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 * Checks whether given mem is same or in the root_mem_cgroup's
1143 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1144 struct mem_cgroup *memcg)
1146 if (root_memcg != memcg) {
1147 return (root_memcg->use_hierarchy &&
1148 css_is_ancestor(&memcg->css, &root_memcg->css));
1154 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1157 struct mem_cgroup *curr = NULL;
1158 struct task_struct *p;
1160 p = find_lock_task_mm(task);
1162 curr = try_get_mem_cgroup_from_mm(p->mm);
1166 * All threads may have already detached their mm's, but the oom
1167 * killer still needs to detect if they have already been oom
1168 * killed to prevent needlessly killing additional tasks.
1171 curr = mem_cgroup_from_task(task);
1173 css_get(&curr->css);
1179 * We should check use_hierarchy of "memcg" not "curr". Because checking
1180 * use_hierarchy of "curr" here make this function true if hierarchy is
1181 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1182 * hierarchy(even if use_hierarchy is disabled in "memcg").
1184 ret = mem_cgroup_same_or_subtree(memcg, curr);
1185 css_put(&curr->css);
1189 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1191 unsigned long inactive_ratio;
1192 int nid = zone_to_nid(zone);
1193 int zid = zone_idx(zone);
1194 unsigned long inactive;
1195 unsigned long active;
1198 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1199 BIT(LRU_INACTIVE_ANON));
1200 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1201 BIT(LRU_ACTIVE_ANON));
1203 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1205 inactive_ratio = int_sqrt(10 * gb);
1209 return inactive * inactive_ratio < active;
1212 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1214 unsigned long active;
1215 unsigned long inactive;
1216 int zid = zone_idx(zone);
1217 int nid = zone_to_nid(zone);
1219 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1220 BIT(LRU_INACTIVE_FILE));
1221 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1222 BIT(LRU_ACTIVE_FILE));
1224 return (active > inactive);
1227 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1230 int nid = zone_to_nid(zone);
1231 int zid = zone_idx(zone);
1232 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1234 return &mz->reclaim_stat;
1237 struct zone_reclaim_stat *
1238 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1240 struct page_cgroup *pc;
1241 struct mem_cgroup_per_zone *mz;
1243 if (mem_cgroup_disabled())
1246 pc = lookup_page_cgroup(page);
1247 if (!PageCgroupUsed(pc))
1249 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1251 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1252 return &mz->reclaim_stat;
1255 #define mem_cgroup_from_res_counter(counter, member) \
1256 container_of(counter, struct mem_cgroup, member)
1259 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1260 * @mem: the memory cgroup
1262 * Returns the maximum amount of memory @mem can be charged with, in
1265 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1267 unsigned long long margin;
1269 margin = res_counter_margin(&memcg->res);
1270 if (do_swap_account)
1271 margin = min(margin, res_counter_margin(&memcg->memsw));
1272 return margin >> PAGE_SHIFT;
1275 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1277 struct cgroup *cgrp = memcg->css.cgroup;
1280 if (cgrp->parent == NULL)
1281 return vm_swappiness;
1283 return memcg->swappiness;
1286 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1291 spin_lock(&memcg->pcp_counter_lock);
1292 for_each_online_cpu(cpu)
1293 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1294 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1295 spin_unlock(&memcg->pcp_counter_lock);
1301 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1308 spin_lock(&memcg->pcp_counter_lock);
1309 for_each_online_cpu(cpu)
1310 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1311 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1312 spin_unlock(&memcg->pcp_counter_lock);
1316 * 2 routines for checking "mem" is under move_account() or not.
1318 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1319 * for avoiding race in accounting. If true,
1320 * pc->mem_cgroup may be overwritten.
1322 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1323 * under hierarchy of moving cgroups. This is for
1324 * waiting at hith-memory prressure caused by "move".
1327 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1329 VM_BUG_ON(!rcu_read_lock_held());
1330 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1333 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1335 struct mem_cgroup *from;
1336 struct mem_cgroup *to;
1339 * Unlike task_move routines, we access mc.to, mc.from not under
1340 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1342 spin_lock(&mc.lock);
1348 ret = mem_cgroup_same_or_subtree(memcg, from)
1349 || mem_cgroup_same_or_subtree(memcg, to);
1351 spin_unlock(&mc.lock);
1355 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1357 if (mc.moving_task && current != mc.moving_task) {
1358 if (mem_cgroup_under_move(memcg)) {
1360 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1361 /* moving charge context might have finished. */
1364 finish_wait(&mc.waitq, &wait);
1372 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1373 * @memcg: The memory cgroup that went over limit
1374 * @p: Task that is going to be killed
1376 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1379 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1381 struct cgroup *task_cgrp;
1382 struct cgroup *mem_cgrp;
1384 * Need a buffer in BSS, can't rely on allocations. The code relies
1385 * on the assumption that OOM is serialized for memory controller.
1386 * If this assumption is broken, revisit this code.
1388 static char memcg_name[PATH_MAX];
1397 mem_cgrp = memcg->css.cgroup;
1398 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1400 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1403 * Unfortunately, we are unable to convert to a useful name
1404 * But we'll still print out the usage information
1411 printk(KERN_INFO "Task in %s killed", memcg_name);
1414 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1422 * Continues from above, so we don't need an KERN_ level
1424 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1427 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1428 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1429 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1430 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1431 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1433 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1434 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1435 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1439 * This function returns the number of memcg under hierarchy tree. Returns
1440 * 1(self count) if no children.
1442 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1445 struct mem_cgroup *iter;
1447 for_each_mem_cgroup_tree(iter, memcg)
1453 * Return the memory (and swap, if configured) limit for a memcg.
1455 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1460 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1461 limit += total_swap_pages << PAGE_SHIFT;
1463 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1465 * If memsw is finite and limits the amount of swap space available
1466 * to this memcg, return that limit.
1468 return min(limit, memsw);
1471 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1473 unsigned long flags)
1475 unsigned long total = 0;
1476 bool noswap = false;
1479 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1481 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1484 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1486 drain_all_stock_async(memcg);
1487 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1489 * Allow limit shrinkers, which are triggered directly
1490 * by userspace, to catch signals and stop reclaim
1491 * after minimal progress, regardless of the margin.
1493 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1495 if (mem_cgroup_margin(memcg))
1498 * If nothing was reclaimed after two attempts, there
1499 * may be no reclaimable pages in this hierarchy.
1508 * test_mem_cgroup_node_reclaimable
1509 * @mem: the target memcg
1510 * @nid: the node ID to be checked.
1511 * @noswap : specify true here if the user wants flle only information.
1513 * This function returns whether the specified memcg contains any
1514 * reclaimable pages on a node. Returns true if there are any reclaimable
1515 * pages in the node.
1517 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1518 int nid, bool noswap)
1520 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1522 if (noswap || !total_swap_pages)
1524 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1529 #if MAX_NUMNODES > 1
1532 * Always updating the nodemask is not very good - even if we have an empty
1533 * list or the wrong list here, we can start from some node and traverse all
1534 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1537 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1541 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1542 * pagein/pageout changes since the last update.
1544 if (!atomic_read(&memcg->numainfo_events))
1546 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1549 /* make a nodemask where this memcg uses memory from */
1550 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1552 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1554 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1555 node_clear(nid, memcg->scan_nodes);
1558 atomic_set(&memcg->numainfo_events, 0);
1559 atomic_set(&memcg->numainfo_updating, 0);
1563 * Selecting a node where we start reclaim from. Because what we need is just
1564 * reducing usage counter, start from anywhere is O,K. Considering
1565 * memory reclaim from current node, there are pros. and cons.
1567 * Freeing memory from current node means freeing memory from a node which
1568 * we'll use or we've used. So, it may make LRU bad. And if several threads
1569 * hit limits, it will see a contention on a node. But freeing from remote
1570 * node means more costs for memory reclaim because of memory latency.
1572 * Now, we use round-robin. Better algorithm is welcomed.
1574 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1578 mem_cgroup_may_update_nodemask(memcg);
1579 node = memcg->last_scanned_node;
1581 node = next_node(node, memcg->scan_nodes);
1582 if (node == MAX_NUMNODES)
1583 node = first_node(memcg->scan_nodes);
1585 * We call this when we hit limit, not when pages are added to LRU.
1586 * No LRU may hold pages because all pages are UNEVICTABLE or
1587 * memcg is too small and all pages are not on LRU. In that case,
1588 * we use curret node.
1590 if (unlikely(node == MAX_NUMNODES))
1591 node = numa_node_id();
1593 memcg->last_scanned_node = node;
1598 * Check all nodes whether it contains reclaimable pages or not.
1599 * For quick scan, we make use of scan_nodes. This will allow us to skip
1600 * unused nodes. But scan_nodes is lazily updated and may not cotain
1601 * enough new information. We need to do double check.
1603 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1608 * quick check...making use of scan_node.
1609 * We can skip unused nodes.
1611 if (!nodes_empty(memcg->scan_nodes)) {
1612 for (nid = first_node(memcg->scan_nodes);
1614 nid = next_node(nid, memcg->scan_nodes)) {
1616 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1621 * Check rest of nodes.
1623 for_each_node_state(nid, N_HIGH_MEMORY) {
1624 if (node_isset(nid, memcg->scan_nodes))
1626 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1633 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1638 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1640 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1644 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1647 unsigned long *total_scanned)
1649 struct mem_cgroup *victim = NULL;
1652 unsigned long excess;
1653 unsigned long nr_scanned;
1654 struct mem_cgroup_reclaim_cookie reclaim = {
1659 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1662 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1667 * If we have not been able to reclaim
1668 * anything, it might because there are
1669 * no reclaimable pages under this hierarchy
1674 * We want to do more targeted reclaim.
1675 * excess >> 2 is not to excessive so as to
1676 * reclaim too much, nor too less that we keep
1677 * coming back to reclaim from this cgroup
1679 if (total >= (excess >> 2) ||
1680 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1685 if (!mem_cgroup_reclaimable(victim, false))
1687 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1689 *total_scanned += nr_scanned;
1690 if (!res_counter_soft_limit_excess(&root_memcg->res))
1693 mem_cgroup_iter_break(root_memcg, victim);
1698 * Check OOM-Killer is already running under our hierarchy.
1699 * If someone is running, return false.
1700 * Has to be called with memcg_oom_lock
1702 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1704 struct mem_cgroup *iter, *failed = NULL;
1706 for_each_mem_cgroup_tree(iter, memcg) {
1707 if (iter->oom_lock) {
1709 * this subtree of our hierarchy is already locked
1710 * so we cannot give a lock.
1713 mem_cgroup_iter_break(memcg, iter);
1716 iter->oom_lock = true;
1723 * OK, we failed to lock the whole subtree so we have to clean up
1724 * what we set up to the failing subtree
1726 for_each_mem_cgroup_tree(iter, memcg) {
1727 if (iter == failed) {
1728 mem_cgroup_iter_break(memcg, iter);
1731 iter->oom_lock = false;
1737 * Has to be called with memcg_oom_lock
1739 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1741 struct mem_cgroup *iter;
1743 for_each_mem_cgroup_tree(iter, memcg)
1744 iter->oom_lock = false;
1748 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1750 struct mem_cgroup *iter;
1752 for_each_mem_cgroup_tree(iter, memcg)
1753 atomic_inc(&iter->under_oom);
1756 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1758 struct mem_cgroup *iter;
1761 * When a new child is created while the hierarchy is under oom,
1762 * mem_cgroup_oom_lock() may not be called. We have to use
1763 * atomic_add_unless() here.
1765 for_each_mem_cgroup_tree(iter, memcg)
1766 atomic_add_unless(&iter->under_oom, -1, 0);
1769 static DEFINE_SPINLOCK(memcg_oom_lock);
1770 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1772 struct oom_wait_info {
1773 struct mem_cgroup *mem;
1777 static int memcg_oom_wake_function(wait_queue_t *wait,
1778 unsigned mode, int sync, void *arg)
1780 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1782 struct oom_wait_info *oom_wait_info;
1784 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1785 oom_wait_memcg = oom_wait_info->mem;
1788 * Both of oom_wait_info->mem and wake_mem are stable under us.
