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 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
659 enum mem_cgroup_events_index idx)
661 unsigned long val = 0;
664 for_each_online_cpu(cpu)
665 val += per_cpu(memcg->stat->events[idx], cpu);
666 #ifdef CONFIG_HOTPLUG_CPU
667 spin_lock(&memcg->pcp_counter_lock);
668 val += memcg->nocpu_base.events[idx];
669 spin_unlock(&memcg->pcp_counter_lock);
674 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
675 bool file, int nr_pages)
680 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
683 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
686 /* pagein of a big page is an event. So, ignore page size */
688 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
690 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
691 nr_pages = -nr_pages; /* for event */
694 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
700 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
701 unsigned int lru_mask)
703 struct mem_cgroup_per_zone *mz;
705 unsigned long ret = 0;
707 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
710 if (BIT(l) & lru_mask)
711 ret += MEM_CGROUP_ZSTAT(mz, l);
717 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
718 int nid, unsigned int lru_mask)
723 for (zid = 0; zid < MAX_NR_ZONES; zid++)
724 total += mem_cgroup_zone_nr_lru_pages(memcg,
730 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
731 unsigned int lru_mask)
736 for_each_node_state(nid, N_HIGH_MEMORY)
737 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
741 static bool mem_cgroup_event_ratelimit(struct mem_cgroup *memcg,
742 enum mem_cgroup_events_target target)
744 unsigned long val, next;
746 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
747 next = __this_cpu_read(memcg->stat->targets[target]);
748 /* from time_after() in jiffies.h */
749 if ((long)next - (long)val < 0) {
751 case MEM_CGROUP_TARGET_THRESH:
752 next = val + THRESHOLDS_EVENTS_TARGET;
754 case MEM_CGROUP_TARGET_SOFTLIMIT:
755 next = val + SOFTLIMIT_EVENTS_TARGET;
757 case MEM_CGROUP_TARGET_NUMAINFO:
758 next = val + NUMAINFO_EVENTS_TARGET;
763 __this_cpu_write(memcg->stat->targets[target], next);
770 * Check events in order.
773 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
776 /* threshold event is triggered in finer grain than soft limit */
777 if (unlikely(mem_cgroup_event_ratelimit(memcg,
778 MEM_CGROUP_TARGET_THRESH))) {
779 bool do_softlimit, do_numainfo;
781 do_softlimit = mem_cgroup_event_ratelimit(memcg,
782 MEM_CGROUP_TARGET_SOFTLIMIT);
784 do_numainfo = mem_cgroup_event_ratelimit(memcg,
785 MEM_CGROUP_TARGET_NUMAINFO);
789 mem_cgroup_threshold(memcg);
790 if (unlikely(do_softlimit))
791 mem_cgroup_update_tree(memcg, page);
793 if (unlikely(do_numainfo))
794 atomic_inc(&memcg->numainfo_events);
800 struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
802 return container_of(cgroup_subsys_state(cont,
803 mem_cgroup_subsys_id), struct mem_cgroup,
807 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
810 * mm_update_next_owner() may clear mm->owner to NULL
811 * if it races with swapoff, page migration, etc.
812 * So this can be called with p == NULL.
817 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
818 struct mem_cgroup, css);
821 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
823 struct mem_cgroup *memcg = NULL;
828 * Because we have no locks, mm->owner's may be being moved to other
829 * cgroup. We use css_tryget() here even if this looks
830 * pessimistic (rather than adding locks here).
834 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
835 if (unlikely(!memcg))
837 } while (!css_tryget(&memcg->css));
843 * mem_cgroup_iter - iterate over memory cgroup hierarchy
844 * @root: hierarchy root
845 * @prev: previously returned memcg, NULL on first invocation
846 * @reclaim: cookie for shared reclaim walks, NULL for full walks
848 * Returns references to children of the hierarchy below @root, or
849 * @root itself, or %NULL after a full round-trip.
851 * Caller must pass the return value in @prev on subsequent
852 * invocations for reference counting, or use mem_cgroup_iter_break()
853 * to cancel a hierarchy walk before the round-trip is complete.
855 * Reclaimers can specify a zone and a priority level in @reclaim to
856 * divide up the memcgs in the hierarchy among all concurrent
857 * reclaimers operating on the same zone and priority.
859 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
860 struct mem_cgroup *prev,
861 struct mem_cgroup_reclaim_cookie *reclaim)
863 struct mem_cgroup *memcg = NULL;
866 if (mem_cgroup_disabled())
870 root = root_mem_cgroup;
872 if (prev && !reclaim)
873 id = css_id(&prev->css);
875 if (prev && prev != root)
878 if (!root->use_hierarchy && root != root_mem_cgroup) {
885 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
886 struct cgroup_subsys_state *css;
889 int nid = zone_to_nid(reclaim->zone);
890 int zid = zone_idx(reclaim->zone);
891 struct mem_cgroup_per_zone *mz;
893 mz = mem_cgroup_zoneinfo(root, nid, zid);
894 iter = &mz->reclaim_iter[reclaim->priority];
895 if (prev && reclaim->generation != iter->generation)
901 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
903 if (css == &root->css || css_tryget(css))
904 memcg = container_of(css,
905 struct mem_cgroup, css);
914 else if (!prev && memcg)
915 reclaim->generation = iter->generation;
925 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
926 * @root: hierarchy root
927 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
929 void mem_cgroup_iter_break(struct mem_cgroup *root,
930 struct mem_cgroup *prev)
933 root = root_mem_cgroup;
934 if (prev && prev != root)
939 * Iteration constructs for visiting all cgroups (under a tree). If
940 * loops are exited prematurely (break), mem_cgroup_iter_break() must
941 * be used for reference counting.
943 #define for_each_mem_cgroup_tree(iter, root) \
944 for (iter = mem_cgroup_iter(root, NULL, NULL); \
946 iter = mem_cgroup_iter(root, iter, NULL))
948 #define for_each_mem_cgroup(iter) \
949 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
951 iter = mem_cgroup_iter(NULL, iter, NULL))
953 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
955 return (memcg == root_mem_cgroup);
958 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
960 struct mem_cgroup *memcg;
966 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
967 if (unlikely(!memcg))
972 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT]);
975 this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT]);
983 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
986 * mem_cgroup_zone_lruvec - get the lru list vector for a zone and memcg
987 * @zone: zone of the wanted lruvec
988 * @mem: memcg of the wanted lruvec
990 * Returns the lru list vector holding pages for the given @zone and
991 * @mem. This can be the global zone lruvec, if the memory controller
994 struct lruvec *mem_cgroup_zone_lruvec(struct zone *zone,
995 struct mem_cgroup *memcg)
997 struct mem_cgroup_per_zone *mz;
999 if (mem_cgroup_disabled())
1000 return &zone->lruvec;
1002 mz = mem_cgroup_zoneinfo(memcg, zone_to_nid(zone), zone_idx(zone));
1007 * Following LRU functions are allowed to be used without PCG_LOCK.
1008 * Operations are called by routine of global LRU independently from memcg.
1009 * What we have to take care of here is validness of pc->mem_cgroup.
1011 * Changes to pc->mem_cgroup happens when
1014 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
1015 * It is added to LRU before charge.
1016 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
1017 * When moving account, the page is not on LRU. It's isolated.
1021 * mem_cgroup_lru_add_list - account for adding an lru page and return lruvec
1022 * @zone: zone of the page
1026 * This function accounts for @page being added to @lru, and returns
1027 * the lruvec for the given @zone and the memcg @page is charged to.
1029 * The callsite is then responsible for physically linking the page to
1030 * the returned lruvec->lists[@lru].
1032 struct lruvec *mem_cgroup_lru_add_list(struct zone *zone, struct page *page,
1035 struct mem_cgroup_per_zone *mz;
1036 struct mem_cgroup *memcg;
1037 struct page_cgroup *pc;
1039 if (mem_cgroup_disabled())
1040 return &zone->lruvec;
1042 pc = lookup_page_cgroup(page);
1043 VM_BUG_ON(PageCgroupAcctLRU(pc));
1046 * SetPageLRU SetPageCgroupUsed
1048 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1050 * Ensure that one of the two sides adds the page to the memcg
1051 * LRU during a race.
1055 * If the page is uncharged, it may be freed soon, but it
1056 * could also be swap cache (readahead, swapoff) that needs to
1057 * be reclaimable in the future. root_mem_cgroup will babysit
1058 * it for the time being.
1060 if (PageCgroupUsed(pc)) {
1061 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1063 memcg = pc->mem_cgroup;
1064 SetPageCgroupAcctLRU(pc);
1066 memcg = root_mem_cgroup;
1067 mz = page_cgroup_zoneinfo(memcg, page);
1068 /* compound_order() is stabilized through lru_lock */
1069 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1074 * mem_cgroup_lru_del_list - account for removing an lru page
1078 * This function accounts for @page being removed from @lru.
1080 * The callsite is then responsible for physically unlinking
1083 void mem_cgroup_lru_del_list(struct page *page, enum lru_list lru)
1085 struct mem_cgroup_per_zone *mz;
1086 struct mem_cgroup *memcg;
1087 struct page_cgroup *pc;
1089 if (mem_cgroup_disabled())
1092 pc = lookup_page_cgroup(page);
1094 * root_mem_cgroup babysits uncharged LRU pages, but
1095 * PageCgroupUsed is cleared when the page is about to get
1096 * freed. PageCgroupAcctLRU remembers whether the
1097 * LRU-accounting happened against pc->mem_cgroup or
1100 if (TestClearPageCgroupAcctLRU(pc)) {
1101 VM_BUG_ON(!pc->mem_cgroup);
1102 memcg = pc->mem_cgroup;
1104 memcg = root_mem_cgroup;
1105 mz = page_cgroup_zoneinfo(memcg, 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);
1110 void mem_cgroup_lru_del(struct page *page)
1112 mem_cgroup_lru_del_list(page, page_lru(page));
1116 * mem_cgroup_lru_move_lists - account for moving a page between lrus
1117 * @zone: zone of the page
1119 * @from: current lru
1122 * This function accounts for @page being moved between the lrus @from
1123 * and @to, and returns the lruvec for the given @zone and the memcg
1124 * @page is charged to.
1126 * The callsite is then responsible for physically relinking
1127 * @page->lru to the returned lruvec->lists[@to].
1129 struct lruvec *mem_cgroup_lru_move_lists(struct zone *zone,
1134 /* XXX: Optimize this, especially for @from == @to */
1135 mem_cgroup_lru_del_list(page, from);
1136 return mem_cgroup_lru_add_list(zone, page, to);
1140 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1141 * while it's linked to lru because the page may be reused after it's fully
1142 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1143 * It's done under lock_page and expected that zone->lru_lock isnever held.
1145 static void mem_cgroup_lru_del_before_commit(struct page *page)
1148 unsigned long flags;
1149 struct zone *zone = page_zone(page);
1150 struct page_cgroup *pc = lookup_page_cgroup(page);
1153 * Doing this check without taking ->lru_lock seems wrong but this
1154 * is safe. Because if page_cgroup's USED bit is unset, the page
1155 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1156 * set, the commit after this will fail, anyway.
1157 * This all charge/uncharge is done under some mutual execustion.
1158 * So, we don't need to taking care of changes in USED bit.
1160 if (likely(!PageLRU(page)))
1163 spin_lock_irqsave(&zone->lru_lock, flags);
1164 lru = page_lru(page);
1166 * The uncharged page could still be registered to the LRU of
1167 * the stale pc->mem_cgroup.
1169 * As pc->mem_cgroup is about to get overwritten, the old LRU
1170 * accounting needs to be taken care of. Let root_mem_cgroup
1171 * babysit the page until the new memcg is responsible for it.
1173 * The PCG_USED bit is guarded by lock_page() as the page is
1174 * swapcache/pagecache.
1176 if (PageLRU(page) && PageCgroupAcctLRU(pc) && !PageCgroupUsed(pc)) {
1177 del_page_from_lru_list(zone, page, lru);
1178 add_page_to_lru_list(zone, page, lru);
1180 spin_unlock_irqrestore(&zone->lru_lock, flags);
1183 static void mem_cgroup_lru_add_after_commit(struct page *page)
1186 unsigned long flags;
1187 struct zone *zone = page_zone(page);
1188 struct page_cgroup *pc = lookup_page_cgroup(page);
1191 * SetPageLRU SetPageCgroupUsed
1193 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1195 * Ensure that one of the two sides adds the page to the memcg
1196 * LRU during a race.
1199 /* taking care of that the page is added to LRU while we commit it */
1200 if (likely(!PageLRU(page)))
1202 spin_lock_irqsave(&zone->lru_lock, flags);
1203 lru = page_lru(page);
1205 * If the page is not on the LRU, someone will soon put it
1206 * there. If it is, and also already accounted for on the
1207 * memcg-side, it must be on the right lruvec as setting
1208 * pc->mem_cgroup and PageCgroupUsed is properly ordered.
1209 * Otherwise, root_mem_cgroup has been babysitting the page
1210 * during the charge. Move it to the new memcg now.
1212 if (PageLRU(page) && !PageCgroupAcctLRU(pc)) {
1213 del_page_from_lru_list(zone, page, lru);
1214 add_page_to_lru_list(zone, page, lru);
1216 spin_unlock_irqrestore(&zone->lru_lock, flags);
1220 * Checks whether given mem is same or in the root_mem_cgroup's
1223 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1224 struct mem_cgroup *memcg)
1226 if (root_memcg != memcg) {
1227 return (root_memcg->use_hierarchy &&
1228 css_is_ancestor(&memcg->css, &root_memcg->css));
1234 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1237 struct mem_cgroup *curr = NULL;
1238 struct task_struct *p;
1240 p = find_lock_task_mm(task);
1242 curr = try_get_mem_cgroup_from_mm(p->mm);
1246 * All threads may have already detached their mm's, but the oom
1247 * killer still needs to detect if they have already been oom
1248 * killed to prevent needlessly killing additional tasks.
