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)
374 static void mem_cgroup_get(struct mem_cgroup *memcg);
375 static void mem_cgroup_put(struct mem_cgroup *memcg);
377 /* Writing them here to avoid exposing memcg's inner layout */
378 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
380 #include <net/sock.h>
383 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
384 void sock_update_memcg(struct sock *sk)
386 if (static_branch(&memcg_socket_limit_enabled)) {
387 struct mem_cgroup *memcg;
389 BUG_ON(!sk->sk_prot->proto_cgroup);
391 /* Socket cloning can throw us here with sk_cgrp already
392 * filled. It won't however, necessarily happen from
393 * process context. So the test for root memcg given
394 * the current task's memcg won't help us in this case.
396 * Respecting the original socket's memcg is a better
397 * decision in this case.
400 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
401 mem_cgroup_get(sk->sk_cgrp->memcg);
406 memcg = mem_cgroup_from_task(current);
407 if (!mem_cgroup_is_root(memcg)) {
408 mem_cgroup_get(memcg);
409 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
414 EXPORT_SYMBOL(sock_update_memcg);
416 void sock_release_memcg(struct sock *sk)
418 if (static_branch(&memcg_socket_limit_enabled) && sk->sk_cgrp) {
419 struct mem_cgroup *memcg;
420 WARN_ON(!sk->sk_cgrp->memcg);
421 memcg = sk->sk_cgrp->memcg;
422 mem_cgroup_put(memcg);
426 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
428 if (!memcg || mem_cgroup_is_root(memcg))
431 return &memcg->tcp_mem.cg_proto;
433 EXPORT_SYMBOL(tcp_proto_cgroup);
434 #endif /* CONFIG_INET */
435 #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
437 static void drain_all_stock_async(struct mem_cgroup *memcg);
439 static struct mem_cgroup_per_zone *
440 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
442 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
445 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
450 static struct mem_cgroup_per_zone *
451 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
453 int nid = page_to_nid(page);
454 int zid = page_zonenum(page);
456 return mem_cgroup_zoneinfo(memcg, nid, zid);
459 static struct mem_cgroup_tree_per_zone *
460 soft_limit_tree_node_zone(int nid, int zid)
462 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
465 static struct mem_cgroup_tree_per_zone *
466 soft_limit_tree_from_page(struct page *page)
468 int nid = page_to_nid(page);
469 int zid = page_zonenum(page);
471 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
475 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
476 struct mem_cgroup_per_zone *mz,
477 struct mem_cgroup_tree_per_zone *mctz,
478 unsigned long long new_usage_in_excess)
480 struct rb_node **p = &mctz->rb_root.rb_node;
481 struct rb_node *parent = NULL;
482 struct mem_cgroup_per_zone *mz_node;
487 mz->usage_in_excess = new_usage_in_excess;
488 if (!mz->usage_in_excess)
492 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
494 if (mz->usage_in_excess < mz_node->usage_in_excess)
497 * We can't avoid mem cgroups that are over their soft
498 * limit by the same amount
500 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
503 rb_link_node(&mz->tree_node, parent, p);
504 rb_insert_color(&mz->tree_node, &mctz->rb_root);
509 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
510 struct mem_cgroup_per_zone *mz,
511 struct mem_cgroup_tree_per_zone *mctz)
515 rb_erase(&mz->tree_node, &mctz->rb_root);
520 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
521 struct mem_cgroup_per_zone *mz,
522 struct mem_cgroup_tree_per_zone *mctz)
524 spin_lock(&mctz->lock);
525 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
526 spin_unlock(&mctz->lock);
530 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
532 unsigned long long excess;
533 struct mem_cgroup_per_zone *mz;
534 struct mem_cgroup_tree_per_zone *mctz;
535 int nid = page_to_nid(page);
536 int zid = page_zonenum(page);
537 mctz = soft_limit_tree_from_page(page);
540 * Necessary to update all ancestors when hierarchy is used.
541 * because their event counter is not touched.
543 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
544 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
545 excess = res_counter_soft_limit_excess(&memcg->res);
547 * We have to update the tree if mz is on RB-tree or
548 * mem is over its softlimit.
550 if (excess || mz->on_tree) {
551 spin_lock(&mctz->lock);
552 /* if on-tree, remove it */
554 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
556 * Insert again. mz->usage_in_excess will be updated.
557 * If excess is 0, no tree ops.
559 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
560 spin_unlock(&mctz->lock);
565 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
568 struct mem_cgroup_per_zone *mz;
569 struct mem_cgroup_tree_per_zone *mctz;
571 for_each_node_state(node, N_POSSIBLE) {
572 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
573 mz = mem_cgroup_zoneinfo(memcg, node, zone);
574 mctz = soft_limit_tree_node_zone(node, zone);
575 mem_cgroup_remove_exceeded(memcg, mz, mctz);
580 static struct mem_cgroup_per_zone *
581 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
583 struct rb_node *rightmost = NULL;
584 struct mem_cgroup_per_zone *mz;
588 rightmost = rb_last(&mctz->rb_root);
590 goto done; /* Nothing to reclaim from */
592 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
594 * Remove the node now but someone else can add it back,
595 * we will to add it back at the end of reclaim to its correct
596 * position in the tree.
598 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
599 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
600 !css_tryget(&mz->mem->css))
606 static struct mem_cgroup_per_zone *
607 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
609 struct mem_cgroup_per_zone *mz;
611 spin_lock(&mctz->lock);
612 mz = __mem_cgroup_largest_soft_limit_node(mctz);
613 spin_unlock(&mctz->lock);
618 * Implementation Note: reading percpu statistics for memcg.
620 * Both of vmstat[] and percpu_counter has threshold and do periodic
621 * synchronization to implement "quick" read. There are trade-off between
622 * reading cost and precision of value. Then, we may have a chance to implement
623 * a periodic synchronizion of counter in memcg's counter.
625 * But this _read() function is used for user interface now. The user accounts
626 * memory usage by memory cgroup and he _always_ requires exact value because
627 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
628 * have to visit all online cpus and make sum. So, for now, unnecessary
629 * synchronization is not implemented. (just implemented for cpu hotplug)
631 * If there are kernel internal actions which can make use of some not-exact
632 * value, and reading all cpu value can be performance bottleneck in some
633 * common workload, threashold and synchonization as vmstat[] should be
636 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
637 enum mem_cgroup_stat_index idx)
643 for_each_online_cpu(cpu)
644 val += per_cpu(memcg->stat->count[idx], cpu);
645 #ifdef CONFIG_HOTPLUG_CPU
646 spin_lock(&memcg->pcp_counter_lock);
647 val += memcg->nocpu_base.count[idx];
648 spin_unlock(&memcg->pcp_counter_lock);
654 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
657 int val = (charge) ? 1 : -1;
658 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
661 void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val)
663 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
666 void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val)
668 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
671 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
672 enum mem_cgroup_events_index idx)
674 unsigned long val = 0;
677 for_each_online_cpu(cpu)
678 val += per_cpu(memcg->stat->events[idx], cpu);
679 #ifdef CONFIG_HOTPLUG_CPU
680 spin_lock(&memcg->pcp_counter_lock);
681 val += memcg->nocpu_base.events[idx];
682 spin_unlock(&memcg->pcp_counter_lock);
687 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
688 bool file, int nr_pages)
693 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
696 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
699 /* pagein of a big page is an event. So, ignore page size */
701 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
703 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
704 nr_pages = -nr_pages; /* for event */
707 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
713 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
714 unsigned int lru_mask)
716 struct mem_cgroup_per_zone *mz;
718 unsigned long ret = 0;
720 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
723 if (BIT(l) & lru_mask)
724 ret += MEM_CGROUP_ZSTAT(mz, l);
730 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
731 int nid, unsigned int lru_mask)
736 for (zid = 0; zid < MAX_NR_ZONES; zid++)
737 total += mem_cgroup_zone_nr_lru_pages(memcg,
743 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
744 unsigned int lru_mask)
749 for_each_node_state(nid, N_HIGH_MEMORY)
750 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
754 static bool __memcg_event_check(struct mem_cgroup *memcg, int target)
756 unsigned long val, next;
758 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
759 next = __this_cpu_read(memcg->stat->targets[target]);
760 /* from time_after() in jiffies.h */
761 return ((long)next - (long)val < 0);
764 static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
766 unsigned long val, next;
768 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
771 case MEM_CGROUP_TARGET_THRESH:
772 next = val + THRESHOLDS_EVENTS_TARGET;
774 case MEM_CGROUP_TARGET_SOFTLIMIT:
775 next = val + SOFTLIMIT_EVENTS_TARGET;
777 case MEM_CGROUP_TARGET_NUMAINFO:
778 next = val + NUMAINFO_EVENTS_TARGET;
784 __this_cpu_write(memcg->stat->targets[target], next);
788 * Check events in order.
791 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
794 /* threshold event is triggered in finer grain than soft limit */
795 if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
796 mem_cgroup_threshold(memcg);
797 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
798 if (unlikely(__memcg_event_check(memcg,
799 MEM_CGROUP_TARGET_SOFTLIMIT))) {
800 mem_cgroup_update_tree(memcg, page);
801 __mem_cgroup_target_update(memcg,
802 MEM_CGROUP_TARGET_SOFTLIMIT);
805 if (unlikely(__memcg_event_check(memcg,
806 MEM_CGROUP_TARGET_NUMAINFO))) {
807 atomic_inc(&memcg->numainfo_events);
808 __mem_cgroup_target_update(memcg,
809 MEM_CGROUP_TARGET_NUMAINFO);
816 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
818 return container_of(cgroup_subsys_state(cont,
819 mem_cgroup_subsys_id), struct mem_cgroup,
823 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
826 * mm_update_next_owner() may clear mm->owner to NULL
827 * if it races with swapoff, page migration, etc.
828 * So this can be called with p == NULL.
833 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
834 struct mem_cgroup, css);
837 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
839 struct mem_cgroup *memcg = NULL;
844 * Because we have no locks, mm->owner's may be being moved to other
845 * cgroup. We use css_tryget() here even if this looks
846 * pessimistic (rather than adding locks here).
850 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
851 if (unlikely(!memcg))
853 } while (!css_tryget(&memcg->css));
859 * mem_cgroup_iter - iterate over memory cgroup hierarchy
860 * @root: hierarchy root
861 * @prev: previously returned memcg, NULL on first invocation
862 * @reclaim: cookie for shared reclaim walks, NULL for full walks
864 * Returns references to children of the hierarchy below @root, or
865 * @root itself, or %NULL after a full round-trip.
867 * Caller must pass the return value in @prev on subsequent
868 * invocations for reference counting, or use mem_cgroup_iter_break()
869 * to cancel a hierarchy walk before the round-trip is complete.
871 * Reclaimers can specify a zone and a priority level in @reclaim to
872 * divide up the memcgs in the hierarchy among all concurrent
873 * reclaimers operating on the same zone and priority.
875 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
876 struct mem_cgroup *prev,
877 struct mem_cgroup_reclaim_cookie *reclaim)
879 struct mem_cgroup *memcg = NULL;
882 if (mem_cgroup_disabled())
886 root = root_mem_cgroup;
888 if (prev && !reclaim)
889 id = css_id(&prev->css);
891 if (prev && prev != root)
894 if (!root->use_hierarchy && root != root_mem_cgroup) {
901 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
902 struct cgroup_subsys_state *css;
905 int nid = zone_to_nid(reclaim->zone);
906 int zid = zone_idx(reclaim->zone);
907 struct mem_cgroup_per_zone *mz;
909 mz = mem_cgroup_zoneinfo(root, nid, zid);
910 iter = &mz->reclaim_iter[reclaim->priority];
911 if (prev && reclaim->generation != iter->generation)
917 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
919 if (css == &root->css || css_tryget(css))
920 memcg = container_of(css,
921 struct mem_cgroup, css);
930 else if (!prev && memcg)
931 reclaim->generation = iter->generation;
941 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
942 * @root: hierarchy root
943 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
945 void mem_cgroup_iter_break(struct mem_cgroup *root,
946 struct mem_cgroup *prev)
949 root = root_mem_cgroup;
950 if (prev && prev != root)
955 * Iteration constructs for visiting all cgroups (under a tree). If
956 * loops are exited prematurely (break), mem_cgroup_iter_break() must
957 * be used for reference counting.
959 #define for_each_mem_cgroup_tree(iter, root) \
960 for (iter = mem_cgroup_iter(root, NULL, NULL); \
962 iter = mem_cgroup_iter(root, iter, NULL))
964 #define for_each_mem_cgroup(iter) \
965 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
967 iter = mem_cgroup_iter(NULL, iter, NULL))
969 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
971 return (memcg == root_mem_cgroup);
974 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
976 struct mem_cgroup *memcg;
982 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
983 if (unlikely(!memcg))
988 mem_cgroup_pgmajfault(memcg, 1);
991 mem_cgroup_pgfault(memcg, 1);
999 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1002 * Following LRU functions are allowed to be used without PCG_LOCK.
1003 * Operations are called by routine of global LRU independently from memcg.
1004 * What we have to take care of here is validness of pc->mem_cgroup.
1006 * Changes to pc->mem_cgroup happens when
1009 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1010 * It is added to LRU before charge.
1011 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1012 * When moving account, the page is not on LRU. It's isolated.
1015 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
1017 struct page_cgroup *pc;
1018 struct mem_cgroup_per_zone *mz;
1020 if (mem_cgroup_disabled())
1022 pc = lookup_page_cgroup(page);
1023 /* can happen while we handle swapcache. */
1024 if (!TestClearPageCgroupAcctLRU(pc))
1026 VM_BUG_ON(!pc->mem_cgroup);
1028 * We don't check PCG_USED bit. It's cleared when the "page" is finally
1029 * removed from global LRU.