1789 * Then we can use css_is_ancestor without taking care of RCU.
1791 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1792 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1794 return autoremove_wake_function(wait, mode, sync, arg);
1797 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1799 /* for filtering, pass "memcg" as argument. */
1800 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1803 static void memcg_oom_recover(struct mem_cgroup *memcg)
1805 if (memcg && atomic_read(&memcg->under_oom))
1806 memcg_wakeup_oom(memcg);
1810 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1812 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1814 struct oom_wait_info owait;
1815 bool locked, need_to_kill;
1818 owait.wait.flags = 0;
1819 owait.wait.func = memcg_oom_wake_function;
1820 owait.wait.private = current;
1821 INIT_LIST_HEAD(&owait.wait.task_list);
1822 need_to_kill = true;
1823 mem_cgroup_mark_under_oom(memcg);
1825 /* At first, try to OOM lock hierarchy under memcg.*/
1826 spin_lock(&memcg_oom_lock);
1827 locked = mem_cgroup_oom_lock(memcg);
1829 * Even if signal_pending(), we can't quit charge() loop without
1830 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1831 * under OOM is always welcomed, use TASK_KILLABLE here.
1833 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1834 if (!locked || memcg->oom_kill_disable)
1835 need_to_kill = false;
1837 mem_cgroup_oom_notify(memcg);
1838 spin_unlock(&memcg_oom_lock);
1841 finish_wait(&memcg_oom_waitq, &owait.wait);
1842 mem_cgroup_out_of_memory(memcg, mask);
1845 finish_wait(&memcg_oom_waitq, &owait.wait);
1847 spin_lock(&memcg_oom_lock);
1849 mem_cgroup_oom_unlock(memcg);
1850 memcg_wakeup_oom(memcg);
1851 spin_unlock(&memcg_oom_lock);
1853 mem_cgroup_unmark_under_oom(memcg);
1855 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1857 /* Give chance to dying process */
1858 schedule_timeout_uninterruptible(1);
1863 * Currently used to update mapped file statistics, but the routine can be
1864 * generalized to update other statistics as well.
1866 * Notes: Race condition
1868 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1869 * it tends to be costly. But considering some conditions, we doesn't need
1870 * to do so _always_.
1872 * Considering "charge", lock_page_cgroup() is not required because all
1873 * file-stat operations happen after a page is attached to radix-tree. There
1874 * are no race with "charge".
1876 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1877 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1878 * if there are race with "uncharge". Statistics itself is properly handled
1881 * Considering "move", this is an only case we see a race. To make the race
1882 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1883 * possibility of race condition. If there is, we take a lock.
1886 void mem_cgroup_update_page_stat(struct page *page,
1887 enum mem_cgroup_page_stat_item idx, int val)
1889 struct mem_cgroup *memcg;
1890 struct page_cgroup *pc = lookup_page_cgroup(page);
1891 bool need_unlock = false;
1892 unsigned long uninitialized_var(flags);
1894 if (mem_cgroup_disabled())
1898 memcg = pc->mem_cgroup;
1899 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1901 /* pc->mem_cgroup is unstable ? */
1902 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
1903 /* take a lock against to access pc->mem_cgroup */
1904 move_lock_page_cgroup(pc, &flags);
1906 memcg = pc->mem_cgroup;
1907 if (!memcg || !PageCgroupUsed(pc))
1912 case MEMCG_NR_FILE_MAPPED:
1914 SetPageCgroupFileMapped(pc);
1915 else if (!page_mapped(page))
1916 ClearPageCgroupFileMapped(pc);
1917 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1923 this_cpu_add(memcg->stat->count[idx], val);
1926 if (unlikely(need_unlock))
1927 move_unlock_page_cgroup(pc, &flags);
1931 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1934 * size of first charge trial. "32" comes from vmscan.c's magic value.
1935 * TODO: maybe necessary to use big numbers in big irons.
1937 #define CHARGE_BATCH 32U
1938 struct memcg_stock_pcp {
1939 struct mem_cgroup *cached; /* this never be root cgroup */
1940 unsigned int nr_pages;
1941 struct work_struct work;
1942 unsigned long flags;
1943 #define FLUSHING_CACHED_CHARGE (0)
1945 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1946 static DEFINE_MUTEX(percpu_charge_mutex);
1949 * Try to consume stocked charge on this cpu. If success, one page is consumed
1950 * from local stock and true is returned. If the stock is 0 or charges from a
1951 * cgroup which is not current target, returns false. This stock will be
1954 static bool consume_stock(struct mem_cgroup *memcg)
1956 struct memcg_stock_pcp *stock;
1959 stock = &get_cpu_var(memcg_stock);
1960 if (memcg == stock->cached && stock->nr_pages)
1962 else /* need to call res_counter_charge */
1964 put_cpu_var(memcg_stock);
1969 * Returns stocks cached in percpu to res_counter and reset cached information.
1971 static void drain_stock(struct memcg_stock_pcp *stock)
1973 struct mem_cgroup *old = stock->cached;
1975 if (stock->nr_pages) {
1976 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
1978 res_counter_uncharge(&old->res, bytes);
1979 if (do_swap_account)
1980 res_counter_uncharge(&old->memsw, bytes);
1981 stock->nr_pages = 0;
1983 stock->cached = NULL;
1987 * This must be called under preempt disabled or must be called by
1988 * a thread which is pinned to local cpu.
1990 static void drain_local_stock(struct work_struct *dummy)
1992 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1994 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1998 * Cache charges(val) which is from res_counter, to local per_cpu area.
1999 * This will be consumed by consume_stock() function, later.
2001 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2003 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2005 if (stock->cached != memcg) { /* reset if necessary */
2007 stock->cached = memcg;
2009 stock->nr_pages += nr_pages;
2010 put_cpu_var(memcg_stock);
2014 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2015 * of the hierarchy under it. sync flag says whether we should block
2016 * until the work is done.
2018 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2022 /* Notify other cpus that system-wide "drain" is running */
2025 for_each_online_cpu(cpu) {
2026 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2027 struct mem_cgroup *memcg;
2029 memcg = stock->cached;
2030 if (!memcg || !stock->nr_pages)
2032 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2034 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2036 drain_local_stock(&stock->work);
2038 schedule_work_on(cpu, &stock->work);
2046 for_each_online_cpu(cpu) {
2047 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2048 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2049 flush_work(&stock->work);
2056 * Tries to drain stocked charges in other cpus. This function is asynchronous
2057 * and just put a work per cpu for draining localy on each cpu. Caller can
2058 * expects some charges will be back to res_counter later but cannot wait for
2061 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2064 * If someone calls draining, avoid adding more kworker runs.
2066 if (!mutex_trylock(&percpu_charge_mutex))
2068 drain_all_stock(root_memcg, false);
2069 mutex_unlock(&percpu_charge_mutex);
2072 /* This is a synchronous drain interface. */
2073 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2075 /* called when force_empty is called */
2076 mutex_lock(&percpu_charge_mutex);
2077 drain_all_stock(root_memcg, true);
2078 mutex_unlock(&percpu_charge_mutex);
2082 * This function drains percpu counter value from DEAD cpu and
2083 * move it to local cpu. Note that this function can be preempted.
2085 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2089 spin_lock(&memcg->pcp_counter_lock);
2090 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2091 long x = per_cpu(memcg->stat->count[i], cpu);
2093 per_cpu(memcg->stat->count[i], cpu) = 0;
2094 memcg->nocpu_base.count[i] += x;
2096 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2097 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2099 per_cpu(memcg->stat->events[i], cpu) = 0;
2100 memcg->nocpu_base.events[i] += x;
2102 /* need to clear ON_MOVE value, works as a kind of lock. */
2103 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2104 spin_unlock(&memcg->pcp_counter_lock);
2107 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2109 int idx = MEM_CGROUP_ON_MOVE;
2111 spin_lock(&memcg->pcp_counter_lock);
2112 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2113 spin_unlock(&memcg->pcp_counter_lock);
2116 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2117 unsigned long action,
2120 int cpu = (unsigned long)hcpu;
2121 struct memcg_stock_pcp *stock;
2122 struct mem_cgroup *iter;
2124 if ((action == CPU_ONLINE)) {
2125 for_each_mem_cgroup(iter)
2126 synchronize_mem_cgroup_on_move(iter, cpu);
2130 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2133 for_each_mem_cgroup(iter)
2134 mem_cgroup_drain_pcp_counter(iter, cpu);
2136 stock = &per_cpu(memcg_stock, cpu);
2142 /* See __mem_cgroup_try_charge() for details */
2144 CHARGE_OK, /* success */
2145 CHARGE_RETRY, /* need to retry but retry is not bad */
2146 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2147 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2148 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2151 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2152 unsigned int nr_pages, bool oom_check)
2154 unsigned long csize = nr_pages * PAGE_SIZE;
2155 struct mem_cgroup *mem_over_limit;
2156 struct res_counter *fail_res;
2157 unsigned long flags = 0;
2160 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2163 if (!do_swap_account)
2165 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2169 res_counter_uncharge(&memcg->res, csize);
2170 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2171 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2173 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2175 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2176 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2178 * Never reclaim on behalf of optional batching, retry with a
2179 * single page instead.
2181 if (nr_pages == CHARGE_BATCH)
2182 return CHARGE_RETRY;
2184 if (!(gfp_mask & __GFP_WAIT))
2185 return CHARGE_WOULDBLOCK;
2187 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2188 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2189 return CHARGE_RETRY;
2191 * Even though the limit is exceeded at this point, reclaim
2192 * may have been able to free some pages. Retry the charge
2193 * before killing the task.
2195 * Only for regular pages, though: huge pages are rather
2196 * unlikely to succeed so close to the limit, and we fall back
2197 * to regular pages anyway in case of failure.
2199 if (nr_pages == 1 && ret)
2200 return CHARGE_RETRY;
2203 * At task move, charge accounts can be doubly counted. So, it's
2204 * better to wait until the end of task_move if something is going on.
2206 if (mem_cgroup_wait_acct_move(mem_over_limit))
2207 return CHARGE_RETRY;
2209 /* If we don't need to call oom-killer at el, return immediately */
2211 return CHARGE_NOMEM;
2213 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2214 return CHARGE_OOM_DIE;
2216 return CHARGE_RETRY;
2220 * Unlike exported interface, "oom" parameter is added. if oom==true,
2221 * oom-killer can be invoked.
2223 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2225 unsigned int nr_pages,
2226 struct mem_cgroup **ptr,
2229 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2230 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2231 struct mem_cgroup *memcg = NULL;
2235 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2236 * in system level. So, allow to go ahead dying process in addition to
2239 if (unlikely(test_thread_flag(TIF_MEMDIE)
2240 || fatal_signal_pending(current)))
2244 * We always charge the cgroup the mm_struct belongs to.
2245 * The mm_struct's mem_cgroup changes on task migration if the
2246 * thread group leader migrates. It's possible that mm is not
2247 * set, if so charge the init_mm (happens for pagecache usage).
2252 if (*ptr) { /* css should be a valid one */
2254 VM_BUG_ON(css_is_removed(&memcg->css));
2255 if (mem_cgroup_is_root(memcg))
2257 if (nr_pages == 1 && consume_stock(memcg))
2259 css_get(&memcg->css);
2261 struct task_struct *p;
2264 p = rcu_dereference(mm->owner);
2266 * Because we don't have task_lock(), "p" can exit.
2267 * In that case, "memcg" can point to root or p can be NULL with
2268 * race with swapoff. Then, we have small risk of mis-accouning.
2269 * But such kind of mis-account by race always happens because
2270 * we don't have cgroup_mutex(). It's overkill and we allo that
2272 * (*) swapoff at el will charge against mm-struct not against
2273 * task-struct. So, mm->owner can be NULL.
