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 static struct mem_cgroup *root_mem_cgroup __read_mostly;
64 #ifdef CONFIG_MEMCG_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_MEMCG_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_SWAP, /* # of pages, swapped out */
91 MEM_CGROUP_STAT_NSTATS,
94 static const char * const mem_cgroup_stat_names[] = {
101 enum mem_cgroup_events_index {
102 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
103 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
104 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
105 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
106 MEM_CGROUP_EVENTS_NSTATS,
109 static const char * const mem_cgroup_events_names[] = {
117 * Per memcg event counter is incremented at every pagein/pageout. With THP,
118 * it will be incremated by the number of pages. This counter is used for
119 * for trigger some periodic events. This is straightforward and better
120 * than using jiffies etc. to handle periodic memcg event.
122 enum mem_cgroup_events_target {
123 MEM_CGROUP_TARGET_THRESH,
124 MEM_CGROUP_TARGET_SOFTLIMIT,
125 MEM_CGROUP_TARGET_NUMAINFO,
128 #define THRESHOLDS_EVENTS_TARGET 128
129 #define SOFTLIMIT_EVENTS_TARGET 1024
130 #define NUMAINFO_EVENTS_TARGET 1024
132 struct mem_cgroup_stat_cpu {
133 long count[MEM_CGROUP_STAT_NSTATS];
134 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
135 unsigned long nr_page_events;
136 unsigned long targets[MEM_CGROUP_NTARGETS];
139 struct mem_cgroup_reclaim_iter {
140 /* css_id of the last scanned hierarchy member */
142 /* scan generation, increased every round-trip */
143 unsigned int generation;
147 * per-zone information in memory controller.
149 struct mem_cgroup_per_zone {
150 struct lruvec lruvec;
151 unsigned long lru_size[NR_LRU_LISTS];
153 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
155 struct rb_node tree_node; /* RB tree node */
156 unsigned long long usage_in_excess;/* Set to the value by which */
157 /* the soft limit is exceeded*/
159 struct mem_cgroup *memcg; /* Back pointer, we cannot */
160 /* use container_of */
163 struct mem_cgroup_per_node {
164 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
167 struct mem_cgroup_lru_info {
168 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
172 * Cgroups above their limits are maintained in a RB-Tree, independent of
173 * their hierarchy representation
176 struct mem_cgroup_tree_per_zone {
177 struct rb_root rb_root;
181 struct mem_cgroup_tree_per_node {
182 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
185 struct mem_cgroup_tree {
186 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
189 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
191 struct mem_cgroup_threshold {
192 struct eventfd_ctx *eventfd;
197 struct mem_cgroup_threshold_ary {
198 /* An array index points to threshold just below or equal to usage. */
199 int current_threshold;
200 /* Size of entries[] */
202 /* Array of thresholds */
203 struct mem_cgroup_threshold entries[0];
206 struct mem_cgroup_thresholds {
207 /* Primary thresholds array */
208 struct mem_cgroup_threshold_ary *primary;
210 * Spare threshold array.
211 * This is needed to make mem_cgroup_unregister_event() "never fail".
212 * It must be able to store at least primary->size - 1 entries.
214 struct mem_cgroup_threshold_ary *spare;
218 struct mem_cgroup_eventfd_list {
219 struct list_head list;
220 struct eventfd_ctx *eventfd;
223 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
224 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
227 * The memory controller data structure. The memory controller controls both
228 * page cache and RSS per cgroup. We would eventually like to provide
229 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
230 * to help the administrator determine what knobs to tune.
232 * TODO: Add a water mark for the memory controller. Reclaim will begin when
233 * we hit the water mark. May be even add a low water mark, such that
234 * no reclaim occurs from a cgroup at it's low water mark, this is
235 * a feature that will be implemented much later in the future.
238 struct cgroup_subsys_state css;
240 * the counter to account for memory usage
242 struct res_counter res;
246 * the counter to account for mem+swap usage.
248 struct res_counter memsw;
251 * rcu_freeing is used only when freeing struct mem_cgroup,
252 * so put it into a union to avoid wasting more memory.
253 * It must be disjoint from the css field. It could be
254 * in a union with the res field, but res plays a much
255 * larger part in mem_cgroup life than memsw, and might
256 * be of interest, even at time of free, when debugging.
257 * So share rcu_head with the less interesting memsw.
259 struct rcu_head rcu_freeing;
261 * We also need some space for a worker in deferred freeing.
262 * By the time we call it, rcu_freeing is no longer in use.
264 struct work_struct work_freeing;
268 * Per cgroup active and inactive list, similar to the
269 * per zone LRU lists.
271 struct mem_cgroup_lru_info info;
272 int last_scanned_node;
274 nodemask_t scan_nodes;
275 atomic_t numainfo_events;
276 atomic_t numainfo_updating;
279 * Should the accounting and control be hierarchical, per subtree?
289 /* OOM-Killer disable */
290 int oom_kill_disable;
292 /* set when res.limit == memsw.limit */
293 bool memsw_is_minimum;
295 /* protect arrays of thresholds */
296 struct mutex thresholds_lock;
298 /* thresholds for memory usage. RCU-protected */
299 struct mem_cgroup_thresholds thresholds;
301 /* thresholds for mem+swap usage. RCU-protected */
302 struct mem_cgroup_thresholds memsw_thresholds;
304 /* For oom notifier event fd */
305 struct list_head oom_notify;
308 * Should we move charges of a task when a task is moved into this
309 * mem_cgroup ? And what type of charges should we move ?
311 unsigned long move_charge_at_immigrate;
313 * set > 0 if pages under this cgroup are moving to other cgroup.
315 atomic_t moving_account;
316 /* taken only while moving_account > 0 */
317 spinlock_t move_lock;
321 struct mem_cgroup_stat_cpu __percpu *stat;
323 * used when a cpu is offlined or other synchronizations
324 * See mem_cgroup_read_stat().
326 struct mem_cgroup_stat_cpu nocpu_base;
327 spinlock_t pcp_counter_lock;
330 struct tcp_memcontrol tcp_mem;
334 /* Stuffs for move charges at task migration. */
336 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
337 * left-shifted bitmap of these types.
340 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
341 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
345 /* "mc" and its members are protected by cgroup_mutex */
346 static struct move_charge_struct {
347 spinlock_t lock; /* for from, to */
348 struct mem_cgroup *from;
349 struct mem_cgroup *to;
350 unsigned long precharge;
351 unsigned long moved_charge;
352 unsigned long moved_swap;
353 struct task_struct *moving_task; /* a task moving charges */
354 wait_queue_head_t waitq; /* a waitq for other context */
356 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
357 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
360 static bool move_anon(void)
362 return test_bit(MOVE_CHARGE_TYPE_ANON,
363 &mc.to->move_charge_at_immigrate);
366 static bool move_file(void)
368 return test_bit(MOVE_CHARGE_TYPE_FILE,
369 &mc.to->move_charge_at_immigrate);
373 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
374 * limit reclaim to prevent infinite loops, if they ever occur.
376 #define MEM_CGROUP_MAX_RECLAIM_LOOPS 100
377 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS 2
380 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
381 MEM_CGROUP_CHARGE_TYPE_ANON,
382 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
383 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
387 /* for encoding cft->private value on file */
390 #define _OOM_TYPE (2)
391 #define MEMFILE_PRIVATE(x, val) ((x) << 16 | (val))
392 #define MEMFILE_TYPE(val) ((val) >> 16 & 0xffff)
393 #define MEMFILE_ATTR(val) ((val) & 0xffff)
394 /* Used for OOM nofiier */
395 #define OOM_CONTROL (0)
398 * Reclaim flags for mem_cgroup_hierarchical_reclaim
400 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
401 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
402 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
403 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
405 static void mem_cgroup_get(struct mem_cgroup *memcg);
406 static void mem_cgroup_put(struct mem_cgroup *memcg);
408 /* Writing them here to avoid exposing memcg's inner layout */
409 #ifdef CONFIG_MEMCG_KMEM
410 #include <net/sock.h>
413 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
414 void sock_update_memcg(struct sock *sk)
416 if (mem_cgroup_sockets_enabled) {
417 struct mem_cgroup *memcg;
418 struct cg_proto *cg_proto;
420 BUG_ON(!sk->sk_prot->proto_cgroup);
422 /* Socket cloning can throw us here with sk_cgrp already
423 * filled. It won't however, necessarily happen from
424 * process context. So the test for root memcg given
425 * the current task's memcg won't help us in this case.
427 * Respecting the original socket's memcg is a better
428 * decision in this case.
431 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
432 mem_cgroup_get(sk->sk_cgrp->memcg);
437 memcg = mem_cgroup_from_task(current);
438 cg_proto = sk->sk_prot->proto_cgroup(memcg);
439 if (!mem_cgroup_is_root(memcg) && memcg_proto_active(cg_proto)) {
440 mem_cgroup_get(memcg);
441 sk->sk_cgrp = cg_proto;
446 EXPORT_SYMBOL(sock_update_memcg);
448 void sock_release_memcg(struct sock *sk)
450 if (mem_cgroup_sockets_enabled && sk->sk_cgrp) {
451 struct mem_cgroup *memcg;
452 WARN_ON(!sk->sk_cgrp->memcg);
453 memcg = sk->sk_cgrp->memcg;
454 mem_cgroup_put(memcg);
459 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
461 if (!memcg || mem_cgroup_is_root(memcg))
464 return &memcg->tcp_mem.cg_proto;
466 EXPORT_SYMBOL(tcp_proto_cgroup);
467 #endif /* CONFIG_INET */
468 #endif /* CONFIG_MEMCG_KMEM */
470 #if defined(CONFIG_INET) && defined(CONFIG_MEMCG_KMEM)
471 static void disarm_sock_keys(struct mem_cgroup *memcg)
473 if (!memcg_proto_activated(&memcg->tcp_mem.cg_proto))
475 static_key_slow_dec(&memcg_socket_limit_enabled);
478 static void disarm_sock_keys(struct mem_cgroup *memcg)
483 static void drain_all_stock_async(struct mem_cgroup *memcg);
485 static struct mem_cgroup_per_zone *
486 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
488 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
491 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
496 static struct mem_cgroup_per_zone *
497 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
499 int nid = page_to_nid(page);
500 int zid = page_zonenum(page);
502 return mem_cgroup_zoneinfo(memcg, nid, zid);
505 static struct mem_cgroup_tree_per_zone *
506 soft_limit_tree_node_zone(int nid, int zid)
508 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
511 static struct mem_cgroup_tree_per_zone *
512 soft_limit_tree_from_page(struct page *page)
514 int nid = page_to_nid(page);
515 int zid = page_zonenum(page);
517 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
521 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
522 struct mem_cgroup_per_zone *mz,
523 struct mem_cgroup_tree_per_zone *mctz,
524 unsigned long long new_usage_in_excess)
526 struct rb_node **p = &mctz->rb_root.rb_node;
527 struct rb_node *parent = NULL;
528 struct mem_cgroup_per_zone *mz_node;
533 mz->usage_in_excess = new_usage_in_excess;
534 if (!mz->usage_in_excess)
538 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
540 if (mz->usage_in_excess < mz_node->usage_in_excess)
543 * We can't avoid mem cgroups that are over their soft
544 * limit by the same amount
546 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
549 rb_link_node(&mz->tree_node, parent, p);
550 rb_insert_color(&mz->tree_node, &mctz->rb_root);
555 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
556 struct mem_cgroup_per_zone *mz,
557 struct mem_cgroup_tree_per_zone *mctz)
561 rb_erase(&mz->tree_node, &mctz->rb_root);
566 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
567 struct mem_cgroup_per_zone *mz,
568 struct mem_cgroup_tree_per_zone *mctz)
570 spin_lock(&mctz->lock);
571 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
572 spin_unlock(&mctz->lock);
576 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
578 unsigned long long excess;
579 struct mem_cgroup_per_zone *mz;
580 struct mem_cgroup_tree_per_zone *mctz;
581 int nid = page_to_nid(page);
582 int zid = page_zonenum(page);
583 mctz = soft_limit_tree_from_page(page);
586 * Necessary to update all ancestors when hierarchy is used.
587 * because their event counter is not touched.
589 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
590 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
591 excess = res_counter_soft_limit_excess(&memcg->res);
593 * We have to update the tree if mz is on RB-tree or
594 * mem is over its softlimit.
596 if (excess || mz->on_tree) {
597 spin_lock(&mctz->lock);
598 /* if on-tree, remove it */
600 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
602 * Insert again. mz->usage_in_excess will be updated.
603 * If excess is 0, no tree ops.
605 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
606 spin_unlock(&mctz->lock);
611 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
614 struct mem_cgroup_per_zone *mz;
615 struct mem_cgroup_tree_per_zone *mctz;
617 for_each_node(node) {
618 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
619 mz = mem_cgroup_zoneinfo(memcg, node, zone);
620 mctz = soft_limit_tree_node_zone(node, zone);
621 mem_cgroup_remove_exceeded(memcg, mz, mctz);
626 static struct mem_cgroup_per_zone *
627 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
629 struct rb_node *rightmost = NULL;
630 struct mem_cgroup_per_zone *mz;
634 rightmost = rb_last(&mctz->rb_root);
636 goto done; /* Nothing to reclaim from */
638 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
640 * Remove the node now but someone else can add it back,
641 * we will to add it back at the end of reclaim to its correct
642 * position in the tree.
644 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
645 if (!res_counter_soft_limit_excess(&mz->memcg->res) ||
646 !css_tryget(&mz->memcg->css))
652 static struct mem_cgroup_per_zone *
653 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
655 struct mem_cgroup_per_zone *mz;
657 spin_lock(&mctz->lock);
658 mz = __mem_cgroup_largest_soft_limit_node(mctz);
659 spin_unlock(&mctz->lock);
664 * Implementation Note: reading percpu statistics for memcg.
666 * Both of vmstat[] and percpu_counter has threshold and do periodic
667 * synchronization to implement "quick" read. There are trade-off between
668 * reading cost and precision of value. Then, we may have a chance to implement
669 * a periodic synchronizion of counter in memcg's counter.
671 * But this _read() function is used for user interface now. The user accounts
672 * memory usage by memory cgroup and he _always_ requires exact value because
673 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
674 * have to visit all online cpus and make sum. So, for now, unnecessary
675 * synchronization is not implemented. (just implemented for cpu hotplug)
677 * If there are kernel internal actions which can make use of some not-exact
678 * value, and reading all cpu value can be performance bottleneck in some
679 * common workload, threashold and synchonization as vmstat[] should be
682 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
683 enum mem_cgroup_stat_index idx)
689 for_each_online_cpu(cpu)
690 val += per_cpu(memcg->stat->count[idx], cpu);
691 #ifdef CONFIG_HOTPLUG_CPU
692 spin_lock(&memcg->pcp_counter_lock);
693 val += memcg->nocpu_base.count[idx];
694 spin_unlock(&memcg->pcp_counter_lock);
700 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
703 int val = (charge) ? 1 : -1;
704 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAP], val);
707 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
708 enum mem_cgroup_events_index idx)
710 unsigned long val = 0;
713 for_each_online_cpu(cpu)
714 val += per_cpu(memcg->stat->events[idx], cpu);
715 #ifdef CONFIG_HOTPLUG_CPU
716 spin_lock(&memcg->pcp_counter_lock);
717 val += memcg->nocpu_base.events[idx];
718 spin_unlock(&memcg->pcp_counter_lock);
723 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
724 bool anon, int nr_pages)
729 * Here, RSS means 'mapped anon' and anon's SwapCache. Shmem/tmpfs is
730 * counted as CACHE even if it's on ANON LRU.
