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>
54 #include <net/tcp_memcontrol.h>
56 #include <asm/uaccess.h>
58 #include <trace/events/vmscan.h>
60 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
61 #define MEM_CGROUP_RECLAIM_RETRIES 5
62 struct mem_cgroup *root_mem_cgroup __read_mostly;
64 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
65 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
66 int do_swap_account __read_mostly;
68 /* for remember boot option*/
69 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
70 static int really_do_swap_account __initdata = 1;
72 static int really_do_swap_account __initdata = 0;
76 #define do_swap_account (0)
81 * Statistics for memory cgroup.
83 enum mem_cgroup_stat_index {
85 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
87 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
88 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
89 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
90 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
91 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
92 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */
93 MEM_CGROUP_STAT_NSTATS,
96 enum mem_cgroup_events_index {
97 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
98 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
99 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
100 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
101 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
102 MEM_CGROUP_EVENTS_NSTATS,
105 * Per memcg event counter is incremented at every pagein/pageout. With THP,
106 * it will be incremated by the number of pages. This counter is used for
107 * for trigger some periodic events. This is straightforward and better
108 * than using jiffies etc. to handle periodic memcg event.
110 enum mem_cgroup_events_target {
111 MEM_CGROUP_TARGET_THRESH,
112 MEM_CGROUP_TARGET_SOFTLIMIT,
113 MEM_CGROUP_TARGET_NUMAINFO,
116 #define THRESHOLDS_EVENTS_TARGET (128)
117 #define SOFTLIMIT_EVENTS_TARGET (1024)
118 #define NUMAINFO_EVENTS_TARGET (1024)
120 struct mem_cgroup_stat_cpu {
121 long count[MEM_CGROUP_STAT_NSTATS];
122 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
123 unsigned long targets[MEM_CGROUP_NTARGETS];
126 struct mem_cgroup_reclaim_iter {
127 /* css_id of the last scanned hierarchy member */
129 /* scan generation, increased every round-trip */
130 unsigned int generation;
134 * per-zone information in memory controller.
136 struct mem_cgroup_per_zone {
138 * spin_lock to protect the per cgroup LRU
140 struct list_head lists[NR_LRU_LISTS];
141 unsigned long count[NR_LRU_LISTS];
143 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
145 struct zone_reclaim_stat reclaim_stat;
146 struct rb_node tree_node; /* RB tree node */
147 unsigned long long usage_in_excess;/* Set to the value by which */
148 /* the soft limit is exceeded*/
150 struct mem_cgroup *mem; /* Back pointer, we cannot */
151 /* use container_of */
153 /* Macro for accessing counter */
154 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
156 struct mem_cgroup_per_node {
157 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
160 struct mem_cgroup_lru_info {
161 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
165 * Cgroups above their limits are maintained in a RB-Tree, independent of
166 * their hierarchy representation
169 struct mem_cgroup_tree_per_zone {
170 struct rb_root rb_root;
174 struct mem_cgroup_tree_per_node {
175 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
178 struct mem_cgroup_tree {
179 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
182 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
184 struct mem_cgroup_threshold {
185 struct eventfd_ctx *eventfd;
190 struct mem_cgroup_threshold_ary {
191 /* An array index points to threshold just below usage. */
192 int current_threshold;
193 /* Size of entries[] */
195 /* Array of thresholds */
196 struct mem_cgroup_threshold entries[0];
199 struct mem_cgroup_thresholds {
200 /* Primary thresholds array */
201 struct mem_cgroup_threshold_ary *primary;
203 * Spare threshold array.
204 * This is needed to make mem_cgroup_unregister_event() "never fail".
205 * It must be able to store at least primary->size - 1 entries.
207 struct mem_cgroup_threshold_ary *spare;
211 struct mem_cgroup_eventfd_list {
212 struct list_head list;
213 struct eventfd_ctx *eventfd;
216 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
217 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
220 * The memory controller data structure. The memory controller controls both
221 * page cache and RSS per cgroup. We would eventually like to provide
222 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
223 * to help the administrator determine what knobs to tune.
225 * TODO: Add a water mark for the memory controller. Reclaim will begin when
226 * we hit the water mark. May be even add a low water mark, such that
227 * no reclaim occurs from a cgroup at it's low water mark, this is
228 * a feature that will be implemented much later in the future.
231 struct cgroup_subsys_state css;
233 * the counter to account for memory usage
235 struct res_counter res;
237 * the counter to account for mem+swap usage.
239 struct res_counter memsw;
241 * Per cgroup active and inactive list, similar to the
242 * per zone LRU lists.
244 struct mem_cgroup_lru_info info;
245 int last_scanned_node;
247 nodemask_t scan_nodes;
248 atomic_t numainfo_events;
249 atomic_t numainfo_updating;
252 * Should the accounting and control be hierarchical, per subtree?
262 /* OOM-Killer disable */
263 int oom_kill_disable;
265 /* set when res.limit == memsw.limit */
266 bool memsw_is_minimum;
268 /* protect arrays of thresholds */
269 struct mutex thresholds_lock;
271 /* thresholds for memory usage. RCU-protected */
272 struct mem_cgroup_thresholds thresholds;
274 /* thresholds for mem+swap usage. RCU-protected */
275 struct mem_cgroup_thresholds memsw_thresholds;
277 /* For oom notifier event fd */
278 struct list_head oom_notify;
281 * Should we move charges of a task when a task is moved into this
282 * mem_cgroup ? And what type of charges should we move ?
284 unsigned long move_charge_at_immigrate;
288 struct mem_cgroup_stat_cpu *stat;
290 * used when a cpu is offlined or other synchronizations
291 * See mem_cgroup_read_stat().
293 struct mem_cgroup_stat_cpu nocpu_base;
294 spinlock_t pcp_counter_lock;
297 struct tcp_memcontrol tcp_mem;
301 /* Stuffs for move charges at task migration. */
303 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
304 * left-shifted bitmap of these types.
307 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
308 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
312 /* "mc" and its members are protected by cgroup_mutex */
313 static struct move_charge_struct {
314 spinlock_t lock; /* for from, to */
315 struct mem_cgroup *from;
316 struct mem_cgroup *to;
317 unsigned long precharge;
318 unsigned long moved_charge;
319 unsigned long moved_swap;
320 struct task_struct *moving_task; /* a task moving charges */
321 wait_queue_head_t waitq; /* a waitq for other context */
323 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
324 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
327 static bool move_anon(void)
329 return test_bit(MOVE_CHARGE_TYPE_ANON,
330 &mc.to->move_charge_at_immigrate);
333 static bool move_file(void)
335 return test_bit(MOVE_CHARGE_TYPE_FILE,
336 &mc.to->move_charge_at_immigrate);
340 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
341 * limit reclaim to prevent infinite loops, if they ever occur.
343 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
344 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
347 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
348 MEM_CGROUP_CHARGE_TYPE_MAPPED,
349 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
350 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
351 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
352 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
356 /* for encoding cft->private value on file */
359 #define _OOM_TYPE (2)
360 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
361 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
362 #define MEMFILE_ATTR(val) ((val) & 0xffff)
363 /* Used for OOM nofiier */
364 #define OOM_CONTROL (0)
367 * Reclaim flags for mem_cgroup_hierarchical_reclaim
369 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
370 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
371 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
372 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
373 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
374 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
376 static void mem_cgroup_get(struct mem_cgroup *memcg);
377 static void mem_cgroup_put(struct mem_cgroup *memcg);
379 /* Writing them here to avoid exposing memcg's inner layout */
380 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
382 #include <net/sock.h>
385 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
386 void sock_update_memcg(struct sock *sk)
388 if (static_branch(&memcg_socket_limit_enabled)) {
389 struct mem_cgroup *memcg;
391 BUG_ON(!sk->sk_prot->proto_cgroup);
393 /* Socket cloning can throw us here with sk_cgrp already
394 * filled. It won't however, necessarily happen from
395 * process context. So the test for root memcg given
396 * the current task's memcg won't help us in this case.
398 * Respecting the original socket's memcg is a better
399 * decision in this case.
402 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
403 mem_cgroup_get(sk->sk_cgrp->memcg);
408 memcg = mem_cgroup_from_task(current);
409 if (!mem_cgroup_is_root(memcg)) {
410 mem_cgroup_get(memcg);
411 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
416 EXPORT_SYMBOL(sock_update_memcg);
418 void sock_release_memcg(struct sock *sk)
420 if (static_branch(&memcg_socket_limit_enabled) && sk->sk_cgrp) {
421 struct mem_cgroup *memcg;
422 WARN_ON(!sk->sk_cgrp->memcg);
423 memcg = sk->sk_cgrp->memcg;
424 mem_cgroup_put(memcg);
428 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
430 if (!memcg || mem_cgroup_is_root(memcg))
433 return &memcg->tcp_mem.cg_proto;
435 EXPORT_SYMBOL(tcp_proto_cgroup);
436 #endif /* CONFIG_INET */
437 #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
439 static void drain_all_stock_async(struct mem_cgroup *memcg);
441 static struct mem_cgroup_per_zone *
442 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
444 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
447 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
452 static struct mem_cgroup_per_zone *
453 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
455 int nid = page_to_nid(page);
456 int zid = page_zonenum(page);
458 return mem_cgroup_zoneinfo(memcg, nid, zid);
461 static struct mem_cgroup_tree_per_zone *
462 soft_limit_tree_node_zone(int nid, int zid)
464 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
467 static struct mem_cgroup_tree_per_zone *
468 soft_limit_tree_from_page(struct page *page)
470 int nid = page_to_nid(page);
471 int zid = page_zonenum(page);
473 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
477 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
478 struct mem_cgroup_per_zone *mz,
479 struct mem_cgroup_tree_per_zone *mctz,
480 unsigned long long new_usage_in_excess)
482 struct rb_node **p = &mctz->rb_root.rb_node;
483 struct rb_node *parent = NULL;
484 struct mem_cgroup_per_zone *mz_node;
489 mz->usage_in_excess = new_usage_in_excess;
490 if (!mz->usage_in_excess)
494 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
496 if (mz->usage_in_excess < mz_node->usage_in_excess)
499 * We can't avoid mem cgroups that are over their soft
500 * limit by the same amount
502 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
505 rb_link_node(&mz->tree_node, parent, p);
506 rb_insert_color(&mz->tree_node, &mctz->rb_root);
511 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
512 struct mem_cgroup_per_zone *mz,
513 struct mem_cgroup_tree_per_zone *mctz)
517 rb_erase(&mz->tree_node, &mctz->rb_root);
522 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
523 struct mem_cgroup_per_zone *mz,
524 struct mem_cgroup_tree_per_zone *mctz)
526 spin_lock(&mctz->lock);
527 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
528 spin_unlock(&mctz->lock);
532 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
534 unsigned long long excess;
535 struct mem_cgroup_per_zone *mz;
536 struct mem_cgroup_tree_per_zone *mctz;
537 int nid = page_to_nid(page);
538 int zid = page_zonenum(page);
539 mctz = soft_limit_tree_from_page(page);
542 * Necessary to update all ancestors when hierarchy is used.
543 * because their event counter is not touched.
545 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
546 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
547 excess = res_counter_soft_limit_excess(&memcg->res);
549 * We have to update the tree if mz is on RB-tree or
550 * mem is over its softlimit.
552 if (excess || mz->on_tree) {
553 spin_lock(&mctz->lock);
554 /* if on-tree, remove it */
556 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
558 * Insert again. mz->usage_in_excess will be updated.
559 * If excess is 0, no tree ops.
561 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
562 spin_unlock(&mctz->lock);
567 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
570 struct mem_cgroup_per_zone *mz;
571 struct mem_cgroup_tree_per_zone *mctz;
573 for_each_node_state(node, N_POSSIBLE) {
574 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
575 mz = mem_cgroup_zoneinfo(memcg, node, zone);
576 mctz = soft_limit_tree_node_zone(node, zone);
577 mem_cgroup_remove_exceeded(memcg, mz, mctz);
582 static struct mem_cgroup_per_zone *
583 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
585 struct rb_node *rightmost = NULL;
586 struct mem_cgroup_per_zone *mz;
590 rightmost = rb_last(&mctz->rb_root);
592 goto done; /* Nothing to reclaim from */
594 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
596 * Remove the node now but someone else can add it back,
597 * we will to add it back at the end of reclaim to its correct
598 * position in the tree.
600 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
601 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
602 !css_tryget(&mz->mem->css))
608 static struct mem_cgroup_per_zone *
609 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
611 struct mem_cgroup_per_zone *mz;
613 spin_lock(&mctz->lock);
614 mz = __mem_cgroup_largest_soft_limit_node(mctz);
615 spin_unlock(&mctz->lock);
620 * Implementation Note: reading percpu statistics for memcg.
622 * Both of vmstat[] and percpu_counter has threshold and do periodic
623 * synchronization to implement "quick" read. There are trade-off between
624 * reading cost and precision of value. Then, we may have a chance to implement
625 * a periodic synchronizion of counter in memcg's counter.
627 * But this _read() function is used for user interface now. The user accounts
628 * memory usage by memory cgroup and he _always_ requires exact value because
629 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
630 * have to visit all online cpus and make sum. So, for now, unnecessary
631 * synchronization is not implemented. (just implemented for cpu hotplug)
633 * If there are kernel internal actions which can make use of some not-exact
634 * value, and reading all cpu value can be performance bottleneck in some
635 * common workload, threashold and synchonization as vmstat[] should be
638 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
639 enum mem_cgroup_stat_index idx)
645 for_each_online_cpu(cpu)
646 val += per_cpu(memcg->stat->count[idx], cpu);
647 #ifdef CONFIG_HOTPLUG_CPU
648 spin_lock(&memcg->pcp_counter_lock);
649 val += memcg->nocpu_base.count[idx];
650 spin_unlock(&memcg->pcp_counter_lock);
656 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
659 int val = (charge) ? 1 : -1;
660 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
663 void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val)
665 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
668 void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val)
670 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
673 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
674 enum mem_cgroup_events_index idx)
676 unsigned long val = 0;
679 for_each_online_cpu(cpu)
680 val += per_cpu(memcg->stat->events[idx], cpu);
681 #ifdef CONFIG_HOTPLUG_CPU
682 spin_lock(&memcg->pcp_counter_lock);
683 val += memcg->nocpu_base.events[idx];
684 spin_unlock(&memcg->pcp_counter_lock);
689 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
690 bool file, int nr_pages)
695 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
698 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
701 /* pagein of a big page is an event. So, ignore page size */
703 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
705 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
706 nr_pages = -nr_pages; /* for event */
709 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
715 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
716 unsigned int lru_mask)
718 struct mem_cgroup_per_zone *mz;
720 unsigned long ret = 0;
722 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
725 if (BIT(l) & lru_mask)
726 ret += MEM_CGROUP_ZSTAT(mz, l);
732 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
733 int nid, unsigned int lru_mask)
738 for (zid = 0; zid < MAX_NR_ZONES; zid++)
739 total += mem_cgroup_zone_nr_lru_pages(memcg,
745 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
746 unsigned int lru_mask)
751 for_each_node_state(nid, N_HIGH_MEMORY)
752 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
756 static bool __memcg_event_check(struct mem_cgroup *memcg, int target)
758 unsigned long val, next;
760 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
761 next = __this_cpu_read(memcg->stat->targets[target]);
762 /* from time_after() in jiffies.h */
763 return ((long)next - (long)val < 0);
766 static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
768 unsigned long val, next;
770 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
773 case MEM_CGROUP_TARGET_THRESH:
774 next = val + THRESHOLDS_EVENTS_TARGET;
776 case MEM_CGROUP_TARGET_SOFTLIMIT:
777 next = val + SOFTLIMIT_EVENTS_TARGET;
779 case MEM_CGROUP_TARGET_NUMAINFO:
780 next = val + NUMAINFO_EVENTS_TARGET;
786 __this_cpu_write(memcg->stat->targets[target], next);
790 * Check events in order.