1251 curr = mem_cgroup_from_task(task);
1253 css_get(&curr->css);
1259 * We should check use_hierarchy of "memcg" not "curr". Because checking
1260 * use_hierarchy of "curr" here make this function true if hierarchy is
1261 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1262 * hierarchy(even if use_hierarchy is disabled in "memcg").
1264 ret = mem_cgroup_same_or_subtree(memcg, curr);
1265 css_put(&curr->css);
1269 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1271 unsigned long inactive_ratio;
1272 int nid = zone_to_nid(zone);
1273 int zid = zone_idx(zone);
1274 unsigned long inactive;
1275 unsigned long active;
1278 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1279 BIT(LRU_INACTIVE_ANON));
1280 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1281 BIT(LRU_ACTIVE_ANON));
1283 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1285 inactive_ratio = int_sqrt(10 * gb);
1289 return inactive * inactive_ratio < active;
1292 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1294 unsigned long active;
1295 unsigned long inactive;
1296 int zid = zone_idx(zone);
1297 int nid = zone_to_nid(zone);
1299 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1300 BIT(LRU_INACTIVE_FILE));
1301 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1302 BIT(LRU_ACTIVE_FILE));
1304 return (active > inactive);
1307 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1310 int nid = zone_to_nid(zone);
1311 int zid = zone_idx(zone);
1312 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1314 return &mz->reclaim_stat;
1317 struct zone_reclaim_stat *
1318 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1320 struct page_cgroup *pc;
1321 struct mem_cgroup_per_zone *mz;
1323 if (mem_cgroup_disabled())
1326 pc = lookup_page_cgroup(page);
1327 if (!PageCgroupUsed(pc))
1329 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1331 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1332 return &mz->reclaim_stat;
1335 #define mem_cgroup_from_res_counter(counter, member) \
1336 container_of(counter, struct mem_cgroup, member)
1339 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1340 * @mem: the memory cgroup
1342 * Returns the maximum amount of memory @mem can be charged with, in
1345 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1347 unsigned long long margin;
1349 margin = res_counter_margin(&memcg->res);
1350 if (do_swap_account)
1351 margin = min(margin, res_counter_margin(&memcg->memsw));
1352 return margin >> PAGE_SHIFT;
1355 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1357 struct cgroup *cgrp = memcg->css.cgroup;
1360 if (cgrp->parent == NULL)
1361 return vm_swappiness;
1363 return memcg->swappiness;
1366 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1371 spin_lock(&memcg->pcp_counter_lock);
1372 for_each_online_cpu(cpu)
1373 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1374 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1375 spin_unlock(&memcg->pcp_counter_lock);
1381 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1388 spin_lock(&memcg->pcp_counter_lock);
1389 for_each_online_cpu(cpu)
1390 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1391 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1392 spin_unlock(&memcg->pcp_counter_lock);
1396 * 2 routines for checking "mem" is under move_account() or not.
1398 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1399 * for avoiding race in accounting. If true,
1400 * pc->mem_cgroup may be overwritten.
1402 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1403 * under hierarchy of moving cgroups. This is for
1404 * waiting at hith-memory prressure caused by "move".
1407 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1409 VM_BUG_ON(!rcu_read_lock_held());
1410 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1413 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1415 struct mem_cgroup *from;
1416 struct mem_cgroup *to;
1419 * Unlike task_move routines, we access mc.to, mc.from not under
1420 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1422 spin_lock(&mc.lock);
1428 ret = mem_cgroup_same_or_subtree(memcg, from)
1429 || mem_cgroup_same_or_subtree(memcg, to);
1431 spin_unlock(&mc.lock);
1435 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1437 if (mc.moving_task && current != mc.moving_task) {
1438 if (mem_cgroup_under_move(memcg)) {
1440 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1441 /* moving charge context might have finished. */
1444 finish_wait(&mc.waitq, &wait);
1452 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1453 * @memcg: The memory cgroup that went over limit
1454 * @p: Task that is going to be killed
1456 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1459 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1461 struct cgroup *task_cgrp;
1462 struct cgroup *mem_cgrp;
1464 * Need a buffer in BSS, can't rely on allocations. The code relies
1465 * on the assumption that OOM is serialized for memory controller.
1466 * If this assumption is broken, revisit this code.
1468 static char memcg_name[PATH_MAX];
1477 mem_cgrp = memcg->css.cgroup;
1478 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1480 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1483 * Unfortunately, we are unable to convert to a useful name
1484 * But we'll still print out the usage information
1491 printk(KERN_INFO "Task in %s killed", memcg_name);
1494 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1502 * Continues from above, so we don't need an KERN_ level
1504 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1507 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1508 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1509 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1510 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1511 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1513 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1514 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1515 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1519 * This function returns the number of memcg under hierarchy tree. Returns
1520 * 1(self count) if no children.
1522 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1525 struct mem_cgroup *iter;
1527 for_each_mem_cgroup_tree(iter, memcg)
1533 * Return the memory (and swap, if configured) limit for a memcg.
1535 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1540 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1541 limit += total_swap_pages << PAGE_SHIFT;
1543 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1545 * If memsw is finite and limits the amount of swap space available
1546 * to this memcg, return that limit.
1548 return min(limit, memsw);
1551 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1553 unsigned long flags)
1555 unsigned long total = 0;
1556 bool noswap = false;
1559 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1561 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1564 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1566 drain_all_stock_async(memcg);
1567 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1569 * Allow limit shrinkers, which are triggered directly
1570 * by userspace, to catch signals and stop reclaim
1571 * after minimal progress, regardless of the margin.
1573 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1575 if (mem_cgroup_margin(memcg))
1578 * If nothing was reclaimed after two attempts, there
1579 * may be no reclaimable pages in this hierarchy.
1588 * test_mem_cgroup_node_reclaimable
1589 * @mem: the target memcg
1590 * @nid: the node ID to be checked.
1591 * @noswap : specify true here if the user wants flle only information.
1593 * This function returns whether the specified memcg contains any
1594 * reclaimable pages on a node. Returns true if there are any reclaimable
1595 * pages in the node.
1597 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1598 int nid, bool noswap)
1600 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1602 if (noswap || !total_swap_pages)
1604 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1609 #if MAX_NUMNODES > 1
1612 * Always updating the nodemask is not very good - even if we have an empty
1613 * list or the wrong list here, we can start from some node and traverse all
1614 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1617 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1621 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1622 * pagein/pageout changes since the last update.
1624 if (!atomic_read(&memcg->numainfo_events))
1626 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1629 /* make a nodemask where this memcg uses memory from */
1630 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1632 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1634 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1635 node_clear(nid, memcg->scan_nodes);
1638 atomic_set(&memcg->numainfo_events, 0);
1639 atomic_set(&memcg->numainfo_updating, 0);
1643 * Selecting a node where we start reclaim from. Because what we need is just
1644 * reducing usage counter, start from anywhere is O,K. Considering
1645 * memory reclaim from current node, there are pros. and cons.
1647 * Freeing memory from current node means freeing memory from a node which
1648 * we'll use or we've used. So, it may make LRU bad. And if several threads
1649 * hit limits, it will see a contention on a node. But freeing from remote
1650 * node means more costs for memory reclaim because of memory latency.
1652 * Now, we use round-robin. Better algorithm is welcomed.
1654 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1658 mem_cgroup_may_update_nodemask(memcg);
1659 node = memcg->last_scanned_node;
1661 node = next_node(node, memcg->scan_nodes);
1662 if (node == MAX_NUMNODES)
1663 node = first_node(memcg->scan_nodes);
1665 * We call this when we hit limit, not when pages are added to LRU.
1666 * No LRU may hold pages because all pages are UNEVICTABLE or
1667 * memcg is too small and all pages are not on LRU. In that case,
1668 * we use curret node.
1670 if (unlikely(node == MAX_NUMNODES))
1671 node = numa_node_id();
1673 memcg->last_scanned_node = node;
1678 * Check all nodes whether it contains reclaimable pages or not.
1679 * For quick scan, we make use of scan_nodes. This will allow us to skip
1680 * unused nodes. But scan_nodes is lazily updated and may not cotain
1681 * enough new information. We need to do double check.
1683 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1688 * quick check...making use of scan_node.
1689 * We can skip unused nodes.
1691 if (!nodes_empty(memcg->scan_nodes)) {
1692 for (nid = first_node(memcg->scan_nodes);
1694 nid = next_node(nid, memcg->scan_nodes)) {
1696 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1701 * Check rest of nodes.
1703 for_each_node_state(nid, N_HIGH_MEMORY) {
1704 if (node_isset(nid, memcg->scan_nodes))
1706 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1713 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1718 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1720 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1724 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1727 unsigned long *total_scanned)
1729 struct mem_cgroup *victim = NULL;
1732 unsigned long excess;
1733 unsigned long nr_scanned;
1734 struct mem_cgroup_reclaim_cookie reclaim = {
1739 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1742 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1747 * If we have not been able to reclaim
1748 * anything, it might because there are
1749 * no reclaimable pages under this hierarchy
1754 * We want to do more targeted reclaim.
1755 * excess >> 2 is not to excessive so as to
1756 * reclaim too much, nor too less that we keep
1757 * coming back to reclaim from this cgroup
1759 if (total >= (excess >> 2) ||
1760 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1765 if (!mem_cgroup_reclaimable(victim, false))
1767 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1769 *total_scanned += nr_scanned;
1770 if (!res_counter_soft_limit_excess(&root_memcg->res))
1773 mem_cgroup_iter_break(root_memcg, victim);
1778 * Check OOM-Killer is already running under our hierarchy.
1779 * If someone is running, return false.
1780 * Has to be called with memcg_oom_lock
1782 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1784 struct mem_cgroup *iter, *failed = NULL;
1786 for_each_mem_cgroup_tree(iter, memcg) {
1787 if (iter->oom_lock) {
1789 * this subtree of our hierarchy is already locked
1790 * so we cannot give a lock.
1793 mem_cgroup_iter_break(memcg, iter);
1796 iter->oom_lock = true;
1803 * OK, we failed to lock the whole subtree so we have to clean up
1804 * what we set up to the failing subtree
1806 for_each_mem_cgroup_tree(iter, memcg) {
1807 if (iter == failed) {
1808 mem_cgroup_iter_break(memcg, iter);
1811 iter->oom_lock = false;
1817 * Has to be called with memcg_oom_lock
1819 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1821 struct mem_cgroup *iter;
1823 for_each_mem_cgroup_tree(iter, memcg)
1824 iter->oom_lock = false;
1828 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1830 struct mem_cgroup *iter;
1832 for_each_mem_cgroup_tree(iter, memcg)
1833 atomic_inc(&iter->under_oom);
1836 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1838 struct mem_cgroup *iter;
1841 * When a new child is created while the hierarchy is under oom,
1842 * mem_cgroup_oom_lock() may not be called. We have to use
1843 * atomic_add_unless() here.
1845 for_each_mem_cgroup_tree(iter, memcg)
1846 atomic_add_unless(&iter->under_oom, -1, 0);
1849 static DEFINE_SPINLOCK(memcg_oom_lock);
1850 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1852 struct oom_wait_info {
1853 struct mem_cgroup *mem;
1857 static int memcg_oom_wake_function(wait_queue_t *wait,
1858 unsigned mode, int sync, void *arg)
1860 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1862 struct oom_wait_info *oom_wait_info;
1864 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1865 oom_wait_memcg = oom_wait_info->mem;
1868 * Both of oom_wait_info->mem and wake_mem are stable under us.
1869 * Then we can use css_is_ancestor without taking care of RCU.
1871 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1872 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1874 return autoremove_wake_function(wait, mode, sync, arg);
1877 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1879 /* for filtering, pass "memcg" as argument. */
1880 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1883 static void memcg_oom_recover(struct mem_cgroup *memcg)
1885 if (memcg && atomic_read(&memcg->under_oom))
1886 memcg_wakeup_oom(memcg);
1890 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1892 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1894 struct oom_wait_info owait;
1895 bool locked, need_to_kill;
1898 owait.wait.flags = 0;
1899 owait.wait.func = memcg_oom_wake_function;
1900 owait.wait.private = current;
1901 INIT_LIST_HEAD(&owait.wait.task_list);
1902 need_to_kill = true;
1903 mem_cgroup_mark_under_oom(memcg);
1905 /* At first, try to OOM lock hierarchy under memcg.*/
1906 spin_lock(&memcg_oom_lock);
1907 locked = mem_cgroup_oom_lock(memcg);
1909 * Even if signal_pending(), we can't quit charge() loop without
1910 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1911 * under OOM is always welcomed, use TASK_KILLABLE here.
1913 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1914 if (!locked || memcg->oom_kill_disable)
1915 need_to_kill = false;
1917 mem_cgroup_oom_notify(memcg);
1918 spin_unlock(&memcg_oom_lock);
1921 finish_wait(&memcg_oom_waitq, &owait.wait);
1922 mem_cgroup_out_of_memory(memcg, mask);
1925 finish_wait(&memcg_oom_waitq, &owait.wait);
1927 spin_lock(&memcg_oom_lock);
1929 mem_cgroup_oom_unlock(memcg);
1930 memcg_wakeup_oom(memcg);
1931 spin_unlock(&memcg_oom_lock);
1933 mem_cgroup_unmark_under_oom(memcg);
1935 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1937 /* Give chance to dying process */
1938 schedule_timeout_uninterruptible(1);
1943 * Currently used to update mapped file statistics, but the routine can be
1944 * generalized to update other statistics as well.