2275 memcg = mem_cgroup_from_task(p);
2276 if (!memcg || mem_cgroup_is_root(memcg)) {
2280 if (nr_pages == 1 && consume_stock(memcg)) {
2282 * It seems dagerous to access memcg without css_get().
2283 * But considering how consume_stok works, it's not
2284 * necessary. If consume_stock success, some charges
2285 * from this memcg are cached on this cpu. So, we
2286 * don't need to call css_get()/css_tryget() before
2287 * calling consume_stock().
2292 /* after here, we may be blocked. we need to get refcnt */
2293 if (!css_tryget(&memcg->css)) {
2303 /* If killed, bypass charge */
2304 if (fatal_signal_pending(current)) {
2305 css_put(&memcg->css);
2310 if (oom && !nr_oom_retries) {
2312 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2315 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2319 case CHARGE_RETRY: /* not in OOM situation but retry */
2321 css_put(&memcg->css);
2324 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2325 css_put(&memcg->css);
2327 case CHARGE_NOMEM: /* OOM routine works */
2329 css_put(&memcg->css);
2332 /* If oom, we never return -ENOMEM */
2335 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2336 css_put(&memcg->css);
2339 } while (ret != CHARGE_OK);
2341 if (batch > nr_pages)
2342 refill_stock(memcg, batch - nr_pages);
2343 css_put(&memcg->css);
2356 * Somemtimes we have to undo a charge we got by try_charge().
2357 * This function is for that and do uncharge, put css's refcnt.
2358 * gotten by try_charge().
2360 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2361 unsigned int nr_pages)
2363 if (!mem_cgroup_is_root(memcg)) {
2364 unsigned long bytes = nr_pages * PAGE_SIZE;
2366 res_counter_uncharge(&memcg->res, bytes);
2367 if (do_swap_account)
2368 res_counter_uncharge(&memcg->memsw, bytes);
2373 * A helper function to get mem_cgroup from ID. must be called under
2374 * rcu_read_lock(). The caller must check css_is_removed() or some if
2375 * it's concern. (dropping refcnt from swap can be called against removed
2378 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2380 struct cgroup_subsys_state *css;
2382 /* ID 0 is unused ID */
2385 css = css_lookup(&mem_cgroup_subsys, id);
2388 return container_of(css, struct mem_cgroup, css);
2391 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2393 struct mem_cgroup *memcg = NULL;
2394 struct page_cgroup *pc;
2398 VM_BUG_ON(!PageLocked(page));
2400 pc = lookup_page_cgroup(page);
2401 lock_page_cgroup(pc);
2402 if (PageCgroupUsed(pc)) {
2403 memcg = pc->mem_cgroup;
2404 if (memcg && !css_tryget(&memcg->css))
2406 } else if (PageSwapCache(page)) {
2407 ent.val = page_private(page);
2408 id = lookup_swap_cgroup_id(ent);
2410 memcg = mem_cgroup_lookup(id);
2411 if (memcg && !css_tryget(&memcg->css))
2415 unlock_page_cgroup(pc);
2419 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2421 unsigned int nr_pages,
2422 struct page_cgroup *pc,
2423 enum charge_type ctype)
2425 lock_page_cgroup(pc);
2426 if (unlikely(PageCgroupUsed(pc))) {
2427 unlock_page_cgroup(pc);
2428 __mem_cgroup_cancel_charge(memcg, nr_pages);
2432 * we don't need page_cgroup_lock about tail pages, becase they are not
2433 * accessed by any other context at this point.
2435 pc->mem_cgroup = memcg;
2437 * We access a page_cgroup asynchronously without lock_page_cgroup().
2438 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2439 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2440 * before USED bit, we need memory barrier here.
2441 * See mem_cgroup_add_lru_list(), etc.
2445 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2446 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2447 SetPageCgroupCache(pc);
2448 SetPageCgroupUsed(pc);
2450 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2451 ClearPageCgroupCache(pc);
2452 SetPageCgroupUsed(pc);
2458 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2459 unlock_page_cgroup(pc);
2461 * "charge_statistics" updated event counter. Then, check it.
2462 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2463 * if they exceeds softlimit.
2465 memcg_check_events(memcg, page);
2468 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2470 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2471 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2473 * Because tail pages are not marked as "used", set it. We're under
2474 * zone->lru_lock, 'splitting on pmd' and compound_lock.
2475 * charge/uncharge will be never happen and move_account() is done under
2476 * compound_lock(), so we don't have to take care of races.
2478 void mem_cgroup_split_huge_fixup(struct page *head)
2480 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2481 struct page_cgroup *pc;
2484 if (mem_cgroup_disabled())
2486 for (i = 1; i < HPAGE_PMD_NR; i++) {
2488 pc->mem_cgroup = head_pc->mem_cgroup;
2489 smp_wmb();/* see __commit_charge() */
2491 * LRU flags cannot be copied because we need to add tail
2492 * page to LRU by generic call and our hooks will be called.
2494 pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2497 if (PageCgroupAcctLRU(head_pc)) {
2499 struct mem_cgroup_per_zone *mz;
2501 * We hold lru_lock, then, reduce counter directly.
2503 lru = page_lru(head);
2504 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2505 MEM_CGROUP_ZSTAT(mz, lru) -= HPAGE_PMD_NR - 1;
2511 * mem_cgroup_move_account - move account of the page
2513 * @nr_pages: number of regular pages (>1 for huge pages)
2514 * @pc: page_cgroup of the page.
2515 * @from: mem_cgroup which the page is moved from.
2516 * @to: mem_cgroup which the page is moved to. @from != @to.
2517 * @uncharge: whether we should call uncharge and css_put against @from.
2519 * The caller must confirm following.
2520 * - page is not on LRU (isolate_page() is useful.)
2521 * - compound_lock is held when nr_pages > 1
2523 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2524 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2525 * true, this function does "uncharge" from old cgroup, but it doesn't if
2526 * @uncharge is false, so a caller should do "uncharge".
2528 static int mem_cgroup_move_account(struct page *page,
2529 unsigned int nr_pages,
2530 struct page_cgroup *pc,
2531 struct mem_cgroup *from,
2532 struct mem_cgroup *to,
2535 unsigned long flags;
2538 VM_BUG_ON(from == to);
2539 VM_BUG_ON(PageLRU(page));
2541 * The page is isolated from LRU. So, collapse function
2542 * will not handle this page. But page splitting can happen.
2543 * Do this check under compound_page_lock(). The caller should
2547 if (nr_pages > 1 && !PageTransHuge(page))
2550 lock_page_cgroup(pc);
2553 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2556 move_lock_page_cgroup(pc, &flags);
2558 if (PageCgroupFileMapped(pc)) {
2559 /* Update mapped_file data for mem_cgroup */
2561 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2562 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2565 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2567 /* This is not "cancel", but cancel_charge does all we need. */
2568 __mem_cgroup_cancel_charge(from, nr_pages);
2570 /* caller should have done css_get */
2571 pc->mem_cgroup = to;
2572 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2574 * We charges against "to" which may not have any tasks. Then, "to"
2575 * can be under rmdir(). But in current implementation, caller of
2576 * this function is just force_empty() and move charge, so it's
2577 * guaranteed that "to" is never removed. So, we don't check rmdir
2580 move_unlock_page_cgroup(pc, &flags);
2583 unlock_page_cgroup(pc);
2587 memcg_check_events(to, page);
2588 memcg_check_events(from, page);
2594 * move charges to its parent.
2597 static int mem_cgroup_move_parent(struct page *page,
2598 struct page_cgroup *pc,
2599 struct mem_cgroup *child,
2602 struct cgroup *cg = child->css.cgroup;
2603 struct cgroup *pcg = cg->parent;
2604 struct mem_cgroup *parent;
2605 unsigned int nr_pages;
2606 unsigned long uninitialized_var(flags);
2614 if (!get_page_unless_zero(page))
2616 if (isolate_lru_page(page))
2619 nr_pages = hpage_nr_pages(page);
2621 parent = mem_cgroup_from_cont(pcg);
2622 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2627 flags = compound_lock_irqsave(page);
2629 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2631 __mem_cgroup_cancel_charge(parent, nr_pages);
2634 compound_unlock_irqrestore(page, flags);
2636 putback_lru_page(page);
2644 * Charge the memory controller for page usage.
2646 * 0 if the charge was successful
2647 * < 0 if the cgroup is over its limit
2649 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2650 gfp_t gfp_mask, enum charge_type ctype)
2652 struct mem_cgroup *memcg = NULL;
2653 unsigned int nr_pages = 1;
2654 struct page_cgroup *pc;
2658 if (PageTransHuge(page)) {
2659 nr_pages <<= compound_order(page);
2660 VM_BUG_ON(!PageTransHuge(page));
2662 * Never OOM-kill a process for a huge page. The
2663 * fault handler will fall back to regular pages.
2668 pc = lookup_page_cgroup(page);
2669 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2673 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2677 int mem_cgroup_newpage_charge(struct page *page,
2678 struct mm_struct *mm, gfp_t gfp_mask)
2680 if (mem_cgroup_disabled())
2682 VM_BUG_ON(page_mapped(page));
2683 VM_BUG_ON(page->mapping && !PageAnon(page));
2685 return mem_cgroup_charge_common(page, mm, gfp_mask,
2686 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2690 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2691 enum charge_type ctype);
2694 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2695 enum charge_type ctype)
2697 struct page_cgroup *pc = lookup_page_cgroup(page);
2698 struct zone *zone = page_zone(page);
2699 unsigned long flags;
2700 bool removed = false;
2703 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2704 * is already on LRU. It means the page may on some other page_cgroup's
2705 * LRU. Take care of it.
2707 spin_lock_irqsave(&zone->lru_lock, flags);
2708 if (PageLRU(page)) {
2709 del_page_from_lru_list(zone, page, page_lru(page));
2713 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2715 add_page_to_lru_list(zone, page, page_lru(page));
2718 spin_unlock_irqrestore(&zone->lru_lock, flags);
2722 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2725 struct mem_cgroup *memcg = NULL;
2726 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
2729 if (mem_cgroup_disabled())
2731 if (PageCompound(page))
2736 if (!page_is_file_cache(page))
2737 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2739 if (!PageSwapCache(page)) {
2740 ret = mem_cgroup_charge_common(page, mm, gfp_mask, type);
2741 WARN_ON_ONCE(PageLRU(page));
2742 } else { /* page is swapcache/shmem */
2743 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2745 __mem_cgroup_commit_charge_swapin(page, memcg, type);
2751 * While swap-in, try_charge -> commit or cancel, the page is locked.
2752 * And when try_charge() successfully returns, one refcnt to memcg without
2753 * struct page_cgroup is acquired. This refcnt will be consumed by
2754 * "commit()" or removed by "cancel()"
2756 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2758 gfp_t mask, struct mem_cgroup **memcgp)
2760 struct mem_cgroup *memcg;
2765 if (mem_cgroup_disabled())
2768 if (!do_swap_account)
2771 * A racing thread's fault, or swapoff, may have already updated
2772 * the pte, and even removed page from swap cache: in those cases
2773 * do_swap_page()'s pte_same() test will fail; but there's also a
2774 * KSM case which does need to charge the page.
2776 if (!PageSwapCache(page))
2778 memcg = try_get_mem_cgroup_from_page(page);
2782 ret = __mem_cgroup_try_charge(NULL, mask, 1, memcgp, true);
2783 css_put(&memcg->css);
2788 return __mem_cgroup_try_charge(mm, mask, 1, memcgp, true);
2792 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *memcg,
2793 enum charge_type ctype)
2795 if (mem_cgroup_disabled())
2799 cgroup_exclude_rmdir(&memcg->css);
2801 __mem_cgroup_commit_charge_lrucare(page, memcg, ctype);
2803 * Now swap is on-memory. This means this page may be
2804 * counted both as mem and swap....double count.
2805 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2806 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2807 * may call delete_from_swap_cache() before reach here.
