1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/export.h>
37 #include <linux/mutex.h>
38 #include <linux/rbtree.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
54 #include <net/tcp_memcontrol.h>
56 #include <asm/uaccess.h>
58 #include <trace/events/vmscan.h>
60 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
61 #define MEM_CGROUP_RECLAIM_RETRIES 5
62 struct mem_cgroup *root_mem_cgroup __read_mostly;
64 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
65 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
66 int do_swap_account __read_mostly;
68 /* for remember boot option*/
69 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
70 static int really_do_swap_account __initdata = 1;
72 static int really_do_swap_account __initdata = 0;
76 #define do_swap_account (0)
81 * Statistics for memory cgroup.
83 enum mem_cgroup_stat_index {
85 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
87 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
88 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
89 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
90 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
91 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
92 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */
93 MEM_CGROUP_STAT_NSTATS,
96 enum mem_cgroup_events_index {
97 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
98 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
99 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
100 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
101 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
102 MEM_CGROUP_EVENTS_NSTATS,
105 * Per memcg event counter is incremented at every pagein/pageout. With THP,
106 * it will be incremated by the number of pages. This counter is used for
107 * for trigger some periodic events. This is straightforward and better
108 * than using jiffies etc. to handle periodic memcg event.
110 enum mem_cgroup_events_target {
111 MEM_CGROUP_TARGET_THRESH,
112 MEM_CGROUP_TARGET_SOFTLIMIT,
113 MEM_CGROUP_TARGET_NUMAINFO,
116 #define THRESHOLDS_EVENTS_TARGET (128)
117 #define SOFTLIMIT_EVENTS_TARGET (1024)
118 #define NUMAINFO_EVENTS_TARGET (1024)
120 struct mem_cgroup_stat_cpu {
121 long count[MEM_CGROUP_STAT_NSTATS];
122 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
123 unsigned long targets[MEM_CGROUP_NTARGETS];
126 struct mem_cgroup_reclaim_iter {
127 /* css_id of the last scanned hierarchy member */
129 /* scan generation, increased every round-trip */
130 unsigned int generation;
134 * per-zone information in memory controller.
136 struct mem_cgroup_per_zone {
137 struct lruvec lruvec;
138 unsigned long count[NR_LRU_LISTS];
140 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
142 struct zone_reclaim_stat reclaim_stat;
143 struct rb_node tree_node; /* RB tree node */
144 unsigned long long usage_in_excess;/* Set to the value by which */
145 /* the soft limit is exceeded*/
147 struct mem_cgroup *mem; /* Back pointer, we cannot */
148 /* use container_of */
150 /* Macro for accessing counter */
151 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
153 struct mem_cgroup_per_node {
154 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
157 struct mem_cgroup_lru_info {
158 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
162 * Cgroups above their limits are maintained in a RB-Tree, independent of
163 * their hierarchy representation
166 struct mem_cgroup_tree_per_zone {
167 struct rb_root rb_root;
171 struct mem_cgroup_tree_per_node {
172 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
175 struct mem_cgroup_tree {
176 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
179 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
181 struct mem_cgroup_threshold {
182 struct eventfd_ctx *eventfd;
187 struct mem_cgroup_threshold_ary {
188 /* An array index points to threshold just below usage. */
189 int current_threshold;
190 /* Size of entries[] */
192 /* Array of thresholds */
193 struct mem_cgroup_threshold entries[0];
196 struct mem_cgroup_thresholds {
197 /* Primary thresholds array */
198 struct mem_cgroup_threshold_ary *primary;
200 * Spare threshold array.
201 * This is needed to make mem_cgroup_unregister_event() "never fail".
202 * It must be able to store at least primary->size - 1 entries.
204 struct mem_cgroup_threshold_ary *spare;
208 struct mem_cgroup_eventfd_list {
209 struct list_head list;
210 struct eventfd_ctx *eventfd;
213 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
214 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
217 * The memory controller data structure. The memory controller controls both
218 * page cache and RSS per cgroup. We would eventually like to provide
219 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
220 * to help the administrator determine what knobs to tune.
222 * TODO: Add a water mark for the memory controller. Reclaim will begin when
223 * we hit the water mark. May be even add a low water mark, such that
224 * no reclaim occurs from a cgroup at it's low water mark, this is
225 * a feature that will be implemented much later in the future.
228 struct cgroup_subsys_state css;
230 * the counter to account for memory usage
232 struct res_counter res;
234 * the counter to account for mem+swap usage.
236 struct res_counter memsw;
238 * Per cgroup active and inactive list, similar to the
239 * per zone LRU lists.
241 struct mem_cgroup_lru_info info;
242 int last_scanned_node;
244 nodemask_t scan_nodes;
245 atomic_t numainfo_events;
246 atomic_t numainfo_updating;
249 * Should the accounting and control be hierarchical, per subtree?
259 /* OOM-Killer disable */
260 int oom_kill_disable;
262 /* set when res.limit == memsw.limit */
263 bool memsw_is_minimum;
265 /* protect arrays of thresholds */
266 struct mutex thresholds_lock;
268 /* thresholds for memory usage. RCU-protected */
269 struct mem_cgroup_thresholds thresholds;
271 /* thresholds for mem+swap usage. RCU-protected */
272 struct mem_cgroup_thresholds memsw_thresholds;
274 /* For oom notifier event fd */
275 struct list_head oom_notify;
278 * Should we move charges of a task when a task is moved into this
279 * mem_cgroup ? And what type of charges should we move ?
281 unsigned long move_charge_at_immigrate;
285 struct mem_cgroup_stat_cpu *stat;
287 * used when a cpu is offlined or other synchronizations
288 * See mem_cgroup_read_stat().
290 struct mem_cgroup_stat_cpu nocpu_base;
291 spinlock_t pcp_counter_lock;
294 struct tcp_memcontrol tcp_mem;
298 /* Stuffs for move charges at task migration. */
300 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
301 * left-shifted bitmap of these types.
304 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
305 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
309 /* "mc" and its members are protected by cgroup_mutex */
310 static struct move_charge_struct {
311 spinlock_t lock; /* for from, to */
312 struct mem_cgroup *from;
313 struct mem_cgroup *to;
314 unsigned long precharge;
315 unsigned long moved_charge;
316 unsigned long moved_swap;
317 struct task_struct *moving_task; /* a task moving charges */
318 wait_queue_head_t waitq; /* a waitq for other context */
320 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
321 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
324 static bool move_anon(void)
326 return test_bit(MOVE_CHARGE_TYPE_ANON,
327 &mc.to->move_charge_at_immigrate);
330 static bool move_file(void)
332 return test_bit(MOVE_CHARGE_TYPE_FILE,
333 &mc.to->move_charge_at_immigrate);
337 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
338 * limit reclaim to prevent infinite loops, if they ever occur.
340 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
341 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
344 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
345 MEM_CGROUP_CHARGE_TYPE_MAPPED,
346 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
347 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
348 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
349 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
353 /* for encoding cft->private value on file */
356 #define _OOM_TYPE (2)
357 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
358 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
359 #define MEMFILE_ATTR(val) ((val) & 0xffff)
360 /* Used for OOM nofiier */
361 #define OOM_CONTROL (0)
364 * Reclaim flags for mem_cgroup_hierarchical_reclaim
366 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
367 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
368 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
369 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
371 static void mem_cgroup_get(struct mem_cgroup *memcg);
372 static void mem_cgroup_put(struct mem_cgroup *memcg);
374 /* Writing them here to avoid exposing memcg's inner layout */
375 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
377 #include <net/sock.h>
380 static bool mem_cgroup_is_root(struct mem_cgroup *memcg);
381 void sock_update_memcg(struct sock *sk)
383 if (static_branch(&memcg_socket_limit_enabled)) {
384 struct mem_cgroup *memcg;
386 BUG_ON(!sk->sk_prot->proto_cgroup);
388 /* Socket cloning can throw us here with sk_cgrp already
389 * filled. It won't however, necessarily happen from
390 * process context. So the test for root memcg given
391 * the current task's memcg won't help us in this case.
393 * Respecting the original socket's memcg is a better
394 * decision in this case.
397 BUG_ON(mem_cgroup_is_root(sk->sk_cgrp->memcg));
398 mem_cgroup_get(sk->sk_cgrp->memcg);
403 memcg = mem_cgroup_from_task(current);
404 if (!mem_cgroup_is_root(memcg)) {
405 mem_cgroup_get(memcg);
406 sk->sk_cgrp = sk->sk_prot->proto_cgroup(memcg);
411 EXPORT_SYMBOL(sock_update_memcg);
413 void sock_release_memcg(struct sock *sk)
415 if (static_branch(&memcg_socket_limit_enabled) && sk->sk_cgrp) {
416 struct mem_cgroup *memcg;
417 WARN_ON(!sk->sk_cgrp->memcg);
418 memcg = sk->sk_cgrp->memcg;
419 mem_cgroup_put(memcg);
423 struct cg_proto *tcp_proto_cgroup(struct mem_cgroup *memcg)
425 if (!memcg || mem_cgroup_is_root(memcg))
428 return &memcg->tcp_mem.cg_proto;
430 EXPORT_SYMBOL(tcp_proto_cgroup);
431 #endif /* CONFIG_INET */
432 #endif /* CONFIG_CGROUP_MEM_RES_CTLR_KMEM */
434 static void drain_all_stock_async(struct mem_cgroup *memcg);
436 static struct mem_cgroup_per_zone *
437 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
439 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
442 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
447 static struct mem_cgroup_per_zone *
448 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
450 int nid = page_to_nid(page);
451 int zid = page_zonenum(page);
453 return mem_cgroup_zoneinfo(memcg, nid, zid);
456 static struct mem_cgroup_tree_per_zone *
457 soft_limit_tree_node_zone(int nid, int zid)
459 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
462 static struct mem_cgroup_tree_per_zone *
463 soft_limit_tree_from_page(struct page *page)
465 int nid = page_to_nid(page);
466 int zid = page_zonenum(page);
468 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
472 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
473 struct mem_cgroup_per_zone *mz,
474 struct mem_cgroup_tree_per_zone *mctz,
475 unsigned long long new_usage_in_excess)
477 struct rb_node **p = &mctz->rb_root.rb_node;
478 struct rb_node *parent = NULL;
479 struct mem_cgroup_per_zone *mz_node;
484 mz->usage_in_excess = new_usage_in_excess;
485 if (!mz->usage_in_excess)
489 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
491 if (mz->usage_in_excess < mz_node->usage_in_excess)
494 * We can't avoid mem cgroups that are over their soft
495 * limit by the same amount
497 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
500 rb_link_node(&mz->tree_node, parent, p);
501 rb_insert_color(&mz->tree_node, &mctz->rb_root);
506 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
507 struct mem_cgroup_per_zone *mz,
508 struct mem_cgroup_tree_per_zone *mctz)
512 rb_erase(&mz->tree_node, &mctz->rb_root);
517 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
518 struct mem_cgroup_per_zone *mz,
519 struct mem_cgroup_tree_per_zone *mctz)
521 spin_lock(&mctz->lock);
522 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
523 spin_unlock(&mctz->lock);
527 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
529 unsigned long long excess;
530 struct mem_cgroup_per_zone *mz;
531 struct mem_cgroup_tree_per_zone *mctz;
532 int nid = page_to_nid(page);
533 int zid = page_zonenum(page);
534 mctz = soft_limit_tree_from_page(page);
537 * Necessary to update all ancestors when hierarchy is used.
538 * because their event counter is not touched.
540 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
541 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
542 excess = res_counter_soft_limit_excess(&memcg->res);
544 * We have to update the tree if mz is on RB-tree or
545 * mem is over its softlimit.
547 if (excess || mz->on_tree) {
548 spin_lock(&mctz->lock);
549 /* if on-tree, remove it */
551 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
553 * Insert again. mz->usage_in_excess will be updated.
554 * If excess is 0, no tree ops.
556 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
557 spin_unlock(&mctz->lock);
562 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
565 struct mem_cgroup_per_zone *mz;
566 struct mem_cgroup_tree_per_zone *mctz;
568 for_each_node_state(node, N_POSSIBLE) {
569 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
570 mz = mem_cgroup_zoneinfo(memcg, node, zone);
571 mctz = soft_limit_tree_node_zone(node, zone);
572 mem_cgroup_remove_exceeded(memcg, mz, mctz);
577 static struct mem_cgroup_per_zone *
578 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
580 struct rb_node *rightmost = NULL;
581 struct mem_cgroup_per_zone *mz;
585 rightmost = rb_last(&mctz->rb_root);
587 goto done; /* Nothing to reclaim from */
589 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
591 * Remove the node now but someone else can add it back,
592 * we will to add it back at the end of reclaim to its correct
593 * position in the tree.
595 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
596 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
597 !css_tryget(&mz->mem->css))
603 static struct mem_cgroup_per_zone *
604 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
606 struct mem_cgroup_per_zone *mz;
608 spin_lock(&mctz->lock);
609 mz = __mem_cgroup_largest_soft_limit_node(mctz);
610 spin_unlock(&mctz->lock);
615 * Implementation Note: reading percpu statistics for memcg.
617 * Both of vmstat[] and percpu_counter has threshold and do periodic
618 * synchronization to implement "quick" read. There are trade-off between
619 * reading cost and precision of value. Then, we may have a chance to implement
620 * a periodic synchronizion of counter in memcg's counter.
622 * But this _read() function is used for user interface now. The user accounts
623 * memory usage by memory cgroup and he _always_ requires exact value because
624 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
625 * have to visit all online cpus and make sum. So, for now, unnecessary
626 * synchronization is not implemented. (just implemented for cpu hotplug)
628 * If there are kernel internal actions which can make use of some not-exact
629 * value, and reading all cpu value can be performance bottleneck in some
630 * common workload, threashold and synchonization as vmstat[] should be
633 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
634 enum mem_cgroup_stat_index idx)
640 for_each_online_cpu(cpu)
641 val += per_cpu(memcg->stat->count[idx], cpu);
642 #ifdef CONFIG_HOTPLUG_CPU
643 spin_lock(&memcg->pcp_counter_lock);
644 val += memcg->nocpu_base.count[idx];
645 spin_unlock(&memcg->pcp_counter_lock);
651 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
654 int val = (charge) ? 1 : -1;
655 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
658 void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val)
660 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
663 void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val)
665 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
668 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
669 enum mem_cgroup_events_index idx)
671 unsigned long val = 0;
674 for_each_online_cpu(cpu)
675 val += per_cpu(memcg->stat->events[idx], cpu);
676 #ifdef CONFIG_HOTPLUG_CPU
677 spin_lock(&memcg->pcp_counter_lock);
678 val += memcg->nocpu_base.events[idx];
679 spin_unlock(&memcg->pcp_counter_lock);
684 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
685 bool file, int nr_pages)
690 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
693 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
696 /* pagein of a big page is an event. So, ignore page size */
698 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
700 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
701 nr_pages = -nr_pages; /* for event */
704 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
710 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
711 unsigned int lru_mask)
713 struct mem_cgroup_per_zone *mz;
715 unsigned long ret = 0;
717 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
720 if (BIT(l) & lru_mask)
721 ret += MEM_CGROUP_ZSTAT(mz, l);
727 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
728 int nid, unsigned int lru_mask)
733 for (zid = 0; zid < MAX_NR_ZONES; zid++)
734 total += mem_cgroup_zone_nr_lru_pages(memcg,
740 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
741 unsigned int lru_mask)
746 for_each_node_state(nid, N_HIGH_MEMORY)
747 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
751 static bool __memcg_event_check(struct mem_cgroup *memcg, int target)
753 unsigned long val, next;
755 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
756 next = __this_cpu_read(memcg->stat->targets[target]);
757 /* from time_after() in jiffies.h */
758 return ((long)next - (long)val < 0);
761 static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
763 unsigned long val, next;
765 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
768 case MEM_CGROUP_TARGET_THRESH:
769 next = val + THRESHOLDS_EVENTS_TARGET;
771 case MEM_CGROUP_TARGET_SOFTLIMIT:
772 next = val + SOFTLIMIT_EVENTS_TARGET;
774 case MEM_CGROUP_TARGET_NUMAINFO:
775 next = val + NUMAINFO_EVENTS_TARGET;
781 __this_cpu_write(memcg->stat->targets[target], next);
785 * Check events in order.