733 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
736 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
739 /* pagein of a big page is an event. So, ignore page size */
741 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
743 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
744 nr_pages = -nr_pages; /* for event */
747 __this_cpu_add(memcg->stat->nr_page_events, nr_pages);
753 mem_cgroup_get_lru_size(struct lruvec *lruvec, enum lru_list lru)
755 struct mem_cgroup_per_zone *mz;
757 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
758 return mz->lru_size[lru];
762 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
763 unsigned int lru_mask)
765 struct mem_cgroup_per_zone *mz;
767 unsigned long ret = 0;
769 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
772 if (BIT(lru) & lru_mask)
773 ret += mz->lru_size[lru];
779 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
780 int nid, unsigned int lru_mask)
785 for (zid = 0; zid < MAX_NR_ZONES; zid++)
786 total += mem_cgroup_zone_nr_lru_pages(memcg,
792 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
793 unsigned int lru_mask)
798 for_each_node_state(nid, N_HIGH_MEMORY)
799 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
803 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
804 enum mem_cgroup_events_target target)
806 unsigned long val, next;
808 val = __this_cpu_read(memcg->stat->nr_page_events);
809 next = __this_cpu_read(memcg->stat->targets[target]);
810 /* from time_after() in jiffies.h */
811 if ((long)next - (long)val < 0) {
813 case MEM_CGROUP_TARGET_THRESH:
814 next = val + THRESHOLDS_EVENTS_TARGET;
816 case MEM_CGROUP_TARGET_SOFTLIMIT:
817 next = val + SOFTLIMIT_EVENTS_TARGET;
819 case MEM_CGROUP_TARGET_NUMAINFO:
820 next = val + NUMAINFO_EVENTS_TARGET;
825 __this_cpu_write(memcg->stat->targets[target], next);
832 * Check events in order.
835 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
838 /* threshold event is triggered in finer grain than soft limit */
839 if (unlikely(mem_cgroup_event_ratelimit(memcg,
840 MEM_CGROUP_TARGET_THRESH))) {
842 bool do_numainfo __maybe_unused;
844 do_softlimit = mem_cgroup_event_ratelimit(memcg,
845 MEM_CGROUP_TARGET_SOFTLIMIT);
847 do_numainfo = mem_cgroup_event_ratelimit(memcg,
848 MEM_CGROUP_TARGET_NUMAINFO);
852 mem_cgroup_threshold(memcg);
853 if (unlikely(do_softlimit))
854 mem_cgroup_update_tree(memcg, page);
856 if (unlikely(do_numainfo))
857 atomic_inc(&memcg->numainfo_events);
863 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
865 return container_of(cgroup_subsys_state(cont,
866 mem_cgroup_subsys_id), struct mem_cgroup,
870 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
873 * mm_update_next_owner() may clear mm->owner to NULL
874 * if it races with swapoff, page migration, etc.
875 * So this can be called with p == NULL.
880 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
881 struct mem_cgroup, css);
884 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
886 struct mem_cgroup *memcg = NULL;
891 * Because we have no locks, mm->owner's may be being moved to other
892 * cgroup. We use css_tryget() here even if this looks
893 * pessimistic (rather than adding locks here).
897 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
898 if (unlikely(!memcg))
900 } while (!css_tryget(&memcg->css));
906 * mem_cgroup_iter - iterate over memory cgroup hierarchy
907 * @root: hierarchy root
908 * @prev: previously returned memcg, NULL on first invocation
909 * @reclaim: cookie for shared reclaim walks, NULL for full walks
911 * Returns references to children of the hierarchy below @root, or
912 * @root itself, or %NULL after a full round-trip.
914 * Caller must pass the return value in @prev on subsequent
915 * invocations for reference counting, or use mem_cgroup_iter_break()
916 * to cancel a hierarchy walk before the round-trip is complete.
918 * Reclaimers can specify a zone and a priority level in @reclaim to
919 * divide up the memcgs in the hierarchy among all concurrent
920 * reclaimers operating on the same zone and priority.
922 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
923 struct mem_cgroup *prev,
924 struct mem_cgroup_reclaim_cookie *reclaim)
926 struct mem_cgroup *memcg = NULL;
929 if (mem_cgroup_disabled())
933 root = root_mem_cgroup;
935 if (prev && !reclaim)
936 id = css_id(&prev->css);
938 if (prev && prev != root)
941 if (!root->use_hierarchy && root != root_mem_cgroup) {
948 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
949 struct cgroup_subsys_state *css;
952 int nid = zone_to_nid(reclaim->zone);
953 int zid = zone_idx(reclaim->zone);
954 struct mem_cgroup_per_zone *mz;
956 mz = mem_cgroup_zoneinfo(root, nid, zid);
957 iter = &mz->reclaim_iter[reclaim->priority];
958 if (prev && reclaim->generation != iter->generation)
964 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
966 if (css == &root->css || css_tryget(css))
967 memcg = container_of(css,
968 struct mem_cgroup, css);
977 else if (!prev && memcg)
978 reclaim->generation = iter->generation;
988 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
989 * @root: hierarchy root
990 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
992 void mem_cgroup_iter_break(struct mem_cgroup *root,
993 struct mem_cgroup *prev)
996 root = root_mem_cgroup;
997 if (prev && prev != root)
1002 * Iteration constructs for visiting all cgroups (under a tree). If
1003 * loops are exited prematurely (break), mem_cgroup_iter_break() must
1004 * be used for reference counting.
1006 #define for_each_mem_cgroup_tree(iter, root) \
1007 for (iter = mem_cgroup_iter(root, NULL, NULL); \
1009 iter = mem_cgroup_iter(root, iter, NULL))
1011 #define for_each_mem_cgroup(iter) \
1012 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
1014 iter = mem_cgroup_iter(NULL, iter, NULL))
1016 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
1018 return (memcg == root_mem_cgroup);
1021 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
1023 struct mem_cgroup *memcg;
1029 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
1030 if (unlikely(!memcg))
1035 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
1038 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
1046 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
1049 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
1050 * @zone: zone of the wanted lruvec
1051 * @memcg: memcg of the wanted lruvec
1053 * Returns the lru list vector holding pages for the given @zone and
1054 * @mem. This can be the global zone lruvec, if the memory controller
1057 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
1058 struct mem_cgroup *memcg)
1060 struct mem_cgroup_per_zone *mz;
1062 if (mem_cgroup_disabled())
1063 return &zone->lruvec;
1065 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1070 * Following LRU functions are allowed to be used without PCG_LOCK.
1071 * Operations are called by routine of global LRU independently from memcg.
1072 * What we have to take care of here is validness of pc->mem_cgroup.
1074 * Changes to pc->mem_cgroup happens when
1077 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1078 * It is added to LRU before charge.
1079 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1080 * When moving account, the page is not on LRU. It's isolated.
1084 * mem_cgroup_page_lruvec - return lruvec for adding an lru page
1086 * @zone: zone of the page
1088 struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct zone *zone)
1090 struct mem_cgroup_per_zone *mz;
1091 struct mem_cgroup *memcg;
1092 struct page_cgroup *pc;
1094 if (mem_cgroup_disabled())
1095 return &zone->lruvec;
1097 pc = lookup_page_cgroup(page);
1098 memcg = pc->mem_cgroup;
1101 * Surreptitiously switch any uncharged offlist page to root:
1102 * an uncharged page off lru does nothing to secure
1103 * its former mem_cgroup from sudden removal.
1105 * Our caller holds lru_lock, and PageCgroupUsed is updated
1106 * under page_cgroup lock: between them, they make all uses
1107 * of pc->mem_cgroup safe.
1109 if (!PageLRU(page) && !PageCgroupUsed(pc) && memcg != root_mem_cgroup)
1110 pc->mem_cgroup = memcg = root_mem_cgroup;
1112 mz = page_cgroup_zoneinfo(memcg, page);
1117 * mem_cgroup_update_lru_size - account for adding or removing an lru page
1118 * @lruvec: mem_cgroup per zone lru vector
1119 * @lru: index of lru list the page is sitting on
1120 * @nr_pages: positive when adding or negative when removing
1122 * This function must be called when a page is added to or removed from an
1125 void mem_cgroup_update_lru_size(struct lruvec *lruvec, enum lru_list lru,
1128 struct mem_cgroup_per_zone *mz;
1129 unsigned long *lru_size;
1131 if (mem_cgroup_disabled())
1134 mz = container_of(lruvec, struct mem_cgroup_per_zone, lruvec);
1135 lru_size = mz->lru_size + lru;
1136 *lru_size += nr_pages;
1137 VM_BUG_ON((long)(*lru_size) < 0);
1141 * Checks whether given mem is same or in the root_mem_cgroup's
1144 bool __mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1145 struct mem_cgroup *memcg)
1147 if (root_memcg == memcg)
1149 if (!root_memcg->use_hierarchy || !memcg)
1151 return css_is_ancestor(&memcg->css, &root_memcg->css);
1154 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1155 struct mem_cgroup *memcg)
1160 ret = __mem_cgroup_same_or_subtree(root_memcg, memcg);
1165 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1168 struct mem_cgroup *curr = NULL;
1169 struct task_struct *p;
1171 p = find_lock_task_mm(task);
1173 curr = try_get_mem_cgroup_from_mm(p->mm);
1177 * All threads may have already detached their mm's, but the oom
1178 * killer still needs to detect if they have already been oom
1179 * killed to prevent needlessly killing additional tasks.
1182 curr = mem_cgroup_from_task(task);
1184 css_get(&curr->css);
1190 * We should check use_hierarchy of "memcg" not "curr". Because checking
1191 * use_hierarchy of "curr" here make this function true if hierarchy is
1192 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1193 * hierarchy(even if use_hierarchy is disabled in "memcg").
1195 ret = mem_cgroup_same_or_subtree(memcg, curr);
1196 css_put(&curr->css);
1200 int mem_cgroup_inactive_anon_is_low(struct lruvec *lruvec)
1202 unsigned long inactive_ratio;
1203 unsigned long inactive;
1204 unsigned long active;
1207 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_ANON);
1208 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_ANON);
1210 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1212 inactive_ratio = int_sqrt(10 * gb);
1216 return inactive * inactive_ratio < active;
1219 int mem_cgroup_inactive_file_is_low(struct lruvec *lruvec)
1221 unsigned long active;
1222 unsigned long inactive;
1224 inactive = mem_cgroup_get_lru_size(lruvec, LRU_INACTIVE_FILE);
1225 active = mem_cgroup_get_lru_size(lruvec, LRU_ACTIVE_FILE);
1227 return (active > inactive);
1230 #define mem_cgroup_from_res_counter(counter, member) \
1231 container_of(counter, struct mem_cgroup, member)
1234 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1235 * @memcg: the memory cgroup
1237 * Returns the maximum amount of memory @mem can be charged with, in
1240 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1242 unsigned long long margin;
1244 margin = res_counter_margin(&memcg->res);
1245 if (do_swap_account)
1246 margin = min(margin, res_counter_margin(&memcg->memsw));
1247 return margin >> PAGE_SHIFT;
1250 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1252 struct cgroup *cgrp = memcg->css.cgroup;
1255 if (cgrp->parent == NULL)
1256 return vm_swappiness;
1258 return memcg->swappiness;
1262 * memcg->moving_account is used for checking possibility that some thread is
1263 * calling move_account(). When a thread on CPU-A starts moving pages under
1264 * a memcg, other threads should check memcg->moving_account under
1265 * rcu_read_lock(), like this:
1269 * memcg->moving_account+1 if (memcg->mocing_account)
1271 * synchronize_rcu() update something.
1276 /* for quick checking without looking up memcg */
1277 atomic_t memcg_moving __read_mostly;
1279 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1281 atomic_inc(&memcg_moving);
1282 atomic_inc(&memcg->moving_account);
1286 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1289 * Now, mem_cgroup_clear_mc() may call this function with NULL.
1290 * We check NULL in callee rather than caller.
1293 atomic_dec(&memcg_moving);
1294 atomic_dec(&memcg->moving_account);
1299 * 2 routines for checking "mem" is under move_account() or not.
1301 * mem_cgroup_stolen() - checking whether a cgroup is mc.from or not. This
1302 * is used for avoiding races in accounting. If true,
1303 * pc->mem_cgroup may be overwritten.
1305 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1306 * under hierarchy of moving cgroups. This is for
1307 * waiting at hith-memory prressure caused by "move".
1310 static bool mem_cgroup_stolen(struct mem_cgroup *memcg)
1312 VM_BUG_ON(!rcu_read_lock_held());
1313 return atomic_read(&memcg->moving_account) > 0;
1316 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1318 struct mem_cgroup *from;
1319 struct mem_cgroup *to;
1322 * Unlike task_move routines, we access mc.to, mc.from not under
1323 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1325 spin_lock(&mc.lock);
1331 ret = mem_cgroup_same_or_subtree(memcg, from)
1332 || mem_cgroup_same_or_subtree(memcg, to);
1334 spin_unlock(&mc.lock);
1338 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1340 if (mc.moving_task && current != mc.moving_task) {
1341 if (mem_cgroup_under_move(memcg)) {
1343 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1344 /* moving charge context might have finished. */
1347 finish_wait(&mc.waitq, &wait);
1355 * Take this lock when
1356 * - a code tries to modify page's memcg while it's USED.
1357 * - a code tries to modify page state accounting in a memcg.
1358 * see mem_cgroup_stolen(), too.
1360 static void move_lock_mem_cgroup(struct mem_cgroup *memcg,
1361 unsigned long *flags)
1363 spin_lock_irqsave(&memcg->move_lock, *flags);
1366 static void move_unlock_mem_cgroup(struct mem_cgroup *memcg,
1367 unsigned long *flags)
1369 spin_unlock_irqrestore(&memcg->move_lock, *flags);
1373 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1374 * @memcg: The memory cgroup that went over limit
1375 * @p: Task that is going to be killed
1377 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1380 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1382 struct cgroup *task_cgrp;
1383 struct cgroup *mem_cgrp;
1385 * Need a buffer in BSS, can't rely on allocations. The code relies
1386 * on the assumption that OOM is serialized for memory controller.
1387 * If this assumption is broken, revisit this code.
1389 static char memcg_name[PATH_MAX];
1397 mem_cgrp = memcg->css.cgroup;
1398 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1400 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1403 * Unfortunately, we are unable to convert to a useful name
1404 * But we'll still print out the usage information
1411 printk(KERN_INFO "Task in %s killed", memcg_name);
1414 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1422 * Continues from above, so we don't need an KERN_ level
1424 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1427 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1428 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1429 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1430 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1431 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1433 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1434 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1435 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1439 * This function returns the number of memcg under hierarchy tree. Returns
1440 * 1(self count) if no children.
1442 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1445 struct mem_cgroup *iter;
1447 for_each_mem_cgroup_tree(iter, memcg)
1453 * Return the memory (and swap, if configured) limit for a memcg.
1455 static u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1460 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1461 limit += total_swap_pages << PAGE_SHIFT;
1463 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1465 * If memsw is finite and limits the amount of swap space available
1466 * to this memcg, return that limit.