2809 if (do_swap_account && PageSwapCache(page)) {
2810 swp_entry_t ent = {.val = page_private(page)};
2811 struct mem_cgroup *swap_memcg;
2814 id = swap_cgroup_record(ent, 0);
2816 swap_memcg = mem_cgroup_lookup(id);
2819 * This recorded memcg can be obsolete one. So, avoid
2820 * calling css_tryget
2822 if (!mem_cgroup_is_root(swap_memcg))
2823 res_counter_uncharge(&swap_memcg->memsw,
2825 mem_cgroup_swap_statistics(swap_memcg, false);
2826 mem_cgroup_put(swap_memcg);
2831 * At swapin, we may charge account against cgroup which has no tasks.
2832 * So, rmdir()->pre_destroy() can be called while we do this charge.
2833 * In that case, we need to call pre_destroy() again. check it here.
2835 cgroup_release_and_wakeup_rmdir(&memcg->css);
2838 void mem_cgroup_commit_charge_swapin(struct page *page,
2839 struct mem_cgroup *memcg)
2841 __mem_cgroup_commit_charge_swapin(page, memcg,
2842 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2845 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2847 if (mem_cgroup_disabled())
2851 __mem_cgroup_cancel_charge(memcg, 1);
2854 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2855 unsigned int nr_pages,
2856 const enum charge_type ctype)
2858 struct memcg_batch_info *batch = NULL;
2859 bool uncharge_memsw = true;
2861 /* If swapout, usage of swap doesn't decrease */
2862 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2863 uncharge_memsw = false;
2865 batch = ¤t->memcg_batch;
2867 * In usual, we do css_get() when we remember memcg pointer.
2868 * But in this case, we keep res->usage until end of a series of
2869 * uncharges. Then, it's ok to ignore memcg's refcnt.
2872 batch->memcg = memcg;
2874 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2875 * In those cases, all pages freed continuously can be expected to be in
2876 * the same cgroup and we have chance to coalesce uncharges.
2877 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2878 * because we want to do uncharge as soon as possible.
2881 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2882 goto direct_uncharge;
2885 goto direct_uncharge;
2888 * In typical case, batch->memcg == mem. This means we can
2889 * merge a series of uncharges to an uncharge of res_counter.
2890 * If not, we uncharge res_counter ony by one.
2892 if (batch->memcg != memcg)
2893 goto direct_uncharge;
2894 /* remember freed charge and uncharge it later */
2897 batch->memsw_nr_pages++;
2900 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2902 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2903 if (unlikely(batch->memcg != memcg))
2904 memcg_oom_recover(memcg);
2909 * uncharge if !page_mapped(page)
2911 static struct mem_cgroup *
2912 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2914 struct mem_cgroup *memcg = NULL;
2915 unsigned int nr_pages = 1;
2916 struct page_cgroup *pc;
2918 if (mem_cgroup_disabled())
2921 if (PageSwapCache(page))
2924 if (PageTransHuge(page)) {
2925 nr_pages <<= compound_order(page);
2926 VM_BUG_ON(!PageTransHuge(page));
2929 * Check if our page_cgroup is valid
2931 pc = lookup_page_cgroup(page);
2932 if (unlikely(!PageCgroupUsed(pc)))
2935 lock_page_cgroup(pc);
2937 memcg = pc->mem_cgroup;
2939 if (!PageCgroupUsed(pc))
2943 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2944 case MEM_CGROUP_CHARGE_TYPE_DROP:
2945 /* See mem_cgroup_prepare_migration() */
2946 if (page_mapped(page) || PageCgroupMigration(pc))
2949 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2950 if (!PageAnon(page)) { /* Shared memory */
2951 if (page->mapping && !page_is_file_cache(page))
2953 } else if (page_mapped(page)) /* Anon */
2960 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
2962 ClearPageCgroupUsed(pc);
2964 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2965 * freed from LRU. This is safe because uncharged page is expected not
2966 * to be reused (freed soon). Exception is SwapCache, it's handled by
2967 * special functions.
2970 unlock_page_cgroup(pc);
2972 * even after unlock, we have memcg->res.usage here and this memcg
2973 * will never be freed.
2975 memcg_check_events(memcg, page);
2976 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2977 mem_cgroup_swap_statistics(memcg, true);
2978 mem_cgroup_get(memcg);
2980 if (!mem_cgroup_is_root(memcg))
2981 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
2986 unlock_page_cgroup(pc);
2990 void mem_cgroup_uncharge_page(struct page *page)
2993 if (page_mapped(page))
2995 VM_BUG_ON(page->mapping && !PageAnon(page));
2996 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2999 void mem_cgroup_uncharge_cache_page(struct page *page)
3001 VM_BUG_ON(page_mapped(page));
3002 VM_BUG_ON(page->mapping);
3003 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3007 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3008 * In that cases, pages are freed continuously and we can expect pages
3009 * are in the same memcg. All these calls itself limits the number of
3010 * pages freed at once, then uncharge_start/end() is called properly.
3011 * This may be called prural(2) times in a context,
3014 void mem_cgroup_uncharge_start(void)
3016 current->memcg_batch.do_batch++;
3017 /* We can do nest. */
3018 if (current->memcg_batch.do_batch == 1) {
3019 current->memcg_batch.memcg = NULL;
3020 current->memcg_batch.nr_pages = 0;
3021 current->memcg_batch.memsw_nr_pages = 0;
3025 void mem_cgroup_uncharge_end(void)
3027 struct memcg_batch_info *batch = ¤t->memcg_batch;
3029 if (!batch->do_batch)
3033 if (batch->do_batch) /* If stacked, do nothing. */
3039 * This "batch->memcg" is valid without any css_get/put etc...
3040 * bacause we hide charges behind us.
3042 if (batch->nr_pages)
3043 res_counter_uncharge(&batch->memcg->res,
3044 batch->nr_pages * PAGE_SIZE);
3045 if (batch->memsw_nr_pages)
3046 res_counter_uncharge(&batch->memcg->memsw,
3047 batch->memsw_nr_pages * PAGE_SIZE);
3048 memcg_oom_recover(batch->memcg);
3049 /* forget this pointer (for sanity check) */
3050 batch->memcg = NULL;
3054 * A function for resetting pc->mem_cgroup for newly allocated pages.
3055 * This function should be called if the newpage will be added to LRU
3056 * before start accounting.
3058 void mem_cgroup_reset_owner(struct page *newpage)
3060 struct page_cgroup *pc;
3062 if (mem_cgroup_disabled())
3065 pc = lookup_page_cgroup(newpage);
3066 VM_BUG_ON(PageCgroupUsed(pc));
3067 pc->mem_cgroup = root_mem_cgroup;
3072 * called after __delete_from_swap_cache() and drop "page" account.
3073 * memcg information is recorded to swap_cgroup of "ent"
3076 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3078 struct mem_cgroup *memcg;
3079 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3081 if (!swapout) /* this was a swap cache but the swap is unused ! */
3082 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3084 memcg = __mem_cgroup_uncharge_common(page, ctype);
3087 * record memcg information, if swapout && memcg != NULL,
3088 * mem_cgroup_get() was called in uncharge().
3090 if (do_swap_account && swapout && memcg)
3091 swap_cgroup_record(ent, css_id(&memcg->css));
3095 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3097 * called from swap_entry_free(). remove record in swap_cgroup and
3098 * uncharge "memsw" account.
3100 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3102 struct mem_cgroup *memcg;
3105 if (!do_swap_account)
3108 id = swap_cgroup_record(ent, 0);
3110 memcg = mem_cgroup_lookup(id);
3113 * We uncharge this because swap is freed.
3114 * This memcg can be obsolete one. We avoid calling css_tryget
3116 if (!mem_cgroup_is_root(memcg))
3117 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3118 mem_cgroup_swap_statistics(memcg, false);
3119 mem_cgroup_put(memcg);
3125 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3126 * @entry: swap entry to be moved
3127 * @from: mem_cgroup which the entry is moved from
3128 * @to: mem_cgroup which the entry is moved to
3129 * @need_fixup: whether we should fixup res_counters and refcounts.
3131 * It succeeds only when the swap_cgroup's record for this entry is the same
3132 * as the mem_cgroup's id of @from.
3134 * Returns 0 on success, -EINVAL on failure.
3136 * The caller must have charged to @to, IOW, called res_counter_charge() about
3137 * both res and memsw, and called css_get().
3139 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3140 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3142 unsigned short old_id, new_id;
3144 old_id = css_id(&from->css);
3145 new_id = css_id(&to->css);
3147 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3148 mem_cgroup_swap_statistics(from, false);
3149 mem_cgroup_swap_statistics(to, true);
3151 * This function is only called from task migration context now.
3152 * It postpones res_counter and refcount handling till the end
3153 * of task migration(mem_cgroup_clear_mc()) for performance
3154 * improvement. But we cannot postpone mem_cgroup_get(to)
3155 * because if the process that has been moved to @to does
3156 * swap-in, the refcount of @to might be decreased to 0.
3160 if (!mem_cgroup_is_root(from))
3161 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3162 mem_cgroup_put(from);
3164 * we charged both to->res and to->memsw, so we should
3167 if (!mem_cgroup_is_root(to))
3168 res_counter_uncharge(&to->res, PAGE_SIZE);
3175 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3176 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3183 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3186 int mem_cgroup_prepare_migration(struct page *page,
3187 struct page *newpage, struct mem_cgroup **memcgp, gfp_t gfp_mask)
3189 struct mem_cgroup *memcg = NULL;
3190 struct page_cgroup *pc;
3191 enum charge_type ctype;
3196 VM_BUG_ON(PageTransHuge(page));
3197 if (mem_cgroup_disabled())
3200 pc = lookup_page_cgroup(page);
3201 lock_page_cgroup(pc);
3202 if (PageCgroupUsed(pc)) {
3203 memcg = pc->mem_cgroup;
3204 css_get(&memcg->css);
3206 * At migrating an anonymous page, its mapcount goes down
3207 * to 0 and uncharge() will be called. But, even if it's fully
3208 * unmapped, migration may fail and this page has to be
3209 * charged again. We set MIGRATION flag here and delay uncharge
3210 * until end_migration() is called
3212 * Corner Case Thinking
3214 * When the old page was mapped as Anon and it's unmap-and-freed
3215 * while migration was ongoing.
3216 * If unmap finds the old page, uncharge() of it will be delayed
3217 * until end_migration(). If unmap finds a new page, it's
3218 * uncharged when it make mapcount to be 1->0. If unmap code
3219 * finds swap_migration_entry, the new page will not be mapped
3220 * and end_migration() will find it(mapcount==0).
3223 * When the old page was mapped but migraion fails, the kernel
3224 * remaps it. A charge for it is kept by MIGRATION flag even
3225 * if mapcount goes down to 0. We can do remap successfully
3226 * without charging it again.
3229 * The "old" page is under lock_page() until the end of
3230 * migration, so, the old page itself will not be swapped-out.
3231 * If the new page is swapped out before end_migraton, our
3232 * hook to usual swap-out path will catch the event.
3235 SetPageCgroupMigration(pc);
3237 unlock_page_cgroup(pc);
3239 * If the page is not charged at this point,
3246 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, memcgp, false);
3247 css_put(&memcg->css);/* drop extra refcnt */
3248 if (ret || *memcgp == NULL) {
3249 if (PageAnon(page)) {
3250 lock_page_cgroup(pc);
3251 ClearPageCgroupMigration(pc);
3252 unlock_page_cgroup(pc);
3254 * The old page may be fully unmapped while we kept it.
3256 mem_cgroup_uncharge_page(page);
3261 * We charge new page before it's used/mapped. So, even if unlock_page()
3262 * is called before end_migration, we can catch all events on this new
3263 * page. In the case new page is migrated but not remapped, new page's
3264 * mapcount will be finally 0 and we call uncharge in end_migration().
3266 pc = lookup_page_cgroup(newpage);
3268 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3269 else if (page_is_file_cache(page))
3270 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3272 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3273 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3277 /* remove redundant charge if migration failed*/
3278 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3279 struct page *oldpage, struct page *newpage, bool migration_ok)
3281 struct page *used, *unused;
3282 struct page_cgroup *pc;
3286 /* blocks rmdir() */
3287 cgroup_exclude_rmdir(&memcg->css);
3288 if (!migration_ok) {
3296 * We disallowed uncharge of pages under migration because mapcount
3297 * of the page goes down to zero, temporarly.