788 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
791 /* threshold event is triggered in finer grain than soft limit */
792 if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
793 mem_cgroup_threshold(memcg);
794 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
795 if (unlikely(__memcg_event_check(memcg,
796 MEM_CGROUP_TARGET_SOFTLIMIT))) {
797 mem_cgroup_update_tree(memcg, page);
798 __mem_cgroup_target_update(memcg,
799 MEM_CGROUP_TARGET_SOFTLIMIT);
802 if (unlikely(__memcg_event_check(memcg,
803 MEM_CGROUP_TARGET_NUMAINFO))) {
804 atomic_inc(&memcg->numainfo_events);
805 __mem_cgroup_target_update(memcg,
806 MEM_CGROUP_TARGET_NUMAINFO);
813 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
815 return container_of(cgroup_subsys_state(cont,
816 mem_cgroup_subsys_id), struct mem_cgroup,
820 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
823 * mm_update_next_owner() may clear mm->owner to NULL
824 * if it races with swapoff, page migration, etc.
825 * So this can be called with p == NULL.
830 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
831 struct mem_cgroup, css);
834 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
836 struct mem_cgroup *memcg = NULL;
841 * Because we have no locks, mm->owner's may be being moved to other
842 * cgroup. We use css_tryget() here even if this looks
843 * pessimistic (rather than adding locks here).
847 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
848 if (unlikely(!memcg))
850 } while (!css_tryget(&memcg->css));
856 * mem_cgroup_iter - iterate over memory cgroup hierarchy
857 * @root: hierarchy root
858 * @prev: previously returned memcg, NULL on first invocation
859 * @reclaim: cookie for shared reclaim walks, NULL for full walks
861 * Returns references to children of the hierarchy below @root, or
862 * @root itself, or %NULL after a full round-trip.
864 * Caller must pass the return value in @prev on subsequent
865 * invocations for reference counting, or use mem_cgroup_iter_break()
866 * to cancel a hierarchy walk before the round-trip is complete.
868 * Reclaimers can specify a zone and a priority level in @reclaim to
869 * divide up the memcgs in the hierarchy among all concurrent
870 * reclaimers operating on the same zone and priority.
872 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
873 struct mem_cgroup *prev,
874 struct mem_cgroup_reclaim_cookie *reclaim)
876 struct mem_cgroup *memcg = NULL;
879 if (mem_cgroup_disabled())
883 root = root_mem_cgroup;
885 if (prev && !reclaim)
886 id = css_id(&prev->css);
888 if (prev && prev != root)
891 if (!root->use_hierarchy && root != root_mem_cgroup) {
898 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
899 struct cgroup_subsys_state *css;
902 int nid = zone_to_nid(reclaim->zone);
903 int zid = zone_idx(reclaim->zone);
904 struct mem_cgroup_per_zone *mz;
906 mz = mem_cgroup_zoneinfo(root, nid, zid);
907 iter = &mz->reclaim_iter[reclaim->priority];
908 if (prev && reclaim->generation != iter->generation)
914 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
916 if (css == &root->css || css_tryget(css))
917 memcg = container_of(css,
918 struct mem_cgroup, css);
927 else if (!prev && memcg)
928 reclaim->generation = iter->generation;
938 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
939 * @root: hierarchy root
940 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
942 void mem_cgroup_iter_break(struct mem_cgroup *root,
943 struct mem_cgroup *prev)
946 root = root_mem_cgroup;
947 if (prev && prev != root)
952 * Iteration constructs for visiting all cgroups (under a tree). If
953 * loops are exited prematurely (break), mem_cgroup_iter_break() must
954 * be used for reference counting.
956 #define for_each_mem_cgroup_tree(iter, root) \
957 for (iter = mem_cgroup_iter(root, NULL, NULL); \
959 iter = mem_cgroup_iter(root, iter, NULL))
961 #define for_each_mem_cgroup(iter) \
962 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
964 iter = mem_cgroup_iter(NULL, iter, NULL))
966 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
968 return (memcg == root_mem_cgroup);
971 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
973 struct mem_cgroup *memcg;
979 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
980 if (unlikely(!memcg))
985 mem_cgroup_pgmajfault(memcg, 1);
988 mem_cgroup_pgfault(memcg, 1);
996 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
999 * Following LRU functions are allowed to be used without PCG_LOCK.
1000 * Operations are called by routine of global LRU independently from memcg.
1001 * What we have to take care of here is validness of pc->mem_cgroup.
1003 * Changes to pc->mem_cgroup happens when
1006 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1007 * It is added to LRU before charge.
1008 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1009 * When moving account, the page is not on LRU. It's isolated.
1012 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
1014 struct page_cgroup *pc;
1015 struct mem_cgroup_per_zone *mz;
1017 if (mem_cgroup_disabled())
1019 pc = lookup_page_cgroup(page);
1020 /* can happen while we handle swapcache. */
1021 if (!TestClearPageCgroupAcctLRU(pc))
1023 VM_BUG_ON(!pc->mem_cgroup);
1025 * We don't check PCG_USED bit. It's cleared when the "page" is finally
1026 * removed from global LRU.
1028 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1029 /* huge page split is done under lru_lock. so, we have no races. */
1030 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
1031 VM_BUG_ON(list_empty(&pc->lru));
1032 list_del_init(&pc->lru);
1035 void mem_cgroup_del_lru(struct page *page)
1037 mem_cgroup_del_lru_list(page, page_lru(page));
1041 * Writeback is about to end against a page which has been marked for immediate
1042 * reclaim. If it still appears to be reclaimable, move it to the tail of the
1045 void mem_cgroup_rotate_reclaimable_page(struct page *page)
1047 struct mem_cgroup_per_zone *mz;
1048 struct page_cgroup *pc;
1049 enum lru_list lru = page_lru(page);
1051 if (mem_cgroup_disabled())
1054 pc = lookup_page_cgroup(page);
1055 /* unused page is not rotated. */
1056 if (!PageCgroupUsed(pc))
1058 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1060 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1061 list_move_tail(&pc->lru, &mz->lruvec.lists[lru]);
1064 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1066 struct mem_cgroup_per_zone *mz;
1067 struct page_cgroup *pc;
1069 if (mem_cgroup_disabled())
1072 pc = lookup_page_cgroup(page);
1073 /* unused page is not rotated. */
1074 if (!PageCgroupUsed(pc))
1076 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1078 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1079 list_move(&pc->lru, &mz->lruvec.lists[lru]);
1082 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1084 struct page_cgroup *pc;
1085 struct mem_cgroup_per_zone *mz;
1087 if (mem_cgroup_disabled())
1089 pc = lookup_page_cgroup(page);
1090 VM_BUG_ON(PageCgroupAcctLRU(pc));
1093 * SetPageLRU SetPageCgroupUsed
1095 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1097 * Ensure that one of the two sides adds the page to the memcg
1098 * LRU during a race.
1101 if (!PageCgroupUsed(pc))
1103 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1105 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1106 /* huge page split is done under lru_lock. so, we have no races. */
1107 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1108 SetPageCgroupAcctLRU(pc);
1109 list_add(&pc->lru, &mz->lruvec.lists[lru]);
1113 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1114 * while it's linked to lru because the page may be reused after it's fully
1115 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1116 * It's done under lock_page and expected that zone->lru_lock isnever held.
1118 static void mem_cgroup_lru_del_before_commit(struct page *page)
1120 unsigned long flags;
1121 struct zone *zone = page_zone(page);
1122 struct page_cgroup *pc = lookup_page_cgroup(page);
1125 * Doing this check without taking ->lru_lock seems wrong but this
1126 * is safe. Because if page_cgroup's USED bit is unset, the page
1127 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1128 * set, the commit after this will fail, anyway.
1129 * This all charge/uncharge is done under some mutual execustion.
1130 * So, we don't need to taking care of changes in USED bit.
1132 if (likely(!PageLRU(page)))
1135 spin_lock_irqsave(&zone->lru_lock, flags);
1137 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1138 * is guarded by lock_page() because the page is SwapCache.
1140 if (!PageCgroupUsed(pc))
1141 mem_cgroup_del_lru_list(page, page_lru(page));
1142 spin_unlock_irqrestore(&zone->lru_lock, flags);
1145 static void mem_cgroup_lru_add_after_commit(struct page *page)
1147 unsigned long flags;
1148 struct zone *zone = page_zone(page);
1149 struct page_cgroup *pc = lookup_page_cgroup(page);
1152 * SetPageLRU SetPageCgroupUsed
1154 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1156 * Ensure that one of the two sides adds the page to the memcg
1157 * LRU during a race.
1160 /* taking care of that the page is added to LRU while we commit it */
1161 if (likely(!PageLRU(page)))
1163 spin_lock_irqsave(&zone->lru_lock, flags);
1164 /* link when the page is linked to LRU but page_cgroup isn't */
1165 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1166 mem_cgroup_add_lru_list(page, page_lru(page));
1167 spin_unlock_irqrestore(&zone->lru_lock, flags);
1171 void mem_cgroup_move_lists(struct page *page,
1172 enum lru_list from, enum lru_list to)
1174 if (mem_cgroup_disabled())
1176 mem_cgroup_del_lru_list(page, from);
1177 mem_cgroup_add_lru_list(page, to);
1181 * Checks whether given mem is same or in the root_mem_cgroup's
1184 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1185 struct mem_cgroup *memcg)
1187 if (root_memcg != memcg) {
1188 return (root_memcg->use_hierarchy &&
1189 css_is_ancestor(&memcg->css, &root_memcg->css));
1195 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1198 struct mem_cgroup *curr = NULL;
1199 struct task_struct *p;
1201 p = find_lock_task_mm(task);
1204 curr = try_get_mem_cgroup_from_mm(p->mm);
1209 * We should check use_hierarchy of "memcg" not "curr". Because checking
1210 * use_hierarchy of "curr" here make this function true if hierarchy is
1211 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1212 * hierarchy(even if use_hierarchy is disabled in "memcg").
1214 ret = mem_cgroup_same_or_subtree(memcg, curr);
1215 css_put(&curr->css);
1219 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1221 unsigned long inactive_ratio;
1222 int nid = zone_to_nid(zone);
1223 int zid = zone_idx(zone);
1224 unsigned long inactive;
1225 unsigned long active;
1228 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1229 BIT(LRU_INACTIVE_ANON));
1230 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1231 BIT(LRU_ACTIVE_ANON));
1233 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1235 inactive_ratio = int_sqrt(10 * gb);
1239 return inactive * inactive_ratio < active;
1242 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1244 unsigned long active;
1245 unsigned long inactive;
1246 int zid = zone_idx(zone);
1247 int nid = zone_to_nid(zone);
1249 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1250 BIT(LRU_INACTIVE_FILE));
1251 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1252 BIT(LRU_ACTIVE_FILE));
1254 return (active > inactive);
1257 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1260 int nid = zone_to_nid(zone);
1261 int zid = zone_idx(zone);
1262 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1264 return &mz->reclaim_stat;
1267 struct zone_reclaim_stat *
1268 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1270 struct page_cgroup *pc;
1271 struct mem_cgroup_per_zone *mz;
1273 if (mem_cgroup_disabled())
1276 pc = lookup_page_cgroup(page);
1277 if (!PageCgroupUsed(pc))
1279 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1281 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1282 return &mz->reclaim_stat;
1285 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1286 struct list_head *dst,
1287 unsigned long *scanned, int order,
1288 isolate_mode_t mode,
1290 struct mem_cgroup *mem_cont,
1291 int active, int file)
1293 unsigned long nr_taken = 0;
1297 struct list_head *src;
1298 struct page_cgroup *pc, *tmp;
1299 int nid = zone_to_nid(z);
1300 int zid = zone_idx(z);
1301 struct mem_cgroup_per_zone *mz;
1302 int lru = LRU_FILE * file + active;
1306 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1307 src = &mz->lruvec.lists[lru];
1310 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1311 if (scan >= nr_to_scan)
1314 if (unlikely(!PageCgroupUsed(pc)))
1317 page = lookup_cgroup_page(pc);
1319 if (unlikely(!PageLRU(page)))
1323 ret = __isolate_lru_page(page, mode, file);
1326 list_move(&page->lru, dst);
1327 mem_cgroup_del_lru(page);
1328 nr_taken += hpage_nr_pages(page);
1331 /* we don't affect global LRU but rotate in our LRU */
1332 mem_cgroup_rotate_lru_list(page, page_lru(page));
1341 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1347 #define mem_cgroup_from_res_counter(counter, member) \
1348 container_of(counter, struct mem_cgroup, member)
1351 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1352 * @mem: the memory cgroup
1354 * Returns the maximum amount of memory @mem can be charged with, in
1357 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1359 unsigned long long margin;
1361 margin = res_counter_margin(&memcg->res);
1362 if (do_swap_account)
1363 margin = min(margin, res_counter_margin(&memcg->memsw));
1364 return margin >> PAGE_SHIFT;
1367 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1369 struct cgroup *cgrp = memcg->css.cgroup;
1372 if (cgrp->parent == NULL)
1373 return vm_swappiness;
1375 return memcg->swappiness;
1378 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1383 spin_lock(&memcg->pcp_counter_lock);
1384 for_each_online_cpu(cpu)
1385 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1386 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1387 spin_unlock(&memcg->pcp_counter_lock);
1393 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1400 spin_lock(&memcg->pcp_counter_lock);
1401 for_each_online_cpu(cpu)
1402 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1403 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1404 spin_unlock(&memcg->pcp_counter_lock);
1408 * 2 routines for checking "mem" is under move_account() or not.
1410 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1411 * for avoiding race in accounting. If true,
1412 * pc->mem_cgroup may be overwritten.
1414 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1415 * under hierarchy of moving cgroups. This is for
1416 * waiting at hith-memory prressure caused by "move".
