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);
2007 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
2010 unsigned long *total_scanned)
2012 struct mem_cgroup *victim = NULL;
2015 unsigned long excess;
2016 unsigned long nr_scanned;
2017 struct mem_cgroup_reclaim_cookie reclaim = {
2022 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
2025 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
2030 * If we have not been able to reclaim
2031 * anything, it might because there are
2032 * no reclaimable pages under this hierarchy
2037 * We want to do more targeted reclaim.
2038 * excess >> 2 is not to excessive so as to
2039 * reclaim too much, nor too less that we keep
2040 * coming back to reclaim from this cgroup
2042 if (total >= (excess >> 2) ||
2043 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
2048 if (!mem_cgroup_reclaimable(victim, false))
2050 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
2052 *total_scanned += nr_scanned;
2053 if (!res_counter_soft_limit_excess(&root_memcg->res))
2056 mem_cgroup_iter_break(root_memcg, victim);
2061 * Check OOM-Killer is already running under our hierarchy.
2062 * If someone is running, return false.
2063 * Has to be called with memcg_oom_lock
2065 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
2067 struct mem_cgroup *iter, *failed = NULL;
2069 for_each_mem_cgroup_tree(iter, memcg) {
2070 if (iter->oom_lock) {
2072 * this subtree of our hierarchy is already locked
2073 * so we cannot give a lock.
2076 mem_cgroup_iter_break(memcg, iter);
2079 iter->oom_lock = true;
2086 * OK, we failed to lock the whole subtree so we have to clean up
2087 * what we set up to the failing subtree
2089 for_each_mem_cgroup_tree(iter, memcg) {
2090 if (iter == failed) {
2091 mem_cgroup_iter_break(memcg, iter);
2094 iter->oom_lock = false;
2100 * Has to be called with memcg_oom_lock
2102 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
2104 struct mem_cgroup *iter;
2106 for_each_mem_cgroup_tree(iter, memcg)
2107 iter->oom_lock = false;
2111 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
2113 struct mem_cgroup *iter;
2115 for_each_mem_cgroup_tree(iter, memcg)
2116 atomic_inc(&iter->under_oom);
2119 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
2121 struct mem_cgroup *iter;
2124 * When a new child is created while the hierarchy is under oom,
2125 * mem_cgroup_oom_lock() may not be called. We have to use
2126 * atomic_add_unless() here.
2128 for_each_mem_cgroup_tree(iter, memcg)
2129 atomic_add_unless(&iter->under_oom, -1, 0);
2132 static DEFINE_SPINLOCK(memcg_oom_lock);
2133 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
2135 struct oom_wait_info {
2136 struct mem_cgroup *memcg;
2140 static int memcg_oom_wake_function(wait_queue_t *wait,
2141 unsigned mode, int sync, void *arg)
2143 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
2144 struct mem_cgroup *oom_wait_memcg;
2145 struct oom_wait_info *oom_wait_info;
2147 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
2148 oom_wait_memcg = oom_wait_info->memcg;
2151 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
2152 * Then we can use css_is_ancestor without taking care of RCU.
2154 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
2155 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
2157 return autoremove_wake_function(wait, mode, sync, arg);
2160 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
2162 /* for filtering, pass "memcg" as argument. */
2163 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
2166 static void memcg_oom_recover(struct mem_cgroup *memcg)
2168 if (memcg && atomic_read(&memcg->under_oom))
2169 memcg_wakeup_oom(memcg);
2173 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
2175 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
2178 struct oom_wait_info owait;
2179 bool locked, need_to_kill;
2181 owait.memcg = memcg;
2182 owait.wait.flags = 0;
2183 owait.wait.func = memcg_oom_wake_function;
2184 owait.wait.private = current;
2185 INIT_LIST_HEAD(&owait.wait.task_list);
2186 need_to_kill = true;
2187 mem_cgroup_mark_under_oom(memcg);
2189 /* At first, try to OOM lock hierarchy under memcg.*/
2190 spin_lock(&memcg_oom_lock);
2191 locked = mem_cgroup_oom_lock(memcg);
2193 * Even if signal_pending(), we can't quit charge() loop without
2194 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
2195 * under OOM is always welcomed, use TASK_KILLABLE here.
2197 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
2198 if (!locked || memcg->oom_kill_disable)
2199 need_to_kill = false;
2201 mem_cgroup_oom_notify(memcg);
2202 spin_unlock(&memcg_oom_lock);
2205 finish_wait(&memcg_oom_waitq, &owait.wait);
2206 mem_cgroup_out_of_memory(memcg, mask, order);
2209 finish_wait(&memcg_oom_waitq, &owait.wait);
2211 spin_lock(&memcg_oom_lock);
2213 mem_cgroup_oom_unlock(memcg);
2214 memcg_wakeup_oom(memcg);
2215 spin_unlock(&memcg_oom_lock);
2217 mem_cgroup_unmark_under_oom(memcg);
2219 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2221 /* Give chance to dying process */
2222 schedule_timeout_uninterruptible(1);
2227 * Currently used to update mapped file statistics, but the routine can be
2228 * generalized to update other statistics as well.
2230 * Notes: Race condition
2232 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2233 * it tends to be costly. But considering some conditions, we doesn't need
2234 * to do so _always_.
2236 * Considering "charge", lock_page_cgroup() is not required because all
2237 * file-stat operations happen after a page is attached to radix-tree. There
2238 * are no race with "charge".
2240 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2241 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2242 * if there are race with "uncharge". Statistics itself is properly handled
2245 * Considering "move", this is an only case we see a race. To make the race
2246 * small, we check mm->moving_account and detect there are possibility of race
2247 * If there is, we take a lock.
2250 void __mem_cgroup_begin_update_page_stat(struct page *page,
2251 bool *locked, unsigned long *flags)
2253 struct mem_cgroup *memcg;
2254 struct page_cgroup *pc;
2256 pc = lookup_page_cgroup(page);
2258 memcg = pc->mem_cgroup;
2259 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2262 * If this memory cgroup is not under account moving, we don't
2263 * need to take move_lock_mem_cgroup(). Because we already hold
2264 * rcu_read_lock(), any calls to move_account will be delayed until
2265 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2267 if (!mem_cgroup_stolen(memcg))
2270 move_lock_mem_cgroup(memcg, flags);
2271 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2272 move_unlock_mem_cgroup(memcg, flags);
2278 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2280 struct page_cgroup *pc = lookup_page_cgroup(page);
2283 * It's guaranteed that pc->mem_cgroup never changes while
2284 * lock is held because a routine modifies pc->mem_cgroup
2285 * should take move_lock_mem_cgroup().
2287 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2290 void mem_cgroup_update_page_stat(struct page *page,
2291 enum mem_cgroup_page_stat_item idx, int val)
2293 struct mem_cgroup *memcg;
2294 struct page_cgroup *pc = lookup_page_cgroup(page);
2295 unsigned long uninitialized_var(flags);
2297 if (mem_cgroup_disabled())
2300 memcg = pc->mem_cgroup;
2301 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2305 case MEMCG_NR_FILE_MAPPED:
2306 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2312 this_cpu_add(memcg->stat->count[idx], val);
2316 * size of first charge trial. "32" comes from vmscan.c's magic value.
2317 * TODO: maybe necessary to use big numbers in big irons.
2319 #define CHARGE_BATCH 32U
2320 struct memcg_stock_pcp {
2321 struct mem_cgroup *cached; /* this never be root cgroup */
2322 unsigned int nr_pages;
2323 struct work_struct work;
2324 unsigned long flags;
2325 #define FLUSHING_CACHED_CHARGE 0
2327 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2328 static DEFINE_MUTEX(percpu_charge_mutex);
2331 * consume_stock: Try to consume stocked charge on this cpu.
2332 * @memcg: memcg to consume from.
2333 * @nr_pages: how many pages to charge.
2335 * The charges will only happen if @memcg matches the current cpu's memcg
2336 * stock, and at least @nr_pages are available in that stock. Failure to
2337 * service an allocation will refill the stock.
2339 * returns true if successful, false otherwise.
2341 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2343 struct memcg_stock_pcp *stock;
2346 if (nr_pages > CHARGE_BATCH)
2349 stock = &get_cpu_var(memcg_stock);
2350 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2351 stock->nr_pages -= nr_pages;
2352 else /* need to call res_counter_charge */
2354 put_cpu_var(memcg_stock);
2359 * Returns stocks cached in percpu to res_counter and reset cached information.
2361 static void drain_stock(struct memcg_stock_pcp *stock)
2363 struct mem_cgroup *old = stock->cached;
2365 if (stock->nr_pages) {
2366 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2368 res_counter_uncharge(&old->res, bytes);
2369 if (do_swap_account)
2370 res_counter_uncharge(&old->memsw, bytes);
2371 stock->nr_pages = 0;
2373 stock->cached = NULL;
2377 * This must be called under preempt disabled or must be called by
2378 * a thread which is pinned to local cpu.
2380 static void drain_local_stock(struct work_struct *dummy)
2382 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2384 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2387 static void __init memcg_stock_init(void)
2391 for_each_possible_cpu(cpu) {
2392 struct memcg_stock_pcp *stock =
2393 &per_cpu(memcg_stock, cpu);
2394 INIT_WORK(&stock->work, drain_local_stock);
2399 * Cache charges(val) which is from res_counter, to local per_cpu area.
2400 * This will be consumed by consume_stock() function, later.
2402 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2404 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2406 if (stock->cached != memcg) { /* reset if necessary */
2408 stock->cached = memcg;
2410 stock->nr_pages += nr_pages;
2411 put_cpu_var(memcg_stock);
2415 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2416 * of the hierarchy under it. sync flag says whether we should block
2417 * until the work is done.
2419 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2423 /* Notify other cpus that system-wide "drain" is running */
2426 for_each_online_cpu(cpu) {
2427 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2428 struct mem_cgroup *memcg;
2430 memcg = stock->cached;
2431 if (!memcg || !stock->nr_pages)
2433 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2435 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2437 drain_local_stock(&stock->work);
2439 schedule_work_on(cpu, &stock->work);
2447 for_each_online_cpu(cpu) {
2448 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2449 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2450 flush_work(&stock->work);
2457 * Tries to drain stocked charges in other cpus. This function is asynchronous
2458 * and just put a work per cpu for draining localy on each cpu. Caller can
2459 * expects some charges will be back to res_counter later but cannot wait for
2462 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2465 * If someone calls draining, avoid adding more kworker runs.
2467 if (!mutex_trylock(&percpu_charge_mutex))
2469 drain_all_stock(root_memcg, false);
2470 mutex_unlock(&percpu_charge_mutex);
2473 /* This is a synchronous drain interface. */
2474 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2476 /* called when force_empty is called */
2477 mutex_lock(&percpu_charge_mutex);
2478 drain_all_stock(root_memcg, true);
2479 mutex_unlock(&percpu_charge_mutex);
2483 * This function drains percpu counter value from DEAD cpu and
2484 * move it to local cpu. Note that this function can be preempted.
2486 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2490 spin_lock(&memcg->pcp_counter_lock);
2491 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2492 long x = per_cpu(memcg->stat->count[i], cpu);
2494 per_cpu(memcg->stat->count[i], cpu) = 0;
2495 memcg->nocpu_base.count[i] += x;
2497 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2498 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2500 per_cpu(memcg->stat->events[i], cpu) = 0;
2501 memcg->nocpu_base.events[i] += x;
2503 spin_unlock(&memcg->pcp_counter_lock);
2506 static int memcg_cpu_hotplug_callback(struct notifier_block *nb,
2507 unsigned long action,
2510 int cpu = (unsigned long)hcpu;
2511 struct memcg_stock_pcp *stock;
2512 struct mem_cgroup *iter;
2514 if (action == CPU_ONLINE)
2517 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2520 for_each_mem_cgroup(iter)
2521 mem_cgroup_drain_pcp_counter(iter, cpu);
2523 stock = &per_cpu(memcg_stock, cpu);
2529 /* See __mem_cgroup_try_charge() for details */
2531 CHARGE_OK, /* success */
2532 CHARGE_RETRY, /* need to retry but retry is not bad */
2533 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2534 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2535 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2538 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2539 unsigned int nr_pages, unsigned int min_pages,
2542 unsigned long csize = nr_pages * PAGE_SIZE;
2543 struct mem_cgroup *mem_over_limit;
2544 struct res_counter *fail_res;
2545 unsigned long flags = 0;
2548 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2551 if (!do_swap_account)
2553 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2557 res_counter_uncharge(&memcg->res, csize);
2558 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2559 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2561 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2563 * Never reclaim on behalf of optional batching, retry with a
2564 * single page instead.
2566 if (nr_pages > min_pages)
2567 return CHARGE_RETRY;
2569 if (!(gfp_mask & __GFP_WAIT))
2570 return CHARGE_WOULDBLOCK;
2572 if (gfp_mask & __GFP_NORETRY)
2573 return CHARGE_NOMEM;
2575 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2576 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2577 return CHARGE_RETRY;
2579 * Even though the limit is exceeded at this point, reclaim
2580 * may have been able to free some pages. Retry the charge
2581 * before killing the task.
2583 * Only for regular pages, though: huge pages are rather
2584 * unlikely to succeed so close to the limit, and we fall back
2585 * to regular pages anyway in case of failure.
2587 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2588 return CHARGE_RETRY;
2591 * At task move, charge accounts can be doubly counted. So, it's
2592 * better to wait until the end of task_move if something is going on.
2594 if (mem_cgroup_wait_acct_move(mem_over_limit))
2595 return CHARGE_RETRY;
2597 /* If we don't need to call oom-killer at el, return immediately */
2599 return CHARGE_NOMEM;
2601 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2602 return CHARGE_OOM_DIE;
2604 return CHARGE_RETRY;
2608 * __mem_cgroup_try_charge() does
2609 * 1. detect memcg to be charged against from passed *mm and *ptr,
2610 * 2. update res_counter
2611 * 3. call memory reclaim if necessary.
2613 * In some special case, if the task is fatal, fatal_signal_pending() or
2614 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2615 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2616 * as possible without any hazards. 2: all pages should have a valid
2617 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2618 * pointer, that is treated as a charge to root_mem_cgroup.
2620 * So __mem_cgroup_try_charge() will return
2621 * 0 ... on success, filling *ptr with a valid memcg pointer.
2622 * -ENOMEM ... charge failure because of resource limits.
2623 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2625 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2626 * the oom-killer can be invoked.
2628 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2630 unsigned int nr_pages,
2631 struct mem_cgroup **ptr,
2634 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2635 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2636 struct mem_cgroup *memcg = NULL;
2640 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2641 * in system level. So, allow to go ahead dying process in addition to
2644 if (unlikely(test_thread_flag(TIF_MEMDIE)
2645 || fatal_signal_pending(current)))
2649 * We always charge the cgroup the mm_struct belongs to.
2650 * The mm_struct's mem_cgroup changes on task migration if the
2651 * thread group leader migrates. It's possible that mm is not
2652 * set, if so charge the root memcg (happens for pagecache usage).
2655 *ptr = root_mem_cgroup;
2657 if (*ptr) { /* css should be a valid one */
2659 if (mem_cgroup_is_root(memcg))
2661 if (consume_stock(memcg, nr_pages))
2663 css_get(&memcg->css);
2665 struct task_struct *p;
2668 p = rcu_dereference(mm->owner);
2670 * Because we don't have task_lock(), "p" can exit.
2671 * In that case, "memcg" can point to root or p can be NULL with
2672 * race with swapoff. Then, we have small risk of mis-accouning.
