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
89 * Statistics for memory cgroup.
91 enum mem_cgroup_stat_index {
93 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
95 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
96 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
97 MEM_CGROUP_STAT_RSS_HUGE, /* # of pages charged as anon huge */
98 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
99 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
100 MEM_CGROUP_STAT_NSTATS,
103 static const char * const mem_cgroup_stat_names[] = {
111 enum mem_cgroup_events_index {
112 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
113 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
114 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
115 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
116 MEM_CGROUP_EVENTS_NSTATS,
119 static const char * const mem_cgroup_events_names[] = {
126 static const char * const mem_cgroup_lru_names[] = {
135 * Per memcg event counter is incremented at every pagein/pageout. With THP,
136 * it will be incremated by the number of pages. This counter is used for
137 * for trigger some periodic events. This is straightforward and better
138 * than using jiffies etc. to handle periodic memcg event.
140 enum mem_cgroup_events_target {
141 MEM_CGROUP_TARGET_THRESH,
142 MEM_CGROUP_TARGET_SOFTLIMIT,
143 MEM_CGROUP_TARGET_NUMAINFO,
146 #define THRESHOLDS_EVENTS_TARGET 128
147 #define SOFTLIMIT_EVENTS_TARGET 1024
148 #define NUMAINFO_EVENTS_TARGET 1024
150 struct mem_cgroup_stat_cpu {
151 long count[MEM_CGROUP_STAT_NSTATS];
152 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
153 unsigned long nr_page_events;
154 unsigned long targets[MEM_CGROUP_NTARGETS];
157 struct mem_cgroup_reclaim_iter {
159 * last scanned hierarchy member. Valid only if last_dead_count
160 * matches memcg->dead_count of the hierarchy root group.
162 struct mem_cgroup *last_visited;
163 unsigned long last_dead_count;
165 /* scan generation, increased every round-trip */
166 unsigned int generation;
170 * per-zone information in memory controller.
172 struct mem_cgroup_per_zone {
173 struct lruvec lruvec;
174 unsigned long lru_size[NR_LRU_LISTS];
176 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
178 struct rb_node tree_node; /* RB tree node */
179 unsigned long long usage_in_excess;/* Set to the value by which */
180 /* the soft limit is exceeded*/
182 struct mem_cgroup *memcg; /* Back pointer, we cannot */
183 /* use container_of */
186 struct mem_cgroup_per_node {
187 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
191 * Cgroups above their limits are maintained in a RB-Tree, independent of
192 * their hierarchy representation
195 struct mem_cgroup_tree_per_zone {
196 struct rb_root rb_root;
200 struct mem_cgroup_tree_per_node {
201 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
204 struct mem_cgroup_tree {
205 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
208 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
210 struct mem_cgroup_threshold {
211 struct eventfd_ctx *eventfd;
216 struct mem_cgroup_threshold_ary {
217 /* An array index points to threshold just below or equal to usage. */
218 int current_threshold;
219 /* Size of entries[] */
221 /* Array of thresholds */
222 struct mem_cgroup_threshold entries[0];
225 struct mem_cgroup_thresholds {
226 /* Primary thresholds array */
227 struct mem_cgroup_threshold_ary *primary;
229 * Spare threshold array.
230 * This is needed to make mem_cgroup_unregister_event() "never fail".
231 * It must be able to store at least primary->size - 1 entries.
233 struct mem_cgroup_threshold_ary *spare;
237 struct mem_cgroup_eventfd_list {
238 struct list_head list;
239 struct eventfd_ctx *eventfd;
242 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
243 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
246 * The memory controller data structure. The memory controller controls both
247 * page cache and RSS per cgroup. We would eventually like to provide
248 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
249 * to help the administrator determine what knobs to tune.
251 * TODO: Add a water mark for the memory controller. Reclaim will begin when
252 * we hit the water mark. May be even add a low water mark, such that
253 * no reclaim occurs from a cgroup at it's low water mark, this is
254 * a feature that will be implemented much later in the future.
257 struct cgroup_subsys_state css;
259 * the counter to account for memory usage
261 struct res_counter res;
263 /* vmpressure notifications */
264 struct vmpressure vmpressure;
267 * the counter to account for mem+swap usage.
269 struct res_counter memsw;
272 * the counter to account for kernel memory usage.
274 struct res_counter kmem;
276 * Should the accounting and control be hierarchical, per subtree?
279 unsigned long kmem_account_flags; /* See KMEM_ACCOUNTED_*, below */
285 /* OOM-Killer disable */
286 int oom_kill_disable;
288 /* set when res.limit == memsw.limit */
289 bool memsw_is_minimum;
291 /* protect arrays of thresholds */
292 struct mutex thresholds_lock;
294 /* thresholds for memory usage. RCU-protected */
295 struct mem_cgroup_thresholds thresholds;
297 /* thresholds for mem+swap usage. RCU-protected */
298 struct mem_cgroup_thresholds memsw_thresholds;
300 /* For oom notifier event fd */
301 struct list_head oom_notify;
304 * Should we move charges of a task when a task is moved into this
305 * mem_cgroup ? And what type of charges should we move ?
307 unsigned long move_charge_at_immigrate;
309 * set > 0 if pages under this cgroup are moving to other cgroup.
311 atomic_t moving_account;
312 /* taken only while moving_account > 0 */
313 spinlock_t move_lock;
317 struct mem_cgroup_stat_cpu __percpu *stat;
319 * used when a cpu is offlined or other synchronizations
320 * See mem_cgroup_read_stat().
322 struct mem_cgroup_stat_cpu nocpu_base;
323 spinlock_t pcp_counter_lock;
326 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
327 struct tcp_memcontrol tcp_mem;
329 #if defined(CONFIG_MEMCG_KMEM)
330 /* analogous to slab_common's slab_caches list. per-memcg */
331 struct list_head memcg_slab_caches;
332 /* Not a spinlock, we can take a lot of time walking the list */
333 struct mutex slab_caches_mutex;
334 /* Index in the kmem_cache->memcg_params->memcg_caches array */
338 int last_scanned_node;
340 nodemask_t scan_nodes;
341 atomic_t numainfo_events;
342 atomic_t numainfo_updating;
345 struct mem_cgroup_per_node *nodeinfo[0];
346 /* WARNING: nodeinfo must be the last member here */
349 static size_t memcg_size(void)
351 return sizeof(struct mem_cgroup) +
352 nr_node_ids * sizeof(struct mem_cgroup_per_node);
355 /* internal only representation about the status of kmem accounting. */
357 KMEM_ACCOUNTED_ACTIVE = 0, /* accounted by this cgroup itself */
358 KMEM_ACCOUNTED_ACTIVATED, /* static key enabled. */
359 KMEM_ACCOUNTED_DEAD, /* dead memcg with pending kmem charges */
362 /* We account when limit is on, but only after call sites are patched */
363 #define KMEM_ACCOUNTED_MASK \
364 ((1 << KMEM_ACCOUNTED_ACTIVE) | (1 << KMEM_ACCOUNTED_ACTIVATED))
366 #ifdef CONFIG_MEMCG_KMEM
367 static inline void memcg_kmem_set_active(struct mem_cgroup *memcg)
369 set_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
372 static bool memcg_kmem_is_active(struct mem_cgroup *memcg)
374 return test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags);
377 static void memcg_kmem_set_activated(struct mem_cgroup *memcg)
379 set_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
382 static void memcg_kmem_clear_activated(struct mem_cgroup *memcg)
384 clear_bit(KMEM_ACCOUNTED_ACTIVATED, &memcg->kmem_account_flags);
387 static void memcg_kmem_mark_dead(struct mem_cgroup *memcg)
390 * Our caller must use css_get() first, because memcg_uncharge_kmem()
391 * will call css_put() if it sees the memcg is dead.
394 if (test_bit(KMEM_ACCOUNTED_ACTIVE, &memcg->kmem_account_flags))
395 set_bit(KMEM_ACCOUNTED_DEAD, &memcg->kmem_account_flags);
398 static bool memcg_kmem_test_and_clear_dead(struct mem_cgroup *memcg)
400 return test_and_clear_bit(KMEM_ACCOUNTED_DEAD,
401 &memcg->kmem_account_flags);
405 /* Stuffs for move charges at task migration. */
407 * Types of charges to be moved. "move_charge_at_immitgrate" and
408 * "immigrate_flags" are treated as a left-shifted bitmap of these types.
411 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
412 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
416 /* "mc" and its members are protected by cgroup_mutex */
417 static struct move_charge_struct {
418 spinlock_t lock; /* for from, to */
419 struct mem_cgroup *from;
420 struct mem_cgroup *to;
421 unsigned long immigrate_flags;
422 unsigned long precharge;
423 unsigned long moved_charge;
424 unsigned long moved_swap;
425 struct task_struct *moving_task; /* a task moving charges */
426 wait_queue_head_t waitq; /* a waitq for other context */
428 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
429 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
432 static bool move_anon(void)
434 return test_bit(MOVE_CHARGE_TYPE_ANON, &mc.immigrate_flags);
437 static bool move_file(void)
439 return test_bit(MOVE_CHARGE_TYPE_FILE, &mc.immigrate_flags);
443 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
444 * limit reclaim to prevent infinite loops, if they ever occur.
446 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
447 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
450 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
451 MEM_CGROUP_CHARGE_TYPE_ANON,
452 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
453 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
457 /* for encoding cft->private value on file */
465 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
466 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
467 #define MEMFILE_ATTR(val) ((val) & 0xffff)
468 /* Used for OOM nofiier */
469 #define OOM_CONTROL (0)
472 * Reclaim flags for mem_cgroup_hierarchical_reclaim
474 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
475 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
476 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
477 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
480 * The memcg_create_mutex will be held whenever a new cgroup is created.
481 * As a consequence, any change that needs to protect against new child cgroups
482 * appearing has to hold it as well.
484 static DEFINE_MUTEX(memcg_create_mutex);
486 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
488 return s ? container_of(s, struct mem_cgroup, css) : NULL;
491 /* Some nice accessors for the vmpressure. */
492 struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg)
495 memcg = root_mem_cgroup;
496 return &memcg->vmpressure;
499 struct cgroup_subsys_state *vmpressure_to_css(struct vmpressure *vmpr)
501 return &container_of(vmpr, struct mem_cgroup, vmpressure)->css;
504 struct vmpressure *css_to_vmpressure(struct cgroup_subsys_state *css)
506 return &mem_cgroup_from_css(css)->vmpressure;
509 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
511 return (memcg == root_mem_cgroup);
514 /* Writing them here to avoid exposing memcg's inner layout */
515 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
517 void sock_update_memcg(struct sock *sk)
519 if (mem_cgroup_sockets_enabled) {
520 struct mem_cgroup *memcg;
521 struct cg_proto *cg_proto;
523 BUG_ON(!sk->sk_prot->proto_cgroup);
525 /* Socket cloning can throw us here with sk_cgrp already
526 * filled. It won't however, necessarily happen from
527 * process context. So the test for root memcg given
528 * the current task's memcg won't help us in this case.
530 * Respecting the original socket's memcg is a better
531 * decision in this case.
534 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
535 css_get(&sk->sk_cgrp->memcg->css);
540 memcg = mem_cgroup_from_task(current);
541 cg_proto = sk->sk_prot->proto_cgroup(memcg);
542 if (!mem_cgroup_is_root(memcg) &&
543 memcg_proto_active(cg_proto) && css_tryget(&memcg->css)) {
544 sk->sk_cgrp = cg_proto;
549 EXPORT_SYMBOL(sock_update_memcg);
551 void sock_release_memcg(struct sock *sk)
553 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
554 struct mem_cgroup *memcg;
555 WARN_ON(!sk->sk_cgrp->memcg);
556 memcg = sk->sk_cgrp->memcg;
557 css_put(&sk->sk_cgrp->memcg->css);
561 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
563 if (!memcg || mem_cgroup_is_root(memcg))
566 return &memcg->tcp_mem.cg_proto;
568 EXPORT_SYMBOL(tcp_proto_cgroup);
570 static void disarm_sock_keys(struct mem_cgroup *memcg)
572 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
574 static_key_slow_dec(&memcg_socket_limit_enabled);
577 static void disarm_sock_keys(struct mem_cgroup *memcg)
582 #ifdef CONFIG_MEMCG_KMEM
584 * This will be the memcg's index in each cache's ->memcg_params->memcg_caches.
585 * There are two main reasons for not using the css_id for this:
586 * 1) this works better in sparse environments, where we have a lot of memcgs,
587 * but only a few kmem-limited. Or also, if we have, for instance, 200
588 * memcgs, and none but the 200th is kmem-limited, we'd have to have a
589 * 200 entry array for that.
591 * 2) In order not to violate the cgroup API, we would like to do all memory
592 * allocation in ->create(). At that point, we haven't yet allocated the
593 * css_id. Having a separate index prevents us from messing with the cgroup
596 * The current size of the caches array is stored in
597 * memcg_limited_groups_array_size. It will double each time we have to
600 static DEFINE_IDA(kmem_limited_groups);
601 int memcg_limited_groups_array_size;
604 * MIN_SIZE is different than 1, because we would like to avoid going through
605 * the alloc/free process all the time. In a small machine, 4 kmem-limited
606 * cgroups is a reasonable guess. In the future, it could be a parameter or
607 * tunable, but that is strictly not necessary.
609 * MAX_SIZE should be as large as the number of css_ids. Ideally, we could get
610 * this constant directly from cgroup, but it is understandable that this is
611 * better kept as an internal representation in cgroup.c. In any case, the
612 * css_id space is not getting any smaller, and we don't have to necessarily
613 * increase ours as well if it increases.
615 #define MEMCG_CACHES_MIN_SIZE 4
616 #define MEMCG_CACHES_MAX_SIZE 65535
619 * A lot of the calls to the cache allocation functions are expected to be
620 * inlined by the compiler. Since the calls to memcg_kmem_get_cache are
621 * conditional to this static branch, we'll have to allow modules that does
622 * kmem_cache_alloc and the such to see this symbol as well
624 struct static_key memcg_kmem_enabled_key;
625 EXPORT_SYMBOL(memcg_kmem_enabled_key);
627 static void disarm_kmem_keys(struct mem_cgroup *memcg)
629 if (memcg_kmem_is_active(memcg)) {
630 static_key_slow_dec(&memcg_kmem_enabled_key);
631 ida_simple_remove(&kmem_limited_groups, memcg->kmemcg_id);
634 * This check can't live in kmem destruction function,
635 * since the charges will outlive the cgroup
637 WARN_ON(res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0);
640 static void disarm_kmem_keys(struct mem_cgroup *memcg)
643 #endif /* CONFIG_MEMCG_KMEM */
645 static void disarm_static_keys(struct mem_cgroup *memcg)
647 disarm_sock_keys(memcg);
648 disarm_kmem_keys(memcg);
651 static void drain_all_stock_async(struct mem_cgroup *memcg);
653 static struct mem_cgroup_per_zone *
654 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
656 VM_BUG_ON((unsigned)nid >= nr_node_ids);
657 return &memcg->nodeinfo[nid]->zoneinfo[zid];
660 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
665 static struct mem_cgroup_per_zone *
666 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
668 int nid = page_to_nid(page);
669 int zid = page_zonenum(page);
671 return mem_cgroup_zoneinfo(memcg, nid, zid);
674 static struct mem_cgroup_tree_per_zone *
675 soft_limit_tree_node_zone(int nid, int zid)
677 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
680 static struct mem_cgroup_tree_per_zone *
681 soft_limit_tree_from_page(struct page *page)
683 int nid = page_to_nid(page);
684 int zid = page_zonenum(page);
686 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
690 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
691 struct mem_cgroup_per_zone *mz,
692 struct mem_cgroup_tree_per_zone *mctz,
693 unsigned long long new_usage_in_excess)
695 struct rb_node **p = &mctz->rb_root.rb_node;
696 struct rb_node *parent = NULL;
697 struct mem_cgroup_per_zone *mz_node;
702 mz->usage_in_excess = new_usage_in_excess;
703 if (!mz->usage_in_excess)
707 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
709 if (mz->usage_in_excess < mz_node->usage_in_excess)
712 * We can't avoid mem cgroups that are over their soft
713 * limit by the same amount
715 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
718 rb_link_node(&mz->tree_node, parent, p);
719 rb_insert_color(&mz->tree_node, &mctz->rb_root);
724 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
725 struct mem_cgroup_per_zone *mz,
726 struct mem_cgroup_tree_per_zone *mctz)
730 rb_erase(&mz->tree_node, &mctz->rb_root);
735 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
736 struct mem_cgroup_per_zone *mz,
737 struct mem_cgroup_tree_per_zone *mctz)
739 spin_lock(&mctz->lock);
740 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
741 spin_unlock(&mctz->lock);
745 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
747 unsigned long long excess;
748 struct mem_cgroup_per_zone *mz;
749 struct mem_cgroup_tree_per_zone *mctz;
750 int nid = page_to_nid(page);
751 int zid = page_zonenum(page);
752 mctz = soft_limit_tree_from_page(page);
755 * Necessary to update all ancestors when hierarchy is used.
756 * because their event counter is not touched.
758 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
759 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
760 excess = res_counter_soft_limit_excess(&memcg->res);
762 * We have to update the tree if mz is on RB-tree or
763 * mem is over its softlimit.
765 if (excess || mz->on_tree) {
766 spin_lock(&mctz->lock);
767 /* if on-tree, remove it */
769 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
771 * Insert again. mz->usage_in_excess will be updated.
772 * If excess is 0, no tree ops.
774 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
775 spin_unlock(&mctz->lock);
780 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
783 struct mem_cgroup_per_zone *mz;
784 struct mem_cgroup_tree_per_zone *mctz;
786 for_each_node(node) {
787 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
788 mz = mem_cgroup_zoneinfo(memcg, node, zone);
789 mctz = soft_limit_tree_node_zone(node, zone);
790 mem_cgroup_remove_exceeded(memcg, mz, mctz);
795 static struct mem_cgroup_per_zone *
796 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
798 struct rb_node *rightmost = NULL;
799 struct mem_cgroup_per_zone *mz;
803 rightmost = rb_last(&mctz->rb_root);
805 goto done; /* Nothing to reclaim from */
807 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
809 * Remove the node now but someone else can add it back,
810 * we will to add it back at the end of reclaim to its correct
811 * position in the tree.
813 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
814 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
815 !css_tryget(&mz->memcg->css))
821 static struct mem_cgroup_per_zone *
822 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
824 struct mem_cgroup_per_zone *mz;
826 spin_lock(&mctz->lock);
827 mz = __mem_cgroup_largest_soft_limit_node(mctz);
828 spin_unlock(&mctz->lock);
833 * Implementation Note: reading percpu statistics for memcg.
835 * Both of vmstat[] and percpu_counter has threshold and do periodic
836 * synchronization to implement "quick" read. There are trade-off between
837 * reading cost and precision of value. Then, we may have a chance to implement
838 * a periodic synchronizion of counter in memcg's counter.