1419 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1421 VM_BUG_ON(!rcu_read_lock_held());
1422 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1425 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1427 struct mem_cgroup *from;
1428 struct mem_cgroup *to;
1431 * Unlike task_move routines, we access mc.to, mc.from not under
1432 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1434 spin_lock(&mc.lock);
1440 ret = mem_cgroup_same_or_subtree(memcg, from)
1441 || mem_cgroup_same_or_subtree(memcg, to);
1443 spin_unlock(&mc.lock);
1447 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1449 if (mc.moving_task && current != mc.moving_task) {
1450 if (mem_cgroup_under_move(memcg)) {
1452 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1453 /* moving charge context might have finished. */
1456 finish_wait(&mc.waitq, &wait);
1464 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1465 * @memcg: The memory cgroup that went over limit
1466 * @p: Task that is going to be killed
1468 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1471 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1473 struct cgroup *task_cgrp;
1474 struct cgroup *mem_cgrp;
1476 * Need a buffer in BSS, can't rely on allocations. The code relies
1477 * on the assumption that OOM is serialized for memory controller.
1478 * If this assumption is broken, revisit this code.
1480 static char memcg_name[PATH_MAX];
1489 mem_cgrp = memcg->css.cgroup;
1490 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1492 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1495 * Unfortunately, we are unable to convert to a useful name
1496 * But we'll still print out the usage information
1503 printk(KERN_INFO "Task in %s killed", memcg_name);
1506 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1514 * Continues from above, so we don't need an KERN_ level
1516 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1519 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1520 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1521 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1522 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1523 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1525 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1526 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1527 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1531 * This function returns the number of memcg under hierarchy tree. Returns
1532 * 1(self count) if no children.
1534 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1537 struct mem_cgroup *iter;
1539 for_each_mem_cgroup_tree(iter, memcg)
1545 * Return the memory (and swap, if configured) limit for a memcg.
1547 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1552 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1553 limit += total_swap_pages << PAGE_SHIFT;
1555 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1557 * If memsw is finite and limits the amount of swap space available
1558 * to this memcg, return that limit.
1560 return min(limit, memsw);
1563 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1565 unsigned long flags)
1567 unsigned long total = 0;
1568 bool noswap = false;
1571 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1573 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1576 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1578 drain_all_stock_async(memcg);
1579 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1581 * Allow limit shrinkers, which are triggered directly
1582 * by userspace, to catch signals and stop reclaim
1583 * after minimal progress, regardless of the margin.
1585 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1587 if (mem_cgroup_margin(memcg))
1590 * If nothing was reclaimed after two attempts, there
1591 * may be no reclaimable pages in this hierarchy.
1600 * test_mem_cgroup_node_reclaimable
1601 * @mem: the target memcg
1602 * @nid: the node ID to be checked.
1603 * @noswap : specify true here if the user wants flle only information.
1605 * This function returns whether the specified memcg contains any
1606 * reclaimable pages on a node. Returns true if there are any reclaimable
1607 * pages in the node.
1609 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1610 int nid, bool noswap)
1612 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1614 if (noswap || !total_swap_pages)
1616 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1621 #if MAX_NUMNODES > 1
1624 * Always updating the nodemask is not very good - even if we have an empty
1625 * list or the wrong list here, we can start from some node and traverse all
1626 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1629 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1633 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1634 * pagein/pageout changes since the last update.
1636 if (!atomic_read(&memcg->numainfo_events))
1638 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1641 /* make a nodemask where this memcg uses memory from */
1642 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1644 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1646 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1647 node_clear(nid, memcg->scan_nodes);
1650 atomic_set(&memcg->numainfo_events, 0);
1651 atomic_set(&memcg->numainfo_updating, 0);
1655 * Selecting a node where we start reclaim from. Because what we need is just
1656 * reducing usage counter, start from anywhere is O,K. Considering
1657 * memory reclaim from current node, there are pros. and cons.
1659 * Freeing memory from current node means freeing memory from a node which
1660 * we'll use or we've used. So, it may make LRU bad. And if several threads
1661 * hit limits, it will see a contention on a node. But freeing from remote
1662 * node means more costs for memory reclaim because of memory latency.
1664 * Now, we use round-robin. Better algorithm is welcomed.
1666 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1670 mem_cgroup_may_update_nodemask(memcg);
1671 node = memcg->last_scanned_node;
1673 node = next_node(node, memcg->scan_nodes);
1674 if (node == MAX_NUMNODES)
1675 node = first_node(memcg->scan_nodes);
1677 * We call this when we hit limit, not when pages are added to LRU.
1678 * No LRU may hold pages because all pages are UNEVICTABLE or
1679 * memcg is too small and all pages are not on LRU. In that case,
1680 * we use curret node.
1682 if (unlikely(node == MAX_NUMNODES))
1683 node = numa_node_id();
1685 memcg->last_scanned_node = node;
1690 * Check all nodes whether it contains reclaimable pages or not.
1691 * For quick scan, we make use of scan_nodes. This will allow us to skip
1692 * unused nodes. But scan_nodes is lazily updated and may not cotain
1693 * enough new information. We need to do double check.
1695 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1700 * quick check...making use of scan_node.
1701 * We can skip unused nodes.
1703 if (!nodes_empty(memcg->scan_nodes)) {
1704 for (nid = first_node(memcg->scan_nodes);
1706 nid = next_node(nid, memcg->scan_nodes)) {
1708 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1713 * Check rest of nodes.
1715 for_each_node_state(nid, N_HIGH_MEMORY) {
1716 if (node_isset(nid, memcg->scan_nodes))
1718 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1725 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1730 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1732 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1736 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1739 unsigned long *total_scanned)
1741 struct mem_cgroup *victim = NULL;
1744 unsigned long excess;
1745 unsigned long nr_scanned;
1746 struct mem_cgroup_reclaim_cookie reclaim = {
1751 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1754 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1759 * If we have not been able to reclaim
1760 * anything, it might because there are
1761 * no reclaimable pages under this hierarchy
1766 * We want to do more targeted reclaim.
1767 * excess >> 2 is not to excessive so as to
1768 * reclaim too much, nor too less that we keep
1769 * coming back to reclaim from this cgroup
1771 if (total >= (excess >> 2) ||
1772 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1777 if (!mem_cgroup_reclaimable(victim, false))
1779 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1781 *total_scanned += nr_scanned;
1782 if (!res_counter_soft_limit_excess(&root_memcg->res))
1785 mem_cgroup_iter_break(root_memcg, victim);
1790 * Check OOM-Killer is already running under our hierarchy.
1791 * If someone is running, return false.
1792 * Has to be called with memcg_oom_lock
1794 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1796 struct mem_cgroup *iter, *failed = NULL;
1798 for_each_mem_cgroup_tree(iter, memcg) {
1799 if (iter->oom_lock) {
1801 * this subtree of our hierarchy is already locked
1802 * so we cannot give a lock.
1805 mem_cgroup_iter_break(memcg, iter);
1808 iter->oom_lock = true;
1815 * OK, we failed to lock the whole subtree so we have to clean up
1816 * what we set up to the failing subtree
1818 for_each_mem_cgroup_tree(iter, memcg) {
1819 if (iter == failed) {
1820 mem_cgroup_iter_break(memcg, iter);
1823 iter->oom_lock = false;
1829 * Has to be called with memcg_oom_lock
1831 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1833 struct mem_cgroup *iter;
1835 for_each_mem_cgroup_tree(iter, memcg)
1836 iter->oom_lock = false;
1840 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1842 struct mem_cgroup *iter;
1844 for_each_mem_cgroup_tree(iter, memcg)
1845 atomic_inc(&iter->under_oom);
1848 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1850 struct mem_cgroup *iter;
1853 * When a new child is created while the hierarchy is under oom,
1854 * mem_cgroup_oom_lock() may not be called. We have to use
1855 * atomic_add_unless() here.
1857 for_each_mem_cgroup_tree(iter, memcg)
1858 atomic_add_unless(&iter->under_oom, -1, 0);
1861 static DEFINE_SPINLOCK(memcg_oom_lock);
1862 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1864 struct oom_wait_info {
1865 struct mem_cgroup *mem;
1869 static int memcg_oom_wake_function(wait_queue_t *wait,
1870 unsigned mode, int sync, void *arg)
1872 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1874 struct oom_wait_info *oom_wait_info;
1876 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1877 oom_wait_memcg = oom_wait_info->mem;
1880 * Both of oom_wait_info->mem and wake_mem are stable under us.
1881 * Then we can use css_is_ancestor without taking care of RCU.
1883 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1884 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1886 return autoremove_wake_function(wait, mode, sync, arg);
1889 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1891 /* for filtering, pass "memcg" as argument. */
1892 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1895 static void memcg_oom_recover(struct mem_cgroup *memcg)
1897 if (memcg && atomic_read(&memcg->under_oom))
1898 memcg_wakeup_oom(memcg);
1902 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1904 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1906 struct oom_wait_info owait;
1907 bool locked, need_to_kill;
1910 owait.wait.flags = 0;
1911 owait.wait.func = memcg_oom_wake_function;
1912 owait.wait.private = current;
1913 INIT_LIST_HEAD(&owait.wait.task_list);
1914 need_to_kill = true;
1915 mem_cgroup_mark_under_oom(memcg);
1917 /* At first, try to OOM lock hierarchy under memcg.*/
1918 spin_lock(&memcg_oom_lock);
1919 locked = mem_cgroup_oom_lock(memcg);
1921 * Even if signal_pending(), we can't quit charge() loop without
1922 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1923 * under OOM is always welcomed, use TASK_KILLABLE here.
1925 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1926 if (!locked || memcg->oom_kill_disable)
1927 need_to_kill = false;
1929 mem_cgroup_oom_notify(memcg);
1930 spin_unlock(&memcg_oom_lock);
1933 finish_wait(&memcg_oom_waitq, &owait.wait);
1934 mem_cgroup_out_of_memory(memcg, mask);
1937 finish_wait(&memcg_oom_waitq, &owait.wait);
1939 spin_lock(&memcg_oom_lock);
1941 mem_cgroup_oom_unlock(memcg);
1942 memcg_wakeup_oom(memcg);
1943 spin_unlock(&memcg_oom_lock);
1945 mem_cgroup_unmark_under_oom(memcg);
1947 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1949 /* Give chance to dying process */
1950 schedule_timeout_uninterruptible(1);
1955 * Currently used to update mapped file statistics, but the routine can be
1956 * generalized to update other statistics as well.
1958 * Notes: Race condition
1960 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1961 * it tends to be costly. But considering some conditions, we doesn't need
1962 * to do so _always_.
1964 * Considering "charge", lock_page_cgroup() is not required because all
1965 * file-stat operations happen after a page is attached to radix-tree. There
1966 * are no race with "charge".
1968 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1969 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1970 * if there are race with "uncharge". Statistics itself is properly handled
1973 * Considering "move", this is an only case we see a race. To make the race
1974 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1975 * possibility of race condition. If there is, we take a lock.
1978 void mem_cgroup_update_page_stat(struct page *page,
1979 enum mem_cgroup_page_stat_item idx, int val)
1981 struct mem_cgroup *memcg;
1982 struct page_cgroup *pc = lookup_page_cgroup(page);
1983 bool need_unlock = false;
1984 unsigned long uninitialized_var(flags);
1990 memcg = pc->mem_cgroup;
1991 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1993 /* pc->mem_cgroup is unstable ? */
1994 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
1995 /* take a lock against to access pc->mem_cgroup */
1996 move_lock_page_cgroup(pc, &flags);
1998 memcg = pc->mem_cgroup;
1999 if (!memcg || !PageCgroupUsed(pc))
2004 case MEMCG_NR_FILE_MAPPED:
2006 SetPageCgroupFileMapped(pc);
2007 else if (!page_mapped(page))
2008 ClearPageCgroupFileMapped(pc);
2009 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2015 this_cpu_add(memcg->stat->count[idx], val);
2018 if (unlikely(need_unlock))
2019 move_unlock_page_cgroup(pc, &flags);
2023 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2026 * size of first charge trial. "32" comes from vmscan.c's magic value.
2027 * TODO: maybe necessary to use big numbers in big irons.
2029 #define CHARGE_BATCH 32U
2030 struct memcg_stock_pcp {
2031 struct mem_cgroup *cached; /* this never be root cgroup */
2032 unsigned int nr_pages;
2033 struct work_struct work;
2034 unsigned long flags;
2035 #define FLUSHING_CACHED_CHARGE (0)
2037 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2038 static DEFINE_MUTEX(percpu_charge_mutex);
2041 * Try to consume stocked charge on this cpu. If success, one page is consumed
2042 * from local stock and true is returned. If the stock is 0 or charges from a
2043 * cgroup which is not current target, returns false. This stock will be
2046 static bool consume_stock(struct mem_cgroup *memcg)
2048 struct memcg_stock_pcp *stock;
2051 stock = &get_cpu_var(memcg_stock);
2052 if (memcg == stock->cached && stock->nr_pages)
2054 else /* need to call res_counter_charge */
2056 put_cpu_var(memcg_stock);
2061 * Returns stocks cached in percpu to res_counter and reset cached information.
2063 static void drain_stock(struct memcg_stock_pcp *stock)
2065 struct mem_cgroup *old = stock->cached;
2067 if (stock->nr_pages) {
2068 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2070 res_counter_uncharge(&old->res, bytes);
2071 if (do_swap_account)
2072 res_counter_uncharge(&old->memsw, bytes);
2073 stock->nr_pages = 0;
2075 stock->cached = NULL;
2079 * This must be called under preempt disabled or must be called by
2080 * a thread which is pinned to local cpu.
2082 static void drain_local_stock(struct work_struct *dummy)
2084 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2086 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2090 * Cache charges(val) which is from res_counter, to local per_cpu area.
2091 * This will be consumed by consume_stock() function, later.
2093 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2095 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2097 if (stock->cached != memcg) { /* reset if necessary */
2099 stock->cached = memcg;
2101 stock->nr_pages += nr_pages;
2102 put_cpu_var(memcg_stock);
2106 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2107 * of the hierarchy under it. sync flag says whether we should block
2108 * until the work is done.
2110 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2114 /* Notify other cpus that system-wide "drain" is running */
2117 for_each_online_cpu(cpu) {
2118 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2119 struct mem_cgroup *memcg;
2121 memcg = stock->cached;
2122 if (!memcg || !stock->nr_pages)
2124 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2126 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2128 drain_local_stock(&stock->work);
2130 schedule_work_on(cpu, &stock->work);
2138 for_each_online_cpu(cpu) {
2139 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2140 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2141 flush_work(&stock->work);
2148 * Tries to drain stocked charges in other cpus. This function is asynchronous
2149 * and just put a work per cpu for draining localy on each cpu. Caller can
2150 * expects some charges will be back to res_counter later but cannot wait for
2153 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2156 * If someone calls draining, avoid adding more kworker runs.
