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 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/export.h>
37 #include <linux/mutex.h>
38 #include <linux/rbtree.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
55 #include <net/tcp_memcontrol.h>
57 #include <asm/uaccess.h>
59 #include <trace/events/vmscan.h>
61 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
62 EXPORT_SYMBOL(mem_cgroup_subsys);
64 #define MEM_CGROUP_RECLAIM_RETRIES 5
65 static struct mem_cgroup *root_mem_cgroup __read_mostly;
67 #ifdef CONFIG_MEMCG_SWAP
68 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
69 int do_swap_account __read_mostly;
71 /* for remember boot option*/
72 #ifdef CONFIG_MEMCG_SWAP_ENABLED
73 static int really_do_swap_account __initdata = 1;
75 static int really_do_swap_account __initdata = 0;
79 #define do_swap_account 0
84 * Statistics for memory cgroup.
86 enum mem_cgroup_stat_index {
88 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
90 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
91 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
92 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
93 MEM_CGROUP_STAT_SWAP, /* # of pages, swapped out */
94 MEM_CGROUP_STAT_NSTATS,
97 static const char * const mem_cgroup_stat_names[] = {
104 enum mem_cgroup_events_index {
105 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
106 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
107 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
108 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
109 MEM_CGROUP_EVENTS_NSTATS,
112 static const char * const mem_cgroup_events_names[] = {
120 * Per memcg event counter is incremented at every pagein/pageout. With THP,
121 * it will be incremated by the number of pages. This counter is used for
122 * for trigger some periodic events. This is straightforward and better
123 * than using jiffies etc. to handle periodic memcg event.
125 enum mem_cgroup_events_target {
126 MEM_CGROUP_TARGET_THRESH,
127 MEM_CGROUP_TARGET_SOFTLIMIT,
128 MEM_CGROUP_TARGET_NUMAINFO,
131 #define THRESHOLDS_EVENTS_TARGET 128
132 #define SOFTLIMIT_EVENTS_TARGET 1024
133 #define NUMAINFO_EVENTS_TARGET 1024
135 struct mem_cgroup_stat_cpu {
136 long count[MEM_CGROUP_STAT_NSTATS];
137 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
138 unsigned long nr_page_events;
139 unsigned long targets[MEM_CGROUP_NTARGETS];
142 struct mem_cgroup_reclaim_iter {
143 /* css_id of the last scanned hierarchy member */
145 /* scan generation, increased every round-trip */
146 unsigned int generation;
150 * per-zone information in memory controller.
152 struct mem_cgroup_per_zone {
153 struct lruvec lruvec;
154 unsigned long lru_size[NR_LRU_LISTS];
156 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
158 struct rb_node tree_node; /* RB tree node */
159 unsigned long long usage_in_excess;/* Set to the value by which */
160 /* the soft limit is exceeded*/
162 struct mem_cgroup *memcg; /* Back pointer, we cannot */
163 /* use container_of */
166 struct mem_cgroup_per_node {
167 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
170 struct mem_cgroup_lru_info {
171 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
175 * Cgroups above their limits are maintained in a RB-Tree, independent of
176 * their hierarchy representation
179 struct mem_cgroup_tree_per_zone {
180 struct rb_root rb_root;
184 struct mem_cgroup_tree_per_node {
185 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
188 struct mem_cgroup_tree {
189 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
192 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
194 struct mem_cgroup_threshold {
195 struct eventfd_ctx *eventfd;
200 struct mem_cgroup_threshold_ary {
201 /* An array index points to threshold just below or equal to usage. */
202 int current_threshold;
203 /* Size of entries[] */
205 /* Array of thresholds */
206 struct mem_cgroup_threshold entries[0];
209 struct mem_cgroup_thresholds {
210 /* Primary thresholds array */
211 struct mem_cgroup_threshold_ary *primary;
213 * Spare threshold array.
214 * This is needed to make mem_cgroup_unregister_event() "never fail".
215 * It must be able to store at least primary->size - 1 entries.
217 struct mem_cgroup_threshold_ary *spare;
221 struct mem_cgroup_eventfd_list {
222 struct list_head list;
223 struct eventfd_ctx *eventfd;
226 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
227 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
230 * The memory controller data structure. The memory controller controls both
231 * page cache and RSS per cgroup. We would eventually like to provide
232 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
233 * to help the administrator determine what knobs to tune.
235 * TODO: Add a water mark for the memory controller. Reclaim will begin when
236 * we hit the water mark. May be even add a low water mark, such that
237 * no reclaim occurs from a cgroup at it's low water mark, this is
238 * a feature that will be implemented much later in the future.
241 struct cgroup_subsys_state css;
243 * the counter to account for memory usage
245 struct res_counter res;
249 * the counter to account for mem+swap usage.
251 struct res_counter memsw;
254 * rcu_freeing is used only when freeing struct mem_cgroup,
255 * so put it into a union to avoid wasting more memory.
256 * It must be disjoint from the css field. It could be
257 * in a union with the res field, but res plays a much
258 * larger part in mem_cgroup life than memsw, and might
259 * be of interest, even at time of free, when debugging.
260 * So share rcu_head with the less interesting memsw.
262 struct rcu_head rcu_freeing;
264 * We also need some space for a worker in deferred freeing.
265 * By the time we call it, rcu_freeing is no longer in use.
267 struct work_struct work_freeing;
271 * Per cgroup active and inactive list, similar to the
272 * per zone LRU lists.
274 struct mem_cgroup_lru_info info;
275 int last_scanned_node;
277 nodemask_t scan_nodes;
278 atomic_t numainfo_events;
279 atomic_t numainfo_updating;
282 * Should the accounting and control be hierarchical, per subtree?
292 /* OOM-Killer disable */
293 int oom_kill_disable;
295 /* set when res.limit == memsw.limit */
296 bool memsw_is_minimum;
298 /* protect arrays of thresholds */
299 struct mutex thresholds_lock;
301 /* thresholds for memory usage. RCU-protected */
302 struct mem_cgroup_thresholds thresholds;
304 /* thresholds for mem+swap usage. RCU-protected */
305 struct mem_cgroup_thresholds memsw_thresholds;
307 /* For oom notifier event fd */
308 struct list_head oom_notify;
311 * Should we move charges of a task when a task is moved into this
312 * mem_cgroup ? And what type of charges should we move ?
314 unsigned long move_charge_at_immigrate;
316 * set > 0 if pages under this cgroup are moving to other cgroup.
318 atomic_t moving_account;
319 /* taken only while moving_account > 0 */
320 spinlock_t move_lock;
324 struct mem_cgroup_stat_cpu __percpu *stat;
326 * used when a cpu is offlined or other synchronizations
327 * See mem_cgroup_read_stat().
329 struct mem_cgroup_stat_cpu nocpu_base;
330 spinlock_t pcp_counter_lock;
332 #if defined(CONFIG_MEMCG_KMEM) && defined(CONFIG_INET)
333 struct tcp_memcontrol tcp_mem;
337 /* Stuffs for move charges at task migration. */
339 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
340 * left-shifted bitmap of these types.
343 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
344 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
348 /* "mc" and its members are protected by cgroup_mutex */
349 static struct move_charge_struct {
350 spinlock_t lock; /* for from, to */
351 struct mem_cgroup *from;
352 struct mem_cgroup *to;
353 unsigned long precharge;
354 unsigned long moved_charge;
355 unsigned long moved_swap;
356 struct task_struct *moving_task; /* a task moving charges */
357 wait_queue_head_t waitq; /* a waitq for other context */
359 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
360 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
363 static bool move_anon(void)
365 return test_bit(MOVE_CHARGE_TYPE_ANON,
366 &mc.to->move_charge_at_immigrate);
369 static bool move_file(void)
371 return test_bit(MOVE_CHARGE_TYPE_FILE,
372 &mc.to->move_charge_at_immigrate);
376 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
377 * limit reclaim to prevent infinite loops, if they ever occur.
379 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
380 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
383 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
384 MEM_CGROUP_CHARGE_TYPE_ANON,
385 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
386 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
390 /* for encoding cft->private value on file */
397 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
398 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
399 #define MEMFILE_ATTR(val) ((val) & 0xffff)
400 /* Used for OOM nofiier */
401 #define OOM_CONTROL (0)
404 * Reclaim flags for mem_cgroup_hierarchical_reclaim
406 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
407 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
408 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
409 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
411 static void mem_cgroup_get(struct mem_cgroup *memcg);
412 static void mem_cgroup_put(struct mem_cgroup *memcg);
415 struct mem_cgroup *mem_cgroup_from_css(struct cgroup_subsys_state *s)
417 return container_of(s, struct mem_cgroup, css);
420 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
422 return (memcg == root_mem_cgroup);
425 /* Writing them here to avoid exposing memcg's inner layout */
426 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
428 void sock_update_memcg(struct sock *sk)
430 if (mem_cgroup_sockets_enabled) {
431 struct mem_cgroup *memcg;
432 struct cg_proto *cg_proto;
434 BUG_ON(!sk->sk_prot->proto_cgroup);
436 /* Socket cloning can throw us here with sk_cgrp already
437 * filled. It won't however, necessarily happen from
438 * process context. So the test for root memcg given
439 * the current task's memcg won't help us in this case.
441 * Respecting the original socket's memcg is a better
442 * decision in this case.
445 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
446 mem_cgroup_get(sk->sk_cgrp->memcg);
451 memcg = mem_cgroup_from_task(current);
452 cg_proto = sk->sk_prot->proto_cgroup(memcg);
453 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
454 mem_cgroup_get(memcg);
455 sk->sk_cgrp = cg_proto;
460 EXPORT_SYMBOL(sock_update_memcg);
462 void sock_release_memcg(struct sock *sk)
464 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
465 struct mem_cgroup *memcg;
466 WARN_ON(!sk->sk_cgrp->memcg);
467 memcg = sk->sk_cgrp->memcg;
468 mem_cgroup_put(memcg);
472 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
474 if (!memcg || mem_cgroup_is_root(memcg))
477 return &memcg->tcp_mem.cg_proto;
479 EXPORT_SYMBOL(tcp_proto_cgroup);
481 static void disarm_sock_keys(struct mem_cgroup *memcg)
483 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
485 static_key_slow_dec(&memcg_socket_limit_enabled);
488 static void disarm_sock_keys(struct mem_cgroup *memcg)
493 static void drain_all_stock_async(struct mem_cgroup *memcg);
495 static struct mem_cgroup_per_zone *
496 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
498 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
501 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
506 static struct mem_cgroup_per_zone *
507 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
509 int nid = page_to_nid(page);
510 int zid = page_zonenum(page);
512 return mem_cgroup_zoneinfo(memcg, nid, zid);
515 static struct mem_cgroup_tree_per_zone *
516 soft_limit_tree_node_zone(int nid, int zid)
518 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
521 static struct mem_cgroup_tree_per_zone *
522 soft_limit_tree_from_page(struct page *page)
524 int nid = page_to_nid(page);
525 int zid = page_zonenum(page);
527 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
531 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
532 struct mem_cgroup_per_zone *mz,
533 struct mem_cgroup_tree_per_zone *mctz,
534 unsigned long long new_usage_in_excess)
536 struct rb_node **p = &mctz->rb_root.rb_node;
537 struct rb_node *parent = NULL;
538 struct mem_cgroup_per_zone *mz_node;
543 mz->usage_in_excess = new_usage_in_excess;
544 if (!mz->usage_in_excess)
548 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
550 if (mz->usage_in_excess < mz_node->usage_in_excess)
553 * We can't avoid mem cgroups that are over their soft
554 * limit by the same amount
556 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
559 rb_link_node(&mz->tree_node, parent, p);
560 rb_insert_color(&mz->tree_node, &mctz->rb_root);
565 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
566 struct mem_cgroup_per_zone *mz,
567 struct mem_cgroup_tree_per_zone *mctz)
571 rb_erase(&mz->tree_node, &mctz->rb_root);
576 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
577 struct mem_cgroup_per_zone *mz,
578 struct mem_cgroup_tree_per_zone *mctz)
580 spin_lock(&mctz->lock);
581 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
582 spin_unlock(&mctz->lock);
586 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
588 unsigned long long excess;
589 struct mem_cgroup_per_zone *mz;
590 struct mem_cgroup_tree_per_zone *mctz;
591 int nid = page_to_nid(page);
592 int zid = page_zonenum(page);
593 mctz = soft_limit_tree_from_page(page);
596 * Necessary to update all ancestors when hierarchy is used.
597 * because their event counter is not touched.
599 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
600 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
601 excess = res_counter_soft_limit_excess(&memcg->res);
603 * We have to update the tree if mz is on RB-tree or
604 * mem is over its softlimit.
606 if (excess || mz->on_tree) {
607 spin_lock(&mctz->lock);
608 /* if on-tree, remove it */
610 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
612 * Insert again. mz->usage_in_excess will be updated.
613 * If excess is 0, no tree ops.
615 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
616 spin_unlock(&mctz->lock);
621 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
624 struct mem_cgroup_per_zone *mz;
625 struct mem_cgroup_tree_per_zone *mctz;
627 for_each_node(node) {
628 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
629 mz = mem_cgroup_zoneinfo(memcg, node, zone);
630 mctz = soft_limit_tree_node_zone(node, zone);
631 mem_cgroup_remove_exceeded(memcg, mz, mctz);
636 static struct mem_cgroup_per_zone *
637 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
639 struct rb_node *rightmost = NULL;
640 struct mem_cgroup_per_zone *mz;
644 rightmost = rb_last(&mctz->rb_root);
646 goto done; /* Nothing to reclaim from */
648 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
650 * Remove the node now but someone else can add it back,
651 * we will to add it back at the end of reclaim to its correct
652 * position in the tree.
654 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
655 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
656 !css_tryget(&mz->memcg->css))
662 static struct mem_cgroup_per_zone *
663 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
665 struct mem_cgroup_per_zone *mz;
667 spin_lock(&mctz->lock);
668 mz = __mem_cgroup_largest_soft_limit_node(mctz);
669 spin_unlock(&mctz->lock);
674 * Implementation Note: reading percpu statistics for memcg.
676 * Both of vmstat[] and percpu_counter has threshold and do periodic
677 * synchronization to implement "quick" read. There are trade-off between
678 * reading cost and precision of value. Then, we may have a chance to implement
679 * a periodic synchronizion of counter in memcg's counter.
681 * But this _read() function is used for user interface now. The user accounts
682 * memory usage by memory cgroup and he _always_ requires exact value because
683 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
684 * have to visit all online cpus and make sum. So, for now, unnecessary
685 * synchronization is not implemented. (just implemented for cpu hotplug)
687 * If there are kernel internal actions which can make use of some not-exact
688 * value, and reading all cpu value can be performance bottleneck in some
689 * common workload, threashold and synchonization as vmstat[] should be
692 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
693 enum mem_cgroup_stat_index idx)
699 for_each_online_cpu(cpu)
700 val += per_cpu(memcg->stat->count[idx], cpu);
701 #ifdef CONFIG_HOTPLUG_CPU
702 spin_lock(&memcg->pcp_counter_lock);
703 val += memcg->nocpu_base.count[idx];
704 spin_unlock(&memcg->pcp_counter_lock);
710 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
713 int val = (charge) ? 1 : -1;
714 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
717 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
718 enum mem_cgroup_events_index idx)
720 unsigned long val = 0;
723 for_each_online_cpu(cpu)
724 val += per_cpu(memcg->stat->events[idx], cpu);
725 #ifdef CONFIG_HOTPLUG_CPU
726 spin_lock(&memcg->pcp_counter_lock);
727 val += memcg->nocpu_base.events[idx];
728 spin_unlock(&memcg->pcp_counter_lock);
733 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
734 bool anon, int nr_pages)
739 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
740 * counted as CACHE even if it's on ANON LRU.
