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];
127 * per-zone information in memory controller.
129 struct mem_cgroup_per_zone {
131 * spin_lock to protect the per cgroup LRU
133 struct list_head lists[NR_LRU_LISTS];
134 unsigned long count[NR_LRU_LISTS];
136 struct zone_reclaim_stat reclaim_stat;
137 struct rb_node tree_node; /* RB tree node */
138 unsigned long long usage_in_excess;/* Set to the value by which */
139 /* the soft limit is exceeded*/
141 struct mem_cgroup *mem; /* Back pointer, we cannot */
142 /* use container_of */
144 /* Macro for accessing counter */
145 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
147 struct mem_cgroup_per_node {
148 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
151 struct mem_cgroup_lru_info {
152 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
156 * Cgroups above their limits are maintained in a RB-Tree, independent of
157 * their hierarchy representation
160 struct mem_cgroup_tree_per_zone {
161 struct rb_root rb_root;
165 struct mem_cgroup_tree_per_node {
166 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
169 struct mem_cgroup_tree {
170 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
173 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
175 struct mem_cgroup_threshold {
176 struct eventfd_ctx *eventfd;
181 struct mem_cgroup_threshold_ary {
182 /* An array index points to threshold just below usage. */
183 int current_threshold;
184 /* Size of entries[] */
186 /* Array of thresholds */
187 struct mem_cgroup_threshold entries[0];
190 struct mem_cgroup_thresholds {
191 /* Primary thresholds array */
192 struct mem_cgroup_threshold_ary *primary;
194 * Spare threshold array.
195 * This is needed to make mem_cgroup_unregister_event() "never fail".
196 * It must be able to store at least primary->size - 1 entries.
198 struct mem_cgroup_threshold_ary *spare;
202 struct mem_cgroup_eventfd_list {
203 struct list_head list;
204 struct eventfd_ctx *eventfd;
207 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
208 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
211 * The memory controller data structure. The memory controller controls both
212 * page cache and RSS per cgroup. We would eventually like to provide
213 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
214 * to help the administrator determine what knobs to tune.
216 * TODO: Add a water mark for the memory controller. Reclaim will begin when
217 * we hit the water mark. May be even add a low water mark, such that
218 * no reclaim occurs from a cgroup at it's low water mark, this is
219 * a feature that will be implemented much later in the future.
222 struct cgroup_subsys_state css;
224 * the counter to account for memory usage
226 struct res_counter res;
228 * the counter to account for mem+swap usage.
230 struct res_counter memsw;
232 * Per cgroup active and inactive list, similar to the
233 * per zone LRU lists.
235 struct mem_cgroup_lru_info info;
237 * While reclaiming in a hierarchy, we cache the last child we
240 int last_scanned_child;
241 int last_scanned_node;
243 nodemask_t scan_nodes;
244 atomic_t numainfo_events;
245 atomic_t numainfo_updating;
248 * Should the accounting and control be hierarchical, per subtree?
258 /* OOM-Killer disable */
259 int oom_kill_disable;
261 /* set when res.limit == memsw.limit */
262 bool memsw_is_minimum;
264 /* protect arrays of thresholds */
265 struct mutex thresholds_lock;
267 /* thresholds for memory usage. RCU-protected */
268 struct mem_cgroup_thresholds thresholds;
270 /* thresholds for mem+swap usage. RCU-protected */
271 struct mem_cgroup_thresholds memsw_thresholds;
273 /* For oom notifier event fd */
274 struct list_head oom_notify;
277 * Should we move charges of a task when a task is moved into this
278 * mem_cgroup ? And what type of charges should we move ?
280 unsigned long move_charge_at_immigrate;
284 struct mem_cgroup_stat_cpu *stat;
286 * used when a cpu is offlined or other synchronizations
287 * See mem_cgroup_read_stat().
289 struct mem_cgroup_stat_cpu nocpu_base;
290 spinlock_t pcp_counter_lock;
293 struct tcp_memcontrol tcp_mem;
297 /* Stuffs for move charges at task migration. */
299 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
300 * left-shifted bitmap of these types.
303 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
304 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
308 /* "mc" and its members are protected by cgroup_mutex */
309 static struct move_charge_struct {
310 spinlock_t lock; /* for from, to */
311 struct mem_cgroup *from;
312 struct mem_cgroup *to;
313 unsigned long precharge;
314 unsigned long moved_charge;
315 unsigned long moved_swap;
316 struct task_struct *moving_task; /* a task moving charges */
317 wait_queue_head_t waitq; /* a waitq for other context */
319 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
320 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
323 static bool move_anon(void)
325 return test_bit(MOVE_CHARGE_TYPE_ANON,
326 &mc.to->move_charge_at_immigrate);
329 static bool move_file(void)
331 return test_bit(MOVE_CHARGE_TYPE_FILE,
332 &mc.to->move_charge_at_immigrate);
336 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
337 * limit reclaim to prevent infinite loops, if they ever occur.
339 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
340 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
343 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
344 MEM_CGROUP_CHARGE_TYPE_MAPPED,
345 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
346 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
347 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
348 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
352 /* for encoding cft->private value on file */
355 #define _OOM_TYPE (2)
356 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
357 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
358 #define MEMFILE_ATTR(val) ((val) & 0xffff)
359 /* Used for OOM nofiier */
360 #define OOM_CONTROL (0)
363 * Reclaim flags for mem_cgroup_hierarchical_reclaim
365 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
366 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
367 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
368 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
369 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
370 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
372 static void mem_cgroup_get(struct mem_cgroup *memcg);
373 static void mem_cgroup_put(struct mem_cgroup *memcg);
375 /* Writing them here to avoid exposing memcg's inner layout */
376 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
378 #include <net/sock.h>
381 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
382 void sock_update_memcg(struct sock *sk)
384 if (static_branch(&memcg_socket_limit_enabled)) {
385 struct mem_cgroup *memcg;
387 BUG_ON(!sk->sk_prot->proto_cgroup);
389 /* Socket cloning can throw us here with sk_cgrp already
390 * filled. It won't however, necessarily happen from
391 * process context. So the test for root memcg given
392 * the current task's memcg won't help us in this case.
394 * Respecting the original socket's memcg is a better
395 * decision in this case.
398 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
399 mem_cgroup_get(sk->sk_cgrp->memcg);
404 memcg = mem_cgroup_from_task(current);
405 if (!mem_cgroup_is_root(memcg)) {
406 mem_cgroup_get(memcg);
407 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
412 EXPORT_SYMBOL(sock_update_memcg);
414 void sock_release_memcg(struct sock *sk)
416 if (static_branch(&memcg_socket_limit_enabled) && sk->sk_cgrp) {
417 struct mem_cgroup *memcg;
418 WARN_ON(!sk->sk_cgrp->memcg);
419 memcg = sk->sk_cgrp->memcg;
420 mem_cgroup_put(memcg);
424 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
426 if (!memcg || mem_cgroup_is_root(memcg))
429 return &memcg->tcp_mem.cg_proto;
431 EXPORT_SYMBOL(tcp_proto_cgroup);
432 #endif /* CONFIG_INET */
433 #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
435 static void drain_all_stock_async(struct mem_cgroup *memcg);
437 static struct mem_cgroup_per_zone *
438 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
440 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
443 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
448 static struct mem_cgroup_per_zone *
449 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
451 int nid = page_to_nid(page);
452 int zid = page_zonenum(page);
454 return mem_cgroup_zoneinfo(memcg, nid, zid);
457 static struct mem_cgroup_tree_per_zone *
458 soft_limit_tree_node_zone(int nid, int zid)
460 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
463 static struct mem_cgroup_tree_per_zone *
464 soft_limit_tree_from_page(struct page *page)
466 int nid = page_to_nid(page);
467 int zid = page_zonenum(page);
469 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
473 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
474 struct mem_cgroup_per_zone *mz,
475 struct mem_cgroup_tree_per_zone *mctz,
476 unsigned long long new_usage_in_excess)
478 struct rb_node **p = &mctz->rb_root.rb_node;
479 struct rb_node *parent = NULL;
480 struct mem_cgroup_per_zone *mz_node;
485 mz->usage_in_excess = new_usage_in_excess;
486 if (!mz->usage_in_excess)
490 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
492 if (mz->usage_in_excess < mz_node->usage_in_excess)
495 * We can't avoid mem cgroups that are over their soft
496 * limit by the same amount
498 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
501 rb_link_node(&mz->tree_node, parent, p);
502 rb_insert_color(&mz->tree_node, &mctz->rb_root);
507 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
508 struct mem_cgroup_per_zone *mz,
509 struct mem_cgroup_tree_per_zone *mctz)
513 rb_erase(&mz->tree_node, &mctz->rb_root);
518 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
519 struct mem_cgroup_per_zone *mz,
520 struct mem_cgroup_tree_per_zone *mctz)
522 spin_lock(&mctz->lock);
523 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
524 spin_unlock(&mctz->lock);
528 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
530 unsigned long long excess;
531 struct mem_cgroup_per_zone *mz;
532 struct mem_cgroup_tree_per_zone *mctz;
533 int nid = page_to_nid(page);
534 int zid = page_zonenum(page);
535 mctz = soft_limit_tree_from_page(page);
538 * Necessary to update all ancestors when hierarchy is used.
539 * because their event counter is not touched.
541 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
542 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
543 excess = res_counter_soft_limit_excess(&memcg->res);
545 * We have to update the tree if mz is on RB-tree or
546 * mem is over its softlimit.
548 if (excess || mz->on_tree) {
549 spin_lock(&mctz->lock);
550 /* if on-tree, remove it */
552 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
554 * Insert again. mz->usage_in_excess will be updated.
555 * If excess is 0, no tree ops.
557 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
558 spin_unlock(&mctz->lock);
563 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
566 struct mem_cgroup_per_zone *mz;
567 struct mem_cgroup_tree_per_zone *mctz;
569 for_each_node_state(node, N_POSSIBLE) {
570 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
571 mz = mem_cgroup_zoneinfo(memcg, node, zone);
572 mctz = soft_limit_tree_node_zone(node, zone);
573 mem_cgroup_remove_exceeded(memcg, mz, mctz);
578 static struct mem_cgroup_per_zone *
579 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
581 struct rb_node *rightmost = NULL;
582 struct mem_cgroup_per_zone *mz;
586 rightmost = rb_last(&mctz->rb_root);
588 goto done; /* Nothing to reclaim from */
590 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
592 * Remove the node now but someone else can add it back,
593 * we will to add it back at the end of reclaim to its correct
594 * position in the tree.
596 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
597 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
598 !css_tryget(&mz->mem->css))
604 static struct mem_cgroup_per_zone *
605 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
607 struct mem_cgroup_per_zone *mz;
609 spin_lock(&mctz->lock);
610 mz = __mem_cgroup_largest_soft_limit_node(mctz);
611 spin_unlock(&mctz->lock);
616 * Implementation Note: reading percpu statistics for memcg.
618 * Both of vmstat[] and percpu_counter has threshold and do periodic
619 * synchronization to implement "quick" read. There are trade-off between
620 * reading cost and precision of value. Then, we may have a chance to implement
621 * a periodic synchronizion of counter in memcg's counter.
623 * But this _read() function is used for user interface now. The user accounts
624 * memory usage by memory cgroup and he _always_ requires exact value because
625 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
626 * have to visit all online cpus and make sum. So, for now, unnecessary
627 * synchronization is not implemented. (just implemented for cpu hotplug)
629 * If there are kernel internal actions which can make use of some not-exact
630 * value, and reading all cpu value can be performance bottleneck in some
631 * common workload, threashold and synchonization as vmstat[] should be
634 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
635 enum mem_cgroup_stat_index idx)
641 for_each_online_cpu(cpu)
642 val += per_cpu(memcg->stat->count[idx], cpu);
643 #ifdef CONFIG_HOTPLUG_CPU
644 spin_lock(&memcg->pcp_counter_lock);
645 val += memcg->nocpu_base.count[idx];
646 spin_unlock(&memcg->pcp_counter_lock);
652 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
655 int val = (charge) ? 1 : -1;
656 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
659 void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val)
661 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
664 void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val)
666 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
669 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
670 enum mem_cgroup_events_index idx)
672 unsigned long val = 0;
675 for_each_online_cpu(cpu)
676 val += per_cpu(memcg->stat->events[idx], cpu);
677 #ifdef CONFIG_HOTPLUG_CPU
678 spin_lock(&memcg->pcp_counter_lock);
679 val += memcg->nocpu_base.events[idx];
680 spin_unlock(&memcg->pcp_counter_lock);
685 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
686 bool file, int nr_pages)
691 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
694 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
697 /* pagein of a big page is an event. So, ignore page size */
699 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
701 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
702 nr_pages = -nr_pages; /* for event */
705 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
711 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
712 unsigned int lru_mask)
714 struct mem_cgroup_per_zone *mz;
716 unsigned long ret = 0;
718 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
721 if (BIT(l) & lru_mask)
722 ret += MEM_CGROUP_ZSTAT(mz, l);
728 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
729 int nid, unsigned int lru_mask)
734 for (zid = 0; zid < MAX_NR_ZONES; zid++)
735 total += mem_cgroup_zone_nr_lru_pages(memcg,
741 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
742 unsigned int lru_mask)
747 for_each_node_state(nid, N_HIGH_MEMORY)
748 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
752 static bool __memcg_event_check(struct mem_cgroup *memcg, int target)
754 unsigned long val, next;
756 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
757 next = __this_cpu_read(memcg->stat->targets[target]);
758 /* from time_after() in jiffies.h */
759 return ((long)next - (long)val < 0);
762 static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
764 unsigned long val, next;
766 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
769 case MEM_CGROUP_TARGET_THRESH:
770 next = val + THRESHOLDS_EVENTS_TARGET;
772 case MEM_CGROUP_TARGET_SOFTLIMIT:
773 next = val + SOFTLIMIT_EVENTS_TARGET;
775 case MEM_CGROUP_TARGET_NUMAINFO:
776 next = val + NUMAINFO_EVENTS_TARGET;
782 __this_cpu_write(memcg->stat->targets[target], next);
786 * Check events in order.
789 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
792 /* threshold event is triggered in finer grain than soft limit */
793 if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
794 mem_cgroup_threshold(memcg);
795 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
796 if (unlikely(__memcg_event_check(memcg,
797 MEM_CGROUP_TARGET_SOFTLIMIT))) {
798 mem_cgroup_update_tree(memcg, page);
799 __mem_cgroup_target_update(memcg,
800 MEM_CGROUP_TARGET_SOFTLIMIT);
803 if (unlikely(__memcg_event_check(memcg,
804 MEM_CGROUP_TARGET_NUMAINFO))) {
805 atomic_inc(&memcg->numainfo_events);
806 __mem_cgroup_target_update(memcg,
807 MEM_CGROUP_TARGET_NUMAINFO);
814 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
816 return container_of(cgroup_subsys_state(cont,
817 mem_cgroup_subsys_id), struct mem_cgroup,
821 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
824 * mm_update_next_owner() may clear mm->owner to NULL
825 * if it races with swapoff, page migration, etc.