793 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
796 /* threshold event is triggered in finer grain than soft limit */
797 if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
798 mem_cgroup_threshold(memcg);
799 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
800 if (unlikely(__memcg_event_check(memcg,
801 MEM_CGROUP_TARGET_SOFTLIMIT))) {
802 mem_cgroup_update_tree(memcg, page);
803 __mem_cgroup_target_update(memcg,
804 MEM_CGROUP_TARGET_SOFTLIMIT);
807 if (unlikely(__memcg_event_check(memcg,
808 MEM_CGROUP_TARGET_NUMAINFO))) {
809 atomic_inc(&memcg->numainfo_events);
810 __mem_cgroup_target_update(memcg,
811 MEM_CGROUP_TARGET_NUMAINFO);
818 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
820 return container_of(cgroup_subsys_state(cont,
821 mem_cgroup_subsys_id), struct mem_cgroup,
825 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
828 * mm_update_next_owner() may clear mm->owner to NULL
829 * if it races with swapoff, page migration, etc.
830 * So this can be called with p == NULL.
835 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
836 struct mem_cgroup, css);
839 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
841 struct mem_cgroup *memcg = NULL;
846 * Because we have no locks, mm->owner's may be being moved to other
847 * cgroup. We use css_tryget() here even if this looks
848 * pessimistic (rather than adding locks here).
852 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
853 if (unlikely(!memcg))
855 } while (!css_tryget(&memcg->css));
860 struct mem_cgroup_reclaim_cookie {
863 unsigned int generation;
866 static struct mem_cgroup *
867 mem_cgroup_iter(struct mem_cgroup *root,
868 struct mem_cgroup *prev,
869 struct mem_cgroup_reclaim_cookie *reclaim)
871 struct mem_cgroup *memcg = NULL;
875 root = root_mem_cgroup;
877 if (prev && !reclaim)
878 id = css_id(&prev->css);
880 if (prev && prev != root)
883 if (!root->use_hierarchy && root != root_mem_cgroup) {
890 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
891 struct cgroup_subsys_state *css;
894 int nid = zone_to_nid(reclaim->zone);
895 int zid = zone_idx(reclaim->zone);
896 struct mem_cgroup_per_zone *mz;
898 mz = mem_cgroup_zoneinfo(root, nid, zid);
899 iter = &mz->reclaim_iter[reclaim->priority];
900 if (prev && reclaim->generation != iter->generation)
906 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
908 if (css == &root->css || css_tryget(css))
909 memcg = container_of(css,
910 struct mem_cgroup, css);
919 else if (!prev && memcg)
920 reclaim->generation = iter->generation;
929 static void mem_cgroup_iter_break(struct mem_cgroup *root,
930 struct mem_cgroup *prev)
933 root = root_mem_cgroup;
934 if (prev && prev != root)
939 * Iteration constructs for visiting all cgroups (under a tree). If
940 * loops are exited prematurely (break), mem_cgroup_iter_break() must
941 * be used for reference counting.
943 #define for_each_mem_cgroup_tree(iter, root) \
944 for (iter = mem_cgroup_iter(root, NULL, NULL); \
946 iter = mem_cgroup_iter(root, iter, NULL))
948 #define for_each_mem_cgroup(iter) \
949 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
951 iter = mem_cgroup_iter(NULL, iter, NULL))
953 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
955 return (memcg == root_mem_cgroup);
958 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
960 struct mem_cgroup *memcg;
966 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
967 if (unlikely(!memcg))
972 mem_cgroup_pgmajfault(memcg, 1);
975 mem_cgroup_pgfault(memcg, 1);
983 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
986 * Following LRU functions are allowed to be used without PCG_LOCK.
987 * Operations are called by routine of global LRU independently from memcg.
988 * What we have to take care of here is validness of pc->mem_cgroup.
990 * Changes to pc->mem_cgroup happens when
993 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
994 * It is added to LRU before charge.
995 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
996 * When moving account, the page is not on LRU. It's isolated.
999 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
1001 struct page_cgroup *pc;
1002 struct mem_cgroup_per_zone *mz;
1004 if (mem_cgroup_disabled())
1006 pc = lookup_page_cgroup(page);
1007 /* can happen while we handle swapcache. */
1008 if (!TestClearPageCgroupAcctLRU(pc))
1010 VM_BUG_ON(!pc->mem_cgroup);
1012 * We don't check PCG_USED bit. It's cleared when the "page" is finally
1013 * removed from global LRU.
1015 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1016 /* huge page split is done under lru_lock. so, we have no races. */
1017 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
1018 if (mem_cgroup_is_root(pc->mem_cgroup))
1020 VM_BUG_ON(list_empty(&pc->lru));
1021 list_del_init(&pc->lru);
1024 void mem_cgroup_del_lru(struct page *page)
1026 mem_cgroup_del_lru_list(page, page_lru(page));
1030 * Writeback is about to end against a page which has been marked for immediate
1031 * reclaim. If it still appears to be reclaimable, move it to the tail of the
1034 void mem_cgroup_rotate_reclaimable_page(struct page *page)
1036 struct mem_cgroup_per_zone *mz;
1037 struct page_cgroup *pc;
1038 enum lru_list lru = page_lru(page);
1040 if (mem_cgroup_disabled())
1043 pc = lookup_page_cgroup(page);
1044 /* unused or root page is not rotated. */
1045 if (!PageCgroupUsed(pc))
1047 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1049 if (mem_cgroup_is_root(pc->mem_cgroup))
1051 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1052 list_move_tail(&pc->lru, &mz->lists[lru]);
1055 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1057 struct mem_cgroup_per_zone *mz;
1058 struct page_cgroup *pc;
1060 if (mem_cgroup_disabled())
1063 pc = lookup_page_cgroup(page);
1064 /* unused or root page is not rotated. */
1065 if (!PageCgroupUsed(pc))
1067 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1069 if (mem_cgroup_is_root(pc->mem_cgroup))
1071 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1072 list_move(&pc->lru, &mz->lists[lru]);
1075 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1077 struct page_cgroup *pc;
1078 struct mem_cgroup_per_zone *mz;
1080 if (mem_cgroup_disabled())
1082 pc = lookup_page_cgroup(page);
1083 VM_BUG_ON(PageCgroupAcctLRU(pc));
1086 * SetPageLRU SetPageCgroupUsed
1088 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1090 * Ensure that one of the two sides adds the page to the memcg
1091 * LRU during a race.
1094 if (!PageCgroupUsed(pc))
1096 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1098 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1099 /* huge page split is done under lru_lock. so, we have no races. */
1100 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1101 SetPageCgroupAcctLRU(pc);
1102 if (mem_cgroup_is_root(pc->mem_cgroup))
1104 list_add(&pc->lru, &mz->lists[lru]);
1108 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1109 * while it's linked to lru because the page may be reused after it's fully
1110 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1111 * It's done under lock_page and expected that zone->lru_lock isnever held.
1113 static void mem_cgroup_lru_del_before_commit(struct page *page)
1115 unsigned long flags;
1116 struct zone *zone = page_zone(page);
1117 struct page_cgroup *pc = lookup_page_cgroup(page);
1120 * Doing this check without taking ->lru_lock seems wrong but this
1121 * is safe. Because if page_cgroup's USED bit is unset, the page
1122 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1123 * set, the commit after this will fail, anyway.
1124 * This all charge/uncharge is done under some mutual execustion.
1125 * So, we don't need to taking care of changes in USED bit.
1127 if (likely(!PageLRU(page)))
1130 spin_lock_irqsave(&zone->lru_lock, flags);
1132 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1133 * is guarded by lock_page() because the page is SwapCache.
1135 if (!PageCgroupUsed(pc))
1136 mem_cgroup_del_lru_list(page, page_lru(page));
1137 spin_unlock_irqrestore(&zone->lru_lock, flags);
1140 static void mem_cgroup_lru_add_after_commit(struct page *page)
1142 unsigned long flags;
1143 struct zone *zone = page_zone(page);
1144 struct page_cgroup *pc = lookup_page_cgroup(page);
1147 * SetPageLRU SetPageCgroupUsed
1149 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1151 * Ensure that one of the two sides adds the page to the memcg
1152 * LRU during a race.
1155 /* taking care of that the page is added to LRU while we commit it */
1156 if (likely(!PageLRU(page)))
1158 spin_lock_irqsave(&zone->lru_lock, flags);
1159 /* link when the page is linked to LRU but page_cgroup isn't */
1160 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1161 mem_cgroup_add_lru_list(page, page_lru(page));
1162 spin_unlock_irqrestore(&zone->lru_lock, flags);
1166 void mem_cgroup_move_lists(struct page *page,
1167 enum lru_list from, enum lru_list to)
1169 if (mem_cgroup_disabled())
1171 mem_cgroup_del_lru_list(page, from);
1172 mem_cgroup_add_lru_list(page, to);
1176 * Checks whether given mem is same or in the root_mem_cgroup's
1179 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1180 struct mem_cgroup *memcg)
1182 if (root_memcg != memcg) {
1183 return (root_memcg->use_hierarchy &&
1184 css_is_ancestor(&memcg->css, &root_memcg->css));
1190 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1193 struct mem_cgroup *curr = NULL;
1194 struct task_struct *p;
1196 p = find_lock_task_mm(task);
1199 curr = try_get_mem_cgroup_from_mm(p->mm);
1204 * We should check use_hierarchy of "memcg" not "curr". Because checking
1205 * use_hierarchy of "curr" here make this function true if hierarchy is
1206 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1207 * hierarchy(even if use_hierarchy is disabled in "memcg").
1209 ret = mem_cgroup_same_or_subtree(memcg, curr);
1210 css_put(&curr->css);
1214 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1216 unsigned long inactive_ratio;
1217 int nid = zone_to_nid(zone);
1218 int zid = zone_idx(zone);
1219 unsigned long inactive;
1220 unsigned long active;
1223 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1224 BIT(LRU_INACTIVE_ANON));
1225 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1226 BIT(LRU_ACTIVE_ANON));
1228 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1230 inactive_ratio = int_sqrt(10 * gb);
1234 return inactive * inactive_ratio < active;
1237 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1239 unsigned long active;
1240 unsigned long inactive;
1241 int zid = zone_idx(zone);
1242 int nid = zone_to_nid(zone);
1244 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1245 BIT(LRU_INACTIVE_FILE));
1246 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1247 BIT(LRU_ACTIVE_FILE));
1249 return (active > inactive);
1252 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1255 int nid = zone_to_nid(zone);
1256 int zid = zone_idx(zone);
1257 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1259 return &mz->reclaim_stat;
1262 struct zone_reclaim_stat *
1263 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1265 struct page_cgroup *pc;
1266 struct mem_cgroup_per_zone *mz;
1268 if (mem_cgroup_disabled())
1271 pc = lookup_page_cgroup(page);
1272 if (!PageCgroupUsed(pc))
1274 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1276 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1277 return &mz->reclaim_stat;
1280 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1281 struct list_head *dst,
1282 unsigned long *scanned, int order,
1283 isolate_mode_t mode,
1285 struct mem_cgroup *mem_cont,
1286 int active, int file)
1288 unsigned long nr_taken = 0;
1292 struct list_head *src;
1293 struct page_cgroup *pc, *tmp;
1294 int nid = zone_to_nid(z);
1295 int zid = zone_idx(z);
1296 struct mem_cgroup_per_zone *mz;
1297 int lru = LRU_FILE * file + active;
1301 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1302 src = &mz->lists[lru];
1305 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1306 if (scan >= nr_to_scan)
1309 if (unlikely(!PageCgroupUsed(pc)))
1312 page = lookup_cgroup_page(pc);
1314 if (unlikely(!PageLRU(page)))
1318 ret = __isolate_lru_page(page, mode, file);
1321 list_move(&page->lru, dst);
1322 mem_cgroup_del_lru(page);
1323 nr_taken += hpage_nr_pages(page);
1326 /* we don't affect global LRU but rotate in our LRU */
1327 mem_cgroup_rotate_lru_list(page, page_lru(page));
1336 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1342 #define mem_cgroup_from_res_counter(counter, member) \
1343 container_of(counter, struct mem_cgroup, member)
1346 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1347 * @mem: the memory cgroup
1349 * Returns the maximum amount of memory @mem can be charged with, in
1352 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1354 unsigned long long margin;
1356 margin = res_counter_margin(&memcg->res);
1357 if (do_swap_account)
1358 margin = min(margin, res_counter_margin(&memcg->memsw));
1359 return margin >> PAGE_SHIFT;
1362 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1364 struct cgroup *cgrp = memcg->css.cgroup;
1367 if (cgrp->parent == NULL)
1368 return vm_swappiness;
1370 return memcg->swappiness;
1373 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1378 spin_lock(&memcg->pcp_counter_lock);
1379 for_each_online_cpu(cpu)
1380 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1381 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1382 spin_unlock(&memcg->pcp_counter_lock);
1388 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1395 spin_lock(&memcg->pcp_counter_lock);
1396 for_each_online_cpu(cpu)
1397 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1398 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1399 spin_unlock(&memcg->pcp_counter_lock);
1403 * 2 routines for checking "mem" is under move_account() or not.
1405 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1406 * for avoiding race in accounting. If true,
1407 * pc->mem_cgroup may be overwritten.
1409 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1410 * under hierarchy of moving cgroups. This is for
1411 * waiting at hith-memory prressure caused by "move".
1414 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1416 VM_BUG_ON(!rcu_read_lock_held());
1417 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1420 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1422 struct mem_cgroup *from;
1423 struct mem_cgroup *to;
1426 * Unlike task_move routines, we access mc.to, mc.from not under
1427 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1429 spin_lock(&mc.lock);
1435 ret = mem_cgroup_same_or_subtree(memcg, from)
1436 || mem_cgroup_same_or_subtree(memcg, to);
1438 spin_unlock(&mc.lock);
1442 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1444 if (mc.moving_task && current != mc.moving_task) {
1445 if (mem_cgroup_under_move(memcg)) {
1447 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1448 /* moving charge context might have finished. */
1451 finish_wait(&mc.waitq, &wait);
1459 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1460 * @memcg: The memory cgroup that went over limit
1461 * @p: Task that is going to be killed
1463 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1466 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1468 struct cgroup *task_cgrp;
1469 struct cgroup *mem_cgrp;
1471 * Need a buffer in BSS, can't rely on allocations. The code relies
1472 * on the assumption that OOM is serialized for memory controller.