743 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
746 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
749 /* pagein of a big page is an event. So, ignore page size */
751 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
753 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
754 nr_pages = -nr_pages; /* for event */
757 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
763 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
765 struct mem_cgroup_per_zone *mz;
767 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
768 return mz->lru_size[lru];
772 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
773 unsigned int lru_mask)
775 struct mem_cgroup_per_zone *mz;
777 unsigned long ret = 0;
779 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
782 if (BIT(lru) & lru_mask)
783 ret += mz->lru_size[lru];
789 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
790 int nid, unsigned int lru_mask)
795 for (zid = 0; zid < MAX_NR_ZONES; zid++)
796 total += mem_cgroup_zone_nr_lru_pages(memcg,
802 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
803 unsigned int lru_mask)
808 for_each_node_state(nid, N_MEMORY)
809 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
813 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
814 enum mem_cgroup_events_target target)
816 unsigned long val, next;
818 val = __this_cpu_read(memcg->stat->nr_page_events);
819 next = __this_cpu_read(memcg->stat->targets[target]);
820 /* from time_after() in jiffies.h */
821 if ((long)next - (long)val < 0) {
823 case MEM_CGROUP_TARGET_THRESH:
824 next = val + THRESHOLDS_EVENTS_TARGET;
826 case MEM_CGROUP_TARGET_SOFTLIMIT:
827 next = val + SOFTLIMIT_EVENTS_TARGET;
829 case MEM_CGROUP_TARGET_NUMAINFO:
830 next = val + NUMAINFO_EVENTS_TARGET;
835 __this_cpu_write(memcg->stat->targets[target], next);
842 * Check events in order.
845 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
848 /* threshold event is triggered in finer grain than soft limit */
849 if (unlikely(mem_cgroup_event_ratelimit(memcg,
850 MEM_CGROUP_TARGET_THRESH))) {
852 bool do_numainfo __maybe_unused;
854 do_softlimit = mem_cgroup_event_ratelimit(memcg,
855 MEM_CGROUP_TARGET_SOFTLIMIT);
857 do_numainfo = mem_cgroup_event_ratelimit(memcg,
858 MEM_CGROUP_TARGET_NUMAINFO);
862 mem_cgroup_threshold(memcg);
863 if (unlikely(do_softlimit))
864 mem_cgroup_update_tree(memcg, page);
866 if (unlikely(do_numainfo))
867 atomic_inc(&memcg->numainfo_events);
873 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
875 return mem_cgroup_from_css(
876 cgroup_subsys_state(cont, mem_cgroup_subsys_id));
879 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
882 * mm_update_next_owner() may clear mm->owner to NULL
883 * if it races with swapoff, page migration, etc.
884 * So this can be called with p == NULL.
889 return mem_cgroup_from_css(task_subsys_state(p, mem_cgroup_subsys_id));
892 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
894 struct mem_cgroup *memcg = NULL;
899 * Because we have no locks, mm->owner's may be being moved to other
900 * cgroup. We use css_tryget() here even if this looks
901 * pessimistic (rather than adding locks here).
905 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
906 if (unlikely(!memcg))
908 } while (!css_tryget(&memcg->css));
914 * mem_cgroup_iter - iterate over memory cgroup hierarchy
915 * @root: hierarchy root
916 * @prev: previously returned memcg, NULL on first invocation
917 * @reclaim: cookie for shared reclaim walks, NULL for full walks
919 * Returns references to children of the hierarchy below @root, or
920 * @root itself, or %NULL after a full round-trip.
922 * Caller must pass the return value in @prev on subsequent
923 * invocations for reference counting, or use mem_cgroup_iter_break()
924 * to cancel a hierarchy walk before the round-trip is complete.
926 * Reclaimers can specify a zone and a priority level in @reclaim to
927 * divide up the memcgs in the hierarchy among all concurrent
928 * reclaimers operating on the same zone and priority.
930 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
931 struct mem_cgroup *prev,
932 struct mem_cgroup_reclaim_cookie *reclaim)
934 struct mem_cgroup *memcg = NULL;
937 if (mem_cgroup_disabled())
941 root = root_mem_cgroup;
943 if (prev && !reclaim)
944 id = css_id(&prev->css);
946 if (prev && prev != root)
949 if (!root->use_hierarchy && root != root_mem_cgroup) {
956 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
957 struct cgroup_subsys_state *css;
960 int nid = zone_to_nid(reclaim->zone);
961 int zid = zone_idx(reclaim->zone);
962 struct mem_cgroup_per_zone *mz;
964 mz = mem_cgroup_zoneinfo(root, nid, zid);
965 iter = &mz->reclaim_iter[reclaim->priority];
966 if (prev && reclaim->generation != iter->generation)
972 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
974 if (css == &root->css || css_tryget(css))
975 memcg = mem_cgroup_from_css(css);
984 else if (!prev && memcg)
985 reclaim->generation = iter->generation;
995 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
996 * @root: hierarchy root
997 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
999 void mem_cgroup_iter_break(struct mem_cgroup *root,
1000 struct mem_cgroup *prev)
1003 root = root_mem_cgroup;
1004 if (prev && prev != root)
1005 css_put(&prev->css);
1009 * Iteration constructs for visiting all cgroups (under a tree). If
1010 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1011 * be used for reference counting.
1013 #define for_each_mem_cgroup_tree(iter, root) \
1014 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1016 iter = mem_cgroup_iter(root, iter, NULL))
1018 #define for_each_mem_cgroup(iter) \
1019 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1021 iter = mem_cgroup_iter(NULL, iter, NULL))
1023 void __mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1025 struct mem_cgroup *memcg;
1028 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1029 if (unlikely(!memcg))
1034 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1037 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1045 EXPORT_SYMBOL(__mem_cgroup_count_vm_event);
1048 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1049 * @zone: zone of the wanted lruvec
1050 * @memcg: memcg of the wanted lruvec
1052 * Returns the lru list vector holding pages for the given @zone and
1053 * @mem. This can be the global zone lruvec, if the memory controller
1056 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1057 struct mem_cgroup *memcg)
1059 struct mem_cgroup_per_zone *mz;
1060 struct lruvec *lruvec;
1062 if (mem_cgroup_disabled()) {
1063 lruvec = &zone->lruvec;
1067 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1068 lruvec = &mz->lruvec;
1071 * Since a node can be onlined after the mem_cgroup was created,
1072 * we have to be prepared to initialize lruvec->zone here;
1073 * and if offlined then reonlined, we need to reinitialize it.
1075 if (unlikely(lruvec->zone != zone))
1076 lruvec->zone = zone;
1081 * Following LRU functions are allowed to be used without PCG_LOCK.
1082 * Operations are called by routine of global LRU independently from memcg.
1083 * What we have to take care of here is validness of pc->mem_cgroup.
1085 * Changes to pc->mem_cgroup happens when
1088 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1089 * It is added to LRU before charge.
1090 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1091 * When moving account, the page is not on LRU. It's isolated.
1095 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1097 * @zone: zone of the page
1099 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1101 struct mem_cgroup_per_zone *mz;
1102 struct mem_cgroup *memcg;
1103 struct page_cgroup *pc;
1104 struct lruvec *lruvec;
1106 if (mem_cgroup_disabled()) {
1107 lruvec = &zone->lruvec;
1111 pc = lookup_page_cgroup(page);
1112 memcg = pc->mem_cgroup;
1115 * Surreptitiously switch any uncharged offlist page to root:
1116 * an uncharged page off lru does nothing to secure
1117 * its former mem_cgroup from sudden removal.
1119 * Our caller holds lru_lock, and PageCgroupUsed is updated
1120 * under page_cgroup lock: between them, they make all uses
1121 * of pc->mem_cgroup safe.
1123 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1124 pc->mem_cgroup = memcg = root_mem_cgroup;
1126 mz = page_cgroup_zoneinfo(memcg, page);
1127 lruvec = &mz->lruvec;
1130 * Since a node can be onlined after the mem_cgroup was created,
1131 * we have to be prepared to initialize lruvec->zone here;
1132 * and if offlined then reonlined, we need to reinitialize it.
1134 if (unlikely(lruvec->zone != zone))
1135 lruvec->zone = zone;
1140 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1141 * @lruvec: mem_cgroup per zone lru vector
1142 * @lru: index of lru list the page is sitting on
1143 * @nr_pages: positive when adding or negative when removing
1145 * This function must be called when a page is added to or removed from an
1148 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1151 struct mem_cgroup_per_zone *mz;
1152 unsigned long *lru_size;
1154 if (mem_cgroup_disabled())
1157 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1158 lru_size = mz->lru_size + lru;
1159 *lru_size += nr_pages;
1160 VM_BUG_ON((long)(*lru_size) < 0);
1164 * Checks whether given mem is same or in the root_mem_cgroup's
1167 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1168 struct mem_cgroup *memcg)
1170 if (root_memcg == memcg)
1172 if (!root_memcg->use_hierarchy || !memcg)
1174 return css_is_ancestor(&memcg->css, &root_memcg->css);
1177 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1178 struct mem_cgroup *memcg)
1183 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1188 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1191 struct mem_cgroup *curr = NULL;
1192 struct task_struct *p;
1194 p = find_lock_task_mm(task);
1196 curr = try_get_mem_cgroup_from_mm(p->mm);
1200 * All threads may have already detached their mm's, but the oom
1201 * killer still needs to detect if they have already been oom
1202 * killed to prevent needlessly killing additional tasks.
1205 curr = mem_cgroup_from_task(task);
1207 css_get(&curr->css);
1213 * We should check use_hierarchy of "memcg" not "curr". Because checking
1214 * use_hierarchy of "curr" here make this function true if hierarchy is
1215 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1216 * hierarchy(even if use_hierarchy is disabled in "memcg").
1218 ret = mem_cgroup_same_or_subtree(memcg, curr);
1219 css_put(&curr->css);
1223 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1225 unsigned long inactive_ratio;
1226 unsigned long inactive;
1227 unsigned long active;
1230 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1231 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1233 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1235 inactive_ratio = int_sqrt(10 * gb);
1239 return inactive * inactive_ratio < active;
1242 int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1244 unsigned long active;
1245 unsigned long inactive;
1247 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_FILE);
1248 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_FILE);
1250 return (active > inactive);
1253 #define mem_cgroup_from_res_counter(counter, member) \
1254 container_of(counter, struct mem_cgroup, member)
1257 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1258 * @memcg: the memory cgroup
1260 * Returns the maximum amount of memory @mem can be charged with, in
1263 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1265 unsigned long long margin;
1267 margin = res_counter_margin(&memcg->res);
1268 if (do_swap_account)
1269 margin = min(margin, res_counter_margin(&memcg->memsw));
1270 return margin >> PAGE_SHIFT;
1273 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1275 struct cgroup *cgrp = memcg->css.cgroup;
1278 if (cgrp->parent == NULL)
1279 return vm_swappiness;
1281 return memcg->swappiness;
1285 * memcg->moving_account is used for checking possibility that some thread is
1286 * calling move_account(). When a thread on CPU-A starts moving pages under
1287 * a memcg, other threads should check memcg->moving_account under
1288 * rcu_read_lock(), like this:
1292 * memcg->moving_account+1 if (memcg->mocing_account)
1294 * synchronize_rcu() update something.
1299 /* for quick checking without looking up memcg */
1300 atomic_t memcg_moving __read_mostly;
1302 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1304 atomic_inc(&memcg_moving);
1305 atomic_inc(&memcg->moving_account);
1309 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1312 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1313 * We check NULL in callee rather than caller.
1316 atomic_dec(&memcg_moving);
1317 atomic_dec(&memcg->moving_account);
1322 * 2 routines for checking "mem" is under move_account() or not.
1324 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1325 * is used for avoiding races in accounting. If true,
1326 * pc->mem_cgroup may be overwritten.
1328 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1329 * under hierarchy of moving cgroups. This is for
1330 * waiting at hith-memory prressure caused by "move".
1333 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1335 VM_BUG_ON(!rcu_read_lock_held());
1336 return atomic_read(&memcg->moving_account) > 0;
1339 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1341 struct mem_cgroup *from;
1342 struct mem_cgroup *to;
1345 * Unlike task_move routines, we access mc.to, mc.from not under
1346 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1348 spin_lock(&mc.lock);
1354 ret = mem_cgroup_same_or_subtree(memcg, from)
1355 || mem_cgroup_same_or_subtree(memcg, to);
1357 spin_unlock(&mc.lock);
1361 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1363 if (mc.moving_task && current != mc.moving_task) {
1364 if (mem_cgroup_under_move(memcg)) {
1366 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1367 /* moving charge context might have finished. */
1370 finish_wait(&mc.waitq, &wait);
1378 * Take this lock when
1379 * - a code tries to modify page's memcg while it's USED.
1380 * - a code tries to modify page state accounting in a memcg.
1381 * see mem_cgroup_stolen(), too.
1383 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1384 unsigned long *flags)
1386 spin_lock_irqsave(&memcg->move_lock, *flags);
1389 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1390 unsigned long *flags)
1392 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1396 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1397 * @memcg: The memory cgroup that went over limit
1398 * @p: Task that is going to be killed
1400 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1403 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1405 struct cgroup *task_cgrp;
1406 struct cgroup *mem_cgrp;
1408 * Need a buffer in BSS, can't rely on allocations. The code relies
1409 * on the assumption that OOM is serialized for memory controller.
1410 * If this assumption is broken, revisit this code.
1412 static char memcg_name[PATH_MAX];
1420 mem_cgrp = memcg->css.cgroup;
1421 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1423 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1426 * Unfortunately, we are unable to convert to a useful name
1427 * But we'll still print out the usage information
1434 printk(KERN_INFO "Task in %s killed", memcg_name);
1437 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1445 * Continues from above, so we don't need an KERN_ level
1447 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1450 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1451 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1452 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1453 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1454 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1456 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1457 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1458 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1462 * This function returns the number of memcg under hierarchy tree. Returns
1463 * 1(self count) if no children.
1465 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1468 struct mem_cgroup *iter;
1470 for_each_mem_cgroup_tree(iter, memcg)
1476 * Return the memory (and swap, if configured) limit for a memcg.
1478 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1482 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1485 * Do not consider swap space if we cannot swap due to swappiness
1487 if (mem_cgroup_swappiness(memcg)) {
1490 limit += total_swap_pages << PAGE_SHIFT;
1491 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1494 * If memsw is finite and limits the amount of swap space
1495 * available to this memcg, return that limit.
