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 if (mem_cgroup_is_root(pc->mem_cgroup))
1036 VM_BUG_ON(list_empty(&pc->lru));
1037 list_del_init(&pc->lru);
1040 void mem_cgroup_del_lru(struct page *page)
1042 mem_cgroup_del_lru_list(page, page_lru(page));
1046 * Writeback is about to end against a page which has been marked for immediate
1047 * reclaim. If it still appears to be reclaimable, move it to the tail of the
1050 void mem_cgroup_rotate_reclaimable_page(struct page *page)
1052 struct mem_cgroup_per_zone *mz;
1053 struct page_cgroup *pc;
1054 enum lru_list lru = page_lru(page);
1056 if (mem_cgroup_disabled())
1059 pc = lookup_page_cgroup(page);
1060 /* unused or root page is not rotated. */
1061 if (!PageCgroupUsed(pc))
1063 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1065 if (mem_cgroup_is_root(pc->mem_cgroup))
1067 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1068 list_move_tail(&pc->lru, &mz->lists[lru]);
1071 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1073 struct mem_cgroup_per_zone *mz;
1074 struct page_cgroup *pc;
1076 if (mem_cgroup_disabled())
1079 pc = lookup_page_cgroup(page);
1080 /* unused or root page is not rotated. */
1081 if (!PageCgroupUsed(pc))
1083 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1085 if (mem_cgroup_is_root(pc->mem_cgroup))
1087 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1088 list_move(&pc->lru, &mz->lists[lru]);
1091 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1093 struct page_cgroup *pc;
1094 struct mem_cgroup_per_zone *mz;
1096 if (mem_cgroup_disabled())
1098 pc = lookup_page_cgroup(page);
1099 VM_BUG_ON(PageCgroupAcctLRU(pc));
1102 * SetPageLRU SetPageCgroupUsed
1104 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1106 * Ensure that one of the two sides adds the page to the memcg
1107 * LRU during a race.
1110 if (!PageCgroupUsed(pc))
1112 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1114 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1115 /* huge page split is done under lru_lock. so, we have no races. */
1116 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1117 SetPageCgroupAcctLRU(pc);
1118 if (mem_cgroup_is_root(pc->mem_cgroup))
1120 list_add(&pc->lru, &mz->lists[lru]);
1124 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1125 * while it's linked to lru because the page may be reused after it's fully
1126 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1127 * It's done under lock_page and expected that zone->lru_lock isnever held.
1129 static void mem_cgroup_lru_del_before_commit(struct page *page)
1131 unsigned long flags;
1132 struct zone *zone = page_zone(page);
1133 struct page_cgroup *pc = lookup_page_cgroup(page);
1136 * Doing this check without taking ->lru_lock seems wrong but this
1137 * is safe. Because if page_cgroup's USED bit is unset, the page
1138 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1139 * set, the commit after this will fail, anyway.
1140 * This all charge/uncharge is done under some mutual execustion.
1141 * So, we don't need to taking care of changes in USED bit.
1143 if (likely(!PageLRU(page)))
1146 spin_lock_irqsave(&zone->lru_lock, flags);
1148 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1149 * is guarded by lock_page() because the page is SwapCache.
1151 if (!PageCgroupUsed(pc))
1152 mem_cgroup_del_lru_list(page, page_lru(page));
1153 spin_unlock_irqrestore(&zone->lru_lock, flags);
1156 static void mem_cgroup_lru_add_after_commit(struct page *page)
1158 unsigned long flags;
1159 struct zone *zone = page_zone(page);
1160 struct page_cgroup *pc = lookup_page_cgroup(page);
1163 * SetPageLRU SetPageCgroupUsed
1165 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1167 * Ensure that one of the two sides adds the page to the memcg
1168 * LRU during a race.
1171 /* taking care of that the page is added to LRU while we commit it */
1172 if (likely(!PageLRU(page)))
1174 spin_lock_irqsave(&zone->lru_lock, flags);
1175 /* link when the page is linked to LRU but page_cgroup isn't */
1176 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1177 mem_cgroup_add_lru_list(page, page_lru(page));
1178 spin_unlock_irqrestore(&zone->lru_lock, flags);
1182 void mem_cgroup_move_lists(struct page *page,
1183 enum lru_list from, enum lru_list to)
1185 if (mem_cgroup_disabled())
1187 mem_cgroup_del_lru_list(page, from);
1188 mem_cgroup_add_lru_list(page, to);
1192 * Checks whether given mem is same or in the root_mem_cgroup's
1195 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1196 struct mem_cgroup *memcg)
1198 if (root_memcg != memcg) {
1199 return (root_memcg->use_hierarchy &&
1200 css_is_ancestor(&memcg->css, &root_memcg->css));
1206 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1209 struct mem_cgroup *curr = NULL;
1210 struct task_struct *p;
1212 p = find_lock_task_mm(task);
1215 curr = try_get_mem_cgroup_from_mm(p->mm);
1220 * We should check use_hierarchy of "memcg" not "curr". Because checking
1221 * use_hierarchy of "curr" here make this function true if hierarchy is
1222 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1223 * hierarchy(even if use_hierarchy is disabled in "memcg").
1225 ret = mem_cgroup_same_or_subtree(memcg, curr);
1226 css_put(&curr->css);
1230 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1232 unsigned long inactive_ratio;
1233 int nid = zone_to_nid(zone);
1234 int zid = zone_idx(zone);
1235 unsigned long inactive;
1236 unsigned long active;
1239 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1240 BIT(LRU_INACTIVE_ANON));
1241 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1242 BIT(LRU_ACTIVE_ANON));
1244 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1246 inactive_ratio = int_sqrt(10 * gb);
1250 return inactive * inactive_ratio < active;
1253 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1255 unsigned long active;
1256 unsigned long inactive;
1257 int zid = zone_idx(zone);
1258 int nid = zone_to_nid(zone);
1260 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1261 BIT(LRU_INACTIVE_FILE));
1262 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1263 BIT(LRU_ACTIVE_FILE));
1265 return (active > inactive);
1268 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1271 int nid = zone_to_nid(zone);
1272 int zid = zone_idx(zone);
1273 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1275 return &mz->reclaim_stat;
1278 struct zone_reclaim_stat *
1279 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1281 struct page_cgroup *pc;
1282 struct mem_cgroup_per_zone *mz;
1284 if (mem_cgroup_disabled())
1287 pc = lookup_page_cgroup(page);
1288 if (!PageCgroupUsed(pc))
1290 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1292 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1293 return &mz->reclaim_stat;
1296 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1297 struct list_head *dst,
1298 unsigned long *scanned, int order,
1299 isolate_mode_t mode,
1301 struct mem_cgroup *mem_cont,
1302 int active, int file)
1304 unsigned long nr_taken = 0;
1308 struct list_head *src;
1309 struct page_cgroup *pc, *tmp;
1310 int nid = zone_to_nid(z);
1311 int zid = zone_idx(z);
1312 struct mem_cgroup_per_zone *mz;
1313 int lru = LRU_FILE * file + active;
1317 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1318 src = &mz->lists[lru];
1321 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1322 if (scan >= nr_to_scan)
1325 if (unlikely(!PageCgroupUsed(pc)))
1328 page = lookup_cgroup_page(pc);
1330 if (unlikely(!PageLRU(page)))
1334 ret = __isolate_lru_page(page, mode, file);
1337 list_move(&page->lru, dst);
1338 mem_cgroup_del_lru(page);
1339 nr_taken += hpage_nr_pages(page);
1342 /* we don't affect global LRU but rotate in our LRU */
1343 mem_cgroup_rotate_lru_list(page, page_lru(page));
1352 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1358 #define mem_cgroup_from_res_counter(counter, member) \
1359 container_of(counter, struct mem_cgroup, member)
1362 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1363 * @mem: the memory cgroup
1365 * Returns the maximum amount of memory @mem can be charged with, in
1368 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1370 unsigned long long margin;
1372 margin = res_counter_margin(&memcg->res);
1373 if (do_swap_account)
1374 margin = min(margin, res_counter_margin(&memcg->memsw));
1375 return margin >> PAGE_SHIFT;
1378 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1380 struct cgroup *cgrp = memcg->css.cgroup;
1383 if (cgrp->parent == NULL)
1384 return vm_swappiness;
1386 return memcg->swappiness;
1389 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1394 spin_lock(&memcg->pcp_counter_lock);
1395 for_each_online_cpu(cpu)
1396 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1397 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1398 spin_unlock(&memcg->pcp_counter_lock);
1404 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1411 spin_lock(&memcg->pcp_counter_lock);
1412 for_each_online_cpu(cpu)
1413 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1414 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1415 spin_unlock(&memcg->pcp_counter_lock);
1419 * 2 routines for checking "mem" is under move_account() or not.
1421 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1422 * for avoiding race in accounting. If true,
1423 * pc->mem_cgroup may be overwritten.
1425 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1426 * under hierarchy of moving cgroups. This is for
1427 * waiting at hith-memory prressure caused by "move".
1430 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1432 VM_BUG_ON(!rcu_read_lock_held());
1433 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1436 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1438 struct mem_cgroup *from;
1439 struct mem_cgroup *to;
1442 * Unlike task_move routines, we access mc.to, mc.from not under
1443 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1445 spin_lock(&mc.lock);
1451 ret = mem_cgroup_same_or_subtree(memcg, from)
1452 || mem_cgroup_same_or_subtree(memcg, to);
1454 spin_unlock(&mc.lock);
1458 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1460 if (mc.moving_task && current != mc.moving_task) {
1461 if (mem_cgroup_under_move(memcg)) {
1463 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1464 /* moving charge context might have finished. */
1467 finish_wait(&mc.waitq, &wait);
1475 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1476 * @memcg: The memory cgroup that went over limit
1477 * @p: Task that is going to be killed
1479 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1482 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1484 struct cgroup *task_cgrp;
1485 struct cgroup *mem_cgrp;
1487 * Need a buffer in BSS, can't rely on allocations. The code relies
1488 * on the assumption that OOM is serialized for memory controller.
1489 * If this assumption is broken, revisit this code.
1491 static char memcg_name[PATH_MAX];
1500 mem_cgrp = memcg->css.cgroup;
1501 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1503 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1506 * Unfortunately, we are unable to convert to a useful name
1507 * But we'll still print out the usage information
1514 printk(KERN_INFO "Task in %s killed", memcg_name);
1517 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1525 * Continues from above, so we don't need an KERN_ level
1527 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1530 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1531 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1532 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1533 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1534 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1536 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1537 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1538 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1542 * This function returns the number of memcg under hierarchy tree. Returns
1543 * 1(self count) if no children.
1545 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1548 struct mem_cgroup *iter;
1550 for_each_mem_cgroup_tree(iter, memcg)
1556 * Return the memory (and swap, if configured) limit for a memcg.
1558 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1563 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1564 limit += total_swap_pages << PAGE_SHIFT;
1566 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1568 * If memsw is finite and limits the amount of swap space available
1569 * to this memcg, return that limit.
1571 return min(limit, memsw);
1574 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1576 unsigned long flags)
1578 unsigned long total = 0;
1579 bool noswap = false;
1582 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1584 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1587 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1589 drain_all_stock_async(memcg);
1590 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1592 * Allow limit shrinkers, which are triggered directly
1593 * by userspace, to catch signals and stop reclaim
1594 * after minimal progress, regardless of the margin.
1596 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1598 if (mem_cgroup_margin(memcg))
1601 * If nothing was reclaimed after two attempts, there
1602 * may be no reclaimable pages in this hierarchy.
1611 * test_mem_cgroup_node_reclaimable
1612 * @mem: the target memcg
1613 * @nid: the node ID to be checked.
1614 * @noswap : specify true here if the user wants flle only information.
1616 * This function returns whether the specified memcg contains any
1617 * reclaimable pages on a node. Returns true if there are any reclaimable
1618 * pages in the node.
1620 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1621 int nid, bool noswap)
1623 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1625 if (noswap || !total_swap_pages)
1627 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1632 #if MAX_NUMNODES > 1
1635 * Always updating the nodemask is not very good - even if we have an empty
1636 * list or the wrong list here, we can start from some node and traverse all
1637 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1640 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1644 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1645 * pagein/pageout changes since the last update.