1031 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1032 /* huge page split is done under lru_lock. so, we have no races. */
1033 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
1034 VM_BUG_ON(list_empty(&pc->lru));
1035 list_del_init(&pc->lru);
1038 void mem_cgroup_del_lru(struct page *page)
1040 mem_cgroup_del_lru_list(page, page_lru(page));
1044 * Writeback is about to end against a page which has been marked for immediate
1045 * reclaim. If it still appears to be reclaimable, move it to the tail of the
1048 void mem_cgroup_rotate_reclaimable_page(struct page *page)
1050 struct mem_cgroup_per_zone *mz;
1051 struct page_cgroup *pc;
1052 enum lru_list lru = page_lru(page);
1054 if (mem_cgroup_disabled())
1057 pc = lookup_page_cgroup(page);
1058 /* unused page is not rotated. */
1059 if (!PageCgroupUsed(pc))
1061 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1063 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1064 list_move_tail(&pc->lru, &mz->lists[lru]);
1067 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1069 struct mem_cgroup_per_zone *mz;
1070 struct page_cgroup *pc;
1072 if (mem_cgroup_disabled())
1075 pc = lookup_page_cgroup(page);
1076 /* unused page is not rotated. */
1077 if (!PageCgroupUsed(pc))
1079 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1081 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1082 list_move(&pc->lru, &mz->lists[lru]);
1085 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1087 struct page_cgroup *pc;
1088 struct mem_cgroup_per_zone *mz;
1090 if (mem_cgroup_disabled())
1092 pc = lookup_page_cgroup(page);
1093 VM_BUG_ON(PageCgroupAcctLRU(pc));
1096 * SetPageLRU SetPageCgroupUsed
1098 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1100 * Ensure that one of the two sides adds the page to the memcg
1101 * LRU during a race.
1104 if (!PageCgroupUsed(pc))
1106 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1108 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1109 /* huge page split is done under lru_lock. so, we have no races. */
1110 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1111 SetPageCgroupAcctLRU(pc);
1112 list_add(&pc->lru, &mz->lists[lru]);
1116 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1117 * while it's linked to lru because the page may be reused after it's fully
1118 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1119 * It's done under lock_page and expected that zone->lru_lock isnever held.
1121 static void mem_cgroup_lru_del_before_commit(struct page *page)
1123 unsigned long flags;
1124 struct zone *zone = page_zone(page);
1125 struct page_cgroup *pc = lookup_page_cgroup(page);
1128 * Doing this check without taking ->lru_lock seems wrong but this
1129 * is safe. Because if page_cgroup's USED bit is unset, the page
1130 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1131 * set, the commit after this will fail, anyway.
1132 * This all charge/uncharge is done under some mutual execustion.
1133 * So, we don't need to taking care of changes in USED bit.
1135 if (likely(!PageLRU(page)))
1138 spin_lock_irqsave(&zone->lru_lock, flags);
1140 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1141 * is guarded by lock_page() because the page is SwapCache.
1143 if (!PageCgroupUsed(pc))
1144 mem_cgroup_del_lru_list(page, page_lru(page));
1145 spin_unlock_irqrestore(&zone->lru_lock, flags);
1148 static void mem_cgroup_lru_add_after_commit(struct page *page)
1150 unsigned long flags;
1151 struct zone *zone = page_zone(page);
1152 struct page_cgroup *pc = lookup_page_cgroup(page);
1155 * SetPageLRU SetPageCgroupUsed
1157 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1159 * Ensure that one of the two sides adds the page to the memcg
1160 * LRU during a race.
1163 /* taking care of that the page is added to LRU while we commit it */
1164 if (likely(!PageLRU(page)))
1166 spin_lock_irqsave(&zone->lru_lock, flags);
1167 /* link when the page is linked to LRU but page_cgroup isn't */
1168 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1169 mem_cgroup_add_lru_list(page, page_lru(page));
1170 spin_unlock_irqrestore(&zone->lru_lock, flags);
1174 void mem_cgroup_move_lists(struct page *page,
1175 enum lru_list from, enum lru_list to)
1177 if (mem_cgroup_disabled())
1179 mem_cgroup_del_lru_list(page, from);
1180 mem_cgroup_add_lru_list(page, to);
1184 * Checks whether given mem is same or in the root_mem_cgroup's
1187 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1188 struct mem_cgroup *memcg)
1190 if (root_memcg != memcg) {
1191 return (root_memcg->use_hierarchy &&
1192 css_is_ancestor(&memcg->css, &root_memcg->css));
1198 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1201 struct mem_cgroup *curr = NULL;
1202 struct task_struct *p;
1204 p = find_lock_task_mm(task);
1207 curr = try_get_mem_cgroup_from_mm(p->mm);
1212 * We should check use_hierarchy of "memcg" not "curr". Because checking
1213 * use_hierarchy of "curr" here make this function true if hierarchy is
1214 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1215 * hierarchy(even if use_hierarchy is disabled in "memcg").
1217 ret = mem_cgroup_same_or_subtree(memcg, curr);
1218 css_put(&curr->css);
1222 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1224 unsigned long inactive_ratio;
1225 int nid = zone_to_nid(zone);
1226 int zid = zone_idx(zone);
1227 unsigned long inactive;
1228 unsigned long active;
1231 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1232 BIT(LRU_INACTIVE_ANON));
1233 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1234 BIT(LRU_ACTIVE_ANON));
1236 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1238 inactive_ratio = int_sqrt(10 * gb);
1242 return inactive * inactive_ratio < active;
1245 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1247 unsigned long active;
1248 unsigned long inactive;
1249 int zid = zone_idx(zone);
1250 int nid = zone_to_nid(zone);
1252 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1253 BIT(LRU_INACTIVE_FILE));
1254 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1255 BIT(LRU_ACTIVE_FILE));
1257 return (active > inactive);
1260 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1263 int nid = zone_to_nid(zone);
1264 int zid = zone_idx(zone);
1265 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1267 return &mz->reclaim_stat;
1270 struct zone_reclaim_stat *
1271 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1273 struct page_cgroup *pc;
1274 struct mem_cgroup_per_zone *mz;
1276 if (mem_cgroup_disabled())
1279 pc = lookup_page_cgroup(page);
1280 if (!PageCgroupUsed(pc))
1282 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1284 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1285 return &mz->reclaim_stat;
1288 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1289 struct list_head *dst,
1290 unsigned long *scanned, int order,
1291 isolate_mode_t mode,
1293 struct mem_cgroup *mem_cont,
1294 int active, int file)
1296 unsigned long nr_taken = 0;
1300 struct list_head *src;
1301 struct page_cgroup *pc, *tmp;
1302 int nid = zone_to_nid(z);
1303 int zid = zone_idx(z);
1304 struct mem_cgroup_per_zone *mz;
1305 int lru = LRU_FILE * file + active;
1309 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1310 src = &mz->lists[lru];
1313 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1314 if (scan >= nr_to_scan)
1317 if (unlikely(!PageCgroupUsed(pc)))
1320 page = lookup_cgroup_page(pc);
1322 if (unlikely(!PageLRU(page)))
1326 ret = __isolate_lru_page(page, mode, file);
1329 list_move(&page->lru, dst);
1330 mem_cgroup_del_lru(page);
1331 nr_taken += hpage_nr_pages(page);
1334 /* we don't affect global LRU but rotate in our LRU */
1335 mem_cgroup_rotate_lru_list(page, page_lru(page));
1344 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1350 #define mem_cgroup_from_res_counter(counter, member) \
1351 container_of(counter, struct mem_cgroup, member)
1354 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1355 * @mem: the memory cgroup
1357 * Returns the maximum amount of memory @mem can be charged with, in
1360 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1362 unsigned long long margin;
1364 margin = res_counter_margin(&memcg->res);
1365 if (do_swap_account)
1366 margin = min(margin, res_counter_margin(&memcg->memsw));
1367 return margin >> PAGE_SHIFT;
1370 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1372 struct cgroup *cgrp = memcg->css.cgroup;
1375 if (cgrp->parent == NULL)
1376 return vm_swappiness;
1378 return memcg->swappiness;
1381 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1386 spin_lock(&memcg->pcp_counter_lock);
1387 for_each_online_cpu(cpu)
1388 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1389 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1390 spin_unlock(&memcg->pcp_counter_lock);
1396 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1403 spin_lock(&memcg->pcp_counter_lock);
1404 for_each_online_cpu(cpu)
1405 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1406 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1407 spin_unlock(&memcg->pcp_counter_lock);
1411 * 2 routines for checking "mem" is under move_account() or not.
1413 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1414 * for avoiding race in accounting. If true,
1415 * pc->mem_cgroup may be overwritten.
1417 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1418 * under hierarchy of moving cgroups. This is for
1419 * waiting at hith-memory prressure caused by "move".
1422 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1424 VM_BUG_ON(!rcu_read_lock_held());
1425 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1428 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1430 struct mem_cgroup *from;
1431 struct mem_cgroup *to;
1434 * Unlike task_move routines, we access mc.to, mc.from not under
1435 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1437 spin_lock(&mc.lock);
1443 ret = mem_cgroup_same_or_subtree(memcg, from)
1444 || mem_cgroup_same_or_subtree(memcg, to);
1446 spin_unlock(&mc.lock);
1450 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1452 if (mc.moving_task && current != mc.moving_task) {
1453 if (mem_cgroup_under_move(memcg)) {
1455 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1456 /* moving charge context might have finished. */
1459 finish_wait(&mc.waitq, &wait);
1467 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1468 * @memcg: The memory cgroup that went over limit
1469 * @p: Task that is going to be killed
1471 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1474 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1476 struct cgroup *task_cgrp;
1477 struct cgroup *mem_cgrp;
1479 * Need a buffer in BSS, can't rely on allocations. The code relies
1480 * on the assumption that OOM is serialized for memory controller.
1481 * If this assumption is broken, revisit this code.
1483 static char memcg_name[PATH_MAX];
1492 mem_cgrp = memcg->css.cgroup;
1493 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1495 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1498 * Unfortunately, we are unable to convert to a useful name
1499 * But we'll still print out the usage information
1506 printk(KERN_INFO "Task in %s killed", memcg_name);
1509 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1517 * Continues from above, so we don't need an KERN_ level
1519 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1522 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1523 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1524 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1525 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1526 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1528 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1529 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1530 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1534 * This function returns the number of memcg under hierarchy tree. Returns
1535 * 1(self count) if no children.
1537 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1540 struct mem_cgroup *iter;
1542 for_each_mem_cgroup_tree(iter, memcg)
1548 * Return the memory (and swap, if configured) limit for a memcg.
1550 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1555 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1556 limit += total_swap_pages << PAGE_SHIFT;
1558 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1560 * If memsw is finite and limits the amount of swap space available
1561 * to this memcg, return that limit.
1563 return min(limit, memsw);
1566 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1568 unsigned long flags)
1570 unsigned long total = 0;
1571 bool noswap = false;
1574 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1576 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1579 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1581 drain_all_stock_async(memcg);
1582 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1584 * Allow limit shrinkers, which are triggered directly
1585 * by userspace, to catch signals and stop reclaim
1586 * after minimal progress, regardless of the margin.
1588 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1590 if (mem_cgroup_margin(memcg))
1593 * If nothing was reclaimed after two attempts, there
1594 * may be no reclaimable pages in this hierarchy.
1603 * test_mem_cgroup_node_reclaimable
1604 * @mem: the target memcg
1605 * @nid: the node ID to be checked.
1606 * @noswap : specify true here if the user wants flle only information.
1608 * This function returns whether the specified memcg contains any
1609 * reclaimable pages on a node. Returns true if there are any reclaimable
1610 * pages in the node.
1612 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1613 int nid, bool noswap)
1615 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1617 if (noswap || !total_swap_pages)
1619 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1624 #if MAX_NUMNODES > 1
1627 * Always updating the nodemask is not very good - even if we have an empty
1628 * list or the wrong list here, we can start from some node and traverse all
1629 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1632 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1636 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1637 * pagein/pageout changes since the last update.
1639 if (!atomic_read(&memcg->numainfo_events))
1641 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1644 /* make a nodemask where this memcg uses memory from */
1645 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1647 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1649 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1650 node_clear(nid, memcg->scan_nodes);
1653 atomic_set(&memcg->numainfo_events, 0);
1654 atomic_set(&memcg->numainfo_updating, 0);
1658 * Selecting a node where we start reclaim from. Because what we need is just
1659 * reducing usage counter, start from anywhere is O,K. Considering
1660 * memory reclaim from current node, there are pros. and cons.
1662 * Freeing memory from current node means freeing memory from a node which
1663 * we'll use or we've used. So, it may make LRU bad. And if several threads
1664 * hit limits, it will see a contention on a node. But freeing from remote
1665 * node means more costs for memory reclaim because of memory latency.
1667 * Now, we use round-robin. Better algorithm is welcomed.
1669 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1673 mem_cgroup_may_update_nodemask(memcg);
1674 node = memcg->last_scanned_node;
1676 node = next_node(node, memcg->scan_nodes);
1677 if (node == MAX_NUMNODES)
1678 node = first_node(memcg->scan_nodes);
1680 * We call this when we hit limit, not when pages are added to LRU.
1681 * No LRU may hold pages because all pages are UNEVICTABLE or
1682 * memcg is too small and all pages are not on LRU. In that case,
1683 * we use curret node.
1685 if (unlikely(node == MAX_NUMNODES))
1686 node = numa_node_id();
1688 memcg->last_scanned_node = node;
1693 * Check all nodes whether it contains reclaimable pages or not.
1694 * For quick scan, we make use of scan_nodes. This will allow us to skip
1695 * unused nodes. But scan_nodes is lazily updated and may not cotain
1696 * enough new information. We need to do double check.
1698 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1703 * quick check...making use of scan_node.
1704 * We can skip unused nodes.
1706 if (!nodes_empty(memcg->scan_nodes)) {
1707 for (nid = first_node(memcg->scan_nodes);
1709 nid = next_node(nid, memcg->scan_nodes)) {
1711 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1716 * Check rest of nodes.
1718 for_each_node_state(nid, N_HIGH_MEMORY) {
1719 if (node_isset(nid, memcg->scan_nodes))
1721 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1728 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1733 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1735 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1739 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1742 unsigned long *total_scanned)
1744 struct mem_cgroup *victim = NULL;
1747 unsigned long excess;
1748 unsigned long nr_scanned;
1749 struct mem_cgroup_reclaim_cookie reclaim = {
1754 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1757 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1762 * If we have not been able to reclaim
1763 * anything, it might because there are
1764 * no reclaimable pages under this hierarchy
1769 * We want to do more targeted reclaim.
