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 <asm/uaccess.h>
56 #include <trace/events/vmscan.h>
58 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
59 #define MEM_CGROUP_RECLAIM_RETRIES 5
60 struct mem_cgroup *root_mem_cgroup __read_mostly;
62 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
63 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
64 int do_swap_account __read_mostly;
66 /* for remember boot option*/
67 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
68 static int really_do_swap_account __initdata = 1;
70 static int really_do_swap_account __initdata = 0;
74 #define do_swap_account (0)
79 * Statistics for memory cgroup.
81 enum mem_cgroup_stat_index {
83 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
85 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
86 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
87 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
88 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
89 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
90 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */
91 MEM_CGROUP_STAT_NSTATS,
94 enum mem_cgroup_events_index {
95 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
96 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
97 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
98 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
99 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
100 MEM_CGROUP_EVENTS_NSTATS,
103 * Per memcg event counter is incremented at every pagein/pageout. With THP,
104 * it will be incremated by the number of pages. This counter is used for
105 * for trigger some periodic events. This is straightforward and better
106 * than using jiffies etc. to handle periodic memcg event.
108 enum mem_cgroup_events_target {
109 MEM_CGROUP_TARGET_THRESH,
110 MEM_CGROUP_TARGET_SOFTLIMIT,
111 MEM_CGROUP_TARGET_NUMAINFO,
114 #define THRESHOLDS_EVENTS_TARGET (128)
115 #define SOFTLIMIT_EVENTS_TARGET (1024)
116 #define NUMAINFO_EVENTS_TARGET (1024)
118 struct mem_cgroup_stat_cpu {
119 long count[MEM_CGROUP_STAT_NSTATS];
120 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
121 unsigned long targets[MEM_CGROUP_NTARGETS];
125 * per-zone information in memory controller.
127 struct mem_cgroup_per_zone {
129 * spin_lock to protect the per cgroup LRU
131 struct list_head lists[NR_LRU_LISTS];
132 unsigned long count[NR_LRU_LISTS];
134 struct zone_reclaim_stat reclaim_stat;
135 struct rb_node tree_node; /* RB tree node */
136 unsigned long long usage_in_excess;/* Set to the value by which */
137 /* the soft limit is exceeded*/
139 struct mem_cgroup *mem; /* Back pointer, we cannot */
140 /* use container_of */
142 /* Macro for accessing counter */
143 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
145 struct mem_cgroup_per_node {
146 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
149 struct mem_cgroup_lru_info {
150 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
154 * Cgroups above their limits are maintained in a RB-Tree, independent of
155 * their hierarchy representation
158 struct mem_cgroup_tree_per_zone {
159 struct rb_root rb_root;
163 struct mem_cgroup_tree_per_node {
164 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
167 struct mem_cgroup_tree {
168 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
171 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
173 struct mem_cgroup_threshold {
174 struct eventfd_ctx *eventfd;
179 struct mem_cgroup_threshold_ary {
180 /* An array index points to threshold just below usage. */
181 int current_threshold;
182 /* Size of entries[] */
184 /* Array of thresholds */
185 struct mem_cgroup_threshold entries[0];
188 struct mem_cgroup_thresholds {
189 /* Primary thresholds array */
190 struct mem_cgroup_threshold_ary *primary;
192 * Spare threshold array.
193 * This is needed to make mem_cgroup_unregister_event() "never fail".
194 * It must be able to store at least primary->size - 1 entries.
196 struct mem_cgroup_threshold_ary *spare;
200 struct mem_cgroup_eventfd_list {
201 struct list_head list;
202 struct eventfd_ctx *eventfd;
205 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
206 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
212 SCAN_BY_SHRINK, /* not recorded now */
231 unsigned long stats[NR_SCAN_CONTEXT][NR_SCANSTATS];
232 unsigned long rootstats[NR_SCAN_CONTEXT][NR_SCANSTATS];
235 const char *scanstat_string[NR_SCANSTATS] = {
237 "scanned_anon_pages",
238 "scanned_file_pages",
240 "rotated_anon_pages",
241 "rotated_file_pages",
247 #define SCANSTAT_WORD_LIMIT "_by_limit"
248 #define SCANSTAT_WORD_SYSTEM "_by_system"
249 #define SCANSTAT_WORD_HIERARCHY "_under_hierarchy"
253 * The memory controller data structure. The memory controller controls both
254 * page cache and RSS per cgroup. We would eventually like to provide
255 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
256 * to help the administrator determine what knobs to tune.
258 * TODO: Add a water mark for the memory controller. Reclaim will begin when
259 * we hit the water mark. May be even add a low water mark, such that
260 * no reclaim occurs from a cgroup at it's low water mark, this is
261 * a feature that will be implemented much later in the future.
264 struct cgroup_subsys_state css;
266 * the counter to account for memory usage
268 struct res_counter res;
270 * the counter to account for mem+swap usage.
272 struct res_counter memsw;
274 * Per cgroup active and inactive list, similar to the
275 * per zone LRU lists.
277 struct mem_cgroup_lru_info info;
279 * While reclaiming in a hierarchy, we cache the last child we
282 int last_scanned_child;
283 int last_scanned_node;
285 nodemask_t scan_nodes;
286 atomic_t numainfo_events;
287 atomic_t numainfo_updating;
290 * Should the accounting and control be hierarchical, per subtree?
300 /* OOM-Killer disable */
301 int oom_kill_disable;
303 /* set when res.limit == memsw.limit */
304 bool memsw_is_minimum;
306 /* protect arrays of thresholds */
307 struct mutex thresholds_lock;
309 /* thresholds for memory usage. RCU-protected */
310 struct mem_cgroup_thresholds thresholds;
312 /* thresholds for mem+swap usage. RCU-protected */
313 struct mem_cgroup_thresholds memsw_thresholds;
315 /* For oom notifier event fd */
316 struct list_head oom_notify;
317 /* For recording LRU-scan statistics */
318 struct scanstat scanstat;
320 * Should we move charges of a task when a task is moved into this
321 * mem_cgroup ? And what type of charges should we move ?
323 unsigned long move_charge_at_immigrate;
327 struct mem_cgroup_stat_cpu *stat;
329 * used when a cpu is offlined or other synchronizations
330 * See mem_cgroup_read_stat().
332 struct mem_cgroup_stat_cpu nocpu_base;
333 spinlock_t pcp_counter_lock;
336 /* Stuffs for move charges at task migration. */
338 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
339 * left-shifted bitmap of these types.
342 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
343 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
347 /* "mc" and its members are protected by cgroup_mutex */
348 static struct move_charge_struct {
349 spinlock_t lock; /* for from, to */
350 struct mem_cgroup *from;
351 struct mem_cgroup *to;
352 unsigned long precharge;
353 unsigned long moved_charge;
354 unsigned long moved_swap;
355 struct task_struct *moving_task; /* a task moving charges */
356 wait_queue_head_t waitq; /* a waitq for other context */
358 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
359 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
362 static bool move_anon(void)
364 return test_bit(MOVE_CHARGE_TYPE_ANON,
365 &mc.to->move_charge_at_immigrate);
368 static bool move_file(void)
370 return test_bit(MOVE_CHARGE_TYPE_FILE,
371 &mc.to->move_charge_at_immigrate);
375 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
376 * limit reclaim to prevent infinite loops, if they ever occur.
378 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
379 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
382 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
383 MEM_CGROUP_CHARGE_TYPE_MAPPED,
384 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
385 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
386 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
387 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
391 /* for encoding cft->private value on file */
394 #define _OOM_TYPE (2)
395 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
396 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
397 #define MEMFILE_ATTR(val) ((val) & 0xffff)
398 /* Used for OOM nofiier */
399 #define OOM_CONTROL (0)
402 * Reclaim flags for mem_cgroup_hierarchical_reclaim
404 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
405 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
406 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
407 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
408 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
409 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
411 static void mem_cgroup_get(struct mem_cgroup *memcg);
412 static void mem_cgroup_put(struct mem_cgroup *memcg);
413 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg);
414 static void drain_all_stock_async(struct mem_cgroup *memcg);
416 static struct mem_cgroup_per_zone *
417 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
419 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
422 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
427 static struct mem_cgroup_per_zone *
428 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
430 int nid = page_to_nid(page);
431 int zid = page_zonenum(page);
433 return mem_cgroup_zoneinfo(memcg, nid, zid);
436 static struct mem_cgroup_tree_per_zone *
437 soft_limit_tree_node_zone(int nid, int zid)
439 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
442 static struct mem_cgroup_tree_per_zone *
443 soft_limit_tree_from_page(struct page *page)
445 int nid = page_to_nid(page);
446 int zid = page_zonenum(page);
448 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
452 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
453 struct mem_cgroup_per_zone *mz,
454 struct mem_cgroup_tree_per_zone *mctz,
455 unsigned long long new_usage_in_excess)
457 struct rb_node **p = &mctz->rb_root.rb_node;
458 struct rb_node *parent = NULL;
459 struct mem_cgroup_per_zone *mz_node;
464 mz->usage_in_excess = new_usage_in_excess;
465 if (!mz->usage_in_excess)
469 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
471 if (mz->usage_in_excess < mz_node->usage_in_excess)
474 * We can't avoid mem cgroups that are over their soft
475 * limit by the same amount
477 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
480 rb_link_node(&mz->tree_node, parent, p);
481 rb_insert_color(&mz->tree_node, &mctz->rb_root);
486 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
487 struct mem_cgroup_per_zone *mz,
488 struct mem_cgroup_tree_per_zone *mctz)
492 rb_erase(&mz->tree_node, &mctz->rb_root);
497 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
498 struct mem_cgroup_per_zone *mz,
499 struct mem_cgroup_tree_per_zone *mctz)
501 spin_lock(&mctz->lock);
502 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
503 spin_unlock(&mctz->lock);
507 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
509 unsigned long long excess;
510 struct mem_cgroup_per_zone *mz;
511 struct mem_cgroup_tree_per_zone *mctz;
512 int nid = page_to_nid(page);
513 int zid = page_zonenum(page);
514 mctz = soft_limit_tree_from_page(page);
517 * Necessary to update all ancestors when hierarchy is used.
518 * because their event counter is not touched.
520 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
521 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
522 excess = res_counter_soft_limit_excess(&memcg->res);
524 * We have to update the tree if mz is on RB-tree or
525 * mem is over its softlimit.
527 if (excess || mz->on_tree) {
528 spin_lock(&mctz->lock);
529 /* if on-tree, remove it */
531 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
533 * Insert again. mz->usage_in_excess will be updated.
534 * If excess is 0, no tree ops.
536 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
537 spin_unlock(&mctz->lock);
542 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
545 struct mem_cgroup_per_zone *mz;
546 struct mem_cgroup_tree_per_zone *mctz;
548 for_each_node_state(node, N_POSSIBLE) {
549 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
550 mz = mem_cgroup_zoneinfo(memcg, node, zone);
551 mctz = soft_limit_tree_node_zone(node, zone);
552 mem_cgroup_remove_exceeded(memcg, mz, mctz);
557 static struct mem_cgroup_per_zone *
558 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
560 struct rb_node *rightmost = NULL;
561 struct mem_cgroup_per_zone *mz;
565 rightmost = rb_last(&mctz->rb_root);
567 goto done; /* Nothing to reclaim from */
569 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
571 * Remove the node now but someone else can add it back,
572 * we will to add it back at the end of reclaim to its correct
573 * position in the tree.
575 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
576 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
577 !css_tryget(&mz->mem->css))
583 static struct mem_cgroup_per_zone *
584 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
586 struct mem_cgroup_per_zone *mz;
588 spin_lock(&mctz->lock);
589 mz = __mem_cgroup_largest_soft_limit_node(mctz);
590 spin_unlock(&mctz->lock);
595 * Implementation Note: reading percpu statistics for memcg.
597 * Both of vmstat[] and percpu_counter has threshold and do periodic
598 * synchronization to implement "quick" read. There are trade-off between
599 * reading cost and precision of value. Then, we may have a chance to implement
600 * a periodic synchronizion of counter in memcg's counter.
602 * But this _read() function is used for user interface now. The user accounts
603 * memory usage by memory cgroup and he _always_ requires exact value because
604 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
605 * have to visit all online cpus and make sum. So, for now, unnecessary
606 * synchronization is not implemented. (just implemented for cpu hotplug)
608 * If there are kernel internal actions which can make use of some not-exact
609 * value, and reading all cpu value can be performance bottleneck in some
610 * common workload, threashold and synchonization as vmstat[] should be
613 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
614 enum mem_cgroup_stat_index idx)
620 for_each_online_cpu(cpu)
621 val += per_cpu(memcg->stat->count[idx], cpu);
622 #ifdef CONFIG_HOTPLUG_CPU
623 spin_lock(&memcg->pcp_counter_lock);
624 val += memcg->nocpu_base.count[idx];
625 spin_unlock(&memcg->pcp_counter_lock);
631 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
634 int val = (charge) ? 1 : -1;
635 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
638 void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val)
640 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
643 void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val)
645 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
648 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
649 enum mem_cgroup_events_index idx)
651 unsigned long val = 0;
654 for_each_online_cpu(cpu)
655 val += per_cpu(memcg->stat->events[idx], cpu);
656 #ifdef CONFIG_HOTPLUG_CPU
657 spin_lock(&memcg->pcp_counter_lock);
658 val += memcg->nocpu_base.events[idx];
659 spin_unlock(&memcg->pcp_counter_lock);
664 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
665 bool file, int nr_pages)
670 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
673 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
676 /* pagein of a big page is an event. So, ignore page size */
678 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
680 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
681 nr_pages = -nr_pages; /* for event */
684 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
690 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
691 unsigned int lru_mask)
693 struct mem_cgroup_per_zone *mz;
695 unsigned long ret = 0;
697 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
700 if (BIT(l) & lru_mask)
701 ret += MEM_CGROUP_ZSTAT(mz, l);
707 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
708 int nid, unsigned int lru_mask)
713 for (zid = 0; zid < MAX_NR_ZONES; zid++)
714 total += mem_cgroup_zone_nr_lru_pages(memcg,
720 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
721 unsigned int lru_mask)
726 for_each_node_state(nid, N_HIGH_MEMORY)
727 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
731 static bool __memcg_event_check(struct mem_cgroup *memcg, int target)
733 unsigned long val, next;
735 val = this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
736 next = this_cpu_read(memcg->stat->targets[target]);
737 /* from time_after() in jiffies.h */
738 return ((long)next - (long)val < 0);
741 static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
743 unsigned long val, next;
745 val = this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
748 case MEM_CGROUP_TARGET_THRESH:
749 next = val + THRESHOLDS_EVENTS_TARGET;
751 case MEM_CGROUP_TARGET_SOFTLIMIT:
752 next = val + SOFTLIMIT_EVENTS_TARGET;
754 case MEM_CGROUP_TARGET_NUMAINFO:
755 next = val + NUMAINFO_EVENTS_TARGET;
761 this_cpu_write(memcg->stat->targets[target], next);
765 * Check events in order.