1946 * Notes: Race condition
1948 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1949 * it tends to be costly. But considering some conditions, we doesn't need
1950 * to do so _always_.
1952 * Considering "charge", lock_page_cgroup() is not required because all
1953 * file-stat operations happen after a page is attached to radix-tree. There
1954 * are no race with "charge".
1956 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1957 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1958 * if there are race with "uncharge". Statistics itself is properly handled
1961 * Considering "move", this is an only case we see a race. To make the race
1962 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1963 * possibility of race condition. If there is, we take a lock.
1966 void mem_cgroup_update_page_stat(struct page *page,
1967 enum mem_cgroup_page_stat_item idx, int val)
1969 struct mem_cgroup *memcg;
1970 struct page_cgroup *pc = lookup_page_cgroup(page);
1971 bool need_unlock = false;
1972 unsigned long uninitialized_var(flags);
1974 if (mem_cgroup_disabled())
1978 memcg = pc->mem_cgroup;
1979 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1981 /* pc->mem_cgroup is unstable ? */
1982 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
1983 /* take a lock against to access pc->mem_cgroup */
1984 move_lock_page_cgroup(pc, &flags);
1986 memcg = pc->mem_cgroup;
1987 if (!memcg || !PageCgroupUsed(pc))
1992 case MEMCG_NR_FILE_MAPPED:
1994 SetPageCgroupFileMapped(pc);
1995 else if (!page_mapped(page))
1996 ClearPageCgroupFileMapped(pc);
1997 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2003 this_cpu_add(memcg->stat->count[idx], val);
2006 if (unlikely(need_unlock))
2007 move_unlock_page_cgroup(pc, &flags);
2011 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2014 * size of first charge trial. "32" comes from vmscan.c's magic value.
2015 * TODO: maybe necessary to use big numbers in big irons.
2017 #define CHARGE_BATCH 32U
2018 struct memcg_stock_pcp {
2019 struct mem_cgroup *cached; /* this never be root cgroup */
2020 unsigned int nr_pages;
2021 struct work_struct work;
2022 unsigned long flags;
2023 #define FLUSHING_CACHED_CHARGE (0)
2025 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2026 static DEFINE_MUTEX(percpu_charge_mutex);
2029 * Try to consume stocked charge on this cpu. If success, one page is consumed
2030 * from local stock and true is returned. If the stock is 0 or charges from a
2031 * cgroup which is not current target, returns false. This stock will be
2034 static bool consume_stock(struct mem_cgroup *memcg)
2036 struct memcg_stock_pcp *stock;
2039 stock = &get_cpu_var(memcg_stock);
2040 if (memcg == stock->cached && stock->nr_pages)
2042 else /* need to call res_counter_charge */
2044 put_cpu_var(memcg_stock);
2049 * Returns stocks cached in percpu to res_counter and reset cached information.
2051 static void drain_stock(struct memcg_stock_pcp *stock)
2053 struct mem_cgroup *old = stock->cached;
2055 if (stock->nr_pages) {
2056 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2058 res_counter_uncharge(&old->res, bytes);
2059 if (do_swap_account)
2060 res_counter_uncharge(&old->memsw, bytes);
2061 stock->nr_pages = 0;
2063 stock->cached = NULL;
2067 * This must be called under preempt disabled or must be called by
2068 * a thread which is pinned to local cpu.
2070 static void drain_local_stock(struct work_struct *dummy)
2072 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2074 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2078 * Cache charges(val) which is from res_counter, to local per_cpu area.
2079 * This will be consumed by consume_stock() function, later.
2081 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2083 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2085 if (stock->cached != memcg) { /* reset if necessary */
2087 stock->cached = memcg;
2089 stock->nr_pages += nr_pages;
2090 put_cpu_var(memcg_stock);
2094 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2095 * of the hierarchy under it. sync flag says whether we should block
2096 * until the work is done.
2098 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2102 /* Notify other cpus that system-wide "drain" is running */
2105 for_each_online_cpu(cpu) {
2106 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2107 struct mem_cgroup *memcg;
2109 memcg = stock->cached;
2110 if (!memcg || !stock->nr_pages)
2112 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2114 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2116 drain_local_stock(&stock->work);
2118 schedule_work_on(cpu, &stock->work);
2126 for_each_online_cpu(cpu) {
2127 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2128 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2129 flush_work(&stock->work);
2136 * Tries to drain stocked charges in other cpus. This function is asynchronous
2137 * and just put a work per cpu for draining localy on each cpu. Caller can
2138 * expects some charges will be back to res_counter later but cannot wait for
2141 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2144 * If someone calls draining, avoid adding more kworker runs.
2146 if (!mutex_trylock(&percpu_charge_mutex))
2148 drain_all_stock(root_memcg, false);
2149 mutex_unlock(&percpu_charge_mutex);
2152 /* This is a synchronous drain interface. */
2153 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2155 /* called when force_empty is called */
2156 mutex_lock(&percpu_charge_mutex);
2157 drain_all_stock(root_memcg, true);
2158 mutex_unlock(&percpu_charge_mutex);
2162 * This function drains percpu counter value from DEAD cpu and
2163 * move it to local cpu. Note that this function can be preempted.
2165 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2169 spin_lock(&memcg->pcp_counter_lock);
2170 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2171 long x = per_cpu(memcg->stat->count[i], cpu);
2173 per_cpu(memcg->stat->count[i], cpu) = 0;
2174 memcg->nocpu_base.count[i] += x;
2176 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2177 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2179 per_cpu(memcg->stat->events[i], cpu) = 0;
2180 memcg->nocpu_base.events[i] += x;
2182 /* need to clear ON_MOVE value, works as a kind of lock. */
2183 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2184 spin_unlock(&memcg->pcp_counter_lock);
2187 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2189 int idx = MEM_CGROUP_ON_MOVE;
2191 spin_lock(&memcg->pcp_counter_lock);
2192 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2193 spin_unlock(&memcg->pcp_counter_lock);
2196 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2197 unsigned long action,
2200 int cpu = (unsigned long)hcpu;
2201 struct memcg_stock_pcp *stock;
2202 struct mem_cgroup *iter;
2204 if ((action == CPU_ONLINE)) {
2205 for_each_mem_cgroup(iter)
2206 synchronize_mem_cgroup_on_move(iter, cpu);
2210 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2213 for_each_mem_cgroup(iter)
2214 mem_cgroup_drain_pcp_counter(iter, cpu);
2216 stock = &per_cpu(memcg_stock, cpu);
2222 /* See __mem_cgroup_try_charge() for details */
2224 CHARGE_OK, /* success */
2225 CHARGE_RETRY, /* need to retry but retry is not bad */
2226 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2227 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2228 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2231 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2232 unsigned int nr_pages, bool oom_check)
2234 unsigned long csize = nr_pages * PAGE_SIZE;
2235 struct mem_cgroup *mem_over_limit;
2236 struct res_counter *fail_res;
2237 unsigned long flags = 0;
2240 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2243 if (!do_swap_account)
2245 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2249 res_counter_uncharge(&memcg->res, csize);
2250 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2251 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2253 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2255 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2256 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2258 * Never reclaim on behalf of optional batching, retry with a
2259 * single page instead.
2261 if (nr_pages == CHARGE_BATCH)
2262 return CHARGE_RETRY;
2264 if (!(gfp_mask & __GFP_WAIT))
2265 return CHARGE_WOULDBLOCK;
2267 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2268 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2269 return CHARGE_RETRY;
2271 * Even though the limit is exceeded at this point, reclaim
2272 * may have been able to free some pages. Retry the charge
2273 * before killing the task.
2275 * Only for regular pages, though: huge pages are rather
2276 * unlikely to succeed so close to the limit, and we fall back
2277 * to regular pages anyway in case of failure.
2279 if (nr_pages == 1 && ret)
2280 return CHARGE_RETRY;
2283 * At task move, charge accounts can be doubly counted. So, it's
2284 * better to wait until the end of task_move if something is going on.
2286 if (mem_cgroup_wait_acct_move(mem_over_limit))
2287 return CHARGE_RETRY;
2289 /* If we don't need to call oom-killer at el, return immediately */
2291 return CHARGE_NOMEM;
2293 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2294 return CHARGE_OOM_DIE;
2296 return CHARGE_RETRY;
2300 * Unlike exported interface, "oom" parameter is added. if oom==true,
2301 * oom-killer can be invoked.
2303 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2305 unsigned int nr_pages,
2306 struct mem_cgroup **ptr,
2309 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2310 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2311 struct mem_cgroup *memcg = NULL;
2315 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2316 * in system level. So, allow to go ahead dying process in addition to
2319 if (unlikely(test_thread_flag(TIF_MEMDIE)
2320 || fatal_signal_pending(current)))
2324 * We always charge the cgroup the mm_struct belongs to.
2325 * The mm_struct's mem_cgroup changes on task migration if the
2326 * thread group leader migrates. It's possible that mm is not
2327 * set, if so charge the init_mm (happens for pagecache usage).
2332 if (*ptr) { /* css should be a valid one */
2334 VM_BUG_ON(css_is_removed(&memcg->css));
2335 if (mem_cgroup_is_root(memcg))
2337 if (nr_pages == 1 && consume_stock(memcg))
2339 css_get(&memcg->css);
2341 struct task_struct *p;
2344 p = rcu_dereference(mm->owner);
2346 * Because we don't have task_lock(), "p" can exit.
2347 * In that case, "memcg" can point to root or p can be NULL with
2348 * race with swapoff. Then, we have small risk of mis-accouning.
2349 * But such kind of mis-account by race always happens because
2350 * we don't have cgroup_mutex(). It's overkill and we allo that
2352 * (*) swapoff at el will charge against mm-struct not against
2353 * task-struct. So, mm->owner can be NULL.
2355 memcg = mem_cgroup_from_task(p);
2356 if (!memcg || mem_cgroup_is_root(memcg)) {
2360 if (nr_pages == 1 && consume_stock(memcg)) {
2362 * It seems dagerous to access memcg without css_get().
2363 * But considering how consume_stok works, it's not
2364 * necessary. If consume_stock success, some charges
2365 * from this memcg are cached on this cpu. So, we
2366 * don't need to call css_get()/css_tryget() before
2367 * calling consume_stock().
2372 /* after here, we may be blocked. we need to get refcnt */
2373 if (!css_tryget(&memcg->css)) {
2383 /* If killed, bypass charge */
2384 if (fatal_signal_pending(current)) {
2385 css_put(&memcg->css);
2390 if (oom && !nr_oom_retries) {
2392 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2395 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2399 case CHARGE_RETRY: /* not in OOM situation but retry */
2401 css_put(&memcg->css);
2404 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2405 css_put(&memcg->css);
2407 case CHARGE_NOMEM: /* OOM routine works */
2409 css_put(&memcg->css);
2412 /* If oom, we never return -ENOMEM */
2415 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2416 css_put(&memcg->css);
2419 } while (ret != CHARGE_OK);
2421 if (batch > nr_pages)
2422 refill_stock(memcg, batch - nr_pages);
2423 css_put(&memcg->css);
2436 * Somemtimes we have to undo a charge we got by try_charge().
2437 * This function is for that and do uncharge, put css's refcnt.
2438 * gotten by try_charge().
2440 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2441 unsigned int nr_pages)
2443 if (!mem_cgroup_is_root(memcg)) {
2444 unsigned long bytes = nr_pages * PAGE_SIZE;
2446 res_counter_uncharge(&memcg->res, bytes);
2447 if (do_swap_account)
2448 res_counter_uncharge(&memcg->memsw, bytes);
2453 * A helper function to get mem_cgroup from ID. must be called under
2454 * rcu_read_lock(). The caller must check css_is_removed() or some if
2455 * it's concern. (dropping refcnt from swap can be called against removed
2458 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2460 struct cgroup_subsys_state *css;
2462 /* ID 0 is unused ID */
2465 css = css_lookup(&mem_cgroup_subsys, id);
2468 return container_of(css, struct mem_cgroup, css);
2471 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2473 struct mem_cgroup *memcg = NULL;
2474 struct page_cgroup *pc;
2478 VM_BUG_ON(!PageLocked(page));
2480 pc = lookup_page_cgroup(page);
2481 lock_page_cgroup(pc);
2482 if (PageCgroupUsed(pc)) {
2483 memcg = pc->mem_cgroup;
2484 if (memcg && !css_tryget(&memcg->css))
2486 } else if (PageSwapCache(page)) {
2487 ent.val = page_private(page);
2488 id = lookup_swap_cgroup_id(ent);
2490 memcg = mem_cgroup_lookup(id);
2491 if (memcg && !css_tryget(&memcg->css))
2495 unlock_page_cgroup(pc);
2499 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2501 unsigned int nr_pages,
2502 struct page_cgroup *pc,
2503 enum charge_type ctype)
2505 lock_page_cgroup(pc);
2506 if (unlikely(PageCgroupUsed(pc))) {
2507 unlock_page_cgroup(pc);
2508 __mem_cgroup_cancel_charge(memcg, nr_pages);
2512 * we don't need page_cgroup_lock about tail pages, becase they are not
2513 * accessed by any other context at this point.
2515 pc->mem_cgroup = memcg;
2517 * We access a page_cgroup asynchronously without lock_page_cgroup().