1770 * excess >> 2 is not to excessive so as to
1771 * reclaim too much, nor too less that we keep
1772 * coming back to reclaim from this cgroup
1774 if (total >= (excess >> 2) ||
1775 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1780 if (!mem_cgroup_reclaimable(victim, false))
1782 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1784 *total_scanned += nr_scanned;
1785 if (!res_counter_soft_limit_excess(&root_memcg->res))
1788 mem_cgroup_iter_break(root_memcg, victim);
1793 * Check OOM-Killer is already running under our hierarchy.
1794 * If someone is running, return false.
1795 * Has to be called with memcg_oom_lock
1797 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1799 struct mem_cgroup *iter, *failed = NULL;
1801 for_each_mem_cgroup_tree(iter, memcg) {
1802 if (iter->oom_lock) {
1804 * this subtree of our hierarchy is already locked
1805 * so we cannot give a lock.
1808 mem_cgroup_iter_break(memcg, iter);
1811 iter->oom_lock = true;
1818 * OK, we failed to lock the whole subtree so we have to clean up
1819 * what we set up to the failing subtree
1821 for_each_mem_cgroup_tree(iter, memcg) {
1822 if (iter == failed) {
1823 mem_cgroup_iter_break(memcg, iter);
1826 iter->oom_lock = false;
1832 * Has to be called with memcg_oom_lock
1834 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1836 struct mem_cgroup *iter;
1838 for_each_mem_cgroup_tree(iter, memcg)
1839 iter->oom_lock = false;
1843 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1845 struct mem_cgroup *iter;
1847 for_each_mem_cgroup_tree(iter, memcg)
1848 atomic_inc(&iter->under_oom);
1851 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1853 struct mem_cgroup *iter;
1856 * When a new child is created while the hierarchy is under oom,
1857 * mem_cgroup_oom_lock() may not be called. We have to use
1858 * atomic_add_unless() here.
1860 for_each_mem_cgroup_tree(iter, memcg)
1861 atomic_add_unless(&iter->under_oom, -1, 0);
1864 static DEFINE_SPINLOCK(memcg_oom_lock);
1865 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1867 struct oom_wait_info {
1868 struct mem_cgroup *mem;
1872 static int memcg_oom_wake_function(wait_queue_t *wait,
1873 unsigned mode, int sync, void *arg)
1875 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1877 struct oom_wait_info *oom_wait_info;
1879 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1880 oom_wait_memcg = oom_wait_info->mem;
1883 * Both of oom_wait_info->mem and wake_mem are stable under us.
1884 * Then we can use css_is_ancestor without taking care of RCU.
1886 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1887 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1889 return autoremove_wake_function(wait, mode, sync, arg);
1892 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1894 /* for filtering, pass "memcg" as argument. */
1895 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1898 static void memcg_oom_recover(struct mem_cgroup *memcg)
1900 if (memcg && atomic_read(&memcg->under_oom))
1901 memcg_wakeup_oom(memcg);
1905 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1907 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1909 struct oom_wait_info owait;
1910 bool locked, need_to_kill;
1913 owait.wait.flags = 0;
1914 owait.wait.func = memcg_oom_wake_function;
1915 owait.wait.private = current;
1916 INIT_LIST_HEAD(&owait.wait.task_list);
1917 need_to_kill = true;
1918 mem_cgroup_mark_under_oom(memcg);
1920 /* At first, try to OOM lock hierarchy under memcg.*/
1921 spin_lock(&memcg_oom_lock);
1922 locked = mem_cgroup_oom_lock(memcg);
1924 * Even if signal_pending(), we can't quit charge() loop without
1925 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1926 * under OOM is always welcomed, use TASK_KILLABLE here.
1928 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1929 if (!locked || memcg->oom_kill_disable)
1930 need_to_kill = false;
1932 mem_cgroup_oom_notify(memcg);
1933 spin_unlock(&memcg_oom_lock);
1936 finish_wait(&memcg_oom_waitq, &owait.wait);
1937 mem_cgroup_out_of_memory(memcg, mask);
1940 finish_wait(&memcg_oom_waitq, &owait.wait);
1942 spin_lock(&memcg_oom_lock);
1944 mem_cgroup_oom_unlock(memcg);
1945 memcg_wakeup_oom(memcg);
1946 spin_unlock(&memcg_oom_lock);
1948 mem_cgroup_unmark_under_oom(memcg);
1950 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1952 /* Give chance to dying process */
1953 schedule_timeout_uninterruptible(1);
1958 * Currently used to update mapped file statistics, but the routine can be
1959 * generalized to update other statistics as well.
1961 * Notes: Race condition
1963 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1964 * it tends to be costly. But considering some conditions, we doesn't need
1965 * to do so _always_.
1967 * Considering "charge", lock_page_cgroup() is not required because all
1968 * file-stat operations happen after a page is attached to radix-tree. There
1969 * are no race with "charge".
1971 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1972 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1973 * if there are race with "uncharge". Statistics itself is properly handled
1976 * Considering "move", this is an only case we see a race. To make the race
1977 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1978 * possibility of race condition. If there is, we take a lock.
1981 void mem_cgroup_update_page_stat(struct page *page,
1982 enum mem_cgroup_page_stat_item idx, int val)
1984 struct mem_cgroup *memcg;
1985 struct page_cgroup *pc = lookup_page_cgroup(page);
1986 bool need_unlock = false;
1987 unsigned long uninitialized_var(flags);
1993 memcg = pc->mem_cgroup;
1994 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1996 /* pc->mem_cgroup is unstable ? */
1997 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
1998 /* take a lock against to access pc->mem_cgroup */
1999 move_lock_page_cgroup(pc, &flags);
2001 memcg = pc->mem_cgroup;
2002 if (!memcg || !PageCgroupUsed(pc))
2007 case MEMCG_NR_FILE_MAPPED:
2009 SetPageCgroupFileMapped(pc);
2010 else if (!page_mapped(page))
2011 ClearPageCgroupFileMapped(pc);
2012 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2018 this_cpu_add(memcg->stat->count[idx], val);
2021 if (unlikely(need_unlock))
2022 move_unlock_page_cgroup(pc, &flags);
2026 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2029 * size of first charge trial. "32" comes from vmscan.c's magic value.
2030 * TODO: maybe necessary to use big numbers in big irons.
2032 #define CHARGE_BATCH 32U
2033 struct memcg_stock_pcp {
2034 struct mem_cgroup *cached; /* this never be root cgroup */
2035 unsigned int nr_pages;
2036 struct work_struct work;
2037 unsigned long flags;
2038 #define FLUSHING_CACHED_CHARGE (0)
2040 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2041 static DEFINE_MUTEX(percpu_charge_mutex);
2044 * Try to consume stocked charge on this cpu. If success, one page is consumed
2045 * from local stock and true is returned. If the stock is 0 or charges from a
2046 * cgroup which is not current target, returns false. This stock will be
2049 static bool consume_stock(struct mem_cgroup *memcg)
2051 struct memcg_stock_pcp *stock;
2054 stock = &get_cpu_var(memcg_stock);
2055 if (memcg == stock->cached && stock->nr_pages)
2057 else /* need to call res_counter_charge */
2059 put_cpu_var(memcg_stock);
2064 * Returns stocks cached in percpu to res_counter and reset cached information.
2066 static void drain_stock(struct memcg_stock_pcp *stock)
2068 struct mem_cgroup *old = stock->cached;
2070 if (stock->nr_pages) {
2071 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2073 res_counter_uncharge(&old->res, bytes);
2074 if (do_swap_account)
2075 res_counter_uncharge(&old->memsw, bytes);
2076 stock->nr_pages = 0;
2078 stock->cached = NULL;
2082 * This must be called under preempt disabled or must be called by
2083 * a thread which is pinned to local cpu.
2085 static void drain_local_stock(struct work_struct *dummy)
2087 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2089 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2093 * Cache charges(val) which is from res_counter, to local per_cpu area.
2094 * This will be consumed by consume_stock() function, later.
2096 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2098 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2100 if (stock->cached != memcg) { /* reset if necessary */
2102 stock->cached = memcg;
2104 stock->nr_pages += nr_pages;
2105 put_cpu_var(memcg_stock);
2109 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2110 * of the hierarchy under it. sync flag says whether we should block
2111 * until the work is done.
2113 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2117 /* Notify other cpus that system-wide "drain" is running */
2120 for_each_online_cpu(cpu) {
2121 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2122 struct mem_cgroup *memcg;
2124 memcg = stock->cached;
2125 if (!memcg || !stock->nr_pages)
2127 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2129 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2131 drain_local_stock(&stock->work);
2133 schedule_work_on(cpu, &stock->work);
2141 for_each_online_cpu(cpu) {
2142 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2143 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2144 flush_work(&stock->work);
2151 * Tries to drain stocked charges in other cpus. This function is asynchronous
2152 * and just put a work per cpu for draining localy on each cpu. Caller can
2153 * expects some charges will be back to res_counter later but cannot wait for
2156 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2159 * If someone calls draining, avoid adding more kworker runs.
2161 if (!mutex_trylock(&percpu_charge_mutex))
2163 drain_all_stock(root_memcg, false);
2164 mutex_unlock(&percpu_charge_mutex);
2167 /* This is a synchronous drain interface. */
2168 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2170 /* called when force_empty is called */
2171 mutex_lock(&percpu_charge_mutex);
2172 drain_all_stock(root_memcg, true);
2173 mutex_unlock(&percpu_charge_mutex);
2177 * This function drains percpu counter value from DEAD cpu and
2178 * move it to local cpu. Note that this function can be preempted.
2180 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2184 spin_lock(&memcg->pcp_counter_lock);
2185 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2186 long x = per_cpu(memcg->stat->count[i], cpu);
2188 per_cpu(memcg->stat->count[i], cpu) = 0;
2189 memcg->nocpu_base.count[i] += x;
2191 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2192 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2194 per_cpu(memcg->stat->events[i], cpu) = 0;
2195 memcg->nocpu_base.events[i] += x;
2197 /* need to clear ON_MOVE value, works as a kind of lock. */
2198 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2199 spin_unlock(&memcg->pcp_counter_lock);
2202 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2204 int idx = MEM_CGROUP_ON_MOVE;
2206 spin_lock(&memcg->pcp_counter_lock);
2207 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2208 spin_unlock(&memcg->pcp_counter_lock);
2211 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2212 unsigned long action,
2215 int cpu = (unsigned long)hcpu;
2216 struct memcg_stock_pcp *stock;
2217 struct mem_cgroup *iter;
2219 if ((action == CPU_ONLINE)) {
2220 for_each_mem_cgroup(iter)
2221 synchronize_mem_cgroup_on_move(iter, cpu);
2225 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2228 for_each_mem_cgroup(iter)
2229 mem_cgroup_drain_pcp_counter(iter, cpu);
2231 stock = &per_cpu(memcg_stock, cpu);
2237 /* See __mem_cgroup_try_charge() for details */
2239 CHARGE_OK, /* success */
2240 CHARGE_RETRY, /* need to retry but retry is not bad */
2241 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2242 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2243 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2246 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2247 unsigned int nr_pages, bool oom_check)
2249 unsigned long csize = nr_pages * PAGE_SIZE;
2250 struct mem_cgroup *mem_over_limit;
2251 struct res_counter *fail_res;
2252 unsigned long flags = 0;
2255 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2258 if (!do_swap_account)
2260 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2264 res_counter_uncharge(&memcg->res, csize);
2265 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2266 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2268 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2270 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2271 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2273 * Never reclaim on behalf of optional batching, retry with a
2274 * single page instead.
2276 if (nr_pages == CHARGE_BATCH)
2277 return CHARGE_RETRY;
2279 if (!(gfp_mask & __GFP_WAIT))
2280 return CHARGE_WOULDBLOCK;
2282 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2283 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2284 return CHARGE_RETRY;
2286 * Even though the limit is exceeded at this point, reclaim
2287 * may have been able to free some pages. Retry the charge
2288 * before killing the task.
2290 * Only for regular pages, though: huge pages are rather
2291 * unlikely to succeed so close to the limit, and we fall back
2292 * to regular pages anyway in case of failure.
2294 if (nr_pages == 1 && ret)
2295 return CHARGE_RETRY;
2298 * At task move, charge accounts can be doubly counted. So, it's
2299 * better to wait until the end of task_move if something is going on.
2301 if (mem_cgroup_wait_acct_move(mem_over_limit))
2302 return CHARGE_RETRY;
2304 /* If we don't need to call oom-killer at el, return immediately */
2306 return CHARGE_NOMEM;
2308 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2309 return CHARGE_OOM_DIE;
2311 return CHARGE_RETRY;
2315 * Unlike exported interface, "oom" parameter is added. if oom==true,
2316 * oom-killer can be invoked.
2318 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2320 unsigned int nr_pages,
2321 struct mem_cgroup **ptr,
2324 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2325 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2326 struct mem_cgroup *memcg = NULL;
2330 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2331 * in system level. So, allow to go ahead dying process in addition to
2334 if (unlikely(test_thread_flag(TIF_MEMDIE)
2335 || fatal_signal_pending(current)))
2339 * We always charge the cgroup the mm_struct belongs to.
2340 * The mm_struct's mem_cgroup changes on task migration if the
2341 * thread group leader migrates. It's possible that mm is not
2342 * set, if so charge the init_mm (happens for pagecache usage).
2347 if (*ptr) { /* css should be a valid one */
2349 VM_BUG_ON(css_is_removed(&memcg->css));
2350 if (mem_cgroup_is_root(memcg))
2352 if (nr_pages == 1 && consume_stock(memcg))
2354 css_get(&memcg->css);
2356 struct task_struct *p;
2359 p = rcu_dereference(mm->owner);
2361 * Because we don't have task_lock(), "p" can exit.
2362 * In that case, "memcg" can point to root or p can be NULL with
2363 * race with swapoff. Then, we have small risk of mis-accouning.