1647 if (!atomic_read(&memcg->numainfo_events))
1649 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1652 /* make a nodemask where this memcg uses memory from */
1653 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1655 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1657 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1658 node_clear(nid, memcg->scan_nodes);
1661 atomic_set(&memcg->numainfo_events, 0);
1662 atomic_set(&memcg->numainfo_updating, 0);
1666 * Selecting a node where we start reclaim from. Because what we need is just
1667 * reducing usage counter, start from anywhere is O,K. Considering
1668 * memory reclaim from current node, there are pros. and cons.
1670 * Freeing memory from current node means freeing memory from a node which
1671 * we'll use or we've used. So, it may make LRU bad. And if several threads
1672 * hit limits, it will see a contention on a node. But freeing from remote
1673 * node means more costs for memory reclaim because of memory latency.
1675 * Now, we use round-robin. Better algorithm is welcomed.
1677 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1681 mem_cgroup_may_update_nodemask(memcg);
1682 node = memcg->last_scanned_node;
1684 node = next_node(node, memcg->scan_nodes);
1685 if (node == MAX_NUMNODES)
1686 node = first_node(memcg->scan_nodes);
1688 * We call this when we hit limit, not when pages are added to LRU.
1689 * No LRU may hold pages because all pages are UNEVICTABLE or
1690 * memcg is too small and all pages are not on LRU. In that case,
1691 * we use curret node.
1693 if (unlikely(node == MAX_NUMNODES))
1694 node = numa_node_id();
1696 memcg->last_scanned_node = node;
1701 * Check all nodes whether it contains reclaimable pages or not.
1702 * For quick scan, we make use of scan_nodes. This will allow us to skip
1703 * unused nodes. But scan_nodes is lazily updated and may not cotain
1704 * enough new information. We need to do double check.
1706 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1711 * quick check...making use of scan_node.
1712 * We can skip unused nodes.
1714 if (!nodes_empty(memcg->scan_nodes)) {
1715 for (nid = first_node(memcg->scan_nodes);
1717 nid = next_node(nid, memcg->scan_nodes)) {
1719 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1724 * Check rest of nodes.
1726 for_each_node_state(nid, N_HIGH_MEMORY) {
1727 if (node_isset(nid, memcg->scan_nodes))
1729 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1736 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1741 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1743 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1747 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1750 unsigned long *total_scanned)
1752 struct mem_cgroup *victim = NULL;
1755 unsigned long excess;
1756 unsigned long nr_scanned;
1757 struct mem_cgroup_reclaim_cookie reclaim = {
1762 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1765 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1770 * If we have not been able to reclaim
1771 * anything, it might because there are
1772 * no reclaimable pages under this hierarchy
1777 * We want to do more targeted reclaim.
1778 * excess >> 2 is not to excessive so as to
1779 * reclaim too much, nor too less that we keep
1780 * coming back to reclaim from this cgroup
1782 if (total >= (excess >> 2) ||
1783 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1788 if (!mem_cgroup_reclaimable(victim, false))
1790 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1792 *total_scanned += nr_scanned;
1793 if (!res_counter_soft_limit_excess(&root_memcg->res))
1796 mem_cgroup_iter_break(root_memcg, victim);
1801 * Check OOM-Killer is already running under our hierarchy.
1802 * If someone is running, return false.
1803 * Has to be called with memcg_oom_lock
1805 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1807 struct mem_cgroup *iter, *failed = NULL;
1809 for_each_mem_cgroup_tree(iter, memcg) {
1810 if (iter->oom_lock) {
1812 * this subtree of our hierarchy is already locked
1813 * so we cannot give a lock.
1816 mem_cgroup_iter_break(memcg, iter);
1819 iter->oom_lock = true;
1826 * OK, we failed to lock the whole subtree so we have to clean up
1827 * what we set up to the failing subtree
1829 for_each_mem_cgroup_tree(iter, memcg) {
1830 if (iter == failed) {
1831 mem_cgroup_iter_break(memcg, iter);
1834 iter->oom_lock = false;
1840 * Has to be called with memcg_oom_lock
1842 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1844 struct mem_cgroup *iter;
1846 for_each_mem_cgroup_tree(iter, memcg)
1847 iter->oom_lock = false;
1851 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1853 struct mem_cgroup *iter;
1855 for_each_mem_cgroup_tree(iter, memcg)
1856 atomic_inc(&iter->under_oom);
1859 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1861 struct mem_cgroup *iter;
1864 * When a new child is created while the hierarchy is under oom,
1865 * mem_cgroup_oom_lock() may not be called. We have to use
1866 * atomic_add_unless() here.
1868 for_each_mem_cgroup_tree(iter, memcg)
1869 atomic_add_unless(&iter->under_oom, -1, 0);
1872 static DEFINE_SPINLOCK(memcg_oom_lock);
1873 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1875 struct oom_wait_info {
1876 struct mem_cgroup *mem;
1880 static int memcg_oom_wake_function(wait_queue_t *wait,
1881 unsigned mode, int sync, void *arg)
1883 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1885 struct oom_wait_info *oom_wait_info;
1887 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1888 oom_wait_memcg = oom_wait_info->mem;
1891 * Both of oom_wait_info->mem and wake_mem are stable under us.
1892 * Then we can use css_is_ancestor without taking care of RCU.
1894 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1895 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1897 return autoremove_wake_function(wait, mode, sync, arg);
1900 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1902 /* for filtering, pass "memcg" as argument. */
1903 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1906 static void memcg_oom_recover(struct mem_cgroup *memcg)
1908 if (memcg && atomic_read(&memcg->under_oom))
1909 memcg_wakeup_oom(memcg);
1913 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1915 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1917 struct oom_wait_info owait;
1918 bool locked, need_to_kill;
1921 owait.wait.flags = 0;
1922 owait.wait.func = memcg_oom_wake_function;
1923 owait.wait.private = current;
1924 INIT_LIST_HEAD(&owait.wait.task_list);
1925 need_to_kill = true;
1926 mem_cgroup_mark_under_oom(memcg);
1928 /* At first, try to OOM lock hierarchy under memcg.*/
1929 spin_lock(&memcg_oom_lock);
1930 locked = mem_cgroup_oom_lock(memcg);
1932 * Even if signal_pending(), we can't quit charge() loop without
1933 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1934 * under OOM is always welcomed, use TASK_KILLABLE here.
1936 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1937 if (!locked || memcg->oom_kill_disable)
1938 need_to_kill = false;
1940 mem_cgroup_oom_notify(memcg);
1941 spin_unlock(&memcg_oom_lock);
1944 finish_wait(&memcg_oom_waitq, &owait.wait);
1945 mem_cgroup_out_of_memory(memcg, mask);
1948 finish_wait(&memcg_oom_waitq, &owait.wait);
1950 spin_lock(&memcg_oom_lock);
1952 mem_cgroup_oom_unlock(memcg);
1953 memcg_wakeup_oom(memcg);
1954 spin_unlock(&memcg_oom_lock);
1956 mem_cgroup_unmark_under_oom(memcg);
1958 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1960 /* Give chance to dying process */
1961 schedule_timeout_uninterruptible(1);
1966 * Currently used to update mapped file statistics, but the routine can be
1967 * generalized to update other statistics as well.
1969 * Notes: Race condition
1971 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1972 * it tends to be costly. But considering some conditions, we doesn't need
1973 * to do so _always_.
1975 * Considering "charge", lock_page_cgroup() is not required because all
1976 * file-stat operations happen after a page is attached to radix-tree. There
1977 * are no race with "charge".
1979 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1980 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1981 * if there are race with "uncharge". Statistics itself is properly handled
1984 * Considering "move", this is an only case we see a race. To make the race
1985 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1986 * possibility of race condition. If there is, we take a lock.
1989 void mem_cgroup_update_page_stat(struct page *page,
1990 enum mem_cgroup_page_stat_item idx, int val)
1992 struct mem_cgroup *memcg;
1993 struct page_cgroup *pc = lookup_page_cgroup(page);
1994 bool need_unlock = false;
1995 unsigned long uninitialized_var(flags);
2001 memcg = pc->mem_cgroup;
2002 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2004 /* pc->mem_cgroup is unstable ? */
2005 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
2006 /* take a lock against to access pc->mem_cgroup */
2007 move_lock_page_cgroup(pc, &flags);
2009 memcg = pc->mem_cgroup;
2010 if (!memcg || !PageCgroupUsed(pc))
2015 case MEMCG_NR_FILE_MAPPED:
2017 SetPageCgroupFileMapped(pc);
2018 else if (!page_mapped(page))
2019 ClearPageCgroupFileMapped(pc);
2020 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2026 this_cpu_add(memcg->stat->count[idx], val);
2029 if (unlikely(need_unlock))
2030 move_unlock_page_cgroup(pc, &flags);
2034 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2037 * size of first charge trial. "32" comes from vmscan.c's magic value.
2038 * TODO: maybe necessary to use big numbers in big irons.
2040 #define CHARGE_BATCH 32U
2041 struct memcg_stock_pcp {
2042 struct mem_cgroup *cached; /* this never be root cgroup */
2043 unsigned int nr_pages;
2044 struct work_struct work;
2045 unsigned long flags;
2046 #define FLUSHING_CACHED_CHARGE (0)
2048 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2049 static DEFINE_MUTEX(percpu_charge_mutex);
2052 * Try to consume stocked charge on this cpu. If success, one page is consumed
2053 * from local stock and true is returned. If the stock is 0 or charges from a
2054 * cgroup which is not current target, returns false. This stock will be
2057 static bool consume_stock(struct mem_cgroup *memcg)
2059 struct memcg_stock_pcp *stock;
2062 stock = &get_cpu_var(memcg_stock);
2063 if (memcg == stock->cached && stock->nr_pages)
2065 else /* need to call res_counter_charge */
2067 put_cpu_var(memcg_stock);
2072 * Returns stocks cached in percpu to res_counter and reset cached information.
2074 static void drain_stock(struct memcg_stock_pcp *stock)
2076 struct mem_cgroup *old = stock->cached;
2078 if (stock->nr_pages) {
2079 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2081 res_counter_uncharge(&old->res, bytes);
2082 if (do_swap_account)
2083 res_counter_uncharge(&old->memsw, bytes);
2084 stock->nr_pages = 0;
2086 stock->cached = NULL;
2090 * This must be called under preempt disabled or must be called by
2091 * a thread which is pinned to local cpu.
2093 static void drain_local_stock(struct work_struct *dummy)
2095 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2097 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2101 * Cache charges(val) which is from res_counter, to local per_cpu area.
2102 * This will be consumed by consume_stock() function, later.
2104 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2106 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2108 if (stock->cached != memcg) { /* reset if necessary */
2110 stock->cached = memcg;
2112 stock->nr_pages += nr_pages;
2113 put_cpu_var(memcg_stock);
2117 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2118 * of the hierarchy under it. sync flag says whether we should block
2119 * until the work is done.
2121 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2125 /* Notify other cpus that system-wide "drain" is running */
2128 for_each_online_cpu(cpu) {
2129 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2130 struct mem_cgroup *memcg;
2132 memcg = stock->cached;
2133 if (!memcg || !stock->nr_pages)
2135 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2137 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2139 drain_local_stock(&stock->work);
2141 schedule_work_on(cpu, &stock->work);
2149 for_each_online_cpu(cpu) {
2150 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2151 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2152 flush_work(&stock->work);
2159 * Tries to drain stocked charges in other cpus. This function is asynchronous
2160 * and just put a work per cpu for draining localy on each cpu. Caller can
2161 * expects some charges will be back to res_counter later but cannot wait for
2164 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2167 * If someone calls draining, avoid adding more kworker runs.
2169 if (!mutex_trylock(&percpu_charge_mutex))
2171 drain_all_stock(root_memcg, false);
2172 mutex_unlock(&percpu_charge_mutex);
2175 /* This is a synchronous drain interface. */
2176 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2178 /* called when force_empty is called */
2179 mutex_lock(&percpu_charge_mutex);
2180 drain_all_stock(root_memcg, true);
2181 mutex_unlock(&percpu_charge_mutex);
2185 * This function drains percpu counter value from DEAD cpu and
2186 * move it to local cpu. Note that this function can be preempted.
