264 VM_BUG_ON(!current->mm);
3265 current->memcg_kmem_skip_account++;
3268 static inline void memcg_resume_kmem_account(void)
3270 VM_BUG_ON(!current->mm);
3271 current->memcg_kmem_skip_account--;
3274 static void kmem_cache_destroy_work_func(struct work_struct *w)
3276 struct kmem_cache *cachep;
3277 struct memcg_cache_params *p;
3279 p = container_of(w, struct memcg_cache_params, destroy);
3281 cachep = memcg_params_to_cache(p);
3284 * If we get down to 0 after shrink, we could delete right away.
3285 * However, memcg_release_pages() already puts us back in the workqueue
3286 * in that case. If we proceed deleting, we'll get a dangling
3287 * reference, and removing the object from the workqueue in that case
3288 * is unnecessary complication. We are not a fast path.
3290 * Note that this case is fundamentally different from racing with
3291 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3292 * kmem_cache_shrink, not only we would be reinserting a dead cache
3293 * into the queue, but doing so from inside the worker racing to
3296 * So if we aren't down to zero, we'll just schedule a worker and try
3299 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3300 kmem_cache_shrink(cachep);
3301 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3304 kmem_cache_destroy(cachep);
3307 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3309 if (!cachep->memcg_params->dead)
3313 * There are many ways in which we can get here.
3315 * We can get to a memory-pressure situation while the delayed work is
3316 * still pending to run. The vmscan shrinkers can then release all
3317 * cache memory and get us to destruction. If this is the case, we'll
3318 * be executed twice, which is a bug (the second time will execute over
3319 * bogus data). In this case, cancelling the work should be fine.
3321 * But we can also get here from the worker itself, if
3322 * kmem_cache_shrink is enough to shake all the remaining objects and
3323 * get the page count to 0. In this case, we'll deadlock if we try to
3324 * cancel the work (the worker runs with an internal lock held, which
3325 * is the same lock we would hold for cancel_work_sync().)
3327 * Since we can't possibly know who got us here, just refrain from
3328 * running if there is already work pending
3330 if (work_pending(&cachep->memcg_params->destroy))
3333 * We have to defer the actual destroying to a workqueue, because
3334 * we might currently be in a context that cannot sleep.
3336 schedule_work(&cachep->memcg_params->destroy);
3340 * This lock protects updaters, not readers. We want readers to be as fast as
3341 * they can, and they will either see NULL or a valid cache value. Our model
3342 * allow them to see NULL, in which case the root memcg will be selected.
3344 * We need this lock because multiple allocations to the same cache from a non
3345 * will span more than one worker. Only one of them can create the cache.
3347 static DEFINE_MUTEX(memcg_cache_mutex);
3350 * Called with memcg_cache_mutex held
3352 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3353 struct kmem_cache *s)
3355 struct kmem_cache *new;
3356 static char *tmp_name = NULL;
3358 lockdep_assert_held(&memcg_cache_mutex);
3361 * kmem_cache_create_memcg duplicates the given name and
3362 * cgroup_name for this name requires RCU context.
3363 * This static temporary buffer is used to prevent from
3364 * pointless shortliving allocation.
3367 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3373 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3374 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3377 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3378 (s->flags & ~SLAB_PANIC), s->ctor, s);
3381 new->allocflags |= __GFP_KMEMCG;
3386 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3387 struct kmem_cache *cachep)
3389 struct kmem_cache *new_cachep;
3392 BUG_ON(!memcg_can_account_kmem(memcg));
3394 idx = memcg_cache_id(memcg);
3396 mutex_lock(&memcg_cache_mutex);
3397 new_cachep = cachep->memcg_params->memcg_caches[idx];
3399 css_put(&memcg->css);
3403 new_cachep = kmem_cache_dup(memcg, cachep);
3404 if (new_cachep == NULL) {
3405 new_cachep = cachep;
3406 css_put(&memcg->css);
3410 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3412 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3414 * the readers won't lock, make sure everybody sees the updated value,
3415 * so they won't put stuff in the queue again for no reason
3419 mutex_unlock(&memcg_cache_mutex);
3423 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3425 struct kmem_cache *c;
3428 if (!s->memcg_params)
3430 if (!s->memcg_params->is_root_cache)
3434 * If the cache is being destroyed, we trust that there is no one else
3435 * requesting objects from it. Even if there are, the sanity checks in
3436 * kmem_cache_destroy should caught this ill-case.
3438 * Still, we don't want anyone else freeing memcg_caches under our
3439 * noses, which can happen if a new memcg comes to life. As usual,
3440 * we'll take the set_limit_mutex to protect ourselves against this.
3442 mutex_lock(&set_limit_mutex);
3443 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3444 c = s->memcg_params->memcg_caches[i];
3449 * We will now manually delete the caches, so to avoid races
3450 * we need to cancel all pending destruction workers and
3451 * proceed with destruction ourselves.
3453 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3454 * and that could spawn the workers again: it is likely that
3455 * the cache still have active pages until this very moment.
3456 * This would lead us back to mem_cgroup_destroy_cache.
3458 * But that will not execute at all if the "dead" flag is not
3459 * set, so flip it down to guarantee we are in control.
3461 c->memcg_params->dead = false;
3462 cancel_work_sync(&c->memcg_params->destroy);
3463 kmem_cache_destroy(c);
3465 mutex_unlock(&set_limit_mutex);
3468 struct create_work {
3469 struct mem_cgroup *memcg;
3470 struct kmem_cache *cachep;
3471 struct work_struct work;
3474 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3476 struct kmem_cache *cachep;
3477 struct memcg_cache_params *params;
3479 if (!memcg_kmem_is_active(memcg))
3482 mutex_lock(&memcg->slab_caches_mutex);
3483 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3484 cachep = memcg_params_to_cache(params);
3485 cachep->memcg_params->dead = true;
3486 schedule_work(&cachep->memcg_params->destroy);
3488 mutex_unlock(&memcg->slab_caches_mutex);
3491 static void memcg_create_cache_work_func(struct work_struct *w)
3493 struct create_work *cw;
3495 cw = container_of(w, struct create_work, work);
3496 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3501 * Enqueue the creation of a per-memcg kmem_cache.
3503 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3504 struct kmem_cache *cachep)
3506 struct create_work *cw;
3508 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3510 css_put(&memcg->css);
3515 cw->cachep = cachep;
3517 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3518 schedule_work(&cw->work);
3521 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3522 struct kmem_cache *cachep)
3525 * We need to stop accounting when we kmalloc, because if the
3526 * corresponding kmalloc cache is not yet created, the first allocation
3527 * in __memcg_create_cache_enqueue will recurse.
3529 * However, it is better to enclose the whole function. Depending on
3530 * the debugging options enabled, INIT_WORK(), for instance, can
3531 * trigger an allocation. This too, will make us recurse. Because at
3532 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3533 * the safest choice is to do it like this, wrapping the whole function.
3535 memcg_stop_kmem_account();
3536 __memcg_create_cache_enqueue(memcg, cachep);
3537 memcg_resume_kmem_account();
3540 * Return the kmem_cache we're supposed to use for a slab allocation.
3541 * We try to use the current memcg's version of the cache.
3543 * If the cache does not exist yet, if we are the first user of it,
3544 * we either create it immediately, if possible, or create it asynchronously
3546 * In the latter case, we will let the current allocation go through with
3547 * the original cache.
3549 * Can't be called in interrupt context or from kernel threads.
3550 * This function needs to be called with rcu_read_lock() held.
3552 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3555 struct mem_cgroup *memcg;
3558 VM_BUG_ON(!cachep->memcg_params);
3559 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3561 if (!current->mm || current->memcg_kmem_skip_account)
3565 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3567 if (!memcg_can_account_kmem(memcg))
3570 idx = memcg_cache_id(memcg);
3573 * barrier to mare sure we're always seeing the up to date value. The
3574 * code updating memcg_caches will issue a write barrier to match this.
3576 read_barrier_depends();
3577 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3578 cachep = cachep->memcg_params->memcg_caches[idx];
3582 /* The corresponding put will be done in the workqueue. */
3583 if (!css_tryget(&memcg->css))
3588 * If we are in a safe context (can wait, and not in interrupt
3589 * context), we could be be predictable and return right away.
3590 * This would guarantee that the allocation being performed
3591 * already belongs in the new cache.
3593 * However, there are some clashes that can arrive from locking.
3594 * For instance, because we acquire the slab_mutex while doing
3595 * kmem_cache_dup, this means no further allocation could happen
3596 * with the slab_mutex held.
3598 * Also, because cache creation issue get_online_cpus(), this
3599 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3600 * that ends up reversed during cpu hotplug. (cpuset allocates
3601 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3602 * better to defer everything.
3604 memcg_create_cache_enqueue(memcg, cachep);
3610 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3613 * We need to verify if the allocation against current->mm->owner's memcg is
3614 * possible for the given order. But the page is not allocated yet, so we'll
3615 * need a further commit step to do the final arrangements.
3617 * It is possible for the task to switch cgroups in this mean time, so at
3618 * commit time, we can't rely on task conversion any longer. We'll then use
3619 * the handle argument to return to the caller which cgroup we should commit
3620 * against. We could also return the memcg directly and avoid the pointer
3621 * passing, but a boolean return value gives better semantics considering
3622 * the compiled-out case as well.
3624 * Returning true means the allocation is possible.
3627 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3629 struct mem_cgroup *memcg;
3635 * Disabling accounting is only relevant for some specific memcg
3636 * internal allocations. Therefore we would initially not have such
3637 * check here, since direct calls to the page allocator that are marked
3638 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3639 * concerned with cache allocations, and by having this test at
3640 * memcg_kmem_get_cache, we are already able to relay the allocation to
3641 * the root cache and bypass the memcg cache altogether.
3643 * There is one exception, though: the SLUB allocator does not create
3644 * large order caches, but rather service large kmallocs directly from
3645 * the page allocator. Therefore, the following sequence when backed by
3646 * the SLUB allocator:
3648 * memcg_stop_kmem_account();
3649 * kmalloc(<large_number>)
3650 * memcg_resume_kmem_account();
3652 * would effectively ignore the fact that we should skip accounting,
3653 * since it will drive us directly to this function without passing
3654 * through the cache selector memcg_kmem_get_cache. Such large
3655 * allocations are extremely rare but can happen, for instance, for the
3656 * cache arrays. We bring this test here.