840 * But this _read() function is used for user interface now. The user accounts
841 * memory usage by memory cgroup and he _always_ requires exact value because
842 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
843 * have to visit all online cpus and make sum. So, for now, unnecessary
844 * synchronization is not implemented. (just implemented for cpu hotplug)
846 * If there are kernel internal actions which can make use of some not-exact
847 * value, and reading all cpu value can be performance bottleneck in some
848 * common workload, threashold and synchonization as vmstat[] should be
851 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
852 enum mem_cgroup_stat_index idx)
858 for_each_online_cpu(cpu)
859 val += per_cpu(memcg->stat->count[idx], cpu);
860 #ifdef CONFIG_HOTPLUG_CPU
861 spin_lock(&memcg->pcp_counter_lock);
862 val += memcg->nocpu_base.count[idx];
863 spin_unlock(&memcg->pcp_counter_lock);
869 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
872 int val = (charge) ? 1 : -1;
873 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
876 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
877 enum mem_cgroup_events_index idx)
879 unsigned long val = 0;
882 for_each_online_cpu(cpu)
883 val += per_cpu(memcg->stat->events[idx], cpu);
884 #ifdef CONFIG_HOTPLUG_CPU
885 spin_lock(&memcg->pcp_counter_lock);
886 val += memcg->nocpu_base.events[idx];
887 spin_unlock(&memcg->pcp_counter_lock);
892 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
894 bool anon, int nr_pages)
899 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
900 * counted as CACHE even if it's on ANON LRU.
903 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
906 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
909 if (PageTransHuge(page))
910 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
913 /* pagein of a big page is an event. So, ignore page size */
915 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
917 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
918 nr_pages = -nr_pages; /* for event */
921 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
927 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
929 struct mem_cgroup_per_zone *mz;
931 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
932 return mz->lru_size[lru];
936 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
937 unsigned int lru_mask)
939 struct mem_cgroup_per_zone *mz;
941 unsigned long ret = 0;
943 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
946 if (BIT(lru) & lru_mask)
947 ret += mz->lru_size[lru];
953 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
954 int nid, unsigned int lru_mask)
959 for (zid = 0; zid < MAX_NR_ZONES; zid++)
960 total += mem_cgroup_zone_nr_lru_pages(memcg,
966 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
967 unsigned int lru_mask)
972 for_each_node_state(nid, N_MEMORY)
973 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
977 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
978 enum mem_cgroup_events_target target)
980 unsigned long val, next;
982 val = __this_cpu_read(memcg->stat->nr_page_events);
983 next = __this_cpu_read(memcg->stat->targets[target]);
984 /* from time_after() in jiffies.h */
985 if ((long)next - (long)val < 0) {
987 case MEM_CGROUP_TARGET_THRESH:
988 next = val + THRESHOLDS_EVENTS_TARGET;
990 case MEM_CGROUP_TARGET_SOFTLIMIT:
991 next = val + SOFTLIMIT_EVENTS_TARGET;
993 case MEM_CGROUP_TARGET_NUMAINFO:
994 next = val + NUMAINFO_EVENTS_TARGET;
999 __this_cpu_write(memcg->stat->targets[target], next);
1006 * Check events in order.
1009 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
1012 /* threshold event is triggered in finer grain than soft limit */
1013 if (unlikely(mem_cgroup_event_ratelimit(memcg,
1014 MEM_CGROUP_TARGET_THRESH))) {
1016 bool do_numainfo __maybe_unused;
1018 do_softlimit = mem_cgroup_event_ratelimit(memcg,
1019 MEM_CGROUP_TARGET_SOFTLIMIT);
1020 #if MAX_NUMNODES > 1
1021 do_numainfo = mem_cgroup_event_ratelimit(memcg,
1022 MEM_CGROUP_TARGET_NUMAINFO);
1026 mem_cgroup_threshold(memcg);
1027 if (unlikely(do_softlimit))
1028 mem_cgroup_update_tree(memcg, page);
1029 #if MAX_NUMNODES > 1
1030 if (unlikely(do_numainfo))
1031 atomic_inc(&memcg->numainfo_events);
1037 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
1040 * mm_update_next_owner() may clear mm->owner to NULL
1041 * if it races with swapoff, page migration, etc.
1042 * So this can be called with p == NULL.
1047 return mem_cgroup_from_css(task_css(p, mem_cgroup_subsys_id));
1050 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
1052 struct mem_cgroup *memcg = NULL;
1057 * Because we have no locks, mm->owner's may be being moved to other
1058 * cgroup. We use css_tryget() here even if this looks
1059 * pessimistic (rather than adding locks here).
1063 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1064 if (unlikely(!memcg))
1066 } while (!css_tryget(&memcg->css));
1072 * Returns a next (in a pre-order walk) alive memcg (with elevated css
1073 * ref. count) or NULL if the whole root's subtree has been visited.
1075 * helper function to be used by mem_cgroup_iter
1077 static struct mem_cgroup *__mem_cgroup_iter_next(struct mem_cgroup *root,
1078 struct mem_cgroup *last_visited)
1080 struct cgroup_subsys_state *prev_css, *next_css;
1082 prev_css = last_visited ? &last_visited->css : NULL;
1084 next_css = css_next_descendant_pre(prev_css, &root->css);
1087 * Even if we found a group we have to make sure it is
1088 * alive. css && !memcg means that the groups should be
1089 * skipped and we should continue the tree walk.
1090 * last_visited css is safe to use because it is
1091 * protected by css_get and the tree walk is rcu safe.
1094 struct mem_cgroup *mem = mem_cgroup_from_css(next_css);
1096 if (css_tryget(&mem->css))
1099 prev_css = next_css;
1107 static void mem_cgroup_iter_invalidate(struct mem_cgroup *root)
1110 * When a group in the hierarchy below root is destroyed, the
1111 * hierarchy iterator can no longer be trusted since it might
1112 * have pointed to the destroyed group. Invalidate it.
1114 atomic_inc(&root->dead_count);
1117 static struct mem_cgroup *
1118 mem_cgroup_iter_load(struct mem_cgroup_reclaim_iter *iter,
1119 struct mem_cgroup *root,
1122 struct mem_cgroup *position = NULL;
1124 * A cgroup destruction happens in two stages: offlining and
1125 * release. They are separated by a RCU grace period.
1127 * If the iterator is valid, we may still race with an
1128 * offlining. The RCU lock ensures the object won't be
1129 * released, tryget will fail if we lost the race.
1131 *sequence = atomic_read(&root->dead_count);
1132 if (iter->last_dead_count == *sequence) {
1134 position = iter->last_visited;
1135 if (position && !css_tryget(&position->css))
1141 static void mem_cgroup_iter_update(struct mem_cgroup_reclaim_iter *iter,
1142 struct mem_cgroup *last_visited,
1143 struct mem_cgroup *new_position,
1147 css_put(&last_visited->css);
1149 * We store the sequence count from the time @last_visited was
1150 * loaded successfully instead of rereading it here so that we
1151 * don't lose destruction events in between. We could have
1152 * raced with the destruction of @new_position after all.
1154 iter->last_visited = new_position;
1156 iter->last_dead_count = sequence;
1160 * mem_cgroup_iter - iterate over memory cgroup hierarchy
1161 * @root: hierarchy root
1162 * @prev: previously returned memcg, NULL on first invocation
1163 * @reclaim: cookie for shared reclaim walks, NULL for full walks
1165 * Returns references to children of the hierarchy below @root, or
1166 * @root itself, or %NULL after a full round-trip.
1168 * Caller must pass the return value in @prev on subsequent
1169 * invocations for reference counting, or use mem_cgroup_iter_break()
1170 * to cancel a hierarchy walk before the round-trip is complete.
1172 * Reclaimers can specify a zone and a priority level in @reclaim to
1173 * divide up the memcgs in the hierarchy among all concurrent
1174 * reclaimers operating on the same zone and priority.
1176 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
1177 struct mem_cgroup *prev,
1178 struct mem_cgroup_reclaim_cookie *reclaim)
1180 struct mem_cgroup *memcg = NULL;
1181 struct mem_cgroup *last_visited = NULL;
1183 if (mem_cgroup_disabled())
1187 root = root_mem_cgroup;
1189 if (prev && !reclaim)
1190 last_visited = prev;
1192 if (!root->use_hierarchy && root != root_mem_cgroup) {
1200 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
1201 int uninitialized_var(seq);
1204 int nid = zone_to_nid(reclaim->zone);
1205 int zid = zone_idx(reclaim->zone);
1206 struct mem_cgroup_per_zone *mz;
1208 mz = mem_cgroup_zoneinfo(root, nid, zid);
1209 iter = &mz->reclaim_iter[reclaim->priority];
1210 if (prev && reclaim->generation != iter->generation) {
1211 iter->last_visited = NULL;
1215 last_visited = mem_cgroup_iter_load(iter, root, &seq);
1218 memcg = __mem_cgroup_iter_next(root, last_visited);
1221 mem_cgroup_iter_update(iter, last_visited, memcg, seq);
1225 else if (!prev && memcg)
1226 reclaim->generation = iter->generation;
1235 if (prev && prev != root)
1236 css_put(&prev->css);
1242 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
1243 * @root: hierarchy root
1244 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
1246 void mem_cgroup_iter_break(struct mem_cgroup *root,
1247 struct mem_cgroup *prev)
1250 root = root_mem_cgroup;
1251 if (prev && prev != root)
1252 css_put(&prev->css);
1256 * Iteration constructs for visiting all cgroups (under a tree). If
1257 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1258 * be used for reference counting.
1260 #define for_each_mem_cgroup_tree(iter, root) \
1261 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1263 iter = mem_cgroup_iter(root, iter, NULL))
1265 #define for_each_mem_cgroup(iter) \
1266 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1268 iter = mem_cgroup_iter(NULL, iter, NULL))
1270 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1272 struct mem_cgroup *memcg;
1275 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1276 if (unlikely(!memcg))
1281 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1284 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1292 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1295 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1296 * @zone: zone of the wanted lruvec
1297 * @memcg: memcg of the wanted lruvec
1299 * Returns the lru list vector holding pages for the given @zone and
1300 * @mem. This can be the global zone lruvec, if the memory controller
1303 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1304 struct mem_cgroup *memcg)
1306 struct mem_cgroup_per_zone *mz;
1307 struct lruvec *lruvec;
1309 if (mem_cgroup_disabled()) {
1310 lruvec = &zone->lruvec;
1314 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1315 lruvec = &mz->lruvec;
1318 * Since a node can be onlined after the mem_cgroup was created,
1319 * we have to be prepared to initialize lruvec->zone here;
1320 * and if offlined then reonlined, we need to reinitialize it.
1322 if (unlikely(lruvec->zone != zone))
1323 lruvec->zone = zone;
1328 * Following LRU functions are allowed to be used without PCG_LOCK.
1329 * Operations are called by routine of global LRU independently from memcg.
1330 * What we have to take care of here is validness of pc->mem_cgroup.
1332 * Changes to pc->mem_cgroup happens when
1335 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1336 * It is added to LRU before charge.
1337 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1338 * When moving account, the page is not on LRU. It's isolated.
1342 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1344 * @zone: zone of the page
1346 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1348 struct mem_cgroup_per_zone *mz;
1349 struct mem_cgroup *memcg;
1350 struct page_cgroup *pc;
1351 struct lruvec *lruvec;
1353 if (mem_cgroup_disabled()) {
1354 lruvec = &zone->lruvec;
1358 pc = lookup_page_cgroup(page);
1359 memcg = pc->mem_cgroup;
1362 * Surreptitiously switch any uncharged offlist page to root:
1363 * an uncharged page off lru does nothing to secure
1364 * its former mem_cgroup from sudden removal.
1366 * Our caller holds lru_lock, and PageCgroupUsed is updated
1367 * under page_cgroup lock: between them, they make all uses
1368 * of pc->mem_cgroup safe.
1370 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1371 pc->mem_cgroup = memcg = root_mem_cgroup;
1373 mz = page_cgroup_zoneinfo(memcg, page);
1374 lruvec = &mz->lruvec;
1377 * Since a node can be onlined after the mem_cgroup was created,
1378 * we have to be prepared to initialize lruvec->zone here;
1379 * and if offlined then reonlined, we need to reinitialize it.
1381 if (unlikely(lruvec->zone != zone))
1382 lruvec->zone = zone;
1387 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1388 * @lruvec: mem_cgroup per zone lru vector
1389 * @lru: index of lru list the page is sitting on
1390 * @nr_pages: positive when adding or negative when removing
1392 * This function must be called when a page is added to or removed from an
1395 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1398 struct mem_cgroup_per_zone *mz;
1399 unsigned long *lru_size;
1401 if (mem_cgroup_disabled())
1404 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1405 lru_size = mz->lru_size + lru;
1406 *lru_size += nr_pages;
1407 VM_BUG_ON((long)(*lru_size) < 0);
1411 * Checks whether given mem is same or in the root_mem_cgroup's
1414 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1415 struct mem_cgroup *memcg)
1417 if (root_memcg == memcg)
1419 if (!root_memcg->use_hierarchy || !memcg)
1421 return css_is_ancestor(&memcg->css, &root_memcg->css);
1424 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1425 struct mem_cgroup *memcg)
1430 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1435 bool task_in_mem_cgroup(struct task_struct *task,
1436 const struct mem_cgroup *memcg)
1438 struct mem_cgroup *curr = NULL;
1439 struct task_struct *p;
1442 p = find_lock_task_mm(task);
1444 curr = try_get_mem_cgroup_from_mm(p->mm);
1448 * All threads may have already detached their mm's, but the oom
1449 * killer still needs to detect if they have already been oom
1450 * killed to prevent needlessly killing additional tasks.
1453 curr = mem_cgroup_from_task(task);
1455 css_get(&curr->css);
1461 * We should check use_hierarchy of "memcg" not "curr". Because checking
1462 * use_hierarchy of "curr" here make this function true if hierarchy is
1463 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1464 * hierarchy(even if use_hierarchy is disabled in "memcg").
1466 ret = mem_cgroup_same_or_subtree(memcg, curr);
1467 css_put(&curr->css);
1471 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1473 unsigned long inactive_ratio;
1474 unsigned long inactive;
1475 unsigned long active;
1478 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1479 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1481 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1483 inactive_ratio = int_sqrt(10 * gb);
1487 return inactive * inactive_ratio < active;
1490 #define mem_cgroup_from_res_counter(counter, member) \
1491 container_of(counter, struct mem_cgroup, member)
1494 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1495 * @memcg: the memory cgroup
1497 * Returns the maximum amount of memory @mem can be charged with, in
1500 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1502 unsigned long long margin;
1504 margin = res_counter_margin(&memcg->res);
1505 if (do_swap_account)
1506 margin = min(margin, res_counter_margin(&memcg->memsw));
1507 return margin >> PAGE_SHIFT;
1510 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1513 if (!css_parent(&memcg->css))
1514 return vm_swappiness;
1516 return memcg->swappiness;
1520 * memcg->moving_account is used for checking possibility that some thread is
1521 * calling move_account(). When a thread on CPU-A starts moving pages under
1522 * a memcg, other threads should check memcg->moving_account under
1523 * rcu_read_lock(), like this:
1527 * memcg->moving_account+1 if (memcg->mocing_account)
1529 * synchronize_rcu() update something.
1534 /* for quick checking without looking up memcg */
1535 atomic_t memcg_moving __read_mostly;
1537 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1539 atomic_inc(&memcg_moving);
1540 atomic_inc(&memcg->moving_account);
1544 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1547 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1548 * We check NULL in callee rather than caller.
1551 atomic_dec(&memcg_moving);
1552 atomic_dec(&memcg->moving_account);
1557 * 2 routines for checking "mem" is under move_account() or not.
1559 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1560 * is used for avoiding races in accounting. If true,
1561 * pc->mem_cgroup may be overwritten.
1563 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1564 * under hierarchy of moving cgroups. This is for
1565 * waiting at hith-memory prressure caused by "move".
1568 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1570 VM_BUG_ON(!rcu_read_lock_held());
1571 return atomic_read(&memcg->moving_account) > 0;
1574 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1576 struct mem_cgroup *from;
1577 struct mem_cgroup *to;
1580 * Unlike task_move routines, we access mc.to, mc.from not under
1581 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1583 spin_lock(&mc.lock);
1589 ret = mem_cgroup_same_or_subtree(memcg, from)
1590 || mem_cgroup_same_or_subtree(memcg, to);
1592 spin_unlock(&mc.lock);
1596 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1598 if (mc.moving_task && current != mc.moving_task) {
1599 if (mem_cgroup_under_move(memcg)) {
1601 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1602 /* moving charge context might have finished. */
1605 finish_wait(&mc.waitq, &wait);
1613 * Take this lock when
1614 * - a code tries to modify page's memcg while it's USED.
1615 * - a code tries to modify page state accounting in a memcg.
1616 * see mem_cgroup_stolen(), too.
1618 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1619 unsigned long *flags)
1621 spin_lock_irqsave(&memcg->move_lock, *flags);
1624 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1625 unsigned long *flags)
1627 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1630 #define K(x) ((x) << (PAGE_SHIFT-10))
1632 * mem_cgroup_print_oom_info: Print OOM information relevant to memory controller.
1633 * @memcg: The memory cgroup that went over limit
1634 * @p: Task that is going to be killed
1636 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1639 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1641 struct cgroup *task_cgrp;
1642 struct cgroup *mem_cgrp;
1644 * Need a buffer in BSS, can't rely on allocations. The code relies
1645 * on the assumption that OOM is serialized for memory controller.
1646 * If this assumption is broken, revisit this code.
1648 static char memcg_name[PATH_MAX];
1650 struct mem_cgroup *iter;
1658 mem_cgrp = memcg->css.cgroup;
1659 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1661 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1664 * Unfortunately, we are unable to convert to a useful name
1665 * But we'll still print out the usage information
1672 pr_info("Task in %s killed", memcg_name);
1675 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1683 * Continues from above, so we don't need an KERN_ level
1685 pr_cont(" as a result of limit of %s\n", memcg_name);
1688 pr_info("memory: usage %llukB, limit %llukB, failcnt %llu\n",
1689 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1690 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1691 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1692 pr_info("memory+swap: usage %llukB, limit %llukB, failcnt %llu\n",
1693 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1694 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1695 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1696 pr_info("kmem: usage %llukB, limit %llukB, failcnt %llu\n",
1697 res_counter_read_u64(&memcg->kmem, RES_USAGE) >> 10,
1698 res_counter_read_u64(&memcg->kmem, RES_LIMIT) >> 10,
1699 res_counter_read_u64(&memcg->kmem, RES_FAILCNT));
1701 for_each_mem_cgroup_tree(iter, memcg) {
1702 pr_info("Memory cgroup stats");
1705 ret = cgroup_path(iter->css.cgroup, memcg_name, PATH_MAX);
1707 pr_cont(" for %s", memcg_name);
1711 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
1712 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
1714 pr_cont(" %s:%ldKB", mem_cgroup_stat_names[i],
1715 K(mem_cgroup_read_stat(iter, i)));
1718 for (i = 0; i < NR_LRU_LISTS; i++)
1719 pr_cont(" %s:%luKB", mem_cgroup_lru_names[i],
1720 K(mem_cgroup_nr_lru_pages(iter, BIT(i))));
1727 * This function returns the number of memcg under hierarchy tree. Returns
1728 * 1(self count) if no children.
1730 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1733 struct mem_cgroup *iter;
1735 for_each_mem_cgroup_tree(iter, memcg)
1741 * Return the memory (and swap, if configured) limit for a memcg.