2518 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2519 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2520 * before USED bit, we need memory barrier here.
2521 * See mem_cgroup_add_lru_list(), etc.
2525 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2526 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2527 SetPageCgroupCache(pc);
2528 SetPageCgroupUsed(pc);
2530 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2531 ClearPageCgroupCache(pc);
2532 SetPageCgroupUsed(pc);
2538 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2539 unlock_page_cgroup(pc);
2541 * "charge_statistics" updated event counter. Then, check it.
2542 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2543 * if they exceeds softlimit.
2545 memcg_check_events(memcg, page);
2548 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2550 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2551 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2553 * Because tail pages are not marked as "used", set it. We're under
2554 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2555 * charge/uncharge will be never happen and move_account() is done under
2556 * compound_lock(), so we don't have to take care of races.
2558 void mem_cgroup_split_huge_fixup(struct page *head)
2560 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2561 struct page_cgroup *pc;
2564 if (mem_cgroup_disabled())
2566 for (i = 1; i < HPAGE_PMD_NR; i++) {
2568 pc->mem_cgroup = head_pc->mem_cgroup;
2569 smp_wmb();/* see __commit_charge() */
2571 * LRU flags cannot be copied because we need to add tail
2572 * page to LRU by generic call and our hooks will be called.
2574 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2577 if (PageCgroupAcctLRU(head_pc)) {
2579 struct mem_cgroup_per_zone *mz;
2581 * We hold lru_lock, then, reduce counter directly.
2583 lru = page_lru(head);
2584 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2585 MEM_CGROUP_ZSTAT(mz, lru) -= HPAGE_PMD_NR - 1;
2591 * mem_cgroup_move_account - move account of the page
2593 * @nr_pages: number of regular pages (>1 for huge pages)
2594 * @pc: page_cgroup of the page.
2595 * @from: mem_cgroup which the page is moved from.
2596 * @to: mem_cgroup which the page is moved to. @from != @to.
2597 * @uncharge: whether we should call uncharge and css_put against @from.
2599 * The caller must confirm following.
2600 * - page is not on LRU (isolate_page() is useful.)
2601 * - compound_lock is held when nr_pages > 1
2603 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2604 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2605 * true, this function does "uncharge" from old cgroup, but it doesn't if
2606 * @uncharge is false, so a caller should do "uncharge".
2608 static int mem_cgroup_move_account(struct page *page,
2609 unsigned int nr_pages,
2610 struct page_cgroup *pc,
2611 struct mem_cgroup *from,
2612 struct mem_cgroup *to,
2615 unsigned long flags;
2618 VM_BUG_ON(from == to);
2619 VM_BUG_ON(PageLRU(page));
2621 * The page is isolated from LRU. So, collapse function
2622 * will not handle this page. But page splitting can happen.
2623 * Do this check under compound_page_lock(). The caller should
2627 if (nr_pages > 1 && !PageTransHuge(page))
2630 lock_page_cgroup(pc);
2633 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2636 move_lock_page_cgroup(pc, &flags);
2638 if (PageCgroupFileMapped(pc)) {
2639 /* Update mapped_file data for mem_cgroup */
2641 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2642 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2645 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2647 /* This is not "cancel", but cancel_charge does all we need. */
2648 __mem_cgroup_cancel_charge(from, nr_pages);
2650 /* caller should have done css_get */
2651 pc->mem_cgroup = to;
2652 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2654 * We charges against "to" which may not have any tasks. Then, "to"
2655 * can be under rmdir(). But in current implementation, caller of
2656 * this function is just force_empty() and move charge, so it's
2657 * guaranteed that "to" is never removed. So, we don't check rmdir
2660 move_unlock_page_cgroup(pc, &flags);
2663 unlock_page_cgroup(pc);
2667 memcg_check_events(to, page);
2668 memcg_check_events(from, page);
2674 * move charges to its parent.
2677 static int mem_cgroup_move_parent(struct page *page,
2678 struct page_cgroup *pc,
2679 struct mem_cgroup *child,
2682 struct cgroup *cg = child->css.cgroup;
2683 struct cgroup *pcg = cg->parent;
2684 struct mem_cgroup *parent;
2685 unsigned int nr_pages;
2686 unsigned long uninitialized_var(flags);
2694 if (!get_page_unless_zero(page))
2696 if (isolate_lru_page(page))
2699 nr_pages = hpage_nr_pages(page);
2701 parent = mem_cgroup_from_cont(pcg);
2702 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2707 flags = compound_lock_irqsave(page);
2709 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2711 __mem_cgroup_cancel_charge(parent, nr_pages);
2714 compound_unlock_irqrestore(page, flags);
2716 putback_lru_page(page);
2724 * Charge the memory controller for page usage.
2726 * 0 if the charge was successful
2727 * < 0 if the cgroup is over its limit
2729 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2730 gfp_t gfp_mask, enum charge_type ctype)
2732 struct mem_cgroup *memcg = NULL;
2733 unsigned int nr_pages = 1;
2734 struct page_cgroup *pc;
2738 if (PageTransHuge(page)) {
2739 nr_pages <<= compound_order(page);
2740 VM_BUG_ON(!PageTransHuge(page));
2742 * Never OOM-kill a process for a huge page. The
2743 * fault handler will fall back to regular pages.
2748 pc = lookup_page_cgroup(page);
2749 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2753 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2757 int mem_cgroup_newpage_charge(struct page *page,
2758 struct mm_struct *mm, gfp_t gfp_mask)
2760 if (mem_cgroup_disabled())
2762 VM_BUG_ON(page_mapped(page));
2763 VM_BUG_ON(page->mapping && !PageAnon(page));
2765 return mem_cgroup_charge_common(page, mm, gfp_mask,
2766 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2770 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2771 enum charge_type ctype);
2774 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2775 enum charge_type ctype)
2777 struct page_cgroup *pc = lookup_page_cgroup(page);
2779 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2780 * is already on LRU. It means the page may on some other page_cgroup's
2781 * LRU. Take care of it.
2783 mem_cgroup_lru_del_before_commit(page);
2784 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2785 mem_cgroup_lru_add_after_commit(page);
2789 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2792 struct mem_cgroup *memcg = NULL;
2795 if (mem_cgroup_disabled())
2797 if (PageCompound(page))
2803 if (page_is_file_cache(page)) {
2804 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2809 * FUSE reuses pages without going through the final
2810 * put that would remove them from the LRU list, make
2811 * sure that they get relinked properly.
2813 __mem_cgroup_commit_charge_lrucare(page, memcg,
2814 MEM_CGROUP_CHARGE_TYPE_CACHE);
2818 if (PageSwapCache(page)) {
2819 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2821 __mem_cgroup_commit_charge_swapin(page, memcg,
2822 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2824 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2825 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2831 * While swap-in, try_charge -> commit or cancel, the page is locked.
2832 * And when try_charge() successfully returns, one refcnt to memcg without
2833 * struct page_cgroup is acquired. This refcnt will be consumed by
2834 * "commit()" or removed by "cancel()"
2836 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2838 gfp_t mask, struct mem_cgroup **memcgp)
2840 struct mem_cgroup *memcg;
2845 if (mem_cgroup_disabled())
2848 if (!do_swap_account)
2851 * A racing thread's fault, or swapoff, may have already updated
2852 * the pte, and even removed page from swap cache: in those cases
2853 * do_swap_page()'s pte_same() test will fail; but there's also a
2854 * KSM case which does need to charge the page.
2856 if (!PageSwapCache(page))
2858 memcg = try_get_mem_cgroup_from_page(page);
2862 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2863 css_put(&memcg->css);
2868 return __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2872 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2873 enum charge_type ctype)
2875 if (mem_cgroup_disabled())
2879 cgroup_exclude_rmdir(&memcg->css);
2881 __mem_cgroup_commit_charge_lrucare(page, memcg, ctype);
2883 * Now swap is on-memory. This means this page may be
2884 * counted both as mem and swap....double count.
2885 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2886 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2887 * may call delete_from_swap_cache() before reach here.
2889 if (do_swap_account && PageSwapCache(page)) {
2890 swp_entry_t ent = {.val = page_private(page)};
2891 struct mem_cgroup *swap_memcg;
2894 id = swap_cgroup_record(ent, 0);
2896 swap_memcg = mem_cgroup_lookup(id);
2899 * This recorded memcg can be obsolete one. So, avoid
2900 * calling css_tryget
2902 if (!mem_cgroup_is_root(swap_memcg))
2903 res_counter_uncharge(&swap_memcg->memsw,
2905 mem_cgroup_swap_statistics(swap_memcg, false);
2906 mem_cgroup_put(swap_memcg);
2911 * At swapin, we may charge account against cgroup which has no tasks.
2912 * So, rmdir()->pre_destroy() can be called while we do this charge.
2913 * In that case, we need to call pre_destroy() again. check it here.
2915 cgroup_release_and_wakeup_rmdir(&memcg->css);
2918 void mem_cgroup_commit_charge_swapin(struct page *page,
2919 struct mem_cgroup *memcg)
2921 __mem_cgroup_commit_charge_swapin(page, memcg,
2922 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2925 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2927 if (mem_cgroup_disabled())
2931 __mem_cgroup_cancel_charge(memcg, 1);
2934 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2935 unsigned int nr_pages,
2936 const enum charge_type ctype)
2938 struct memcg_batch_info *batch = NULL;
2939 bool uncharge_memsw = true;
2941 /* If swapout, usage of swap doesn't decrease */
2942 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2943 uncharge_memsw = false;
2945 batch = ¤t->memcg_batch;
2947 * In usual, we do css_get() when we remember memcg pointer.
2948 * But in this case, we keep res->usage until end of a series of
2949 * uncharges. Then, it's ok to ignore memcg's refcnt.
2952 batch->memcg = memcg;
2954 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2955 * In those cases, all pages freed continuously can be expected to be in
2956 * the same cgroup and we have chance to coalesce uncharges.
2957 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2958 * because we want to do uncharge as soon as possible.
2961 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2962 goto direct_uncharge;
2965 goto direct_uncharge;
2968 * In typical case, batch->memcg == mem. This means we can
2969 * merge a series of uncharges to an uncharge of res_counter.
2970 * If not, we uncharge res_counter ony by one.
2972 if (batch->memcg != memcg)
2973 goto direct_uncharge;
2974 /* remember freed charge and uncharge it later */
2977 batch->memsw_nr_pages++;
2980 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2982 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2983 if (unlikely(batch->memcg != memcg))
2984 memcg_oom_recover(memcg);
2989 * uncharge if !page_mapped(page)
2991 static struct mem_cgroup *
2992 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2994 struct mem_cgroup *memcg = NULL;
2995 unsigned int nr_pages = 1;
2996 struct page_cgroup *pc;
2998 if (mem_cgroup_disabled())
3001 if (PageSwapCache(page))
3004 if (PageTransHuge(page)) {
3005 nr_pages <<= compound_order(page);
3006 VM_BUG_ON(!PageTransHuge(page));
3009 * Check if our page_cgroup is valid
3011 pc = lookup_page_cgroup(page);
3012 if (unlikely(!PageCgroupUsed(pc)))
3015 lock_page_cgroup(pc);
3017 memcg = pc->mem_cgroup;
3019 if (!PageCgroupUsed(pc))
3023 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3024 case MEM_CGROUP_CHARGE_TYPE_DROP:
3025 /* See mem_cgroup_prepare_migration() */
3026 if (page_mapped(page) || PageCgroupMigration(pc))
3029 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3030 if (!PageAnon(page)) { /* Shared memory */
3031 if (page->mapping && !page_is_file_cache(page))
3033 } else if (page_mapped(page)) /* Anon */
3040 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3042 ClearPageCgroupUsed(pc);
3044 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3045 * freed from LRU. This is safe because uncharged page is expected not
3046 * to be reused (freed soon). Exception is SwapCache, it's handled by
3047 * special functions.
3050 unlock_page_cgroup(pc);
3052 * even after unlock, we have memcg->res.usage here and this memcg
3053 * will never be freed.
3055 memcg_check_events(memcg, page);
3056 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3057 mem_cgroup_swap_statistics(memcg, true);
3058 mem_cgroup_get(memcg);
3060 if (!mem_cgroup_is_root(memcg))
3061 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3066 unlock_page_cgroup(pc);
3070 void mem_cgroup_uncharge_page(struct page *page)
3073 if (page_mapped(page))
3075 VM_BUG_ON(page->mapping && !PageAnon(page));
3076 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3079 void mem_cgroup_uncharge_cache_page(struct page *page)
3081 VM_BUG_ON(page_mapped(page));
3082 VM_BUG_ON(page->mapping);
3083 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3087 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3088 * In that cases, pages are freed continuously and we can expect pages
3089 * are in the same memcg. All these calls itself limits the number of
3090 * pages freed at once, then uncharge_start/end() is called properly.
3091 * This may be called prural(2) times in a context,
3094 void mem_cgroup_uncharge_start(void)
3096 current->memcg_batch.do_batch++;
3097 /* We can do nest. */
3098 if (current->memcg_batch.do_batch == 1) {
3099 current->memcg_batch.memcg = NULL;
3100 current->memcg_batch.nr_pages = 0;
3101 current->memcg_batch.memsw_nr_pages = 0;
3105 void mem_cgroup_uncharge_end(void)
3107 struct memcg_batch_info *batch = ¤t->memcg_batch;
3109 if (!batch->do_batch)
3113 if (batch->do_batch) /* If stacked, do nothing. */
3119 * This "batch->memcg" is valid without any css_get/put etc...