768 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
770 /* threshold event is triggered in finer grain than soft limit */
771 if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
772 mem_cgroup_threshold(memcg);
773 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
774 if (unlikely(__memcg_event_check(memcg,
775 MEM_CGROUP_TARGET_SOFTLIMIT))) {
776 mem_cgroup_update_tree(memcg, page);
777 __mem_cgroup_target_update(memcg,
778 MEM_CGROUP_TARGET_SOFTLIMIT);
781 if (unlikely(__memcg_event_check(memcg,
782 MEM_CGROUP_TARGET_NUMAINFO))) {
783 atomic_inc(&memcg->numainfo_events);
784 __mem_cgroup_target_update(memcg,
785 MEM_CGROUP_TARGET_NUMAINFO);
791 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
793 return container_of(cgroup_subsys_state(cont,
794 mem_cgroup_subsys_id), struct mem_cgroup,
798 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
801 * mm_update_next_owner() may clear mm->owner to NULL
802 * if it races with swapoff, page migration, etc.
803 * So this can be called with p == NULL.
808 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
809 struct mem_cgroup, css);
812 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
814 struct mem_cgroup *memcg = NULL;
819 * Because we have no locks, mm->owner's may be being moved to other
820 * cgroup. We use css_tryget() here even if this looks
821 * pessimistic (rather than adding locks here).
825 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
826 if (unlikely(!memcg))
828 } while (!css_tryget(&memcg->css));
833 /* The caller has to guarantee "mem" exists before calling this */
834 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *memcg)
836 struct cgroup_subsys_state *css;
839 if (!memcg) /* ROOT cgroup has the smallest ID */
840 return root_mem_cgroup; /*css_put/get against root is ignored*/
841 if (!memcg->use_hierarchy) {
842 if (css_tryget(&memcg->css))
848 * searching a memory cgroup which has the smallest ID under given
849 * ROOT cgroup. (ID >= 1)
851 css = css_get_next(&mem_cgroup_subsys, 1, &memcg->css, &found);
852 if (css && css_tryget(css))
853 memcg = container_of(css, struct mem_cgroup, css);
860 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
861 struct mem_cgroup *root,
864 int nextid = css_id(&iter->css) + 1;
867 struct cgroup_subsys_state *css;
869 hierarchy_used = iter->use_hierarchy;
872 /* If no ROOT, walk all, ignore hierarchy */
873 if (!cond || (root && !hierarchy_used))
877 root = root_mem_cgroup;
883 css = css_get_next(&mem_cgroup_subsys, nextid,
885 if (css && css_tryget(css))
886 iter = container_of(css, struct mem_cgroup, css);
888 /* If css is NULL, no more cgroups will be found */
890 } while (css && !iter);
895 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
896 * be careful that "break" loop is not allowed. We have reference count.
897 * Instead of that modify "cond" to be false and "continue" to exit the loop.
899 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
900 for (iter = mem_cgroup_start_loop(root);\
902 iter = mem_cgroup_get_next(iter, root, cond))
904 #define for_each_mem_cgroup_tree(iter, root) \
905 for_each_mem_cgroup_tree_cond(iter, root, true)
907 #define for_each_mem_cgroup_all(iter) \
908 for_each_mem_cgroup_tree_cond(iter, NULL, true)
911 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
913 return (memcg == root_mem_cgroup);
916 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
918 struct mem_cgroup *memcg;
924 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
925 if (unlikely(!memcg))
930 mem_cgroup_pgmajfault(memcg, 1);
933 mem_cgroup_pgfault(memcg, 1);
941 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
944 * Following LRU functions are allowed to be used without PCG_LOCK.
945 * Operations are called by routine of global LRU independently from memcg.
946 * What we have to take care of here is validness of pc->mem_cgroup.
948 * Changes to pc->mem_cgroup happens when
951 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
952 * It is added to LRU before charge.
953 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
954 * When moving account, the page is not on LRU. It's isolated.
957 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
959 struct page_cgroup *pc;
960 struct mem_cgroup_per_zone *mz;
962 if (mem_cgroup_disabled())
964 pc = lookup_page_cgroup(page);
965 /* can happen while we handle swapcache. */
966 if (!TestClearPageCgroupAcctLRU(pc))
968 VM_BUG_ON(!pc->mem_cgroup);
970 * We don't check PCG_USED bit. It's cleared when the "page" is finally
971 * removed from global LRU.
973 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
974 /* huge page split is done under lru_lock. so, we have no races. */
975 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
976 if (mem_cgroup_is_root(pc->mem_cgroup))
978 VM_BUG_ON(list_empty(&pc->lru));
979 list_del_init(&pc->lru);
982 void mem_cgroup_del_lru(struct page *page)
984 mem_cgroup_del_lru_list(page, page_lru(page));
988 * Writeback is about to end against a page which has been marked for immediate
989 * reclaim. If it still appears to be reclaimable, move it to the tail of the
992 void mem_cgroup_rotate_reclaimable_page(struct page *page)
994 struct mem_cgroup_per_zone *mz;
995 struct page_cgroup *pc;
996 enum lru_list lru = page_lru(page);
998 if (mem_cgroup_disabled())
1001 pc = lookup_page_cgroup(page);
1002 /* unused or root page is not rotated. */
1003 if (!PageCgroupUsed(pc))
1005 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1007 if (mem_cgroup_is_root(pc->mem_cgroup))
1009 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1010 list_move_tail(&pc->lru, &mz->lists[lru]);
1013 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1015 struct mem_cgroup_per_zone *mz;
1016 struct page_cgroup *pc;
1018 if (mem_cgroup_disabled())
1021 pc = lookup_page_cgroup(page);
1022 /* unused or root page is not rotated. */
1023 if (!PageCgroupUsed(pc))
1025 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1027 if (mem_cgroup_is_root(pc->mem_cgroup))
1029 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1030 list_move(&pc->lru, &mz->lists[lru]);
1033 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1035 struct page_cgroup *pc;
1036 struct mem_cgroup_per_zone *mz;
1038 if (mem_cgroup_disabled())
1040 pc = lookup_page_cgroup(page);
1041 VM_BUG_ON(PageCgroupAcctLRU(pc));
1042 if (!PageCgroupUsed(pc))
1044 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1046 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1047 /* huge page split is done under lru_lock. so, we have no races. */
1048 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1049 SetPageCgroupAcctLRU(pc);
1050 if (mem_cgroup_is_root(pc->mem_cgroup))
1052 list_add(&pc->lru, &mz->lists[lru]);
1056 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1057 * while it's linked to lru because the page may be reused after it's fully
1058 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1059 * It's done under lock_page and expected that zone->lru_lock isnever held.
1061 static void mem_cgroup_lru_del_before_commit(struct page *page)
1063 unsigned long flags;
1064 struct zone *zone = page_zone(page);
1065 struct page_cgroup *pc = lookup_page_cgroup(page);
1068 * Doing this check without taking ->lru_lock seems wrong but this
1069 * is safe. Because if page_cgroup's USED bit is unset, the page
1070 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1071 * set, the commit after this will fail, anyway.
1072 * This all charge/uncharge is done under some mutual execustion.
1073 * So, we don't need to taking care of changes in USED bit.
1075 if (likely(!PageLRU(page)))
1078 spin_lock_irqsave(&zone->lru_lock, flags);
1080 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1081 * is guarded by lock_page() because the page is SwapCache.
1083 if (!PageCgroupUsed(pc))
1084 mem_cgroup_del_lru_list(page, page_lru(page));
1085 spin_unlock_irqrestore(&zone->lru_lock, flags);
1088 static void mem_cgroup_lru_add_after_commit(struct page *page)
1090 unsigned long flags;
1091 struct zone *zone = page_zone(page);
1092 struct page_cgroup *pc = lookup_page_cgroup(page);
1094 /* taking care of that the page is added to LRU while we commit it */
1095 if (likely(!PageLRU(page)))
1097 spin_lock_irqsave(&zone->lru_lock, flags);
1098 /* link when the page is linked to LRU but page_cgroup isn't */
1099 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1100 mem_cgroup_add_lru_list(page, page_lru(page));
1101 spin_unlock_irqrestore(&zone->lru_lock, flags);
1105 void mem_cgroup_move_lists(struct page *page,
1106 enum lru_list from, enum lru_list to)
1108 if (mem_cgroup_disabled())
1110 mem_cgroup_del_lru_list(page, from);
1111 mem_cgroup_add_lru_list(page, to);
1115 * Checks whether given mem is same or in the root_mem's
1118 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1119 struct mem_cgroup *memcg)
1121 if (root_memcg != memcg) {
1122 return (root_memcg->use_hierarchy &&
1123 css_is_ancestor(&memcg->css, &root_memcg->css));
1129 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1132 struct mem_cgroup *curr = NULL;
1133 struct task_struct *p;
1135 p = find_lock_task_mm(task);
1138 curr = try_get_mem_cgroup_from_mm(p->mm);
1143 * We should check use_hierarchy of "memcg" not "curr". Because checking
1144 * use_hierarchy of "curr" here make this function true if hierarchy is
1145 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1146 * hierarchy(even if use_hierarchy is disabled in "memcg").
1148 ret = mem_cgroup_same_or_subtree(memcg, curr);
1149 css_put(&curr->css);
1153 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1155 unsigned long active;
1156 unsigned long inactive;
1158 unsigned long inactive_ratio;
1160 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
1161 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
1163 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1165 inactive_ratio = int_sqrt(10 * gb);
1169 if (present_pages) {
1170 present_pages[0] = inactive;
1171 present_pages[1] = active;
1174 return inactive_ratio;
1177 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1179 unsigned long active;
1180 unsigned long inactive;
1181 unsigned long present_pages[2];
1182 unsigned long inactive_ratio;
1184 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1186 inactive = present_pages[0];
1187 active = present_pages[1];
1189 if (inactive * inactive_ratio < active)
1195 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1197 unsigned long active;
1198 unsigned long inactive;
1200 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
1201 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
1203 return (active > inactive);
1206 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1209 int nid = zone_to_nid(zone);
1210 int zid = zone_idx(zone);
1211 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1213 return &mz->reclaim_stat;
1216 struct zone_reclaim_stat *
1217 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1219 struct page_cgroup *pc;
1220 struct mem_cgroup_per_zone *mz;
1222 if (mem_cgroup_disabled())
1225 pc = lookup_page_cgroup(page);
1226 if (!PageCgroupUsed(pc))
1228 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1230 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1231 return &mz->reclaim_stat;
1234 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1235 struct list_head *dst,
1236 unsigned long *scanned, int order,
1237 isolate_mode_t mode,
1239 struct mem_cgroup *mem_cont,
1240 int active, int file)
1242 unsigned long nr_taken = 0;
1246 struct list_head *src;
1247 struct page_cgroup *pc, *tmp;
1248 int nid = zone_to_nid(z);
1249 int zid = zone_idx(z);
1250 struct mem_cgroup_per_zone *mz;
1251 int lru = LRU_FILE * file + active;
1255 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1256 src = &mz->lists[lru];
1259 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1260 if (scan >= nr_to_scan)
1263 if (unlikely(!PageCgroupUsed(pc)))
1266 page = lookup_cgroup_page(pc);
1268 if (unlikely(!PageLRU(page)))
1272 ret = __isolate_lru_page(page, mode, file);
1275 list_move(&page->lru, dst);
1276 mem_cgroup_del_lru(page);
1277 nr_taken += hpage_nr_pages(page);
1280 /* we don't affect global LRU but rotate in our LRU */
1281 mem_cgroup_rotate_lru_list(page, page_lru(page));
1290 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1296 #define mem_cgroup_from_res_counter(counter, member) \
1297 container_of(counter, struct mem_cgroup, member)
1300 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1301 * @mem: the memory cgroup
1303 * Returns the maximum amount of memory @mem can be charged with, in
1306 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1308 unsigned long long margin;
1310 margin = res_counter_margin(&memcg->res);
1311 if (do_swap_account)
1312 margin = min(margin, res_counter_margin(&memcg->memsw));
1313 return margin >> PAGE_SHIFT;
1316 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1318 struct cgroup *cgrp = memcg->css.cgroup;
1321 if (cgrp->parent == NULL)
1322 return vm_swappiness;
1324 return memcg->swappiness;
1327 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1332 spin_lock(&memcg->pcp_counter_lock);
1333 for_each_online_cpu(cpu)
1334 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1335 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1336 spin_unlock(&memcg->pcp_counter_lock);
1342 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1349 spin_lock(&memcg->pcp_counter_lock);
1350 for_each_online_cpu(cpu)
1351 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1352 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1353 spin_unlock(&memcg->pcp_counter_lock);
1357 * 2 routines for checking "mem" is under move_account() or not.
1359 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1360 * for avoiding race in accounting. If true,
1361 * pc->mem_cgroup may be overwritten.
1363 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1364 * under hierarchy of moving cgroups. This is for
1365 * waiting at hith-memory prressure caused by "move".
1368 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1370 VM_BUG_ON(!rcu_read_lock_held());
1371 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1374 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1376 struct mem_cgroup *from;
1377 struct mem_cgroup *to;
1380 * Unlike task_move routines, we access mc.to, mc.from not under
1381 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1383 spin_lock(&mc.lock);
1389 ret = mem_cgroup_same_or_subtree(memcg, from)
1390 || mem_cgroup_same_or_subtree(memcg, to);
1392 spin_unlock(&mc.lock);
1396 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1398 if (mc.moving_task && current != mc.moving_task) {
1399 if (mem_cgroup_under_move(memcg)) {
1401 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1402 /* moving charge context might have finished. */
1405 finish_wait(&mc.waitq, &wait);
1413 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1414 * @memcg: The memory cgroup that went over limit
1415 * @p: Task that is going to be killed
1417 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1420 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1422 struct cgroup *task_cgrp;
1423 struct cgroup *mem_cgrp;
1425 * Need a buffer in BSS, can't rely on allocations. The code relies
1426 * on the assumption that OOM is serialized for memory controller.
1427 * If this assumption is broken, revisit this code.
1429 static char memcg_name[PATH_MAX];
1438 mem_cgrp = memcg->css.cgroup;
1439 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1441 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1444 * Unfortunately, we are unable to convert to a useful name
1445 * But we'll still print out the usage information
1452 printk(KERN_INFO "Task in %s killed", memcg_name);
1455 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1463 * Continues from above, so we don't need an KERN_ level
1465 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1468 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1469 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1470 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1471 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1472 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1474 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1475 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1476 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1480 * This function returns the number of memcg under hierarchy tree. Returns
1481 * 1(self count) if no children.
1483 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1486 struct mem_cgroup *iter;
1488 for_each_mem_cgroup_tree(iter, memcg)
1494 * Return the memory (and swap, if configured) limit for a memcg.
1496 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1501 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1502 limit += total_swap_pages << PAGE_SHIFT;
1504 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1506 * If memsw is finite and limits the amount of swap space available
1507 * to this memcg, return that limit.
1509 return min(limit, memsw);
1513 * Visit the first child (need not be the first child as per the ordering
1514 * of the cgroup list, since we track last_scanned_child) of @mem and use
1515 * that to reclaim free pages from.