1743 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1747 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1750 * Do not consider swap space if we cannot swap due to swappiness
1752 if (mem_cgroup_swappiness(memcg)) {
1755 limit += total_swap_pages << PAGE_SHIFT;
1756 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1759 * If memsw is finite and limits the amount of swap space
1760 * available to this memcg, return that limit.
1762 limit = min(limit, memsw);
1768 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1771 struct mem_cgroup *iter;
1772 unsigned long chosen_points = 0;
1773 unsigned long totalpages;
1774 unsigned int points = 0;
1775 struct task_struct *chosen = NULL;
1778 * If current has a pending SIGKILL or is exiting, then automatically
1779 * select it. The goal is to allow it to allocate so that it may
1780 * quickly exit and free its memory.
1782 if (fatal_signal_pending(current) || current->flags & PF_EXITING) {
1783 set_thread_flag(TIF_MEMDIE);
1787 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1788 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1789 for_each_mem_cgroup_tree(iter, memcg) {
1790 struct css_task_iter it;
1791 struct task_struct *task;
1793 css_task_iter_start(&iter->css, &it);
1794 while ((task = css_task_iter_next(&it))) {
1795 switch (oom_scan_process_thread(task, totalpages, NULL,
1797 case OOM_SCAN_SELECT:
1799 put_task_struct(chosen);
1801 chosen_points = ULONG_MAX;
1802 get_task_struct(chosen);
1804 case OOM_SCAN_CONTINUE:
1806 case OOM_SCAN_ABORT:
1807 css_task_iter_end(&it);
1808 mem_cgroup_iter_break(memcg, iter);
1810 put_task_struct(chosen);
1815 points = oom_badness(task, memcg, NULL, totalpages);
1816 if (points > chosen_points) {
1818 put_task_struct(chosen);
1820 chosen_points = points;
1821 get_task_struct(chosen);
1824 css_task_iter_end(&it);
1829 points = chosen_points * 1000 / totalpages;
1830 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1831 NULL, "Memory cgroup out of memory");
1834 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1836 unsigned long flags)
1838 unsigned long total = 0;
1839 bool noswap = false;
1842 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1844 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1847 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1849 drain_all_stock_async(memcg);
1850 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1852 * Allow limit shrinkers, which are triggered directly
1853 * by userspace, to catch signals and stop reclaim
1854 * after minimal progress, regardless of the margin.
1856 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1858 if (mem_cgroup_margin(memcg))
1861 * If nothing was reclaimed after two attempts, there
1862 * may be no reclaimable pages in this hierarchy.
1871 * test_mem_cgroup_node_reclaimable
1872 * @memcg: the target memcg
1873 * @nid: the node ID to be checked.
1874 * @noswap : specify true here if the user wants flle only information.
1876 * This function returns whether the specified memcg contains any
1877 * reclaimable pages on a node. Returns true if there are any reclaimable
1878 * pages in the node.
1880 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1881 int nid, bool noswap)
1883 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1885 if (noswap || !total_swap_pages)
1887 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1892 #if MAX_NUMNODES > 1
1895 * Always updating the nodemask is not very good - even if we have an empty
1896 * list or the wrong list here, we can start from some node and traverse all
1897 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1900 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1904 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1905 * pagein/pageout changes since the last update.
1907 if (!atomic_read(&memcg->numainfo_events))
1909 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1912 /* make a nodemask where this memcg uses memory from */
1913 memcg->scan_nodes = node_states[N_MEMORY];
1915 for_each_node_mask(nid, node_states[N_MEMORY]) {
1917 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1918 node_clear(nid, memcg->scan_nodes);
1921 atomic_set(&memcg->numainfo_events, 0);
1922 atomic_set(&memcg->numainfo_updating, 0);
1926 * Selecting a node where we start reclaim from. Because what we need is just
1927 * reducing usage counter, start from anywhere is O,K. Considering
1928 * memory reclaim from current node, there are pros. and cons.
1930 * Freeing memory from current node means freeing memory from a node which
1931 * we'll use or we've used. So, it may make LRU bad. And if several threads
1932 * hit limits, it will see a contention on a node. But freeing from remote
1933 * node means more costs for memory reclaim because of memory latency.
1935 * Now, we use round-robin. Better algorithm is welcomed.
1937 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1941 mem_cgroup_may_update_nodemask(memcg);
1942 node = memcg->last_scanned_node;
1944 node = next_node(node, memcg->scan_nodes);
1945 if (node == MAX_NUMNODES)
1946 node = first_node(memcg->scan_nodes);
1948 * We call this when we hit limit, not when pages are added to LRU.
1949 * No LRU may hold pages because all pages are UNEVICTABLE or
1950 * memcg is too small and all pages are not on LRU. In that case,
1951 * we use curret node.
1953 if (unlikely(node == MAX_NUMNODES))
1954 node = numa_node_id();
1956 memcg->last_scanned_node = node;
1961 * Check all nodes whether it contains reclaimable pages or not.
1962 * For quick scan, we make use of scan_nodes. This will allow us to skip
1963 * unused nodes. But scan_nodes is lazily updated and may not cotain
1964 * enough new information. We need to do double check.
1966 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1971 * quick check...making use of scan_node.
1972 * We can skip unused nodes.
1974 if (!nodes_empty(memcg->scan_nodes)) {
1975 for (nid = first_node(memcg->scan_nodes);
1977 nid = next_node(nid, memcg->scan_nodes)) {
1979 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1984 * Check rest of nodes.
1986 for_each_node_state(nid, N_MEMORY) {
1987 if (node_isset(nid, memcg->scan_nodes))
1989 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1996 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
2001 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
2003 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
2008 * A group is eligible for the soft limit reclaim if it is
2009 * a) is over its soft limit
2010 * b) any parent up the hierarchy is over its soft limit
2012 bool mem_cgroup_soft_reclaim_eligible(struct mem_cgroup *memcg)
2014 struct mem_cgroup *parent = memcg;
2016 if (res_counter_soft_limit_excess(&memcg->res))
2020 * If any parent up the hierarchy is over its soft limit then we
2021 * have to obey and reclaim from this group as well.
2023 while((parent = parent_mem_cgroup(parent))) {
2024 if (res_counter_soft_limit_excess(&parent->res))
2032 * Check OOM-Killer is already running under our hierarchy.
2033 * If someone is running, return false.
2034 * Has to be called with memcg_oom_lock
2036 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
2038 struct mem_cgroup *iter, *failed = NULL;
2040 for_each_mem_cgroup_tree(iter, memcg) {
2041 if (iter->oom_lock) {
2043 * this subtree of our hierarchy is already locked
2044 * so we cannot give a lock.
2047 mem_cgroup_iter_break(memcg, iter);
2050 iter->oom_lock = true;
2057 * OK, we failed to lock the whole subtree so we have to clean up
2058 * what we set up to the failing subtree
2060 for_each_mem_cgroup_tree(iter, memcg) {
2061 if (iter == failed) {
2062 mem_cgroup_iter_break(memcg, iter);
2065 iter->oom_lock = false;
2071 * Has to be called with memcg_oom_lock
2073 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2075 struct mem_cgroup *iter;
2077 for_each_mem_cgroup_tree(iter, memcg)
2078 iter->oom_lock = false;
2082 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2084 struct mem_cgroup *iter;
2086 for_each_mem_cgroup_tree(iter, memcg)
2087 atomic_inc(&iter->under_oom);
2090 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2092 struct mem_cgroup *iter;
2095 * When a new child is created while the hierarchy is under oom,
2096 * mem_cgroup_oom_lock() may not be called. We have to use
2097 * atomic_add_unless() here.
2099 for_each_mem_cgroup_tree(iter, memcg)
2100 atomic_add_unless(&iter->under_oom, -1, 0);
2103 static DEFINE_SPINLOCK(memcg_oom_lock);
2104 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2106 struct oom_wait_info {
2107 struct mem_cgroup *memcg;
2111 static int memcg_oom_wake_function(wait_queue_t *wait,
2112 unsigned mode, int sync, void *arg)
2114 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2115 struct mem_cgroup *oom_wait_memcg;
2116 struct oom_wait_info *oom_wait_info;
2118 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2119 oom_wait_memcg = oom_wait_info->memcg;
2122 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2123 * Then we can use css_is_ancestor without taking care of RCU.
2125 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2126 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2128 return autoremove_wake_function(wait, mode, sync, arg);
2131 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2133 /* for filtering, pass "memcg" as argument. */
2134 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2137 static void memcg_oom_recover(struct mem_cgroup *memcg)
2139 if (memcg && atomic_read(&memcg->under_oom))
2140 memcg_wakeup_oom(memcg);
2144 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2146 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2149 struct oom_wait_info owait;
2150 bool locked, need_to_kill;
2152 owait.memcg = memcg;
2153 owait.wait.flags = 0;
2154 owait.wait.func = memcg_oom_wake_function;
2155 owait.wait.private = current;
2156 INIT_LIST_HEAD(&owait.wait.task_list);
2157 need_to_kill = true;
2158 mem_cgroup_mark_under_oom(memcg);
2160 /* At first, try to OOM lock hierarchy under memcg.*/
2161 spin_lock(&memcg_oom_lock);
2162 locked = mem_cgroup_oom_lock(memcg);
2164 * Even if signal_pending(), we can't quit charge() loop without
2165 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2166 * under OOM is always welcomed, use TASK_KILLABLE here.
2168 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2169 if (!locked || memcg->oom_kill_disable)
2170 need_to_kill = false;
2172 mem_cgroup_oom_notify(memcg);
2173 spin_unlock(&memcg_oom_lock);
2176 finish_wait(&memcg_oom_waitq, &owait.wait);
2177 mem_cgroup_out_of_memory(memcg, mask, order);
2180 finish_wait(&memcg_oom_waitq, &owait.wait);
2182 spin_lock(&memcg_oom_lock);
2184 mem_cgroup_oom_unlock(memcg);
2185 memcg_wakeup_oom(memcg);
2186 spin_unlock(&memcg_oom_lock);
2188 mem_cgroup_unmark_under_oom(memcg);
2190 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2192 /* Give chance to dying process */
2193 schedule_timeout_uninterruptible(1);
2198 * Currently used to update mapped file statistics, but the routine can be
2199 * generalized to update other statistics as well.
2201 * Notes: Race condition
2203 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2204 * it tends to be costly. But considering some conditions, we doesn't need
2205 * to do so _always_.
2207 * Considering "charge", lock_page_cgroup() is not required because all
2208 * file-stat operations happen after a page is attached to radix-tree. There
2209 * are no race with "charge".
2211 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2212 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2213 * if there are race with "uncharge". Statistics itself is properly handled
2216 * Considering "move", this is an only case we see a race. To make the race
2217 * small, we check mm->moving_account and detect there are possibility of race
2218 * If there is, we take a lock.
2221 void __mem_cgroup_begin_update_page_stat(struct page *page,
2222 bool *locked, unsigned long *flags)
2224 struct mem_cgroup *memcg;
2225 struct page_cgroup *pc;
2227 pc = lookup_page_cgroup(page);
2229 memcg = pc->mem_cgroup;
2230 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2233 * If this memory cgroup is not under account moving, we don't
2234 * need to take move_lock_mem_cgroup(). Because we already hold
2235 * rcu_read_lock(), any calls to move_account will be delayed until
2236 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2238 if (!mem_cgroup_stolen(memcg))
2241 move_lock_mem_cgroup(memcg, flags);
2242 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2243 move_unlock_mem_cgroup(memcg, flags);
2249 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2251 struct page_cgroup *pc = lookup_page_cgroup(page);
2254 * It's guaranteed that pc->mem_cgroup never changes while
2255 * lock is held because a routine modifies pc->mem_cgroup
2256 * should take move_lock_mem_cgroup().
2258 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2261 void mem_cgroup_update_page_stat(struct page *page,
2262 enum mem_cgroup_page_stat_item idx, int val)
2264 struct mem_cgroup *memcg;
2265 struct page_cgroup *pc = lookup_page_cgroup(page);
2266 unsigned long uninitialized_var(flags);
2268 if (mem_cgroup_disabled())
2271 memcg = pc->mem_cgroup;
2272 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2276 case MEMCG_NR_FILE_MAPPED:
2277 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2283 this_cpu_add(memcg->stat->count[idx], val);
2287 * size of first charge trial. "32" comes from vmscan.c's magic value.
2288 * TODO: maybe necessary to use big numbers in big irons.
2290 #define CHARGE_BATCH 32U
2291 struct memcg_stock_pcp {
2292 struct mem_cgroup *cached; /* this never be root cgroup */
2293 unsigned int nr_pages;
2294 struct work_struct work;
2295 unsigned long flags;
2296 #define FLUSHING_CACHED_CHARGE 0
2298 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2299 static DEFINE_MUTEX(percpu_charge_mutex);
2302 * consume_stock: Try to consume stocked charge on this cpu.
2303 * @memcg: memcg to consume from.
2304 * @nr_pages: how many pages to charge.
2306 * The charges will only happen if @memcg matches the current cpu's memcg
2307 * stock, and at least @nr_pages are available in that stock. Failure to
2308 * service an allocation will refill the stock.
2310 * returns true if successful, false otherwise.
2312 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2314 struct memcg_stock_pcp *stock;
2317 if (nr_pages > CHARGE_BATCH)
2320 stock = &get_cpu_var(memcg_stock);
2321 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2322 stock->nr_pages -= nr_pages;
2323 else /* need to call res_counter_charge */
2325 put_cpu_var(memcg_stock);
2330 * Returns stocks cached in percpu to res_counter and reset cached information.
2332 static void drain_stock(struct memcg_stock_pcp *stock)
2334 struct mem_cgroup *old = stock->cached;
2336 if (stock->nr_pages) {
2337 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2339 res_counter_uncharge(&old->res, bytes);
2340 if (do_swap_account)
2341 res_counter_uncharge(&old->memsw, bytes);
2342 stock->nr_pages = 0;
2344 stock->cached = NULL;
2348 * This must be called under preempt disabled or must be called by
2349 * a thread which is pinned to local cpu.
2351 static void drain_local_stock(struct work_struct *dummy)
2353 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2355 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2358 static void __init memcg_stock_init(void)
2362 for_each_possible_cpu(cpu) {
2363 struct memcg_stock_pcp *stock =
2364 &per_cpu(memcg_stock, cpu);
2365 INIT_WORK(&stock->work, drain_local_stock);
2370 * Cache charges(val) which is from res_counter, to local per_cpu area.
2371 * This will be consumed by consume_stock() function, later.
2373 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2375 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2377 if (stock->cached != memcg) { /* reset if necessary */
2379 stock->cached = memcg;
2381 stock->nr_pages += nr_pages;
2382 put_cpu_var(memcg_stock);
2386 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2387 * of the hierarchy under it. sync flag says whether we should block
2388 * until the work is done.
2390 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2394 /* Notify other cpus that system-wide "drain" is running */
2397 for_each_online_cpu(cpu) {
2398 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2399 struct mem_cgroup *memcg;
2401 memcg = stock->cached;
2402 if (!memcg || !stock->nr_pages)
2404 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2406 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2408 drain_local_stock(&stock->work);
2410 schedule_work_on(cpu, &stock->work);
2418 for_each_online_cpu(cpu) {
2419 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2420 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2421 flush_work(&stock->work);
2428 * Tries to drain stocked charges in other cpus. This function is asynchronous
2429 * and just put a work per cpu for draining localy on each cpu. Caller can
2430 * expects some charges will be back to res_counter later but cannot wait for
2433 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2436 * If someone calls draining, avoid adding more kworker runs.
2438 if (!mutex_trylock(&percpu_charge_mutex))
2440 drain_all_stock(root_memcg, false);
2441 mutex_unlock(&percpu_charge_mutex);
2444 /* This is a synchronous drain interface. */
2445 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2447 /* called when force_empty is called */
2448 mutex_lock(&percpu_charge_mutex);
2449 drain_all_stock(root_memcg, true);
2450 mutex_unlock(&percpu_charge_mutex);
2454 * This function drains percpu counter value from DEAD cpu and
2455 * move it to local cpu. Note that this function can be preempted.
2457 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2461 spin_lock(&memcg->pcp_counter_lock);
2462 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2463 long x = per_cpu(memcg->stat->count[i], cpu);
2465 per_cpu(memcg->stat->count[i], cpu) = 0;
2466 memcg->nocpu_base.count[i] += x;
2468 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2469 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2471 per_cpu(memcg->stat->events[i], cpu) = 0;
2472 memcg->nocpu_base.events[i] += x;
2474 spin_unlock(&memcg->pcp_counter_lock);
2477 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2478 unsigned long action,
2481 int cpu = (unsigned long)hcpu;
2482 struct memcg_stock_pcp *stock;
2483 struct mem_cgroup *iter;
2485 if (action == CPU_ONLINE)
2488 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2491 for_each_mem_cgroup(iter)
2492 mem_cgroup_drain_pcp_counter(iter, cpu);
2494 stock = &per_cpu(memcg_stock, cpu);
2500 /* See __mem_cgroup_try_charge() for details */
2502 CHARGE_OK, /* success */
2503 CHARGE_RETRY, /* need to retry but retry is not bad */
2504 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2505 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2506 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2509 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2510 unsigned int nr_pages, unsigned int min_pages,
2513 unsigned long csize = nr_pages * PAGE_SIZE;
2514 struct mem_cgroup *mem_over_limit;
2515 struct res_counter *fail_res;
2516 unsigned long flags = 0;
2519 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2522 if (!do_swap_account)
2524 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2528 res_counter_uncharge(&memcg->res, csize);
2529 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2530 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2532 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2534 * Never reclaim on behalf of optional batching, retry with a
2535 * single page instead.
2537 if (nr_pages > min_pages)
2538 return CHARGE_RETRY;
2540 if (!(gfp_mask & __GFP_WAIT))
2541 return CHARGE_WOULDBLOCK;
2543 if (gfp_mask & __GFP_NORETRY)
2544 return CHARGE_NOMEM;
2546 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2547 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2548 return CHARGE_RETRY;
2550 * Even though the limit is exceeded at this point, reclaim
2551 * may have been able to free some pages. Retry the charge
2552 * before killing the task.
2554 * Only for regular pages, though: huge pages are rather
2555 * unlikely to succeed so close to the limit, and we fall back
2556 * to regular pages anyway in case of failure.
2558 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2559 return CHARGE_RETRY;
2562 * At task move, charge accounts can be doubly counted. So, it's
2563 * better to wait until the end of task_move if something is going on.
2565 if (mem_cgroup_wait_acct_move(mem_over_limit))
2566 return CHARGE_RETRY;
2568 /* If we don't need to call oom-killer at el, return immediately */
2570 return CHARGE_NOMEM;
2572 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2573 return CHARGE_OOM_DIE;
2575 return CHARGE_RETRY;
2579 * __mem_cgroup_try_charge() does
2580 * 1. detect memcg to be charged against from passed *mm and *ptr,
2581 * 2. update res_counter
2582 * 3. call memory reclaim if necessary.
2584 * In some special case, if the task is fatal, fatal_signal_pending() or
2585 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2586 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2587 * as possible without any hazards. 2: all pages should have a valid
2588 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2589 * pointer, that is treated as a charge to root_mem_cgroup.
2591 * So __mem_cgroup_try_charge() will return
2592 * 0 ... on success, filling *ptr with a valid memcg pointer.