1473 * If this assumption is broken, revisit this code.
1475 static char memcg_name[PATH_MAX];
1484 mem_cgrp = memcg->css.cgroup;
1485 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1487 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1490 * Unfortunately, we are unable to convert to a useful name
1491 * But we'll still print out the usage information
1498 printk(KERN_INFO "Task in %s killed", memcg_name);
1501 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1509 * Continues from above, so we don't need an KERN_ level
1511 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1514 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1515 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1516 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1517 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1518 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1520 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1521 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1522 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1526 * This function returns the number of memcg under hierarchy tree. Returns
1527 * 1(self count) if no children.
1529 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1532 struct mem_cgroup *iter;
1534 for_each_mem_cgroup_tree(iter, memcg)
1540 * Return the memory (and swap, if configured) limit for a memcg.
1542 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1547 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1548 limit += total_swap_pages << PAGE_SHIFT;
1550 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1552 * If memsw is finite and limits the amount of swap space available
1553 * to this memcg, return that limit.
1555 return min(limit, memsw);
1559 * test_mem_cgroup_node_reclaimable
1560 * @mem: the target memcg
1561 * @nid: the node ID to be checked.
1562 * @noswap : specify true here if the user wants flle only information.
1564 * This function returns whether the specified memcg contains any
1565 * reclaimable pages on a node. Returns true if there are any reclaimable
1566 * pages in the node.
1568 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1569 int nid, bool noswap)
1571 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1573 if (noswap || !total_swap_pages)
1575 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1580 #if MAX_NUMNODES > 1
1583 * Always updating the nodemask is not very good - even if we have an empty
1584 * list or the wrong list here, we can start from some node and traverse all
1585 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1588 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1592 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1593 * pagein/pageout changes since the last update.
1595 if (!atomic_read(&memcg->numainfo_events))
1597 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1600 /* make a nodemask where this memcg uses memory from */
1601 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1603 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1605 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1606 node_clear(nid, memcg->scan_nodes);
1609 atomic_set(&memcg->numainfo_events, 0);
1610 atomic_set(&memcg->numainfo_updating, 0);
1614 * Selecting a node where we start reclaim from. Because what we need is just
1615 * reducing usage counter, start from anywhere is O,K. Considering
1616 * memory reclaim from current node, there are pros. and cons.
1618 * Freeing memory from current node means freeing memory from a node which
1619 * we'll use or we've used. So, it may make LRU bad. And if several threads
1620 * hit limits, it will see a contention on a node. But freeing from remote
1621 * node means more costs for memory reclaim because of memory latency.
1623 * Now, we use round-robin. Better algorithm is welcomed.
1625 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1629 mem_cgroup_may_update_nodemask(memcg);
1630 node = memcg->last_scanned_node;
1632 node = next_node(node, memcg->scan_nodes);
1633 if (node == MAX_NUMNODES)
1634 node = first_node(memcg->scan_nodes);
1636 * We call this when we hit limit, not when pages are added to LRU.
1637 * No LRU may hold pages because all pages are UNEVICTABLE or
1638 * memcg is too small and all pages are not on LRU. In that case,
1639 * we use curret node.
1641 if (unlikely(node == MAX_NUMNODES))
1642 node = numa_node_id();
1644 memcg->last_scanned_node = node;
1649 * Check all nodes whether it contains reclaimable pages or not.
1650 * For quick scan, we make use of scan_nodes. This will allow us to skip
1651 * unused nodes. But scan_nodes is lazily updated and may not cotain
1652 * enough new information. We need to do double check.
1654 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1659 * quick check...making use of scan_node.
1660 * We can skip unused nodes.
1662 if (!nodes_empty(memcg->scan_nodes)) {
1663 for (nid = first_node(memcg->scan_nodes);
1665 nid = next_node(nid, memcg->scan_nodes)) {
1667 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1672 * Check rest of nodes.
1674 for_each_node_state(nid, N_HIGH_MEMORY) {
1675 if (node_isset(nid, memcg->scan_nodes))
1677 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1684 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1689 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1691 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1696 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1697 * we reclaimed from, so that we don't end up penalizing one child extensively
1698 * based on its position in the children list.
1700 * root_memcg is the original ancestor that we've been reclaim from.
1702 * We give up and return to the caller when we visit root_memcg twice.
1703 * (other groups can be removed while we're walking....)
1705 * If shrink==true, for avoiding to free too much, this returns immedieately.
1707 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_memcg,
1710 unsigned long reclaim_options,
1711 unsigned long *total_scanned)
1713 struct mem_cgroup *victim = NULL;
1716 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1717 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1718 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1719 unsigned long excess;
1720 unsigned long nr_scanned;
1721 struct mem_cgroup_reclaim_cookie reclaim = {
1726 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1728 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1729 if (!check_soft && !shrink && root_memcg->memsw_is_minimum)
1733 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1737 * We are not draining per cpu cached charges during
1738 * soft limit reclaim because global reclaim doesn't
1739 * care about charges. It tries to free some memory and
1740 * charges will not give any.
1742 if (!check_soft && loop >= 1)
1743 drain_all_stock_async(root_memcg);
1746 * If we have not been able to reclaim
1747 * anything, it might because there are
1748 * no reclaimable pages under this hierarchy
1750 if (!check_soft || !total)
1753 * We want to do more targeted reclaim.
1754 * excess >> 2 is not to excessive so as to
1755 * reclaim too much, nor too less that we keep
1756 * coming back to reclaim from this cgroup
1758 if (total >= (excess >> 2) ||
1759 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1764 if (!mem_cgroup_reclaimable(victim, noswap)) {
1765 /* this cgroup's local usage == 0 */
1768 /* we use swappiness of local cgroup */
1770 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1771 noswap, zone, &nr_scanned);
1772 *total_scanned += nr_scanned;
1774 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1778 * At shrinking usage, we can't check we should stop here or
1779 * reclaim more. It's depends on callers. last_scanned_child
1780 * will work enough for keeping fairness under tree.
1785 if (!res_counter_soft_limit_excess(&root_memcg->res))
1787 } else if (mem_cgroup_margin(root_memcg))
1790 mem_cgroup_iter_break(root_memcg, victim);
1795 * Check OOM-Killer is already running under our hierarchy.
1796 * If someone is running, return false.
1797 * Has to be called with memcg_oom_lock
1799 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1801 struct mem_cgroup *iter, *failed = NULL;
1803 for_each_mem_cgroup_tree(iter, memcg) {
1804 if (iter->oom_lock) {
1806 * this subtree of our hierarchy is already locked
1807 * so we cannot give a lock.
1810 mem_cgroup_iter_break(memcg, iter);
1813 iter->oom_lock = true;
1820 * OK, we failed to lock the whole subtree so we have to clean up
1821 * what we set up to the failing subtree
1823 for_each_mem_cgroup_tree(iter, memcg) {
1824 if (iter == failed) {
1825 mem_cgroup_iter_break(memcg, iter);
1828 iter->oom_lock = false;
1834 * Has to be called with memcg_oom_lock
1836 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1838 struct mem_cgroup *iter;
1840 for_each_mem_cgroup_tree(iter, memcg)
1841 iter->oom_lock = false;
1845 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1847 struct mem_cgroup *iter;
1849 for_each_mem_cgroup_tree(iter, memcg)
1850 atomic_inc(&iter->under_oom);
1853 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1855 struct mem_cgroup *iter;
1858 * When a new child is created while the hierarchy is under oom,
1859 * mem_cgroup_oom_lock() may not be called. We have to use
1860 * atomic_add_unless() here.
1862 for_each_mem_cgroup_tree(iter, memcg)
1863 atomic_add_unless(&iter->under_oom, -1, 0);
1866 static DEFINE_SPINLOCK(memcg_oom_lock);
1867 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1869 struct oom_wait_info {
1870 struct mem_cgroup *mem;
1874 static int memcg_oom_wake_function(wait_queue_t *wait,
1875 unsigned mode, int sync, void *arg)
1877 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1879 struct oom_wait_info *oom_wait_info;
1881 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1882 oom_wait_memcg = oom_wait_info->mem;
1885 * Both of oom_wait_info->mem and wake_mem are stable under us.
1886 * Then we can use css_is_ancestor without taking care of RCU.
1888 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1889 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1891 return autoremove_wake_function(wait, mode, sync, arg);
1894 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1896 /* for filtering, pass "memcg" as argument. */
1897 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1900 static void memcg_oom_recover(struct mem_cgroup *memcg)
1902 if (memcg && atomic_read(&memcg->under_oom))
1903 memcg_wakeup_oom(memcg);
1907 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1909 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1911 struct oom_wait_info owait;
1912 bool locked, need_to_kill;
1915 owait.wait.flags = 0;
1916 owait.wait.func = memcg_oom_wake_function;
1917 owait.wait.private = current;
1918 INIT_LIST_HEAD(&owait.wait.task_list);
1919 need_to_kill = true;
1920 mem_cgroup_mark_under_oom(memcg);
1922 /* At first, try to OOM lock hierarchy under memcg.*/
1923 spin_lock(&memcg_oom_lock);
1924 locked = mem_cgroup_oom_lock(memcg);
1926 * Even if signal_pending(), we can't quit charge() loop without
1927 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1928 * under OOM is always welcomed, use TASK_KILLABLE here.
1930 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1931 if (!locked || memcg->oom_kill_disable)
1932 need_to_kill = false;
1934 mem_cgroup_oom_notify(memcg);
1935 spin_unlock(&memcg_oom_lock);
1938 finish_wait(&memcg_oom_waitq, &owait.wait);
1939 mem_cgroup_out_of_memory(memcg, mask);
1942 finish_wait(&memcg_oom_waitq, &owait.wait);
1944 spin_lock(&memcg_oom_lock);
1946 mem_cgroup_oom_unlock(memcg);
1947 memcg_wakeup_oom(memcg);
1948 spin_unlock(&memcg_oom_lock);
1950 mem_cgroup_unmark_under_oom(memcg);
1952 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1954 /* Give chance to dying process */
1955 schedule_timeout_uninterruptible(1);
1960 * Currently used to update mapped file statistics, but the routine can be
1961 * generalized to update other statistics as well.
1963 * Notes: Race condition
1965 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1966 * it tends to be costly. But considering some conditions, we doesn't need
1967 * to do so _always_.
1969 * Considering "charge", lock_page_cgroup() is not required because all
1970 * file-stat operations happen after a page is attached to radix-tree. There
1971 * are no race with "charge".
1973 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1974 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1975 * if there are race with "uncharge". Statistics itself is properly handled
1978 * Considering "move", this is an only case we see a race. To make the race
1979 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1980 * possibility of race condition. If there is, we take a lock.
1983 void mem_cgroup_update_page_stat(struct page *page,
1984 enum mem_cgroup_page_stat_item idx, int val)
1986 struct mem_cgroup *memcg;
1987 struct page_cgroup *pc = lookup_page_cgroup(page);
1988 bool need_unlock = false;
1989 unsigned long uninitialized_var(flags);
1995 memcg = pc->mem_cgroup;
1996 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1998 /* pc->mem_cgroup is unstable ? */
1999 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
2000 /* take a lock against to access pc->mem_cgroup */
2001 move_lock_page_cgroup(pc, &flags);
2003 memcg = pc->mem_cgroup;
2004 if (!memcg || !PageCgroupUsed(pc))
2009 case MEMCG_NR_FILE_MAPPED:
2011 SetPageCgroupFileMapped(pc);
2012 else if (!page_mapped(page))
2013 ClearPageCgroupFileMapped(pc);
2014 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2020 this_cpu_add(memcg->stat->count[idx], val);
2023 if (unlikely(need_unlock))
2024 move_unlock_page_cgroup(pc, &flags);
2028 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2031 * size of first charge trial. "32" comes from vmscan.c's magic value.
2032 * TODO: maybe necessary to use big numbers in big irons.
2034 #define CHARGE_BATCH 32U
2035 struct memcg_stock_pcp {
2036 struct mem_cgroup *cached; /* this never be root cgroup */
2037 unsigned int nr_pages;
2038 struct work_struct work;
2039 unsigned long flags;
2040 #define FLUSHING_CACHED_CHARGE (0)
2042 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2043 static DEFINE_MUTEX(percpu_charge_mutex);
2046 * Try to consume stocked charge on this cpu. If success, one page is consumed
2047 * from local stock and true is returned. If the stock is 0 or charges from a
2048 * cgroup which is not current target, returns false. This stock will be
2051 static bool consume_stock(struct mem_cgroup *memcg)
2053 struct memcg_stock_pcp *stock;
2056 stock = &get_cpu_var(memcg_stock);
2057 if (memcg == stock->cached && stock->nr_pages)
2059 else /* need to call res_counter_charge */
2061 put_cpu_var(memcg_stock);
2066 * Returns stocks cached in percpu to res_counter and reset cached information.
2068 static void drain_stock(struct memcg_stock_pcp *stock)
2070 struct mem_cgroup *old = stock->cached;
2072 if (stock->nr_pages) {
2073 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2075 res_counter_uncharge(&old->res, bytes);
2076 if (do_swap_account)
2077 res_counter_uncharge(&old->memsw, bytes);
2078 stock->nr_pages = 0;
2080 stock->cached = NULL;
2084 * This must be called under preempt disabled or must be called by
2085 * a thread which is pinned to local cpu.
2087 static void drain_local_stock(struct work_struct *dummy)
2089 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2091 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2095 * Cache charges(val) which is from res_counter, to local per_cpu area.
2096 * This will be consumed by consume_stock() function, later.
2098 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2100 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2102 if (stock->cached != memcg) { /* reset if necessary */
2104 stock->cached = memcg;
2106 stock->nr_pages += nr_pages;
2107 put_cpu_var(memcg_stock);
2111 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2112 * of the hierarchy under it. sync flag says whether we should block
2113 * until the work is done.
2115 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2119 /* Notify other cpus that system-wide "drain" is running */
2122 for_each_online_cpu(cpu) {
2123 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2124 struct mem_cgroup *memcg;
2126 memcg = stock->cached;
2127 if (!memcg || !stock->nr_pages)
2129 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2131 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2133 drain_local_stock(&stock->work);
2135 schedule_work_on(cpu, &stock->work);
2143 for_each_online_cpu(cpu) {
2144 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2145 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2146 flush_work(&stock->work);
2153 * Tries to drain stocked charges in other cpus. This function is asynchronous
2154 * and just put a work per cpu for draining localy on each cpu. Caller can
2155 * expects some charges will be back to res_counter later but cannot wait for
2158 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2161 * If someone calls draining, avoid adding more kworker runs.
