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/mutex.h>
37 #include <linux/rbtree.h>
38 #include <linux/shmem_fs.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 *mem);
206 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
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 *mem);
412 static void mem_cgroup_put(struct mem_cgroup *mem);
413 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
414 static void drain_all_stock_async(struct mem_cgroup *mem);
416 static struct mem_cgroup_per_zone *
417 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
419 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
422 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
427 static struct mem_cgroup_per_zone *
428 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
430 int nid = page_to_nid(page);
431 int zid = page_zonenum(page);
433 return mem_cgroup_zoneinfo(mem, 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 *mem,
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 *mem,
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 *mem,
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(mem, mz, mctz);
503 spin_unlock(&mctz->lock);
507 static void mem_cgroup_update_tree(struct mem_cgroup *mem, 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 (; mem; mem = parent_mem_cgroup(mem)) {
521 mz = mem_cgroup_zoneinfo(mem, nid, zid);
522 excess = res_counter_soft_limit_excess(&mem->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(mem, 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(mem, mz, mctz, excess);
537 spin_unlock(&mctz->lock);
542 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
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(mem, node, zone);
551 mctz = soft_limit_tree_node_zone(node, zone);
552 mem_cgroup_remove_exceeded(mem, 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 *mem,
614 enum mem_cgroup_stat_index idx)
620 for_each_online_cpu(cpu)
621 val += per_cpu(mem->stat->count[idx], cpu);
622 #ifdef CONFIG_HOTPLUG_CPU
623 spin_lock(&mem->pcp_counter_lock);
624 val += mem->nocpu_base.count[idx];
625 spin_unlock(&mem->pcp_counter_lock);
631 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
634 int val = (charge) ? 1 : -1;
635 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
638 void mem_cgroup_pgfault(struct mem_cgroup *mem, int val)
640 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
643 void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val)
645 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
648 static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
649 enum mem_cgroup_events_index idx)
651 unsigned long val = 0;
654 for_each_online_cpu(cpu)
655 val += per_cpu(mem->stat->events[idx], cpu);
656 #ifdef CONFIG_HOTPLUG_CPU
657 spin_lock(&mem->pcp_counter_lock);
658 val += mem->nocpu_base.events[idx];
659 spin_unlock(&mem->pcp_counter_lock);
664 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
665 bool file, int nr_pages)
670 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
672 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
674 /* pagein of a big page is an event. So, ignore page size */
676 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
678 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
679 nr_pages = -nr_pages; /* for event */
682 __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
688 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *mem, int nid, int zid,
689 unsigned int lru_mask)
691 struct mem_cgroup_per_zone *mz;
693 unsigned long ret = 0;
695 mz = mem_cgroup_zoneinfo(mem, nid, zid);
698 if (BIT(l) & lru_mask)
699 ret += MEM_CGROUP_ZSTAT(mz, l);
705 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *mem,
706 int nid, unsigned int lru_mask)
711 for (zid = 0; zid < MAX_NR_ZONES; zid++)
712 total += mem_cgroup_zone_nr_lru_pages(mem, nid, zid, lru_mask);
717 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *mem,
718 unsigned int lru_mask)
723 for_each_node_state(nid, N_HIGH_MEMORY)
724 total += mem_cgroup_node_nr_lru_pages(mem, nid, lru_mask);
728 static bool __memcg_event_check(struct mem_cgroup *mem, int target)
730 unsigned long val, next;
732 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
733 next = this_cpu_read(mem->stat->targets[target]);
734 /* from time_after() in jiffies.h */
735 return ((long)next - (long)val < 0);
738 static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target)
740 unsigned long val, next;
742 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
745 case MEM_CGROUP_TARGET_THRESH:
746 next = val + THRESHOLDS_EVENTS_TARGET;
748 case MEM_CGROUP_TARGET_SOFTLIMIT:
749 next = val + SOFTLIMIT_EVENTS_TARGET;
751 case MEM_CGROUP_TARGET_NUMAINFO:
752 next = val + NUMAINFO_EVENTS_TARGET;
758 this_cpu_write(mem->stat->targets[target], next);
762 * Check events in order.
765 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
767 /* threshold event is triggered in finer grain than soft limit */
768 if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) {
769 mem_cgroup_threshold(mem);
770 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH);
771 if (unlikely(__memcg_event_check(mem,
772 MEM_CGROUP_TARGET_SOFTLIMIT))) {
773 mem_cgroup_update_tree(mem, page);
774 __mem_cgroup_target_update(mem,
775 MEM_CGROUP_TARGET_SOFTLIMIT);
778 if (unlikely(__memcg_event_check(mem,
779 MEM_CGROUP_TARGET_NUMAINFO))) {
780 atomic_inc(&mem->numainfo_events);
781 __mem_cgroup_target_update(mem,
782 MEM_CGROUP_TARGET_NUMAINFO);
788 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
790 return container_of(cgroup_subsys_state(cont,
791 mem_cgroup_subsys_id), struct mem_cgroup,
795 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
798 * mm_update_next_owner() may clear mm->owner to NULL
799 * if it races with swapoff, page migration, etc.
800 * So this can be called with p == NULL.