2364 * But such kind of mis-account by race always happens because
2365 * we don't have cgroup_mutex(). It's overkill and we allo that
2367 * (*) swapoff at el will charge against mm-struct not against
2368 * task-struct. So, mm->owner can be NULL.
2370 memcg = mem_cgroup_from_task(p);
2371 if (!memcg || mem_cgroup_is_root(memcg)) {
2375 if (nr_pages == 1 && consume_stock(memcg)) {
2377 * It seems dagerous to access memcg without css_get().
2378 * But considering how consume_stok works, it's not
2379 * necessary. If consume_stock success, some charges
2380 * from this memcg are cached on this cpu. So, we
2381 * don't need to call css_get()/css_tryget() before
2382 * calling consume_stock().
2387 /* after here, we may be blocked. we need to get refcnt */
2388 if (!css_tryget(&memcg->css)) {
2398 /* If killed, bypass charge */
2399 if (fatal_signal_pending(current)) {
2400 css_put(&memcg->css);
2405 if (oom && !nr_oom_retries) {
2407 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2410 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2414 case CHARGE_RETRY: /* not in OOM situation but retry */
2416 css_put(&memcg->css);
2419 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2420 css_put(&memcg->css);
2422 case CHARGE_NOMEM: /* OOM routine works */
2424 css_put(&memcg->css);
2427 /* If oom, we never return -ENOMEM */
2430 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2431 css_put(&memcg->css);
2434 } while (ret != CHARGE_OK);
2436 if (batch > nr_pages)
2437 refill_stock(memcg, batch - nr_pages);
2438 css_put(&memcg->css);
2451 * Somemtimes we have to undo a charge we got by try_charge().
2452 * This function is for that and do uncharge, put css's refcnt.
2453 * gotten by try_charge().
2455 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2456 unsigned int nr_pages)
2458 if (!mem_cgroup_is_root(memcg)) {
2459 unsigned long bytes = nr_pages * PAGE_SIZE;
2461 res_counter_uncharge(&memcg->res, bytes);
2462 if (do_swap_account)
2463 res_counter_uncharge(&memcg->memsw, bytes);
2468 * A helper function to get mem_cgroup from ID. must be called under
2469 * rcu_read_lock(). The caller must check css_is_removed() or some if
2470 * it's concern. (dropping refcnt from swap can be called against removed
2473 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2475 struct cgroup_subsys_state *css;
2477 /* ID 0 is unused ID */
2480 css = css_lookup(&mem_cgroup_subsys, id);
2483 return container_of(css, struct mem_cgroup, css);
2486 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2488 struct mem_cgroup *memcg = NULL;
2489 struct page_cgroup *pc;
2493 VM_BUG_ON(!PageLocked(page));
2495 pc = lookup_page_cgroup(page);
2496 lock_page_cgroup(pc);
2497 if (PageCgroupUsed(pc)) {
2498 memcg = pc->mem_cgroup;
2499 if (memcg && !css_tryget(&memcg->css))
2501 } else if (PageSwapCache(page)) {
2502 ent.val = page_private(page);
2503 id = lookup_swap_cgroup(ent);
2505 memcg = mem_cgroup_lookup(id);
2506 if (memcg && !css_tryget(&memcg->css))
2510 unlock_page_cgroup(pc);
2514 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2516 unsigned int nr_pages,
2517 struct page_cgroup *pc,
2518 enum charge_type ctype)
2520 lock_page_cgroup(pc);
2521 if (unlikely(PageCgroupUsed(pc))) {
2522 unlock_page_cgroup(pc);
2523 __mem_cgroup_cancel_charge(memcg, nr_pages);
2527 * we don't need page_cgroup_lock about tail pages, becase they are not
2528 * accessed by any other context at this point.
2530 pc->mem_cgroup = memcg;
2532 * We access a page_cgroup asynchronously without lock_page_cgroup().
2533 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2534 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2535 * before USED bit, we need memory barrier here.
2536 * See mem_cgroup_add_lru_list(), etc.
2540 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2541 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2542 SetPageCgroupCache(pc);
2543 SetPageCgroupUsed(pc);
2545 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2546 ClearPageCgroupCache(pc);
2547 SetPageCgroupUsed(pc);
2553 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2554 unlock_page_cgroup(pc);
2556 * "charge_statistics" updated event counter. Then, check it.
2557 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2558 * if they exceeds softlimit.
2560 memcg_check_events(memcg, page);
2563 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2565 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2566 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2568 * Because tail pages are not marked as "used", set it. We're under
2569 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2571 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2573 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2574 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2575 unsigned long flags;
2577 if (mem_cgroup_disabled())
2580 * We have no races with charge/uncharge but will have races with
2581 * page state accounting.
2583 move_lock_page_cgroup(head_pc, &flags);
2585 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2586 smp_wmb(); /* see __commit_charge() */
2587 if (PageCgroupAcctLRU(head_pc)) {
2589 struct mem_cgroup_per_zone *mz;
2592 * LRU flags cannot be copied because we need to add tail
2593 *.page to LRU by generic call and our hook will be called.
2594 * We hold lru_lock, then, reduce counter directly.
2596 lru = page_lru(head);
2597 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2598 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2600 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2601 move_unlock_page_cgroup(head_pc, &flags);
2606 * mem_cgroup_move_account - move account of the page
2608 * @nr_pages: number of regular pages (>1 for huge pages)
2609 * @pc: page_cgroup of the page.
2610 * @from: mem_cgroup which the page is moved from.
2611 * @to: mem_cgroup which the page is moved to. @from != @to.
2612 * @uncharge: whether we should call uncharge and css_put against @from.
2614 * The caller must confirm following.
2615 * - page is not on LRU (isolate_page() is useful.)
2616 * - compound_lock is held when nr_pages > 1
2618 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2619 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2620 * true, this function does "uncharge" from old cgroup, but it doesn't if
2621 * @uncharge is false, so a caller should do "uncharge".
2623 static int mem_cgroup_move_account(struct page *page,
2624 unsigned int nr_pages,
2625 struct page_cgroup *pc,
2626 struct mem_cgroup *from,
2627 struct mem_cgroup *to,
2630 unsigned long flags;
2633 VM_BUG_ON(from == to);
2634 VM_BUG_ON(PageLRU(page));
2636 * The page is isolated from LRU. So, collapse function
2637 * will not handle this page. But page splitting can happen.
2638 * Do this check under compound_page_lock(). The caller should
2642 if (nr_pages > 1 && !PageTransHuge(page))
2645 lock_page_cgroup(pc);
2648 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2651 move_lock_page_cgroup(pc, &flags);
2653 if (PageCgroupFileMapped(pc)) {
2654 /* Update mapped_file data for mem_cgroup */
2656 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2657 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2660 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2662 /* This is not "cancel", but cancel_charge does all we need. */
2663 __mem_cgroup_cancel_charge(from, nr_pages);
2665 /* caller should have done css_get */
2666 pc->mem_cgroup = to;
2667 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2669 * We charges against "to" which may not have any tasks. Then, "to"
2670 * can be under rmdir(). But in current implementation, caller of
2671 * this function is just force_empty() and move charge, so it's
2672 * guaranteed that "to" is never removed. So, we don't check rmdir
2675 move_unlock_page_cgroup(pc, &flags);
2678 unlock_page_cgroup(pc);
2682 memcg_check_events(to, page);
2683 memcg_check_events(from, page);
2689 * move charges to its parent.
2692 static int mem_cgroup_move_parent(struct page *page,
2693 struct page_cgroup *pc,
2694 struct mem_cgroup *child,
2697 struct cgroup *cg = child->css.cgroup;
2698 struct cgroup *pcg = cg->parent;
2699 struct mem_cgroup *parent;
2700 unsigned int nr_pages;
2701 unsigned long uninitialized_var(flags);
2709 if (!get_page_unless_zero(page))
2711 if (isolate_lru_page(page))
2714 nr_pages = hpage_nr_pages(page);
2716 parent = mem_cgroup_from_cont(pcg);
2717 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2722 flags = compound_lock_irqsave(page);
2724 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2726 __mem_cgroup_cancel_charge(parent, nr_pages);
2729 compound_unlock_irqrestore(page, flags);
2731 putback_lru_page(page);
2739 * Charge the memory controller for page usage.
2741 * 0 if the charge was successful
2742 * < 0 if the cgroup is over its limit
2744 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2745 gfp_t gfp_mask, enum charge_type ctype)
2747 struct mem_cgroup *memcg = NULL;
2748 unsigned int nr_pages = 1;
2749 struct page_cgroup *pc;
2753 if (PageTransHuge(page)) {
2754 nr_pages <<= compound_order(page);
2755 VM_BUG_ON(!PageTransHuge(page));
2757 * Never OOM-kill a process for a huge page. The
2758 * fault handler will fall back to regular pages.
2763 pc = lookup_page_cgroup(page);
2764 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2766 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2770 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2774 int mem_cgroup_newpage_charge(struct page *page,
2775 struct mm_struct *mm, gfp_t gfp_mask)
2777 if (mem_cgroup_disabled())
2780 * If already mapped, we don't have to account.
2781 * If page cache, page->mapping has address_space.
2782 * But page->mapping may have out-of-use anon_vma pointer,
2783 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2786 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2790 return mem_cgroup_charge_common(page, mm, gfp_mask,
2791 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2795 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2796 enum charge_type ctype);
2799 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2800 enum charge_type ctype)
2802 struct page_cgroup *pc = lookup_page_cgroup(page);
2804 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2805 * is already on LRU. It means the page may on some other page_cgroup's
2806 * LRU. Take care of it.
2808 mem_cgroup_lru_del_before_commit(page);
2809 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2810 mem_cgroup_lru_add_after_commit(page);
2814 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2817 struct mem_cgroup *memcg = NULL;
2820 if (mem_cgroup_disabled())
2822 if (PageCompound(page))
2828 if (page_is_file_cache(page)) {
2829 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2834 * FUSE reuses pages without going through the final
2835 * put that would remove them from the LRU list, make
2836 * sure that they get relinked properly.
2838 __mem_cgroup_commit_charge_lrucare(page, memcg,
2839 MEM_CGROUP_CHARGE_TYPE_CACHE);
2843 if (PageSwapCache(page)) {
2844 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2846 __mem_cgroup_commit_charge_swapin(page, memcg,
2847 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2849 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2850 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2856 * While swap-in, try_charge -> commit or cancel, the page is locked.
2857 * And when try_charge() successfully returns, one refcnt to memcg without
2858 * struct page_cgroup is acquired. This refcnt will be consumed by
2859 * "commit()" or removed by "cancel()"
2861 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2863 gfp_t mask, struct mem_cgroup **ptr)
2865 struct mem_cgroup *memcg;
2870 if (mem_cgroup_disabled())
2873 if (!do_swap_account)
2876 * A racing thread's fault, or swapoff, may have already updated
2877 * the pte, and even removed page from swap cache: in those cases
2878 * do_swap_page()'s pte_same() test will fail; but there's also a
2879 * KSM case which does need to charge the page.
2881 if (!PageSwapCache(page))
2883 memcg = try_get_mem_cgroup_from_page(page);
2887 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2888 css_put(&memcg->css);
2893 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2897 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2898 enum charge_type ctype)
2900 if (mem_cgroup_disabled())
2904 cgroup_exclude_rmdir(&ptr->css);
2906 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2908 * Now swap is on-memory. This means this page may be
2909 * counted both as mem and swap....double count.
2910 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2911 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2912 * may call delete_from_swap_cache() before reach here.
2914 if (do_swap_account && PageSwapCache(page)) {
2915 swp_entry_t ent = {.val = page_private(page)};
2917 struct mem_cgroup *memcg;
2919 id = swap_cgroup_record(ent, 0);
2921 memcg = mem_cgroup_lookup(id);
2924 * This recorded memcg can be obsolete one. So, avoid
2925 * calling css_tryget
2927 if (!mem_cgroup_is_root(memcg))
2928 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2929 mem_cgroup_swap_statistics(memcg, false);
2930 mem_cgroup_put(memcg);
2935 * At swapin, we may charge account against cgroup which has no tasks.
2936 * So, rmdir()->pre_destroy() can be called while we do this charge.
2937 * In that case, we need to call pre_destroy() again. check it here.
2939 cgroup_release_and_wakeup_rmdir(&ptr->css);
2942 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2944 __mem_cgroup_commit_charge_swapin(page, ptr,
2945 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2948 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2950 if (mem_cgroup_disabled())
2954 __mem_cgroup_cancel_charge(memcg, 1);
2957 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2958 unsigned int nr_pages,
2959 const enum charge_type ctype)
2961 struct memcg_batch_info *batch = NULL;
2962 bool uncharge_memsw = true;
2964 /* If swapout, usage of swap doesn't decrease */
2965 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2966 uncharge_memsw = false;
2968 batch = ¤t->memcg_batch;
2970 * In usual, we do css_get() when we remember memcg pointer.
2971 * But in this case, we keep res->usage until end of a series of
2972 * uncharges. Then, it's ok to ignore memcg's refcnt.
2975 batch->memcg = memcg;
2977 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2978 * In those cases, all pages freed continuously can be expected to be in
2979 * the same cgroup and we have chance to coalesce uncharges.
2980 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2981 * because we want to do uncharge as soon as possible.
2984 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2985 goto direct_uncharge;
2988 goto direct_uncharge;
2991 * In typical case, batch->memcg == mem. This means we can
2992 * merge a series of uncharges to an uncharge of res_counter.
2993 * If not, we uncharge res_counter ony by one.