2163 if (!mutex_trylock(&percpu_charge_mutex))
2165 drain_all_stock(root_memcg, false);
2166 mutex_unlock(&percpu_charge_mutex);
2169 /* This is a synchronous drain interface. */
2170 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2172 /* called when force_empty is called */
2173 mutex_lock(&percpu_charge_mutex);
2174 drain_all_stock(root_memcg, true);
2175 mutex_unlock(&percpu_charge_mutex);
2179 * This function drains percpu counter value from DEAD cpu and
2180 * move it to local cpu. Note that this function can be preempted.
2182 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2186 spin_lock(&memcg->pcp_counter_lock);
2187 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2188 long x = per_cpu(memcg->stat->count[i], cpu);
2190 per_cpu(memcg->stat->count[i], cpu) = 0;
2191 memcg->nocpu_base.count[i] += x;
2193 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2194 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2196 per_cpu(memcg->stat->events[i], cpu) = 0;
2197 memcg->nocpu_base.events[i] += x;
2199 /* need to clear ON_MOVE value, works as a kind of lock. */
2200 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2201 spin_unlock(&memcg->pcp_counter_lock);
2204 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2206 int idx = MEM_CGROUP_ON_MOVE;
2208 spin_lock(&memcg->pcp_counter_lock);
2209 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2210 spin_unlock(&memcg->pcp_counter_lock);
2213 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2214 unsigned long action,
2217 int cpu = (unsigned long)hcpu;
2218 struct memcg_stock_pcp *stock;
2219 struct mem_cgroup *iter;
2221 if ((action == CPU_ONLINE)) {
2222 for_each_mem_cgroup(iter)
2223 synchronize_mem_cgroup_on_move(iter, cpu);
2227 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2230 for_each_mem_cgroup(iter)
2231 mem_cgroup_drain_pcp_counter(iter, cpu);
2233 stock = &per_cpu(memcg_stock, cpu);
2239 /* See __mem_cgroup_try_charge() for details */
2241 CHARGE_OK, /* success */
2242 CHARGE_RETRY, /* need to retry but retry is not bad */
2243 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2244 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2245 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2248 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2249 unsigned int nr_pages, bool oom_check)
2251 unsigned long csize = nr_pages * PAGE_SIZE;
2252 struct mem_cgroup *mem_over_limit;
2253 struct res_counter *fail_res;
2254 unsigned long flags = 0;
2257 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2260 if (!do_swap_account)
2262 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2266 res_counter_uncharge(&memcg->res, csize);
2267 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2268 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2270 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2272 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2273 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2275 * Never reclaim on behalf of optional batching, retry with a
2276 * single page instead.
2278 if (nr_pages == CHARGE_BATCH)
2279 return CHARGE_RETRY;
2281 if (!(gfp_mask & __GFP_WAIT))
2282 return CHARGE_WOULDBLOCK;
2284 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2285 gfp_mask, flags, NULL);
2286 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2287 return CHARGE_RETRY;
2289 * Even though the limit is exceeded at this point, reclaim
2290 * may have been able to free some pages. Retry the charge
2291 * before killing the task.
2293 * Only for regular pages, though: huge pages are rather
2294 * unlikely to succeed so close to the limit, and we fall back
2295 * to regular pages anyway in case of failure.
2297 if (nr_pages == 1 && ret)
2298 return CHARGE_RETRY;
2301 * At task move, charge accounts can be doubly counted. So, it's
2302 * better to wait until the end of task_move if something is going on.
2304 if (mem_cgroup_wait_acct_move(mem_over_limit))
2305 return CHARGE_RETRY;
2307 /* If we don't need to call oom-killer at el, return immediately */
2309 return CHARGE_NOMEM;
2311 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2312 return CHARGE_OOM_DIE;
2314 return CHARGE_RETRY;
2318 * Unlike exported interface, "oom" parameter is added. if oom==true,
2319 * oom-killer can be invoked.
2321 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2323 unsigned int nr_pages,
2324 struct mem_cgroup **ptr,
2327 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2328 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2329 struct mem_cgroup *memcg = NULL;
2333 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2334 * in system level. So, allow to go ahead dying process in addition to
2337 if (unlikely(test_thread_flag(TIF_MEMDIE)
2338 || fatal_signal_pending(current)))
2342 * We always charge the cgroup the mm_struct belongs to.
2343 * The mm_struct's mem_cgroup changes on task migration if the
2344 * thread group leader migrates. It's possible that mm is not
2345 * set, if so charge the init_mm (happens for pagecache usage).
2350 if (*ptr) { /* css should be a valid one */
2352 VM_BUG_ON(css_is_removed(&memcg->css));
2353 if (mem_cgroup_is_root(memcg))
2355 if (nr_pages == 1 && consume_stock(memcg))
2357 css_get(&memcg->css);
2359 struct task_struct *p;
2362 p = rcu_dereference(mm->owner);
2364 * Because we don't have task_lock(), "p" can exit.
2365 * In that case, "memcg" can point to root or p can be NULL with
2366 * race with swapoff. Then, we have small risk of mis-accouning.
2367 * But such kind of mis-account by race always happens because
2368 * we don't have cgroup_mutex(). It's overkill and we allo that
2370 * (*) swapoff at el will charge against mm-struct not against
2371 * task-struct. So, mm->owner can be NULL.
2373 memcg = mem_cgroup_from_task(p);
2374 if (!memcg || mem_cgroup_is_root(memcg)) {
2378 if (nr_pages == 1 && consume_stock(memcg)) {
2380 * It seems dagerous to access memcg without css_get().
2381 * But considering how consume_stok works, it's not
2382 * necessary. If consume_stock success, some charges
2383 * from this memcg are cached on this cpu. So, we
2384 * don't need to call css_get()/css_tryget() before
2385 * calling consume_stock().
2390 /* after here, we may be blocked. we need to get refcnt */
2391 if (!css_tryget(&memcg->css)) {
2401 /* If killed, bypass charge */
2402 if (fatal_signal_pending(current)) {
2403 css_put(&memcg->css);
2408 if (oom && !nr_oom_retries) {
2410 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2413 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2417 case CHARGE_RETRY: /* not in OOM situation but retry */
2419 css_put(&memcg->css);
2422 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2423 css_put(&memcg->css);
2425 case CHARGE_NOMEM: /* OOM routine works */
2427 css_put(&memcg->css);
2430 /* If oom, we never return -ENOMEM */
2433 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2434 css_put(&memcg->css);
2437 } while (ret != CHARGE_OK);
2439 if (batch > nr_pages)
2440 refill_stock(memcg, batch - nr_pages);
2441 css_put(&memcg->css);
2454 * Somemtimes we have to undo a charge we got by try_charge().
2455 * This function is for that and do uncharge, put css's refcnt.
2456 * gotten by try_charge().
2458 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2459 unsigned int nr_pages)
2461 if (!mem_cgroup_is_root(memcg)) {
2462 unsigned long bytes = nr_pages * PAGE_SIZE;
2464 res_counter_uncharge(&memcg->res, bytes);
2465 if (do_swap_account)
2466 res_counter_uncharge(&memcg->memsw, bytes);
2471 * A helper function to get mem_cgroup from ID. must be called under
2472 * rcu_read_lock(). The caller must check css_is_removed() or some if
2473 * it's concern. (dropping refcnt from swap can be called against removed
2476 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2478 struct cgroup_subsys_state *css;
2480 /* ID 0 is unused ID */
2483 css = css_lookup(&mem_cgroup_subsys, id);
2486 return container_of(css, struct mem_cgroup, css);
2489 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2491 struct mem_cgroup *memcg = NULL;
2492 struct page_cgroup *pc;
2496 VM_BUG_ON(!PageLocked(page));
2498 pc = lookup_page_cgroup(page);
2499 lock_page_cgroup(pc);
2500 if (PageCgroupUsed(pc)) {
2501 memcg = pc->mem_cgroup;
2502 if (memcg && !css_tryget(&memcg->css))
2504 } else if (PageSwapCache(page)) {
2505 ent.val = page_private(page);
2506 id = lookup_swap_cgroup(ent);
2508 memcg = mem_cgroup_lookup(id);
2509 if (memcg && !css_tryget(&memcg->css))
2513 unlock_page_cgroup(pc);
2517 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2519 unsigned int nr_pages,
2520 struct page_cgroup *pc,
2521 enum charge_type ctype)
2523 lock_page_cgroup(pc);
2524 if (unlikely(PageCgroupUsed(pc))) {
2525 unlock_page_cgroup(pc);
2526 __mem_cgroup_cancel_charge(memcg, nr_pages);
2530 * we don't need page_cgroup_lock about tail pages, becase they are not
2531 * accessed by any other context at this point.
2533 pc->mem_cgroup = memcg;
2535 * We access a page_cgroup asynchronously without lock_page_cgroup().
2536 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2537 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2538 * before USED bit, we need memory barrier here.
2539 * See mem_cgroup_add_lru_list(), etc.
2543 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2544 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2545 SetPageCgroupCache(pc);
2546 SetPageCgroupUsed(pc);
2548 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2549 ClearPageCgroupCache(pc);
2550 SetPageCgroupUsed(pc);
2556 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2557 unlock_page_cgroup(pc);
2559 * "charge_statistics" updated event counter. Then, check it.
2560 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2561 * if they exceeds softlimit.
2563 memcg_check_events(memcg, page);
2566 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2568 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2569 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2571 * Because tail pages are not marked as "used", set it. We're under
2572 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2574 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2576 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2577 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2578 unsigned long flags;
2580 if (mem_cgroup_disabled())
2583 * We have no races with charge/uncharge but will have races with
2584 * page state accounting.
2586 move_lock_page_cgroup(head_pc, &flags);
2588 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2589 smp_wmb(); /* see __commit_charge() */
2590 if (PageCgroupAcctLRU(head_pc)) {
2592 struct mem_cgroup_per_zone *mz;
2595 * LRU flags cannot be copied because we need to add tail
2596 *.page to LRU by generic call and our hook will be called.
2597 * We hold lru_lock, then, reduce counter directly.
2599 lru = page_lru(head);
2600 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2601 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2603 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2604 move_unlock_page_cgroup(head_pc, &flags);
2609 * mem_cgroup_move_account - move account of the page
2611 * @nr_pages: number of regular pages (>1 for huge pages)
2612 * @pc: page_cgroup of the page.
2613 * @from: mem_cgroup which the page is moved from.
2614 * @to: mem_cgroup which the page is moved to. @from != @to.
2615 * @uncharge: whether we should call uncharge and css_put against @from.
2617 * The caller must confirm following.
2618 * - page is not on LRU (isolate_page() is useful.)
2619 * - compound_lock is held when nr_pages > 1
2621 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2622 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2623 * true, this function does "uncharge" from old cgroup, but it doesn't if
2624 * @uncharge is false, so a caller should do "uncharge".
2626 static int mem_cgroup_move_account(struct page *page,
2627 unsigned int nr_pages,
2628 struct page_cgroup *pc,
2629 struct mem_cgroup *from,
2630 struct mem_cgroup *to,
2633 unsigned long flags;
2636 VM_BUG_ON(from == to);
2637 VM_BUG_ON(PageLRU(page));
2639 * The page is isolated from LRU. So, collapse function
2640 * will not handle this page. But page splitting can happen.
2641 * Do this check under compound_page_lock(). The caller should
2645 if (nr_pages > 1 && !PageTransHuge(page))
2648 lock_page_cgroup(pc);
2651 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2654 move_lock_page_cgroup(pc, &flags);
2656 if (PageCgroupFileMapped(pc)) {
2657 /* Update mapped_file data for mem_cgroup */
2659 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2660 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2663 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2665 /* This is not "cancel", but cancel_charge does all we need. */
2666 __mem_cgroup_cancel_charge(from, nr_pages);
2668 /* caller should have done css_get */
2669 pc->mem_cgroup = to;
2670 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2672 * We charges against "to" which may not have any tasks. Then, "to"
2673 * can be under rmdir(). But in current implementation, caller of
2674 * this function is just force_empty() and move charge, so it's
2675 * guaranteed that "to" is never removed. So, we don't check rmdir
2678 move_unlock_page_cgroup(pc, &flags);
2681 unlock_page_cgroup(pc);
2685 memcg_check_events(to, page);
2686 memcg_check_events(from, page);
2692 * move charges to its parent.
2695 static int mem_cgroup_move_parent(struct page *page,
2696 struct page_cgroup *pc,
2697 struct mem_cgroup *child,
2700 struct cgroup *cg = child->css.cgroup;
2701 struct cgroup *pcg = cg->parent;
2702 struct mem_cgroup *parent;
2703 unsigned int nr_pages;
2704 unsigned long uninitialized_var(flags);
2712 if (!get_page_unless_zero(page))
2714 if (isolate_lru_page(page))
2717 nr_pages = hpage_nr_pages(page);
2719 parent = mem_cgroup_from_cont(pcg);
2720 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2725 flags = compound_lock_irqsave(page);
2727 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2729 __mem_cgroup_cancel_charge(parent, nr_pages);
2732 compound_unlock_irqrestore(page, flags);
2734 putback_lru_page(page);
2742 * Charge the memory controller for page usage.
2744 * 0 if the charge was successful
2745 * < 0 if the cgroup is over its limit
2747 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2748 gfp_t gfp_mask, enum charge_type ctype)
2750 struct mem_cgroup *memcg = NULL;
2751 unsigned int nr_pages = 1;
2752 struct page_cgroup *pc;
2756 if (PageTransHuge(page)) {
2757 nr_pages <<= compound_order(page);
2758 VM_BUG_ON(!PageTransHuge(page));
2760 * Never OOM-kill a process for a huge page. The
2761 * fault handler will fall back to regular pages.
2766 pc = lookup_page_cgroup(page);
2767 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2769 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2773 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2777 int mem_cgroup_newpage_charge(struct page *page,
2778 struct mm_struct *mm, gfp_t gfp_mask)
2780 if (mem_cgroup_disabled())
2783 * If already mapped, we don't have to account.
2784 * If page cache, page->mapping has address_space.
2785 * But page->mapping may have out-of-use anon_vma pointer,
2786 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2789 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2793 return mem_cgroup_charge_common(page, mm, gfp_mask,
2794 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2798 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2799 enum charge_type ctype);
2802 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2803 enum charge_type ctype)
2805 struct page_cgroup *pc = lookup_page_cgroup(page);
2807 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2808 * is already on LRU. It means the page may on some other page_cgroup's
2809 * LRU. Take care of it.
2811 mem_cgroup_lru_del_before_commit(page);
2812 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2813 mem_cgroup_lru_add_after_commit(page);
2817 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2820 struct mem_cgroup *memcg = NULL;
2823 if (mem_cgroup_disabled())
2825 if (PageCompound(page))
2831 if (page_is_file_cache(page)) {
2832 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2837 * FUSE reuses pages without going through the final
2838 * put that would remove them from the LRU list, make
2839 * sure that they get relinked properly.
2841 __mem_cgroup_commit_charge_lrucare(page, memcg,
2842 MEM_CGROUP_CHARGE_TYPE_CACHE);
2846 if (PageSwapCache(page)) {
2847 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2849 __mem_cgroup_commit_charge_swapin(page, memcg,
2850 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2852 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2853 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2859 * While swap-in, try_charge -> commit or cancel, the page is locked.