1468 return min(limit, memsw);
1471 void mem_cgroup_out_of_memory(struct mem_cgroup *memcg, gfp_t gfp_mask,
1474 struct mem_cgroup *iter;
1475 unsigned long chosen_points = 0;
1476 unsigned long totalpages;
1477 unsigned int points = 0;
1478 struct task_struct *chosen = NULL;
1481 * If current has a pending SIGKILL, then automatically select it. The
1482 * goal is to allow it to allocate so that it may quickly exit and free
1485 if (fatal_signal_pending(current)) {
1486 set_thread_flag(TIF_MEMDIE);
1490 check_panic_on_oom(CONSTRAINT_MEMCG, gfp_mask, order, NULL);
1491 totalpages = mem_cgroup_get_limit(memcg) >> PAGE_SHIFT ? : 1;
1492 for_each_mem_cgroup_tree(iter, memcg) {
1493 struct cgroup *cgroup = iter->css.cgroup;
1494 struct cgroup_iter it;
1495 struct task_struct *task;
1497 cgroup_iter_start(cgroup, &it);
1498 while ((task = cgroup_iter_next(cgroup, &it))) {
1499 switch (oom_scan_process_thread(task, totalpages, NULL,
1501 case OOM_SCAN_SELECT:
1503 put_task_struct(chosen);
1505 chosen_points = ULONG_MAX;
1506 get_task_struct(chosen);
1508 case OOM_SCAN_CONTINUE:
1510 case OOM_SCAN_ABORT:
1511 cgroup_iter_end(cgroup, &it);
1512 mem_cgroup_iter_break(memcg, iter);
1514 put_task_struct(chosen);
1519 points = oom_badness(task, memcg, NULL, totalpages);
1520 if (points > chosen_points) {
1522 put_task_struct(chosen);
1524 chosen_points = points;
1525 get_task_struct(chosen);
1528 cgroup_iter_end(cgroup, &it);
1533 points = chosen_points * 1000 / totalpages;
1534 oom_kill_process(chosen, gfp_mask, order, points, totalpages, memcg,
1535 NULL, "Memory cgroup out of memory");
1538 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1540 unsigned long flags)
1542 unsigned long total = 0;
1543 bool noswap = false;
1546 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1548 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1551 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1553 drain_all_stock_async(memcg);
1554 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1556 * Allow limit shrinkers, which are triggered directly
1557 * by userspace, to catch signals and stop reclaim
1558 * after minimal progress, regardless of the margin.
1560 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1562 if (mem_cgroup_margin(memcg))
1565 * If nothing was reclaimed after two attempts, there
1566 * may be no reclaimable pages in this hierarchy.
1575 * test_mem_cgroup_node_reclaimable
1576 * @memcg: the target memcg
1577 * @nid: the node ID to be checked.
1578 * @noswap : specify true here if the user wants flle only information.
1580 * This function returns whether the specified memcg contains any
1581 * reclaimable pages on a node. Returns true if there are any reclaimable
1582 * pages in the node.
1584 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1585 int nid, bool noswap)
1587 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1589 if (noswap || !total_swap_pages)
1591 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1596 #if MAX_NUMNODES > 1
1599 * Always updating the nodemask is not very good - even if we have an empty
1600 * list or the wrong list here, we can start from some node and traverse all
1601 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1604 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1608 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1609 * pagein/pageout changes since the last update.
1611 if (!atomic_read(&memcg->numainfo_events))
1613 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1616 /* make a nodemask where this memcg uses memory from */
1617 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1619 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1621 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1622 node_clear(nid, memcg->scan_nodes);
1625 atomic_set(&memcg->numainfo_events, 0);
1626 atomic_set(&memcg->numainfo_updating, 0);
1630 * Selecting a node where we start reclaim from. Because what we need is just
1631 * reducing usage counter, start from anywhere is O,K. Considering
1632 * memory reclaim from current node, there are pros. and cons.
1634 * Freeing memory from current node means freeing memory from a node which
1635 * we'll use or we've used. So, it may make LRU bad. And if several threads
1636 * hit limits, it will see a contention on a node. But freeing from remote
1637 * node means more costs for memory reclaim because of memory latency.
1639 * Now, we use round-robin. Better algorithm is welcomed.
1641 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1645 mem_cgroup_may_update_nodemask(memcg);
1646 node = memcg->last_scanned_node;
1648 node = next_node(node, memcg->scan_nodes);
1649 if (node == MAX_NUMNODES)
1650 node = first_node(memcg->scan_nodes);
1652 * We call this when we hit limit, not when pages are added to LRU.
1653 * No LRU may hold pages because all pages are UNEVICTABLE or
1654 * memcg is too small and all pages are not on LRU. In that case,
1655 * we use curret node.
1657 if (unlikely(node == MAX_NUMNODES))
1658 node = numa_node_id();
1660 memcg->last_scanned_node = node;
1665 * Check all nodes whether it contains reclaimable pages or not.
1666 * For quick scan, we make use of scan_nodes. This will allow us to skip
1667 * unused nodes. But scan_nodes is lazily updated and may not cotain
1668 * enough new information. We need to do double check.
1670 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1675 * quick check...making use of scan_node.
1676 * We can skip unused nodes.
1678 if (!nodes_empty(memcg->scan_nodes)) {
1679 for (nid = first_node(memcg->scan_nodes);
1681 nid = next_node(nid, memcg->scan_nodes)) {
1683 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1688 * Check rest of nodes.
1690 for_each_node_state(nid, N_HIGH_MEMORY) {
1691 if (node_isset(nid, memcg->scan_nodes))
1693 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1700 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1705 static bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1707 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1711 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1714 unsigned long *total_scanned)
1716 struct mem_cgroup *victim = NULL;
1719 unsigned long excess;
1720 unsigned long nr_scanned;
1721 struct mem_cgroup_reclaim_cookie reclaim = {
1726 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1729 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1734 * If we have not been able to reclaim
1735 * anything, it might because there are
1736 * no reclaimable pages under this hierarchy
1741 * We want to do more targeted reclaim.
1742 * excess >> 2 is not to excessive so as to
1743 * reclaim too much, nor too less that we keep
1744 * coming back to reclaim from this cgroup
1746 if (total >= (excess >> 2) ||
1747 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1752 if (!mem_cgroup_reclaimable(victim, false))
1754 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1756 *total_scanned += nr_scanned;
1757 if (!res_counter_soft_limit_excess(&root_memcg->res))
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 *memcg;
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;
1848 struct mem_cgroup *oom_wait_memcg;
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->memcg;
1855 * Both of oom_wait_info->memcg and wake_memcg 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 static bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask,
1882 struct oom_wait_info owait;
1883 bool locked, need_to_kill;
1885 owait.memcg = memcg;
1886 owait.wait.flags = 0;
1887 owait.wait.func = memcg_oom_wake_function;
1888 owait.wait.private = current;
1889 INIT_LIST_HEAD(&owait.wait.task_list);
1890 need_to_kill = true;
1891 mem_cgroup_mark_under_oom(memcg);
1893 /* At first, try to OOM lock hierarchy under memcg.*/
1894 spin_lock(&memcg_oom_lock);
1895 locked = mem_cgroup_oom_lock(memcg);
1897 * Even if signal_pending(), we can't quit charge() loop without
1898 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1899 * under OOM is always welcomed, use TASK_KILLABLE here.
1901 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1902 if (!locked || memcg->oom_kill_disable)
1903 need_to_kill = false;
1905 mem_cgroup_oom_notify(memcg);
1906 spin_unlock(&memcg_oom_lock);
1909 finish_wait(&memcg_oom_waitq, &owait.wait);
1910 mem_cgroup_out_of_memory(memcg, mask, order);
1913 finish_wait(&memcg_oom_waitq, &owait.wait);
1915 spin_lock(&memcg_oom_lock);
1917 mem_cgroup_oom_unlock(memcg);
1918 memcg_wakeup_oom(memcg);
1919 spin_unlock(&memcg_oom_lock);
1921 mem_cgroup_unmark_under_oom(memcg);
1923 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1925 /* Give chance to dying process */
1926 schedule_timeout_uninterruptible(1);
1931 * Currently used to update mapped file statistics, but the routine can be
1932 * generalized to update other statistics as well.
1934 * Notes: Race condition
1936 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1937 * it tends to be costly. But considering some conditions, we doesn't need
1938 * to do so _always_.
1940 * Considering "charge", lock_page_cgroup() is not required because all
1941 * file-stat operations happen after a page is attached to radix-tree. There
1942 * are no race with "charge".
1944 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1945 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1946 * if there are race with "uncharge". Statistics itself is properly handled
1949 * Considering "move", this is an only case we see a race. To make the race
1950 * small, we check mm->moving_account and detect there are possibility of race
1951 * If there is, we take a lock.
1954 void __mem_cgroup_begin_update_page_stat(struct page *page,
1955 bool *locked, unsigned long *flags)
1957 struct mem_cgroup *memcg;
1958 struct page_cgroup *pc;
1960 pc = lookup_page_cgroup(page);
1962 memcg = pc->mem_cgroup;
1963 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1966 * If this memory cgroup is not under account moving, we don't
1967 * need to take move_lock_mem_cgroup(). Because we already hold
1968 * rcu_read_lock(), any calls to move_account will be delayed until
1969 * rcu_read_unlock() if mem_cgroup_stolen() == true.
1971 if (!mem_cgroup_stolen(memcg))
1974 move_lock_mem_cgroup(memcg, flags);
1975 if (memcg != pc->mem_cgroup || !PageCgroupUsed(pc)) {
1976 move_unlock_mem_cgroup(memcg, flags);
1982 void __mem_cgroup_end_update_page_stat(struct page *page, unsigned long *flags)
1984 struct page_cgroup *pc = lookup_page_cgroup(page);
1987 * It's guaranteed that pc->mem_cgroup never changes while
1988 * lock is held because a routine modifies pc->mem_cgroup
1989 * should take move_lock_mem_cgroup().
1991 move_unlock_mem_cgroup(pc->mem_cgroup, flags);
1994 void mem_cgroup_update_page_stat(struct page *page,
1995 enum mem_cgroup_page_stat_item idx, int val)
1997 struct mem_cgroup *memcg;
1998 struct page_cgroup *pc = lookup_page_cgroup(page);
1999 unsigned long uninitialized_var(flags);
2001 if (mem_cgroup_disabled())
2004 memcg = pc->mem_cgroup;
2005 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2009 case MEMCG_NR_FILE_MAPPED:
2010 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2016 this_cpu_add(memcg->stat->count[idx], val);
2020 * size of first charge trial. "32" comes from vmscan.c's magic value.
2021 * TODO: maybe necessary to use big numbers in big irons.
2023 #define CHARGE_BATCH 32U
2024 struct memcg_stock_pcp {
2025 struct mem_cgroup *cached; /* this never be root cgroup */
2026 unsigned int nr_pages;
2027 struct work_struct work;
2028 unsigned long flags;
2029 #define FLUSHING_CACHED_CHARGE 0
2031 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2032 static DEFINE_MUTEX(percpu_charge_mutex);
2035 * Try to consume stocked charge on this cpu. If success, one page is consumed
2036 * from local stock and true is returned. If the stock is 0 or charges from a
2037 * cgroup which is not current target, returns false. This stock will be
2040 static bool consume_stock(struct mem_cgroup *memcg)
2042 struct memcg_stock_pcp *stock;
2045 stock = &get_cpu_var(memcg_stock);
2046 if (memcg == stock->cached && stock->nr_pages)
2048 else /* need to call res_counter_charge */
2050 put_cpu_var(memcg_stock);
2055 * Returns stocks cached in percpu to res_counter and reset cached information.
2057 static void drain_stock(struct memcg_stock_pcp *stock)
2059 struct mem_cgroup *old = stock->cached;
2061 if (stock->nr_pages) {
2062 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2064 res_counter_uncharge(&old->res, bytes);
2065 if (do_swap_account)
2066 res_counter_uncharge(&old->memsw, bytes);
2067 stock->nr_pages = 0;
2069 stock->cached = NULL;
2073 * This must be called under preempt disabled or must be called by
2074 * a thread which is pinned to local cpu.
2076 static void drain_local_stock(struct work_struct *dummy)
2078 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2080 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2084 * Cache charges(val) which is from res_counter, to local per_cpu area.
2085 * This will be consumed by consume_stock() function, later.
2087 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2089 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2091 if (stock->cached != memcg) { /* reset if necessary */
2093 stock->cached = memcg;
2095 stock->nr_pages += nr_pages;
2096 put_cpu_var(memcg_stock);
2100 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2101 * of the hierarchy under it. sync flag says whether we should block
2102 * until the work is done.
2104 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2108 /* Notify other cpus that system-wide "drain" is running */
2111 for_each_online_cpu(cpu) {
2112 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2113 struct mem_cgroup *memcg;
2115 memcg = stock->cached;
2116 if (!memcg || !stock->nr_pages)
2118 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2120 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2122 drain_local_stock(&stock->work);
2124 schedule_work_on(cpu, &stock->work);
2132 for_each_online_cpu(cpu) {
2133 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2134 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2135 flush_work(&stock->work);
2142 * Tries to drain stocked charges in other cpus. This function is asynchronous
2143 * and just put a work per cpu for draining localy on each cpu. Caller can
2144 * expects some charges will be back to res_counter later but cannot wait for
2147 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2150 * If someone calls draining, avoid adding more kworker runs.
2152 if (!mutex_trylock(&percpu_charge_mutex))
2154 drain_all_stock(root_memcg, false);
2155 mutex_unlock(&percpu_charge_mutex);
2158 /* This is a synchronous drain interface. */
2159 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2161 /* called when force_empty is called */
2162 mutex_lock(&percpu_charge_mutex);
2163 drain_all_stock(root_memcg, true);
2164 mutex_unlock(&percpu_charge_mutex);
2168 * This function drains percpu counter value from DEAD cpu and
2169 * move it to local cpu. Note that this function can be preempted.
2171 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2175 spin_lock(&memcg->pcp_counter_lock);
2176 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
2177 long x = per_cpu(memcg->stat->count[i], cpu);
2179 per_cpu(memcg->stat->count[i], cpu) = 0;
2180 memcg->nocpu_base.count[i] += x;
2182 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2183 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2185 per_cpu(memcg->stat->events[i], cpu) = 0;
2186 memcg->nocpu_base.events[i] += x;
2188 spin_unlock(&memcg->pcp_counter_lock);
2191 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2192 unsigned long action,
2195 int cpu = (unsigned long)hcpu;
2196 struct memcg_stock_pcp *stock;
2197 struct mem_cgroup *iter;
2199 if (action == CPU_ONLINE)
2202 if (action != CPU_DEAD && action != CPU_DEAD_FROZEN)
2205 for_each_mem_cgroup(iter)
2206 mem_cgroup_drain_pcp_counter(iter, cpu);
2208 stock = &per_cpu(memcg_stock, cpu);
2214 /* See __mem_cgroup_try_charge() for details */
2216 CHARGE_OK, /* success */
2217 CHARGE_RETRY, /* need to retry but retry is not bad */
2218 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2219 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2220 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2223 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2224 unsigned int nr_pages, bool oom_check)
2226 unsigned long csize = nr_pages * PAGE_SIZE;
2227 struct mem_cgroup *mem_over_limit;
2228 struct res_counter *fail_res;
2229 unsigned long flags = 0;
2232 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2235 if (!do_swap_account)
2237 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2241 res_counter_uncharge(&memcg->res, csize);
2242 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2243 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2245 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2247 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2248 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2250 * Never reclaim on behalf of optional batching, retry with a
2251 * single page instead.
2253 if (nr_pages == CHARGE_BATCH)
2254 return CHARGE_RETRY;
2256 if (!(gfp_mask & __GFP_WAIT))
2257 return CHARGE_WOULDBLOCK;
2259 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2260 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2261 return CHARGE_RETRY;
2263 * Even though the limit is exceeded at this point, reclaim
2264 * may have been able to free some pages. Retry the charge
2265 * before killing the task.
2267 * Only for regular pages, though: huge pages are rather
2268 * unlikely to succeed so close to the limit, and we fall back
2269 * to regular pages anyway in case of failure.
2271 if (nr_pages == 1 && ret)
2272 return CHARGE_RETRY;
2275 * At task move, charge accounts can be doubly counted. So, it's
2276 * better to wait until the end of task_move if something is going on.
2278 if (mem_cgroup_wait_acct_move(mem_over_limit))
2279 return CHARGE_RETRY;
2281 /* If we don't need to call oom-killer at el, return immediately */
2283 return CHARGE_NOMEM;
2285 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask, get_order(csize)))
2286 return CHARGE_OOM_DIE;
2288 return CHARGE_RETRY;
2292 * __mem_cgroup_try_charge() does
2293 * 1. detect memcg to be charged against from passed *mm and *ptr,
2294 * 2. update res_counter
2295 * 3. call memory reclaim if necessary.