805 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
806 struct mem_cgroup, css);
809 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
811 struct mem_cgroup *mem = NULL;
816 * Because we have no locks, mm->owner's may be being moved to other
817 * cgroup. We use css_tryget() here even if this looks
818 * pessimistic (rather than adding locks here).
822 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
825 } while (!css_tryget(&mem->css));
830 /* The caller has to guarantee "mem" exists before calling this */
831 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
833 struct cgroup_subsys_state *css;
836 if (!mem) /* ROOT cgroup has the smallest ID */
837 return root_mem_cgroup; /*css_put/get against root is ignored*/
838 if (!mem->use_hierarchy) {
839 if (css_tryget(&mem->css))
845 * searching a memory cgroup which has the smallest ID under given
846 * ROOT cgroup. (ID >= 1)
848 css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
849 if (css && css_tryget(css))
850 mem = container_of(css, struct mem_cgroup, css);
857 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
858 struct mem_cgroup *root,
861 int nextid = css_id(&iter->css) + 1;
864 struct cgroup_subsys_state *css;
866 hierarchy_used = iter->use_hierarchy;
869 /* If no ROOT, walk all, ignore hierarchy */
870 if (!cond || (root && !hierarchy_used))
874 root = root_mem_cgroup;
880 css = css_get_next(&mem_cgroup_subsys, nextid,
882 if (css && css_tryget(css))
883 iter = container_of(css, struct mem_cgroup, css);
885 /* If css is NULL, no more cgroups will be found */
887 } while (css && !iter);
892 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
893 * be careful that "break" loop is not allowed. We have reference count.
894 * Instead of that modify "cond" to be false and "continue" to exit the loop.
896 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
897 for (iter = mem_cgroup_start_loop(root);\
899 iter = mem_cgroup_get_next(iter, root, cond))
901 #define for_each_mem_cgroup_tree(iter, root) \
902 for_each_mem_cgroup_tree_cond(iter, root, true)
904 #define for_each_mem_cgroup_all(iter) \
905 for_each_mem_cgroup_tree_cond(iter, NULL, true)
908 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
910 return (mem == root_mem_cgroup);
913 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
915 struct mem_cgroup *mem;
921 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
927 mem_cgroup_pgmajfault(mem, 1);
930 mem_cgroup_pgfault(mem, 1);
938 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
941 * Following LRU functions are allowed to be used without PCG_LOCK.
942 * Operations are called by routine of global LRU independently from memcg.
943 * What we have to take care of here is validness of pc->mem_cgroup.
945 * Changes to pc->mem_cgroup happens when
948 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
949 * It is added to LRU before charge.
950 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
951 * When moving account, the page is not on LRU. It's isolated.
954 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
956 struct page_cgroup *pc;
957 struct mem_cgroup_per_zone *mz;
959 if (mem_cgroup_disabled())
961 pc = lookup_page_cgroup(page);
962 /* can happen while we handle swapcache. */
963 if (!TestClearPageCgroupAcctLRU(pc))
965 VM_BUG_ON(!pc->mem_cgroup);
967 * We don't check PCG_USED bit. It's cleared when the "page" is finally
968 * removed from global LRU.
970 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
971 /* huge page split is done under lru_lock. so, we have no races. */
972 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
973 if (mem_cgroup_is_root(pc->mem_cgroup))
975 VM_BUG_ON(list_empty(&pc->lru));
976 list_del_init(&pc->lru);
979 void mem_cgroup_del_lru(struct page *page)
981 mem_cgroup_del_lru_list(page, page_lru(page));
985 * Writeback is about to end against a page which has been marked for immediate
986 * reclaim. If it still appears to be reclaimable, move it to the tail of the
989 void mem_cgroup_rotate_reclaimable_page(struct page *page)
991 struct mem_cgroup_per_zone *mz;
992 struct page_cgroup *pc;
993 enum lru_list lru = page_lru(page);
995 if (mem_cgroup_disabled())
998 pc = lookup_page_cgroup(page);
999 /* unused or root page is not rotated. */
1000 if (!PageCgroupUsed(pc))
1002 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1004 if (mem_cgroup_is_root(pc->mem_cgroup))
1006 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1007 list_move_tail(&pc->lru, &mz->lists[lru]);
1010 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1012 struct mem_cgroup_per_zone *mz;
1013 struct page_cgroup *pc;
1015 if (mem_cgroup_disabled())
1018 pc = lookup_page_cgroup(page);
1019 /* unused or root page is not rotated. */
1020 if (!PageCgroupUsed(pc))
1022 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1024 if (mem_cgroup_is_root(pc->mem_cgroup))
1026 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1027 list_move(&pc->lru, &mz->lists[lru]);
1030 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1032 struct page_cgroup *pc;
1033 struct mem_cgroup_per_zone *mz;
1035 if (mem_cgroup_disabled())
1037 pc = lookup_page_cgroup(page);
1038 VM_BUG_ON(PageCgroupAcctLRU(pc));
1039 if (!PageCgroupUsed(pc))
1041 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1043 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1044 /* huge page split is done under lru_lock. so, we have no races. */
1045 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1046 SetPageCgroupAcctLRU(pc);
1047 if (mem_cgroup_is_root(pc->mem_cgroup))
1049 list_add(&pc->lru, &mz->lists[lru]);
1053 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1054 * while it's linked to lru because the page may be reused after it's fully
1055 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1056 * It's done under lock_page and expected that zone->lru_lock isnever held.