2860 * And when try_charge() successfully returns, one refcnt to memcg without
2861 * struct page_cgroup is acquired. This refcnt will be consumed by
2862 * "commit()" or removed by "cancel()"
2864 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2866 gfp_t mask, struct mem_cgroup **ptr)
2868 struct mem_cgroup *memcg;
2873 if (mem_cgroup_disabled())
2876 if (!do_swap_account)
2879 * A racing thread's fault, or swapoff, may have already updated
2880 * the pte, and even removed page from swap cache: in those cases
2881 * do_swap_page()'s pte_same() test will fail; but there's also a
2882 * KSM case which does need to charge the page.
2884 if (!PageSwapCache(page))
2886 memcg = try_get_mem_cgroup_from_page(page);
2890 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2891 css_put(&memcg->css);
2896 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2900 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2901 enum charge_type ctype)
2903 if (mem_cgroup_disabled())
2907 cgroup_exclude_rmdir(&ptr->css);
2909 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2911 * Now swap is on-memory. This means this page may be
2912 * counted both as mem and swap....double count.
2913 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2914 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2915 * may call delete_from_swap_cache() before reach here.
2917 if (do_swap_account && PageSwapCache(page)) {
2918 swp_entry_t ent = {.val = page_private(page)};
2920 struct mem_cgroup *memcg;
2922 id = swap_cgroup_record(ent, 0);
2924 memcg = mem_cgroup_lookup(id);
2927 * This recorded memcg can be obsolete one. So, avoid
2928 * calling css_tryget
2930 if (!mem_cgroup_is_root(memcg))
2931 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2932 mem_cgroup_swap_statistics(memcg, false);
2933 mem_cgroup_put(memcg);
2938 * At swapin, we may charge account against cgroup which has no tasks.
2939 * So, rmdir()->pre_destroy() can be called while we do this charge.
2940 * In that case, we need to call pre_destroy() again. check it here.
2942 cgroup_release_and_wakeup_rmdir(&ptr->css);
2945 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2947 __mem_cgroup_commit_charge_swapin(page, ptr,
2948 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2951 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2953 if (mem_cgroup_disabled())
2957 __mem_cgroup_cancel_charge(memcg, 1);
2960 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2961 unsigned int nr_pages,
2962 const enum charge_type ctype)
2964 struct memcg_batch_info *batch = NULL;
2965 bool uncharge_memsw = true;
2967 /* If swapout, usage of swap doesn't decrease */
2968 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2969 uncharge_memsw = false;
2971 batch = ¤t->memcg_batch;
2973 * In usual, we do css_get() when we remember memcg pointer.
2974 * But in this case, we keep res->usage until end of a series of
2975 * uncharges. Then, it's ok to ignore memcg's refcnt.
2978 batch->memcg = memcg;
2980 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2981 * In those cases, all pages freed continuously can be expected to be in
2982 * the same cgroup and we have chance to coalesce uncharges.
2983 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2984 * because we want to do uncharge as soon as possible.
2987 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2988 goto direct_uncharge;
2991 goto direct_uncharge;
2994 * In typical case, batch->memcg == mem. This means we can
2995 * merge a series of uncharges to an uncharge of res_counter.
2996 * If not, we uncharge res_counter ony by one.
2998 if (batch->memcg != memcg)
2999 goto direct_uncharge;
3000 /* remember freed charge and uncharge it later */
3003 batch->memsw_nr_pages++;
3006 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3008 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3009 if (unlikely(batch->memcg != memcg))
3010 memcg_oom_recover(memcg);
3015 * uncharge if !page_mapped(page)
3017 static struct mem_cgroup *
3018 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3020 struct mem_cgroup *memcg = NULL;
3021 unsigned int nr_pages = 1;
3022 struct page_cgroup *pc;
3024 if (mem_cgroup_disabled())
3027 if (PageSwapCache(page))
3030 if (PageTransHuge(page)) {
3031 nr_pages <<= compound_order(page);
3032 VM_BUG_ON(!PageTransHuge(page));
3035 * Check if our page_cgroup is valid
3037 pc = lookup_page_cgroup(page);
3038 if (unlikely(!pc || !PageCgroupUsed(pc)))
3041 lock_page_cgroup(pc);
3043 memcg = pc->mem_cgroup;
3045 if (!PageCgroupUsed(pc))
3049 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3050 case MEM_CGROUP_CHARGE_TYPE_DROP:
3051 /* See mem_cgroup_prepare_migration() */
3052 if (page_mapped(page) || PageCgroupMigration(pc))
3055 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3056 if (!PageAnon(page)) { /* Shared memory */
3057 if (page->mapping && !page_is_file_cache(page))
3059 } else if (page_mapped(page)) /* Anon */
3066 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3068 ClearPageCgroupUsed(pc);
3070 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3071 * freed from LRU. This is safe because uncharged page is expected not
3072 * to be reused (freed soon). Exception is SwapCache, it's handled by
3073 * special functions.
3076 unlock_page_cgroup(pc);
3078 * even after unlock, we have memcg->res.usage here and this memcg
3079 * will never be freed.
3081 memcg_check_events(memcg, page);
3082 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3083 mem_cgroup_swap_statistics(memcg, true);
3084 mem_cgroup_get(memcg);
3086 if (!mem_cgroup_is_root(memcg))
3087 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3092 unlock_page_cgroup(pc);
3096 void mem_cgroup_uncharge_page(struct page *page)
3099 if (page_mapped(page))
3101 if (page->mapping && !PageAnon(page))
3103 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3106 void mem_cgroup_uncharge_cache_page(struct page *page)
3108 VM_BUG_ON(page_mapped(page));
3109 VM_BUG_ON(page->mapping);
3110 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3114 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3115 * In that cases, pages are freed continuously and we can expect pages
3116 * are in the same memcg. All these calls itself limits the number of
3117 * pages freed at once, then uncharge_start/end() is called properly.
3118 * This may be called prural(2) times in a context,
3121 void mem_cgroup_uncharge_start(void)
3123 current->memcg_batch.do_batch++;
3124 /* We can do nest. */
3125 if (current->memcg_batch.do_batch == 1) {
3126 current->memcg_batch.memcg = NULL;
3127 current->memcg_batch.nr_pages = 0;
3128 current->memcg_batch.memsw_nr_pages = 0;
3132 void mem_cgroup_uncharge_end(void)
3134 struct memcg_batch_info *batch = ¤t->memcg_batch;
3136 if (!batch->do_batch)
3140 if (batch->do_batch) /* If stacked, do nothing. */
3146 * This "batch->memcg" is valid without any css_get/put etc...
3147 * bacause we hide charges behind us.
3149 if (batch->nr_pages)
3150 res_counter_uncharge(&batch->memcg->res,
3151 batch->nr_pages * PAGE_SIZE);
3152 if (batch->memsw_nr_pages)
3153 res_counter_uncharge(&batch->memcg->memsw,
3154 batch->memsw_nr_pages * PAGE_SIZE);
3155 memcg_oom_recover(batch->memcg);
3156 /* forget this pointer (for sanity check) */
3157 batch->memcg = NULL;
3162 * called after __delete_from_swap_cache() and drop "page" account.
3163 * memcg information is recorded to swap_cgroup of "ent"
3166 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3168 struct mem_cgroup *memcg;
3169 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3171 if (!swapout) /* this was a swap cache but the swap is unused ! */
3172 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3174 memcg = __mem_cgroup_uncharge_common(page, ctype);
3177 * record memcg information, if swapout && memcg != NULL,
3178 * mem_cgroup_get() was called in uncharge().
3180 if (do_swap_account && swapout && memcg)
3181 swap_cgroup_record(ent, css_id(&memcg->css));
3185 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3187 * called from swap_entry_free(). remove record in swap_cgroup and
3188 * uncharge "memsw" account.
3190 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3192 struct mem_cgroup *memcg;
3195 if (!do_swap_account)
3198 id = swap_cgroup_record(ent, 0);
3200 memcg = mem_cgroup_lookup(id);
3203 * We uncharge this because swap is freed.
3204 * This memcg can be obsolete one. We avoid calling css_tryget
3206 if (!mem_cgroup_is_root(memcg))
3207 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3208 mem_cgroup_swap_statistics(memcg, false);
3209 mem_cgroup_put(memcg);
3215 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3216 * @entry: swap entry to be moved
3217 * @from: mem_cgroup which the entry is moved from
3218 * @to: mem_cgroup which the entry is moved to
3219 * @need_fixup: whether we should fixup res_counters and refcounts.
3221 * It succeeds only when the swap_cgroup's record for this entry is the same
3222 * as the mem_cgroup's id of @from.
3224 * Returns 0 on success, -EINVAL on failure.
3226 * The caller must have charged to @to, IOW, called res_counter_charge() about
3227 * both res and memsw, and called css_get().
3229 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3230 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3232 unsigned short old_id, new_id;
3234 old_id = css_id(&from->css);
3235 new_id = css_id(&to->css);
3237 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3238 mem_cgroup_swap_statistics(from, false);
3239 mem_cgroup_swap_statistics(to, true);
3241 * This function is only called from task migration context now.
3242 * It postpones res_counter and refcount handling till the end
3243 * of task migration(mem_cgroup_clear_mc()) for performance
3244 * improvement. But we cannot postpone mem_cgroup_get(to)
3245 * because if the process that has been moved to @to does
3246 * swap-in, the refcount of @to might be decreased to 0.
3250 if (!mem_cgroup_is_root(from))
3251 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3252 mem_cgroup_put(from);
3254 * we charged both to->res and to->memsw, so we should
3257 if (!mem_cgroup_is_root(to))
3258 res_counter_uncharge(&to->res, PAGE_SIZE);
3265 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3266 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3273 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3276 int mem_cgroup_prepare_migration(struct page *page,
3277 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3279 struct mem_cgroup *memcg = NULL;
3280 struct page_cgroup *pc;
3281 enum charge_type ctype;
3286 VM_BUG_ON(PageTransHuge(page));
3287 if (mem_cgroup_disabled())
3290 pc = lookup_page_cgroup(page);
3291 lock_page_cgroup(pc);
3292 if (PageCgroupUsed(pc)) {
3293 memcg = pc->mem_cgroup;
3294 css_get(&memcg->css);
3296 * At migrating an anonymous page, its mapcount goes down
3297 * to 0 and uncharge() will be called. But, even if it's fully
3298 * unmapped, migration may fail and this page has to be
3299 * charged again. We set MIGRATION flag here and delay uncharge
3300 * until end_migration() is called
3302 * Corner Case Thinking
3304 * When the old page was mapped as Anon and it's unmap-and-freed
3305 * while migration was ongoing.
3306 * If unmap finds the old page, uncharge() of it will be delayed
3307 * until end_migration(). If unmap finds a new page, it's
3308 * uncharged when it make mapcount to be 1->0. If unmap code
3309 * finds swap_migration_entry, the new page will not be mapped
3310 * and end_migration() will find it(mapcount==0).
3313 * When the old page was mapped but migraion fails, the kernel
3314 * remaps it. A charge for it is kept by MIGRATION flag even
3315 * if mapcount goes down to 0. We can do remap successfully
3316 * without charging it again.
3319 * The "old" page is under lock_page() until the end of
3320 * migration, so, the old page itself will not be swapped-out.
3321 * If the new page is swapped out before end_migraton, our
3322 * hook to usual swap-out path will catch the event.
3325 SetPageCgroupMigration(pc);
3327 unlock_page_cgroup(pc);
3329 * If the page is not charged at this point,
3336 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3337 css_put(&memcg->css);/* drop extra refcnt */
3338 if (ret || *ptr == NULL) {
3339 if (PageAnon(page)) {
3340 lock_page_cgroup(pc);
3341 ClearPageCgroupMigration(pc);
3342 unlock_page_cgroup(pc);
3344 * The old page may be fully unmapped while we kept it.
3346 mem_cgroup_uncharge_page(page);
3351 * We charge new page before it's used/mapped. So, even if unlock_page()
3352 * is called before end_migration, we can catch all events on this new
3353 * page. In the case new page is migrated but not remapped, new page's
3354 * mapcount will be finally 0 and we call uncharge in end_migration().
3356 pc = lookup_page_cgroup(newpage);
3358 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3359 else if (page_is_file_cache(page))
3360 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3362 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3363 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3367 /* remove redundant charge if migration failed*/
3368 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3369 struct page *oldpage, struct page *newpage, bool migration_ok)
3371 struct page *used, *unused;
3372 struct page_cgroup *pc;
3376 /* blocks rmdir() */
3377 cgroup_exclude_rmdir(&memcg->css);
3378 if (!migration_ok) {
3386 * We disallowed uncharge of pages under migration because mapcount
3387 * of the page goes down to zero, temporarly.
3388 * Clear the flag and check the page should be charged.
3390 pc = lookup_page_cgroup(oldpage);
3391 lock_page_cgroup(pc);
3392 ClearPageCgroupMigration(pc);
3393 unlock_page_cgroup(pc);
3395 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3398 * If a page is a file cache, radix-tree replacement is very atomic
3399 * and we can skip this check. When it was an Anon page, its mapcount
3400 * goes down to 0. But because we added MIGRATION flage, it's not
3401 * uncharged yet. There are several case but page->mapcount check
3402 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3403 * check. (see prepare_charge() also)
3406 mem_cgroup_uncharge_page(used);
3408 * At migration, we may charge account against cgroup which has no
3410 * So, rmdir()->pre_destroy() can be called while we do this charge.
3411 * In that case, we need to call pre_destroy() again. check it here.
3413 cgroup_release_and_wakeup_rmdir(&memcg->css);
3417 * At replace page cache, newpage is not under any memcg but it's on
3418 * LRU. So, this function doesn't touch res_counter but handles LRU
3419 * in correct way. Both pages are locked so we cannot race with uncharge.
3421 void mem_cgroup_replace_page_cache(struct page *oldpage,
3422 struct page *newpage)
3424 struct mem_cgroup *memcg;
3425 struct page_cgroup *pc;
3427 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3428 unsigned long flags;
3430 if (mem_cgroup_disabled())
3433 pc = lookup_page_cgroup(oldpage);
3434 /* fix accounting on old pages */
3435 lock_page_cgroup(pc);
3436 memcg = pc->mem_cgroup;
3437 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1);
3438 ClearPageCgroupUsed(pc);
3439 unlock_page_cgroup(pc);
3441 if (PageSwapBacked(oldpage))
3442 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3444 zone = page_zone(newpage);
3445 pc = lookup_page_cgroup(newpage);
3447 * Even if newpage->mapping was NULL before starting replacement,
3448 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3449 * LRU while we overwrite pc->mem_cgroup.