2995 if (batch->memcg != memcg)
2996 goto direct_uncharge;
2997 /* remember freed charge and uncharge it later */
3000 batch->memsw_nr_pages++;
3003 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3005 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3006 if (unlikely(batch->memcg != memcg))
3007 memcg_oom_recover(memcg);
3012 * uncharge if !page_mapped(page)
3014 static struct mem_cgroup *
3015 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3017 struct mem_cgroup *memcg = NULL;
3018 unsigned int nr_pages = 1;
3019 struct page_cgroup *pc;
3021 if (mem_cgroup_disabled())
3024 if (PageSwapCache(page))
3027 if (PageTransHuge(page)) {
3028 nr_pages <<= compound_order(page);
3029 VM_BUG_ON(!PageTransHuge(page));
3032 * Check if our page_cgroup is valid
3034 pc = lookup_page_cgroup(page);
3035 if (unlikely(!pc || !PageCgroupUsed(pc)))
3038 lock_page_cgroup(pc);
3040 memcg = pc->mem_cgroup;
3042 if (!PageCgroupUsed(pc))
3046 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3047 case MEM_CGROUP_CHARGE_TYPE_DROP:
3048 /* See mem_cgroup_prepare_migration() */
3049 if (page_mapped(page) || PageCgroupMigration(pc))
3052 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3053 if (!PageAnon(page)) { /* Shared memory */
3054 if (page->mapping && !page_is_file_cache(page))
3056 } else if (page_mapped(page)) /* Anon */
3063 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3065 ClearPageCgroupUsed(pc);
3067 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3068 * freed from LRU. This is safe because uncharged page is expected not
3069 * to be reused (freed soon). Exception is SwapCache, it's handled by
3070 * special functions.
3073 unlock_page_cgroup(pc);
3075 * even after unlock, we have memcg->res.usage here and this memcg
3076 * will never be freed.
3078 memcg_check_events(memcg, page);
3079 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3080 mem_cgroup_swap_statistics(memcg, true);
3081 mem_cgroup_get(memcg);
3083 if (!mem_cgroup_is_root(memcg))
3084 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3089 unlock_page_cgroup(pc);
3093 void mem_cgroup_uncharge_page(struct page *page)
3096 if (page_mapped(page))
3098 if (page->mapping && !PageAnon(page))
3100 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3103 void mem_cgroup_uncharge_cache_page(struct page *page)
3105 VM_BUG_ON(page_mapped(page));
3106 VM_BUG_ON(page->mapping);
3107 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3111 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3112 * In that cases, pages are freed continuously and we can expect pages
3113 * are in the same memcg. All these calls itself limits the number of
3114 * pages freed at once, then uncharge_start/end() is called properly.
3115 * This may be called prural(2) times in a context,
3118 void mem_cgroup_uncharge_start(void)
3120 current->memcg_batch.do_batch++;
3121 /* We can do nest. */
3122 if (current->memcg_batch.do_batch == 1) {
3123 current->memcg_batch.memcg = NULL;
3124 current->memcg_batch.nr_pages = 0;
3125 current->memcg_batch.memsw_nr_pages = 0;
3129 void mem_cgroup_uncharge_end(void)
3131 struct memcg_batch_info *batch = ¤t->memcg_batch;
3133 if (!batch->do_batch)
3137 if (batch->do_batch) /* If stacked, do nothing. */
3143 * This "batch->memcg" is valid without any css_get/put etc...
3144 * bacause we hide charges behind us.
3146 if (batch->nr_pages)
3147 res_counter_uncharge(&batch->memcg->res,
3148 batch->nr_pages * PAGE_SIZE);
3149 if (batch->memsw_nr_pages)
3150 res_counter_uncharge(&batch->memcg->memsw,
3151 batch->memsw_nr_pages * PAGE_SIZE);
3152 memcg_oom_recover(batch->memcg);
3153 /* forget this pointer (for sanity check) */
3154 batch->memcg = NULL;
3159 * called after __delete_from_swap_cache() and drop "page" account.
3160 * memcg information is recorded to swap_cgroup of "ent"
3163 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3165 struct mem_cgroup *memcg;
3166 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3168 if (!swapout) /* this was a swap cache but the swap is unused ! */
3169 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3171 memcg = __mem_cgroup_uncharge_common(page, ctype);
3174 * record memcg information, if swapout && memcg != NULL,
3175 * mem_cgroup_get() was called in uncharge().
3177 if (do_swap_account && swapout && memcg)
3178 swap_cgroup_record(ent, css_id(&memcg->css));
3182 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3184 * called from swap_entry_free(). remove record in swap_cgroup and
3185 * uncharge "memsw" account.
3187 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3189 struct mem_cgroup *memcg;
3192 if (!do_swap_account)
3195 id = swap_cgroup_record(ent, 0);
3197 memcg = mem_cgroup_lookup(id);
3200 * We uncharge this because swap is freed.
3201 * This memcg can be obsolete one. We avoid calling css_tryget
3203 if (!mem_cgroup_is_root(memcg))
3204 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3205 mem_cgroup_swap_statistics(memcg, false);
3206 mem_cgroup_put(memcg);
3212 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3213 * @entry: swap entry to be moved
3214 * @from: mem_cgroup which the entry is moved from
3215 * @to: mem_cgroup which the entry is moved to
3216 * @need_fixup: whether we should fixup res_counters and refcounts.
3218 * It succeeds only when the swap_cgroup's record for this entry is the same
3219 * as the mem_cgroup's id of @from.
3221 * Returns 0 on success, -EINVAL on failure.
3223 * The caller must have charged to @to, IOW, called res_counter_charge() about
3224 * both res and memsw, and called css_get().
3226 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3227 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3229 unsigned short old_id, new_id;
3231 old_id = css_id(&from->css);
3232 new_id = css_id(&to->css);
3234 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3235 mem_cgroup_swap_statistics(from, false);
3236 mem_cgroup_swap_statistics(to, true);
3238 * This function is only called from task migration context now.
3239 * It postpones res_counter and refcount handling till the end
3240 * of task migration(mem_cgroup_clear_mc()) for performance
3241 * improvement. But we cannot postpone mem_cgroup_get(to)
3242 * because if the process that has been moved to @to does
3243 * swap-in, the refcount of @to might be decreased to 0.
3247 if (!mem_cgroup_is_root(from))
3248 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3249 mem_cgroup_put(from);
3251 * we charged both to->res and to->memsw, so we should
3254 if (!mem_cgroup_is_root(to))
3255 res_counter_uncharge(&to->res, PAGE_SIZE);
3262 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3263 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3270 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3273 int mem_cgroup_prepare_migration(struct page *page,
3274 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3276 struct mem_cgroup *memcg = NULL;
3277 struct page_cgroup *pc;
3278 enum charge_type ctype;
3283 VM_BUG_ON(PageTransHuge(page));
3284 if (mem_cgroup_disabled())
3287 pc = lookup_page_cgroup(page);
3288 lock_page_cgroup(pc);
3289 if (PageCgroupUsed(pc)) {
3290 memcg = pc->mem_cgroup;
3291 css_get(&memcg->css);
3293 * At migrating an anonymous page, its mapcount goes down
3294 * to 0 and uncharge() will be called. But, even if it's fully
3295 * unmapped, migration may fail and this page has to be
3296 * charged again. We set MIGRATION flag here and delay uncharge
3297 * until end_migration() is called
3299 * Corner Case Thinking
3301 * When the old page was mapped as Anon and it's unmap-and-freed
3302 * while migration was ongoing.
3303 * If unmap finds the old page, uncharge() of it will be delayed
3304 * until end_migration(). If unmap finds a new page, it's
3305 * uncharged when it make mapcount to be 1->0. If unmap code
3306 * finds swap_migration_entry, the new page will not be mapped
3307 * and end_migration() will find it(mapcount==0).
3310 * When the old page was mapped but migraion fails, the kernel
3311 * remaps it. A charge for it is kept by MIGRATION flag even
3312 * if mapcount goes down to 0. We can do remap successfully
3313 * without charging it again.
3316 * The "old" page is under lock_page() until the end of
3317 * migration, so, the old page itself will not be swapped-out.
3318 * If the new page is swapped out before end_migraton, our
3319 * hook to usual swap-out path will catch the event.
3322 SetPageCgroupMigration(pc);
3324 unlock_page_cgroup(pc);
3326 * If the page is not charged at this point,
3333 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3334 css_put(&memcg->css);/* drop extra refcnt */
3335 if (ret || *ptr == NULL) {
3336 if (PageAnon(page)) {
3337 lock_page_cgroup(pc);
3338 ClearPageCgroupMigration(pc);
3339 unlock_page_cgroup(pc);
3341 * The old page may be fully unmapped while we kept it.
3343 mem_cgroup_uncharge_page(page);
3348 * We charge new page before it's used/mapped. So, even if unlock_page()
3349 * is called before end_migration, we can catch all events on this new
3350 * page. In the case new page is migrated but not remapped, new page's
3351 * mapcount will be finally 0 and we call uncharge in end_migration().
3353 pc = lookup_page_cgroup(newpage);
3355 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3356 else if (page_is_file_cache(page))
3357 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3359 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3360 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3364 /* remove redundant charge if migration failed*/
3365 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3366 struct page *oldpage, struct page *newpage, bool migration_ok)
3368 struct page *used, *unused;
3369 struct page_cgroup *pc;
3373 /* blocks rmdir() */
3374 cgroup_exclude_rmdir(&memcg->css);
3375 if (!migration_ok) {
3383 * We disallowed uncharge of pages under migration because mapcount
3384 * of the page goes down to zero, temporarly.
3385 * Clear the flag and check the page should be charged.
3387 pc = lookup_page_cgroup(oldpage);
3388 lock_page_cgroup(pc);
3389 ClearPageCgroupMigration(pc);
3390 unlock_page_cgroup(pc);
3392 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3395 * If a page is a file cache, radix-tree replacement is very atomic
3396 * and we can skip this check. When it was an Anon page, its mapcount
3397 * goes down to 0. But because we added MIGRATION flage, it's not
3398 * uncharged yet. There are several case but page->mapcount check
3399 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3400 * check. (see prepare_charge() also)
3403 mem_cgroup_uncharge_page(used);
3405 * At migration, we may charge account against cgroup which has no
3407 * So, rmdir()->pre_destroy() can be called while we do this charge.
3408 * In that case, we need to call pre_destroy() again. check it here.
3410 cgroup_release_and_wakeup_rmdir(&memcg->css);
3414 * At replace page cache, newpage is not under any memcg but it's on
3415 * LRU. So, this function doesn't touch res_counter but handles LRU
3416 * in correct way. Both pages are locked so we cannot race with uncharge.
3418 void mem_cgroup_replace_page_cache(struct page *oldpage,
3419 struct page *newpage)
3421 struct mem_cgroup *memcg;
3422 struct page_cgroup *pc;
3424 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3425 unsigned long flags;
3427 if (mem_cgroup_disabled())
3430 pc = lookup_page_cgroup(oldpage);
3431 /* fix accounting on old pages */
3432 lock_page_cgroup(pc);
3433 memcg = pc->mem_cgroup;
3434 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1);
3435 ClearPageCgroupUsed(pc);
3436 unlock_page_cgroup(pc);
3438 if (PageSwapBacked(oldpage))
3439 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3441 zone = page_zone(newpage);
3442 pc = lookup_page_cgroup(newpage);
3444 * Even if newpage->mapping was NULL before starting replacement,
3445 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3446 * LRU while we overwrite pc->mem_cgroup.
3448 spin_lock_irqsave(&zone->lru_lock, flags);
3449 if (PageLRU(newpage))
3450 del_page_from_lru_list(zone, newpage, page_lru(newpage));
3451 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type);
3452 if (PageLRU(newpage))
3453 add_page_to_lru_list(zone, newpage, page_lru(newpage));
3454 spin_unlock_irqrestore(&zone->lru_lock, flags);
3457 #ifdef CONFIG_DEBUG_VM
3458 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3460 struct page_cgroup *pc;
3462 pc = lookup_page_cgroup(page);
3463 if (likely(pc) && PageCgroupUsed(pc))
3468 bool mem_cgroup_bad_page_check(struct page *page)
3470 if (mem_cgroup_disabled())
3473 return lookup_page_cgroup_used(page) != NULL;
3476 void mem_cgroup_print_bad_page(struct page *page)
3478 struct page_cgroup *pc;
3480 pc = lookup_page_cgroup_used(page);
3485 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3486 pc, pc->flags, pc->mem_cgroup);
3488 path = kmalloc(PATH_MAX, GFP_KERNEL);
3491 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3496 printk(KERN_CONT "(%s)\n",
3497 (ret < 0) ? "cannot get the path" : path);
3503 static DEFINE_MUTEX(set_limit_mutex);
3505 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3506 unsigned long long val)
3509 u64 memswlimit, memlimit;
3511 int children = mem_cgroup_count_children(memcg);
3512 u64 curusage, oldusage;
3516 * For keeping hierarchical_reclaim simple, how long we should retry
3517 * is depends on callers. We set our retry-count to be function
3518 * of # of children which we should visit in this loop.
3520 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3522 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3525 while (retry_count) {
3526 if (signal_pending(current)) {
3531 * Rather than hide all in some function, I do this in
3532 * open coded manner. You see what this really does.
3533 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3535 mutex_lock(&set_limit_mutex);
3536 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3537 if (memswlimit < val) {
3539 mutex_unlock(&set_limit_mutex);
3543 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3547 ret = res_counter_set_limit(&memcg->res, val);
3549 if (memswlimit == val)
3550 memcg->memsw_is_minimum = true;
3552 memcg->memsw_is_minimum = false;
3554 mutex_unlock(&set_limit_mutex);
3559 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3560 MEM_CGROUP_RECLAIM_SHRINK);
3561 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3562 /* Usage is reduced ? */
3563 if (curusage >= oldusage)
3566 oldusage = curusage;
3568 if (!ret && enlarge)
3569 memcg_oom_recover(memcg);
3574 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3575 unsigned long long val)
3578 u64 memlimit, memswlimit, oldusage, curusage;
3579 int children = mem_cgroup_count_children(memcg);
3583 /* see mem_cgroup_resize_res_limit */
3584 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3585 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3586 while (retry_count) {
3587 if (signal_pending(current)) {
3592 * Rather than hide all in some function, I do this in
3593 * open coded manner. You see what this really does.