1058 static void mem_cgroup_lru_del_before_commit(struct page *page)
1060 unsigned long flags;
1061 struct zone *zone = page_zone(page);
1062 struct page_cgroup *pc = lookup_page_cgroup(page);
1065 * Doing this check without taking ->lru_lock seems wrong but this
1066 * is safe. Because if page_cgroup's USED bit is unset, the page
1067 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1068 * set, the commit after this will fail, anyway.
1069 * This all charge/uncharge is done under some mutual execustion.
1070 * So, we don't need to taking care of changes in USED bit.
1072 if (likely(!PageLRU(page)))
1075 spin_lock_irqsave(&zone->lru_lock, flags);
1077 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1078 * is guarded by lock_page() because the page is SwapCache.
1080 if (!PageCgroupUsed(pc))
1081 mem_cgroup_del_lru_list(page, page_lru(page));
1082 spin_unlock_irqrestore(&zone->lru_lock, flags);
1085 static void mem_cgroup_lru_add_after_commit(struct page *page)
1087 unsigned long flags;
1088 struct zone *zone = page_zone(page);
1089 struct page_cgroup *pc = lookup_page_cgroup(page);
1091 /* taking care of that the page is added to LRU while we commit it */
1092 if (likely(!PageLRU(page)))
1094 spin_lock_irqsave(&zone->lru_lock, flags);
1095 /* link when the page is linked to LRU but page_cgroup isn't */
1096 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1097 mem_cgroup_add_lru_list(page, page_lru(page));
1098 spin_unlock_irqrestore(&zone->lru_lock, flags);
1102 void mem_cgroup_move_lists(struct page *page,
1103 enum lru_list from, enum lru_list to)
1105 if (mem_cgroup_disabled())
1107 mem_cgroup_del_lru_list(page, from);
1108 mem_cgroup_add_lru_list(page, to);
1111 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
1114 struct mem_cgroup *curr = NULL;
1115 struct task_struct *p;
1117 p = find_lock_task_mm(task);
1120 curr = try_get_mem_cgroup_from_mm(p->mm);
1125 * We should check use_hierarchy of "mem" not "curr". Because checking
1126 * use_hierarchy of "curr" here make this function true if hierarchy is
1127 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
1128 * hierarchy(even if use_hierarchy is disabled in "mem").
1130 if (mem->use_hierarchy)
1131 ret = css_is_ancestor(&curr->css, &mem->css);
1133 ret = (curr == mem);
1134 css_put(&curr->css);
1138 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1140 unsigned long active;
1141 unsigned long inactive;
1143 unsigned long inactive_ratio;
1145 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
1146 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
1148 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1150 inactive_ratio = int_sqrt(10 * gb);
1154 if (present_pages) {
1155 present_pages[0] = inactive;
1156 present_pages[1] = active;
1159 return inactive_ratio;
1162 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1164 unsigned long active;
1165 unsigned long inactive;
1166 unsigned long present_pages[2];
1167 unsigned long inactive_ratio;
1169 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1171 inactive = present_pages[0];
1172 active = present_pages[1];
1174 if (inactive * inactive_ratio < active)
1180 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1182 unsigned long active;
1183 unsigned long inactive;
1185 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
1186 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
1188 return (active > inactive);
1191 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1194 int nid = zone_to_nid(zone);
1195 int zid = zone_idx(zone);
1196 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1198 return &mz->reclaim_stat;
1201 struct zone_reclaim_stat *
1202 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1204 struct page_cgroup *pc;
1205 struct mem_cgroup_per_zone *mz;
1207 if (mem_cgroup_disabled())
1210 pc = lookup_page_cgroup(page);
1211 if (!PageCgroupUsed(pc))
1213 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1215 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1216 return &mz->reclaim_stat;
1219 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1220 struct list_head *dst,
1221 unsigned long *scanned, int order,
1222 int mode, struct zone *z,
1223 struct mem_cgroup *mem_cont,
1224 int active, int file)
1226 unsigned long nr_taken = 0;
1230 struct list_head *src;
1231 struct page_cgroup *pc, *tmp;
1232 int nid = zone_to_nid(z);
1233 int zid = zone_idx(z);
1234 struct mem_cgroup_per_zone *mz;
1235 int lru = LRU_FILE * file + active;
1239 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1240 src = &mz->lists[lru];
1243 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1244 if (scan >= nr_to_scan)
1247 if (unlikely(!PageCgroupUsed(pc)))
1250 page = lookup_cgroup_page(pc);
1252 if (unlikely(!PageLRU(page)))
1256 ret = __isolate_lru_page(page, mode, file);
1259 list_move(&page->lru, dst);
1260 mem_cgroup_del_lru(page);
1261 nr_taken += hpage_nr_pages(page);
1264 /* we don't affect global LRU but rotate in our LRU */
1265 mem_cgroup_rotate_lru_list(page, page_lru(page));
1274 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1280 #define mem_cgroup_from_res_counter(counter, member) \
1281 container_of(counter, struct mem_cgroup, member)
1284 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1285 * @mem: the memory cgroup
1287 * Returns the maximum amount of memory @mem can be charged with, in
1290 static unsigned long mem_cgroup_margin(struct mem_cgroup *mem)
1292 unsigned long long margin;
1294 margin = res_counter_margin(&mem->res);
1295 if (do_swap_account)
1296 margin = min(margin, res_counter_margin(&mem->memsw));
1297 return margin >> PAGE_SHIFT;
1300 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1302 struct cgroup *cgrp = memcg->css.cgroup;
1305 if (cgrp->parent == NULL)
1306 return vm_swappiness;
1308 return memcg->swappiness;
1311 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1316 spin_lock(&mem->pcp_counter_lock);
1317 for_each_online_cpu(cpu)
1318 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1319 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1320 spin_unlock(&mem->pcp_counter_lock);
1326 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1333 spin_lock(&mem->pcp_counter_lock);
1334 for_each_online_cpu(cpu)
1335 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1336 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1337 spin_unlock(&mem->pcp_counter_lock);
1341 * 2 routines for checking "mem" is under move_account() or not.