826 * So this can be called with p == NULL.
831 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
832 struct mem_cgroup, css);
835 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
837 struct mem_cgroup *memcg = NULL;
842 * Because we have no locks, mm->owner's may be being moved to other
843 * cgroup. We use css_tryget() here even if this looks
844 * pessimistic (rather than adding locks here).
848 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
849 if (unlikely(!memcg))
851 } while (!css_tryget(&memcg->css));
856 static struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
857 struct mem_cgroup *prev,
860 struct mem_cgroup *memcg = NULL;
864 root = root_mem_cgroup;
866 if (prev && !reclaim)
867 id = css_id(&prev->css);
869 if (prev && prev != root)
872 if (!root->use_hierarchy && root != root_mem_cgroup) {
879 struct cgroup_subsys_state *css;
882 id = root->last_scanned_child;
885 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
887 if (css == &root->css || css_tryget(css))
888 memcg = container_of(css,
889 struct mem_cgroup, css);
895 root->last_scanned_child = id;
903 static void mem_cgroup_iter_break(struct mem_cgroup *root,
904 struct mem_cgroup *prev)
907 root = root_mem_cgroup;
908 if (prev && prev != root)
913 * Iteration constructs for visiting all cgroups (under a tree). If
914 * loops are exited prematurely (break), mem_cgroup_iter_break() must
915 * be used for reference counting.
917 #define for_each_mem_cgroup_tree(iter, root) \
918 for (iter = mem_cgroup_iter(root, NULL, false); \
920 iter = mem_cgroup_iter(root, iter, false))
922 #define for_each_mem_cgroup(iter) \
923 for (iter = mem_cgroup_iter(NULL, NULL, false); \
925 iter = mem_cgroup_iter(NULL, iter, false))
927 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
929 return (memcg == root_mem_cgroup);
932 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
934 struct mem_cgroup *memcg;
940 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
941 if (unlikely(!memcg))
946 mem_cgroup_pgmajfault(memcg, 1);
949 mem_cgroup_pgfault(memcg, 1);
957 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
960 * Following LRU functions are allowed to be used without PCG_LOCK.
961 * Operations are called by routine of global LRU independently from memcg.
962 * What we have to take care of here is validness of pc->mem_cgroup.
964 * Changes to pc->mem_cgroup happens when
967 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
968 * It is added to LRU before charge.
969 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
970 * When moving account, the page is not on LRU. It's isolated.
973 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
975 struct page_cgroup *pc;
976 struct mem_cgroup_per_zone *mz;
978 if (mem_cgroup_disabled())
980 pc = lookup_page_cgroup(page);
981 /* can happen while we handle swapcache. */
982 if (!TestClearPageCgroupAcctLRU(pc))
984 VM_BUG_ON(!pc->mem_cgroup);
986 * We don't check PCG_USED bit. It's cleared when the "page" is finally
987 * removed from global LRU.
989 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
990 /* huge page split is done under lru_lock. so, we have no races. */
991 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
992 if (mem_cgroup_is_root(pc->mem_cgroup))
994 VM_BUG_ON(list_empty(&pc->lru));
995 list_del_init(&pc->lru);
998 void mem_cgroup_del_lru(struct page *page)
1000 mem_cgroup_del_lru_list(page, page_lru(page));
1004 * Writeback is about to end against a page which has been marked for immediate
1005 * reclaim. If it still appears to be reclaimable, move it to the tail of the
1008 void mem_cgroup_rotate_reclaimable_page(struct page *page)
1010 struct mem_cgroup_per_zone *mz;
1011 struct page_cgroup *pc;
1012 enum lru_list lru = page_lru(page);
1014 if (mem_cgroup_disabled())
1017 pc = lookup_page_cgroup(page);
1018 /* unused or root page is not rotated. */
1019 if (!PageCgroupUsed(pc))
1021 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1023 if (mem_cgroup_is_root(pc->mem_cgroup))
1025 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1026 list_move_tail(&pc->lru, &mz->lists[lru]);
1029 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1031 struct mem_cgroup_per_zone *mz;
1032 struct page_cgroup *pc;
1034 if (mem_cgroup_disabled())
1037 pc = lookup_page_cgroup(page);
1038 /* unused or root page is not rotated. */
1039 if (!PageCgroupUsed(pc))
1041 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1043 if (mem_cgroup_is_root(pc->mem_cgroup))
1045 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1046 list_move(&pc->lru, &mz->lists[lru]);
1049 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1051 struct page_cgroup *pc;
1052 struct mem_cgroup_per_zone *mz;
1054 if (mem_cgroup_disabled())
1056 pc = lookup_page_cgroup(page);
1057 VM_BUG_ON(PageCgroupAcctLRU(pc));
1060 * SetPageLRU SetPageCgroupUsed
1062 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1064 * Ensure that one of the two sides adds the page to the memcg
1065 * LRU during a race.
1068 if (!PageCgroupUsed(pc))
1070 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1072 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1073 /* huge page split is done under lru_lock. so, we have no races. */
1074 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1075 SetPageCgroupAcctLRU(pc);
1076 if (mem_cgroup_is_root(pc->mem_cgroup))
1078 list_add(&pc->lru, &mz->lists[lru]);
1082 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1083 * while it's linked to lru because the page may be reused after it's fully
1084 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1085 * It's done under lock_page and expected that zone->lru_lock isnever held.
1087 static void mem_cgroup_lru_del_before_commit(struct page *page)
1089 unsigned long flags;
1090 struct zone *zone = page_zone(page);
1091 struct page_cgroup *pc = lookup_page_cgroup(page);
1094 * Doing this check without taking ->lru_lock seems wrong but this
1095 * is safe. Because if page_cgroup's USED bit is unset, the page
1096 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1097 * set, the commit after this will fail, anyway.
1098 * This all charge/uncharge is done under some mutual execustion.
1099 * So, we don't need to taking care of changes in USED bit.
1101 if (likely(!PageLRU(page)))
1104 spin_lock_irqsave(&zone->lru_lock, flags);
1106 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1107 * is guarded by lock_page() because the page is SwapCache.
1109 if (!PageCgroupUsed(pc))
1110 mem_cgroup_del_lru_list(page, page_lru(page));
1111 spin_unlock_irqrestore(&zone->lru_lock, flags);
1114 static void mem_cgroup_lru_add_after_commit(struct page *page)
1116 unsigned long flags;
1117 struct zone *zone = page_zone(page);
1118 struct page_cgroup *pc = lookup_page_cgroup(page);
1121 * SetPageLRU SetPageCgroupUsed
1123 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1125 * Ensure that one of the two sides adds the page to the memcg
1126 * LRU during a race.
1129 /* taking care of that the page is added to LRU while we commit it */
1130 if (likely(!PageLRU(page)))
1132 spin_lock_irqsave(&zone->lru_lock, flags);
1133 /* link when the page is linked to LRU but page_cgroup isn't */
1134 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1135 mem_cgroup_add_lru_list(page, page_lru(page));
1136 spin_unlock_irqrestore(&zone->lru_lock, flags);
1140 void mem_cgroup_move_lists(struct page *page,
1141 enum lru_list from, enum lru_list to)
1143 if (mem_cgroup_disabled())
1145 mem_cgroup_del_lru_list(page, from);
1146 mem_cgroup_add_lru_list(page, to);
1150 * Checks whether given mem is same or in the root_mem_cgroup's
1153 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1154 struct mem_cgroup *memcg)
1156 if (root_memcg != memcg) {
1157 return (root_memcg->use_hierarchy &&
1158 css_is_ancestor(&memcg->css, &root_memcg->css));
1164 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1167 struct mem_cgroup *curr = NULL;
1168 struct task_struct *p;
1170 p = find_lock_task_mm(task);
1173 curr = try_get_mem_cgroup_from_mm(p->mm);
1178 * We should check use_hierarchy of "memcg" not "curr". Because checking
1179 * use_hierarchy of "curr" here make this function true if hierarchy is
1180 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1181 * hierarchy(even if use_hierarchy is disabled in "memcg").
1183 ret = mem_cgroup_same_or_subtree(memcg, curr);
1184 css_put(&curr->css);
1188 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1190 unsigned long inactive_ratio;
1191 int nid = zone_to_nid(zone);
1192 int zid = zone_idx(zone);
1193 unsigned long inactive;
1194 unsigned long active;
1197 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1198 BIT(LRU_INACTIVE_ANON));
1199 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1200 BIT(LRU_ACTIVE_ANON));
1202 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1204 inactive_ratio = int_sqrt(10 * gb);
1208 return inactive * inactive_ratio < active;
1211 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1213 unsigned long active;
1214 unsigned long inactive;
1215 int zid = zone_idx(zone);
1216 int nid = zone_to_nid(zone);
1218 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1219 BIT(LRU_INACTIVE_FILE));
1220 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1221 BIT(LRU_ACTIVE_FILE));
1223 return (active > inactive);
1226 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1229 int nid = zone_to_nid(zone);
1230 int zid = zone_idx(zone);
1231 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1233 return &mz->reclaim_stat;
1236 struct zone_reclaim_stat *
1237 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1239 struct page_cgroup *pc;
1240 struct mem_cgroup_per_zone *mz;
1242 if (mem_cgroup_disabled())
1245 pc = lookup_page_cgroup(page);
1246 if (!PageCgroupUsed(pc))
1248 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1250 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1251 return &mz->reclaim_stat;
1254 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1255 struct list_head *dst,
1256 unsigned long *scanned, int order,
1257 isolate_mode_t mode,
1259 struct mem_cgroup *mem_cont,
1260 int active, int file)
1262 unsigned long nr_taken = 0;
1266 struct list_head *src;
1267 struct page_cgroup *pc, *tmp;
1268 int nid = zone_to_nid(z);
1269 int zid = zone_idx(z);
1270 struct mem_cgroup_per_zone *mz;
1271 int lru = LRU_FILE * file + active;
1275 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1276 src = &mz->lists[lru];
1279 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1280 if (scan >= nr_to_scan)
1283 if (unlikely(!PageCgroupUsed(pc)))
1286 page = lookup_cgroup_page(pc);
1288 if (unlikely(!PageLRU(page)))
1292 ret = __isolate_lru_page(page, mode, file);
1295 list_move(&page->lru, dst);
1296 mem_cgroup_del_lru(page);
1297 nr_taken += hpage_nr_pages(page);
1300 /* we don't affect global LRU but rotate in our LRU */
1301 mem_cgroup_rotate_lru_list(page, page_lru(page));
1310 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1316 #define mem_cgroup_from_res_counter(counter, member) \
1317 container_of(counter, struct mem_cgroup, member)
1320 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1321 * @mem: the memory cgroup
1323 * Returns the maximum amount of memory @mem can be charged with, in
1326 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1328 unsigned long long margin;
1330 margin = res_counter_margin(&memcg->res);
1331 if (do_swap_account)
1332 margin = min(margin, res_counter_margin(&memcg->memsw));
1333 return margin >> PAGE_SHIFT;
1336 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1338 struct cgroup *cgrp = memcg->css.cgroup;
1341 if (cgrp->parent == NULL)
1342 return vm_swappiness;
1344 return memcg->swappiness;
1347 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1352 spin_lock(&memcg->pcp_counter_lock);
1353 for_each_online_cpu(cpu)
1354 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1355 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1356 spin_unlock(&memcg->pcp_counter_lock);
1362 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1369 spin_lock(&memcg->pcp_counter_lock);
1370 for_each_online_cpu(cpu)
1371 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1372 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1373 spin_unlock(&memcg->pcp_counter_lock);
1377 * 2 routines for checking "mem" is under move_account() or not.
1379 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1380 * for avoiding race in accounting. If true,
1381 * pc->mem_cgroup may be overwritten.
1383 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1384 * under hierarchy of moving cgroups. This is for
1385 * waiting at hith-memory prressure caused by "move".
1388 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1390 VM_BUG_ON(!rcu_read_lock_held());
1391 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1394 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1396 struct mem_cgroup *from;
1397 struct mem_cgroup *to;
1400 * Unlike task_move routines, we access mc.to, mc.from not under
1401 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1403 spin_lock(&mc.lock);
1409 ret = mem_cgroup_same_or_subtree(memcg, from)
1410 || mem_cgroup_same_or_subtree(memcg, to);
1412 spin_unlock(&mc.lock);
1416 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1418 if (mc.moving_task && current != mc.moving_task) {
1419 if (mem_cgroup_under_move(memcg)) {
1421 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1422 /* moving charge context might have finished. */
1425 finish_wait(&mc.waitq, &wait);
1433 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1434 * @memcg: The memory cgroup that went over limit
1435 * @p: Task that is going to be killed
1437 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1440 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1442 struct cgroup *task_cgrp;
1443 struct cgroup *mem_cgrp;
1445 * Need a buffer in BSS, can't rely on allocations. The code relies
1446 * on the assumption that OOM is serialized for memory controller.
1447 * If this assumption is broken, revisit this code.
1449 static char memcg_name[PATH_MAX];
1458 mem_cgrp = memcg->css.cgroup;
1459 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1461 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1464 * Unfortunately, we are unable to convert to a useful name
1465 * But we'll still print out the usage information
1472 printk(KERN_INFO "Task in %s killed", memcg_name);
1475 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1483 * Continues from above, so we don't need an KERN_ level
1485 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1488 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1489 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1490 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1491 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1492 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1494 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1495 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1496 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1500 * This function returns the number of memcg under hierarchy tree. Returns
1501 * 1(self count) if no children.
1503 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1506 struct mem_cgroup *iter;
1508 for_each_mem_cgroup_tree(iter, memcg)
1514 * Return the memory (and swap, if configured) limit for a memcg.
1516 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1521 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1522 limit += total_swap_pages << PAGE_SHIFT;
1524 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1526 * If memsw is finite and limits the amount of swap space available
1527 * to this memcg, return that limit.
1529 return min(limit, memsw);
1533 * test_mem_cgroup_node_reclaimable
1534 * @mem: the target memcg
1535 * @nid: the node ID to be checked.
1536 * @noswap : specify true here if the user wants flle only information.
1538 * This function returns whether the specified memcg contains any
1539 * reclaimable pages on a node. Returns true if there are any reclaimable
1540 * pages in the node.
1542 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1543 int nid, bool noswap)
1545 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1547 if (noswap || !total_swap_pages)
1549 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1554 #if MAX_NUMNODES > 1
1557 * Always updating the nodemask is not very good - even if we have an empty
1558 * list or the wrong list here, we can start from some node and traverse all
1559 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1562 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1566 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1567 * pagein/pageout changes since the last update.