2593 * -ENOMEM ... charge failure because of resource limits.
2594 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2596 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2597 * the oom-killer can be invoked.
2599 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2601 unsigned int nr_pages,
2602 struct mem_cgroup **ptr,
2605 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2606 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2607 struct mem_cgroup *memcg = NULL;
2611 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2612 * in system level. So, allow to go ahead dying process in addition to
2615 if (unlikely(test_thread_flag(TIF_MEMDIE)
2616 || fatal_signal_pending(current)))
2620 * We always charge the cgroup the mm_struct belongs to.
2621 * The mm_struct's mem_cgroup changes on task migration if the
2622 * thread group leader migrates. It's possible that mm is not
2623 * set, if so charge the root memcg (happens for pagecache usage).
2626 *ptr = root_mem_cgroup;
2628 if (*ptr) { /* css should be a valid one */
2630 if (mem_cgroup_is_root(memcg))
2632 if (consume_stock(memcg, nr_pages))
2634 css_get(&memcg->css);
2636 struct task_struct *p;
2639 p = rcu_dereference(mm->owner);
2641 * Because we don't have task_lock(), "p" can exit.
2642 * In that case, "memcg" can point to root or p can be NULL with
2643 * race with swapoff. Then, we have small risk of mis-accouning.
2644 * But such kind of mis-account by race always happens because
2645 * we don't have cgroup_mutex(). It's overkill and we allo that
2647 * (*) swapoff at el will charge against mm-struct not against
2648 * task-struct. So, mm->owner can be NULL.
2650 memcg = mem_cgroup_from_task(p);
2652 memcg = root_mem_cgroup;
2653 if (mem_cgroup_is_root(memcg)) {
2657 if (consume_stock(memcg, nr_pages)) {
2659 * It seems dagerous to access memcg without css_get().
2660 * But considering how consume_stok works, it's not
2661 * necessary. If consume_stock success, some charges
2662 * from this memcg are cached on this cpu. So, we
2663 * don't need to call css_get()/css_tryget() before
2664 * calling consume_stock().
2669 /* after here, we may be blocked. we need to get refcnt */
2670 if (!css_tryget(&memcg->css)) {
2680 /* If killed, bypass charge */
2681 if (fatal_signal_pending(current)) {
2682 css_put(&memcg->css);
2687 if (oom && !nr_oom_retries) {
2689 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2692 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2697 case CHARGE_RETRY: /* not in OOM situation but retry */
2699 css_put(&memcg->css);
2702 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2703 css_put(&memcg->css);
2705 case CHARGE_NOMEM: /* OOM routine works */
2707 css_put(&memcg->css);
2710 /* If oom, we never return -ENOMEM */
2713 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2714 css_put(&memcg->css);
2717 } while (ret != CHARGE_OK);
2719 if (batch > nr_pages)
2720 refill_stock(memcg, batch - nr_pages);
2721 css_put(&memcg->css);
2729 *ptr = root_mem_cgroup;
2734 * Somemtimes we have to undo a charge we got by try_charge().
2735 * This function is for that and do uncharge, put css's refcnt.
2736 * gotten by try_charge().
2738 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2739 unsigned int nr_pages)
2741 if (!mem_cgroup_is_root(memcg)) {
2742 unsigned long bytes = nr_pages * PAGE_SIZE;
2744 res_counter_uncharge(&memcg->res, bytes);
2745 if (do_swap_account)
2746 res_counter_uncharge(&memcg->memsw, bytes);
2751 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2752 * This is useful when moving usage to parent cgroup.
2754 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2755 unsigned int nr_pages)
2757 unsigned long bytes = nr_pages * PAGE_SIZE;
2759 if (mem_cgroup_is_root(memcg))
2762 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2763 if (do_swap_account)
2764 res_counter_uncharge_until(&memcg->memsw,
2765 memcg->memsw.parent, bytes);
2769 * A helper function to get mem_cgroup from ID. must be called under
2770 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2771 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2772 * called against removed memcg.)
2774 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2776 struct cgroup_subsys_state *css;
2778 /* ID 0 is unused ID */
2781 css = css_lookup(&mem_cgroup_subsys, id);
2784 return mem_cgroup_from_css(css);
2787 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2789 struct mem_cgroup *memcg = NULL;
2790 struct page_cgroup *pc;
2794 VM_BUG_ON(!PageLocked(page));
2796 pc = lookup_page_cgroup(page);
2797 lock_page_cgroup(pc);
2798 if (PageCgroupUsed(pc)) {
2799 memcg = pc->mem_cgroup;
2800 if (memcg && !css_tryget(&memcg->css))
2802 } else if (PageSwapCache(page)) {
2803 ent.val = page_private(page);
2804 id = lookup_swap_cgroup_id(ent);
2806 memcg = mem_cgroup_lookup(id);
2807 if (memcg && !css_tryget(&memcg->css))
2811 unlock_page_cgroup(pc);
2815 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2817 unsigned int nr_pages,
2818 enum charge_type ctype,
2821 struct page_cgroup *pc = lookup_page_cgroup(page);
2822 struct zone *uninitialized_var(zone);
2823 struct lruvec *lruvec;
2824 bool was_on_lru = false;
2827 lock_page_cgroup(pc);
2828 VM_BUG_ON(PageCgroupUsed(pc));
2830 * we don't need page_cgroup_lock about tail pages, becase they are not
2831 * accessed by any other context at this point.
2835 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2836 * may already be on some other mem_cgroup's LRU. Take care of it.
2839 zone = page_zone(page);
2840 spin_lock_irq(&zone->lru_lock);
2841 if (PageLRU(page)) {
2842 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2844 del_page_from_lru_list(page, lruvec, page_lru(page));
2849 pc->mem_cgroup = memcg;
2851 * We access a page_cgroup asynchronously without lock_page_cgroup().
2852 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2853 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2854 * before USED bit, we need memory barrier here.
2855 * See mem_cgroup_add_lru_list(), etc.
2858 SetPageCgroupUsed(pc);
2862 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2863 VM_BUG_ON(PageLRU(page));
2865 add_page_to_lru_list(page, lruvec, page_lru(page));
2867 spin_unlock_irq(&zone->lru_lock);
2870 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2875 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2876 unlock_page_cgroup(pc);
2879 * "charge_statistics" updated event counter. Then, check it.
2880 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2881 * if they exceeds softlimit.
2883 memcg_check_events(memcg, page);
2886 static DEFINE_MUTEX(set_limit_mutex);
2888 #ifdef CONFIG_MEMCG_KMEM
2889 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2891 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2892 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2896 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2897 * in the memcg_cache_params struct.
2899 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2901 struct kmem_cache *cachep;
2903 VM_BUG_ON(p->is_root_cache);
2904 cachep = p->root_cache;
2905 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2908 #ifdef CONFIG_SLABINFO
2909 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2910 struct cftype *cft, struct seq_file *m)
2912 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2913 struct memcg_cache_params *params;
2915 if (!memcg_can_account_kmem(memcg))
2918 print_slabinfo_header(m);
2920 mutex_lock(&memcg->slab_caches_mutex);
2921 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2922 cache_show(memcg_params_to_cache(params), m);
2923 mutex_unlock(&memcg->slab_caches_mutex);
2929 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2931 struct res_counter *fail_res;
2932 struct mem_cgroup *_memcg;
2936 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2941 * Conditions under which we can wait for the oom_killer. Those are
2942 * the same conditions tested by the core page allocator
2944 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2947 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2950 if (ret == -EINTR) {
2952 * __mem_cgroup_try_charge() chosed to bypass to root due to
2953 * OOM kill or fatal signal. Since our only options are to
2954 * either fail the allocation or charge it to this cgroup, do
2955 * it as a temporary condition. But we can't fail. From a
2956 * kmem/slab perspective, the cache has already been selected,
2957 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2960 * This condition will only trigger if the task entered
2961 * memcg_charge_kmem in a sane state, but was OOM-killed during
2962 * __mem_cgroup_try_charge() above. Tasks that were already
2963 * dying when the allocation triggers should have been already
2964 * directed to the root cgroup in memcontrol.h
2966 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2967 if (do_swap_account)
2968 res_counter_charge_nofail(&memcg->memsw, size,
2972 res_counter_uncharge(&memcg->kmem, size);
2977 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
2979 res_counter_uncharge(&memcg->res, size);
2980 if (do_swap_account)
2981 res_counter_uncharge(&memcg->memsw, size);
2984 if (res_counter_uncharge(&memcg->kmem, size))
2988 * Releases a reference taken in kmem_cgroup_css_offline in case
2989 * this last uncharge is racing with the offlining code or it is
2990 * outliving the memcg existence.
2992 * The memory barrier imposed by test&clear is paired with the
2993 * explicit one in memcg_kmem_mark_dead().
2995 if (memcg_kmem_test_and_clear_dead(memcg))
2996 css_put(&memcg->css);
2999 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3004 mutex_lock(&memcg->slab_caches_mutex);
3005 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3006 mutex_unlock(&memcg->slab_caches_mutex);
3010 * helper for acessing a memcg's index. It will be used as an index in the
3011 * child cache array in kmem_cache, and also to derive its name. This function
3012 * will return -1 when this is not a kmem-limited memcg.
3014 int memcg_cache_id(struct mem_cgroup *memcg)
3016 return memcg ? memcg->kmemcg_id : -1;
3020 * This ends up being protected by the set_limit mutex, during normal
3021 * operation, because that is its main call site.
3023 * But when we create a new cache, we can call this as well if its parent
3024 * is kmem-limited. That will have to hold set_limit_mutex as well.
3026 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3030 num = ida_simple_get(&kmem_limited_groups,
3031 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3035 * After this point, kmem_accounted (that we test atomically in
3036 * the beginning of this conditional), is no longer 0. This
3037 * guarantees only one process will set the following boolean
3038 * to true. We don't need test_and_set because we're protected
3039 * by the set_limit_mutex anyway.
3041 memcg_kmem_set_activated(memcg);
3043 ret = memcg_update_all_caches(num+1);
3045 ida_simple_remove(&kmem_limited_groups, num);
3046 memcg_kmem_clear_activated(memcg);
3050 memcg->kmemcg_id = num;
3051 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3052 mutex_init(&memcg->slab_caches_mutex);
3056 static size_t memcg_caches_array_size(int num_groups)
3059 if (num_groups <= 0)
3062 size = 2 * num_groups;
3063 if (size < MEMCG_CACHES_MIN_SIZE)
3064 size = MEMCG_CACHES_MIN_SIZE;
3065 else if (size > MEMCG_CACHES_MAX_SIZE)
3066 size = MEMCG_CACHES_MAX_SIZE;
3072 * We should update the current array size iff all caches updates succeed. This
3073 * can only be done from the slab side. The slab mutex needs to be held when
3076 void memcg_update_array_size(int num)
3078 if (num > memcg_limited_groups_array_size)
3079 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3082 static void kmem_cache_destroy_work_func(struct work_struct *w);
3084 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3086 struct memcg_cache_params *cur_params = s->memcg_params;
3088 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3090 if (num_groups > memcg_limited_groups_array_size) {
3092 ssize_t size = memcg_caches_array_size(num_groups);
3094 size *= sizeof(void *);
3095 size += offsetof(struct memcg_cache_params, memcg_caches);
3097 s->memcg_params = kzalloc(size, GFP_KERNEL);
3098 if (!s->memcg_params) {
3099 s->memcg_params = cur_params;
3103 s->memcg_params->is_root_cache = true;
3106 * There is the chance it will be bigger than
3107 * memcg_limited_groups_array_size, if we failed an allocation
3108 * in a cache, in which case all caches updated before it, will
3109 * have a bigger array.
3111 * But if that is the case, the data after
3112 * memcg_limited_groups_array_size is certainly unused
3114 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3115 if (!cur_params->memcg_caches[i])
3117 s->memcg_params->memcg_caches[i] =
3118 cur_params->memcg_caches[i];
3122 * Ideally, we would wait until all caches succeed, and only
3123 * then free the old one. But this is not worth the extra
3124 * pointer per-cache we'd have to have for this.
3126 * It is not a big deal if some caches are left with a size
3127 * bigger than the others. And all updates will reset this
3135 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3136 struct kmem_cache *root_cache)
3140 if (!memcg_kmem_enabled())
3144 size = offsetof(struct memcg_cache_params, memcg_caches);
3145 size += memcg_limited_groups_array_size * sizeof(void *);
3147 size = sizeof(struct memcg_cache_params);
3149 s->memcg_params = kzalloc(size, GFP_KERNEL);
3150 if (!s->memcg_params)
3154 s->memcg_params->memcg = memcg;
3155 s->memcg_params->root_cache = root_cache;
3156 INIT_WORK(&s->memcg_params->destroy,
3157 kmem_cache_destroy_work_func);
3159 s->memcg_params->is_root_cache = true;
3164 void memcg_release_cache(struct kmem_cache *s)
3166 struct kmem_cache *root;
3167 struct mem_cgroup *memcg;
3171 * This happens, for instance, when a root cache goes away before we
3174 if (!s->memcg_params)
3177 if (s->memcg_params->is_root_cache)
3180 memcg = s->memcg_params->memcg;
3181 id = memcg_cache_id(memcg);
3183 root = s->memcg_params->root_cache;
3184 root->memcg_params->memcg_caches[id] = NULL;
3186 mutex_lock(&memcg->slab_caches_mutex);
3187 list_del(&s->memcg_params->list);
3188 mutex_unlock(&memcg->slab_caches_mutex);
3190 css_put(&memcg->css);
3192 kfree(s->memcg_params);
3196 * During the creation a new cache, we need to disable our accounting mechanism
3197 * altogether. This is true even if we are not creating, but rather just
3198 * enqueing new caches to be created.
3200 * This is because that process will trigger allocations; some visible, like
3201 * explicit kmallocs to auxiliary data structures, name strings and internal
3202 * cache structures; some well concealed, like INIT_WORK() that can allocate
3203 * objects during debug.
3205 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3206 * to it. This may not be a bounded recursion: since the first cache creation
3207 * failed to complete (waiting on the allocation), we'll just try to create the
3208 * cache again, failing at the same point.
3210 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3211 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3212 * inside the following two functions.
3214 static inline void memcg_stop_kmem_account(void)
3216 VM_BUG_ON(!current->mm);
3217 current->memcg_kmem_skip_account++;
3220 static inline void memcg_resume_kmem_account(void)
3222 VM_BUG_ON(!current->mm);
3223 current->memcg_kmem_skip_account--;
3226 static void kmem_cache_destroy_work_func(struct work_struct *w)
3228 struct kmem_cache *cachep;
3229 struct memcg_cache_params *p;
3231 p = container_of(w, struct memcg_cache_params, destroy);
3233 cachep = memcg_params_to_cache(p);
3236 * If we get down to 0 after shrink, we could delete right away.
3237 * However, memcg_release_pages() already puts us back in the workqueue
3238 * in that case. If we proceed deleting, we'll get a dangling
3239 * reference, and removing the object from the workqueue in that case
3240 * is unnecessary complication. We are not a fast path.
3242 * Note that this case is fundamentally different from racing with
3243 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3244 * kmem_cache_shrink, not only we would be reinserting a dead cache
3245 * into the queue, but doing so from inside the worker racing to
3248 * So if we aren't down to zero, we'll just schedule a worker and try
3251 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3252 kmem_cache_shrink(cachep);
3253 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3256 kmem_cache_destroy(cachep);
3259 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3261 if (!cachep->memcg_params->dead)
3265 * There are many ways in which we can get here.
3267 * We can get to a memory-pressure situation while the delayed work is
3268 * still pending to run. The vmscan shrinkers can then release all
3269 * cache memory and get us to destruction. If this is the case, we'll
3270 * be executed twice, which is a bug (the second time will execute over
3271 * bogus data). In this case, cancelling the work should be fine.
3273 * But we can also get here from the worker itself, if
3274 * kmem_cache_shrink is enough to shake all the remaining objects and
3275 * get the page count to 0. In this case, we'll deadlock if we try to
3276 * cancel the work (the worker runs with an internal lock held, which
3277 * is the same lock we would hold for cancel_work_sync().)
3279 * Since we can't possibly know who got us here, just refrain from
3280 * running if there is already work pending
3282 if (work_pending(&cachep->memcg_params->destroy))
3285 * We have to defer the actual destroying to a workqueue, because
3286 * we might currently be in a context that cannot sleep.
3288 schedule_work(&cachep->memcg_params->destroy);
3292 * This lock protects updaters, not readers. We want readers to be as fast as
3293 * they can, and they will either see NULL or a valid cache value. Our model
3294 * allow them to see NULL, in which case the root memcg will be selected.
3296 * We need this lock because multiple allocations to the same cache from a non
3297 * will span more than one worker. Only one of them can create the cache.
3299 static DEFINE_MUTEX(memcg_cache_mutex);
3302 * Called with memcg_cache_mutex held
3304 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3305 struct kmem_cache *s)
3307 struct kmem_cache *new;
3308 static char *tmp_name = NULL;
3310 lockdep_assert_held(&memcg_cache_mutex);
3313 * kmem_cache_create_memcg duplicates the given name and
3314 * cgroup_name for this name requires RCU context.
3315 * This static temporary buffer is used to prevent from
3316 * pointless shortliving allocation.
3319 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3325 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3326 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3329 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3330 (s->flags & ~SLAB_PANIC), s->ctor, s);
3333 new->allocflags |= __GFP_KMEMCG;
3338 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3339 struct kmem_cache *cachep)
3341 struct kmem_cache *new_cachep;
3344 BUG_ON(!memcg_can_account_kmem(memcg));
3346 idx = memcg_cache_id(memcg);
3348 mutex_lock(&memcg_cache_mutex);
3349 new_cachep = cachep->memcg_params->memcg_caches[idx];
3351 css_put(&memcg->css);
3355 new_cachep = kmem_cache_dup(memcg, cachep);
3356 if (new_cachep == NULL) {
3357 new_cachep = cachep;
3358 css_put(&memcg->css);
3362 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3364 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3366 * the readers won't lock, make sure everybody sees the updated value,
3367 * so they won't put stuff in the queue again for no reason
3371 mutex_unlock(&memcg_cache_mutex);
3375 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3377 struct kmem_cache *c;
3380 if (!s->memcg_params)
3382 if (!s->memcg_params->is_root_cache)
3386 * If the cache is being destroyed, we trust that there is no one else
3387 * requesting objects from it. Even if there are, the sanity checks in
3388 * kmem_cache_destroy should caught this ill-case.
3390 * Still, we don't want anyone else freeing memcg_caches under our
3391 * noses, which can happen if a new memcg comes to life. As usual,
3392 * we'll take the set_limit_mutex to protect ourselves against this.
3394 mutex_lock(&set_limit_mutex);
3395 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3396 c = s->memcg_params->memcg_caches[i];
3401 * We will now manually delete the caches, so to avoid races
3402 * we need to cancel all pending destruction workers and
3403 * proceed with destruction ourselves.
3405 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3406 * and that could spawn the workers again: it is likely that
3407 * the cache still have active pages until this very moment.
3408 * This would lead us back to mem_cgroup_destroy_cache.
3410 * But that will not execute at all if the "dead" flag is not
3411 * set, so flip it down to guarantee we are in control.