3658 if (!current->mm || current->memcg_kmem_skip_account)
3661 memcg = try_get_mem_cgroup_from_mm(current->mm);
3664 * very rare case described in mem_cgroup_from_task. Unfortunately there
3665 * isn't much we can do without complicating this too much, and it would
3666 * be gfp-dependent anyway. Just let it go
3668 if (unlikely(!memcg))
3671 if (!memcg_can_account_kmem(memcg)) {
3672 css_put(&memcg->css);
3676 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3680 css_put(&memcg->css);
3684 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3687 struct page_cgroup *pc;
3689 VM_BUG_ON(mem_cgroup_is_root(memcg));
3691 /* The page allocation failed. Revert */
3693 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3697 pc = lookup_page_cgroup(page);
3698 lock_page_cgroup(pc);
3699 pc->mem_cgroup = memcg;
3700 SetPageCgroupUsed(pc);
3701 unlock_page_cgroup(pc);
3704 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3706 struct mem_cgroup *memcg = NULL;
3707 struct page_cgroup *pc;
3710 pc = lookup_page_cgroup(page);
3712 * Fast unlocked return. Theoretically might have changed, have to
3713 * check again after locking.
3715 if (!PageCgroupUsed(pc))
3718 lock_page_cgroup(pc);
3719 if (PageCgroupUsed(pc)) {
3720 memcg = pc->mem_cgroup;
3721 ClearPageCgroupUsed(pc);
3723 unlock_page_cgroup(pc);
3726 * We trust that only if there is a memcg associated with the page, it
3727 * is a valid allocation
3732 VM_BUG_ON(mem_cgroup_is_root(memcg));
3733 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3736 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3739 #endif /* CONFIG_MEMCG_KMEM */
3741 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3743 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3745 * Because tail pages are not marked as "used", set it. We're under
3746 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3747 * charge/uncharge will be never happen and move_account() is done under
3748 * compound_lock(), so we don't have to take care of races.
3750 void mem_cgroup_split_huge_fixup(struct page *head)
3752 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3753 struct page_cgroup *pc;
3754 struct mem_cgroup *memcg;
3757 if (mem_cgroup_disabled())
3760 memcg = head_pc->mem_cgroup;
3761 for (i = 1; i < HPAGE_PMD_NR; i++) {
3763 pc->mem_cgroup = memcg;
3764 smp_wmb();/* see __commit_charge() */
3765 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3767 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3770 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3773 void mem_cgroup_move_account_page_stat(struct mem_cgroup *from,
3774 struct mem_cgroup *to,
3775 unsigned int nr_pages,
3776 enum mem_cgroup_stat_index idx)
3778 /* Update stat data for mem_cgroup */
3780 WARN_ON_ONCE(from->stat->count[idx] < nr_pages);
3781 __this_cpu_add(from->stat->count[idx], -nr_pages);
3782 __this_cpu_add(to->stat->count[idx], nr_pages);
3787 * mem_cgroup_move_account - move account of the page
3789 * @nr_pages: number of regular pages (>1 for huge pages)
3790 * @pc: page_cgroup of the page.
3791 * @from: mem_cgroup which the page is moved from.
3792 * @to: mem_cgroup which the page is moved to. @from != @to.
3794 * The caller must confirm following.
3795 * - page is not on LRU (isolate_page() is useful.)
3796 * - compound_lock is held when nr_pages > 1
3798 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3801 static int mem_cgroup_move_account(struct page *page,
3802 unsigned int nr_pages,
3803 struct page_cgroup *pc,
3804 struct mem_cgroup *from,
3805 struct mem_cgroup *to)
3807 unsigned long flags;
3809 bool anon = PageAnon(page);
3811 VM_BUG_ON(from == to);
3812 VM_BUG_ON(PageLRU(page));
3814 * The page is isolated from LRU. So, collapse function
3815 * will not handle this page. But page splitting can happen.
3816 * Do this check under compound_page_lock(). The caller should
3820 if (nr_pages > 1 && !PageTransHuge(page))
3823 lock_page_cgroup(pc);
3826 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3829 move_lock_mem_cgroup(from, &flags);
3831 if (!anon && page_mapped(page))
3832 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3833 MEM_CGROUP_STAT_FILE_MAPPED);
3835 if (PageWriteback(page))
3836 mem_cgroup_move_account_page_stat(from, to, nr_pages,
3837 MEM_CGROUP_STAT_WRITEBACK);
3839 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3841 /* caller should have done css_get */
3842 pc->mem_cgroup = to;
3843 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3844 move_unlock_mem_cgroup(from, &flags);
3847 unlock_page_cgroup(pc);
3851 memcg_check_events(to, page);
3852 memcg_check_events(from, page);
3858 * mem_cgroup_move_parent - moves page to the parent group
3859 * @page: the page to move
3860 * @pc: page_cgroup of the page
3861 * @child: page's cgroup
3863 * move charges to its parent or the root cgroup if the group has no
3864 * parent (aka use_hierarchy==0).
3865 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3866 * mem_cgroup_move_account fails) the failure is always temporary and
3867 * it signals a race with a page removal/uncharge or migration. In the
3868 * first case the page is on the way out and it will vanish from the LRU
3869 * on the next attempt and the call should be retried later.
3870 * Isolation from the LRU fails only if page has been isolated from
3871 * the LRU since we looked at it and that usually means either global
3872 * reclaim or migration going on. The page will either get back to the
3874 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3875 * (!PageCgroupUsed) or moved to a different group. The page will
3876 * disappear in the next attempt.
3878 static int mem_cgroup_move_parent(struct page *page,
3879 struct page_cgroup *pc,
3880 struct mem_cgroup *child)
3882 struct mem_cgroup *parent;
3883 unsigned int nr_pages;
3884 unsigned long uninitialized_var(flags);
3887 VM_BUG_ON(mem_cgroup_is_root(child));
3890 if (!get_page_unless_zero(page))
3892 if (isolate_lru_page(page))
3895 nr_pages = hpage_nr_pages(page);
3897 parent = parent_mem_cgroup(child);
3899 * If no parent, move charges to root cgroup.
3902 parent = root_mem_cgroup;
3905 VM_BUG_ON(!PageTransHuge(page));
3906 flags = compound_lock_irqsave(page);
3909 ret = mem_cgroup_move_account(page, nr_pages,
3912 __mem_cgroup_cancel_local_charge(child, nr_pages);
3915 compound_unlock_irqrestore(page, flags);
3916 putback_lru_page(page);
3924 * Charge the memory controller for page usage.
3926 * 0 if the charge was successful
3927 * < 0 if the cgroup is over its limit
3929 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3930 gfp_t gfp_mask, enum charge_type ctype)
3932 struct mem_cgroup *memcg = NULL;
3933 unsigned int nr_pages = 1;
3937 if (PageTransHuge(page)) {
3938 nr_pages <<= compound_order(page);
3939 VM_BUG_ON(!PageTransHuge(page));
3941 * Never OOM-kill a process for a huge page. The
3942 * fault handler will fall back to regular pages.
3947 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3950 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3954 int mem_cgroup_newpage_charge(struct page *page,
3955 struct mm_struct *mm, gfp_t gfp_mask)
3957 if (mem_cgroup_disabled())
3959 VM_BUG_ON(page_mapped(page));
3960 VM_BUG_ON(page->mapping && !PageAnon(page));
3962 return mem_cgroup_charge_common(page, mm, gfp_mask,
3963 MEM_CGROUP_CHARGE_TYPE_ANON);
3967 * While swap-in, try_charge -> commit or cancel, the page is locked.
3968 * And when try_charge() successfully returns, one refcnt to memcg without
3969 * struct page_cgroup is acquired. This refcnt will be consumed by
3970 * "commit()" or removed by "cancel()"
3972 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3975 struct mem_cgroup **memcgp)
3977 struct mem_cgroup *memcg;
3978 struct page_cgroup *pc;
3981 pc = lookup_page_cgroup(page);
3983 * Every swap fault against a single page tries to charge the
3984 * page, bail as early as possible. shmem_unuse() encounters
3985 * already charged pages, too. The USED bit is protected by
3986 * the page lock, which serializes swap cache removal, which
3987 * in turn serializes uncharging.
3989 if (PageCgroupUsed(pc))
3991 if (!do_swap_account)
3993 memcg = try_get_mem_cgroup_from_page(page);
3997 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3998 css_put(&memcg->css);
4003 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
4009 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
4010 gfp_t gfp_mask, struct mem_cgroup **memcgp)
4013 if (mem_cgroup_disabled())
4016 * A racing thread's fault, or swapoff, may have already
4017 * updated the pte, and even removed page from swap cache: in
4018 * those cases unuse_pte()'s pte_same() test will fail; but
4019 * there's also a KSM case which does need to charge the page.
4021 if (!PageSwapCache(page)) {
4024 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
4029 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
4032 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4034 if (mem_cgroup_disabled())
4038 __mem_cgroup_cancel_charge(memcg, 1);
4042 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4043 enum charge_type ctype)
4045 if (mem_cgroup_disabled())
4050 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4052 * Now swap is on-memory. This means this page may be
4053 * counted both as mem and swap....double count.
4054 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4055 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4056 * may call delete_from_swap_cache() before reach here.
4058 if (do_swap_account && PageSwapCache(page)) {
4059 swp_entry_t ent = {.val = page_private(page)};
4060 mem_cgroup_uncharge_swap(ent);
4064 void mem_cgroup_commit_charge_swapin(struct page *page,
4065 struct mem_cgroup *memcg)
4067 __mem_cgroup_commit_charge_swapin(page, memcg,
4068 MEM_CGROUP_CHARGE_TYPE_ANON);
4071 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4074 struct mem_cgroup *memcg = NULL;
4075 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4078 if (mem_cgroup_disabled())
4080 if (PageCompound(page))
4083 if (!PageSwapCache(page))
4084 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4085 else { /* page is swapcache/shmem */
4086 ret = __mem_cgroup_try_charge_swapin(mm, page,
4089 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4094 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4095 unsigned int nr_pages,
4096 const enum charge_type ctype)
4098 struct memcg_batch_info *batch = NULL;
4099 bool uncharge_memsw = true;
4101 /* If swapout, usage of swap doesn't decrease */
4102 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4103 uncharge_memsw = false;
4105 batch = ¤t->memcg_batch;
4107 * In usual, we do css_get() when we remember memcg pointer.
4108 * But in this case, we keep res->usage until end of a series of
4109 * uncharges. Then, it's ok to ignore memcg's refcnt.