2188 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2192 spin_lock(&memcg->pcp_counter_lock);
2193 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2194 long x = per_cpu(memcg->stat->count[i], cpu);
2196 per_cpu(memcg->stat->count[i], cpu) = 0;
2197 memcg->nocpu_base.count[i] += x;
2199 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2200 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2202 per_cpu(memcg->stat->events[i], cpu) = 0;
2203 memcg->nocpu_base.events[i] += x;
2205 /* need to clear ON_MOVE value, works as a kind of lock. */
2206 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2207 spin_unlock(&memcg->pcp_counter_lock);
2210 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2212 int idx = MEM_CGROUP_ON_MOVE;
2214 spin_lock(&memcg->pcp_counter_lock);
2215 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2216 spin_unlock(&memcg->pcp_counter_lock);
2219 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2220 unsigned long action,
2223 int cpu = (unsigned long)hcpu;
2224 struct memcg_stock_pcp *stock;
2225 struct mem_cgroup *iter;
2227 if ((action == CPU_ONLINE)) {
2228 for_each_mem_cgroup(iter)
2229 synchronize_mem_cgroup_on_move(iter, cpu);
2233 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2236 for_each_mem_cgroup(iter)
2237 mem_cgroup_drain_pcp_counter(iter, cpu);
2239 stock = &per_cpu(memcg_stock, cpu);
2245 /* See __mem_cgroup_try_charge() for details */
2247 CHARGE_OK, /* success */
2248 CHARGE_RETRY, /* need to retry but retry is not bad */
2249 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2250 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2251 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2254 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2255 unsigned int nr_pages, bool oom_check)
2257 unsigned long csize = nr_pages * PAGE_SIZE;
2258 struct mem_cgroup *mem_over_limit;
2259 struct res_counter *fail_res;
2260 unsigned long flags = 0;
2263 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2266 if (!do_swap_account)
2268 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2272 res_counter_uncharge(&memcg->res, csize);
2273 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2274 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2276 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2278 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2279 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2281 * Never reclaim on behalf of optional batching, retry with a
2282 * single page instead.
2284 if (nr_pages == CHARGE_BATCH)
2285 return CHARGE_RETRY;
2287 if (!(gfp_mask & __GFP_WAIT))
2288 return CHARGE_WOULDBLOCK;
2290 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2291 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2292 return CHARGE_RETRY;
2294 * Even though the limit is exceeded at this point, reclaim
2295 * may have been able to free some pages. Retry the charge
2296 * before killing the task.
2298 * Only for regular pages, though: huge pages are rather
2299 * unlikely to succeed so close to the limit, and we fall back
2300 * to regular pages anyway in case of failure.
2302 if (nr_pages == 1 && ret)
2303 return CHARGE_RETRY;
2306 * At task move, charge accounts can be doubly counted. So, it's
2307 * better to wait until the end of task_move if something is going on.
2309 if (mem_cgroup_wait_acct_move(mem_over_limit))
2310 return CHARGE_RETRY;
2312 /* If we don't need to call oom-killer at el, return immediately */
2314 return CHARGE_NOMEM;
2316 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2317 return CHARGE_OOM_DIE;
2319 return CHARGE_RETRY;
2323 * Unlike exported interface, "oom" parameter is added. if oom==true,
2324 * oom-killer can be invoked.
2326 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2328 unsigned int nr_pages,
2329 struct mem_cgroup **ptr,
2332 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2333 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2334 struct mem_cgroup *memcg = NULL;
2338 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2339 * in system level. So, allow to go ahead dying process in addition to
2342 if (unlikely(test_thread_flag(TIF_MEMDIE)
2343 || fatal_signal_pending(current)))
2347 * We always charge the cgroup the mm_struct belongs to.
2348 * The mm_struct's mem_cgroup changes on task migration if the
2349 * thread group leader migrates. It's possible that mm is not
2350 * set, if so charge the init_mm (happens for pagecache usage).
2355 if (*ptr) { /* css should be a valid one */
2357 VM_BUG_ON(css_is_removed(&memcg->css));
2358 if (mem_cgroup_is_root(memcg))
2360 if (nr_pages == 1 && consume_stock(memcg))
2362 css_get(&memcg->css);
2364 struct task_struct *p;
2367 p = rcu_dereference(mm->owner);
2369 * Because we don't have task_lock(), "p" can exit.
2370 * In that case, "memcg" can point to root or p can be NULL with
2371 * race with swapoff. Then, we have small risk of mis-accouning.
2372 * But such kind of mis-account by race always happens because
2373 * we don't have cgroup_mutex(). It's overkill and we allo that
2375 * (*) swapoff at el will charge against mm-struct not against
2376 * task-struct. So, mm->owner can be NULL.
2378 memcg = mem_cgroup_from_task(p);
2379 if (!memcg || mem_cgroup_is_root(memcg)) {
2383 if (nr_pages == 1 && consume_stock(memcg)) {
2385 * It seems dagerous to access memcg without css_get().
2386 * But considering how consume_stok works, it's not
2387 * necessary. If consume_stock success, some charges
2388 * from this memcg are cached on this cpu. So, we
2389 * don't need to call css_get()/css_tryget() before
2390 * calling consume_stock().
2395 /* after here, we may be blocked. we need to get refcnt */
2396 if (!css_tryget(&memcg->css)) {
2406 /* If killed, bypass charge */
2407 if (fatal_signal_pending(current)) {
2408 css_put(&memcg->css);
2413 if (oom && !nr_oom_retries) {
2415 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2418 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2422 case CHARGE_RETRY: /* not in OOM situation but retry */
2424 css_put(&memcg->css);
2427 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2428 css_put(&memcg->css);
2430 case CHARGE_NOMEM: /* OOM routine works */
2432 css_put(&memcg->css);
2435 /* If oom, we never return -ENOMEM */
2438 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2439 css_put(&memcg->css);
2442 } while (ret != CHARGE_OK);
2444 if (batch > nr_pages)
2445 refill_stock(memcg, batch - nr_pages);
2446 css_put(&memcg->css);
2459 * Somemtimes we have to undo a charge we got by try_charge().
2460 * This function is for that and do uncharge, put css's refcnt.
2461 * gotten by try_charge().
2463 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2464 unsigned int nr_pages)
2466 if (!mem_cgroup_is_root(memcg)) {
2467 unsigned long bytes = nr_pages * PAGE_SIZE;
2469 res_counter_uncharge(&memcg->res, bytes);
2470 if (do_swap_account)
2471 res_counter_uncharge(&memcg->memsw, bytes);
2476 * A helper function to get mem_cgroup from ID. must be called under
2477 * rcu_read_lock(). The caller must check css_is_removed() or some if
2478 * it's concern. (dropping refcnt from swap can be called against removed
2481 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2483 struct cgroup_subsys_state *css;
2485 /* ID 0 is unused ID */
2488 css = css_lookup(&mem_cgroup_subsys, id);
2491 return container_of(css, struct mem_cgroup, css);
2494 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2496 struct mem_cgroup *memcg = NULL;
2497 struct page_cgroup *pc;
2501 VM_BUG_ON(!PageLocked(page));
2503 pc = lookup_page_cgroup(page);
2504 lock_page_cgroup(pc);
2505 if (PageCgroupUsed(pc)) {
2506 memcg = pc->mem_cgroup;
2507 if (memcg && !css_tryget(&memcg->css))
2509 } else if (PageSwapCache(page)) {
2510 ent.val = page_private(page);
2511 id = lookup_swap_cgroup(ent);
2513 memcg = mem_cgroup_lookup(id);
2514 if (memcg && !css_tryget(&memcg->css))
2518 unlock_page_cgroup(pc);
2522 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2524 unsigned int nr_pages,
2525 struct page_cgroup *pc,
2526 enum charge_type ctype)
2528 lock_page_cgroup(pc);
2529 if (unlikely(PageCgroupUsed(pc))) {
2530 unlock_page_cgroup(pc);
2531 __mem_cgroup_cancel_charge(memcg, nr_pages);
2535 * we don't need page_cgroup_lock about tail pages, becase they are not
2536 * accessed by any other context at this point.
2538 pc->mem_cgroup = memcg;
2540 * We access a page_cgroup asynchronously without lock_page_cgroup().
2541 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2542 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2543 * before USED bit, we need memory barrier here.
2544 * See mem_cgroup_add_lru_list(), etc.
2548 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2549 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2550 SetPageCgroupCache(pc);
2551 SetPageCgroupUsed(pc);
2553 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2554 ClearPageCgroupCache(pc);
2555 SetPageCgroupUsed(pc);
2561 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2562 unlock_page_cgroup(pc);
2564 * "charge_statistics" updated event counter. Then, check it.
2565 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2566 * if they exceeds softlimit.
2568 memcg_check_events(memcg, page);
2571 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2573 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2574 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2576 * Because tail pages are not marked as "used", set it. We're under
2577 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2579 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2581 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2582 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2583 unsigned long flags;
2585 if (mem_cgroup_disabled())
2588 * We have no races with charge/uncharge but will have races with
2589 * page state accounting.
2591 move_lock_page_cgroup(head_pc, &flags);
2593 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2594 smp_wmb(); /* see __commit_charge() */
2595 if (PageCgroupAcctLRU(head_pc)) {
2597 struct mem_cgroup_per_zone *mz;
2600 * LRU flags cannot be copied because we need to add tail
2601 *.page to LRU by generic call and our hook will be called.
2602 * We hold lru_lock, then, reduce counter directly.
2604 lru = page_lru(head);
2605 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2606 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2608 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2609 move_unlock_page_cgroup(head_pc, &flags);
2614 * mem_cgroup_move_account - move account of the page
2616 * @nr_pages: number of regular pages (>1 for huge pages)
2617 * @pc: page_cgroup of the page.
2618 * @from: mem_cgroup which the page is moved from.
2619 * @to: mem_cgroup which the page is moved to. @from != @to.
2620 * @uncharge: whether we should call uncharge and css_put against @from.
2622 * The caller must confirm following.
2623 * - page is not on LRU (isolate_page() is useful.)
2624 * - compound_lock is held when nr_pages > 1
2626 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2627 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2628 * true, this function does "uncharge" from old cgroup, but it doesn't if
2629 * @uncharge is false, so a caller should do "uncharge".
2631 static int mem_cgroup_move_account(struct page *page,
2632 unsigned int nr_pages,
2633 struct page_cgroup *pc,
2634 struct mem_cgroup *from,
2635 struct mem_cgroup *to,
2638 unsigned long flags;
2641 VM_BUG_ON(from == to);
2642 VM_BUG_ON(PageLRU(page));
2644 * The page is isolated from LRU. So, collapse function
2645 * will not handle this page. But page splitting can happen.
2646 * Do this check under compound_page_lock(). The caller should
2650 if (nr_pages > 1 && !PageTransHuge(page))
2653 lock_page_cgroup(pc);
2656 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2659 move_lock_page_cgroup(pc, &flags);
2661 if (PageCgroupFileMapped(pc)) {
2662 /* Update mapped_file data for mem_cgroup */
2664 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2665 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2668 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2670 /* This is not "cancel", but cancel_charge does all we need. */
2671 __mem_cgroup_cancel_charge(from, nr_pages);
2673 /* caller should have done css_get */
2674 pc->mem_cgroup = to;
2675 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2677 * We charges against "to" which may not have any tasks. Then, "to"
2678 * can be under rmdir(). But in current implementation, caller of
2679 * this function is just force_empty() and move charge, so it's
2680 * guaranteed that "to" is never removed. So, we don't check rmdir
2683 move_unlock_page_cgroup(pc, &flags);
2686 unlock_page_cgroup(pc);
2690 memcg_check_events(to, page);
2691 memcg_check_events(from, page);
2697 * move charges to its parent.
2700 static int mem_cgroup_move_parent(struct page *page,
2701 struct page_cgroup *pc,
2702 struct mem_cgroup *child,
2705 struct cgroup *cg = child->css.cgroup;
2706 struct cgroup *pcg = cg->parent;
2707 struct mem_cgroup *parent;
2708 unsigned int nr_pages;
2709 unsigned long uninitialized_var(flags);
2717 if (!get_page_unless_zero(page))
2719 if (isolate_lru_page(page))
2722 nr_pages = hpage_nr_pages(page);
2724 parent = mem_cgroup_from_cont(pcg);
2725 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2730 flags = compound_lock_irqsave(page);
2732 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2734 __mem_cgroup_cancel_charge(parent, nr_pages);
2737 compound_unlock_irqrestore(page, flags);
2739 putback_lru_page(page);
2747 * Charge the memory controller for page usage.
2749 * 0 if the charge was successful
2750 * < 0 if the cgroup is over its limit
2752 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2753 gfp_t gfp_mask, enum charge_type ctype)
2755 struct mem_cgroup *memcg = NULL;
2756 unsigned int nr_pages = 1;
2757 struct page_cgroup *pc;
2761 if (PageTransHuge(page)) {
2762 nr_pages <<= compound_order(page);
2763 VM_BUG_ON(!PageTransHuge(page));
2765 * Never OOM-kill a process for a huge page. The
2766 * fault handler will fall back to regular pages.