2673 * But such kind of mis-account by race always happens because
2674 * we don't have cgroup_mutex(). It's overkill and we allo that
2676 * (*) swapoff at el will charge against mm-struct not against
2677 * task-struct. So, mm->owner can be NULL.
2679 memcg = mem_cgroup_from_task(p);
2681 memcg = root_mem_cgroup;
2682 if (mem_cgroup_is_root(memcg)) {
2686 if (consume_stock(memcg, nr_pages)) {
2688 * It seems dagerous to access memcg without css_get().
2689 * But considering how consume_stok works, it's not
2690 * necessary. If consume_stock success, some charges
2691 * from this memcg are cached on this cpu. So, we
2692 * don't need to call css_get()/css_tryget() before
2693 * calling consume_stock().
2698 /* after here, we may be blocked. we need to get refcnt */
2699 if (!css_tryget(&memcg->css)) {
2709 /* If killed, bypass charge */
2710 if (fatal_signal_pending(current)) {
2711 css_put(&memcg->css);
2716 if (oom && !nr_oom_retries) {
2718 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2721 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2726 case CHARGE_RETRY: /* not in OOM situation but retry */
2728 css_put(&memcg->css);
2731 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2732 css_put(&memcg->css);
2734 case CHARGE_NOMEM: /* OOM routine works */
2736 css_put(&memcg->css);
2739 /* If oom, we never return -ENOMEM */
2742 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2743 css_put(&memcg->css);
2746 } while (ret != CHARGE_OK);
2748 if (batch > nr_pages)
2749 refill_stock(memcg, batch - nr_pages);
2750 css_put(&memcg->css);
2758 *ptr = root_mem_cgroup;
2763 * Somemtimes we have to undo a charge we got by try_charge().
2764 * This function is for that and do uncharge, put css's refcnt.
2765 * gotten by try_charge().
2767 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2768 unsigned int nr_pages)
2770 if (!mem_cgroup_is_root(memcg)) {
2771 unsigned long bytes = nr_pages * PAGE_SIZE;
2773 res_counter_uncharge(&memcg->res, bytes);
2774 if (do_swap_account)
2775 res_counter_uncharge(&memcg->memsw, bytes);
2780 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2781 * This is useful when moving usage to parent cgroup.
2783 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2784 unsigned int nr_pages)
2786 unsigned long bytes = nr_pages * PAGE_SIZE;
2788 if (mem_cgroup_is_root(memcg))
2791 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2792 if (do_swap_account)
2793 res_counter_uncharge_until(&memcg->memsw,
2794 memcg->memsw.parent, bytes);
2798 * A helper function to get mem_cgroup from ID. must be called under
2799 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2800 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2801 * called against removed memcg.)
2803 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2805 struct cgroup_subsys_state *css;
2807 /* ID 0 is unused ID */
2810 css = css_lookup(&mem_cgroup_subsys, id);
2813 return mem_cgroup_from_css(css);
2816 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2818 struct mem_cgroup *memcg = NULL;
2819 struct page_cgroup *pc;
2823 VM_BUG_ON(!PageLocked(page));
2825 pc = lookup_page_cgroup(page);
2826 lock_page_cgroup(pc);
2827 if (PageCgroupUsed(pc)) {
2828 memcg = pc->mem_cgroup;
2829 if (memcg && !css_tryget(&memcg->css))
2831 } else if (PageSwapCache(page)) {
2832 ent.val = page_private(page);
2833 id = lookup_swap_cgroup_id(ent);
2835 memcg = mem_cgroup_lookup(id);
2836 if (memcg && !css_tryget(&memcg->css))
2840 unlock_page_cgroup(pc);
2844 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2846 unsigned int nr_pages,
2847 enum charge_type ctype,
2850 struct page_cgroup *pc = lookup_page_cgroup(page);
2851 struct zone *uninitialized_var(zone);
2852 struct lruvec *lruvec;
2853 bool was_on_lru = false;
2856 lock_page_cgroup(pc);
2857 VM_BUG_ON(PageCgroupUsed(pc));
2859 * we don't need page_cgroup_lock about tail pages, becase they are not
2860 * accessed by any other context at this point.
2864 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2865 * may already be on some other mem_cgroup's LRU. Take care of it.
2868 zone = page_zone(page);
2869 spin_lock_irq(&zone->lru_lock);
2870 if (PageLRU(page)) {
2871 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2873 del_page_from_lru_list(page, lruvec, page_lru(page));
2878 pc->mem_cgroup = memcg;
2880 * We access a page_cgroup asynchronously without lock_page_cgroup().
2881 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2882 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2883 * before USED bit, we need memory barrier here.
2884 * See mem_cgroup_add_lru_list(), etc.
2887 SetPageCgroupUsed(pc);
2891 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2892 VM_BUG_ON(PageLRU(page));
2894 add_page_to_lru_list(page, lruvec, page_lru(page));
2896 spin_unlock_irq(&zone->lru_lock);
2899 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2904 mem_cgroup_charge_statistics(memcg, page, anon, nr_pages);
2905 unlock_page_cgroup(pc);
2908 * "charge_statistics" updated event counter. Then, check it.
2909 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2910 * if they exceeds softlimit.
2912 memcg_check_events(memcg, page);
2915 static DEFINE_MUTEX(set_limit_mutex);
2917 #ifdef CONFIG_MEMCG_KMEM
2918 static inline bool memcg_can_account_kmem(struct mem_cgroup *memcg)
2920 return !mem_cgroup_disabled() && !mem_cgroup_is_root(memcg) &&
2921 (memcg->kmem_account_flags & KMEM_ACCOUNTED_MASK);
2925 * This is a bit cumbersome, but it is rarely used and avoids a backpointer
2926 * in the memcg_cache_params struct.
2928 static struct kmem_cache *memcg_params_to_cache(struct memcg_cache_params *p)
2930 struct kmem_cache *cachep;
2932 VM_BUG_ON(p->is_root_cache);
2933 cachep = p->root_cache;
2934 return cachep->memcg_params->memcg_caches[memcg_cache_id(p->memcg)];
2937 #ifdef CONFIG_SLABINFO
2938 static int mem_cgroup_slabinfo_read(struct cgroup_subsys_state *css,
2939 struct cftype *cft, struct seq_file *m)
2941 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
2942 struct memcg_cache_params *params;
2944 if (!memcg_can_account_kmem(memcg))
2947 print_slabinfo_header(m);
2949 mutex_lock(&memcg->slab_caches_mutex);
2950 list_for_each_entry(params, &memcg->memcg_slab_caches, list)
2951 cache_show(memcg_params_to_cache(params), m);
2952 mutex_unlock(&memcg->slab_caches_mutex);
2958 static int memcg_charge_kmem(struct mem_cgroup *memcg, gfp_t gfp, u64 size)
2960 struct res_counter *fail_res;
2961 struct mem_cgroup *_memcg;
2965 ret = res_counter_charge(&memcg->kmem, size, &fail_res);
2970 * Conditions under which we can wait for the oom_killer. Those are
2971 * the same conditions tested by the core page allocator
2973 may_oom = (gfp & __GFP_FS) && !(gfp & __GFP_NORETRY);
2976 ret = __mem_cgroup_try_charge(NULL, gfp, size >> PAGE_SHIFT,
2979 if (ret == -EINTR) {
2981 * __mem_cgroup_try_charge() chosed to bypass to root due to
2982 * OOM kill or fatal signal. Since our only options are to
2983 * either fail the allocation or charge it to this cgroup, do
2984 * it as a temporary condition. But we can't fail. From a
2985 * kmem/slab perspective, the cache has already been selected,
2986 * by mem_cgroup_kmem_get_cache(), so it is too late to change
2989 * This condition will only trigger if the task entered
2990 * memcg_charge_kmem in a sane state, but was OOM-killed during
2991 * __mem_cgroup_try_charge() above. Tasks that were already
2992 * dying when the allocation triggers should have been already
2993 * directed to the root cgroup in memcontrol.h
2995 res_counter_charge_nofail(&memcg->res, size, &fail_res);
2996 if (do_swap_account)
2997 res_counter_charge_nofail(&memcg->memsw, size,
3001 res_counter_uncharge(&memcg->kmem, size);
3006 static void memcg_uncharge_kmem(struct mem_cgroup *memcg, u64 size)
3008 res_counter_uncharge(&memcg->res, size);
3009 if (do_swap_account)
3010 res_counter_uncharge(&memcg->memsw, size);
3013 if (res_counter_uncharge(&memcg->kmem, size))
3017 * Releases a reference taken in kmem_cgroup_css_offline in case
3018 * this last uncharge is racing with the offlining code or it is
3019 * outliving the memcg existence.
3021 * The memory barrier imposed by test&clear is paired with the
3022 * explicit one in memcg_kmem_mark_dead().
3024 if (memcg_kmem_test_and_clear_dead(memcg))
3025 css_put(&memcg->css);
3028 void memcg_cache_list_add(struct mem_cgroup *memcg, struct kmem_cache *cachep)
3033 mutex_lock(&memcg->slab_caches_mutex);
3034 list_add(&cachep->memcg_params->list, &memcg->memcg_slab_caches);
3035 mutex_unlock(&memcg->slab_caches_mutex);
3039 * helper for acessing a memcg's index. It will be used as an index in the
3040 * child cache array in kmem_cache, and also to derive its name. This function
3041 * will return -1 when this is not a kmem-limited memcg.
3043 int memcg_cache_id(struct mem_cgroup *memcg)
3045 return memcg ? memcg->kmemcg_id : -1;
3049 * This ends up being protected by the set_limit mutex, during normal
3050 * operation, because that is its main call site.
3052 * But when we create a new cache, we can call this as well if its parent
3053 * is kmem-limited. That will have to hold set_limit_mutex as well.
3055 int memcg_update_cache_sizes(struct mem_cgroup *memcg)
3059 num = ida_simple_get(&kmem_limited_groups,
3060 0, MEMCG_CACHES_MAX_SIZE, GFP_KERNEL);
3064 * After this point, kmem_accounted (that we test atomically in
3065 * the beginning of this conditional), is no longer 0. This
3066 * guarantees only one process will set the following boolean
3067 * to true. We don't need test_and_set because we're protected
3068 * by the set_limit_mutex anyway.
3070 memcg_kmem_set_activated(memcg);
3072 ret = memcg_update_all_caches(num+1);
3074 ida_simple_remove(&kmem_limited_groups, num);
3075 memcg_kmem_clear_activated(memcg);
3079 memcg->kmemcg_id = num;
3080 INIT_LIST_HEAD(&memcg->memcg_slab_caches);
3081 mutex_init(&memcg->slab_caches_mutex);
3085 static size_t memcg_caches_array_size(int num_groups)
3088 if (num_groups <= 0)
3091 size = 2 * num_groups;
3092 if (size < MEMCG_CACHES_MIN_SIZE)
3093 size = MEMCG_CACHES_MIN_SIZE;
3094 else if (size > MEMCG_CACHES_MAX_SIZE)
3095 size = MEMCG_CACHES_MAX_SIZE;
3101 * We should update the current array size iff all caches updates succeed. This
3102 * can only be done from the slab side. The slab mutex needs to be held when
3105 void memcg_update_array_size(int num)
3107 if (num > memcg_limited_groups_array_size)
3108 memcg_limited_groups_array_size = memcg_caches_array_size(num);
3111 static void kmem_cache_destroy_work_func(struct work_struct *w);
3113 int memcg_update_cache_size(struct kmem_cache *s, int num_groups)
3115 struct memcg_cache_params *cur_params = s->memcg_params;
3117 VM_BUG_ON(s->memcg_params && !s->memcg_params->is_root_cache);
3119 if (num_groups > memcg_limited_groups_array_size) {
3121 ssize_t size = memcg_caches_array_size(num_groups);
3123 size *= sizeof(void *);
3124 size += offsetof(struct memcg_cache_params, memcg_caches);
3126 s->memcg_params = kzalloc(size, GFP_KERNEL);
3127 if (!s->memcg_params) {
3128 s->memcg_params = cur_params;
3132 s->memcg_params->is_root_cache = true;
3135 * There is the chance it will be bigger than
3136 * memcg_limited_groups_array_size, if we failed an allocation
3137 * in a cache, in which case all caches updated before it, will
3138 * have a bigger array.
3140 * But if that is the case, the data after
3141 * memcg_limited_groups_array_size is certainly unused
3143 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3144 if (!cur_params->memcg_caches[i])
3146 s->memcg_params->memcg_caches[i] =
3147 cur_params->memcg_caches[i];
3151 * Ideally, we would wait until all caches succeed, and only
3152 * then free the old one. But this is not worth the extra
3153 * pointer per-cache we'd have to have for this.
3155 * It is not a big deal if some caches are left with a size
3156 * bigger than the others. And all updates will reset this
3164 int memcg_register_cache(struct mem_cgroup *memcg, struct kmem_cache *s,
3165 struct kmem_cache *root_cache)
3169 if (!memcg_kmem_enabled())
3173 size = offsetof(struct memcg_cache_params, memcg_caches);
3174 size += memcg_limited_groups_array_size * sizeof(void *);
3176 size = sizeof(struct memcg_cache_params);
3178 s->memcg_params = kzalloc(size, GFP_KERNEL);
3179 if (!s->memcg_params)
3183 s->memcg_params->memcg = memcg;
3184 s->memcg_params->root_cache = root_cache;
3185 INIT_WORK(&s->memcg_params->destroy,
3186 kmem_cache_destroy_work_func);
3188 s->memcg_params->is_root_cache = true;
3193 void memcg_release_cache(struct kmem_cache *s)
3195 struct kmem_cache *root;
3196 struct mem_cgroup *memcg;
3200 * This happens, for instance, when a root cache goes away before we
3203 if (!s->memcg_params)
3206 if (s->memcg_params->is_root_cache)
3209 memcg = s->memcg_params->memcg;
3210 id = memcg_cache_id(memcg);
3212 root = s->memcg_params->root_cache;
3213 root->memcg_params->memcg_caches[id] = NULL;
3215 mutex_lock(&memcg->slab_caches_mutex);
3216 list_del(&s->memcg_params->list);
3217 mutex_unlock(&memcg->slab_caches_mutex);
3219 css_put(&memcg->css);
3221 kfree(s->memcg_params);
3225 * During the creation a new cache, we need to disable our accounting mechanism
3226 * altogether. This is true even if we are not creating, but rather just
3227 * enqueing new caches to be created.
3229 * This is because that process will trigger allocations; some visible, like
3230 * explicit kmallocs to auxiliary data structures, name strings and internal
3231 * cache structures; some well concealed, like INIT_WORK() that can allocate
3232 * objects during debug.
3234 * If any allocation happens during memcg_kmem_get_cache, we will recurse back
3235 * to it. This may not be a bounded recursion: since the first cache creation
3236 * failed to complete (waiting on the allocation), we'll just try to create the
3237 * cache again, failing at the same point.
3239 * memcg_kmem_get_cache is prepared to abort after seeing a positive count of
3240 * memcg_kmem_skip_account. So we enclose anything that might allocate memory
3241 * inside the following two functions.
3243 static inline void memcg_stop_kmem_account(void)
3245 VM_BUG_ON(!current->mm);
3246 current->memcg_kmem_skip_account++;
3249 static inline void memcg_resume_kmem_account(void)
3251 VM_BUG_ON(!current->mm);
3252 current->memcg_kmem_skip_account--;
3255 static void kmem_cache_destroy_work_func(struct work_struct *w)
3257 struct kmem_cache *cachep;
3258 struct memcg_cache_params *p;
3260 p = container_of(w, struct memcg_cache_params, destroy);
3262 cachep = memcg_params_to_cache(p);
3265 * If we get down to 0 after shrink, we could delete right away.