3120 * bacause we hide charges behind us.
3122 if (batch->nr_pages)
3123 res_counter_uncharge(&batch->memcg->res,
3124 batch->nr_pages * PAGE_SIZE);
3125 if (batch->memsw_nr_pages)
3126 res_counter_uncharge(&batch->memcg->memsw,
3127 batch->memsw_nr_pages * PAGE_SIZE);
3128 memcg_oom_recover(batch->memcg);
3129 /* forget this pointer (for sanity check) */
3130 batch->memcg = NULL;
3135 * called after __delete_from_swap_cache() and drop "page" account.
3136 * memcg information is recorded to swap_cgroup of "ent"
3139 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3141 struct mem_cgroup *memcg;
3142 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3144 if (!swapout) /* this was a swap cache but the swap is unused ! */
3145 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3147 memcg = __mem_cgroup_uncharge_common(page, ctype);
3150 * record memcg information, if swapout && memcg != NULL,
3151 * mem_cgroup_get() was called in uncharge().
3153 if (do_swap_account && swapout && memcg)
3154 swap_cgroup_record(ent, css_id(&memcg->css));
3158 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3160 * called from swap_entry_free(). remove record in swap_cgroup and
3161 * uncharge "memsw" account.
3163 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3165 struct mem_cgroup *memcg;
3168 if (!do_swap_account)
3171 id = swap_cgroup_record(ent, 0);
3173 memcg = mem_cgroup_lookup(id);
3176 * We uncharge this because swap is freed.
3177 * This memcg can be obsolete one. We avoid calling css_tryget
3179 if (!mem_cgroup_is_root(memcg))
3180 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3181 mem_cgroup_swap_statistics(memcg, false);
3182 mem_cgroup_put(memcg);
3188 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3189 * @entry: swap entry to be moved
3190 * @from: mem_cgroup which the entry is moved from
3191 * @to: mem_cgroup which the entry is moved to
3192 * @need_fixup: whether we should fixup res_counters and refcounts.
3194 * It succeeds only when the swap_cgroup's record for this entry is the same
3195 * as the mem_cgroup's id of @from.
3197 * Returns 0 on success, -EINVAL on failure.
3199 * The caller must have charged to @to, IOW, called res_counter_charge() about
3200 * both res and memsw, and called css_get().
3202 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3203 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3205 unsigned short old_id, new_id;
3207 old_id = css_id(&from->css);
3208 new_id = css_id(&to->css);
3210 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3211 mem_cgroup_swap_statistics(from, false);
3212 mem_cgroup_swap_statistics(to, true);
3214 * This function is only called from task migration context now.
3215 * It postpones res_counter and refcount handling till the end
3216 * of task migration(mem_cgroup_clear_mc()) for performance
3217 * improvement. But we cannot postpone mem_cgroup_get(to)
3218 * because if the process that has been moved to @to does
3219 * swap-in, the refcount of @to might be decreased to 0.
3223 if (!mem_cgroup_is_root(from))
3224 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3225 mem_cgroup_put(from);
3227 * we charged both to->res and to->memsw, so we should
3230 if (!mem_cgroup_is_root(to))
3231 res_counter_uncharge(&to->res, PAGE_SIZE);
3238 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3239 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3246 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3249 int mem_cgroup_prepare_migration(struct page *page,
3250 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask)
3252 struct mem_cgroup *memcg = NULL;
3253 struct page_cgroup *pc;
3254 enum charge_type ctype;
3259 VM_BUG_ON(PageTransHuge(page));
3260 if (mem_cgroup_disabled())
3263 pc = lookup_page_cgroup(page);
3264 lock_page_cgroup(pc);
3265 if (PageCgroupUsed(pc)) {
3266 memcg = pc->mem_cgroup;
3267 css_get(&memcg->css);
3269 * At migrating an anonymous page, its mapcount goes down
3270 * to 0 and uncharge() will be called. But, even if it's fully
3271 * unmapped, migration may fail and this page has to be
3272 * charged again. We set MIGRATION flag here and delay uncharge
3273 * until end_migration() is called
3275 * Corner Case Thinking
3277 * When the old page was mapped as Anon and it's unmap-and-freed
3278 * while migration was ongoing.
3279 * If unmap finds the old page, uncharge() of it will be delayed
3280 * until end_migration(). If unmap finds a new page, it's
3281 * uncharged when it make mapcount to be 1->0. If unmap code
3282 * finds swap_migration_entry, the new page will not be mapped
3283 * and end_migration() will find it(mapcount==0).
3286 * When the old page was mapped but migraion fails, the kernel
3287 * remaps it. A charge for it is kept by MIGRATION flag even
3288 * if mapcount goes down to 0. We can do remap successfully
3289 * without charging it again.
3292 * The "old" page is under lock_page() until the end of
3293 * migration, so, the old page itself will not be swapped-out.
3294 * If the new page is swapped out before end_migraton, our
3295 * hook to usual swap-out path will catch the event.
3298 SetPageCgroupMigration(pc);
3300 unlock_page_cgroup(pc);
3302 * If the page is not charged at this point,
3309 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false);
3310 css_put(&memcg->css);/* drop extra refcnt */
3311 if (ret || *memcgp == NULL) {
3312 if (PageAnon(page)) {
3313 lock_page_cgroup(pc);
3314 ClearPageCgroupMigration(pc);
3315 unlock_page_cgroup(pc);
3317 * The old page may be fully unmapped while we kept it.
3319 mem_cgroup_uncharge_page(page);
3324 * We charge new page before it's used/mapped. So, even if unlock_page()
3325 * is called before end_migration, we can catch all events on this new
3326 * page. In the case new page is migrated but not remapped, new page's
3327 * mapcount will be finally 0 and we call uncharge in end_migration().
3329 pc = lookup_page_cgroup(newpage);
3331 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3332 else if (page_is_file_cache(page))
3333 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3335 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3336 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3340 /* remove redundant charge if migration failed*/
3341 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3342 struct page *oldpage, struct page *newpage, bool migration_ok)
3344 struct page *used, *unused;
3345 struct page_cgroup *pc;
3349 /* blocks rmdir() */
3350 cgroup_exclude_rmdir(&memcg->css);
3351 if (!migration_ok) {
3359 * We disallowed uncharge of pages under migration because mapcount
3360 * of the page goes down to zero, temporarly.
3361 * Clear the flag and check the page should be charged.
3363 pc = lookup_page_cgroup(oldpage);
3364 lock_page_cgroup(pc);
3365 ClearPageCgroupMigration(pc);
3366 unlock_page_cgroup(pc);
3368 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3371 * If a page is a file cache, radix-tree replacement is very atomic
3372 * and we can skip this check. When it was an Anon page, its mapcount
3373 * goes down to 0. But because we added MIGRATION flage, it's not
3374 * uncharged yet. There are several case but page->mapcount check
3375 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3376 * check. (see prepare_charge() also)
3379 mem_cgroup_uncharge_page(used);
3381 * At migration, we may charge account against cgroup which has no
3383 * So, rmdir()->pre_destroy() can be called while we do this charge.
3384 * In that case, we need to call pre_destroy() again. check it here.
3386 cgroup_release_and_wakeup_rmdir(&memcg->css);
3390 * At replace page cache, newpage is not under any memcg but it's on
3391 * LRU. So, this function doesn't touch res_counter but handles LRU
3392 * in correct way. Both pages are locked so we cannot race with uncharge.
3394 void mem_cgroup_replace_page_cache(struct page *oldpage,
3395 struct page *newpage)
3397 struct mem_cgroup *memcg;
3398 struct page_cgroup *pc;
3400 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3401 unsigned long flags;
3403 if (mem_cgroup_disabled())
3406 pc = lookup_page_cgroup(oldpage);
3407 /* fix accounting on old pages */
3408 lock_page_cgroup(pc);
3409 memcg = pc->mem_cgroup;
3410 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1);
3411 ClearPageCgroupUsed(pc);
3412 unlock_page_cgroup(pc);
3414 if (PageSwapBacked(oldpage))
3415 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3417 zone = page_zone(newpage);
3418 pc = lookup_page_cgroup(newpage);
3420 * Even if newpage->mapping was NULL before starting replacement,
3421 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3422 * LRU while we overwrite pc->mem_cgroup.
3424 spin_lock_irqsave(&zone->lru_lock, flags);
3425 if (PageLRU(newpage))
3426 del_page_from_lru_list(zone, newpage, page_lru(newpage));
3427 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type);
3428 if (PageLRU(newpage))
3429 add_page_to_lru_list(zone, newpage, page_lru(newpage));
3430 spin_unlock_irqrestore(&zone->lru_lock, flags);
3433 #ifdef CONFIG_DEBUG_VM
3434 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3436 struct page_cgroup *pc;
3438 pc = lookup_page_cgroup(page);
3440 * Can be NULL while feeding pages into the page allocator for
3441 * the first time, i.e. during boot or memory hotplug;
3442 * or when mem_cgroup_disabled().
3444 if (likely(pc) && PageCgroupUsed(pc))
3449 bool mem_cgroup_bad_page_check(struct page *page)
3451 if (mem_cgroup_disabled())
3454 return lookup_page_cgroup_used(page) != NULL;
3457 void mem_cgroup_print_bad_page(struct page *page)
3459 struct page_cgroup *pc;
3461 pc = lookup_page_cgroup_used(page);
3466 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3467 pc, pc->flags, pc->mem_cgroup);
3469 path = kmalloc(PATH_MAX, GFP_KERNEL);
3472 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3477 printk(KERN_CONT "(%s)\n",
3478 (ret < 0) ? "cannot get the path" : path);
3484 static DEFINE_MUTEX(set_limit_mutex);
3486 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3487 unsigned long long val)
3490 u64 memswlimit, memlimit;
3492 int children = mem_cgroup_count_children(memcg);
3493 u64 curusage, oldusage;
3497 * For keeping hierarchical_reclaim simple, how long we should retry
3498 * is depends on callers. We set our retry-count to be function
3499 * of # of children which we should visit in this loop.
3501 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3503 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3506 while (retry_count) {
3507 if (signal_pending(current)) {
3512 * Rather than hide all in some function, I do this in
3513 * open coded manner. You see what this really does.
3514 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3516 mutex_lock(&set_limit_mutex);
3517 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3518 if (memswlimit < val) {
3520 mutex_unlock(&set_limit_mutex);
3524 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3528 ret = res_counter_set_limit(&memcg->res, val);
3530 if (memswlimit == val)
3531 memcg->memsw_is_minimum = true;
3533 memcg->memsw_is_minimum = false;
3535 mutex_unlock(&set_limit_mutex);
3540 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3541 MEM_CGROUP_RECLAIM_SHRINK);
3542 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3543 /* Usage is reduced ? */
3544 if (curusage >= oldusage)
3547 oldusage = curusage;
3549 if (!ret && enlarge)
3550 memcg_oom_recover(memcg);
3555 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3556 unsigned long long val)
3559 u64 memlimit, memswlimit, oldusage, curusage;
3560 int children = mem_cgroup_count_children(memcg);
3564 /* see mem_cgroup_resize_res_limit */
3565 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3566 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3567 while (retry_count) {
3568 if (signal_pending(current)) {
3573 * Rather than hide all in some function, I do this in
3574 * open coded manner. You see what this really does.
3575 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3577 mutex_lock(&set_limit_mutex);
3578 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3579 if (memlimit > val) {
3581 mutex_unlock(&set_limit_mutex);
3584 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3585 if (memswlimit < val)
3587 ret = res_counter_set_limit(&memcg->memsw, val);
3589 if (memlimit == val)
3590 memcg->memsw_is_minimum = true;
3592 memcg->memsw_is_minimum = false;
3594 mutex_unlock(&set_limit_mutex);
3599 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3600 MEM_CGROUP_RECLAIM_NOSWAP |
3601 MEM_CGROUP_RECLAIM_SHRINK);
3602 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3603 /* Usage is reduced ? */
3604 if (curusage >= oldusage)
3607 oldusage = curusage;
3609 if (!ret && enlarge)
3610 memcg_oom_recover(memcg);
3614 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3616 unsigned long *total_scanned)
3618 unsigned long nr_reclaimed = 0;
3619 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3620 unsigned long reclaimed;
3622 struct mem_cgroup_tree_per_zone *mctz;
3623 unsigned long long excess;
3624 unsigned long nr_scanned;
3629 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3631 * This loop can run a while, specially if mem_cgroup's continuously
3632 * keep exceeding their soft limit and putting the system under
3639 mz = mem_cgroup_largest_soft_limit_node(mctz);
3644 reclaimed = mem_cgroup_soft_reclaim(mz->mem, zone,
3645 gfp_mask, &nr_scanned);
3646 nr_reclaimed += reclaimed;
3647 *total_scanned += nr_scanned;
3648 spin_lock(&mctz->lock);
3651 * If we failed to reclaim anything from this memory cgroup
3652 * it is time to move on to the next cgroup
3658 * Loop until we find yet another one.
3660 * By the time we get the soft_limit lock
3661 * again, someone might have aded the
3662 * group back on the RB tree. Iterate to
3663 * make sure we get a different mem.
3664 * mem_cgroup_largest_soft_limit_node returns
3665 * NULL if no other cgroup is present on
3669 __mem_cgroup_largest_soft_limit_node(mctz);
3671 css_put(&next_mz->mem->css);
3672 else /* next_mz == NULL or other memcg */
3676 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3677 excess = res_counter_soft_limit_excess(&mz->mem->res);
3679 * One school of thought says that we should not add
3680 * back the node to the tree if reclaim returns 0.