2158 if (!mutex_trylock(&percpu_charge_mutex))
2160 drain_all_stock(root_memcg, false);
2161 mutex_unlock(&percpu_charge_mutex);
2164 /* This is a synchronous drain interface. */
2165 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2167 /* called when force_empty is called */
2168 mutex_lock(&percpu_charge_mutex);
2169 drain_all_stock(root_memcg, true);
2170 mutex_unlock(&percpu_charge_mutex);
2174 * This function drains percpu counter value from DEAD cpu and
2175 * move it to local cpu. Note that this function can be preempted.
2177 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2181 spin_lock(&memcg->pcp_counter_lock);
2182 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2183 long x = per_cpu(memcg->stat->count[i], cpu);
2185 per_cpu(memcg->stat->count[i], cpu) = 0;
2186 memcg->nocpu_base.count[i] += x;
2188 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2189 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2191 per_cpu(memcg->stat->events[i], cpu) = 0;
2192 memcg->nocpu_base.events[i] += x;
2194 /* need to clear ON_MOVE value, works as a kind of lock. */
2195 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2196 spin_unlock(&memcg->pcp_counter_lock);
2199 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2201 int idx = MEM_CGROUP_ON_MOVE;
2203 spin_lock(&memcg->pcp_counter_lock);
2204 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2205 spin_unlock(&memcg->pcp_counter_lock);
2208 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2209 unsigned long action,
2212 int cpu = (unsigned long)hcpu;
2213 struct memcg_stock_pcp *stock;
2214 struct mem_cgroup *iter;
2216 if ((action == CPU_ONLINE)) {
2217 for_each_mem_cgroup(iter)
2218 synchronize_mem_cgroup_on_move(iter, cpu);
2222 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2225 for_each_mem_cgroup(iter)
2226 mem_cgroup_drain_pcp_counter(iter, cpu);
2228 stock = &per_cpu(memcg_stock, cpu);
2234 /* See __mem_cgroup_try_charge() for details */
2236 CHARGE_OK, /* success */
2237 CHARGE_RETRY, /* need to retry but retry is not bad */
2238 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2239 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2240 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2243 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2244 unsigned int nr_pages, bool oom_check)
2246 unsigned long csize = nr_pages * PAGE_SIZE;
2247 struct mem_cgroup *mem_over_limit;
2248 struct res_counter *fail_res;
2249 unsigned long flags = 0;
2252 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2255 if (!do_swap_account)
2257 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2261 res_counter_uncharge(&memcg->res, csize);
2262 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2263 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2265 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2267 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2268 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2270 * Never reclaim on behalf of optional batching, retry with a
2271 * single page instead.
2273 if (nr_pages == CHARGE_BATCH)
2274 return CHARGE_RETRY;
2276 if (!(gfp_mask & __GFP_WAIT))
2277 return CHARGE_WOULDBLOCK;
2279 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2280 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2281 return CHARGE_RETRY;
2283 * Even though the limit is exceeded at this point, reclaim
2284 * may have been able to free some pages. Retry the charge
2285 * before killing the task.
2287 * Only for regular pages, though: huge pages are rather
2288 * unlikely to succeed so close to the limit, and we fall back
2289 * to regular pages anyway in case of failure.
2291 if (nr_pages == 1 && ret)
2292 return CHARGE_RETRY;
2295 * At task move, charge accounts can be doubly counted. So, it's
2296 * better to wait until the end of task_move if something is going on.
2298 if (mem_cgroup_wait_acct_move(mem_over_limit))
2299 return CHARGE_RETRY;
2301 /* If we don't need to call oom-killer at el, return immediately */
2303 return CHARGE_NOMEM;
2305 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2306 return CHARGE_OOM_DIE;
2308 return CHARGE_RETRY;
2312 * Unlike exported interface, "oom" parameter is added. if oom==true,
2313 * oom-killer can be invoked.
2315 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2317 unsigned int nr_pages,
2318 struct mem_cgroup **ptr,
2321 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2322 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2323 struct mem_cgroup *memcg = NULL;
2327 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2328 * in system level. So, allow to go ahead dying process in addition to
2331 if (unlikely(test_thread_flag(TIF_MEMDIE)
2332 || fatal_signal_pending(current)))
2336 * We always charge the cgroup the mm_struct belongs to.
2337 * The mm_struct's mem_cgroup changes on task migration if the
2338 * thread group leader migrates. It's possible that mm is not
2339 * set, if so charge the init_mm (happens for pagecache usage).
2344 if (*ptr) { /* css should be a valid one */
2346 VM_BUG_ON(css_is_removed(&memcg->css));
2347 if (mem_cgroup_is_root(memcg))
2349 if (nr_pages == 1 && consume_stock(memcg))
2351 css_get(&memcg->css);
2353 struct task_struct *p;
2356 p = rcu_dereference(mm->owner);
2358 * Because we don't have task_lock(), "p" can exit.
2359 * In that case, "memcg" can point to root or p can be NULL with
2360 * race with swapoff. Then, we have small risk of mis-accouning.
2361 * But such kind of mis-account by race always happens because
2362 * we don't have cgroup_mutex(). It's overkill and we allo that
2364 * (*) swapoff at el will charge against mm-struct not against
2365 * task-struct. So, mm->owner can be NULL.
2367 memcg = mem_cgroup_from_task(p);
2368 if (!memcg || mem_cgroup_is_root(memcg)) {
2372 if (nr_pages == 1 && consume_stock(memcg)) {
2374 * It seems dagerous to access memcg without css_get().
2375 * But considering how consume_stok works, it's not
2376 * necessary. If consume_stock success, some charges
2377 * from this memcg are cached on this cpu. So, we
2378 * don't need to call css_get()/css_tryget() before
2379 * calling consume_stock().
2384 /* after here, we may be blocked. we need to get refcnt */
2385 if (!css_tryget(&memcg->css)) {
2395 /* If killed, bypass charge */
2396 if (fatal_signal_pending(current)) {
2397 css_put(&memcg->css);
2402 if (oom && !nr_oom_retries) {
2404 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2407 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2411 case CHARGE_RETRY: /* not in OOM situation but retry */
2413 css_put(&memcg->css);
2416 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2417 css_put(&memcg->css);
2419 case CHARGE_NOMEM: /* OOM routine works */
2421 css_put(&memcg->css);
2424 /* If oom, we never return -ENOMEM */
2427 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2428 css_put(&memcg->css);
2431 } while (ret != CHARGE_OK);
2433 if (batch > nr_pages)
2434 refill_stock(memcg, batch - nr_pages);
2435 css_put(&memcg->css);
2448 * Somemtimes we have to undo a charge we got by try_charge().
2449 * This function is for that and do uncharge, put css's refcnt.
2450 * gotten by try_charge().
2452 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2453 unsigned int nr_pages)
2455 if (!mem_cgroup_is_root(memcg)) {
2456 unsigned long bytes = nr_pages * PAGE_SIZE;
2458 res_counter_uncharge(&memcg->res, bytes);
2459 if (do_swap_account)
2460 res_counter_uncharge(&memcg->memsw, bytes);
2465 * A helper function to get mem_cgroup from ID. must be called under
2466 * rcu_read_lock(). The caller must check css_is_removed() or some if
2467 * it's concern. (dropping refcnt from swap can be called against removed
2470 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2472 struct cgroup_subsys_state *css;
2474 /* ID 0 is unused ID */
2477 css = css_lookup(&mem_cgroup_subsys, id);
2480 return container_of(css, struct mem_cgroup, css);
2483 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2485 struct mem_cgroup *memcg = NULL;
2486 struct page_cgroup *pc;
2490 VM_BUG_ON(!PageLocked(page));
2492 pc = lookup_page_cgroup(page);
2493 lock_page_cgroup(pc);
2494 if (PageCgroupUsed(pc)) {
2495 memcg = pc->mem_cgroup;
2496 if (memcg && !css_tryget(&memcg->css))
2498 } else if (PageSwapCache(page)) {
2499 ent.val = page_private(page);
2500 id = lookup_swap_cgroup(ent);
2502 memcg = mem_cgroup_lookup(id);
2503 if (memcg && !css_tryget(&memcg->css))
2507 unlock_page_cgroup(pc);
2511 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2513 unsigned int nr_pages,
2514 struct page_cgroup *pc,
2515 enum charge_type ctype)
2517 lock_page_cgroup(pc);
2518 if (unlikely(PageCgroupUsed(pc))) {
2519 unlock_page_cgroup(pc);
2520 __mem_cgroup_cancel_charge(memcg, nr_pages);
2524 * we don't need page_cgroup_lock about tail pages, becase they are not
2525 * accessed by any other context at this point.
2527 pc->mem_cgroup = memcg;
2529 * We access a page_cgroup asynchronously without lock_page_cgroup().
2530 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2531 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2532 * before USED bit, we need memory barrier here.
2533 * See mem_cgroup_add_lru_list(), etc.
2537 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2538 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2539 SetPageCgroupCache(pc);
2540 SetPageCgroupUsed(pc);
2542 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2543 ClearPageCgroupCache(pc);
2544 SetPageCgroupUsed(pc);
2550 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2551 unlock_page_cgroup(pc);
2553 * "charge_statistics" updated event counter. Then, check it.
2554 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2555 * if they exceeds softlimit.
2557 memcg_check_events(memcg, page);
2560 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2562 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2563 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2565 * Because tail pages are not marked as "used", set it. We're under
2566 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2568 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2570 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2571 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2572 unsigned long flags;
2574 if (mem_cgroup_disabled())
2577 * We have no races with charge/uncharge but will have races with
2578 * page state accounting.
2580 move_lock_page_cgroup(head_pc, &flags);
2582 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2583 smp_wmb(); /* see __commit_charge() */
2584 if (PageCgroupAcctLRU(head_pc)) {
2586 struct mem_cgroup_per_zone *mz;
2589 * LRU flags cannot be copied because we need to add tail
2590 *.page to LRU by generic call and our hook will be called.
2591 * We hold lru_lock, then, reduce counter directly.
2593 lru = page_lru(head);
2594 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2595 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2597 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2598 move_unlock_page_cgroup(head_pc, &flags);
2603 * mem_cgroup_move_account - move account of the page
2605 * @nr_pages: number of regular pages (>1 for huge pages)
2606 * @pc: page_cgroup of the page.
2607 * @from: mem_cgroup which the page is moved from.
2608 * @to: mem_cgroup which the page is moved to. @from != @to.
2609 * @uncharge: whether we should call uncharge and css_put against @from.
2611 * The caller must confirm following.
2612 * - page is not on LRU (isolate_page() is useful.)
2613 * - compound_lock is held when nr_pages > 1
2615 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2616 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2617 * true, this function does "uncharge" from old cgroup, but it doesn't if
2618 * @uncharge is false, so a caller should do "uncharge".
2620 static int mem_cgroup_move_account(struct page *page,
2621 unsigned int nr_pages,
2622 struct page_cgroup *pc,
2623 struct mem_cgroup *from,
2624 struct mem_cgroup *to,
2627 unsigned long flags;
2630 VM_BUG_ON(from == to);
2631 VM_BUG_ON(PageLRU(page));
2633 * The page is isolated from LRU. So, collapse function
2634 * will not handle this page. But page splitting can happen.
2635 * Do this check under compound_page_lock(). The caller should
2639 if (nr_pages > 1 && !PageTransHuge(page))
2642 lock_page_cgroup(pc);
2645 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2648 move_lock_page_cgroup(pc, &flags);
2650 if (PageCgroupFileMapped(pc)) {
2651 /* Update mapped_file data for mem_cgroup */
2653 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2654 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2657 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2659 /* This is not "cancel", but cancel_charge does all we need. */
2660 __mem_cgroup_cancel_charge(from, nr_pages);
2662 /* caller should have done css_get */
2663 pc->mem_cgroup = to;
2664 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2666 * We charges against "to" which may not have any tasks. Then, "to"
2667 * can be under rmdir(). But in current implementation, caller of
2668 * this function is just force_empty() and move charge, so it's
2669 * guaranteed that "to" is never removed. So, we don't check rmdir
2672 move_unlock_page_cgroup(pc, &flags);
2675 unlock_page_cgroup(pc);
2679 memcg_check_events(to, page);
2680 memcg_check_events(from, page);
2686 * move charges to its parent.
2689 static int mem_cgroup_move_parent(struct page *page,
2690 struct page_cgroup *pc,
2691 struct mem_cgroup *child,
2694 struct cgroup *cg = child->css.cgroup;
2695 struct cgroup *pcg = cg->parent;
2696 struct mem_cgroup *parent;
2697 unsigned int nr_pages;
2698 unsigned long uninitialized_var(flags);
2706 if (!get_page_unless_zero(page))
2708 if (isolate_lru_page(page))
2711 nr_pages = hpage_nr_pages(page);
2713 parent = mem_cgroup_from_cont(pcg);
2714 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2719 flags = compound_lock_irqsave(page);
2721 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2723 __mem_cgroup_cancel_charge(parent, nr_pages);
2726 compound_unlock_irqrestore(page, flags);
2728 putback_lru_page(page);
2736 * Charge the memory controller for page usage.
2738 * 0 if the charge was successful
2739 * < 0 if the cgroup is over its limit
2741 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2742 gfp_t gfp_mask, enum charge_type ctype)
2744 struct mem_cgroup *memcg = NULL;
2745 unsigned int nr_pages = 1;
2746 struct page_cgroup *pc;
2750 if (PageTransHuge(page)) {
2751 nr_pages <<= compound_order(page);
2752 VM_BUG_ON(!PageTransHuge(page));
2754 * Never OOM-kill a process for a huge page. The
2755 * fault handler will fall back to regular pages.
2760 pc = lookup_page_cgroup(page);
2761 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2763 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2767 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2771 int mem_cgroup_newpage_charge(struct page *page,
2772 struct mm_struct *mm, gfp_t gfp_mask)
2774 if (mem_cgroup_disabled())
2777 * If already mapped, we don't have to account.
2778 * If page cache, page->mapping has address_space.
2779 * But page->mapping may have out-of-use anon_vma pointer,
2780 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2783 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2787 return mem_cgroup_charge_common(page, mm, gfp_mask,
2788 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2792 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2793 enum charge_type ctype);
2796 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2797 enum charge_type ctype)
2799 struct page_cgroup *pc = lookup_page_cgroup(page);
2801 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2802 * is already on LRU. It means the page may on some other page_cgroup's
2803 * LRU. Take care of it.
2805 mem_cgroup_lru_del_before_commit(page);
2806 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2807 mem_cgroup_lru_add_after_commit(page);
2811 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2814 struct mem_cgroup *memcg = NULL;
2817 if (mem_cgroup_disabled())
2819 if (PageCompound(page))
2825 if (page_is_file_cache(page)) {
2826 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2831 * FUSE reuses pages without going through the final
2832 * put that would remove them from the LRU list, make
2833 * sure that they get relinked properly.
2835 __mem_cgroup_commit_charge_lrucare(page, memcg,
2836 MEM_CGROUP_CHARGE_TYPE_CACHE);
2840 if (PageSwapCache(page)) {
2841 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2843 __mem_cgroup_commit_charge_swapin(page, memcg,
2844 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2846 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2847 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2853 * While swap-in, try_charge -> commit or cancel, the page is locked.