2771 pc = lookup_page_cgroup(page);
2772 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2774 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2778 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2782 int mem_cgroup_newpage_charge(struct page *page,
2783 struct mm_struct *mm, gfp_t gfp_mask)
2785 if (mem_cgroup_disabled())
2788 * If already mapped, we don't have to account.
2789 * If page cache, page->mapping has address_space.
2790 * But page->mapping may have out-of-use anon_vma pointer,
2791 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2794 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2798 return mem_cgroup_charge_common(page, mm, gfp_mask,
2799 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2803 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2804 enum charge_type ctype);
2807 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2808 enum charge_type ctype)
2810 struct page_cgroup *pc = lookup_page_cgroup(page);
2812 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2813 * is already on LRU. It means the page may on some other page_cgroup's
2814 * LRU. Take care of it.
2816 mem_cgroup_lru_del_before_commit(page);
2817 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2818 mem_cgroup_lru_add_after_commit(page);
2822 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2825 struct mem_cgroup *memcg = NULL;
2828 if (mem_cgroup_disabled())
2830 if (PageCompound(page))
2836 if (page_is_file_cache(page)) {
2837 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2842 * FUSE reuses pages without going through the final
2843 * put that would remove them from the LRU list, make
2844 * sure that they get relinked properly.
2846 __mem_cgroup_commit_charge_lrucare(page, memcg,
2847 MEM_CGROUP_CHARGE_TYPE_CACHE);
2851 if (PageSwapCache(page)) {
2852 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2854 __mem_cgroup_commit_charge_swapin(page, memcg,
2855 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2857 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2858 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2864 * While swap-in, try_charge -> commit or cancel, the page is locked.
2865 * And when try_charge() successfully returns, one refcnt to memcg without
2866 * struct page_cgroup is acquired. This refcnt will be consumed by
2867 * "commit()" or removed by "cancel()"
2869 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2871 gfp_t mask, struct mem_cgroup **ptr)
2873 struct mem_cgroup *memcg;
2878 if (mem_cgroup_disabled())
2881 if (!do_swap_account)
2884 * A racing thread's fault, or swapoff, may have already updated
2885 * the pte, and even removed page from swap cache: in those cases
2886 * do_swap_page()'s pte_same() test will fail; but there's also a
2887 * KSM case which does need to charge the page.
2889 if (!PageSwapCache(page))
2891 memcg = try_get_mem_cgroup_from_page(page);
2895 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2896 css_put(&memcg->css);
2901 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2905 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2906 enum charge_type ctype)
2908 if (mem_cgroup_disabled())
2912 cgroup_exclude_rmdir(&ptr->css);
2914 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2916 * Now swap is on-memory. This means this page may be
2917 * counted both as mem and swap....double count.
2918 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2919 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2920 * may call delete_from_swap_cache() before reach here.
2922 if (do_swap_account && PageSwapCache(page)) {
2923 swp_entry_t ent = {.val = page_private(page)};
2925 struct mem_cgroup *memcg;
2927 id = swap_cgroup_record(ent, 0);
2929 memcg = mem_cgroup_lookup(id);
2932 * This recorded memcg can be obsolete one. So, avoid
2933 * calling css_tryget
2935 if (!mem_cgroup_is_root(memcg))
2936 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2937 mem_cgroup_swap_statistics(memcg, false);
2938 mem_cgroup_put(memcg);
2943 * At swapin, we may charge account against cgroup which has no tasks.
2944 * So, rmdir()->pre_destroy() can be called while we do this charge.
2945 * In that case, we need to call pre_destroy() again. check it here.
2947 cgroup_release_and_wakeup_rmdir(&ptr->css);
2950 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2952 __mem_cgroup_commit_charge_swapin(page, ptr,
2953 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2956 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2958 if (mem_cgroup_disabled())
2962 __mem_cgroup_cancel_charge(memcg, 1);
2965 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2966 unsigned int nr_pages,
2967 const enum charge_type ctype)
2969 struct memcg_batch_info *batch = NULL;
2970 bool uncharge_memsw = true;
2972 /* If swapout, usage of swap doesn't decrease */
2973 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2974 uncharge_memsw = false;
2976 batch = ¤t->memcg_batch;
2978 * In usual, we do css_get() when we remember memcg pointer.
2979 * But in this case, we keep res->usage until end of a series of
2980 * uncharges. Then, it's ok to ignore memcg's refcnt.
2983 batch->memcg = memcg;
2985 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2986 * In those cases, all pages freed continuously can be expected to be in
2987 * the same cgroup and we have chance to coalesce uncharges.
2988 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2989 * because we want to do uncharge as soon as possible.
2992 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2993 goto direct_uncharge;
2996 goto direct_uncharge;
2999 * In typical case, batch->memcg == mem. This means we can
3000 * merge a series of uncharges to an uncharge of res_counter.
3001 * If not, we uncharge res_counter ony by one.
3003 if (batch->memcg != memcg)
3004 goto direct_uncharge;
3005 /* remember freed charge and uncharge it later */
3008 batch->memsw_nr_pages++;
3011 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3013 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3014 if (unlikely(batch->memcg != memcg))
3015 memcg_oom_recover(memcg);
3020 * uncharge if !page_mapped(page)
3022 static struct mem_cgroup *
3023 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3025 struct mem_cgroup *memcg = NULL;
3026 unsigned int nr_pages = 1;
3027 struct page_cgroup *pc;
3029 if (mem_cgroup_disabled())
3032 if (PageSwapCache(page))
3035 if (PageTransHuge(page)) {
3036 nr_pages <<= compound_order(page);
3037 VM_BUG_ON(!PageTransHuge(page));
3040 * Check if our page_cgroup is valid
3042 pc = lookup_page_cgroup(page);
3043 if (unlikely(!pc || !PageCgroupUsed(pc)))
3046 lock_page_cgroup(pc);
3048 memcg = pc->mem_cgroup;
3050 if (!PageCgroupUsed(pc))
3054 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3055 case MEM_CGROUP_CHARGE_TYPE_DROP:
3056 /* See mem_cgroup_prepare_migration() */
3057 if (page_mapped(page) || PageCgroupMigration(pc))
3060 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3061 if (!PageAnon(page)) { /* Shared memory */
3062 if (page->mapping && !page_is_file_cache(page))
3064 } else if (page_mapped(page)) /* Anon */
3071 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3073 ClearPageCgroupUsed(pc);
3075 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3076 * freed from LRU. This is safe because uncharged page is expected not
3077 * to be reused (freed soon). Exception is SwapCache, it's handled by
3078 * special functions.
3081 unlock_page_cgroup(pc);
3083 * even after unlock, we have memcg->res.usage here and this memcg
3084 * will never be freed.
3086 memcg_check_events(memcg, page);
3087 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3088 mem_cgroup_swap_statistics(memcg, true);
3089 mem_cgroup_get(memcg);
3091 if (!mem_cgroup_is_root(memcg))
3092 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3097 unlock_page_cgroup(pc);
3101 void mem_cgroup_uncharge_page(struct page *page)
3104 if (page_mapped(page))
3106 if (page->mapping && !PageAnon(page))
3108 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3111 void mem_cgroup_uncharge_cache_page(struct page *page)
3113 VM_BUG_ON(page_mapped(page));
3114 VM_BUG_ON(page->mapping);
3115 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3119 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3120 * In that cases, pages are freed continuously and we can expect pages
3121 * are in the same memcg. All these calls itself limits the number of
3122 * pages freed at once, then uncharge_start/end() is called properly.
3123 * This may be called prural(2) times in a context,
3126 void mem_cgroup_uncharge_start(void)
3128 current->memcg_batch.do_batch++;
3129 /* We can do nest. */
3130 if (current->memcg_batch.do_batch == 1) {
3131 current->memcg_batch.memcg = NULL;
3132 current->memcg_batch.nr_pages = 0;
3133 current->memcg_batch.memsw_nr_pages = 0;
3137 void mem_cgroup_uncharge_end(void)
3139 struct memcg_batch_info *batch = ¤t->memcg_batch;
3141 if (!batch->do_batch)
3145 if (batch->do_batch) /* If stacked, do nothing. */
3151 * This "batch->memcg" is valid without any css_get/put etc...
3152 * bacause we hide charges behind us.
3154 if (batch->nr_pages)
3155 res_counter_uncharge(&batch->memcg->res,
3156 batch->nr_pages * PAGE_SIZE);
3157 if (batch->memsw_nr_pages)
3158 res_counter_uncharge(&batch->memcg->memsw,
3159 batch->memsw_nr_pages * PAGE_SIZE);
3160 memcg_oom_recover(batch->memcg);
3161 /* forget this pointer (for sanity check) */
3162 batch->memcg = NULL;
3167 * called after __delete_from_swap_cache() and drop "page" account.
3168 * memcg information is recorded to swap_cgroup of "ent"
3171 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3173 struct mem_cgroup *memcg;
3174 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3176 if (!swapout) /* this was a swap cache but the swap is unused ! */
3177 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3179 memcg = __mem_cgroup_uncharge_common(page, ctype);
3182 * record memcg information, if swapout && memcg != NULL,
3183 * mem_cgroup_get() was called in uncharge().
3185 if (do_swap_account && swapout && memcg)
3186 swap_cgroup_record(ent, css_id(&memcg->css));
3190 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3192 * called from swap_entry_free(). remove record in swap_cgroup and
3193 * uncharge "memsw" account.
3195 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3197 struct mem_cgroup *memcg;
3200 if (!do_swap_account)
3203 id = swap_cgroup_record(ent, 0);
3205 memcg = mem_cgroup_lookup(id);
3208 * We uncharge this because swap is freed.
3209 * This memcg can be obsolete one. We avoid calling css_tryget
3211 if (!mem_cgroup_is_root(memcg))
3212 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3213 mem_cgroup_swap_statistics(memcg, false);
3214 mem_cgroup_put(memcg);
3220 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3221 * @entry: swap entry to be moved
3222 * @from: mem_cgroup which the entry is moved from
3223 * @to: mem_cgroup which the entry is moved to
3224 * @need_fixup: whether we should fixup res_counters and refcounts.
3226 * It succeeds only when the swap_cgroup's record for this entry is the same
3227 * as the mem_cgroup's id of @from.
3229 * Returns 0 on success, -EINVAL on failure.
3231 * The caller must have charged to @to, IOW, called res_counter_charge() about
3232 * both res and memsw, and called css_get().
3234 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3235 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3237 unsigned short old_id, new_id;
3239 old_id = css_id(&from->css);
3240 new_id = css_id(&to->css);
3242 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3243 mem_cgroup_swap_statistics(from, false);
3244 mem_cgroup_swap_statistics(to, true);
3246 * This function is only called from task migration context now.
3247 * It postpones res_counter and refcount handling till the end
3248 * of task migration(mem_cgroup_clear_mc()) for performance
3249 * improvement. But we cannot postpone mem_cgroup_get(to)
3250 * because if the process that has been moved to @to does
3251 * swap-in, the refcount of @to might be decreased to 0.
3255 if (!mem_cgroup_is_root(from))
3256 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3257 mem_cgroup_put(from);
3259 * we charged both to->res and to->memsw, so we should
3262 if (!mem_cgroup_is_root(to))
3263 res_counter_uncharge(&to->res, PAGE_SIZE);
3270 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3271 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3278 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3281 int mem_cgroup_prepare_migration(struct page *page,
3282 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3284 struct mem_cgroup *memcg = NULL;
3285 struct page_cgroup *pc;
3286 enum charge_type ctype;
3291 VM_BUG_ON(PageTransHuge(page));
3292 if (mem_cgroup_disabled())
3295 pc = lookup_page_cgroup(page);
3296 lock_page_cgroup(pc);
3297 if (PageCgroupUsed(pc)) {
3298 memcg = pc->mem_cgroup;
3299 css_get(&memcg->css);
3301 * At migrating an anonymous page, its mapcount goes down
3302 * to 0 and uncharge() will be called. But, even if it's fully
3303 * unmapped, migration may fail and this page has to be
3304 * charged again. We set MIGRATION flag here and delay uncharge
3305 * until end_migration() is called
3307 * Corner Case Thinking
3309 * When the old page was mapped as Anon and it's unmap-and-freed
3310 * while migration was ongoing.