1569 if (!atomic_read(&memcg->numainfo_events))
1571 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1574 /* make a nodemask where this memcg uses memory from */
1575 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1577 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1579 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1580 node_clear(nid, memcg->scan_nodes);
1583 atomic_set(&memcg->numainfo_events, 0);
1584 atomic_set(&memcg->numainfo_updating, 0);
1588 * Selecting a node where we start reclaim from. Because what we need is just
1589 * reducing usage counter, start from anywhere is O,K. Considering
1590 * memory reclaim from current node, there are pros. and cons.
1592 * Freeing memory from current node means freeing memory from a node which
1593 * we'll use or we've used. So, it may make LRU bad. And if several threads
1594 * hit limits, it will see a contention on a node. But freeing from remote
1595 * node means more costs for memory reclaim because of memory latency.
1597 * Now, we use round-robin. Better algorithm is welcomed.
1599 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1603 mem_cgroup_may_update_nodemask(memcg);
1604 node = memcg->last_scanned_node;
1606 node = next_node(node, memcg->scan_nodes);
1607 if (node == MAX_NUMNODES)
1608 node = first_node(memcg->scan_nodes);
1610 * We call this when we hit limit, not when pages are added to LRU.
1611 * No LRU may hold pages because all pages are UNEVICTABLE or
1612 * memcg is too small and all pages are not on LRU. In that case,
1613 * we use curret node.
1615 if (unlikely(node == MAX_NUMNODES))
1616 node = numa_node_id();
1618 memcg->last_scanned_node = node;
1623 * Check all nodes whether it contains reclaimable pages or not.
1624 * For quick scan, we make use of scan_nodes. This will allow us to skip
1625 * unused nodes. But scan_nodes is lazily updated and may not cotain
1626 * enough new information. We need to do double check.
1628 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1633 * quick check...making use of scan_node.
1634 * We can skip unused nodes.
1636 if (!nodes_empty(memcg->scan_nodes)) {
1637 for (nid = first_node(memcg->scan_nodes);
1639 nid = next_node(nid, memcg->scan_nodes)) {
1641 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1646 * Check rest of nodes.
1648 for_each_node_state(nid, N_HIGH_MEMORY) {
1649 if (node_isset(nid, memcg->scan_nodes))
1651 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1658 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1663 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1665 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1670 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1671 * we reclaimed from, so that we don't end up penalizing one child extensively
1672 * based on its position in the children list.
1674 * root_memcg is the original ancestor that we've been reclaim from.
1676 * We give up and return to the caller when we visit root_memcg twice.
1677 * (other groups can be removed while we're walking....)
1679 * If shrink==true, for avoiding to free too much, this returns immedieately.
1681 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_memcg,
1684 unsigned long reclaim_options,
1685 unsigned long *total_scanned)
1687 struct mem_cgroup *victim = NULL;
1690 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1691 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1692 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1693 unsigned long excess;
1694 unsigned long nr_scanned;
1696 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1698 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1699 if (!check_soft && !shrink && root_memcg->memsw_is_minimum)
1703 victim = mem_cgroup_iter(root_memcg, victim, true);
1707 * We are not draining per cpu cached charges during
1708 * soft limit reclaim because global reclaim doesn't
1709 * care about charges. It tries to free some memory and
1710 * charges will not give any.
1712 if (!check_soft && loop >= 1)
1713 drain_all_stock_async(root_memcg);
1716 * If we have not been able to reclaim
1717 * anything, it might because there are
1718 * no reclaimable pages under this hierarchy
1720 if (!check_soft || !total)
1723 * We want to do more targeted reclaim.
1724 * excess >> 2 is not to excessive so as to
1725 * reclaim too much, nor too less that we keep
1726 * coming back to reclaim from this cgroup
1728 if (total >= (excess >> 2) ||
1729 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1734 if (!mem_cgroup_reclaimable(victim, noswap)) {
1735 /* this cgroup's local usage == 0 */
1738 /* we use swappiness of local cgroup */
1740 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1741 noswap, zone, &nr_scanned);
1742 *total_scanned += nr_scanned;
1744 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1748 * At shrinking usage, we can't check we should stop here or
1749 * reclaim more. It's depends on callers. last_scanned_child
1750 * will work enough for keeping fairness under tree.
1755 if (!res_counter_soft_limit_excess(&root_memcg->res))
1757 } else if (mem_cgroup_margin(root_memcg))
1760 mem_cgroup_iter_break(root_memcg, victim);
1765 * Check OOM-Killer is already running under our hierarchy.
1766 * If someone is running, return false.
1767 * Has to be called with memcg_oom_lock
1769 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1771 struct mem_cgroup *iter, *failed = NULL;
1773 for_each_mem_cgroup_tree(iter, memcg) {
1774 if (iter->oom_lock) {
1776 * this subtree of our hierarchy is already locked
1777 * so we cannot give a lock.
1780 mem_cgroup_iter_break(memcg, iter);
1783 iter->oom_lock = true;
1790 * OK, we failed to lock the whole subtree so we have to clean up
1791 * what we set up to the failing subtree
1793 for_each_mem_cgroup_tree(iter, memcg) {
1794 if (iter == failed) {
1795 mem_cgroup_iter_break(memcg, iter);
1798 iter->oom_lock = false;
1804 * Has to be called with memcg_oom_lock
1806 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1808 struct mem_cgroup *iter;
1810 for_each_mem_cgroup_tree(iter, memcg)
1811 iter->oom_lock = false;
1815 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1817 struct mem_cgroup *iter;
1819 for_each_mem_cgroup_tree(iter, memcg)
1820 atomic_inc(&iter->under_oom);
1823 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1825 struct mem_cgroup *iter;
1828 * When a new child is created while the hierarchy is under oom,
1829 * mem_cgroup_oom_lock() may not be called. We have to use
1830 * atomic_add_unless() here.
1832 for_each_mem_cgroup_tree(iter, memcg)
1833 atomic_add_unless(&iter->under_oom, -1, 0);
1836 static DEFINE_SPINLOCK(memcg_oom_lock);
1837 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1839 struct oom_wait_info {
1840 struct mem_cgroup *mem;
1844 static int memcg_oom_wake_function(wait_queue_t *wait,
1845 unsigned mode, int sync, void *arg)
1847 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1849 struct oom_wait_info *oom_wait_info;
1851 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1852 oom_wait_memcg = oom_wait_info->mem;
1855 * Both of oom_wait_info->mem and wake_mem are stable under us.
1856 * Then we can use css_is_ancestor without taking care of RCU.
1858 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1859 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1861 return autoremove_wake_function(wait, mode, sync, arg);
1864 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1866 /* for filtering, pass "memcg" as argument. */
1867 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1870 static void memcg_oom_recover(struct mem_cgroup *memcg)
1872 if (memcg && atomic_read(&memcg->under_oom))
1873 memcg_wakeup_oom(memcg);
1877 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1879 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1881 struct oom_wait_info owait;
1882 bool locked, need_to_kill;
1885 owait.wait.flags = 0;
1886 owait.wait.func = memcg_oom_wake_function;
1887 owait.wait.private = current;
1888 INIT_LIST_HEAD(&owait.wait.task_list);
1889 need_to_kill = true;
1890 mem_cgroup_mark_under_oom(memcg);
1892 /* At first, try to OOM lock hierarchy under memcg.*/
1893 spin_lock(&memcg_oom_lock);
1894 locked = mem_cgroup_oom_lock(memcg);
1896 * Even if signal_pending(), we can't quit charge() loop without
1897 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1898 * under OOM is always welcomed, use TASK_KILLABLE here.
1900 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1901 if (!locked || memcg->oom_kill_disable)
1902 need_to_kill = false;
1904 mem_cgroup_oom_notify(memcg);
1905 spin_unlock(&memcg_oom_lock);
1908 finish_wait(&memcg_oom_waitq, &owait.wait);
1909 mem_cgroup_out_of_memory(memcg, mask);
1912 finish_wait(&memcg_oom_waitq, &owait.wait);
1914 spin_lock(&memcg_oom_lock);
1916 mem_cgroup_oom_unlock(memcg);
1917 memcg_wakeup_oom(memcg);
1918 spin_unlock(&memcg_oom_lock);
1920 mem_cgroup_unmark_under_oom(memcg);
1922 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1924 /* Give chance to dying process */
1925 schedule_timeout_uninterruptible(1);
1930 * Currently used to update mapped file statistics, but the routine can be
1931 * generalized to update other statistics as well.
1933 * Notes: Race condition
1935 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1936 * it tends to be costly. But considering some conditions, we doesn't need
1937 * to do so _always_.
1939 * Considering "charge", lock_page_cgroup() is not required because all
1940 * file-stat operations happen after a page is attached to radix-tree. There
1941 * are no race with "charge".
1943 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1944 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1945 * if there are race with "uncharge". Statistics itself is properly handled
1948 * Considering "move", this is an only case we see a race. To make the race
1949 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1950 * possibility of race condition. If there is, we take a lock.
1953 void mem_cgroup_update_page_stat(struct page *page,
1954 enum mem_cgroup_page_stat_item idx, int val)
1956 struct mem_cgroup *memcg;
1957 struct page_cgroup *pc = lookup_page_cgroup(page);
1958 bool need_unlock = false;
1959 unsigned long uninitialized_var(flags);
1965 memcg = pc->mem_cgroup;
1966 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1968 /* pc->mem_cgroup is unstable ? */
1969 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
1970 /* take a lock against to access pc->mem_cgroup */
1971 move_lock_page_cgroup(pc, &flags);
1973 memcg = pc->mem_cgroup;
1974 if (!memcg || !PageCgroupUsed(pc))
1979 case MEMCG_NR_FILE_MAPPED:
1981 SetPageCgroupFileMapped(pc);
1982 else if (!page_mapped(page))
1983 ClearPageCgroupFileMapped(pc);
1984 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1990 this_cpu_add(memcg->stat->count[idx], val);
1993 if (unlikely(need_unlock))
1994 move_unlock_page_cgroup(pc, &flags);
1998 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2001 * size of first charge trial. "32" comes from vmscan.c's magic value.
2002 * TODO: maybe necessary to use big numbers in big irons.
2004 #define CHARGE_BATCH 32U
2005 struct memcg_stock_pcp {
2006 struct mem_cgroup *cached; /* this never be root cgroup */
2007 unsigned int nr_pages;
2008 struct work_struct work;
2009 unsigned long flags;
2010 #define FLUSHING_CACHED_CHARGE (0)
2012 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2013 static DEFINE_MUTEX(percpu_charge_mutex);
2016 * Try to consume stocked charge on this cpu. If success, one page is consumed
2017 * from local stock and true is returned. If the stock is 0 or charges from a
2018 * cgroup which is not current target, returns false. This stock will be
2021 static bool consume_stock(struct mem_cgroup *memcg)
2023 struct memcg_stock_pcp *stock;
2026 stock = &get_cpu_var(memcg_stock);
2027 if (memcg == stock->cached && stock->nr_pages)
2029 else /* need to call res_counter_charge */
2031 put_cpu_var(memcg_stock);
2036 * Returns stocks cached in percpu to res_counter and reset cached information.
2038 static void drain_stock(struct memcg_stock_pcp *stock)
2040 struct mem_cgroup *old = stock->cached;
2042 if (stock->nr_pages) {
2043 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2045 res_counter_uncharge(&old->res, bytes);
2046 if (do_swap_account)
2047 res_counter_uncharge(&old->memsw, bytes);
2048 stock->nr_pages = 0;
2050 stock->cached = NULL;
2054 * This must be called under preempt disabled or must be called by
2055 * a thread which is pinned to local cpu.
2057 static void drain_local_stock(struct work_struct *dummy)
2059 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2061 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2065 * Cache charges(val) which is from res_counter, to local per_cpu area.
2066 * This will be consumed by consume_stock() function, later.
2068 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2070 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2072 if (stock->cached != memcg) { /* reset if necessary */
2074 stock->cached = memcg;
2076 stock->nr_pages += nr_pages;
2077 put_cpu_var(memcg_stock);
2081 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2082 * of the hierarchy under it. sync flag says whether we should block
2083 * until the work is done.
2085 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2089 /* Notify other cpus that system-wide "drain" is running */
2092 for_each_online_cpu(cpu) {
2093 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2094 struct mem_cgroup *memcg;
2096 memcg = stock->cached;
2097 if (!memcg || !stock->nr_pages)
2099 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2101 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2103 drain_local_stock(&stock->work);
2105 schedule_work_on(cpu, &stock->work);
2113 for_each_online_cpu(cpu) {
2114 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2115 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2116 flush_work(&stock->work);
2123 * Tries to drain stocked charges in other cpus. This function is asynchronous
2124 * and just put a work per cpu for draining localy on each cpu. Caller can
2125 * expects some charges will be back to res_counter later but cannot wait for
2128 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2131 * If someone calls draining, avoid adding more kworker runs.
2133 if (!mutex_trylock(&percpu_charge_mutex))
2135 drain_all_stock(root_memcg, false);
2136 mutex_unlock(&percpu_charge_mutex);
2139 /* This is a synchronous drain interface. */
2140 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2142 /* called when force_empty is called */
2143 mutex_lock(&percpu_charge_mutex);
2144 drain_all_stock(root_memcg, true);
2145 mutex_unlock(&percpu_charge_mutex);
2149 * This function drains percpu counter value from DEAD cpu and
2150 * move it to local cpu. Note that this function can be preempted.
2152 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2156 spin_lock(&memcg->pcp_counter_lock);
2157 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2158 long x = per_cpu(memcg->stat->count[i], cpu);
2160 per_cpu(memcg->stat->count[i], cpu) = 0;
2161 memcg->nocpu_base.count[i] += x;
2163 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2164 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2166 per_cpu(memcg->stat->events[i], cpu) = 0;
2167 memcg->nocpu_base.events[i] += x;
2169 /* need to clear ON_MOVE value, works as a kind of lock. */
2170 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2171 spin_unlock(&memcg->pcp_counter_lock);
2174 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2176 int idx = MEM_CGROUP_ON_MOVE;
2178 spin_lock(&memcg->pcp_counter_lock);
2179 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2180 spin_unlock(&memcg->pcp_counter_lock);
2183 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2184 unsigned long action,
2187 int cpu = (unsigned long)hcpu;
2188 struct memcg_stock_pcp *stock;
2189 struct mem_cgroup *iter;
2191 if ((action == CPU_ONLINE)) {
2192 for_each_mem_cgroup(iter)
2193 synchronize_mem_cgroup_on_move(iter, cpu);
2197 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2200 for_each_mem_cgroup(iter)
2201 mem_cgroup_drain_pcp_counter(iter, cpu);
2203 stock = &per_cpu(memcg_stock, cpu);
2209 /* See __mem_cgroup_try_charge() for details */
2211 CHARGE_OK, /* success */
2212 CHARGE_RETRY, /* need to retry but retry is not bad */
2213 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2214 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2215 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2218 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2219 unsigned int nr_pages, bool oom_check)
2221 unsigned long csize = nr_pages * PAGE_SIZE;
2222 struct mem_cgroup *mem_over_limit;
2223 struct res_counter *fail_res;
2224 unsigned long flags = 0;
2227 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2230 if (!do_swap_account)
2232 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2236 res_counter_uncharge(&memcg->res, csize);
2237 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2238 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2240 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2242 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2243 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2245 * Never reclaim on behalf of optional batching, retry with a
2246 * single page instead.