2297 * In some special case, if the task is fatal, fatal_signal_pending() or
2298 * has TIF_MEMDIE, this function returns -EINTR while writing root_mem_cgroup
2299 * to *ptr. There are two reasons for this. 1: fatal threads should quit as soon
2300 * as possible without any hazards. 2: all pages should have a valid
2301 * pc->mem_cgroup. If mm is NULL and the caller doesn't pass a valid memcg
2302 * pointer, that is treated as a charge to root_mem_cgroup.
2304 * So __mem_cgroup_try_charge() will return
2305 * 0 ... on success, filling *ptr with a valid memcg pointer.
2306 * -ENOMEM ... charge failure because of resource limits.
2307 * -EINTR ... if thread is fatal. *ptr is filled with root_mem_cgroup.
2309 * Unlike the exported interface, an "oom" parameter is added. if oom==true,
2310 * the oom-killer can be invoked.
2312 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2314 unsigned int nr_pages,
2315 struct mem_cgroup **ptr,
2318 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2319 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2320 struct mem_cgroup *memcg = NULL;
2324 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2325 * in system level. So, allow to go ahead dying process in addition to
2328 if (unlikely(test_thread_flag(TIF_MEMDIE)
2329 || fatal_signal_pending(current)))
2333 * We always charge the cgroup the mm_struct belongs to.
2334 * The mm_struct's mem_cgroup changes on task migration if the
2335 * thread group leader migrates. It's possible that mm is not
2336 * set, if so charge the root memcg (happens for pagecache usage).
2339 *ptr = root_mem_cgroup;
2341 if (*ptr) { /* css should be a valid one */
2343 VM_BUG_ON(css_is_removed(&memcg->css));
2344 if (mem_cgroup_is_root(memcg))
2346 if (nr_pages == 1 && consume_stock(memcg))
2348 css_get(&memcg->css);
2350 struct task_struct *p;
2353 p = rcu_dereference(mm->owner);
2355 * Because we don't have task_lock(), "p" can exit.
2356 * In that case, "memcg" can point to root or p can be NULL with
2357 * race with swapoff. Then, we have small risk of mis-accouning.
2358 * But such kind of mis-account by race always happens because
2359 * we don't have cgroup_mutex(). It's overkill and we allo that
2361 * (*) swapoff at el will charge against mm-struct not against
2362 * task-struct. So, mm->owner can be NULL.
2364 memcg = mem_cgroup_from_task(p);
2366 memcg = root_mem_cgroup;
2367 if (mem_cgroup_is_root(memcg)) {
2371 if (nr_pages == 1 && consume_stock(memcg)) {
2373 * It seems dagerous to access memcg without css_get().
2374 * But considering how consume_stok works, it's not
2375 * necessary. If consume_stock success, some charges
2376 * from this memcg are cached on this cpu. So, we
2377 * don't need to call css_get()/css_tryget() before
2378 * calling consume_stock().
2383 /* after here, we may be blocked. we need to get refcnt */
2384 if (!css_tryget(&memcg->css)) {
2394 /* If killed, bypass charge */
2395 if (fatal_signal_pending(current)) {
2396 css_put(&memcg->css);
2401 if (oom && !nr_oom_retries) {
2403 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2406 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2410 case CHARGE_RETRY: /* not in OOM situation but retry */
2412 css_put(&memcg->css);
2415 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2416 css_put(&memcg->css);
2418 case CHARGE_NOMEM: /* OOM routine works */
2420 css_put(&memcg->css);
2423 /* If oom, we never return -ENOMEM */
2426 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2427 css_put(&memcg->css);
2430 } while (ret != CHARGE_OK);
2432 if (batch > nr_pages)
2433 refill_stock(memcg, batch - nr_pages);
2434 css_put(&memcg->css);
2442 *ptr = root_mem_cgroup;
2447 * Somemtimes we have to undo a charge we got by try_charge().
2448 * This function is for that and do uncharge, put css's refcnt.
2449 * gotten by try_charge().
2451 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2452 unsigned int nr_pages)
2454 if (!mem_cgroup_is_root(memcg)) {
2455 unsigned long bytes = nr_pages * PAGE_SIZE;
2457 res_counter_uncharge(&memcg->res, bytes);
2458 if (do_swap_account)
2459 res_counter_uncharge(&memcg->memsw, bytes);
2464 * Cancel chrages in this cgroup....doesn't propagate to parent cgroup.
2465 * This is useful when moving usage to parent cgroup.
2467 static void __mem_cgroup_cancel_local_charge(struct mem_cgroup *memcg,
2468 unsigned int nr_pages)
2470 unsigned long bytes = nr_pages * PAGE_SIZE;
2472 if (mem_cgroup_is_root(memcg))
2475 res_counter_uncharge_until(&memcg->res, memcg->res.parent, bytes);
2476 if (do_swap_account)
2477 res_counter_uncharge_until(&memcg->memsw,
2478 memcg->memsw.parent, bytes);
2482 * A helper function to get mem_cgroup from ID. must be called under
2483 * rcu_read_lock(). The caller must check css_is_removed() or some if
2484 * it's concern. (dropping refcnt from swap can be called against removed
2487 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2489 struct cgroup_subsys_state *css;
2491 /* ID 0 is unused ID */
2494 css = css_lookup(&mem_cgroup_subsys, id);
2497 return container_of(css, struct mem_cgroup, css);
2500 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2502 struct mem_cgroup *memcg = NULL;
2503 struct page_cgroup *pc;
2507 VM_BUG_ON(!PageLocked(page));
2509 pc = lookup_page_cgroup(page);
2510 lock_page_cgroup(pc);
2511 if (PageCgroupUsed(pc)) {
2512 memcg = pc->mem_cgroup;
2513 if (memcg && !css_tryget(&memcg->css))
2515 } else if (PageSwapCache(page)) {
2516 ent.val = page_private(page);
2517 id = lookup_swap_cgroup_id(ent);
2519 memcg = mem_cgroup_lookup(id);
2520 if (memcg && !css_tryget(&memcg->css))
2524 unlock_page_cgroup(pc);
2528 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2530 unsigned int nr_pages,
2531 enum charge_type ctype,
2534 struct page_cgroup *pc = lookup_page_cgroup(page);
2535 struct zone *uninitialized_var(zone);
2536 struct lruvec *lruvec;
2537 bool was_on_lru = false;
2540 lock_page_cgroup(pc);
2541 VM_BUG_ON(PageCgroupUsed(pc));
2543 * we don't need page_cgroup_lock about tail pages, becase they are not
2544 * accessed by any other context at this point.
2548 * In some cases, SwapCache and FUSE(splice_buf->radixtree), the page
2549 * may already be on some other mem_cgroup's LRU. Take care of it.
2552 zone = page_zone(page);
2553 spin_lock_irq(&zone->lru_lock);
2554 if (PageLRU(page)) {
2555 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2557 del_page_from_lru_list(page, lruvec, page_lru(page));
2562 pc->mem_cgroup = memcg;
2564 * We access a page_cgroup asynchronously without lock_page_cgroup().
2565 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2566 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2567 * before USED bit, we need memory barrier here.
2568 * See mem_cgroup_add_lru_list(), etc.
2571 SetPageCgroupUsed(pc);
2575 lruvec = mem_cgroup_zone_lruvec(zone, pc->mem_cgroup);
2576 VM_BUG_ON(PageLRU(page));
2578 add_page_to_lru_list(page, lruvec, page_lru(page));
2580 spin_unlock_irq(&zone->lru_lock);
2583 if (ctype == MEM_CGROUP_CHARGE_TYPE_ANON)
2588 mem_cgroup_charge_statistics(memcg, anon, nr_pages);
2589 unlock_page_cgroup(pc);
2592 * "charge_statistics" updated event counter. Then, check it.
2593 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2594 * if they exceeds softlimit.
2596 memcg_check_events(memcg, page);
2599 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2601 #define PCGF_NOCOPY_AT_SPLIT (1 << PCG_LOCK | 1 << PCG_MIGRATION)
2603 * Because tail pages are not marked as "used", set it. We're under
2604 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2605 * charge/uncharge will be never happen and move_account() is done under
2606 * compound_lock(), so we don't have to take care of races.
2608 void mem_cgroup_split_huge_fixup(struct page *head)
2610 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2611 struct page_cgroup *pc;
2614 if (mem_cgroup_disabled())
2616 for (i = 1; i < HPAGE_PMD_NR; i++) {
2618 pc->mem_cgroup = head_pc->mem_cgroup;
2619 smp_wmb();/* see __commit_charge() */
2620 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2623 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2626 * mem_cgroup_move_account - move account of the page
2628 * @nr_pages: number of regular pages (>1 for huge pages)
2629 * @pc: page_cgroup of the page.
2630 * @from: mem_cgroup which the page is moved from.
2631 * @to: mem_cgroup which the page is moved to. @from != @to.
2633 * The caller must confirm following.
2634 * - page is not on LRU (isolate_page() is useful.)
2635 * - compound_lock is held when nr_pages > 1
2637 * This function doesn't do "charge" to new cgroup and doesn't do "uncharge"
2640 static int mem_cgroup_move_account(struct page *page,
2641 unsigned int nr_pages,
2642 struct page_cgroup *pc,
2643 struct mem_cgroup *from,
2644 struct mem_cgroup *to)
2646 unsigned long flags;
2648 bool anon = PageAnon(page);
2650 VM_BUG_ON(from == to);
2651 VM_BUG_ON(PageLRU(page));
2653 * The page is isolated from LRU. So, collapse function
2654 * will not handle this page. But page splitting can happen.
2655 * Do this check under compound_page_lock(). The caller should
2659 if (nr_pages > 1 && !PageTransHuge(page))
2662 lock_page_cgroup(pc);
2665 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2668 move_lock_mem_cgroup(from, &flags);
2670 if (!anon && page_mapped(page)) {
2671 /* Update mapped_file data for mem_cgroup */
2673 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2674 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2677 mem_cgroup_charge_statistics(from, anon, -nr_pages);
2679 /* caller should have done css_get */
2680 pc->mem_cgroup = to;
2681 mem_cgroup_charge_statistics(to, anon, nr_pages);
2683 * We charges against "to" which may not have any tasks. Then, "to"
2684 * can be under rmdir(). But in current implementation, caller of
2685 * this function is just force_empty() and move charge, so it's
2686 * guaranteed that "to" is never removed. So, we don't check rmdir
2689 move_unlock_mem_cgroup(from, &flags);
2692 unlock_page_cgroup(pc);
2696 memcg_check_events(to, page);
2697 memcg_check_events(from, page);
2703 * move charges to its parent.
2706 static int mem_cgroup_move_parent(struct page *page,
2707 struct page_cgroup *pc,
2708 struct mem_cgroup *child)
2710 struct mem_cgroup *parent;
2711 unsigned int nr_pages;
2712 unsigned long uninitialized_var(flags);
2716 if (mem_cgroup_is_root(child))
2720 if (!get_page_unless_zero(page))
2722 if (isolate_lru_page(page))
2725 nr_pages = hpage_nr_pages(page);
2727 parent = parent_mem_cgroup(child);
2729 * If no parent, move charges to root cgroup.
2732 parent = root_mem_cgroup;
2735 flags = compound_lock_irqsave(page);
2737 ret = mem_cgroup_move_account(page, nr_pages,
2740 __mem_cgroup_cancel_local_charge(child, nr_pages);
2743 compound_unlock_irqrestore(page, flags);
2744 putback_lru_page(page);
2752 * Charge the memory controller for page usage.
2754 * 0 if the charge was successful
2755 * < 0 if the cgroup is over its limit
2757 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2758 gfp_t gfp_mask, enum charge_type ctype)
2760 struct mem_cgroup *memcg = NULL;
2761 unsigned int nr_pages = 1;
2765 if (PageTransHuge(page)) {
2766 nr_pages <<= compound_order(page);
2767 VM_BUG_ON(!PageTransHuge(page));
2769 * Never OOM-kill a process for a huge page. The
2770 * fault handler will fall back to regular pages.
2775 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2778 __mem_cgroup_commit_charge(memcg, page, nr_pages, ctype, false);
2782 int mem_cgroup_newpage_charge(struct page *page,
2783 struct mm_struct *mm, gfp_t gfp_mask)
2785 if (mem_cgroup_disabled())
2787 VM_BUG_ON(page_mapped(page));
2788 VM_BUG_ON(page->mapping && !PageAnon(page));
2790 return mem_cgroup_charge_common(page, mm, gfp_mask,
2791 MEM_CGROUP_CHARGE_TYPE_ANON);
2795 * While swap-in, try_charge -> commit or cancel, the page is locked.
2796 * And when try_charge() successfully returns, one refcnt to memcg without
2797 * struct page_cgroup is acquired. This refcnt will be consumed by
2798 * "commit()" or removed by "cancel()"
2800 static int __mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2803 struct mem_cgroup **memcgp)
2805 struct mem_cgroup *memcg;
2806 struct page_cgroup *pc;
2809 pc = lookup_page_cgroup(page);
2811 * Every swap fault against a single page tries to charge the
2812 * page, bail as early as possible. shmem_unuse() encounters
2813 * already charged pages, too. The USED bit is protected by
2814 * the page lock, which serializes swap cache removal, which
2815 * in turn serializes uncharging.
2817 if (PageCgroupUsed(pc))
2819 if (!do_swap_account)
2822 * A racing thread's fault, or swapoff, may have already updated
2823 * the pte, and even removed page from swap cache: in those cases
2824 * do_swap_page()'s pte_same() test will fail; but there's also a
2825 * KSM case which does need to charge the page.
2827 if (!PageSwapCache(page))
2829 memcg = try_get_mem_cgroup_from_page(page);
2833 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2834 css_put(&memcg->css);
2839 ret = __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2845 int mem_cgroup_try_charge_swapin(struct mm_struct *mm, struct page *page,
2846 gfp_t gfp_mask, struct mem_cgroup **memcgp)
2849 if (mem_cgroup_disabled())
2851 return __mem_cgroup_try_charge_swapin(mm, page, gfp_mask, memcgp);
2854 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2856 if (mem_cgroup_disabled())
2860 __mem_cgroup_cancel_charge(memcg, 1);
2864 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2865 enum charge_type ctype)
2867 if (mem_cgroup_disabled())
2871 cgroup_exclude_rmdir(&memcg->css);
2873 __mem_cgroup_commit_charge(memcg, page, 1, ctype, true);
2875 * Now swap is on-memory. This means this page may be
2876 * counted both as mem and swap....double count.
2877 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2878 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2879 * may call delete_from_swap_cache() before reach here.
2881 if (do_swap_account && PageSwapCache(page)) {
2882 swp_entry_t ent = {.val = page_private(page)};
2883 mem_cgroup_uncharge_swap(ent);
2886 * At swapin, we may charge account against cgroup which has no tasks.
2887 * So, rmdir()->pre_destroy() can be called while we do this charge.
2888 * In that case, we need to call pre_destroy() again. check it here.
2890 cgroup_release_and_wakeup_rmdir(&memcg->css);
2893 void mem_cgroup_commit_charge_swapin(struct page *page,
2894 struct mem_cgroup *memcg)
2896 __mem_cgroup_commit_charge_swapin(page, memcg,
2897 MEM_CGROUP_CHARGE_TYPE_ANON);
2900 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2903 struct mem_cgroup *memcg = NULL;
2904 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2907 if (mem_cgroup_disabled())
2909 if (PageCompound(page))
2912 if (!PageSwapCache(page))
2913 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2914 else { /* page is swapcache/shmem */
2915 ret = __mem_cgroup_try_charge_swapin(mm, page,
2918 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2923 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2924 unsigned int nr_pages,
2925 const enum charge_type ctype)
2927 struct memcg_batch_info *batch = NULL;
2928 bool uncharge_memsw = true;
2930 /* If swapout, usage of swap doesn't decrease */
2931 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2932 uncharge_memsw = false;
2934 batch = ¤t->memcg_batch;
2936 * In usual, we do css_get() when we remember memcg pointer.