3311 * If unmap finds the old page, uncharge() of it will be delayed
3312 * until end_migration(). If unmap finds a new page, it's
3313 * uncharged when it make mapcount to be 1->0. If unmap code
3314 * finds swap_migration_entry, the new page will not be mapped
3315 * and end_migration() will find it(mapcount==0).
3318 * When the old page was mapped but migraion fails, the kernel
3319 * remaps it. A charge for it is kept by MIGRATION flag even
3320 * if mapcount goes down to 0. We can do remap successfully
3321 * without charging it again.
3324 * The "old" page is under lock_page() until the end of
3325 * migration, so, the old page itself will not be swapped-out.
3326 * If the new page is swapped out before end_migraton, our
3327 * hook to usual swap-out path will catch the event.
3330 SetPageCgroupMigration(pc);
3332 unlock_page_cgroup(pc);
3334 * If the page is not charged at this point,
3341 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3342 css_put(&memcg->css);/* drop extra refcnt */
3343 if (ret || *ptr == NULL) {
3344 if (PageAnon(page)) {
3345 lock_page_cgroup(pc);
3346 ClearPageCgroupMigration(pc);
3347 unlock_page_cgroup(pc);
3349 * The old page may be fully unmapped while we kept it.
3351 mem_cgroup_uncharge_page(page);
3356 * We charge new page before it's used/mapped. So, even if unlock_page()
3357 * is called before end_migration, we can catch all events on this new
3358 * page. In the case new page is migrated but not remapped, new page's
3359 * mapcount will be finally 0 and we call uncharge in end_migration().
3361 pc = lookup_page_cgroup(newpage);
3363 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3364 else if (page_is_file_cache(page))
3365 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3367 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3368 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3372 /* remove redundant charge if migration failed*/
3373 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3374 struct page *oldpage, struct page *newpage, bool migration_ok)
3376 struct page *used, *unused;
3377 struct page_cgroup *pc;
3381 /* blocks rmdir() */
3382 cgroup_exclude_rmdir(&memcg->css);
3383 if (!migration_ok) {
3391 * We disallowed uncharge of pages under migration because mapcount
3392 * of the page goes down to zero, temporarly.
3393 * Clear the flag and check the page should be charged.
3395 pc = lookup_page_cgroup(oldpage);
3396 lock_page_cgroup(pc);
3397 ClearPageCgroupMigration(pc);
3398 unlock_page_cgroup(pc);
3400 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3403 * If a page is a file cache, radix-tree replacement is very atomic
3404 * and we can skip this check. When it was an Anon page, its mapcount
3405 * goes down to 0. But because we added MIGRATION flage, it's not
3406 * uncharged yet. There are several case but page->mapcount check
3407 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3408 * check. (see prepare_charge() also)
3411 mem_cgroup_uncharge_page(used);
3413 * At migration, we may charge account against cgroup which has no
3415 * So, rmdir()->pre_destroy() can be called while we do this charge.
3416 * In that case, we need to call pre_destroy() again. check it here.
3418 cgroup_release_and_wakeup_rmdir(&memcg->css);
3422 * At replace page cache, newpage is not under any memcg but it's on
3423 * LRU. So, this function doesn't touch res_counter but handles LRU
3424 * in correct way. Both pages are locked so we cannot race with uncharge.
3426 void mem_cgroup_replace_page_cache(struct page *oldpage,
3427 struct page *newpage)
3429 struct mem_cgroup *memcg;
3430 struct page_cgroup *pc;
3432 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3433 unsigned long flags;
3435 if (mem_cgroup_disabled())
3438 pc = lookup_page_cgroup(oldpage);
3439 /* fix accounting on old pages */
3440 lock_page_cgroup(pc);
3441 memcg = pc->mem_cgroup;
3442 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1);
3443 ClearPageCgroupUsed(pc);
3444 unlock_page_cgroup(pc);
3446 if (PageSwapBacked(oldpage))
3447 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3449 zone = page_zone(newpage);
3450 pc = lookup_page_cgroup(newpage);
3452 * Even if newpage->mapping was NULL before starting replacement,
3453 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3454 * LRU while we overwrite pc->mem_cgroup.
3456 spin_lock_irqsave(&zone->lru_lock, flags);
3457 if (PageLRU(newpage))
3458 del_page_from_lru_list(zone, newpage, page_lru(newpage));
3459 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type);
3460 if (PageLRU(newpage))
3461 add_page_to_lru_list(zone, newpage, page_lru(newpage));
3462 spin_unlock_irqrestore(&zone->lru_lock, flags);
3465 #ifdef CONFIG_DEBUG_VM
3466 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3468 struct page_cgroup *pc;
3470 pc = lookup_page_cgroup(page);
3471 if (likely(pc) && PageCgroupUsed(pc))
3476 bool mem_cgroup_bad_page_check(struct page *page)
3478 if (mem_cgroup_disabled())
3481 return lookup_page_cgroup_used(page) != NULL;
3484 void mem_cgroup_print_bad_page(struct page *page)
3486 struct page_cgroup *pc;
3488 pc = lookup_page_cgroup_used(page);
3493 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3494 pc, pc->flags, pc->mem_cgroup);
3496 path = kmalloc(PATH_MAX, GFP_KERNEL);
3499 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3504 printk(KERN_CONT "(%s)\n",
3505 (ret < 0) ? "cannot get the path" : path);
3511 static DEFINE_MUTEX(set_limit_mutex);
3513 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3514 unsigned long long val)
3517 u64 memswlimit, memlimit;
3519 int children = mem_cgroup_count_children(memcg);
3520 u64 curusage, oldusage;
3524 * For keeping hierarchical_reclaim simple, how long we should retry
3525 * is depends on callers. We set our retry-count to be function
3526 * of # of children which we should visit in this loop.
3528 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3530 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3533 while (retry_count) {
3534 if (signal_pending(current)) {
3539 * Rather than hide all in some function, I do this in
3540 * open coded manner. You see what this really does.
3541 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3543 mutex_lock(&set_limit_mutex);
3544 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3545 if (memswlimit < val) {
3547 mutex_unlock(&set_limit_mutex);
3551 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3555 ret = res_counter_set_limit(&memcg->res, val);
3557 if (memswlimit == val)
3558 memcg->memsw_is_minimum = true;
3560 memcg->memsw_is_minimum = false;
3562 mutex_unlock(&set_limit_mutex);
3567 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3568 MEM_CGROUP_RECLAIM_SHRINK);
3569 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3570 /* Usage is reduced ? */
3571 if (curusage >= oldusage)
3574 oldusage = curusage;
3576 if (!ret && enlarge)
3577 memcg_oom_recover(memcg);
3582 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3583 unsigned long long val)
3586 u64 memlimit, memswlimit, oldusage, curusage;
3587 int children = mem_cgroup_count_children(memcg);
3591 /* see mem_cgroup_resize_res_limit */
3592 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3593 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3594 while (retry_count) {
3595 if (signal_pending(current)) {
3600 * Rather than hide all in some function, I do this in
3601 * open coded manner. You see what this really does.
3602 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3604 mutex_lock(&set_limit_mutex);
3605 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3606 if (memlimit > val) {
3608 mutex_unlock(&set_limit_mutex);
3611 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3612 if (memswlimit < val)
3614 ret = res_counter_set_limit(&memcg->memsw, val);
3616 if (memlimit == val)
3617 memcg->memsw_is_minimum = true;
3619 memcg->memsw_is_minimum = false;
3621 mutex_unlock(&set_limit_mutex);
3626 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3627 MEM_CGROUP_RECLAIM_NOSWAP |
3628 MEM_CGROUP_RECLAIM_SHRINK);
3629 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3630 /* Usage is reduced ? */
3631 if (curusage >= oldusage)
3634 oldusage = curusage;
3636 if (!ret && enlarge)
3637 memcg_oom_recover(memcg);
3641 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3643 unsigned long *total_scanned)
3645 unsigned long nr_reclaimed = 0;
3646 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3647 unsigned long reclaimed;
3649 struct mem_cgroup_tree_per_zone *mctz;
3650 unsigned long long excess;
3651 unsigned long nr_scanned;
3656 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3658 * This loop can run a while, specially if mem_cgroup's continuously
3659 * keep exceeding their soft limit and putting the system under
3666 mz = mem_cgroup_largest_soft_limit_node(mctz);
3671 reclaimed = mem_cgroup_soft_reclaim(mz->mem, zone,
3672 gfp_mask, &nr_scanned);
3673 nr_reclaimed += reclaimed;
3674 *total_scanned += nr_scanned;
3675 spin_lock(&mctz->lock);
3678 * If we failed to reclaim anything from this memory cgroup
3679 * it is time to move on to the next cgroup
3685 * Loop until we find yet another one.
3687 * By the time we get the soft_limit lock
3688 * again, someone might have aded the
3689 * group back on the RB tree. Iterate to
3690 * make sure we get a different mem.
3691 * mem_cgroup_largest_soft_limit_node returns
3692 * NULL if no other cgroup is present on
3696 __mem_cgroup_largest_soft_limit_node(mctz);
3698 css_put(&next_mz->mem->css);
3699 else /* next_mz == NULL or other memcg */
3703 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3704 excess = res_counter_soft_limit_excess(&mz->mem->res);
3706 * One school of thought says that we should not add
3707 * back the node to the tree if reclaim returns 0.
3708 * But our reclaim could return 0, simply because due
3709 * to priority we are exposing a smaller subset of
3710 * memory to reclaim from. Consider this as a longer
3713 /* If excess == 0, no tree ops */
3714 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3715 spin_unlock(&mctz->lock);
3716 css_put(&mz->mem->css);
3719 * Could not reclaim anything and there are no more
3720 * mem cgroups to try or we seem to be looping without
3721 * reclaiming anything.
3723 if (!nr_reclaimed &&
3725 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3727 } while (!nr_reclaimed);
3729 css_put(&next_mz->mem->css);
3730 return nr_reclaimed;
3734 * This routine traverse page_cgroup in given list and drop them all.
3735 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3737 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3738 int node, int zid, enum lru_list lru)
3741 struct mem_cgroup_per_zone *mz;
3742 struct page_cgroup *pc, *busy;
3743 unsigned long flags, loop;
3744 struct list_head *list;
3747 zone = &NODE_DATA(node)->node_zones[zid];
3748 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3749 list = &mz->lists[lru];
3751 loop = MEM_CGROUP_ZSTAT(mz, lru);
3752 /* give some margin against EBUSY etc...*/
3759 spin_lock_irqsave(&zone->lru_lock, flags);
3760 if (list_empty(list)) {
3761 spin_unlock_irqrestore(&zone->lru_lock, flags);
3764 pc = list_entry(list->prev, struct page_cgroup, lru);
3766 list_move(&pc->lru, list);
3768 spin_unlock_irqrestore(&zone->lru_lock, flags);
3771 spin_unlock_irqrestore(&zone->lru_lock, flags);
3773 page = lookup_cgroup_page(pc);
3775 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3779 if (ret == -EBUSY || ret == -EINVAL) {
3780 /* found lock contention or "pc" is obsolete. */
3787 if (!ret && !list_empty(list))
3793 * make mem_cgroup's charge to be 0 if there is no task.
3794 * This enables deleting this mem_cgroup.
3796 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3799 int node, zid, shrink;
3800 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3801 struct cgroup *cgrp = memcg->css.cgroup;
3803 css_get(&memcg->css);
3806 /* should free all ? */
3812 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3815 if (signal_pending(current))
3817 /* This is for making all *used* pages to be on LRU. */
3818 lru_add_drain_all();
3819 drain_all_stock_sync(memcg);
3821 mem_cgroup_start_move(memcg);
3822 for_each_node_state(node, N_HIGH_MEMORY) {
3823 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3826 ret = mem_cgroup_force_empty_list(memcg,
3835 mem_cgroup_end_move(memcg);
3836 memcg_oom_recover(memcg);
3837 /* it seems parent cgroup doesn't have enough mem */
3841 /* "ret" should also be checked to ensure all lists are empty. */
3842 } while (memcg->res.usage > 0 || ret);
3844 css_put(&memcg->css);
3848 /* returns EBUSY if there is a task or if we come here twice. */
3849 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3853 /* we call try-to-free pages for make this cgroup empty */
3854 lru_add_drain_all();
3855 /* try to free all pages in this cgroup */
3857 while (nr_retries && memcg->res.usage > 0) {
3860 if (signal_pending(current)) {
3864 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3868 /* maybe some writeback is necessary */
3869 congestion_wait(BLK_RW_ASYNC, HZ/10);
3874 /* try move_account...there may be some *locked* pages. */
3878 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3880 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3884 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3886 return mem_cgroup_from_cont(cont)->use_hierarchy;
3889 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3893 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3894 struct cgroup *parent = cont->parent;
3895 struct mem_cgroup *parent_memcg = NULL;
3898 parent_memcg = mem_cgroup_from_cont(parent);
3902 * If parent's use_hierarchy is set, we can't make any modifications
3903 * in the child subtrees. If it is unset, then the change can
3904 * occur, provided the current cgroup has no children.