1497 limit = min(limit, memsw);
1503 static void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1506 struct mem_cgroup *iter;
1507 unsigned long chosen_points = 0;
1508 unsigned long totalpages;
1509 unsigned int points = 0;
1510 struct task_struct *chosen = NULL;
1513 * If current has a pending SIGKILL, then automatically select it. The
1514 * goal is to allow it to allocate so that it may quickly exit and free
1517 if (fatal_signal_pending(current)) {
1518 set_thread_flag(TIF_MEMDIE);
1522 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1523 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1524 for_each_mem_cgroup_tree(iter, memcg) {
1525 struct cgroup *cgroup = iter->css.cgroup;
1526 struct cgroup_iter it;
1527 struct task_struct *task;
1529 cgroup_iter_start(cgroup, &it);
1530 while ((task = cgroup_iter_next(cgroup, &it))) {
1531 switch (oom_scan_process_thread(task, totalpages, NULL,
1533 case OOM_SCAN_SELECT:
1535 put_task_struct(chosen);
1537 chosen_points = ULONG_MAX;
1538 get_task_struct(chosen);
1540 case OOM_SCAN_CONTINUE:
1542 case OOM_SCAN_ABORT:
1543 cgroup_iter_end(cgroup, &it);
1544 mem_cgroup_iter_break(memcg, iter);
1546 put_task_struct(chosen);
1551 points = oom_badness(task, memcg, NULL, totalpages);
1552 if (points > chosen_points) {
1554 put_task_struct(chosen);
1556 chosen_points = points;
1557 get_task_struct(chosen);
1560 cgroup_iter_end(cgroup, &it);
1565 points = chosen_points * 1000 / totalpages;
1566 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1567 NULL, "Memory cgroup out of memory");
1570 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1572 unsigned long flags)
1574 unsigned long total = 0;
1575 bool noswap = false;
1578 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1580 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1583 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1585 drain_all_stock_async(memcg);
1586 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1588 * Allow limit shrinkers, which are triggered directly
1589 * by userspace, to catch signals and stop reclaim
1590 * after minimal progress, regardless of the margin.
1592 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1594 if (mem_cgroup_margin(memcg))
1597 * If nothing was reclaimed after two attempts, there
1598 * may be no reclaimable pages in this hierarchy.
1607 * test_mem_cgroup_node_reclaimable
1608 * @memcg: the target memcg
1609 * @nid: the node ID to be checked.
1610 * @noswap : specify true here if the user wants flle only information.
1612 * This function returns whether the specified memcg contains any
1613 * reclaimable pages on a node. Returns true if there are any reclaimable
1614 * pages in the node.
1616 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1617 int nid, bool noswap)
1619 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1621 if (noswap || !total_swap_pages)
1623 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1628 #if MAX_NUMNODES > 1
1631 * Always updating the nodemask is not very good - even if we have an empty
1632 * list or the wrong list here, we can start from some node and traverse all
1633 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1636 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1640 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1641 * pagein/pageout changes since the last update.
1643 if (!atomic_read(&memcg->numainfo_events))
1645 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1648 /* make a nodemask where this memcg uses memory from */
1649 memcg->scan_nodes = node_states[N_MEMORY];
1651 for_each_node_mask(nid, node_states[N_MEMORY]) {
1653 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1654 node_clear(nid, memcg->scan_nodes);
1657 atomic_set(&memcg->numainfo_events, 0);
1658 atomic_set(&memcg->numainfo_updating, 0);
1662 * Selecting a node where we start reclaim from. Because what we need is just
1663 * reducing usage counter, start from anywhere is O,K. Considering
1664 * memory reclaim from current node, there are pros. and cons.
1666 * Freeing memory from current node means freeing memory from a node which
1667 * we'll use or we've used. So, it may make LRU bad. And if several threads
1668 * hit limits, it will see a contention on a node. But freeing from remote
1669 * node means more costs for memory reclaim because of memory latency.
1671 * Now, we use round-robin. Better algorithm is welcomed.
1673 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1677 mem_cgroup_may_update_nodemask(memcg);
1678 node = memcg->last_scanned_node;
1680 node = next_node(node, memcg->scan_nodes);
1681 if (node == MAX_NUMNODES)
1682 node = first_node(memcg->scan_nodes);
1684 * We call this when we hit limit, not when pages are added to LRU.
1685 * No LRU may hold pages because all pages are UNEVICTABLE or
1686 * memcg is too small and all pages are not on LRU. In that case,
1687 * we use curret node.
1689 if (unlikely(node == MAX_NUMNODES))
1690 node = numa_node_id();
1692 memcg->last_scanned_node = node;
1697 * Check all nodes whether it contains reclaimable pages or not.
1698 * For quick scan, we make use of scan_nodes. This will allow us to skip
1699 * unused nodes. But scan_nodes is lazily updated and may not cotain
1700 * enough new information. We need to do double check.
1702 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1707 * quick check...making use of scan_node.
1708 * We can skip unused nodes.
1710 if (!nodes_empty(memcg->scan_nodes)) {
1711 for (nid = first_node(memcg->scan_nodes);
1713 nid = next_node(nid, memcg->scan_nodes)) {
1715 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1720 * Check rest of nodes.
1722 for_each_node_state(nid, N_MEMORY) {
1723 if (node_isset(nid, memcg->scan_nodes))
1725 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1732 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1737 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1739 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1743 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1746 unsigned long *total_scanned)
1748 struct mem_cgroup *victim = NULL;
1751 unsigned long excess;
1752 unsigned long nr_scanned;
1753 struct mem_cgroup_reclaim_cookie reclaim = {
1758 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1761 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1766 * If we have not been able to reclaim
1767 * anything, it might because there are
1768 * no reclaimable pages under this hierarchy
1773 * We want to do more targeted reclaim.
1774 * excess >> 2 is not to excessive so as to
1775 * reclaim too much, nor too less that we keep
1776 * coming back to reclaim from this cgroup
1778 if (total >= (excess >> 2) ||
1779 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1784 if (!mem_cgroup_reclaimable(victim, false))
1786 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1788 *total_scanned += nr_scanned;
1789 if (!res_counter_soft_limit_excess(&root_memcg->res))
1792 mem_cgroup_iter_break(root_memcg, victim);
1797 * Check OOM-Killer is already running under our hierarchy.
1798 * If someone is running, return false.
1799 * Has to be called with memcg_oom_lock
1801 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1803 struct mem_cgroup *iter, *failed = NULL;
1805 for_each_mem_cgroup_tree(iter, memcg) {
1806 if (iter->oom_lock) {
1808 * this subtree of our hierarchy is already locked
1809 * so we cannot give a lock.
1812 mem_cgroup_iter_break(memcg, iter);
1815 iter->oom_lock = true;
1822 * OK, we failed to lock the whole subtree so we have to clean up
1823 * what we set up to the failing subtree
1825 for_each_mem_cgroup_tree(iter, memcg) {
1826 if (iter == failed) {
1827 mem_cgroup_iter_break(memcg, iter);
1830 iter->oom_lock = false;
1836 * Has to be called with memcg_oom_lock
1838 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1840 struct mem_cgroup *iter;
1842 for_each_mem_cgroup_tree(iter, memcg)
1843 iter->oom_lock = false;
1847 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1849 struct mem_cgroup *iter;
1851 for_each_mem_cgroup_tree(iter, memcg)
1852 atomic_inc(&iter->under_oom);
1855 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1857 struct mem_cgroup *iter;
1860 * When a new child is created while the hierarchy is under oom,
1861 * mem_cgroup_oom_lock() may not be called. We have to use
1862 * atomic_add_unless() here.
1864 for_each_mem_cgroup_tree(iter, memcg)
1865 atomic_add_unless(&iter->under_oom, -1, 0);
1868 static DEFINE_SPINLOCK(memcg_oom_lock);
1869 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1871 struct oom_wait_info {
1872 struct mem_cgroup *memcg;
1876 static int memcg_oom_wake_function(wait_queue_t *wait,
1877 unsigned mode, int sync, void *arg)
1879 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg;
1880 struct mem_cgroup *oom_wait_memcg;
1881 struct oom_wait_info *oom_wait_info;
1883 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1884 oom_wait_memcg = oom_wait_info->memcg;
1887 * Both of oom_wait_info->memcg and wake_memcg are stable under us.
1888 * Then we can use css_is_ancestor without taking care of RCU.
1890 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1891 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1893 return autoremove_wake_function(wait, mode, sync, arg);
1896 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1898 /* for filtering, pass "memcg" as argument. */
1899 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1902 static void memcg_oom_recover(struct mem_cgroup *memcg)
1904 if (memcg && atomic_read(&memcg->under_oom))
1905 memcg_wakeup_oom(memcg);
1909 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1911 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
1914 struct oom_wait_info owait;
1915 bool locked, need_to_kill;
1917 owait.memcg = memcg;
1918 owait.wait.flags = 0;
1919 owait.wait.func = memcg_oom_wake_function;
1920 owait.wait.private = current;
1921 INIT_LIST_HEAD(&owait.wait.task_list);
1922 need_to_kill = true;
1923 mem_cgroup_mark_under_oom(memcg);
1925 /* At first, try to OOM lock hierarchy under memcg.*/
1926 spin_lock(&memcg_oom_lock);
1927 locked = mem_cgroup_oom_lock(memcg);
1929 * Even if signal_pending(), we can't quit charge() loop without
1930 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1931 * under OOM is always welcomed, use TASK_KILLABLE here.
1933 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1934 if (!locked || memcg->oom_kill_disable)
1935 need_to_kill = false;
1937 mem_cgroup_oom_notify(memcg);
1938 spin_unlock(&memcg_oom_lock);
1941 finish_wait(&memcg_oom_waitq, &owait.wait);
1942 mem_cgroup_out_of_memory(memcg, mask, order);
1945 finish_wait(&memcg_oom_waitq, &owait.wait);
1947 spin_lock(&memcg_oom_lock);
1949 mem_cgroup_oom_unlock(memcg);
1950 memcg_wakeup_oom(memcg);
1951 spin_unlock(&memcg_oom_lock);
1953 mem_cgroup_unmark_under_oom(memcg);
1955 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1957 /* Give chance to dying process */
1958 schedule_timeout_uninterruptible(1);
1963 * Currently used to update mapped file statistics, but the routine can be
1964 * generalized to update other statistics as well.
1966 * Notes: Race condition
1968 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1969 * it tends to be costly. But considering some conditions, we doesn't need
1970 * to do so _always_.
1972 * Considering "charge", lock_page_cgroup() is not required because all
1973 * file-stat operations happen after a page is attached to radix-tree. There
1974 * are no race with "charge".
1976 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1977 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1978 * if there are race with "uncharge". Statistics itself is properly handled
1981 * Considering "move", this is an only case we see a race. To make the race
1982 * small, we check mm->moving_account and detect there are possibility of race
1983 * If there is, we take a lock.
1986 void __mem_cgroup_begin_update_page_stat(struct page *page,
1987 bool *locked, unsigned long *flags)
1989 struct mem_cgroup *memcg;
1990 struct page_cgroup *pc;
1992 pc = lookup_page_cgroup(page);
1994 memcg = pc->mem_cgroup;
1995 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1998 * If this memory cgroup is not under account moving, we don't
1999 * need to take move_lock_mem_cgroup(). Because we already hold
2000 * rcu_read_lock(), any calls to move_account will be delayed until
2001 * rcu_read_unlock() if mem_cgroup_stolen() == true.
2003 if (!mem_cgroup_stolen(memcg))
2006 move_lock_mem_cgroup(memcg, flags);
2007 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
2008 move_unlock_mem_cgroup(memcg, flags);
2014 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
2016 struct page_cgroup *pc = lookup_page_cgroup(page);
2019 * It's guaranteed that pc->mem_cgroup never changes while
2020 * lock is held because a routine modifies pc->mem_cgroup
2021 * should take move_lock_mem_cgroup().
2023 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
2026 void mem_cgroup_update_page_stat(struct page *page,
2027 enum mem_cgroup_page_stat_item idx, int val)
2029 struct mem_cgroup *memcg;
2030 struct page_cgroup *pc = lookup_page_cgroup(page);
2031 unsigned long uninitialized_var(flags);
2033 if (mem_cgroup_disabled())
2036 memcg = pc->mem_cgroup;
2037 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2041 case MEMCG_NR_FILE_MAPPED:
2042 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2048 this_cpu_add(memcg->stat->count[idx], val);
2052 * size of first charge trial. "32" comes from vmscan.c's magic value.
2053 * TODO: maybe necessary to use big numbers in big irons.
2055 #define CHARGE_BATCH 32U
2056 struct memcg_stock_pcp {
2057 struct mem_cgroup *cached; /* this never be root cgroup */
2058 unsigned int nr_pages;
2059 struct work_struct work;
2060 unsigned long flags;
2061 #define FLUSHING_CACHED_CHARGE 0
2063 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2064 static DEFINE_MUTEX(percpu_charge_mutex);
2067 * consume_stock: Try to consume stocked charge on this cpu.
2068 * @memcg: memcg to consume from.
2069 * @nr_pages: how many pages to charge.
2071 * The charges will only happen if @memcg matches the current cpu's memcg
2072 * stock, and at least @nr_pages are available in that stock. Failure to
2073 * service an allocation will refill the stock.
2075 * returns true if successful, false otherwise.
2077 static bool consume_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2079 struct memcg_stock_pcp *stock;
2082 if (nr_pages > CHARGE_BATCH)
2085 stock = &get_cpu_var(memcg_stock);
2086 if (memcg == stock->cached && stock->nr_pages >= nr_pages)
2087 stock->nr_pages -= nr_pages;
2088 else /* need to call res_counter_charge */
2090 put_cpu_var(memcg_stock);
2095 * Returns stocks cached in percpu to res_counter and reset cached information.
2097 static void drain_stock(struct memcg_stock_pcp *stock)
2099 struct mem_cgroup *old = stock->cached;
2101 if (stock->nr_pages) {
2102 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2104 res_counter_uncharge(&old->res, bytes);
2105 if (do_swap_account)
2106 res_counter_uncharge(&old->memsw, bytes);
2107 stock->nr_pages = 0;
2109 stock->cached = NULL;
2113 * This must be called under preempt disabled or must be called by
2114 * a thread which is pinned to local cpu.
2116 static void drain_local_stock(struct work_struct *dummy)
2118 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2120 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2124 * Cache charges(val) which is from res_counter, to local per_cpu area.
2125 * This will be consumed by consume_stock() function, later.
2127 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2129 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2131 if (stock->cached != memcg) { /* reset if necessary */
2133 stock->cached = memcg;
2135 stock->nr_pages += nr_pages;
2136 put_cpu_var(memcg_stock);
2140 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2141 * of the hierarchy under it. sync flag says whether we should block
2142 * until the work is done.
2144 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2148 /* Notify other cpus that system-wide "drain" is running */
2151 for_each_online_cpu(cpu) {
2152 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2153 struct mem_cgroup *memcg;
2155 memcg = stock->cached;
2156 if (!memcg || !stock->nr_pages)
2158 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2160 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2162 drain_local_stock(&stock->work);
2164 schedule_work_on(cpu, &stock->work);
2172 for_each_online_cpu(cpu) {
2173 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2174 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2175 flush_work(&stock->work);
2182 * Tries to drain stocked charges in other cpus. This function is asynchronous
2183 * and just put a work per cpu for draining localy on each cpu. Caller can
2184 * expects some charges will be back to res_counter later but cannot wait for
2187 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2190 * If someone calls draining, avoid adding more kworker runs.
2192 if (!mutex_trylock(&percpu_charge_mutex))
2194 drain_all_stock(root_memcg, false);
2195 mutex_unlock(&percpu_charge_mutex);
2198 /* This is a synchronous drain interface. */
2199 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2201 /* called when force_empty is called */
2202 mutex_lock(&percpu_charge_mutex);
2203 drain_all_stock(root_memcg, true);
2204 mutex_unlock(&percpu_charge_mutex);
2208 * This function drains percpu counter value from DEAD cpu and
2209 * move it to local cpu. Note that this function can be preempted.