2854 * And when try_charge() successfully returns, one refcnt to memcg without
2855 * struct page_cgroup is acquired. This refcnt will be consumed by
2856 * "commit()" or removed by "cancel()"
2858 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2860 gfp_t mask, struct mem_cgroup **ptr)
2862 struct mem_cgroup *memcg;
2867 if (mem_cgroup_disabled())
2870 if (!do_swap_account)
2873 * A racing thread's fault, or swapoff, may have already updated
2874 * the pte, and even removed page from swap cache: in those cases
2875 * do_swap_page()'s pte_same() test will fail; but there's also a
2876 * KSM case which does need to charge the page.
2878 if (!PageSwapCache(page))
2880 memcg = try_get_mem_cgroup_from_page(page);
2884 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2885 css_put(&memcg->css);
2890 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2894 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2895 enum charge_type ctype)
2897 if (mem_cgroup_disabled())
2901 cgroup_exclude_rmdir(&ptr->css);
2903 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2905 * Now swap is on-memory. This means this page may be
2906 * counted both as mem and swap....double count.
2907 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2908 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2909 * may call delete_from_swap_cache() before reach here.
2911 if (do_swap_account && PageSwapCache(page)) {
2912 swp_entry_t ent = {.val = page_private(page)};
2914 struct mem_cgroup *memcg;
2916 id = swap_cgroup_record(ent, 0);
2918 memcg = mem_cgroup_lookup(id);
2921 * This recorded memcg can be obsolete one. So, avoid
2922 * calling css_tryget
2924 if (!mem_cgroup_is_root(memcg))
2925 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2926 mem_cgroup_swap_statistics(memcg, false);
2927 mem_cgroup_put(memcg);
2932 * At swapin, we may charge account against cgroup which has no tasks.
2933 * So, rmdir()->pre_destroy() can be called while we do this charge.
2934 * In that case, we need to call pre_destroy() again. check it here.
2936 cgroup_release_and_wakeup_rmdir(&ptr->css);
2939 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2941 __mem_cgroup_commit_charge_swapin(page, ptr,
2942 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2945 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2947 if (mem_cgroup_disabled())
2951 __mem_cgroup_cancel_charge(memcg, 1);
2954 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2955 unsigned int nr_pages,
2956 const enum charge_type ctype)
2958 struct memcg_batch_info *batch = NULL;
2959 bool uncharge_memsw = true;
2961 /* If swapout, usage of swap doesn't decrease */
2962 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2963 uncharge_memsw = false;
2965 batch = ¤t->memcg_batch;
2967 * In usual, we do css_get() when we remember memcg pointer.
2968 * But in this case, we keep res->usage until end of a series of
2969 * uncharges. Then, it's ok to ignore memcg's refcnt.
2972 batch->memcg = memcg;
2974 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2975 * In those cases, all pages freed continuously can be expected to be in
2976 * the same cgroup and we have chance to coalesce uncharges.
2977 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2978 * because we want to do uncharge as soon as possible.
2981 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2982 goto direct_uncharge;
2985 goto direct_uncharge;
2988 * In typical case, batch->memcg == mem. This means we can
2989 * merge a series of uncharges to an uncharge of res_counter.
2990 * If not, we uncharge res_counter ony by one.
2992 if (batch->memcg != memcg)
2993 goto direct_uncharge;
2994 /* remember freed charge and uncharge it later */
2997 batch->memsw_nr_pages++;
3000 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3002 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3003 if (unlikely(batch->memcg != memcg))
3004 memcg_oom_recover(memcg);
3009 * uncharge if !page_mapped(page)
3011 static struct mem_cgroup *
3012 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3014 struct mem_cgroup *memcg = NULL;
3015 unsigned int nr_pages = 1;
3016 struct page_cgroup *pc;
3018 if (mem_cgroup_disabled())
3021 if (PageSwapCache(page))
3024 if (PageTransHuge(page)) {
3025 nr_pages <<= compound_order(page);
3026 VM_BUG_ON(!PageTransHuge(page));
3029 * Check if our page_cgroup is valid
3031 pc = lookup_page_cgroup(page);
3032 if (unlikely(!pc || !PageCgroupUsed(pc)))
3035 lock_page_cgroup(pc);
3037 memcg = pc->mem_cgroup;
3039 if (!PageCgroupUsed(pc))
3043 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3044 case MEM_CGROUP_CHARGE_TYPE_DROP:
3045 /* See mem_cgroup_prepare_migration() */
3046 if (page_mapped(page) || PageCgroupMigration(pc))
3049 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3050 if (!PageAnon(page)) { /* Shared memory */
3051 if (page->mapping && !page_is_file_cache(page))
3053 } else if (page_mapped(page)) /* Anon */
3060 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3062 ClearPageCgroupUsed(pc);
3064 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3065 * freed from LRU. This is safe because uncharged page is expected not
3066 * to be reused (freed soon). Exception is SwapCache, it's handled by
3067 * special functions.
3070 unlock_page_cgroup(pc);
3072 * even after unlock, we have memcg->res.usage here and this memcg
3073 * will never be freed.
3075 memcg_check_events(memcg, page);
3076 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3077 mem_cgroup_swap_statistics(memcg, true);
3078 mem_cgroup_get(memcg);
3080 if (!mem_cgroup_is_root(memcg))
3081 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3086 unlock_page_cgroup(pc);
3090 void mem_cgroup_uncharge_page(struct page *page)
3093 if (page_mapped(page))
3095 if (page->mapping && !PageAnon(page))
3097 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3100 void mem_cgroup_uncharge_cache_page(struct page *page)
3102 VM_BUG_ON(page_mapped(page));
3103 VM_BUG_ON(page->mapping);
3104 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3108 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3109 * In that cases, pages are freed continuously and we can expect pages
3110 * are in the same memcg. All these calls itself limits the number of
3111 * pages freed at once, then uncharge_start/end() is called properly.
3112 * This may be called prural(2) times in a context,
3115 void mem_cgroup_uncharge_start(void)
3117 current->memcg_batch.do_batch++;
3118 /* We can do nest. */
3119 if (current->memcg_batch.do_batch == 1) {
3120 current->memcg_batch.memcg = NULL;
3121 current->memcg_batch.nr_pages = 0;
3122 current->memcg_batch.memsw_nr_pages = 0;
3126 void mem_cgroup_uncharge_end(void)
3128 struct memcg_batch_info *batch = ¤t->memcg_batch;
3130 if (!batch->do_batch)
3134 if (batch->do_batch) /* If stacked, do nothing. */
3140 * This "batch->memcg" is valid without any css_get/put etc...
3141 * bacause we hide charges behind us.
3143 if (batch->nr_pages)
3144 res_counter_uncharge(&batch->memcg->res,
3145 batch->nr_pages * PAGE_SIZE);
3146 if (batch->memsw_nr_pages)
3147 res_counter_uncharge(&batch->memcg->memsw,
3148 batch->memsw_nr_pages * PAGE_SIZE);
3149 memcg_oom_recover(batch->memcg);
3150 /* forget this pointer (for sanity check) */
3151 batch->memcg = NULL;
3156 * called after __delete_from_swap_cache() and drop "page" account.
3157 * memcg information is recorded to swap_cgroup of "ent"
3160 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3162 struct mem_cgroup *memcg;
3163 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3165 if (!swapout) /* this was a swap cache but the swap is unused ! */
3166 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3168 memcg = __mem_cgroup_uncharge_common(page, ctype);
3171 * record memcg information, if swapout && memcg != NULL,
3172 * mem_cgroup_get() was called in uncharge().
3174 if (do_swap_account && swapout && memcg)
3175 swap_cgroup_record(ent, css_id(&memcg->css));
3179 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3181 * called from swap_entry_free(). remove record in swap_cgroup and
3182 * uncharge "memsw" account.
3184 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3186 struct mem_cgroup *memcg;
3189 if (!do_swap_account)
3192 id = swap_cgroup_record(ent, 0);
3194 memcg = mem_cgroup_lookup(id);
3197 * We uncharge this because swap is freed.
3198 * This memcg can be obsolete one. We avoid calling css_tryget
3200 if (!mem_cgroup_is_root(memcg))
3201 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3202 mem_cgroup_swap_statistics(memcg, false);
3203 mem_cgroup_put(memcg);
3209 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3210 * @entry: swap entry to be moved
3211 * @from: mem_cgroup which the entry is moved from
3212 * @to: mem_cgroup which the entry is moved to
3213 * @need_fixup: whether we should fixup res_counters and refcounts.
3215 * It succeeds only when the swap_cgroup's record for this entry is the same
3216 * as the mem_cgroup's id of @from.
3218 * Returns 0 on success, -EINVAL on failure.
3220 * The caller must have charged to @to, IOW, called res_counter_charge() about
3221 * both res and memsw, and called css_get().
3223 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3224 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3226 unsigned short old_id, new_id;
3228 old_id = css_id(&from->css);
3229 new_id = css_id(&to->css);
3231 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3232 mem_cgroup_swap_statistics(from, false);
3233 mem_cgroup_swap_statistics(to, true);
3235 * This function is only called from task migration context now.
3236 * It postpones res_counter and refcount handling till the end
3237 * of task migration(mem_cgroup_clear_mc()) for performance
3238 * improvement. But we cannot postpone mem_cgroup_get(to)
3239 * because if the process that has been moved to @to does
3240 * swap-in, the refcount of @to might be decreased to 0.
3244 if (!mem_cgroup_is_root(from))
3245 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3246 mem_cgroup_put(from);
3248 * we charged both to->res and to->memsw, so we should
3251 if (!mem_cgroup_is_root(to))
3252 res_counter_uncharge(&to->res, PAGE_SIZE);
3259 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3260 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3267 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3270 int mem_cgroup_prepare_migration(struct page *page,
3271 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3273 struct mem_cgroup *memcg = NULL;
3274 struct page_cgroup *pc;
3275 enum charge_type ctype;
3280 VM_BUG_ON(PageTransHuge(page));
3281 if (mem_cgroup_disabled())
3284 pc = lookup_page_cgroup(page);
3285 lock_page_cgroup(pc);
3286 if (PageCgroupUsed(pc)) {
3287 memcg = pc->mem_cgroup;
3288 css_get(&memcg->css);
3290 * At migrating an anonymous page, its mapcount goes down
3291 * to 0 and uncharge() will be called. But, even if it's fully
3292 * unmapped, migration may fail and this page has to be
3293 * charged again. We set MIGRATION flag here and delay uncharge
3294 * until end_migration() is called
3296 * Corner Case Thinking
3298 * When the old page was mapped as Anon and it's unmap-and-freed
3299 * while migration was ongoing.
3300 * If unmap finds the old page, uncharge() of it will be delayed
3301 * until end_migration(). If unmap finds a new page, it's
3302 * uncharged when it make mapcount to be 1->0. If unmap code
3303 * finds swap_migration_entry, the new page will not be mapped
3304 * and end_migration() will find it(mapcount==0).
3307 * When the old page was mapped but migraion fails, the kernel
3308 * remaps it. A charge for it is kept by MIGRATION flag even
3309 * if mapcount goes down to 0. We can do remap successfully
3310 * without charging it again.
3313 * The "old" page is under lock_page() until the end of
3314 * migration, so, the old page itself will not be swapped-out.
3315 * If the new page is swapped out before end_migraton, our
3316 * hook to usual swap-out path will catch the event.
3319 SetPageCgroupMigration(pc);
3321 unlock_page_cgroup(pc);
3323 * If the page is not charged at this point,
3330 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3331 css_put(&memcg->css);/* drop extra refcnt */
3332 if (ret || *ptr == NULL) {
3333 if (PageAnon(page)) {
3334 lock_page_cgroup(pc);
3335 ClearPageCgroupMigration(pc);
3336 unlock_page_cgroup(pc);
3338 * The old page may be fully unmapped while we kept it.
3340 mem_cgroup_uncharge_page(page);
3345 * We charge new page before it's used/mapped. So, even if unlock_page()
3346 * is called before end_migration, we can catch all events on this new
3347 * page. In the case new page is migrated but not remapped, new page's
3348 * mapcount will be finally 0 and we call uncharge in end_migration().
3350 pc = lookup_page_cgroup(newpage);
3352 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3353 else if (page_is_file_cache(page))
3354 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3356 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3357 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3361 /* remove redundant charge if migration failed*/
3362 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3363 struct page *oldpage, struct page *newpage, bool migration_ok)
3365 struct page *used, *unused;
3366 struct page_cgroup *pc;
3370 /* blocks rmdir() */
3371 cgroup_exclude_rmdir(&memcg->css);
3372 if (!migration_ok) {
3380 * We disallowed uncharge of pages under migration because mapcount
3381 * of the page goes down to zero, temporarly.
3382 * Clear the flag and check the page should be charged.
3384 pc = lookup_page_cgroup(oldpage);
3385 lock_page_cgroup(pc);
3386 ClearPageCgroupMigration(pc);
3387 unlock_page_cgroup(pc);
3389 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3392 * If a page is a file cache, radix-tree replacement is very atomic
3393 * and we can skip this check. When it was an Anon page, its mapcount
3394 * goes down to 0. But because we added MIGRATION flage, it's not
3395 * uncharged yet. There are several case but page->mapcount check
3396 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3397 * check. (see prepare_charge() also)
3400 mem_cgroup_uncharge_page(used);
3402 * At migration, we may charge account against cgroup which has no
3404 * So, rmdir()->pre_destroy() can be called while we do this charge.
3405 * In that case, we need to call pre_destroy() again. check it here.
3407 cgroup_release_and_wakeup_rmdir(&memcg->css);
3411 * At replace page cache, newpage is not under any memcg but it's on
3412 * LRU. So, this function doesn't touch res_counter but handles LRU
3413 * in correct way. Both pages are locked so we cannot race with uncharge.
3415 void mem_cgroup_replace_page_cache(struct page *oldpage,
3416 struct page *newpage)
3418 struct mem_cgroup *memcg;
3419 struct page_cgroup *pc;
3421 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3422 unsigned long flags;
3424 if (mem_cgroup_disabled())
3427 pc = lookup_page_cgroup(oldpage);
3428 /* fix accounting on old pages */
3429 lock_page_cgroup(pc);
3430 memcg = pc->mem_cgroup;
3431 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1);
3432 ClearPageCgroupUsed(pc);
3433 unlock_page_cgroup(pc);
3435 if (PageSwapBacked(oldpage))
3436 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3438 zone = page_zone(newpage);
3439 pc = lookup_page_cgroup(newpage);
3441 * Even if newpage->mapping was NULL before starting replacement,
3442 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3443 * LRU while we overwrite pc->mem_cgroup.