3298 * Clear the flag and check the page should be charged.
3300 pc = lookup_page_cgroup(oldpage);
3301 lock_page_cgroup(pc);
3302 ClearPageCgroupMigration(pc);
3303 unlock_page_cgroup(pc);
3305 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3308 * If a page is a file cache, radix-tree replacement is very atomic
3309 * and we can skip this check. When it was an Anon page, its mapcount
3310 * goes down to 0. But because we added MIGRATION flage, it's not
3311 * uncharged yet. There are several case but page->mapcount check
3312 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3313 * check. (see prepare_charge() also)
3316 mem_cgroup_uncharge_page(used);
3318 * At migration, we may charge account against cgroup which has no
3320 * So, rmdir()->pre_destroy() can be called while we do this charge.
3321 * In that case, we need to call pre_destroy() again. check it here.
3323 cgroup_release_and_wakeup_rmdir(&memcg->css);
3327 * At replace page cache, newpage is not under any memcg but it's on
3328 * LRU. So, this function doesn't touch res_counter but handles LRU
3329 * in correct way. Both pages are locked so we cannot race with uncharge.
3331 void mem_cgroup_replace_page_cache(struct page *oldpage,
3332 struct page *newpage)
3334 struct mem_cgroup *memcg;
3335 struct page_cgroup *pc;
3336 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3338 if (mem_cgroup_disabled())
3341 pc = lookup_page_cgroup(oldpage);
3342 /* fix accounting on old pages */
3343 lock_page_cgroup(pc);
3344 memcg = pc->mem_cgroup;
3345 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1);
3346 ClearPageCgroupUsed(pc);
3347 unlock_page_cgroup(pc);
3349 if (PageSwapBacked(oldpage))
3350 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3353 * Even if newpage->mapping was NULL before starting replacement,
3354 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3355 * LRU while we overwrite pc->mem_cgroup.
3357 __mem_cgroup_commit_charge_lrucare(newpage, memcg, type);
3360 #ifdef CONFIG_DEBUG_VM
3361 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3363 struct page_cgroup *pc;
3365 pc = lookup_page_cgroup(page);
3367 * Can be NULL while feeding pages into the page allocator for
3368 * the first time, i.e. during boot or memory hotplug;
3369 * or when mem_cgroup_disabled().
3371 if (likely(pc) && PageCgroupUsed(pc))
3376 bool mem_cgroup_bad_page_check(struct page *page)
3378 if (mem_cgroup_disabled())
3381 return lookup_page_cgroup_used(page) != NULL;
3384 void mem_cgroup_print_bad_page(struct page *page)
3386 struct page_cgroup *pc;
3388 pc = lookup_page_cgroup_used(page);
3393 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3394 pc, pc->flags, pc->mem_cgroup);
3396 path = kmalloc(PATH_MAX, GFP_KERNEL);
3399 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3404 printk(KERN_CONT "(%s)\n",
3405 (ret < 0) ? "cannot get the path" : path);
3411 static DEFINE_MUTEX(set_limit_mutex);
3413 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3414 unsigned long long val)
3417 u64 memswlimit, memlimit;
3419 int children = mem_cgroup_count_children(memcg);
3420 u64 curusage, oldusage;
3424 * For keeping hierarchical_reclaim simple, how long we should retry
3425 * is depends on callers. We set our retry-count to be function
3426 * of # of children which we should visit in this loop.
3428 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3430 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3433 while (retry_count) {
3434 if (signal_pending(current)) {
3439 * Rather than hide all in some function, I do this in
3440 * open coded manner. You see what this really does.
3441 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3443 mutex_lock(&set_limit_mutex);
3444 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3445 if (memswlimit < val) {
3447 mutex_unlock(&set_limit_mutex);
3451 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3455 ret = res_counter_set_limit(&memcg->res, val);
3457 if (memswlimit == val)
3458 memcg->memsw_is_minimum = true;
3460 memcg->memsw_is_minimum = false;
3462 mutex_unlock(&set_limit_mutex);
3467 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3468 MEM_CGROUP_RECLAIM_SHRINK);
3469 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3470 /* Usage is reduced ? */
3471 if (curusage >= oldusage)
3474 oldusage = curusage;
3476 if (!ret && enlarge)
3477 memcg_oom_recover(memcg);
3482 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3483 unsigned long long val)
3486 u64 memlimit, memswlimit, oldusage, curusage;
3487 int children = mem_cgroup_count_children(memcg);
3491 /* see mem_cgroup_resize_res_limit */
3492 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3493 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3494 while (retry_count) {
3495 if (signal_pending(current)) {
3500 * Rather than hide all in some function, I do this in
3501 * open coded manner. You see what this really does.
3502 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3504 mutex_lock(&set_limit_mutex);
3505 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3506 if (memlimit > val) {
3508 mutex_unlock(&set_limit_mutex);
3511 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3512 if (memswlimit < val)
3514 ret = res_counter_set_limit(&memcg->memsw, val);
3516 if (memlimit == val)
3517 memcg->memsw_is_minimum = true;
3519 memcg->memsw_is_minimum = false;
3521 mutex_unlock(&set_limit_mutex);
3526 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3527 MEM_CGROUP_RECLAIM_NOSWAP |
3528 MEM_CGROUP_RECLAIM_SHRINK);
3529 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3530 /* Usage is reduced ? */
3531 if (curusage >= oldusage)
3534 oldusage = curusage;
3536 if (!ret && enlarge)
3537 memcg_oom_recover(memcg);
3541 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3543 unsigned long *total_scanned)
3545 unsigned long nr_reclaimed = 0;
3546 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3547 unsigned long reclaimed;
3549 struct mem_cgroup_tree_per_zone *mctz;
3550 unsigned long long excess;
3551 unsigned long nr_scanned;
3556 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3558 * This loop can run a while, specially if mem_cgroup's continuously
3559 * keep exceeding their soft limit and putting the system under
3566 mz = mem_cgroup_largest_soft_limit_node(mctz);
3571 reclaimed = mem_cgroup_soft_reclaim(mz->mem, zone,
3572 gfp_mask, &nr_scanned);
3573 nr_reclaimed += reclaimed;
3574 *total_scanned += nr_scanned;
3575 spin_lock(&mctz->lock);
3578 * If we failed to reclaim anything from this memory cgroup
3579 * it is time to move on to the next cgroup
3585 * Loop until we find yet another one.
3587 * By the time we get the soft_limit lock
3588 * again, someone might have aded the
3589 * group back on the RB tree. Iterate to
3590 * make sure we get a different mem.
3591 * mem_cgroup_largest_soft_limit_node returns
3592 * NULL if no other cgroup is present on
3596 __mem_cgroup_largest_soft_limit_node(mctz);
3598 css_put(&next_mz->mem->css);
3599 else /* next_mz == NULL or other memcg */
3603 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3604 excess = res_counter_soft_limit_excess(&mz->mem->res);
3606 * One school of thought says that we should not add
3607 * back the node to the tree if reclaim returns 0.
3608 * But our reclaim could return 0, simply because due
3609 * to priority we are exposing a smaller subset of
3610 * memory to reclaim from. Consider this as a longer
3613 /* If excess == 0, no tree ops */
3614 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3615 spin_unlock(&mctz->lock);
3616 css_put(&mz->mem->css);
3619 * Could not reclaim anything and there are no more
3620 * mem cgroups to try or we seem to be looping without
3621 * reclaiming anything.
3623 if (!nr_reclaimed &&
3625 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3627 } while (!nr_reclaimed);
3629 css_put(&next_mz->mem->css);
3630 return nr_reclaimed;
3634 * This routine traverse page_cgroup in given list and drop them all.
3635 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3637 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3638 int node, int zid, enum lru_list lru)
3640 struct mem_cgroup_per_zone *mz;
3641 unsigned long flags, loop;
3642 struct list_head *list;
3647 zone = &NODE_DATA(node)->node_zones[zid];
3648 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3649 list = &mz->lruvec.lists[lru];
3651 loop = MEM_CGROUP_ZSTAT(mz, lru);
3652 /* give some margin against EBUSY etc...*/
3656 struct page_cgroup *pc;
3660 spin_lock_irqsave(&zone->lru_lock, flags);
3661 if (list_empty(list)) {
3662 spin_unlock_irqrestore(&zone->lru_lock, flags);
3665 page = list_entry(list->prev, struct page, lru);
3667 list_move(&page->lru, list);
3669 spin_unlock_irqrestore(&zone->lru_lock, flags);
3672 spin_unlock_irqrestore(&zone->lru_lock, flags);
3674 pc = lookup_page_cgroup(page);
3676 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3680 if (ret == -EBUSY || ret == -EINVAL) {
3681 /* found lock contention or "pc" is obsolete. */
3688 if (!ret && !list_empty(list))
3694 * make mem_cgroup's charge to be 0 if there is no task.
3695 * This enables deleting this mem_cgroup.
3697 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3700 int node, zid, shrink;
3701 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3702 struct cgroup *cgrp = memcg->css.cgroup;
3704 css_get(&memcg->css);
3707 /* should free all ? */
3713 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3716 if (signal_pending(current))
3718 /* This is for making all *used* pages to be on LRU. */
3719 lru_add_drain_all();
3720 drain_all_stock_sync(memcg);
3722 mem_cgroup_start_move(memcg);
3723 for_each_node_state(node, N_HIGH_MEMORY) {
3724 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3727 ret = mem_cgroup_force_empty_list(memcg,
3736 mem_cgroup_end_move(memcg);
3737 memcg_oom_recover(memcg);
3738 /* it seems parent cgroup doesn't have enough mem */
3742 /* "ret" should also be checked to ensure all lists are empty. */
3743 } while (memcg->res.usage > 0 || ret);
3745 css_put(&memcg->css);
3749 /* returns EBUSY if there is a task or if we come here twice. */
3750 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3754 /* we call try-to-free pages for make this cgroup empty */
3755 lru_add_drain_all();
3756 /* try to free all pages in this cgroup */
3758 while (nr_retries && memcg->res.usage > 0) {
3761 if (signal_pending(current)) {
3765 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3769 /* maybe some writeback is necessary */
3770 congestion_wait(BLK_RW_ASYNC, HZ/10);
3775 /* try move_account...there may be some *locked* pages. */
3779 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3781 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3785 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3787 return mem_cgroup_from_cont(cont)->use_hierarchy;
3790 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3794 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3795 struct cgroup *parent = cont->parent;
3796 struct mem_cgroup *parent_memcg = NULL;
3799 parent_memcg = mem_cgroup_from_cont(parent);
3803 * If parent's use_hierarchy is set, we can't make any modifications
3804 * in the child subtrees. If it is unset, then the change can
3805 * occur, provided the current cgroup has no children.
3807 * For the root cgroup, parent_mem is NULL, we allow value to be
3808 * set if there are no children.
3810 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3811 (val == 1 || val == 0)) {
3812 if (list_empty(&cont->children))
3813 memcg->use_hierarchy = val;
3824 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3825 enum mem_cgroup_stat_index idx)
3827 struct mem_cgroup *iter;
3830 /* Per-cpu values can be negative, use a signed accumulator */
3831 for_each_mem_cgroup_tree(iter, memcg)
3832 val += mem_cgroup_read_stat(iter, idx);
3834 if (val < 0) /* race ? */
3839 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3843 if (!mem_cgroup_is_root(memcg)) {
3845 return res_counter_read_u64(&memcg->res, RES_USAGE);
3847 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3850 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3851 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3854 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3856 return val << PAGE_SHIFT;
3859 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3861 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3865 type = MEMFILE_TYPE(cft->private);
3866 name = MEMFILE_ATTR(cft->private);
3869 if (name == RES_USAGE)
3870 val = mem_cgroup_usage(memcg, false);
3872 val = res_counter_read_u64(&memcg->res, name);
3875 if (name == RES_USAGE)
3876 val = mem_cgroup_usage(memcg, true);
3878 val = res_counter_read_u64(&memcg->memsw, name);
3887 * The user of this function is...