1343 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1344 * for avoiding race in accounting. If true,
1345 * pc->mem_cgroup may be overwritten.
1347 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1348 * under hierarchy of moving cgroups. This is for
1349 * waiting at hith-memory prressure caused by "move".
1352 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1354 VM_BUG_ON(!rcu_read_lock_held());
1355 return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1358 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1360 struct mem_cgroup *from;
1361 struct mem_cgroup *to;
1364 * Unlike task_move routines, we access mc.to, mc.from not under
1365 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1367 spin_lock(&mc.lock);
1372 if (from == mem || to == mem
1373 || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1374 || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1377 spin_unlock(&mc.lock);
1381 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1383 if (mc.moving_task && current != mc.moving_task) {
1384 if (mem_cgroup_under_move(mem)) {
1386 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1387 /* moving charge context might have finished. */
1390 finish_wait(&mc.waitq, &wait);
1398 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1399 * @memcg: The memory cgroup that went over limit
1400 * @p: Task that is going to be killed
1402 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1405 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1407 struct cgroup *task_cgrp;
1408 struct cgroup *mem_cgrp;
1410 * Need a buffer in BSS, can't rely on allocations. The code relies
1411 * on the assumption that OOM is serialized for memory controller.
1412 * If this assumption is broken, revisit this code.
1414 static char memcg_name[PATH_MAX];
1423 mem_cgrp = memcg->css.cgroup;
1424 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1426 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1429 * Unfortunately, we are unable to convert to a useful name
1430 * But we'll still print out the usage information
1437 printk(KERN_INFO "Task in %s killed", memcg_name);
1440 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1448 * Continues from above, so we don't need an KERN_ level
1450 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1453 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1454 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1455 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1456 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1457 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1459 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1460 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1461 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1465 * This function returns the number of memcg under hierarchy tree. Returns
1466 * 1(self count) if no children.
1468 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1471 struct mem_cgroup *iter;
1473 for_each_mem_cgroup_tree(iter, mem)
1479 * Return the memory (and swap, if configured) limit for a memcg.
1481 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1486 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1487 limit += total_swap_pages << PAGE_SHIFT;
1489 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1491 * If memsw is finite and limits the amount of swap space available
1492 * to this memcg, return that limit.
1494 return min(limit, memsw);
1498 * Visit the first child (need not be the first child as per the ordering
1499 * of the cgroup list, since we track last_scanned_child) of @mem and use
1500 * that to reclaim free pages from.
1502 static struct mem_cgroup *
1503 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1505 struct mem_cgroup *ret = NULL;
1506 struct cgroup_subsys_state *css;
1509 if (!root_mem->use_hierarchy) {
1510 css_get(&root_mem->css);
1516 nextid = root_mem->last_scanned_child + 1;
1517 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1519 if (css && css_tryget(css))
1520 ret = container_of(css, struct mem_cgroup, css);
1523 /* Updates scanning parameter */
1525 /* this means start scan from ID:1 */
1526 root_mem->last_scanned_child = 0;
1528 root_mem->last_scanned_child = found;
1535 * test_mem_cgroup_node_reclaimable
1536 * @mem: the target memcg
1537 * @nid: the node ID to be checked.
1538 * @noswap : specify true here if the user wants flle only information.
1540 * This function returns whether the specified memcg contains any
1541 * reclaimable pages on a node. Returns true if there are any reclaimable
1542 * pages in the node.
1544 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *mem,
1545 int nid, bool noswap)
1547 if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_FILE))
1549 if (noswap || !total_swap_pages)
1551 if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_ANON))
1556 #if MAX_NUMNODES > 1
1559 * Always updating the nodemask is not very good - even if we have an empty
1560 * list or the wrong list here, we can start from some node and traverse all
1561 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1564 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
1568 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1569 * pagein/pageout changes since the last update.