3413 c->memcg_params->dead = false;
3414 cancel_work_sync(&c->memcg_params->destroy);
3415 kmem_cache_destroy(c);
3417 mutex_unlock(&set_limit_mutex);
3420 struct create_work {
3421 struct mem_cgroup *memcg;
3422 struct kmem_cache *cachep;
3423 struct work_struct work;
3426 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3428 struct kmem_cache *cachep;
3429 struct memcg_cache_params *params;
3431 if (!memcg_kmem_is_active(memcg))
3434 mutex_lock(&memcg->slab_caches_mutex);
3435 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3436 cachep = memcg_params_to_cache(params);
3437 cachep->memcg_params->dead = true;
3438 schedule_work(&cachep->memcg_params->destroy);
3440 mutex_unlock(&memcg->slab_caches_mutex);
3443 static void memcg_create_cache_work_func(struct work_struct *w)
3445 struct create_work *cw;
3447 cw = container_of(w, struct create_work, work);
3448 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3453 * Enqueue the creation of a per-memcg kmem_cache.
3455 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3456 struct kmem_cache *cachep)
3458 struct create_work *cw;
3460 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3462 css_put(&memcg->css);
3467 cw->cachep = cachep;
3469 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3470 schedule_work(&cw->work);
3473 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3474 struct kmem_cache *cachep)
3477 * We need to stop accounting when we kmalloc, because if the
3478 * corresponding kmalloc cache is not yet created, the first allocation
3479 * in __memcg_create_cache_enqueue will recurse.
3481 * However, it is better to enclose the whole function. Depending on
3482 * the debugging options enabled, INIT_WORK(), for instance, can
3483 * trigger an allocation. This too, will make us recurse. Because at
3484 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3485 * the safest choice is to do it like this, wrapping the whole function.
3487 memcg_stop_kmem_account();
3488 __memcg_create_cache_enqueue(memcg, cachep);
3489 memcg_resume_kmem_account();
3492 * Return the kmem_cache we're supposed to use for a slab allocation.
3493 * We try to use the current memcg's version of the cache.
3495 * If the cache does not exist yet, if we are the first user of it,
3496 * we either create it immediately, if possible, or create it asynchronously
3498 * In the latter case, we will let the current allocation go through with
3499 * the original cache.
3501 * Can't be called in interrupt context or from kernel threads.
3502 * This function needs to be called with rcu_read_lock() held.
3504 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3507 struct mem_cgroup *memcg;
3510 VM_BUG_ON(!cachep->memcg_params);
3511 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3513 if (!current->mm || current->memcg_kmem_skip_account)
3517 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3519 if (!memcg_can_account_kmem(memcg))
3522 idx = memcg_cache_id(memcg);
3525 * barrier to mare sure we're always seeing the up to date value. The
3526 * code updating memcg_caches will issue a write barrier to match this.
3528 read_barrier_depends();
3529 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3530 cachep = cachep->memcg_params->memcg_caches[idx];
3534 /* The corresponding put will be done in the workqueue. */
3535 if (!css_tryget(&memcg->css))
3540 * If we are in a safe context (can wait, and not in interrupt
3541 * context), we could be be predictable and return right away.
3542 * This would guarantee that the allocation being performed
3543 * already belongs in the new cache.
3545 * However, there are some clashes that can arrive from locking.
3546 * For instance, because we acquire the slab_mutex while doing
3547 * kmem_cache_dup, this means no further allocation could happen
3548 * with the slab_mutex held.
3550 * Also, because cache creation issue get_online_cpus(), this
3551 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3552 * that ends up reversed during cpu hotplug. (cpuset allocates
3553 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3554 * better to defer everything.
3556 memcg_create_cache_enqueue(memcg, cachep);
3562 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3565 * We need to verify if the allocation against current->mm->owner's memcg is
3566 * possible for the given order. But the page is not allocated yet, so we'll
3567 * need a further commit step to do the final arrangements.
3569 * It is possible for the task to switch cgroups in this mean time, so at
3570 * commit time, we can't rely on task conversion any longer. We'll then use
3571 * the handle argument to return to the caller which cgroup we should commit
3572 * against. We could also return the memcg directly and avoid the pointer
3573 * passing, but a boolean return value gives better semantics considering
3574 * the compiled-out case as well.
3576 * Returning true means the allocation is possible.
3579 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3581 struct mem_cgroup *memcg;
3587 * Disabling accounting is only relevant for some specific memcg
3588 * internal allocations. Therefore we would initially not have such
3589 * check here, since direct calls to the page allocator that are marked
3590 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3591 * concerned with cache allocations, and by having this test at
3592 * memcg_kmem_get_cache, we are already able to relay the allocation to
3593 * the root cache and bypass the memcg cache altogether.
3595 * There is one exception, though: the SLUB allocator does not create
3596 * large order caches, but rather service large kmallocs directly from
3597 * the page allocator. Therefore, the following sequence when backed by
3598 * the SLUB allocator:
3600 * memcg_stop_kmem_account();
3601 * kmalloc(<large_number>)
3602 * memcg_resume_kmem_account();
3604 * would effectively ignore the fact that we should skip accounting,
3605 * since it will drive us directly to this function without passing
3606 * through the cache selector memcg_kmem_get_cache. Such large
3607 * allocations are extremely rare but can happen, for instance, for the
3608 * cache arrays. We bring this test here.
3610 if (!current->mm || current->memcg_kmem_skip_account)
3613 memcg = try_get_mem_cgroup_from_mm(current->mm);
3616 * very rare case described in mem_cgroup_from_task. Unfortunately there
3617 * isn't much we can do without complicating this too much, and it would
3618 * be gfp-dependent anyway. Just let it go
3620 if (unlikely(!memcg))
3623 if (!memcg_can_account_kmem(memcg)) {
3624 css_put(&memcg->css);
3628 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3632 css_put(&memcg->css);
3636 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3639 struct page_cgroup *pc;
3641 VM_BUG_ON(mem_cgroup_is_root(memcg));
3643 /* The page allocation failed. Revert */
3645 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3649 pc = lookup_page_cgroup(page);
3650 lock_page_cgroup(pc);
3651 pc->mem_cgroup = memcg;
3652 SetPageCgroupUsed(pc);
3653 unlock_page_cgroup(pc);
3656 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3658 struct mem_cgroup *memcg = NULL;
3659 struct page_cgroup *pc;
3662 pc = lookup_page_cgroup(page);
3664 * Fast unlocked return. Theoretically might have changed, have to
3665 * check again after locking.
3667 if (!PageCgroupUsed(pc))
3670 lock_page_cgroup(pc);
3671 if (PageCgroupUsed(pc)) {
3672 memcg = pc->mem_cgroup;
3673 ClearPageCgroupUsed(pc);
3675 unlock_page_cgroup(pc);
3678 * We trust that only if there is a memcg associated with the page, it
3679 * is a valid allocation
3684 VM_BUG_ON(mem_cgroup_is_root(memcg));
3685 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3688 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3691 #endif /* CONFIG_MEMCG_KMEM */
3693 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3695 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3697 * Because tail pages are not marked as "used", set it. We're under
3698 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3699 * charge/uncharge will be never happen and move_account() is done under
3700 * compound_lock(), so we don't have to take care of races.
3702 void mem_cgroup_split_huge_fixup(struct page *head)
3704 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3705 struct page_cgroup *pc;
3706 struct mem_cgroup *memcg;
3709 if (mem_cgroup_disabled())
3712 memcg = head_pc->mem_cgroup;
3713 for (i = 1; i < HPAGE_PMD_NR; i++) {
3715 pc->mem_cgroup = memcg;
3716 smp_wmb();/* see __commit_charge() */
3717 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3719 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3722 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3725 * mem_cgroup_move_account - move account of the page
3727 * @nr_pages: number of regular pages (>1 for huge pages)
3728 * @pc: page_cgroup of the page.
3729 * @from: mem_cgroup which the page is moved from.
3730 * @to: mem_cgroup which the page is moved to. @from != @to.
3732 * The caller must confirm following.
3733 * - page is not on LRU (isolate_page() is useful.)
3734 * - compound_lock is held when nr_pages > 1
3736 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3739 static int mem_cgroup_move_account(struct page *page,
3740 unsigned int nr_pages,
3741 struct page_cgroup *pc,
3742 struct mem_cgroup *from,
3743 struct mem_cgroup *to)
3745 unsigned long flags;
3747 bool anon = PageAnon(page);
3749 VM_BUG_ON(from == to);
3750 VM_BUG_ON(PageLRU(page));
3752 * The page is isolated from LRU. So, collapse function
3753 * will not handle this page. But page splitting can happen.
3754 * Do this check under compound_page_lock(). The caller should
3758 if (nr_pages > 1 && !PageTransHuge(page))
3761 lock_page_cgroup(pc);
3764 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3767 move_lock_mem_cgroup(from, &flags);
3769 if (!anon && page_mapped(page)) {
3770 /* Update mapped_file data for mem_cgroup */
3772 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3773 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3776 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3778 /* caller should have done css_get */
3779 pc->mem_cgroup = to;
3780 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3781 move_unlock_mem_cgroup(from, &flags);
3784 unlock_page_cgroup(pc);
3788 memcg_check_events(to, page);
3789 memcg_check_events(from, page);
3795 * mem_cgroup_move_parent - moves page to the parent group
3796 * @page: the page to move
3797 * @pc: page_cgroup of the page
3798 * @child: page's cgroup
3800 * move charges to its parent or the root cgroup if the group has no
3801 * parent (aka use_hierarchy==0).
3802 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3803 * mem_cgroup_move_account fails) the failure is always temporary and
3804 * it signals a race with a page removal/uncharge or migration. In the
3805 * first case the page is on the way out and it will vanish from the LRU
3806 * on the next attempt and the call should be retried later.
3807 * Isolation from the LRU fails only if page has been isolated from
3808 * the LRU since we looked at it and that usually means either global
3809 * reclaim or migration going on. The page will either get back to the
3811 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3812 * (!PageCgroupUsed) or moved to a different group. The page will
3813 * disappear in the next attempt.
3815 static int mem_cgroup_move_parent(struct page *page,
3816 struct page_cgroup *pc,
3817 struct mem_cgroup *child)
3819 struct mem_cgroup *parent;
3820 unsigned int nr_pages;
3821 unsigned long uninitialized_var(flags);
3824 VM_BUG_ON(mem_cgroup_is_root(child));
3827 if (!get_page_unless_zero(page))
3829 if (isolate_lru_page(page))
3832 nr_pages = hpage_nr_pages(page);
3834 parent = parent_mem_cgroup(child);
3836 * If no parent, move charges to root cgroup.
3839 parent = root_mem_cgroup;
3842 VM_BUG_ON(!PageTransHuge(page));
3843 flags = compound_lock_irqsave(page);
3846 ret = mem_cgroup_move_account(page, nr_pages,
3849 __mem_cgroup_cancel_local_charge(child, nr_pages);
3852 compound_unlock_irqrestore(page, flags);
3853 putback_lru_page(page);
3861 * Charge the memory controller for page usage.
3863 * 0 if the charge was successful
3864 * < 0 if the cgroup is over its limit
3866 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3867 gfp_t gfp_mask, enum charge_type ctype)
3869 struct mem_cgroup *memcg = NULL;
3870 unsigned int nr_pages = 1;
3874 if (PageTransHuge(page)) {
3875 nr_pages <<= compound_order(page);
3876 VM_BUG_ON(!PageTransHuge(page));
3878 * Never OOM-kill a process for a huge page. The
3879 * fault handler will fall back to regular pages.
3884 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3887 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3891 int mem_cgroup_newpage_charge(struct page *page,
3892 struct mm_struct *mm, gfp_t gfp_mask)
3894 if (mem_cgroup_disabled())
3896 VM_BUG_ON(page_mapped(page));
3897 VM_BUG_ON(page->mapping && !PageAnon(page));
3899 return mem_cgroup_charge_common(page, mm, gfp_mask,
3900 MEM_CGROUP_CHARGE_TYPE_ANON);
3904 * While swap-in, try_charge -> commit or cancel, the page is locked.
3905 * And when try_charge() successfully returns, one refcnt to memcg without
3906 * struct page_cgroup is acquired. This refcnt will be consumed by
3907 * "commit()" or removed by "cancel()"
3909 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3912 struct mem_cgroup **memcgp)
3914 struct mem_cgroup *memcg;
3915 struct page_cgroup *pc;
3918 pc = lookup_page_cgroup(page);
3920 * Every swap fault against a single page tries to charge the
3921 * page, bail as early as possible. shmem_unuse() encounters
3922 * already charged pages, too. The USED bit is protected by
3923 * the page lock, which serializes swap cache removal, which
3924 * in turn serializes uncharging.
3926 if (PageCgroupUsed(pc))
3928 if (!do_swap_account)
3930 memcg = try_get_mem_cgroup_from_page(page);
3934 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3935 css_put(&memcg->css);
3940 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3946 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3947 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3950 if (mem_cgroup_disabled())
3953 * A racing thread's fault, or swapoff, may have already
3954 * updated the pte, and even removed page from swap cache: in
3955 * those cases unuse_pte()'s pte_same() test will fail; but
3956 * there's also a KSM case which does need to charge the page.
3958 if (!PageSwapCache(page)) {
3961 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3966 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3969 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3971 if (mem_cgroup_disabled())
3975 __mem_cgroup_cancel_charge(memcg, 1);
3979 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
3980 enum charge_type ctype)
3982 if (mem_cgroup_disabled())
3987 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
3989 * Now swap is on-memory. This means this page may be
3990 * counted both as mem and swap....double count.
3991 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
3992 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
3993 * may call delete_from_swap_cache() before reach here.
3995 if (do_swap_account && PageSwapCache(page)) {
3996 swp_entry_t ent = {.val = page_private(page)};
3997 mem_cgroup_uncharge_swap(ent);
4001 void mem_cgroup_commit_charge_swapin(struct page *page,
4002 struct mem_cgroup *memcg)
4004 __mem_cgroup_commit_charge_swapin(page, memcg,
4005 MEM_CGROUP_CHARGE_TYPE_ANON);
4008 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4011 struct mem_cgroup *memcg = NULL;
4012 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4015 if (mem_cgroup_disabled())
4017 if (PageCompound(page))
4020 if (!PageSwapCache(page))
4021 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4022 else { /* page is swapcache/shmem */
4023 ret = __mem_cgroup_try_charge_swapin(mm, page,
4026 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4031 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4032 unsigned int nr_pages,
4033 const enum charge_type ctype)
4035 struct memcg_batch_info *batch = NULL;
4036 bool uncharge_memsw = true;
4038 /* If swapout, usage of swap doesn't decrease */
4039 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4040 uncharge_memsw = false;
4042 batch = ¤t->memcg_batch;
4044 * In usual, we do css_get() when we remember memcg pointer.
4045 * But in this case, we keep res->usage until end of a series of
4046 * uncharges. Then, it's ok to ignore memcg's refcnt.
4049 batch->memcg = memcg;
4051 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4052 * In those cases, all pages freed continuously can be expected to be in
4053 * the same cgroup and we have chance to coalesce uncharges.
4054 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4055 * because we want to do uncharge as soon as possible.
4058 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4059 goto direct_uncharge;
4062 goto direct_uncharge;
4065 * In typical case, batch->memcg == mem. This means we can
4066 * merge a series of uncharges to an uncharge of res_counter.
4067 * If not, we uncharge res_counter ony by one.
4069 if (batch->memcg != memcg)
4070 goto direct_uncharge;
4071 /* remember freed charge and uncharge it later */
4074 batch->memsw_nr_pages++;
4077 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4079 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4080 if (unlikely(batch->memcg != memcg))
4081 memcg_oom_recover(memcg);
4085 * uncharge if !page_mapped(page)
4087 static struct mem_cgroup *
4088 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4091 struct mem_cgroup *memcg = NULL;
4092 unsigned int nr_pages = 1;
4093 struct page_cgroup *pc;
4096 if (mem_cgroup_disabled())
4099 if (PageTransHuge(page)) {
4100 nr_pages <<= compound_order(page);
4101 VM_BUG_ON(!PageTransHuge(page));
4104 * Check if our page_cgroup is valid
4106 pc = lookup_page_cgroup(page);
4107 if (unlikely(!PageCgroupUsed(pc)))
4110 lock_page_cgroup(pc);
4112 memcg = pc->mem_cgroup;
4114 if (!PageCgroupUsed(pc))
4117 anon = PageAnon(page);
4120 case MEM_CGROUP_CHARGE_TYPE_ANON:
4122 * Generally PageAnon tells if it's the anon statistics to be
4123 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4124 * used before page reached the stage of being marked PageAnon.
4128 case MEM_CGROUP_CHARGE_TYPE_DROP:
4129 /* See mem_cgroup_prepare_migration() */
4130 if (page_mapped(page))
4133 * Pages under migration may not be uncharged. But
4134 * end_migration() /must/ be the one uncharging the
4135 * unused post-migration page and so it has to call
4136 * here with the migration bit still set. See the
4137 * res_counter handling below.
4139 if (!end_migration && PageCgroupMigration(pc))
4142 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4143 if (!PageAnon(page)) { /* Shared memory */
4144 if (page->mapping && !page_is_file_cache(page))
4146 } else if (page_mapped(page)) /* Anon */
4153 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4155 ClearPageCgroupUsed(pc);
4157 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4158 * freed from LRU. This is safe because uncharged page is expected not
4159 * to be reused (freed soon). Exception is SwapCache, it's handled by
4160 * special functions.
4163 unlock_page_cgroup(pc);
4165 * even after unlock, we have memcg->res.usage here and this memcg
4166 * will never be freed, so it's safe to call css_get().
4168 memcg_check_events(memcg, page);
4169 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4170 mem_cgroup_swap_statistics(memcg, true);
4171 css_get(&memcg->css);
4174 * Migration does not charge the res_counter for the
4175 * replacement page, so leave it alone when phasing out the
4176 * page that is unused after the migration.
4178 if (!end_migration && !mem_cgroup_is_root(memcg))
4179 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4184 unlock_page_cgroup(pc);
4188 void mem_cgroup_uncharge_page(struct page *page)
4191 if (page_mapped(page))
4193 VM_BUG_ON(page->mapping && !PageAnon(page));
4195 * If the page is in swap cache, uncharge should be deferred
4196 * to the swap path, which also properly accounts swap usage
4197 * and handles memcg lifetime.
4199 * Note that this check is not stable and reclaim may add the
4200 * page to swap cache at any time after this. However, if the
4201 * page is not in swap cache by the time page->mapcount hits
4202 * 0, there won't be any page table references to the swap
4203 * slot, and reclaim will free it and not actually write the
4206 if (PageSwapCache(page))
4208 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4211 void mem_cgroup_uncharge_cache_page(struct page *page)
4213 VM_BUG_ON(page_mapped(page));
4214 VM_BUG_ON(page->mapping);
4215 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4219 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4220 * In that cases, pages are freed continuously and we can expect pages
4221 * are in the same memcg. All these calls itself limits the number of
4222 * pages freed at once, then uncharge_start/end() is called properly.