3594 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3596 mutex_lock(&set_limit_mutex);
3597 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3598 if (memlimit > val) {
3600 mutex_unlock(&set_limit_mutex);
3603 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3604 if (memswlimit < val)
3606 ret = res_counter_set_limit(&memcg->memsw, val);
3608 if (memlimit == val)
3609 memcg->memsw_is_minimum = true;
3611 memcg->memsw_is_minimum = false;
3613 mutex_unlock(&set_limit_mutex);
3618 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3619 MEM_CGROUP_RECLAIM_NOSWAP |
3620 MEM_CGROUP_RECLAIM_SHRINK);
3621 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3622 /* Usage is reduced ? */
3623 if (curusage >= oldusage)
3626 oldusage = curusage;
3628 if (!ret && enlarge)
3629 memcg_oom_recover(memcg);
3633 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3635 unsigned long *total_scanned)
3637 unsigned long nr_reclaimed = 0;
3638 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3639 unsigned long reclaimed;
3641 struct mem_cgroup_tree_per_zone *mctz;
3642 unsigned long long excess;
3643 unsigned long nr_scanned;
3648 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3650 * This loop can run a while, specially if mem_cgroup's continuously
3651 * keep exceeding their soft limit and putting the system under
3658 mz = mem_cgroup_largest_soft_limit_node(mctz);
3663 reclaimed = mem_cgroup_soft_reclaim(mz->mem, zone,
3664 gfp_mask, &nr_scanned);
3665 nr_reclaimed += reclaimed;
3666 *total_scanned += nr_scanned;
3667 spin_lock(&mctz->lock);
3670 * If we failed to reclaim anything from this memory cgroup
3671 * it is time to move on to the next cgroup
3677 * Loop until we find yet another one.
3679 * By the time we get the soft_limit lock
3680 * again, someone might have aded the
3681 * group back on the RB tree. Iterate to
3682 * make sure we get a different mem.
3683 * mem_cgroup_largest_soft_limit_node returns
3684 * NULL if no other cgroup is present on
3688 __mem_cgroup_largest_soft_limit_node(mctz);
3690 css_put(&next_mz->mem->css);
3691 else /* next_mz == NULL or other memcg */
3695 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3696 excess = res_counter_soft_limit_excess(&mz->mem->res);
3698 * One school of thought says that we should not add
3699 * back the node to the tree if reclaim returns 0.
3700 * But our reclaim could return 0, simply because due
3701 * to priority we are exposing a smaller subset of
3702 * memory to reclaim from. Consider this as a longer
3705 /* If excess == 0, no tree ops */
3706 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3707 spin_unlock(&mctz->lock);
3708 css_put(&mz->mem->css);
3711 * Could not reclaim anything and there are no more
3712 * mem cgroups to try or we seem to be looping without
3713 * reclaiming anything.
3715 if (!nr_reclaimed &&
3717 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3719 } while (!nr_reclaimed);
3721 css_put(&next_mz->mem->css);
3722 return nr_reclaimed;
3726 * This routine traverse page_cgroup in given list and drop them all.
3727 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3729 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3730 int node, int zid, enum lru_list lru)
3733 struct mem_cgroup_per_zone *mz;
3734 struct page_cgroup *pc, *busy;
3735 unsigned long flags, loop;
3736 struct list_head *list;
3739 zone = &NODE_DATA(node)->node_zones[zid];
3740 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3741 list = &mz->lists[lru];
3743 loop = MEM_CGROUP_ZSTAT(mz, lru);
3744 /* give some margin against EBUSY etc...*/
3751 spin_lock_irqsave(&zone->lru_lock, flags);
3752 if (list_empty(list)) {
3753 spin_unlock_irqrestore(&zone->lru_lock, flags);
3756 pc = list_entry(list->prev, struct page_cgroup, lru);
3758 list_move(&pc->lru, list);
3760 spin_unlock_irqrestore(&zone->lru_lock, flags);
3763 spin_unlock_irqrestore(&zone->lru_lock, flags);
3765 page = lookup_cgroup_page(pc);
3767 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3771 if (ret == -EBUSY || ret == -EINVAL) {
3772 /* found lock contention or "pc" is obsolete. */
3779 if (!ret && !list_empty(list))
3785 * make mem_cgroup's charge to be 0 if there is no task.
3786 * This enables deleting this mem_cgroup.
3788 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3791 int node, zid, shrink;
3792 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3793 struct cgroup *cgrp = memcg->css.cgroup;
3795 css_get(&memcg->css);
3798 /* should free all ? */
3804 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3807 if (signal_pending(current))
3809 /* This is for making all *used* pages to be on LRU. */
3810 lru_add_drain_all();
3811 drain_all_stock_sync(memcg);
3813 mem_cgroup_start_move(memcg);
3814 for_each_node_state(node, N_HIGH_MEMORY) {
3815 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3818 ret = mem_cgroup_force_empty_list(memcg,
3827 mem_cgroup_end_move(memcg);
3828 memcg_oom_recover(memcg);
3829 /* it seems parent cgroup doesn't have enough mem */
3833 /* "ret" should also be checked to ensure all lists are empty. */
3834 } while (memcg->res.usage > 0 || ret);
3836 css_put(&memcg->css);
3840 /* returns EBUSY if there is a task or if we come here twice. */
3841 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3845 /* we call try-to-free pages for make this cgroup empty */
3846 lru_add_drain_all();
3847 /* try to free all pages in this cgroup */
3849 while (nr_retries && memcg->res.usage > 0) {
3852 if (signal_pending(current)) {
3856 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3860 /* maybe some writeback is necessary */
3861 congestion_wait(BLK_RW_ASYNC, HZ/10);
3866 /* try move_account...there may be some *locked* pages. */
3870 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3872 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3876 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3878 return mem_cgroup_from_cont(cont)->use_hierarchy;
3881 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3885 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3886 struct cgroup *parent = cont->parent;
3887 struct mem_cgroup *parent_memcg = NULL;
3890 parent_memcg = mem_cgroup_from_cont(parent);
3894 * If parent's use_hierarchy is set, we can't make any modifications
3895 * in the child subtrees. If it is unset, then the change can
3896 * occur, provided the current cgroup has no children.
3898 * For the root cgroup, parent_mem is NULL, we allow value to be
3899 * set if there are no children.
3901 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3902 (val == 1 || val == 0)) {
3903 if (list_empty(&cont->children))
3904 memcg->use_hierarchy = val;
3915 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3916 enum mem_cgroup_stat_index idx)
3918 struct mem_cgroup *iter;
3921 /* Per-cpu values can be negative, use a signed accumulator */
3922 for_each_mem_cgroup_tree(iter, memcg)
3923 val += mem_cgroup_read_stat(iter, idx);
3925 if (val < 0) /* race ? */
3930 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3934 if (!mem_cgroup_is_root(memcg)) {
3936 return res_counter_read_u64(&memcg->res, RES_USAGE);
3938 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3941 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3942 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3945 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3947 return val << PAGE_SHIFT;
3950 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3952 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3956 type = MEMFILE_TYPE(cft->private);
3957 name = MEMFILE_ATTR(cft->private);
3960 if (name == RES_USAGE)
3961 val = mem_cgroup_usage(memcg, false);
3963 val = res_counter_read_u64(&memcg->res, name);
3966 if (name == RES_USAGE)
3967 val = mem_cgroup_usage(memcg, true);
3969 val = res_counter_read_u64(&memcg->memsw, name);
3978 * The user of this function is...
3981 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3984 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3986 unsigned long long val;
3989 type = MEMFILE_TYPE(cft->private);
3990 name = MEMFILE_ATTR(cft->private);
3993 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3997 /* This function does all necessary parse...reuse it */
3998 ret = res_counter_memparse_write_strategy(buffer, &val);
4002 ret = mem_cgroup_resize_limit(memcg, val);
4004 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4006 case RES_SOFT_LIMIT:
4007 ret = res_counter_memparse_write_strategy(buffer, &val);
4011 * For memsw, soft limits are hard to implement in terms
4012 * of semantics, for now, we support soft limits for
4013 * control without swap
4016 ret = res_counter_set_soft_limit(&memcg->res, val);
4021 ret = -EINVAL; /* should be BUG() ? */
4027 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4028 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4030 struct cgroup *cgroup;
4031 unsigned long long min_limit, min_memsw_limit, tmp;
4033 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4034 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4035 cgroup = memcg->css.cgroup;
4036 if (!memcg->use_hierarchy)
4039 while (cgroup->parent) {
4040 cgroup = cgroup->parent;
4041 memcg = mem_cgroup_from_cont(cgroup);
4042 if (!memcg->use_hierarchy)
4044 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4045 min_limit = min(min_limit, tmp);
4046 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4047 min_memsw_limit = min(min_memsw_limit, tmp);
4050 *mem_limit = min_limit;
4051 *memsw_limit = min_memsw_limit;
4055 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4057 struct mem_cgroup *memcg;
4060 memcg = mem_cgroup_from_cont(cont);
4061 type = MEMFILE_TYPE(event);
4062 name = MEMFILE_ATTR(event);
4066 res_counter_reset_max(&memcg->res);
4068 res_counter_reset_max(&memcg->memsw);
4072 res_counter_reset_failcnt(&memcg->res);
4074 res_counter_reset_failcnt(&memcg->memsw);
4081 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4084 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4088 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4089 struct cftype *cft, u64 val)
4091 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4093 if (val >= (1 << NR_MOVE_TYPE))
4096 * We check this value several times in both in can_attach() and
4097 * attach(), so we need cgroup lock to prevent this value from being
4101 memcg->move_charge_at_immigrate = val;
4107 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4108 struct cftype *cft, u64 val)
4115 /* For read statistics */
4133 struct mcs_total_stat {
4134 s64 stat[NR_MCS_STAT];
4140 } memcg_stat_strings[NR_MCS_STAT] = {
4141 {"cache", "total_cache"},
4142 {"rss", "total_rss"},
4143 {"mapped_file", "total_mapped_file"},
4144 {"pgpgin", "total_pgpgin"},
4145 {"pgpgout", "total_pgpgout"},
4146 {"swap", "total_swap"},
4147 {"pgfault", "total_pgfault"},
4148 {"pgmajfault", "total_pgmajfault"},
4149 {"inactive_anon", "total_inactive_anon"},
4150 {"active_anon", "total_active_anon"},
4151 {"inactive_file", "total_inactive_file"},
4152 {"active_file", "total_active_file"},
4153 {"unevictable", "total_unevictable"}
4158 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4163 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4164 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4165 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4166 s->stat[MCS_RSS] += val * PAGE_SIZE;
4167 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4168 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4169 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4170 s->stat[MCS_PGPGIN] += val;
4171 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4172 s->stat[MCS_PGPGOUT] += val;
4173 if (do_swap_account) {
4174 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4175 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4177 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4178 s->stat[MCS_PGFAULT] += val;
4179 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4180 s->stat[MCS_PGMAJFAULT] += val;
4183 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4184 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4185 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4186 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4187 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4188 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4189 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4190 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4191 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4192 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4196 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4198 struct mem_cgroup *iter;
4200 for_each_mem_cgroup_tree(iter, memcg)
4201 mem_cgroup_get_local_stat(iter, s);
4205 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4208 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4209 unsigned long node_nr;
4210 struct cgroup *cont = m->private;
4211 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4213 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4214 seq_printf(m, "total=%lu", total_nr);
4215 for_each_node_state(nid, N_HIGH_MEMORY) {
4216 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4217 seq_printf(m, " N%d=%lu", nid, node_nr);
4221 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4222 seq_printf(m, "file=%lu", file_nr);
4223 for_each_node_state(nid, N_HIGH_MEMORY) {
4224 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4226 seq_printf(m, " N%d=%lu", nid, node_nr);
4230 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4231 seq_printf(m, "anon=%lu", anon_nr);
4232 for_each_node_state(nid, N_HIGH_MEMORY) {
4233 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4235 seq_printf(m, " N%d=%lu", nid, node_nr);
4239 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4240 seq_printf(m, "unevictable=%lu", unevictable_nr);
4241 for_each_node_state(nid, N_HIGH_MEMORY) {
4242 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4243 BIT(LRU_UNEVICTABLE));
4244 seq_printf(m, " N%d=%lu", nid, node_nr);
4249 #endif /* CONFIG_NUMA */
4251 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4252 struct cgroup_map_cb *cb)
4254 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4255 struct mcs_total_stat mystat;
4258 memset(&mystat, 0, sizeof(mystat));
4259 mem_cgroup_get_local_stat(mem_cont, &mystat);
4262 for (i = 0; i < NR_MCS_STAT; i++) {
4263 if (i == MCS_SWAP && !do_swap_account)
4265 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4268 /* Hierarchical information */
4270 unsigned long long limit, memsw_limit;
4271 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4272 cb->fill(cb, "hierarchical_memory_limit", limit);
4273 if (do_swap_account)
4274 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4277 memset(&mystat, 0, sizeof(mystat));
4278 mem_cgroup_get_total_stat(mem_cont, &mystat);
4279 for (i = 0; i < NR_MCS_STAT; i++) {
4280 if (i == MCS_SWAP && !do_swap_account)
4282 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4285 #ifdef CONFIG_DEBUG_VM
4288 struct mem_cgroup_per_zone *mz;
4289 unsigned long recent_rotated[2] = {0, 0};
4290 unsigned long recent_scanned[2] = {0, 0};
4292 for_each_online_node(nid)
4293 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4294 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4296 recent_rotated[0] +=
4297 mz->reclaim_stat.recent_rotated[0];
4298 recent_rotated[1] +=
4299 mz->reclaim_stat.recent_rotated[1];
4300 recent_scanned[0] +=
4301 mz->reclaim_stat.recent_scanned[0];
4302 recent_scanned[1] +=
4303 mz->reclaim_stat.recent_scanned[1];
4305 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4306 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4307 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4308 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4315 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4317 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4319 return mem_cgroup_swappiness(memcg);
4322 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4325 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4326 struct mem_cgroup *parent;
4331 if (cgrp->parent == NULL)
4334 parent = mem_cgroup_from_cont(cgrp->parent);
4338 /* If under hierarchy, only empty-root can set this value */
4339 if ((parent->use_hierarchy) ||
4340 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4345 memcg->swappiness = val;
4352 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4354 struct mem_cgroup_threshold_ary *t;
4360 t = rcu_dereference(memcg->thresholds.primary);
4362 t = rcu_dereference(memcg->memsw_thresholds.primary);
4367 usage = mem_cgroup_usage(memcg, swap);
4370 * current_threshold points to threshold just below usage.