2248 if (nr_pages == CHARGE_BATCH)
2249 return CHARGE_RETRY;
2251 if (!(gfp_mask & __GFP_WAIT))
2252 return CHARGE_WOULDBLOCK;
2254 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2255 gfp_mask, flags, NULL);
2256 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2257 return CHARGE_RETRY;
2259 * Even though the limit is exceeded at this point, reclaim
2260 * may have been able to free some pages. Retry the charge
2261 * before killing the task.
2263 * Only for regular pages, though: huge pages are rather
2264 * unlikely to succeed so close to the limit, and we fall back
2265 * to regular pages anyway in case of failure.
2267 if (nr_pages == 1 && ret)
2268 return CHARGE_RETRY;
2271 * At task move, charge accounts can be doubly counted. So, it's
2272 * better to wait until the end of task_move if something is going on.
2274 if (mem_cgroup_wait_acct_move(mem_over_limit))
2275 return CHARGE_RETRY;
2277 /* If we don't need to call oom-killer at el, return immediately */
2279 return CHARGE_NOMEM;
2281 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2282 return CHARGE_OOM_DIE;
2284 return CHARGE_RETRY;
2288 * Unlike exported interface, "oom" parameter is added. if oom==true,
2289 * oom-killer can be invoked.
2291 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2293 unsigned int nr_pages,
2294 struct mem_cgroup **ptr,
2297 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2298 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2299 struct mem_cgroup *memcg = NULL;
2303 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2304 * in system level. So, allow to go ahead dying process in addition to
2307 if (unlikely(test_thread_flag(TIF_MEMDIE)
2308 || fatal_signal_pending(current)))
2312 * We always charge the cgroup the mm_struct belongs to.
2313 * The mm_struct's mem_cgroup changes on task migration if the
2314 * thread group leader migrates. It's possible that mm is not
2315 * set, if so charge the init_mm (happens for pagecache usage).
2320 if (*ptr) { /* css should be a valid one */
2322 VM_BUG_ON(css_is_removed(&memcg->css));
2323 if (mem_cgroup_is_root(memcg))
2325 if (nr_pages == 1 && consume_stock(memcg))
2327 css_get(&memcg->css);
2329 struct task_struct *p;
2332 p = rcu_dereference(mm->owner);
2334 * Because we don't have task_lock(), "p" can exit.
2335 * In that case, "memcg" can point to root or p can be NULL with
2336 * race with swapoff. Then, we have small risk of mis-accouning.
2337 * But such kind of mis-account by race always happens because
2338 * we don't have cgroup_mutex(). It's overkill and we allo that
2340 * (*) swapoff at el will charge against mm-struct not against
2341 * task-struct. So, mm->owner can be NULL.
2343 memcg = mem_cgroup_from_task(p);
2344 if (!memcg || mem_cgroup_is_root(memcg)) {
2348 if (nr_pages == 1 && consume_stock(memcg)) {
2350 * It seems dagerous to access memcg without css_get().
2351 * But considering how consume_stok works, it's not
2352 * necessary. If consume_stock success, some charges
2353 * from this memcg are cached on this cpu. So, we
2354 * don't need to call css_get()/css_tryget() before
2355 * calling consume_stock().
2360 /* after here, we may be blocked. we need to get refcnt */
2361 if (!css_tryget(&memcg->css)) {
2371 /* If killed, bypass charge */
2372 if (fatal_signal_pending(current)) {
2373 css_put(&memcg->css);
2378 if (oom && !nr_oom_retries) {
2380 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2383 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2387 case CHARGE_RETRY: /* not in OOM situation but retry */
2389 css_put(&memcg->css);
2392 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2393 css_put(&memcg->css);
2395 case CHARGE_NOMEM: /* OOM routine works */
2397 css_put(&memcg->css);
2400 /* If oom, we never return -ENOMEM */
2403 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2404 css_put(&memcg->css);
2407 } while (ret != CHARGE_OK);
2409 if (batch > nr_pages)
2410 refill_stock(memcg, batch - nr_pages);
2411 css_put(&memcg->css);
2424 * Somemtimes we have to undo a charge we got by try_charge().
2425 * This function is for that and do uncharge, put css's refcnt.
2426 * gotten by try_charge().
2428 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2429 unsigned int nr_pages)
2431 if (!mem_cgroup_is_root(memcg)) {
2432 unsigned long bytes = nr_pages * PAGE_SIZE;
2434 res_counter_uncharge(&memcg->res, bytes);
2435 if (do_swap_account)
2436 res_counter_uncharge(&memcg->memsw, bytes);
2441 * A helper function to get mem_cgroup from ID. must be called under
2442 * rcu_read_lock(). The caller must check css_is_removed() or some if
2443 * it's concern. (dropping refcnt from swap can be called against removed
2446 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2448 struct cgroup_subsys_state *css;
2450 /* ID 0 is unused ID */
2453 css = css_lookup(&mem_cgroup_subsys, id);
2456 return container_of(css, struct mem_cgroup, css);
2459 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2461 struct mem_cgroup *memcg = NULL;
2462 struct page_cgroup *pc;
2466 VM_BUG_ON(!PageLocked(page));
2468 pc = lookup_page_cgroup(page);
2469 lock_page_cgroup(pc);
2470 if (PageCgroupUsed(pc)) {
2471 memcg = pc->mem_cgroup;
2472 if (memcg && !css_tryget(&memcg->css))
2474 } else if (PageSwapCache(page)) {
2475 ent.val = page_private(page);
2476 id = lookup_swap_cgroup(ent);
2478 memcg = mem_cgroup_lookup(id);
2479 if (memcg && !css_tryget(&memcg->css))
2483 unlock_page_cgroup(pc);
2487 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2489 unsigned int nr_pages,
2490 struct page_cgroup *pc,
2491 enum charge_type ctype)
2493 lock_page_cgroup(pc);
2494 if (unlikely(PageCgroupUsed(pc))) {
2495 unlock_page_cgroup(pc);
2496 __mem_cgroup_cancel_charge(memcg, nr_pages);
2500 * we don't need page_cgroup_lock about tail pages, becase they are not
2501 * accessed by any other context at this point.
2503 pc->mem_cgroup = memcg;
2505 * We access a page_cgroup asynchronously without lock_page_cgroup().
2506 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2507 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2508 * before USED bit, we need memory barrier here.
2509 * See mem_cgroup_add_lru_list(), etc.
2513 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2514 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2515 SetPageCgroupCache(pc);
2516 SetPageCgroupUsed(pc);
2518 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2519 ClearPageCgroupCache(pc);
2520 SetPageCgroupUsed(pc);
2526 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2527 unlock_page_cgroup(pc);
2529 * "charge_statistics" updated event counter. Then, check it.
2530 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2531 * if they exceeds softlimit.
2533 memcg_check_events(memcg, page);
2536 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2538 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2539 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2541 * Because tail pages are not marked as "used", set it. We're under
2542 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2544 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2546 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2547 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2548 unsigned long flags;
2550 if (mem_cgroup_disabled())
2553 * We have no races with charge/uncharge but will have races with
2554 * page state accounting.
2556 move_lock_page_cgroup(head_pc, &flags);
2558 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2559 smp_wmb(); /* see __commit_charge() */
2560 if (PageCgroupAcctLRU(head_pc)) {
2562 struct mem_cgroup_per_zone *mz;
2565 * LRU flags cannot be copied because we need to add tail
2566 *.page to LRU by generic call and our hook will be called.
2567 * We hold lru_lock, then, reduce counter directly.
2569 lru = page_lru(head);
2570 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2571 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2573 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2574 move_unlock_page_cgroup(head_pc, &flags);
2579 * mem_cgroup_move_account - move account of the page
2581 * @nr_pages: number of regular pages (>1 for huge pages)
2582 * @pc: page_cgroup of the page.
2583 * @from: mem_cgroup which the page is moved from.
2584 * @to: mem_cgroup which the page is moved to. @from != @to.
2585 * @uncharge: whether we should call uncharge and css_put against @from.
2587 * The caller must confirm following.
2588 * - page is not on LRU (isolate_page() is useful.)
2589 * - compound_lock is held when nr_pages > 1
2591 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2592 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2593 * true, this function does "uncharge" from old cgroup, but it doesn't if
2594 * @uncharge is false, so a caller should do "uncharge".
2596 static int mem_cgroup_move_account(struct page *page,
2597 unsigned int nr_pages,
2598 struct page_cgroup *pc,
2599 struct mem_cgroup *from,
2600 struct mem_cgroup *to,
2603 unsigned long flags;
2606 VM_BUG_ON(from == to);
2607 VM_BUG_ON(PageLRU(page));
2609 * The page is isolated from LRU. So, collapse function
2610 * will not handle this page. But page splitting can happen.
2611 * Do this check under compound_page_lock(). The caller should
2615 if (nr_pages > 1 && !PageTransHuge(page))
2618 lock_page_cgroup(pc);
2621 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2624 move_lock_page_cgroup(pc, &flags);
2626 if (PageCgroupFileMapped(pc)) {
2627 /* Update mapped_file data for mem_cgroup */
2629 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2630 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2633 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2635 /* This is not "cancel", but cancel_charge does all we need. */
2636 __mem_cgroup_cancel_charge(from, nr_pages);
2638 /* caller should have done css_get */
2639 pc->mem_cgroup = to;
2640 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2642 * We charges against "to" which may not have any tasks. Then, "to"
2643 * can be under rmdir(). But in current implementation, caller of
2644 * this function is just force_empty() and move charge, so it's
2645 * guaranteed that "to" is never removed. So, we don't check rmdir
2648 move_unlock_page_cgroup(pc, &flags);
2651 unlock_page_cgroup(pc);
2655 memcg_check_events(to, page);
2656 memcg_check_events(from, page);
2662 * move charges to its parent.
2665 static int mem_cgroup_move_parent(struct page *page,
2666 struct page_cgroup *pc,
2667 struct mem_cgroup *child,
2670 struct cgroup *cg = child->css.cgroup;
2671 struct cgroup *pcg = cg->parent;
2672 struct mem_cgroup *parent;
2673 unsigned int nr_pages;
2674 unsigned long uninitialized_var(flags);
2682 if (!get_page_unless_zero(page))
2684 if (isolate_lru_page(page))
2687 nr_pages = hpage_nr_pages(page);
2689 parent = mem_cgroup_from_cont(pcg);
2690 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2695 flags = compound_lock_irqsave(page);
2697 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2699 __mem_cgroup_cancel_charge(parent, nr_pages);
2702 compound_unlock_irqrestore(page, flags);
2704 putback_lru_page(page);
2712 * Charge the memory controller for page usage.
2714 * 0 if the charge was successful
2715 * < 0 if the cgroup is over its limit
2717 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2718 gfp_t gfp_mask, enum charge_type ctype)
2720 struct mem_cgroup *memcg = NULL;
2721 unsigned int nr_pages = 1;
2722 struct page_cgroup *pc;
2726 if (PageTransHuge(page)) {
2727 nr_pages <<= compound_order(page);
2728 VM_BUG_ON(!PageTransHuge(page));
2730 * Never OOM-kill a process for a huge page. The
2731 * fault handler will fall back to regular pages.
2736 pc = lookup_page_cgroup(page);
2737 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2739 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2743 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2747 int mem_cgroup_newpage_charge(struct page *page,
2748 struct mm_struct *mm, gfp_t gfp_mask)
2750 if (mem_cgroup_disabled())
2753 * If already mapped, we don't have to account.
2754 * If page cache, page->mapping has address_space.
2755 * But page->mapping may have out-of-use anon_vma pointer,
2756 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2759 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2763 return mem_cgroup_charge_common(page, mm, gfp_mask,
2764 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2768 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2769 enum charge_type ctype);
2772 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2773 enum charge_type ctype)
2775 struct page_cgroup *pc = lookup_page_cgroup(page);
2777 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2778 * is already on LRU. It means the page may on some other page_cgroup's
2779 * LRU. Take care of it.
2781 mem_cgroup_lru_del_before_commit(page);
2782 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2783 mem_cgroup_lru_add_after_commit(page);
2787 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2790 struct mem_cgroup *memcg = NULL;
2793 if (mem_cgroup_disabled())
2795 if (PageCompound(page))
2801 if (page_is_file_cache(page)) {
2802 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2807 * FUSE reuses pages without going through the final
2808 * put that would remove them from the LRU list, make
2809 * sure that they get relinked properly.
2811 __mem_cgroup_commit_charge_lrucare(page, memcg,
2812 MEM_CGROUP_CHARGE_TYPE_CACHE);
2816 if (PageSwapCache(page)) {
2817 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2819 __mem_cgroup_commit_charge_swapin(page, memcg,
2820 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2822 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2823 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2829 * While swap-in, try_charge -> commit or cancel, the page is locked.
2830 * And when try_charge() successfully returns, one refcnt to memcg without
2831 * struct page_cgroup is acquired. This refcnt will be consumed by
2832 * "commit()" or removed by "cancel()"
2834 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2836 gfp_t mask, struct mem_cgroup **ptr)
2838 struct mem_cgroup *memcg;
2843 if (mem_cgroup_disabled())
2846 if (!do_swap_account)
2849 * A racing thread's fault, or swapoff, may have already updated
2850 * the pte, and even removed page from swap cache: in those cases
2851 * do_swap_page()'s pte_same() test will fail; but there's also a
2852 * KSM case which does need to charge the page.
2854 if (!PageSwapCache(page))
2856 memcg = try_get_mem_cgroup_from_page(page);
2860 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2861 css_put(&memcg->css);
2866 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2870 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2871 enum charge_type ctype)
2873 if (mem_cgroup_disabled())
2877 cgroup_exclude_rmdir(&ptr->css);
2879 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2881 * Now swap is on-memory. This means this page may be
2882 * counted both as mem and swap....double count.
2883 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2884 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2885 * may call delete_from_swap_cache() before reach here.