3445 spin_lock_irqsave(&zone->lru_lock, flags);
3446 if (PageLRU(newpage))
3447 del_page_from_lru_list(zone, newpage, page_lru(newpage));
3448 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type);
3449 if (PageLRU(newpage))
3450 add_page_to_lru_list(zone, newpage, page_lru(newpage));
3451 spin_unlock_irqrestore(&zone->lru_lock, flags);
3454 #ifdef CONFIG_DEBUG_VM
3455 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3457 struct page_cgroup *pc;
3459 pc = lookup_page_cgroup(page);
3460 if (likely(pc) && PageCgroupUsed(pc))
3465 bool mem_cgroup_bad_page_check(struct page *page)
3467 if (mem_cgroup_disabled())
3470 return lookup_page_cgroup_used(page) != NULL;
3473 void mem_cgroup_print_bad_page(struct page *page)
3475 struct page_cgroup *pc;
3477 pc = lookup_page_cgroup_used(page);
3482 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3483 pc, pc->flags, pc->mem_cgroup);
3485 path = kmalloc(PATH_MAX, GFP_KERNEL);
3488 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3493 printk(KERN_CONT "(%s)\n",
3494 (ret < 0) ? "cannot get the path" : path);
3500 static DEFINE_MUTEX(set_limit_mutex);
3502 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3503 unsigned long long val)
3506 u64 memswlimit, memlimit;
3508 int children = mem_cgroup_count_children(memcg);
3509 u64 curusage, oldusage;
3513 * For keeping hierarchical_reclaim simple, how long we should retry
3514 * is depends on callers. We set our retry-count to be function
3515 * of # of children which we should visit in this loop.
3517 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3519 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3522 while (retry_count) {
3523 if (signal_pending(current)) {
3528 * Rather than hide all in some function, I do this in
3529 * open coded manner. You see what this really does.
3530 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3532 mutex_lock(&set_limit_mutex);
3533 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3534 if (memswlimit < val) {
3536 mutex_unlock(&set_limit_mutex);
3540 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3544 ret = res_counter_set_limit(&memcg->res, val);
3546 if (memswlimit == val)
3547 memcg->memsw_is_minimum = true;
3549 memcg->memsw_is_minimum = false;
3551 mutex_unlock(&set_limit_mutex);
3556 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3557 MEM_CGROUP_RECLAIM_SHRINK);
3558 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3559 /* Usage is reduced ? */
3560 if (curusage >= oldusage)
3563 oldusage = curusage;
3565 if (!ret && enlarge)
3566 memcg_oom_recover(memcg);
3571 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3572 unsigned long long val)
3575 u64 memlimit, memswlimit, oldusage, curusage;
3576 int children = mem_cgroup_count_children(memcg);
3580 /* see mem_cgroup_resize_res_limit */
3581 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3582 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3583 while (retry_count) {
3584 if (signal_pending(current)) {
3589 * Rather than hide all in some function, I do this in
3590 * open coded manner. You see what this really does.
3591 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3593 mutex_lock(&set_limit_mutex);
3594 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3595 if (memlimit > val) {
3597 mutex_unlock(&set_limit_mutex);
3600 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3601 if (memswlimit < val)
3603 ret = res_counter_set_limit(&memcg->memsw, val);
3605 if (memlimit == val)
3606 memcg->memsw_is_minimum = true;
3608 memcg->memsw_is_minimum = false;
3610 mutex_unlock(&set_limit_mutex);
3615 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3616 MEM_CGROUP_RECLAIM_NOSWAP |
3617 MEM_CGROUP_RECLAIM_SHRINK);
3618 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3619 /* Usage is reduced ? */
3620 if (curusage >= oldusage)
3623 oldusage = curusage;
3625 if (!ret && enlarge)
3626 memcg_oom_recover(memcg);
3630 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3632 unsigned long *total_scanned)
3634 unsigned long nr_reclaimed = 0;
3635 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3636 unsigned long reclaimed;
3638 struct mem_cgroup_tree_per_zone *mctz;
3639 unsigned long long excess;
3640 unsigned long nr_scanned;
3645 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3647 * This loop can run a while, specially if mem_cgroup's continuously
3648 * keep exceeding their soft limit and putting the system under
3655 mz = mem_cgroup_largest_soft_limit_node(mctz);
3660 reclaimed = mem_cgroup_soft_reclaim(mz->mem, zone,
3661 gfp_mask, &nr_scanned);
3662 nr_reclaimed += reclaimed;
3663 *total_scanned += nr_scanned;
3664 spin_lock(&mctz->lock);
3667 * If we failed to reclaim anything from this memory cgroup
3668 * it is time to move on to the next cgroup
3674 * Loop until we find yet another one.
3676 * By the time we get the soft_limit lock
3677 * again, someone might have aded the
3678 * group back on the RB tree. Iterate to
3679 * make sure we get a different mem.
3680 * mem_cgroup_largest_soft_limit_node returns
3681 * NULL if no other cgroup is present on
3685 __mem_cgroup_largest_soft_limit_node(mctz);
3687 css_put(&next_mz->mem->css);
3688 else /* next_mz == NULL or other memcg */
3692 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3693 excess = res_counter_soft_limit_excess(&mz->mem->res);
3695 * One school of thought says that we should not add
3696 * back the node to the tree if reclaim returns 0.
3697 * But our reclaim could return 0, simply because due
3698 * to priority we are exposing a smaller subset of
3699 * memory to reclaim from. Consider this as a longer
3702 /* If excess == 0, no tree ops */
3703 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3704 spin_unlock(&mctz->lock);
3705 css_put(&mz->mem->css);
3708 * Could not reclaim anything and there are no more
3709 * mem cgroups to try or we seem to be looping without
3710 * reclaiming anything.
3712 if (!nr_reclaimed &&
3714 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3716 } while (!nr_reclaimed);
3718 css_put(&next_mz->mem->css);
3719 return nr_reclaimed;
3723 * This routine traverse page_cgroup in given list and drop them all.
3724 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3726 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3727 int node, int zid, enum lru_list lru)
3730 struct mem_cgroup_per_zone *mz;
3731 struct page_cgroup *pc, *busy;
3732 unsigned long flags, loop;
3733 struct list_head *list;
3736 zone = &NODE_DATA(node)->node_zones[zid];
3737 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3738 list = &mz->lruvec.lists[lru];
3740 loop = MEM_CGROUP_ZSTAT(mz, lru);
3741 /* give some margin against EBUSY etc...*/
3748 spin_lock_irqsave(&zone->lru_lock, flags);
3749 if (list_empty(list)) {
3750 spin_unlock_irqrestore(&zone->lru_lock, flags);
3753 pc = list_entry(list->prev, struct page_cgroup, lru);
3755 list_move(&pc->lru, list);
3757 spin_unlock_irqrestore(&zone->lru_lock, flags);
3760 spin_unlock_irqrestore(&zone->lru_lock, flags);
3762 page = lookup_cgroup_page(pc);
3764 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3768 if (ret == -EBUSY || ret == -EINVAL) {
3769 /* found lock contention or "pc" is obsolete. */
3776 if (!ret && !list_empty(list))
3782 * make mem_cgroup's charge to be 0 if there is no task.
3783 * This enables deleting this mem_cgroup.
3785 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3788 int node, zid, shrink;
3789 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3790 struct cgroup *cgrp = memcg->css.cgroup;
3792 css_get(&memcg->css);
3795 /* should free all ? */
3801 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3804 if (signal_pending(current))
3806 /* This is for making all *used* pages to be on LRU. */
3807 lru_add_drain_all();
3808 drain_all_stock_sync(memcg);
3810 mem_cgroup_start_move(memcg);
3811 for_each_node_state(node, N_HIGH_MEMORY) {
3812 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3815 ret = mem_cgroup_force_empty_list(memcg,
3824 mem_cgroup_end_move(memcg);
3825 memcg_oom_recover(memcg);
3826 /* it seems parent cgroup doesn't have enough mem */
3830 /* "ret" should also be checked to ensure all lists are empty. */
3831 } while (memcg->res.usage > 0 || ret);
3833 css_put(&memcg->css);
3837 /* returns EBUSY if there is a task or if we come here twice. */
3838 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3842 /* we call try-to-free pages for make this cgroup empty */
3843 lru_add_drain_all();
3844 /* try to free all pages in this cgroup */
3846 while (nr_retries && memcg->res.usage > 0) {
3849 if (signal_pending(current)) {
3853 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3857 /* maybe some writeback is necessary */
3858 congestion_wait(BLK_RW_ASYNC, HZ/10);
3863 /* try move_account...there may be some *locked* pages. */
3867 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3869 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3873 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3875 return mem_cgroup_from_cont(cont)->use_hierarchy;
3878 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3882 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3883 struct cgroup *parent = cont->parent;
3884 struct mem_cgroup *parent_memcg = NULL;
3887 parent_memcg = mem_cgroup_from_cont(parent);
3891 * If parent's use_hierarchy is set, we can't make any modifications
3892 * in the child subtrees. If it is unset, then the change can
3893 * occur, provided the current cgroup has no children.
3895 * For the root cgroup, parent_mem is NULL, we allow value to be
3896 * set if there are no children.
3898 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3899 (val == 1 || val == 0)) {
3900 if (list_empty(&cont->children))
3901 memcg->use_hierarchy = val;
3912 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3913 enum mem_cgroup_stat_index idx)
3915 struct mem_cgroup *iter;
3918 /* Per-cpu values can be negative, use a signed accumulator */
3919 for_each_mem_cgroup_tree(iter, memcg)
3920 val += mem_cgroup_read_stat(iter, idx);
3922 if (val < 0) /* race ? */
3927 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3931 if (!mem_cgroup_is_root(memcg)) {
3933 return res_counter_read_u64(&memcg->res, RES_USAGE);
3935 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3938 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3939 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3942 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3944 return val << PAGE_SHIFT;
3947 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3949 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3953 type = MEMFILE_TYPE(cft->private);
3954 name = MEMFILE_ATTR(cft->private);
3957 if (name == RES_USAGE)
3958 val = mem_cgroup_usage(memcg, false);
3960 val = res_counter_read_u64(&memcg->res, name);
3963 if (name == RES_USAGE)
3964 val = mem_cgroup_usage(memcg, true);
3966 val = res_counter_read_u64(&memcg->memsw, name);
3975 * The user of this function is...
3978 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3981 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3983 unsigned long long val;
3986 type = MEMFILE_TYPE(cft->private);
3987 name = MEMFILE_ATTR(cft->private);
3990 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3994 /* This function does all necessary parse...reuse it */
3995 ret = res_counter_memparse_write_strategy(buffer, &val);
3999 ret = mem_cgroup_resize_limit(memcg, val);
4001 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4003 case RES_SOFT_LIMIT:
4004 ret = res_counter_memparse_write_strategy(buffer, &val);
4008 * For memsw, soft limits are hard to implement in terms
4009 * of semantics, for now, we support soft limits for
4010 * control without swap
4013 ret = res_counter_set_soft_limit(&memcg->res, val);
4018 ret = -EINVAL; /* should be BUG() ? */
4024 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4025 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4027 struct cgroup *cgroup;
4028 unsigned long long min_limit, min_memsw_limit, tmp;
4030 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4031 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4032 cgroup = memcg->css.cgroup;
4033 if (!memcg->use_hierarchy)
4036 while (cgroup->parent) {
4037 cgroup = cgroup->parent;
4038 memcg = mem_cgroup_from_cont(cgroup);
4039 if (!memcg->use_hierarchy)
4041 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4042 min_limit = min(min_limit, tmp);
4043 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4044 min_memsw_limit = min(min_memsw_limit, tmp);
4047 *mem_limit = min_limit;
4048 *memsw_limit = min_memsw_limit;
4052 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4054 struct mem_cgroup *memcg;
4057 memcg = mem_cgroup_from_cont(cont);
4058 type = MEMFILE_TYPE(event);
4059 name = MEMFILE_ATTR(event);
4063 res_counter_reset_max(&memcg->res);
4065 res_counter_reset_max(&memcg->memsw);
4069 res_counter_reset_failcnt(&memcg->res);
4071 res_counter_reset_failcnt(&memcg->memsw);
4078 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4081 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4085 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4086 struct cftype *cft, u64 val)
4088 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4090 if (val >= (1 << NR_MOVE_TYPE))
4093 * We check this value several times in both in can_attach() and
4094 * attach(), so we need cgroup lock to prevent this value from being
4098 memcg->move_charge_at_immigrate = val;
4104 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4105 struct cftype *cft, u64 val)
4112 /* For read statistics */
4130 struct mcs_total_stat {
4131 s64 stat[NR_MCS_STAT];
4137 } memcg_stat_strings[NR_MCS_STAT] = {
4138 {"cache", "total_cache"},
4139 {"rss", "total_rss"},
4140 {"mapped_file", "total_mapped_file"},
4141 {"pgpgin", "total_pgpgin"},
4142 {"pgpgout", "total_pgpgout"},
4143 {"swap", "total_swap"},
4144 {"pgfault", "total_pgfault"},
4145 {"pgmajfault", "total_pgmajfault"},
4146 {"inactive_anon", "total_inactive_anon"},
4147 {"active_anon", "total_active_anon"},
4148 {"inactive_file", "total_inactive_file"},
4149 {"active_file", "total_active_file"},
4150 {"unevictable", "total_unevictable"}
4155 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4160 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4161 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4162 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4163 s->stat[MCS_RSS] += val * PAGE_SIZE;
4164 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4165 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4166 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4167 s->stat[MCS_PGPGIN] += val;
4168 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4169 s->stat[MCS_PGPGOUT] += val;
4170 if (do_swap_account) {
4171 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4172 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4174 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4175 s->stat[MCS_PGFAULT] += val;
4176 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4177 s->stat[MCS_PGMAJFAULT] += val;
4180 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4181 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4182 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4183 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4184 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4185 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4186 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4187 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4188 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4189 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4193 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4195 struct mem_cgroup *iter;
4197 for_each_mem_cgroup_tree(iter, memcg)
4198 mem_cgroup_get_local_stat(iter, s);
4202 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4205 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4206 unsigned long node_nr;
4207 struct cgroup *cont = m->private;
4208 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4210 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4211 seq_printf(m, "total=%lu", total_nr);
4212 for_each_node_state(nid, N_HIGH_MEMORY) {
4213 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4214 seq_printf(m, " N%d=%lu", nid, node_nr);
4218 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4219 seq_printf(m, "file=%lu", file_nr);
4220 for_each_node_state(nid, N_HIGH_MEMORY) {
4221 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4223 seq_printf(m, " N%d=%lu", nid, node_nr);
4227 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4228 seq_printf(m, "anon=%lu", anon_nr);
4229 for_each_node_state(nid, N_HIGH_MEMORY) {
4230 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4232 seq_printf(m, " N%d=%lu", nid, node_nr);
4236 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4237 seq_printf(m, "unevictable=%lu", unevictable_nr);
4238 for_each_node_state(nid, N_HIGH_MEMORY) {
4239 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4240 BIT(LRU_UNEVICTABLE));
4241 seq_printf(m, " N%d=%lu", nid, node_nr);
4246 #endif /* CONFIG_NUMA */
4248 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4249 struct cgroup_map_cb *cb)
4251 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4252 struct mcs_total_stat mystat;
4255 memset(&mystat, 0, sizeof(mystat));
4256 mem_cgroup_get_local_stat(mem_cont, &mystat);
4259 for (i = 0; i < NR_MCS_STAT; i++) {
4260 if (i == MCS_SWAP && !do_swap_account)
4262 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4265 /* Hierarchical information */
4267 unsigned long long limit, memsw_limit;
4268 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4269 cb->fill(cb, "hierarchical_memory_limit", limit);
4270 if (do_swap_account)
4271 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4274 memset(&mystat, 0, sizeof(mystat));
4275 mem_cgroup_get_total_stat(mem_cont, &mystat);
4276 for (i = 0; i < NR_MCS_STAT; i++) {
4277 if (i == MCS_SWAP && !do_swap_account)
4279 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4282 #ifdef CONFIG_DEBUG_VM
4285 struct mem_cgroup_per_zone *mz;
4286 unsigned long recent_rotated[2] = {0, 0};
4287 unsigned long recent_scanned[2] = {0, 0};
4289 for_each_online_node(nid)
4290 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4291 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4293 recent_rotated[0] +=
4294 mz->reclaim_stat.recent_rotated[0];
4295 recent_rotated[1] +=
4296 mz->reclaim_stat.recent_rotated[1];
4297 recent_scanned[0] +=
4298 mz->reclaim_stat.recent_scanned[0];
4299 recent_scanned[1] +=
4300 mz->reclaim_stat.recent_scanned[1];
4302 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4303 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4304 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4305 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4312 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4314 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4316 return mem_cgroup_swappiness(memcg);
4319 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4322 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4323 struct mem_cgroup *parent;
4328 if (cgrp->parent == NULL)
4331 parent = mem_cgroup_from_cont(cgrp->parent);
4335 /* If under hierarchy, only empty-root can set this value */
4336 if ((parent->use_hierarchy) ||
4337 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4342 memcg->swappiness = val;
4349 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4351 struct mem_cgroup_threshold_ary *t;
4357 t = rcu_dereference(memcg->thresholds.primary);
4359 t = rcu_dereference(memcg->memsw_thresholds.primary);
4364 usage = mem_cgroup_usage(memcg, swap);
4367 * current_threshold points to threshold just below usage.