4371 * If it's not true, a threshold was crossed after last
4372 * call of __mem_cgroup_threshold().
4374 i = t->current_threshold;
4377 * Iterate backward over array of thresholds starting from
4378 * current_threshold and check if a threshold is crossed.
4379 * If none of thresholds below usage is crossed, we read
4380 * only one element of the array here.
4382 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4383 eventfd_signal(t->entries[i].eventfd, 1);
4385 /* i = current_threshold + 1 */
4389 * Iterate forward over array of thresholds starting from
4390 * current_threshold+1 and check if a threshold is crossed.
4391 * If none of thresholds above usage is crossed, we read
4392 * only one element of the array here.
4394 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4395 eventfd_signal(t->entries[i].eventfd, 1);
4397 /* Update current_threshold */
4398 t->current_threshold = i - 1;
4403 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4406 __mem_cgroup_threshold(memcg, false);
4407 if (do_swap_account)
4408 __mem_cgroup_threshold(memcg, true);
4410 memcg = parent_mem_cgroup(memcg);
4414 static int compare_thresholds(const void *a, const void *b)
4416 const struct mem_cgroup_threshold *_a = a;
4417 const struct mem_cgroup_threshold *_b = b;
4419 return _a->threshold - _b->threshold;
4422 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4424 struct mem_cgroup_eventfd_list *ev;
4426 list_for_each_entry(ev, &memcg->oom_notify, list)
4427 eventfd_signal(ev->eventfd, 1);
4431 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4433 struct mem_cgroup *iter;
4435 for_each_mem_cgroup_tree(iter, memcg)
4436 mem_cgroup_oom_notify_cb(iter);
4439 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4440 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4442 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4443 struct mem_cgroup_thresholds *thresholds;
4444 struct mem_cgroup_threshold_ary *new;
4445 int type = MEMFILE_TYPE(cft->private);
4446 u64 threshold, usage;
4449 ret = res_counter_memparse_write_strategy(args, &threshold);
4453 mutex_lock(&memcg->thresholds_lock);
4456 thresholds = &memcg->thresholds;
4457 else if (type == _MEMSWAP)
4458 thresholds = &memcg->memsw_thresholds;
4462 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4464 /* Check if a threshold crossed before adding a new one */
4465 if (thresholds->primary)
4466 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4468 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4470 /* Allocate memory for new array of thresholds */
4471 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4479 /* Copy thresholds (if any) to new array */
4480 if (thresholds->primary) {
4481 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4482 sizeof(struct mem_cgroup_threshold));
4485 /* Add new threshold */
4486 new->entries[size - 1].eventfd = eventfd;
4487 new->entries[size - 1].threshold = threshold;
4489 /* Sort thresholds. Registering of new threshold isn't time-critical */
4490 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4491 compare_thresholds, NULL);
4493 /* Find current threshold */
4494 new->current_threshold = -1;
4495 for (i = 0; i < size; i++) {
4496 if (new->entries[i].threshold < usage) {
4498 * new->current_threshold will not be used until
4499 * rcu_assign_pointer(), so it's safe to increment
4502 ++new->current_threshold;
4506 /* Free old spare buffer and save old primary buffer as spare */
4507 kfree(thresholds->spare);
4508 thresholds->spare = thresholds->primary;
4510 rcu_assign_pointer(thresholds->primary, new);
4512 /* To be sure that nobody uses thresholds */
4516 mutex_unlock(&memcg->thresholds_lock);
4521 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4522 struct cftype *cft, struct eventfd_ctx *eventfd)
4524 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4525 struct mem_cgroup_thresholds *thresholds;
4526 struct mem_cgroup_threshold_ary *new;
4527 int type = MEMFILE_TYPE(cft->private);
4531 mutex_lock(&memcg->thresholds_lock);
4533 thresholds = &memcg->thresholds;
4534 else if (type == _MEMSWAP)
4535 thresholds = &memcg->memsw_thresholds;
4540 * Something went wrong if we trying to unregister a threshold
4541 * if we don't have thresholds
4543 BUG_ON(!thresholds);
4545 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4547 /* Check if a threshold crossed before removing */
4548 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4550 /* Calculate new number of threshold */
4552 for (i = 0; i < thresholds->primary->size; i++) {
4553 if (thresholds->primary->entries[i].eventfd != eventfd)
4557 new = thresholds->spare;
4559 /* Set thresholds array to NULL if we don't have thresholds */
4568 /* Copy thresholds and find current threshold */
4569 new->current_threshold = -1;
4570 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4571 if (thresholds->primary->entries[i].eventfd == eventfd)
4574 new->entries[j] = thresholds->primary->entries[i];
4575 if (new->entries[j].threshold < usage) {
4577 * new->current_threshold will not be used
4578 * until rcu_assign_pointer(), so it's safe to increment
4581 ++new->current_threshold;
4587 /* Swap primary and spare array */
4588 thresholds->spare = thresholds->primary;
4589 rcu_assign_pointer(thresholds->primary, new);
4591 /* To be sure that nobody uses thresholds */
4594 mutex_unlock(&memcg->thresholds_lock);
4597 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4598 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4600 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4601 struct mem_cgroup_eventfd_list *event;
4602 int type = MEMFILE_TYPE(cft->private);
4604 BUG_ON(type != _OOM_TYPE);
4605 event = kmalloc(sizeof(*event), GFP_KERNEL);
4609 spin_lock(&memcg_oom_lock);
4611 event->eventfd = eventfd;
4612 list_add(&event->list, &memcg->oom_notify);
4614 /* already in OOM ? */
4615 if (atomic_read(&memcg->under_oom))
4616 eventfd_signal(eventfd, 1);
4617 spin_unlock(&memcg_oom_lock);
4622 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4623 struct cftype *cft, struct eventfd_ctx *eventfd)
4625 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4626 struct mem_cgroup_eventfd_list *ev, *tmp;
4627 int type = MEMFILE_TYPE(cft->private);
4629 BUG_ON(type != _OOM_TYPE);
4631 spin_lock(&memcg_oom_lock);
4633 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4634 if (ev->eventfd == eventfd) {
4635 list_del(&ev->list);
4640 spin_unlock(&memcg_oom_lock);
4643 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4644 struct cftype *cft, struct cgroup_map_cb *cb)
4646 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4648 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4650 if (atomic_read(&memcg->under_oom))
4651 cb->fill(cb, "under_oom", 1);
4653 cb->fill(cb, "under_oom", 0);
4657 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4658 struct cftype *cft, u64 val)
4660 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4661 struct mem_cgroup *parent;
4663 /* cannot set to root cgroup and only 0 and 1 are allowed */
4664 if (!cgrp->parent || !((val == 0) || (val == 1)))
4667 parent = mem_cgroup_from_cont(cgrp->parent);
4670 /* oom-kill-disable is a flag for subhierarchy. */
4671 if ((parent->use_hierarchy) ||
4672 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4676 memcg->oom_kill_disable = val;
4678 memcg_oom_recover(memcg);
4684 static const struct file_operations mem_control_numa_stat_file_operations = {
4686 .llseek = seq_lseek,
4687 .release = single_release,
4690 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4692 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4694 file->f_op = &mem_control_numa_stat_file_operations;
4695 return single_open(file, mem_control_numa_stat_show, cont);
4697 #endif /* CONFIG_NUMA */
4699 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4700 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4703 * Part of this would be better living in a separate allocation
4704 * function, leaving us with just the cgroup tree population work.
4705 * We, however, depend on state such as network's proto_list that
4706 * is only initialized after cgroup creation. I found the less
4707 * cumbersome way to deal with it to defer it all to populate time
4709 return mem_cgroup_sockets_init(cont, ss);
4712 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4713 struct cgroup *cont)
4715 mem_cgroup_sockets_destroy(cont, ss);
4718 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4723 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4724 struct cgroup *cont)
4729 static struct cftype mem_cgroup_files[] = {
4731 .name = "usage_in_bytes",
4732 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4733 .read_u64 = mem_cgroup_read,
4734 .register_event = mem_cgroup_usage_register_event,
4735 .unregister_event = mem_cgroup_usage_unregister_event,
4738 .name = "max_usage_in_bytes",
4739 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4740 .trigger = mem_cgroup_reset,
4741 .read_u64 = mem_cgroup_read,
4744 .name = "limit_in_bytes",
4745 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4746 .write_string = mem_cgroup_write,
4747 .read_u64 = mem_cgroup_read,
4750 .name = "soft_limit_in_bytes",
4751 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4752 .write_string = mem_cgroup_write,
4753 .read_u64 = mem_cgroup_read,
4757 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4758 .trigger = mem_cgroup_reset,
4759 .read_u64 = mem_cgroup_read,
4763 .read_map = mem_control_stat_show,
4766 .name = "force_empty",
4767 .trigger = mem_cgroup_force_empty_write,
4770 .name = "use_hierarchy",
4771 .write_u64 = mem_cgroup_hierarchy_write,
4772 .read_u64 = mem_cgroup_hierarchy_read,
4775 .name = "swappiness",
4776 .read_u64 = mem_cgroup_swappiness_read,
4777 .write_u64 = mem_cgroup_swappiness_write,
4780 .name = "move_charge_at_immigrate",
4781 .read_u64 = mem_cgroup_move_charge_read,
4782 .write_u64 = mem_cgroup_move_charge_write,
4785 .name = "oom_control",
4786 .read_map = mem_cgroup_oom_control_read,
4787 .write_u64 = mem_cgroup_oom_control_write,
4788 .register_event = mem_cgroup_oom_register_event,
4789 .unregister_event = mem_cgroup_oom_unregister_event,
4790 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4794 .name = "numa_stat",
4795 .open = mem_control_numa_stat_open,
4801 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4802 static struct cftype memsw_cgroup_files[] = {
4804 .name = "memsw.usage_in_bytes",
4805 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4806 .read_u64 = mem_cgroup_read,
4807 .register_event = mem_cgroup_usage_register_event,
4808 .unregister_event = mem_cgroup_usage_unregister_event,
4811 .name = "memsw.max_usage_in_bytes",
4812 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4813 .trigger = mem_cgroup_reset,
4814 .read_u64 = mem_cgroup_read,
4817 .name = "memsw.limit_in_bytes",
4818 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4819 .write_string = mem_cgroup_write,
4820 .read_u64 = mem_cgroup_read,
4823 .name = "memsw.failcnt",
4824 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4825 .trigger = mem_cgroup_reset,
4826 .read_u64 = mem_cgroup_read,
4830 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4832 if (!do_swap_account)
4834 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4835 ARRAY_SIZE(memsw_cgroup_files));
4838 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4844 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4846 struct mem_cgroup_per_node *pn;
4847 struct mem_cgroup_per_zone *mz;
4849 int zone, tmp = node;
4851 * This routine is called against possible nodes.
4852 * But it's BUG to call kmalloc() against offline node.
4854 * TODO: this routine can waste much memory for nodes which will
4855 * never be onlined. It's better to use memory hotplug callback
4858 if (!node_state(node, N_NORMAL_MEMORY))
4860 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4864 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4865 mz = &pn->zoneinfo[zone];
4867 INIT_LIST_HEAD(&mz->lists[l]);
4868 mz->usage_in_excess = 0;
4869 mz->on_tree = false;
4872 memcg->info.nodeinfo[node] = pn;
4876 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4878 kfree(memcg->info.nodeinfo[node]);
4881 static struct mem_cgroup *mem_cgroup_alloc(void)
4883 struct mem_cgroup *mem;
4884 int size = sizeof(struct mem_cgroup);
4886 /* Can be very big if MAX_NUMNODES is very big */
4887 if (size < PAGE_SIZE)
4888 mem = kzalloc(size, GFP_KERNEL);
4890 mem = vzalloc(size);
4895 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4898 spin_lock_init(&mem->pcp_counter_lock);
4902 if (size < PAGE_SIZE)
4910 * At destroying mem_cgroup, references from swap_cgroup can remain.
4911 * (scanning all at force_empty is too costly...)
4913 * Instead of clearing all references at force_empty, we remember
4914 * the number of reference from swap_cgroup and free mem_cgroup when
4915 * it goes down to 0.
4917 * Removal of cgroup itself succeeds regardless of refs from swap.
4920 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4924 mem_cgroup_remove_from_trees(memcg);
4925 free_css_id(&mem_cgroup_subsys, &memcg->css);
4927 for_each_node_state(node, N_POSSIBLE)
4928 free_mem_cgroup_per_zone_info(memcg, node);
4930 free_percpu(memcg->stat);
4931 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4937 static void mem_cgroup_get(struct mem_cgroup *memcg)
4939 atomic_inc(&memcg->refcnt);
4942 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4944 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4945 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4946 __mem_cgroup_free(memcg);
4948 mem_cgroup_put(parent);
4952 static void mem_cgroup_put(struct mem_cgroup *memcg)
4954 __mem_cgroup_put(memcg, 1);
4958 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4960 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4962 if (!memcg->res.parent)
4964 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4966 EXPORT_SYMBOL(parent_mem_cgroup);
4968 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4969 static void __init enable_swap_cgroup(void)
4971 if (!mem_cgroup_disabled() && really_do_swap_account)
4972 do_swap_account = 1;
4975 static void __init enable_swap_cgroup(void)
4980 static int mem_cgroup_soft_limit_tree_init(void)
4982 struct mem_cgroup_tree_per_node *rtpn;
4983 struct mem_cgroup_tree_per_zone *rtpz;
4984 int tmp, node, zone;
4986 for_each_node_state(node, N_POSSIBLE) {
4988 if (!node_state(node, N_NORMAL_MEMORY))
4990 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4994 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4996 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4997 rtpz = &rtpn->rb_tree_per_zone[zone];
4998 rtpz->rb_root = RB_ROOT;
4999 spin_lock_init(&rtpz->lock);
5005 static struct cgroup_subsys_state * __ref
5006 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5008 struct mem_cgroup *memcg, *parent;
5009 long error = -ENOMEM;
5012 memcg = mem_cgroup_alloc();
5014 return ERR_PTR(error);
5016 for_each_node_state(node, N_POSSIBLE)
5017 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5021 if (cont->parent == NULL) {
5023 enable_swap_cgroup();
5025 if (mem_cgroup_soft_limit_tree_init())
5027 root_mem_cgroup = memcg;
5028 for_each_possible_cpu(cpu) {
5029 struct memcg_stock_pcp *stock =
5030 &per_cpu(memcg_stock, cpu);
5031 INIT_WORK(&stock->work, drain_local_stock);
5033 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5035 parent = mem_cgroup_from_cont(cont->parent);
5036 memcg->use_hierarchy = parent->use_hierarchy;
5037 memcg->oom_kill_disable = parent->oom_kill_disable;
5040 if (parent && parent->use_hierarchy) {
5041 res_counter_init(&memcg->res, &parent->res);
5042 res_counter_init(&memcg->memsw, &parent->memsw);
5044 * We increment refcnt of the parent to ensure that we can
5045 * safely access it on res_counter_charge/uncharge.