2887 if (do_swap_account && PageSwapCache(page)) {
2888 swp_entry_t ent = {.val = page_private(page)};
2890 struct mem_cgroup *memcg;
2892 id = swap_cgroup_record(ent, 0);
2894 memcg = mem_cgroup_lookup(id);
2897 * This recorded memcg can be obsolete one. So, avoid
2898 * calling css_tryget
2900 if (!mem_cgroup_is_root(memcg))
2901 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2902 mem_cgroup_swap_statistics(memcg, false);
2903 mem_cgroup_put(memcg);
2908 * At swapin, we may charge account against cgroup which has no tasks.
2909 * So, rmdir()->pre_destroy() can be called while we do this charge.
2910 * In that case, we need to call pre_destroy() again. check it here.
2912 cgroup_release_and_wakeup_rmdir(&ptr->css);
2915 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2917 __mem_cgroup_commit_charge_swapin(page, ptr,
2918 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2921 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2923 if (mem_cgroup_disabled())
2927 __mem_cgroup_cancel_charge(memcg, 1);
2930 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2931 unsigned int nr_pages,
2932 const enum charge_type ctype)
2934 struct memcg_batch_info *batch = NULL;
2935 bool uncharge_memsw = true;
2937 /* If swapout, usage of swap doesn't decrease */
2938 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2939 uncharge_memsw = false;
2941 batch = ¤t->memcg_batch;
2943 * In usual, we do css_get() when we remember memcg pointer.
2944 * But in this case, we keep res->usage until end of a series of
2945 * uncharges. Then, it's ok to ignore memcg's refcnt.
2948 batch->memcg = memcg;
2950 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2951 * In those cases, all pages freed continuously can be expected to be in
2952 * the same cgroup and we have chance to coalesce uncharges.
2953 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2954 * because we want to do uncharge as soon as possible.
2957 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2958 goto direct_uncharge;
2961 goto direct_uncharge;
2964 * In typical case, batch->memcg == mem. This means we can
2965 * merge a series of uncharges to an uncharge of res_counter.
2966 * If not, we uncharge res_counter ony by one.
2968 if (batch->memcg != memcg)
2969 goto direct_uncharge;
2970 /* remember freed charge and uncharge it later */
2973 batch->memsw_nr_pages++;
2976 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2978 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2979 if (unlikely(batch->memcg != memcg))
2980 memcg_oom_recover(memcg);
2985 * uncharge if !page_mapped(page)
2987 static struct mem_cgroup *
2988 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2990 struct mem_cgroup *memcg = NULL;
2991 unsigned int nr_pages = 1;
2992 struct page_cgroup *pc;
2994 if (mem_cgroup_disabled())
2997 if (PageSwapCache(page))
3000 if (PageTransHuge(page)) {
3001 nr_pages <<= compound_order(page);
3002 VM_BUG_ON(!PageTransHuge(page));
3005 * Check if our page_cgroup is valid
3007 pc = lookup_page_cgroup(page);
3008 if (unlikely(!pc || !PageCgroupUsed(pc)))
3011 lock_page_cgroup(pc);
3013 memcg = pc->mem_cgroup;
3015 if (!PageCgroupUsed(pc))
3019 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3020 case MEM_CGROUP_CHARGE_TYPE_DROP:
3021 /* See mem_cgroup_prepare_migration() */
3022 if (page_mapped(page) || PageCgroupMigration(pc))
3025 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3026 if (!PageAnon(page)) { /* Shared memory */
3027 if (page->mapping && !page_is_file_cache(page))
3029 } else if (page_mapped(page)) /* Anon */
3036 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3038 ClearPageCgroupUsed(pc);
3040 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3041 * freed from LRU. This is safe because uncharged page is expected not
3042 * to be reused (freed soon). Exception is SwapCache, it's handled by
3043 * special functions.
3046 unlock_page_cgroup(pc);
3048 * even after unlock, we have memcg->res.usage here and this memcg
3049 * will never be freed.
3051 memcg_check_events(memcg, page);
3052 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3053 mem_cgroup_swap_statistics(memcg, true);
3054 mem_cgroup_get(memcg);
3056 if (!mem_cgroup_is_root(memcg))
3057 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3062 unlock_page_cgroup(pc);
3066 void mem_cgroup_uncharge_page(struct page *page)
3069 if (page_mapped(page))
3071 if (page->mapping && !PageAnon(page))
3073 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3076 void mem_cgroup_uncharge_cache_page(struct page *page)
3078 VM_BUG_ON(page_mapped(page));
3079 VM_BUG_ON(page->mapping);
3080 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3084 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3085 * In that cases, pages are freed continuously and we can expect pages
3086 * are in the same memcg. All these calls itself limits the number of
3087 * pages freed at once, then uncharge_start/end() is called properly.
3088 * This may be called prural(2) times in a context,
3091 void mem_cgroup_uncharge_start(void)
3093 current->memcg_batch.do_batch++;
3094 /* We can do nest. */
3095 if (current->memcg_batch.do_batch == 1) {
3096 current->memcg_batch.memcg = NULL;
3097 current->memcg_batch.nr_pages = 0;
3098 current->memcg_batch.memsw_nr_pages = 0;
3102 void mem_cgroup_uncharge_end(void)
3104 struct memcg_batch_info *batch = ¤t->memcg_batch;
3106 if (!batch->do_batch)
3110 if (batch->do_batch) /* If stacked, do nothing. */
3116 * This "batch->memcg" is valid without any css_get/put etc...
3117 * bacause we hide charges behind us.
3119 if (batch->nr_pages)
3120 res_counter_uncharge(&batch->memcg->res,
3121 batch->nr_pages * PAGE_SIZE);
3122 if (batch->memsw_nr_pages)
3123 res_counter_uncharge(&batch->memcg->memsw,
3124 batch->memsw_nr_pages * PAGE_SIZE);
3125 memcg_oom_recover(batch->memcg);
3126 /* forget this pointer (for sanity check) */
3127 batch->memcg = NULL;
3132 * called after __delete_from_swap_cache() and drop "page" account.
3133 * memcg information is recorded to swap_cgroup of "ent"
3136 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3138 struct mem_cgroup *memcg;
3139 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3141 if (!swapout) /* this was a swap cache but the swap is unused ! */
3142 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3144 memcg = __mem_cgroup_uncharge_common(page, ctype);
3147 * record memcg information, if swapout && memcg != NULL,
3148 * mem_cgroup_get() was called in uncharge().
3150 if (do_swap_account && swapout && memcg)
3151 swap_cgroup_record(ent, css_id(&memcg->css));
3155 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3157 * called from swap_entry_free(). remove record in swap_cgroup and
3158 * uncharge "memsw" account.
3160 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3162 struct mem_cgroup *memcg;
3165 if (!do_swap_account)
3168 id = swap_cgroup_record(ent, 0);
3170 memcg = mem_cgroup_lookup(id);
3173 * We uncharge this because swap is freed.
3174 * This memcg can be obsolete one. We avoid calling css_tryget
3176 if (!mem_cgroup_is_root(memcg))
3177 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3178 mem_cgroup_swap_statistics(memcg, false);
3179 mem_cgroup_put(memcg);
3185 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3186 * @entry: swap entry to be moved
3187 * @from: mem_cgroup which the entry is moved from
3188 * @to: mem_cgroup which the entry is moved to
3189 * @need_fixup: whether we should fixup res_counters and refcounts.
3191 * It succeeds only when the swap_cgroup's record for this entry is the same
3192 * as the mem_cgroup's id of @from.
3194 * Returns 0 on success, -EINVAL on failure.
3196 * The caller must have charged to @to, IOW, called res_counter_charge() about
3197 * both res and memsw, and called css_get().
3199 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3200 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3202 unsigned short old_id, new_id;
3204 old_id = css_id(&from->css);
3205 new_id = css_id(&to->css);
3207 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3208 mem_cgroup_swap_statistics(from, false);
3209 mem_cgroup_swap_statistics(to, true);
3211 * This function is only called from task migration context now.
3212 * It postpones res_counter and refcount handling till the end
3213 * of task migration(mem_cgroup_clear_mc()) for performance
3214 * improvement. But we cannot postpone mem_cgroup_get(to)
3215 * because if the process that has been moved to @to does
3216 * swap-in, the refcount of @to might be decreased to 0.
3220 if (!mem_cgroup_is_root(from))
3221 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3222 mem_cgroup_put(from);
3224 * we charged both to->res and to->memsw, so we should
3227 if (!mem_cgroup_is_root(to))
3228 res_counter_uncharge(&to->res, PAGE_SIZE);
3235 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3236 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3243 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3246 int mem_cgroup_prepare_migration(struct page *page,
3247 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3249 struct mem_cgroup *memcg = NULL;
3250 struct page_cgroup *pc;
3251 enum charge_type ctype;
3256 VM_BUG_ON(PageTransHuge(page));
3257 if (mem_cgroup_disabled())
3260 pc = lookup_page_cgroup(page);
3261 lock_page_cgroup(pc);
3262 if (PageCgroupUsed(pc)) {
3263 memcg = pc->mem_cgroup;
3264 css_get(&memcg->css);
3266 * At migrating an anonymous page, its mapcount goes down
3267 * to 0 and uncharge() will be called. But, even if it's fully
3268 * unmapped, migration may fail and this page has to be
3269 * charged again. We set MIGRATION flag here and delay uncharge
3270 * until end_migration() is called
3272 * Corner Case Thinking
3274 * When the old page was mapped as Anon and it's unmap-and-freed
3275 * while migration was ongoing.
3276 * If unmap finds the old page, uncharge() of it will be delayed
3277 * until end_migration(). If unmap finds a new page, it's
3278 * uncharged when it make mapcount to be 1->0. If unmap code
3279 * finds swap_migration_entry, the new page will not be mapped
3280 * and end_migration() will find it(mapcount==0).
3283 * When the old page was mapped but migraion fails, the kernel
3284 * remaps it. A charge for it is kept by MIGRATION flag even
3285 * if mapcount goes down to 0. We can do remap successfully
3286 * without charging it again.
3289 * The "old" page is under lock_page() until the end of
3290 * migration, so, the old page itself will not be swapped-out.
3291 * If the new page is swapped out before end_migraton, our
3292 * hook to usual swap-out path will catch the event.
3295 SetPageCgroupMigration(pc);
3297 unlock_page_cgroup(pc);
3299 * If the page is not charged at this point,
3306 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3307 css_put(&memcg->css);/* drop extra refcnt */
3308 if (ret || *ptr == NULL) {
3309 if (PageAnon(page)) {
3310 lock_page_cgroup(pc);
3311 ClearPageCgroupMigration(pc);
3312 unlock_page_cgroup(pc);
3314 * The old page may be fully unmapped while we kept it.
3316 mem_cgroup_uncharge_page(page);
3321 * We charge new page before it's used/mapped. So, even if unlock_page()
3322 * is called before end_migration, we can catch all events on this new
3323 * page. In the case new page is migrated but not remapped, new page's
3324 * mapcount will be finally 0 and we call uncharge in end_migration().
3326 pc = lookup_page_cgroup(newpage);
3328 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3329 else if (page_is_file_cache(page))
3330 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3332 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3333 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3337 /* remove redundant charge if migration failed*/
3338 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3339 struct page *oldpage, struct page *newpage, bool migration_ok)
3341 struct page *used, *unused;
3342 struct page_cgroup *pc;
3346 /* blocks rmdir() */
3347 cgroup_exclude_rmdir(&memcg->css);
3348 if (!migration_ok) {
3356 * We disallowed uncharge of pages under migration because mapcount
3357 * of the page goes down to zero, temporarly.
3358 * Clear the flag and check the page should be charged.
3360 pc = lookup_page_cgroup(oldpage);
3361 lock_page_cgroup(pc);
3362 ClearPageCgroupMigration(pc);
3363 unlock_page_cgroup(pc);
3365 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3368 * If a page is a file cache, radix-tree replacement is very atomic
3369 * and we can skip this check. When it was an Anon page, its mapcount
3370 * goes down to 0. But because we added MIGRATION flage, it's not
3371 * uncharged yet. There are several case but page->mapcount check
3372 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3373 * check. (see prepare_charge() also)
3376 mem_cgroup_uncharge_page(used);
3378 * At migration, we may charge account against cgroup which has no
3380 * So, rmdir()->pre_destroy() can be called while we do this charge.
3381 * In that case, we need to call pre_destroy() again. check it here.
3383 cgroup_release_and_wakeup_rmdir(&memcg->css);
3387 * At replace page cache, newpage is not under any memcg but it's on
3388 * LRU. So, this function doesn't touch res_counter but handles LRU
3389 * in correct way. Both pages are locked so we cannot race with uncharge.
3391 void mem_cgroup_replace_page_cache(struct page *oldpage,
3392 struct page *newpage)
3394 struct mem_cgroup *memcg;
3395 struct page_cgroup *pc;
3397 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3398 unsigned long flags;
3400 if (mem_cgroup_disabled())
3403 pc = lookup_page_cgroup(oldpage);
3404 /* fix accounting on old pages */
3405 lock_page_cgroup(pc);
3406 memcg = pc->mem_cgroup;
3407 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1);
3408 ClearPageCgroupUsed(pc);
3409 unlock_page_cgroup(pc);
3411 if (PageSwapBacked(oldpage))
3412 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3414 zone = page_zone(newpage);
3415 pc = lookup_page_cgroup(newpage);
3417 * Even if newpage->mapping was NULL before starting replacement,
3418 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3419 * LRU while we overwrite pc->mem_cgroup.
3421 spin_lock_irqsave(&zone->lru_lock, flags);
3422 if (PageLRU(newpage))
3423 del_page_from_lru_list(zone, newpage, page_lru(newpage));
3424 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type);
3425 if (PageLRU(newpage))
3426 add_page_to_lru_list(zone, newpage, page_lru(newpage));
3427 spin_unlock_irqrestore(&zone->lru_lock, flags);
3430 #ifdef CONFIG_DEBUG_VM
3431 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3433 struct page_cgroup *pc;
3435 pc = lookup_page_cgroup(page);
3436 if (likely(pc) && PageCgroupUsed(pc))
3441 bool mem_cgroup_bad_page_check(struct page *page)
3443 if (mem_cgroup_disabled())
3446 return lookup_page_cgroup_used(page) != NULL;
3449 void mem_cgroup_print_bad_page(struct page *page)
3451 struct page_cgroup *pc;
3453 pc = lookup_page_cgroup_used(page);
3458 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3459 pc, pc->flags, pc->mem_cgroup);
3461 path = kmalloc(PATH_MAX, GFP_KERNEL);
3464 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3469 printk(KERN_CONT "(%s)\n",
3470 (ret < 0) ? "cannot get the path" : path);
3476 static DEFINE_MUTEX(set_limit_mutex);
3478 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3479 unsigned long long val)
3482 u64 memswlimit, memlimit;
3484 int children = mem_cgroup_count_children(memcg);
3485 u64 curusage, oldusage;
3489 * For keeping hierarchical_reclaim simple, how long we should retry
3490 * is depends on callers. We set our retry-count to be function
3491 * of # of children which we should visit in this loop.