4112 batch->memcg = memcg;
4114 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4115 * In those cases, all pages freed continuously can be expected to be in
4116 * the same cgroup and we have chance to coalesce uncharges.
4117 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4118 * because we want to do uncharge as soon as possible.
4121 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4122 goto direct_uncharge;
4125 goto direct_uncharge;
4128 * In typical case, batch->memcg == mem. This means we can
4129 * merge a series of uncharges to an uncharge of res_counter.
4130 * If not, we uncharge res_counter ony by one.
4132 if (batch->memcg != memcg)
4133 goto direct_uncharge;
4134 /* remember freed charge and uncharge it later */
4137 batch->memsw_nr_pages++;
4140 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4142 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4143 if (unlikely(batch->memcg != memcg))
4144 memcg_oom_recover(memcg);
4148 * uncharge if !page_mapped(page)
4150 static struct mem_cgroup *
4151 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4154 struct mem_cgroup *memcg = NULL;
4155 unsigned int nr_pages = 1;
4156 struct page_cgroup *pc;
4159 if (mem_cgroup_disabled())
4162 if (PageTransHuge(page)) {
4163 nr_pages <<= compound_order(page);
4164 VM_BUG_ON(!PageTransHuge(page));
4167 * Check if our page_cgroup is valid
4169 pc = lookup_page_cgroup(page);
4170 if (unlikely(!PageCgroupUsed(pc)))
4173 lock_page_cgroup(pc);
4175 memcg = pc->mem_cgroup;
4177 if (!PageCgroupUsed(pc))
4180 anon = PageAnon(page);
4183 case MEM_CGROUP_CHARGE_TYPE_ANON:
4185 * Generally PageAnon tells if it's the anon statistics to be
4186 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4187 * used before page reached the stage of being marked PageAnon.
4191 case MEM_CGROUP_CHARGE_TYPE_DROP:
4192 /* See mem_cgroup_prepare_migration() */
4193 if (page_mapped(page))
4196 * Pages under migration may not be uncharged. But
4197 * end_migration() /must/ be the one uncharging the
4198 * unused post-migration page and so it has to call
4199 * here with the migration bit still set. See the
4200 * res_counter handling below.
4202 if (!end_migration && PageCgroupMigration(pc))
4205 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4206 if (!PageAnon(page)) { /* Shared memory */
4207 if (page->mapping && !page_is_file_cache(page))
4209 } else if (page_mapped(page)) /* Anon */
4216 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4218 ClearPageCgroupUsed(pc);
4220 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4221 * freed from LRU. This is safe because uncharged page is expected not
4222 * to be reused (freed soon). Exception is SwapCache, it's handled by
4223 * special functions.
4226 unlock_page_cgroup(pc);
4228 * even after unlock, we have memcg->res.usage here and this memcg
4229 * will never be freed, so it's safe to call css_get().
4231 memcg_check_events(memcg, page);
4232 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4233 mem_cgroup_swap_statistics(memcg, true);
4234 css_get(&memcg->css);
4237 * Migration does not charge the res_counter for the
4238 * replacement page, so leave it alone when phasing out the
4239 * page that is unused after the migration.
4241 if (!end_migration && !mem_cgroup_is_root(memcg))
4242 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4247 unlock_page_cgroup(pc);
4251 void mem_cgroup_uncharge_page(struct page *page)
4254 if (page_mapped(page))
4256 VM_BUG_ON(page->mapping && !PageAnon(page));
4258 * If the page is in swap cache, uncharge should be deferred
4259 * to the swap path, which also properly accounts swap usage
4260 * and handles memcg lifetime.
4262 * Note that this check is not stable and reclaim may add the
4263 * page to swap cache at any time after this. However, if the
4264 * page is not in swap cache by the time page->mapcount hits
4265 * 0, there won't be any page table references to the swap
4266 * slot, and reclaim will free it and not actually write the
4269 if (PageSwapCache(page))
4271 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4274 void mem_cgroup_uncharge_cache_page(struct page *page)
4276 VM_BUG_ON(page_mapped(page));
4277 VM_BUG_ON(page->mapping);
4278 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4282 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4283 * In that cases, pages are freed continuously and we can expect pages
4284 * are in the same memcg. All these calls itself limits the number of
4285 * pages freed at once, then uncharge_start/end() is called properly.
4286 * This may be called prural(2) times in a context,
4289 void mem_cgroup_uncharge_start(void)
4291 current->memcg_batch.do_batch++;
4292 /* We can do nest. */
4293 if (current->memcg_batch.do_batch == 1) {
4294 current->memcg_batch.memcg = NULL;
4295 current->memcg_batch.nr_pages = 0;
4296 current->memcg_batch.memsw_nr_pages = 0;
4300 void mem_cgroup_uncharge_end(void)
4302 struct memcg_batch_info *batch = ¤t->memcg_batch;
4304 if (!batch->do_batch)
4308 if (batch->do_batch) /* If stacked, do nothing. */
4314 * This "batch->memcg" is valid without any css_get/put etc...
4315 * bacause we hide charges behind us.
4317 if (batch->nr_pages)
4318 res_counter_uncharge(&batch->memcg->res,
4319 batch->nr_pages * PAGE_SIZE);
4320 if (batch->memsw_nr_pages)
4321 res_counter_uncharge(&batch->memcg->memsw,
4322 batch->memsw_nr_pages * PAGE_SIZE);
4323 memcg_oom_recover(batch->memcg);
4324 /* forget this pointer (for sanity check) */
4325 batch->memcg = NULL;
4330 * called after __delete_from_swap_cache() and drop "page" account.
4331 * memcg information is recorded to swap_cgroup of "ent"
4334 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4336 struct mem_cgroup *memcg;
4337 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4339 if (!swapout) /* this was a swap cache but the swap is unused ! */
4340 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4342 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4345 * record memcg information, if swapout && memcg != NULL,
4346 * css_get() was called in uncharge().
4348 if (do_swap_account && swapout && memcg)
4349 swap_cgroup_record(ent, css_id(&memcg->css));
4353 #ifdef CONFIG_MEMCG_SWAP
4355 * called from swap_entry_free(). remove record in swap_cgroup and
4356 * uncharge "memsw" account.
4358 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4360 struct mem_cgroup *memcg;
4363 if (!do_swap_account)
4366 id = swap_cgroup_record(ent, 0);
4368 memcg = mem_cgroup_lookup(id);
4371 * We uncharge this because swap is freed.
4372 * This memcg can be obsolete one. We avoid calling css_tryget
4374 if (!mem_cgroup_is_root(memcg))
4375 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4376 mem_cgroup_swap_statistics(memcg, false);
4377 css_put(&memcg->css);
4383 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4384 * @entry: swap entry to be moved
4385 * @from: mem_cgroup which the entry is moved from
4386 * @to: mem_cgroup which the entry is moved to
4388 * It succeeds only when the swap_cgroup's record for this entry is the same
4389 * as the mem_cgroup's id of @from.
4391 * Returns 0 on success, -EINVAL on failure.
4393 * The caller must have charged to @to, IOW, called res_counter_charge() about
4394 * both res and memsw, and called css_get().
4396 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4397 struct mem_cgroup *from, struct mem_cgroup *to)
4399 unsigned short old_id, new_id;
4401 old_id = css_id(&from->css);
4402 new_id = css_id(&to->css);
4404 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4405 mem_cgroup_swap_statistics(from, false);
4406 mem_cgroup_swap_statistics(to, true);
4408 * This function is only called from task migration context now.
4409 * It postpones res_counter and refcount handling till the end
4410 * of task migration(mem_cgroup_clear_mc()) for performance
4411 * improvement. But we cannot postpone css_get(to) because if
4412 * the process that has been moved to @to does swap-in, the
4413 * refcount of @to might be decreased to 0.
4415 * We are in attach() phase, so the cgroup is guaranteed to be
4416 * alive, so we can just call css_get().
4424 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4425 struct mem_cgroup *from, struct mem_cgroup *to)
4432 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4435 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4436 struct mem_cgroup **memcgp)
4438 struct mem_cgroup *memcg = NULL;
4439 unsigned int nr_pages = 1;
4440 struct page_cgroup *pc;
4441 enum charge_type ctype;
4445 if (mem_cgroup_disabled())
4448 if (PageTransHuge(page))
4449 nr_pages <<= compound_order(page);
4451 pc = lookup_page_cgroup(page);
4452 lock_page_cgroup(pc);
4453 if (PageCgroupUsed(pc)) {
4454 memcg = pc->mem_cgroup;
4455 css_get(&memcg->css);
4457 * At migrating an anonymous page, its mapcount goes down
4458 * to 0 and uncharge() will be called. But, even if it's fully
4459 * unmapped, migration may fail and this page has to be
4460 * charged again. We set MIGRATION flag here and delay uncharge
4461 * until end_migration() is called
4463 * Corner Case Thinking
4465 * When the old page was mapped as Anon and it's unmap-and-freed
4466 * while migration was ongoing.
4467 * If unmap finds the old page, uncharge() of it will be delayed
4468 * until end_migration(). If unmap finds a new page, it's
4469 * uncharged when it make mapcount to be 1->0. If unmap code
4470 * finds swap_migration_entry, the new page will not be mapped
4471 * and end_migration() will find it(mapcount==0).
4474 * When the old page was mapped but migraion fails, the kernel
4475 * remaps it. A charge for it is kept by MIGRATION flag even
4476 * if mapcount goes down to 0. We can do remap successfully
4477 * without charging it again.
4480 * The "old" page is under lock_page() until the end of
4481 * migration, so, the old page itself will not be swapped-out.
4482 * If the new page is swapped out before end_migraton, our
4483 * hook to usual swap-out path will catch the event.
4486 SetPageCgroupMigration(pc);
4488 unlock_page_cgroup(pc);
4490 * If the page is not charged at this point,
4498 * We charge new page before it's used/mapped. So, even if unlock_page()
4499 * is called before end_migration, we can catch all events on this new
4500 * page. In the case new page is migrated but not remapped, new page's
4501 * mapcount will be finally 0 and we call uncharge in end_migration().
4504 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4506 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4508 * The page is committed to the memcg, but it's not actually
4509 * charged to the res_counter since we plan on replacing the
4510 * old one and only one page is going to be left afterwards.
4512 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4515 /* remove redundant charge if migration failed*/
4516 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4517 struct page *oldpage, struct page *newpage, bool migration_ok)
4519 struct page *used, *unused;
4520 struct page_cgroup *pc;
4526 if (!migration_ok) {
4533 anon = PageAnon(used);
4534 __mem_cgroup_uncharge_common(unused,
4535 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4536 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4538 css_put(&memcg->css);
4540 * We disallowed uncharge of pages under migration because mapcount
4541 * of the page goes down to zero, temporarly.