3906 * For the root cgroup, parent_mem is NULL, we allow value to be
3907 * set if there are no children.
3909 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3910 (val == 1 || val == 0)) {
3911 if (list_empty(&cont->children))
3912 memcg->use_hierarchy = val;
3923 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3924 enum mem_cgroup_stat_index idx)
3926 struct mem_cgroup *iter;
3929 /* Per-cpu values can be negative, use a signed accumulator */
3930 for_each_mem_cgroup_tree(iter, memcg)
3931 val += mem_cgroup_read_stat(iter, idx);
3933 if (val < 0) /* race ? */
3938 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3942 if (!mem_cgroup_is_root(memcg)) {
3944 return res_counter_read_u64(&memcg->res, RES_USAGE);
3946 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3949 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3950 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3953 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3955 return val << PAGE_SHIFT;
3958 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3960 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3964 type = MEMFILE_TYPE(cft->private);
3965 name = MEMFILE_ATTR(cft->private);
3968 if (name == RES_USAGE)
3969 val = mem_cgroup_usage(memcg, false);
3971 val = res_counter_read_u64(&memcg->res, name);
3974 if (name == RES_USAGE)
3975 val = mem_cgroup_usage(memcg, true);
3977 val = res_counter_read_u64(&memcg->memsw, name);
3986 * The user of this function is...
3989 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3992 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3994 unsigned long long val;
3997 type = MEMFILE_TYPE(cft->private);
3998 name = MEMFILE_ATTR(cft->private);
4001 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4005 /* This function does all necessary parse...reuse it */
4006 ret = res_counter_memparse_write_strategy(buffer, &val);
4010 ret = mem_cgroup_resize_limit(memcg, val);
4012 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4014 case RES_SOFT_LIMIT:
4015 ret = res_counter_memparse_write_strategy(buffer, &val);
4019 * For memsw, soft limits are hard to implement in terms
4020 * of semantics, for now, we support soft limits for
4021 * control without swap
4024 ret = res_counter_set_soft_limit(&memcg->res, val);
4029 ret = -EINVAL; /* should be BUG() ? */
4035 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4036 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4038 struct cgroup *cgroup;
4039 unsigned long long min_limit, min_memsw_limit, tmp;
4041 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4042 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4043 cgroup = memcg->css.cgroup;
4044 if (!memcg->use_hierarchy)
4047 while (cgroup->parent) {
4048 cgroup = cgroup->parent;
4049 memcg = mem_cgroup_from_cont(cgroup);
4050 if (!memcg->use_hierarchy)
4052 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4053 min_limit = min(min_limit, tmp);
4054 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4055 min_memsw_limit = min(min_memsw_limit, tmp);
4058 *mem_limit = min_limit;
4059 *memsw_limit = min_memsw_limit;
4063 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4065 struct mem_cgroup *memcg;
4068 memcg = mem_cgroup_from_cont(cont);
4069 type = MEMFILE_TYPE(event);
4070 name = MEMFILE_ATTR(event);
4074 res_counter_reset_max(&memcg->res);
4076 res_counter_reset_max(&memcg->memsw);
4080 res_counter_reset_failcnt(&memcg->res);
4082 res_counter_reset_failcnt(&memcg->memsw);
4089 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4092 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4096 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4097 struct cftype *cft, u64 val)
4099 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4101 if (val >= (1 << NR_MOVE_TYPE))
4104 * We check this value several times in both in can_attach() and
4105 * attach(), so we need cgroup lock to prevent this value from being
4109 memcg->move_charge_at_immigrate = val;
4115 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4116 struct cftype *cft, u64 val)
4123 /* For read statistics */
4141 struct mcs_total_stat {
4142 s64 stat[NR_MCS_STAT];
4148 } memcg_stat_strings[NR_MCS_STAT] = {
4149 {"cache", "total_cache"},
4150 {"rss", "total_rss"},
4151 {"mapped_file", "total_mapped_file"},
4152 {"pgpgin", "total_pgpgin"},
4153 {"pgpgout", "total_pgpgout"},
4154 {"swap", "total_swap"},
4155 {"pgfault", "total_pgfault"},
4156 {"pgmajfault", "total_pgmajfault"},
4157 {"inactive_anon", "total_inactive_anon"},
4158 {"active_anon", "total_active_anon"},
4159 {"inactive_file", "total_inactive_file"},
4160 {"active_file", "total_active_file"},
4161 {"unevictable", "total_unevictable"}
4166 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4171 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4172 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4173 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4174 s->stat[MCS_RSS] += val * PAGE_SIZE;
4175 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4176 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4177 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4178 s->stat[MCS_PGPGIN] += val;
4179 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4180 s->stat[MCS_PGPGOUT] += val;
4181 if (do_swap_account) {
4182 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4183 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4185 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4186 s->stat[MCS_PGFAULT] += val;
4187 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4188 s->stat[MCS_PGMAJFAULT] += val;
4191 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4192 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4193 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4194 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4195 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4196 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4197 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4198 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4199 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4200 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4204 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4206 struct mem_cgroup *iter;
4208 for_each_mem_cgroup_tree(iter, memcg)
4209 mem_cgroup_get_local_stat(iter, s);
4213 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4216 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4217 unsigned long node_nr;
4218 struct cgroup *cont = m->private;
4219 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4221 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4222 seq_printf(m, "total=%lu", total_nr);
4223 for_each_node_state(nid, N_HIGH_MEMORY) {
4224 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4225 seq_printf(m, " N%d=%lu", nid, node_nr);
4229 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4230 seq_printf(m, "file=%lu", file_nr);
4231 for_each_node_state(nid, N_HIGH_MEMORY) {
4232 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4234 seq_printf(m, " N%d=%lu", nid, node_nr);
4238 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4239 seq_printf(m, "anon=%lu", anon_nr);
4240 for_each_node_state(nid, N_HIGH_MEMORY) {
4241 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4243 seq_printf(m, " N%d=%lu", nid, node_nr);
4247 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4248 seq_printf(m, "unevictable=%lu", unevictable_nr);
4249 for_each_node_state(nid, N_HIGH_MEMORY) {
4250 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4251 BIT(LRU_UNEVICTABLE));
4252 seq_printf(m, " N%d=%lu", nid, node_nr);
4257 #endif /* CONFIG_NUMA */
4259 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4260 struct cgroup_map_cb *cb)
4262 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4263 struct mcs_total_stat mystat;
4266 memset(&mystat, 0, sizeof(mystat));
4267 mem_cgroup_get_local_stat(mem_cont, &mystat);
4270 for (i = 0; i < NR_MCS_STAT; i++) {
4271 if (i == MCS_SWAP && !do_swap_account)
4273 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4276 /* Hierarchical information */
4278 unsigned long long limit, memsw_limit;
4279 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4280 cb->fill(cb, "hierarchical_memory_limit", limit);
4281 if (do_swap_account)
4282 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4285 memset(&mystat, 0, sizeof(mystat));
4286 mem_cgroup_get_total_stat(mem_cont, &mystat);
4287 for (i = 0; i < NR_MCS_STAT; i++) {
4288 if (i == MCS_SWAP && !do_swap_account)
4290 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4293 #ifdef CONFIG_DEBUG_VM
4296 struct mem_cgroup_per_zone *mz;
4297 unsigned long recent_rotated[2] = {0, 0};
4298 unsigned long recent_scanned[2] = {0, 0};
4300 for_each_online_node(nid)
4301 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4302 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4304 recent_rotated[0] +=
4305 mz->reclaim_stat.recent_rotated[0];
4306 recent_rotated[1] +=
4307 mz->reclaim_stat.recent_rotated[1];
4308 recent_scanned[0] +=
4309 mz->reclaim_stat.recent_scanned[0];
4310 recent_scanned[1] +=
4311 mz->reclaim_stat.recent_scanned[1];
4313 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4314 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4315 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4316 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4323 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4325 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4327 return mem_cgroup_swappiness(memcg);
4330 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4333 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4334 struct mem_cgroup *parent;
4339 if (cgrp->parent == NULL)
4342 parent = mem_cgroup_from_cont(cgrp->parent);
4346 /* If under hierarchy, only empty-root can set this value */
4347 if ((parent->use_hierarchy) ||
4348 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4353 memcg->swappiness = val;
4360 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4362 struct mem_cgroup_threshold_ary *t;
4368 t = rcu_dereference(memcg->thresholds.primary);
4370 t = rcu_dereference(memcg->memsw_thresholds.primary);
4375 usage = mem_cgroup_usage(memcg, swap);
4378 * current_threshold points to threshold just below usage.
4379 * If it's not true, a threshold was crossed after last
4380 * call of __mem_cgroup_threshold().
4382 i = t->current_threshold;
4385 * Iterate backward over array of thresholds starting from
4386 * current_threshold and check if a threshold is crossed.
4387 * If none of thresholds below usage is crossed, we read
4388 * only one element of the array here.
4390 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4391 eventfd_signal(t->entries[i].eventfd, 1);
4393 /* i = current_threshold + 1 */
4397 * Iterate forward over array of thresholds starting from
4398 * current_threshold+1 and check if a threshold is crossed.
4399 * If none of thresholds above usage is crossed, we read
4400 * only one element of the array here.