3493 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3495 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3498 while (retry_count) {
3499 if (signal_pending(current)) {
3504 * Rather than hide all in some function, I do this in
3505 * open coded manner. You see what this really does.
3506 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3508 mutex_lock(&set_limit_mutex);
3509 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3510 if (memswlimit < val) {
3512 mutex_unlock(&set_limit_mutex);
3516 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3520 ret = res_counter_set_limit(&memcg->res, val);
3522 if (memswlimit == val)
3523 memcg->memsw_is_minimum = true;
3525 memcg->memsw_is_minimum = false;
3527 mutex_unlock(&set_limit_mutex);
3532 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3533 MEM_CGROUP_RECLAIM_SHRINK,
3535 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3536 /* Usage is reduced ? */
3537 if (curusage >= oldusage)
3540 oldusage = curusage;
3542 if (!ret && enlarge)
3543 memcg_oom_recover(memcg);
3548 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3549 unsigned long long val)
3552 u64 memlimit, memswlimit, oldusage, curusage;
3553 int children = mem_cgroup_count_children(memcg);
3557 /* see mem_cgroup_resize_res_limit */
3558 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3559 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3560 while (retry_count) {
3561 if (signal_pending(current)) {
3566 * Rather than hide all in some function, I do this in
3567 * open coded manner. You see what this really does.
3568 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3570 mutex_lock(&set_limit_mutex);
3571 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3572 if (memlimit > val) {
3574 mutex_unlock(&set_limit_mutex);
3577 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3578 if (memswlimit < val)
3580 ret = res_counter_set_limit(&memcg->memsw, val);
3582 if (memlimit == val)
3583 memcg->memsw_is_minimum = true;
3585 memcg->memsw_is_minimum = false;
3587 mutex_unlock(&set_limit_mutex);
3592 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3593 MEM_CGROUP_RECLAIM_NOSWAP |
3594 MEM_CGROUP_RECLAIM_SHRINK,
3596 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3597 /* Usage is reduced ? */
3598 if (curusage >= oldusage)
3601 oldusage = curusage;
3603 if (!ret && enlarge)
3604 memcg_oom_recover(memcg);
3608 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3610 unsigned long *total_scanned)
3612 unsigned long nr_reclaimed = 0;
3613 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3614 unsigned long reclaimed;
3616 struct mem_cgroup_tree_per_zone *mctz;
3617 unsigned long long excess;
3618 unsigned long nr_scanned;
3623 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3625 * This loop can run a while, specially if mem_cgroup's continuously
3626 * keep exceeding their soft limit and putting the system under
3633 mz = mem_cgroup_largest_soft_limit_node(mctz);
3638 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3640 MEM_CGROUP_RECLAIM_SOFT,
3642 nr_reclaimed += reclaimed;
3643 *total_scanned += nr_scanned;
3644 spin_lock(&mctz->lock);
3647 * If we failed to reclaim anything from this memory cgroup
3648 * it is time to move on to the next cgroup
3654 * Loop until we find yet another one.
3656 * By the time we get the soft_limit lock
3657 * again, someone might have aded the
3658 * group back on the RB tree. Iterate to
3659 * make sure we get a different mem.
3660 * mem_cgroup_largest_soft_limit_node returns
3661 * NULL if no other cgroup is present on
3665 __mem_cgroup_largest_soft_limit_node(mctz);
3667 css_put(&next_mz->mem->css);
3668 else /* next_mz == NULL or other memcg */
3672 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3673 excess = res_counter_soft_limit_excess(&mz->mem->res);
3675 * One school of thought says that we should not add
3676 * back the node to the tree if reclaim returns 0.
3677 * But our reclaim could return 0, simply because due
3678 * to priority we are exposing a smaller subset of
3679 * memory to reclaim from. Consider this as a longer
3682 /* If excess == 0, no tree ops */
3683 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3684 spin_unlock(&mctz->lock);
3685 css_put(&mz->mem->css);
3688 * Could not reclaim anything and there are no more
3689 * mem cgroups to try or we seem to be looping without
3690 * reclaiming anything.
3692 if (!nr_reclaimed &&
3694 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3696 } while (!nr_reclaimed);
3698 css_put(&next_mz->mem->css);
3699 return nr_reclaimed;
3703 * This routine traverse page_cgroup in given list and drop them all.
3704 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3706 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3707 int node, int zid, enum lru_list lru)
3710 struct mem_cgroup_per_zone *mz;
3711 struct page_cgroup *pc, *busy;
3712 unsigned long flags, loop;
3713 struct list_head *list;
3716 zone = &NODE_DATA(node)->node_zones[zid];
3717 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3718 list = &mz->lists[lru];
3720 loop = MEM_CGROUP_ZSTAT(mz, lru);
3721 /* give some margin against EBUSY etc...*/
3728 spin_lock_irqsave(&zone->lru_lock, flags);
3729 if (list_empty(list)) {
3730 spin_unlock_irqrestore(&zone->lru_lock, flags);
3733 pc = list_entry(list->prev, struct page_cgroup, lru);
3735 list_move(&pc->lru, list);
3737 spin_unlock_irqrestore(&zone->lru_lock, flags);
3740 spin_unlock_irqrestore(&zone->lru_lock, flags);
3742 page = lookup_cgroup_page(pc);
3744 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3748 if (ret == -EBUSY || ret == -EINVAL) {
3749 /* found lock contention or "pc" is obsolete. */
3756 if (!ret && !list_empty(list))
3762 * make mem_cgroup's charge to be 0 if there is no task.
3763 * This enables deleting this mem_cgroup.
3765 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3768 int node, zid, shrink;
3769 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3770 struct cgroup *cgrp = memcg->css.cgroup;
3772 css_get(&memcg->css);
3775 /* should free all ? */
3781 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3784 if (signal_pending(current))
3786 /* This is for making all *used* pages to be on LRU. */
3787 lru_add_drain_all();
3788 drain_all_stock_sync(memcg);
3790 mem_cgroup_start_move(memcg);
3791 for_each_node_state(node, N_HIGH_MEMORY) {
3792 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3795 ret = mem_cgroup_force_empty_list(memcg,
3804 mem_cgroup_end_move(memcg);
3805 memcg_oom_recover(memcg);
3806 /* it seems parent cgroup doesn't have enough mem */
3810 /* "ret" should also be checked to ensure all lists are empty. */
3811 } while (memcg->res.usage > 0 || ret);
3813 css_put(&memcg->css);
3817 /* returns EBUSY if there is a task or if we come here twice. */
3818 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3822 /* we call try-to-free pages for make this cgroup empty */
3823 lru_add_drain_all();
3824 /* try to free all pages in this cgroup */
3826 while (nr_retries && memcg->res.usage > 0) {
3829 if (signal_pending(current)) {
3833 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3837 /* maybe some writeback is necessary */
3838 congestion_wait(BLK_RW_ASYNC, HZ/10);
3843 /* try move_account...there may be some *locked* pages. */
3847 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3849 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3853 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3855 return mem_cgroup_from_cont(cont)->use_hierarchy;
3858 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3862 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3863 struct cgroup *parent = cont->parent;
3864 struct mem_cgroup *parent_memcg = NULL;
3867 parent_memcg = mem_cgroup_from_cont(parent);
3871 * If parent's use_hierarchy is set, we can't make any modifications
3872 * in the child subtrees. If it is unset, then the change can
3873 * occur, provided the current cgroup has no children.
3875 * For the root cgroup, parent_mem is NULL, we allow value to be
3876 * set if there are no children.
3878 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3879 (val == 1 || val == 0)) {
3880 if (list_empty(&cont->children))
3881 memcg->use_hierarchy = val;
3892 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3893 enum mem_cgroup_stat_index idx)
3895 struct mem_cgroup *iter;
3898 /* Per-cpu values can be negative, use a signed accumulator */
3899 for_each_mem_cgroup_tree(iter, memcg)
3900 val += mem_cgroup_read_stat(iter, idx);
3902 if (val < 0) /* race ? */
3907 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3911 if (!mem_cgroup_is_root(memcg)) {
3913 return res_counter_read_u64(&memcg->res, RES_USAGE);
3915 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3918 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3919 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3922 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3924 return val << PAGE_SHIFT;
3927 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3929 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3933 type = MEMFILE_TYPE(cft->private);
3934 name = MEMFILE_ATTR(cft->private);
3937 if (name == RES_USAGE)
3938 val = mem_cgroup_usage(memcg, false);
3940 val = res_counter_read_u64(&memcg->res, name);
3943 if (name == RES_USAGE)
3944 val = mem_cgroup_usage(memcg, true);
3946 val = res_counter_read_u64(&memcg->memsw, name);
3955 * The user of this function is...
3958 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3961 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3963 unsigned long long val;
3966 type = MEMFILE_TYPE(cft->private);
3967 name = MEMFILE_ATTR(cft->private);
3970 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3974 /* This function does all necessary parse...reuse it */
3975 ret = res_counter_memparse_write_strategy(buffer, &val);
3979 ret = mem_cgroup_resize_limit(memcg, val);
3981 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3983 case RES_SOFT_LIMIT:
3984 ret = res_counter_memparse_write_strategy(buffer, &val);
3988 * For memsw, soft limits are hard to implement in terms
3989 * of semantics, for now, we support soft limits for
3990 * control without swap
3993 ret = res_counter_set_soft_limit(&memcg->res, val);
3998 ret = -EINVAL; /* should be BUG() ? */
4004 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4005 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4007 struct cgroup *cgroup;
4008 unsigned long long min_limit, min_memsw_limit, tmp;
4010 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4011 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4012 cgroup = memcg->css.cgroup;
4013 if (!memcg->use_hierarchy)
4016 while (cgroup->parent) {
4017 cgroup = cgroup->parent;
4018 memcg = mem_cgroup_from_cont(cgroup);
4019 if (!memcg->use_hierarchy)
4021 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4022 min_limit = min(min_limit, tmp);
4023 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4024 min_memsw_limit = min(min_memsw_limit, tmp);
4027 *mem_limit = min_limit;
4028 *memsw_limit = min_memsw_limit;
4032 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4034 struct mem_cgroup *memcg;
4037 memcg = mem_cgroup_from_cont(cont);
4038 type = MEMFILE_TYPE(event);
4039 name = MEMFILE_ATTR(event);
4043 res_counter_reset_max(&memcg->res);
4045 res_counter_reset_max(&memcg->memsw);
4049 res_counter_reset_failcnt(&memcg->res);
4051 res_counter_reset_failcnt(&memcg->memsw);
4058 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4061 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4065 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4066 struct cftype *cft, u64 val)
4068 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4070 if (val >= (1 << NR_MOVE_TYPE))
4073 * We check this value several times in both in can_attach() and
4074 * attach(), so we need cgroup lock to prevent this value from being
4078 memcg->move_charge_at_immigrate = val;
4084 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4085 struct cftype *cft, u64 val)
4092 /* For read statistics */
4110 struct mcs_total_stat {
4111 s64 stat[NR_MCS_STAT];
4117 } memcg_stat_strings[NR_MCS_STAT] = {
4118 {"cache", "total_cache"},
4119 {"rss", "total_rss"},
4120 {"mapped_file", "total_mapped_file"},
4121 {"pgpgin", "total_pgpgin"},
4122 {"pgpgout", "total_pgpgout"},
4123 {"swap", "total_swap"},
4124 {"pgfault", "total_pgfault"},
4125 {"pgmajfault", "total_pgmajfault"},
4126 {"inactive_anon", "total_inactive_anon"},
4127 {"active_anon", "total_active_anon"},
4128 {"inactive_file", "total_inactive_file"},
4129 {"active_file", "total_active_file"},
4130 {"unevictable", "total_unevictable"}
4135 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4140 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4141 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4142 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4143 s->stat[MCS_RSS] += val * PAGE_SIZE;
4144 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4145 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4146 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4147 s->stat[MCS_PGPGIN] += val;
4148 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4149 s->stat[MCS_PGPGOUT] += val;
4150 if (do_swap_account) {
4151 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4152 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4154 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4155 s->stat[MCS_PGFAULT] += val;
4156 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4157 s->stat[MCS_PGMAJFAULT] += val;
4160 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4161 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4162 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4163 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4164 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4165 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4166 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4167 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4168 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4169 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4173 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4175 struct mem_cgroup *iter;
4177 for_each_mem_cgroup_tree(iter, memcg)
4178 mem_cgroup_get_local_stat(iter, s);
4182 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4185 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4186 unsigned long node_nr;
4187 struct cgroup *cont = m->private;
4188 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4190 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4191 seq_printf(m, "total=%lu", total_nr);
4192 for_each_node_state(nid, N_HIGH_MEMORY) {
4193 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4194 seq_printf(m, " N%d=%lu", nid, node_nr);
4198 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4199 seq_printf(m, "file=%lu", file_nr);
4200 for_each_node_state(nid, N_HIGH_MEMORY) {
4201 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4203 seq_printf(m, " N%d=%lu", nid, node_nr);
4207 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4208 seq_printf(m, "anon=%lu", anon_nr);
4209 for_each_node_state(nid, N_HIGH_MEMORY) {
4210 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4212 seq_printf(m, " N%d=%lu", nid, node_nr);
4216 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4217 seq_printf(m, "unevictable=%lu", unevictable_nr);
4218 for_each_node_state(nid, N_HIGH_MEMORY) {
4219 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4220 BIT(LRU_UNEVICTABLE));
4221 seq_printf(m, " N%d=%lu", nid, node_nr);
4226 #endif /* CONFIG_NUMA */
4228 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4229 struct cgroup_map_cb *cb)
4231 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4232 struct mcs_total_stat mystat;
4235 memset(&mystat, 0, sizeof(mystat));
4236 mem_cgroup_get_local_stat(mem_cont, &mystat);
4239 for (i = 0; i < NR_MCS_STAT; i++) {
4240 if (i == MCS_SWAP && !do_swap_account)
4242 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4245 /* Hierarchical information */
4247 unsigned long long limit, memsw_limit;
4248 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4249 cb->fill(cb, "hierarchical_memory_limit", limit);
4250 if (do_swap_account)
4251 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4254 memset(&mystat, 0, sizeof(mystat));
4255 mem_cgroup_get_total_stat(mem_cont, &mystat);
4256 for (i = 0; i < NR_MCS_STAT; i++) {
4257 if (i == MCS_SWAP && !do_swap_account)
4259 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4262 #ifdef CONFIG_DEBUG_VM
4265 struct mem_cgroup_per_zone *mz;
4266 unsigned long recent_rotated[2] = {0, 0};
4267 unsigned long recent_scanned[2] = {0, 0};
4269 for_each_online_node(nid)
4270 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4271 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4273 recent_rotated[0] +=
4274 mz->reclaim_stat.recent_rotated[0];
4275 recent_rotated[1] +=
4276 mz->reclaim_stat.recent_rotated[1];
4277 recent_scanned[0] +=
4278 mz->reclaim_stat.recent_scanned[0];
4279 recent_scanned[1] +=
4280 mz->reclaim_stat.recent_scanned[1];
4282 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4283 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4284 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4285 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4292 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4294 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4296 return mem_cgroup_swappiness(memcg);
4299 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4302 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4303 struct mem_cgroup *parent;
4308 if (cgrp->parent == NULL)
4311 parent = mem_cgroup_from_cont(cgrp->parent);
4315 /* If under hierarchy, only empty-root can set this value */
4316 if ((parent->use_hierarchy) ||
4317 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4322 memcg->swappiness = val;
4329 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4331 struct mem_cgroup_threshold_ary *t;
4337 t = rcu_dereference(memcg->thresholds.primary);
4339 t = rcu_dereference(memcg->memsw_thresholds.primary);
4344 usage = mem_cgroup_usage(memcg, swap);
4347 * current_threshold points to threshold just below usage.