4368 * If it's not true, a threshold was crossed after last
4369 * call of __mem_cgroup_threshold().
4371 i = t->current_threshold;
4374 * Iterate backward over array of thresholds starting from
4375 * current_threshold and check if a threshold is crossed.
4376 * If none of thresholds below usage is crossed, we read
4377 * only one element of the array here.
4379 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4380 eventfd_signal(t->entries[i].eventfd, 1);
4382 /* i = current_threshold + 1 */
4386 * Iterate forward over array of thresholds starting from
4387 * current_threshold+1 and check if a threshold is crossed.
4388 * If none of thresholds above usage is crossed, we read
4389 * only one element of the array here.
4391 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4392 eventfd_signal(t->entries[i].eventfd, 1);
4394 /* Update current_threshold */
4395 t->current_threshold = i - 1;
4400 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4403 __mem_cgroup_threshold(memcg, false);
4404 if (do_swap_account)
4405 __mem_cgroup_threshold(memcg, true);
4407 memcg = parent_mem_cgroup(memcg);
4411 static int compare_thresholds(const void *a, const void *b)
4413 const struct mem_cgroup_threshold *_a = a;
4414 const struct mem_cgroup_threshold *_b = b;
4416 return _a->threshold - _b->threshold;
4419 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4421 struct mem_cgroup_eventfd_list *ev;
4423 list_for_each_entry(ev, &memcg->oom_notify, list)
4424 eventfd_signal(ev->eventfd, 1);
4428 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4430 struct mem_cgroup *iter;
4432 for_each_mem_cgroup_tree(iter, memcg)
4433 mem_cgroup_oom_notify_cb(iter);
4436 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4437 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4439 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4440 struct mem_cgroup_thresholds *thresholds;
4441 struct mem_cgroup_threshold_ary *new;
4442 int type = MEMFILE_TYPE(cft->private);
4443 u64 threshold, usage;
4446 ret = res_counter_memparse_write_strategy(args, &threshold);
4450 mutex_lock(&memcg->thresholds_lock);
4453 thresholds = &memcg->thresholds;
4454 else if (type == _MEMSWAP)
4455 thresholds = &memcg->memsw_thresholds;
4459 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4461 /* Check if a threshold crossed before adding a new one */
4462 if (thresholds->primary)
4463 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4465 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4467 /* Allocate memory for new array of thresholds */
4468 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4476 /* Copy thresholds (if any) to new array */
4477 if (thresholds->primary) {
4478 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4479 sizeof(struct mem_cgroup_threshold));
4482 /* Add new threshold */
4483 new->entries[size - 1].eventfd = eventfd;
4484 new->entries[size - 1].threshold = threshold;
4486 /* Sort thresholds. Registering of new threshold isn't time-critical */
4487 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4488 compare_thresholds, NULL);
4490 /* Find current threshold */
4491 new->current_threshold = -1;
4492 for (i = 0; i < size; i++) {
4493 if (new->entries[i].threshold < usage) {
4495 * new->current_threshold will not be used until
4496 * rcu_assign_pointer(), so it's safe to increment
4499 ++new->current_threshold;
4503 /* Free old spare buffer and save old primary buffer as spare */
4504 kfree(thresholds->spare);
4505 thresholds->spare = thresholds->primary;
4507 rcu_assign_pointer(thresholds->primary, new);
4509 /* To be sure that nobody uses thresholds */
4513 mutex_unlock(&memcg->thresholds_lock);
4518 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4519 struct cftype *cft, struct eventfd_ctx *eventfd)
4521 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4522 struct mem_cgroup_thresholds *thresholds;
4523 struct mem_cgroup_threshold_ary *new;
4524 int type = MEMFILE_TYPE(cft->private);
4528 mutex_lock(&memcg->thresholds_lock);
4530 thresholds = &memcg->thresholds;
4531 else if (type == _MEMSWAP)
4532 thresholds = &memcg->memsw_thresholds;
4537 * Something went wrong if we trying to unregister a threshold
4538 * if we don't have thresholds
4540 BUG_ON(!thresholds);
4542 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4544 /* Check if a threshold crossed before removing */
4545 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4547 /* Calculate new number of threshold */
4549 for (i = 0; i < thresholds->primary->size; i++) {
4550 if (thresholds->primary->entries[i].eventfd != eventfd)
4554 new = thresholds->spare;
4556 /* Set thresholds array to NULL if we don't have thresholds */
4565 /* Copy thresholds and find current threshold */
4566 new->current_threshold = -1;
4567 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4568 if (thresholds->primary->entries[i].eventfd == eventfd)
4571 new->entries[j] = thresholds->primary->entries[i];
4572 if (new->entries[j].threshold < usage) {
4574 * new->current_threshold will not be used
4575 * until rcu_assign_pointer(), so it's safe to increment
4578 ++new->current_threshold;
4584 /* Swap primary and spare array */
4585 thresholds->spare = thresholds->primary;
4586 rcu_assign_pointer(thresholds->primary, new);
4588 /* To be sure that nobody uses thresholds */
4591 mutex_unlock(&memcg->thresholds_lock);
4594 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4595 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4597 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4598 struct mem_cgroup_eventfd_list *event;
4599 int type = MEMFILE_TYPE(cft->private);
4601 BUG_ON(type != _OOM_TYPE);
4602 event = kmalloc(sizeof(*event), GFP_KERNEL);
4606 spin_lock(&memcg_oom_lock);
4608 event->eventfd = eventfd;
4609 list_add(&event->list, &memcg->oom_notify);
4611 /* already in OOM ? */
4612 if (atomic_read(&memcg->under_oom))
4613 eventfd_signal(eventfd, 1);
4614 spin_unlock(&memcg_oom_lock);
4619 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4620 struct cftype *cft, struct eventfd_ctx *eventfd)
4622 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4623 struct mem_cgroup_eventfd_list *ev, *tmp;
4624 int type = MEMFILE_TYPE(cft->private);
4626 BUG_ON(type != _OOM_TYPE);
4628 spin_lock(&memcg_oom_lock);
4630 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4631 if (ev->eventfd == eventfd) {
4632 list_del(&ev->list);
4637 spin_unlock(&memcg_oom_lock);
4640 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4641 struct cftype *cft, struct cgroup_map_cb *cb)
4643 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4645 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4647 if (atomic_read(&memcg->under_oom))
4648 cb->fill(cb, "under_oom", 1);
4650 cb->fill(cb, "under_oom", 0);
4654 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4655 struct cftype *cft, u64 val)
4657 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4658 struct mem_cgroup *parent;
4660 /* cannot set to root cgroup and only 0 and 1 are allowed */
4661 if (!cgrp->parent || !((val == 0) || (val == 1)))
4664 parent = mem_cgroup_from_cont(cgrp->parent);
4667 /* oom-kill-disable is a flag for subhierarchy. */
4668 if ((parent->use_hierarchy) ||
4669 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4673 memcg->oom_kill_disable = val;
4675 memcg_oom_recover(memcg);
4681 static const struct file_operations mem_control_numa_stat_file_operations = {
4683 .llseek = seq_lseek,
4684 .release = single_release,
4687 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4689 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4691 file->f_op = &mem_control_numa_stat_file_operations;
4692 return single_open(file, mem_control_numa_stat_show, cont);
4694 #endif /* CONFIG_NUMA */
4696 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4697 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4700 * Part of this would be better living in a separate allocation
4701 * function, leaving us with just the cgroup tree population work.
4702 * We, however, depend on state such as network's proto_list that
4703 * is only initialized after cgroup creation. I found the less
4704 * cumbersome way to deal with it to defer it all to populate time
4706 return mem_cgroup_sockets_init(cont, ss);
4709 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4710 struct cgroup *cont)
4712 mem_cgroup_sockets_destroy(cont, ss);
4715 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4720 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4721 struct cgroup *cont)
4726 static struct cftype mem_cgroup_files[] = {
4728 .name = "usage_in_bytes",
4729 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4730 .read_u64 = mem_cgroup_read,
4731 .register_event = mem_cgroup_usage_register_event,
4732 .unregister_event = mem_cgroup_usage_unregister_event,
4735 .name = "max_usage_in_bytes",
4736 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4737 .trigger = mem_cgroup_reset,
4738 .read_u64 = mem_cgroup_read,
4741 .name = "limit_in_bytes",
4742 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4743 .write_string = mem_cgroup_write,
4744 .read_u64 = mem_cgroup_read,
4747 .name = "soft_limit_in_bytes",
4748 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4749 .write_string = mem_cgroup_write,
4750 .read_u64 = mem_cgroup_read,
4754 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4755 .trigger = mem_cgroup_reset,
4756 .read_u64 = mem_cgroup_read,
4760 .read_map = mem_control_stat_show,
4763 .name = "force_empty",
4764 .trigger = mem_cgroup_force_empty_write,
4767 .name = "use_hierarchy",
4768 .write_u64 = mem_cgroup_hierarchy_write,
4769 .read_u64 = mem_cgroup_hierarchy_read,
4772 .name = "swappiness",
4773 .read_u64 = mem_cgroup_swappiness_read,
4774 .write_u64 = mem_cgroup_swappiness_write,
4777 .name = "move_charge_at_immigrate",
4778 .read_u64 = mem_cgroup_move_charge_read,
4779 .write_u64 = mem_cgroup_move_charge_write,
4782 .name = "oom_control",
4783 .read_map = mem_cgroup_oom_control_read,
4784 .write_u64 = mem_cgroup_oom_control_write,
4785 .register_event = mem_cgroup_oom_register_event,
4786 .unregister_event = mem_cgroup_oom_unregister_event,
4787 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4791 .name = "numa_stat",
4792 .open = mem_control_numa_stat_open,
4798 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4799 static struct cftype memsw_cgroup_files[] = {
4801 .name = "memsw.usage_in_bytes",
4802 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4803 .read_u64 = mem_cgroup_read,
4804 .register_event = mem_cgroup_usage_register_event,
4805 .unregister_event = mem_cgroup_usage_unregister_event,
4808 .name = "memsw.max_usage_in_bytes",
4809 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4810 .trigger = mem_cgroup_reset,
4811 .read_u64 = mem_cgroup_read,
4814 .name = "memsw.limit_in_bytes",
4815 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4816 .write_string = mem_cgroup_write,
4817 .read_u64 = mem_cgroup_read,
4820 .name = "memsw.failcnt",
4821 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4822 .trigger = mem_cgroup_reset,
4823 .read_u64 = mem_cgroup_read,
4827 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4829 if (!do_swap_account)
4831 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4832 ARRAY_SIZE(memsw_cgroup_files));
4835 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4841 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4843 struct mem_cgroup_per_node *pn;
4844 struct mem_cgroup_per_zone *mz;
4846 int zone, tmp = node;
4848 * This routine is called against possible nodes.
4849 * But it's BUG to call kmalloc() against offline node.
4851 * TODO: this routine can waste much memory for nodes which will
4852 * never be onlined. It's better to use memory hotplug callback
4855 if (!node_state(node, N_NORMAL_MEMORY))
4857 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4861 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4862 mz = &pn->zoneinfo[zone];
4864 INIT_LIST_HEAD(&mz->lruvec.lists[l]);
4865 mz->usage_in_excess = 0;
4866 mz->on_tree = false;
4869 memcg->info.nodeinfo[node] = pn;
4873 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4875 kfree(memcg->info.nodeinfo[node]);
4878 static struct mem_cgroup *mem_cgroup_alloc(void)
4880 struct mem_cgroup *mem;
4881 int size = sizeof(struct mem_cgroup);
4883 /* Can be very big if MAX_NUMNODES is very big */
4884 if (size < PAGE_SIZE)
4885 mem = kzalloc(size, GFP_KERNEL);
4887 mem = vzalloc(size);
4892 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4895 spin_lock_init(&mem->pcp_counter_lock);
4899 if (size < PAGE_SIZE)
4907 * At destroying mem_cgroup, references from swap_cgroup can remain.
4908 * (scanning all at force_empty is too costly...)
4910 * Instead of clearing all references at force_empty, we remember
4911 * the number of reference from swap_cgroup and free mem_cgroup when
4912 * it goes down to 0.
4914 * Removal of cgroup itself succeeds regardless of refs from swap.