1571 if (!atomic_read(&mem->numainfo_events))
1573 if (atomic_inc_return(&mem->numainfo_updating) > 1)
1576 /* make a nodemask where this memcg uses memory from */
1577 mem->scan_nodes = node_states[N_HIGH_MEMORY];
1579 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1581 if (!test_mem_cgroup_node_reclaimable(mem, nid, false))
1582 node_clear(nid, mem->scan_nodes);
1585 atomic_set(&mem->numainfo_events, 0);
1586 atomic_set(&mem->numainfo_updating, 0);
1590 * Selecting a node where we start reclaim from. Because what we need is just
1591 * reducing usage counter, start from anywhere is O,K. Considering
1592 * memory reclaim from current node, there are pros. and cons.
1594 * Freeing memory from current node means freeing memory from a node which
1595 * we'll use or we've used. So, it may make LRU bad. And if several threads
1596 * hit limits, it will see a contention on a node. But freeing from remote
1597 * node means more costs for memory reclaim because of memory latency.
1599 * Now, we use round-robin. Better algorithm is welcomed.
1601 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1605 mem_cgroup_may_update_nodemask(mem);
1606 node = mem->last_scanned_node;
1608 node = next_node(node, mem->scan_nodes);
1609 if (node == MAX_NUMNODES)
1610 node = first_node(mem->scan_nodes);
1612 * We call this when we hit limit, not when pages are added to LRU.
1613 * No LRU may hold pages because all pages are UNEVICTABLE or
1614 * memcg is too small and all pages are not on LRU. In that case,
1615 * we use curret node.
1617 if (unlikely(node == MAX_NUMNODES))
1618 node = numa_node_id();
1620 mem->last_scanned_node = node;
1625 * Check all nodes whether it contains reclaimable pages or not.
1626 * For quick scan, we make use of scan_nodes. This will allow us to skip
1627 * unused nodes. But scan_nodes is lazily updated and may not cotain
1628 * enough new information. We need to do double check.
1630 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1635 * quick check...making use of scan_node.
1636 * We can skip unused nodes.
1638 if (!nodes_empty(mem->scan_nodes)) {
1639 for (nid = first_node(mem->scan_nodes);
1641 nid = next_node(nid, mem->scan_nodes)) {
1643 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1648 * Check rest of nodes.
1650 for_each_node_state(nid, N_HIGH_MEMORY) {
1651 if (node_isset(nid, mem->scan_nodes))
1653 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1660 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1665 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1667 return test_mem_cgroup_node_reclaimable(mem, 0, noswap);
1671 static void __mem_cgroup_record_scanstat(unsigned long *stats,
1672 struct memcg_scanrecord *rec)
1675 stats[SCAN] += rec->nr_scanned[0] + rec->nr_scanned[1];
1676 stats[SCAN_ANON] += rec->nr_scanned[0];
1677 stats[SCAN_FILE] += rec->nr_scanned[1];
1679 stats[ROTATE] += rec->nr_rotated[0] + rec->nr_rotated[1];
1680 stats[ROTATE_ANON] += rec->nr_rotated[0];
1681 stats[ROTATE_FILE] += rec->nr_rotated[1];
1683 stats[FREED] += rec->nr_freed[0] + rec->nr_freed[1];
1684 stats[FREED_ANON] += rec->nr_freed[0];
1685 stats[FREED_FILE] += rec->nr_freed[1];
1687 stats[ELAPSED] += rec->elapsed;
1690 static void mem_cgroup_record_scanstat(struct memcg_scanrecord *rec)
1692 struct mem_cgroup *mem;
1693 int context = rec->context;
1695 if (context >= NR_SCAN_CONTEXT)
1699 spin_lock(&mem->scanstat.lock);
1700 __mem_cgroup_record_scanstat(mem->scanstat.stats[context], rec);
1701 spin_unlock(&mem->scanstat.lock);
1704 spin_lock(&mem->scanstat.lock);
1705 __mem_cgroup_record_scanstat(mem->scanstat.rootstats[context], rec);
1706 spin_unlock(&mem->scanstat.lock);
1710 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1711 * we reclaimed from, so that we don't end up penalizing one child extensively
1712 * based on its position in the children list.
1714 * root_mem is the original ancestor that we've been reclaim from.
1716 * We give up and return to the caller when we visit root_mem twice.
1717 * (other groups can be removed while we're walking....)
1719 * If shrink==true, for avoiding to free too much, this returns immedieately.
1721 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1724 unsigned long reclaim_options,
1725 unsigned long *total_scanned)
1727 struct mem_cgroup *victim;
1730 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1731 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1732 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1733 struct memcg_scanrecord rec;
1734 unsigned long excess;
1735 unsigned long scanned;
1737 excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1739 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1740 if (!check_soft && !shrink && root_mem->memsw_is_minimum)
1744 rec.context = SCAN_BY_SHRINK;
1745 else if (check_soft)
1746 rec.context = SCAN_BY_SYSTEM;
1748 rec.context = SCAN_BY_LIMIT;
1750 rec.root = root_mem;
1753 victim = mem_cgroup_select_victim(root_mem);
1754 if (victim == root_mem) {
1757 * We are not draining per cpu cached charges during
1758 * soft limit reclaim because global reclaim doesn't
1759 * care about charges. It tries to free some memory and
1760 * charges will not give any.