1517 static struct mem_cgroup *
1518 mem_cgroup_select_victim(struct mem_cgroup *root_memcg)
1520 struct mem_cgroup *ret = NULL;
1521 struct cgroup_subsys_state *css;
1524 if (!root_memcg->use_hierarchy) {
1525 css_get(&root_memcg->css);
1531 nextid = root_memcg->last_scanned_child + 1;
1532 css = css_get_next(&mem_cgroup_subsys, nextid, &root_memcg->css,
1534 if (css && css_tryget(css))
1535 ret = container_of(css, struct mem_cgroup, css);
1538 /* Updates scanning parameter */
1540 /* this means start scan from ID:1 */
1541 root_memcg->last_scanned_child = 0;
1543 root_memcg->last_scanned_child = found;
1550 * test_mem_cgroup_node_reclaimable
1551 * @mem: the target memcg
1552 * @nid: the node ID to be checked.
1553 * @noswap : specify true here if the user wants flle only information.
1555 * This function returns whether the specified memcg contains any
1556 * reclaimable pages on a node. Returns true if there are any reclaimable
1557 * pages in the node.
1559 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1560 int nid, bool noswap)
1562 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1564 if (noswap || !total_swap_pages)
1566 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1571 #if MAX_NUMNODES > 1
1574 * Always updating the nodemask is not very good - even if we have an empty
1575 * list or the wrong list here, we can start from some node and traverse all
1576 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1579 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1583 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1584 * pagein/pageout changes since the last update.
1586 if (!atomic_read(&memcg->numainfo_events))
1588 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1591 /* make a nodemask where this memcg uses memory from */
1592 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1594 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1596 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1597 node_clear(nid, memcg->scan_nodes);
1600 atomic_set(&memcg->numainfo_events, 0);
1601 atomic_set(&memcg->numainfo_updating, 0);
1605 * Selecting a node where we start reclaim from. Because what we need is just
1606 * reducing usage counter, start from anywhere is O,K. Considering
1607 * memory reclaim from current node, there are pros. and cons.
1609 * Freeing memory from current node means freeing memory from a node which
1610 * we'll use or we've used. So, it may make LRU bad. And if several threads
1611 * hit limits, it will see a contention on a node. But freeing from remote
1612 * node means more costs for memory reclaim because of memory latency.
1614 * Now, we use round-robin. Better algorithm is welcomed.
1616 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1620 mem_cgroup_may_update_nodemask(memcg);
1621 node = memcg->last_scanned_node;
1623 node = next_node(node, memcg->scan_nodes);
1624 if (node == MAX_NUMNODES)
1625 node = first_node(memcg->scan_nodes);
1627 * We call this when we hit limit, not when pages are added to LRU.
1628 * No LRU may hold pages because all pages are UNEVICTABLE or
1629 * memcg is too small and all pages are not on LRU. In that case,
1630 * we use curret node.
1632 if (unlikely(node == MAX_NUMNODES))
1633 node = numa_node_id();
1635 memcg->last_scanned_node = node;
1640 * Check all nodes whether it contains reclaimable pages or not.
1641 * For quick scan, we make use of scan_nodes. This will allow us to skip
1642 * unused nodes. But scan_nodes is lazily updated and may not cotain
1643 * enough new information. We need to do double check.
1645 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1650 * quick check...making use of scan_node.
1651 * We can skip unused nodes.
1653 if (!nodes_empty(memcg->scan_nodes)) {
1654 for (nid = first_node(memcg->scan_nodes);
1656 nid = next_node(nid, memcg->scan_nodes)) {
1658 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1663 * Check rest of nodes.
1665 for_each_node_state(nid, N_HIGH_MEMORY) {
1666 if (node_isset(nid, memcg->scan_nodes))
1668 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1675 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1680 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1682 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1686 static void __mem_cgroup_record_scanstat(unsigned long *stats,
1687 struct memcg_scanrecord *rec)
1690 stats[SCAN] += rec->nr_scanned[0] + rec->nr_scanned[1];
1691 stats[SCAN_ANON] += rec->nr_scanned[0];
1692 stats[SCAN_FILE] += rec->nr_scanned[1];
1694 stats[ROTATE] += rec->nr_rotated[0] + rec->nr_rotated[1];
1695 stats[ROTATE_ANON] += rec->nr_rotated[0];
1696 stats[ROTATE_FILE] += rec->nr_rotated[1];
1698 stats[FREED] += rec->nr_freed[0] + rec->nr_freed[1];
1699 stats[FREED_ANON] += rec->nr_freed[0];
1700 stats[FREED_FILE] += rec->nr_freed[1];
1702 stats[ELAPSED] += rec->elapsed;
1705 static void mem_cgroup_record_scanstat(struct memcg_scanrecord *rec)
1707 struct mem_cgroup *memcg;
1708 int context = rec->context;
1710 if (context >= NR_SCAN_CONTEXT)
1714 spin_lock(&memcg->scanstat.lock);
1715 __mem_cgroup_record_scanstat(memcg->scanstat.stats[context], rec);
1716 spin_unlock(&memcg->scanstat.lock);
1719 spin_lock(&memcg->scanstat.lock);
1720 __mem_cgroup_record_scanstat(memcg->scanstat.rootstats[context], rec);
1721 spin_unlock(&memcg->scanstat.lock);
1725 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1726 * we reclaimed from, so that we don't end up penalizing one child extensively
1727 * based on its position in the children list.
1729 * root_memcg is the original ancestor that we've been reclaim from.
1731 * We give up and return to the caller when we visit root_memcg twice.
1732 * (other groups can be removed while we're walking....)
1734 * If shrink==true, for avoiding to free too much, this returns immedieately.
1736 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_memcg,
1739 unsigned long reclaim_options,
1740 unsigned long *total_scanned)
1742 struct mem_cgroup *victim;
1745 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1746 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1747 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1748 struct memcg_scanrecord rec;
1749 unsigned long excess;
1750 unsigned long scanned;
1752 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1754 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1755 if (!check_soft && !shrink && root_memcg->memsw_is_minimum)
1759 rec.context = SCAN_BY_SHRINK;
1760 else if (check_soft)
1761 rec.context = SCAN_BY_SYSTEM;
1763 rec.context = SCAN_BY_LIMIT;
1765 rec.root = root_memcg;
1768 victim = mem_cgroup_select_victim(root_memcg);
1769 if (victim == root_memcg) {
1772 * We are not draining per cpu cached charges during
1773 * soft limit reclaim because global reclaim doesn't
1774 * care about charges. It tries to free some memory and
1775 * charges will not give any.
1777 if (!check_soft && loop >= 1)
1778 drain_all_stock_async(root_memcg);
1781 * If we have not been able to reclaim
1782 * anything, it might because there are
1783 * no reclaimable pages under this hierarchy
1785 if (!check_soft || !total) {
1786 css_put(&victim->css);
1790 * We want to do more targeted reclaim.
1791 * excess >> 2 is not to excessive so as to
1792 * reclaim too much, nor too less that we keep
1793 * coming back to reclaim from this cgroup
1795 if (total >= (excess >> 2) ||
1796 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1797 css_put(&victim->css);
1802 if (!mem_cgroup_reclaimable(victim, noswap)) {
1803 /* this cgroup's local usage == 0 */
1804 css_put(&victim->css);
1808 rec.nr_scanned[0] = 0;
1809 rec.nr_scanned[1] = 0;
1810 rec.nr_rotated[0] = 0;
1811 rec.nr_rotated[1] = 0;
1812 rec.nr_freed[0] = 0;
1813 rec.nr_freed[1] = 0;
1815 /* we use swappiness of local cgroup */
1817 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1818 noswap, zone, &rec, &scanned);
1819 *total_scanned += scanned;
1821 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1823 mem_cgroup_record_scanstat(&rec);
1824 css_put(&victim->css);
1826 * At shrinking usage, we can't check we should stop here or
1827 * reclaim more. It's depends on callers. last_scanned_child
1828 * will work enough for keeping fairness under tree.
1834 if (!res_counter_soft_limit_excess(&root_memcg->res))
1836 } else if (mem_cgroup_margin(root_memcg))
1843 * Check OOM-Killer is already running under our hierarchy.
1844 * If someone is running, return false.
1845 * Has to be called with memcg_oom_lock
1847 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1849 struct mem_cgroup *iter, *failed = NULL;
1852 for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1853 if (iter->oom_lock) {
1855 * this subtree of our hierarchy is already locked
1856 * so we cannot give a lock.
1861 iter->oom_lock = true;
1868 * OK, we failed to lock the whole subtree so we have to clean up
1869 * what we set up to the failing subtree
1872 for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1873 if (iter == failed) {
1877 iter->oom_lock = false;
1883 * Has to be called with memcg_oom_lock
1885 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1887 struct mem_cgroup *iter;
1889 for_each_mem_cgroup_tree(iter, memcg)
1890 iter->oom_lock = false;
1894 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1896 struct mem_cgroup *iter;
1898 for_each_mem_cgroup_tree(iter, memcg)
1899 atomic_inc(&iter->under_oom);
1902 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1904 struct mem_cgroup *iter;
1907 * When a new child is created while the hierarchy is under oom,
1908 * mem_cgroup_oom_lock() may not be called. We have to use
1909 * atomic_add_unless() here.
1911 for_each_mem_cgroup_tree(iter, memcg)
1912 atomic_add_unless(&iter->under_oom, -1, 0);
1915 static DEFINE_SPINLOCK(memcg_oom_lock);
1916 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1918 struct oom_wait_info {
1919 struct mem_cgroup *mem;
1923 static int memcg_oom_wake_function(wait_queue_t *wait,
1924 unsigned mode, int sync, void *arg)
1926 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1928 struct oom_wait_info *oom_wait_info;
1930 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1931 oom_wait_memcg = oom_wait_info->mem;
1934 * Both of oom_wait_info->mem and wake_mem are stable under us.
1935 * Then we can use css_is_ancestor without taking care of RCU.
1937 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1938 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1940 return autoremove_wake_function(wait, mode, sync, arg);
1943 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1945 /* for filtering, pass "memcg" as argument. */
1946 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1949 static void memcg_oom_recover(struct mem_cgroup *memcg)
1951 if (memcg && atomic_read(&memcg->under_oom))
1952 memcg_wakeup_oom(memcg);
1956 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1958 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1960 struct oom_wait_info owait;
1961 bool locked, need_to_kill;
1964 owait.wait.flags = 0;
1965 owait.wait.func = memcg_oom_wake_function;
1966 owait.wait.private = current;
1967 INIT_LIST_HEAD(&owait.wait.task_list);
1968 need_to_kill = true;
1969 mem_cgroup_mark_under_oom(memcg);
1971 /* At first, try to OOM lock hierarchy under memcg.*/
1972 spin_lock(&memcg_oom_lock);
1973 locked = mem_cgroup_oom_lock(memcg);
1975 * Even if signal_pending(), we can't quit charge() loop without
1976 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1977 * under OOM is always welcomed, use TASK_KILLABLE here.
1979 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1980 if (!locked || memcg->oom_kill_disable)
1981 need_to_kill = false;
1983 mem_cgroup_oom_notify(memcg);
1984 spin_unlock(&memcg_oom_lock);
1987 finish_wait(&memcg_oom_waitq, &owait.wait);
1988 mem_cgroup_out_of_memory(memcg, mask);
1991 finish_wait(&memcg_oom_waitq, &owait.wait);
1993 spin_lock(&memcg_oom_lock);
1995 mem_cgroup_oom_unlock(memcg);
1996 memcg_wakeup_oom(memcg);
1997 spin_unlock(&memcg_oom_lock);
1999 mem_cgroup_unmark_under_oom(memcg);
2001 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2003 /* Give chance to dying process */
2004 schedule_timeout_uninterruptible(1);
2009 * Currently used to update mapped file statistics, but the routine can be
2010 * generalized to update other statistics as well.
2012 * Notes: Race condition
2014 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2015 * it tends to be costly. But considering some conditions, we doesn't need
2016 * to do so _always_.
2018 * Considering "charge", lock_page_cgroup() is not required because all
2019 * file-stat operations happen after a page is attached to radix-tree. There
2020 * are no race with "charge".
2022 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2023 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2024 * if there are race with "uncharge". Statistics itself is properly handled
2027 * Considering "move", this is an only case we see a race. To make the race
2028 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
2029 * possibility of race condition. If there is, we take a lock.
2032 void mem_cgroup_update_page_stat(struct page *page,
2033 enum mem_cgroup_page_stat_item idx, int val)
2035 struct mem_cgroup *memcg;
2036 struct page_cgroup *pc = lookup_page_cgroup(page);
2037 bool need_unlock = false;
2038 unsigned long uninitialized_var(flags);
2044 memcg = pc->mem_cgroup;
2045 if (unlikely(!memcg || !PageCgroupUsed(pc)))
2047 /* pc->mem_cgroup is unstable ? */
2048 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
2049 /* take a lock against to access pc->mem_cgroup */
2050 move_lock_page_cgroup(pc, &flags);
2052 memcg = pc->mem_cgroup;
2053 if (!memcg || !PageCgroupUsed(pc))
2058 case MEMCG_NR_FILE_MAPPED:
2060 SetPageCgroupFileMapped(pc);
2061 else if (!page_mapped(page))
2062 ClearPageCgroupFileMapped(pc);
2063 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2069 this_cpu_add(memcg->stat->count[idx], val);
2072 if (unlikely(need_unlock))
2073 move_unlock_page_cgroup(pc, &flags);
2077 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2080 * size of first charge trial. "32" comes from vmscan.c's magic value.
2081 * TODO: maybe necessary to use big numbers in big irons.
2083 #define CHARGE_BATCH 32U
2084 struct memcg_stock_pcp {
2085 struct mem_cgroup *cached; /* this never be root cgroup */
2086 unsigned int nr_pages;
2087 struct work_struct work;
2088 unsigned long flags;
2089 #define FLUSHING_CACHED_CHARGE (0)
2091 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2092 static DEFINE_MUTEX(percpu_charge_mutex);
2095 * Try to consume stocked charge on this cpu. If success, one page is consumed
2096 * from local stock and true is returned. If the stock is 0 or charges from a
2097 * cgroup which is not current target, returns false. This stock will be
2100 static bool consume_stock(struct mem_cgroup *memcg)
2102 struct memcg_stock_pcp *stock;
2105 stock = &get_cpu_var(memcg_stock);
2106 if (memcg == stock->cached && stock->nr_pages)
2108 else /* need to call res_counter_charge */
2110 put_cpu_var(memcg_stock);
2115 * Returns stocks cached in percpu to res_counter and reset cached information.
2117 static void drain_stock(struct memcg_stock_pcp *stock)
2119 struct mem_cgroup *old = stock->cached;
2121 if (stock->nr_pages) {
2122 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2124 res_counter_uncharge(&old->res, bytes);
2125 if (do_swap_account)
2126 res_counter_uncharge(&old->memsw, bytes);
2127 stock->nr_pages = 0;
2129 stock->cached = NULL;
2133 * This must be called under preempt disabled or must be called by
2134 * a thread which is pinned to local cpu.
2136 static void drain_local_stock(struct work_struct *dummy)
2138 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2140 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2144 * Cache charges(val) which is from res_counter, to local per_cpu area.
2145 * This will be consumed by consume_stock() function, later.
2147 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2149 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2151 if (stock->cached != memcg) { /* reset if necessary */
2153 stock->cached = memcg;
2155 stock->nr_pages += nr_pages;
2156 put_cpu_var(memcg_stock);
2160 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2161 * of the hierarchy under it. sync flag says whether we should block
2162 * until the work is done.