2937 * But in this case, we keep res->usage until end of a series of
2938 * uncharges. Then, it's ok to ignore memcg's refcnt.
2941 batch->memcg = memcg;
2943 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2944 * In those cases, all pages freed continuously can be expected to be in
2945 * the same cgroup and we have chance to coalesce uncharges.
2946 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2947 * because we want to do uncharge as soon as possible.
2950 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2951 goto direct_uncharge;
2954 goto direct_uncharge;
2957 * In typical case, batch->memcg == mem. This means we can
2958 * merge a series of uncharges to an uncharge of res_counter.
2959 * If not, we uncharge res_counter ony by one.
2961 if (batch->memcg != memcg)
2962 goto direct_uncharge;
2963 /* remember freed charge and uncharge it later */
2966 batch->memsw_nr_pages++;
2969 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2971 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2972 if (unlikely(batch->memcg != memcg))
2973 memcg_oom_recover(memcg);
2977 * uncharge if !page_mapped(page)
2979 static struct mem_cgroup *
2980 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype,
2983 struct mem_cgroup *memcg = NULL;
2984 unsigned int nr_pages = 1;
2985 struct page_cgroup *pc;
2988 if (mem_cgroup_disabled())
2991 VM_BUG_ON(PageSwapCache(page));
2993 if (PageTransHuge(page)) {
2994 nr_pages <<= compound_order(page);
2995 VM_BUG_ON(!PageTransHuge(page));
2998 * Check if our page_cgroup is valid
3000 pc = lookup_page_cgroup(page);
3001 if (unlikely(!PageCgroupUsed(pc)))
3004 lock_page_cgroup(pc);
3006 memcg = pc->mem_cgroup;
3008 if (!PageCgroupUsed(pc))
3011 anon = PageAnon(page);
3014 case MEM_CGROUP_CHARGE_TYPE_ANON:
3016 * Generally PageAnon tells if it's the anon statistics to be
3017 * updated; but sometimes e.g. mem_cgroup_uncharge_page() is
3018 * used before page reached the stage of being marked PageAnon.
3022 case MEM_CGROUP_CHARGE_TYPE_DROP:
3023 /* See mem_cgroup_prepare_migration() */
3024 if (page_mapped(page))
3027 * Pages under migration may not be uncharged. But
3028 * end_migration() /must/ be the one uncharging the
3029 * unused post-migration page and so it has to call
3030 * here with the migration bit still set. See the
3031 * res_counter handling below.
3033 if (!end_migration && PageCgroupMigration(pc))
3036 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3037 if (!PageAnon(page)) { /* Shared memory */
3038 if (page->mapping && !page_is_file_cache(page))
3040 } else if (page_mapped(page)) /* Anon */
3047 mem_cgroup_charge_statistics(memcg, anon, -nr_pages);
3049 ClearPageCgroupUsed(pc);
3051 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3052 * freed from LRU. This is safe because uncharged page is expected not
3053 * to be reused (freed soon). Exception is SwapCache, it's handled by
3054 * special functions.
3057 unlock_page_cgroup(pc);
3059 * even after unlock, we have memcg->res.usage here and this memcg
3060 * will never be freed.
3062 memcg_check_events(memcg, page);
3063 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3064 mem_cgroup_swap_statistics(memcg, true);
3065 mem_cgroup_get(memcg);
3068 * Migration does not charge the res_counter for the
3069 * replacement page, so leave it alone when phasing out the
3070 * page that is unused after the migration.
3072 if (!end_migration && !mem_cgroup_is_root(memcg))
3073 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3078 unlock_page_cgroup(pc);
3082 void mem_cgroup_uncharge_page(struct page *page)
3085 if (page_mapped(page))
3087 VM_BUG_ON(page->mapping && !PageAnon(page));
3088 if (PageSwapCache(page))
3090 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_ANON, false);
3093 void mem_cgroup_uncharge_cache_page(struct page *page)
3095 VM_BUG_ON(page_mapped(page));
3096 VM_BUG_ON(page->mapping);
3097 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE, false);
3101 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3102 * In that cases, pages are freed continuously and we can expect pages
3103 * are in the same memcg. All these calls itself limits the number of
3104 * pages freed at once, then uncharge_start/end() is called properly.
3105 * This may be called prural(2) times in a context,
3108 void mem_cgroup_uncharge_start(void)
3110 current->memcg_batch.do_batch++;
3111 /* We can do nest. */
3112 if (current->memcg_batch.do_batch == 1) {
3113 current->memcg_batch.memcg = NULL;
3114 current->memcg_batch.nr_pages = 0;
3115 current->memcg_batch.memsw_nr_pages = 0;
3119 void mem_cgroup_uncharge_end(void)
3121 struct memcg_batch_info *batch = ¤t->memcg_batch;
3123 if (!batch->do_batch)
3127 if (batch->do_batch) /* If stacked, do nothing. */
3133 * This "batch->memcg" is valid without any css_get/put etc...
3134 * bacause we hide charges behind us.
3136 if (batch->nr_pages)
3137 res_counter_uncharge(&batch->memcg->res,
3138 batch->nr_pages * PAGE_SIZE);
3139 if (batch->memsw_nr_pages)
3140 res_counter_uncharge(&batch->memcg->memsw,
3141 batch->memsw_nr_pages * PAGE_SIZE);
3142 memcg_oom_recover(batch->memcg);
3143 /* forget this pointer (for sanity check) */
3144 batch->memcg = NULL;
3149 * called after __delete_from_swap_cache() and drop "page" account.
3150 * memcg information is recorded to swap_cgroup of "ent"
3153 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3155 struct mem_cgroup *memcg;
3156 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3158 if (!swapout) /* this was a swap cache but the swap is unused ! */
3159 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3161 memcg = __mem_cgroup_uncharge_common(page, ctype, false);
3164 * record memcg information, if swapout && memcg != NULL,
3165 * mem_cgroup_get() was called in uncharge().
3167 if (do_swap_account && swapout && memcg)
3168 swap_cgroup_record(ent, css_id(&memcg->css));
3172 #ifdef CONFIG_MEMCG_SWAP
3174 * called from swap_entry_free(). remove record in swap_cgroup and
3175 * uncharge "memsw" account.
3177 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3179 struct mem_cgroup *memcg;
3182 if (!do_swap_account)
3185 id = swap_cgroup_record(ent, 0);
3187 memcg = mem_cgroup_lookup(id);
3190 * We uncharge this because swap is freed.
3191 * This memcg can be obsolete one. We avoid calling css_tryget
3193 if (!mem_cgroup_is_root(memcg))
3194 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3195 mem_cgroup_swap_statistics(memcg, false);
3196 mem_cgroup_put(memcg);
3202 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3203 * @entry: swap entry to be moved
3204 * @from: mem_cgroup which the entry is moved from
3205 * @to: mem_cgroup which the entry is moved to
3207 * It succeeds only when the swap_cgroup's record for this entry is the same
3208 * as the mem_cgroup's id of @from.
3210 * Returns 0 on success, -EINVAL on failure.
3212 * The caller must have charged to @to, IOW, called res_counter_charge() about
3213 * both res and memsw, and called css_get().
3215 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3216 struct mem_cgroup *from, struct mem_cgroup *to)
3218 unsigned short old_id, new_id;
3220 old_id = css_id(&from->css);
3221 new_id = css_id(&to->css);
3223 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3224 mem_cgroup_swap_statistics(from, false);
3225 mem_cgroup_swap_statistics(to, true);
3227 * This function is only called from task migration context now.
3228 * It postpones res_counter and refcount handling till the end
3229 * of task migration(mem_cgroup_clear_mc()) for performance
3230 * improvement. But we cannot postpone mem_cgroup_get(to)
3231 * because if the process that has been moved to @to does
3232 * swap-in, the refcount of @to might be decreased to 0.
3240 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3241 struct mem_cgroup *from, struct mem_cgroup *to)
3248 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3251 void mem_cgroup_prepare_migration(struct page *page, struct page *newpage,
3252 struct mem_cgroup **memcgp)
3254 struct mem_cgroup *memcg = NULL;
3255 struct page_cgroup *pc;
3256 enum charge_type ctype;
3260 VM_BUG_ON(PageTransHuge(page));
3261 if (mem_cgroup_disabled())
3264 pc = lookup_page_cgroup(page);
3265 lock_page_cgroup(pc);
3266 if (PageCgroupUsed(pc)) {
3267 memcg = pc->mem_cgroup;
3268 css_get(&memcg->css);
3270 * At migrating an anonymous page, its mapcount goes down
3271 * to 0 and uncharge() will be called. But, even if it's fully
3272 * unmapped, migration may fail and this page has to be
3273 * charged again. We set MIGRATION flag here and delay uncharge
3274 * until end_migration() is called
3276 * Corner Case Thinking
3278 * When the old page was mapped as Anon and it's unmap-and-freed
3279 * while migration was ongoing.
3280 * If unmap finds the old page, uncharge() of it will be delayed
3281 * until end_migration(). If unmap finds a new page, it's
3282 * uncharged when it make mapcount to be 1->0. If unmap code
3283 * finds swap_migration_entry, the new page will not be mapped
3284 * and end_migration() will find it(mapcount==0).
3287 * When the old page was mapped but migraion fails, the kernel
3288 * remaps it. A charge for it is kept by MIGRATION flag even
3289 * if mapcount goes down to 0. We can do remap successfully
3290 * without charging it again.
3293 * The "old" page is under lock_page() until the end of
3294 * migration, so, the old page itself will not be swapped-out.
3295 * If the new page is swapped out before end_migraton, our
3296 * hook to usual swap-out path will catch the event.
3299 SetPageCgroupMigration(pc);
3301 unlock_page_cgroup(pc);
3303 * If the page is not charged at this point,
3311 * We charge new page before it's used/mapped. So, even if unlock_page()
3312 * is called before end_migration, we can catch all events on this new
3313 * page. In the case new page is migrated but not remapped, new page's
3314 * mapcount will be finally 0 and we call uncharge in end_migration().
3317 ctype = MEM_CGROUP_CHARGE_TYPE_ANON;
3319 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3321 * The page is committed to the memcg, but it's not actually
3322 * charged to the res_counter since we plan on replacing the
3323 * old one and only one page is going to be left afterwards.
3325 __mem_cgroup_commit_charge(memcg, newpage, 1, ctype, false);
3328 /* remove redundant charge if migration failed*/
3329 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3330 struct page *oldpage, struct page *newpage, bool migration_ok)
3332 struct page *used, *unused;
3333 struct page_cgroup *pc;
3338 /* blocks rmdir() */
3339 cgroup_exclude_rmdir(&memcg->css);
3340 if (!migration_ok) {
3347 anon = PageAnon(used);
3348 __mem_cgroup_uncharge_common(unused,
3349 anon ? MEM_CGROUP_CHARGE_TYPE_ANON
3350 : MEM_CGROUP_CHARGE_TYPE_CACHE,
3352 css_put(&memcg->css);
3354 * We disallowed uncharge of pages under migration because mapcount
3355 * of the page goes down to zero, temporarly.
3356 * Clear the flag and check the page should be charged.
3358 pc = lookup_page_cgroup(oldpage);
3359 lock_page_cgroup(pc);
3360 ClearPageCgroupMigration(pc);
3361 unlock_page_cgroup(pc);
3364 * If a page is a file cache, radix-tree replacement is very atomic
3365 * and we can skip this check. When it was an Anon page, its mapcount
3366 * goes down to 0. But because we added MIGRATION flage, it's not
3367 * uncharged yet. There are several case but page->mapcount check
3368 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3369 * check. (see prepare_charge() also)
3372 mem_cgroup_uncharge_page(used);
3374 * At migration, we may charge account against cgroup which has no
3376 * So, rmdir()->pre_destroy() can be called while we do this charge.
3377 * In that case, we need to call pre_destroy() again. check it here.
3379 cgroup_release_and_wakeup_rmdir(&memcg->css);
3383 * At replace page cache, newpage is not under any memcg but it's on
3384 * LRU. So, this function doesn't touch res_counter but handles LRU
3385 * in correct way. Both pages are locked so we cannot race with uncharge.
3387 void mem_cgroup_replace_page_cache(struct page *oldpage,
3388 struct page *newpage)
3390 struct mem_cgroup *memcg = NULL;
3391 struct page_cgroup *pc;
3392 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3394 if (mem_cgroup_disabled())
3397 pc = lookup_page_cgroup(oldpage);
3398 /* fix accounting on old pages */
3399 lock_page_cgroup(pc);
3400 if (PageCgroupUsed(pc)) {
3401 memcg = pc->mem_cgroup;
3402 mem_cgroup_charge_statistics(memcg, false, -1);
3403 ClearPageCgroupUsed(pc);
3405 unlock_page_cgroup(pc);
3408 * When called from shmem_replace_page(), in some cases the
3409 * oldpage has already been charged, and in some cases not.
3414 * Even if newpage->mapping was NULL before starting replacement,
3415 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3416 * LRU while we overwrite pc->mem_cgroup.
3418 __mem_cgroup_commit_charge(memcg, newpage, 1, type, true);
3421 #ifdef CONFIG_DEBUG_VM
3422 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3424 struct page_cgroup *pc;
3426 pc = lookup_page_cgroup(page);
3428 * Can be NULL while feeding pages into the page allocator for
3429 * the first time, i.e. during boot or memory hotplug;
3430 * or when mem_cgroup_disabled().
3432 if (likely(pc) && PageCgroupUsed(pc))
3437 bool mem_cgroup_bad_page_check(struct page *page)
3439 if (mem_cgroup_disabled())
3442 return lookup_page_cgroup_used(page) != NULL;
3445 void mem_cgroup_print_bad_page(struct page *page)
3447 struct page_cgroup *pc;
3449 pc = lookup_page_cgroup_used(page);
3451 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p\n",
3452 pc, pc->flags, pc->mem_cgroup);
3457 static DEFINE_MUTEX(set_limit_mutex);
3459 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3460 unsigned long long val)
3463 u64 memswlimit, memlimit;
3465 int children = mem_cgroup_count_children(memcg);
3466 u64 curusage, oldusage;
3470 * For keeping hierarchical_reclaim simple, how long we should retry
3471 * is depends on callers. We set our retry-count to be function
3472 * of # of children which we should visit in this loop.
3474 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3476 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3479 while (retry_count) {
3480 if (signal_pending(current)) {
3485 * Rather than hide all in some function, I do this in
3486 * open coded manner. You see what this really does.
3487 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3489 mutex_lock(&set_limit_mutex);
3490 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3491 if (memswlimit < val) {
3493 mutex_unlock(&set_limit_mutex);
3497 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3501 ret = res_counter_set_limit(&memcg->res, val);
3503 if (memswlimit == val)
3504 memcg->memsw_is_minimum = true;
3506 memcg->memsw_is_minimum = false;
3508 mutex_unlock(&set_limit_mutex);
3513 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3514 MEM_CGROUP_RECLAIM_SHRINK);
3515 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3516 /* Usage is reduced ? */
3517 if (curusage >= oldusage)
3520 oldusage = curusage;
3522 if (!ret && enlarge)
3523 memcg_oom_recover(memcg);
3528 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3529 unsigned long long val)
3532 u64 memlimit, memswlimit, oldusage, curusage;
3533 int children = mem_cgroup_count_children(memcg);
3537 /* see mem_cgroup_resize_res_limit */
3538 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3539 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3540 while (retry_count) {
3541 if (signal_pending(current)) {
3546 * Rather than hide all in some function, I do this in
3547 * open coded manner. You see what this really does.