4348 * If it's not true, a threshold was crossed after last
4349 * call of __mem_cgroup_threshold().
4351 i = t->current_threshold;
4354 * Iterate backward over array of thresholds starting from
4355 * current_threshold and check if a threshold is crossed.
4356 * If none of thresholds below usage is crossed, we read
4357 * only one element of the array here.
4359 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4360 eventfd_signal(t->entries[i].eventfd, 1);
4362 /* i = current_threshold + 1 */
4366 * Iterate forward over array of thresholds starting from
4367 * current_threshold+1 and check if a threshold is crossed.
4368 * If none of thresholds above usage is crossed, we read
4369 * only one element of the array here.
4371 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4372 eventfd_signal(t->entries[i].eventfd, 1);
4374 /* Update current_threshold */
4375 t->current_threshold = i - 1;
4380 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4383 __mem_cgroup_threshold(memcg, false);
4384 if (do_swap_account)
4385 __mem_cgroup_threshold(memcg, true);
4387 memcg = parent_mem_cgroup(memcg);
4391 static int compare_thresholds(const void *a, const void *b)
4393 const struct mem_cgroup_threshold *_a = a;
4394 const struct mem_cgroup_threshold *_b = b;
4396 return _a->threshold - _b->threshold;
4399 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4401 struct mem_cgroup_eventfd_list *ev;
4403 list_for_each_entry(ev, &memcg->oom_notify, list)
4404 eventfd_signal(ev->eventfd, 1);
4408 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4410 struct mem_cgroup *iter;
4412 for_each_mem_cgroup_tree(iter, memcg)
4413 mem_cgroup_oom_notify_cb(iter);
4416 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4417 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4419 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4420 struct mem_cgroup_thresholds *thresholds;
4421 struct mem_cgroup_threshold_ary *new;
4422 int type = MEMFILE_TYPE(cft->private);
4423 u64 threshold, usage;
4426 ret = res_counter_memparse_write_strategy(args, &threshold);
4430 mutex_lock(&memcg->thresholds_lock);
4433 thresholds = &memcg->thresholds;
4434 else if (type == _MEMSWAP)
4435 thresholds = &memcg->memsw_thresholds;
4439 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4441 /* Check if a threshold crossed before adding a new one */
4442 if (thresholds->primary)
4443 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4445 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4447 /* Allocate memory for new array of thresholds */
4448 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4456 /* Copy thresholds (if any) to new array */
4457 if (thresholds->primary) {
4458 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4459 sizeof(struct mem_cgroup_threshold));
4462 /* Add new threshold */
4463 new->entries[size - 1].eventfd = eventfd;
4464 new->entries[size - 1].threshold = threshold;
4466 /* Sort thresholds. Registering of new threshold isn't time-critical */
4467 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4468 compare_thresholds, NULL);
4470 /* Find current threshold */
4471 new->current_threshold = -1;
4472 for (i = 0; i < size; i++) {
4473 if (new->entries[i].threshold < usage) {
4475 * new->current_threshold will not be used until
4476 * rcu_assign_pointer(), so it's safe to increment
4479 ++new->current_threshold;
4483 /* Free old spare buffer and save old primary buffer as spare */
4484 kfree(thresholds->spare);
4485 thresholds->spare = thresholds->primary;
4487 rcu_assign_pointer(thresholds->primary, new);
4489 /* To be sure that nobody uses thresholds */
4493 mutex_unlock(&memcg->thresholds_lock);
4498 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4499 struct cftype *cft, struct eventfd_ctx *eventfd)
4501 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4502 struct mem_cgroup_thresholds *thresholds;
4503 struct mem_cgroup_threshold_ary *new;
4504 int type = MEMFILE_TYPE(cft->private);
4508 mutex_lock(&memcg->thresholds_lock);
4510 thresholds = &memcg->thresholds;
4511 else if (type == _MEMSWAP)
4512 thresholds = &memcg->memsw_thresholds;
4517 * Something went wrong if we trying to unregister a threshold
4518 * if we don't have thresholds
4520 BUG_ON(!thresholds);
4522 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4524 /* Check if a threshold crossed before removing */
4525 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4527 /* Calculate new number of threshold */
4529 for (i = 0; i < thresholds->primary->size; i++) {
4530 if (thresholds->primary->entries[i].eventfd != eventfd)
4534 new = thresholds->spare;
4536 /* Set thresholds array to NULL if we don't have thresholds */
4545 /* Copy thresholds and find current threshold */
4546 new->current_threshold = -1;
4547 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4548 if (thresholds->primary->entries[i].eventfd == eventfd)
4551 new->entries[j] = thresholds->primary->entries[i];
4552 if (new->entries[j].threshold < usage) {
4554 * new->current_threshold will not be used
4555 * until rcu_assign_pointer(), so it's safe to increment
4558 ++new->current_threshold;
4564 /* Swap primary and spare array */
4565 thresholds->spare = thresholds->primary;
4566 rcu_assign_pointer(thresholds->primary, new);
4568 /* To be sure that nobody uses thresholds */
4571 mutex_unlock(&memcg->thresholds_lock);
4574 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4575 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4577 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4578 struct mem_cgroup_eventfd_list *event;
4579 int type = MEMFILE_TYPE(cft->private);
4581 BUG_ON(type != _OOM_TYPE);
4582 event = kmalloc(sizeof(*event), GFP_KERNEL);
4586 spin_lock(&memcg_oom_lock);
4588 event->eventfd = eventfd;
4589 list_add(&event->list, &memcg->oom_notify);
4591 /* already in OOM ? */
4592 if (atomic_read(&memcg->under_oom))
4593 eventfd_signal(eventfd, 1);
4594 spin_unlock(&memcg_oom_lock);
4599 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4600 struct cftype *cft, struct eventfd_ctx *eventfd)
4602 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4603 struct mem_cgroup_eventfd_list *ev, *tmp;
4604 int type = MEMFILE_TYPE(cft->private);
4606 BUG_ON(type != _OOM_TYPE);
4608 spin_lock(&memcg_oom_lock);
4610 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4611 if (ev->eventfd == eventfd) {
4612 list_del(&ev->list);
4617 spin_unlock(&memcg_oom_lock);
4620 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4621 struct cftype *cft, struct cgroup_map_cb *cb)
4623 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4625 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4627 if (atomic_read(&memcg->under_oom))
4628 cb->fill(cb, "under_oom", 1);
4630 cb->fill(cb, "under_oom", 0);
4634 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4635 struct cftype *cft, u64 val)
4637 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4638 struct mem_cgroup *parent;
4640 /* cannot set to root cgroup and only 0 and 1 are allowed */
4641 if (!cgrp->parent || !((val == 0) || (val == 1)))
4644 parent = mem_cgroup_from_cont(cgrp->parent);
4647 /* oom-kill-disable is a flag for subhierarchy. */
4648 if ((parent->use_hierarchy) ||
4649 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4653 memcg->oom_kill_disable = val;
4655 memcg_oom_recover(memcg);
4661 static const struct file_operations mem_control_numa_stat_file_operations = {
4663 .llseek = seq_lseek,
4664 .release = single_release,
4667 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4669 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4671 file->f_op = &mem_control_numa_stat_file_operations;
4672 return single_open(file, mem_control_numa_stat_show, cont);
4674 #endif /* CONFIG_NUMA */
4676 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4677 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4680 * Part of this would be better living in a separate allocation
4681 * function, leaving us with just the cgroup tree population work.
4682 * We, however, depend on state such as network's proto_list that
4683 * is only initialized after cgroup creation. I found the less
4684 * cumbersome way to deal with it to defer it all to populate time
4686 return mem_cgroup_sockets_init(cont, ss);
4689 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4690 struct cgroup *cont)
4692 mem_cgroup_sockets_destroy(cont, ss);
4695 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4700 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4701 struct cgroup *cont)
4706 static struct cftype mem_cgroup_files[] = {
4708 .name = "usage_in_bytes",
4709 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4710 .read_u64 = mem_cgroup_read,
4711 .register_event = mem_cgroup_usage_register_event,
4712 .unregister_event = mem_cgroup_usage_unregister_event,
4715 .name = "max_usage_in_bytes",
4716 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4717 .trigger = mem_cgroup_reset,
4718 .read_u64 = mem_cgroup_read,
4721 .name = "limit_in_bytes",
4722 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4723 .write_string = mem_cgroup_write,
4724 .read_u64 = mem_cgroup_read,
4727 .name = "soft_limit_in_bytes",
4728 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4729 .write_string = mem_cgroup_write,
4730 .read_u64 = mem_cgroup_read,
4734 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4735 .trigger = mem_cgroup_reset,
4736 .read_u64 = mem_cgroup_read,
4740 .read_map = mem_control_stat_show,
4743 .name = "force_empty",
4744 .trigger = mem_cgroup_force_empty_write,
4747 .name = "use_hierarchy",
4748 .write_u64 = mem_cgroup_hierarchy_write,
4749 .read_u64 = mem_cgroup_hierarchy_read,
4752 .name = "swappiness",
4753 .read_u64 = mem_cgroup_swappiness_read,
4754 .write_u64 = mem_cgroup_swappiness_write,
4757 .name = "move_charge_at_immigrate",
4758 .read_u64 = mem_cgroup_move_charge_read,
4759 .write_u64 = mem_cgroup_move_charge_write,
4762 .name = "oom_control",
4763 .read_map = mem_cgroup_oom_control_read,
4764 .write_u64 = mem_cgroup_oom_control_write,
4765 .register_event = mem_cgroup_oom_register_event,
4766 .unregister_event = mem_cgroup_oom_unregister_event,
4767 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4771 .name = "numa_stat",
4772 .open = mem_control_numa_stat_open,
4778 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4779 static struct cftype memsw_cgroup_files[] = {
4781 .name = "memsw.usage_in_bytes",
4782 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4783 .read_u64 = mem_cgroup_read,
4784 .register_event = mem_cgroup_usage_register_event,
4785 .unregister_event = mem_cgroup_usage_unregister_event,
4788 .name = "memsw.max_usage_in_bytes",
4789 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4790 .trigger = mem_cgroup_reset,
4791 .read_u64 = mem_cgroup_read,
4794 .name = "memsw.limit_in_bytes",
4795 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4796 .write_string = mem_cgroup_write,
4797 .read_u64 = mem_cgroup_read,
4800 .name = "memsw.failcnt",
4801 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4802 .trigger = mem_cgroup_reset,
4803 .read_u64 = mem_cgroup_read,
4807 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4809 if (!do_swap_account)
4811 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4812 ARRAY_SIZE(memsw_cgroup_files));
4815 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4821 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4823 struct mem_cgroup_per_node *pn;
4824 struct mem_cgroup_per_zone *mz;
4826 int zone, tmp = node;
4828 * This routine is called against possible nodes.
4829 * But it's BUG to call kmalloc() against offline node.
4831 * TODO: this routine can waste much memory for nodes which will
4832 * never be onlined. It's better to use memory hotplug callback
4835 if (!node_state(node, N_NORMAL_MEMORY))
4837 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4841 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4842 mz = &pn->zoneinfo[zone];
4844 INIT_LIST_HEAD(&mz->lists[l]);
4845 mz->usage_in_excess = 0;
4846 mz->on_tree = false;
4849 memcg->info.nodeinfo[node] = pn;
4853 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4855 kfree(memcg->info.nodeinfo[node]);
4858 static struct mem_cgroup *mem_cgroup_alloc(void)
4860 struct mem_cgroup *mem;
4861 int size = sizeof(struct mem_cgroup);
4863 /* Can be very big if MAX_NUMNODES is very big */
4864 if (size < PAGE_SIZE)
4865 mem = kzalloc(size, GFP_KERNEL);
4867 mem = vzalloc(size);
4872 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4875 spin_lock_init(&mem->pcp_counter_lock);
4879 if (size < PAGE_SIZE)
4887 * At destroying mem_cgroup, references from swap_cgroup can remain.
4888 * (scanning all at force_empty is too costly...)
4890 * Instead of clearing all references at force_empty, we remember
4891 * the number of reference from swap_cgroup and free mem_cgroup when
4892 * it goes down to 0.
4894 * Removal of cgroup itself succeeds regardless of refs from swap.