1762 if (!check_soft && loop >= 1)
1763 drain_all_stock_async(root_mem);
1766 * If we have not been able to reclaim
1767 * anything, it might because there are
1768 * no reclaimable pages under this hierarchy
1770 if (!check_soft || !total) {
1771 css_put(&victim->css);
1775 * We want to do more targeted reclaim.
1776 * excess >> 2 is not to excessive so as to
1777 * reclaim too much, nor too less that we keep
1778 * coming back to reclaim from this cgroup
1780 if (total >= (excess >> 2) ||
1781 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1782 css_put(&victim->css);
1787 if (!mem_cgroup_reclaimable(victim, noswap)) {
1788 /* this cgroup's local usage == 0 */
1789 css_put(&victim->css);
1793 rec.nr_scanned[0] = 0;
1794 rec.nr_scanned[1] = 0;
1795 rec.nr_rotated[0] = 0;
1796 rec.nr_rotated[1] = 0;
1797 rec.nr_freed[0] = 0;
1798 rec.nr_freed[1] = 0;
1800 /* we use swappiness of local cgroup */
1802 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1803 noswap, zone, &rec, &scanned);
1804 *total_scanned += scanned;
1806 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1808 mem_cgroup_record_scanstat(&rec);
1809 css_put(&victim->css);
1811 * At shrinking usage, we can't check we should stop here or
1812 * reclaim more. It's depends on callers. last_scanned_child
1813 * will work enough for keeping fairness under tree.
1819 if (!res_counter_soft_limit_excess(&root_mem->res))
1821 } else if (mem_cgroup_margin(root_mem))
1828 * Check OOM-Killer is already running under our hierarchy.
1829 * If someone is running, return false.
1830 * Has to be called with memcg_oom_lock
1832 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1834 int lock_count = -1;
1835 struct mem_cgroup *iter, *failed = NULL;
1838 for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1839 bool locked = iter->oom_lock;
1841 iter->oom_lock = true;
1842 if (lock_count == -1)
1843 lock_count = iter->oom_lock;
1844 else if (lock_count != locked) {
1846 * this subtree of our hierarchy is already locked
1847 * so we cannot give a lock.
1859 * OK, we failed to lock the whole subtree so we have to clean up
1860 * what we set up to the failing subtree
1863 for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1864 if (iter == failed) {
1868 iter->oom_lock = false;
1875 * Has to be called with memcg_oom_lock
1877 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1879 struct mem_cgroup *iter;
1881 for_each_mem_cgroup_tree(iter, mem)
1882 iter->oom_lock = false;
1886 static void mem_cgroup_mark_under_oom(struct mem_cgroup *mem)
1888 struct mem_cgroup *iter;
1890 for_each_mem_cgroup_tree(iter, mem)
1891 atomic_inc(&iter->under_oom);
1894 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *mem)
1896 struct mem_cgroup *iter;
1899 * When a new child is created while the hierarchy is under oom,
1900 * mem_cgroup_oom_lock() may not be called. We have to use
1901 * atomic_add_unless() here.
1903 for_each_mem_cgroup_tree(iter, mem)
1904 atomic_add_unless(&iter->under_oom, -1, 0);
1907 static DEFINE_SPINLOCK(memcg_oom_lock);
1908 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1910 struct oom_wait_info {
1911 struct mem_cgroup *mem;
1915 static int memcg_oom_wake_function(wait_queue_t *wait,
1916 unsigned mode, int sync, void *arg)
1918 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1919 struct oom_wait_info *oom_wait_info;
1921 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1923 if (oom_wait_info->mem == wake_mem)
1925 /* if no hierarchy, no match */
1926 if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1929 * Both of oom_wait_info->mem and wake_mem are stable under us.
1930 * Then we can use css_is_ancestor without taking care of RCU.
1932 if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1933 !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1937 return autoremove_wake_function(wait, mode, sync, arg);
1940 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1942 /* for filtering, pass "mem" as argument. */
1943 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1946 static void memcg_oom_recover(struct mem_cgroup *mem)
1948 if (mem && atomic_read(&mem->under_oom))
1949 memcg_wakeup_oom(mem);
1953 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1955 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1957 struct oom_wait_info owait;
1958 bool locked, need_to_kill;
1961 owait.wait.flags = 0;
1962 owait.wait.func = memcg_oom_wake_function;
1963 owait.wait.private = current;
1964 INIT_LIST_HEAD(&owait.wait.task_list);
1965 need_to_kill = true;
1966 mem_cgroup_mark_under_oom(mem);
1968 /* At first, try to OOM lock hierarchy under mem.*/
1969 spin_lock(&memcg_oom_lock);
1970 locked = mem_cgroup_oom_lock(mem);
1972 * Even if signal_pending(), we can't quit charge() loop without
1973 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1974 * under OOM is always welcomed, use TASK_KILLABLE here.
1976 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1977 if (!locked || mem->oom_kill_disable)
1978 need_to_kill = false;
1980 mem_cgroup_oom_notify(mem);
1981 spin_unlock(&memcg_oom_lock);
1984 finish_wait(&memcg_oom_waitq, &owait.wait);
1985 mem_cgroup_out_of_memory(mem, mask);
1988 finish_wait(&memcg_oom_waitq, &owait.wait);
1990 spin_lock(&memcg_oom_lock);
1992 mem_cgroup_oom_unlock(mem);
1993 memcg_wakeup_oom(mem);
1994 spin_unlock(&memcg_oom_lock);
1996 mem_cgroup_unmark_under_oom(mem);
1998 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
2000 /* Give chance to dying process */
2001 schedule_timeout(1);
2006 * Currently used to update mapped file statistics, but the routine can be
2007 * generalized to update other statistics as well.