4402 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4403 eventfd_signal(t->entries[i].eventfd, 1);
4405 /* Update current_threshold */
4406 t->current_threshold = i - 1;
4411 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4414 __mem_cgroup_threshold(memcg, false);
4415 if (do_swap_account)
4416 __mem_cgroup_threshold(memcg, true);
4418 memcg = parent_mem_cgroup(memcg);
4422 static int compare_thresholds(const void *a, const void *b)
4424 const struct mem_cgroup_threshold *_a = a;
4425 const struct mem_cgroup_threshold *_b = b;
4427 return _a->threshold - _b->threshold;
4430 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4432 struct mem_cgroup_eventfd_list *ev;
4434 list_for_each_entry(ev, &memcg->oom_notify, list)
4435 eventfd_signal(ev->eventfd, 1);
4439 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4441 struct mem_cgroup *iter;
4443 for_each_mem_cgroup_tree(iter, memcg)
4444 mem_cgroup_oom_notify_cb(iter);
4447 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4448 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4450 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4451 struct mem_cgroup_thresholds *thresholds;
4452 struct mem_cgroup_threshold_ary *new;
4453 int type = MEMFILE_TYPE(cft->private);
4454 u64 threshold, usage;
4457 ret = res_counter_memparse_write_strategy(args, &threshold);
4461 mutex_lock(&memcg->thresholds_lock);
4464 thresholds = &memcg->thresholds;
4465 else if (type == _MEMSWAP)
4466 thresholds = &memcg->memsw_thresholds;
4470 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4472 /* Check if a threshold crossed before adding a new one */
4473 if (thresholds->primary)
4474 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4476 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4478 /* Allocate memory for new array of thresholds */
4479 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4487 /* Copy thresholds (if any) to new array */
4488 if (thresholds->primary) {
4489 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4490 sizeof(struct mem_cgroup_threshold));
4493 /* Add new threshold */
4494 new->entries[size - 1].eventfd = eventfd;
4495 new->entries[size - 1].threshold = threshold;
4497 /* Sort thresholds. Registering of new threshold isn't time-critical */
4498 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4499 compare_thresholds, NULL);
4501 /* Find current threshold */
4502 new->current_threshold = -1;
4503 for (i = 0; i < size; i++) {
4504 if (new->entries[i].threshold < usage) {
4506 * new->current_threshold will not be used until
4507 * rcu_assign_pointer(), so it's safe to increment
4510 ++new->current_threshold;
4514 /* Free old spare buffer and save old primary buffer as spare */
4515 kfree(thresholds->spare);
4516 thresholds->spare = thresholds->primary;
4518 rcu_assign_pointer(thresholds->primary, new);
4520 /* To be sure that nobody uses thresholds */
4524 mutex_unlock(&memcg->thresholds_lock);
4529 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4530 struct cftype *cft, struct eventfd_ctx *eventfd)
4532 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4533 struct mem_cgroup_thresholds *thresholds;
4534 struct mem_cgroup_threshold_ary *new;
4535 int type = MEMFILE_TYPE(cft->private);
4539 mutex_lock(&memcg->thresholds_lock);
4541 thresholds = &memcg->thresholds;
4542 else if (type == _MEMSWAP)
4543 thresholds = &memcg->memsw_thresholds;
4548 * Something went wrong if we trying to unregister a threshold
4549 * if we don't have thresholds
4551 BUG_ON(!thresholds);
4553 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4555 /* Check if a threshold crossed before removing */
4556 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4558 /* Calculate new number of threshold */
4560 for (i = 0; i < thresholds->primary->size; i++) {
4561 if (thresholds->primary->entries[i].eventfd != eventfd)
4565 new = thresholds->spare;
4567 /* Set thresholds array to NULL if we don't have thresholds */
4576 /* Copy thresholds and find current threshold */
4577 new->current_threshold = -1;
4578 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4579 if (thresholds->primary->entries[i].eventfd == eventfd)
4582 new->entries[j] = thresholds->primary->entries[i];
4583 if (new->entries[j].threshold < usage) {
4585 * new->current_threshold will not be used
4586 * until rcu_assign_pointer(), so it's safe to increment
4589 ++new->current_threshold;
4595 /* Swap primary and spare array */
4596 thresholds->spare = thresholds->primary;
4597 rcu_assign_pointer(thresholds->primary, new);
4599 /* To be sure that nobody uses thresholds */
4602 mutex_unlock(&memcg->thresholds_lock);
4605 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4606 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4608 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4609 struct mem_cgroup_eventfd_list *event;
4610 int type = MEMFILE_TYPE(cft->private);
4612 BUG_ON(type != _OOM_TYPE);
4613 event = kmalloc(sizeof(*event), GFP_KERNEL);
4617 spin_lock(&memcg_oom_lock);
4619 event->eventfd = eventfd;
4620 list_add(&event->list, &memcg->oom_notify);
4622 /* already in OOM ? */
4623 if (atomic_read(&memcg->under_oom))
4624 eventfd_signal(eventfd, 1);
4625 spin_unlock(&memcg_oom_lock);
4630 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4631 struct cftype *cft, struct eventfd_ctx *eventfd)
4633 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4634 struct mem_cgroup_eventfd_list *ev, *tmp;
4635 int type = MEMFILE_TYPE(cft->private);
4637 BUG_ON(type != _OOM_TYPE);
4639 spin_lock(&memcg_oom_lock);
4641 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4642 if (ev->eventfd == eventfd) {
4643 list_del(&ev->list);
4648 spin_unlock(&memcg_oom_lock);
4651 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4652 struct cftype *cft, struct cgroup_map_cb *cb)
4654 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4656 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4658 if (atomic_read(&memcg->under_oom))
4659 cb->fill(cb, "under_oom", 1);
4661 cb->fill(cb, "under_oom", 0);
4665 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4666 struct cftype *cft, u64 val)
4668 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4669 struct mem_cgroup *parent;
4671 /* cannot set to root cgroup and only 0 and 1 are allowed */
4672 if (!cgrp->parent || !((val == 0) || (val == 1)))
4675 parent = mem_cgroup_from_cont(cgrp->parent);
4678 /* oom-kill-disable is a flag for subhierarchy. */
4679 if ((parent->use_hierarchy) ||
4680 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4684 memcg->oom_kill_disable = val;
4686 memcg_oom_recover(memcg);
4692 static const struct file_operations mem_control_numa_stat_file_operations = {
4694 .llseek = seq_lseek,
4695 .release = single_release,
4698 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4700 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4702 file->f_op = &mem_control_numa_stat_file_operations;
4703 return single_open(file, mem_control_numa_stat_show, cont);
4705 #endif /* CONFIG_NUMA */
4707 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4708 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4711 * Part of this would be better living in a separate allocation
4712 * function, leaving us with just the cgroup tree population work.
4713 * We, however, depend on state such as network's proto_list that
4714 * is only initialized after cgroup creation. I found the less
4715 * cumbersome way to deal with it to defer it all to populate time
4717 return mem_cgroup_sockets_init(cont, ss);
4720 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4721 struct cgroup *cont)
4723 mem_cgroup_sockets_destroy(cont, ss);
4726 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4731 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4732 struct cgroup *cont)
4737 static struct cftype mem_cgroup_files[] = {
4739 .name = "usage_in_bytes",
4740 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4741 .read_u64 = mem_cgroup_read,
4742 .register_event = mem_cgroup_usage_register_event,
4743 .unregister_event = mem_cgroup_usage_unregister_event,
4746 .name = "max_usage_in_bytes",
4747 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4748 .trigger = mem_cgroup_reset,
4749 .read_u64 = mem_cgroup_read,
4752 .name = "limit_in_bytes",
4753 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4754 .write_string = mem_cgroup_write,
4755 .read_u64 = mem_cgroup_read,
4758 .name = "soft_limit_in_bytes",
4759 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4760 .write_string = mem_cgroup_write,
4761 .read_u64 = mem_cgroup_read,
4765 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4766 .trigger = mem_cgroup_reset,
4767 .read_u64 = mem_cgroup_read,
4771 .read_map = mem_control_stat_show,
4774 .name = "force_empty",
4775 .trigger = mem_cgroup_force_empty_write,
4778 .name = "use_hierarchy",
4779 .write_u64 = mem_cgroup_hierarchy_write,
4780 .read_u64 = mem_cgroup_hierarchy_read,
4783 .name = "swappiness",
4784 .read_u64 = mem_cgroup_swappiness_read,
4785 .write_u64 = mem_cgroup_swappiness_write,
4788 .name = "move_charge_at_immigrate",
4789 .read_u64 = mem_cgroup_move_charge_read,
4790 .write_u64 = mem_cgroup_move_charge_write,
4793 .name = "oom_control",
4794 .read_map = mem_cgroup_oom_control_read,
4795 .write_u64 = mem_cgroup_oom_control_write,
4796 .register_event = mem_cgroup_oom_register_event,
4797 .unregister_event = mem_cgroup_oom_unregister_event,
4798 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4802 .name = "numa_stat",
4803 .open = mem_control_numa_stat_open,
4809 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4810 static struct cftype memsw_cgroup_files[] = {
4812 .name = "memsw.usage_in_bytes",
4813 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4814 .read_u64 = mem_cgroup_read,
4815 .register_event = mem_cgroup_usage_register_event,
4816 .unregister_event = mem_cgroup_usage_unregister_event,
4819 .name = "memsw.max_usage_in_bytes",
4820 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4821 .trigger = mem_cgroup_reset,
4822 .read_u64 = mem_cgroup_read,
4825 .name = "memsw.limit_in_bytes",
4826 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4827 .write_string = mem_cgroup_write,
4828 .read_u64 = mem_cgroup_read,
4831 .name = "memsw.failcnt",
4832 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4833 .trigger = mem_cgroup_reset,
4834 .read_u64 = mem_cgroup_read,
4838 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4840 if (!do_swap_account)
4842 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4843 ARRAY_SIZE(memsw_cgroup_files));
4846 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4852 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4854 struct mem_cgroup_per_node *pn;
4855 struct mem_cgroup_per_zone *mz;
4857 int zone, tmp = node;
4859 * This routine is called against possible nodes.
4860 * But it's BUG to call kmalloc() against offline node.
4862 * TODO: this routine can waste much memory for nodes which will
4863 * never be onlined. It's better to use memory hotplug callback
4866 if (!node_state(node, N_NORMAL_MEMORY))
4868 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4872 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4873 mz = &pn->zoneinfo[zone];
4875 INIT_LIST_HEAD(&mz->lists[l]);
4876 mz->usage_in_excess = 0;
4877 mz->on_tree = false;
4880 memcg->info.nodeinfo[node] = pn;
4884 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4886 kfree(memcg->info.nodeinfo[node]);
4889 static struct mem_cgroup *mem_cgroup_alloc(void)
4891 struct mem_cgroup *mem;
4892 int size = sizeof(struct mem_cgroup);
4894 /* Can be very big if MAX_NUMNODES is very big */
4895 if (size < PAGE_SIZE)
4896 mem = kzalloc(size, GFP_KERNEL);
4898 mem = vzalloc(size);
4903 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4906 spin_lock_init(&mem->pcp_counter_lock);
4910 if (size < PAGE_SIZE)
4918 * At destroying mem_cgroup, references from swap_cgroup can remain.
4919 * (scanning all at force_empty is too costly...)
4921 * Instead of clearing all references at force_empty, we remember
4922 * the number of reference from swap_cgroup and free mem_cgroup when
4923 * it goes down to 0.
4925 * Removal of cgroup itself succeeds regardless of refs from swap.
4928 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4932 mem_cgroup_remove_from_trees(memcg);
4933 free_css_id(&mem_cgroup_subsys, &memcg->css);
4935 for_each_node_state(node, N_POSSIBLE)
4936 free_mem_cgroup_per_zone_info(memcg, node);
4938 free_percpu(memcg->stat);
4939 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4945 static void mem_cgroup_get(struct mem_cgroup *memcg)
4947 atomic_inc(&memcg->refcnt);
4950 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4952 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4953 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4954 __mem_cgroup_free(memcg);
4956 mem_cgroup_put(parent);
4960 static void mem_cgroup_put(struct mem_cgroup *memcg)
4962 __mem_cgroup_put(memcg, 1);
4966 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4968 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4970 if (!memcg->res.parent)
4972 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4974 EXPORT_SYMBOL(parent_mem_cgroup);
4976 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4977 static void __init enable_swap_cgroup(void)
4979 if (!mem_cgroup_disabled() && really_do_swap_account)
4980 do_swap_account = 1;
4983 static void __init enable_swap_cgroup(void)
4988 static int mem_cgroup_soft_limit_tree_init(void)
4990 struct mem_cgroup_tree_per_node *rtpn;
4991 struct mem_cgroup_tree_per_zone *rtpz;
4992 int tmp, node, zone;
4994 for_each_node_state(node, N_POSSIBLE) {
4996 if (!node_state(node, N_NORMAL_MEMORY))
4998 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5002 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5004 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5005 rtpz = &rtpn->rb_tree_per_zone[zone];
5006 rtpz->rb_root = RB_ROOT;
5007 spin_lock_init(&rtpz->lock);
5013 static struct cgroup_subsys_state * __ref
5014 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5016 struct mem_cgroup *memcg, *parent;
5017 long error = -ENOMEM;
5020 memcg = mem_cgroup_alloc();
5022 return ERR_PTR(error);
5024 for_each_node_state(node, N_POSSIBLE)
5025 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5029 if (cont->parent == NULL) {
5031 enable_swap_cgroup();
5033 if (mem_cgroup_soft_limit_tree_init())
5035 root_mem_cgroup = memcg;
5036 for_each_possible_cpu(cpu) {
5037 struct memcg_stock_pcp *stock =
5038 &per_cpu(memcg_stock, cpu);
5039 INIT_WORK(&stock->work, drain_local_stock);
5041 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5043 parent = mem_cgroup_from_cont(cont->parent);
5044 memcg->use_hierarchy = parent->use_hierarchy;
5045 memcg->oom_kill_disable = parent->oom_kill_disable;
5048 if (parent && parent->use_hierarchy) {
5049 res_counter_init(&memcg->res, &parent->res);
5050 res_counter_init(&memcg->memsw, &parent->memsw);
5052 * We increment refcnt of the parent to ensure that we can
5053 * safely access it on res_counter_charge/uncharge.
5054 * This refcnt will be decremented when freeing this
5055 * mem_cgroup(see mem_cgroup_put).