3681 * But our reclaim could return 0, simply because due
3682 * to priority we are exposing a smaller subset of
3683 * memory to reclaim from. Consider this as a longer
3686 /* If excess == 0, no tree ops */
3687 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3688 spin_unlock(&mctz->lock);
3689 css_put(&mz->mem->css);
3692 * Could not reclaim anything and there are no more
3693 * mem cgroups to try or we seem to be looping without
3694 * reclaiming anything.
3696 if (!nr_reclaimed &&
3698 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3700 } while (!nr_reclaimed);
3702 css_put(&next_mz->mem->css);
3703 return nr_reclaimed;
3707 * This routine traverse page_cgroup in given list and drop them all.
3708 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3710 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3711 int node, int zid, enum lru_list lru)
3713 struct mem_cgroup_per_zone *mz;
3714 unsigned long flags, loop;
3715 struct list_head *list;
3720 zone = &NODE_DATA(node)->node_zones[zid];
3721 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3722 list = &mz->lruvec.lists[lru];
3724 loop = MEM_CGROUP_ZSTAT(mz, lru);
3725 /* give some margin against EBUSY etc...*/
3729 struct page_cgroup *pc;
3733 spin_lock_irqsave(&zone->lru_lock, flags);
3734 if (list_empty(list)) {
3735 spin_unlock_irqrestore(&zone->lru_lock, flags);
3738 page = list_entry(list->prev, struct page, lru);
3740 list_move(&page->lru, list);
3742 spin_unlock_irqrestore(&zone->lru_lock, flags);
3745 spin_unlock_irqrestore(&zone->lru_lock, flags);
3747 pc = lookup_page_cgroup(page);
3749 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3753 if (ret == -EBUSY || ret == -EINVAL) {
3754 /* found lock contention or "pc" is obsolete. */
3761 if (!ret && !list_empty(list))
3767 * make mem_cgroup's charge to be 0 if there is no task.
3768 * This enables deleting this mem_cgroup.
3770 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3773 int node, zid, shrink;
3774 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3775 struct cgroup *cgrp = memcg->css.cgroup;
3777 css_get(&memcg->css);
3780 /* should free all ? */
3786 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3789 if (signal_pending(current))
3791 /* This is for making all *used* pages to be on LRU. */
3792 lru_add_drain_all();
3793 drain_all_stock_sync(memcg);
3795 mem_cgroup_start_move(memcg);
3796 for_each_node_state(node, N_HIGH_MEMORY) {
3797 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3800 ret = mem_cgroup_force_empty_list(memcg,
3809 mem_cgroup_end_move(memcg);
3810 memcg_oom_recover(memcg);
3811 /* it seems parent cgroup doesn't have enough mem */
3815 /* "ret" should also be checked to ensure all lists are empty. */
3816 } while (memcg->res.usage > 0 || ret);
3818 css_put(&memcg->css);
3822 /* returns EBUSY if there is a task or if we come here twice. */
3823 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3827 /* we call try-to-free pages for make this cgroup empty */
3828 lru_add_drain_all();
3829 /* try to free all pages in this cgroup */
3831 while (nr_retries && memcg->res.usage > 0) {
3834 if (signal_pending(current)) {
3838 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3842 /* maybe some writeback is necessary */
3843 congestion_wait(BLK_RW_ASYNC, HZ/10);
3848 /* try move_account...there may be some *locked* pages. */
3852 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3854 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3858 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3860 return mem_cgroup_from_cont(cont)->use_hierarchy;
3863 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3867 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3868 struct cgroup *parent = cont->parent;
3869 struct mem_cgroup *parent_memcg = NULL;
3872 parent_memcg = mem_cgroup_from_cont(parent);
3876 * If parent's use_hierarchy is set, we can't make any modifications
3877 * in the child subtrees. If it is unset, then the change can
3878 * occur, provided the current cgroup has no children.
3880 * For the root cgroup, parent_mem is NULL, we allow value to be
3881 * set if there are no children.
3883 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3884 (val == 1 || val == 0)) {
3885 if (list_empty(&cont->children))
3886 memcg->use_hierarchy = val;
3897 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3898 enum mem_cgroup_stat_index idx)
3900 struct mem_cgroup *iter;
3903 /* Per-cpu values can be negative, use a signed accumulator */
3904 for_each_mem_cgroup_tree(iter, memcg)
3905 val += mem_cgroup_read_stat(iter, idx);
3907 if (val < 0) /* race ? */
3912 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3916 if (!mem_cgroup_is_root(memcg)) {
3918 return res_counter_read_u64(&memcg->res, RES_USAGE);
3920 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3923 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3924 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3927 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3929 return val << PAGE_SHIFT;
3932 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3934 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3938 type = MEMFILE_TYPE(cft->private);
3939 name = MEMFILE_ATTR(cft->private);
3942 if (name == RES_USAGE)
3943 val = mem_cgroup_usage(memcg, false);
3945 val = res_counter_read_u64(&memcg->res, name);
3948 if (name == RES_USAGE)
3949 val = mem_cgroup_usage(memcg, true);
3951 val = res_counter_read_u64(&memcg->memsw, name);
3960 * The user of this function is...
3963 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3966 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3968 unsigned long long val;
3971 type = MEMFILE_TYPE(cft->private);
3972 name = MEMFILE_ATTR(cft->private);
3975 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3979 /* This function does all necessary parse...reuse it */
3980 ret = res_counter_memparse_write_strategy(buffer, &val);
3984 ret = mem_cgroup_resize_limit(memcg, val);
3986 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3988 case RES_SOFT_LIMIT:
3989 ret = res_counter_memparse_write_strategy(buffer, &val);
3993 * For memsw, soft limits are hard to implement in terms
3994 * of semantics, for now, we support soft limits for
3995 * control without swap
3998 ret = res_counter_set_soft_limit(&memcg->res, val);
4003 ret = -EINVAL; /* should be BUG() ? */
4009 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4010 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4012 struct cgroup *cgroup;
4013 unsigned long long min_limit, min_memsw_limit, tmp;
4015 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4016 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4017 cgroup = memcg->css.cgroup;
4018 if (!memcg->use_hierarchy)
4021 while (cgroup->parent) {
4022 cgroup = cgroup->parent;
4023 memcg = mem_cgroup_from_cont(cgroup);
4024 if (!memcg->use_hierarchy)
4026 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4027 min_limit = min(min_limit, tmp);
4028 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4029 min_memsw_limit = min(min_memsw_limit, tmp);
4032 *mem_limit = min_limit;
4033 *memsw_limit = min_memsw_limit;
4037 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4039 struct mem_cgroup *memcg;
4042 memcg = mem_cgroup_from_cont(cont);
4043 type = MEMFILE_TYPE(event);
4044 name = MEMFILE_ATTR(event);
4048 res_counter_reset_max(&memcg->res);
4050 res_counter_reset_max(&memcg->memsw);
4054 res_counter_reset_failcnt(&memcg->res);
4056 res_counter_reset_failcnt(&memcg->memsw);
4063 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4066 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4070 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4071 struct cftype *cft, u64 val)
4073 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4075 if (val >= (1 << NR_MOVE_TYPE))
4078 * We check this value several times in both in can_attach() and
4079 * attach(), so we need cgroup lock to prevent this value from being
4083 memcg->move_charge_at_immigrate = val;
4089 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4090 struct cftype *cft, u64 val)
4097 /* For read statistics */
4115 struct mcs_total_stat {
4116 s64 stat[NR_MCS_STAT];
4122 } memcg_stat_strings[NR_MCS_STAT] = {
4123 {"cache", "total_cache"},
4124 {"rss", "total_rss"},
4125 {"mapped_file", "total_mapped_file"},
4126 {"pgpgin", "total_pgpgin"},
4127 {"pgpgout", "total_pgpgout"},
4128 {"swap", "total_swap"},
4129 {"pgfault", "total_pgfault"},
4130 {"pgmajfault", "total_pgmajfault"},
4131 {"inactive_anon", "total_inactive_anon"},
4132 {"active_anon", "total_active_anon"},
4133 {"inactive_file", "total_inactive_file"},
4134 {"active_file", "total_active_file"},
4135 {"unevictable", "total_unevictable"}
4140 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4145 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4146 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4147 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4148 s->stat[MCS_RSS] += val * PAGE_SIZE;
4149 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4150 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4151 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4152 s->stat[MCS_PGPGIN] += val;
4153 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4154 s->stat[MCS_PGPGOUT] += val;
4155 if (do_swap_account) {
4156 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4157 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4159 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4160 s->stat[MCS_PGFAULT] += val;
4161 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4162 s->stat[MCS_PGMAJFAULT] += val;
4165 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4166 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4167 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4168 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4169 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4170 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4171 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4172 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4173 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4174 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4178 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4180 struct mem_cgroup *iter;
4182 for_each_mem_cgroup_tree(iter, memcg)
4183 mem_cgroup_get_local_stat(iter, s);
4187 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4190 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4191 unsigned long node_nr;
4192 struct cgroup *cont = m->private;
4193 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4195 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4196 seq_printf(m, "total=%lu", total_nr);
4197 for_each_node_state(nid, N_HIGH_MEMORY) {
4198 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4199 seq_printf(m, " N%d=%lu", nid, node_nr);
4203 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4204 seq_printf(m, "file=%lu", file_nr);
4205 for_each_node_state(nid, N_HIGH_MEMORY) {
4206 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4208 seq_printf(m, " N%d=%lu", nid, node_nr);
4212 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4213 seq_printf(m, "anon=%lu", anon_nr);
4214 for_each_node_state(nid, N_HIGH_MEMORY) {
4215 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4217 seq_printf(m, " N%d=%lu", nid, node_nr);
4221 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4222 seq_printf(m, "unevictable=%lu", unevictable_nr);
4223 for_each_node_state(nid, N_HIGH_MEMORY) {
4224 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4225 BIT(LRU_UNEVICTABLE));
4226 seq_printf(m, " N%d=%lu", nid, node_nr);
4231 #endif /* CONFIG_NUMA */
4233 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4234 struct cgroup_map_cb *cb)
4236 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4237 struct mcs_total_stat mystat;
4240 memset(&mystat, 0, sizeof(mystat));
4241 mem_cgroup_get_local_stat(mem_cont, &mystat);
4244 for (i = 0; i < NR_MCS_STAT; i++) {
4245 if (i == MCS_SWAP && !do_swap_account)
4247 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4250 /* Hierarchical information */
4252 unsigned long long limit, memsw_limit;
4253 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4254 cb->fill(cb, "hierarchical_memory_limit", limit);
4255 if (do_swap_account)
4256 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4259 memset(&mystat, 0, sizeof(mystat));
4260 mem_cgroup_get_total_stat(mem_cont, &mystat);
4261 for (i = 0; i < NR_MCS_STAT; i++) {
4262 if (i == MCS_SWAP && !do_swap_account)
4264 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4267 #ifdef CONFIG_DEBUG_VM
4270 struct mem_cgroup_per_zone *mz;
4271 unsigned long recent_rotated[2] = {0, 0};
4272 unsigned long recent_scanned[2] = {0, 0};
4274 for_each_online_node(nid)
4275 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4276 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4278 recent_rotated[0] +=
4279 mz->reclaim_stat.recent_rotated[0];
4280 recent_rotated[1] +=
4281 mz->reclaim_stat.recent_rotated[1];
4282 recent_scanned[0] +=
4283 mz->reclaim_stat.recent_scanned[0];
4284 recent_scanned[1] +=
4285 mz->reclaim_stat.recent_scanned[1];
4287 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4288 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4289 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4290 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4297 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4299 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4301 return mem_cgroup_swappiness(memcg);
4304 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4307 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4308 struct mem_cgroup *parent;
4313 if (cgrp->parent == NULL)
4316 parent = mem_cgroup_from_cont(cgrp->parent);
4320 /* If under hierarchy, only empty-root can set this value */
4321 if ((parent->use_hierarchy) ||
4322 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4327 memcg->swappiness = val;
4334 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4336 struct mem_cgroup_threshold_ary *t;
4342 t = rcu_dereference(memcg->thresholds.primary);
4344 t = rcu_dereference(memcg->memsw_thresholds.primary);
4349 usage = mem_cgroup_usage(memcg, swap);
4352 * current_threshold points to threshold just below usage.
4353 * If it's not true, a threshold was crossed after last
4354 * call of __mem_cgroup_threshold().
4356 i = t->current_threshold;
4359 * Iterate backward over array of thresholds starting from
4360 * current_threshold and check if a threshold is crossed.
4361 * If none of thresholds below usage is crossed, we read
4362 * only one element of the array here.
4364 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4365 eventfd_signal(t->entries[i].eventfd, 1);
4367 /* i = current_threshold + 1 */
4371 * Iterate forward over array of thresholds starting from
4372 * current_threshold+1 and check if a threshold is crossed.
4373 * If none of thresholds above usage is crossed, we read
4374 * only one element of the array here.