3266 * However, memcg_release_pages() already puts us back in the workqueue
3267 * in that case. If we proceed deleting, we'll get a dangling
3268 * reference, and removing the object from the workqueue in that case
3269 * is unnecessary complication. We are not a fast path.
3271 * Note that this case is fundamentally different from racing with
3272 * shrink_slab(): if memcg_cgroup_destroy_cache() is called in
3273 * kmem_cache_shrink, not only we would be reinserting a dead cache
3274 * into the queue, but doing so from inside the worker racing to
3277 * So if we aren't down to zero, we'll just schedule a worker and try
3280 if (atomic_read(&cachep->memcg_params->nr_pages) != 0) {
3281 kmem_cache_shrink(cachep);
3282 if (atomic_read(&cachep->memcg_params->nr_pages) == 0)
3285 kmem_cache_destroy(cachep);
3288 void mem_cgroup_destroy_cache(struct kmem_cache *cachep)
3290 if (!cachep->memcg_params->dead)
3294 * There are many ways in which we can get here.
3296 * We can get to a memory-pressure situation while the delayed work is
3297 * still pending to run. The vmscan shrinkers can then release all
3298 * cache memory and get us to destruction. If this is the case, we'll
3299 * be executed twice, which is a bug (the second time will execute over
3300 * bogus data). In this case, cancelling the work should be fine.
3302 * But we can also get here from the worker itself, if
3303 * kmem_cache_shrink is enough to shake all the remaining objects and
3304 * get the page count to 0. In this case, we'll deadlock if we try to
3305 * cancel the work (the worker runs with an internal lock held, which
3306 * is the same lock we would hold for cancel_work_sync().)
3308 * Since we can't possibly know who got us here, just refrain from
3309 * running if there is already work pending
3311 if (work_pending(&cachep->memcg_params->destroy))
3314 * We have to defer the actual destroying to a workqueue, because
3315 * we might currently be in a context that cannot sleep.
3317 schedule_work(&cachep->memcg_params->destroy);
3321 * This lock protects updaters, not readers. We want readers to be as fast as
3322 * they can, and they will either see NULL or a valid cache value. Our model
3323 * allow them to see NULL, in which case the root memcg will be selected.
3325 * We need this lock because multiple allocations to the same cache from a non
3326 * will span more than one worker. Only one of them can create the cache.
3328 static DEFINE_MUTEX(memcg_cache_mutex);
3331 * Called with memcg_cache_mutex held
3333 static struct kmem_cache *kmem_cache_dup(struct mem_cgroup *memcg,
3334 struct kmem_cache *s)
3336 struct kmem_cache *new;
3337 static char *tmp_name = NULL;
3339 lockdep_assert_held(&memcg_cache_mutex);
3342 * kmem_cache_create_memcg duplicates the given name and
3343 * cgroup_name for this name requires RCU context.
3344 * This static temporary buffer is used to prevent from
3345 * pointless shortliving allocation.
3348 tmp_name = kmalloc(PATH_MAX, GFP_KERNEL);
3354 snprintf(tmp_name, PATH_MAX, "%s(%d:%s)", s->name,
3355 memcg_cache_id(memcg), cgroup_name(memcg->css.cgroup));
3358 new = kmem_cache_create_memcg(memcg, tmp_name, s->object_size, s->align,
3359 (s->flags & ~SLAB_PANIC), s->ctor, s);
3362 new->allocflags |= __GFP_KMEMCG;
3367 static struct kmem_cache *memcg_create_kmem_cache(struct mem_cgroup *memcg,
3368 struct kmem_cache *cachep)
3370 struct kmem_cache *new_cachep;
3373 BUG_ON(!memcg_can_account_kmem(memcg));
3375 idx = memcg_cache_id(memcg);
3377 mutex_lock(&memcg_cache_mutex);
3378 new_cachep = cachep->memcg_params->memcg_caches[idx];
3380 css_put(&memcg->css);
3384 new_cachep = kmem_cache_dup(memcg, cachep);
3385 if (new_cachep == NULL) {
3386 new_cachep = cachep;
3387 css_put(&memcg->css);
3391 atomic_set(&new_cachep->memcg_params->nr_pages , 0);
3393 cachep->memcg_params->memcg_caches[idx] = new_cachep;
3395 * the readers won't lock, make sure everybody sees the updated value,
3396 * so they won't put stuff in the queue again for no reason
3400 mutex_unlock(&memcg_cache_mutex);
3404 void kmem_cache_destroy_memcg_children(struct kmem_cache *s)
3406 struct kmem_cache *c;
3409 if (!s->memcg_params)
3411 if (!s->memcg_params->is_root_cache)
3415 * If the cache is being destroyed, we trust that there is no one else
3416 * requesting objects from it. Even if there are, the sanity checks in
3417 * kmem_cache_destroy should caught this ill-case.
3419 * Still, we don't want anyone else freeing memcg_caches under our
3420 * noses, which can happen if a new memcg comes to life. As usual,
3421 * we'll take the set_limit_mutex to protect ourselves against this.
3423 mutex_lock(&set_limit_mutex);
3424 for (i = 0; i < memcg_limited_groups_array_size; i++) {
3425 c = s->memcg_params->memcg_caches[i];
3430 * We will now manually delete the caches, so to avoid races
3431 * we need to cancel all pending destruction workers and
3432 * proceed with destruction ourselves.
3434 * kmem_cache_destroy() will call kmem_cache_shrink internally,
3435 * and that could spawn the workers again: it is likely that
3436 * the cache still have active pages until this very moment.
3437 * This would lead us back to mem_cgroup_destroy_cache.
3439 * But that will not execute at all if the "dead" flag is not
3440 * set, so flip it down to guarantee we are in control.
3442 c->memcg_params->dead = false;
3443 cancel_work_sync(&c->memcg_params->destroy);
3444 kmem_cache_destroy(c);
3446 mutex_unlock(&set_limit_mutex);
3449 struct create_work {
3450 struct mem_cgroup *memcg;
3451 struct kmem_cache *cachep;
3452 struct work_struct work;
3455 static void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3457 struct kmem_cache *cachep;
3458 struct memcg_cache_params *params;
3460 if (!memcg_kmem_is_active(memcg))
3463 mutex_lock(&memcg->slab_caches_mutex);
3464 list_for_each_entry(params, &memcg->memcg_slab_caches, list) {
3465 cachep = memcg_params_to_cache(params);
3466 cachep->memcg_params->dead = true;
3467 schedule_work(&cachep->memcg_params->destroy);
3469 mutex_unlock(&memcg->slab_caches_mutex);
3472 static void memcg_create_cache_work_func(struct work_struct *w)
3474 struct create_work *cw;
3476 cw = container_of(w, struct create_work, work);
3477 memcg_create_kmem_cache(cw->memcg, cw->cachep);
3482 * Enqueue the creation of a per-memcg kmem_cache.
3484 static void __memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3485 struct kmem_cache *cachep)
3487 struct create_work *cw;
3489 cw = kmalloc(sizeof(struct create_work), GFP_NOWAIT);
3491 css_put(&memcg->css);
3496 cw->cachep = cachep;
3498 INIT_WORK(&cw->work, memcg_create_cache_work_func);
3499 schedule_work(&cw->work);
3502 static void memcg_create_cache_enqueue(struct mem_cgroup *memcg,
3503 struct kmem_cache *cachep)
3506 * We need to stop accounting when we kmalloc, because if the
3507 * corresponding kmalloc cache is not yet created, the first allocation
3508 * in __memcg_create_cache_enqueue will recurse.
3510 * However, it is better to enclose the whole function. Depending on
3511 * the debugging options enabled, INIT_WORK(), for instance, can
3512 * trigger an allocation. This too, will make us recurse. Because at
3513 * this point we can't allow ourselves back into memcg_kmem_get_cache,
3514 * the safest choice is to do it like this, wrapping the whole function.
3516 memcg_stop_kmem_account();
3517 __memcg_create_cache_enqueue(memcg, cachep);
3518 memcg_resume_kmem_account();
3521 * Return the kmem_cache we're supposed to use for a slab allocation.
3522 * We try to use the current memcg's version of the cache.
3524 * If the cache does not exist yet, if we are the first user of it,
3525 * we either create it immediately, if possible, or create it asynchronously
3527 * In the latter case, we will let the current allocation go through with
3528 * the original cache.
3530 * Can't be called in interrupt context or from kernel threads.
3531 * This function needs to be called with rcu_read_lock() held.
3533 struct kmem_cache *__memcg_kmem_get_cache(struct kmem_cache *cachep,
3536 struct mem_cgroup *memcg;
3539 VM_BUG_ON(!cachep->memcg_params);
3540 VM_BUG_ON(!cachep->memcg_params->is_root_cache);
3542 if (!current->mm || current->memcg_kmem_skip_account)
3546 memcg = mem_cgroup_from_task(rcu_dereference(current->mm->owner));
3548 if (!memcg_can_account_kmem(memcg))
3551 idx = memcg_cache_id(memcg);
3554 * barrier to mare sure we're always seeing the up to date value. The
3555 * code updating memcg_caches will issue a write barrier to match this.
3557 read_barrier_depends();
3558 if (likely(cachep->memcg_params->memcg_caches[idx])) {
3559 cachep = cachep->memcg_params->memcg_caches[idx];
3563 /* The corresponding put will be done in the workqueue. */
3564 if (!css_tryget(&memcg->css))
3569 * If we are in a safe context (can wait, and not in interrupt
3570 * context), we could be be predictable and return right away.
3571 * This would guarantee that the allocation being performed
3572 * already belongs in the new cache.
3574 * However, there are some clashes that can arrive from locking.
3575 * For instance, because we acquire the slab_mutex while doing
3576 * kmem_cache_dup, this means no further allocation could happen
3577 * with the slab_mutex held.
3579 * Also, because cache creation issue get_online_cpus(), this
3580 * creates a lock chain: memcg_slab_mutex -> cpu_hotplug_mutex,
3581 * that ends up reversed during cpu hotplug. (cpuset allocates
3582 * a bunch of GFP_KERNEL memory during cpuup). Due to all that,
3583 * better to defer everything.
3585 memcg_create_cache_enqueue(memcg, cachep);
3591 EXPORT_SYMBOL(__memcg_kmem_get_cache);
3594 * We need to verify if the allocation against current->mm->owner's memcg is
3595 * possible for the given order. But the page is not allocated yet, so we'll
3596 * need a further commit step to do the final arrangements.
3598 * It is possible for the task to switch cgroups in this mean time, so at
3599 * commit time, we can't rely on task conversion any longer. We'll then use
3600 * the handle argument to return to the caller which cgroup we should commit
3601 * against. We could also return the memcg directly and avoid the pointer
3602 * passing, but a boolean return value gives better semantics considering
3603 * the compiled-out case as well.
3605 * Returning true means the allocation is possible.
3608 __memcg_kmem_newpage_charge(gfp_t gfp, struct mem_cgroup **_memcg, int order)
3610 struct mem_cgroup *memcg;
3616 * Disabling accounting is only relevant for some specific memcg
3617 * internal allocations. Therefore we would initially not have such
3618 * check here, since direct calls to the page allocator that are marked
3619 * with GFP_KMEMCG only happen outside memcg core. We are mostly
3620 * concerned with cache allocations, and by having this test at
3621 * memcg_kmem_get_cache, we are already able to relay the allocation to
3622 * the root cache and bypass the memcg cache altogether.
3624 * There is one exception, though: the SLUB allocator does not create
3625 * large order caches, but rather service large kmallocs directly from
3626 * the page allocator. Therefore, the following sequence when backed by
3627 * the SLUB allocator:
3629 * memcg_stop_kmem_account();
3630 * kmalloc(<large_number>)
3631 * memcg_resume_kmem_account();
3633 * would effectively ignore the fact that we should skip accounting,
3634 * since it will drive us directly to this function without passing
3635 * through the cache selector memcg_kmem_get_cache. Such large
3636 * allocations are extremely rare but can happen, for instance, for the
3637 * cache arrays. We bring this test here.
3639 if (!current->mm || current->memcg_kmem_skip_account)
3642 memcg = try_get_mem_cgroup_from_mm(current->mm);
3645 * very rare case described in mem_cgroup_from_task. Unfortunately there
3646 * isn't much we can do without complicating this too much, and it would
3647 * be gfp-dependent anyway. Just let it go
3649 if (unlikely(!memcg))
3652 if (!memcg_can_account_kmem(memcg)) {
3653 css_put(&memcg->css);
3657 ret = memcg_charge_kmem(memcg, gfp, PAGE_SIZE << order);
3661 css_put(&memcg->css);
3665 void __memcg_kmem_commit_charge(struct page *page, struct mem_cgroup *memcg,
3668 struct page_cgroup *pc;
3670 VM_BUG_ON(mem_cgroup_is_root(memcg));
3672 /* The page allocation failed. Revert */
3674 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3678 pc = lookup_page_cgroup(page);
3679 lock_page_cgroup(pc);
3680 pc->mem_cgroup = memcg;
3681 SetPageCgroupUsed(pc);
3682 unlock_page_cgroup(pc);
3685 void __memcg_kmem_uncharge_pages(struct page *page, int order)
3687 struct mem_cgroup *memcg = NULL;
3688 struct page_cgroup *pc;
3691 pc = lookup_page_cgroup(page);
3693 * Fast unlocked return. Theoretically might have changed, have to
3694 * check again after locking.
3696 if (!PageCgroupUsed(pc))
3699 lock_page_cgroup(pc);
3700 if (PageCgroupUsed(pc)) {
3701 memcg = pc->mem_cgroup;
3702 ClearPageCgroupUsed(pc);
3704 unlock_page_cgroup(pc);
3707 * We trust that only if there is a memcg associated with the page, it
3708 * is a valid allocation
3713 VM_BUG_ON(mem_cgroup_is_root(memcg));
3714 memcg_uncharge_kmem(memcg, PAGE_SIZE << order);
3717 static inline void mem_cgroup_destroy_all_caches(struct mem_cgroup *memcg)
3720 #endif /* CONFIG_MEMCG_KMEM */
3722 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3724 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
3726 * Because tail pages are not marked as "used", set it. We're under
3727 * zone->lru_lock, 'splitting on pmd' and compound_lock.
3728 * charge/uncharge will be never happen and move_account() is done under
3729 * compound_lock(), so we don't have to take care of races.
3731 void mem_cgroup_split_huge_fixup(struct page *head)
3733 struct page_cgroup *head_pc = lookup_page_cgroup(head);
3734 struct page_cgroup *pc;
3735 struct mem_cgroup *memcg;
3738 if (mem_cgroup_disabled())
3741 memcg = head_pc->mem_cgroup;
3742 for (i = 1; i < HPAGE_PMD_NR; i++) {
3744 pc->mem_cgroup = memcg;
3745 smp_wmb();/* see __commit_charge() */
3746 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
3748 __this_cpu_sub(memcg->stat->count[MEM_CGROUP_STAT_RSS_HUGE],
3751 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3754 * mem_cgroup_move_account - move account of the page
3756 * @nr_pages: number of regular pages (>1 for huge pages)
3757 * @pc: page_cgroup of the page.
3758 * @from: mem_cgroup which the page is moved from.
3759 * @to: mem_cgroup which the page is moved to. @from != @to.
3761 * The caller must confirm following.
3762 * - page is not on LRU (isolate_page() is useful.)