3548 * We have to guarantee memcg->res.limit <= memcg->memsw.limit.
3550 mutex_lock(&set_limit_mutex);
3551 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3552 if (memlimit > val) {
3554 mutex_unlock(&set_limit_mutex);
3557 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3558 if (memswlimit < val)
3560 ret = res_counter_set_limit(&memcg->memsw, val);
3562 if (memlimit == val)
3563 memcg->memsw_is_minimum = true;
3565 memcg->memsw_is_minimum = false;
3567 mutex_unlock(&set_limit_mutex);
3572 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3573 MEM_CGROUP_RECLAIM_NOSWAP |
3574 MEM_CGROUP_RECLAIM_SHRINK);
3575 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3576 /* Usage is reduced ? */
3577 if (curusage >= oldusage)
3580 oldusage = curusage;
3582 if (!ret && enlarge)
3583 memcg_oom_recover(memcg);
3587 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3589 unsigned long *total_scanned)
3591 unsigned long nr_reclaimed = 0;
3592 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3593 unsigned long reclaimed;
3595 struct mem_cgroup_tree_per_zone *mctz;
3596 unsigned long long excess;
3597 unsigned long nr_scanned;
3602 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3604 * This loop can run a while, specially if mem_cgroup's continuously
3605 * keep exceeding their soft limit and putting the system under
3612 mz = mem_cgroup_largest_soft_limit_node(mctz);
3617 reclaimed = mem_cgroup_soft_reclaim(mz->memcg, zone,
3618 gfp_mask, &nr_scanned);
3619 nr_reclaimed += reclaimed;
3620 *total_scanned += nr_scanned;
3621 spin_lock(&mctz->lock);
3624 * If we failed to reclaim anything from this memory cgroup
3625 * it is time to move on to the next cgroup
3631 * Loop until we find yet another one.
3633 * By the time we get the soft_limit lock
3634 * again, someone might have aded the
3635 * group back on the RB tree. Iterate to
3636 * make sure we get a different mem.
3637 * mem_cgroup_largest_soft_limit_node returns
3638 * NULL if no other cgroup is present on
3642 __mem_cgroup_largest_soft_limit_node(mctz);
3644 css_put(&next_mz->memcg->css);
3645 else /* next_mz == NULL or other memcg */
3649 __mem_cgroup_remove_exceeded(mz->memcg, mz, mctz);
3650 excess = res_counter_soft_limit_excess(&mz->memcg->res);
3652 * One school of thought says that we should not add
3653 * back the node to the tree if reclaim returns 0.
3654 * But our reclaim could return 0, simply because due
3655 * to priority we are exposing a smaller subset of
3656 * memory to reclaim from. Consider this as a longer
3659 /* If excess == 0, no tree ops */
3660 __mem_cgroup_insert_exceeded(mz->memcg, mz, mctz, excess);
3661 spin_unlock(&mctz->lock);
3662 css_put(&mz->memcg->css);
3665 * Could not reclaim anything and there are no more
3666 * mem cgroups to try or we seem to be looping without
3667 * reclaiming anything.
3669 if (!nr_reclaimed &&
3671 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3673 } while (!nr_reclaimed);
3675 css_put(&next_mz->memcg->css);
3676 return nr_reclaimed;
3680 * Traverse a specified page_cgroup list and try to drop them all. This doesn't
3681 * reclaim the pages page themselves - it just removes the page_cgroups.
3682 * Returns true if some page_cgroups were not freed, indicating that the caller
3683 * must retry this operation.
3685 static bool mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3686 int node, int zid, enum lru_list lru)
3688 struct mem_cgroup_per_zone *mz;
3689 unsigned long flags, loop;
3690 struct list_head *list;
3694 zone = &NODE_DATA(node)->node_zones[zid];
3695 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3696 list = &mz->lruvec.lists[lru];
3698 loop = mz->lru_size[lru];
3699 /* give some margin against EBUSY etc...*/
3703 struct page_cgroup *pc;
3706 spin_lock_irqsave(&zone->lru_lock, flags);
3707 if (list_empty(list)) {
3708 spin_unlock_irqrestore(&zone->lru_lock, flags);
3711 page = list_entry(list->prev, struct page, lru);
3713 list_move(&page->lru, list);
3715 spin_unlock_irqrestore(&zone->lru_lock, flags);
3718 spin_unlock_irqrestore(&zone->lru_lock, flags);
3720 pc = lookup_page_cgroup(page);
3722 if (mem_cgroup_move_parent(page, pc, memcg)) {
3723 /* found lock contention or "pc" is obsolete. */
3729 return !list_empty(list);
3733 * make mem_cgroup's charge to be 0 if there is no task.
3734 * This enables deleting this mem_cgroup.
3736 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3739 int node, zid, shrink;
3740 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3741 struct cgroup *cgrp = memcg->css.cgroup;
3743 css_get(&memcg->css);
3746 /* should free all ? */
3752 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3754 /* This is for making all *used* pages to be on LRU. */
3755 lru_add_drain_all();
3756 drain_all_stock_sync(memcg);
3758 mem_cgroup_start_move(memcg);
3759 for_each_node_state(node, N_HIGH_MEMORY) {
3760 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3763 ret = mem_cgroup_force_empty_list(memcg,
3772 mem_cgroup_end_move(memcg);
3773 memcg_oom_recover(memcg);
3775 /* "ret" should also be checked to ensure all lists are empty. */
3776 } while (res_counter_read_u64(&memcg->res, RES_USAGE) > 0 || ret);
3778 css_put(&memcg->css);
3782 /* returns EBUSY if there is a task or if we come here twice. */
3783 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3787 /* we call try-to-free pages for make this cgroup empty */
3788 lru_add_drain_all();
3789 /* try to free all pages in this cgroup */
3791 while (nr_retries && res_counter_read_u64(&memcg->res, RES_USAGE) > 0) {
3794 if (signal_pending(current)) {
3798 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3802 /* maybe some writeback is necessary */
3803 congestion_wait(BLK_RW_ASYNC, HZ/10);
3808 /* try move_account...there may be some *locked* pages. */
3812 static int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3814 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3818 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3820 return mem_cgroup_from_cont(cont)->use_hierarchy;
3823 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3827 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3828 struct cgroup *parent = cont->parent;
3829 struct mem_cgroup *parent_memcg = NULL;
3832 parent_memcg = mem_cgroup_from_cont(parent);
3836 if (memcg->use_hierarchy == val)
3840 * If parent's use_hierarchy is set, we can't make any modifications
3841 * in the child subtrees. If it is unset, then the change can
3842 * occur, provided the current cgroup has no children.
3844 * For the root cgroup, parent_mem is NULL, we allow value to be
3845 * set if there are no children.
3847 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3848 (val == 1 || val == 0)) {
3849 if (list_empty(&cont->children))
3850 memcg->use_hierarchy = val;
3863 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3864 enum mem_cgroup_stat_index idx)
3866 struct mem_cgroup *iter;
3869 /* Per-cpu values can be negative, use a signed accumulator */
3870 for_each_mem_cgroup_tree(iter, memcg)
3871 val += mem_cgroup_read_stat(iter, idx);
3873 if (val < 0) /* race ? */
3878 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3882 if (!mem_cgroup_is_root(memcg)) {
3884 return res_counter_read_u64(&memcg->res, RES_USAGE);
3886 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3889 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3890 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3893 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAP);
3895 return val << PAGE_SHIFT;
3898 static ssize_t mem_cgroup_read(struct cgroup *cont, struct cftype *cft,
3899 struct file *file, char __user *buf,
3900 size_t nbytes, loff_t *ppos)
3902 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3905 int type, name, len;
3907 type = MEMFILE_TYPE(cft->private);
3908 name = MEMFILE_ATTR(cft->private);
3910 if (!do_swap_account && type == _MEMSWAP)
3915 if (name == RES_USAGE)
3916 val = mem_cgroup_usage(memcg, false);
3918 val = res_counter_read_u64(&memcg->res, name);
3921 if (name == RES_USAGE)
3922 val = mem_cgroup_usage(memcg, true);
3924 val = res_counter_read_u64(&memcg->memsw, name);
3930 len = scnprintf(str, sizeof(str), "%llu\n", (unsigned long long)val);
3931 return simple_read_from_buffer(buf, nbytes, ppos, str, len);
3934 * The user of this function is...
3937 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3940 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3942 unsigned long long val;
3945 type = MEMFILE_TYPE(cft->private);
3946 name = MEMFILE_ATTR(cft->private);
3948 if (!do_swap_account && type == _MEMSWAP)
3953 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3957 /* This function does all necessary parse...reuse it */
3958 ret = res_counter_memparse_write_strategy(buffer, &val);
3962 ret = mem_cgroup_resize_limit(memcg, val);
3964 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3966 case RES_SOFT_LIMIT:
3967 ret = res_counter_memparse_write_strategy(buffer, &val);
3971 * For memsw, soft limits are hard to implement in terms
3972 * of semantics, for now, we support soft limits for
3973 * control without swap
3976 ret = res_counter_set_soft_limit(&memcg->res, val);
3981 ret = -EINVAL; /* should be BUG() ? */
3987 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3988 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3990 struct cgroup *cgroup;
3991 unsigned long long min_limit, min_memsw_limit, tmp;
3993 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3994 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3995 cgroup = memcg->css.cgroup;
3996 if (!memcg->use_hierarchy)
3999 while (cgroup->parent) {
4000 cgroup = cgroup->parent;
4001 memcg = mem_cgroup_from_cont(cgroup);
4002 if (!memcg->use_hierarchy)
4004 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4005 min_limit = min(min_limit, tmp);
4006 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4007 min_memsw_limit = min(min_memsw_limit, tmp);
4010 *mem_limit = min_limit;
4011 *memsw_limit = min_memsw_limit;
4014 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4016 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4019 type = MEMFILE_TYPE(event);
4020 name = MEMFILE_ATTR(event);
4022 if (!do_swap_account && type == _MEMSWAP)
4028 res_counter_reset_max(&memcg->res);
4030 res_counter_reset_max(&memcg->memsw);
4034 res_counter_reset_failcnt(&memcg->res);
4036 res_counter_reset_failcnt(&memcg->memsw);
4043 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4046 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4050 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4051 struct cftype *cft, u64 val)
4053 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4055 if (val >= (1 << NR_MOVE_TYPE))
4058 * We check this value several times in both in can_attach() and
4059 * attach(), so we need cgroup lock to prevent this value from being
4063 memcg->move_charge_at_immigrate = val;
4069 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4070 struct cftype *cft, u64 val)
4077 static int memcg_numa_stat_show(struct cgroup *cont, struct cftype *cft,
4081 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4082 unsigned long node_nr;
4083 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4085 total_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL);
4086 seq_printf(m, "total=%lu", total_nr);
4087 for_each_node_state(nid, N_HIGH_MEMORY) {
4088 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL);
4089 seq_printf(m, " N%d=%lu", nid, node_nr);
4093 file_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_FILE);
4094 seq_printf(m, "file=%lu", file_nr);
4095 for_each_node_state(nid, N_HIGH_MEMORY) {
4096 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4098 seq_printf(m, " N%d=%lu", nid, node_nr);
4102 anon_nr = mem_cgroup_nr_lru_pages(memcg, LRU_ALL_ANON);
4103 seq_printf(m, "anon=%lu", anon_nr);
4104 for_each_node_state(nid, N_HIGH_MEMORY) {
4105 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4107 seq_printf(m, " N%d=%lu", nid, node_nr);
4111 unevictable_nr = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4112 seq_printf(m, "unevictable=%lu", unevictable_nr);
4113 for_each_node_state(nid, N_HIGH_MEMORY) {
4114 node_nr = mem_cgroup_node_nr_lru_pages(memcg, nid,
4115 BIT(LRU_UNEVICTABLE));
4116 seq_printf(m, " N%d=%lu", nid, node_nr);
4121 #endif /* CONFIG_NUMA */
4123 static const char * const mem_cgroup_lru_names[] = {
4131 static inline void mem_cgroup_lru_names_not_uptodate(void)
4133 BUILD_BUG_ON(ARRAY_SIZE(mem_cgroup_lru_names) != NR_LRU_LISTS);
4136 static int memcg_stat_show(struct cgroup *cont, struct cftype *cft,
4139 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4140 struct mem_cgroup *mi;
4143 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4144 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4146 seq_printf(m, "%s %ld\n", mem_cgroup_stat_names[i],
4147 mem_cgroup_read_stat(memcg, i) * PAGE_SIZE);
4150 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++)
4151 seq_printf(m, "%s %lu\n", mem_cgroup_events_names[i],
4152 mem_cgroup_read_events(memcg, i));
4154 for (i = 0; i < NR_LRU_LISTS; i++)
4155 seq_printf(m, "%s %lu\n", mem_cgroup_lru_names[i],
4156 mem_cgroup_nr_lru_pages(memcg, BIT(i)) * PAGE_SIZE);
4158 /* Hierarchical information */
4160 unsigned long long limit, memsw_limit;
4161 memcg_get_hierarchical_limit(memcg, &limit, &memsw_limit);
4162 seq_printf(m, "hierarchical_memory_limit %llu\n", limit);
4163 if (do_swap_account)
4164 seq_printf(m, "hierarchical_memsw_limit %llu\n",
4168 for (i = 0; i < MEM_CGROUP_STAT_NSTATS; i++) {
4171 if (i == MEM_CGROUP_STAT_SWAP && !do_swap_account)
4173 for_each_mem_cgroup_tree(mi, memcg)
4174 val += mem_cgroup_read_stat(mi, i) * PAGE_SIZE;
4175 seq_printf(m, "total_%s %lld\n", mem_cgroup_stat_names[i], val);
4178 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
4179 unsigned long long val = 0;
4181 for_each_mem_cgroup_tree(mi, memcg)
4182 val += mem_cgroup_read_events(mi, i);
4183 seq_printf(m, "total_%s %llu\n",
4184 mem_cgroup_events_names[i], val);
4187 for (i = 0; i < NR_LRU_LISTS; i++) {
4188 unsigned long long val = 0;
4190 for_each_mem_cgroup_tree(mi, memcg)
4191 val += mem_cgroup_nr_lru_pages(mi, BIT(i)) * PAGE_SIZE;
4192 seq_printf(m, "total_%s %llu\n", mem_cgroup_lru_names[i], val);
4195 #ifdef CONFIG_DEBUG_VM
4198 struct mem_cgroup_per_zone *mz;
4199 struct zone_reclaim_stat *rstat;
4200 unsigned long recent_rotated[2] = {0, 0};
4201 unsigned long recent_scanned[2] = {0, 0};
4203 for_each_online_node(nid)
4204 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4205 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
4206 rstat = &mz->lruvec.reclaim_stat;
4208 recent_rotated[0] += rstat->recent_rotated[0];
4209 recent_rotated[1] += rstat->recent_rotated[1];
4210 recent_scanned[0] += rstat->recent_scanned[0];
4211 recent_scanned[1] += rstat->recent_scanned[1];
4213 seq_printf(m, "recent_rotated_anon %lu\n", recent_rotated[0]);
4214 seq_printf(m, "recent_rotated_file %lu\n", recent_rotated[1]);
4215 seq_printf(m, "recent_scanned_anon %lu\n", recent_scanned[0]);
4216 seq_printf(m, "recent_scanned_file %lu\n", recent_scanned[1]);
4223 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4225 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4227 return mem_cgroup_swappiness(memcg);
4230 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4233 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4234 struct mem_cgroup *parent;
4239 if (cgrp->parent == NULL)
4242 parent = mem_cgroup_from_cont(cgrp->parent);
4246 /* If under hierarchy, only empty-root can set this value */
4247 if ((parent->use_hierarchy) ||
4248 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4253 memcg->swappiness = val;
4260 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4262 struct mem_cgroup_threshold_ary *t;
4268 t = rcu_dereference(memcg->thresholds.primary);
4270 t = rcu_dereference(memcg->memsw_thresholds.primary);
4275 usage = mem_cgroup_usage(memcg, swap);
4278 * current_threshold points to threshold just below or equal to usage.