4376 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4377 eventfd_signal(t->entries[i].eventfd, 1);
4379 /* Update current_threshold */
4380 t->current_threshold = i - 1;
4385 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4388 __mem_cgroup_threshold(memcg, false);
4389 if (do_swap_account)
4390 __mem_cgroup_threshold(memcg, true);
4392 memcg = parent_mem_cgroup(memcg);
4396 static int compare_thresholds(const void *a, const void *b)
4398 const struct mem_cgroup_threshold *_a = a;
4399 const struct mem_cgroup_threshold *_b = b;
4401 return _a->threshold - _b->threshold;
4404 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4406 struct mem_cgroup_eventfd_list *ev;
4408 list_for_each_entry(ev, &memcg->oom_notify, list)
4409 eventfd_signal(ev->eventfd, 1);
4413 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4415 struct mem_cgroup *iter;
4417 for_each_mem_cgroup_tree(iter, memcg)
4418 mem_cgroup_oom_notify_cb(iter);
4421 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4422 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4424 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4425 struct mem_cgroup_thresholds *thresholds;
4426 struct mem_cgroup_threshold_ary *new;
4427 int type = MEMFILE_TYPE(cft->private);
4428 u64 threshold, usage;
4431 ret = res_counter_memparse_write_strategy(args, &threshold);
4435 mutex_lock(&memcg->thresholds_lock);
4438 thresholds = &memcg->thresholds;
4439 else if (type == _MEMSWAP)
4440 thresholds = &memcg->memsw_thresholds;
4444 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4446 /* Check if a threshold crossed before adding a new one */
4447 if (thresholds->primary)
4448 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4450 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4452 /* Allocate memory for new array of thresholds */
4453 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4461 /* Copy thresholds (if any) to new array */
4462 if (thresholds->primary) {
4463 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4464 sizeof(struct mem_cgroup_threshold));
4467 /* Add new threshold */
4468 new->entries[size - 1].eventfd = eventfd;
4469 new->entries[size - 1].threshold = threshold;
4471 /* Sort thresholds. Registering of new threshold isn't time-critical */
4472 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4473 compare_thresholds, NULL);
4475 /* Find current threshold */
4476 new->current_threshold = -1;
4477 for (i = 0; i < size; i++) {
4478 if (new->entries[i].threshold < usage) {
4480 * new->current_threshold will not be used until
4481 * rcu_assign_pointer(), so it's safe to increment
4484 ++new->current_threshold;
4488 /* Free old spare buffer and save old primary buffer as spare */
4489 kfree(thresholds->spare);
4490 thresholds->spare = thresholds->primary;
4492 rcu_assign_pointer(thresholds->primary, new);
4494 /* To be sure that nobody uses thresholds */
4498 mutex_unlock(&memcg->thresholds_lock);
4503 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4504 struct cftype *cft, struct eventfd_ctx *eventfd)
4506 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4507 struct mem_cgroup_thresholds *thresholds;
4508 struct mem_cgroup_threshold_ary *new;
4509 int type = MEMFILE_TYPE(cft->private);
4513 mutex_lock(&memcg->thresholds_lock);
4515 thresholds = &memcg->thresholds;
4516 else if (type == _MEMSWAP)
4517 thresholds = &memcg->memsw_thresholds;
4522 * Something went wrong if we trying to unregister a threshold
4523 * if we don't have thresholds
4525 BUG_ON(!thresholds);
4527 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4529 /* Check if a threshold crossed before removing */
4530 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4532 /* Calculate new number of threshold */
4534 for (i = 0; i < thresholds->primary->size; i++) {
4535 if (thresholds->primary->entries[i].eventfd != eventfd)
4539 new = thresholds->spare;
4541 /* Set thresholds array to NULL if we don't have thresholds */
4550 /* Copy thresholds and find current threshold */
4551 new->current_threshold = -1;
4552 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4553 if (thresholds->primary->entries[i].eventfd == eventfd)
4556 new->entries[j] = thresholds->primary->entries[i];
4557 if (new->entries[j].threshold < usage) {
4559 * new->current_threshold will not be used
4560 * until rcu_assign_pointer(), so it's safe to increment
4563 ++new->current_threshold;
4569 /* Swap primary and spare array */
4570 thresholds->spare = thresholds->primary;
4571 rcu_assign_pointer(thresholds->primary, new);
4573 /* To be sure that nobody uses thresholds */
4576 mutex_unlock(&memcg->thresholds_lock);
4579 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4580 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4582 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4583 struct mem_cgroup_eventfd_list *event;
4584 int type = MEMFILE_TYPE(cft->private);
4586 BUG_ON(type != _OOM_TYPE);
4587 event = kmalloc(sizeof(*event), GFP_KERNEL);
4591 spin_lock(&memcg_oom_lock);
4593 event->eventfd = eventfd;
4594 list_add(&event->list, &memcg->oom_notify);
4596 /* already in OOM ? */
4597 if (atomic_read(&memcg->under_oom))
4598 eventfd_signal(eventfd, 1);
4599 spin_unlock(&memcg_oom_lock);
4604 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4605 struct cftype *cft, struct eventfd_ctx *eventfd)
4607 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4608 struct mem_cgroup_eventfd_list *ev, *tmp;
4609 int type = MEMFILE_TYPE(cft->private);
4611 BUG_ON(type != _OOM_TYPE);
4613 spin_lock(&memcg_oom_lock);
4615 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4616 if (ev->eventfd == eventfd) {
4617 list_del(&ev->list);
4622 spin_unlock(&memcg_oom_lock);
4625 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4626 struct cftype *cft, struct cgroup_map_cb *cb)
4628 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4630 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4632 if (atomic_read(&memcg->under_oom))
4633 cb->fill(cb, "under_oom", 1);
4635 cb->fill(cb, "under_oom", 0);
4639 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4640 struct cftype *cft, u64 val)
4642 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4643 struct mem_cgroup *parent;
4645 /* cannot set to root cgroup and only 0 and 1 are allowed */
4646 if (!cgrp->parent || !((val == 0) || (val == 1)))
4649 parent = mem_cgroup_from_cont(cgrp->parent);
4652 /* oom-kill-disable is a flag for subhierarchy. */
4653 if ((parent->use_hierarchy) ||
4654 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4658 memcg->oom_kill_disable = val;
4660 memcg_oom_recover(memcg);
4666 static const struct file_operations mem_control_numa_stat_file_operations = {
4668 .llseek = seq_lseek,
4669 .release = single_release,
4672 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4674 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4676 file->f_op = &mem_control_numa_stat_file_operations;
4677 return single_open(file, mem_control_numa_stat_show, cont);
4679 #endif /* CONFIG_NUMA */
4681 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4682 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4685 * Part of this would be better living in a separate allocation
4686 * function, leaving us with just the cgroup tree population work.
4687 * We, however, depend on state such as network's proto_list that
4688 * is only initialized after cgroup creation. I found the less
4689 * cumbersome way to deal with it to defer it all to populate time
4691 return mem_cgroup_sockets_init(cont, ss);
4694 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4695 struct cgroup *cont)
4697 mem_cgroup_sockets_destroy(cont, ss);
4700 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4705 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4706 struct cgroup *cont)
4711 static struct cftype mem_cgroup_files[] = {
4713 .name = "usage_in_bytes",
4714 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4715 .read_u64 = mem_cgroup_read,
4716 .register_event = mem_cgroup_usage_register_event,
4717 .unregister_event = mem_cgroup_usage_unregister_event,
4720 .name = "max_usage_in_bytes",
4721 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4722 .trigger = mem_cgroup_reset,
4723 .read_u64 = mem_cgroup_read,
4726 .name = "limit_in_bytes",
4727 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4728 .write_string = mem_cgroup_write,
4729 .read_u64 = mem_cgroup_read,
4732 .name = "soft_limit_in_bytes",
4733 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4734 .write_string = mem_cgroup_write,
4735 .read_u64 = mem_cgroup_read,
4739 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4740 .trigger = mem_cgroup_reset,
4741 .read_u64 = mem_cgroup_read,
4745 .read_map = mem_control_stat_show,
4748 .name = "force_empty",
4749 .trigger = mem_cgroup_force_empty_write,
4752 .name = "use_hierarchy",
4753 .write_u64 = mem_cgroup_hierarchy_write,
4754 .read_u64 = mem_cgroup_hierarchy_read,
4757 .name = "swappiness",
4758 .read_u64 = mem_cgroup_swappiness_read,
4759 .write_u64 = mem_cgroup_swappiness_write,
4762 .name = "move_charge_at_immigrate",
4763 .read_u64 = mem_cgroup_move_charge_read,
4764 .write_u64 = mem_cgroup_move_charge_write,
4767 .name = "oom_control",
4768 .read_map = mem_cgroup_oom_control_read,
4769 .write_u64 = mem_cgroup_oom_control_write,
4770 .register_event = mem_cgroup_oom_register_event,
4771 .unregister_event = mem_cgroup_oom_unregister_event,
4772 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4776 .name = "numa_stat",
4777 .open = mem_control_numa_stat_open,
4783 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4784 static struct cftype memsw_cgroup_files[] = {
4786 .name = "memsw.usage_in_bytes",
4787 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4788 .read_u64 = mem_cgroup_read,
4789 .register_event = mem_cgroup_usage_register_event,
4790 .unregister_event = mem_cgroup_usage_unregister_event,
4793 .name = "memsw.max_usage_in_bytes",
4794 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4795 .trigger = mem_cgroup_reset,
4796 .read_u64 = mem_cgroup_read,
4799 .name = "memsw.limit_in_bytes",
4800 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4801 .write_string = mem_cgroup_write,
4802 .read_u64 = mem_cgroup_read,
4805 .name = "memsw.failcnt",
4806 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4807 .trigger = mem_cgroup_reset,
4808 .read_u64 = mem_cgroup_read,
4812 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4814 if (!do_swap_account)
4816 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4817 ARRAY_SIZE(memsw_cgroup_files));
4820 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4826 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4828 struct mem_cgroup_per_node *pn;
4829 struct mem_cgroup_per_zone *mz;
4831 int zone, tmp = node;
4833 * This routine is called against possible nodes.
4834 * But it's BUG to call kmalloc() against offline node.
4836 * TODO: this routine can waste much memory for nodes which will
4837 * never be onlined. It's better to use memory hotplug callback
4840 if (!node_state(node, N_NORMAL_MEMORY))
4842 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4846 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4847 mz = &pn->zoneinfo[zone];
4849 INIT_LIST_HEAD(&mz->lruvec.lists[l]);
4850 mz->usage_in_excess = 0;
4851 mz->on_tree = false;
4854 memcg->info.nodeinfo[node] = pn;
4858 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4860 kfree(memcg->info.nodeinfo[node]);
4863 static struct mem_cgroup *mem_cgroup_alloc(void)
4865 struct mem_cgroup *mem;
4866 int size = sizeof(struct mem_cgroup);
4868 /* Can be very big if MAX_NUMNODES is very big */
4869 if (size < PAGE_SIZE)
4870 mem = kzalloc(size, GFP_KERNEL);
4872 mem = vzalloc(size);
4877 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4880 spin_lock_init(&mem->pcp_counter_lock);
4884 if (size < PAGE_SIZE)
4892 * At destroying mem_cgroup, references from swap_cgroup can remain.
4893 * (scanning all at force_empty is too costly...)
4895 * Instead of clearing all references at force_empty, we remember
4896 * the number of reference from swap_cgroup and free mem_cgroup when
4897 * it goes down to 0.
4899 * Removal of cgroup itself succeeds regardless of refs from swap.
4902 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4906 mem_cgroup_remove_from_trees(memcg);
4907 free_css_id(&mem_cgroup_subsys, &memcg->css);
4909 for_each_node_state(node, N_POSSIBLE)
4910 free_mem_cgroup_per_zone_info(memcg, node);
4912 free_percpu(memcg->stat);
4913 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4919 static void mem_cgroup_get(struct mem_cgroup *memcg)
4921 atomic_inc(&memcg->refcnt);
4924 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4926 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4927 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4928 __mem_cgroup_free(memcg);
4930 mem_cgroup_put(parent);
4934 static void mem_cgroup_put(struct mem_cgroup *memcg)
4936 __mem_cgroup_put(memcg, 1);
4940 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4942 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4944 if (!memcg->res.parent)
4946 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4948 EXPORT_SYMBOL(parent_mem_cgroup);
4950 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4951 static void __init enable_swap_cgroup(void)
4953 if (!mem_cgroup_disabled() && really_do_swap_account)
4954 do_swap_account = 1;
4957 static void __init enable_swap_cgroup(void)
4962 static int mem_cgroup_soft_limit_tree_init(void)
4964 struct mem_cgroup_tree_per_node *rtpn;
4965 struct mem_cgroup_tree_per_zone *rtpz;
4966 int tmp, node, zone;
4968 for_each_node_state(node, N_POSSIBLE) {
4970 if (!node_state(node, N_NORMAL_MEMORY))
4972 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4976 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4978 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4979 rtpz = &rtpn->rb_tree_per_zone[zone];
4980 rtpz->rb_root = RB_ROOT;
4981 spin_lock_init(&rtpz->lock);
4987 for_each_node_state(node, N_POSSIBLE) {
4988 if (!soft_limit_tree.rb_tree_per_node[node])
4990 kfree(soft_limit_tree.rb_tree_per_node[node]);
4991 soft_limit_tree.rb_tree_per_node[node] = NULL;
4997 static struct cgroup_subsys_state * __ref
4998 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5000 struct mem_cgroup *memcg, *parent;
5001 long error = -ENOMEM;
5004 memcg = mem_cgroup_alloc();
5006 return ERR_PTR(error);
5008 for_each_node_state(node, N_POSSIBLE)
5009 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5013 if (cont->parent == NULL) {
5015 enable_swap_cgroup();
5017 if (mem_cgroup_soft_limit_tree_init())
5019 root_mem_cgroup = memcg;
5020 for_each_possible_cpu(cpu) {
5021 struct memcg_stock_pcp *stock =
5022 &per_cpu(memcg_stock, cpu);
5023 INIT_WORK(&stock->work, drain_local_stock);
5025 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5027 parent = mem_cgroup_from_cont(cont->parent);
5028 memcg->use_hierarchy = parent->use_hierarchy;
5029 memcg->oom_kill_disable = parent->oom_kill_disable;
5032 if (parent && parent->use_hierarchy) {
5033 res_counter_init(&memcg->res, &parent->res);
5034 res_counter_init(&memcg->memsw, &parent->memsw);
5036 * We increment refcnt of the parent to ensure that we can
5037 * safely access it on res_counter_charge/uncharge.