2164 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2168 /* Notify other cpus that system-wide "drain" is running */
2171 for_each_online_cpu(cpu) {
2172 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2173 struct mem_cgroup *memcg;
2175 memcg = stock->cached;
2176 if (!memcg || !stock->nr_pages)
2178 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2180 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2182 drain_local_stock(&stock->work);
2184 schedule_work_on(cpu, &stock->work);
2192 for_each_online_cpu(cpu) {
2193 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2194 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2195 flush_work(&stock->work);
2202 * Tries to drain stocked charges in other cpus. This function is asynchronous
2203 * and just put a work per cpu for draining localy on each cpu. Caller can
2204 * expects some charges will be back to res_counter later but cannot wait for
2207 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2210 * If someone calls draining, avoid adding more kworker runs.
2212 if (!mutex_trylock(&percpu_charge_mutex))
2214 drain_all_stock(root_memcg, false);
2215 mutex_unlock(&percpu_charge_mutex);
2218 /* This is a synchronous drain interface. */
2219 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2221 /* called when force_empty is called */
2222 mutex_lock(&percpu_charge_mutex);
2223 drain_all_stock(root_memcg, true);
2224 mutex_unlock(&percpu_charge_mutex);
2228 * This function drains percpu counter value from DEAD cpu and
2229 * move it to local cpu. Note that this function can be preempted.
2231 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2235 spin_lock(&memcg->pcp_counter_lock);
2236 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2237 long x = per_cpu(memcg->stat->count[i], cpu);
2239 per_cpu(memcg->stat->count[i], cpu) = 0;
2240 memcg->nocpu_base.count[i] += x;
2242 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2243 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2245 per_cpu(memcg->stat->events[i], cpu) = 0;
2246 memcg->nocpu_base.events[i] += x;
2248 /* need to clear ON_MOVE value, works as a kind of lock. */
2249 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2250 spin_unlock(&memcg->pcp_counter_lock);
2253 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2255 int idx = MEM_CGROUP_ON_MOVE;
2257 spin_lock(&memcg->pcp_counter_lock);
2258 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2259 spin_unlock(&memcg->pcp_counter_lock);
2262 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2263 unsigned long action,
2266 int cpu = (unsigned long)hcpu;
2267 struct memcg_stock_pcp *stock;
2268 struct mem_cgroup *iter;
2270 if ((action == CPU_ONLINE)) {
2271 for_each_mem_cgroup_all(iter)
2272 synchronize_mem_cgroup_on_move(iter, cpu);
2276 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2279 for_each_mem_cgroup_all(iter)
2280 mem_cgroup_drain_pcp_counter(iter, cpu);
2282 stock = &per_cpu(memcg_stock, cpu);
2288 /* See __mem_cgroup_try_charge() for details */
2290 CHARGE_OK, /* success */
2291 CHARGE_RETRY, /* need to retry but retry is not bad */
2292 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2293 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2294 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2297 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2298 unsigned int nr_pages, bool oom_check)
2300 unsigned long csize = nr_pages * PAGE_SIZE;
2301 struct mem_cgroup *mem_over_limit;
2302 struct res_counter *fail_res;
2303 unsigned long flags = 0;
2306 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2309 if (!do_swap_account)
2311 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2315 res_counter_uncharge(&memcg->res, csize);
2316 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2317 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2319 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2321 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2322 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2324 * Never reclaim on behalf of optional batching, retry with a
2325 * single page instead.
2327 if (nr_pages == CHARGE_BATCH)
2328 return CHARGE_RETRY;
2330 if (!(gfp_mask & __GFP_WAIT))
2331 return CHARGE_WOULDBLOCK;
2333 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2334 gfp_mask, flags, NULL);
2335 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2336 return CHARGE_RETRY;
2338 * Even though the limit is exceeded at this point, reclaim
2339 * may have been able to free some pages. Retry the charge
2340 * before killing the task.
2342 * Only for regular pages, though: huge pages are rather
2343 * unlikely to succeed so close to the limit, and we fall back
2344 * to regular pages anyway in case of failure.
2346 if (nr_pages == 1 && ret)
2347 return CHARGE_RETRY;
2350 * At task move, charge accounts can be doubly counted. So, it's
2351 * better to wait until the end of task_move if something is going on.
2353 if (mem_cgroup_wait_acct_move(mem_over_limit))
2354 return CHARGE_RETRY;
2356 /* If we don't need to call oom-killer at el, return immediately */
2358 return CHARGE_NOMEM;
2360 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2361 return CHARGE_OOM_DIE;
2363 return CHARGE_RETRY;
2367 * Unlike exported interface, "oom" parameter is added. if oom==true,
2368 * oom-killer can be invoked.
2370 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2372 unsigned int nr_pages,
2373 struct mem_cgroup **ptr,
2376 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2377 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2378 struct mem_cgroup *memcg = NULL;
2382 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2383 * in system level. So, allow to go ahead dying process in addition to
2386 if (unlikely(test_thread_flag(TIF_MEMDIE)
2387 || fatal_signal_pending(current)))
2391 * We always charge the cgroup the mm_struct belongs to.
2392 * The mm_struct's mem_cgroup changes on task migration if the
2393 * thread group leader migrates. It's possible that mm is not
2394 * set, if so charge the init_mm (happens for pagecache usage).
2399 if (*ptr) { /* css should be a valid one */
2401 VM_BUG_ON(css_is_removed(&memcg->css));
2402 if (mem_cgroup_is_root(memcg))
2404 if (nr_pages == 1 && consume_stock(memcg))
2406 css_get(&memcg->css);
2408 struct task_struct *p;
2411 p = rcu_dereference(mm->owner);
2413 * Because we don't have task_lock(), "p" can exit.
2414 * In that case, "memcg" can point to root or p can be NULL with
2415 * race with swapoff. Then, we have small risk of mis-accouning.
2416 * But such kind of mis-account by race always happens because
2417 * we don't have cgroup_mutex(). It's overkill and we allo that
2419 * (*) swapoff at el will charge against mm-struct not against
2420 * task-struct. So, mm->owner can be NULL.
2422 memcg = mem_cgroup_from_task(p);
2423 if (!memcg || mem_cgroup_is_root(memcg)) {
2427 if (nr_pages == 1 && consume_stock(memcg)) {
2429 * It seems dagerous to access memcg without css_get().
2430 * But considering how consume_stok works, it's not
2431 * necessary. If consume_stock success, some charges
2432 * from this memcg are cached on this cpu. So, we
2433 * don't need to call css_get()/css_tryget() before
2434 * calling consume_stock().
2439 /* after here, we may be blocked. we need to get refcnt */
2440 if (!css_tryget(&memcg->css)) {
2450 /* If killed, bypass charge */
2451 if (fatal_signal_pending(current)) {
2452 css_put(&memcg->css);
2457 if (oom && !nr_oom_retries) {
2459 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2462 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2466 case CHARGE_RETRY: /* not in OOM situation but retry */
2468 css_put(&memcg->css);
2471 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2472 css_put(&memcg->css);
2474 case CHARGE_NOMEM: /* OOM routine works */
2476 css_put(&memcg->css);
2479 /* If oom, we never return -ENOMEM */
2482 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2483 css_put(&memcg->css);
2486 } while (ret != CHARGE_OK);
2488 if (batch > nr_pages)
2489 refill_stock(memcg, batch - nr_pages);
2490 css_put(&memcg->css);
2503 * Somemtimes we have to undo a charge we got by try_charge().
2504 * This function is for that and do uncharge, put css's refcnt.
2505 * gotten by try_charge().
2507 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2508 unsigned int nr_pages)
2510 if (!mem_cgroup_is_root(memcg)) {
2511 unsigned long bytes = nr_pages * PAGE_SIZE;
2513 res_counter_uncharge(&memcg->res, bytes);
2514 if (do_swap_account)
2515 res_counter_uncharge(&memcg->memsw, bytes);
2520 * A helper function to get mem_cgroup from ID. must be called under
2521 * rcu_read_lock(). The caller must check css_is_removed() or some if
2522 * it's concern. (dropping refcnt from swap can be called against removed
2525 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2527 struct cgroup_subsys_state *css;
2529 /* ID 0 is unused ID */
2532 css = css_lookup(&mem_cgroup_subsys, id);
2535 return container_of(css, struct mem_cgroup, css);
2538 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2540 struct mem_cgroup *memcg = NULL;
2541 struct page_cgroup *pc;
2545 VM_BUG_ON(!PageLocked(page));
2547 pc = lookup_page_cgroup(page);
2548 lock_page_cgroup(pc);
2549 if (PageCgroupUsed(pc)) {
2550 memcg = pc->mem_cgroup;
2551 if (memcg && !css_tryget(&memcg->css))
2553 } else if (PageSwapCache(page)) {
2554 ent.val = page_private(page);
2555 id = lookup_swap_cgroup(ent);
2557 memcg = mem_cgroup_lookup(id);
2558 if (memcg && !css_tryget(&memcg->css))
2562 unlock_page_cgroup(pc);
2566 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2568 unsigned int nr_pages,
2569 struct page_cgroup *pc,
2570 enum charge_type ctype)
2572 lock_page_cgroup(pc);
2573 if (unlikely(PageCgroupUsed(pc))) {
2574 unlock_page_cgroup(pc);
2575 __mem_cgroup_cancel_charge(memcg, nr_pages);
2579 * we don't need page_cgroup_lock about tail pages, becase they are not
2580 * accessed by any other context at this point.
2582 pc->mem_cgroup = memcg;
2584 * We access a page_cgroup asynchronously without lock_page_cgroup().
2585 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2586 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2587 * before USED bit, we need memory barrier here.
2588 * See mem_cgroup_add_lru_list(), etc.
2592 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2593 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2594 SetPageCgroupCache(pc);
2595 SetPageCgroupUsed(pc);
2597 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2598 ClearPageCgroupCache(pc);
2599 SetPageCgroupUsed(pc);
2605 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2606 unlock_page_cgroup(pc);
2608 * "charge_statistics" updated event counter. Then, check it.
2609 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2610 * if they exceeds softlimit.
2612 memcg_check_events(memcg, page);
2615 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2617 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2618 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2620 * Because tail pages are not marked as "used", set it. We're under
2621 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2623 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2625 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2626 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2627 unsigned long flags;
2629 if (mem_cgroup_disabled())
2632 * We have no races with charge/uncharge but will have races with
2633 * page state accounting.
2635 move_lock_page_cgroup(head_pc, &flags);
2637 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2638 smp_wmb(); /* see __commit_charge() */
2639 if (PageCgroupAcctLRU(head_pc)) {
2641 struct mem_cgroup_per_zone *mz;
2644 * LRU flags cannot be copied because we need to add tail
2645 *.page to LRU by generic call and our hook will be called.
2646 * We hold lru_lock, then, reduce counter directly.
2648 lru = page_lru(head);
2649 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2650 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2652 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2653 move_unlock_page_cgroup(head_pc, &flags);
2658 * mem_cgroup_move_account - move account of the page
2660 * @nr_pages: number of regular pages (>1 for huge pages)
2661 * @pc: page_cgroup of the page.
2662 * @from: mem_cgroup which the page is moved from.
2663 * @to: mem_cgroup which the page is moved to. @from != @to.
2664 * @uncharge: whether we should call uncharge and css_put against @from.
2666 * The caller must confirm following.
2667 * - page is not on LRU (isolate_page() is useful.)
2668 * - compound_lock is held when nr_pages > 1
2670 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2671 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2672 * true, this function does "uncharge" from old cgroup, but it doesn't if
2673 * @uncharge is false, so a caller should do "uncharge".
2675 static int mem_cgroup_move_account(struct page *page,
2676 unsigned int nr_pages,
2677 struct page_cgroup *pc,
2678 struct mem_cgroup *from,
2679 struct mem_cgroup *to,
2682 unsigned long flags;
2685 VM_BUG_ON(from == to);
2686 VM_BUG_ON(PageLRU(page));
2688 * The page is isolated from LRU. So, collapse function
2689 * will not handle this page. But page splitting can happen.
2690 * Do this check under compound_page_lock(). The caller should
2694 if (nr_pages > 1 && !PageTransHuge(page))
2697 lock_page_cgroup(pc);
2700 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2703 move_lock_page_cgroup(pc, &flags);
2705 if (PageCgroupFileMapped(pc)) {
2706 /* Update mapped_file data for mem_cgroup */
2708 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2709 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2712 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2714 /* This is not "cancel", but cancel_charge does all we need. */
2715 __mem_cgroup_cancel_charge(from, nr_pages);
2717 /* caller should have done css_get */
2718 pc->mem_cgroup = to;
2719 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2721 * We charges against "to" which may not have any tasks. Then, "to"
2722 * can be under rmdir(). But in current implementation, caller of
2723 * this function is just force_empty() and move charge, so it's
2724 * guaranteed that "to" is never removed. So, we don't check rmdir
2727 move_unlock_page_cgroup(pc, &flags);
2730 unlock_page_cgroup(pc);
2734 memcg_check_events(to, page);
2735 memcg_check_events(from, page);
2741 * move charges to its parent.
2744 static int mem_cgroup_move_parent(struct page *page,
2745 struct page_cgroup *pc,
2746 struct mem_cgroup *child,
2749 struct cgroup *cg = child->css.cgroup;
2750 struct cgroup *pcg = cg->parent;
2751 struct mem_cgroup *parent;
2752 unsigned int nr_pages;
2753 unsigned long uninitialized_var(flags);
2761 if (!get_page_unless_zero(page))
2763 if (isolate_lru_page(page))
2766 nr_pages = hpage_nr_pages(page);
2768 parent = mem_cgroup_from_cont(pcg);
2769 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2774 flags = compound_lock_irqsave(page);
2776 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2778 __mem_cgroup_cancel_charge(parent, nr_pages);
2781 compound_unlock_irqrestore(page, flags);
2783 putback_lru_page(page);
2791 * Charge the memory controller for page usage.
2793 * 0 if the charge was successful
2794 * < 0 if the cgroup is over its limit
2796 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2797 gfp_t gfp_mask, enum charge_type ctype)
2799 struct mem_cgroup *memcg = NULL;
2800 unsigned int nr_pages = 1;
2801 struct page_cgroup *pc;
2805 if (PageTransHuge(page)) {
2806 nr_pages <<= compound_order(page);
2807 VM_BUG_ON(!PageTransHuge(page));
2809 * Never OOM-kill a process for a huge page. The
2810 * fault handler will fall back to regular pages.
2815 pc = lookup_page_cgroup(page);
2816 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2818 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2822 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2826 int mem_cgroup_newpage_charge(struct page *page,
2827 struct mm_struct *mm, gfp_t gfp_mask)
2829 if (mem_cgroup_disabled())
2832 * If already mapped, we don't have to account.
2833 * If page cache, page->mapping has address_space.
2834 * But page->mapping may have out-of-use anon_vma pointer,
2835 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2838 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2842 return mem_cgroup_charge_common(page, mm, gfp_mask,
2843 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2847 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2848 enum charge_type ctype);
2851 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2852 enum charge_type ctype)
2854 struct page_cgroup *pc = lookup_page_cgroup(page);
2856 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2857 * is already on LRU. It means the page may on some other page_cgroup's
2858 * LRU. Take care of it.