2009 * Notes: Race condition
2011 * We usually use page_cgroup_lock() for accessing page_cgroup member but
2012 * it tends to be costly. But considering some conditions, we doesn't need
2013 * to do so _always_.
2015 * Considering "charge", lock_page_cgroup() is not required because all
2016 * file-stat operations happen after a page is attached to radix-tree. There
2017 * are no race with "charge".
2019 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
2020 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
2021 * if there are race with "uncharge". Statistics itself is properly handled
2024 * Considering "move", this is an only case we see a race. To make the race
2025 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
2026 * possibility of race condition. If there is, we take a lock.
2029 void mem_cgroup_update_page_stat(struct page *page,
2030 enum mem_cgroup_page_stat_item idx, int val)
2032 struct mem_cgroup *mem;
2033 struct page_cgroup *pc = lookup_page_cgroup(page);
2034 bool need_unlock = false;
2035 unsigned long uninitialized_var(flags);
2041 mem = pc->mem_cgroup;
2042 if (unlikely(!mem || !PageCgroupUsed(pc)))
2044 /* pc->mem_cgroup is unstable ? */
2045 if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
2046 /* take a lock against to access pc->mem_cgroup */
2047 move_lock_page_cgroup(pc, &flags);
2049 mem = pc->mem_cgroup;
2050 if (!mem || !PageCgroupUsed(pc))
2055 case MEMCG_NR_FILE_MAPPED:
2057 SetPageCgroupFileMapped(pc);
2058 else if (!page_mapped(page))
2059 ClearPageCgroupFileMapped(pc);
2060 idx = MEM_CGROUP_STAT_FILE_MAPPED;
2066 this_cpu_add(mem->stat->count[idx], val);
2069 if (unlikely(need_unlock))
2070 move_unlock_page_cgroup(pc, &flags);
2074 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2077 * size of first charge trial. "32" comes from vmscan.c's magic value.
2078 * TODO: maybe necessary to use big numbers in big irons.
2080 #define CHARGE_BATCH 32U
2081 struct memcg_stock_pcp {
2082 struct mem_cgroup *cached; /* this never be root cgroup */
2083 unsigned int nr_pages;
2084 struct work_struct work;
2085 unsigned long flags;
2086 #define FLUSHING_CACHED_CHARGE (0)
2088 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2089 static DEFINE_MUTEX(percpu_charge_mutex);
2092 * Try to consume stocked charge on this cpu. If success, one page is consumed
2093 * from local stock and true is returned. If the stock is 0 or charges from a
2094 * cgroup which is not current target, returns false. This stock will be
2097 static bool consume_stock(struct mem_cgroup *mem)
2099 struct memcg_stock_pcp *stock;
2102 stock = &get_cpu_var(memcg_stock);
2103 if (mem == stock->cached && stock->nr_pages)
2105 else /* need to call res_counter_charge */
2107 put_cpu_var(memcg_stock);
2112 * Returns stocks cached in percpu to res_counter and reset cached information.
2114 static void drain_stock(struct memcg_stock_pcp *stock)
2116 struct mem_cgroup *old = stock->cached;
2118 if (stock->nr_pages) {
2119 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2121 res_counter_uncharge(&old->res, bytes);
2122 if (do_swap_account)
2123 res_counter_uncharge(&old->memsw, bytes);
2124 stock->nr_pages = 0;
2126 stock->cached = NULL;
2130 * This must be called under preempt disabled or must be called by
2131 * a thread which is pinned to local cpu.
2133 static void drain_local_stock(struct work_struct *dummy)
2135 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2137 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2141 * Cache charges(val) which is from res_counter, to local per_cpu area.
2142 * This will be consumed by consume_stock() function, later.
2144 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
2146 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2148 if (stock->cached != mem) { /* reset if necessary */
2150 stock->cached = mem;
2152 stock->nr_pages += nr_pages;
2153 put_cpu_var(memcg_stock);
2157 * Drains all per-CPU charge caches for given root_mem resp. subtree
2158 * of the hierarchy under it. sync flag says whether we should block
2159 * until the work is done.
2161 static void drain_all_stock(struct mem_cgroup *root_mem, bool sync)
2165 /* Notify other cpus that system-wide "drain" is running */
2168 * Get a hint for avoiding draining charges on the current cpu,
2169 * which must be exhausted by our charging. It is not required that
2170 * this be a precise check, so we use raw_smp_processor_id() instead of
2171 * getcpu()/putcpu().