3890 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3893 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3895 unsigned long long val;
3898 type = MEMFILE_TYPE(cft->private);
3899 name = MEMFILE_ATTR(cft->private);
3902 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3906 /* This function does all necessary parse...reuse it */
3907 ret = res_counter_memparse_write_strategy(buffer, &val);
3911 ret = mem_cgroup_resize_limit(memcg, val);
3913 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3915 case RES_SOFT_LIMIT:
3916 ret = res_counter_memparse_write_strategy(buffer, &val);
3920 * For memsw, soft limits are hard to implement in terms
3921 * of semantics, for now, we support soft limits for
3922 * control without swap
3925 ret = res_counter_set_soft_limit(&memcg->res, val);
3930 ret = -EINVAL; /* should be BUG() ? */
3936 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3937 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3939 struct cgroup *cgroup;
3940 unsigned long long min_limit, min_memsw_limit, tmp;
3942 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3943 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3944 cgroup = memcg->css.cgroup;
3945 if (!memcg->use_hierarchy)
3948 while (cgroup->parent) {
3949 cgroup = cgroup->parent;
3950 memcg = mem_cgroup_from_cont(cgroup);
3951 if (!memcg->use_hierarchy)
3953 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3954 min_limit = min(min_limit, tmp);
3955 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3956 min_memsw_limit = min(min_memsw_limit, tmp);
3959 *mem_limit = min_limit;
3960 *memsw_limit = min_memsw_limit;
3964 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3966 struct mem_cgroup *memcg;
3969 memcg = mem_cgroup_from_cont(cont);
3970 type = MEMFILE_TYPE(event);
3971 name = MEMFILE_ATTR(event);
3975 res_counter_reset_max(&memcg->res);
3977 res_counter_reset_max(&memcg->memsw);
3981 res_counter_reset_failcnt(&memcg->res);
3983 res_counter_reset_failcnt(&memcg->memsw);
3990 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3993 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3997 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3998 struct cftype *cft, u64 val)
4000 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4002 if (val >= (1 << NR_MOVE_TYPE))
4005 * We check this value several times in both in can_attach() and
4006 * attach(), so we need cgroup lock to prevent this value from being
4010 memcg->move_charge_at_immigrate = val;
4016 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4017 struct cftype *cft, u64 val)
4024 /* For read statistics */
4042 struct mcs_total_stat {
4043 s64 stat[NR_MCS_STAT];
4049 } memcg_stat_strings[NR_MCS_STAT] = {
4050 {"cache", "total_cache"},
4051 {"rss", "total_rss"},
4052 {"mapped_file", "total_mapped_file"},
4053 {"pgpgin", "total_pgpgin"},
4054 {"pgpgout", "total_pgpgout"},
4055 {"swap", "total_swap"},
4056 {"pgfault", "total_pgfault"},
4057 {"pgmajfault", "total_pgmajfault"},
4058 {"inactive_anon", "total_inactive_anon"},
4059 {"active_anon", "total_active_anon"},
4060 {"inactive_file", "total_inactive_file"},
4061 {"active_file", "total_active_file"},
4062 {"unevictable", "total_unevictable"}
4067 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4072 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4073 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4074 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4075 s->stat[MCS_RSS] += val * PAGE_SIZE;
4076 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4077 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4078 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4079 s->stat[MCS_PGPGIN] += val;
4080 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4081 s->stat[MCS_PGPGOUT] += val;
4082 if (do_swap_account) {
4083 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4084 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4086 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4087 s->stat[MCS_PGFAULT] += val;
4088 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4089 s->stat[MCS_PGMAJFAULT] += val;
4092 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4093 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4094 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4095 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4096 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4097 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4098 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4099 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4100 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4101 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4105 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4107 struct mem_cgroup *iter;
4109 for_each_mem_cgroup_tree(iter, memcg)
4110 mem_cgroup_get_local_stat(iter, s);
4114 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4117 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4118 unsigned long node_nr;
4119 struct cgroup *cont = m->private;
4120 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4122 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4123 seq_printf(m, "total=%lu", total_nr);
4124 for_each_node_state(nid, N_HIGH_MEMORY) {
4125 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4126 seq_printf(m, " N%d=%lu", nid, node_nr);
4130 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4131 seq_printf(m, "file=%lu", file_nr);
4132 for_each_node_state(nid, N_HIGH_MEMORY) {
4133 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4135 seq_printf(m, " N%d=%lu", nid, node_nr);
4139 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4140 seq_printf(m, "anon=%lu", anon_nr);
4141 for_each_node_state(nid, N_HIGH_MEMORY) {
4142 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4144 seq_printf(m, " N%d=%lu", nid, node_nr);
4148 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4149 seq_printf(m, "unevictable=%lu", unevictable_nr);
4150 for_each_node_state(nid, N_HIGH_MEMORY) {
4151 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4152 BIT(LRU_UNEVICTABLE));
4153 seq_printf(m, " N%d=%lu", nid, node_nr);
4158 #endif /* CONFIG_NUMA */
4160 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4161 struct cgroup_map_cb *cb)
4163 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4164 struct mcs_total_stat mystat;
4167 memset(&mystat, 0, sizeof(mystat));
4168 mem_cgroup_get_local_stat(mem_cont, &mystat);
4171 for (i = 0; i < NR_MCS_STAT; i++) {
4172 if (i == MCS_SWAP && !do_swap_account)
4174 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4177 /* Hierarchical information */
4179 unsigned long long limit, memsw_limit;
4180 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4181 cb->fill(cb, "hierarchical_memory_limit", limit);
4182 if (do_swap_account)
4183 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4186 memset(&mystat, 0, sizeof(mystat));
4187 mem_cgroup_get_total_stat(mem_cont, &mystat);
4188 for (i = 0; i < NR_MCS_STAT; i++) {
4189 if (i == MCS_SWAP && !do_swap_account)
4191 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4194 #ifdef CONFIG_DEBUG_VM
4197 struct mem_cgroup_per_zone *mz;
4198 unsigned long recent_rotated[2] = {0, 0};
4199 unsigned long recent_scanned[2] = {0, 0};
4201 for_each_online_node(nid)
4202 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4203 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4205 recent_rotated[0] +=
4206 mz->reclaim_stat.recent_rotated[0];
4207 recent_rotated[1] +=
4208 mz->reclaim_stat.recent_rotated[1];
4209 recent_scanned[0] +=
4210 mz->reclaim_stat.recent_scanned[0];
4211 recent_scanned[1] +=
4212 mz->reclaim_stat.recent_scanned[1];
4214 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4215 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4216 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4217 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4224 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4226 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4228 return mem_cgroup_swappiness(memcg);
4231 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4234 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4235 struct mem_cgroup *parent;
4240 if (cgrp->parent == NULL)
4243 parent = mem_cgroup_from_cont(cgrp->parent);
4247 /* If under hierarchy, only empty-root can set this value */
4248 if ((parent->use_hierarchy) ||
4249 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4254 memcg->swappiness = val;
4261 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4263 struct mem_cgroup_threshold_ary *t;
4269 t = rcu_dereference(memcg->thresholds.primary);
4271 t = rcu_dereference(memcg->memsw_thresholds.primary);
4276 usage = mem_cgroup_usage(memcg, swap);
4279 * current_threshold points to threshold just below usage.
4280 * If it's not true, a threshold was crossed after last
4281 * call of __mem_cgroup_threshold().
4283 i = t->current_threshold;
4286 * Iterate backward over array of thresholds starting from
4287 * current_threshold and check if a threshold is crossed.
4288 * If none of thresholds below usage is crossed, we read
4289 * only one element of the array here.
4291 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4292 eventfd_signal(t->entries[i].eventfd, 1);
4294 /* i = current_threshold + 1 */
4298 * Iterate forward over array of thresholds starting from
4299 * current_threshold+1 and check if a threshold is crossed.
4300 * If none of thresholds above usage is crossed, we read
4301 * only one element of the array here.
4303 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4304 eventfd_signal(t->entries[i].eventfd, 1);
4306 /* Update current_threshold */
4307 t->current_threshold = i - 1;
4312 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4315 __mem_cgroup_threshold(memcg, false);
4316 if (do_swap_account)
4317 __mem_cgroup_threshold(memcg, true);
4319 memcg = parent_mem_cgroup(memcg);
4323 static int compare_thresholds(const void *a, const void *b)
4325 const struct mem_cgroup_threshold *_a = a;
4326 const struct mem_cgroup_threshold *_b = b;
4328 return _a->threshold - _b->threshold;
4331 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4333 struct mem_cgroup_eventfd_list *ev;
4335 list_for_each_entry(ev, &memcg->oom_notify, list)
4336 eventfd_signal(ev->eventfd, 1);
4340 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4342 struct mem_cgroup *iter;
4344 for_each_mem_cgroup_tree(iter, memcg)
4345 mem_cgroup_oom_notify_cb(iter);
4348 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4349 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4351 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4352 struct mem_cgroup_thresholds *thresholds;
4353 struct mem_cgroup_threshold_ary *new;
4354 int type = MEMFILE_TYPE(cft->private);
4355 u64 threshold, usage;
4358 ret = res_counter_memparse_write_strategy(args, &threshold);
4362 mutex_lock(&memcg->thresholds_lock);
4365 thresholds = &memcg->thresholds;
4366 else if (type == _MEMSWAP)
4367 thresholds = &memcg->memsw_thresholds;
4371 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4373 /* Check if a threshold crossed before adding a new one */
4374 if (thresholds->primary)
4375 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4377 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4379 /* Allocate memory for new array of thresholds */
4380 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4388 /* Copy thresholds (if any) to new array */
4389 if (thresholds->primary) {
4390 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4391 sizeof(struct mem_cgroup_threshold));
4394 /* Add new threshold */
4395 new->entries[size - 1].eventfd = eventfd;
4396 new->entries[size - 1].threshold = threshold;
4398 /* Sort thresholds. Registering of new threshold isn't time-critical */
4399 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4400 compare_thresholds, NULL);
4402 /* Find current threshold */
4403 new->current_threshold = -1;
4404 for (i = 0; i < size; i++) {
4405 if (new->entries[i].threshold < usage) {
4407 * new->current_threshold will not be used until
4408 * rcu_assign_pointer(), so it's safe to increment
4411 ++new->current_threshold;
4415 /* Free old spare buffer and save old primary buffer as spare */
4416 kfree(thresholds->spare);
4417 thresholds->spare = thresholds->primary;
4419 rcu_assign_pointer(thresholds->primary, new);
4421 /* To be sure that nobody uses thresholds */
4425 mutex_unlock(&memcg->thresholds_lock);
4430 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4431 struct cftype *cft, struct eventfd_ctx *eventfd)
4433 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4434 struct mem_cgroup_thresholds *thresholds;
4435 struct mem_cgroup_threshold_ary *new;
4436 int type = MEMFILE_TYPE(cft->private);
4440 mutex_lock(&memcg->thresholds_lock);
4442 thresholds = &memcg->thresholds;
4443 else if (type == _MEMSWAP)
4444 thresholds = &memcg->memsw_thresholds;
4449 * Something went wrong if we trying to unregister a threshold
4450 * if we don't have thresholds
4452 BUG_ON(!thresholds);
4454 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4456 /* Check if a threshold crossed before removing */
4457 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4459 /* Calculate new number of threshold */
4461 for (i = 0; i < thresholds->primary->size; i++) {
4462 if (thresholds->primary->entries[i].eventfd != eventfd)
4466 new = thresholds->spare;
4468 /* Set thresholds array to NULL if we don't have thresholds */
4477 /* Copy thresholds and find current threshold */
4478 new->current_threshold = -1;
4479 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4480 if (thresholds->primary->entries[i].eventfd == eventfd)
4483 new->entries[j] = thresholds->primary->entries[i];
4484 if (new->entries[j].threshold < usage) {
4486 * new->current_threshold will not be used
4487 * until rcu_assign_pointer(), so it's safe to increment
4490 ++new->current_threshold;
4496 /* Swap primary and spare array */
4497 thresholds->spare = thresholds->primary;
4498 rcu_assign_pointer(thresholds->primary, new);
4500 /* To be sure that nobody uses thresholds */
4503 mutex_unlock(&memcg->thresholds_lock);
4506 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4507 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4509 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4510 struct mem_cgroup_eventfd_list *event;
4511 int type = MEMFILE_TYPE(cft->private);
4513 BUG_ON(type != _OOM_TYPE);
4514 event = kmalloc(sizeof(*event), GFP_KERNEL);
4518 spin_lock(&memcg_oom_lock);
4520 event->eventfd = eventfd;
4521 list_add(&event->list, &memcg->oom_notify);
4523 /* already in OOM ? */
4524 if (atomic_read(&memcg->under_oom))
4525 eventfd_signal(eventfd, 1);
4526 spin_unlock(&memcg_oom_lock);
4531 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4532 struct cftype *cft, struct eventfd_ctx *eventfd)
4534 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4535 struct mem_cgroup_eventfd_list *ev, *tmp;
4536 int type = MEMFILE_TYPE(cft->private);
4538 BUG_ON(type != _OOM_TYPE);
4540 spin_lock(&memcg_oom_lock);
4542 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4543 if (ev->eventfd == eventfd) {
4544 list_del(&ev->list);
4549 spin_unlock(&memcg_oom_lock);
4552 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4553 struct cftype *cft, struct cgroup_map_cb *cb)
4555 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4557 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4559 if (atomic_read(&memcg->under_oom))
4560 cb->fill(cb, "under_oom", 1);
4562 cb->fill(cb, "under_oom", 0);
4566 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4567 struct cftype *cft, u64 val)
4569 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4570 struct mem_cgroup *parent;
4572 /* cannot set to root cgroup and only 0 and 1 are allowed */
4573 if (!cgrp->parent || !((val == 0) || (val == 1)))
4576 parent = mem_cgroup_from_cont(cgrp->parent);
4579 /* oom-kill-disable is a flag for subhierarchy. */
4580 if ((parent->use_hierarchy) ||
4581 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4585 memcg->oom_kill_disable = val;
4587 memcg_oom_recover(memcg);
4593 static const struct file_operations mem_control_numa_stat_file_operations = {
4595 .llseek = seq_lseek,
4596 .release = single_release,
4599 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4601 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4603 file->f_op = &mem_control_numa_stat_file_operations;
4604 return single_open(file, mem_control_numa_stat_show, cont);
4606 #endif /* CONFIG_NUMA */
4608 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_KMEM
4609 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4612 * Part of this would be better living in a separate allocation
4613 * function, leaving us with just the cgroup tree population work.