2860 mem_cgroup_lru_del_before_commit(page);
2861 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2862 mem_cgroup_lru_add_after_commit(page);
2866 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2869 struct mem_cgroup *memcg = NULL;
2872 if (mem_cgroup_disabled())
2874 if (PageCompound(page))
2880 if (page_is_file_cache(page)) {
2881 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2886 * FUSE reuses pages without going through the final
2887 * put that would remove them from the LRU list, make
2888 * sure that they get relinked properly.
2890 __mem_cgroup_commit_charge_lrucare(page, memcg,
2891 MEM_CGROUP_CHARGE_TYPE_CACHE);
2895 if (PageSwapCache(page)) {
2896 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2898 __mem_cgroup_commit_charge_swapin(page, memcg,
2899 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2901 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2902 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2908 * While swap-in, try_charge -> commit or cancel, the page is locked.
2909 * And when try_charge() successfully returns, one refcnt to memcg without
2910 * struct page_cgroup is acquired. This refcnt will be consumed by
2911 * "commit()" or removed by "cancel()"
2913 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2915 gfp_t mask, struct mem_cgroup **ptr)
2917 struct mem_cgroup *memcg;
2922 if (mem_cgroup_disabled())
2925 if (!do_swap_account)
2928 * A racing thread's fault, or swapoff, may have already updated
2929 * the pte, and even removed page from swap cache: in those cases
2930 * do_swap_page()'s pte_same() test will fail; but there's also a
2931 * KSM case which does need to charge the page.
2933 if (!PageSwapCache(page))
2935 memcg = try_get_mem_cgroup_from_page(page);
2939 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2940 css_put(&memcg->css);
2945 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2949 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2950 enum charge_type ctype)
2952 if (mem_cgroup_disabled())
2956 cgroup_exclude_rmdir(&ptr->css);
2958 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2960 * Now swap is on-memory. This means this page may be
2961 * counted both as mem and swap....double count.
2962 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2963 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2964 * may call delete_from_swap_cache() before reach here.
2966 if (do_swap_account && PageSwapCache(page)) {
2967 swp_entry_t ent = {.val = page_private(page)};
2969 struct mem_cgroup *memcg;
2971 id = swap_cgroup_record(ent, 0);
2973 memcg = mem_cgroup_lookup(id);
2976 * This recorded memcg can be obsolete one. So, avoid
2977 * calling css_tryget
2979 if (!mem_cgroup_is_root(memcg))
2980 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2981 mem_cgroup_swap_statistics(memcg, false);
2982 mem_cgroup_put(memcg);
2987 * At swapin, we may charge account against cgroup which has no tasks.
2988 * So, rmdir()->pre_destroy() can be called while we do this charge.
2989 * In that case, we need to call pre_destroy() again. check it here.
2991 cgroup_release_and_wakeup_rmdir(&ptr->css);
2994 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2996 __mem_cgroup_commit_charge_swapin(page, ptr,
2997 MEM_CGROUP_CHARGE_TYPE_MAPPED);
3000 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
3002 if (mem_cgroup_disabled())
3006 __mem_cgroup_cancel_charge(memcg, 1);
3009 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
3010 unsigned int nr_pages,
3011 const enum charge_type ctype)
3013 struct memcg_batch_info *batch = NULL;
3014 bool uncharge_memsw = true;
3016 /* If swapout, usage of swap doesn't decrease */
3017 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3018 uncharge_memsw = false;
3020 batch = ¤t->memcg_batch;
3022 * In usual, we do css_get() when we remember memcg pointer.
3023 * But in this case, we keep res->usage until end of a series of
3024 * uncharges. Then, it's ok to ignore memcg's refcnt.
3027 batch->memcg = memcg;
3029 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3030 * In those cases, all pages freed continuously can be expected to be in
3031 * the same cgroup and we have chance to coalesce uncharges.
3032 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3033 * because we want to do uncharge as soon as possible.
3036 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3037 goto direct_uncharge;
3040 goto direct_uncharge;
3043 * In typical case, batch->memcg == mem. This means we can
3044 * merge a series of uncharges to an uncharge of res_counter.
3045 * If not, we uncharge res_counter ony by one.
3047 if (batch->memcg != memcg)
3048 goto direct_uncharge;
3049 /* remember freed charge and uncharge it later */
3052 batch->memsw_nr_pages++;
3055 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
3057 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
3058 if (unlikely(batch->memcg != memcg))
3059 memcg_oom_recover(memcg);
3064 * uncharge if !page_mapped(page)
3066 static struct mem_cgroup *
3067 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3069 struct mem_cgroup *memcg = NULL;
3070 unsigned int nr_pages = 1;
3071 struct page_cgroup *pc;
3073 if (mem_cgroup_disabled())
3076 if (PageSwapCache(page))
3079 if (PageTransHuge(page)) {
3080 nr_pages <<= compound_order(page);
3081 VM_BUG_ON(!PageTransHuge(page));
3084 * Check if our page_cgroup is valid
3086 pc = lookup_page_cgroup(page);
3087 if (unlikely(!pc || !PageCgroupUsed(pc)))
3090 lock_page_cgroup(pc);
3092 memcg = pc->mem_cgroup;
3094 if (!PageCgroupUsed(pc))
3098 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3099 case MEM_CGROUP_CHARGE_TYPE_DROP:
3100 /* See mem_cgroup_prepare_migration() */
3101 if (page_mapped(page) || PageCgroupMigration(pc))
3104 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3105 if (!PageAnon(page)) { /* Shared memory */
3106 if (page->mapping && !page_is_file_cache(page))
3108 } else if (page_mapped(page)) /* Anon */
3115 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3117 ClearPageCgroupUsed(pc);
3119 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3120 * freed from LRU. This is safe because uncharged page is expected not
3121 * to be reused (freed soon). Exception is SwapCache, it's handled by
3122 * special functions.
3125 unlock_page_cgroup(pc);
3127 * even after unlock, we have mem->res.usage here and this memcg
3128 * will never be freed.
3130 memcg_check_events(memcg, page);
3131 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3132 mem_cgroup_swap_statistics(memcg, true);
3133 mem_cgroup_get(memcg);
3135 if (!mem_cgroup_is_root(memcg))
3136 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3141 unlock_page_cgroup(pc);
3145 void mem_cgroup_uncharge_page(struct page *page)
3148 if (page_mapped(page))
3150 if (page->mapping && !PageAnon(page))
3152 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3155 void mem_cgroup_uncharge_cache_page(struct page *page)
3157 VM_BUG_ON(page_mapped(page));
3158 VM_BUG_ON(page->mapping);
3159 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3163 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3164 * In that cases, pages are freed continuously and we can expect pages
3165 * are in the same memcg. All these calls itself limits the number of
3166 * pages freed at once, then uncharge_start/end() is called properly.
3167 * This may be called prural(2) times in a context,
3170 void mem_cgroup_uncharge_start(void)
3172 current->memcg_batch.do_batch++;
3173 /* We can do nest. */
3174 if (current->memcg_batch.do_batch == 1) {
3175 current->memcg_batch.memcg = NULL;
3176 current->memcg_batch.nr_pages = 0;
3177 current->memcg_batch.memsw_nr_pages = 0;
3181 void mem_cgroup_uncharge_end(void)
3183 struct memcg_batch_info *batch = ¤t->memcg_batch;
3185 if (!batch->do_batch)
3189 if (batch->do_batch) /* If stacked, do nothing. */
3195 * This "batch->memcg" is valid without any css_get/put etc...
3196 * bacause we hide charges behind us.
3198 if (batch->nr_pages)
3199 res_counter_uncharge(&batch->memcg->res,
3200 batch->nr_pages * PAGE_SIZE);
3201 if (batch->memsw_nr_pages)
3202 res_counter_uncharge(&batch->memcg->memsw,
3203 batch->memsw_nr_pages * PAGE_SIZE);
3204 memcg_oom_recover(batch->memcg);
3205 /* forget this pointer (for sanity check) */
3206 batch->memcg = NULL;
3211 * called after __delete_from_swap_cache() and drop "page" account.
3212 * memcg information is recorded to swap_cgroup of "ent"
3215 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3217 struct mem_cgroup *memcg;
3218 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3220 if (!swapout) /* this was a swap cache but the swap is unused ! */
3221 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3223 memcg = __mem_cgroup_uncharge_common(page, ctype);
3226 * record memcg information, if swapout && memcg != NULL,
3227 * mem_cgroup_get() was called in uncharge().
3229 if (do_swap_account && swapout && memcg)
3230 swap_cgroup_record(ent, css_id(&memcg->css));
3234 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3236 * called from swap_entry_free(). remove record in swap_cgroup and
3237 * uncharge "memsw" account.
3239 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3241 struct mem_cgroup *memcg;
3244 if (!do_swap_account)
3247 id = swap_cgroup_record(ent, 0);
3249 memcg = mem_cgroup_lookup(id);
3252 * We uncharge this because swap is freed.
3253 * This memcg can be obsolete one. We avoid calling css_tryget
3255 if (!mem_cgroup_is_root(memcg))
3256 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3257 mem_cgroup_swap_statistics(memcg, false);
3258 mem_cgroup_put(memcg);
3264 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3265 * @entry: swap entry to be moved
3266 * @from: mem_cgroup which the entry is moved from
3267 * @to: mem_cgroup which the entry is moved to
3268 * @need_fixup: whether we should fixup res_counters and refcounts.
3270 * It succeeds only when the swap_cgroup's record for this entry is the same
3271 * as the mem_cgroup's id of @from.
3273 * Returns 0 on success, -EINVAL on failure.
3275 * The caller must have charged to @to, IOW, called res_counter_charge() about
3276 * both res and memsw, and called css_get().
3278 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3279 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3281 unsigned short old_id, new_id;
3283 old_id = css_id(&from->css);
3284 new_id = css_id(&to->css);
3286 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3287 mem_cgroup_swap_statistics(from, false);
3288 mem_cgroup_swap_statistics(to, true);
3290 * This function is only called from task migration context now.
3291 * It postpones res_counter and refcount handling till the end
3292 * of task migration(mem_cgroup_clear_mc()) for performance
3293 * improvement. But we cannot postpone mem_cgroup_get(to)
3294 * because if the process that has been moved to @to does
3295 * swap-in, the refcount of @to might be decreased to 0.
3299 if (!mem_cgroup_is_root(from))
3300 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3301 mem_cgroup_put(from);
3303 * we charged both to->res and to->memsw, so we should
3306 if (!mem_cgroup_is_root(to))
3307 res_counter_uncharge(&to->res, PAGE_SIZE);
3314 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3315 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3322 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3325 int mem_cgroup_prepare_migration(struct page *page,
3326 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3328 struct mem_cgroup *memcg = NULL;
3329 struct page_cgroup *pc;
3330 enum charge_type ctype;
3335 VM_BUG_ON(PageTransHuge(page));
3336 if (mem_cgroup_disabled())
3339 pc = lookup_page_cgroup(page);
3340 lock_page_cgroup(pc);
3341 if (PageCgroupUsed(pc)) {
3342 memcg = pc->mem_cgroup;
3343 css_get(&memcg->css);
3345 * At migrating an anonymous page, its mapcount goes down
3346 * to 0 and uncharge() will be called. But, even if it's fully
3347 * unmapped, migration may fail and this page has to be
3348 * charged again. We set MIGRATION flag here and delay uncharge
3349 * until end_migration() is called
3351 * Corner Case Thinking
3353 * When the old page was mapped as Anon and it's unmap-and-freed
3354 * while migration was ongoing.
3355 * If unmap finds the old page, uncharge() of it will be delayed
3356 * until end_migration(). If unmap finds a new page, it's
3357 * uncharged when it make mapcount to be 1->0. If unmap code
3358 * finds swap_migration_entry, the new page will not be mapped
3359 * and end_migration() will find it(mapcount==0).
3362 * When the old page was mapped but migraion fails, the kernel
3363 * remaps it. A charge for it is kept by MIGRATION flag even
3364 * if mapcount goes down to 0. We can do remap successfully
3365 * without charging it again.
3368 * The "old" page is under lock_page() until the end of
3369 * migration, so, the old page itself will not be swapped-out.
3370 * If the new page is swapped out before end_migraton, our
3371 * hook to usual swap-out path will catch the event.
3374 SetPageCgroupMigration(pc);
3376 unlock_page_cgroup(pc);
3378 * If the page is not charged at this point,
3385 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3386 css_put(&memcg->css);/* drop extra refcnt */
3387 if (ret || *ptr == NULL) {
3388 if (PageAnon(page)) {
3389 lock_page_cgroup(pc);
3390 ClearPageCgroupMigration(pc);
3391 unlock_page_cgroup(pc);
3393 * The old page may be fully unmapped while we kept it.
3395 mem_cgroup_uncharge_page(page);
3400 * We charge new page before it's used/mapped. So, even if unlock_page()
3401 * is called before end_migration, we can catch all events on this new
3402 * page. In the case new page is migrated but not remapped, new page's
3403 * mapcount will be finally 0 and we call uncharge in end_migration().
3405 pc = lookup_page_cgroup(newpage);
3407 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3408 else if (page_is_file_cache(page))
3409 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3411 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3412 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3416 /* remove redundant charge if migration failed*/
3417 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3418 struct page *oldpage, struct page *newpage, bool migration_ok)
3420 struct page *used, *unused;
3421 struct page_cgroup *pc;
3425 /* blocks rmdir() */
3426 cgroup_exclude_rmdir(&memcg->css);
3427 if (!migration_ok) {
3435 * We disallowed uncharge of pages under migration because mapcount
3436 * of the page goes down to zero, temporarly.
3437 * Clear the flag and check the page should be charged.
3439 pc = lookup_page_cgroup(oldpage);
3440 lock_page_cgroup(pc);
3441 ClearPageCgroupMigration(pc);
3442 unlock_page_cgroup(pc);
3444 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3447 * If a page is a file cache, radix-tree replacement is very atomic
3448 * and we can skip this check. When it was an Anon page, its mapcount
3449 * goes down to 0. But because we added MIGRATION flage, it's not
3450 * uncharged yet. There are several case but page->mapcount check
3451 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3452 * check. (see prepare_charge() also)
3455 mem_cgroup_uncharge_page(used);
3457 * At migration, we may charge account against cgroup which has no
3459 * So, rmdir()->pre_destroy() can be called while we do this charge.
3460 * In that case, we need to call pre_destroy() again. check it here.
3462 cgroup_release_and_wakeup_rmdir(&memcg->css);
3465 #ifdef CONFIG_DEBUG_VM
3466 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3468 struct page_cgroup *pc;
3470 pc = lookup_page_cgroup(page);
3471 if (likely(pc) && PageCgroupUsed(pc))
3476 bool mem_cgroup_bad_page_check(struct page *page)
3478 if (mem_cgroup_disabled())
3481 return lookup_page_cgroup_used(page) != NULL;
3484 void mem_cgroup_print_bad_page(struct page *page)
3486 struct page_cgroup *pc;
3488 pc = lookup_page_cgroup_used(page);
3493 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3494 pc, pc->flags, pc->mem_cgroup);
3496 path = kmalloc(PATH_MAX, GFP_KERNEL);
3499 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3504 printk(KERN_CONT "(%s)\n",
3505 (ret < 0) ? "cannot get the path" : path);
3511 static DEFINE_MUTEX(set_limit_mutex);
3513 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3514 unsigned long long val)
3517 u64 memswlimit, memlimit;
3519 int children = mem_cgroup_count_children(memcg);
3520 u64 curusage, oldusage;
3524 * For keeping hierarchical_reclaim simple, how long we should retry
3525 * is depends on callers. We set our retry-count to be function
3526 * of # of children which we should visit in this loop.