4542 * Clear the flag and check the page should be charged.
4544 pc = lookup_page_cgroup(oldpage);
4545 lock_page_cgroup(pc);
4546 ClearPageCgroupMigration(pc);
4547 unlock_page_cgroup(pc);
4550 * If a page is a file cache, radix-tree replacement is very atomic
4551 * and we can skip this check. When it was an Anon page, its mapcount
4552 * goes down to 0. But because we added MIGRATION flage, it's not
4553 * uncharged yet. There are several case but page->mapcount check
4554 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4555 * check. (see prepare_charge() also)
4558 mem_cgroup_uncharge_page(used);
4562 * At replace page cache, newpage is not under any memcg but it's on
4563 * LRU. So, this function doesn't touch res_counter but handles LRU
4564 * in correct way. Both pages are locked so we cannot race with uncharge.
4566 void mem_cgroup_replace_page_cache(struct page *oldpage,
4567 struct page *newpage)
4569 struct mem_cgroup *memcg = NULL;
4570 struct page_cgroup *pc;
4571 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4573 if (mem_cgroup_disabled())
4576 pc = lookup_page_cgroup(oldpage);
4577 /* fix accounting on old pages */
4578 lock_page_cgroup(pc);
4579 if (PageCgroupUsed(pc)) {
4580 memcg = pc->mem_cgroup;
4581 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4582 ClearPageCgroupUsed(pc);
4584 unlock_page_cgroup(pc);
4587 * When called from shmem_replace_page(), in some cases the
4588 * oldpage has already been charged, and in some cases not.
4593 * Even if newpage->mapping was NULL before starting replacement,
4594 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4595 * LRU while we overwrite pc->mem_cgroup.
4597 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4600 #ifdef CONFIG_DEBUG_VM
4601 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4603 struct page_cgroup *pc;
4605 pc = lookup_page_cgroup(page);
4607 * Can be NULL while feeding pages into the page allocator for
4608 * the first time, i.e. during boot or memory hotplug;
4609 * or when mem_cgroup_disabled().
4611 if (likely(pc) && PageCgroupUsed(pc))
4616 bool mem_cgroup_bad_page_check(struct page *page)
4618 if (mem_cgroup_disabled())
4621 return lookup_page_cgroup_used(page) != NULL;
4624 void mem_cgroup_print_bad_page(struct page *page)
4626 struct page_cgroup *pc;
4628 pc = lookup_page_cgroup_used(page);
4630 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4631 pc, pc->flags, pc->mem_cgroup);
4636 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4637 unsigned long long val)
4640 u64 memswlimit, memlimit;
4642 int children = mem_cgroup_count_children(memcg);
4643 u64 curusage, oldusage;
4647 * For keeping hierarchical_reclaim simple, how long we should retry
4648 * is depends on callers. We set our retry-count to be function
4649 * of # of children which we should visit in this loop.
4651 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4653 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4656 while (retry_count) {
4657 if (signal_pending(current)) {
4662 * Rather than hide all in some function, I do this in
4663 * open coded manner. You see what this really does.
4664 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4666 mutex_lock(&set_limit_mutex);
4667 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4668 if (memswlimit < val) {
4670 mutex_unlock(&set_limit_mutex);
4674 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4678 ret = res_counter_set_limit(&memcg->res, val);
4680 if (memswlimit == val)
4681 memcg->memsw_is_minimum = true;
4683 memcg->memsw_is_minimum = false;
4685 mutex_unlock(&set_limit_mutex);
4690 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4691 MEM_CGROUP_RECLAIM_SHRINK);
4692 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4693 /* Usage is reduced ? */
4694 if (curusage >= oldusage)
4697 oldusage = curusage;
4699 if (!ret && enlarge)
4700 memcg_oom_recover(memcg);
4705 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4706 unsigned long long val)
4709 u64 memlimit, memswlimit, oldusage, curusage;
4710 int children = mem_cgroup_count_children(memcg);
4714 /* see mem_cgroup_resize_res_limit */
4715 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4716 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4717 while (retry_count) {
4718 if (signal_pending(current)) {
4723 * Rather than hide all in some function, I do this in
4724 * open coded manner. You see what this really does.
4725 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4727 mutex_lock(&set_limit_mutex);
4728 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4729 if (memlimit > val) {
4731 mutex_unlock(&set_limit_mutex);
4734 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4735 if (memswlimit < val)
4737 ret = res_counter_set_limit(&memcg->memsw, val);
4739 if (memlimit == val)
4740 memcg->memsw_is_minimum = true;
4742 memcg->memsw_is_minimum = false;
4744 mutex_unlock(&set_limit_mutex);
4749 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4750 MEM_CGROUP_RECLAIM_NOSWAP |
4751 MEM_CGROUP_RECLAIM_SHRINK);
4752 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4753 /* Usage is reduced ? */
4754 if (curusage >= oldusage)
4757 oldusage = curusage;
4759 if (!ret && enlarge)
4760 memcg_oom_recover(memcg);
4765 * mem_cgroup_force_empty_list - clears LRU of a group
4766 * @memcg: group to clear
4769 * @lru: lru to to clear
4771 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4772 * reclaim the pages page themselves - pages are moved to the parent (or root)
4775 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4776 int node, int zid, enum lru_list lru)
4778 struct lruvec *lruvec;
4779 unsigned long flags;
4780 struct list_head *list;
4784 zone = &NODE_DATA(node)->node_zones[zid];
4785 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4786 list = &lruvec->lists[lru];
4790 struct page_cgroup *pc;
4793 spin_lock_irqsave(&zone->lru_lock, flags);
4794 if (list_empty(list)) {
4795 spin_unlock_irqrestore(&zone->lru_lock, flags);
4798 page = list_entry(list->prev, struct page, lru);
4800 list_move(&page->lru, list);
4802 spin_unlock_irqrestore(&zone->lru_lock, flags);
4805 spin_unlock_irqrestore(&zone->lru_lock, flags);
4807 pc = lookup_page_cgroup(page);
4809 if (mem_cgroup_move_parent(page, pc, memcg)) {
4810 /* found lock contention or "pc" is obsolete. */
4815 } while (!list_empty(list));
4819 * make mem_cgroup's charge to be 0 if there is no task by moving
4820 * all the charges and pages to the parent.
4821 * This enables deleting this mem_cgroup.
4823 * Caller is responsible for holding css reference on the memcg.
4825 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4831 /* This is for making all *used* pages to be on LRU. */
4832 lru_add_drain_all();
4833 drain_all_stock_sync(memcg);
4834 mem_cgroup_start_move(memcg);
4835 for_each_node_state(node, N_MEMORY) {
4836 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4839 mem_cgroup_force_empty_list(memcg,
4844 mem_cgroup_end_move(memcg);
4845 memcg_oom_recover(memcg);
4849 * Kernel memory may not necessarily be trackable to a specific
4850 * process. So they are not migrated, and therefore we can't
4851 * expect their value to drop to 0 here.
4852 * Having res filled up with kmem only is enough.
4854 * This is a safety check because mem_cgroup_force_empty_list
4855 * could have raced with mem_cgroup_replace_page_cache callers
4856 * so the lru seemed empty but the page could have been added
4857 * right after the check. RES_USAGE should be safe as we always
4858 * charge before adding to the LRU.
4860 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4861 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4862 } while (usage > 0);
4866 * This mainly exists for tests during the setting of set of use_hierarchy.
4867 * Since this is the very setting we are changing, the current hierarchy value
4870 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4872 struct cgroup_subsys_state *pos;
4874 /* bounce at first found */
4875 css_for_each_child(pos, &memcg->css)
4881 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4882 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4883 * from mem_cgroup_count_children(), in the sense that we don't really care how
4884 * many children we have; we only need to know if we have any. It also counts
4885 * any memcg without hierarchy as infertile.
4887 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4889 return memcg->use_hierarchy && __memcg_has_children(memcg);
4893 * Reclaims as many pages from the given memcg as possible and moves
4894 * the rest to the parent.
4896 * Caller is responsible for holding css reference for memcg.
4898 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4900 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4901 struct cgroup *cgrp = memcg->css.cgroup;
4903 /* returns EBUSY if there is a task or if we come here twice. */
4904 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4907 /* we call try-to-free pages for make this cgroup empty */
4908 lru_add_drain_all();
4909 /* try to free all pages in this cgroup */
4910 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4913 if (signal_pending(current))
4916 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4920 /* maybe some writeback is necessary */
4921 congestion_wait(BLK_RW_ASYNC, HZ/10);
4926 mem_cgroup_reparent_charges(memcg);
4931 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4934 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4936 if (mem_cgroup_is_root(memcg))
4938 return mem_cgroup_force_empty(memcg);
4941 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4944 return mem_cgroup_from_css(css)->use_hierarchy;
4947 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4948 struct cftype *cft, u64 val)
4951 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4952 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4954 mutex_lock(&memcg_create_mutex);
4956 if (memcg->use_hierarchy == val)
4960 * If parent's use_hierarchy is set, we can't make any modifications
4961 * in the child subtrees. If it is unset, then the change can
4962 * occur, provided the current cgroup has no children.
4964 * For the root cgroup, parent_mem is NULL, we allow value to be
4965 * set if there are no children.
4967 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4968 (val == 1 || val == 0)) {
4969 if (!__memcg_has_children(memcg))
4970 memcg->use_hierarchy = val;
4977 mutex_unlock(&memcg_create_mutex);
4983 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4984 enum mem_cgroup_stat_index idx)
4986 struct mem_cgroup *iter;
4989 /* Per-cpu values can be negative, use a signed accumulator */
4990 for_each_mem_cgroup_tree(iter, memcg)
4991 val += mem_cgroup_read_stat(iter, idx);
4993 if (val < 0) /* race ? */
4998 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5002 if (!mem_cgroup_is_root(memcg)) {
5004 return res_counter_read_u64(&memcg->res, RES_USAGE);
5006 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5010 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5011 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5013 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5014 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5017 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5019 return val << PAGE_SHIFT;
5022 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
5023 struct cftype *cft, struct file *file,
5024 char __user *buf, size_t nbytes, loff_t *ppos)
5026 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5032 type = MEMFILE_TYPE(cft->private);
5033 name = MEMFILE_ATTR(cft->private);
5037 if (name == RES_USAGE)
5038 val = mem_cgroup_usage(memcg, false);
5040 val = res_counter_read_u64(&memcg->res, name);
5043 if (name == RES_USAGE)
5044 val = mem_cgroup_usage(memcg, true);
5046 val = res_counter_read_u64(&memcg->memsw, name);
5049 val = res_counter_read_u64(&memcg->kmem, name);
5055 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5056 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5059 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
5062 #ifdef CONFIG_MEMCG_KMEM
5063 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5065 * For simplicity, we won't allow this to be disabled. It also can't
5066 * be changed if the cgroup has children already, or if tasks had
5069 * If tasks join before we set the limit, a person looking at
5070 * kmem.usage_in_bytes will have no way to determine when it took
5071 * place, which makes the value quite meaningless.