4917 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4921 mem_cgroup_remove_from_trees(memcg);
4922 free_css_id(&mem_cgroup_subsys, &memcg->css);
4924 for_each_node_state(node, N_POSSIBLE)
4925 free_mem_cgroup_per_zone_info(memcg, node);
4927 free_percpu(memcg->stat);
4928 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4934 static void mem_cgroup_get(struct mem_cgroup *memcg)
4936 atomic_inc(&memcg->refcnt);
4939 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4941 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4942 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4943 __mem_cgroup_free(memcg);
4945 mem_cgroup_put(parent);
4949 static void mem_cgroup_put(struct mem_cgroup *memcg)
4951 __mem_cgroup_put(memcg, 1);
4955 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4957 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4959 if (!memcg->res.parent)
4961 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4963 EXPORT_SYMBOL(parent_mem_cgroup);
4965 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4966 static void __init enable_swap_cgroup(void)
4968 if (!mem_cgroup_disabled() && really_do_swap_account)
4969 do_swap_account = 1;
4972 static void __init enable_swap_cgroup(void)
4977 static int mem_cgroup_soft_limit_tree_init(void)
4979 struct mem_cgroup_tree_per_node *rtpn;
4980 struct mem_cgroup_tree_per_zone *rtpz;
4981 int tmp, node, zone;
4983 for_each_node_state(node, N_POSSIBLE) {
4985 if (!node_state(node, N_NORMAL_MEMORY))
4987 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4991 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4993 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4994 rtpz = &rtpn->rb_tree_per_zone[zone];
4995 rtpz->rb_root = RB_ROOT;
4996 spin_lock_init(&rtpz->lock);
5002 static struct cgroup_subsys_state * __ref
5003 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5005 struct mem_cgroup *memcg, *parent;
5006 long error = -ENOMEM;
5009 memcg = mem_cgroup_alloc();
5011 return ERR_PTR(error);
5013 for_each_node_state(node, N_POSSIBLE)
5014 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5018 if (cont->parent == NULL) {
5020 enable_swap_cgroup();
5022 if (mem_cgroup_soft_limit_tree_init())
5024 root_mem_cgroup = memcg;
5025 for_each_possible_cpu(cpu) {
5026 struct memcg_stock_pcp *stock =
5027 &per_cpu(memcg_stock, cpu);
5028 INIT_WORK(&stock->work, drain_local_stock);
5030 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5032 parent = mem_cgroup_from_cont(cont->parent);
5033 memcg->use_hierarchy = parent->use_hierarchy;
5034 memcg->oom_kill_disable = parent->oom_kill_disable;
5037 if (parent && parent->use_hierarchy) {
5038 res_counter_init(&memcg->res, &parent->res);
5039 res_counter_init(&memcg->memsw, &parent->memsw);
5041 * We increment refcnt of the parent to ensure that we can
5042 * safely access it on res_counter_charge/uncharge.
5043 * This refcnt will be decremented when freeing this
5044 * mem_cgroup(see mem_cgroup_put).
5046 mem_cgroup_get(parent);
5048 res_counter_init(&memcg->res, NULL);
5049 res_counter_init(&memcg->memsw, NULL);
5051 memcg->last_scanned_node = MAX_NUMNODES;
5052 INIT_LIST_HEAD(&memcg->oom_notify);
5055 memcg->swappiness = mem_cgroup_swappiness(parent);
5056 atomic_set(&memcg->refcnt, 1);
5057 memcg->move_charge_at_immigrate = 0;
5058 mutex_init(&memcg->thresholds_lock);
5061 __mem_cgroup_free(memcg);
5062 return ERR_PTR(error);
5065 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5066 struct cgroup *cont)
5068 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5070 return mem_cgroup_force_empty(memcg, false);
5073 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5074 struct cgroup *cont)
5076 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5078 kmem_cgroup_destroy(ss, cont);
5080 mem_cgroup_put(memcg);
5083 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5084 struct cgroup *cont)
5088 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5089 ARRAY_SIZE(mem_cgroup_files));
5092 ret = register_memsw_files(cont, ss);
5095 ret = register_kmem_files(cont, ss);
5101 /* Handlers for move charge at task migration. */
5102 #define PRECHARGE_COUNT_AT_ONCE 256
5103 static int mem_cgroup_do_precharge(unsigned long count)
5106 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5107 struct mem_cgroup *memcg = mc.to;
5109 if (mem_cgroup_is_root(memcg)) {
5110 mc.precharge += count;
5111 /* we don't need css_get for root */
5114 /* try to charge at once */
5116 struct res_counter *dummy;
5118 * "memcg" cannot be under rmdir() because we've already checked
5119 * by cgroup_lock_live_cgroup() that it is not removed and we
5120 * are still under the same cgroup_mutex. So we can postpone
5123 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5125 if (do_swap_account && res_counter_charge(&memcg->memsw,
5126 PAGE_SIZE * count, &dummy)) {
5127 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5130 mc.precharge += count;
5134 /* fall back to one by one charge */
5136 if (signal_pending(current)) {
5140 if (!batch_count--) {
5141 batch_count = PRECHARGE_COUNT_AT_ONCE;
5144 ret = __mem_cgroup_try_charge(NULL,
5145 GFP_KERNEL, 1, &memcg, false);
5147 /* mem_cgroup_clear_mc() will do uncharge later */
5155 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5156 * @vma: the vma the pte to be checked belongs
5157 * @addr: the address corresponding to the pte to be checked
5158 * @ptent: the pte to be checked
5159 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5162 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5163 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5164 * move charge. if @target is not NULL, the page is stored in target->page
5165 * with extra refcnt got(Callers should handle it).
5166 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5167 * target for charge migration. if @target is not NULL, the entry is stored
5170 * Called with pte lock held.
5177 enum mc_target_type {
5178 MC_TARGET_NONE, /* not used */
5183 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5184 unsigned long addr, pte_t ptent)
5186 struct page *page = vm_normal_page(vma, addr, ptent);
5188 if (!page || !page_mapped(page))
5190 if (PageAnon(page)) {
5191 /* we don't move shared anon */
5192 if (!move_anon() || page_mapcount(page) > 2)
5194 } else if (!move_file())
5195 /* we ignore mapcount for file pages */
5197 if (!get_page_unless_zero(page))
5203 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5204 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5207 struct page *page = NULL;
5208 swp_entry_t ent = pte_to_swp_entry(ptent);
5210 if (!move_anon() || non_swap_entry(ent))
5212 usage_count = mem_cgroup_count_swap_user(ent, &page);
5213 if (usage_count > 1) { /* we don't move shared anon */
5218 if (do_swap_account)
5219 entry->val = ent.val;
5224 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5225 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5227 struct page *page = NULL;
5228 struct inode *inode;
5229 struct address_space *mapping;
5232 if (!vma->vm_file) /* anonymous vma */
5237 inode = vma->vm_file->f_path.dentry->d_inode;
5238 mapping = vma->vm_file->f_mapping;
5239 if (pte_none(ptent))
5240 pgoff = linear_page_index(vma, addr);
5241 else /* pte_file(ptent) is true */
5242 pgoff = pte_to_pgoff(ptent);
5244 /* page is moved even if it's not RSS of this task(page-faulted). */
5245 page = find_get_page(mapping, pgoff);
5248 /* shmem/tmpfs may report page out on swap: account for that too. */
5249 if (radix_tree_exceptional_entry(page)) {
5250 swp_entry_t swap = radix_to_swp_entry(page);
5251 if (do_swap_account)
5253 page = find_get_page(&swapper_space, swap.val);
5259 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5260 unsigned long addr, pte_t ptent, union mc_target *target)
5262 struct page *page = NULL;
5263 struct page_cgroup *pc;
5265 swp_entry_t ent = { .val = 0 };
5267 if (pte_present(ptent))
5268 page = mc_handle_present_pte(vma, addr, ptent);
5269 else if (is_swap_pte(ptent))
5270 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5271 else if (pte_none(ptent) || pte_file(ptent))
5272 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5274 if (!page && !ent.val)
5277 pc = lookup_page_cgroup(page);
5279 * Do only loose check w/o page_cgroup lock.
5280 * mem_cgroup_move_account() checks the pc is valid or not under
5283 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5284 ret = MC_TARGET_PAGE;
5286 target->page = page;
5288 if (!ret || !target)
5291 /* There is a swap entry and a page doesn't exist or isn't charged */
5292 if (ent.val && !ret &&
5293 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5294 ret = MC_TARGET_SWAP;
5301 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5302 unsigned long addr, unsigned long end,
5303 struct mm_walk *walk)
5305 struct vm_area_struct *vma = walk->private;
5309 split_huge_page_pmd(walk->mm, pmd);
5311 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5312 for (; addr != end; pte++, addr += PAGE_SIZE)
5313 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5314 mc.precharge++; /* increment precharge temporarily */
5315 pte_unmap_unlock(pte - 1, ptl);
5321 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5323 unsigned long precharge;
5324 struct vm_area_struct *vma;
5326 down_read(&mm->mmap_sem);
5327 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5328 struct mm_walk mem_cgroup_count_precharge_walk = {
5329 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5333 if (is_vm_hugetlb_page(vma))
5335 walk_page_range(vma->vm_start, vma->vm_end,
5336 &mem_cgroup_count_precharge_walk);
5338 up_read(&mm->mmap_sem);
5340 precharge = mc.precharge;
5346 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5348 unsigned long precharge = mem_cgroup_count_precharge(mm);
5350 VM_BUG_ON(mc.moving_task);
5351 mc.moving_task = current;
5352 return mem_cgroup_do_precharge(precharge);
5355 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5356 static void __mem_cgroup_clear_mc(void)
5358 struct mem_cgroup *from = mc.from;
5359 struct mem_cgroup *to = mc.to;
5361 /* we must uncharge all the leftover precharges from mc.to */
5363 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5367 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5368 * we must uncharge here.
5370 if (mc.moved_charge) {
5371 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5372 mc.moved_charge = 0;
5374 /* we must fixup refcnts and charges */
5375 if (mc.moved_swap) {
5376 /* uncharge swap account from the old cgroup */
5377 if (!mem_cgroup_is_root(mc.from))
5378 res_counter_uncharge(&mc.from->memsw,
5379 PAGE_SIZE * mc.moved_swap);
5380 __mem_cgroup_put(mc.from, mc.moved_swap);
5382 if (!mem_cgroup_is_root(mc.to)) {
5384 * we charged both to->res and to->memsw, so we should
5387 res_counter_uncharge(&mc.to->res,
5388 PAGE_SIZE * mc.moved_swap);
5390 /* we've already done mem_cgroup_get(mc.to) */
5393 memcg_oom_recover(from);
5394 memcg_oom_recover(to);
5395 wake_up_all(&mc.waitq);
5398 static void mem_cgroup_clear_mc(void)
5400 struct mem_cgroup *from = mc.from;
5403 * we must clear moving_task before waking up waiters at the end of
5406 mc.moving_task = NULL;
5407 __mem_cgroup_clear_mc();
5408 spin_lock(&mc.lock);
5411 spin_unlock(&mc.lock);
5412 mem_cgroup_end_move(from);
5415 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5416 struct cgroup *cgroup,
5417 struct cgroup_taskset *tset)
5419 struct task_struct *p = cgroup_taskset_first(tset);
5421 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5423 if (memcg->move_charge_at_immigrate) {
5424 struct mm_struct *mm;
5425 struct mem_cgroup *from = mem_cgroup_from_task(p);
5427 VM_BUG_ON(from == memcg);
5429 mm = get_task_mm(p);
5432 /* We move charges only when we move a owner of the mm */
5433 if (mm->owner == p) {
5436 VM_BUG_ON(mc.precharge);
5437 VM_BUG_ON(mc.moved_charge);
5438 VM_BUG_ON(mc.moved_swap);
5439 mem_cgroup_start_move(from);
5440 spin_lock(&mc.lock);
5443 spin_unlock(&mc.lock);
5444 /* We set mc.moving_task later */
5446 ret = mem_cgroup_precharge_mc(mm);
5448 mem_cgroup_clear_mc();
5455 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5456 struct cgroup *cgroup,
5457 struct cgroup_taskset *tset)
5459 mem_cgroup_clear_mc();
5462 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5463 unsigned long addr, unsigned long end,
5464 struct mm_walk *walk)
5467 struct vm_area_struct *vma = walk->private;
5471 split_huge_page_pmd(walk->mm, pmd);
5473 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5474 for (; addr != end; addr += PAGE_SIZE) {
5475 pte_t ptent = *(pte++);
5476 union mc_target target;
5479 struct page_cgroup *pc;
5485 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5487 case MC_TARGET_PAGE:
5489 if (isolate_lru_page(page))
5491 pc = lookup_page_cgroup(page);
5492 if (!mem_cgroup_move_account(page, 1, pc,
5493 mc.from, mc.to, false)) {
5495 /* we uncharge from mc.from later. */
5498 putback_lru_page(page);
5499 put: /* is_target_pte_for_mc() gets the page */
5502 case MC_TARGET_SWAP:
5504 if (!mem_cgroup_move_swap_account(ent,
5505 mc.from, mc.to, false)) {
5507 /* we fixup refcnts and charges later. */
5515 pte_unmap_unlock(pte - 1, ptl);
5520 * We have consumed all precharges we got in can_attach().
5521 * We try charge one by one, but don't do any additional
5522 * charges to mc.to if we have failed in charge once in attach()
5525 ret = mem_cgroup_do_precharge(1);
5533 static void mem_cgroup_move_charge(struct mm_struct *mm)
5535 struct vm_area_struct *vma;
5537 lru_add_drain_all();
5539 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5541 * Someone who are holding the mmap_sem might be waiting in
5542 * waitq. So we cancel all extra charges, wake up all waiters,
5543 * and retry. Because we cancel precharges, we might not be able
5544 * to move enough charges, but moving charge is a best-effort
5545 * feature anyway, so it wouldn't be a big problem.
5547 __mem_cgroup_clear_mc();
5551 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5553 struct mm_walk mem_cgroup_move_charge_walk = {
5554 .pmd_entry = mem_cgroup_move_charge_pte_range,
5558 if (is_vm_hugetlb_page(vma))
5560 ret = walk_page_range(vma->vm_start, vma->vm_end,
5561 &mem_cgroup_move_charge_walk);
5564 * means we have consumed all precharges and failed in
5565 * doing additional charge. Just abandon here.
5569 up_read(&mm->mmap_sem);
5572 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5573 struct cgroup *cont,
5574 struct cgroup_taskset *tset)
5576 struct task_struct *p = cgroup_taskset_first(tset);
5577 struct mm_struct *mm = get_task_mm(p);
5581 mem_cgroup_move_charge(mm);
5586 mem_cgroup_clear_mc();
5588 #else /* !CONFIG_MMU */
5589 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5590 struct cgroup *cgroup,
5591 struct cgroup_taskset *tset)
5595 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5596 struct cgroup *cgroup,
5597 struct cgroup_taskset *tset)
5600 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5601 struct cgroup *cont,
5602 struct cgroup_taskset *tset)
5607 struct cgroup_subsys mem_cgroup_subsys = {
5609 .subsys_id = mem_cgroup_subsys_id,
5610 .create = mem_cgroup_create,
5611 .pre_destroy = mem_cgroup_pre_destroy,
5612 .destroy = mem_cgroup_destroy,
5613 .populate = mem_cgroup_populate,
5614 .can_attach = mem_cgroup_can_attach,
5615 .cancel_attach = mem_cgroup_cancel_attach,
5616 .attach = mem_cgroup_move_task,
5621 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5622 static int __init enable_swap_account(char *s)
5624 /* consider enabled if no parameter or 1 is given */
5625 if (!strcmp(s, "1"))
5626 really_do_swap_account = 1;
5627 else if (!strcmp(s, "0"))
5628 really_do_swap_account = 0;
5631 __setup("swapaccount=", enable_swap_account);