3528 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3530 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3533 while (retry_count) {
3534 if (signal_pending(current)) {
3539 * Rather than hide all in some function, I do this in
3540 * open coded manner. You see what this really does.
3541 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3543 mutex_lock(&set_limit_mutex);
3544 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3545 if (memswlimit < val) {
3547 mutex_unlock(&set_limit_mutex);
3551 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3555 ret = res_counter_set_limit(&memcg->res, val);
3557 if (memswlimit == val)
3558 memcg->memsw_is_minimum = true;
3560 memcg->memsw_is_minimum = false;
3562 mutex_unlock(&set_limit_mutex);
3567 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3568 MEM_CGROUP_RECLAIM_SHRINK,
3570 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3571 /* Usage is reduced ? */
3572 if (curusage >= oldusage)
3575 oldusage = curusage;
3577 if (!ret && enlarge)
3578 memcg_oom_recover(memcg);
3583 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3584 unsigned long long val)
3587 u64 memlimit, memswlimit, oldusage, curusage;
3588 int children = mem_cgroup_count_children(memcg);
3592 /* see mem_cgroup_resize_res_limit */
3593 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3594 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3595 while (retry_count) {
3596 if (signal_pending(current)) {
3601 * Rather than hide all in some function, I do this in
3602 * open coded manner. You see what this really does.
3603 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3605 mutex_lock(&set_limit_mutex);
3606 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3607 if (memlimit > val) {
3609 mutex_unlock(&set_limit_mutex);
3612 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3613 if (memswlimit < val)
3615 ret = res_counter_set_limit(&memcg->memsw, val);
3617 if (memlimit == val)
3618 memcg->memsw_is_minimum = true;
3620 memcg->memsw_is_minimum = false;
3622 mutex_unlock(&set_limit_mutex);
3627 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3628 MEM_CGROUP_RECLAIM_NOSWAP |
3629 MEM_CGROUP_RECLAIM_SHRINK,
3631 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3632 /* Usage is reduced ? */
3633 if (curusage >= oldusage)
3636 oldusage = curusage;
3638 if (!ret && enlarge)
3639 memcg_oom_recover(memcg);
3643 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3645 unsigned long *total_scanned)
3647 unsigned long nr_reclaimed = 0;
3648 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3649 unsigned long reclaimed;
3651 struct mem_cgroup_tree_per_zone *mctz;
3652 unsigned long long excess;
3653 unsigned long nr_scanned;
3658 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3660 * This loop can run a while, specially if mem_cgroup's continuously
3661 * keep exceeding their soft limit and putting the system under
3668 mz = mem_cgroup_largest_soft_limit_node(mctz);
3673 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3675 MEM_CGROUP_RECLAIM_SOFT,
3677 nr_reclaimed += reclaimed;
3678 *total_scanned += nr_scanned;
3679 spin_lock(&mctz->lock);
3682 * If we failed to reclaim anything from this memory cgroup
3683 * it is time to move on to the next cgroup
3689 * Loop until we find yet another one.
3691 * By the time we get the soft_limit lock
3692 * again, someone might have aded the
3693 * group back on the RB tree. Iterate to
3694 * make sure we get a different mem.
3695 * mem_cgroup_largest_soft_limit_node returns
3696 * NULL if no other cgroup is present on
3700 __mem_cgroup_largest_soft_limit_node(mctz);
3702 css_put(&next_mz->mem->css);
3703 else /* next_mz == NULL or other memcg */
3707 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3708 excess = res_counter_soft_limit_excess(&mz->mem->res);
3710 * One school of thought says that we should not add
3711 * back the node to the tree if reclaim returns 0.
3712 * But our reclaim could return 0, simply because due
3713 * to priority we are exposing a smaller subset of
3714 * memory to reclaim from. Consider this as a longer
3717 /* If excess == 0, no tree ops */
3718 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3719 spin_unlock(&mctz->lock);
3720 css_put(&mz->mem->css);
3723 * Could not reclaim anything and there are no more
3724 * mem cgroups to try or we seem to be looping without
3725 * reclaiming anything.
3727 if (!nr_reclaimed &&
3729 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3731 } while (!nr_reclaimed);
3733 css_put(&next_mz->mem->css);
3734 return nr_reclaimed;
3738 * This routine traverse page_cgroup in given list and drop them all.
3739 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3741 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3742 int node, int zid, enum lru_list lru)
3745 struct mem_cgroup_per_zone *mz;
3746 struct page_cgroup *pc, *busy;
3747 unsigned long flags, loop;
3748 struct list_head *list;
3751 zone = &NODE_DATA(node)->node_zones[zid];
3752 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3753 list = &mz->lists[lru];
3755 loop = MEM_CGROUP_ZSTAT(mz, lru);
3756 /* give some margin against EBUSY etc...*/
3763 spin_lock_irqsave(&zone->lru_lock, flags);
3764 if (list_empty(list)) {
3765 spin_unlock_irqrestore(&zone->lru_lock, flags);
3768 pc = list_entry(list->prev, struct page_cgroup, lru);
3770 list_move(&pc->lru, list);
3772 spin_unlock_irqrestore(&zone->lru_lock, flags);
3775 spin_unlock_irqrestore(&zone->lru_lock, flags);
3777 page = lookup_cgroup_page(pc);
3779 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3783 if (ret == -EBUSY || ret == -EINVAL) {
3784 /* found lock contention or "pc" is obsolete. */
3791 if (!ret && !list_empty(list))
3797 * make mem_cgroup's charge to be 0 if there is no task.
3798 * This enables deleting this mem_cgroup.
3800 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3803 int node, zid, shrink;
3804 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3805 struct cgroup *cgrp = memcg->css.cgroup;
3807 css_get(&memcg->css);
3810 /* should free all ? */
3816 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3819 if (signal_pending(current))
3821 /* This is for making all *used* pages to be on LRU. */
3822 lru_add_drain_all();
3823 drain_all_stock_sync(memcg);
3825 mem_cgroup_start_move(memcg);
3826 for_each_node_state(node, N_HIGH_MEMORY) {
3827 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3830 ret = mem_cgroup_force_empty_list(memcg,
3839 mem_cgroup_end_move(memcg);
3840 memcg_oom_recover(memcg);
3841 /* it seems parent cgroup doesn't have enough mem */
3845 /* "ret" should also be checked to ensure all lists are empty. */
3846 } while (memcg->res.usage > 0 || ret);
3848 css_put(&memcg->css);
3852 /* returns EBUSY if there is a task or if we come here twice. */
3853 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3857 /* we call try-to-free pages for make this cgroup empty */
3858 lru_add_drain_all();
3859 /* try to free all pages in this cgroup */
3861 while (nr_retries && memcg->res.usage > 0) {
3862 struct memcg_scanrecord rec;
3865 if (signal_pending(current)) {
3869 rec.context = SCAN_BY_SHRINK;
3872 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3876 /* maybe some writeback is necessary */
3877 congestion_wait(BLK_RW_ASYNC, HZ/10);
3882 /* try move_account...there may be some *locked* pages. */
3886 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3888 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3892 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3894 return mem_cgroup_from_cont(cont)->use_hierarchy;
3897 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3901 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3902 struct cgroup *parent = cont->parent;
3903 struct mem_cgroup *parent_memcg = NULL;
3906 parent_memcg = mem_cgroup_from_cont(parent);
3910 * If parent's use_hierarchy is set, we can't make any modifications
3911 * in the child subtrees. If it is unset, then the change can
3912 * occur, provided the current cgroup has no children.
3914 * For the root cgroup, parent_mem is NULL, we allow value to be
3915 * set if there are no children.
3917 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3918 (val == 1 || val == 0)) {
3919 if (list_empty(&cont->children))
3920 memcg->use_hierarchy = val;
3931 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3932 enum mem_cgroup_stat_index idx)
3934 struct mem_cgroup *iter;
3937 /* Per-cpu values can be negative, use a signed accumulator */
3938 for_each_mem_cgroup_tree(iter, memcg)
3939 val += mem_cgroup_read_stat(iter, idx);
3941 if (val < 0) /* race ? */
3946 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3950 if (!mem_cgroup_is_root(memcg)) {
3952 return res_counter_read_u64(&memcg->res, RES_USAGE);
3954 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3957 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3958 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3961 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3963 return val << PAGE_SHIFT;
3966 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3968 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3972 type = MEMFILE_TYPE(cft->private);
3973 name = MEMFILE_ATTR(cft->private);
3976 if (name == RES_USAGE)
3977 val = mem_cgroup_usage(memcg, false);
3979 val = res_counter_read_u64(&memcg->res, name);
3982 if (name == RES_USAGE)
3983 val = mem_cgroup_usage(memcg, true);
3985 val = res_counter_read_u64(&memcg->memsw, name);
3994 * The user of this function is...
3997 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4000 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4002 unsigned long long val;
4005 type = MEMFILE_TYPE(cft->private);
4006 name = MEMFILE_ATTR(cft->private);
4009 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4013 /* This function does all necessary parse...reuse it */
4014 ret = res_counter_memparse_write_strategy(buffer, &val);
4018 ret = mem_cgroup_resize_limit(memcg, val);
4020 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4022 case RES_SOFT_LIMIT:
4023 ret = res_counter_memparse_write_strategy(buffer, &val);
4027 * For memsw, soft limits are hard to implement in terms
4028 * of semantics, for now, we support soft limits for
4029 * control without swap
4032 ret = res_counter_set_soft_limit(&memcg->res, val);
4037 ret = -EINVAL; /* should be BUG() ? */
4043 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4044 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4046 struct cgroup *cgroup;
4047 unsigned long long min_limit, min_memsw_limit, tmp;
4049 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4050 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4051 cgroup = memcg->css.cgroup;
4052 if (!memcg->use_hierarchy)
4055 while (cgroup->parent) {
4056 cgroup = cgroup->parent;
4057 memcg = mem_cgroup_from_cont(cgroup);
4058 if (!memcg->use_hierarchy)
4060 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4061 min_limit = min(min_limit, tmp);
4062 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4063 min_memsw_limit = min(min_memsw_limit, tmp);
4066 *mem_limit = min_limit;
4067 *memsw_limit = min_memsw_limit;
4071 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4073 struct mem_cgroup *memcg;
4076 memcg = mem_cgroup_from_cont(cont);
4077 type = MEMFILE_TYPE(event);
4078 name = MEMFILE_ATTR(event);
4082 res_counter_reset_max(&memcg->res);
4084 res_counter_reset_max(&memcg->memsw);
4088 res_counter_reset_failcnt(&memcg->res);
4090 res_counter_reset_failcnt(&memcg->memsw);
4097 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4100 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4104 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4105 struct cftype *cft, u64 val)
4107 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4109 if (val >= (1 << NR_MOVE_TYPE))
4112 * We check this value several times in both in can_attach() and
4113 * attach(), so we need cgroup lock to prevent this value from being
4117 memcg->move_charge_at_immigrate = val;
4123 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4124 struct cftype *cft, u64 val)
4131 /* For read statistics */
4149 struct mcs_total_stat {
4150 s64 stat[NR_MCS_STAT];
4156 } memcg_stat_strings[NR_MCS_STAT] = {
4157 {"cache", "total_cache"},
4158 {"rss", "total_rss"},
4159 {"mapped_file", "total_mapped_file"},
4160 {"pgpgin", "total_pgpgin"},
4161 {"pgpgout", "total_pgpgout"},
4162 {"swap", "total_swap"},
4163 {"pgfault", "total_pgfault"},
4164 {"pgmajfault", "total_pgmajfault"},
4165 {"inactive_anon", "total_inactive_anon"},
4166 {"active_anon", "total_active_anon"},
4167 {"inactive_file", "total_inactive_file"},
4168 {"active_file", "total_active_file"},
4169 {"unevictable", "total_unevictable"}
4174 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4179 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4180 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4181 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4182 s->stat[MCS_RSS] += val * PAGE_SIZE;
4183 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4184 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4185 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4186 s->stat[MCS_PGPGIN] += val;
4187 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4188 s->stat[MCS_PGPGOUT] += val;
4189 if (do_swap_account) {
4190 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4191 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4193 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4194 s->stat[MCS_PGFAULT] += val;
4195 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4196 s->stat[MCS_PGMAJFAULT] += val;
4199 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4200 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4201 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4202 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4203 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4204 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4205 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4206 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4207 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4208 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4212 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4214 struct mem_cgroup *iter;
4216 for_each_mem_cgroup_tree(iter, memcg)
4217 mem_cgroup_get_local_stat(iter, s);
4221 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4224 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4225 unsigned long node_nr;
4226 struct cgroup *cont = m->private;
4227 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4229 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4230 seq_printf(m, "total=%lu", total_nr);
4231 for_each_node_state(nid, N_HIGH_MEMORY) {
4232 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4233 seq_printf(m, " N%d=%lu", nid, node_nr);
4237 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4238 seq_printf(m, "file=%lu", file_nr);
4239 for_each_node_state(nid, N_HIGH_MEMORY) {
4240 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4242 seq_printf(m, " N%d=%lu", nid, node_nr);
4246 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4247 seq_printf(m, "anon=%lu", anon_nr);
4248 for_each_node_state(nid, N_HIGH_MEMORY) {
4249 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4251 seq_printf(m, " N%d=%lu", nid, node_nr);
4255 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4256 seq_printf(m, "unevictable=%lu", unevictable_nr);
4257 for_each_node_state(nid, N_HIGH_MEMORY) {
4258 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4259 BIT(LRU_UNEVICTABLE));
4260 seq_printf(m, " N%d=%lu", nid, node_nr);
4265 #endif /* CONFIG_NUMA */
4267 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4268 struct cgroup_map_cb *cb)
4270 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4271 struct mcs_total_stat mystat;
4274 memset(&mystat, 0, sizeof(mystat));
4275 mem_cgroup_get_local_stat(mem_cont, &mystat);
4278 for (i = 0; i < NR_MCS_STAT; i++) {
4279 if (i == MCS_SWAP && !do_swap_account)
4281 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4284 /* Hierarchical information */
4286 unsigned long long limit, memsw_limit;
4287 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4288 cb->fill(cb, "hierarchical_memory_limit", limit);
4289 if (do_swap_account)
4290 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4293 memset(&mystat, 0, sizeof(mystat));
4294 mem_cgroup_get_total_stat(mem_cont, &mystat);
4295 for (i = 0; i < NR_MCS_STAT; i++) {
4296 if (i == MCS_SWAP && !do_swap_account)
4298 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4301 #ifdef CONFIG_DEBUG_VM
4302 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4306 struct mem_cgroup_per_zone *mz;
4307 unsigned long recent_rotated[2] = {0, 0};
4308 unsigned long recent_scanned[2] = {0, 0};
4310 for_each_online_node(nid)
4311 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4312 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4314 recent_rotated[0] +=
4315 mz->reclaim_stat.recent_rotated[0];
4316 recent_rotated[1] +=
4317 mz->reclaim_stat.recent_rotated[1];
4318 recent_scanned[0] +=
4319 mz->reclaim_stat.recent_scanned[0];
4320 recent_scanned[1] +=
4321 mz->reclaim_stat.recent_scanned[1];
4323 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4324 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4325 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4326 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4333 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4335 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4337 return mem_cgroup_swappiness(memcg);
4340 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4343 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4344 struct mem_cgroup *parent;
4349 if (cgrp->parent == NULL)
4352 parent = mem_cgroup_from_cont(cgrp->parent);
4356 /* If under hierarchy, only empty-root can set this value */
4357 if ((parent->use_hierarchy) ||
4358 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4363 memcg->swappiness = val;
4370 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4372 struct mem_cgroup_threshold_ary *t;
4378 t = rcu_dereference(memcg->thresholds.primary);
4380 t = rcu_dereference(memcg->memsw_thresholds.primary);
4385 usage = mem_cgroup_usage(memcg, swap);
4388 * current_threshold points to threshold just below usage.