5073 * After it first became limited, changes in the value of the limit are
5074 * of course permitted.
5076 mutex_lock(&memcg_create_mutex);
5077 mutex_lock(&set_limit_mutex);
5078 if (!memcg->kmem_account_flags && val != RES_COUNTER_MAX) {
5079 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
5083 ret = res_counter_set_limit(&memcg->kmem, val);
5086 ret = memcg_update_cache_sizes(memcg);
5088 res_counter_set_limit(&memcg->kmem, RES_COUNTER_MAX);
5091 static_key_slow_inc(&memcg_kmem_enabled_key);
5093 * setting the active bit after the inc will guarantee no one
5094 * starts accounting before all call sites are patched
5096 memcg_kmem_set_active(memcg);
5098 ret = res_counter_set_limit(&memcg->kmem, val);
5100 mutex_unlock(&set_limit_mutex);
5101 mutex_unlock(&memcg_create_mutex);
5106 #ifdef CONFIG_MEMCG_KMEM
5107 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5110 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5114 memcg->kmem_account_flags = parent->kmem_account_flags;
5116 * When that happen, we need to disable the static branch only on those
5117 * memcgs that enabled it. To achieve this, we would be forced to
5118 * complicate the code by keeping track of which memcgs were the ones
5119 * that actually enabled limits, and which ones got it from its
5122 * It is a lot simpler just to do static_key_slow_inc() on every child
5123 * that is accounted.
5125 if (!memcg_kmem_is_active(memcg))
5129 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5130 * memcg is active already. If the later initialization fails then the
5131 * cgroup core triggers the cleanup so we do not have to do it here.
5133 static_key_slow_inc(&memcg_kmem_enabled_key);
5135 mutex_lock(&set_limit_mutex);
5136 memcg_stop_kmem_account();
5137 ret = memcg_update_cache_sizes(memcg);
5138 memcg_resume_kmem_account();
5139 mutex_unlock(&set_limit_mutex);
5143 #endif /* CONFIG_MEMCG_KMEM */
5146 * The user of this function is...
5149 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5152 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5155 unsigned long long val;
5158 type = MEMFILE_TYPE(cft->private);
5159 name = MEMFILE_ATTR(cft->private);
5163 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5167 /* This function does all necessary parse...reuse it */
5168 ret = res_counter_memparse_write_strategy(buffer, &val);
5172 ret = mem_cgroup_resize_limit(memcg, val);
5173 else if (type == _MEMSWAP)
5174 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5175 else if (type == _KMEM)
5176 ret = memcg_update_kmem_limit(css, val);
5180 case RES_SOFT_LIMIT:
5181 ret = res_counter_memparse_write_strategy(buffer, &val);
5185 * For memsw, soft limits are hard to implement in terms
5186 * of semantics, for now, we support soft limits for
5187 * control without swap
5190 ret = res_counter_set_soft_limit(&memcg->res, val);
5195 ret = -EINVAL; /* should be BUG() ? */
5201 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5202 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5204 unsigned long long min_limit, min_memsw_limit, tmp;
5206 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5207 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5208 if (!memcg->use_hierarchy)
5211 while (css_parent(&memcg->css)) {
5212 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5213 if (!memcg->use_hierarchy)
5215 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5216 min_limit = min(min_limit, tmp);
5217 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5218 min_memsw_limit = min(min_memsw_limit, tmp);
5221 *mem_limit = min_limit;
5222 *memsw_limit = min_memsw_limit;
5225 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5227 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5231 type = MEMFILE_TYPE(event);
5232 name = MEMFILE_ATTR(event);
5237 res_counter_reset_max(&memcg->res);
5238 else if (type == _MEMSWAP)
5239 res_counter_reset_max(&memcg->memsw);
5240 else if (type == _KMEM)
5241 res_counter_reset_max(&memcg->kmem);
5247 res_counter_reset_failcnt(&memcg->res);
5248 else if (type == _MEMSWAP)
5249 res_counter_reset_failcnt(&memcg->memsw);
5250 else if (type == _KMEM)
5251 res_counter_reset_failcnt(&memcg->kmem);
5260 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5263 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5267 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5268 struct cftype *cft, u64 val)
5270 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5272 if (val >= (1 << NR_MOVE_TYPE))
5276 * No kind of locking is needed in here, because ->can_attach() will
5277 * check this value once in the beginning of the process, and then carry
5278 * on with stale data. This means that changes to this value will only
5279 * affect task migrations starting after the change.
5281 memcg->move_charge_at_immigrate = val;
5285 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5286 struct cftype *cft, u64 val)
5293 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5294 struct cftype *cft, struct seq_file *m)
5297 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5298 unsigned long node_nr;
5299 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5301 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5302 seq_printf(m, "total=%lu", total_nr);
5303 for_each_node_state(nid, N_MEMORY) {
5304 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5305 seq_printf(m, " N%d=%lu", nid, node_nr);
5309 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5310 seq_printf(m, "file=%lu", file_nr);
5311 for_each_node_state(nid, N_MEMORY) {
5312 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5314 seq_printf(m, " N%d=%lu", nid, node_nr);
5318 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5319 seq_printf(m, "anon=%lu", anon_nr);
5320 for_each_node_state(nid, N_MEMORY) {
5321 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5323 seq_printf(m, " N%d=%lu", nid, node_nr);
5327 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5328 seq_printf(m, "unevictable=%lu", unevictable_nr);
5329 for_each_node_state(nid, N_MEMORY) {
5330 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5331 BIT(LRU_UNEVICTABLE));
5332 seq_printf(m, " N%d=%lu", nid, node_nr);
5337 #endif /* CONFIG_NUMA */
5339 static inline void mem_cgroup_lru_names_not_uptodate(void)
5341 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5344 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5347 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5348 struct mem_cgroup *mi;
5351 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5352 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5354 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5355 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5358 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5359 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5360 mem_cgroup_read_events(memcg, i));
5362 for (i = 0; i < NR_LRU_LISTS; i++)
5363 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5364 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5366 /* Hierarchical information */
5368 unsigned long long limit, memsw_limit;
5369 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5370 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5371 if (do_swap_account)
5372 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5376 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5379 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5381 for_each_mem_cgroup_tree(mi, memcg)
5382 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5383 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5386 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5387 unsigned long long val = 0;
5389 for_each_mem_cgroup_tree(mi, memcg)
5390 val += mem_cgroup_read_events(mi, i);
5391 seq_printf(m, "total_%s %llu\n",
5392 mem_cgroup_events_names[i], val);
5395 for (i = 0; i < NR_LRU_LISTS; i++) {
5396 unsigned long long val = 0;
5398 for_each_mem_cgroup_tree(mi, memcg)
5399 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5400 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5403 #ifdef CONFIG_DEBUG_VM
5406 struct mem_cgroup_per_zone *mz;
5407 struct zone_reclaim_stat *rstat;
5408 unsigned long recent_rotated[2] = {0, 0};
5409 unsigned long recent_scanned[2] = {0, 0};
5411 for_each_online_node(nid)
5412 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5413 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5414 rstat = &mz->lruvec.reclaim_stat;
5416 recent_rotated[0] += rstat->recent_rotated[0];
5417 recent_rotated[1] += rstat->recent_rotated[1];
5418 recent_scanned[0] += rstat->recent_scanned[0];
5419 recent_scanned[1] += rstat->recent_scanned[1];
5421 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5422 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5423 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5424 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5431 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5434 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5436 return mem_cgroup_swappiness(memcg);
5439 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5440 struct cftype *cft, u64 val)
5442 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5443 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5445 if (val > 100 || !parent)
5448 mutex_lock(&memcg_create_mutex);
5450 /* If under hierarchy, only empty-root can set this value */
5451 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5452 mutex_unlock(&memcg_create_mutex);
5456 memcg->swappiness = val;
5458 mutex_unlock(&memcg_create_mutex);
5463 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5465 struct mem_cgroup_threshold_ary *t;
5471 t = rcu_dereference(memcg->thresholds.primary);
5473 t = rcu_dereference(memcg->memsw_thresholds.primary);
5478 usage = mem_cgroup_usage(memcg, swap);
5481 * current_threshold points to threshold just below or equal to usage.
5482 * If it's not true, a threshold was crossed after last
5483 * call of __mem_cgroup_threshold().
5485 i = t->current_threshold;
5488 * Iterate backward over array of thresholds starting from
5489 * current_threshold and check if a threshold is crossed.
5490 * If none of thresholds below usage is crossed, we read
5491 * only one element of the array here.
5493 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5494 eventfd_signal(t->entries[i].eventfd, 1);
5496 /* i = current_threshold + 1 */
5500 * Iterate forward over array of thresholds starting from
5501 * current_threshold+1 and check if a threshold is crossed.
5502 * If none of thresholds above usage is crossed, we read
5503 * only one element of the array here.