3451 spin_lock_irqsave(&zone->lru_lock, flags);
3452 if (PageLRU(newpage))
3453 del_page_from_lru_list(zone, newpage, page_lru(newpage));
3454 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type);
3455 if (PageLRU(newpage))
3456 add_page_to_lru_list(zone, newpage, page_lru(newpage));
3457 spin_unlock_irqrestore(&zone->lru_lock, flags);
3460 #ifdef CONFIG_DEBUG_VM
3461 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3463 struct page_cgroup *pc;
3465 pc = lookup_page_cgroup(page);
3466 if (likely(pc) && PageCgroupUsed(pc))
3471 bool mem_cgroup_bad_page_check(struct page *page)
3473 if (mem_cgroup_disabled())
3476 return lookup_page_cgroup_used(page) != NULL;
3479 void mem_cgroup_print_bad_page(struct page *page)
3481 struct page_cgroup *pc;
3483 pc = lookup_page_cgroup_used(page);
3488 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3489 pc, pc->flags, pc->mem_cgroup);
3491 path = kmalloc(PATH_MAX, GFP_KERNEL);
3494 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3499 printk(KERN_CONT "(%s)\n",
3500 (ret < 0) ? "cannot get the path" : path);
3506 static DEFINE_MUTEX(set_limit_mutex);
3508 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3509 unsigned long long val)
3512 u64 memswlimit, memlimit;
3514 int children = mem_cgroup_count_children(memcg);
3515 u64 curusage, oldusage;
3519 * For keeping hierarchical_reclaim simple, how long we should retry
3520 * is depends on callers. We set our retry-count to be function
3521 * of # of children which we should visit in this loop.
3523 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3525 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3528 while (retry_count) {
3529 if (signal_pending(current)) {
3534 * Rather than hide all in some function, I do this in
3535 * open coded manner. You see what this really does.
3536 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3538 mutex_lock(&set_limit_mutex);
3539 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3540 if (memswlimit < val) {
3542 mutex_unlock(&set_limit_mutex);
3546 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3550 ret = res_counter_set_limit(&memcg->res, val);
3552 if (memswlimit == val)
3553 memcg->memsw_is_minimum = true;
3555 memcg->memsw_is_minimum = false;
3557 mutex_unlock(&set_limit_mutex);
3562 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3563 MEM_CGROUP_RECLAIM_SHRINK,
3565 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3566 /* Usage is reduced ? */
3567 if (curusage >= oldusage)
3570 oldusage = curusage;
3572 if (!ret && enlarge)
3573 memcg_oom_recover(memcg);
3578 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3579 unsigned long long val)
3582 u64 memlimit, memswlimit, oldusage, curusage;
3583 int children = mem_cgroup_count_children(memcg);
3587 /* see mem_cgroup_resize_res_limit */
3588 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3589 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3590 while (retry_count) {
3591 if (signal_pending(current)) {
3596 * Rather than hide all in some function, I do this in
3597 * open coded manner. You see what this really does.
3598 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3600 mutex_lock(&set_limit_mutex);
3601 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3602 if (memlimit > val) {
3604 mutex_unlock(&set_limit_mutex);
3607 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3608 if (memswlimit < val)
3610 ret = res_counter_set_limit(&memcg->memsw, val);
3612 if (memlimit == val)
3613 memcg->memsw_is_minimum = true;
3615 memcg->memsw_is_minimum = false;
3617 mutex_unlock(&set_limit_mutex);
3622 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3623 MEM_CGROUP_RECLAIM_NOSWAP |
3624 MEM_CGROUP_RECLAIM_SHRINK,
3626 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3627 /* Usage is reduced ? */
3628 if (curusage >= oldusage)
3631 oldusage = curusage;
3633 if (!ret && enlarge)
3634 memcg_oom_recover(memcg);
3638 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3640 unsigned long *total_scanned)
3642 unsigned long nr_reclaimed = 0;
3643 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3644 unsigned long reclaimed;
3646 struct mem_cgroup_tree_per_zone *mctz;
3647 unsigned long long excess;
3648 unsigned long nr_scanned;
3653 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3655 * This loop can run a while, specially if mem_cgroup's continuously
3656 * keep exceeding their soft limit and putting the system under
3663 mz = mem_cgroup_largest_soft_limit_node(mctz);
3668 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3670 MEM_CGROUP_RECLAIM_SOFT,
3672 nr_reclaimed += reclaimed;
3673 *total_scanned += nr_scanned;
3674 spin_lock(&mctz->lock);
3677 * If we failed to reclaim anything from this memory cgroup
3678 * it is time to move on to the next cgroup
3684 * Loop until we find yet another one.
3686 * By the time we get the soft_limit lock
3687 * again, someone might have aded the
3688 * group back on the RB tree. Iterate to
3689 * make sure we get a different mem.
3690 * mem_cgroup_largest_soft_limit_node returns
3691 * NULL if no other cgroup is present on
3695 __mem_cgroup_largest_soft_limit_node(mctz);
3697 css_put(&next_mz->mem->css);
3698 else /* next_mz == NULL or other memcg */
3702 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3703 excess = res_counter_soft_limit_excess(&mz->mem->res);
3705 * One school of thought says that we should not add
3706 * back the node to the tree if reclaim returns 0.
3707 * But our reclaim could return 0, simply because due
3708 * to priority we are exposing a smaller subset of
3709 * memory to reclaim from. Consider this as a longer
3712 /* If excess == 0, no tree ops */
3713 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3714 spin_unlock(&mctz->lock);
3715 css_put(&mz->mem->css);
3718 * Could not reclaim anything and there are no more
3719 * mem cgroups to try or we seem to be looping without
3720 * reclaiming anything.
3722 if (!nr_reclaimed &&
3724 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3726 } while (!nr_reclaimed);
3728 css_put(&next_mz->mem->css);
3729 return nr_reclaimed;
3733 * This routine traverse page_cgroup in given list and drop them all.
3734 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3736 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3737 int node, int zid, enum lru_list lru)
3740 struct mem_cgroup_per_zone *mz;
3741 struct page_cgroup *pc, *busy;
3742 unsigned long flags, loop;
3743 struct list_head *list;
3746 zone = &NODE_DATA(node)->node_zones[zid];
3747 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3748 list = &mz->lists[lru];
3750 loop = MEM_CGROUP_ZSTAT(mz, lru);
3751 /* give some margin against EBUSY etc...*/
3758 spin_lock_irqsave(&zone->lru_lock, flags);
3759 if (list_empty(list)) {
3760 spin_unlock_irqrestore(&zone->lru_lock, flags);
3763 pc = list_entry(list->prev, struct page_cgroup, lru);
3765 list_move(&pc->lru, list);
3767 spin_unlock_irqrestore(&zone->lru_lock, flags);
3770 spin_unlock_irqrestore(&zone->lru_lock, flags);
3772 page = lookup_cgroup_page(pc);
3774 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3778 if (ret == -EBUSY || ret == -EINVAL) {
3779 /* found lock contention or "pc" is obsolete. */
3786 if (!ret && !list_empty(list))
3792 * make mem_cgroup's charge to be 0 if there is no task.
3793 * This enables deleting this mem_cgroup.
3795 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3798 int node, zid, shrink;
3799 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3800 struct cgroup *cgrp = memcg->css.cgroup;
3802 css_get(&memcg->css);
3805 /* should free all ? */
3811 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3814 if (signal_pending(current))
3816 /* This is for making all *used* pages to be on LRU. */
3817 lru_add_drain_all();
3818 drain_all_stock_sync(memcg);
3820 mem_cgroup_start_move(memcg);
3821 for_each_node_state(node, N_HIGH_MEMORY) {
3822 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3825 ret = mem_cgroup_force_empty_list(memcg,
3834 mem_cgroup_end_move(memcg);
3835 memcg_oom_recover(memcg);
3836 /* it seems parent cgroup doesn't have enough mem */
3840 /* "ret" should also be checked to ensure all lists are empty. */
3841 } while (memcg->res.usage > 0 || ret);
3843 css_put(&memcg->css);
3847 /* returns EBUSY if there is a task or if we come here twice. */
3848 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3852 /* we call try-to-free pages for make this cgroup empty */
3853 lru_add_drain_all();
3854 /* try to free all pages in this cgroup */
3856 while (nr_retries && memcg->res.usage > 0) {
3859 if (signal_pending(current)) {
3863 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3867 /* maybe some writeback is necessary */
3868 congestion_wait(BLK_RW_ASYNC, HZ/10);
3873 /* try move_account...there may be some *locked* pages. */
3877 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3879 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3883 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3885 return mem_cgroup_from_cont(cont)->use_hierarchy;
3888 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3892 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3893 struct cgroup *parent = cont->parent;
3894 struct mem_cgroup *parent_memcg = NULL;
3897 parent_memcg = mem_cgroup_from_cont(parent);
3901 * If parent's use_hierarchy is set, we can't make any modifications
3902 * in the child subtrees. If it is unset, then the change can
3903 * occur, provided the current cgroup has no children.
3905 * For the root cgroup, parent_mem is NULL, we allow value to be
3906 * set if there are no children.
3908 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3909 (val == 1 || val == 0)) {
3910 if (list_empty(&cont->children))
3911 memcg->use_hierarchy = val;
3922 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3923 enum mem_cgroup_stat_index idx)
3925 struct mem_cgroup *iter;
3928 /* Per-cpu values can be negative, use a signed accumulator */
3929 for_each_mem_cgroup_tree(iter, memcg)
3930 val += mem_cgroup_read_stat(iter, idx);
3932 if (val < 0) /* race ? */
3937 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3941 if (!mem_cgroup_is_root(memcg)) {
3943 return res_counter_read_u64(&memcg->res, RES_USAGE);
3945 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3948 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3949 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3952 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3954 return val << PAGE_SHIFT;
3957 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3959 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3963 type = MEMFILE_TYPE(cft->private);
3964 name = MEMFILE_ATTR(cft->private);
3967 if (name == RES_USAGE)
3968 val = mem_cgroup_usage(memcg, false);
3970 val = res_counter_read_u64(&memcg->res, name);
3973 if (name == RES_USAGE)
3974 val = mem_cgroup_usage(memcg, true);
3976 val = res_counter_read_u64(&memcg->memsw, name);
3985 * The user of this function is...
3988 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3991 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3993 unsigned long long val;
3996 type = MEMFILE_TYPE(cft->private);
3997 name = MEMFILE_ATTR(cft->private);
4000 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4004 /* This function does all necessary parse...reuse it */
4005 ret = res_counter_memparse_write_strategy(buffer, &val);
4009 ret = mem_cgroup_resize_limit(memcg, val);
4011 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4013 case RES_SOFT_LIMIT:
4014 ret = res_counter_memparse_write_strategy(buffer, &val);
4018 * For memsw, soft limits are hard to implement in terms
4019 * of semantics, for now, we support soft limits for
4020 * control without swap
4023 ret = res_counter_set_soft_limit(&memcg->res, val);
4028 ret = -EINVAL; /* should be BUG() ? */
4034 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4035 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4037 struct cgroup *cgroup;
4038 unsigned long long min_limit, min_memsw_limit, tmp;
4040 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4041 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4042 cgroup = memcg->css.cgroup;
4043 if (!memcg->use_hierarchy)
4046 while (cgroup->parent) {
4047 cgroup = cgroup->parent;
4048 memcg = mem_cgroup_from_cont(cgroup);
4049 if (!memcg->use_hierarchy)
4051 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4052 min_limit = min(min_limit, tmp);
4053 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4054 min_memsw_limit = min(min_memsw_limit, tmp);
4057 *mem_limit = min_limit;
4058 *memsw_limit = min_memsw_limit;
4062 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4064 struct mem_cgroup *memcg;
4067 memcg = mem_cgroup_from_cont(cont);
4068 type = MEMFILE_TYPE(event);
4069 name = MEMFILE_ATTR(event);
4073 res_counter_reset_max(&memcg->res);
4075 res_counter_reset_max(&memcg->memsw);
4079 res_counter_reset_failcnt(&memcg->res);
4081 res_counter_reset_failcnt(&memcg->memsw);
4088 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4091 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4095 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4096 struct cftype *cft, u64 val)
4098 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4100 if (val >= (1 << NR_MOVE_TYPE))
4103 * We check this value several times in both in can_attach() and
4104 * attach(), so we need cgroup lock to prevent this value from being
4108 memcg->move_charge_at_immigrate = val;
4114 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4115 struct cftype *cft, u64 val)
4122 /* For read statistics */
4140 struct mcs_total_stat {
4141 s64 stat[NR_MCS_STAT];
4147 } memcg_stat_strings[NR_MCS_STAT] = {
4148 {"cache", "total_cache"},
4149 {"rss", "total_rss"},
4150 {"mapped_file", "total_mapped_file"},
4151 {"pgpgin", "total_pgpgin"},
4152 {"pgpgout", "total_pgpgout"},
4153 {"swap", "total_swap"},
4154 {"pgfault", "total_pgfault"},
4155 {"pgmajfault", "total_pgmajfault"},
4156 {"inactive_anon", "total_inactive_anon"},
4157 {"active_anon", "total_active_anon"},
4158 {"inactive_file", "total_inactive_file"},
4159 {"active_file", "total_active_file"},
4160 {"unevictable", "total_unevictable"}
4165 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4170 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4171 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4172 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4173 s->stat[MCS_RSS] += val * PAGE_SIZE;
4174 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4175 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4176 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4177 s->stat[MCS_PGPGIN] += val;
4178 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4179 s->stat[MCS_PGPGOUT] += val;
4180 if (do_swap_account) {
4181 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4182 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4184 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4185 s->stat[MCS_PGFAULT] += val;
4186 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4187 s->stat[MCS_PGMAJFAULT] += val;
4190 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4191 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4192 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4193 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4194 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4195 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4196 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4197 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4198 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4199 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4203 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4205 struct mem_cgroup *iter;
4207 for_each_mem_cgroup_tree(iter, memcg)
4208 mem_cgroup_get_local_stat(iter, s);
4212 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4215 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4216 unsigned long node_nr;
4217 struct cgroup *cont = m->private;
4218 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4220 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4221 seq_printf(m, "total=%lu", total_nr);
4222 for_each_node_state(nid, N_HIGH_MEMORY) {
4223 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4224 seq_printf(m, " N%d=%lu", nid, node_nr);
4228 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4229 seq_printf(m, "file=%lu", file_nr);
4230 for_each_node_state(nid, N_HIGH_MEMORY) {
4231 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4233 seq_printf(m, " N%d=%lu", nid, node_nr);
4237 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4238 seq_printf(m, "anon=%lu", anon_nr);
4239 for_each_node_state(nid, N_HIGH_MEMORY) {
4240 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4242 seq_printf(m, " N%d=%lu", nid, node_nr);
4246 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4247 seq_printf(m, "unevictable=%lu", unevictable_nr);
4248 for_each_node_state(nid, N_HIGH_MEMORY) {
4249 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4250 BIT(LRU_UNEVICTABLE));
4251 seq_printf(m, " N%d=%lu", nid, node_nr);
4256 #endif /* CONFIG_NUMA */
4258 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4259 struct cgroup_map_cb *cb)
4261 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4262 struct mcs_total_stat mystat;
4265 memset(&mystat, 0, sizeof(mystat));
4266 mem_cgroup_get_local_stat(mem_cont, &mystat);
4269 for (i = 0; i < NR_MCS_STAT; i++) {
4270 if (i == MCS_SWAP && !do_swap_account)
4272 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4275 /* Hierarchical information */
4277 unsigned long long limit, memsw_limit;
4278 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4279 cb->fill(cb, "hierarchical_memory_limit", limit);
4280 if (do_swap_account)
4281 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4284 memset(&mystat, 0, sizeof(mystat));
4285 mem_cgroup_get_total_stat(mem_cont, &mystat);
4286 for (i = 0; i < NR_MCS_STAT; i++) {
4287 if (i == MCS_SWAP && !do_swap_account)
4289 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4292 #ifdef CONFIG_DEBUG_VM
4295 struct mem_cgroup_per_zone *mz;
4296 unsigned long recent_rotated[2] = {0, 0};
4297 unsigned long recent_scanned[2] = {0, 0};
4299 for_each_online_node(nid)
4300 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4301 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4303 recent_rotated[0] +=
4304 mz->reclaim_stat.recent_rotated[0];
4305 recent_rotated[1] +=
4306 mz->reclaim_stat.recent_rotated[1];
4307 recent_scanned[0] +=
4308 mz->reclaim_stat.recent_scanned[0];
4309 recent_scanned[1] +=
4310 mz->reclaim_stat.recent_scanned[1];
4312 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4313 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4314 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4315 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4322 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4324 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4326 return mem_cgroup_swappiness(memcg);
4329 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4332 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4333 struct mem_cgroup *parent;
4338 if (cgrp->parent == NULL)
4341 parent = mem_cgroup_from_cont(cgrp->parent);
4345 /* If under hierarchy, only empty-root can set this value */
4346 if ((parent->use_hierarchy) ||
4347 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4352 memcg->swappiness = val;
4359 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4361 struct mem_cgroup_threshold_ary *t;
4367 t = rcu_dereference(memcg->thresholds.primary);
4369 t = rcu_dereference(memcg->memsw_thresholds.primary);
4374 usage = mem_cgroup_usage(memcg, swap);
4377 * current_threshold points to threshold just below usage.