5038 * This refcnt will be decremented when freeing this
5039 * mem_cgroup(see mem_cgroup_put).
5041 mem_cgroup_get(parent);
5043 res_counter_init(&memcg->res, NULL);
5044 res_counter_init(&memcg->memsw, NULL);
5046 memcg->last_scanned_node = MAX_NUMNODES;
5047 INIT_LIST_HEAD(&memcg->oom_notify);
5050 memcg->swappiness = mem_cgroup_swappiness(parent);
5051 atomic_set(&memcg->refcnt, 1);
5052 memcg->move_charge_at_immigrate = 0;
5053 mutex_init(&memcg->thresholds_lock);
5056 __mem_cgroup_free(memcg);
5057 return ERR_PTR(error);
5060 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5061 struct cgroup *cont)
5063 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5065 return mem_cgroup_force_empty(memcg, false);
5068 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5069 struct cgroup *cont)
5071 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5073 kmem_cgroup_destroy(ss, cont);
5075 mem_cgroup_put(memcg);
5078 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5079 struct cgroup *cont)
5083 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5084 ARRAY_SIZE(mem_cgroup_files));
5087 ret = register_memsw_files(cont, ss);
5090 ret = register_kmem_files(cont, ss);
5096 /* Handlers for move charge at task migration. */
5097 #define PRECHARGE_COUNT_AT_ONCE 256
5098 static int mem_cgroup_do_precharge(unsigned long count)
5101 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5102 struct mem_cgroup *memcg = mc.to;
5104 if (mem_cgroup_is_root(memcg)) {
5105 mc.precharge += count;
5106 /* we don't need css_get for root */
5109 /* try to charge at once */
5111 struct res_counter *dummy;
5113 * "memcg" cannot be under rmdir() because we've already checked
5114 * by cgroup_lock_live_cgroup() that it is not removed and we
5115 * are still under the same cgroup_mutex. So we can postpone
5118 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5120 if (do_swap_account && res_counter_charge(&memcg->memsw,
5121 PAGE_SIZE * count, &dummy)) {
5122 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5125 mc.precharge += count;
5129 /* fall back to one by one charge */
5131 if (signal_pending(current)) {
5135 if (!batch_count--) {
5136 batch_count = PRECHARGE_COUNT_AT_ONCE;
5139 ret = __mem_cgroup_try_charge(NULL,
5140 GFP_KERNEL, 1, &memcg, false);
5142 /* mem_cgroup_clear_mc() will do uncharge later */
5150 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5151 * @vma: the vma the pte to be checked belongs
5152 * @addr: the address corresponding to the pte to be checked
5153 * @ptent: the pte to be checked
5154 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5157 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5158 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5159 * move charge. if @target is not NULL, the page is stored in target->page
5160 * with extra refcnt got(Callers should handle it).
5161 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5162 * target for charge migration. if @target is not NULL, the entry is stored
5165 * Called with pte lock held.
5172 enum mc_target_type {
5173 MC_TARGET_NONE, /* not used */
5178 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5179 unsigned long addr, pte_t ptent)
5181 struct page *page = vm_normal_page(vma, addr, ptent);
5183 if (!page || !page_mapped(page))
5185 if (PageAnon(page)) {
5186 /* we don't move shared anon */
5187 if (!move_anon() || page_mapcount(page) > 2)
5189 } else if (!move_file())
5190 /* we ignore mapcount for file pages */
5192 if (!get_page_unless_zero(page))
5198 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5199 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5202 struct page *page = NULL;
5203 swp_entry_t ent = pte_to_swp_entry(ptent);
5205 if (!move_anon() || non_swap_entry(ent))
5207 usage_count = mem_cgroup_count_swap_user(ent, &page);
5208 if (usage_count > 1) { /* we don't move shared anon */
5213 if (do_swap_account)
5214 entry->val = ent.val;
5219 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5220 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5222 struct page *page = NULL;
5223 struct inode *inode;
5224 struct address_space *mapping;
5227 if (!vma->vm_file) /* anonymous vma */
5232 inode = vma->vm_file->f_path.dentry->d_inode;
5233 mapping = vma->vm_file->f_mapping;
5234 if (pte_none(ptent))
5235 pgoff = linear_page_index(vma, addr);
5236 else /* pte_file(ptent) is true */
5237 pgoff = pte_to_pgoff(ptent);
5239 /* page is moved even if it's not RSS of this task(page-faulted). */
5240 page = find_get_page(mapping, pgoff);
5243 /* shmem/tmpfs may report page out on swap: account for that too. */
5244 if (radix_tree_exceptional_entry(page)) {
5245 swp_entry_t swap = radix_to_swp_entry(page);
5246 if (do_swap_account)
5248 page = find_get_page(&swapper_space, swap.val);
5254 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5255 unsigned long addr, pte_t ptent, union mc_target *target)
5257 struct page *page = NULL;
5258 struct page_cgroup *pc;
5260 swp_entry_t ent = { .val = 0 };
5262 if (pte_present(ptent))
5263 page = mc_handle_present_pte(vma, addr, ptent);
5264 else if (is_swap_pte(ptent))
5265 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5266 else if (pte_none(ptent) || pte_file(ptent))
5267 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5269 if (!page && !ent.val)
5272 pc = lookup_page_cgroup(page);
5274 * Do only loose check w/o page_cgroup lock.
5275 * mem_cgroup_move_account() checks the pc is valid or not under
5278 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5279 ret = MC_TARGET_PAGE;
5281 target->page = page;
5283 if (!ret || !target)
5286 /* There is a swap entry and a page doesn't exist or isn't charged */
5287 if (ent.val && !ret &&
5288 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5289 ret = MC_TARGET_SWAP;
5296 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5297 unsigned long addr, unsigned long end,
5298 struct mm_walk *walk)
5300 struct vm_area_struct *vma = walk->private;
5304 split_huge_page_pmd(walk->mm, pmd);
5306 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5307 for (; addr != end; pte++, addr += PAGE_SIZE)
5308 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5309 mc.precharge++; /* increment precharge temporarily */
5310 pte_unmap_unlock(pte - 1, ptl);
5316 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5318 unsigned long precharge;
5319 struct vm_area_struct *vma;
5321 down_read(&mm->mmap_sem);
5322 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5323 struct mm_walk mem_cgroup_count_precharge_walk = {
5324 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5328 if (is_vm_hugetlb_page(vma))
5330 walk_page_range(vma->vm_start, vma->vm_end,
5331 &mem_cgroup_count_precharge_walk);
5333 up_read(&mm->mmap_sem);
5335 precharge = mc.precharge;
5341 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5343 unsigned long precharge = mem_cgroup_count_precharge(mm);
5345 VM_BUG_ON(mc.moving_task);
5346 mc.moving_task = current;
5347 return mem_cgroup_do_precharge(precharge);
5350 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5351 static void __mem_cgroup_clear_mc(void)
5353 struct mem_cgroup *from = mc.from;
5354 struct mem_cgroup *to = mc.to;
5356 /* we must uncharge all the leftover precharges from mc.to */
5358 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5362 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5363 * we must uncharge here.
5365 if (mc.moved_charge) {
5366 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5367 mc.moved_charge = 0;
5369 /* we must fixup refcnts and charges */
5370 if (mc.moved_swap) {
5371 /* uncharge swap account from the old cgroup */
5372 if (!mem_cgroup_is_root(mc.from))
5373 res_counter_uncharge(&mc.from->memsw,
5374 PAGE_SIZE * mc.moved_swap);
5375 __mem_cgroup_put(mc.from, mc.moved_swap);
5377 if (!mem_cgroup_is_root(mc.to)) {
5379 * we charged both to->res and to->memsw, so we should
5382 res_counter_uncharge(&mc.to->res,
5383 PAGE_SIZE * mc.moved_swap);
5385 /* we've already done mem_cgroup_get(mc.to) */
5388 memcg_oom_recover(from);
5389 memcg_oom_recover(to);
5390 wake_up_all(&mc.waitq);
5393 static void mem_cgroup_clear_mc(void)
5395 struct mem_cgroup *from = mc.from;
5398 * we must clear moving_task before waking up waiters at the end of
5401 mc.moving_task = NULL;
5402 __mem_cgroup_clear_mc();
5403 spin_lock(&mc.lock);
5406 spin_unlock(&mc.lock);
5407 mem_cgroup_end_move(from);
5410 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5411 struct cgroup *cgroup,
5412 struct cgroup_taskset *tset)
5414 struct task_struct *p = cgroup_taskset_first(tset);
5416 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5418 if (memcg->move_charge_at_immigrate) {
5419 struct mm_struct *mm;
5420 struct mem_cgroup *from = mem_cgroup_from_task(p);
5422 VM_BUG_ON(from == memcg);
5424 mm = get_task_mm(p);
5427 /* We move charges only when we move a owner of the mm */
5428 if (mm->owner == p) {
5431 VM_BUG_ON(mc.precharge);
5432 VM_BUG_ON(mc.moved_charge);
5433 VM_BUG_ON(mc.moved_swap);
5434 mem_cgroup_start_move(from);
5435 spin_lock(&mc.lock);
5438 spin_unlock(&mc.lock);
5439 /* We set mc.moving_task later */
5441 ret = mem_cgroup_precharge_mc(mm);
5443 mem_cgroup_clear_mc();
5450 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5451 struct cgroup *cgroup,
5452 struct cgroup_taskset *tset)
5454 mem_cgroup_clear_mc();
5457 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5458 unsigned long addr, unsigned long end,
5459 struct mm_walk *walk)
5462 struct vm_area_struct *vma = walk->private;
5466 split_huge_page_pmd(walk->mm, pmd);
5468 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5469 for (; addr != end; addr += PAGE_SIZE) {
5470 pte_t ptent = *(pte++);
5471 union mc_target target;
5474 struct page_cgroup *pc;
5480 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5482 case MC_TARGET_PAGE:
5484 if (isolate_lru_page(page))
5486 pc = lookup_page_cgroup(page);
5487 if (!mem_cgroup_move_account(page, 1, pc,
5488 mc.from, mc.to, false)) {
5490 /* we uncharge from mc.from later. */
5493 putback_lru_page(page);
5494 put: /* is_target_pte_for_mc() gets the page */
5497 case MC_TARGET_SWAP:
5499 if (!mem_cgroup_move_swap_account(ent,
5500 mc.from, mc.to, false)) {
5502 /* we fixup refcnts and charges later. */
5510 pte_unmap_unlock(pte - 1, ptl);
5515 * We have consumed all precharges we got in can_attach().
5516 * We try charge one by one, but don't do any additional
5517 * charges to mc.to if we have failed in charge once in attach()
5520 ret = mem_cgroup_do_precharge(1);
5528 static void mem_cgroup_move_charge(struct mm_struct *mm)
5530 struct vm_area_struct *vma;
5532 lru_add_drain_all();
5534 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5536 * Someone who are holding the mmap_sem might be waiting in
5537 * waitq. So we cancel all extra charges, wake up all waiters,
5538 * and retry. Because we cancel precharges, we might not be able
5539 * to move enough charges, but moving charge is a best-effort
5540 * feature anyway, so it wouldn't be a big problem.
5542 __mem_cgroup_clear_mc();
5546 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5548 struct mm_walk mem_cgroup_move_charge_walk = {
5549 .pmd_entry = mem_cgroup_move_charge_pte_range,
5553 if (is_vm_hugetlb_page(vma))
5555 ret = walk_page_range(vma->vm_start, vma->vm_end,
5556 &mem_cgroup_move_charge_walk);
5559 * means we have consumed all precharges and failed in
5560 * doing additional charge. Just abandon here.
5564 up_read(&mm->mmap_sem);
5567 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5568 struct cgroup *cont,
5569 struct cgroup_taskset *tset)
5571 struct task_struct *p = cgroup_taskset_first(tset);
5572 struct mm_struct *mm = get_task_mm(p);
5576 mem_cgroup_move_charge(mm);
5581 mem_cgroup_clear_mc();
5583 #else /* !CONFIG_MMU */
5584 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5585 struct cgroup *cgroup,
5586 struct cgroup_taskset *tset)
5590 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5591 struct cgroup *cgroup,
5592 struct cgroup_taskset *tset)
5595 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5596 struct cgroup *cont,
5597 struct cgroup_taskset *tset)
5602 struct cgroup_subsys mem_cgroup_subsys = {
5604 .subsys_id = mem_cgroup_subsys_id,
5605 .create = mem_cgroup_create,
5606 .pre_destroy = mem_cgroup_pre_destroy,
5607 .destroy = mem_cgroup_destroy,
5608 .populate = mem_cgroup_populate,
5609 .can_attach = mem_cgroup_can_attach,
5610 .cancel_attach = mem_cgroup_cancel_attach,
5611 .attach = mem_cgroup_move_task,
5616 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5617 static int __init enable_swap_account(char *s)
5619 /* consider enabled if no parameter or 1 is given */
5620 if (!strcmp(s, "1"))
5621 really_do_swap_account = 1;
5622 else if (!strcmp(s, "0"))
5623 really_do_swap_account = 0;
5626 __setup("swapaccount=", enable_swap_account);