3763 * - compound_lock is held when nr_pages > 1
3765 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
3768 static int mem_cgroup_move_account(struct page *page,
3769 unsigned int nr_pages,
3770 struct page_cgroup *pc,
3771 struct mem_cgroup *from,
3772 struct mem_cgroup *to)
3774 unsigned long flags;
3776 bool anon = PageAnon(page);
3778 VM_BUG_ON(from == to);
3779 VM_BUG_ON(PageLRU(page));
3781 * The page is isolated from LRU. So, collapse function
3782 * will not handle this page. But page splitting can happen.
3783 * Do this check under compound_page_lock(). The caller should
3787 if (nr_pages > 1 && !PageTransHuge(page))
3790 lock_page_cgroup(pc);
3793 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
3796 move_lock_mem_cgroup(from, &flags);
3798 if (!anon && page_mapped(page)) {
3799 /* Update mapped_file data for mem_cgroup */
3801 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3802 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
3805 mem_cgroup_charge_statistics(from, page, anon, -nr_pages);
3807 /* caller should have done css_get */
3808 pc->mem_cgroup = to;
3809 mem_cgroup_charge_statistics(to, page, anon, nr_pages);
3810 move_unlock_mem_cgroup(from, &flags);
3813 unlock_page_cgroup(pc);
3817 memcg_check_events(to, page);
3818 memcg_check_events(from, page);
3824 * mem_cgroup_move_parent - moves page to the parent group
3825 * @page: the page to move
3826 * @pc: page_cgroup of the page
3827 * @child: page's cgroup
3829 * move charges to its parent or the root cgroup if the group has no
3830 * parent (aka use_hierarchy==0).
3831 * Although this might fail (get_page_unless_zero, isolate_lru_page or
3832 * mem_cgroup_move_account fails) the failure is always temporary and
3833 * it signals a race with a page removal/uncharge or migration. In the
3834 * first case the page is on the way out and it will vanish from the LRU
3835 * on the next attempt and the call should be retried later.
3836 * Isolation from the LRU fails only if page has been isolated from
3837 * the LRU since we looked at it and that usually means either global
3838 * reclaim or migration going on. The page will either get back to the
3840 * Finaly mem_cgroup_move_account fails only if the page got uncharged
3841 * (!PageCgroupUsed) or moved to a different group. The page will
3842 * disappear in the next attempt.
3844 static int mem_cgroup_move_parent(struct page *page,
3845 struct page_cgroup *pc,
3846 struct mem_cgroup *child)
3848 struct mem_cgroup *parent;
3849 unsigned int nr_pages;
3850 unsigned long uninitialized_var(flags);
3853 VM_BUG_ON(mem_cgroup_is_root(child));
3856 if (!get_page_unless_zero(page))
3858 if (isolate_lru_page(page))
3861 nr_pages = hpage_nr_pages(page);
3863 parent = parent_mem_cgroup(child);
3865 * If no parent, move charges to root cgroup.
3868 parent = root_mem_cgroup;
3871 VM_BUG_ON(!PageTransHuge(page));
3872 flags = compound_lock_irqsave(page);
3875 ret = mem_cgroup_move_account(page, nr_pages,
3878 __mem_cgroup_cancel_local_charge(child, nr_pages);
3881 compound_unlock_irqrestore(page, flags);
3882 putback_lru_page(page);
3890 * Charge the memory controller for page usage.
3892 * 0 if the charge was successful
3893 * < 0 if the cgroup is over its limit
3895 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
3896 gfp_t gfp_mask, enum charge_type ctype)
3898 struct mem_cgroup *memcg = NULL;
3899 unsigned int nr_pages = 1;
3903 if (PageTransHuge(page)) {
3904 nr_pages <<= compound_order(page);
3905 VM_BUG_ON(!PageTransHuge(page));
3907 * Never OOM-kill a process for a huge page. The
3908 * fault handler will fall back to regular pages.
3913 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
3916 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
3920 int mem_cgroup_newpage_charge(struct page *page,
3921 struct mm_struct *mm, gfp_t gfp_mask)
3923 if (mem_cgroup_disabled())
3925 VM_BUG_ON(page_mapped(page));
3926 VM_BUG_ON(page->mapping && !PageAnon(page));
3928 return mem_cgroup_charge_common(page, mm, gfp_mask,
3929 MEM_CGROUP_CHARGE_TYPE_ANON);
3933 * While swap-in, try_charge -> commit or cancel, the page is locked.
3934 * And when try_charge() successfully returns, one refcnt to memcg without
3935 * struct page_cgroup is acquired. This refcnt will be consumed by
3936 * "commit()" or removed by "cancel()"
3938 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
3941 struct mem_cgroup **memcgp)
3943 struct mem_cgroup *memcg;
3944 struct page_cgroup *pc;
3947 pc = lookup_page_cgroup(page);
3949 * Every swap fault against a single page tries to charge the
3950 * page, bail as early as possible. shmem_unuse() encounters
3951 * already charged pages, too. The USED bit is protected by
3952 * the page lock, which serializes swap cache removal, which
3953 * in turn serializes uncharging.
3955 if (PageCgroupUsed(pc))
3957 if (!do_swap_account)
3959 memcg = try_get_mem_cgroup_from_page(page);
3963 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
3964 css_put(&memcg->css);
3969 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
3975 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
3976 gfp_t gfp_mask, struct mem_cgroup **memcgp)
3979 if (mem_cgroup_disabled())
3982 * A racing thread's fault, or swapoff, may have already
3983 * updated the pte, and even removed page from swap cache: in
3984 * those cases unuse_pte()'s pte_same() test will fail; but
3985 * there's also a KSM case which does need to charge the page.
3987 if (!PageSwapCache(page)) {
3990 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
3995 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
3998 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
4000 if (mem_cgroup_disabled())
4004 __mem_cgroup_cancel_charge(memcg, 1);
4008 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
4009 enum charge_type ctype)
4011 if (mem_cgroup_disabled())
4016 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
4018 * Now swap is on-memory. This means this page may be
4019 * counted both as mem and swap....double count.
4020 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
4021 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
4022 * may call delete_from_swap_cache() before reach here.
4024 if (do_swap_account && PageSwapCache(page)) {
4025 swp_entry_t ent = {.val = page_private(page)};
4026 mem_cgroup_uncharge_swap(ent);
4030 void mem_cgroup_commit_charge_swapin(struct page *page,
4031 struct mem_cgroup *memcg)
4033 __mem_cgroup_commit_charge_swapin(page, memcg,
4034 MEM_CGROUP_CHARGE_TYPE_ANON);
4037 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
4040 struct mem_cgroup *memcg = NULL;
4041 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4044 if (mem_cgroup_disabled())
4046 if (PageCompound(page))
4049 if (!PageSwapCache(page))
4050 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
4051 else { /* page is swapcache/shmem */
4052 ret = __mem_cgroup_try_charge_swapin(mm, page,
4055 __mem_cgroup_commit_charge_swapin(page, memcg, type);
4060 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
4061 unsigned int nr_pages,
4062 const enum charge_type ctype)
4064 struct memcg_batch_info *batch = NULL;
4065 bool uncharge_memsw = true;
4067 /* If swapout, usage of swap doesn't decrease */
4068 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
4069 uncharge_memsw = false;
4071 batch = ¤t->memcg_batch;
4073 * In usual, we do css_get() when we remember memcg pointer.
4074 * But in this case, we keep res->usage until end of a series of
4075 * uncharges. Then, it's ok to ignore memcg's refcnt.
4078 batch->memcg = memcg;
4080 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
4081 * In those cases, all pages freed continuously can be expected to be in
4082 * the same cgroup and we have chance to coalesce uncharges.
4083 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
4084 * because we want to do uncharge as soon as possible.
4087 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
4088 goto direct_uncharge;
4091 goto direct_uncharge;
4094 * In typical case, batch->memcg == mem. This means we can
4095 * merge a series of uncharges to an uncharge of res_counter.
4096 * If not, we uncharge res_counter ony by one.
4098 if (batch->memcg != memcg)
4099 goto direct_uncharge;
4100 /* remember freed charge and uncharge it later */
4103 batch->memsw_nr_pages++;
4106 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
4108 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
4109 if (unlikely(batch->memcg != memcg))
4110 memcg_oom_recover(memcg);
4114 * uncharge if !page_mapped(page)
4116 static struct mem_cgroup *
4117 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
4120 struct mem_cgroup *memcg = NULL;
4121 unsigned int nr_pages = 1;
4122 struct page_cgroup *pc;
4125 if (mem_cgroup_disabled())
4128 if (PageTransHuge(page)) {
4129 nr_pages <<= compound_order(page);
4130 VM_BUG_ON(!PageTransHuge(page));
4133 * Check if our page_cgroup is valid
4135 pc = lookup_page_cgroup(page);
4136 if (unlikely(!PageCgroupUsed(pc)))
4139 lock_page_cgroup(pc);
4141 memcg = pc->mem_cgroup;
4143 if (!PageCgroupUsed(pc))
4146 anon = PageAnon(page);
4149 case MEM_CGROUP_CHARGE_TYPE_ANON:
4151 * Generally PageAnon tells if it's the anon statistics to be
4152 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
4153 * used before page reached the stage of being marked PageAnon.
4157 case MEM_CGROUP_CHARGE_TYPE_DROP:
4158 /* See mem_cgroup_prepare_migration() */
4159 if (page_mapped(page))
4162 * Pages under migration may not be uncharged. But
4163 * end_migration() /must/ be the one uncharging the
4164 * unused post-migration page and so it has to call
4165 * here with the migration bit still set. See the
4166 * res_counter handling below.
4168 if (!end_migration && PageCgroupMigration(pc))
4171 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
4172 if (!PageAnon(page)) { /* Shared memory */
4173 if (page->mapping && !page_is_file_cache(page))
4175 } else if (page_mapped(page)) /* Anon */
4182 mem_cgroup_charge_statistics(memcg, page, anon, -nr_pages);
4184 ClearPageCgroupUsed(pc);
4186 * pc->mem_cgroup is not cleared here. It will be accessed when it's
4187 * freed from LRU. This is safe because uncharged page is expected not
4188 * to be reused (freed soon). Exception is SwapCache, it's handled by
4189 * special functions.
4192 unlock_page_cgroup(pc);
4194 * even after unlock, we have memcg->res.usage here and this memcg
4195 * will never be freed, so it's safe to call css_get().
4197 memcg_check_events(memcg, page);
4198 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
4199 mem_cgroup_swap_statistics(memcg, true);
4200 css_get(&memcg->css);
4203 * Migration does not charge the res_counter for the
4204 * replacement page, so leave it alone when phasing out the
4205 * page that is unused after the migration.
4207 if (!end_migration && !mem_cgroup_is_root(memcg))
4208 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
4213 unlock_page_cgroup(pc);
4217 void mem_cgroup_uncharge_page(struct page *page)
4220 if (page_mapped(page))
4222 VM_BUG_ON(page->mapping && !PageAnon(page));
4224 * If the page is in swap cache, uncharge should be deferred
4225 * to the swap path, which also properly accounts swap usage
4226 * and handles memcg lifetime.
4228 * Note that this check is not stable and reclaim may add the
4229 * page to swap cache at any time after this. However, if the
4230 * page is not in swap cache by the time page->mapcount hits
4231 * 0, there won't be any page table references to the swap
4232 * slot, and reclaim will free it and not actually write the
4235 if (PageSwapCache(page))
4237 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
4240 void mem_cgroup_uncharge_cache_page(struct page *page)
4242 VM_BUG_ON(page_mapped(page));
4243 VM_BUG_ON(page->mapping);
4244 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
4248 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
4249 * In that cases, pages are freed continuously and we can expect pages
4250 * are in the same memcg. All these calls itself limits the number of
4251 * pages freed at once, then uncharge_start/end() is called properly.
4252 * This may be called prural(2) times in a context,
4255 void mem_cgroup_uncharge_start(void)
4257 current->memcg_batch.do_batch++;
4258 /* We can do nest. */
4259 if (current->memcg_batch.do_batch == 1) {
4260 current->memcg_batch.memcg = NULL;
4261 current->memcg_batch.nr_pages = 0;
4262 current->memcg_batch.memsw_nr_pages = 0;
4266 void mem_cgroup_uncharge_end(void)
4268 struct memcg_batch_info *batch = ¤t->memcg_batch;
4270 if (!batch->do_batch)
4274 if (batch->do_batch) /* If stacked, do nothing. */
4280 * This "batch->memcg" is valid without any css_get/put etc...
4281 * bacause we hide charges behind us.
4283 if (batch->nr_pages)
4284 res_counter_uncharge(&batch->memcg->res,
4285 batch->nr_pages * PAGE_SIZE);
4286 if (batch->memsw_nr_pages)
4287 res_counter_uncharge(&batch->memcg->memsw,
4288 batch->memsw_nr_pages * PAGE_SIZE);
4289 memcg_oom_recover(batch->memcg);
4290 /* forget this pointer (for sanity check) */
4291 batch->memcg = NULL;
4296 * called after __delete_from_swap_cache() and drop "page" account.
4297 * memcg information is recorded to swap_cgroup of "ent"
4300 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
4302 struct mem_cgroup *memcg;
4303 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
4305 if (!swapout) /* this was a swap cache but the swap is unused ! */
4306 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
4308 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
4311 * record memcg information, if swapout && memcg != NULL,
4312 * css_get() was called in uncharge().
4314 if (do_swap_account && swapout && memcg)
4315 swap_cgroup_record(ent, css_id(&memcg->css));
4319 #ifdef CONFIG_MEMCG_SWAP
4321 * called from swap_entry_free(). remove record in swap_cgroup and
4322 * uncharge "memsw" account.
4324 void mem_cgroup_uncharge_swap(swp_entry_t ent)
4326 struct mem_cgroup *memcg;
4329 if (!do_swap_account)
4332 id = swap_cgroup_record(ent, 0);
4334 memcg = mem_cgroup_lookup(id);
4337 * We uncharge this because swap is freed.
4338 * This memcg can be obsolete one. We avoid calling css_tryget
4340 if (!mem_cgroup_is_root(memcg))
4341 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
4342 mem_cgroup_swap_statistics(memcg, false);
4343 css_put(&memcg->css);
4349 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
4350 * @entry: swap entry to be moved
4351 * @from: mem_cgroup which the entry is moved from
4352 * @to: mem_cgroup which the entry is moved to
4354 * It succeeds only when the swap_cgroup's record for this entry is the same
4355 * as the mem_cgroup's id of @from.
4357 * Returns 0 on success, -EINVAL on failure.
4359 * The caller must have charged to @to, IOW, called res_counter_charge() about
4360 * both res and memsw, and called css_get().
4362 static int mem_cgroup_move_swap_account(swp_entry_t entry,
4363 struct mem_cgroup *from, struct mem_cgroup *to)
4365 unsigned short old_id, new_id;
4367 old_id = css_id(&from->css);
4368 new_id = css_id(&to->css);
4370 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
4371 mem_cgroup_swap_statistics(from, false);
4372 mem_cgroup_swap_statistics(to, true);
4374 * This function is only called from task migration context now.
4375 * It postpones res_counter and refcount handling till the end
4376 * of task migration(mem_cgroup_clear_mc()) for performance
4377 * improvement. But we cannot postpone css_get(to) because if
4378 * the process that has been moved to @to does swap-in, the
4379 * refcount of @to might be decreased to 0.
4381 * We are in attach() phase, so the cgroup is guaranteed to be
4382 * alive, so we can just call css_get().