2211 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2215 spin_lock(&memcg->pcp_counter_lock);
2216 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2217 long x = per_cpu(memcg->stat->count[i], cpu);
2219 per_cpu(memcg->stat->count[i], cpu) = 0;
2220 memcg->nocpu_base.count[i] += x;
2222 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2223 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2225 per_cpu(memcg->stat->events[i], cpu) = 0;
2226 memcg->nocpu_base.events[i] += x;
2228 spin_unlock(&memcg->pcp_counter_lock);
2231 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2232 unsigned long action,
2235 int cpu = (unsigned long)hcpu;
2236 struct memcg_stock_pcp *stock;
2237 struct mem_cgroup *iter;
2239 if (action == CPU_ONLINE)
2242 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2245 for_each_mem_cgroup(iter)
2246 mem_cgroup_drain_pcp_counter(iter, cpu);
2248 stock = &per_cpu(memcg_stock, cpu);
2254 /* See __mem_cgroup_try_charge() for details */
2256 CHARGE_OK, /* success */
2257 CHARGE_RETRY, /* need to retry but retry is not bad */
2258 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2259 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2260 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2263 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2264 unsigned int nr_pages, unsigned int min_pages,
2267 unsigned long csize = nr_pages * PAGE_SIZE;
2268 struct mem_cgroup *mem_over_limit;
2269 struct res_counter *fail_res;
2270 unsigned long flags = 0;
2273 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2276 if (!do_swap_account)
2278 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2282 res_counter_uncharge(&memcg->res, csize);
2283 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2284 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2286 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2288 * Never reclaim on behalf of optional batching, retry with a
2289 * single page instead.
2291 if (nr_pages > min_pages)
2292 return CHARGE_RETRY;
2294 if (!(gfp_mask & __GFP_WAIT))
2295 return CHARGE_WOULDBLOCK;
2297 if (gfp_mask & __GFP_NORETRY)
2298 return CHARGE_NOMEM;
2300 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2301 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2302 return CHARGE_RETRY;
2304 * Even though the limit is exceeded at this point, reclaim
2305 * may have been able to free some pages. Retry the charge
2306 * before killing the task.
2308 * Only for regular pages, though: huge pages are rather
2309 * unlikely to succeed so close to the limit, and we fall back
2310 * to regular pages anyway in case of failure.
2312 if (nr_pages <= (1 << PAGE_ALLOC_COSTLY_ORDER) && ret)
2313 return CHARGE_RETRY;
2316 * At task move, charge accounts can be doubly counted. So, it's
2317 * better to wait until the end of task_move if something is going on.
2319 if (mem_cgroup_wait_acct_move(mem_over_limit))
2320 return CHARGE_RETRY;
2322 /* If we don't need to call oom-killer at el, return immediately */
2324 return CHARGE_NOMEM;
2326 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2327 return CHARGE_OOM_DIE;
2329 return CHARGE_RETRY;
2333 * __mem_cgroup_try_charge() does
2334 * 1. detect memcg to be charged against from passed *mm and *ptr,
2335 * 2. update res_counter
2336 * 3. call memory reclaim if necessary.
2338 * In some special case, if the task is fatal, fatal_signal_pending() or
2339 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2340 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2341 * as possible without any hazards. 2: all pages should have a valid
2342 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2343 * pointer, that is treated as a charge to root_mem_cgroup.
2345 * So __mem_cgroup_try_charge() will return
2346 * 0 ... on success, filling *ptr with a valid memcg pointer.
2347 * -ENOMEM ... charge failure because of resource limits.
2348 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2350 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2351 * the oom-killer can be invoked.
2353 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2355 unsigned int nr_pages,
2356 struct mem_cgroup **ptr,
2359 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2360 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2361 struct mem_cgroup *memcg = NULL;
2365 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2366 * in system level. So, allow to go ahead dying process in addition to
2369 if (unlikely(test_thread_flag(TIF_MEMDIE)
2370 || fatal_signal_pending(current)))
2374 * We always charge the cgroup the mm_struct belongs to.
2375 * The mm_struct's mem_cgroup changes on task migration if the
2376 * thread group leader migrates. It's possible that mm is not
2377 * set, if so charge the root memcg (happens for pagecache usage).
2380 *ptr = root_mem_cgroup;
2382 if (*ptr) { /* css should be a valid one */
2384 if (mem_cgroup_is_root(memcg))
2386 if (consume_stock(memcg, nr_pages))
2388 css_get(&memcg->css);
2390 struct task_struct *p;
2393 p = rcu_dereference(mm->owner);
2395 * Because we don't have task_lock(), "p" can exit.
2396 * In that case, "memcg" can point to root or p can be NULL with
2397 * race with swapoff. Then, we have small risk of mis-accouning.
2398 * But such kind of mis-account by race always happens because
2399 * we don't have cgroup_mutex(). It's overkill and we allo that
2401 * (*) swapoff at el will charge against mm-struct not against
2402 * task-struct. So, mm->owner can be NULL.
2404 memcg = mem_cgroup_from_task(p);
2406 memcg = root_mem_cgroup;
2407 if (mem_cgroup_is_root(memcg)) {
2411 if (consume_stock(memcg, nr_pages)) {
2413 * It seems dagerous to access memcg without css_get().
2414 * But considering how consume_stok works, it's not
2415 * necessary. If consume_stock success, some charges
2416 * from this memcg are cached on this cpu. So, we
2417 * don't need to call css_get()/css_tryget() before
2418 * calling consume_stock().
2423 /* after here, we may be blocked. we need to get refcnt */
2424 if (!css_tryget(&memcg->css)) {
2434 /* If killed, bypass charge */
2435 if (fatal_signal_pending(current)) {
2436 css_put(&memcg->css);
2441 if (oom && !nr_oom_retries) {
2443 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2446 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, nr_pages,
2451 case CHARGE_RETRY: /* not in OOM situation but retry */
2453 css_put(&memcg->css);
2456 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2457 css_put(&memcg->css);
2459 case CHARGE_NOMEM: /* OOM routine works */
2461 css_put(&memcg->css);
2464 /* If oom, we never return -ENOMEM */
2467 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2468 css_put(&memcg->css);
2471 } while (ret != CHARGE_OK);
2473 if (batch > nr_pages)
2474 refill_stock(memcg, batch - nr_pages);
2475 css_put(&memcg->css);
2483 *ptr = root_mem_cgroup;
2488 * Somemtimes we have to undo a charge we got by try_charge().
2489 * This function is for that and do uncharge, put css's refcnt.
2490 * gotten by try_charge().
2492 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2493 unsigned int nr_pages)
2495 if (!mem_cgroup_is_root(memcg)) {
2496 unsigned long bytes = nr_pages * PAGE_SIZE;
2498 res_counter_uncharge(&memcg->res, bytes);
2499 if (do_swap_account)
2500 res_counter_uncharge(&memcg->memsw, bytes);
2505 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2506 * This is useful when moving usage to parent cgroup.
2508 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2509 unsigned int nr_pages)
2511 unsigned long bytes = nr_pages * PAGE_SIZE;
2513 if (mem_cgroup_is_root(memcg))
2516 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2517 if (do_swap_account)
2518 res_counter_uncharge_until(&memcg->memsw,
2519 memcg->memsw.parent, bytes);
2523 * A helper function to get mem_cgroup from ID. must be called under
2524 * rcu_read_lock(). The caller is responsible for calling css_tryget if
2525 * the mem_cgroup is used for charging. (dropping refcnt from swap can be
2526 * called against removed memcg.)
2528 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2530 struct cgroup_subsys_state *css;
2532 /* ID 0 is unused ID */
2535 css = css_lookup(&mem_cgroup_subsys, id);
2538 return mem_cgroup_from_css(css);
2541 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2543 struct mem_cgroup *memcg = NULL;
2544 struct page_cgroup *pc;
2548 VM_BUG_ON(!PageLocked(page));
2550 pc = lookup_page_cgroup(page);
2551 lock_page_cgroup(pc);
2552 if (PageCgroupUsed(pc)) {
2553 memcg = pc->mem_cgroup;
2554 if (memcg && !css_tryget(&memcg->css))
2556 } else if (PageSwapCache(page)) {
2557 ent.val = page_private(page);
2558 id = lookup_swap_cgroup_id(ent);
2560 memcg = mem_cgroup_lookup(id);
2561 if (memcg && !css_tryget(&memcg->css))
2565 unlock_page_cgroup(pc);
2569 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2571 unsigned int nr_pages,
2572 enum charge_type ctype,
2575 struct page_cgroup *pc = lookup_page_cgroup(page);
2576 struct zone *uninitialized_var(zone);
2577 struct lruvec *lruvec;
2578 bool was_on_lru = false;
2581 lock_page_cgroup(pc);
2582 VM_BUG_ON(PageCgroupUsed(pc));
2584 * we don't need page_cgroup_lock about tail pages, becase they are not
2585 * accessed by any other context at this point.
2589 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2590 * may already be on some other mem_cgroup's LRU. Take care of it.
2593 zone = page_zone(page);
2594 spin_lock_irq(&zone->lru_lock);
2595 if (PageLRU(page)) {
2596 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2598 del_page_from_lru_list(page, lruvec, page_lru(page));
2603 pc->mem_cgroup = memcg;
2605 * We access a page_cgroup asynchronously without lock_page_cgroup().
2606 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2607 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2608 * before USED bit, we need memory barrier here.
2609 * See mem_cgroup_add_lru_list(), etc.
2612 SetPageCgroupUsed(pc);
2616 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2617 VM_BUG_ON(PageLRU(page));
2619 add_page_to_lru_list(page, lruvec, page_lru(page));
2621 spin_unlock_irq(&zone->lru_lock);
2624 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2629 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2630 unlock_page_cgroup(pc);
2633 * "charge_statistics" updated event counter. Then, check it.
2634 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2635 * if they exceeds softlimit.
2637 memcg_check_events(memcg, page);
2640 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2642 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
2644 * Because tail pages are not marked as "used", set it. We're under
2645 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2646 * charge/uncharge will be never happen and move_account() is done under
2647 * compound_lock(), so we don't have to take care of races.
2649 void mem_cgroup_split_huge_fixup(struct page *head)
2651 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2652 struct page_cgroup *pc;
2655 if (mem_cgroup_disabled())
2657 for (i = 1; i < HPAGE_PMD_NR; i++) {
2659 pc->mem_cgroup = head_pc->mem_cgroup;
2660 smp_wmb();/* see __commit_charge() */
2661 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2664 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2667 * mem_cgroup_move_account - move account of the page
2669 * @nr_pages: number of regular pages (>1 for huge pages)
2670 * @pc: page_cgroup of the page.
2671 * @from: mem_cgroup which the page is moved from.
2672 * @to: mem_cgroup which the page is moved to. @from != @to.
2674 * The caller must confirm following.
2675 * - page is not on LRU (isolate_page() is useful.)
2676 * - compound_lock is held when nr_pages > 1
2678 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2681 static int mem_cgroup_move_account(struct page *page,
2682 unsigned int nr_pages,
2683 struct page_cgroup *pc,
2684 struct mem_cgroup *from,
2685 struct mem_cgroup *to)
2687 unsigned long flags;
2689 bool anon = PageAnon(page);
2691 VM_BUG_ON(from == to);
2692 VM_BUG_ON(PageLRU(page));
2694 * The page is isolated from LRU. So, collapse function
2695 * will not handle this page. But page splitting can happen.
2696 * Do this check under compound_page_lock(). The caller should
2700 if (nr_pages > 1 && !PageTransHuge(page))
2703 lock_page_cgroup(pc);
2706 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2709 move_lock_mem_cgroup(from, &flags);
2711 if (!anon && page_mapped(page)) {
2712 /* Update mapped_file data for mem_cgroup */
2714 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2715 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2718 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2720 /* caller should have done css_get */
2721 pc->mem_cgroup = to;
2722 mem_cgroup_charge_statistics(to, anon, nr_pages);
2723 move_unlock_mem_cgroup(from, &flags);
2726 unlock_page_cgroup(pc);
2730 memcg_check_events(to, page);
2731 memcg_check_events(from, page);
2737 * mem_cgroup_move_parent - moves page to the parent group
2738 * @page: the page to move
2739 * @pc: page_cgroup of the page
2740 * @child: page's cgroup
2742 * move charges to its parent or the root cgroup if the group has no
2743 * parent (aka use_hierarchy==0).
2744 * Although this might fail (get_page_unless_zero, isolate_lru_page or
2745 * mem_cgroup_move_account fails) the failure is always temporary and
2746 * it signals a race with a page removal/uncharge or migration. In the
2747 * first case the page is on the way out and it will vanish from the LRU
2748 * on the next attempt and the call should be retried later.
2749 * Isolation from the LRU fails only if page has been isolated from
2750 * the LRU since we looked at it and that usually means either global
2751 * reclaim or migration going on. The page will either get back to the
2753 * Finaly mem_cgroup_move_account fails only if the page got uncharged
2754 * (!PageCgroupUsed) or moved to a different group. The page will
2755 * disappear in the next attempt.
2757 static int mem_cgroup_move_parent(struct page *page,
2758 struct page_cgroup *pc,
2759 struct mem_cgroup *child)
2761 struct mem_cgroup *parent;
2762 unsigned int nr_pages;
2763 unsigned long uninitialized_var(flags);
2766 VM_BUG_ON(mem_cgroup_is_root(child));
2769 if (!get_page_unless_zero(page))
2771 if (isolate_lru_page(page))
2774 nr_pages = hpage_nr_pages(page);
2776 parent = parent_mem_cgroup(child);
2778 * If no parent, move charges to root cgroup.
2781 parent = root_mem_cgroup;
2784 VM_BUG_ON(!PageTransHuge(page));
2785 flags = compound_lock_irqsave(page);
2788 ret = mem_cgroup_move_account(page, nr_pages,
2791 __mem_cgroup_cancel_local_charge(child, nr_pages);
2794 compound_unlock_irqrestore(page, flags);
2795 putback_lru_page(page);
2803 * Charge the memory controller for page usage.
2805 * 0 if the charge was successful
2806 * < 0 if the cgroup is over its limit
2808 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2809 gfp_t gfp_mask, enum charge_type ctype)
2811 struct mem_cgroup *memcg = NULL;
2812 unsigned int nr_pages = 1;
2816 if (PageTransHuge(page)) {
2817 nr_pages <<= compound_order(page);
2818 VM_BUG_ON(!PageTransHuge(page));
2820 * Never OOM-kill a process for a huge page. The
2821 * fault handler will fall back to regular pages.
2826 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2829 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
2833 int mem_cgroup_newpage_charge(struct page *page,
2834 struct mm_struct *mm, gfp_t gfp_mask)
2836 if (mem_cgroup_disabled())
2838 VM_BUG_ON(page_mapped(page));
2839 VM_BUG_ON(page->mapping && !PageAnon(page));
2841 return mem_cgroup_charge_common(page, mm, gfp_mask,
2842 MEM_CGROUP_CHARGE_TYPE_ANON);
2846 * While swap-in, try_charge -> commit or cancel, the page is locked.
2847 * And when try_charge() successfully returns, one refcnt to memcg without
2848 * struct page_cgroup is acquired. This refcnt will be consumed by
2849 * "commit()" or removed by "cancel()"
2851 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2854 struct mem_cgroup **memcgp)
2856 struct mem_cgroup *memcg;
2857 struct page_cgroup *pc;
2860 pc = lookup_page_cgroup(page);
2862 * Every swap fault against a single page tries to charge the
2863 * page, bail as early as possible. shmem_unuse() encounters
2864 * already charged pages, too. The USED bit is protected by
2865 * the page lock, which serializes swap cache removal, which
2866 * in turn serializes uncharging.