4897 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4901 mem_cgroup_remove_from_trees(memcg);
4902 free_css_id(&mem_cgroup_subsys, &memcg->css);
4904 for_each_node_state(node, N_POSSIBLE)
4905 free_mem_cgroup_per_zone_info(memcg, node);
4907 free_percpu(memcg->stat);
4908 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4914 static void mem_cgroup_get(struct mem_cgroup *memcg)
4916 atomic_inc(&memcg->refcnt);
4919 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4921 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4922 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4923 __mem_cgroup_free(memcg);
4925 mem_cgroup_put(parent);
4929 static void mem_cgroup_put(struct mem_cgroup *memcg)
4931 __mem_cgroup_put(memcg, 1);
4935 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4937 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4939 if (!memcg->res.parent)
4941 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4943 EXPORT_SYMBOL(parent_mem_cgroup);
4945 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4946 static void __init enable_swap_cgroup(void)
4948 if (!mem_cgroup_disabled() && really_do_swap_account)
4949 do_swap_account = 1;
4952 static void __init enable_swap_cgroup(void)
4957 static int mem_cgroup_soft_limit_tree_init(void)
4959 struct mem_cgroup_tree_per_node *rtpn;
4960 struct mem_cgroup_tree_per_zone *rtpz;
4961 int tmp, node, zone;
4963 for_each_node_state(node, N_POSSIBLE) {
4965 if (!node_state(node, N_NORMAL_MEMORY))
4967 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4971 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4973 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4974 rtpz = &rtpn->rb_tree_per_zone[zone];
4975 rtpz->rb_root = RB_ROOT;
4976 spin_lock_init(&rtpz->lock);
4982 static struct cgroup_subsys_state * __ref
4983 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4985 struct mem_cgroup *memcg, *parent;
4986 long error = -ENOMEM;
4989 memcg = mem_cgroup_alloc();
4991 return ERR_PTR(error);
4993 for_each_node_state(node, N_POSSIBLE)
4994 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4998 if (cont->parent == NULL) {
5000 enable_swap_cgroup();
5002 if (mem_cgroup_soft_limit_tree_init())
5004 root_mem_cgroup = memcg;
5005 for_each_possible_cpu(cpu) {
5006 struct memcg_stock_pcp *stock =
5007 &per_cpu(memcg_stock, cpu);
5008 INIT_WORK(&stock->work, drain_local_stock);
5010 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5012 parent = mem_cgroup_from_cont(cont->parent);
5013 memcg->use_hierarchy = parent->use_hierarchy;
5014 memcg->oom_kill_disable = parent->oom_kill_disable;
5017 if (parent && parent->use_hierarchy) {
5018 res_counter_init(&memcg->res, &parent->res);
5019 res_counter_init(&memcg->memsw, &parent->memsw);
5021 * We increment refcnt of the parent to ensure that we can
5022 * safely access it on res_counter_charge/uncharge.
5023 * This refcnt will be decremented when freeing this
5024 * mem_cgroup(see mem_cgroup_put).
5026 mem_cgroup_get(parent);
5028 res_counter_init(&memcg->res, NULL);
5029 res_counter_init(&memcg->memsw, NULL);
5031 memcg->last_scanned_child = 0;
5032 memcg->last_scanned_node = MAX_NUMNODES;
5033 INIT_LIST_HEAD(&memcg->oom_notify);
5036 memcg->swappiness = mem_cgroup_swappiness(parent);
5037 atomic_set(&memcg->refcnt, 1);
5038 memcg->move_charge_at_immigrate = 0;
5039 mutex_init(&memcg->thresholds_lock);
5042 __mem_cgroup_free(memcg);
5043 return ERR_PTR(error);
5046 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5047 struct cgroup *cont)
5049 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5051 return mem_cgroup_force_empty(memcg, false);
5054 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5055 struct cgroup *cont)
5057 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5059 kmem_cgroup_destroy(ss, cont);
5061 mem_cgroup_put(memcg);
5064 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5065 struct cgroup *cont)
5069 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5070 ARRAY_SIZE(mem_cgroup_files));
5073 ret = register_memsw_files(cont, ss);
5076 ret = register_kmem_files(cont, ss);
5082 /* Handlers for move charge at task migration. */
5083 #define PRECHARGE_COUNT_AT_ONCE 256
5084 static int mem_cgroup_do_precharge(unsigned long count)
5087 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5088 struct mem_cgroup *memcg = mc.to;
5090 if (mem_cgroup_is_root(memcg)) {
5091 mc.precharge += count;
5092 /* we don't need css_get for root */
5095 /* try to charge at once */
5097 struct res_counter *dummy;
5099 * "memcg" cannot be under rmdir() because we've already checked
5100 * by cgroup_lock_live_cgroup() that it is not removed and we
5101 * are still under the same cgroup_mutex. So we can postpone
5104 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5106 if (do_swap_account && res_counter_charge(&memcg->memsw,
5107 PAGE_SIZE * count, &dummy)) {
5108 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5111 mc.precharge += count;
5115 /* fall back to one by one charge */
5117 if (signal_pending(current)) {
5121 if (!batch_count--) {
5122 batch_count = PRECHARGE_COUNT_AT_ONCE;
5125 ret = __mem_cgroup_try_charge(NULL,
5126 GFP_KERNEL, 1, &memcg, false);
5128 /* mem_cgroup_clear_mc() will do uncharge later */
5136 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5137 * @vma: the vma the pte to be checked belongs
5138 * @addr: the address corresponding to the pte to be checked
5139 * @ptent: the pte to be checked
5140 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5143 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5144 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5145 * move charge. if @target is not NULL, the page is stored in target->page
5146 * with extra refcnt got(Callers should handle it).
5147 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5148 * target for charge migration. if @target is not NULL, the entry is stored
5151 * Called with pte lock held.
5158 enum mc_target_type {
5159 MC_TARGET_NONE, /* not used */
5164 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5165 unsigned long addr, pte_t ptent)
5167 struct page *page = vm_normal_page(vma, addr, ptent);
5169 if (!page || !page_mapped(page))
5171 if (PageAnon(page)) {
5172 /* we don't move shared anon */
5173 if (!move_anon() || page_mapcount(page) > 2)
5175 } else if (!move_file())
5176 /* we ignore mapcount for file pages */
5178 if (!get_page_unless_zero(page))
5184 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5185 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5188 struct page *page = NULL;
5189 swp_entry_t ent = pte_to_swp_entry(ptent);
5191 if (!move_anon() || non_swap_entry(ent))
5193 usage_count = mem_cgroup_count_swap_user(ent, &page);
5194 if (usage_count > 1) { /* we don't move shared anon */
5199 if (do_swap_account)
5200 entry->val = ent.val;
5205 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5206 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5208 struct page *page = NULL;
5209 struct inode *inode;
5210 struct address_space *mapping;
5213 if (!vma->vm_file) /* anonymous vma */
5218 inode = vma->vm_file->f_path.dentry->d_inode;
5219 mapping = vma->vm_file->f_mapping;
5220 if (pte_none(ptent))
5221 pgoff = linear_page_index(vma, addr);
5222 else /* pte_file(ptent) is true */
5223 pgoff = pte_to_pgoff(ptent);
5225 /* page is moved even if it's not RSS of this task(page-faulted). */
5226 page = find_get_page(mapping, pgoff);
5229 /* shmem/tmpfs may report page out on swap: account for that too. */
5230 if (radix_tree_exceptional_entry(page)) {
5231 swp_entry_t swap = radix_to_swp_entry(page);
5232 if (do_swap_account)
5234 page = find_get_page(&swapper_space, swap.val);
5240 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5241 unsigned long addr, pte_t ptent, union mc_target *target)
5243 struct page *page = NULL;
5244 struct page_cgroup *pc;
5246 swp_entry_t ent = { .val = 0 };
5248 if (pte_present(ptent))
5249 page = mc_handle_present_pte(vma, addr, ptent);
5250 else if (is_swap_pte(ptent))
5251 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5252 else if (pte_none(ptent) || pte_file(ptent))
5253 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5255 if (!page && !ent.val)
5258 pc = lookup_page_cgroup(page);
5260 * Do only loose check w/o page_cgroup lock.
5261 * mem_cgroup_move_account() checks the pc is valid or not under
5264 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5265 ret = MC_TARGET_PAGE;
5267 target->page = page;
5269 if (!ret || !target)
5272 /* There is a swap entry and a page doesn't exist or isn't charged */
5273 if (ent.val && !ret &&
5274 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5275 ret = MC_TARGET_SWAP;
5282 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5283 unsigned long addr, unsigned long end,
5284 struct mm_walk *walk)
5286 struct vm_area_struct *vma = walk->private;
5290 split_huge_page_pmd(walk->mm, pmd);
5292 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5293 for (; addr != end; pte++, addr += PAGE_SIZE)
5294 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5295 mc.precharge++; /* increment precharge temporarily */
5296 pte_unmap_unlock(pte - 1, ptl);
5302 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5304 unsigned long precharge;
5305 struct vm_area_struct *vma;
5307 down_read(&mm->mmap_sem);
5308 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5309 struct mm_walk mem_cgroup_count_precharge_walk = {
5310 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5314 if (is_vm_hugetlb_page(vma))
5316 walk_page_range(vma->vm_start, vma->vm_end,
5317 &mem_cgroup_count_precharge_walk);
5319 up_read(&mm->mmap_sem);
5321 precharge = mc.precharge;
5327 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5329 unsigned long precharge = mem_cgroup_count_precharge(mm);
5331 VM_BUG_ON(mc.moving_task);
5332 mc.moving_task = current;
5333 return mem_cgroup_do_precharge(precharge);
5336 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5337 static void __mem_cgroup_clear_mc(void)
5339 struct mem_cgroup *from = mc.from;
5340 struct mem_cgroup *to = mc.to;
5342 /* we must uncharge all the leftover precharges from mc.to */
5344 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5348 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5349 * we must uncharge here.
5351 if (mc.moved_charge) {
5352 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5353 mc.moved_charge = 0;
5355 /* we must fixup refcnts and charges */
5356 if (mc.moved_swap) {
5357 /* uncharge swap account from the old cgroup */
5358 if (!mem_cgroup_is_root(mc.from))
5359 res_counter_uncharge(&mc.from->memsw,
5360 PAGE_SIZE * mc.moved_swap);
5361 __mem_cgroup_put(mc.from, mc.moved_swap);
5363 if (!mem_cgroup_is_root(mc.to)) {
5365 * we charged both to->res and to->memsw, so we should
5368 res_counter_uncharge(&mc.to->res,
5369 PAGE_SIZE * mc.moved_swap);
5371 /* we've already done mem_cgroup_get(mc.to) */
5374 memcg_oom_recover(from);
5375 memcg_oom_recover(to);
5376 wake_up_all(&mc.waitq);
5379 static void mem_cgroup_clear_mc(void)
5381 struct mem_cgroup *from = mc.from;
5384 * we must clear moving_task before waking up waiters at the end of
5387 mc.moving_task = NULL;
5388 __mem_cgroup_clear_mc();
5389 spin_lock(&mc.lock);
5392 spin_unlock(&mc.lock);
5393 mem_cgroup_end_move(from);
5396 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5397 struct cgroup *cgroup,
5398 struct cgroup_taskset *tset)
5400 struct task_struct *p = cgroup_taskset_first(tset);
5402 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5404 if (memcg->move_charge_at_immigrate) {
5405 struct mm_struct *mm;
5406 struct mem_cgroup *from = mem_cgroup_from_task(p);
5408 VM_BUG_ON(from == memcg);
5410 mm = get_task_mm(p);
5413 /* We move charges only when we move a owner of the mm */
5414 if (mm->owner == p) {
5417 VM_BUG_ON(mc.precharge);
5418 VM_BUG_ON(mc.moved_charge);
5419 VM_BUG_ON(mc.moved_swap);
5420 mem_cgroup_start_move(from);
5421 spin_lock(&mc.lock);
5424 spin_unlock(&mc.lock);
5425 /* We set mc.moving_task later */
5427 ret = mem_cgroup_precharge_mc(mm);
5429 mem_cgroup_clear_mc();
5436 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5437 struct cgroup *cgroup,
5438 struct cgroup_taskset *tset)
5440 mem_cgroup_clear_mc();
5443 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5444 unsigned long addr, unsigned long end,
5445 struct mm_walk *walk)
5448 struct vm_area_struct *vma = walk->private;
5452 split_huge_page_pmd(walk->mm, pmd);
5454 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5455 for (; addr != end; addr += PAGE_SIZE) {
5456 pte_t ptent = *(pte++);
5457 union mc_target target;
5460 struct page_cgroup *pc;
5466 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5468 case MC_TARGET_PAGE:
5470 if (isolate_lru_page(page))
5472 pc = lookup_page_cgroup(page);
5473 if (!mem_cgroup_move_account(page, 1, pc,
5474 mc.from, mc.to, false)) {
5476 /* we uncharge from mc.from later. */
5479 putback_lru_page(page);
5480 put: /* is_target_pte_for_mc() gets the page */
5483 case MC_TARGET_SWAP:
5485 if (!mem_cgroup_move_swap_account(ent,
5486 mc.from, mc.to, false)) {
5488 /* we fixup refcnts and charges later. */
5496 pte_unmap_unlock(pte - 1, ptl);
5501 * We have consumed all precharges we got in can_attach().
5502 * We try charge one by one, but don't do any additional
5503 * charges to mc.to if we have failed in charge once in attach()
5506 ret = mem_cgroup_do_precharge(1);
5514 static void mem_cgroup_move_charge(struct mm_struct *mm)
5516 struct vm_area_struct *vma;
5518 lru_add_drain_all();
5520 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5522 * Someone who are holding the mmap_sem might be waiting in
5523 * waitq. So we cancel all extra charges, wake up all waiters,
5524 * and retry. Because we cancel precharges, we might not be able
5525 * to move enough charges, but moving charge is a best-effort
5526 * feature anyway, so it wouldn't be a big problem.
5528 __mem_cgroup_clear_mc();
5532 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5534 struct mm_walk mem_cgroup_move_charge_walk = {
5535 .pmd_entry = mem_cgroup_move_charge_pte_range,
5539 if (is_vm_hugetlb_page(vma))
5541 ret = walk_page_range(vma->vm_start, vma->vm_end,
5542 &mem_cgroup_move_charge_walk);
5545 * means we have consumed all precharges and failed in
5546 * doing additional charge. Just abandon here.
5550 up_read(&mm->mmap_sem);
5553 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5554 struct cgroup *cont,
5555 struct cgroup_taskset *tset)
5557 struct task_struct *p = cgroup_taskset_first(tset);
5558 struct mm_struct *mm = get_task_mm(p);
5562 mem_cgroup_move_charge(mm);
5567 mem_cgroup_clear_mc();
5569 #else /* !CONFIG_MMU */
5570 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5571 struct cgroup *cgroup,
5572 struct cgroup_taskset *tset)
5576 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5577 struct cgroup *cgroup,
5578 struct cgroup_taskset *tset)
5581 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5582 struct cgroup *cont,
5583 struct cgroup_taskset *tset)
5588 struct cgroup_subsys mem_cgroup_subsys = {
5590 .subsys_id = mem_cgroup_subsys_id,
5591 .create = mem_cgroup_create,
5592 .pre_destroy = mem_cgroup_pre_destroy,
5593 .destroy = mem_cgroup_destroy,
5594 .populate = mem_cgroup_populate,
5595 .can_attach = mem_cgroup_can_attach,
5596 .cancel_attach = mem_cgroup_cancel_attach,
5597 .attach = mem_cgroup_move_task,
5602 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5603 static int __init enable_swap_account(char *s)
5605 /* consider enabled if no parameter or 1 is given */
5606 if (!strcmp(s, "1"))
5607 really_do_swap_account = 1;
5608 else if (!strcmp(s, "0"))
5609 really_do_swap_account = 0;
5612 __setup("swapaccount=", enable_swap_account);