4390 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
4391 struct mem_cgroup *from, struct mem_cgroup *to)
4398 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
4401 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
4402 struct mem_cgroup **memcgp)
4404 struct mem_cgroup *memcg = NULL;
4405 unsigned int nr_pages = 1;
4406 struct page_cgroup *pc;
4407 enum charge_type ctype;
4411 if (mem_cgroup_disabled())
4414 if (PageTransHuge(page))
4415 nr_pages <<= compound_order(page);
4417 pc = lookup_page_cgroup(page);
4418 lock_page_cgroup(pc);
4419 if (PageCgroupUsed(pc)) {
4420 memcg = pc->mem_cgroup;
4421 css_get(&memcg->css);
4423 * At migrating an anonymous page, its mapcount goes down
4424 * to 0 and uncharge() will be called. But, even if it's fully
4425 * unmapped, migration may fail and this page has to be
4426 * charged again. We set MIGRATION flag here and delay uncharge
4427 * until end_migration() is called
4429 * Corner Case Thinking
4431 * When the old page was mapped as Anon and it's unmap-and-freed
4432 * while migration was ongoing.
4433 * If unmap finds the old page, uncharge() of it will be delayed
4434 * until end_migration(). If unmap finds a new page, it's
4435 * uncharged when it make mapcount to be 1->0. If unmap code
4436 * finds swap_migration_entry, the new page will not be mapped
4437 * and end_migration() will find it(mapcount==0).
4440 * When the old page was mapped but migraion fails, the kernel
4441 * remaps it. A charge for it is kept by MIGRATION flag even
4442 * if mapcount goes down to 0. We can do remap successfully
4443 * without charging it again.
4446 * The "old" page is under lock_page() until the end of
4447 * migration, so, the old page itself will not be swapped-out.
4448 * If the new page is swapped out before end_migraton, our
4449 * hook to usual swap-out path will catch the event.
4452 SetPageCgroupMigration(pc);
4454 unlock_page_cgroup(pc);
4456 * If the page is not charged at this point,
4464 * We charge new page before it's used/mapped. So, even if unlock_page()
4465 * is called before end_migration, we can catch all events on this new
4466 * page. In the case new page is migrated but not remapped, new page's
4467 * mapcount will be finally 0 and we call uncharge in end_migration().
4470 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
4472 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
4474 * The page is committed to the memcg, but it's not actually
4475 * charged to the res_counter since we plan on replacing the
4476 * old one and only one page is going to be left afterwards.
4478 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
4481 /* remove redundant charge if migration failed*/
4482 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
4483 struct page *oldpage, struct page *newpage, bool migration_ok)
4485 struct page *used, *unused;
4486 struct page_cgroup *pc;
4492 if (!migration_ok) {
4499 anon = PageAnon(used);
4500 __mem_cgroup_uncharge_common(unused,
4501 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
4502 : MEM_CGROUP_CHARGE_TYPE_CACHE,
4504 css_put(&memcg->css);
4506 * We disallowed uncharge of pages under migration because mapcount
4507 * of the page goes down to zero, temporarly.
4508 * Clear the flag and check the page should be charged.
4510 pc = lookup_page_cgroup(oldpage);
4511 lock_page_cgroup(pc);
4512 ClearPageCgroupMigration(pc);
4513 unlock_page_cgroup(pc);
4516 * If a page is a file cache, radix-tree replacement is very atomic
4517 * and we can skip this check. When it was an Anon page, its mapcount
4518 * goes down to 0. But because we added MIGRATION flage, it's not
4519 * uncharged yet. There are several case but page->mapcount check
4520 * and USED bit check in mem_cgroup_uncharge_page() will do enough
4521 * check. (see prepare_charge() also)
4524 mem_cgroup_uncharge_page(used);
4528 * At replace page cache, newpage is not under any memcg but it's on
4529 * LRU. So, this function doesn't touch res_counter but handles LRU
4530 * in correct way. Both pages are locked so we cannot race with uncharge.
4532 void mem_cgroup_replace_page_cache(struct page *oldpage,
4533 struct page *newpage)
4535 struct mem_cgroup *memcg = NULL;
4536 struct page_cgroup *pc;
4537 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
4539 if (mem_cgroup_disabled())
4542 pc = lookup_page_cgroup(oldpage);
4543 /* fix accounting on old pages */
4544 lock_page_cgroup(pc);
4545 if (PageCgroupUsed(pc)) {
4546 memcg = pc->mem_cgroup;
4547 mem_cgroup_charge_statistics(memcg, oldpage, false, -1);
4548 ClearPageCgroupUsed(pc);
4550 unlock_page_cgroup(pc);
4553 * When called from shmem_replace_page(), in some cases the
4554 * oldpage has already been charged, and in some cases not.
4559 * Even if newpage->mapping was NULL before starting replacement,
4560 * the newpage may be on LRU(or pagevec for LRU) already. We lock
4561 * LRU while we overwrite pc->mem_cgroup.
4563 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
4566 #ifdef CONFIG_DEBUG_VM
4567 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
4569 struct page_cgroup *pc;
4571 pc = lookup_page_cgroup(page);
4573 * Can be NULL while feeding pages into the page allocator for
4574 * the first time, i.e. during boot or memory hotplug;
4575 * or when mem_cgroup_disabled().
4577 if (likely(pc) && PageCgroupUsed(pc))
4582 bool mem_cgroup_bad_page_check(struct page *page)
4584 if (mem_cgroup_disabled())
4587 return lookup_page_cgroup_used(page) != NULL;
4590 void mem_cgroup_print_bad_page(struct page *page)
4592 struct page_cgroup *pc;
4594 pc = lookup_page_cgroup_used(page);
4596 pr_alert("pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
4597 pc, pc->flags, pc->mem_cgroup);
4602 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
4603 unsigned long long val)
4606 u64 memswlimit, memlimit;
4608 int children = mem_cgroup_count_children(memcg);
4609 u64 curusage, oldusage;
4613 * For keeping hierarchical_reclaim simple, how long we should retry
4614 * is depends on callers. We set our retry-count to be function
4615 * of # of children which we should visit in this loop.
4617 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
4619 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4622 while (retry_count) {
4623 if (signal_pending(current)) {
4628 * Rather than hide all in some function, I do this in
4629 * open coded manner. You see what this really does.
4630 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4632 mutex_lock(&set_limit_mutex);
4633 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4634 if (memswlimit < val) {
4636 mutex_unlock(&set_limit_mutex);
4640 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4644 ret = res_counter_set_limit(&memcg->res, val);
4646 if (memswlimit == val)
4647 memcg->memsw_is_minimum = true;
4649 memcg->memsw_is_minimum = false;
4651 mutex_unlock(&set_limit_mutex);
4656 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4657 MEM_CGROUP_RECLAIM_SHRINK);
4658 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
4659 /* Usage is reduced ? */
4660 if (curusage >= oldusage)
4663 oldusage = curusage;
4665 if (!ret && enlarge)
4666 memcg_oom_recover(memcg);
4671 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
4672 unsigned long long val)
4675 u64 memlimit, memswlimit, oldusage, curusage;
4676 int children = mem_cgroup_count_children(memcg);
4680 /* see mem_cgroup_resize_res_limit */
4681 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
4682 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4683 while (retry_count) {
4684 if (signal_pending(current)) {
4689 * Rather than hide all in some function, I do this in
4690 * open coded manner. You see what this really does.
4691 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
4693 mutex_lock(&set_limit_mutex);
4694 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4695 if (memlimit > val) {
4697 mutex_unlock(&set_limit_mutex);
4700 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4701 if (memswlimit < val)
4703 ret = res_counter_set_limit(&memcg->memsw, val);
4705 if (memlimit == val)
4706 memcg->memsw_is_minimum = true;
4708 memcg->memsw_is_minimum = false;
4710 mutex_unlock(&set_limit_mutex);
4715 mem_cgroup_reclaim(memcg, GFP_KERNEL,
4716 MEM_CGROUP_RECLAIM_NOSWAP |
4717 MEM_CGROUP_RECLAIM_SHRINK);
4718 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
4719 /* Usage is reduced ? */
4720 if (curusage >= oldusage)
4723 oldusage = curusage;
4725 if (!ret && enlarge)
4726 memcg_oom_recover(memcg);
4730 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
4732 unsigned long *total_scanned)
4734 unsigned long nr_reclaimed = 0;
4735 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
4736 unsigned long reclaimed;
4738 struct mem_cgroup_tree_per_zone *mctz;
4739 unsigned long long excess;
4740 unsigned long nr_scanned;
4745 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
4747 * This loop can run a while, specially if mem_cgroup's continuously
4748 * keep exceeding their soft limit and putting the system under
4755 mz = mem_cgroup_largest_soft_limit_node(mctz);
4760 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
4761 gfp_mask, &nr_scanned);
4762 nr_reclaimed += reclaimed;
4763 *total_scanned += nr_scanned;
4764 spin_lock(&mctz->lock);
4767 * If we failed to reclaim anything from this memory cgroup
4768 * it is time to move on to the next cgroup
4774 * Loop until we find yet another one.
4776 * By the time we get the soft_limit lock
4777 * again, someone might have aded the
4778 * group back on the RB tree. Iterate to
4779 * make sure we get a different mem.
4780 * mem_cgroup_largest_soft_limit_node returns
4781 * NULL if no other cgroup is present on
4785 __mem_cgroup_largest_soft_limit_node(mctz);
4787 css_put(&next_mz->memcg->css);
4788 else /* next_mz == NULL or other memcg */
4792 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
4793 excess = res_counter_soft_limit_excess(&mz->memcg->res);
4795 * One school of thought says that we should not add
4796 * back the node to the tree if reclaim returns 0.
4797 * But our reclaim could return 0, simply because due
4798 * to priority we are exposing a smaller subset of
4799 * memory to reclaim from. Consider this as a longer
4802 /* If excess == 0, no tree ops */
4803 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
4804 spin_unlock(&mctz->lock);
4805 css_put(&mz->memcg->css);
4808 * Could not reclaim anything and there are no more
4809 * mem cgroups to try or we seem to be looping without
4810 * reclaiming anything.
4812 if (!nr_reclaimed &&
4814 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
4816 } while (!nr_reclaimed);
4818 css_put(&next_mz->memcg->css);
4819 return nr_reclaimed;
4823 * mem_cgroup_force_empty_list - clears LRU of a group
4824 * @memcg: group to clear
4827 * @lru: lru to to clear
4829 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
4830 * reclaim the pages page themselves - pages are moved to the parent (or root)
4833 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
4834 int node, int zid, enum lru_list lru)
4836 struct lruvec *lruvec;
4837 unsigned long flags;
4838 struct list_head *list;
4842 zone = &NODE_DATA(node)->node_zones[zid];
4843 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
4844 list = &lruvec->lists[lru];
4848 struct page_cgroup *pc;
4851 spin_lock_irqsave(&zone->lru_lock, flags);
4852 if (list_empty(list)) {
4853 spin_unlock_irqrestore(&zone->lru_lock, flags);
4856 page = list_entry(list->prev, struct page, lru);
4858 list_move(&page->lru, list);
4860 spin_unlock_irqrestore(&zone->lru_lock, flags);
4863 spin_unlock_irqrestore(&zone->lru_lock, flags);
4865 pc = lookup_page_cgroup(page);
4867 if (mem_cgroup_move_parent(page, pc, memcg)) {
4868 /* found lock contention or "pc" is obsolete. */
4873 } while (!list_empty(list));
4877 * make mem_cgroup's charge to be 0 if there is no task by moving
4878 * all the charges and pages to the parent.
4879 * This enables deleting this mem_cgroup.
4881 * Caller is responsible for holding css reference on the memcg.
4883 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
4889 /* This is for making all *used* pages to be on LRU. */
4890 lru_add_drain_all();
4891 drain_all_stock_sync(memcg);
4892 mem_cgroup_start_move(memcg);
4893 for_each_node_state(node, N_MEMORY) {
4894 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4897 mem_cgroup_force_empty_list(memcg,
4902 mem_cgroup_end_move(memcg);
4903 memcg_oom_recover(memcg);
4907 * Kernel memory may not necessarily be trackable to a specific
4908 * process. So they are not migrated, and therefore we can't
4909 * expect their value to drop to 0 here.
4910 * Having res filled up with kmem only is enough.
4912 * This is a safety check because mem_cgroup_force_empty_list
4913 * could have raced with mem_cgroup_replace_page_cache callers
4914 * so the lru seemed empty but the page could have been added
4915 * right after the check. RES_USAGE should be safe as we always
4916 * charge before adding to the LRU.
4918 usage = res_counter_read_u64(&memcg->res, RES_USAGE) -
4919 res_counter_read_u64(&memcg->kmem, RES_USAGE);
4920 } while (usage > 0);
4924 * This mainly exists for tests during the setting of set of use_hierarchy.
4925 * Since this is the very setting we are changing, the current hierarchy value
4928 static inline bool __memcg_has_children(struct mem_cgroup *memcg)
4930 struct cgroup_subsys_state *pos;
4932 /* bounce at first found */
4933 css_for_each_child(pos, &memcg->css)
4939 * Must be called with memcg_create_mutex held, unless the cgroup is guaranteed
4940 * to be already dead (as in mem_cgroup_force_empty, for instance). This is
4941 * from mem_cgroup_count_children(), in the sense that we don't really care how
4942 * many children we have; we only need to know if we have any. It also counts
4943 * any memcg without hierarchy as infertile.
4945 static inline bool memcg_has_children(struct mem_cgroup *memcg)
4947 return memcg->use_hierarchy && __memcg_has_children(memcg);
4951 * Reclaims as many pages from the given memcg as possible and moves
4952 * the rest to the parent.
4954 * Caller is responsible for holding css reference for memcg.
4956 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
4958 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
4959 struct cgroup *cgrp = memcg->css.cgroup;
4961 /* returns EBUSY if there is a task or if we come here twice. */
4962 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
4965 /* we call try-to-free pages for make this cgroup empty */
4966 lru_add_drain_all();
4967 /* try to free all pages in this cgroup */
4968 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
4971 if (signal_pending(current))
4974 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
4978 /* maybe some writeback is necessary */
4979 congestion_wait(BLK_RW_ASYNC, HZ/10);
4984 mem_cgroup_reparent_charges(memcg);
4989 static int mem_cgroup_force_empty_write(struct cgroup_subsys_state *css,
4992 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
4995 if (mem_cgroup_is_root(memcg))
4997 css_get(&memcg->css);
4998 ret = mem_cgroup_force_empty(memcg);
4999 css_put(&memcg->css);
5005 static u64 mem_cgroup_hierarchy_read(struct cgroup_subsys_state *css,
5008 return mem_cgroup_from_css(css)->use_hierarchy;
5011 static int mem_cgroup_hierarchy_write(struct cgroup_subsys_state *css,
5012 struct cftype *cft, u64 val)
5015 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5016 struct mem_cgroup *parent_memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5018 mutex_lock(&memcg_create_mutex);
5020 if (memcg->use_hierarchy == val)
5024 * If parent's use_hierarchy is set, we can't make any modifications
5025 * in the child subtrees. If it is unset, then the change can
5026 * occur, provided the current cgroup has no children.
5028 * For the root cgroup, parent_mem is NULL, we allow value to be
5029 * set if there are no children.
5031 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
5032 (val == 1 || val == 0)) {
5033 if (!__memcg_has_children(memcg))
5034 memcg->use_hierarchy = val;
5041 mutex_unlock(&memcg_create_mutex);
5047 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
5048 enum mem_cgroup_stat_index idx)
5050 struct mem_cgroup *iter;
5053 /* Per-cpu values can be negative, use a signed accumulator */
5054 for_each_mem_cgroup_tree(iter, memcg)
5055 val += mem_cgroup_read_stat(iter, idx);
5057 if (val < 0) /* race ? */
5062 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
5066 if (!mem_cgroup_is_root(memcg)) {
5068 return res_counter_read_u64(&memcg->res, RES_USAGE);
5070 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
5074 * Transparent hugepages are still accounted for in MEM_CGROUP_STAT_RSS
5075 * as well as in MEM_CGROUP_STAT_RSS_HUGE.