5046 * This refcnt will be decremented when freeing this
5047 * mem_cgroup(see mem_cgroup_put).
5049 mem_cgroup_get(parent);
5051 res_counter_init(&memcg->res, NULL);
5052 res_counter_init(&memcg->memsw, NULL);
5054 memcg->last_scanned_node = MAX_NUMNODES;
5055 INIT_LIST_HEAD(&memcg->oom_notify);
5058 memcg->swappiness = mem_cgroup_swappiness(parent);
5059 atomic_set(&memcg->refcnt, 1);
5060 memcg->move_charge_at_immigrate = 0;
5061 mutex_init(&memcg->thresholds_lock);
5064 __mem_cgroup_free(memcg);
5065 return ERR_PTR(error);
5068 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5069 struct cgroup *cont)
5071 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5073 return mem_cgroup_force_empty(memcg, false);
5076 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5077 struct cgroup *cont)
5079 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5081 kmem_cgroup_destroy(ss, cont);
5083 mem_cgroup_put(memcg);
5086 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5087 struct cgroup *cont)
5091 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5092 ARRAY_SIZE(mem_cgroup_files));
5095 ret = register_memsw_files(cont, ss);
5098 ret = register_kmem_files(cont, ss);
5104 /* Handlers for move charge at task migration. */
5105 #define PRECHARGE_COUNT_AT_ONCE 256
5106 static int mem_cgroup_do_precharge(unsigned long count)
5109 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5110 struct mem_cgroup *memcg = mc.to;
5112 if (mem_cgroup_is_root(memcg)) {
5113 mc.precharge += count;
5114 /* we don't need css_get for root */
5117 /* try to charge at once */
5119 struct res_counter *dummy;
5121 * "memcg" cannot be under rmdir() because we've already checked
5122 * by cgroup_lock_live_cgroup() that it is not removed and we
5123 * are still under the same cgroup_mutex. So we can postpone
5126 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5128 if (do_swap_account && res_counter_charge(&memcg->memsw,
5129 PAGE_SIZE * count, &dummy)) {
5130 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5133 mc.precharge += count;
5137 /* fall back to one by one charge */
5139 if (signal_pending(current)) {
5143 if (!batch_count--) {
5144 batch_count = PRECHARGE_COUNT_AT_ONCE;
5147 ret = __mem_cgroup_try_charge(NULL,
5148 GFP_KERNEL, 1, &memcg, false);
5150 /* mem_cgroup_clear_mc() will do uncharge later */
5158 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5159 * @vma: the vma the pte to be checked belongs
5160 * @addr: the address corresponding to the pte to be checked
5161 * @ptent: the pte to be checked
5162 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5165 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5166 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5167 * move charge. if @target is not NULL, the page is stored in target->page
5168 * with extra refcnt got(Callers should handle it).
5169 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5170 * target for charge migration. if @target is not NULL, the entry is stored
5173 * Called with pte lock held.
5180 enum mc_target_type {
5181 MC_TARGET_NONE, /* not used */
5186 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5187 unsigned long addr, pte_t ptent)
5189 struct page *page = vm_normal_page(vma, addr, ptent);
5191 if (!page || !page_mapped(page))
5193 if (PageAnon(page)) {
5194 /* we don't move shared anon */
5195 if (!move_anon() || page_mapcount(page) > 2)
5197 } else if (!move_file())
5198 /* we ignore mapcount for file pages */
5200 if (!get_page_unless_zero(page))
5206 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5207 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5210 struct page *page = NULL;
5211 swp_entry_t ent = pte_to_swp_entry(ptent);
5213 if (!move_anon() || non_swap_entry(ent))
5215 usage_count = mem_cgroup_count_swap_user(ent, &page);
5216 if (usage_count > 1) { /* we don't move shared anon */
5221 if (do_swap_account)
5222 entry->val = ent.val;
5227 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5228 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5230 struct page *page = NULL;
5231 struct inode *inode;
5232 struct address_space *mapping;
5235 if (!vma->vm_file) /* anonymous vma */
5240 inode = vma->vm_file->f_path.dentry->d_inode;
5241 mapping = vma->vm_file->f_mapping;
5242 if (pte_none(ptent))
5243 pgoff = linear_page_index(vma, addr);
5244 else /* pte_file(ptent) is true */
5245 pgoff = pte_to_pgoff(ptent);
5247 /* page is moved even if it's not RSS of this task(page-faulted). */
5248 page = find_get_page(mapping, pgoff);
5251 /* shmem/tmpfs may report page out on swap: account for that too. */
5252 if (radix_tree_exceptional_entry(page)) {
5253 swp_entry_t swap = radix_to_swp_entry(page);
5254 if (do_swap_account)
5256 page = find_get_page(&swapper_space, swap.val);
5262 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5263 unsigned long addr, pte_t ptent, union mc_target *target)
5265 struct page *page = NULL;
5266 struct page_cgroup *pc;
5268 swp_entry_t ent = { .val = 0 };
5270 if (pte_present(ptent))
5271 page = mc_handle_present_pte(vma, addr, ptent);
5272 else if (is_swap_pte(ptent))
5273 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5274 else if (pte_none(ptent) || pte_file(ptent))
5275 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5277 if (!page && !ent.val)
5280 pc = lookup_page_cgroup(page);
5282 * Do only loose check w/o page_cgroup lock.
5283 * mem_cgroup_move_account() checks the pc is valid or not under
5286 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5287 ret = MC_TARGET_PAGE;
5289 target->page = page;
5291 if (!ret || !target)
5294 /* There is a swap entry and a page doesn't exist or isn't charged */
5295 if (ent.val && !ret &&
5296 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5297 ret = MC_TARGET_SWAP;
5304 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5305 unsigned long addr, unsigned long end,
5306 struct mm_walk *walk)
5308 struct vm_area_struct *vma = walk->private;
5312 split_huge_page_pmd(walk->mm, pmd);
5314 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5315 for (; addr != end; pte++, addr += PAGE_SIZE)
5316 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5317 mc.precharge++; /* increment precharge temporarily */
5318 pte_unmap_unlock(pte - 1, ptl);
5324 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5326 unsigned long precharge;
5327 struct vm_area_struct *vma;
5329 down_read(&mm->mmap_sem);
5330 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5331 struct mm_walk mem_cgroup_count_precharge_walk = {
5332 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5336 if (is_vm_hugetlb_page(vma))
5338 walk_page_range(vma->vm_start, vma->vm_end,
5339 &mem_cgroup_count_precharge_walk);
5341 up_read(&mm->mmap_sem);
5343 precharge = mc.precharge;
5349 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5351 unsigned long precharge = mem_cgroup_count_precharge(mm);
5353 VM_BUG_ON(mc.moving_task);
5354 mc.moving_task = current;
5355 return mem_cgroup_do_precharge(precharge);
5358 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5359 static void __mem_cgroup_clear_mc(void)
5361 struct mem_cgroup *from = mc.from;
5362 struct mem_cgroup *to = mc.to;
5364 /* we must uncharge all the leftover precharges from mc.to */
5366 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5370 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5371 * we must uncharge here.
5373 if (mc.moved_charge) {
5374 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5375 mc.moved_charge = 0;
5377 /* we must fixup refcnts and charges */
5378 if (mc.moved_swap) {
5379 /* uncharge swap account from the old cgroup */
5380 if (!mem_cgroup_is_root(mc.from))
5381 res_counter_uncharge(&mc.from->memsw,
5382 PAGE_SIZE * mc.moved_swap);
5383 __mem_cgroup_put(mc.from, mc.moved_swap);
5385 if (!mem_cgroup_is_root(mc.to)) {
5387 * we charged both to->res and to->memsw, so we should
5390 res_counter_uncharge(&mc.to->res,
5391 PAGE_SIZE * mc.moved_swap);
5393 /* we've already done mem_cgroup_get(mc.to) */
5396 memcg_oom_recover(from);
5397 memcg_oom_recover(to);
5398 wake_up_all(&mc.waitq);
5401 static void mem_cgroup_clear_mc(void)
5403 struct mem_cgroup *from = mc.from;
5406 * we must clear moving_task before waking up waiters at the end of
5409 mc.moving_task = NULL;
5410 __mem_cgroup_clear_mc();
5411 spin_lock(&mc.lock);
5414 spin_unlock(&mc.lock);
5415 mem_cgroup_end_move(from);
5418 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5419 struct cgroup *cgroup,
5420 struct cgroup_taskset *tset)
5422 struct task_struct *p = cgroup_taskset_first(tset);
5424 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5426 if (memcg->move_charge_at_immigrate) {
5427 struct mm_struct *mm;
5428 struct mem_cgroup *from = mem_cgroup_from_task(p);
5430 VM_BUG_ON(from == memcg);
5432 mm = get_task_mm(p);
5435 /* We move charges only when we move a owner of the mm */
5436 if (mm->owner == p) {
5439 VM_BUG_ON(mc.precharge);
5440 VM_BUG_ON(mc.moved_charge);
5441 VM_BUG_ON(mc.moved_swap);
5442 mem_cgroup_start_move(from);
5443 spin_lock(&mc.lock);
5446 spin_unlock(&mc.lock);
5447 /* We set mc.moving_task later */
5449 ret = mem_cgroup_precharge_mc(mm);
5451 mem_cgroup_clear_mc();
5458 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5459 struct cgroup *cgroup,
5460 struct cgroup_taskset *tset)
5462 mem_cgroup_clear_mc();
5465 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5466 unsigned long addr, unsigned long end,
5467 struct mm_walk *walk)
5470 struct vm_area_struct *vma = walk->private;
5474 split_huge_page_pmd(walk->mm, pmd);
5476 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5477 for (; addr != end; addr += PAGE_SIZE) {
5478 pte_t ptent = *(pte++);
5479 union mc_target target;
5482 struct page_cgroup *pc;
5488 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5490 case MC_TARGET_PAGE:
5492 if (isolate_lru_page(page))
5494 pc = lookup_page_cgroup(page);
5495 if (!mem_cgroup_move_account(page, 1, pc,
5496 mc.from, mc.to, false)) {
5498 /* we uncharge from mc.from later. */
5501 putback_lru_page(page);
5502 put: /* is_target_pte_for_mc() gets the page */
5505 case MC_TARGET_SWAP:
5507 if (!mem_cgroup_move_swap_account(ent,
5508 mc.from, mc.to, false)) {
5510 /* we fixup refcnts and charges later. */
5518 pte_unmap_unlock(pte - 1, ptl);
5523 * We have consumed all precharges we got in can_attach().
5524 * We try charge one by one, but don't do any additional
5525 * charges to mc.to if we have failed in charge once in attach()
5528 ret = mem_cgroup_do_precharge(1);
5536 static void mem_cgroup_move_charge(struct mm_struct *mm)
5538 struct vm_area_struct *vma;
5540 lru_add_drain_all();
5542 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5544 * Someone who are holding the mmap_sem might be waiting in
5545 * waitq. So we cancel all extra charges, wake up all waiters,
5546 * and retry. Because we cancel precharges, we might not be able
5547 * to move enough charges, but moving charge is a best-effort
5548 * feature anyway, so it wouldn't be a big problem.
5550 __mem_cgroup_clear_mc();
5554 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5556 struct mm_walk mem_cgroup_move_charge_walk = {
5557 .pmd_entry = mem_cgroup_move_charge_pte_range,
5561 if (is_vm_hugetlb_page(vma))
5563 ret = walk_page_range(vma->vm_start, vma->vm_end,
5564 &mem_cgroup_move_charge_walk);
5567 * means we have consumed all precharges and failed in
5568 * doing additional charge. Just abandon here.
5572 up_read(&mm->mmap_sem);
5575 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5576 struct cgroup *cont,
5577 struct cgroup_taskset *tset)
5579 struct task_struct *p = cgroup_taskset_first(tset);
5580 struct mm_struct *mm = get_task_mm(p);
5584 mem_cgroup_move_charge(mm);
5589 mem_cgroup_clear_mc();
5591 #else /* !CONFIG_MMU */
5592 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5593 struct cgroup *cgroup,
5594 struct cgroup_taskset *tset)
5598 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5599 struct cgroup *cgroup,
5600 struct cgroup_taskset *tset)
5603 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5604 struct cgroup *cont,
5605 struct cgroup_taskset *tset)
5610 struct cgroup_subsys mem_cgroup_subsys = {
5612 .subsys_id = mem_cgroup_subsys_id,
5613 .create = mem_cgroup_create,
5614 .pre_destroy = mem_cgroup_pre_destroy,
5615 .destroy = mem_cgroup_destroy,
5616 .populate = mem_cgroup_populate,
5617 .can_attach = mem_cgroup_can_attach,
5618 .cancel_attach = mem_cgroup_cancel_attach,
5619 .attach = mem_cgroup_move_task,
5624 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5625 static int __init enable_swap_account(char *s)
5627 /* consider enabled if no parameter or 1 is given */
5628 if (!strcmp(s, "1"))
5629 really_do_swap_account = 1;
5630 else if (!strcmp(s, "0"))
5631 really_do_swap_account = 0;
5634 __setup("swapaccount=", enable_swap_account);