5505 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5506 eventfd_signal(t->entries[i].eventfd, 1);
5508 /* Update current_threshold */
5509 t->current_threshold = i - 1;
5514 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5517 __mem_cgroup_threshold(memcg, false);
5518 if (do_swap_account)
5519 __mem_cgroup_threshold(memcg, true);
5521 memcg = parent_mem_cgroup(memcg);
5525 static int compare_thresholds(const void *a, const void *b)
5527 const struct mem_cgroup_threshold *_a = a;
5528 const struct mem_cgroup_threshold *_b = b;
5530 if (_a->threshold > _b->threshold)
5533 if (_a->threshold < _b->threshold)
5539 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5541 struct mem_cgroup_eventfd_list *ev;
5543 list_for_each_entry(ev, &memcg->oom_notify, list)
5544 eventfd_signal(ev->eventfd, 1);
5548 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5550 struct mem_cgroup *iter;
5552 for_each_mem_cgroup_tree(iter, memcg)
5553 mem_cgroup_oom_notify_cb(iter);
5556 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5557 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5559 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5560 struct mem_cgroup_thresholds *thresholds;
5561 struct mem_cgroup_threshold_ary *new;
5562 enum res_type type = MEMFILE_TYPE(cft->private);
5563 u64 threshold, usage;
5566 ret = res_counter_memparse_write_strategy(args, &threshold);
5570 mutex_lock(&memcg->thresholds_lock);
5573 thresholds = &memcg->thresholds;
5574 else if (type == _MEMSWAP)
5575 thresholds = &memcg->memsw_thresholds;
5579 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5581 /* Check if a threshold crossed before adding a new one */
5582 if (thresholds->primary)
5583 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5585 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5587 /* Allocate memory for new array of thresholds */
5588 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5596 /* Copy thresholds (if any) to new array */
5597 if (thresholds->primary) {
5598 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5599 sizeof(struct mem_cgroup_threshold));
5602 /* Add new threshold */
5603 new->entries[size - 1].eventfd = eventfd;
5604 new->entries[size - 1].threshold = threshold;
5606 /* Sort thresholds. Registering of new threshold isn't time-critical */
5607 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5608 compare_thresholds, NULL);
5610 /* Find current threshold */
5611 new->current_threshold = -1;
5612 for (i = 0; i < size; i++) {
5613 if (new->entries[i].threshold <= usage) {
5615 * new->current_threshold will not be used until
5616 * rcu_assign_pointer(), so it's safe to increment
5619 ++new->current_threshold;
5624 /* Free old spare buffer and save old primary buffer as spare */
5625 kfree(thresholds->spare);
5626 thresholds->spare = thresholds->primary;
5628 rcu_assign_pointer(thresholds->primary, new);
5630 /* To be sure that nobody uses thresholds */
5634 mutex_unlock(&memcg->thresholds_lock);
5639 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5640 struct cftype *cft, struct eventfd_ctx *eventfd)
5642 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5643 struct mem_cgroup_thresholds *thresholds;
5644 struct mem_cgroup_threshold_ary *new;
5645 enum res_type type = MEMFILE_TYPE(cft->private);
5649 mutex_lock(&memcg->thresholds_lock);
5651 thresholds = &memcg->thresholds;
5652 else if (type == _MEMSWAP)
5653 thresholds = &memcg->memsw_thresholds;
5657 if (!thresholds->primary)
5660 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5662 /* Check if a threshold crossed before removing */
5663 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5665 /* Calculate new number of threshold */
5667 for (i = 0; i < thresholds->primary->size; i++) {
5668 if (thresholds->primary->entries[i].eventfd != eventfd)
5672 new = thresholds->spare;
5674 /* Set thresholds array to NULL if we don't have thresholds */
5683 /* Copy thresholds and find current threshold */
5684 new->current_threshold = -1;
5685 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5686 if (thresholds->primary->entries[i].eventfd == eventfd)
5689 new->entries[j] = thresholds->primary->entries[i];
5690 if (new->entries[j].threshold <= usage) {
5692 * new->current_threshold will not be used
5693 * until rcu_assign_pointer(), so it's safe to increment
5696 ++new->current_threshold;
5702 /* Swap primary and spare array */
5703 thresholds->spare = thresholds->primary;
5704 /* If all events are unregistered, free the spare array */
5706 kfree(thresholds->spare);
5707 thresholds->spare = NULL;
5710 rcu_assign_pointer(thresholds->primary, new);
5712 /* To be sure that nobody uses thresholds */
5715 mutex_unlock(&memcg->thresholds_lock);
5718 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5719 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5721 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5722 struct mem_cgroup_eventfd_list *event;
5723 enum res_type type = MEMFILE_TYPE(cft->private);
5725 BUG_ON(type != _OOM_TYPE);
5726 event = kmalloc(sizeof(*event), GFP_KERNEL);
5730 spin_lock(&memcg_oom_lock);
5732 event->eventfd = eventfd;
5733 list_add(&event->list, &memcg->oom_notify);
5735 /* already in OOM ? */
5736 if (atomic_read(&memcg->under_oom))
5737 eventfd_signal(eventfd, 1);
5738 spin_unlock(&memcg_oom_lock);
5743 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5744 struct cftype *cft, struct eventfd_ctx *eventfd)
5746 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5747 struct mem_cgroup_eventfd_list *ev, *tmp;
5748 enum res_type type = MEMFILE_TYPE(cft->private);
5750 BUG_ON(type != _OOM_TYPE);
5752 spin_lock(&memcg_oom_lock);
5754 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5755 if (ev->eventfd == eventfd) {
5756 list_del(&ev->list);
5761 spin_unlock(&memcg_oom_lock);
5764 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5765 struct cftype *cft, struct cgroup_map_cb *cb)
5767 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5769 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5771 if (atomic_read(&memcg->under_oom))
5772 cb->fill(cb, "under_oom", 1);
5774 cb->fill(cb, "under_oom", 0);
5778 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5779 struct cftype *cft, u64 val)
5781 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5782 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5784 /* cannot set to root cgroup and only 0 and 1 are allowed */
5785 if (!parent || !((val == 0) || (val == 1)))
5788 mutex_lock(&memcg_create_mutex);
5789 /* oom-kill-disable is a flag for subhierarchy. */
5790 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5791 mutex_unlock(&memcg_create_mutex);
5794 memcg->oom_kill_disable = val;
5796 memcg_oom_recover(memcg);
5797 mutex_unlock(&memcg_create_mutex);
5801 #ifdef CONFIG_MEMCG_KMEM
5802 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5806 memcg->kmemcg_id = -1;
5807 ret = memcg_propagate_kmem(memcg);
5811 return mem_cgroup_sockets_init(memcg, ss);
5814 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5816 mem_cgroup_sockets_destroy(memcg);
5819 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5821 if (!memcg_kmem_is_active(memcg))
5825 * kmem charges can outlive the cgroup. In the case of slab
5826 * pages, for instance, a page contain objects from various
5827 * processes. As we prevent from taking a reference for every
5828 * such allocation we have to be careful when doing uncharge
5829 * (see memcg_uncharge_kmem) and here during offlining.
5831 * The idea is that that only the _last_ uncharge which sees
5832 * the dead memcg will drop the last reference. An additional
5833 * reference is taken here before the group is marked dead
5834 * which is then paired with css_put during uncharge resp. here.
5836 * Although this might sound strange as this path is called from
5837 * css_offline() when the referencemight have dropped down to 0
5838 * and shouldn't be incremented anymore (css_tryget would fail)
5839 * we do not have other options because of the kmem allocations
5842 css_get(&memcg->css);
5844 memcg_kmem_mark_dead(memcg);
5846 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5849 if (memcg_kmem_test_and_clear_dead(memcg))
5850 css_put(&memcg->css);
5853 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5858 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5862 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5867 static struct cftype mem_cgroup_files[] = {
5869 .name = "usage_in_bytes",
5870 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5871 .read = mem_cgroup_read,
5872 .register_event = mem_cgroup_usage_register_event,
5873 .unregister_event = mem_cgroup_usage_unregister_event,
5876 .name = "max_usage_in_bytes",
5877 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5878 .trigger = mem_cgroup_reset,
5879 .read = mem_cgroup_read,
5882 .name = "limit_in_bytes",
5883 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5884 .write_string = mem_cgroup_write,
5885 .read = mem_cgroup_read,
5888 .name = "soft_limit_in_bytes",
5889 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5890 .write_string = mem_cgroup_write,
5891 .read = mem_cgroup_read,
5895 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5896 .trigger = mem_cgroup_reset,
5897 .read = mem_cgroup_read,
5901 .read_seq_string = memcg_stat_show,
5904 .name = "force_empty",
5905 .trigger = mem_cgroup_force_empty_write,
5908 .name = "use_hierarchy",
5909 .flags = CFTYPE_INSANE,
5910 .write_u64 = mem_cgroup_hierarchy_write,
5911 .read_u64 = mem_cgroup_hierarchy_read,
5914 .name = "swappiness",
5915 .read_u64 = mem_cgroup_swappiness_read,
5916 .write_u64 = mem_cgroup_swappiness_write,
5919 .name = "move_charge_at_immigrate",
5920 .read_u64 = mem_cgroup_move_charge_read,
5921 .write_u64 = mem_cgroup_move_charge_write,
5924 .name = "oom_control",
5925 .read_map = mem_cgroup_oom_control_read,
5926 .write_u64 = mem_cgroup_oom_control_write,
5927 .register_event = mem_cgroup_oom_register_event,
5928 .unregister_event = mem_cgroup_oom_unregister_event,
5929 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5932 .name = "pressure_level",
5933 .register_event = vmpressure_register_event,
5934 .unregister_event = vmpressure_unregister_event,
5938 .name = "numa_stat",
5939 .read_seq_string = memcg_numa_stat_show,
5942 #ifdef CONFIG_MEMCG_KMEM
5944 .name = "kmem.limit_in_bytes",
5945 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5946 .write_string = mem_cgroup_write,
5947 .read = mem_cgroup_read,
5950 .name = "kmem.usage_in_bytes",
5951 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5952 .read = mem_cgroup_read,
5955 .name = "kmem.failcnt",
5956 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5957 .trigger = mem_cgroup_reset,
5958 .read = mem_cgroup_read,
5961 .name = "kmem.max_usage_in_bytes",
5962 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5963 .trigger = mem_cgroup_reset,
5964 .read = mem_cgroup_read,
5966 #ifdef CONFIG_SLABINFO
5968 .name = "kmem.slabinfo",
5969 .read_seq_string = mem_cgroup_slabinfo_read,
5973 { }, /* terminate */
5976 #ifdef CONFIG_MEMCG_SWAP
5977 static struct cftype memsw_cgroup_files[] = {
5979 .name = "memsw.usage_in_bytes",
5980 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5981 .read = mem_cgroup_read,
5982 .register_event = mem_cgroup_usage_register_event,
5983 .unregister_event = mem_cgroup_usage_unregister_event,
5986 .name = "memsw.max_usage_in_bytes",
5987 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5988 .trigger = mem_cgroup_reset,
5989 .read = mem_cgroup_read,
5992 .name = "memsw.limit_in_bytes",
5993 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5994 .write_string = mem_cgroup_write,
5995 .read = mem_cgroup_read,
5998 .name = "memsw.failcnt",
5999 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6000 .trigger = mem_cgroup_reset,
6001 .read = mem_cgroup_read,
6003 { }, /* terminate */
6006 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6008 struct mem_cgroup_per_node *pn;
6009 struct mem_cgroup_per_zone *mz;
6010 int zone, tmp = node;
6012 * This routine is called against possible nodes.