4378 * If it's not true, a threshold was crossed after last
4379 * call of __mem_cgroup_threshold().
4381 i = t->current_threshold;
4384 * Iterate backward over array of thresholds starting from
4385 * current_threshold and check if a threshold is crossed.
4386 * If none of thresholds below usage is crossed, we read
4387 * only one element of the array here.
4389 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4390 eventfd_signal(t->entries[i].eventfd, 1);
4392 /* i = current_threshold + 1 */
4396 * Iterate forward over array of thresholds starting from
4397 * current_threshold+1 and check if a threshold is crossed.
4398 * If none of thresholds above usage is crossed, we read
4399 * only one element of the array here.
4401 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4402 eventfd_signal(t->entries[i].eventfd, 1);
4404 /* Update current_threshold */
4405 t->current_threshold = i - 1;
4410 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4413 __mem_cgroup_threshold(memcg, false);
4414 if (do_swap_account)
4415 __mem_cgroup_threshold(memcg, true);
4417 memcg = parent_mem_cgroup(memcg);
4421 static int compare_thresholds(const void *a, const void *b)
4423 const struct mem_cgroup_threshold *_a = a;
4424 const struct mem_cgroup_threshold *_b = b;
4426 return _a->threshold - _b->threshold;
4429 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4431 struct mem_cgroup_eventfd_list *ev;
4433 list_for_each_entry(ev, &memcg->oom_notify, list)
4434 eventfd_signal(ev->eventfd, 1);
4438 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4440 struct mem_cgroup *iter;
4442 for_each_mem_cgroup_tree(iter, memcg)
4443 mem_cgroup_oom_notify_cb(iter);
4446 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4447 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4449 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4450 struct mem_cgroup_thresholds *thresholds;
4451 struct mem_cgroup_threshold_ary *new;
4452 int type = MEMFILE_TYPE(cft->private);
4453 u64 threshold, usage;
4456 ret = res_counter_memparse_write_strategy(args, &threshold);
4460 mutex_lock(&memcg->thresholds_lock);
4463 thresholds = &memcg->thresholds;
4464 else if (type == _MEMSWAP)
4465 thresholds = &memcg->memsw_thresholds;
4469 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4471 /* Check if a threshold crossed before adding a new one */
4472 if (thresholds->primary)
4473 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4475 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4477 /* Allocate memory for new array of thresholds */
4478 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4486 /* Copy thresholds (if any) to new array */
4487 if (thresholds->primary) {
4488 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4489 sizeof(struct mem_cgroup_threshold));
4492 /* Add new threshold */
4493 new->entries[size - 1].eventfd = eventfd;
4494 new->entries[size - 1].threshold = threshold;
4496 /* Sort thresholds. Registering of new threshold isn't time-critical */
4497 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4498 compare_thresholds, NULL);
4500 /* Find current threshold */
4501 new->current_threshold = -1;
4502 for (i = 0; i < size; i++) {
4503 if (new->entries[i].threshold < usage) {
4505 * new->current_threshold will not be used until
4506 * rcu_assign_pointer(), so it's safe to increment
4509 ++new->current_threshold;
4513 /* Free old spare buffer and save old primary buffer as spare */
4514 kfree(thresholds->spare);
4515 thresholds->spare = thresholds->primary;
4517 rcu_assign_pointer(thresholds->primary, new);
4519 /* To be sure that nobody uses thresholds */
4523 mutex_unlock(&memcg->thresholds_lock);
4528 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4529 struct cftype *cft, struct eventfd_ctx *eventfd)
4531 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4532 struct mem_cgroup_thresholds *thresholds;
4533 struct mem_cgroup_threshold_ary *new;
4534 int type = MEMFILE_TYPE(cft->private);
4538 mutex_lock(&memcg->thresholds_lock);
4540 thresholds = &memcg->thresholds;
4541 else if (type == _MEMSWAP)
4542 thresholds = &memcg->memsw_thresholds;
4547 * Something went wrong if we trying to unregister a threshold
4548 * if we don't have thresholds
4550 BUG_ON(!thresholds);
4552 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4554 /* Check if a threshold crossed before removing */
4555 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4557 /* Calculate new number of threshold */
4559 for (i = 0; i < thresholds->primary->size; i++) {
4560 if (thresholds->primary->entries[i].eventfd != eventfd)
4564 new = thresholds->spare;
4566 /* Set thresholds array to NULL if we don't have thresholds */
4575 /* Copy thresholds and find current threshold */
4576 new->current_threshold = -1;
4577 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4578 if (thresholds->primary->entries[i].eventfd == eventfd)
4581 new->entries[j] = thresholds->primary->entries[i];
4582 if (new->entries[j].threshold < usage) {
4584 * new->current_threshold will not be used
4585 * until rcu_assign_pointer(), so it's safe to increment
4588 ++new->current_threshold;
4594 /* Swap primary and spare array */
4595 thresholds->spare = thresholds->primary;
4596 rcu_assign_pointer(thresholds->primary, new);
4598 /* To be sure that nobody uses thresholds */
4601 mutex_unlock(&memcg->thresholds_lock);
4604 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4605 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4607 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4608 struct mem_cgroup_eventfd_list *event;
4609 int type = MEMFILE_TYPE(cft->private);
4611 BUG_ON(type != _OOM_TYPE);
4612 event = kmalloc(sizeof(*event), GFP_KERNEL);
4616 spin_lock(&memcg_oom_lock);
4618 event->eventfd = eventfd;
4619 list_add(&event->list, &memcg->oom_notify);
4621 /* already in OOM ? */
4622 if (atomic_read(&memcg->under_oom))
4623 eventfd_signal(eventfd, 1);
4624 spin_unlock(&memcg_oom_lock);
4629 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4630 struct cftype *cft, struct eventfd_ctx *eventfd)
4632 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4633 struct mem_cgroup_eventfd_list *ev, *tmp;
4634 int type = MEMFILE_TYPE(cft->private);
4636 BUG_ON(type != _OOM_TYPE);
4638 spin_lock(&memcg_oom_lock);
4640 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4641 if (ev->eventfd == eventfd) {
4642 list_del(&ev->list);
4647 spin_unlock(&memcg_oom_lock);
4650 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4651 struct cftype *cft, struct cgroup_map_cb *cb)
4653 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4655 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4657 if (atomic_read(&memcg->under_oom))
4658 cb->fill(cb, "under_oom", 1);
4660 cb->fill(cb, "under_oom", 0);
4664 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4665 struct cftype *cft, u64 val)
4667 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4668 struct mem_cgroup *parent;
4670 /* cannot set to root cgroup and only 0 and 1 are allowed */
4671 if (!cgrp->parent || !((val == 0) || (val == 1)))
4674 parent = mem_cgroup_from_cont(cgrp->parent);
4677 /* oom-kill-disable is a flag for subhierarchy. */
4678 if ((parent->use_hierarchy) ||
4679 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4683 memcg->oom_kill_disable = val;
4685 memcg_oom_recover(memcg);
4691 static const struct file_operations mem_control_numa_stat_file_operations = {
4693 .llseek = seq_lseek,
4694 .release = single_release,
4697 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4699 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4701 file->f_op = &mem_control_numa_stat_file_operations;
4702 return single_open(file, mem_control_numa_stat_show, cont);
4704 #endif /* CONFIG_NUMA */
4706 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4707 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4710 * Part of this would be better living in a separate allocation
4711 * function, leaving us with just the cgroup tree population work.
4712 * We, however, depend on state such as network's proto_list that
4713 * is only initialized after cgroup creation. I found the less
4714 * cumbersome way to deal with it to defer it all to populate time
4716 return mem_cgroup_sockets_init(cont, ss);
4719 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4720 struct cgroup *cont)
4722 mem_cgroup_sockets_destroy(cont, ss);
4725 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4730 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4731 struct cgroup *cont)
4736 static struct cftype mem_cgroup_files[] = {
4738 .name = "usage_in_bytes",
4739 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4740 .read_u64 = mem_cgroup_read,
4741 .register_event = mem_cgroup_usage_register_event,
4742 .unregister_event = mem_cgroup_usage_unregister_event,
4745 .name = "max_usage_in_bytes",
4746 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4747 .trigger = mem_cgroup_reset,
4748 .read_u64 = mem_cgroup_read,
4751 .name = "limit_in_bytes",
4752 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4753 .write_string = mem_cgroup_write,
4754 .read_u64 = mem_cgroup_read,
4757 .name = "soft_limit_in_bytes",
4758 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4759 .write_string = mem_cgroup_write,
4760 .read_u64 = mem_cgroup_read,
4764 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4765 .trigger = mem_cgroup_reset,
4766 .read_u64 = mem_cgroup_read,
4770 .read_map = mem_control_stat_show,
4773 .name = "force_empty",
4774 .trigger = mem_cgroup_force_empty_write,
4777 .name = "use_hierarchy",
4778 .write_u64 = mem_cgroup_hierarchy_write,
4779 .read_u64 = mem_cgroup_hierarchy_read,
4782 .name = "swappiness",
4783 .read_u64 = mem_cgroup_swappiness_read,
4784 .write_u64 = mem_cgroup_swappiness_write,
4787 .name = "move_charge_at_immigrate",
4788 .read_u64 = mem_cgroup_move_charge_read,
4789 .write_u64 = mem_cgroup_move_charge_write,
4792 .name = "oom_control",
4793 .read_map = mem_cgroup_oom_control_read,
4794 .write_u64 = mem_cgroup_oom_control_write,
4795 .register_event = mem_cgroup_oom_register_event,
4796 .unregister_event = mem_cgroup_oom_unregister_event,
4797 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4801 .name = "numa_stat",
4802 .open = mem_control_numa_stat_open,
4808 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4809 static struct cftype memsw_cgroup_files[] = {
4811 .name = "memsw.usage_in_bytes",
4812 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4813 .read_u64 = mem_cgroup_read,
4814 .register_event = mem_cgroup_usage_register_event,
4815 .unregister_event = mem_cgroup_usage_unregister_event,
4818 .name = "memsw.max_usage_in_bytes",
4819 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4820 .trigger = mem_cgroup_reset,
4821 .read_u64 = mem_cgroup_read,
4824 .name = "memsw.limit_in_bytes",
4825 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4826 .write_string = mem_cgroup_write,
4827 .read_u64 = mem_cgroup_read,
4830 .name = "memsw.failcnt",
4831 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4832 .trigger = mem_cgroup_reset,
4833 .read_u64 = mem_cgroup_read,
4837 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4839 if (!do_swap_account)
4841 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4842 ARRAY_SIZE(memsw_cgroup_files));
4845 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4851 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4853 struct mem_cgroup_per_node *pn;
4854 struct mem_cgroup_per_zone *mz;
4856 int zone, tmp = node;
4858 * This routine is called against possible nodes.
4859 * But it's BUG to call kmalloc() against offline node.
4861 * TODO: this routine can waste much memory for nodes which will
4862 * never be onlined. It's better to use memory hotplug callback
4865 if (!node_state(node, N_NORMAL_MEMORY))
4867 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4871 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4872 mz = &pn->zoneinfo[zone];
4874 INIT_LIST_HEAD(&mz->lists[l]);
4875 mz->usage_in_excess = 0;
4876 mz->on_tree = false;
4879 memcg->info.nodeinfo[node] = pn;
4883 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4885 kfree(memcg->info.nodeinfo[node]);
4888 static struct mem_cgroup *mem_cgroup_alloc(void)
4890 struct mem_cgroup *mem;
4891 int size = sizeof(struct mem_cgroup);
4893 /* Can be very big if MAX_NUMNODES is very big */
4894 if (size < PAGE_SIZE)
4895 mem = kzalloc(size, GFP_KERNEL);
4897 mem = vzalloc(size);
4902 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4905 spin_lock_init(&mem->pcp_counter_lock);
4909 if (size < PAGE_SIZE)
4917 * At destroying mem_cgroup, references from swap_cgroup can remain.
4918 * (scanning all at force_empty is too costly...)
4920 * Instead of clearing all references at force_empty, we remember
4921 * the number of reference from swap_cgroup and free mem_cgroup when
4922 * it goes down to 0.
4924 * Removal of cgroup itself succeeds regardless of refs from swap.