4279 * If it's not true, a threshold was crossed after last
4280 * call of __mem_cgroup_threshold().
4282 i = t->current_threshold;
4285 * Iterate backward over array of thresholds starting from
4286 * current_threshold and check if a threshold is crossed.
4287 * If none of thresholds below usage is crossed, we read
4288 * only one element of the array here.
4290 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4291 eventfd_signal(t->entries[i].eventfd, 1);
4293 /* i = current_threshold + 1 */
4297 * Iterate forward over array of thresholds starting from
4298 * current_threshold+1 and check if a threshold is crossed.
4299 * If none of thresholds above usage is crossed, we read
4300 * only one element of the array here.
4302 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4303 eventfd_signal(t->entries[i].eventfd, 1);
4305 /* Update current_threshold */
4306 t->current_threshold = i - 1;
4311 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4314 __mem_cgroup_threshold(memcg, false);
4315 if (do_swap_account)
4316 __mem_cgroup_threshold(memcg, true);
4318 memcg = parent_mem_cgroup(memcg);
4322 static int compare_thresholds(const void *a, const void *b)
4324 const struct mem_cgroup_threshold *_a = a;
4325 const struct mem_cgroup_threshold *_b = b;
4327 return _a->threshold - _b->threshold;
4330 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4332 struct mem_cgroup_eventfd_list *ev;
4334 list_for_each_entry(ev, &memcg->oom_notify, list)
4335 eventfd_signal(ev->eventfd, 1);
4339 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4341 struct mem_cgroup *iter;
4343 for_each_mem_cgroup_tree(iter, memcg)
4344 mem_cgroup_oom_notify_cb(iter);
4347 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4348 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4350 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4351 struct mem_cgroup_thresholds *thresholds;
4352 struct mem_cgroup_threshold_ary *new;
4353 int type = MEMFILE_TYPE(cft->private);
4354 u64 threshold, usage;
4357 ret = res_counter_memparse_write_strategy(args, &threshold);
4361 mutex_lock(&memcg->thresholds_lock);
4364 thresholds = &memcg->thresholds;
4365 else if (type == _MEMSWAP)
4366 thresholds = &memcg->memsw_thresholds;
4370 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4372 /* Check if a threshold crossed before adding a new one */
4373 if (thresholds->primary)
4374 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4376 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4378 /* Allocate memory for new array of thresholds */
4379 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4387 /* Copy thresholds (if any) to new array */
4388 if (thresholds->primary) {
4389 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4390 sizeof(struct mem_cgroup_threshold));
4393 /* Add new threshold */
4394 new->entries[size - 1].eventfd = eventfd;
4395 new->entries[size - 1].threshold = threshold;
4397 /* Sort thresholds. Registering of new threshold isn't time-critical */
4398 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4399 compare_thresholds, NULL);
4401 /* Find current threshold */
4402 new->current_threshold = -1;
4403 for (i = 0; i < size; i++) {
4404 if (new->entries[i].threshold <= usage) {
4406 * new->current_threshold will not be used until
4407 * rcu_assign_pointer(), so it's safe to increment
4410 ++new->current_threshold;
4415 /* Free old spare buffer and save old primary buffer as spare */
4416 kfree(thresholds->spare);
4417 thresholds->spare = thresholds->primary;
4419 rcu_assign_pointer(thresholds->primary, new);
4421 /* To be sure that nobody uses thresholds */
4425 mutex_unlock(&memcg->thresholds_lock);
4430 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4431 struct cftype *cft, struct eventfd_ctx *eventfd)
4433 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4434 struct mem_cgroup_thresholds *thresholds;
4435 struct mem_cgroup_threshold_ary *new;
4436 int type = MEMFILE_TYPE(cft->private);
4440 mutex_lock(&memcg->thresholds_lock);
4442 thresholds = &memcg->thresholds;
4443 else if (type == _MEMSWAP)
4444 thresholds = &memcg->memsw_thresholds;
4448 if (!thresholds->primary)
4451 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4453 /* Check if a threshold crossed before removing */
4454 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4456 /* Calculate new number of threshold */
4458 for (i = 0; i < thresholds->primary->size; i++) {
4459 if (thresholds->primary->entries[i].eventfd != eventfd)
4463 new = thresholds->spare;
4465 /* Set thresholds array to NULL if we don't have thresholds */
4474 /* Copy thresholds and find current threshold */
4475 new->current_threshold = -1;
4476 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4477 if (thresholds->primary->entries[i].eventfd == eventfd)
4480 new->entries[j] = thresholds->primary->entries[i];
4481 if (new->entries[j].threshold <= usage) {
4483 * new->current_threshold will not be used
4484 * until rcu_assign_pointer(), so it's safe to increment
4487 ++new->current_threshold;
4493 /* Swap primary and spare array */
4494 thresholds->spare = thresholds->primary;
4495 /* If all events are unregistered, free the spare array */
4497 kfree(thresholds->spare);
4498 thresholds->spare = NULL;
4501 rcu_assign_pointer(thresholds->primary, new);
4503 /* To be sure that nobody uses thresholds */
4506 mutex_unlock(&memcg->thresholds_lock);
4509 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4510 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4512 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4513 struct mem_cgroup_eventfd_list *event;
4514 int type = MEMFILE_TYPE(cft->private);
4516 BUG_ON(type != _OOM_TYPE);
4517 event = kmalloc(sizeof(*event), GFP_KERNEL);
4521 spin_lock(&memcg_oom_lock);
4523 event->eventfd = eventfd;
4524 list_add(&event->list, &memcg->oom_notify);
4526 /* already in OOM ? */
4527 if (atomic_read(&memcg->under_oom))
4528 eventfd_signal(eventfd, 1);
4529 spin_unlock(&memcg_oom_lock);
4534 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4535 struct cftype *cft, struct eventfd_ctx *eventfd)
4537 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4538 struct mem_cgroup_eventfd_list *ev, *tmp;
4539 int type = MEMFILE_TYPE(cft->private);
4541 BUG_ON(type != _OOM_TYPE);
4543 spin_lock(&memcg_oom_lock);
4545 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4546 if (ev->eventfd == eventfd) {
4547 list_del(&ev->list);
4552 spin_unlock(&memcg_oom_lock);
4555 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4556 struct cftype *cft, struct cgroup_map_cb *cb)
4558 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4560 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4562 if (atomic_read(&memcg->under_oom))
4563 cb->fill(cb, "under_oom", 1);
4565 cb->fill(cb, "under_oom", 0);
4569 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4570 struct cftype *cft, u64 val)
4572 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4573 struct mem_cgroup *parent;
4575 /* cannot set to root cgroup and only 0 and 1 are allowed */
4576 if (!cgrp->parent || !((val == 0) || (val == 1)))
4579 parent = mem_cgroup_from_cont(cgrp->parent);
4582 /* oom-kill-disable is a flag for subhierarchy. */
4583 if ((parent->use_hierarchy) ||
4584 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4588 memcg->oom_kill_disable = val;
4590 memcg_oom_recover(memcg);
4595 #ifdef CONFIG_MEMCG_KMEM
4596 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4598 return mem_cgroup_sockets_init(memcg, ss);
4601 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4603 mem_cgroup_sockets_destroy(memcg);
4606 static int memcg_init_kmem(struct mem_cgroup *memcg, struct cgroup_subsys *ss)
4611 static void kmem_cgroup_destroy(struct mem_cgroup *memcg)
4616 static struct cftype mem_cgroup_files[] = {
4618 .name = "usage_in_bytes",
4619 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4620 .read = mem_cgroup_read,
4621 .register_event = mem_cgroup_usage_register_event,
4622 .unregister_event = mem_cgroup_usage_unregister_event,
4625 .name = "max_usage_in_bytes",
4626 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4627 .trigger = mem_cgroup_reset,
4628 .read = mem_cgroup_read,
4631 .name = "limit_in_bytes",
4632 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4633 .write_string = mem_cgroup_write,
4634 .read = mem_cgroup_read,
4637 .name = "soft_limit_in_bytes",
4638 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4639 .write_string = mem_cgroup_write,
4640 .read = mem_cgroup_read,
4644 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4645 .trigger = mem_cgroup_reset,
4646 .read = mem_cgroup_read,
4650 .read_seq_string = memcg_stat_show,
4653 .name = "force_empty",
4654 .trigger = mem_cgroup_force_empty_write,
4657 .name = "use_hierarchy",
4658 .write_u64 = mem_cgroup_hierarchy_write,
4659 .read_u64 = mem_cgroup_hierarchy_read,
4662 .name = "swappiness",
4663 .read_u64 = mem_cgroup_swappiness_read,
4664 .write_u64 = mem_cgroup_swappiness_write,
4667 .name = "move_charge_at_immigrate",
4668 .read_u64 = mem_cgroup_move_charge_read,
4669 .write_u64 = mem_cgroup_move_charge_write,
4672 .name = "oom_control",
4673 .read_map = mem_cgroup_oom_control_read,
4674 .write_u64 = mem_cgroup_oom_control_write,
4675 .register_event = mem_cgroup_oom_register_event,
4676 .unregister_event = mem_cgroup_oom_unregister_event,
4677 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4681 .name = "numa_stat",
4682 .read_seq_string = memcg_numa_stat_show,
4685 #ifdef CONFIG_MEMCG_SWAP
4687 .name = "memsw.usage_in_bytes",
4688 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4689 .read = mem_cgroup_read,
4690 .register_event = mem_cgroup_usage_register_event,
4691 .unregister_event = mem_cgroup_usage_unregister_event,
4694 .name = "memsw.max_usage_in_bytes",
4695 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4696 .trigger = mem_cgroup_reset,
4697 .read = mem_cgroup_read,
4700 .name = "memsw.limit_in_bytes",
4701 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4702 .write_string = mem_cgroup_write,
4703 .read = mem_cgroup_read,
4706 .name = "memsw.failcnt",
4707 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4708 .trigger = mem_cgroup_reset,
4709 .read = mem_cgroup_read,
4712 { }, /* terminate */
4715 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4717 struct mem_cgroup_per_node *pn;
4718 struct mem_cgroup_per_zone *mz;
4719 int zone, tmp = node;
4721 * This routine is called against possible nodes.
4722 * But it's BUG to call kmalloc() against offline node.
4724 * TODO: this routine can waste much memory for nodes which will
4725 * never be onlined. It's better to use memory hotplug callback
4728 if (!node_state(node, N_NORMAL_MEMORY))
4730 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4734 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4735 mz = &pn->zoneinfo[zone];
4736 lruvec_init(&mz->lruvec, &NODE_DATA(node)->node_zones[zone]);
4737 mz->usage_in_excess = 0;
4738 mz->on_tree = false;
4741 memcg->info.nodeinfo[node] = pn;
4745 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4747 kfree(memcg->info.nodeinfo[node]);
4750 static struct mem_cgroup *mem_cgroup_alloc(void)
4752 struct mem_cgroup *memcg;
4753 int size = sizeof(struct mem_cgroup);
4755 /* Can be very big if MAX_NUMNODES is very big */
4756 if (size < PAGE_SIZE)
4757 memcg = kzalloc(size, GFP_KERNEL);
4759 memcg = vzalloc(size);
4764 memcg->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4767 spin_lock_init(&memcg->pcp_counter_lock);
4771 if (size < PAGE_SIZE)
4779 * Helpers for freeing a kmalloc()ed/vzalloc()ed mem_cgroup by RCU,
4780 * but in process context. The work_freeing structure is overlaid
4781 * on the rcu_freeing structure, which itself is overlaid on memsw.
4783 static void free_work(struct work_struct *work)
4785 struct mem_cgroup *memcg;
4786 int size = sizeof(struct mem_cgroup);
4788 memcg = container_of(work, struct mem_cgroup, work_freeing);
4790 * We need to make sure that (at least for now), the jump label
4791 * destruction code runs outside of the cgroup lock. This is because
4792 * get_online_cpus(), which is called from the static_branch update,
4793 * can't be called inside the cgroup_lock. cpusets are the ones
4794 * enforcing this dependency, so if they ever change, we might as well.
4796 * schedule_work() will guarantee this happens. Be careful if you need
4797 * to move this code around, and make sure it is outside
4800 disarm_sock_keys(memcg);
4801 if (size < PAGE_SIZE)
4807 static void free_rcu(struct rcu_head *rcu_head)
4809 struct mem_cgroup *memcg;
4811 memcg = container_of(rcu_head, struct mem_cgroup, rcu_freeing);
4812 INIT_WORK(&memcg->work_freeing, free_work);
4813 schedule_work(&memcg->work_freeing);
4817 * At destroying mem_cgroup, references from swap_cgroup can remain.
4818 * (scanning all at force_empty is too costly...)
4820 * Instead of clearing all references at force_empty, we remember
4821 * the number of reference from swap_cgroup and free mem_cgroup when
4822 * it goes down to 0.
4824 * Removal of cgroup itself succeeds regardless of refs from swap.
4827 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4831 mem_cgroup_remove_from_trees(memcg);
4832 free_css_id(&mem_cgroup_subsys, &memcg->css);
4835 free_mem_cgroup_per_zone_info(memcg, node);
4837 free_percpu(memcg->stat);
4838 call_rcu(&memcg->rcu_freeing, free_rcu);
4841 static void mem_cgroup_get(struct mem_cgroup *memcg)
4843 atomic_inc(&memcg->refcnt);
4846 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4848 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4849 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4850 __mem_cgroup_free(memcg);
4852 mem_cgroup_put(parent);
4856 static void mem_cgroup_put(struct mem_cgroup *memcg)
4858 __mem_cgroup_put(memcg, 1);
4862 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4864 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4866 if (!memcg->res.parent)
4868 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4870 EXPORT_SYMBOL(parent_mem_cgroup);
4872 #ifdef CONFIG_MEMCG_SWAP
4873 static void __init enable_swap_cgroup(void)
4875 if (!mem_cgroup_disabled() && really_do_swap_account)
4876 do_swap_account = 1;
4879 static void __init enable_swap_cgroup(void)
4884 static int mem_cgroup_soft_limit_tree_init(void)
4886 struct mem_cgroup_tree_per_node *rtpn;
4887 struct mem_cgroup_tree_per_zone *rtpz;
4888 int tmp, node, zone;
4890 for_each_node(node) {
4892 if (!node_state(node, N_NORMAL_MEMORY))
4894 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4898 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4900 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4901 rtpz = &rtpn->rb_tree_per_zone[zone];
4902 rtpz->rb_root = RB_ROOT;
4903 spin_lock_init(&rtpz->lock);
4909 for_each_node(node) {
4910 if (!soft_limit_tree.rb_tree_per_node[node])
4912 kfree(soft_limit_tree.rb_tree_per_node[node]);
4913 soft_limit_tree.rb_tree_per_node[node] = NULL;
4919 static struct cgroup_subsys_state * __ref
4920 mem_cgroup_create(struct cgroup *cont)
4922 struct mem_cgroup *memcg, *parent;
4923 long error = -ENOMEM;
4926 memcg = mem_cgroup_alloc();
4928 return ERR_PTR(error);
4931 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4935 if (cont->parent == NULL) {
4937 enable_swap_cgroup();
4939 if (mem_cgroup_soft_limit_tree_init())
4941 root_mem_cgroup = memcg;
4942 for_each_possible_cpu(cpu) {
4943 struct memcg_stock_pcp *stock =
4944 &per_cpu(memcg_stock, cpu);
4945 INIT_WORK(&stock->work, drain_local_stock);
4947 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4949 parent = mem_cgroup_from_cont(cont->parent);
4950 memcg->use_hierarchy = parent->use_hierarchy;
4951 memcg->oom_kill_disable = parent->oom_kill_disable;
4954 if (parent && parent->use_hierarchy) {
4955 res_counter_init(&memcg->res, &parent->res);
4956 res_counter_init(&memcg->memsw, &parent->memsw);
4958 * We increment refcnt of the parent to ensure that we can
4959 * safely access it on res_counter_charge/uncharge.