6013 * But it's BUG to call kmalloc() against offline node.
6015 * TODO: this routine can waste much memory for nodes which will
6016 * never be onlined. It's better to use memory hotplug callback
6019 if (!node_state(node, N_NORMAL_MEMORY))
6021 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6025 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6026 mz = &pn->zoneinfo[zone];
6027 lruvec_init(&mz->lruvec);
6028 mz->usage_in_excess = 0;
6029 mz->on_tree = false;
6032 memcg->nodeinfo[node] = pn;
6036 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6038 kfree(memcg->nodeinfo[node]);
6041 static struct mem_cgroup *mem_cgroup_alloc(void)
6043 struct mem_cgroup *memcg;
6044 size_t size = memcg_size();
6046 /* Can be very big if nr_node_ids is very big */
6047 if (size < PAGE_SIZE)
6048 memcg = kzalloc(size, GFP_KERNEL);
6050 memcg = vzalloc(size);
6055 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6058 spin_lock_init(&memcg->pcp_counter_lock);
6062 if (size < PAGE_SIZE)
6070 * At destroying mem_cgroup, references from swap_cgroup can remain.
6071 * (scanning all at force_empty is too costly...)
6073 * Instead of clearing all references at force_empty, we remember
6074 * the number of reference from swap_cgroup and free mem_cgroup when
6075 * it goes down to 0.
6077 * Removal of cgroup itself succeeds regardless of refs from swap.
6080 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6083 size_t size = memcg_size();
6085 mem_cgroup_remove_from_trees(memcg);
6086 free_css_id(&mem_cgroup_subsys, &memcg->css);
6089 free_mem_cgroup_per_zone_info(memcg, node);
6091 free_percpu(memcg->stat);
6094 * We need to make sure that (at least for now), the jump label
6095 * destruction code runs outside of the cgroup lock. This is because
6096 * get_online_cpus(), which is called from the static_branch update,
6097 * can't be called inside the cgroup_lock. cpusets are the ones
6098 * enforcing this dependency, so if they ever change, we might as well.
6100 * schedule_work() will guarantee this happens. Be careful if you need
6101 * to move this code around, and make sure it is outside
6104 disarm_static_keys(memcg);
6105 if (size < PAGE_SIZE)
6112 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6114 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6116 if (!memcg->res.parent)
6118 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6120 EXPORT_SYMBOL(parent_mem_cgroup);
6122 static void __init mem_cgroup_soft_limit_tree_init(void)
6124 struct mem_cgroup_tree_per_node *rtpn;
6125 struct mem_cgroup_tree_per_zone *rtpz;
6126 int tmp, node, zone;
6128 for_each_node(node) {
6130 if (!node_state(node, N_NORMAL_MEMORY))
6132 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6135 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6137 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6138 rtpz = &rtpn->rb_tree_per_zone[zone];
6139 rtpz->rb_root = RB_ROOT;
6140 spin_lock_init(&rtpz->lock);
6145 static struct cgroup_subsys_state * __ref
6146 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6148 struct mem_cgroup *memcg;
6149 long error = -ENOMEM;
6152 memcg = mem_cgroup_alloc();
6154 return ERR_PTR(error);
6157 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6161 if (parent_css == NULL) {
6162 root_mem_cgroup = memcg;
6163 res_counter_init(&memcg->res, NULL);
6164 res_counter_init(&memcg->memsw, NULL);
6165 res_counter_init(&memcg->kmem, NULL);
6168 memcg->last_scanned_node = MAX_NUMNODES;
6169 INIT_LIST_HEAD(&memcg->oom_notify);
6170 memcg->move_charge_at_immigrate = 0;
6171 mutex_init(&memcg->thresholds_lock);
6172 spin_lock_init(&memcg->move_lock);
6173 vmpressure_init(&memcg->vmpressure);
6178 __mem_cgroup_free(memcg);
6179 return ERR_PTR(error);
6183 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6185 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6186 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6192 mutex_lock(&memcg_create_mutex);
6194 memcg->use_hierarchy = parent->use_hierarchy;
6195 memcg->oom_kill_disable = parent->oom_kill_disable;
6196 memcg->swappiness = mem_cgroup_swappiness(parent);
6198 if (parent->use_hierarchy) {
6199 res_counter_init(&memcg->res, &parent->res);
6200 res_counter_init(&memcg->memsw, &parent->memsw);
6201 res_counter_init(&memcg->kmem, &parent->kmem);
6204 * No need to take a reference to the parent because cgroup
6205 * core guarantees its existence.
6208 res_counter_init(&memcg->res, NULL);
6209 res_counter_init(&memcg->memsw, NULL);
6210 res_counter_init(&memcg->kmem, NULL);
6212 * Deeper hierachy with use_hierarchy == false doesn't make
6213 * much sense so let cgroup subsystem know about this
6214 * unfortunate state in our controller.
6216 if (parent != root_mem_cgroup)
6217 mem_cgroup_subsys.broken_hierarchy = true;
6220 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6221 mutex_unlock(&memcg_create_mutex);
6226 * Announce all parents that a group from their hierarchy is gone.
6228 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6230 struct mem_cgroup *parent = memcg;
6232 while ((parent = parent_mem_cgroup(parent)))
6233 mem_cgroup_iter_invalidate(parent);
6236 * if the root memcg is not hierarchical we have to check it
6239 if (!root_mem_cgroup->use_hierarchy)
6240 mem_cgroup_iter_invalidate(root_mem_cgroup);
6243 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6245 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6247 kmem_cgroup_css_offline(memcg);
6249 mem_cgroup_invalidate_reclaim_iterators(memcg);
6250 mem_cgroup_reparent_charges(memcg);
6251 mem_cgroup_destroy_all_caches(memcg);
6252 vmpressure_cleanup(&memcg->vmpressure);
6255 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6257 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6259 memcg_destroy_kmem(memcg);
6260 __mem_cgroup_free(memcg);
6264 /* Handlers for move charge at task migration. */
6265 #define PRECHARGE_COUNT_AT_ONCE 256
6266 static int mem_cgroup_do_precharge(unsigned long count)
6269 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6270 struct mem_cgroup *memcg = mc.to;
6272 if (mem_cgroup_is_root(memcg)) {
6273 mc.precharge += count;
6274 /* we don't need css_get for root */
6277 /* try to charge at once */
6279 struct res_counter *dummy;
6281 * "memcg" cannot be under rmdir() because we've already checked
6282 * by cgroup_lock_live_cgroup() that it is not removed and we
6283 * are still under the same cgroup_mutex. So we can postpone
6286 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6288 if (do_swap_account && res_counter_charge(&memcg->memsw,
6289 PAGE_SIZE * count, &dummy)) {
6290 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6293 mc.precharge += count;
6297 /* fall back to one by one charge */
6299 if (signal_pending(current)) {
6303 if (!batch_count--) {
6304 batch_count = PRECHARGE_COUNT_AT_ONCE;
6307 ret = __mem_cgroup_try_charge(NULL,
6308 GFP_KERNEL, 1, &memcg, false);
6310 /* mem_cgroup_clear_mc() will do uncharge later */
6318 * get_mctgt_type - get target type of moving charge
6319 * @vma: the vma the pte to be checked belongs
6320 * @addr: the address corresponding to the pte to be checked
6321 * @ptent: the pte to be checked
6322 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6325 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6326 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6327 * move charge. if @target is not NULL, the page is stored in target->page
6328 * with extra refcnt got(Callers should handle it).
6329 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6330 * target for charge migration. if @target is not NULL, the entry is stored
6333 * Called with pte lock held.
6340 enum mc_target_type {
6346 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6347 unsigned long addr, pte_t ptent)
6349 struct page *page = vm_normal_page(vma, addr, ptent);
6351 if (!page || !page_mapped(page))
6353 if (PageAnon(page)) {
6354 /* we don't move shared anon */
6357 } else if (!move_file())
6358 /* we ignore mapcount for file pages */
6360 if (!get_page_unless_zero(page))
6367 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6368 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6370 struct page *page = NULL;
6371 swp_entry_t ent = pte_to_swp_entry(ptent);
6373 if (!move_anon() || non_swap_entry(ent))
6376 * Because lookup_swap_cache() updates some statistics counter,
6377 * we call find_get_page() with swapper_space directly.
6379 page = find_get_page(swap_address_space(ent), ent.val);
6380 if (do_swap_account)
6381 entry->val = ent.val;
6386 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6387 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6393 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6394 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6396 struct page *page = NULL;
6397 struct address_space *mapping;
6400 if (!vma->vm_file) /* anonymous vma */
6405 mapping = vma->vm_file->f_mapping;
6406 if (pte_none(ptent))
6407 pgoff = linear_page_index(vma, addr);
6408 else /* pte_file(ptent) is true */
6409 pgoff = pte_to_pgoff(ptent);
6411 /* page is moved even if it's not RSS of this task(page-faulted). */
6412 page = find_get_page(mapping, pgoff);
6415 /* shmem/tmpfs may report page out on swap: account for that too. */
6416 if (radix_tree_exceptional_entry(page)) {
6417 swp_entry_t swap = radix_to_swp_entry(page);
6418 if (do_swap_account)
6420 page = find_get_page(swap_address_space(swap), swap.val);
6426 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6427 unsigned long addr, pte_t ptent, union mc_target *target)
6429 struct page *page = NULL;
6430 struct page_cgroup *pc;
6431 enum mc_target_type ret = MC_TARGET_NONE;
6432 swp_entry_t ent = { .val = 0 };
6434 if (pte_present(ptent))
6435 page = mc_handle_present_pte(vma, addr, ptent);
6436 else if (is_swap_pte(ptent))
6437 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6438 else if (pte_none(ptent) || pte_file(ptent))
6439 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6441 if (!page && !ent.val)
6444 pc = lookup_page_cgroup(page);
6446 * Do only loose check w/o page_cgroup lock.
6447 * mem_cgroup_move_account() checks the pc is valid or not under
6450 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6451 ret = MC_TARGET_PAGE;
6453 target->page = page;
6455 if (!ret || !target)
6458 /* There is a swap entry and a page doesn't exist or isn't charged */
6459 if (ent.val && !ret &&
6460 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6461 ret = MC_TARGET_SWAP;
6468 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6470 * We don't consider swapping or file mapped pages because THP does not
6471 * support them for now.