4614 * We, however, depend on state such as network's proto_list that
4615 * is only initialized after cgroup creation. I found the less
4616 * cumbersome way to deal with it to defer it all to populate time
4618 return mem_cgroup_sockets_init(cont, ss);
4621 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4622 struct cgroup *cont)
4624 mem_cgroup_sockets_destroy(cont, ss);
4627 static int register_kmem_files(struct cgroup *cont, struct cgroup_subsys *ss)
4632 static void kmem_cgroup_destroy(struct cgroup_subsys *ss,
4633 struct cgroup *cont)
4638 static struct cftype mem_cgroup_files[] = {
4640 .name = "usage_in_bytes",
4641 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4642 .read_u64 = mem_cgroup_read,
4643 .register_event = mem_cgroup_usage_register_event,
4644 .unregister_event = mem_cgroup_usage_unregister_event,
4647 .name = "max_usage_in_bytes",
4648 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4649 .trigger = mem_cgroup_reset,
4650 .read_u64 = mem_cgroup_read,
4653 .name = "limit_in_bytes",
4654 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4655 .write_string = mem_cgroup_write,
4656 .read_u64 = mem_cgroup_read,
4659 .name = "soft_limit_in_bytes",
4660 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4661 .write_string = mem_cgroup_write,
4662 .read_u64 = mem_cgroup_read,
4666 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4667 .trigger = mem_cgroup_reset,
4668 .read_u64 = mem_cgroup_read,
4672 .read_map = mem_control_stat_show,
4675 .name = "force_empty",
4676 .trigger = mem_cgroup_force_empty_write,
4679 .name = "use_hierarchy",
4680 .write_u64 = mem_cgroup_hierarchy_write,
4681 .read_u64 = mem_cgroup_hierarchy_read,
4684 .name = "swappiness",
4685 .read_u64 = mem_cgroup_swappiness_read,
4686 .write_u64 = mem_cgroup_swappiness_write,
4689 .name = "move_charge_at_immigrate",
4690 .read_u64 = mem_cgroup_move_charge_read,
4691 .write_u64 = mem_cgroup_move_charge_write,
4694 .name = "oom_control",
4695 .read_map = mem_cgroup_oom_control_read,
4696 .write_u64 = mem_cgroup_oom_control_write,
4697 .register_event = mem_cgroup_oom_register_event,
4698 .unregister_event = mem_cgroup_oom_unregister_event,
4699 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4703 .name = "numa_stat",
4704 .open = mem_control_numa_stat_open,
4710 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4711 static struct cftype memsw_cgroup_files[] = {
4713 .name = "memsw.usage_in_bytes",
4714 .private = MEMFILE_PRIVATE(_MEMSWAP, 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 = "memsw.max_usage_in_bytes",
4721 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4722 .trigger = mem_cgroup_reset,
4723 .read_u64 = mem_cgroup_read,
4726 .name = "memsw.limit_in_bytes",
4727 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4728 .write_string = mem_cgroup_write,
4729 .read_u64 = mem_cgroup_read,
4732 .name = "memsw.failcnt",
4733 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4734 .trigger = mem_cgroup_reset,
4735 .read_u64 = mem_cgroup_read,
4739 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4741 if (!do_swap_account)
4743 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4744 ARRAY_SIZE(memsw_cgroup_files));
4747 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4753 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4755 struct mem_cgroup_per_node *pn;
4756 struct mem_cgroup_per_zone *mz;
4758 int zone, tmp = node;
4760 * This routine is called against possible nodes.
4761 * But it's BUG to call kmalloc() against offline node.
4763 * TODO: this routine can waste much memory for nodes which will
4764 * never be onlined. It's better to use memory hotplug callback
4767 if (!node_state(node, N_NORMAL_MEMORY))
4769 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4773 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4774 mz = &pn->zoneinfo[zone];
4776 INIT_LIST_HEAD(&mz->lruvec.lists[l]);
4777 mz->usage_in_excess = 0;
4778 mz->on_tree = false;
4781 memcg->info.nodeinfo[node] = pn;
4785 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4787 kfree(memcg->info.nodeinfo[node]);
4790 static struct mem_cgroup *mem_cgroup_alloc(void)
4792 struct mem_cgroup *mem;
4793 int size = sizeof(struct mem_cgroup);
4795 /* Can be very big if MAX_NUMNODES is very big */
4796 if (size < PAGE_SIZE)
4797 mem = kzalloc(size, GFP_KERNEL);
4799 mem = vzalloc(size);
4804 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4807 spin_lock_init(&mem->pcp_counter_lock);
4811 if (size < PAGE_SIZE)
4819 * At destroying mem_cgroup, references from swap_cgroup can remain.
4820 * (scanning all at force_empty is too costly...)
4822 * Instead of clearing all references at force_empty, we remember
4823 * the number of reference from swap_cgroup and free mem_cgroup when
4824 * it goes down to 0.
4826 * Removal of cgroup itself succeeds regardless of refs from swap.
4829 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4833 mem_cgroup_remove_from_trees(memcg);
4834 free_css_id(&mem_cgroup_subsys, &memcg->css);
4836 for_each_node_state(node, N_POSSIBLE)
4837 free_mem_cgroup_per_zone_info(memcg, node);
4839 free_percpu(memcg->stat);
4840 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4846 static void mem_cgroup_get(struct mem_cgroup *memcg)
4848 atomic_inc(&memcg->refcnt);
4851 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4853 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4854 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4855 __mem_cgroup_free(memcg);
4857 mem_cgroup_put(parent);
4861 static void mem_cgroup_put(struct mem_cgroup *memcg)
4863 __mem_cgroup_put(memcg, 1);
4867 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4869 struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4871 if (!memcg->res.parent)
4873 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4875 EXPORT_SYMBOL(parent_mem_cgroup);
4877 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4878 static void __init enable_swap_cgroup(void)
4880 if (!mem_cgroup_disabled() && really_do_swap_account)
4881 do_swap_account = 1;
4884 static void __init enable_swap_cgroup(void)
4889 static int mem_cgroup_soft_limit_tree_init(void)
4891 struct mem_cgroup_tree_per_node *rtpn;
4892 struct mem_cgroup_tree_per_zone *rtpz;
4893 int tmp, node, zone;
4895 for_each_node_state(node, N_POSSIBLE) {
4897 if (!node_state(node, N_NORMAL_MEMORY))
4899 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4903 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4905 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4906 rtpz = &rtpn->rb_tree_per_zone[zone];
4907 rtpz->rb_root = RB_ROOT;
4908 spin_lock_init(&rtpz->lock);
4914 for_each_node_state(node, N_POSSIBLE) {
4915 if (!soft_limit_tree.rb_tree_per_node[node])
4917 kfree(soft_limit_tree.rb_tree_per_node[node]);
4918 soft_limit_tree.rb_tree_per_node[node] = NULL;
4924 static struct cgroup_subsys_state * __ref
4925 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4927 struct mem_cgroup *memcg, *parent;
4928 long error = -ENOMEM;
4931 memcg = mem_cgroup_alloc();
4933 return ERR_PTR(error);
4935 for_each_node_state(node, N_POSSIBLE)
4936 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4940 if (cont->parent == NULL) {
4942 enable_swap_cgroup();
4944 if (mem_cgroup_soft_limit_tree_init())
4946 root_mem_cgroup = memcg;
4947 for_each_possible_cpu(cpu) {
4948 struct memcg_stock_pcp *stock =
4949 &per_cpu(memcg_stock, cpu);
4950 INIT_WORK(&stock->work, drain_local_stock);
4952 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4954 parent = mem_cgroup_from_cont(cont->parent);
4955 memcg->use_hierarchy = parent->use_hierarchy;
4956 memcg->oom_kill_disable = parent->oom_kill_disable;
4959 if (parent && parent->use_hierarchy) {
4960 res_counter_init(&memcg->res, &parent->res);
4961 res_counter_init(&memcg->memsw, &parent->memsw);
4963 * We increment refcnt of the parent to ensure that we can
4964 * safely access it on res_counter_charge/uncharge.
4965 * This refcnt will be decremented when freeing this
4966 * mem_cgroup(see mem_cgroup_put).