2868 if (PageCgroupUsed(pc))
2870 if (!do_swap_account)
2872 memcg = try_get_mem_cgroup_from_page(page);
2876 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2877 css_put(&memcg->css);
2882 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2888 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
2889 gfp_t gfp_mask, struct mem_cgroup **memcgp)
2892 if (mem_cgroup_disabled())
2895 * A racing thread's fault, or swapoff, may have already
2896 * updated the pte, and even removed page from swap cache: in
2897 * those cases unuse_pte()'s pte_same() test will fail; but
2898 * there's also a KSM case which does need to charge the page.
2900 if (!PageSwapCache(page)) {
2903 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, memcgp, true);
2908 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
2911 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2913 if (mem_cgroup_disabled())
2917 __mem_cgroup_cancel_charge(memcg, 1);
2921 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2922 enum charge_type ctype)
2924 if (mem_cgroup_disabled())
2929 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
2931 * Now swap is on-memory. This means this page may be
2932 * counted both as mem and swap....double count.
2933 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2934 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2935 * may call delete_from_swap_cache() before reach here.
2937 if (do_swap_account && PageSwapCache(page)) {
2938 swp_entry_t ent = {.val = page_private(page)};
2939 mem_cgroup_uncharge_swap(ent);
2943 void mem_cgroup_commit_charge_swapin(struct page *page,
2944 struct mem_cgroup *memcg)
2946 __mem_cgroup_commit_charge_swapin(page, memcg,
2947 MEM_CGROUP_CHARGE_TYPE_ANON);
2950 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2953 struct mem_cgroup *memcg = NULL;
2954 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2957 if (mem_cgroup_disabled())
2959 if (PageCompound(page))
2962 if (!PageSwapCache(page))
2963 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2964 else { /* page is swapcache/shmem */
2965 ret = __mem_cgroup_try_charge_swapin(mm, page,
2968 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2973 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2974 unsigned int nr_pages,
2975 const enum charge_type ctype)
2977 struct memcg_batch_info *batch = NULL;
2978 bool uncharge_memsw = true;
2980 /* If swapout, usage of swap doesn't decrease */
2981 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2982 uncharge_memsw = false;
2984 batch = ¤t->memcg_batch;
2986 * In usual, we do css_get() when we remember memcg pointer.
2987 * But in this case, we keep res->usage until end of a series of
2988 * uncharges. Then, it's ok to ignore memcg's refcnt.
2991 batch->memcg = memcg;
2993 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2994 * In those cases, all pages freed continuously can be expected to be in
2995 * the same cgroup and we have chance to coalesce uncharges.
2996 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2997 * because we want to do uncharge as soon as possible.
3000 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3001 goto direct_uncharge;
3004 goto direct_uncharge;
3007 * In typical case, batch->memcg == mem. This means we can
3008 * merge a series of uncharges to an uncharge of res_counter.
3009 * If not, we uncharge res_counter ony by one.
3011 if (batch->memcg != memcg)
3012 goto direct_uncharge;
3013 /* remember freed charge and uncharge it later */
3016 batch->memsw_nr_pages++;
3019 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3021 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3022 if (unlikely(batch->memcg != memcg))
3023 memcg_oom_recover(memcg);
3027 * uncharge if !page_mapped(page)
3029 static struct mem_cgroup *
3030 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
3033 struct mem_cgroup *memcg = NULL;
3034 unsigned int nr_pages = 1;
3035 struct page_cgroup *pc;
3038 if (mem_cgroup_disabled())
3041 VM_BUG_ON(PageSwapCache(page));
3043 if (PageTransHuge(page)) {
3044 nr_pages <<= compound_order(page);
3045 VM_BUG_ON(!PageTransHuge(page));
3048 * Check if our page_cgroup is valid
3050 pc = lookup_page_cgroup(page);
3051 if (unlikely(!PageCgroupUsed(pc)))
3054 lock_page_cgroup(pc);
3056 memcg = pc->mem_cgroup;
3058 if (!PageCgroupUsed(pc))
3061 anon = PageAnon(page);
3064 case MEM_CGROUP_CHARGE_TYPE_ANON:
3066 * Generally PageAnon tells if it's the anon statistics to be
3067 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3068 * used before page reached the stage of being marked PageAnon.
3072 case MEM_CGROUP_CHARGE_TYPE_DROP:
3073 /* See mem_cgroup_prepare_migration() */
3074 if (page_mapped(page))
3077 * Pages under migration may not be uncharged. But
3078 * end_migration() /must/ be the one uncharging the
3079 * unused post-migration page and so it has to call
3080 * here with the migration bit still set. See the
3081 * res_counter handling below.
3083 if (!end_migration && PageCgroupMigration(pc))
3086 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3087 if (!PageAnon(page)) { /* Shared memory */
3088 if (page->mapping && !page_is_file_cache(page))
3090 } else if (page_mapped(page)) /* Anon */
3097 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3099 ClearPageCgroupUsed(pc);
3101 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3102 * freed from LRU. This is safe because uncharged page is expected not
3103 * to be reused (freed soon). Exception is SwapCache, it's handled by
3104 * special functions.
3107 unlock_page_cgroup(pc);
3109 * even after unlock, we have memcg->res.usage here and this memcg
3110 * will never be freed.
3112 memcg_check_events(memcg, page);
3113 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3114 mem_cgroup_swap_statistics(memcg, true);
3115 mem_cgroup_get(memcg);
3118 * Migration does not charge the res_counter for the
3119 * replacement page, so leave it alone when phasing out the
3120 * page that is unused after the migration.
3122 if (!end_migration && !mem_cgroup_is_root(memcg))
3123 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3128 unlock_page_cgroup(pc);
3132 void mem_cgroup_uncharge_page(struct page *page)
3135 if (page_mapped(page))
3137 VM_BUG_ON(page->mapping && !PageAnon(page));
3138 if (PageSwapCache(page))
3140 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
3143 void mem_cgroup_uncharge_cache_page(struct page *page)
3145 VM_BUG_ON(page_mapped(page));
3146 VM_BUG_ON(page->mapping);
3147 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
3151 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3152 * In that cases, pages are freed continuously and we can expect pages
3153 * are in the same memcg. All these calls itself limits the number of
3154 * pages freed at once, then uncharge_start/end() is called properly.
3155 * This may be called prural(2) times in a context,
3158 void mem_cgroup_uncharge_start(void)
3160 current->memcg_batch.do_batch++;
3161 /* We can do nest. */
3162 if (current->memcg_batch.do_batch == 1) {
3163 current->memcg_batch.memcg = NULL;
3164 current->memcg_batch.nr_pages = 0;
3165 current->memcg_batch.memsw_nr_pages = 0;
3169 void mem_cgroup_uncharge_end(void)
3171 struct memcg_batch_info *batch = ¤t->memcg_batch;
3173 if (!batch->do_batch)
3177 if (batch->do_batch) /* If stacked, do nothing. */
3183 * This "batch->memcg" is valid without any css_get/put etc...
3184 * bacause we hide charges behind us.
3186 if (batch->nr_pages)
3187 res_counter_uncharge(&batch->memcg->res,
3188 batch->nr_pages * PAGE_SIZE);
3189 if (batch->memsw_nr_pages)
3190 res_counter_uncharge(&batch->memcg->memsw,
3191 batch->memsw_nr_pages * PAGE_SIZE);
3192 memcg_oom_recover(batch->memcg);
3193 /* forget this pointer (for sanity check) */
3194 batch->memcg = NULL;
3199 * called after __delete_from_swap_cache() and drop "page" account.
3200 * memcg information is recorded to swap_cgroup of "ent"
3203 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3205 struct mem_cgroup *memcg;
3206 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3208 if (!swapout) /* this was a swap cache but the swap is unused ! */
3209 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3211 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
3214 * record memcg information, if swapout && memcg != NULL,
3215 * mem_cgroup_get() was called in uncharge().
3217 if (do_swap_account && swapout && memcg)
3218 swap_cgroup_record(ent, css_id(&memcg->css));
3222 #ifdef CONFIG_MEMCG_SWAP
3224 * called from swap_entry_free(). remove record in swap_cgroup and
3225 * uncharge "memsw" account.
3227 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3229 struct mem_cgroup *memcg;
3232 if (!do_swap_account)
3235 id = swap_cgroup_record(ent, 0);
3237 memcg = mem_cgroup_lookup(id);
3240 * We uncharge this because swap is freed.
3241 * This memcg can be obsolete one. We avoid calling css_tryget
3243 if (!mem_cgroup_is_root(memcg))
3244 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3245 mem_cgroup_swap_statistics(memcg, false);
3246 mem_cgroup_put(memcg);
3252 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3253 * @entry: swap entry to be moved
3254 * @from: mem_cgroup which the entry is moved from
3255 * @to: mem_cgroup which the entry is moved to
3257 * It succeeds only when the swap_cgroup's record for this entry is the same
3258 * as the mem_cgroup's id of @from.
3260 * Returns 0 on success, -EINVAL on failure.
3262 * The caller must have charged to @to, IOW, called res_counter_charge() about
3263 * both res and memsw, and called css_get().
3265 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3266 struct mem_cgroup *from, struct mem_cgroup *to)
3268 unsigned short old_id, new_id;
3270 old_id = css_id(&from->css);
3271 new_id = css_id(&to->css);
3273 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3274 mem_cgroup_swap_statistics(from, false);
3275 mem_cgroup_swap_statistics(to, true);
3277 * This function is only called from task migration context now.
3278 * It postpones res_counter and refcount handling till the end
3279 * of task migration(mem_cgroup_clear_mc()) for performance
3280 * improvement. But we cannot postpone mem_cgroup_get(to)
3281 * because if the process that has been moved to @to does
3282 * swap-in, the refcount of @to might be decreased to 0.
3290 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3291 struct mem_cgroup *from, struct mem_cgroup *to)
3298 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3301 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
3302 struct mem_cgroup **memcgp)
3304 struct mem_cgroup *memcg = NULL;
3305 unsigned int nr_pages = 1;
3306 struct page_cgroup *pc;
3307 enum charge_type ctype;
3311 if (mem_cgroup_disabled())
3314 if (PageTransHuge(page))
3315 nr_pages <<= compound_order(page);
3317 pc = lookup_page_cgroup(page);
3318 lock_page_cgroup(pc);
3319 if (PageCgroupUsed(pc)) {
3320 memcg = pc->mem_cgroup;
3321 css_get(&memcg->css);
3323 * At migrating an anonymous page, its mapcount goes down
3324 * to 0 and uncharge() will be called. But, even if it's fully
3325 * unmapped, migration may fail and this page has to be
3326 * charged again. We set MIGRATION flag here and delay uncharge
3327 * until end_migration() is called
3329 * Corner Case Thinking
3331 * When the old page was mapped as Anon and it's unmap-and-freed
3332 * while migration was ongoing.
3333 * If unmap finds the old page, uncharge() of it will be delayed
3334 * until end_migration(). If unmap finds a new page, it's
3335 * uncharged when it make mapcount to be 1->0. If unmap code
3336 * finds swap_migration_entry, the new page will not be mapped
3337 * and end_migration() will find it(mapcount==0).
3340 * When the old page was mapped but migraion fails, the kernel
3341 * remaps it. A charge for it is kept by MIGRATION flag even
3342 * if mapcount goes down to 0. We can do remap successfully
3343 * without charging it again.
3346 * The "old" page is under lock_page() until the end of
3347 * migration, so, the old page itself will not be swapped-out.
3348 * If the new page is swapped out before end_migraton, our
3349 * hook to usual swap-out path will catch the event.
3352 SetPageCgroupMigration(pc);
3354 unlock_page_cgroup(pc);
3356 * If the page is not charged at this point,
3364 * We charge new page before it's used/mapped. So, even if unlock_page()
3365 * is called before end_migration, we can catch all events on this new
3366 * page. In the case new page is migrated but not remapped, new page's
3367 * mapcount will be finally 0 and we call uncharge in end_migration().
3370 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
3372 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3374 * The page is committed to the memcg, but it's not actually
3375 * charged to the res_counter since we plan on replacing the
3376 * old one and only one page is going to be left afterwards.
3378 __mem_cgroup_commit_charge(memcg, newpage, nr_pages, ctype, false);
3381 /* remove redundant charge if migration failed*/
3382 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3383 struct page *oldpage, struct page *newpage, bool migration_ok)
3385 struct page *used, *unused;
3386 struct page_cgroup *pc;
3392 if (!migration_ok) {
3399 anon = PageAnon(used);
3400 __mem_cgroup_uncharge_common(unused,
3401 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
3402 : MEM_CGROUP_CHARGE_TYPE_CACHE,
3404 css_put(&memcg->css);
3406 * We disallowed uncharge of pages under migration because mapcount
3407 * of the page goes down to zero, temporarly.
3408 * Clear the flag and check the page should be charged.
3410 pc = lookup_page_cgroup(oldpage);
3411 lock_page_cgroup(pc);
3412 ClearPageCgroupMigration(pc);
3413 unlock_page_cgroup(pc);
3416 * If a page is a file cache, radix-tree replacement is very atomic
3417 * and we can skip this check. When it was an Anon page, its mapcount
3418 * goes down to 0. But because we added MIGRATION flage, it's not
3419 * uncharged yet. There are several case but page->mapcount check
3420 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3421 * check. (see prepare_charge() also)
3424 mem_cgroup_uncharge_page(used);
3428 * At replace page cache, newpage is not under any memcg but it's on
3429 * LRU. So, this function doesn't touch res_counter but handles LRU
3430 * in correct way. Both pages are locked so we cannot race with uncharge.
3432 void mem_cgroup_replace_page_cache(struct page *oldpage,
3433 struct page *newpage)
3435 struct mem_cgroup *memcg = NULL;
3436 struct page_cgroup *pc;
3437 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3439 if (mem_cgroup_disabled())
3442 pc = lookup_page_cgroup(oldpage);
3443 /* fix accounting on old pages */
3444 lock_page_cgroup(pc);
3445 if (PageCgroupUsed(pc)) {
3446 memcg = pc->mem_cgroup;
3447 mem_cgroup_charge_statistics(memcg, false, -1);
3448 ClearPageCgroupUsed(pc);
3450 unlock_page_cgroup(pc);
3453 * When called from shmem_replace_page(), in some cases the
3454 * oldpage has already been charged, and in some cases not.
3459 * Even if newpage->mapping was NULL before starting replacement,
3460 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3461 * LRU while we overwrite pc->mem_cgroup.
3463 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3466 #ifdef CONFIG_DEBUG_VM
3467 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3469 struct page_cgroup *pc;
3471 pc = lookup_page_cgroup(page);
3473 * Can be NULL while feeding pages into the page allocator for
3474 * the first time, i.e. during boot or memory hotplug;
3475 * or when mem_cgroup_disabled().