5077 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
5078 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
5081 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
5083 return val << PAGE_SHIFT;
5086 static ssize_t mem_cgroup_read(struct cgroup_subsys_state *css,
5087 struct cftype *cft, struct file *file,
5088 char __user *buf, size_t nbytes, loff_t *ppos)
5090 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5096 type = MEMFILE_TYPE(cft->private);
5097 name = MEMFILE_ATTR(cft->private);
5101 if (name == RES_USAGE)
5102 val = mem_cgroup_usage(memcg, false);
5104 val = res_counter_read_u64(&memcg->res, name);
5107 if (name == RES_USAGE)
5108 val = mem_cgroup_usage(memcg, true);
5110 val = res_counter_read_u64(&memcg->memsw, name);
5113 val = res_counter_read_u64(&memcg->kmem, name);
5119 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
5120 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
5123 static int memcg_update_kmem_limit(struct cgroup_subsys_state *css, u64 val)
5126 #ifdef CONFIG_MEMCG_KMEM
5127 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5129 * For simplicity, we won't allow this to be disabled. It also can't
5130 * be changed if the cgroup has children already, or if tasks had
5133 * If tasks join before we set the limit, a person looking at
5134 * kmem.usage_in_bytes will have no way to determine when it took
5135 * place, which makes the value quite meaningless.
5137 * After it first became limited, changes in the value of the limit are
5138 * of course permitted.
5140 mutex_lock(&memcg_create_mutex);
5141 mutex_lock(&set_limit_mutex);
5142 if (!memcg->kmem_account_flags && val != RESOURCE_MAX) {
5143 if (cgroup_task_count(css->cgroup) || memcg_has_children(memcg)) {
5147 ret = res_counter_set_limit(&memcg->kmem, val);
5150 ret = memcg_update_cache_sizes(memcg);
5152 res_counter_set_limit(&memcg->kmem, RESOURCE_MAX);
5155 static_key_slow_inc(&memcg_kmem_enabled_key);
5157 * setting the active bit after the inc will guarantee no one
5158 * starts accounting before all call sites are patched
5160 memcg_kmem_set_active(memcg);
5162 ret = res_counter_set_limit(&memcg->kmem, val);
5164 mutex_unlock(&set_limit_mutex);
5165 mutex_unlock(&memcg_create_mutex);
5170 #ifdef CONFIG_MEMCG_KMEM
5171 static int memcg_propagate_kmem(struct mem_cgroup *memcg)
5174 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
5178 memcg->kmem_account_flags = parent->kmem_account_flags;
5180 * When that happen, we need to disable the static branch only on those
5181 * memcgs that enabled it. To achieve this, we would be forced to
5182 * complicate the code by keeping track of which memcgs were the ones
5183 * that actually enabled limits, and which ones got it from its
5186 * It is a lot simpler just to do static_key_slow_inc() on every child
5187 * that is accounted.
5189 if (!memcg_kmem_is_active(memcg))
5193 * __mem_cgroup_free() will issue static_key_slow_dec() because this
5194 * memcg is active already. If the later initialization fails then the
5195 * cgroup core triggers the cleanup so we do not have to do it here.
5197 static_key_slow_inc(&memcg_kmem_enabled_key);
5199 mutex_lock(&set_limit_mutex);
5200 memcg_stop_kmem_account();
5201 ret = memcg_update_cache_sizes(memcg);
5202 memcg_resume_kmem_account();
5203 mutex_unlock(&set_limit_mutex);
5207 #endif /* CONFIG_MEMCG_KMEM */
5210 * The user of this function is...
5213 static int mem_cgroup_write(struct cgroup_subsys_state *css, struct cftype *cft,
5216 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5219 unsigned long long val;
5222 type = MEMFILE_TYPE(cft->private);
5223 name = MEMFILE_ATTR(cft->private);
5227 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
5231 /* This function does all necessary parse...reuse it */
5232 ret = res_counter_memparse_write_strategy(buffer, &val);
5236 ret = mem_cgroup_resize_limit(memcg, val);
5237 else if (type == _MEMSWAP)
5238 ret = mem_cgroup_resize_memsw_limit(memcg, val);
5239 else if (type == _KMEM)
5240 ret = memcg_update_kmem_limit(css, val);
5244 case RES_SOFT_LIMIT:
5245 ret = res_counter_memparse_write_strategy(buffer, &val);
5249 * For memsw, soft limits are hard to implement in terms
5250 * of semantics, for now, we support soft limits for
5251 * control without swap
5254 ret = res_counter_set_soft_limit(&memcg->res, val);
5259 ret = -EINVAL; /* should be BUG() ? */
5265 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
5266 unsigned long long *mem_limit, unsigned long long *memsw_limit)
5268 unsigned long long min_limit, min_memsw_limit, tmp;
5270 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
5271 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5272 if (!memcg->use_hierarchy)
5275 while (css_parent(&memcg->css)) {
5276 memcg = mem_cgroup_from_css(css_parent(&memcg->css));
5277 if (!memcg->use_hierarchy)
5279 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
5280 min_limit = min(min_limit, tmp);
5281 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
5282 min_memsw_limit = min(min_memsw_limit, tmp);
5285 *mem_limit = min_limit;
5286 *memsw_limit = min_memsw_limit;
5289 static int mem_cgroup_reset(struct cgroup_subsys_state *css, unsigned int event)
5291 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5295 type = MEMFILE_TYPE(event);
5296 name = MEMFILE_ATTR(event);
5301 res_counter_reset_max(&memcg->res);
5302 else if (type == _MEMSWAP)
5303 res_counter_reset_max(&memcg->memsw);
5304 else if (type == _KMEM)
5305 res_counter_reset_max(&memcg->kmem);
5311 res_counter_reset_failcnt(&memcg->res);
5312 else if (type == _MEMSWAP)
5313 res_counter_reset_failcnt(&memcg->memsw);
5314 else if (type == _KMEM)
5315 res_counter_reset_failcnt(&memcg->kmem);
5324 static u64 mem_cgroup_move_charge_read(struct cgroup_subsys_state *css,
5327 return mem_cgroup_from_css(css)->move_charge_at_immigrate;
5331 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5332 struct cftype *cft, u64 val)
5334 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5336 if (val >= (1 << NR_MOVE_TYPE))
5340 * No kind of locking is needed in here, because ->can_attach() will
5341 * check this value once in the beginning of the process, and then carry
5342 * on with stale data. This means that changes to this value will only
5343 * affect task migrations starting after the change.
5345 memcg->move_charge_at_immigrate = val;
5349 static int mem_cgroup_move_charge_write(struct cgroup_subsys_state *css,
5350 struct cftype *cft, u64 val)
5357 static int memcg_numa_stat_show(struct cgroup_subsys_state *css,
5358 struct cftype *cft, struct seq_file *m)
5361 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
5362 unsigned long node_nr;
5363 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5365 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
5366 seq_printf(m, "total=%lu", total_nr);
5367 for_each_node_state(nid, N_MEMORY) {
5368 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
5369 seq_printf(m, " N%d=%lu", nid, node_nr);
5373 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
5374 seq_printf(m, "file=%lu", file_nr);
5375 for_each_node_state(nid, N_MEMORY) {
5376 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5378 seq_printf(m, " N%d=%lu", nid, node_nr);
5382 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
5383 seq_printf(m, "anon=%lu", anon_nr);
5384 for_each_node_state(nid, N_MEMORY) {
5385 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5387 seq_printf(m, " N%d=%lu", nid, node_nr);
5391 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
5392 seq_printf(m, "unevictable=%lu", unevictable_nr);
5393 for_each_node_state(nid, N_MEMORY) {
5394 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
5395 BIT(LRU_UNEVICTABLE));
5396 seq_printf(m, " N%d=%lu", nid, node_nr);
5401 #endif /* CONFIG_NUMA */
5403 static inline void mem_cgroup_lru_names_not_uptodate(void)
5405 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
5408 static int memcg_stat_show(struct cgroup_subsys_state *css, struct cftype *cft,
5411 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5412 struct mem_cgroup *mi;
5415 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5416 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5418 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
5419 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
5422 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
5423 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
5424 mem_cgroup_read_events(memcg, i));
5426 for (i = 0; i < NR_LRU_LISTS; i++)
5427 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
5428 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
5430 /* Hierarchical information */
5432 unsigned long long limit, memsw_limit;
5433 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
5434 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
5435 if (do_swap_account)
5436 seq_printf(m, "hierarchical_memsw_limit %llu\n",
5440 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
5443 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
5445 for_each_mem_cgroup_tree(mi, memcg)
5446 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
5447 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
5450 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
5451 unsigned long long val = 0;
5453 for_each_mem_cgroup_tree(mi, memcg)
5454 val += mem_cgroup_read_events(mi, i);
5455 seq_printf(m, "total_%s %llu\n",
5456 mem_cgroup_events_names[i], val);
5459 for (i = 0; i < NR_LRU_LISTS; i++) {
5460 unsigned long long val = 0;
5462 for_each_mem_cgroup_tree(mi, memcg)
5463 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
5464 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
5467 #ifdef CONFIG_DEBUG_VM
5470 struct mem_cgroup_per_zone *mz;
5471 struct zone_reclaim_stat *rstat;
5472 unsigned long recent_rotated[2] = {0, 0};
5473 unsigned long recent_scanned[2] = {0, 0};
5475 for_each_online_node(nid)
5476 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
5477 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
5478 rstat = &mz->lruvec.reclaim_stat;
5480 recent_rotated[0] += rstat->recent_rotated[0];
5481 recent_rotated[1] += rstat->recent_rotated[1];
5482 recent_scanned[0] += rstat->recent_scanned[0];
5483 recent_scanned[1] += rstat->recent_scanned[1];
5485 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
5486 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
5487 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
5488 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
5495 static u64 mem_cgroup_swappiness_read(struct cgroup_subsys_state *css,
5498 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5500 return mem_cgroup_swappiness(memcg);
5503 static int mem_cgroup_swappiness_write(struct cgroup_subsys_state *css,
5504 struct cftype *cft, u64 val)
5506 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5507 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5509 if (val > 100 || !parent)
5512 mutex_lock(&memcg_create_mutex);
5514 /* If under hierarchy, only empty-root can set this value */
5515 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5516 mutex_unlock(&memcg_create_mutex);
5520 memcg->swappiness = val;
5522 mutex_unlock(&memcg_create_mutex);
5527 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
5529 struct mem_cgroup_threshold_ary *t;
5535 t = rcu_dereference(memcg->thresholds.primary);
5537 t = rcu_dereference(memcg->memsw_thresholds.primary);
5542 usage = mem_cgroup_usage(memcg, swap);
5545 * current_threshold points to threshold just below or equal to usage.
5546 * If it's not true, a threshold was crossed after last
5547 * call of __mem_cgroup_threshold().
5549 i = t->current_threshold;
5552 * Iterate backward over array of thresholds starting from
5553 * current_threshold and check if a threshold is crossed.
5554 * If none of thresholds below usage is crossed, we read
5555 * only one element of the array here.
5557 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
5558 eventfd_signal(t->entries[i].eventfd, 1);
5560 /* i = current_threshold + 1 */
5564 * Iterate forward over array of thresholds starting from
5565 * current_threshold+1 and check if a threshold is crossed.
5566 * If none of thresholds above usage is crossed, we read
5567 * only one element of the array here.
5569 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
5570 eventfd_signal(t->entries[i].eventfd, 1);
5572 /* Update current_threshold */
5573 t->current_threshold = i - 1;
5578 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
5581 __mem_cgroup_threshold(memcg, false);
5582 if (do_swap_account)
5583 __mem_cgroup_threshold(memcg, true);
5585 memcg = parent_mem_cgroup(memcg);
5589 static int compare_thresholds(const void *a, const void *b)
5591 const struct mem_cgroup_threshold *_a = a;
5592 const struct mem_cgroup_threshold *_b = b;
5594 return _a->threshold - _b->threshold;
5597 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
5599 struct mem_cgroup_eventfd_list *ev;
5601 list_for_each_entry(ev, &memcg->oom_notify, list)
5602 eventfd_signal(ev->eventfd, 1);
5606 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
5608 struct mem_cgroup *iter;
5610 for_each_mem_cgroup_tree(iter, memcg)
5611 mem_cgroup_oom_notify_cb(iter);
5614 static int mem_cgroup_usage_register_event(struct cgroup_subsys_state *css,
5615 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5617 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5618 struct mem_cgroup_thresholds *thresholds;
5619 struct mem_cgroup_threshold_ary *new;
5620 enum res_type type = MEMFILE_TYPE(cft->private);
5621 u64 threshold, usage;
5624 ret = res_counter_memparse_write_strategy(args, &threshold);
5628 mutex_lock(&memcg->thresholds_lock);
5631 thresholds = &memcg->thresholds;
5632 else if (type == _MEMSWAP)
5633 thresholds = &memcg->memsw_thresholds;
5637 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5639 /* Check if a threshold crossed before adding a new one */
5640 if (thresholds->primary)
5641 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5643 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
5645 /* Allocate memory for new array of thresholds */
5646 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
5654 /* Copy thresholds (if any) to new array */
5655 if (thresholds->primary) {
5656 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
5657 sizeof(struct mem_cgroup_threshold));
5660 /* Add new threshold */
5661 new->entries[size - 1].eventfd = eventfd;
5662 new->entries[size - 1].threshold = threshold;
5664 /* Sort thresholds. Registering of new threshold isn't time-critical */
5665 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
5666 compare_thresholds, NULL);
5668 /* Find current threshold */
5669 new->current_threshold = -1;
5670 for (i = 0; i < size; i++) {
5671 if (new->entries[i].threshold <= usage) {
5673 * new->current_threshold will not be used until
5674 * rcu_assign_pointer(), so it's safe to increment
5677 ++new->current_threshold;
5682 /* Free old spare buffer and save old primary buffer as spare */
5683 kfree(thresholds->spare);
5684 thresholds->spare = thresholds->primary;
5686 rcu_assign_pointer(thresholds->primary, new);
5688 /* To be sure that nobody uses thresholds */
5692 mutex_unlock(&memcg->thresholds_lock);
5697 static void mem_cgroup_usage_unregister_event(struct cgroup_subsys_state *css,
5698 struct cftype *cft, struct eventfd_ctx *eventfd)
5700 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5701 struct mem_cgroup_thresholds *thresholds;
5702 struct mem_cgroup_threshold_ary *new;
5703 enum res_type type = MEMFILE_TYPE(cft->private);
5707 mutex_lock(&memcg->thresholds_lock);
5709 thresholds = &memcg->thresholds;
5710 else if (type == _MEMSWAP)
5711 thresholds = &memcg->memsw_thresholds;
5715 if (!thresholds->primary)
5718 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
5720 /* Check if a threshold crossed before removing */
5721 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
5723 /* Calculate new number of threshold */
5725 for (i = 0; i < thresholds->primary->size; i++) {
5726 if (thresholds->primary->entries[i].eventfd != eventfd)
5730 new = thresholds->spare;
5732 /* Set thresholds array to NULL if we don't have thresholds */
5741 /* Copy thresholds and find current threshold */
5742 new->current_threshold = -1;
5743 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
5744 if (thresholds->primary->entries[i].eventfd == eventfd)
5747 new->entries[j] = thresholds->primary->entries[i];
5748 if (new->entries[j].threshold <= usage) {
5750 * new->current_threshold will not be used
5751 * until rcu_assign_pointer(), so it's safe to increment
5754 ++new->current_threshold;
5760 /* Swap primary and spare array */
5761 thresholds->spare = thresholds->primary;
5762 /* If all events are unregistered, free the spare array */
5764 kfree(thresholds->spare);
5765 thresholds->spare = NULL;
5768 rcu_assign_pointer(thresholds->primary, new);
5770 /* To be sure that nobody uses thresholds */
5773 mutex_unlock(&memcg->thresholds_lock);
5776 static int mem_cgroup_oom_register_event(struct cgroup_subsys_state *css,
5777 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
5779 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5780 struct mem_cgroup_eventfd_list *event;
5781 enum res_type type = MEMFILE_TYPE(cft->private);
5783 BUG_ON(type != _OOM_TYPE);
5784 event = kmalloc(sizeof(*event), GFP_KERNEL);
5788 spin_lock(&memcg_oom_lock);
5790 event->eventfd = eventfd;
5791 list_add(&event->list, &memcg->oom_notify);
5793 /* already in OOM ? */
5794 if (atomic_read(&memcg->under_oom))
5795 eventfd_signal(eventfd, 1);
5796 spin_unlock(&memcg_oom_lock);
5801 static void mem_cgroup_oom_unregister_event(struct cgroup_subsys_state *css,
5802 struct cftype *cft, struct eventfd_ctx *eventfd)
5804 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5805 struct mem_cgroup_eventfd_list *ev, *tmp;
5806 enum res_type type = MEMFILE_TYPE(cft->private);
5808 BUG_ON(type != _OOM_TYPE);
5810 spin_lock(&memcg_oom_lock);
5812 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
5813 if (ev->eventfd == eventfd) {
5814 list_del(&ev->list);
5819 spin_unlock(&memcg_oom_lock);
5822 static int mem_cgroup_oom_control_read(struct cgroup_subsys_state *css,
5823 struct cftype *cft, struct cgroup_map_cb *cb)
5825 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5827 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
5829 if (atomic_read(&memcg->under_oom))
5830 cb->fill(cb, "under_oom", 1);
5832 cb->fill(cb, "under_oom", 0);
5836 static int mem_cgroup_oom_control_write(struct cgroup_subsys_state *css,
5837 struct cftype *cft, u64 val)
5839 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
5840 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(&memcg->css));
5842 /* cannot set to root cgroup and only 0 and 1 are allowed */
5843 if (!parent || !((val == 0) || (val == 1)))
5846 mutex_lock(&memcg_create_mutex);
5847 /* oom-kill-disable is a flag for subhierarchy. */
5848 if ((parent->use_hierarchy) || memcg_has_children(memcg)) {
5849 mutex_unlock(&memcg_create_mutex);
5852 memcg->oom_kill_disable = val;
5854 memcg_oom_recover(memcg);
5855 mutex_unlock(&memcg_create_mutex);
5859 #ifdef CONFIG_MEMCG_KMEM
5860 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5864 memcg->kmemcg_id = -1;
5865 ret = memcg_propagate_kmem(memcg);
5869 return mem_cgroup_sockets_init(memcg, ss);
5872 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5874 mem_cgroup_sockets_destroy(memcg);
5877 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5879 if (!memcg_kmem_is_active(memcg))
5883 * kmem charges can outlive the cgroup. In the case of slab
5884 * pages, for instance, a page contain objects from various
5885 * processes. As we prevent from taking a reference for every
5886 * such allocation we have to be careful when doing uncharge
5887 * (see memcg_uncharge_kmem) and here during offlining.