4927 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4931 mem_cgroup_remove_from_trees(memcg);
4932 free_css_id(&mem_cgroup_subsys, &memcg->css);
4934 for_each_node_state(node, N_POSSIBLE)
4935 free_mem_cgroup_per_zone_info(memcg, node);
4937 free_percpu(memcg->stat);
4938 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4944 static void mem_cgroup_get(struct mem_cgroup *memcg)
4946 atomic_inc(&memcg->refcnt);
4949 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4951 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4952 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4953 __mem_cgroup_free(memcg);
4955 mem_cgroup_put(parent);
4959 static void mem_cgroup_put(struct mem_cgroup *memcg)
4961 __mem_cgroup_put(memcg, 1);
4965 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4967 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4969 if (!memcg->res.parent)
4971 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4973 EXPORT_SYMBOL(parent_mem_cgroup);
4975 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4976 static void __init enable_swap_cgroup(void)
4978 if (!mem_cgroup_disabled() && really_do_swap_account)
4979 do_swap_account = 1;
4982 static void __init enable_swap_cgroup(void)
4987 static int mem_cgroup_soft_limit_tree_init(void)
4989 struct mem_cgroup_tree_per_node *rtpn;
4990 struct mem_cgroup_tree_per_zone *rtpz;
4991 int tmp, node, zone;
4993 for_each_node_state(node, N_POSSIBLE) {
4995 if (!node_state(node, N_NORMAL_MEMORY))
4997 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5001 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5003 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5004 rtpz = &rtpn->rb_tree_per_zone[zone];
5005 rtpz->rb_root = RB_ROOT;
5006 spin_lock_init(&rtpz->lock);
5012 static struct cgroup_subsys_state * __ref
5013 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5015 struct mem_cgroup *memcg, *parent;
5016 long error = -ENOMEM;
5019 memcg = mem_cgroup_alloc();
5021 return ERR_PTR(error);
5023 for_each_node_state(node, N_POSSIBLE)
5024 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5028 if (cont->parent == NULL) {
5030 enable_swap_cgroup();
5032 if (mem_cgroup_soft_limit_tree_init())
5034 root_mem_cgroup = memcg;
5035 for_each_possible_cpu(cpu) {
5036 struct memcg_stock_pcp *stock =
5037 &per_cpu(memcg_stock, cpu);
5038 INIT_WORK(&stock->work, drain_local_stock);
5040 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5042 parent = mem_cgroup_from_cont(cont->parent);
5043 memcg->use_hierarchy = parent->use_hierarchy;
5044 memcg->oom_kill_disable = parent->oom_kill_disable;
5047 if (parent && parent->use_hierarchy) {
5048 res_counter_init(&memcg->res, &parent->res);
5049 res_counter_init(&memcg->memsw, &parent->memsw);
5051 * We increment refcnt of the parent to ensure that we can
5052 * safely access it on res_counter_charge/uncharge.
5053 * This refcnt will be decremented when freeing this
5054 * mem_cgroup(see mem_cgroup_put).
5056 mem_cgroup_get(parent);
5058 res_counter_init(&memcg->res, NULL);
5059 res_counter_init(&memcg->memsw, NULL);
5061 memcg->last_scanned_node = MAX_NUMNODES;
5062 INIT_LIST_HEAD(&memcg->oom_notify);
5065 memcg->swappiness = mem_cgroup_swappiness(parent);
5066 atomic_set(&memcg->refcnt, 1);
5067 memcg->move_charge_at_immigrate = 0;
5068 mutex_init(&memcg->thresholds_lock);
5071 __mem_cgroup_free(memcg);
5072 return ERR_PTR(error);
5075 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5076 struct cgroup *cont)
5078 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5080 return mem_cgroup_force_empty(memcg, false);
5083 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5084 struct cgroup *cont)
5086 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5088 kmem_cgroup_destroy(ss, cont);
5090 mem_cgroup_put(memcg);
5093 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5094 struct cgroup *cont)
5098 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5099 ARRAY_SIZE(mem_cgroup_files));
5102 ret = register_memsw_files(cont, ss);
5105 ret = register_kmem_files(cont, ss);
5111 /* Handlers for move charge at task migration. */
5112 #define PRECHARGE_COUNT_AT_ONCE 256
5113 static int mem_cgroup_do_precharge(unsigned long count)
5116 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5117 struct mem_cgroup *memcg = mc.to;
5119 if (mem_cgroup_is_root(memcg)) {
5120 mc.precharge += count;
5121 /* we don't need css_get for root */
5124 /* try to charge at once */
5126 struct res_counter *dummy;
5128 * "memcg" cannot be under rmdir() because we've already checked
5129 * by cgroup_lock_live_cgroup() that it is not removed and we
5130 * are still under the same cgroup_mutex. So we can postpone
5133 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5135 if (do_swap_account && res_counter_charge(&memcg->memsw,
5136 PAGE_SIZE * count, &dummy)) {
5137 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5140 mc.precharge += count;
5144 /* fall back to one by one charge */
5146 if (signal_pending(current)) {
5150 if (!batch_count--) {
5151 batch_count = PRECHARGE_COUNT_AT_ONCE;
5154 ret = __mem_cgroup_try_charge(NULL,
5155 GFP_KERNEL, 1, &memcg, false);
5157 /* mem_cgroup_clear_mc() will do uncharge later */
5165 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5166 * @vma: the vma the pte to be checked belongs
5167 * @addr: the address corresponding to the pte to be checked
5168 * @ptent: the pte to be checked
5169 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5172 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5173 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5174 * move charge. if @target is not NULL, the page is stored in target->page
5175 * with extra refcnt got(Callers should handle it).
5176 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5177 * target for charge migration. if @target is not NULL, the entry is stored
5180 * Called with pte lock held.
5187 enum mc_target_type {
5188 MC_TARGET_NONE, /* not used */
5193 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5194 unsigned long addr, pte_t ptent)
5196 struct page *page = vm_normal_page(vma, addr, ptent);
5198 if (!page || !page_mapped(page))
5200 if (PageAnon(page)) {
5201 /* we don't move shared anon */
5202 if (!move_anon() || page_mapcount(page) > 2)
5204 } else if (!move_file())
5205 /* we ignore mapcount for file pages */
5207 if (!get_page_unless_zero(page))
5213 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5214 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5217 struct page *page = NULL;
5218 swp_entry_t ent = pte_to_swp_entry(ptent);
5220 if (!move_anon() || non_swap_entry(ent))
5222 usage_count = mem_cgroup_count_swap_user(ent, &page);
5223 if (usage_count > 1) { /* we don't move shared anon */
5228 if (do_swap_account)
5229 entry->val = ent.val;
5234 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5235 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5237 struct page *page = NULL;
5238 struct inode *inode;
5239 struct address_space *mapping;
5242 if (!vma->vm_file) /* anonymous vma */
5247 inode = vma->vm_file->f_path.dentry->d_inode;
5248 mapping = vma->vm_file->f_mapping;
5249 if (pte_none(ptent))
5250 pgoff = linear_page_index(vma, addr);
5251 else /* pte_file(ptent) is true */
5252 pgoff = pte_to_pgoff(ptent);
5254 /* page is moved even if it's not RSS of this task(page-faulted). */
5255 page = find_get_page(mapping, pgoff);
5258 /* shmem/tmpfs may report page out on swap: account for that too. */
5259 if (radix_tree_exceptional_entry(page)) {
5260 swp_entry_t swap = radix_to_swp_entry(page);
5261 if (do_swap_account)
5263 page = find_get_page(&swapper_space, swap.val);
5269 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5270 unsigned long addr, pte_t ptent, union mc_target *target)
5272 struct page *page = NULL;
5273 struct page_cgroup *pc;
5275 swp_entry_t ent = { .val = 0 };
5277 if (pte_present(ptent))
5278 page = mc_handle_present_pte(vma, addr, ptent);
5279 else if (is_swap_pte(ptent))
5280 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5281 else if (pte_none(ptent) || pte_file(ptent))
5282 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5284 if (!page && !ent.val)
5287 pc = lookup_page_cgroup(page);
5289 * Do only loose check w/o page_cgroup lock.
5290 * mem_cgroup_move_account() checks the pc is valid or not under
5293 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5294 ret = MC_TARGET_PAGE;
5296 target->page = page;
5298 if (!ret || !target)
5301 /* There is a swap entry and a page doesn't exist or isn't charged */
5302 if (ent.val && !ret &&
5303 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5304 ret = MC_TARGET_SWAP;
5311 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5312 unsigned long addr, unsigned long end,
5313 struct mm_walk *walk)
5315 struct vm_area_struct *vma = walk->private;
5319 split_huge_page_pmd(walk->mm, pmd);
5321 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5322 for (; addr != end; pte++, addr += PAGE_SIZE)
5323 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5324 mc.precharge++; /* increment precharge temporarily */
5325 pte_unmap_unlock(pte - 1, ptl);
5331 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5333 unsigned long precharge;
5334 struct vm_area_struct *vma;
5336 down_read(&mm->mmap_sem);
5337 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5338 struct mm_walk mem_cgroup_count_precharge_walk = {
5339 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5343 if (is_vm_hugetlb_page(vma))
5345 walk_page_range(vma->vm_start, vma->vm_end,
5346 &mem_cgroup_count_precharge_walk);
5348 up_read(&mm->mmap_sem);
5350 precharge = mc.precharge;
5356 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5358 unsigned long precharge = mem_cgroup_count_precharge(mm);
5360 VM_BUG_ON(mc.moving_task);
5361 mc.moving_task = current;
5362 return mem_cgroup_do_precharge(precharge);
5365 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5366 static void __mem_cgroup_clear_mc(void)
5368 struct mem_cgroup *from = mc.from;
5369 struct mem_cgroup *to = mc.to;
5371 /* we must uncharge all the leftover precharges from mc.to */
5373 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5377 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5378 * we must uncharge here.
5380 if (mc.moved_charge) {
5381 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5382 mc.moved_charge = 0;
5384 /* we must fixup refcnts and charges */
5385 if (mc.moved_swap) {
5386 /* uncharge swap account from the old cgroup */
5387 if (!mem_cgroup_is_root(mc.from))
5388 res_counter_uncharge(&mc.from->memsw,
5389 PAGE_SIZE * mc.moved_swap);
5390 __mem_cgroup_put(mc.from, mc.moved_swap);
5392 if (!mem_cgroup_is_root(mc.to)) {
5394 * we charged both to->res and to->memsw, so we should
5397 res_counter_uncharge(&mc.to->res,
5398 PAGE_SIZE * mc.moved_swap);
5400 /* we've already done mem_cgroup_get(mc.to) */
5403 memcg_oom_recover(from);
5404 memcg_oom_recover(to);
5405 wake_up_all(&mc.waitq);
5408 static void mem_cgroup_clear_mc(void)
5410 struct mem_cgroup *from = mc.from;
5413 * we must clear moving_task before waking up waiters at the end of
5416 mc.moving_task = NULL;
5417 __mem_cgroup_clear_mc();
5418 spin_lock(&mc.lock);
5421 spin_unlock(&mc.lock);
5422 mem_cgroup_end_move(from);
5425 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5426 struct cgroup *cgroup,
5427 struct cgroup_taskset *tset)
5429 struct task_struct *p = cgroup_taskset_first(tset);
5431 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5433 if (memcg->move_charge_at_immigrate) {
5434 struct mm_struct *mm;
5435 struct mem_cgroup *from = mem_cgroup_from_task(p);
5437 VM_BUG_ON(from == memcg);
5439 mm = get_task_mm(p);
5442 /* We move charges only when we move a owner of the mm */
5443 if (mm->owner == p) {
5446 VM_BUG_ON(mc.precharge);
5447 VM_BUG_ON(mc.moved_charge);
5448 VM_BUG_ON(mc.moved_swap);
5449 mem_cgroup_start_move(from);
5450 spin_lock(&mc.lock);
5453 spin_unlock(&mc.lock);
5454 /* We set mc.moving_task later */
5456 ret = mem_cgroup_precharge_mc(mm);
5458 mem_cgroup_clear_mc();
5465 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5466 struct cgroup *cgroup,
5467 struct cgroup_taskset *tset)
5469 mem_cgroup_clear_mc();
5472 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5473 unsigned long addr, unsigned long end,
5474 struct mm_walk *walk)
5477 struct vm_area_struct *vma = walk->private;
5481 split_huge_page_pmd(walk->mm, pmd);
5483 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5484 for (; addr != end; addr += PAGE_SIZE) {
5485 pte_t ptent = *(pte++);
5486 union mc_target target;
5489 struct page_cgroup *pc;
5495 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5497 case MC_TARGET_PAGE:
5499 if (isolate_lru_page(page))
5501 pc = lookup_page_cgroup(page);
5502 if (!mem_cgroup_move_account(page, 1, pc,
5503 mc.from, mc.to, false)) {
5505 /* we uncharge from mc.from later. */
5508 putback_lru_page(page);
5509 put: /* is_target_pte_for_mc() gets the page */
5512 case MC_TARGET_SWAP:
5514 if (!mem_cgroup_move_swap_account(ent,
5515 mc.from, mc.to, false)) {
5517 /* we fixup refcnts and charges later. */
5525 pte_unmap_unlock(pte - 1, ptl);
5530 * We have consumed all precharges we got in can_attach().
5531 * We try charge one by one, but don't do any additional
5532 * charges to mc.to if we have failed in charge once in attach()
5535 ret = mem_cgroup_do_precharge(1);
5543 static void mem_cgroup_move_charge(struct mm_struct *mm)
5545 struct vm_area_struct *vma;
5547 lru_add_drain_all();
5549 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5551 * Someone who are holding the mmap_sem might be waiting in
5552 * waitq. So we cancel all extra charges, wake up all waiters,
5553 * and retry. Because we cancel precharges, we might not be able
5554 * to move enough charges, but moving charge is a best-effort
5555 * feature anyway, so it wouldn't be a big problem.
5557 __mem_cgroup_clear_mc();
5561 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5563 struct mm_walk mem_cgroup_move_charge_walk = {
5564 .pmd_entry = mem_cgroup_move_charge_pte_range,
5568 if (is_vm_hugetlb_page(vma))
5570 ret = walk_page_range(vma->vm_start, vma->vm_end,
5571 &mem_cgroup_move_charge_walk);
5574 * means we have consumed all precharges and failed in
5575 * doing additional charge. Just abandon here.
5579 up_read(&mm->mmap_sem);
5582 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5583 struct cgroup *cont,
5584 struct cgroup_taskset *tset)
5586 struct task_struct *p = cgroup_taskset_first(tset);
5587 struct mm_struct *mm = get_task_mm(p);
5591 mem_cgroup_move_charge(mm);
5596 mem_cgroup_clear_mc();
5598 #else /* !CONFIG_MMU */
5599 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5600 struct cgroup *cgroup,
5601 struct cgroup_taskset *tset)
5605 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5606 struct cgroup *cgroup,
5607 struct cgroup_taskset *tset)
5610 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5611 struct cgroup *cont,
5612 struct cgroup_taskset *tset)
5617 struct cgroup_subsys mem_cgroup_subsys = {
5619 .subsys_id = mem_cgroup_subsys_id,
5620 .create = mem_cgroup_create,
5621 .pre_destroy = mem_cgroup_pre_destroy,
5622 .destroy = mem_cgroup_destroy,
5623 .populate = mem_cgroup_populate,
5624 .can_attach = mem_cgroup_can_attach,
5625 .cancel_attach = mem_cgroup_cancel_attach,
5626 .attach = mem_cgroup_move_task,
5631 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5632 static int __init enable_swap_account(char *s)
5634 /* consider enabled if no parameter or 1 is given */
5635 if (!strcmp(s, "1"))
5636 really_do_swap_account = 1;
5637 else if (!strcmp(s, "0"))
5638 really_do_swap_account = 0;
5641 __setup("swapaccount=", enable_swap_account);