3477 if (likely(pc) && PageCgroupUsed(pc))
3482 bool mem_cgroup_bad_page_check(struct page *page)
3484 if (mem_cgroup_disabled())
3487 return lookup_page_cgroup_used(page) != NULL;
3490 void mem_cgroup_print_bad_page(struct page *page)
3492 struct page_cgroup *pc;
3494 pc = lookup_page_cgroup_used(page);
3496 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3497 pc, pc->flags, pc->mem_cgroup);
3502 static DEFINE_MUTEX(set_limit_mutex);
3504 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3505 unsigned long long val)
3508 u64 memswlimit, memlimit;
3510 int children = mem_cgroup_count_children(memcg);
3511 u64 curusage, oldusage;
3515 * For keeping hierarchical_reclaim simple, how long we should retry
3516 * is depends on callers. We set our retry-count to be function
3517 * of # of children which we should visit in this loop.
3519 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3521 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3524 while (retry_count) {
3525 if (signal_pending(current)) {
3530 * Rather than hide all in some function, I do this in
3531 * open coded manner. You see what this really does.
3532 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3534 mutex_lock(&set_limit_mutex);
3535 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3536 if (memswlimit < val) {
3538 mutex_unlock(&set_limit_mutex);
3542 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3546 ret = res_counter_set_limit(&memcg->res, val);
3548 if (memswlimit == val)
3549 memcg->memsw_is_minimum = true;
3551 memcg->memsw_is_minimum = false;
3553 mutex_unlock(&set_limit_mutex);
3558 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3559 MEM_CGROUP_RECLAIM_SHRINK);
3560 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3561 /* Usage is reduced ? */
3562 if (curusage >= oldusage)
3565 oldusage = curusage;
3567 if (!ret && enlarge)
3568 memcg_oom_recover(memcg);
3573 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3574 unsigned long long val)
3577 u64 memlimit, memswlimit, oldusage, curusage;
3578 int children = mem_cgroup_count_children(memcg);
3582 /* see mem_cgroup_resize_res_limit */
3583 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3584 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3585 while (retry_count) {
3586 if (signal_pending(current)) {
3591 * Rather than hide all in some function, I do this in
3592 * open coded manner. You see what this really does.
3593 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3595 mutex_lock(&set_limit_mutex);
3596 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3597 if (memlimit > val) {
3599 mutex_unlock(&set_limit_mutex);
3602 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3603 if (memswlimit < val)
3605 ret = res_counter_set_limit(&memcg->memsw, val);
3607 if (memlimit == val)
3608 memcg->memsw_is_minimum = true;
3610 memcg->memsw_is_minimum = false;
3612 mutex_unlock(&set_limit_mutex);
3617 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3618 MEM_CGROUP_RECLAIM_NOSWAP |
3619 MEM_CGROUP_RECLAIM_SHRINK);
3620 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3621 /* Usage is reduced ? */
3622 if (curusage >= oldusage)
3625 oldusage = curusage;
3627 if (!ret && enlarge)
3628 memcg_oom_recover(memcg);
3632 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3634 unsigned long *total_scanned)
3636 unsigned long nr_reclaimed = 0;
3637 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3638 unsigned long reclaimed;
3640 struct mem_cgroup_tree_per_zone *mctz;
3641 unsigned long long excess;
3642 unsigned long nr_scanned;
3647 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3649 * This loop can run a while, specially if mem_cgroup's continuously
3650 * keep exceeding their soft limit and putting the system under
3657 mz = mem_cgroup_largest_soft_limit_node(mctz);
3662 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3663 gfp_mask, &nr_scanned);
3664 nr_reclaimed += reclaimed;
3665 *total_scanned += nr_scanned;
3666 spin_lock(&mctz->lock);
3669 * If we failed to reclaim anything from this memory cgroup
3670 * it is time to move on to the next cgroup
3676 * Loop until we find yet another one.
3678 * By the time we get the soft_limit lock
3679 * again, someone might have aded the
3680 * group back on the RB tree. Iterate to
3681 * make sure we get a different mem.
3682 * mem_cgroup_largest_soft_limit_node returns
3683 * NULL if no other cgroup is present on
3687 __mem_cgroup_largest_soft_limit_node(mctz);
3689 css_put(&next_mz->memcg->css);
3690 else /* next_mz == NULL or other memcg */
3694 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3695 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3697 * One school of thought says that we should not add
3698 * back the node to the tree if reclaim returns 0.
3699 * But our reclaim could return 0, simply because due
3700 * to priority we are exposing a smaller subset of
3701 * memory to reclaim from. Consider this as a longer
3704 /* If excess == 0, no tree ops */
3705 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3706 spin_unlock(&mctz->lock);
3707 css_put(&mz->memcg->css);
3710 * Could not reclaim anything and there are no more
3711 * mem cgroups to try or we seem to be looping without
3712 * reclaiming anything.
3714 if (!nr_reclaimed &&
3716 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3718 } while (!nr_reclaimed);
3720 css_put(&next_mz->memcg->css);
3721 return nr_reclaimed;
3725 * mem_cgroup_force_empty_list - clears LRU of a group
3726 * @memcg: group to clear
3729 * @lru: lru to to clear
3731 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3732 * reclaim the pages page themselves - pages are moved to the parent (or root)
3735 static void mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3736 int node, int zid, enum lru_list lru)
3738 struct lruvec *lruvec;
3739 unsigned long flags;
3740 struct list_head *list;
3744 zone = &NODE_DATA(node)->node_zones[zid];
3745 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
3746 list = &lruvec->lists[lru];
3750 struct page_cgroup *pc;
3753 spin_lock_irqsave(&zone->lru_lock, flags);
3754 if (list_empty(list)) {
3755 spin_unlock_irqrestore(&zone->lru_lock, flags);
3758 page = list_entry(list->prev, struct page, lru);
3760 list_move(&page->lru, list);
3762 spin_unlock_irqrestore(&zone->lru_lock, flags);
3765 spin_unlock_irqrestore(&zone->lru_lock, flags);
3767 pc = lookup_page_cgroup(page);
3769 if (mem_cgroup_move_parent(page, pc, memcg)) {
3770 /* found lock contention or "pc" is obsolete. */
3775 } while (!list_empty(list));
3779 * make mem_cgroup's charge to be 0 if there is no task by moving
3780 * all the charges and pages to the parent.
3781 * This enables deleting this mem_cgroup.
3783 * Caller is responsible for holding css reference on the memcg.
3785 static void mem_cgroup_reparent_charges(struct mem_cgroup *memcg)
3790 /* This is for making all *used* pages to be on LRU. */
3791 lru_add_drain_all();
3792 drain_all_stock_sync(memcg);
3793 mem_cgroup_start_move(memcg);
3794 for_each_node_state(node, N_MEMORY) {
3795 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3798 mem_cgroup_force_empty_list(memcg,
3803 mem_cgroup_end_move(memcg);
3804 memcg_oom_recover(memcg);
3808 * This is a safety check because mem_cgroup_force_empty_list
3809 * could have raced with mem_cgroup_replace_page_cache callers
3810 * so the lru seemed empty but the page could have been added
3811 * right after the check. RES_USAGE should be safe as we always
3812 * charge before adding to the LRU.
3814 } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0);
3818 * Reclaims as many pages from the given memcg as possible and moves
3819 * the rest to the parent.
3821 * Caller is responsible for holding css reference for memcg.
3823 static int mem_cgroup_force_empty(struct mem_cgroup *memcg)
3825 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3826 struct cgroup *cgrp = memcg->css.cgroup;
3828 /* returns EBUSY if there is a task or if we come here twice. */
3829 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3832 /* we call try-to-free pages for make this cgroup empty */
3833 lru_add_drain_all();
3834 /* try to free all pages in this cgroup */
3835 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
3838 if (signal_pending(current))
3841 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3845 /* maybe some writeback is necessary */
3846 congestion_wait(BLK_RW_ASYNC, HZ/10);
3851 mem_cgroup_reparent_charges(memcg);
3856 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3858 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3861 if (mem_cgroup_is_root(memcg))
3863 css_get(&memcg->css);
3864 ret = mem_cgroup_force_empty(memcg);
3865 css_put(&memcg->css);
3871 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3873 return mem_cgroup_from_cont(cont)->use_hierarchy;
3876 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3880 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3881 struct cgroup *parent = cont->parent;
3882 struct mem_cgroup *parent_memcg = NULL;
3885 parent_memcg = mem_cgroup_from_cont(parent);
3889 if (memcg->use_hierarchy == val)
3893 * If parent's use_hierarchy is set, we can't make any modifications
3894 * in the child subtrees. If it is unset, then the change can
3895 * occur, provided the current cgroup has no children.
3897 * For the root cgroup, parent_mem is NULL, we allow value to be
3898 * set if there are no children.
3900 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3901 (val == 1 || val == 0)) {
3902 if (list_empty(&cont->children))
3903 memcg->use_hierarchy = val;
3916 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3917 enum mem_cgroup_stat_index idx)
3919 struct mem_cgroup *iter;
3922 /* Per-cpu values can be negative, use a signed accumulator */
3923 for_each_mem_cgroup_tree(iter, memcg)
3924 val += mem_cgroup_read_stat(iter, idx);
3926 if (val < 0) /* race ? */
3931 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3935 if (!mem_cgroup_is_root(memcg)) {
3937 return res_counter_read_u64(&memcg->res, RES_USAGE);
3939 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3942 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3943 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3946 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
3948 return val << PAGE_SHIFT;
3951 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
3952 struct file *file, char __user *buf,
3953 size_t nbytes, loff_t *ppos)
3955 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3961 type = MEMFILE_TYPE(cft->private);
3962 name = MEMFILE_ATTR(cft->private);
3964 if (!do_swap_account && type == _MEMSWAP)
3969 if (name == RES_USAGE)
3970 val = mem_cgroup_usage(memcg, false);
3972 val = res_counter_read_u64(&memcg->res, name);
3975 if (name == RES_USAGE)
3976 val = mem_cgroup_usage(memcg, true);
3978 val = res_counter_read_u64(&memcg->memsw, name);
3984 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
3985 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
3988 * The user of this function is...
3991 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3994 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3997 unsigned long long val;
4000 type = MEMFILE_TYPE(cft->private);
4001 name = MEMFILE_ATTR(cft->private);
4003 if (!do_swap_account && type == _MEMSWAP)
4008 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4012 /* This function does all necessary parse...reuse it */
4013 ret = res_counter_memparse_write_strategy(buffer, &val);
4017 ret = mem_cgroup_resize_limit(memcg, val);
4019 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4021 case RES_SOFT_LIMIT:
4022 ret = res_counter_memparse_write_strategy(buffer, &val);
4026 * For memsw, soft limits are hard to implement in terms
4027 * of semantics, for now, we support soft limits for
4028 * control without swap
4031 ret = res_counter_set_soft_limit(&memcg->res, val);
4036 ret = -EINVAL; /* should be BUG() ? */
4042 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4043 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4045 struct cgroup *cgroup;
4046 unsigned long long min_limit, min_memsw_limit, tmp;
4048 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4049 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4050 cgroup = memcg->css.cgroup;
4051 if (!memcg->use_hierarchy)
4054 while (cgroup->parent) {
4055 cgroup = cgroup->parent;
4056 memcg = mem_cgroup_from_cont(cgroup);
4057 if (!memcg->use_hierarchy)
4059 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4060 min_limit = min(min_limit, tmp);
4061 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4062 min_memsw_limit = min(min_memsw_limit, tmp);
4065 *mem_limit = min_limit;
4066 *memsw_limit = min_memsw_limit;
4069 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4071 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4075 type = MEMFILE_TYPE(event);
4076 name = MEMFILE_ATTR(event);
4078 if (!do_swap_account && type == _MEMSWAP)
4084 res_counter_reset_max(&memcg->res);
4086 res_counter_reset_max(&memcg->memsw);
4090 res_counter_reset_failcnt(&memcg->res);
4092 res_counter_reset_failcnt(&memcg->memsw);
4099 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4102 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4106 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4107 struct cftype *cft, u64 val)
4109 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4111 if (val >= (1 << NR_MOVE_TYPE))
4114 * We check this value several times in both in can_attach() and
4115 * attach(), so we need cgroup lock to prevent this value from being
4119 memcg->move_charge_at_immigrate = val;
4125 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4126 struct cftype *cft, u64 val)
4133 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
4137 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4138 unsigned long node_nr;
4139 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4141 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4142 seq_printf(m, "total=%lu", total_nr);
4143 for_each_node_state(nid, N_MEMORY) {
4144 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4145 seq_printf(m, " N%d=%lu", nid, node_nr);
4149 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4150 seq_printf(m, "file=%lu", file_nr);
4151 for_each_node_state(nid, N_MEMORY) {
4152 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4154 seq_printf(m, " N%d=%lu", nid, node_nr);
4158 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4159 seq_printf(m, "anon=%lu", anon_nr);
4160 for_each_node_state(nid, N_MEMORY) {
4161 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4163 seq_printf(m, " N%d=%lu", nid, node_nr);
4167 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4168 seq_printf(m, "unevictable=%lu", unevictable_nr);
4169 for_each_node_state(nid, N_MEMORY) {
4170 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4171 BIT(LRU_UNEVICTABLE));
4172 seq_printf(m, " N%d=%lu", nid, node_nr);
4177 #endif /* CONFIG_NUMA */
4179 static const char * const mem_cgroup_lru_names[] = {
4187 static inline void mem_cgroup_lru_names_not_uptodate(void)
4189 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4192 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
4195 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4196 struct mem_cgroup *mi;
4199 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4200 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4202 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4203 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4206 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4207 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4208 mem_cgroup_read_events(memcg, i));
4210 for (i = 0; i < NR_LRU_LISTS; i++)
4211 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4212 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4214 /* Hierarchical information */
4216 unsigned long long limit, memsw_limit;
4217 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4218 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
4219 if (do_swap_account)
4220 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4224 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4227 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4229 for_each_mem_cgroup_tree(mi, memcg)
4230 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4231 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4234 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4235 unsigned long long val = 0;
4237 for_each_mem_cgroup_tree(mi, memcg)
4238 val += mem_cgroup_read_events(mi, i);
4239 seq_printf(m, "total_%s %llu\n",
4240 mem_cgroup_events_names[i], val);
4243 for (i = 0; i < NR_LRU_LISTS; i++) {
4244 unsigned long long val = 0;
4246 for_each_mem_cgroup_tree(mi, memcg)
4247 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4248 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4251 #ifdef CONFIG_DEBUG_VM
4254 struct mem_cgroup_per_zone *mz;
4255 struct zone_reclaim_stat *rstat;
4256 unsigned long recent_rotated[2] = {0, 0};
4257 unsigned long recent_scanned[2] = {0, 0};
4259 for_each_online_node(nid)
4260 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4261 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4262 rstat = &mz->lruvec.reclaim_stat;
4264 recent_rotated[0] += rstat->recent_rotated[0];
4265 recent_rotated[1] += rstat->recent_rotated[1];
4266 recent_scanned[0] += rstat->recent_scanned[0];
4267 recent_scanned[1] += rstat->recent_scanned[1];
4269 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4270 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4271 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4272 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4279 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4281 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4283 return mem_cgroup_swappiness(memcg);
4286 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4289 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4290 struct mem_cgroup *parent;
4295 if (cgrp->parent == NULL)
4298 parent = mem_cgroup_from_cont(cgrp->parent);
4302 /* If under hierarchy, only empty-root can set this value */
4303 if ((parent->use_hierarchy) ||
4304 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4309 memcg->swappiness = val;
4316 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4318 struct mem_cgroup_threshold_ary *t;
4324 t = rcu_dereference(memcg->thresholds.primary);
4326 t = rcu_dereference(memcg->memsw_thresholds.primary);
4331 usage = mem_cgroup_usage(memcg, swap);
4334 * current_threshold points to threshold just below or equal to usage.