6472 * Caller should make sure that pmd_trans_huge(pmd) is true.
6474 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6475 unsigned long addr, pmd_t pmd, union mc_target *target)
6477 struct page *page = NULL;
6478 struct page_cgroup *pc;
6479 enum mc_target_type ret = MC_TARGET_NONE;
6481 page = pmd_page(pmd);
6482 VM_BUG_ON(!page || !PageHead(page));
6485 pc = lookup_page_cgroup(page);
6486 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6487 ret = MC_TARGET_PAGE;
6490 target->page = page;
6496 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6497 unsigned long addr, pmd_t pmd, union mc_target *target)
6499 return MC_TARGET_NONE;
6503 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6504 unsigned long addr, unsigned long end,
6505 struct mm_walk *walk)
6507 struct vm_area_struct *vma = walk->private;
6511 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6512 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6513 mc.precharge += HPAGE_PMD_NR;
6514 spin_unlock(&vma->vm_mm->page_table_lock);
6518 if (pmd_trans_unstable(pmd))
6520 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6521 for (; addr != end; pte++, addr += PAGE_SIZE)
6522 if (get_mctgt_type(vma, addr, *pte, NULL))
6523 mc.precharge++; /* increment precharge temporarily */
6524 pte_unmap_unlock(pte - 1, ptl);
6530 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6532 unsigned long precharge;
6533 struct vm_area_struct *vma;
6535 down_read(&mm->mmap_sem);
6536 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6537 struct mm_walk mem_cgroup_count_precharge_walk = {
6538 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6542 if (is_vm_hugetlb_page(vma))
6544 walk_page_range(vma->vm_start, vma->vm_end,
6545 &mem_cgroup_count_precharge_walk);
6547 up_read(&mm->mmap_sem);
6549 precharge = mc.precharge;
6555 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6557 unsigned long precharge = mem_cgroup_count_precharge(mm);
6559 VM_BUG_ON(mc.moving_task);
6560 mc.moving_task = current;
6561 return mem_cgroup_do_precharge(precharge);
6564 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6565 static void __mem_cgroup_clear_mc(void)
6567 struct mem_cgroup *from = mc.from;
6568 struct mem_cgroup *to = mc.to;
6571 /* we must uncharge all the leftover precharges from mc.to */
6573 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6577 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6578 * we must uncharge here.
6580 if (mc.moved_charge) {
6581 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6582 mc.moved_charge = 0;
6584 /* we must fixup refcnts and charges */
6585 if (mc.moved_swap) {
6586 /* uncharge swap account from the old cgroup */
6587 if (!mem_cgroup_is_root(mc.from))
6588 res_counter_uncharge(&mc.from->memsw,
6589 PAGE_SIZE * mc.moved_swap);
6591 for (i = 0; i < mc.moved_swap; i++)
6592 css_put(&mc.from->css);
6594 if (!mem_cgroup_is_root(mc.to)) {
6596 * we charged both to->res and to->memsw, so we should
6599 res_counter_uncharge(&mc.to->res,
6600 PAGE_SIZE * mc.moved_swap);
6602 /* we've already done css_get(mc.to) */
6605 memcg_oom_recover(from);
6606 memcg_oom_recover(to);
6607 wake_up_all(&mc.waitq);
6610 static void mem_cgroup_clear_mc(void)
6612 struct mem_cgroup *from = mc.from;
6615 * we must clear moving_task before waking up waiters at the end of
6618 mc.moving_task = NULL;
6619 __mem_cgroup_clear_mc();
6620 spin_lock(&mc.lock);
6623 spin_unlock(&mc.lock);
6624 mem_cgroup_end_move(from);
6627 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6628 struct cgroup_taskset *tset)
6630 struct task_struct *p = cgroup_taskset_first(tset);
6632 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6633 unsigned long move_charge_at_immigrate;
6636 * We are now commited to this value whatever it is. Changes in this
6637 * tunable will only affect upcoming migrations, not the current one.
6638 * So we need to save it, and keep it going.
6640 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6641 if (move_charge_at_immigrate) {
6642 struct mm_struct *mm;
6643 struct mem_cgroup *from = mem_cgroup_from_task(p);
6645 VM_BUG_ON(from == memcg);
6647 mm = get_task_mm(p);
6650 /* We move charges only when we move a owner of the mm */
6651 if (mm->owner == p) {
6654 VM_BUG_ON(mc.precharge);
6655 VM_BUG_ON(mc.moved_charge);
6656 VM_BUG_ON(mc.moved_swap);
6657 mem_cgroup_start_move(from);
6658 spin_lock(&mc.lock);
6661 mc.immigrate_flags = move_charge_at_immigrate;
6662 spin_unlock(&mc.lock);
6663 /* We set mc.moving_task later */
6665 ret = mem_cgroup_precharge_mc(mm);
6667 mem_cgroup_clear_mc();
6674 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6675 struct cgroup_taskset *tset)
6677 mem_cgroup_clear_mc();
6680 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6681 unsigned long addr, unsigned long end,
6682 struct mm_walk *walk)
6685 struct vm_area_struct *vma = walk->private;
6688 enum mc_target_type target_type;
6689 union mc_target target;
6691 struct page_cgroup *pc;
6694 * We don't take compound_lock() here but no race with splitting thp
6696 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6697 * under splitting, which means there's no concurrent thp split,
6698 * - if another thread runs into split_huge_page() just after we
6699 * entered this if-block, the thread must wait for page table lock
6700 * to be unlocked in __split_huge_page_splitting(), where the main
6701 * part of thp split is not executed yet.
6703 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6704 if (mc.precharge < HPAGE_PMD_NR) {
6705 spin_unlock(&vma->vm_mm->page_table_lock);
6708 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6709 if (target_type == MC_TARGET_PAGE) {
6711 if (!isolate_lru_page(page)) {
6712 pc = lookup_page_cgroup(page);
6713 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6714 pc, mc.from, mc.to)) {
6715 mc.precharge -= HPAGE_PMD_NR;
6716 mc.moved_charge += HPAGE_PMD_NR;
6718 putback_lru_page(page);
6722 spin_unlock(&vma->vm_mm->page_table_lock);
6726 if (pmd_trans_unstable(pmd))
6729 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6730 for (; addr != end; addr += PAGE_SIZE) {
6731 pte_t ptent = *(pte++);
6737 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6738 case MC_TARGET_PAGE:
6740 if (isolate_lru_page(page))
6742 pc = lookup_page_cgroup(page);
6743 if (!mem_cgroup_move_account(page, 1, pc,
6746 /* we uncharge from mc.from later. */
6749 putback_lru_page(page);
6750 put: /* get_mctgt_type() gets the page */
6753 case MC_TARGET_SWAP:
6755 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6757 /* we fixup refcnts and charges later. */
6765 pte_unmap_unlock(pte - 1, ptl);
6770 * We have consumed all precharges we got in can_attach().
6771 * We try charge one by one, but don't do any additional
6772 * charges to mc.to if we have failed in charge once in attach()
6775 ret = mem_cgroup_do_precharge(1);
6783 static void mem_cgroup_move_charge(struct mm_struct *mm)
6785 struct vm_area_struct *vma;
6787 lru_add_drain_all();
6789 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6791 * Someone who are holding the mmap_sem might be waiting in
6792 * waitq. So we cancel all extra charges, wake up all waiters,
6793 * and retry. Because we cancel precharges, we might not be able
6794 * to move enough charges, but moving charge is a best-effort
6795 * feature anyway, so it wouldn't be a big problem.
6797 __mem_cgroup_clear_mc();
6801 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6803 struct mm_walk mem_cgroup_move_charge_walk = {
6804 .pmd_entry = mem_cgroup_move_charge_pte_range,
6808 if (is_vm_hugetlb_page(vma))
6810 ret = walk_page_range(vma->vm_start, vma->vm_end,
6811 &mem_cgroup_move_charge_walk);
6814 * means we have consumed all precharges and failed in
6815 * doing additional charge. Just abandon here.
6819 up_read(&mm->mmap_sem);
6822 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6823 struct cgroup_taskset *tset)
6825 struct task_struct *p = cgroup_taskset_first(tset);
6826 struct mm_struct *mm = get_task_mm(p);
6830 mem_cgroup_move_charge(mm);
6834 mem_cgroup_clear_mc();
6836 #else /* !CONFIG_MMU */
6837 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6838 struct cgroup_taskset *tset)
6842 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6843 struct cgroup_taskset *tset)
6846 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6847 struct cgroup_taskset *tset)
6853 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6854 * to verify sane_behavior flag on each mount attempt.
6856 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6859 * use_hierarchy is forced with sane_behavior. cgroup core
6860 * guarantees that @root doesn't have any children, so turning it
6861 * on for the root memcg is enough.
6863 if (cgroup_sane_behavior(root_css->cgroup))
6864 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6867 struct cgroup_subsys mem_cgroup_subsys = {
6869 .subsys_id = mem_cgroup_subsys_id,
6870 .css_alloc = mem_cgroup_css_alloc,
6871 .css_online = mem_cgroup_css_online,
6872 .css_offline = mem_cgroup_css_offline,
6873 .css_free = mem_cgroup_css_free,
6874 .can_attach = mem_cgroup_can_attach,
6875 .cancel_attach = mem_cgroup_cancel_attach,
6876 .attach = mem_cgroup_move_task,
6877 .bind = mem_cgroup_bind,
6878 .base_cftypes = mem_cgroup_files,
6883 #ifdef CONFIG_MEMCG_SWAP
6884 static int __init enable_swap_account(char *s)
6886 if (!strcmp(s, "1"))
6887 really_do_swap_account = 1;
6888 else if (!strcmp(s, "0"))
6889 really_do_swap_account = 0;
6892 __setup("swapaccount=", enable_swap_account);
6894 static void __init memsw_file_init(void)
6896 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6899 static void __init enable_swap_cgroup(void)
6901 if (!mem_cgroup_disabled() && really_do_swap_account) {
6902 do_swap_account = 1;
6908 static void __init enable_swap_cgroup(void)
6914 * subsys_initcall() for memory controller.
6916 * Some parts like hotcpu_notifier() have to be initialized from this context
6917 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6918 * everything that doesn't depend on a specific mem_cgroup structure should
6919 * be initialized from here.
6921 static int __init mem_cgroup_init(void)
6923 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6924 enable_swap_cgroup();
6925 mem_cgroup_soft_limit_tree_init();
6929 subsys_initcall(mem_cgroup_init);