4223 * This may be called prural(2) times in a context,
4226 void mem_cgroup_uncharge_start(void)
4228 current->memcg_batch.do_batch++;
4229 /* We can do nest. */
4230 if (current->memcg_batch.do_batch == 1) {
4231 current->memcg_batch.memcg = NULL;
4232 current->memcg_batch.nr_pages = 0;
4233 current->memcg_batch.memsw_nr_pages = 0;
4237 void mem_cgroup_uncharge_end(void)
4239 struct memcg_batch_info *batch = ¤t->memcg_batch;
4241 if (!batch->do_batch)
4245 if (batch->do_batch) /* If stacked, do nothing. */
4251 * This "batch->memcg" is valid without any css_get/put etc...
4252 * bacause we hide charges behind us.
4254 if (batch->nr_pages)
4255 res_counter_uncharge(&batch->memcg->res,
4256 batch->nr_pages * PAGE_SIZE);
4257 if (batch->memsw_nr_pages)
4258 res_counter_uncharge(&batch->memcg->memsw,
4259 batch->memsw_nr_pages * PAGE_SIZE);
4260 memcg_oom_recover(batch->memcg);
4261 /* forget this pointer (for sanity check) */
4262 batch->memcg = NULL;
4267 * called after __delete_from_swap_cache() and drop "page" account.
4268 * memcg information is recorded to swap_cgroup of "ent"
4271 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4273 struct mem_cgroup *memcg;
4274 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4276 if (!swapout) /* this was a swap cache but the swap is unused ! */
4277 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4279 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4282 * record memcg information, if swapout && memcg != NULL,
4283 * css_get() was called in uncharge().
4285 if (do_swap_account && swapout && memcg)
4286 swap_cgroup_record(ent, css_id(&memcg->css));
4290 #ifdef CONFIG_MEMCG_SWAP
4292 * called from swap_entry_free(). remove record in swap_cgroup and
4293 * uncharge "memsw" account.
4295 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4297 struct mem_cgroup *memcg;
4300 if (!do_swap_account)
4303 id = swap_cgroup_record(ent, 0);
4305 memcg = mem_cgroup_lookup(id);
4308 * We uncharge this because swap is freed.
4309 * This memcg can be obsolete one. We avoid calling css_tryget
4311 if (!mem_cgroup_is_root(memcg))
4312 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4313 mem_cgroup_swap_statistics(memcg, false);
4314 css_put(&memcg->css);
4320 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4321 * @entry: swap entry to be moved
4322 * @from: mem_cgroup which the entry is moved from
4323 * @to: mem_cgroup which the entry is moved to
4325 * It succeeds only when the swap_cgroup's record for this entry is the same
4326 * as the mem_cgroup's id of @from.
4328 * Returns 0 on success, -EINVAL on failure.
4330 * The caller must have charged to @to, IOW, called res_counter_charge() about
4331 * both res and memsw, and called css_get().
4333 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4334 struct mem_cgroup *from, struct mem_cgroup *to)
4336 unsigned short old_id, new_id;
4338 old_id = css_id(&from->css);
4339 new_id = css_id(&to->css);
4341 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4342 mem_cgroup_swap_statistics(from, false);
4343 mem_cgroup_swap_statistics(to, true);
4345 * This function is only called from task migration context now.
4346 * It postpones res_counter and refcount handling till the end
4347 * of task migration(mem_cgroup_clear_mc()) for performance
4348 * improvement. But we cannot postpone css_get(to) because if
4349 * the process that has been moved to @to does swap-in, the
4350 * refcount of @to might be decreased to 0.
4352 * We are in attach() phase, so the cgroup is guaranteed to be
4353 * alive, so we can just call css_get().
4361 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4362 struct mem_cgroup *from, struct mem_cgroup *to)
4369 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4372 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4373 struct mem_cgroup **memcgp)
4375 struct mem_cgroup *memcg = NULL;
4376 unsigned int nr_pages = 1;
4377 struct page_cgroup *pc;
4378 enum charge_type ctype;
4382 if (mem_cgroup_disabled())
4385 if (PageTransHuge(page))
4386 nr_pages <<= compound_order(page);
4388 pc = lookup_page_cgroup(page);
4389 lock_page_cgroup(pc);
4390 if (PageCgroupUsed(pc)) {
4391 memcg = pc->mem_cgroup;
4392 css_get(&memcg->css);
4394 * At migrating an anonymous page, its mapcount goes down
4395 * to 0 and uncharge() will be called. But, even if it's fully
4396 * unmapped, migration may fail and this page has to be
4397 * charged again. We set MIGRATION flag here and delay uncharge
4398 * until end_migration() is called
4400 * Corner Case Thinking
4402 * When the old page was mapped as Anon and it's unmap-and-freed
4403 * while migration was ongoing.
4404 * If unmap finds the old page, uncharge() of it will be delayed
4405 * until end_migration(). If unmap finds a new page, it's
4406 * uncharged when it make mapcount to be 1->0. If unmap code
4407 * finds swap_migration_entry, the new page will not be mapped
4408 * and end_migration() will find it(mapcount==0).
4411 * When the old page was mapped but migraion fails, the kernel
4412 * remaps it. A charge for it is kept by MIGRATION flag even
4413 * if mapcount goes down to 0. We can do remap successfully
4414 * without charging it again.
4417 * The "old" page is under lock_page() until the end of
4418 * migration, so, the old page itself will not be swapped-out.
4419 * If the new page is swapped out before end_migraton, our
4420 * hook to usual swap-out path will catch the event.
4423 SetPageCgroupMigration(pc);
4425 unlock_page_cgroup(pc);
4427 * If the page is not charged at this point,
4435 * We charge new page before it's used/mapped. So, even if unlock_page()
4436 * is called before end_migration, we can catch all events on this new
4437 * page. In the case new page is migrated but not remapped, new page's
4438 * mapcount will be finally 0 and we call uncharge in end_migration().
4441 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4443 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4445 * The page is committed to the memcg, but it's not actually
4446 * charged to the res_counter since we plan on replacing the
4447 * old one and only one page is going to be left afterwards.
4449 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4452 /* remove redundant charge if migration failed*/
4453 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4454 struct page *oldpage, struct page *newpage, bool migration_ok)
4456 struct page *used, *unused;
4457 struct page_cgroup *pc;
4463 if (!migration_ok) {
4470 anon = PageAnon(used);
4471 __mem_cgroup_uncharge_common(unused,
4472 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4473 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4475 css_put(&memcg->css);
4477 * We disallowed uncharge of pages under migration because mapcount
4478 * of the page goes down to zero, temporarly.
4479 * Clear the flag and check the page should be charged.
4481 pc = lookup_page_cgroup(oldpage);
4482 lock_page_cgroup(pc);
4483 ClearPageCgroupMigration(pc);
4484 unlock_page_cgroup(pc);
4487 * If a page is a file cache, radix-tree replacement is very atomic
4488 * and we can skip this check. When it was an Anon page, its mapcount
4489 * goes down to 0. But because we added MIGRATION flage, it's not
4490 * uncharged yet. There are several case but page->mapcount check
4491 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4492 * check. (see prepare_charge() also)
4495 mem_cgroup_uncharge_page(used);
4499 * At replace page cache, newpage is not under any memcg but it's on
4500 * LRU. So, this function doesn't touch res_counter but handles LRU
4501 * in correct way. Both pages are locked so we cannot race with uncharge.
4503 void mem_cgroup_replace_page_cache(struct page *oldpage,
4504 struct page *newpage)
4506 struct mem_cgroup *memcg = NULL;
4507 struct page_cgroup *pc;
4508 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4510 if (mem_cgroup_disabled())
4513 pc = lookup_page_cgroup(oldpage);
4514 /* fix accounting on old pages */
4515 lock_page_cgroup(pc);
4516 if (PageCgroupUsed(pc)) {
4517 memcg = pc->mem_cgroup;
4518 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4519 ClearPageCgroupUsed(pc);
4521 unlock_page_cgroup(pc);
4524 * When called from shmem_replace_page(), in some cases the
4525 * oldpage has already been charged, and in some cases not.
4530 * Even if newpage->mapping was NULL before starting replacement,
4531 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4532 * LRU while we overwrite pc->mem_cgroup.
4534 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4537 #ifdef CONFIG_DEBUG_VM
4538 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4540 struct page_cgroup *pc;
4542 pc = lookup_page_cgroup(page);
4544 * Can be NULL while feeding pages into the page allocator for
4545 * the first time, i.e. during boot or memory hotplug;
4546 * or when mem_cgroup_disabled().
4548 if (likely(pc) && PageCgroupUsed(pc))
4553 bool mem_cgroup_bad_page_check(struct page *page)
4555 if (mem_cgroup_disabled())
4558 return lookup_page_cgroup_used(page) != NULL;
4561 void mem_cgroup_print_bad_page(struct page *page)
4563 struct page_cgroup *pc;
4565 pc = lookup_page_cgroup_used(page);
4567 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4568 pc, pc->flags, pc->mem_cgroup);
4573 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4574 unsigned long long val)
4577 u64 memswlimit, memlimit;
4579 int children = mem_cgroup_count_children(memcg);
4580 u64 curusage, oldusage;
4584 * For keeping hierarchical_reclaim simple, how long we should retry
4585 * is depends on callers. We set our retry-count to be function
4586 * of # of children which we should visit in this loop.
4588 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4590 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4593 while (retry_count) {
4594 if (signal_pending(current)) {
4599 * Rather than hide all in some function, I do this in
4600 * open coded manner. You see what this really does.
4601 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4603 mutex_lock(&set_limit_mutex);
4604 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4605 if (memswlimit < val) {
4607 mutex_unlock(&set_limit_mutex);
4611 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4615 ret = res_counter_set_limit(&memcg->res, val);
4617 if (memswlimit == val)
4618 memcg->memsw_is_minimum = true;
4620 memcg->memsw_is_minimum = false;
4622 mutex_unlock(&set_limit_mutex);
4627 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4628 MEM_CGROUP_RECLAIM_SHRINK);
4629 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4630 /* Usage is reduced ? */
4631 if (curusage >= oldusage)
4634 oldusage = curusage;
4636 if (!ret && enlarge)
4637 memcg_oom_recover(memcg);
4642 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4643 unsigned long long val)
4646 u64 memlimit, memswlimit, oldusage, curusage;
4647 int children = mem_cgroup_count_children(memcg);
4651 /* see mem_cgroup_resize_res_limit */
4652 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4653 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4654 while (retry_count) {
4655 if (signal_pending(current)) {
4660 * Rather than hide all in some function, I do this in
4661 * open coded manner. You see what this really does.
4662 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4664 mutex_lock(&set_limit_mutex);
4665 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4666 if (memlimit > val) {
4668 mutex_unlock(&set_limit_mutex);
4671 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4672 if (memswlimit < val)
4674 ret = res_counter_set_limit(&memcg->memsw, val);
4676 if (memlimit == val)
4677 memcg->memsw_is_minimum = true;
4679 memcg->memsw_is_minimum = false;
4681 mutex_unlock(&set_limit_mutex);
4686 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4687 MEM_CGROUP_RECLAIM_NOSWAP |
4688 MEM_CGROUP_RECLAIM_SHRINK);
4689 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4690 /* Usage is reduced ? */
4691 if (curusage >= oldusage)
4694 oldusage = curusage;
4696 if (!ret && enlarge)
4697 memcg_oom_recover(memcg);
4702 * mem_cgroup_force_empty_list - clears LRU of a group
4703 * @memcg: group to clear
4706 * @lru: lru to to clear
4708 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4709 * reclaim the pages page themselves - pages are moved to the parent (or root)
4712 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4713 int node, int zid, enum lru_list lru)
4715 struct lruvec *lruvec;
4716 unsigned long flags;
4717 struct list_head *list;
4721 zone = &NODE_DATA(node)->node_zones[zid];
4722 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4723 list = &lruvec->lists[lru];
4727 struct page_cgroup *pc;
4730 spin_lock_irqsave(&zone->lru_lock, flags);
4731 if (list_empty(list)) {
4732 spin_unlock_irqrestore(&zone->lru_lock, flags);
4735 page = list_entry(list->prev, struct page, lru);
4737 list_move(&page->lru, list);
4739 spin_unlock_irqrestore(&zone->lru_lock, flags);
4742 spin_unlock_irqrestore(&zone->lru_lock, flags);
4744 pc = lookup_page_cgroup(page);
4746 if (mem_cgroup_move_parent(page, pc, memcg)) {
4747 /* found lock contention or "pc" is obsolete. */
4752 } while (!list_empty(list));
4756 * make mem_cgroup's charge to be 0 if there is no task by moving
4757 * all the charges and pages to the parent.
4758 * This enables deleting this mem_cgroup.
4760 * Caller is responsible for holding css reference on the memcg.
4762 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4768 /* This is for making all *used* pages to be on LRU. */
4769 lru_add_drain_all();
4770 drain_all_stock_sync(memcg);
4771 mem_cgroup_start_move(memcg);
4772 for_each_node_state(node, N_MEMORY) {
4773 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4776 mem_cgroup_force_empty_list(memcg,
4781 mem_cgroup_end_move(memcg);
4782 memcg_oom_recover(memcg);
4786 * Kernel memory may not necessarily be trackable to a specific
4787 * process. So they are not migrated, and therefore we can't
4788 * expect their value to drop to 0 here.
4789 * Having res filled up with kmem only is enough.
4791 * This is a safety check because mem_cgroup_force_empty_list
4792 * could have raced with mem_cgroup_replace_page_cache callers
4793 * so the lru seemed empty but the page could have been added
4794 * right after the check. RES_USAGE should be safe as we always
4795 * charge before adding to the LRU.
4797 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4798 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4799 } while (usage > 0);
4803 * This mainly exists for tests during the setting of set of use_hierarchy.
4804 * Since this is the very setting we are changing, the current hierarchy value
4807 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4809 struct cgroup_subsys_state *pos;
4811 /* bounce at first found */
4812 css_for_each_child(pos, &memcg->css)
4818 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4819 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4820 * from mem_cgroup_count_children(), in the sense that we don't really care how
4821 * many children we have; we only need to know if we have any. It also counts
4822 * any memcg without hierarchy as infertile.
4824 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4826 return memcg->use_hierarchy && __memcg_has_children(memcg);
4830 * Reclaims as many pages from the given memcg as possible and moves
4831 * the rest to the parent.
4833 * Caller is responsible for holding css reference for memcg.
4835 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4837 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4838 struct cgroup *cgrp = memcg->css.cgroup;
4840 /* returns EBUSY if there is a task or if we come here twice. */
4841 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4844 /* we call try-to-free pages for make this cgroup empty */
4845 lru_add_drain_all();
4846 /* try to free all pages in this cgroup */
4847 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4850 if (signal_pending(current))
4853 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4857 /* maybe some writeback is necessary */
4858 congestion_wait(BLK_RW_ASYNC, HZ/10);
4863 mem_cgroup_reparent_charges(memcg);
4868 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4871 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4873 if (mem_cgroup_is_root(memcg))
4875 return mem_cgroup_force_empty(memcg);
4878 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
4881 return mem_cgroup_from_css(css)->use_hierarchy;
4884 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
4885 struct cftype *cft, u64 val)
4888 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4889 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
4891 mutex_lock(&memcg_create_mutex);
4893 if (memcg->use_hierarchy == val)
4897 * If parent's use_hierarchy is set, we can't make any modifications
4898 * in the child subtrees. If it is unset, then the change can
4899 * occur, provided the current cgroup has no children.
4901 * For the root cgroup, parent_mem is NULL, we allow value to be
4902 * set if there are no children.
4904 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
4905 (val == 1 || val == 0)) {
4906 if (!__memcg_has_children(memcg))
4907 memcg->use_hierarchy = val;
4914 mutex_unlock(&memcg_create_mutex);
4920 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
4921 enum mem_cgroup_stat_index idx)
4923 struct mem_cgroup *iter;
4926 /* Per-cpu values can be negative, use a signed accumulator */
4927 for_each_mem_cgroup_tree(iter, memcg)
4928 val += mem_cgroup_read_stat(iter, idx);
4930 if (val < 0) /* race ? */
4935 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
4939 if (!mem_cgroup_is_root(memcg)) {
4941 return res_counter_read_u64(&memcg->res, RES_USAGE);
4943 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
4947 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
4948 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
4950 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
4951 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
4954 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
4956 return val << PAGE_SHIFT;
4959 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
4960 struct cftype *cft, struct file *file,
4961 char __user *buf, size_t nbytes, loff_t *ppos)
4963 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4969 type = MEMFILE_TYPE(cft->private);
4970 name = MEMFILE_ATTR(cft->private);
4974 if (name == RES_USAGE)
4975 val = mem_cgroup_usage(memcg, false);
4977 val = res_counter_read_u64(&memcg->res, name);
4980 if (name == RES_USAGE)
4981 val = mem_cgroup_usage(memcg, true);
4983 val = res_counter_read_u64(&memcg->memsw, name);
4986 val = res_counter_read_u64(&memcg->kmem, name);
4992 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
4993 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
4996 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
4999 #ifdef CONFIG_MEMCG_KMEM
5000 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5002 * For simplicity, we won't allow this to be disabled. It also can't
5003 * be changed if the cgroup has children already, or if tasks had
5006 * If tasks join before we set the limit, a person looking at
5007 * kmem.usage_in_bytes will have no way to determine when it took
5008 * place, which makes the value quite meaningless.
5010 * After it first became limited, changes in the value of the limit are
5011 * of course permitted.
5013 mutex_lock(&memcg_create_mutex);
5014 mutex_lock(&set_limit_mutex);
5015 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
5016 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
5020 ret = res_counter_set_limit(&memcg->kmem, val);
5023 ret = memcg_update_cache_sizes(memcg);
5025 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5028 static_key_slow_inc(&memcg_kmem_enabled_key);
5030 * setting the active bit after the inc will guarantee no one
5031 * starts accounting before all call sites are patched
5033 memcg_kmem_set_active(memcg);
5035 ret = res_counter_set_limit(&memcg->kmem, val);
5037 mutex_unlock(&set_limit_mutex);
5038 mutex_unlock(&memcg_create_mutex);
5043 #ifdef CONFIG_MEMCG_KMEM
5044 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5047 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5051 memcg->kmem_account_flags = parent->kmem_account_flags;
5053 * When that happen, we need to disable the static branch only on those
5054 * memcgs that enabled it. To achieve this, we would be forced to
5055 * complicate the code by keeping track of which memcgs were the ones
5056 * that actually enabled limits, and which ones got it from its
5059 * It is a lot simpler just to do static_key_slow_inc() on every child
5060 * that is accounted.
5062 if (!memcg_kmem_is_active(memcg))
5066 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5067 * memcg is active already. If the later initialization fails then the
5068 * cgroup core triggers the cleanup so we do not have to do it here.
5070 static_key_slow_inc(&memcg_kmem_enabled_key);
5072 mutex_lock(&set_limit_mutex);
5073 memcg_stop_kmem_account();
5074 ret = memcg_update_cache_sizes(memcg);
5075 memcg_resume_kmem_account();
5076 mutex_unlock(&set_limit_mutex);
5080 #endif /* CONFIG_MEMCG_KMEM */
5083 * The user of this function is...