4960 * This refcnt will be decremented when freeing this
4961 * mem_cgroup(see mem_cgroup_put).
4963 mem_cgroup_get(parent);
4965 res_counter_init(&memcg->res, NULL);
4966 res_counter_init(&memcg->memsw, NULL);
4968 memcg->last_scanned_node = MAX_NUMNODES;
4969 INIT_LIST_HEAD(&memcg->oom_notify);
4972 memcg->swappiness = mem_cgroup_swappiness(parent);
4973 atomic_set(&memcg->refcnt, 1);
4974 memcg->move_charge_at_immigrate = 0;
4975 mutex_init(&memcg->thresholds_lock);
4976 spin_lock_init(&memcg->move_lock);
4978 error = memcg_init_kmem(memcg, &mem_cgroup_subsys);
4981 * We call put now because our (and parent's) refcnts
4982 * are already in place. mem_cgroup_put() will internally
4983 * call __mem_cgroup_free, so return directly
4985 mem_cgroup_put(memcg);
4986 return ERR_PTR(error);
4990 __mem_cgroup_free(memcg);
4991 return ERR_PTR(error);
4994 static int mem_cgroup_pre_destroy(struct cgroup *cont)
4996 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4998 return mem_cgroup_force_empty(memcg, false);
5001 static void mem_cgroup_destroy(struct cgroup *cont)
5003 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5005 kmem_cgroup_destroy(memcg);
5007 mem_cgroup_put(memcg);
5011 /* Handlers for move charge at task migration. */
5012 #define PRECHARGE_COUNT_AT_ONCE 256
5013 static int mem_cgroup_do_precharge(unsigned long count)
5016 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5017 struct mem_cgroup *memcg = mc.to;
5019 if (mem_cgroup_is_root(memcg)) {
5020 mc.precharge += count;
5021 /* we don't need css_get for root */
5024 /* try to charge at once */
5026 struct res_counter *dummy;
5028 * "memcg" cannot be under rmdir() because we've already checked
5029 * by cgroup_lock_live_cgroup() that it is not removed and we
5030 * are still under the same cgroup_mutex. So we can postpone
5033 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5035 if (do_swap_account && res_counter_charge(&memcg->memsw,
5036 PAGE_SIZE * count, &dummy)) {
5037 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5040 mc.precharge += count;
5044 /* fall back to one by one charge */
5046 if (signal_pending(current)) {
5050 if (!batch_count--) {
5051 batch_count = PRECHARGE_COUNT_AT_ONCE;
5054 ret = __mem_cgroup_try_charge(NULL,
5055 GFP_KERNEL, 1, &memcg, false);
5057 /* mem_cgroup_clear_mc() will do uncharge later */
5065 * get_mctgt_type - get target type of moving charge
5066 * @vma: the vma the pte to be checked belongs
5067 * @addr: the address corresponding to the pte to be checked
5068 * @ptent: the pte to be checked
5069 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5072 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5073 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5074 * move charge. if @target is not NULL, the page is stored in target->page
5075 * with extra refcnt got(Callers should handle it).
5076 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5077 * target for charge migration. if @target is not NULL, the entry is stored
5080 * Called with pte lock held.
5087 enum mc_target_type {
5093 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5094 unsigned long addr, pte_t ptent)
5096 struct page *page = vm_normal_page(vma, addr, ptent);
5098 if (!page || !page_mapped(page))
5100 if (PageAnon(page)) {
5101 /* we don't move shared anon */
5104 } else if (!move_file())
5105 /* we ignore mapcount for file pages */
5107 if (!get_page_unless_zero(page))
5114 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5115 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5117 struct page *page = NULL;
5118 swp_entry_t ent = pte_to_swp_entry(ptent);
5120 if (!move_anon() || non_swap_entry(ent))
5123 * Because lookup_swap_cache() updates some statistics counter,
5124 * we call find_get_page() with swapper_space directly.
5126 page = find_get_page(&swapper_space, ent.val);
5127 if (do_swap_account)
5128 entry->val = ent.val;
5133 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5134 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5140 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5141 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5143 struct page *page = NULL;
5144 struct address_space *mapping;
5147 if (!vma->vm_file) /* anonymous vma */
5152 mapping = vma->vm_file->f_mapping;
5153 if (pte_none(ptent))
5154 pgoff = linear_page_index(vma, addr);
5155 else /* pte_file(ptent) is true */
5156 pgoff = pte_to_pgoff(ptent);
5158 /* page is moved even if it's not RSS of this task(page-faulted). */
5159 page = find_get_page(mapping, pgoff);
5162 /* shmem/tmpfs may report page out on swap: account for that too. */
5163 if (radix_tree_exceptional_entry(page)) {
5164 swp_entry_t swap = radix_to_swp_entry(page);
5165 if (do_swap_account)
5167 page = find_get_page(&swapper_space, swap.val);
5173 static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
5174 unsigned long addr, pte_t ptent, union mc_target *target)
5176 struct page *page = NULL;
5177 struct page_cgroup *pc;
5178 enum mc_target_type ret = MC_TARGET_NONE;
5179 swp_entry_t ent = { .val = 0 };
5181 if (pte_present(ptent))
5182 page = mc_handle_present_pte(vma, addr, ptent);
5183 else if (is_swap_pte(ptent))
5184 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5185 else if (pte_none(ptent) || pte_file(ptent))
5186 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5188 if (!page && !ent.val)
5191 pc = lookup_page_cgroup(page);
5193 * Do only loose check w/o page_cgroup lock.
5194 * mem_cgroup_move_account() checks the pc is valid or not under
5197 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5198 ret = MC_TARGET_PAGE;
5200 target->page = page;
5202 if (!ret || !target)
5205 /* There is a swap entry and a page doesn't exist or isn't charged */
5206 if (ent.val && !ret &&
5207 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5208 ret = MC_TARGET_SWAP;
5215 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
5217 * We don't consider swapping or file mapped pages because THP does not
5218 * support them for now.
5219 * Caller should make sure that pmd_trans_huge(pmd) is true.
5221 static enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5222 unsigned long addr, pmd_t pmd, union mc_target *target)
5224 struct page *page = NULL;
5225 struct page_cgroup *pc;
5226 enum mc_target_type ret = MC_TARGET_NONE;
5228 page = pmd_page(pmd);
5229 VM_BUG_ON(!page || !PageHead(page));
5232 pc = lookup_page_cgroup(page);
5233 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5234 ret = MC_TARGET_PAGE;
5237 target->page = page;
5243 static inline enum mc_target_type get_mctgt_type_thp(struct vm_area_struct *vma,
5244 unsigned long addr, pmd_t pmd, union mc_target *target)
5246 return MC_TARGET_NONE;
5250 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5251 unsigned long addr, unsigned long end,
5252 struct mm_walk *walk)
5254 struct vm_area_struct *vma = walk->private;
5258 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5259 if (get_mctgt_type_thp(vma, addr, *pmd, NULL) == MC_TARGET_PAGE)
5260 mc.precharge += HPAGE_PMD_NR;
5261 spin_unlock(&vma->vm_mm->page_table_lock);
5265 if (pmd_trans_unstable(pmd))
5267 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5268 for (; addr != end; pte++, addr += PAGE_SIZE)
5269 if (get_mctgt_type(vma, addr, *pte, NULL))
5270 mc.precharge++; /* increment precharge temporarily */
5271 pte_unmap_unlock(pte - 1, ptl);
5277 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5279 unsigned long precharge;
5280 struct vm_area_struct *vma;
5282 down_read(&mm->mmap_sem);
5283 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5284 struct mm_walk mem_cgroup_count_precharge_walk = {
5285 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5289 if (is_vm_hugetlb_page(vma))
5291 walk_page_range(vma->vm_start, vma->vm_end,
5292 &mem_cgroup_count_precharge_walk);
5294 up_read(&mm->mmap_sem);
5296 precharge = mc.precharge;
5302 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5304 unsigned long precharge = mem_cgroup_count_precharge(mm);
5306 VM_BUG_ON(mc.moving_task);
5307 mc.moving_task = current;
5308 return mem_cgroup_do_precharge(precharge);
5311 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5312 static void __mem_cgroup_clear_mc(void)
5314 struct mem_cgroup *from = mc.from;
5315 struct mem_cgroup *to = mc.to;
5317 /* we must uncharge all the leftover precharges from mc.to */
5319 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5323 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5324 * we must uncharge here.
5326 if (mc.moved_charge) {
5327 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5328 mc.moved_charge = 0;
5330 /* we must fixup refcnts and charges */
5331 if (mc.moved_swap) {
5332 /* uncharge swap account from the old cgroup */
5333 if (!mem_cgroup_is_root(mc.from))
5334 res_counter_uncharge(&mc.from->memsw,
5335 PAGE_SIZE * mc.moved_swap);
5336 __mem_cgroup_put(mc.from, mc.moved_swap);
5338 if (!mem_cgroup_is_root(mc.to)) {
5340 * we charged both to->res and to->memsw, so we should
5343 res_counter_uncharge(&mc.to->res,
5344 PAGE_SIZE * mc.moved_swap);
5346 /* we've already done mem_cgroup_get(mc.to) */
5349 memcg_oom_recover(from);
5350 memcg_oom_recover(to);
5351 wake_up_all(&mc.waitq);
5354 static void mem_cgroup_clear_mc(void)
5356 struct mem_cgroup *from = mc.from;
5359 * we must clear moving_task before waking up waiters at the end of
5362 mc.moving_task = NULL;
5363 __mem_cgroup_clear_mc();
5364 spin_lock(&mc.lock);
5367 spin_unlock(&mc.lock);
5368 mem_cgroup_end_move(from);
5371 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5372 struct cgroup_taskset *tset)
5374 struct task_struct *p = cgroup_taskset_first(tset);
5376 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5378 if (memcg->move_charge_at_immigrate) {
5379 struct mm_struct *mm;
5380 struct mem_cgroup *from = mem_cgroup_from_task(p);
5382 VM_BUG_ON(from == memcg);
5384 mm = get_task_mm(p);
5387 /* We move charges only when we move a owner of the mm */
5388 if (mm->owner == p) {
5391 VM_BUG_ON(mc.precharge);
5392 VM_BUG_ON(mc.moved_charge);
5393 VM_BUG_ON(mc.moved_swap);
5394 mem_cgroup_start_move(from);
5395 spin_lock(&mc.lock);
5398 spin_unlock(&mc.lock);
5399 /* We set mc.moving_task later */
5401 ret = mem_cgroup_precharge_mc(mm);
5403 mem_cgroup_clear_mc();
5410 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5411 struct cgroup_taskset *tset)
5413 mem_cgroup_clear_mc();
5416 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5417 unsigned long addr, unsigned long end,
5418 struct mm_walk *walk)
5421 struct vm_area_struct *vma = walk->private;
5424 enum mc_target_type target_type;
5425 union mc_target target;
5427 struct page_cgroup *pc;
5430 * We don't take compound_lock() here but no race with splitting thp
5432 * - if pmd_trans_huge_lock() returns 1, the relevant thp is not
5433 * under splitting, which means there's no concurrent thp split,
5434 * - if another thread runs into split_huge_page() just after we
5435 * entered this if-block, the thread must wait for page table lock
5436 * to be unlocked in __split_huge_page_splitting(), where the main
5437 * part of thp split is not executed yet.
5439 if (pmd_trans_huge_lock(pmd, vma) == 1) {
5440 if (mc.precharge < HPAGE_PMD_NR) {
5441 spin_unlock(&vma->vm_mm->page_table_lock);
5444 target_type = get_mctgt_type_thp(vma, addr, *pmd, &target);
5445 if (target_type == MC_TARGET_PAGE) {
5447 if (!isolate_lru_page(page)) {
5448 pc = lookup_page_cgroup(page);
5449 if (!mem_cgroup_move_account(page, HPAGE_PMD_NR,
5450 pc, mc.from, mc.to)) {
5451 mc.precharge -= HPAGE_PMD_NR;
5452 mc.moved_charge += HPAGE_PMD_NR;
5454 putback_lru_page(page);
5458 spin_unlock(&vma->vm_mm->page_table_lock);
5462 if (pmd_trans_unstable(pmd))
5465 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5466 for (; addr != end; addr += PAGE_SIZE) {
5467 pte_t ptent = *(pte++);
5473 switch (get_mctgt_type(vma, addr, ptent, &target)) {
5474 case MC_TARGET_PAGE:
5476 if (isolate_lru_page(page))
5478 pc = lookup_page_cgroup(page);
5479 if (!mem_cgroup_move_account(page, 1, pc,
5482 /* we uncharge from mc.from later. */
5485 putback_lru_page(page);
5486 put: /* get_mctgt_type() gets the page */
5489 case MC_TARGET_SWAP:
5491 if (!mem_cgroup_move_swap_account(ent, mc.from, mc.to)) {
5493 /* we fixup refcnts and charges later. */
5501 pte_unmap_unlock(pte - 1, ptl);
5506 * We have consumed all precharges we got in can_attach().
5507 * We try charge one by one, but don't do any additional
5508 * charges to mc.to if we have failed in charge once in attach()
5511 ret = mem_cgroup_do_precharge(1);
5519 static void mem_cgroup_move_charge(struct mm_struct *mm)
5521 struct vm_area_struct *vma;
5523 lru_add_drain_all();
5525 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5527 * Someone who are holding the mmap_sem might be waiting in
5528 * waitq. So we cancel all extra charges, wake up all waiters,
5529 * and retry. Because we cancel precharges, we might not be able
5530 * to move enough charges, but moving charge is a best-effort
5531 * feature anyway, so it wouldn't be a big problem.
5533 __mem_cgroup_clear_mc();
5537 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5539 struct mm_walk mem_cgroup_move_charge_walk = {
5540 .pmd_entry = mem_cgroup_move_charge_pte_range,
5544 if (is_vm_hugetlb_page(vma))
5546 ret = walk_page_range(vma->vm_start, vma->vm_end,
5547 &mem_cgroup_move_charge_walk);
5550 * means we have consumed all precharges and failed in
5551 * doing additional charge. Just abandon here.
5555 up_read(&mm->mmap_sem);
5558 static void mem_cgroup_move_task(struct cgroup *cont,
5559 struct cgroup_taskset *tset)
5561 struct task_struct *p = cgroup_taskset_first(tset);
5562 struct mm_struct *mm = get_task_mm(p);
5566 mem_cgroup_move_charge(mm);
5570 mem_cgroup_clear_mc();
5572 #else /* !CONFIG_MMU */
5573 static int mem_cgroup_can_attach(struct cgroup *cgroup,
5574 struct cgroup_taskset *tset)
5578 static void mem_cgroup_cancel_attach(struct cgroup *cgroup,
5579 struct cgroup_taskset *tset)
5582 static void mem_cgroup_move_task(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 .can_attach = mem_cgroup_can_attach,
5595 .cancel_attach = mem_cgroup_cancel_attach,
5596 .attach = mem_cgroup_move_task,
5597 .base_cftypes = mem_cgroup_files,
5600 .__DEPRECATED_clear_css_refs = true,
5603 #ifdef CONFIG_MEMCG_SWAP
5604 static int __init enable_swap_account(char *s)
5606 /* consider enabled if no parameter or 1 is given */
5607 if (!strcmp(s, "1"))
5608 really_do_swap_account = 1;
5609 else if (!strcmp(s, "0"))
5610 really_do_swap_account = 0;
5613 __setup("swapaccount=", enable_swap_account);