4389 * If it's not true, a threshold was crossed after last
4390 * call of __mem_cgroup_threshold().
4392 i = t->current_threshold;
4395 * Iterate backward over array of thresholds starting from
4396 * current_threshold and check if a threshold is crossed.
4397 * If none of thresholds below usage is crossed, we read
4398 * only one element of the array here.
4400 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4401 eventfd_signal(t->entries[i].eventfd, 1);
4403 /* i = current_threshold + 1 */
4407 * Iterate forward over array of thresholds starting from
4408 * current_threshold+1 and check if a threshold is crossed.
4409 * If none of thresholds above usage is crossed, we read
4410 * only one element of the array here.
4412 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4413 eventfd_signal(t->entries[i].eventfd, 1);
4415 /* Update current_threshold */
4416 t->current_threshold = i - 1;
4421 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4424 __mem_cgroup_threshold(memcg, false);
4425 if (do_swap_account)
4426 __mem_cgroup_threshold(memcg, true);
4428 memcg = parent_mem_cgroup(memcg);
4432 static int compare_thresholds(const void *a, const void *b)
4434 const struct mem_cgroup_threshold *_a = a;
4435 const struct mem_cgroup_threshold *_b = b;
4437 return _a->threshold - _b->threshold;
4440 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4442 struct mem_cgroup_eventfd_list *ev;
4444 list_for_each_entry(ev, &memcg->oom_notify, list)
4445 eventfd_signal(ev->eventfd, 1);
4449 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4451 struct mem_cgroup *iter;
4453 for_each_mem_cgroup_tree(iter, memcg)
4454 mem_cgroup_oom_notify_cb(iter);
4457 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4458 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4460 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4461 struct mem_cgroup_thresholds *thresholds;
4462 struct mem_cgroup_threshold_ary *new;
4463 int type = MEMFILE_TYPE(cft->private);
4464 u64 threshold, usage;
4467 ret = res_counter_memparse_write_strategy(args, &threshold);
4471 mutex_lock(&memcg->thresholds_lock);
4474 thresholds = &memcg->thresholds;
4475 else if (type == _MEMSWAP)
4476 thresholds = &memcg->memsw_thresholds;
4480 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4482 /* Check if a threshold crossed before adding a new one */
4483 if (thresholds->primary)
4484 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4486 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4488 /* Allocate memory for new array of thresholds */
4489 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4497 /* Copy thresholds (if any) to new array */
4498 if (thresholds->primary) {
4499 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4500 sizeof(struct mem_cgroup_threshold));
4503 /* Add new threshold */
4504 new->entries[size - 1].eventfd = eventfd;
4505 new->entries[size - 1].threshold = threshold;
4507 /* Sort thresholds. Registering of new threshold isn't time-critical */
4508 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4509 compare_thresholds, NULL);
4511 /* Find current threshold */
4512 new->current_threshold = -1;
4513 for (i = 0; i < size; i++) {
4514 if (new->entries[i].threshold < usage) {
4516 * new->current_threshold will not be used until
4517 * rcu_assign_pointer(), so it's safe to increment
4520 ++new->current_threshold;
4524 /* Free old spare buffer and save old primary buffer as spare */
4525 kfree(thresholds->spare);
4526 thresholds->spare = thresholds->primary;
4528 rcu_assign_pointer(thresholds->primary, new);
4530 /* To be sure that nobody uses thresholds */
4534 mutex_unlock(&memcg->thresholds_lock);
4539 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4540 struct cftype *cft, struct eventfd_ctx *eventfd)
4542 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4543 struct mem_cgroup_thresholds *thresholds;
4544 struct mem_cgroup_threshold_ary *new;
4545 int type = MEMFILE_TYPE(cft->private);
4549 mutex_lock(&memcg->thresholds_lock);
4551 thresholds = &memcg->thresholds;
4552 else if (type == _MEMSWAP)
4553 thresholds = &memcg->memsw_thresholds;
4558 * Something went wrong if we trying to unregister a threshold
4559 * if we don't have thresholds
4561 BUG_ON(!thresholds);
4563 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4565 /* Check if a threshold crossed before removing */
4566 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4568 /* Calculate new number of threshold */
4570 for (i = 0; i < thresholds->primary->size; i++) {
4571 if (thresholds->primary->entries[i].eventfd != eventfd)
4575 new = thresholds->spare;
4577 /* Set thresholds array to NULL if we don't have thresholds */
4586 /* Copy thresholds and find current threshold */
4587 new->current_threshold = -1;
4588 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4589 if (thresholds->primary->entries[i].eventfd == eventfd)
4592 new->entries[j] = thresholds->primary->entries[i];
4593 if (new->entries[j].threshold < usage) {
4595 * new->current_threshold will not be used
4596 * until rcu_assign_pointer(), so it's safe to increment
4599 ++new->current_threshold;
4605 /* Swap primary and spare array */
4606 thresholds->spare = thresholds->primary;
4607 rcu_assign_pointer(thresholds->primary, new);
4609 /* To be sure that nobody uses thresholds */
4612 mutex_unlock(&memcg->thresholds_lock);
4615 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4616 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4618 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4619 struct mem_cgroup_eventfd_list *event;
4620 int type = MEMFILE_TYPE(cft->private);
4622 BUG_ON(type != _OOM_TYPE);
4623 event = kmalloc(sizeof(*event), GFP_KERNEL);
4627 spin_lock(&memcg_oom_lock);
4629 event->eventfd = eventfd;
4630 list_add(&event->list, &memcg->oom_notify);
4632 /* already in OOM ? */
4633 if (atomic_read(&memcg->under_oom))
4634 eventfd_signal(eventfd, 1);
4635 spin_unlock(&memcg_oom_lock);
4640 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4641 struct cftype *cft, struct eventfd_ctx *eventfd)
4643 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4644 struct mem_cgroup_eventfd_list *ev, *tmp;
4645 int type = MEMFILE_TYPE(cft->private);
4647 BUG_ON(type != _OOM_TYPE);
4649 spin_lock(&memcg_oom_lock);
4651 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4652 if (ev->eventfd == eventfd) {
4653 list_del(&ev->list);
4658 spin_unlock(&memcg_oom_lock);
4661 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4662 struct cftype *cft, struct cgroup_map_cb *cb)
4664 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4666 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4668 if (atomic_read(&memcg->under_oom))
4669 cb->fill(cb, "under_oom", 1);
4671 cb->fill(cb, "under_oom", 0);
4675 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4676 struct cftype *cft, u64 val)
4678 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4679 struct mem_cgroup *parent;
4681 /* cannot set to root cgroup and only 0 and 1 are allowed */
4682 if (!cgrp->parent || !((val == 0) || (val == 1)))
4685 parent = mem_cgroup_from_cont(cgrp->parent);
4688 /* oom-kill-disable is a flag for subhierarchy. */
4689 if ((parent->use_hierarchy) ||
4690 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4694 memcg->oom_kill_disable = val;
4696 memcg_oom_recover(memcg);
4702 static const struct file_operations mem_control_numa_stat_file_operations = {
4704 .llseek = seq_lseek,
4705 .release = single_release,
4708 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4710 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4712 file->f_op = &mem_control_numa_stat_file_operations;
4713 return single_open(file, mem_control_numa_stat_show, cont);
4715 #endif /* CONFIG_NUMA */
4717 static int mem_cgroup_vmscan_stat_read(struct cgroup *cgrp,
4719 struct cgroup_map_cb *cb)
4721 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4725 for (i = 0; i < NR_SCANSTATS; i++) {
4726 strcpy(string, scanstat_string[i]);
4727 strcat(string, SCANSTAT_WORD_LIMIT);
4728 cb->fill(cb, string, memcg->scanstat.stats[SCAN_BY_LIMIT][i]);
4731 for (i = 0; i < NR_SCANSTATS; i++) {
4732 strcpy(string, scanstat_string[i]);
4733 strcat(string, SCANSTAT_WORD_SYSTEM);
4734 cb->fill(cb, string, memcg->scanstat.stats[SCAN_BY_SYSTEM][i]);
4737 for (i = 0; i < NR_SCANSTATS; i++) {
4738 strcpy(string, scanstat_string[i]);
4739 strcat(string, SCANSTAT_WORD_LIMIT);
4740 strcat(string, SCANSTAT_WORD_HIERARCHY);
4742 string, memcg->scanstat.rootstats[SCAN_BY_LIMIT][i]);
4744 for (i = 0; i < NR_SCANSTATS; i++) {
4745 strcpy(string, scanstat_string[i]);
4746 strcat(string, SCANSTAT_WORD_SYSTEM);
4747 strcat(string, SCANSTAT_WORD_HIERARCHY);
4749 string, memcg->scanstat.rootstats[SCAN_BY_SYSTEM][i]);
4754 static int mem_cgroup_reset_vmscan_stat(struct cgroup *cgrp,
4757 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4759 spin_lock(&memcg->scanstat.lock);
4760 memset(&memcg->scanstat.stats, 0, sizeof(memcg->scanstat.stats));
4761 memset(&memcg->scanstat.rootstats,
4762 0, sizeof(memcg->scanstat.rootstats));
4763 spin_unlock(&memcg->scanstat.lock);
4768 static struct cftype mem_cgroup_files[] = {
4770 .name = "usage_in_bytes",
4771 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4772 .read_u64 = mem_cgroup_read,
4773 .register_event = mem_cgroup_usage_register_event,
4774 .unregister_event = mem_cgroup_usage_unregister_event,
4777 .name = "max_usage_in_bytes",
4778 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4779 .trigger = mem_cgroup_reset,
4780 .read_u64 = mem_cgroup_read,
4783 .name = "limit_in_bytes",
4784 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4785 .write_string = mem_cgroup_write,
4786 .read_u64 = mem_cgroup_read,
4789 .name = "soft_limit_in_bytes",
4790 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4791 .write_string = mem_cgroup_write,
4792 .read_u64 = mem_cgroup_read,
4796 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4797 .trigger = mem_cgroup_reset,
4798 .read_u64 = mem_cgroup_read,
4802 .read_map = mem_control_stat_show,
4805 .name = "force_empty",
4806 .trigger = mem_cgroup_force_empty_write,
4809 .name = "use_hierarchy",
4810 .write_u64 = mem_cgroup_hierarchy_write,
4811 .read_u64 = mem_cgroup_hierarchy_read,
4814 .name = "swappiness",
4815 .read_u64 = mem_cgroup_swappiness_read,
4816 .write_u64 = mem_cgroup_swappiness_write,
4819 .name = "move_charge_at_immigrate",
4820 .read_u64 = mem_cgroup_move_charge_read,
4821 .write_u64 = mem_cgroup_move_charge_write,
4824 .name = "oom_control",
4825 .read_map = mem_cgroup_oom_control_read,
4826 .write_u64 = mem_cgroup_oom_control_write,
4827 .register_event = mem_cgroup_oom_register_event,
4828 .unregister_event = mem_cgroup_oom_unregister_event,
4829 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4833 .name = "numa_stat",
4834 .open = mem_control_numa_stat_open,
4839 .name = "vmscan_stat",
4840 .read_map = mem_cgroup_vmscan_stat_read,
4841 .trigger = mem_cgroup_reset_vmscan_stat,
4845 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4846 static struct cftype memsw_cgroup_files[] = {
4848 .name = "memsw.usage_in_bytes",
4849 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4850 .read_u64 = mem_cgroup_read,
4851 .register_event = mem_cgroup_usage_register_event,
4852 .unregister_event = mem_cgroup_usage_unregister_event,
4855 .name = "memsw.max_usage_in_bytes",
4856 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4857 .trigger = mem_cgroup_reset,
4858 .read_u64 = mem_cgroup_read,
4861 .name = "memsw.limit_in_bytes",
4862 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4863 .write_string = mem_cgroup_write,
4864 .read_u64 = mem_cgroup_read,
4867 .name = "memsw.failcnt",
4868 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4869 .trigger = mem_cgroup_reset,
4870 .read_u64 = mem_cgroup_read,
4874 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4876 if (!do_swap_account)
4878 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4879 ARRAY_SIZE(memsw_cgroup_files));
4882 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4888 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4890 struct mem_cgroup_per_node *pn;
4891 struct mem_cgroup_per_zone *mz;
4893 int zone, tmp = node;
4895 * This routine is called against possible nodes.
4896 * But it's BUG to call kmalloc() against offline node.
4898 * TODO: this routine can waste much memory for nodes which will
4899 * never be onlined. It's better to use memory hotplug callback
4902 if (!node_state(node, N_NORMAL_MEMORY))
4904 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4908 memcg->info.nodeinfo[node] = pn;
4909 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4910 mz = &pn->zoneinfo[zone];
4912 INIT_LIST_HEAD(&mz->lists[l]);
4913 mz->usage_in_excess = 0;
4914 mz->on_tree = false;
4920 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4922 kfree(memcg->info.nodeinfo[node]);
4925 static struct mem_cgroup *mem_cgroup_alloc(void)
4927 struct mem_cgroup *mem;
4928 int size = sizeof(struct mem_cgroup);
4930 /* Can be very big if MAX_NUMNODES is very big */
4931 if (size < PAGE_SIZE)
4932 mem = kzalloc(size, GFP_KERNEL);
4934 mem = vzalloc(size);
4939 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4942 spin_lock_init(&mem->pcp_counter_lock);
4946 if (size < PAGE_SIZE)
4954 * At destroying mem_cgroup, references from swap_cgroup can remain.
4955 * (scanning all at force_empty is too costly...)
4957 * Instead of clearing all references at force_empty, we remember
4958 * the number of reference from swap_cgroup and free mem_cgroup when
4959 * it goes down to 0.
4961 * Removal of cgroup itself succeeds regardless of refs from swap.