5086 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5089 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5092 unsigned long long val;
5095 type = MEMFILE_TYPE(cft->private);
5096 name = MEMFILE_ATTR(cft->private);
5100 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5104 /* This function does all necessary parse...reuse it */
5105 ret = res_counter_memparse_write_strategy(buffer, &val);
5109 ret = mem_cgroup_resize_limit(memcg, val);
5110 else if (type == _MEMSWAP)
5111 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5112 else if (type == _KMEM)
5113 ret = memcg_update_kmem_limit(css, val);
5117 case RES_SOFT_LIMIT:
5118 ret = res_counter_memparse_write_strategy(buffer, &val);
5122 * For memsw, soft limits are hard to implement in terms
5123 * of semantics, for now, we support soft limits for
5124 * control without swap
5127 ret = res_counter_set_soft_limit(&memcg->res, val);
5132 ret = -EINVAL; /* should be BUG() ? */
5138 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5139 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5141 unsigned long long min_limit, min_memsw_limit, tmp;
5143 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5144 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5145 if (!memcg->use_hierarchy)
5148 while (css_parent(&memcg->css)) {
5149 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5150 if (!memcg->use_hierarchy)
5152 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5153 min_limit = min(min_limit, tmp);
5154 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5155 min_memsw_limit = min(min_memsw_limit, tmp);
5158 *mem_limit = min_limit;
5159 *memsw_limit = min_memsw_limit;
5162 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5164 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5168 type = MEMFILE_TYPE(event);
5169 name = MEMFILE_ATTR(event);
5174 res_counter_reset_max(&memcg->res);
5175 else if (type == _MEMSWAP)
5176 res_counter_reset_max(&memcg->memsw);
5177 else if (type == _KMEM)
5178 res_counter_reset_max(&memcg->kmem);
5184 res_counter_reset_failcnt(&memcg->res);
5185 else if (type == _MEMSWAP)
5186 res_counter_reset_failcnt(&memcg->memsw);
5187 else if (type == _KMEM)
5188 res_counter_reset_failcnt(&memcg->kmem);
5197 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5200 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5204 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5205 struct cftype *cft, u64 val)
5207 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5209 if (val >= (1 << NR_MOVE_TYPE))
5213 * No kind of locking is needed in here, because ->can_attach() will
5214 * check this value once in the beginning of the process, and then carry
5215 * on with stale data. This means that changes to this value will only
5216 * affect task migrations starting after the change.
5218 memcg->move_charge_at_immigrate = val;
5222 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5223 struct cftype *cft, u64 val)
5230 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5231 struct cftype *cft, struct seq_file *m)
5234 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5235 unsigned long node_nr;
5236 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5238 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5239 seq_printf(m, "total=%lu", total_nr);
5240 for_each_node_state(nid, N_MEMORY) {
5241 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5242 seq_printf(m, " N%d=%lu", nid, node_nr);
5246 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5247 seq_printf(m, "file=%lu", file_nr);
5248 for_each_node_state(nid, N_MEMORY) {
5249 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5251 seq_printf(m, " N%d=%lu", nid, node_nr);
5255 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5256 seq_printf(m, "anon=%lu", anon_nr);
5257 for_each_node_state(nid, N_MEMORY) {
5258 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5260 seq_printf(m, " N%d=%lu", nid, node_nr);
5264 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5265 seq_printf(m, "unevictable=%lu", unevictable_nr);
5266 for_each_node_state(nid, N_MEMORY) {
5267 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5268 BIT(LRU_UNEVICTABLE));
5269 seq_printf(m, " N%d=%lu", nid, node_nr);
5274 #endif /* CONFIG_NUMA */
5276 static inline void mem_cgroup_lru_names_not_uptodate(void)
5278 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5281 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5284 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5285 struct mem_cgroup *mi;
5288 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5289 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5291 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5292 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5295 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5296 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5297 mem_cgroup_read_events(memcg, i));
5299 for (i = 0; i < NR_LRU_LISTS; i++)
5300 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5301 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5303 /* Hierarchical information */
5305 unsigned long long limit, memsw_limit;
5306 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5307 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5308 if (do_swap_account)
5309 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5313 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5316 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5318 for_each_mem_cgroup_tree(mi, memcg)
5319 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5320 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5323 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5324 unsigned long long val = 0;
5326 for_each_mem_cgroup_tree(mi, memcg)
5327 val += mem_cgroup_read_events(mi, i);
5328 seq_printf(m, "total_%s %llu\n",
5329 mem_cgroup_events_names[i], val);
5332 for (i = 0; i < NR_LRU_LISTS; i++) {
5333 unsigned long long val = 0;
5335 for_each_mem_cgroup_tree(mi, memcg)
5336 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5337 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5340 #ifdef CONFIG_DEBUG_VM
5343 struct mem_cgroup_per_zone *mz;
5344 struct zone_reclaim_stat *rstat;
5345 unsigned long recent_rotated[2] = {0, 0};
5346 unsigned long recent_scanned[2] = {0, 0};
5348 for_each_online_node(nid)
5349 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5350 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5351 rstat = &mz->lruvec.reclaim_stat;
5353 recent_rotated[0] += rstat->recent_rotated[0];
5354 recent_rotated[1] += rstat->recent_rotated[1];
5355 recent_scanned[0] += rstat->recent_scanned[0];
5356 recent_scanned[1] += rstat->recent_scanned[1];
5358 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5359 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5360 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5361 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5368 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5371 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5373 return mem_cgroup_swappiness(memcg);
5376 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5377 struct cftype *cft, u64 val)
5379 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5380 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5382 if (val > 100 || !parent)
5385 mutex_lock(&memcg_create_mutex);
5387 /* If under hierarchy, only empty-root can set this value */
5388 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5389 mutex_unlock(&memcg_create_mutex);
5393 memcg->swappiness = val;
5395 mutex_unlock(&memcg_create_mutex);
5400 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5402 struct mem_cgroup_threshold_ary *t;
5408 t = rcu_dereference(memcg->thresholds.primary);
5410 t = rcu_dereference(memcg->memsw_thresholds.primary);
5415 usage = mem_cgroup_usage(memcg, swap);
5418 * current_threshold points to threshold just below or equal to usage.
5419 * If it's not true, a threshold was crossed after last
5420 * call of __mem_cgroup_threshold().
5422 i = t->current_threshold;
5425 * Iterate backward over array of thresholds starting from
5426 * current_threshold and check if a threshold is crossed.
5427 * If none of thresholds below usage is crossed, we read
5428 * only one element of the array here.
5430 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5431 eventfd_signal(t->entries[i].eventfd, 1);
5433 /* i = current_threshold + 1 */
5437 * Iterate forward over array of thresholds starting from
5438 * current_threshold+1 and check if a threshold is crossed.
5439 * If none of thresholds above usage is crossed, we read
5440 * only one element of the array here.
5442 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5443 eventfd_signal(t->entries[i].eventfd, 1);
5445 /* Update current_threshold */
5446 t->current_threshold = i - 1;
5451 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5454 __mem_cgroup_threshold(memcg, false);
5455 if (do_swap_account)
5456 __mem_cgroup_threshold(memcg, true);
5458 memcg = parent_mem_cgroup(memcg);
5462 static int compare_thresholds(const void *a, const void *b)
5464 const struct mem_cgroup_threshold *_a = a;
5465 const struct mem_cgroup_threshold *_b = b;
5467 if (_a->threshold > _b->threshold)
5470 if (_a->threshold < _b->threshold)
5476 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5478 struct mem_cgroup_eventfd_list *ev;
5480 list_for_each_entry(ev, &memcg->oom_notify, list)
5481 eventfd_signal(ev->eventfd, 1);
5485 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5487 struct mem_cgroup *iter;
5489 for_each_mem_cgroup_tree(iter, memcg)
5490 mem_cgroup_oom_notify_cb(iter);
5493 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5494 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5496 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5497 struct mem_cgroup_thresholds *thresholds;
5498 struct mem_cgroup_threshold_ary *new;
5499 enum res_type type = MEMFILE_TYPE(cft->private);
5500 u64 threshold, usage;
5503 ret = res_counter_memparse_write_strategy(args, &threshold);
5507 mutex_lock(&memcg->thresholds_lock);
5510 thresholds = &memcg->thresholds;
5511 else if (type == _MEMSWAP)
5512 thresholds = &memcg->memsw_thresholds;
5516 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5518 /* Check if a threshold crossed before adding a new one */
5519 if (thresholds->primary)
5520 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5522 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5524 /* Allocate memory for new array of thresholds */
5525 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5533 /* Copy thresholds (if any) to new array */
5534 if (thresholds->primary) {
5535 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5536 sizeof(struct mem_cgroup_threshold));
5539 /* Add new threshold */
5540 new->entries[size - 1].eventfd = eventfd;
5541 new->entries[size - 1].threshold = threshold;
5543 /* Sort thresholds. Registering of new threshold isn't time-critical */
5544 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5545 compare_thresholds, NULL);
5547 /* Find current threshold */
5548 new->current_threshold = -1;
5549 for (i = 0; i < size; i++) {
5550 if (new->entries[i].threshold <= usage) {
5552 * new->current_threshold will not be used until
5553 * rcu_assign_pointer(), so it's safe to increment
5556 ++new->current_threshold;
5561 /* Free old spare buffer and save old primary buffer as spare */
5562 kfree(thresholds->spare);
5563 thresholds->spare = thresholds->primary;
5565 rcu_assign_pointer(thresholds->primary, new);
5567 /* To be sure that nobody uses thresholds */
5571 mutex_unlock(&memcg->thresholds_lock);
5576 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5577 struct cftype *cft, struct eventfd_ctx *eventfd)
5579 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5580 struct mem_cgroup_thresholds *thresholds;
5581 struct mem_cgroup_threshold_ary *new;
5582 enum res_type type = MEMFILE_TYPE(cft->private);
5586 mutex_lock(&memcg->thresholds_lock);
5588 thresholds = &memcg->thresholds;
5589 else if (type == _MEMSWAP)
5590 thresholds = &memcg->memsw_thresholds;
5594 if (!thresholds->primary)
5597 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5599 /* Check if a threshold crossed before removing */
5600 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5602 /* Calculate new number of threshold */
5604 for (i = 0; i < thresholds->primary->size; i++) {
5605 if (thresholds->primary->entries[i].eventfd != eventfd)
5609 new = thresholds->spare;
5611 /* Set thresholds array to NULL if we don't have thresholds */
5620 /* Copy thresholds and find current threshold */
5621 new->current_threshold = -1;
5622 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5623 if (thresholds->primary->entries[i].eventfd == eventfd)
5626 new->entries[j] = thresholds->primary->entries[i];
5627 if (new->entries[j].threshold <= usage) {
5629 * new->current_threshold will not be used
5630 * until rcu_assign_pointer(), so it's safe to increment
5633 ++new->current_threshold;
5639 /* Swap primary and spare array */
5640 thresholds->spare = thresholds->primary;
5641 /* If all events are unregistered, free the spare array */
5643 kfree(thresholds->spare);
5644 thresholds->spare = NULL;
5647 rcu_assign_pointer(thresholds->primary, new);
5649 /* To be sure that nobody uses thresholds */
5652 mutex_unlock(&memcg->thresholds_lock);
5655 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5656 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5658 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5659 struct mem_cgroup_eventfd_list *event;
5660 enum res_type type = MEMFILE_TYPE(cft->private);
5662 BUG_ON(type != _OOM_TYPE);
5663 event = kmalloc(sizeof(*event), GFP_KERNEL);
5667 spin_lock(&memcg_oom_lock);
5669 event->eventfd = eventfd;
5670 list_add(&event->list, &memcg->oom_notify);
5672 /* already in OOM ? */
5673 if (atomic_read(&memcg->under_oom))
5674 eventfd_signal(eventfd, 1);
5675 spin_unlock(&memcg_oom_lock);
5680 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5681 struct cftype *cft, struct eventfd_ctx *eventfd)
5683 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5684 struct mem_cgroup_eventfd_list *ev, *tmp;
5685 enum res_type type = MEMFILE_TYPE(cft->private);
5687 BUG_ON(type != _OOM_TYPE);
5689 spin_lock(&memcg_oom_lock);
5691 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5692 if (ev->eventfd == eventfd) {
5693 list_del(&ev->list);
5698 spin_unlock(&memcg_oom_lock);
5701 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5702 struct cftype *cft, struct cgroup_map_cb *cb)
5704 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5706 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5708 if (atomic_read(&memcg->under_oom))
5709 cb->fill(cb, "under_oom", 1);
5711 cb->fill(cb, "under_oom", 0);
5715 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5716 struct cftype *cft, u64 val)
5718 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5719 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5721 /* cannot set to root cgroup and only 0 and 1 are allowed */
5722 if (!parent || !((val == 0) || (val == 1)))
5725 mutex_lock(&memcg_create_mutex);
5726 /* oom-kill-disable is a flag for subhierarchy. */
5727 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5728 mutex_unlock(&memcg_create_mutex);
5731 memcg->oom_kill_disable = val;
5733 memcg_oom_recover(memcg);
5734 mutex_unlock(&memcg_create_mutex);
5738 #ifdef CONFIG_MEMCG_KMEM
5739 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5743 memcg->kmemcg_id = -1;
5744 ret = memcg_propagate_kmem(memcg);
5748 return mem_cgroup_sockets_init(memcg, ss);
5751 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5753 mem_cgroup_sockets_destroy(memcg);
5756 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5758 if (!memcg_kmem_is_active(memcg))
5762 * kmem charges can outlive the cgroup. In the case of slab
5763 * pages, for instance, a page contain objects from various
5764 * processes. As we prevent from taking a reference for every
5765 * such allocation we have to be careful when doing uncharge
5766 * (see memcg_uncharge_kmem) and here during offlining.
5768 * The idea is that that only the _last_ uncharge which sees
5769 * the dead memcg will drop the last reference. An additional
5770 * reference is taken here before the group is marked dead
5771 * which is then paired with css_put during uncharge resp. here.
5773 * Although this might sound strange as this path is called from
5774 * css_offline() when the referencemight have dropped down to 0
5775 * and shouldn't be incremented anymore (css_tryget would fail)
5776 * we do not have other options because of the kmem allocations
5779 css_get(&memcg->css);
5781 memcg_kmem_mark_dead(memcg);
5783 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5786 if (memcg_kmem_test_and_clear_dead(memcg))
5787 css_put(&memcg->css);
5790 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5795 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5799 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5804 static struct cftype mem_cgroup_files[] = {
5806 .name = "usage_in_bytes",
5807 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5808 .read = mem_cgroup_read,
5809 .register_event = mem_cgroup_usage_register_event,
5810 .unregister_event = mem_cgroup_usage_unregister_event,
5813 .name = "max_usage_in_bytes",
5814 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5815 .trigger = mem_cgroup_reset,
5816 .read = mem_cgroup_read,
5819 .name = "limit_in_bytes",
5820 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5821 .write_string = mem_cgroup_write,
5822 .read = mem_cgroup_read,
5825 .name = "soft_limit_in_bytes",
5826 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5827 .write_string = mem_cgroup_write,
5828 .read = mem_cgroup_read,
5832 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5833 .trigger = mem_cgroup_reset,
5834 .read = mem_cgroup_read,
5838 .read_seq_string = memcg_stat_show,
5841 .name = "force_empty",
5842 .trigger = mem_cgroup_force_empty_write,
5845 .name = "use_hierarchy",
5846 .flags = CFTYPE_INSANE,
5847 .write_u64 = mem_cgroup_hierarchy_write,
5848 .read_u64 = mem_cgroup_hierarchy_read,
5851 .name = "swappiness",
5852 .read_u64 = mem_cgroup_swappiness_read,
5853 .write_u64 = mem_cgroup_swappiness_write,
5856 .name = "move_charge_at_immigrate",
5857 .read_u64 = mem_cgroup_move_charge_read,
5858 .write_u64 = mem_cgroup_move_charge_write,
5861 .name = "oom_control",
5862 .read_map = mem_cgroup_oom_control_read,
5863 .write_u64 = mem_cgroup_oom_control_write,
5864 .register_event = mem_cgroup_oom_register_event,
5865 .unregister_event = mem_cgroup_oom_unregister_event,
5866 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5869 .name = "pressure_level",
5870 .register_event = vmpressure_register_event,
5871 .unregister_event = vmpressure_unregister_event,
5875 .name = "numa_stat",
5876 .read_seq_string = memcg_numa_stat_show,
5879 #ifdef CONFIG_MEMCG_KMEM
5881 .name = "kmem.limit_in_bytes",
5882 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
5883 .write_string = mem_cgroup_write,
5884 .read = mem_cgroup_read,
5887 .name = "kmem.usage_in_bytes",
5888 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
5889 .read = mem_cgroup_read,
5892 .name = "kmem.failcnt",
5893 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
5894 .trigger = mem_cgroup_reset,
5895 .read = mem_cgroup_read,
5898 .name = "kmem.max_usage_in_bytes",
5899 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
5900 .trigger = mem_cgroup_reset,
5901 .read = mem_cgroup_read,
5903 #ifdef CONFIG_SLABINFO
5905 .name = "kmem.slabinfo",
5906 .read_seq_string = mem_cgroup_slabinfo_read,
5910 { }, /* terminate */
5913 #ifdef CONFIG_MEMCG_SWAP
5914 static struct cftype memsw_cgroup_files[] = {
5916 .name = "memsw.usage_in_bytes",
5917 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
5918 .read = mem_cgroup_read,
5919 .register_event = mem_cgroup_usage_register_event,
5920 .unregister_event = mem_cgroup_usage_unregister_event,
5923 .name = "memsw.max_usage_in_bytes",
5924 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
5925 .trigger = mem_cgroup_reset,
5926 .read = mem_cgroup_read,
5929 .name = "memsw.limit_in_bytes",
5930 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
5931 .write_string = mem_cgroup_write,
5932 .read = mem_cgroup_read,
5935 .name = "memsw.failcnt",
5936 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
5937 .trigger = mem_cgroup_reset,
5938 .read = mem_cgroup_read,
5940 { }, /* terminate */
5943 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5945 struct mem_cgroup_per_node *pn;
5946 struct mem_cgroup_per_zone *mz;
5947 int zone, tmp = node;
5949 * This routine is called against possible nodes.
5950 * But it's BUG to call kmalloc() against offline node.
5952 * TODO: this routine can waste much memory for nodes which will
5953 * never be onlined. It's better to use memory hotplug callback
5956 if (!node_state(node, N_NORMAL_MEMORY))
5958 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
5962 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5963 mz = &pn->zoneinfo[zone];
5964 lruvec_init(&mz->lruvec);
5965 mz->usage_in_excess = 0;
5966 mz->on_tree = false;
5969 memcg->nodeinfo[node] = pn;
5973 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
5975 kfree(memcg->nodeinfo[node]);
5978 static struct mem_cgroup *mem_cgroup_alloc(void)
5980 struct mem_cgroup *memcg;
5981 size_t size = memcg_size();
5983 /* Can be very big if nr_node_ids is very big */
5984 if (size < PAGE_SIZE)
5985 memcg = kzalloc(size, GFP_KERNEL);
5987 memcg = vzalloc(size);
5992 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
5995 spin_lock_init(&memcg->pcp_counter_lock);
5999 if (size < PAGE_SIZE)
6007 * At destroying mem_cgroup, references from swap_cgroup can remain.
6008 * (scanning all at force_empty is too costly...)
6010 * Instead of clearing all references at force_empty, we remember
6011 * the number of reference from swap_cgroup and free mem_cgroup when
6012 * it goes down to 0.