5889 * The idea is that that only the _last_ uncharge which sees
5890 * the dead memcg will drop the last reference. An additional
5891 * reference is taken here before the group is marked dead
5892 * which is then paired with css_put during uncharge resp. here.
5894 * Although this might sound strange as this path is called from
5895 * css_offline() when the referencemight have dropped down to 0
5896 * and shouldn't be incremented anymore (css_tryget would fail)
5897 * we do not have other options because of the kmem allocations
5900 css_get(&memcg->css);
5902 memcg_kmem_mark_dead(memcg);
5904 if (res_counter_read_u64(&memcg->kmem, RES_USAGE) != 0)
5907 if (memcg_kmem_test_and_clear_dead(memcg))
5908 css_put(&memcg->css);
5911 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
5916 static void memcg_destroy_kmem(struct mem_cgroup *memcg)
5920 static void kmem_cgroup_css_offline(struct mem_cgroup *memcg)
5925 static struct cftype mem_cgroup_files[] = {
5927 .name = "usage_in_bytes",
5928 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
5929 .read = mem_cgroup_read,
5930 .register_event = mem_cgroup_usage_register_event,
5931 .unregister_event = mem_cgroup_usage_unregister_event,
5934 .name = "max_usage_in_bytes",
5935 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
5936 .trigger = mem_cgroup_reset,
5937 .read = mem_cgroup_read,
5940 .name = "limit_in_bytes",
5941 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
5942 .write_string = mem_cgroup_write,
5943 .read = mem_cgroup_read,
5946 .name = "soft_limit_in_bytes",
5947 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
5948 .write_string = mem_cgroup_write,
5949 .read = mem_cgroup_read,
5953 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
5954 .trigger = mem_cgroup_reset,
5955 .read = mem_cgroup_read,
5959 .read_seq_string = memcg_stat_show,
5962 .name = "force_empty",
5963 .trigger = mem_cgroup_force_empty_write,
5966 .name = "use_hierarchy",
5967 .flags = CFTYPE_INSANE,
5968 .write_u64 = mem_cgroup_hierarchy_write,
5969 .read_u64 = mem_cgroup_hierarchy_read,
5972 .name = "swappiness",
5973 .read_u64 = mem_cgroup_swappiness_read,
5974 .write_u64 = mem_cgroup_swappiness_write,
5977 .name = "move_charge_at_immigrate",
5978 .read_u64 = mem_cgroup_move_charge_read,
5979 .write_u64 = mem_cgroup_move_charge_write,
5982 .name = "oom_control",
5983 .read_map = mem_cgroup_oom_control_read,
5984 .write_u64 = mem_cgroup_oom_control_write,
5985 .register_event = mem_cgroup_oom_register_event,
5986 .unregister_event = mem_cgroup_oom_unregister_event,
5987 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
5990 .name = "pressure_level",
5991 .register_event = vmpressure_register_event,
5992 .unregister_event = vmpressure_unregister_event,
5996 .name = "numa_stat",
5997 .read_seq_string = memcg_numa_stat_show,
6000 #ifdef CONFIG_MEMCG_KMEM
6002 .name = "kmem.limit_in_bytes",
6003 .private = MEMFILE_PRIVATE(_KMEM, RES_LIMIT),
6004 .write_string = mem_cgroup_write,
6005 .read = mem_cgroup_read,
6008 .name = "kmem.usage_in_bytes",
6009 .private = MEMFILE_PRIVATE(_KMEM, RES_USAGE),
6010 .read = mem_cgroup_read,
6013 .name = "kmem.failcnt",
6014 .private = MEMFILE_PRIVATE(_KMEM, RES_FAILCNT),
6015 .trigger = mem_cgroup_reset,
6016 .read = mem_cgroup_read,
6019 .name = "kmem.max_usage_in_bytes",
6020 .private = MEMFILE_PRIVATE(_KMEM, RES_MAX_USAGE),
6021 .trigger = mem_cgroup_reset,
6022 .read = mem_cgroup_read,
6024 #ifdef CONFIG_SLABINFO
6026 .name = "kmem.slabinfo",
6027 .read_seq_string = mem_cgroup_slabinfo_read,
6031 { }, /* terminate */
6034 #ifdef CONFIG_MEMCG_SWAP
6035 static struct cftype memsw_cgroup_files[] = {
6037 .name = "memsw.usage_in_bytes",
6038 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
6039 .read = mem_cgroup_read,
6040 .register_event = mem_cgroup_usage_register_event,
6041 .unregister_event = mem_cgroup_usage_unregister_event,
6044 .name = "memsw.max_usage_in_bytes",
6045 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
6046 .trigger = mem_cgroup_reset,
6047 .read = mem_cgroup_read,
6050 .name = "memsw.limit_in_bytes",
6051 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
6052 .write_string = mem_cgroup_write,
6053 .read = mem_cgroup_read,
6056 .name = "memsw.failcnt",
6057 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
6058 .trigger = mem_cgroup_reset,
6059 .read = mem_cgroup_read,
6061 { }, /* terminate */
6064 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6066 struct mem_cgroup_per_node *pn;
6067 struct mem_cgroup_per_zone *mz;
6068 int zone, tmp = node;
6070 * This routine is called against possible nodes.
6071 * But it's BUG to call kmalloc() against offline node.
6073 * TODO: this routine can waste much memory for nodes which will
6074 * never be onlined. It's better to use memory hotplug callback
6077 if (!node_state(node, N_NORMAL_MEMORY))
6079 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
6083 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6084 mz = &pn->zoneinfo[zone];
6085 lruvec_init(&mz->lruvec);
6086 mz->usage_in_excess = 0;
6087 mz->on_tree = false;
6090 memcg->nodeinfo[node] = pn;
6094 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
6096 kfree(memcg->nodeinfo[node]);
6099 static struct mem_cgroup *mem_cgroup_alloc(void)
6101 struct mem_cgroup *memcg;
6102 size_t size = memcg_size();
6104 /* Can be very big if nr_node_ids is very big */
6105 if (size < PAGE_SIZE)
6106 memcg = kzalloc(size, GFP_KERNEL);
6108 memcg = vzalloc(size);
6113 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
6116 spin_lock_init(&memcg->pcp_counter_lock);
6120 if (size < PAGE_SIZE)
6128 * At destroying mem_cgroup, references from swap_cgroup can remain.
6129 * (scanning all at force_empty is too costly...)
6131 * Instead of clearing all references at force_empty, we remember
6132 * the number of reference from swap_cgroup and free mem_cgroup when
6133 * it goes down to 0.
6135 * Removal of cgroup itself succeeds regardless of refs from swap.
6138 static void __mem_cgroup_free(struct mem_cgroup *memcg)
6141 size_t size = memcg_size();
6143 mem_cgroup_remove_from_trees(memcg);
6144 free_css_id(&mem_cgroup_subsys, &memcg->css);
6147 free_mem_cgroup_per_zone_info(memcg, node);
6149 free_percpu(memcg->stat);
6152 * We need to make sure that (at least for now), the jump label
6153 * destruction code runs outside of the cgroup lock. This is because
6154 * get_online_cpus(), which is called from the static_branch update,
6155 * can't be called inside the cgroup_lock. cpusets are the ones
6156 * enforcing this dependency, so if they ever change, we might as well.
6158 * schedule_work() will guarantee this happens. Be careful if you need
6159 * to move this code around, and make sure it is outside
6162 disarm_static_keys(memcg);
6163 if (size < PAGE_SIZE)
6170 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
6172 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
6174 if (!memcg->res.parent)
6176 return mem_cgroup_from_res_counter(memcg->res.parent, res);
6178 EXPORT_SYMBOL(parent_mem_cgroup);
6180 static void __init mem_cgroup_soft_limit_tree_init(void)
6182 struct mem_cgroup_tree_per_node *rtpn;
6183 struct mem_cgroup_tree_per_zone *rtpz;
6184 int tmp, node, zone;
6186 for_each_node(node) {
6188 if (!node_state(node, N_NORMAL_MEMORY))
6190 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
6193 soft_limit_tree.rb_tree_per_node[node] = rtpn;
6195 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
6196 rtpz = &rtpn->rb_tree_per_zone[zone];
6197 rtpz->rb_root = RB_ROOT;
6198 spin_lock_init(&rtpz->lock);
6203 static struct cgroup_subsys_state * __ref
6204 mem_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6206 struct mem_cgroup *memcg;
6207 long error = -ENOMEM;
6210 memcg = mem_cgroup_alloc();
6212 return ERR_PTR(error);
6215 if (alloc_mem_cgroup_per_zone_info(memcg, node))
6219 if (parent_css == NULL) {
6220 root_mem_cgroup = memcg;
6221 res_counter_init(&memcg->res, NULL);
6222 res_counter_init(&memcg->memsw, NULL);
6223 res_counter_init(&memcg->kmem, NULL);
6226 memcg->last_scanned_node = MAX_NUMNODES;
6227 INIT_LIST_HEAD(&memcg->oom_notify);
6228 memcg->move_charge_at_immigrate = 0;
6229 mutex_init(&memcg->thresholds_lock);
6230 spin_lock_init(&memcg->move_lock);
6231 vmpressure_init(&memcg->vmpressure);
6236 __mem_cgroup_free(memcg);
6237 return ERR_PTR(error);
6241 mem_cgroup_css_online(struct cgroup_subsys_state *css)
6243 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6244 struct mem_cgroup *parent = mem_cgroup_from_css(css_parent(css));
6250 mutex_lock(&memcg_create_mutex);
6252 memcg->use_hierarchy = parent->use_hierarchy;
6253 memcg->oom_kill_disable = parent->oom_kill_disable;
6254 memcg->swappiness = mem_cgroup_swappiness(parent);
6256 if (parent->use_hierarchy) {
6257 res_counter_init(&memcg->res, &parent->res);
6258 res_counter_init(&memcg->memsw, &parent->memsw);
6259 res_counter_init(&memcg->kmem, &parent->kmem);
6262 * No need to take a reference to the parent because cgroup
6263 * core guarantees its existence.
6266 res_counter_init(&memcg->res, NULL);
6267 res_counter_init(&memcg->memsw, NULL);
6268 res_counter_init(&memcg->kmem, NULL);
6270 * Deeper hierachy with use_hierarchy == false doesn't make
6271 * much sense so let cgroup subsystem know about this
6272 * unfortunate state in our controller.
6274 if (parent != root_mem_cgroup)
6275 mem_cgroup_subsys.broken_hierarchy = true;
6278 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
6279 mutex_unlock(&memcg_create_mutex);
6284 * Announce all parents that a group from their hierarchy is gone.
6286 static void mem_cgroup_invalidate_reclaim_iterators(struct mem_cgroup *memcg)
6288 struct mem_cgroup *parent = memcg;
6290 while ((parent = parent_mem_cgroup(parent)))
6291 mem_cgroup_iter_invalidate(parent);
6294 * if the root memcg is not hierarchical we have to check it
6297 if (!root_mem_cgroup->use_hierarchy)
6298 mem_cgroup_iter_invalidate(root_mem_cgroup);
6301 static void mem_cgroup_css_offline(struct cgroup_subsys_state *css)
6303 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6305 kmem_cgroup_css_offline(memcg);
6307 mem_cgroup_invalidate_reclaim_iterators(memcg);
6308 mem_cgroup_reparent_charges(memcg);
6309 mem_cgroup_destroy_all_caches(memcg);
6310 vmpressure_cleanup(&memcg->vmpressure);
6313 static void mem_cgroup_css_free(struct cgroup_subsys_state *css)
6315 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6317 memcg_destroy_kmem(memcg);
6318 __mem_cgroup_free(memcg);
6322 /* Handlers for move charge at task migration. */
6323 #define PRECHARGE_COUNT_AT_ONCE 256
6324 static int mem_cgroup_do_precharge(unsigned long count)
6327 int batch_count = PRECHARGE_COUNT_AT_ONCE;
6328 struct mem_cgroup *memcg = mc.to;
6330 if (mem_cgroup_is_root(memcg)) {
6331 mc.precharge += count;
6332 /* we don't need css_get for root */
6335 /* try to charge at once */
6337 struct res_counter *dummy;
6339 * "memcg" cannot be under rmdir() because we've already checked
6340 * by cgroup_lock_live_cgroup() that it is not removed and we
6341 * are still under the same cgroup_mutex. So we can postpone
6344 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
6346 if (do_swap_account && res_counter_charge(&memcg->memsw,
6347 PAGE_SIZE * count, &dummy)) {
6348 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
6351 mc.precharge += count;
6355 /* fall back to one by one charge */
6357 if (signal_pending(current)) {
6361 if (!batch_count--) {
6362 batch_count = PRECHARGE_COUNT_AT_ONCE;
6365 ret = __mem_cgroup_try_charge(NULL,
6366 GFP_KERNEL, 1, &memcg, false);
6368 /* mem_cgroup_clear_mc() will do uncharge later */
6376 * get_mctgt_type - get target type of moving charge
6377 * @vma: the vma the pte to be checked belongs
6378 * @addr: the address corresponding to the pte to be checked
6379 * @ptent: the pte to be checked
6380 * @target: the pointer the target page or swap ent will be stored(can be NULL)
6383 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
6384 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
6385 * move charge. if @target is not NULL, the page is stored in target->page
6386 * with extra refcnt got(Callers should handle it).