4964 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4968 mem_cgroup_remove_from_trees(memcg);
4969 free_css_id(&mem_cgroup_subsys, &memcg->css);
4971 for_each_node_state(node, N_POSSIBLE)
4972 free_mem_cgroup_per_zone_info(memcg, node);
4974 free_percpu(memcg->stat);
4975 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4981 static void mem_cgroup_get(struct mem_cgroup *memcg)
4983 atomic_inc(&memcg->refcnt);
4986 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4988 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4989 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4990 __mem_cgroup_free(memcg);
4992 mem_cgroup_put(parent);
4996 static void mem_cgroup_put(struct mem_cgroup *memcg)
4998 __mem_cgroup_put(memcg, 1);
5002 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5004 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
5006 if (!memcg->res.parent)
5008 return mem_cgroup_from_res_counter(memcg->res.parent, res);
5011 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5012 static void __init enable_swap_cgroup(void)
5014 if (!mem_cgroup_disabled() && really_do_swap_account)
5015 do_swap_account = 1;
5018 static void __init enable_swap_cgroup(void)
5023 static int mem_cgroup_soft_limit_tree_init(void)
5025 struct mem_cgroup_tree_per_node *rtpn;
5026 struct mem_cgroup_tree_per_zone *rtpz;
5027 int tmp, node, zone;
5029 for_each_node_state(node, N_POSSIBLE) {
5031 if (!node_state(node, N_NORMAL_MEMORY))
5033 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5037 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5039 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5040 rtpz = &rtpn->rb_tree_per_zone[zone];
5041 rtpz->rb_root = RB_ROOT;
5042 spin_lock_init(&rtpz->lock);
5048 static struct cgroup_subsys_state * __ref
5049 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5051 struct mem_cgroup *memcg, *parent;
5052 long error = -ENOMEM;
5055 memcg = mem_cgroup_alloc();
5057 return ERR_PTR(error);
5059 for_each_node_state(node, N_POSSIBLE)
5060 if (alloc_mem_cgroup_per_zone_info(memcg, node))
5064 if (cont->parent == NULL) {
5066 enable_swap_cgroup();
5068 root_mem_cgroup = memcg;
5069 if (mem_cgroup_soft_limit_tree_init())
5071 for_each_possible_cpu(cpu) {
5072 struct memcg_stock_pcp *stock =
5073 &per_cpu(memcg_stock, cpu);
5074 INIT_WORK(&stock->work, drain_local_stock);
5076 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5078 parent = mem_cgroup_from_cont(cont->parent);
5079 memcg->use_hierarchy = parent->use_hierarchy;
5080 memcg->oom_kill_disable = parent->oom_kill_disable;
5083 if (parent && parent->use_hierarchy) {
5084 res_counter_init(&memcg->res, &parent->res);
5085 res_counter_init(&memcg->memsw, &parent->memsw);
5087 * We increment refcnt of the parent to ensure that we can
5088 * safely access it on res_counter_charge/uncharge.
5089 * This refcnt will be decremented when freeing this
5090 * mem_cgroup(see mem_cgroup_put).
5092 mem_cgroup_get(parent);
5094 res_counter_init(&memcg->res, NULL);
5095 res_counter_init(&memcg->memsw, NULL);
5097 memcg->last_scanned_child = 0;
5098 memcg->last_scanned_node = MAX_NUMNODES;
5099 INIT_LIST_HEAD(&memcg->oom_notify);
5102 memcg->swappiness = mem_cgroup_swappiness(parent);
5103 atomic_set(&memcg->refcnt, 1);
5104 memcg->move_charge_at_immigrate = 0;
5105 mutex_init(&memcg->thresholds_lock);
5106 spin_lock_init(&memcg->scanstat.lock);
5109 __mem_cgroup_free(memcg);
5110 root_mem_cgroup = NULL;
5111 return ERR_PTR(error);
5114 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5115 struct cgroup *cont)
5117 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5119 return mem_cgroup_force_empty(memcg, false);
5122 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5123 struct cgroup *cont)
5125 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5127 mem_cgroup_put(memcg);
5130 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5131 struct cgroup *cont)
5135 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5136 ARRAY_SIZE(mem_cgroup_files));
5139 ret = register_memsw_files(cont, ss);
5144 /* Handlers for move charge at task migration. */
5145 #define PRECHARGE_COUNT_AT_ONCE 256
5146 static int mem_cgroup_do_precharge(unsigned long count)
5149 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5150 struct mem_cgroup *memcg = mc.to;
5152 if (mem_cgroup_is_root(memcg)) {
5153 mc.precharge += count;
5154 /* we don't need css_get for root */
5157 /* try to charge at once */
5159 struct res_counter *dummy;
5161 * "memcg" cannot be under rmdir() because we've already checked
5162 * by cgroup_lock_live_cgroup() that it is not removed and we
5163 * are still under the same cgroup_mutex. So we can postpone
5166 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5168 if (do_swap_account && res_counter_charge(&memcg->memsw,
5169 PAGE_SIZE * count, &dummy)) {
5170 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5173 mc.precharge += count;
5177 /* fall back to one by one charge */
5179 if (signal_pending(current)) {
5183 if (!batch_count--) {
5184 batch_count = PRECHARGE_COUNT_AT_ONCE;
5187 ret = __mem_cgroup_try_charge(NULL,
5188 GFP_KERNEL, 1, &memcg, false);
5190 /* mem_cgroup_clear_mc() will do uncharge later */
5198 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5199 * @vma: the vma the pte to be checked belongs
5200 * @addr: the address corresponding to the pte to be checked
5201 * @ptent: the pte to be checked
5202 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5205 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5206 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5207 * move charge. if @target is not NULL, the page is stored in target->page
5208 * with extra refcnt got(Callers should handle it).
5209 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5210 * target for charge migration. if @target is not NULL, the entry is stored
5213 * Called with pte lock held.
5220 enum mc_target_type {
5221 MC_TARGET_NONE, /* not used */
5226 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5227 unsigned long addr, pte_t ptent)
5229 struct page *page = vm_normal_page(vma, addr, ptent);
5231 if (!page || !page_mapped(page))
5233 if (PageAnon(page)) {
5234 /* we don't move shared anon */
5235 if (!move_anon() || page_mapcount(page) > 2)
5237 } else if (!move_file())
5238 /* we ignore mapcount for file pages */
5240 if (!get_page_unless_zero(page))
5246 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5247 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5250 struct page *page = NULL;
5251 swp_entry_t ent = pte_to_swp_entry(ptent);
5253 if (!move_anon() || non_swap_entry(ent))
5255 usage_count = mem_cgroup_count_swap_user(ent, &page);
5256 if (usage_count > 1) { /* we don't move shared anon */
5261 if (do_swap_account)
5262 entry->val = ent.val;
5267 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5268 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5270 struct page *page = NULL;
5271 struct inode *inode;
5272 struct address_space *mapping;
5275 if (!vma->vm_file) /* anonymous vma */
5280 inode = vma->vm_file->f_path.dentry->d_inode;
5281 mapping = vma->vm_file->f_mapping;
5282 if (pte_none(ptent))
5283 pgoff = linear_page_index(vma, addr);
5284 else /* pte_file(ptent) is true */
5285 pgoff = pte_to_pgoff(ptent);
5287 /* page is moved even if it's not RSS of this task(page-faulted). */
5288 page = find_get_page(mapping, pgoff);
5291 /* shmem/tmpfs may report page out on swap: account for that too. */
5292 if (radix_tree_exceptional_entry(page)) {
5293 swp_entry_t swap = radix_to_swp_entry(page);
5294 if (do_swap_account)
5296 page = find_get_page(&swapper_space, swap.val);
5302 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5303 unsigned long addr, pte_t ptent, union mc_target *target)
5305 struct page *page = NULL;
5306 struct page_cgroup *pc;
5308 swp_entry_t ent = { .val = 0 };
5310 if (pte_present(ptent))
5311 page = mc_handle_present_pte(vma, addr, ptent);
5312 else if (is_swap_pte(ptent))
5313 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5314 else if (pte_none(ptent) || pte_file(ptent))
5315 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5317 if (!page && !ent.val)
5320 pc = lookup_page_cgroup(page);
5322 * Do only loose check w/o page_cgroup lock.
5323 * mem_cgroup_move_account() checks the pc is valid or not under
5326 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5327 ret = MC_TARGET_PAGE;
5329 target->page = page;
5331 if (!ret || !target)
5334 /* There is a swap entry and a page doesn't exist or isn't charged */
5335 if (ent.val && !ret &&
5336 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5337 ret = MC_TARGET_SWAP;
5344 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5345 unsigned long addr, unsigned long end,
5346 struct mm_walk *walk)
5348 struct vm_area_struct *vma = walk->private;
5352 split_huge_page_pmd(walk->mm, pmd);
5354 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5355 for (; addr != end; pte++, addr += PAGE_SIZE)
5356 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5357 mc.precharge++; /* increment precharge temporarily */
5358 pte_unmap_unlock(pte - 1, ptl);
5364 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5366 unsigned long precharge;
5367 struct vm_area_struct *vma;
5369 down_read(&mm->mmap_sem);
5370 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5371 struct mm_walk mem_cgroup_count_precharge_walk = {
5372 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5376 if (is_vm_hugetlb_page(vma))
5378 walk_page_range(vma->vm_start, vma->vm_end,
5379 &mem_cgroup_count_precharge_walk);
5381 up_read(&mm->mmap_sem);
5383 precharge = mc.precharge;
5389 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5391 unsigned long precharge = mem_cgroup_count_precharge(mm);
5393 VM_BUG_ON(mc.moving_task);
5394 mc.moving_task = current;
5395 return mem_cgroup_do_precharge(precharge);
5398 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5399 static void __mem_cgroup_clear_mc(void)
5401 struct mem_cgroup *from = mc.from;
5402 struct mem_cgroup *to = mc.to;
5404 /* we must uncharge all the leftover precharges from mc.to */
5406 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5410 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5411 * we must uncharge here.
5413 if (mc.moved_charge) {
5414 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5415 mc.moved_charge = 0;
5417 /* we must fixup refcnts and charges */
5418 if (mc.moved_swap) {
5419 /* uncharge swap account from the old cgroup */
5420 if (!mem_cgroup_is_root(mc.from))
5421 res_counter_uncharge(&mc.from->memsw,
5422 PAGE_SIZE * mc.moved_swap);
5423 __mem_cgroup_put(mc.from, mc.moved_swap);
5425 if (!mem_cgroup_is_root(mc.to)) {
5427 * we charged both to->res and to->memsw, so we should
5430 res_counter_uncharge(&mc.to->res,
5431 PAGE_SIZE * mc.moved_swap);
5433 /* we've already done mem_cgroup_get(mc.to) */
5436 memcg_oom_recover(from);
5437 memcg_oom_recover(to);
5438 wake_up_all(&mc.waitq);
5441 static void mem_cgroup_clear_mc(void)
5443 struct mem_cgroup *from = mc.from;
5446 * we must clear moving_task before waking up waiters at the end of
5449 mc.moving_task = NULL;
5450 __mem_cgroup_clear_mc();
5451 spin_lock(&mc.lock);
5454 spin_unlock(&mc.lock);
5455 mem_cgroup_end_move(from);
5458 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5459 struct cgroup *cgroup,
5460 struct task_struct *p)
5463 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5465 if (memcg->move_charge_at_immigrate) {
5466 struct mm_struct *mm;
5467 struct mem_cgroup *from = mem_cgroup_from_task(p);
5469 VM_BUG_ON(from == memcg);
5471 mm = get_task_mm(p);
5474 /* We move charges only when we move a owner of the mm */
5475 if (mm->owner == p) {
5478 VM_BUG_ON(mc.precharge);
5479 VM_BUG_ON(mc.moved_charge);
5480 VM_BUG_ON(mc.moved_swap);
5481 mem_cgroup_start_move(from);
5482 spin_lock(&mc.lock);
5485 spin_unlock(&mc.lock);
5486 /* We set mc.moving_task later */
5488 ret = mem_cgroup_precharge_mc(mm);
5490 mem_cgroup_clear_mc();
5497 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5498 struct cgroup *cgroup,
5499 struct task_struct *p)
5501 mem_cgroup_clear_mc();
5504 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5505 unsigned long addr, unsigned long end,
5506 struct mm_walk *walk)
5509 struct vm_area_struct *vma = walk->private;
5513 split_huge_page_pmd(walk->mm, pmd);
5515 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5516 for (; addr != end; addr += PAGE_SIZE) {
5517 pte_t ptent = *(pte++);
5518 union mc_target target;
5521 struct page_cgroup *pc;
5527 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5529 case MC_TARGET_PAGE:
5531 if (isolate_lru_page(page))
5533 pc = lookup_page_cgroup(page);
5534 if (!mem_cgroup_move_account(page, 1, pc,
5535 mc.from, mc.to, false)) {
5537 /* we uncharge from mc.from later. */
5540 putback_lru_page(page);
5541 put: /* is_target_pte_for_mc() gets the page */
5544 case MC_TARGET_SWAP:
5546 if (!mem_cgroup_move_swap_account(ent,
5547 mc.from, mc.to, false)) {
5549 /* we fixup refcnts and charges later. */
5557 pte_unmap_unlock(pte - 1, ptl);
5562 * We have consumed all precharges we got in can_attach().
5563 * We try charge one by one, but don't do any additional
5564 * charges to mc.to if we have failed in charge once in attach()
5567 ret = mem_cgroup_do_precharge(1);
5575 static void mem_cgroup_move_charge(struct mm_struct *mm)
5577 struct vm_area_struct *vma;
5579 lru_add_drain_all();
5581 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5583 * Someone who are holding the mmap_sem might be waiting in
5584 * waitq. So we cancel all extra charges, wake up all waiters,
5585 * and retry. Because we cancel precharges, we might not be able
5586 * to move enough charges, but moving charge is a best-effort
5587 * feature anyway, so it wouldn't be a big problem.
5589 __mem_cgroup_clear_mc();
5593 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5595 struct mm_walk mem_cgroup_move_charge_walk = {
5596 .pmd_entry = mem_cgroup_move_charge_pte_range,
5600 if (is_vm_hugetlb_page(vma))
5602 ret = walk_page_range(vma->vm_start, vma->vm_end,
5603 &mem_cgroup_move_charge_walk);
5606 * means we have consumed all precharges and failed in
5607 * doing additional charge. Just abandon here.
5611 up_read(&mm->mmap_sem);
5614 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5615 struct cgroup *cont,
5616 struct cgroup *old_cont,
5617 struct task_struct *p)
5619 struct mm_struct *mm = get_task_mm(p);
5623 mem_cgroup_move_charge(mm);
5628 mem_cgroup_clear_mc();
5630 #else /* !CONFIG_MMU */
5631 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5632 struct cgroup *cgroup,
5633 struct task_struct *p)
5637 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5638 struct cgroup *cgroup,
5639 struct task_struct *p)
5642 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5643 struct cgroup *cont,
5644 struct cgroup *old_cont,
5645 struct task_struct *p)
5650 struct cgroup_subsys mem_cgroup_subsys = {
5652 .subsys_id = mem_cgroup_subsys_id,
5653 .create = mem_cgroup_create,
5654 .pre_destroy = mem_cgroup_pre_destroy,
5655 .destroy = mem_cgroup_destroy,
5656 .populate = mem_cgroup_populate,
5657 .can_attach = mem_cgroup_can_attach,
5658 .cancel_attach = mem_cgroup_cancel_attach,
5659 .attach = mem_cgroup_move_task,
5664 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5665 static int __init enable_swap_account(char *s)
5667 /* consider enabled if no parameter or 1 is given */
5668 if (!strcmp(s, "1"))
5669 really_do_swap_account = 1;
5670 else if (!strcmp(s, "0"))
5671 really_do_swap_account = 0;
5674 __setup("swapaccount=", enable_swap_account);