5057 mem_cgroup_get(parent);
5059 res_counter_init(&memcg->res, NULL);
5060 res_counter_init(&memcg->memsw, NULL);
5062 memcg->last_scanned_node = MAX_NUMNODES;
5063 INIT_LIST_HEAD(&memcg->oom_notify);
5066 memcg->swappiness = mem_cgroup_swappiness(parent);
5067 atomic_set(&memcg->refcnt, 1);
5068 memcg->move_charge_at_immigrate = 0;
5069 mutex_init(&memcg->thresholds_lock);
5072 __mem_cgroup_free(memcg);
5073 return ERR_PTR(error);
5076 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5077 struct cgroup *cont)
5079 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5081 return mem_cgroup_force_empty(memcg, false);
5084 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5085 struct cgroup *cont)
5087 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5089 kmem_cgroup_destroy(ss, cont);
5091 mem_cgroup_put(memcg);
5094 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5095 struct cgroup *cont)
5099 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5100 ARRAY_SIZE(mem_cgroup_files));
5103 ret = register_memsw_files(cont, ss);
5106 ret = register_kmem_files(cont, ss);
5112 /* Handlers for move charge at task migration. */
5113 #define PRECHARGE_COUNT_AT_ONCE 256
5114 static int mem_cgroup_do_precharge(unsigned long count)
5117 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5118 struct mem_cgroup *memcg = mc.to;
5120 if (mem_cgroup_is_root(memcg)) {
5121 mc.precharge += count;
5122 /* we don't need css_get for root */
5125 /* try to charge at once */
5127 struct res_counter *dummy;
5129 * "memcg" cannot be under rmdir() because we've already checked
5130 * by cgroup_lock_live_cgroup() that it is not removed and we
5131 * are still under the same cgroup_mutex. So we can postpone
5134 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5136 if (do_swap_account && res_counter_charge(&memcg->memsw,
5137 PAGE_SIZE * count, &dummy)) {
5138 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5141 mc.precharge += count;
5145 /* fall back to one by one charge */
5147 if (signal_pending(current)) {
5151 if (!batch_count--) {
5152 batch_count = PRECHARGE_COUNT_AT_ONCE;
5155 ret = __mem_cgroup_try_charge(NULL,
5156 GFP_KERNEL, 1, &memcg, false);
5158 /* mem_cgroup_clear_mc() will do uncharge later */
5166 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5167 * @vma: the vma the pte to be checked belongs
5168 * @addr: the address corresponding to the pte to be checked
5169 * @ptent: the pte to be checked
5170 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5173 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5174 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5175 * move charge. if @target is not NULL, the page is stored in target->page
5176 * with extra refcnt got(Callers should handle it).
5177 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5178 * target for charge migration. if @target is not NULL, the entry is stored
5181 * Called with pte lock held.
5188 enum mc_target_type {
5189 MC_TARGET_NONE, /* not used */
5194 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5195 unsigned long addr, pte_t ptent)
5197 struct page *page = vm_normal_page(vma, addr, ptent);
5199 if (!page || !page_mapped(page))
5201 if (PageAnon(page)) {
5202 /* we don't move shared anon */
5203 if (!move_anon() || page_mapcount(page) > 2)
5205 } else if (!move_file())
5206 /* we ignore mapcount for file pages */
5208 if (!get_page_unless_zero(page))
5214 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5215 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5218 struct page *page = NULL;
5219 swp_entry_t ent = pte_to_swp_entry(ptent);
5221 if (!move_anon() || non_swap_entry(ent))
5223 usage_count = mem_cgroup_count_swap_user(ent, &page);
5224 if (usage_count > 1) { /* we don't move shared anon */
5229 if (do_swap_account)
5230 entry->val = ent.val;
5235 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5236 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5238 struct page *page = NULL;
5239 struct inode *inode;
5240 struct address_space *mapping;
5243 if (!vma->vm_file) /* anonymous vma */
5248 inode = vma->vm_file->f_path.dentry->d_inode;
5249 mapping = vma->vm_file->f_mapping;
5250 if (pte_none(ptent))
5251 pgoff = linear_page_index(vma, addr);
5252 else /* pte_file(ptent) is true */
5253 pgoff = pte_to_pgoff(ptent);
5255 /* page is moved even if it's not RSS of this task(page-faulted). */
5256 page = find_get_page(mapping, pgoff);
5259 /* shmem/tmpfs may report page out on swap: account for that too. */
5260 if (radix_tree_exceptional_entry(page)) {
5261 swp_entry_t swap = radix_to_swp_entry(page);
5262 if (do_swap_account)
5264 page = find_get_page(&swapper_space, swap.val);
5270 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5271 unsigned long addr, pte_t ptent, union mc_target *target)
5273 struct page *page = NULL;
5274 struct page_cgroup *pc;
5276 swp_entry_t ent = { .val = 0 };
5278 if (pte_present(ptent))
5279 page = mc_handle_present_pte(vma, addr, ptent);
5280 else if (is_swap_pte(ptent))
5281 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5282 else if (pte_none(ptent) || pte_file(ptent))
5283 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5285 if (!page && !ent.val)
5288 pc = lookup_page_cgroup(page);
5290 * Do only loose check w/o page_cgroup lock.
5291 * mem_cgroup_move_account() checks the pc is valid or not under
5294 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5295 ret = MC_TARGET_PAGE;
5297 target->page = page;
5299 if (!ret || !target)
5302 /* There is a swap entry and a page doesn't exist or isn't charged */
5303 if (ent.val && !ret &&
5304 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5305 ret = MC_TARGET_SWAP;
5312 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5313 unsigned long addr, unsigned long end,
5314 struct mm_walk *walk)
5316 struct vm_area_struct *vma = walk->private;
5320 split_huge_page_pmd(walk->mm, pmd);
5322 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5323 for (; addr != end; pte++, addr += PAGE_SIZE)
5324 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5325 mc.precharge++; /* increment precharge temporarily */
5326 pte_unmap_unlock(pte - 1, ptl);
5332 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5334 unsigned long precharge;
5335 struct vm_area_struct *vma;
5337 down_read(&mm->mmap_sem);
5338 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5339 struct mm_walk mem_cgroup_count_precharge_walk = {
5340 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5344 if (is_vm_hugetlb_page(vma))
5346 walk_page_range(vma->vm_start, vma->vm_end,
5347 &mem_cgroup_count_precharge_walk);
5349 up_read(&mm->mmap_sem);
5351 precharge = mc.precharge;
5357 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5359 unsigned long precharge = mem_cgroup_count_precharge(mm);
5361 VM_BUG_ON(mc.moving_task);
5362 mc.moving_task = current;
5363 return mem_cgroup_do_precharge(precharge);
5366 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5367 static void __mem_cgroup_clear_mc(void)
5369 struct mem_cgroup *from = mc.from;
5370 struct mem_cgroup *to = mc.to;
5372 /* we must uncharge all the leftover precharges from mc.to */
5374 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5378 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5379 * we must uncharge here.
5381 if (mc.moved_charge) {
5382 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5383 mc.moved_charge = 0;
5385 /* we must fixup refcnts and charges */
5386 if (mc.moved_swap) {
5387 /* uncharge swap account from the old cgroup */
5388 if (!mem_cgroup_is_root(mc.from))
5389 res_counter_uncharge(&mc.from->memsw,
5390 PAGE_SIZE * mc.moved_swap);
5391 __mem_cgroup_put(mc.from, mc.moved_swap);
5393 if (!mem_cgroup_is_root(mc.to)) {
5395 * we charged both to->res and to->memsw, so we should
5398 res_counter_uncharge(&mc.to->res,
5399 PAGE_SIZE * mc.moved_swap);
5401 /* we've already done mem_cgroup_get(mc.to) */
5404 memcg_oom_recover(from);
5405 memcg_oom_recover(to);
5406 wake_up_all(&mc.waitq);
5409 static void mem_cgroup_clear_mc(void)
5411 struct mem_cgroup *from = mc.from;
5414 * we must clear moving_task before waking up waiters at the end of
5417 mc.moving_task = NULL;
5418 __mem_cgroup_clear_mc();
5419 spin_lock(&mc.lock);
5422 spin_unlock(&mc.lock);
5423 mem_cgroup_end_move(from);
5426 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5427 struct cgroup *cgroup,
5428 struct cgroup_taskset *tset)
5430 struct task_struct *p = cgroup_taskset_first(tset);
5432 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5434 if (memcg->move_charge_at_immigrate) {
5435 struct mm_struct *mm;
5436 struct mem_cgroup *from = mem_cgroup_from_task(p);
5438 VM_BUG_ON(from == memcg);
5440 mm = get_task_mm(p);
5443 /* We move charges only when we move a owner of the mm */
5444 if (mm->owner == p) {
5447 VM_BUG_ON(mc.precharge);
5448 VM_BUG_ON(mc.moved_charge);
5449 VM_BUG_ON(mc.moved_swap);
5450 mem_cgroup_start_move(from);
5451 spin_lock(&mc.lock);
5454 spin_unlock(&mc.lock);
5455 /* We set mc.moving_task later */
5457 ret = mem_cgroup_precharge_mc(mm);
5459 mem_cgroup_clear_mc();
5466 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5467 struct cgroup *cgroup,
5468 struct cgroup_taskset *tset)
5470 mem_cgroup_clear_mc();
5473 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5474 unsigned long addr, unsigned long end,
5475 struct mm_walk *walk)
5478 struct vm_area_struct *vma = walk->private;
5482 split_huge_page_pmd(walk->mm, pmd);
5484 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5485 for (; addr != end; addr += PAGE_SIZE) {
5486 pte_t ptent = *(pte++);
5487 union mc_target target;
5490 struct page_cgroup *pc;
5496 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5498 case MC_TARGET_PAGE:
5500 if (isolate_lru_page(page))
5502 pc = lookup_page_cgroup(page);
5503 if (!mem_cgroup_move_account(page, 1, pc,
5504 mc.from, mc.to, false)) {
5506 /* we uncharge from mc.from later. */
5509 putback_lru_page(page);
5510 put: /* is_target_pte_for_mc() gets the page */
5513 case MC_TARGET_SWAP:
5515 if (!mem_cgroup_move_swap_account(ent,
5516 mc.from, mc.to, false)) {
5518 /* we fixup refcnts and charges later. */
5526 pte_unmap_unlock(pte - 1, ptl);
5531 * We have consumed all precharges we got in can_attach().
5532 * We try charge one by one, but don't do any additional
5533 * charges to mc.to if we have failed in charge once in attach()
5536 ret = mem_cgroup_do_precharge(1);
5544 static void mem_cgroup_move_charge(struct mm_struct *mm)
5546 struct vm_area_struct *vma;
5548 lru_add_drain_all();
5550 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5552 * Someone who are holding the mmap_sem might be waiting in
5553 * waitq. So we cancel all extra charges, wake up all waiters,
5554 * and retry. Because we cancel precharges, we might not be able
5555 * to move enough charges, but moving charge is a best-effort
5556 * feature anyway, so it wouldn't be a big problem.
5558 __mem_cgroup_clear_mc();
5562 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5564 struct mm_walk mem_cgroup_move_charge_walk = {
5565 .pmd_entry = mem_cgroup_move_charge_pte_range,
5569 if (is_vm_hugetlb_page(vma))
5571 ret = walk_page_range(vma->vm_start, vma->vm_end,
5572 &mem_cgroup_move_charge_walk);
5575 * means we have consumed all precharges and failed in
5576 * doing additional charge. Just abandon here.
5580 up_read(&mm->mmap_sem);
5583 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5584 struct cgroup *cont,
5585 struct cgroup_taskset *tset)
5587 struct task_struct *p = cgroup_taskset_first(tset);
5588 struct mm_struct *mm = get_task_mm(p);
5592 mem_cgroup_move_charge(mm);
5597 mem_cgroup_clear_mc();
5599 #else /* !CONFIG_MMU */
5600 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5601 struct cgroup *cgroup,
5602 struct cgroup_taskset *tset)
5606 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5607 struct cgroup *cgroup,
5608 struct cgroup_taskset *tset)
5611 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5612 struct cgroup *cont,
5613 struct cgroup_taskset *tset)
5618 struct cgroup_subsys mem_cgroup_subsys = {
5620 .subsys_id = mem_cgroup_subsys_id,
5621 .create = mem_cgroup_create,
5622 .pre_destroy = mem_cgroup_pre_destroy,
5623 .destroy = mem_cgroup_destroy,
5624 .populate = mem_cgroup_populate,
5625 .can_attach = mem_cgroup_can_attach,
5626 .cancel_attach = mem_cgroup_cancel_attach,
5627 .attach = mem_cgroup_move_task,
5632 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5633 static int __init enable_swap_account(char *s)
5635 /* consider enabled if no parameter or 1 is given */
5636 if (!strcmp(s, "1"))
5637 really_do_swap_account = 1;
5638 else if (!strcmp(s, "0"))
5639 really_do_swap_account = 0;
5642 __setup("swapaccount=", enable_swap_account);