6387 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
6388 * target for charge migration. if @target is not NULL, the entry is stored
6391 * Called with pte lock held.
6398 enum mc_target_type {
6404 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
6405 unsigned long addr, pte_t ptent)
6407 struct page *page = vm_normal_page(vma, addr, ptent);
6409 if (!page || !page_mapped(page))
6411 if (PageAnon(page)) {
6412 /* we don't move shared anon */
6415 } else if (!move_file())
6416 /* we ignore mapcount for file pages */
6418 if (!get_page_unless_zero(page))
6425 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6426 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6428 struct page *page = NULL;
6429 swp_entry_t ent = pte_to_swp_entry(ptent);
6431 if (!move_anon() || non_swap_entry(ent))
6434 * Because lookup_swap_cache() updates some statistics counter,
6435 * we call find_get_page() with swapper_space directly.
6437 page = find_get_page(swap_address_space(ent), ent.val);
6438 if (do_swap_account)
6439 entry->val = ent.val;
6444 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
6445 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6451 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
6452 unsigned long addr, pte_t ptent, swp_entry_t *entry)
6454 struct page *page = NULL;
6455 struct address_space *mapping;
6458 if (!vma->vm_file) /* anonymous vma */
6463 mapping = vma->vm_file->f_mapping;
6464 if (pte_none(ptent))
6465 pgoff = linear_page_index(vma, addr);
6466 else /* pte_file(ptent) is true */
6467 pgoff = pte_to_pgoff(ptent);
6469 /* page is moved even if it's not RSS of this task(page-faulted). */
6470 page = find_get_page(mapping, pgoff);
6473 /* shmem/tmpfs may report page out on swap: account for that too. */
6474 if (radix_tree_exceptional_entry(page)) {
6475 swp_entry_t swap = radix_to_swp_entry(page);
6476 if (do_swap_account)
6478 page = find_get_page(swap_address_space(swap), swap.val);
6484 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
6485 unsigned long addr, pte_t ptent, union mc_target *target)
6487 struct page *page = NULL;
6488 struct page_cgroup *pc;
6489 enum mc_target_type ret = MC_TARGET_NONE;
6490 swp_entry_t ent = { .val = 0 };
6492 if (pte_present(ptent))
6493 page = mc_handle_present_pte(vma, addr, ptent);
6494 else if (is_swap_pte(ptent))
6495 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
6496 else if (pte_none(ptent) || pte_file(ptent))
6497 page = mc_handle_file_pte(vma, addr, ptent, &ent);
6499 if (!page && !ent.val)
6502 pc = lookup_page_cgroup(page);
6504 * Do only loose check w/o page_cgroup lock.
6505 * mem_cgroup_move_account() checks the pc is valid or not under
6508 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6509 ret = MC_TARGET_PAGE;
6511 target->page = page;
6513 if (!ret || !target)
6516 /* There is a swap entry and a page doesn't exist or isn't charged */
6517 if (ent.val && !ret &&
6518 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
6519 ret = MC_TARGET_SWAP;
6526 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
6528 * We don't consider swapping or file mapped pages because THP does not
6529 * support them for now.
6530 * Caller should make sure that pmd_trans_huge(pmd) is true.
6532 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6533 unsigned long addr, pmd_t pmd, union mc_target *target)
6535 struct page *page = NULL;
6536 struct page_cgroup *pc;
6537 enum mc_target_type ret = MC_TARGET_NONE;
6539 page = pmd_page(pmd);
6540 VM_BUG_ON(!page || !PageHead(page));
6543 pc = lookup_page_cgroup(page);
6544 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
6545 ret = MC_TARGET_PAGE;
6548 target->page = page;
6554 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
6555 unsigned long addr, pmd_t pmd, union mc_target *target)
6557 return MC_TARGET_NONE;
6561 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
6562 unsigned long addr, unsigned long end,
6563 struct mm_walk *walk)
6565 struct vm_area_struct *vma = walk->private;
6569 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6570 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
6571 mc.precharge += HPAGE_PMD_NR;
6572 spin_unlock(&vma->vm_mm->page_table_lock);
6576 if (pmd_trans_unstable(pmd))
6578 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6579 for (; addr != end; pte++, addr += PAGE_SIZE)
6580 if (get_mctgt_type(vma, addr, *pte, NULL))
6581 mc.precharge++; /* increment precharge temporarily */
6582 pte_unmap_unlock(pte - 1, ptl);
6588 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
6590 unsigned long precharge;
6591 struct vm_area_struct *vma;
6593 down_read(&mm->mmap_sem);
6594 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6595 struct mm_walk mem_cgroup_count_precharge_walk = {
6596 .pmd_entry = mem_cgroup_count_precharge_pte_range,
6600 if (is_vm_hugetlb_page(vma))
6602 walk_page_range(vma->vm_start, vma->vm_end,
6603 &mem_cgroup_count_precharge_walk);
6605 up_read(&mm->mmap_sem);
6607 precharge = mc.precharge;
6613 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
6615 unsigned long precharge = mem_cgroup_count_precharge(mm);
6617 VM_BUG_ON(mc.moving_task);
6618 mc.moving_task = current;
6619 return mem_cgroup_do_precharge(precharge);
6622 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
6623 static void __mem_cgroup_clear_mc(void)
6625 struct mem_cgroup *from = mc.from;
6626 struct mem_cgroup *to = mc.to;
6629 /* we must uncharge all the leftover precharges from mc.to */
6631 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
6635 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
6636 * we must uncharge here.
6638 if (mc.moved_charge) {
6639 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
6640 mc.moved_charge = 0;
6642 /* we must fixup refcnts and charges */
6643 if (mc.moved_swap) {
6644 /* uncharge swap account from the old cgroup */
6645 if (!mem_cgroup_is_root(mc.from))
6646 res_counter_uncharge(&mc.from->memsw,
6647 PAGE_SIZE * mc.moved_swap);
6649 for (i = 0; i < mc.moved_swap; i++)
6650 css_put(&mc.from->css);
6652 if (!mem_cgroup_is_root(mc.to)) {
6654 * we charged both to->res and to->memsw, so we should
6657 res_counter_uncharge(&mc.to->res,
6658 PAGE_SIZE * mc.moved_swap);
6660 /* we've already done css_get(mc.to) */
6663 memcg_oom_recover(from);
6664 memcg_oom_recover(to);
6665 wake_up_all(&mc.waitq);
6668 static void mem_cgroup_clear_mc(void)
6670 struct mem_cgroup *from = mc.from;
6673 * we must clear moving_task before waking up waiters at the end of
6676 mc.moving_task = NULL;
6677 __mem_cgroup_clear_mc();
6678 spin_lock(&mc.lock);
6681 spin_unlock(&mc.lock);
6682 mem_cgroup_end_move(from);
6685 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6686 struct cgroup_taskset *tset)
6688 struct task_struct *p = cgroup_taskset_first(tset);
6690 struct mem_cgroup *memcg = mem_cgroup_from_css(css);
6691 unsigned long move_charge_at_immigrate;
6694 * We are now commited to this value whatever it is. Changes in this
6695 * tunable will only affect upcoming migrations, not the current one.
6696 * So we need to save it, and keep it going.
6698 move_charge_at_immigrate = memcg->move_charge_at_immigrate;
6699 if (move_charge_at_immigrate) {
6700 struct mm_struct *mm;
6701 struct mem_cgroup *from = mem_cgroup_from_task(p);
6703 VM_BUG_ON(from == memcg);
6705 mm = get_task_mm(p);
6708 /* We move charges only when we move a owner of the mm */
6709 if (mm->owner == p) {
6712 VM_BUG_ON(mc.precharge);
6713 VM_BUG_ON(mc.moved_charge);
6714 VM_BUG_ON(mc.moved_swap);
6715 mem_cgroup_start_move(from);
6716 spin_lock(&mc.lock);
6719 mc.immigrate_flags = move_charge_at_immigrate;
6720 spin_unlock(&mc.lock);
6721 /* We set mc.moving_task later */
6723 ret = mem_cgroup_precharge_mc(mm);
6725 mem_cgroup_clear_mc();
6732 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6733 struct cgroup_taskset *tset)
6735 mem_cgroup_clear_mc();
6738 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
6739 unsigned long addr, unsigned long end,
6740 struct mm_walk *walk)
6743 struct vm_area_struct *vma = walk->private;
6746 enum mc_target_type target_type;
6747 union mc_target target;
6749 struct page_cgroup *pc;
6752 * We don't take compound_lock() here but no race with splitting thp
6754 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
6755 * under splitting, which means there's no concurrent thp split,
6756 * - if another thread runs into split_huge_page() just after we
6757 * entered this if-block, the thread must wait for page table lock
6758 * to be unlocked in __split_huge_page_splitting(), where the main
6759 * part of thp split is not executed yet.
6761 if (pmd_trans_huge_lock(pmd, vma) == 1) {
6762 if (mc.precharge < HPAGE_PMD_NR) {
6763 spin_unlock(&vma->vm_mm->page_table_lock);
6766 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
6767 if (target_type == MC_TARGET_PAGE) {
6769 if (!isolate_lru_page(page)) {
6770 pc = lookup_page_cgroup(page);
6771 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
6772 pc, mc.from, mc.to)) {
6773 mc.precharge -= HPAGE_PMD_NR;
6774 mc.moved_charge += HPAGE_PMD_NR;
6776 putback_lru_page(page);
6780 spin_unlock(&vma->vm_mm->page_table_lock);
6784 if (pmd_trans_unstable(pmd))
6787 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
6788 for (; addr != end; addr += PAGE_SIZE) {
6789 pte_t ptent = *(pte++);
6795 switch (get_mctgt_type(vma, addr, ptent, &target)) {
6796 case MC_TARGET_PAGE:
6798 if (isolate_lru_page(page))
6800 pc = lookup_page_cgroup(page);
6801 if (!mem_cgroup_move_account(page, 1, pc,
6804 /* we uncharge from mc.from later. */
6807 putback_lru_page(page);
6808 put: /* get_mctgt_type() gets the page */
6811 case MC_TARGET_SWAP:
6813 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
6815 /* we fixup refcnts and charges later. */
6823 pte_unmap_unlock(pte - 1, ptl);
6828 * We have consumed all precharges we got in can_attach().
6829 * We try charge one by one, but don't do any additional
6830 * charges to mc.to if we have failed in charge once in attach()
6833 ret = mem_cgroup_do_precharge(1);
6841 static void mem_cgroup_move_charge(struct mm_struct *mm)
6843 struct vm_area_struct *vma;
6845 lru_add_drain_all();
6847 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
6849 * Someone who are holding the mmap_sem might be waiting in
6850 * waitq. So we cancel all extra charges, wake up all waiters,
6851 * and retry. Because we cancel precharges, we might not be able
6852 * to move enough charges, but moving charge is a best-effort
6853 * feature anyway, so it wouldn't be a big problem.
6855 __mem_cgroup_clear_mc();
6859 for (vma = mm->mmap; vma; vma = vma->vm_next) {
6861 struct mm_walk mem_cgroup_move_charge_walk = {
6862 .pmd_entry = mem_cgroup_move_charge_pte_range,
6866 if (is_vm_hugetlb_page(vma))
6868 ret = walk_page_range(vma->vm_start, vma->vm_end,
6869 &mem_cgroup_move_charge_walk);
6872 * means we have consumed all precharges and failed in
6873 * doing additional charge. Just abandon here.
6877 up_read(&mm->mmap_sem);
6880 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6881 struct cgroup_taskset *tset)
6883 struct task_struct *p = cgroup_taskset_first(tset);
6884 struct mm_struct *mm = get_task_mm(p);
6888 mem_cgroup_move_charge(mm);
6892 mem_cgroup_clear_mc();
6894 #else /* !CONFIG_MMU */
6895 static int mem_cgroup_can_attach(struct cgroup_subsys_state *css,
6896 struct cgroup_taskset *tset)
6900 static void mem_cgroup_cancel_attach(struct cgroup_subsys_state *css,
6901 struct cgroup_taskset *tset)
6904 static void mem_cgroup_move_task(struct cgroup_subsys_state *css,
6905 struct cgroup_taskset *tset)
6911 * Cgroup retains root cgroups across [un]mount cycles making it necessary
6912 * to verify sane_behavior flag on each mount attempt.
6914 static void mem_cgroup_bind(struct cgroup_subsys_state *root_css)
6917 * use_hierarchy is forced with sane_behavior. cgroup core
6918 * guarantees that @root doesn't have any children, so turning it
6919 * on for the root memcg is enough.
6921 if (cgroup_sane_behavior(root_css->cgroup))
6922 mem_cgroup_from_css(root_css)->use_hierarchy = true;
6925 struct cgroup_subsys mem_cgroup_subsys = {
6927 .subsys_id = mem_cgroup_subsys_id,
6928 .css_alloc = mem_cgroup_css_alloc,
6929 .css_online = mem_cgroup_css_online,
6930 .css_offline = mem_cgroup_css_offline,
6931 .css_free = mem_cgroup_css_free,
6932 .can_attach = mem_cgroup_can_attach,
6933 .cancel_attach = mem_cgroup_cancel_attach,
6934 .attach = mem_cgroup_move_task,
6935 .bind = mem_cgroup_bind,
6936 .base_cftypes = mem_cgroup_files,
6941 #ifdef CONFIG_MEMCG_SWAP
6942 static int __init enable_swap_account(char *s)
6944 if (!strcmp(s, "1"))
6945 really_do_swap_account = 1;
6946 else if (!strcmp(s, "0"))
6947 really_do_swap_account = 0;
6950 __setup("swapaccount=", enable_swap_account);
6952 static void __init memsw_file_init(void)
6954 WARN_ON(cgroup_add_cftypes(&mem_cgroup_subsys, memsw_cgroup_files));
6957 static void __init enable_swap_cgroup(void)
6959 if (!mem_cgroup_disabled() && really_do_swap_account) {
6960 do_swap_account = 1;
6966 static void __init enable_swap_cgroup(void)
6972 * subsys_initcall() for memory controller.
6974 * Some parts like hotcpu_notifier() have to be initialized from this context
6975 * because of lock dependencies (cgroup_lock -> cpu hotplug) but basically
6976 * everything that doesn't depend on a specific mem_cgroup structure should
6977 * be initialized from here.
6979 static int __init mem_cgroup_init(void)
6981 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
6982 enable_swap_cgroup();
6983 mem_cgroup_soft_limit_tree_init();
6987 subsys_initcall(mem_cgroup_init);