4335 * If it's not true, a threshold was crossed after last
4336 * call of __mem_cgroup_threshold().
4338 i = t->current_threshold;
4341 * Iterate backward over array of thresholds starting from
4342 * current_threshold and check if a threshold is crossed.
4343 * If none of thresholds below usage is crossed, we read
4344 * only one element of the array here.
4346 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4347 eventfd_signal(t->entries[i].eventfd, 1);
4349 /* i = current_threshold + 1 */
4353 * Iterate forward over array of thresholds starting from
4354 * current_threshold+1 and check if a threshold is crossed.
4355 * If none of thresholds above usage is crossed, we read
4356 * only one element of the array here.
4358 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4359 eventfd_signal(t->entries[i].eventfd, 1);
4361 /* Update current_threshold */
4362 t->current_threshold = i - 1;
4367 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4370 __mem_cgroup_threshold(memcg, false);
4371 if (do_swap_account)
4372 __mem_cgroup_threshold(memcg, true);
4374 memcg = parent_mem_cgroup(memcg);
4378 static int compare_thresholds(const void *a, const void *b)
4380 const struct mem_cgroup_threshold *_a = a;
4381 const struct mem_cgroup_threshold *_b = b;
4383 return _a->threshold - _b->threshold;
4386 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4388 struct mem_cgroup_eventfd_list *ev;
4390 list_for_each_entry(ev, &memcg->oom_notify, list)
4391 eventfd_signal(ev->eventfd, 1);
4395 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4397 struct mem_cgroup *iter;
4399 for_each_mem_cgroup_tree(iter, memcg)
4400 mem_cgroup_oom_notify_cb(iter);
4403 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4404 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4406 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4407 struct mem_cgroup_thresholds *thresholds;
4408 struct mem_cgroup_threshold_ary *new;
4409 enum res_type type = MEMFILE_TYPE(cft->private);
4410 u64 threshold, usage;
4413 ret = res_counter_memparse_write_strategy(args, &threshold);
4417 mutex_lock(&memcg->thresholds_lock);
4420 thresholds = &memcg->thresholds;
4421 else if (type == _MEMSWAP)
4422 thresholds = &memcg->memsw_thresholds;
4426 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4428 /* Check if a threshold crossed before adding a new one */
4429 if (thresholds->primary)
4430 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4432 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4434 /* Allocate memory for new array of thresholds */
4435 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4443 /* Copy thresholds (if any) to new array */
4444 if (thresholds->primary) {
4445 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4446 sizeof(struct mem_cgroup_threshold));
4449 /* Add new threshold */
4450 new->entries[size - 1].eventfd = eventfd;
4451 new->entries[size - 1].threshold = threshold;
4453 /* Sort thresholds. Registering of new threshold isn't time-critical */
4454 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4455 compare_thresholds, NULL);
4457 /* Find current threshold */
4458 new->current_threshold = -1;
4459 for (i = 0; i < size; i++) {
4460 if (new->entries[i].threshold <= usage) {
4462 * new->current_threshold will not be used until
4463 * rcu_assign_pointer(), so it's safe to increment
4466 ++new->current_threshold;
4471 /* Free old spare buffer and save old primary buffer as spare */
4472 kfree(thresholds->spare);
4473 thresholds->spare = thresholds->primary;
4475 rcu_assign_pointer(thresholds->primary, new);
4477 /* To be sure that nobody uses thresholds */
4481 mutex_unlock(&memcg->thresholds_lock);
4486 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4487 struct cftype *cft, struct eventfd_ctx *eventfd)
4489 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4490 struct mem_cgroup_thresholds *thresholds;
4491 struct mem_cgroup_threshold_ary *new;
4492 enum res_type type = MEMFILE_TYPE(cft->private);
4496 mutex_lock(&memcg->thresholds_lock);
4498 thresholds = &memcg->thresholds;
4499 else if (type == _MEMSWAP)
4500 thresholds = &memcg->memsw_thresholds;
4504 if (!thresholds->primary)
4507 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4509 /* Check if a threshold crossed before removing */
4510 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4512 /* Calculate new number of threshold */
4514 for (i = 0; i < thresholds->primary->size; i++) {
4515 if (thresholds->primary->entries[i].eventfd != eventfd)
4519 new = thresholds->spare;
4521 /* Set thresholds array to NULL if we don't have thresholds */
4530 /* Copy thresholds and find current threshold */
4531 new->current_threshold = -1;
4532 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4533 if (thresholds->primary->entries[i].eventfd == eventfd)
4536 new->entries[j] = thresholds->primary->entries[i];
4537 if (new->entries[j].threshold <= usage) {
4539 * new->current_threshold will not be used
4540 * until rcu_assign_pointer(), so it's safe to increment
4543 ++new->current_threshold;
4549 /* Swap primary and spare array */
4550 thresholds->spare = thresholds->primary;
4551 /* If all events are unregistered, free the spare array */
4553 kfree(thresholds->spare);
4554 thresholds->spare = NULL;
4557 rcu_assign_pointer(thresholds->primary, new);
4559 /* To be sure that nobody uses thresholds */
4562 mutex_unlock(&memcg->thresholds_lock);
4565 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4566 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4568 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4569 struct mem_cgroup_eventfd_list *event;
4570 enum res_type type = MEMFILE_TYPE(cft->private);
4572 BUG_ON(type != _OOM_TYPE);
4573 event = kmalloc(sizeof(*event), GFP_KERNEL);
4577 spin_lock(&memcg_oom_lock);
4579 event->eventfd = eventfd;
4580 list_add(&event->list, &memcg->oom_notify);
4582 /* already in OOM ? */
4583 if (atomic_read(&memcg->under_oom))
4584 eventfd_signal(eventfd, 1);
4585 spin_unlock(&memcg_oom_lock);
4590 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4591 struct cftype *cft, struct eventfd_ctx *eventfd)
4593 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4594 struct mem_cgroup_eventfd_list *ev, *tmp;
4595 enum res_type type = MEMFILE_TYPE(cft->private);
4597 BUG_ON(type != _OOM_TYPE);
4599 spin_lock(&memcg_oom_lock);
4601 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4602 if (ev->eventfd == eventfd) {
4603 list_del(&ev->list);
4608 spin_unlock(&memcg_oom_lock);
4611 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4612 struct cftype *cft, struct cgroup_map_cb *cb)
4614 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4616 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4618 if (atomic_read(&memcg->under_oom))
4619 cb->fill(cb, "under_oom", 1);
4621 cb->fill(cb, "under_oom", 0);
4625 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4626 struct cftype *cft, u64 val)
4628 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4629 struct mem_cgroup *parent;
4631 /* cannot set to root cgroup and only 0 and 1 are allowed */
4632 if (!cgrp->parent || !((val == 0) || (val == 1)))
4635 parent = mem_cgroup_from_cont(cgrp->parent);
4638 /* oom-kill-disable is a flag for subhierarchy. */
4639 if ((parent->use_hierarchy) ||
4640 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4644 memcg->oom_kill_disable = val;
4646 memcg_oom_recover(memcg);
4651 #ifdef CONFIG_MEMCG_KMEM
4652 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4654 return mem_cgroup_sockets_init(memcg, ss);
4657 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4659 mem_cgroup_sockets_destroy(memcg);
4662 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4667 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4672 static struct cftype mem_cgroup_files[] = {
4674 .name = "usage_in_bytes",
4675 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4676 .read = mem_cgroup_read,
4677 .register_event = mem_cgroup_usage_register_event,
4678 .unregister_event = mem_cgroup_usage_unregister_event,
4681 .name = "max_usage_in_bytes",
4682 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4683 .trigger = mem_cgroup_reset,
4684 .read = mem_cgroup_read,
4687 .name = "limit_in_bytes",
4688 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4689 .write_string = mem_cgroup_write,
4690 .read = mem_cgroup_read,
4693 .name = "soft_limit_in_bytes",
4694 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4695 .write_string = mem_cgroup_write,
4696 .read = mem_cgroup_read,
4700 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4701 .trigger = mem_cgroup_reset,
4702 .read = mem_cgroup_read,
4706 .read_seq_string = memcg_stat_show,
4709 .name = "force_empty",
4710 .trigger = mem_cgroup_force_empty_write,
4713 .name = "use_hierarchy",
4714 .write_u64 = mem_cgroup_hierarchy_write,
4715 .read_u64 = mem_cgroup_hierarchy_read,
4718 .name = "swappiness",
4719 .read_u64 = mem_cgroup_swappiness_read,
4720 .write_u64 = mem_cgroup_swappiness_write,
4723 .name = "move_charge_at_immigrate",
4724 .read_u64 = mem_cgroup_move_charge_read,
4725 .write_u64 = mem_cgroup_move_charge_write,
4728 .name = "oom_control",
4729 .read_map = mem_cgroup_oom_control_read,
4730 .write_u64 = mem_cgroup_oom_control_write,
4731 .register_event = mem_cgroup_oom_register_event,
4732 .unregister_event = mem_cgroup_oom_unregister_event,
4733 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4737 .name = "numa_stat",
4738 .read_seq_string = memcg_numa_stat_show,
4741 #ifdef CONFIG_MEMCG_SWAP
4743 .name = "memsw.usage_in_bytes",
4744 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4745 .read = mem_cgroup_read,
4746 .register_event = mem_cgroup_usage_register_event,
4747 .unregister_event = mem_cgroup_usage_unregister_event,
4750 .name = "memsw.max_usage_in_bytes",
4751 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4752 .trigger = mem_cgroup_reset,
4753 .read = mem_cgroup_read,
4756 .name = "memsw.limit_in_bytes",
4757 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4758 .write_string = mem_cgroup_write,
4759 .read = mem_cgroup_read,
4762 .name = "memsw.failcnt",
4763 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4764 .trigger = mem_cgroup_reset,
4765 .read = mem_cgroup_read,
4768 { }, /* terminate */
4771 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4773 struct mem_cgroup_per_node *pn;
4774 struct mem_cgroup_per_zone *mz;
4775 int zone, tmp = node;
4777 * This routine is called against possible nodes.
4778 * But it's BUG to call kmalloc() against offline node.
4780 * TODO: this routine can waste much memory for nodes which will
4781 * never be onlined. It's better to use memory hotplug callback
4784 if (!node_state(node, N_NORMAL_MEMORY))
4786 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4790 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4791 mz = &pn->zoneinfo[zone];
4792 lruvec_init(&mz->lruvec);
4793 mz->usage_in_excess = 0;
4794 mz->on_tree = false;
4797 memcg->info.nodeinfo[node] = pn;
4801 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4803 kfree(memcg->info.nodeinfo[node]);
4806 static struct mem_cgroup *mem_cgroup_alloc(void)
4808 struct mem_cgroup *memcg;
4809 int size = sizeof(struct mem_cgroup);
4811 /* Can be very big if MAX_NUMNODES is very big */
4812 if (size < PAGE_SIZE)
4813 memcg = kzalloc(size, GFP_KERNEL);
4815 memcg = vzalloc(size);
4820 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4823 spin_lock_init(&memcg->pcp_counter_lock);
4827 if (size < PAGE_SIZE)
4835 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
4836 * but in process context. The work_freeing structure is overlaid
4837 * on the rcu_freeing structure, which itself is overlaid on memsw.
4839 static void free_work(struct work_struct *work)
4841 struct mem_cgroup *memcg;
4842 int size = sizeof(struct mem_cgroup);
4844 memcg = container_of(work, struct mem_cgroup, work_freeing);
4846 * We need to make sure that (at least for now), the jump label
4847 * destruction code runs outside of the cgroup lock. This is because
4848 * get_online_cpus(), which is called from the static_branch update,
4849 * can't be called inside the cgroup_lock. cpusets are the ones
4850 * enforcing this dependency, so if they ever change, we might as well.
4852 * schedule_work() will guarantee this happens. Be careful if you need
4853 * to move this code around, and make sure it is outside
4856 disarm_sock_keys(memcg);
4857 if (size < PAGE_SIZE)
4863 static void free_rcu(struct rcu_head *rcu_head)
4865 struct mem_cgroup *memcg;
4867 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4868 INIT_WORK(&memcg->work_freeing, free_work);
4869 schedule_work(&memcg->work_freeing);
4873 * At destroying mem_cgroup, references from swap_cgroup can remain.
4874 * (scanning all at force_empty is too costly...)
4876 * Instead of clearing all references at force_empty, we remember
4877 * the number of reference from swap_cgroup and free mem_cgroup when
4878 * it goes down to 0.
4880 * Removal of cgroup itself succeeds regardless of refs from swap.
4883 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4887 mem_cgroup_remove_from_trees(memcg);
4888 free_css_id(&mem_cgroup_subsys, &memcg->css);
4891 free_mem_cgroup_per_zone_info(memcg, node);
4893 free_percpu(memcg->stat);
4894 call_rcu(&memcg->rcu_freeing, free_rcu);
4897 static void mem_cgroup_get(struct mem_cgroup *memcg)
4899 atomic_inc(&memcg->refcnt);
4902 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4904 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4905 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4906 __mem_cgroup_free(memcg);
4908 mem_cgroup_put(parent);
4912 static void mem_cgroup_put(struct mem_cgroup *memcg)
4914 __mem_cgroup_put(memcg, 1);
4918 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4920 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4922 if (!memcg->res.parent)
4924 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4926 EXPORT_SYMBOL(parent_mem_cgroup);
4928 #ifdef CONFIG_MEMCG_SWAP
4929 static void __init enable_swap_cgroup(void)
4931 if (!mem_cgroup_disabled() && really_do_swap_account)
4932 do_swap_account = 1;
4935 static void __init enable_swap_cgroup(void)
4940 static int mem_cgroup_soft_limit_tree_init(void)
4942 struct mem_cgroup_tree_per_node *rtpn;
4943 struct mem_cgroup_tree_per_zone *rtpz;
4944 int tmp, node, zone;
4946 for_each_node(node) {
4948 if (!node_state(node, N_NORMAL_MEMORY))
4950 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4954 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4956 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4957 rtpz = &rtpn->rb_tree_per_zone[zone];
4958 rtpz->rb_root = RB_ROOT;
4959 spin_lock_init(&rtpz->lock);
4965 for_each_node(node) {
4966 if (!soft_limit_tree.rb_tree_per_node[node])
4968 kfree(soft_limit_tree.rb_tree_per_node[node]);
4969 soft_limit_tree.rb_tree_per_node[node] = NULL;
4975 static struct cgroup_subsys_state * __ref
4976 mem_cgroup_css_alloc(struct cgroup *cont)
4978 struct mem_cgroup *memcg, *parent;
4979 long error = -ENOMEM;
4982 memcg = mem_cgroup_alloc();
4984 return ERR_PTR(error);
4987 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4991 if (cont->parent == NULL) {
4993 enable_swap_cgroup();
4995 if (mem_cgroup_soft_limit_tree_init())
4997 root_mem_cgroup = memcg;
4998 for_each_possible_cpu(cpu) {
4999 struct memcg_stock_pcp *stock =
5000 &per_cpu(memcg_stock, cpu);
5001 INIT_WORK(&stock->work, drain_local_stock);
5003 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5005 parent = mem_cgroup_from_cont(cont->parent);
5006 memcg->use_hierarchy = parent->use_hierarchy;
5007 memcg->oom_kill_disable = parent->oom_kill_disable;
5010 if (parent && parent->use_hierarchy) {
5011 res_counter_init(&memcg->res, &parent->res);
5012 res_counter_init(&memcg->memsw, &parent->memsw);
5014 * We increment refcnt of the parent to ensure that we can
5015 * safely access it on res_counter_charge/uncharge.