4968 mem_cgroup_get(parent);
4970 res_counter_init(&memcg->res, NULL);
4971 res_counter_init(&memcg->memsw, NULL);
4973 memcg->last_scanned_node = MAX_NUMNODES;
4974 INIT_LIST_HEAD(&memcg->oom_notify);
4977 memcg->swappiness = mem_cgroup_swappiness(parent);
4978 atomic_set(&memcg->refcnt, 1);
4979 memcg->move_charge_at_immigrate = 0;
4980 mutex_init(&memcg->thresholds_lock);
4983 __mem_cgroup_free(memcg);
4984 return ERR_PTR(error);
4987 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4988 struct cgroup *cont)
4990 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4992 return mem_cgroup_force_empty(memcg, false);
4995 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4996 struct cgroup *cont)
4998 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5000 kmem_cgroup_destroy(ss, cont);
5002 mem_cgroup_put(memcg);
5005 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5006 struct cgroup *cont)
5010 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5011 ARRAY_SIZE(mem_cgroup_files));
5014 ret = register_memsw_files(cont, ss);
5017 ret = register_kmem_files(cont, ss);
5023 /* Handlers for move charge at task migration. */
5024 #define PRECHARGE_COUNT_AT_ONCE 256
5025 static int mem_cgroup_do_precharge(unsigned long count)
5028 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5029 struct mem_cgroup *memcg = mc.to;
5031 if (mem_cgroup_is_root(memcg)) {
5032 mc.precharge += count;
5033 /* we don't need css_get for root */
5036 /* try to charge at once */
5038 struct res_counter *dummy;
5040 * "memcg" cannot be under rmdir() because we've already checked
5041 * by cgroup_lock_live_cgroup() that it is not removed and we
5042 * are still under the same cgroup_mutex. So we can postpone
5045 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5047 if (do_swap_account && res_counter_charge(&memcg->memsw,
5048 PAGE_SIZE * count, &dummy)) {
5049 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5052 mc.precharge += count;
5056 /* fall back to one by one charge */
5058 if (signal_pending(current)) {
5062 if (!batch_count--) {
5063 batch_count = PRECHARGE_COUNT_AT_ONCE;
5066 ret = __mem_cgroup_try_charge(NULL,
5067 GFP_KERNEL, 1, &memcg, false);
5069 /* mem_cgroup_clear_mc() will do uncharge later */
5077 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5078 * @vma: the vma the pte to be checked belongs
5079 * @addr: the address corresponding to the pte to be checked
5080 * @ptent: the pte to be checked
5081 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5084 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5085 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5086 * move charge. if @target is not NULL, the page is stored in target->page
5087 * with extra refcnt got(Callers should handle it).
5088 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5089 * target for charge migration. if @target is not NULL, the entry is stored
5092 * Called with pte lock held.
5099 enum mc_target_type {
5100 MC_TARGET_NONE, /* not used */
5105 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5106 unsigned long addr, pte_t ptent)
5108 struct page *page = vm_normal_page(vma, addr, ptent);
5110 if (!page || !page_mapped(page))
5112 if (PageAnon(page)) {
5113 /* we don't move shared anon */
5114 if (!move_anon() || page_mapcount(page) > 2)
5116 } else if (!move_file())
5117 /* we ignore mapcount for file pages */
5119 if (!get_page_unless_zero(page))
5125 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5126 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5129 struct page *page = NULL;
5130 swp_entry_t ent = pte_to_swp_entry(ptent);
5132 if (!move_anon() || non_swap_entry(ent))
5134 usage_count = mem_cgroup_count_swap_user(ent, &page);
5135 if (usage_count > 1) { /* we don't move shared anon */
5140 if (do_swap_account)
5141 entry->val = ent.val;
5146 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5147 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5149 struct page *page = NULL;
5150 struct inode *inode;
5151 struct address_space *mapping;
5154 if (!vma->vm_file) /* anonymous vma */
5159 inode = vma->vm_file->f_path.dentry->d_inode;
5160 mapping = vma->vm_file->f_mapping;
5161 if (pte_none(ptent))
5162 pgoff = linear_page_index(vma, addr);
5163 else /* pte_file(ptent) is true */
5164 pgoff = pte_to_pgoff(ptent);
5166 /* page is moved even if it's not RSS of this task(page-faulted). */
5167 page = find_get_page(mapping, pgoff);
5170 /* shmem/tmpfs may report page out on swap: account for that too. */
5171 if (radix_tree_exceptional_entry(page)) {
5172 swp_entry_t swap = radix_to_swp_entry(page);
5173 if (do_swap_account)
5175 page = find_get_page(&swapper_space, swap.val);
5181 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5182 unsigned long addr, pte_t ptent, union mc_target *target)
5184 struct page *page = NULL;
5185 struct page_cgroup *pc;
5187 swp_entry_t ent = { .val = 0 };
5189 if (pte_present(ptent))
5190 page = mc_handle_present_pte(vma, addr, ptent);
5191 else if (is_swap_pte(ptent))
5192 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5193 else if (pte_none(ptent) || pte_file(ptent))
5194 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5196 if (!page && !ent.val)
5199 pc = lookup_page_cgroup(page);
5201 * Do only loose check w/o page_cgroup lock.
5202 * mem_cgroup_move_account() checks the pc is valid or not under
5205 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5206 ret = MC_TARGET_PAGE;
5208 target->page = page;
5210 if (!ret || !target)
5213 /* There is a swap entry and a page doesn't exist or isn't charged */
5214 if (ent.val && !ret &&
5215 css_id(&mc.from->css) == lookup_swap_cgroup_id(ent)) {
5216 ret = MC_TARGET_SWAP;
5223 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5224 unsigned long addr, unsigned long end,
5225 struct mm_walk *walk)
5227 struct vm_area_struct *vma = walk->private;
5231 split_huge_page_pmd(walk->mm, pmd);
5233 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5234 for (; addr != end; pte++, addr += PAGE_SIZE)
5235 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5236 mc.precharge++; /* increment precharge temporarily */
5237 pte_unmap_unlock(pte - 1, ptl);
5243 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5245 unsigned long precharge;
5246 struct vm_area_struct *vma;
5248 down_read(&mm->mmap_sem);
5249 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5250 struct mm_walk mem_cgroup_count_precharge_walk = {
5251 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5255 if (is_vm_hugetlb_page(vma))
5257 walk_page_range(vma->vm_start, vma->vm_end,
5258 &mem_cgroup_count_precharge_walk);
5260 up_read(&mm->mmap_sem);
5262 precharge = mc.precharge;
5268 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5270 unsigned long precharge = mem_cgroup_count_precharge(mm);
5272 VM_BUG_ON(mc.moving_task);
5273 mc.moving_task = current;
5274 return mem_cgroup_do_precharge(precharge);
5277 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5278 static void __mem_cgroup_clear_mc(void)
5280 struct mem_cgroup *from = mc.from;
5281 struct mem_cgroup *to = mc.to;
5283 /* we must uncharge all the leftover precharges from mc.to */
5285 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5289 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5290 * we must uncharge here.
5292 if (mc.moved_charge) {
5293 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5294 mc.moved_charge = 0;
5296 /* we must fixup refcnts and charges */
5297 if (mc.moved_swap) {
5298 /* uncharge swap account from the old cgroup */
5299 if (!mem_cgroup_is_root(mc.from))
5300 res_counter_uncharge(&mc.from->memsw,
5301 PAGE_SIZE * mc.moved_swap);
5302 __mem_cgroup_put(mc.from, mc.moved_swap);
5304 if (!mem_cgroup_is_root(mc.to)) {
5306 * we charged both to->res and to->memsw, so we should
5309 res_counter_uncharge(&mc.to->res,
5310 PAGE_SIZE * mc.moved_swap);
5312 /* we've already done mem_cgroup_get(mc.to) */
5315 memcg_oom_recover(from);
5316 memcg_oom_recover(to);
5317 wake_up_all(&mc.waitq);
5320 static void mem_cgroup_clear_mc(void)
5322 struct mem_cgroup *from = mc.from;
5325 * we must clear moving_task before waking up waiters at the end of
5328 mc.moving_task = NULL;
5329 __mem_cgroup_clear_mc();
5330 spin_lock(&mc.lock);
5333 spin_unlock(&mc.lock);
5334 mem_cgroup_end_move(from);
5337 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5338 struct cgroup *cgroup,
5339 struct cgroup_taskset *tset)
5341 struct task_struct *p = cgroup_taskset_first(tset);
5343 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5345 if (memcg->move_charge_at_immigrate) {
5346 struct mm_struct *mm;
5347 struct mem_cgroup *from = mem_cgroup_from_task(p);
5349 VM_BUG_ON(from == memcg);
5351 mm = get_task_mm(p);
5354 /* We move charges only when we move a owner of the mm */
5355 if (mm->owner == p) {
5358 VM_BUG_ON(mc.precharge);
5359 VM_BUG_ON(mc.moved_charge);
5360 VM_BUG_ON(mc.moved_swap);
5361 mem_cgroup_start_move(from);
5362 spin_lock(&mc.lock);
5365 spin_unlock(&mc.lock);
5366 /* We set mc.moving_task later */
5368 ret = mem_cgroup_precharge_mc(mm);
5370 mem_cgroup_clear_mc();
5377 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5378 struct cgroup *cgroup,
5379 struct cgroup_taskset *tset)
5381 mem_cgroup_clear_mc();
5384 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5385 unsigned long addr, unsigned long end,
5386 struct mm_walk *walk)
5389 struct vm_area_struct *vma = walk->private;
5393 split_huge_page_pmd(walk->mm, pmd);
5395 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5396 for (; addr != end; addr += PAGE_SIZE) {
5397 pte_t ptent = *(pte++);
5398 union mc_target target;
5401 struct page_cgroup *pc;
5407 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5409 case MC_TARGET_PAGE:
5411 if (isolate_lru_page(page))
5413 pc = lookup_page_cgroup(page);
5414 if (!mem_cgroup_move_account(page, 1, pc,
5415 mc.from, mc.to, false)) {
5417 /* we uncharge from mc.from later. */
5420 putback_lru_page(page);
5421 put: /* is_target_pte_for_mc() gets the page */
5424 case MC_TARGET_SWAP:
5426 if (!mem_cgroup_move_swap_account(ent,
5427 mc.from, mc.to, false)) {
5429 /* we fixup refcnts and charges later. */
5437 pte_unmap_unlock(pte - 1, ptl);
5442 * We have consumed all precharges we got in can_attach().
5443 * We try charge one by one, but don't do any additional
5444 * charges to mc.to if we have failed in charge once in attach()
5447 ret = mem_cgroup_do_precharge(1);
5455 static void mem_cgroup_move_charge(struct mm_struct *mm)
5457 struct vm_area_struct *vma;
5459 lru_add_drain_all();
5461 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5463 * Someone who are holding the mmap_sem might be waiting in
5464 * waitq. So we cancel all extra charges, wake up all waiters,
5465 * and retry. Because we cancel precharges, we might not be able
5466 * to move enough charges, but moving charge is a best-effort
5467 * feature anyway, so it wouldn't be a big problem.
5469 __mem_cgroup_clear_mc();
5473 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5475 struct mm_walk mem_cgroup_move_charge_walk = {
5476 .pmd_entry = mem_cgroup_move_charge_pte_range,
5480 if (is_vm_hugetlb_page(vma))
5482 ret = walk_page_range(vma->vm_start, vma->vm_end,
5483 &mem_cgroup_move_charge_walk);
5486 * means we have consumed all precharges and failed in
5487 * doing additional charge. Just abandon here.
5491 up_read(&mm->mmap_sem);
5494 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5495 struct cgroup *cont,
5496 struct cgroup_taskset *tset)
5498 struct task_struct *p = cgroup_taskset_first(tset);
5499 struct mm_struct *mm = get_task_mm(p);
5503 mem_cgroup_move_charge(mm);
5508 mem_cgroup_clear_mc();
5510 #else /* !CONFIG_MMU */
5511 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5512 struct cgroup *cgroup,
5513 struct cgroup_taskset *tset)
5517 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5518 struct cgroup *cgroup,
5519 struct cgroup_taskset *tset)
5522 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5523 struct cgroup *cont,
5524 struct cgroup_taskset *tset)
5529 struct cgroup_subsys mem_cgroup_subsys = {
5531 .subsys_id = mem_cgroup_subsys_id,
5532 .create = mem_cgroup_create,
5533 .pre_destroy = mem_cgroup_pre_destroy,
5534 .destroy = mem_cgroup_destroy,
5535 .populate = mem_cgroup_populate,
5536 .can_attach = mem_cgroup_can_attach,
5537 .cancel_attach = mem_cgroup_cancel_attach,
5538 .attach = mem_cgroup_move_task,
5543 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5544 static int __init enable_swap_account(char *s)
5546 /* consider enabled if no parameter or 1 is given */
5547 if (!strcmp(s, "1"))
5548 really_do_swap_account = 1;
5549 else if (!strcmp(s, "0"))
5550 really_do_swap_account = 0;
5553 __setup("swapaccount=", enable_swap_account);