6014 * Removal of cgroup itself succeeds regardless of refs from swap.
6017 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6020 size_t size = memcg_size();
6022 mem_cgroup_remove_from_trees(memcg);
6023 free_css_id(&mem_cgroup_subsys, &memcg->css);
6026 free_mem_cgroup_per_zone_info(memcg, node);
6028 free_percpu(memcg->stat);
6031 * We need to make sure that (at least for now), the jump label
6032 * destruction code runs outside of the cgroup lock. This is because
6033 * get_online_cpus(), which is called from the static_branch update,
6034 * can't be called inside the cgroup_lock. cpusets are the ones
6035 * enforcing this dependency, so if they ever change, we might as well.
6037 * schedule_work() will guarantee this happens. Be careful if you need
6038 * to move this code around, and make sure it is outside
6041 disarm_static_keys(memcg);
6042 if (size < PAGE_SIZE)
6049 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6051 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6053 if (!memcg->res.parent)
6055 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6057 EXPORT_SYMBOL(parent_mem_cgroup);
6059 static void __init mem_cgroup_soft_limit_tree_init(void)
6061 struct mem_cgroup_tree_per_node *rtpn;
6062 struct mem_cgroup_tree_per_zone *rtpz;
6063 int tmp, node, zone;
6065 for_each_node(node) {
6067 if (!node_state(node, N_NORMAL_MEMORY))
6069 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6072 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6074 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6075 rtpz = &rtpn->rb_tree_per_zone[zone];
6076 rtpz->rb_root = RB_ROOT;
6077 spin_lock_init(&rtpz->lock);
6082 static struct cgroup_subsys_state * __ref
6083 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6085 struct mem_cgroup *memcg;
6086 long error = -ENOMEM;
6089 memcg = mem_cgroup_alloc();
6091 return ERR_PTR(error);
6094 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6098 if (parent_css == NULL) {
6099 root_mem_cgroup = memcg;
6100 res_counter_init(&memcg->res, NULL);
6101 res_counter_init(&memcg->memsw, NULL);
6102 res_counter_init(&memcg->kmem, NULL);
6105 memcg->last_scanned_node = MAX_NUMNODES;
6106 INIT_LIST_HEAD(&memcg->oom_notify);
6107 memcg->move_charge_at_immigrate = 0;
6108 mutex_init(&memcg->thresholds_lock);
6109 spin_lock_init(&memcg->move_lock);
6110 vmpressure_init(&memcg->vmpressure);
6115 __mem_cgroup_free(memcg);
6116 return ERR_PTR(error);
6120 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6122 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6123 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6129 mutex_lock(&memcg_create_mutex);
6131 memcg->use_hierarchy = parent->use_hierarchy;
6132 memcg->oom_kill_disable = parent->oom_kill_disable;
6133 memcg->swappiness = mem_cgroup_swappiness(parent);
6135 if (parent->use_hierarchy) {
6136 res_counter_init(&memcg->res, &parent->res);
6137 res_counter_init(&memcg->memsw, &parent->memsw);
6138 res_counter_init(&memcg->kmem, &parent->kmem);
6141 * No need to take a reference to the parent because cgroup
6142 * core guarantees its existence.
6145 res_counter_init(&memcg->res, NULL);
6146 res_counter_init(&memcg->memsw, NULL);
6147 res_counter_init(&memcg->kmem, NULL);
6149 * Deeper hierachy with use_hierarchy == false doesn't make
6150 * much sense so let cgroup subsystem know about this
6151 * unfortunate state in our controller.
6153 if (parent != root_mem_cgroup)
6154 mem_cgroup_subsys.broken_hierarchy = true;
6157 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6158 mutex_unlock(&memcg_create_mutex);
6163 * Announce all parents that a group from their hierarchy is gone.
6165 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6167 struct mem_cgroup *parent = memcg;
6169 while ((parent = parent_mem_cgroup(parent)))
6170 mem_cgroup_iter_invalidate(parent);
6173 * if the root memcg is not hierarchical we have to check it
6176 if (!root_mem_cgroup->use_hierarchy)
6177 mem_cgroup_iter_invalidate(root_mem_cgroup);
6180 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6182 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6184 kmem_cgroup_css_offline(memcg);
6186 mem_cgroup_invalidate_reclaim_iterators(memcg);
6187 mem_cgroup_reparent_charges(memcg);
6188 mem_cgroup_destroy_all_caches(memcg);
6189 vmpressure_cleanup(&memcg->vmpressure);
6192 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6194 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6196 memcg_destroy_kmem(memcg);
6197 __mem_cgroup_free(memcg);
6201 /* Handlers for move charge at task migration. */
6202 #define PRECHARGE_COUNT_AT_ONCE 256
6203 static int mem_cgroup_do_precharge(unsigned long count)
6206 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6207 struct mem_cgroup *memcg = mc.to;
6209 if (mem_cgroup_is_root(memcg)) {
6210 mc.precharge += count;
6211 /* we don't need css_get for root */
6214 /* try to charge at once */
6216 struct res_counter *dummy;
6218 * "memcg" cannot be under rmdir() because we've already checked
6219 * by cgroup_lock_live_cgroup() that it is not removed and we
6220 * are still under the same cgroup_mutex. So we can postpone
6223 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6225 if (do_swap_account && res_counter_charge(&memcg->memsw,
6226 PAGE_SIZE * count, &dummy)) {
6227 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6230 mc.precharge += count;
6234 /* fall back to one by one charge */
6236 if (signal_pending(current)) {
6240 if (!batch_count--) {
6241 batch_count = PRECHARGE_COUNT_AT_ONCE;
6244 ret = __mem_cgroup_try_charge(NULL,
6245 GFP_KERNEL, 1, &memcg, false);
6247 /* mem_cgroup_clear_mc() will do uncharge later */
6255 * get_mctgt_type - get target type of moving charge
6256 * @vma: the vma the pte to be checked belongs
6257 * @addr: the address corresponding to the pte to be checked
6258 * @ptent: the pte to be checked
6259 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6262 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6263 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6264 * move charge. if @target is not NULL, the page is stored in target->page
6265 * with extra refcnt got(Callers should handle it).
6266 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6267 * target for charge migration. if @target is not NULL, the entry is stored
6270 * Called with pte lock held.
6277 enum mc_target_type {
6283 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6284 unsigned long addr, pte_t ptent)
6286 struct page *page = vm_normal_page(vma, addr, ptent);
6288 if (!page || !page_mapped(page))
6290 if (PageAnon(page)) {
6291 /* we don't move shared anon */
6294 } else if (!move_file())
6295 /* we ignore mapcount for file pages */
6297 if (!get_page_unless_zero(page))
6304 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6305 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6307 struct page *page = NULL;
6308 swp_entry_t ent = pte_to_swp_entry(ptent);
6310 if (!move_anon() || non_swap_entry(ent))
6313 * Because lookup_swap_cache() updates some statistics counter,
6314 * we call find_get_page() with swapper_space directly.
6316 page = find_get_page(swap_address_space(ent), ent.val);
6317 if (do_swap_account)
6318 entry->val = ent.val;
6323 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6324 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6330 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6331 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6333 struct page *page = NULL;
6334 struct address_space *mapping;
6337 if (!vma->vm_file) /* anonymous vma */
6342 mapping = vma->vm_file->f_mapping;
6343 if (pte_none(ptent))
6344 pgoff = linear_page_index(vma, addr);
6345 else /* pte_file(ptent) is true */
6346 pgoff = pte_to_pgoff(ptent);
6348 /* page is moved even if it's not RSS of this task(page-faulted). */
6349 page = find_get_page(mapping, pgoff);
6352 /* shmem/tmpfs may report page out on swap: account for that too. */
6353 if (radix_tree_exceptional_entry(page)) {
6354 swp_entry_t swap = radix_to_swp_entry(page);
6355 if (do_swap_account)
6357 page = find_get_page(swap_address_space(swap), swap.val);
6363 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6364 unsigned long addr, pte_t ptent, union mc_target *target)
6366 struct page *page = NULL;
6367 struct page_cgroup *pc;
6368 enum mc_target_type ret = MC_TARGET_NONE;
6369 swp_entry_t ent = { .val = 0 };
6371 if (pte_present(ptent))
6372 page = mc_handle_present_pte(vma, addr, ptent);
6373 else if (is_swap_pte(ptent))
6374 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6375 else if (pte_none(ptent) || pte_file(ptent))
6376 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6378 if (!page && !ent.val)
6381 pc = lookup_page_cgroup(page);
6383 * Do only loose check w/o page_cgroup lock.
6384 * mem_cgroup_move_account() checks the pc is valid or not under
6387 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6388 ret = MC_TARGET_PAGE;
6390 target->page = page;
6392 if (!ret || !target)
6395 /* There is a swap entry and a page doesn't exist or isn't charged */
6396 if (ent.val && !ret &&
6397 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6398 ret = MC_TARGET_SWAP;
6405 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6407 * We don't consider swapping or file mapped pages because THP does not
6408 * support them for now.
6409 * Caller should make sure that pmd_trans_huge(pmd) is true.
6411 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6412 unsigned long addr, pmd_t pmd, union mc_target *target)
6414 struct page *page = NULL;
6415 struct page_cgroup *pc;
6416 enum mc_target_type ret = MC_TARGET_NONE;
6418 page = pmd_page(pmd);
6419 VM_BUG_ON(!page || !PageHead(page));
6422 pc = lookup_page_cgroup(page);
6423 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6424 ret = MC_TARGET_PAGE;
6427 target->page = page;
6433 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6434 unsigned long addr, pmd_t pmd, union mc_target *target)
6436 return MC_TARGET_NONE;
6440 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6441 unsigned long addr, unsigned long end,
6442 struct mm_walk *walk)
6444 struct vm_area_struct *vma = walk->private;
6448 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6449 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6450 mc.precharge += HPAGE_PMD_NR;
6451 spin_unlock(&vma->vm_mm->page_table_lock);
6455 if (pmd_trans_unstable(pmd))
6457 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6458 for (; addr != end; pte++, addr += PAGE_SIZE)
6459 if (get_mctgt_type(vma, addr, *pte, NULL))
6460 mc.precharge++; /* increment precharge temporarily */
6461 pte_unmap_unlock(pte - 1, ptl);
6467 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6469 unsigned long precharge;
6470 struct vm_area_struct *vma;
6472 down_read(&mm->mmap_sem);
6473 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6474 struct mm_walk mem_cgroup_count_precharge_walk = {
6475 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6479 if (is_vm_hugetlb_page(vma))
6481 walk_page_range(vma->vm_start, vma->vm_end,
6482 &mem_cgroup_count_precharge_walk);
6484 up_read(&mm->mmap_sem);
6486 precharge = mc.precharge;
6492 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6494 unsigned long precharge = mem_cgroup_count_precharge(mm);
6496 VM_BUG_ON(mc.moving_task);
6497 mc.moving_task = current;
6498 return mem_cgroup_do_precharge(precharge);
6501 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6502 static void __mem_cgroup_clear_mc(void)
6504 struct mem_cgroup *from = mc.from;
6505 struct mem_cgroup *to = mc.to;
6508 /* we must uncharge all the leftover precharges from mc.to */
6510 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6514 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6515 * we must uncharge here.
6517 if (mc.moved_charge) {
6518 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6519 mc.moved_charge = 0;
6521 /* we must fixup refcnts and charges */
6522 if (mc.moved_swap) {
6523 /* uncharge swap account from the old cgroup */
6524 if (!mem_cgroup_is_root(mc.from))
6525 res_counter_uncharge(&mc.from->memsw,
6526 PAGE_SIZE * mc.moved_swap);
6528 for (i = 0; i < mc.moved_swap; i++)
6529 css_put(&mc.from->css);
6531 if (!mem_cgroup_is_root(mc.to)) {
6533 * we charged both to->res and to->memsw, so we should
6536 res_counter_uncharge(&mc.to->res,
6537 PAGE_SIZE * mc.moved_swap);
6539 /* we've already done css_get(mc.to) */
6542 memcg_oom_recover(from);
6543 memcg_oom_recover(to);
6544 wake_up_all(&mc.waitq);
6547 static void mem_cgroup_clear_mc(void)
6549 struct mem_cgroup *from = mc.from;
6552 * we must clear moving_task before waking up waiters at the end of
6555 mc.moving_task = NULL;
6556 __mem_cgroup_clear_mc();
6557 spin_lock(&mc.lock);
6560 spin_unlock(&mc.lock);
6561 mem_cgroup_end_move(from);
6564 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6565 struct cgroup_taskset *tset)
6567 struct task_struct *p = cgroup_taskset_first(tset);
6569 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6570 unsigned long move_charge_at_immigrate;
6573 * We are now commited to this value whatever it is. Changes in this
6574 * tunable will only affect upcoming migrations, not the current one.
6575 * So we need to save it, and keep it going.
6577 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6578 if (move_charge_at_immigrate) {
6579 struct mm_struct *mm;
6580 struct mem_cgroup *from = mem_cgroup_from_task(p);
6582 VM_BUG_ON(from == memcg);
6584 mm = get_task_mm(p);
6587 /* We move charges only when we move a owner of the mm */
6588 if (mm->owner == p) {
6591 VM_BUG_ON(mc.precharge);
6592 VM_BUG_ON(mc.moved_charge);
6593 VM_BUG_ON(mc.moved_swap);
6594 mem_cgroup_start_move(from);
6595 spin_lock(&mc.lock);
6598 mc.immigrate_flags = move_charge_at_immigrate;
6599 spin_unlock(&mc.lock);
6600 /* We set mc.moving_task later */
6602 ret = mem_cgroup_precharge_mc(mm);
6604 mem_cgroup_clear_mc();
6611 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6612 struct cgroup_taskset *tset)
6614 mem_cgroup_clear_mc();
6617 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6618 unsigned long addr, unsigned long end,
6619 struct mm_walk *walk)
6622 struct vm_area_struct *vma = walk->private;
6625 enum mc_target_type target_type;
6626 union mc_target target;
6628 struct page_cgroup *pc;
6631 * We don't take compound_lock() here but no race with splitting thp
6633 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6634 * under splitting, which means there's no concurrent thp split,
6635 * - if another thread runs into split_huge_page() just after we
6636 * entered this if-block, the thread must wait for page table lock
6637 * to be unlocked in __split_huge_page_splitting(), where the main
6638 * part of thp split is not executed yet.
6640 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6641 if (mc.precharge < HPAGE_PMD_NR) {
6642 spin_unlock(&vma->vm_mm->page_table_lock);
6645 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6646 if (target_type == MC_TARGET_PAGE) {
6648 if (!isolate_lru_page(page)) {
6649 pc = lookup_page_cgroup(page);
6650 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6651 pc, mc.from, mc.to)) {
6652 mc.precharge -= HPAGE_PMD_NR;
6653 mc.moved_charge += HPAGE_PMD_NR;
6655 putback_lru_page(page);
6659 spin_unlock(&vma->vm_mm->page_table_lock);
6663 if (pmd_trans_unstable(pmd))
6666 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6667 for (; addr != end; addr += PAGE_SIZE) {
6668 pte_t ptent = *(pte++);
6674 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6675 case MC_TARGET_PAGE:
6677 if (isolate_lru_page(page))
6679 pc = lookup_page_cgroup(page);
6680 if (!mem_cgroup_move_account(page, 1, pc,
6683 /* we uncharge from mc.from later. */
6686 putback_lru_page(page);
6687 put: /* get_mctgt_type() gets the page */
6690 case MC_TARGET_SWAP:
6692 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6694 /* we fixup refcnts and charges later. */
6702 pte_unmap_unlock(pte - 1, ptl);
6707 * We have consumed all precharges we got in can_attach().
6708 * We try charge one by one, but don't do any additional
6709 * charges to mc.to if we have failed in charge once in attach()
6712 ret = mem_cgroup_do_precharge(1);
6720 static void mem_cgroup_move_charge(struct mm_struct *mm)
6722 struct vm_area_struct *vma;
6724 lru_add_drain_all();
6726 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6728 * Someone who are holding the mmap_sem might be waiting in
6729 * waitq. So we cancel all extra charges, wake up all waiters,
6730 * and retry. Because we cancel precharges, we might not be able
6731 * to move enough charges, but moving charge is a best-effort
6732 * feature anyway, so it wouldn't be a big problem.
6734 __mem_cgroup_clear_mc();
6738 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6740 struct mm_walk mem_cgroup_move_charge_walk = {
6741 .pmd_entry = mem_cgroup_move_charge_pte_range,
6745 if (is_vm_hugetlb_page(vma))
6747 ret = walk_page_range(vma->vm_start, vma->vm_end,
6748 &mem_cgroup_move_charge_walk);
6751 * means we have consumed all precharges and failed in
6752 * doing additional charge. Just abandon here.
6756 up_read(&mm->mmap_sem);
6759 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6760 struct cgroup_taskset *tset)
6762 struct task_struct *p = cgroup_taskset_first(tset);
6763 struct mm_struct *mm = get_task_mm(p);
6767 mem_cgroup_move_charge(mm);
6771 mem_cgroup_clear_mc();
6773 #else /* !CONFIG_MMU */
6774 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6775 struct cgroup_taskset *tset)
6779 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6780 struct cgroup_taskset *tset)
6783 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6784 struct cgroup_taskset *tset)
6790 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6791 * to verify sane_behavior flag on each mount attempt.
6793 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6796 * use_hierarchy is forced with sane_behavior. cgroup core
6797 * guarantees that @root doesn't have any children, so turning it
6798 * on for the root memcg is enough.
6800 if (cgroup_sane_behavior(root_css->cgroup))
6801 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6804 struct cgroup_subsys mem_cgroup_subsys = {
6806 .subsys_id = mem_cgroup_subsys_id,
6807 .css_alloc = mem_cgroup_css_alloc,
6808 .css_online = mem_cgroup_css_online,
6809 .css_offline = mem_cgroup_css_offline,
6810 .css_free = mem_cgroup_css_free,
6811 .can_attach = mem_cgroup_can_attach,
6812 .cancel_attach = mem_cgroup_cancel_attach,
6813 .attach = mem_cgroup_move_task,
6814 .bind = mem_cgroup_bind,
6815 .base_cftypes = mem_cgroup_files,
6820 #ifdef CONFIG_MEMCG_SWAP
6821 static int __init enable_swap_account(char *s)
6823 if (!strcmp(s, "1"))
6824 really_do_swap_account = 1;
6825 else if (!strcmp(s, "0"))
6826 really_do_swap_account = 0;
6829 __setup("swapaccount=", enable_swap_account);
6831 static void __init memsw_file_init(void)
6833 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6836 static void __init enable_swap_cgroup(void)
6838 if (!mem_cgroup_disabled() && really_do_swap_account) {
6839 do_swap_account = 1;
6845 static void __init enable_swap_cgroup(void)
6851 * subsys_initcall() for memory controller.
6853 * Some parts like hotcpu_notifier() have to be initialized from this context
6854 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6855 * everything that doesn't depend on a specific mem_cgroup structure should
6856 * be initialized from here.
6858 static int __init mem_cgroup_init(void)
6860 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6861 enable_swap_cgroup();
6862 mem_cgroup_soft_limit_tree_init();
6866 subsys_initcall(mem_cgroup_init);