5016 * This refcnt will be decremented when freeing this
5017 * mem_cgroup(see mem_cgroup_put).
5019 mem_cgroup_get(parent);
5021 res_counter_init(&memcg->res, NULL);
5022 res_counter_init(&memcg->memsw, NULL);
5024 * Deeper hierachy with use_hierarchy == false doesn't make
5025 * much sense so let cgroup subsystem know about this
5026 * unfortunate state in our controller.
5028 if (parent && parent != root_mem_cgroup)
5029 mem_cgroup_subsys.broken_hierarchy = true;
5031 memcg->last_scanned_node = MAX_NUMNODES;
5032 INIT_LIST_HEAD(&memcg->oom_notify);
5035 memcg->swappiness = mem_cgroup_swappiness(parent);
5036 atomic_set(&memcg->refcnt, 1);
5037 memcg->move_charge_at_immigrate = 0;
5038 mutex_init(&memcg->thresholds_lock);
5039 spin_lock_init(&memcg->move_lock);
5041 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
5044 * We call put now because our (and parent's) refcnts
5045 * are already in place. mem_cgroup_put() will internally
5046 * call __mem_cgroup_free, so return directly
5048 mem_cgroup_put(memcg);
5049 return ERR_PTR(error);
5053 __mem_cgroup_free(memcg);
5054 return ERR_PTR(error);
5057 static void mem_cgroup_css_offline(struct cgroup *cont)
5059 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5061 mem_cgroup_reparent_charges(memcg);
5064 static void mem_cgroup_css_free(struct cgroup *cont)
5066 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5068 kmem_cgroup_destroy(memcg);
5070 mem_cgroup_put(memcg);
5074 /* Handlers for move charge at task migration. */
5075 #define PRECHARGE_COUNT_AT_ONCE 256
5076 static int mem_cgroup_do_precharge(unsigned long count)
5079 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5080 struct mem_cgroup *memcg = mc.to;
5082 if (mem_cgroup_is_root(memcg)) {
5083 mc.precharge += count;
5084 /* we don't need css_get for root */
5087 /* try to charge at once */
5089 struct res_counter *dummy;
5091 * "memcg" cannot be under rmdir() because we've already checked
5092 * by cgroup_lock_live_cgroup() that it is not removed and we
5093 * are still under the same cgroup_mutex. So we can postpone
5096 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5098 if (do_swap_account && res_counter_charge(&memcg->memsw,
5099 PAGE_SIZE * count, &dummy)) {
5100 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5103 mc.precharge += count;
5107 /* fall back to one by one charge */
5109 if (signal_pending(current)) {
5113 if (!batch_count--) {
5114 batch_count = PRECHARGE_COUNT_AT_ONCE;
5117 ret = __mem_cgroup_try_charge(NULL,
5118 GFP_KERNEL, 1, &memcg, false);
5120 /* mem_cgroup_clear_mc() will do uncharge later */
5128 * get_mctgt_type - get target type of moving charge
5129 * @vma: the vma the pte to be checked belongs
5130 * @addr: the address corresponding to the pte to be checked
5131 * @ptent: the pte to be checked
5132 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5135 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5136 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5137 * move charge. if @target is not NULL, the page is stored in target->page
5138 * with extra refcnt got(Callers should handle it).
5139 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5140 * target for charge migration. if @target is not NULL, the entry is stored
5143 * Called with pte lock held.
5150 enum mc_target_type {
5156 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5157 unsigned long addr, pte_t ptent)
5159 struct page *page = vm_normal_page(vma, addr, ptent);
5161 if (!page || !page_mapped(page))
5163 if (PageAnon(page)) {
5164 /* we don't move shared anon */
5167 } else if (!move_file())
5168 /* we ignore mapcount for file pages */
5170 if (!get_page_unless_zero(page))
5177 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5178 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5180 struct page *page = NULL;
5181 swp_entry_t ent = pte_to_swp_entry(ptent);
5183 if (!move_anon() || non_swap_entry(ent))
5186 * Because lookup_swap_cache() updates some statistics counter,
5187 * we call find_get_page() with swapper_space directly.
5189 page = find_get_page(&swapper_space, ent.val);
5190 if (do_swap_account)
5191 entry->val = ent.val;
5196 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5197 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5203 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5204 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5206 struct page *page = NULL;
5207 struct address_space *mapping;
5210 if (!vma->vm_file) /* anonymous vma */
5215 mapping = vma->vm_file->f_mapping;
5216 if (pte_none(ptent))
5217 pgoff = linear_page_index(vma, addr);
5218 else /* pte_file(ptent) is true */
5219 pgoff = pte_to_pgoff(ptent);
5221 /* page is moved even if it's not RSS of this task(page-faulted). */
5222 page = find_get_page(mapping, pgoff);
5225 /* shmem/tmpfs may report page out on swap: account for that too. */
5226 if (radix_tree_exceptional_entry(page)) {
5227 swp_entry_t swap = radix_to_swp_entry(page);
5228 if (do_swap_account)
5230 page = find_get_page(&swapper_space, swap.val);
5236 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5237 unsigned long addr, pte_t ptent, union mc_target *target)
5239 struct page *page = NULL;
5240 struct page_cgroup *pc;
5241 enum mc_target_type ret = MC_TARGET_NONE;
5242 swp_entry_t ent = { .val = 0 };
5244 if (pte_present(ptent))
5245 page = mc_handle_present_pte(vma, addr, ptent);
5246 else if (is_swap_pte(ptent))
5247 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5248 else if (pte_none(ptent) || pte_file(ptent))
5249 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5251 if (!page && !ent.val)
5254 pc = lookup_page_cgroup(page);
5256 * Do only loose check w/o page_cgroup lock.
5257 * mem_cgroup_move_account() checks the pc is valid or not under
5260 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5261 ret = MC_TARGET_PAGE;
5263 target->page = page;
5265 if (!ret || !target)
5268 /* There is a swap entry and a page doesn't exist or isn't charged */
5269 if (ent.val && !ret &&
5270 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5271 ret = MC_TARGET_SWAP;
5278 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5280 * We don't consider swapping or file mapped pages because THP does not
5281 * support them for now.
5282 * Caller should make sure that pmd_trans_huge(pmd) is true.
5284 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5285 unsigned long addr, pmd_t pmd, union mc_target *target)
5287 struct page *page = NULL;
5288 struct page_cgroup *pc;
5289 enum mc_target_type ret = MC_TARGET_NONE;
5291 page = pmd_page(pmd);
5292 VM_BUG_ON(!page || !PageHead(page));
5295 pc = lookup_page_cgroup(page);
5296 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5297 ret = MC_TARGET_PAGE;
5300 target->page = page;
5306 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5307 unsigned long addr, pmd_t pmd, union mc_target *target)
5309 return MC_TARGET_NONE;
5313 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5314 unsigned long addr, unsigned long end,
5315 struct mm_walk *walk)
5317 struct vm_area_struct *vma = walk->private;
5321 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5322 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5323 mc.precharge += HPAGE_PMD_NR;
5324 spin_unlock(&vma->vm_mm->page_table_lock);
5328 if (pmd_trans_unstable(pmd))
5330 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5331 for (; addr != end; pte++, addr += PAGE_SIZE)
5332 if (get_mctgt_type(vma, addr, *pte, NULL))
5333 mc.precharge++; /* increment precharge temporarily */
5334 pte_unmap_unlock(pte - 1, ptl);
5340 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5342 unsigned long precharge;
5343 struct vm_area_struct *vma;
5345 down_read(&mm->mmap_sem);
5346 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5347 struct mm_walk mem_cgroup_count_precharge_walk = {
5348 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5352 if (is_vm_hugetlb_page(vma))
5354 walk_page_range(vma->vm_start, vma->vm_end,
5355 &mem_cgroup_count_precharge_walk);
5357 up_read(&mm->mmap_sem);
5359 precharge = mc.precharge;
5365 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5367 unsigned long precharge = mem_cgroup_count_precharge(mm);
5369 VM_BUG_ON(mc.moving_task);
5370 mc.moving_task = current;
5371 return mem_cgroup_do_precharge(precharge);
5374 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5375 static void __mem_cgroup_clear_mc(void)
5377 struct mem_cgroup *from = mc.from;
5378 struct mem_cgroup *to = mc.to;
5380 /* we must uncharge all the leftover precharges from mc.to */
5382 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5386 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5387 * we must uncharge here.
5389 if (mc.moved_charge) {
5390 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5391 mc.moved_charge = 0;
5393 /* we must fixup refcnts and charges */
5394 if (mc.moved_swap) {
5395 /* uncharge swap account from the old cgroup */
5396 if (!mem_cgroup_is_root(mc.from))
5397 res_counter_uncharge(&mc.from->memsw,
5398 PAGE_SIZE * mc.moved_swap);
5399 __mem_cgroup_put(mc.from, mc.moved_swap);
5401 if (!mem_cgroup_is_root(mc.to)) {
5403 * we charged both to->res and to->memsw, so we should
5406 res_counter_uncharge(&mc.to->res,
5407 PAGE_SIZE * mc.moved_swap);
5409 /* we've already done mem_cgroup_get(mc.to) */
5412 memcg_oom_recover(from);
5413 memcg_oom_recover(to);
5414 wake_up_all(&mc.waitq);
5417 static void mem_cgroup_clear_mc(void)
5419 struct mem_cgroup *from = mc.from;
5422 * we must clear moving_task before waking up waiters at the end of
5425 mc.moving_task = NULL;
5426 __mem_cgroup_clear_mc();
5427 spin_lock(&mc.lock);
5430 spin_unlock(&mc.lock);
5431 mem_cgroup_end_move(from);
5434 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5435 struct cgroup_taskset *tset)
5437 struct task_struct *p = cgroup_taskset_first(tset);
5439 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5441 if (memcg->move_charge_at_immigrate) {
5442 struct mm_struct *mm;
5443 struct mem_cgroup *from = mem_cgroup_from_task(p);
5445 VM_BUG_ON(from == memcg);
5447 mm = get_task_mm(p);
5450 /* We move charges only when we move a owner of the mm */
5451 if (mm->owner == p) {
5454 VM_BUG_ON(mc.precharge);
5455 VM_BUG_ON(mc.moved_charge);
5456 VM_BUG_ON(mc.moved_swap);
5457 mem_cgroup_start_move(from);
5458 spin_lock(&mc.lock);
5461 spin_unlock(&mc.lock);
5462 /* We set mc.moving_task later */
5464 ret = mem_cgroup_precharge_mc(mm);
5466 mem_cgroup_clear_mc();
5473 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5474 struct cgroup_taskset *tset)
5476 mem_cgroup_clear_mc();
5479 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5480 unsigned long addr, unsigned long end,
5481 struct mm_walk *walk)
5484 struct vm_area_struct *vma = walk->private;
5487 enum mc_target_type target_type;
5488 union mc_target target;
5490 struct page_cgroup *pc;
5493 * We don't take compound_lock() here but no race with splitting thp
5495 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5496 * under splitting, which means there's no concurrent thp split,
5497 * - if another thread runs into split_huge_page() just after we
5498 * entered this if-block, the thread must wait for page table lock
5499 * to be unlocked in __split_huge_page_splitting(), where the main
5500 * part of thp split is not executed yet.
5502 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5503 if (mc.precharge < HPAGE_PMD_NR) {
5504 spin_unlock(&vma->vm_mm->page_table_lock);
5507 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5508 if (target_type == MC_TARGET_PAGE) {
5510 if (!isolate_lru_page(page)) {
5511 pc = lookup_page_cgroup(page);
5512 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5513 pc, mc.from, mc.to)) {
5514 mc.precharge -= HPAGE_PMD_NR;
5515 mc.moved_charge += HPAGE_PMD_NR;
5517 putback_lru_page(page);
5521 spin_unlock(&vma->vm_mm->page_table_lock);
5525 if (pmd_trans_unstable(pmd))
5528 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5529 for (; addr != end; addr += PAGE_SIZE) {
5530 pte_t ptent = *(pte++);
5536 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5537 case MC_TARGET_PAGE:
5539 if (isolate_lru_page(page))
5541 pc = lookup_page_cgroup(page);
5542 if (!mem_cgroup_move_account(page, 1, pc,
5545 /* we uncharge from mc.from later. */
5548 putback_lru_page(page);
5549 put: /* get_mctgt_type() gets the page */
5552 case MC_TARGET_SWAP:
5554 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5556 /* we fixup refcnts and charges later. */
5564 pte_unmap_unlock(pte - 1, ptl);
5569 * We have consumed all precharges we got in can_attach().
5570 * We try charge one by one, but don't do any additional
5571 * charges to mc.to if we have failed in charge once in attach()
5574 ret = mem_cgroup_do_precharge(1);
5582 static void mem_cgroup_move_charge(struct mm_struct *mm)
5584 struct vm_area_struct *vma;
5586 lru_add_drain_all();
5588 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5590 * Someone who are holding the mmap_sem might be waiting in
5591 * waitq. So we cancel all extra charges, wake up all waiters,
5592 * and retry. Because we cancel precharges, we might not be able
5593 * to move enough charges, but moving charge is a best-effort
5594 * feature anyway, so it wouldn't be a big problem.
5596 __mem_cgroup_clear_mc();
5600 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5602 struct mm_walk mem_cgroup_move_charge_walk = {
5603 .pmd_entry = mem_cgroup_move_charge_pte_range,
5607 if (is_vm_hugetlb_page(vma))
5609 ret = walk_page_range(vma->vm_start, vma->vm_end,
5610 &mem_cgroup_move_charge_walk);
5613 * means we have consumed all precharges and failed in
5614 * doing additional charge. Just abandon here.
5618 up_read(&mm->mmap_sem);
5621 static void mem_cgroup_move_task(struct cgroup *cont,
5622 struct cgroup_taskset *tset)
5624 struct task_struct *p = cgroup_taskset_first(tset);
5625 struct mm_struct *mm = get_task_mm(p);
5629 mem_cgroup_move_charge(mm);
5633 mem_cgroup_clear_mc();
5635 #else /* !CONFIG_MMU */
5636 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5637 struct cgroup_taskset *tset)
5641 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5642 struct cgroup_taskset *tset)
5645 static void mem_cgroup_move_task(struct cgroup *cont,
5646 struct cgroup_taskset *tset)
5651 struct cgroup_subsys mem_cgroup_subsys = {
5653 .subsys_id = mem_cgroup_subsys_id,
5654 .css_alloc = mem_cgroup_css_alloc,
5655 .css_offline = mem_cgroup_css_offline,
5656 .css_free = mem_cgroup_css_free,
5657 .can_attach = mem_cgroup_can_attach,
5658 .cancel_attach = mem_cgroup_cancel_attach,
5659 .attach = mem_cgroup_move_task,
5660 .base_cftypes = mem_cgroup_files,
5665 #ifdef CONFIG_MEMCG_SWAP
5666 static int __init enable_swap_account(char *s)
5668 /* consider enabled if no parameter or 1 is given */
5669 if (!strcmp(s, "1"))
5670 really_do_swap_account = 1;
5671 else if (!strcmp(s, "0"))
5672 really_do_swap_account = 0;
5675 __setup("swapaccount=", enable_swap_account);