2173 curcpu = raw_smp_processor_id();
2174 for_each_online_cpu(cpu) {
2175 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2176 struct mem_cgroup *mem;
2178 mem = stock->cached;
2179 if (!mem || !stock->nr_pages)
2181 if (mem != root_mem) {
2182 if (!root_mem->use_hierarchy)
2184 /* check whether "mem" is under tree of "root_mem" */
2185 if (!css_is_ancestor(&mem->css, &root_mem->css))
2188 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2190 drain_local_stock(&stock->work);
2192 schedule_work_on(cpu, &stock->work);
2199 for_each_online_cpu(cpu) {
2200 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2201 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2202 flush_work(&stock->work);
2209 * Tries to drain stocked charges in other cpus. This function is asynchronous
2210 * and just put a work per cpu for draining localy on each cpu. Caller can
2211 * expects some charges will be back to res_counter later but cannot wait for
2214 static void drain_all_stock_async(struct mem_cgroup *root_mem)
2217 * If someone calls draining, avoid adding more kworker runs.
2219 if (!mutex_trylock(&percpu_charge_mutex))
2221 drain_all_stock(root_mem, false);
2222 mutex_unlock(&percpu_charge_mutex);
2225 /* This is a synchronous drain interface. */
2226 static void drain_all_stock_sync(struct mem_cgroup *root_mem)
2228 /* called when force_empty is called */
2229 mutex_lock(&percpu_charge_mutex);
2230 drain_all_stock(root_mem, true);
2231 mutex_unlock(&percpu_charge_mutex);
2235 * This function drains percpu counter value from DEAD cpu and
2236 * move it to local cpu. Note that this function can be preempted.
2238 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2242 spin_lock(&mem->pcp_counter_lock);
2243 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2244 long x = per_cpu(mem->stat->count[i], cpu);
2246 per_cpu(mem->stat->count[i], cpu) = 0;
2247 mem->nocpu_base.count[i] += x;
2249 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2250 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2252 per_cpu(mem->stat->events[i], cpu) = 0;
2253 mem->nocpu_base.events[i] += x;
2255 /* need to clear ON_MOVE value, works as a kind of lock. */
2256 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2257 spin_unlock(&mem->pcp_counter_lock);
2260 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2262 int idx = MEM_CGROUP_ON_MOVE;
2264 spin_lock(&mem->pcp_counter_lock);
2265 per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2266 spin_unlock(&mem->pcp_counter_lock);
2269 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2270 unsigned long action,
2273 int cpu = (unsigned long)hcpu;
2274 struct memcg_stock_pcp *stock;
2275 struct mem_cgroup *iter;
2277 if ((action == CPU_ONLINE)) {
2278 for_each_mem_cgroup_all(iter)
2279 synchronize_mem_cgroup_on_move(iter, cpu);
2283 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2286 for_each_mem_cgroup_all(iter)
2287 mem_cgroup_drain_pcp_counter(iter, cpu);
2289 stock = &per_cpu(memcg_stock, cpu);
2295 /* See __mem_cgroup_try_charge() for details */
2297 CHARGE_OK, /* success */
2298 CHARGE_RETRY, /* need to retry but retry is not bad */
2299 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2300 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2301 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2304 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2305 unsigned int nr_pages, bool oom_check)
2307 unsigned long csize = nr_pages * PAGE_SIZE;
2308 struct mem_cgroup *mem_over_limit;
2309 struct res_counter *fail_res;
2310 unsigned long flags = 0;
2313 ret = res_counter_charge(&mem->res, csize, &fail_res);
2316 if (!do_swap_account)
2318 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2322 res_counter_uncharge(&mem->res, csize);
2323 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2324 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2326 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2328 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2329 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2331 * Never reclaim on behalf of optional batching, retry with a
2332 * single page instead.
2334 if (nr_pages == CHARGE_BATCH)
2335 return CHARGE_RETRY;
2337 if (!(gfp_mask & __GFP_WAIT))
2338 return CHARGE_WOULDBLOCK;
2340 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2341 gfp_mask, flags, NULL);
2342 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2343 return CHARGE_RETRY;
2345 * Even though the limit is exceeded at this point, reclaim
2346 * may have been able to free some pages. Retry the charge
2347 * before killing the task.
2349 * Only for regular pages, though: huge pages are rather
2350 * unlikely to succeed so close to the limit, and we fall back
2351 * to regular pages anyway in case of failure.
2353 if (nr_pages == 1 && ret)
2354 return CHARGE_RETRY;
2357 * At task move, charge accounts can be doubly counted. So, it's
2358 * better to wait until the end of task_move if something is going on.
2360 if (mem_cgroup_wait_acct_move(mem_over_limit))
2361 return CHARGE_RETRY;
2363 /* If we don't need to call oom-killer at el, return immediately */
2365 return CHARGE_NOMEM;
2367 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2368 return CHARGE_OOM_DIE;
2370 return CHARGE_RETRY;
2374 * Unlike exported interface, "oom" parameter is added. if oom==true,
2375 * oom-killer can be invoked.
2377 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2379 unsigned int nr_pages,
2380 struct mem_cgroup **memcg,
2383 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2384 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2385 struct mem_cgroup *mem = NULL;
2389 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2390 * in system level. So, allow to go ahead dying process in addition to
2393 if (unlikely(test_thread_flag(TIF_MEMDIE)
2394 || fatal_signal_pending(current)))
2398 * We always charge the cgroup the mm_struct belongs to.
2399 * The mm_struct's mem_cgroup changes on task migration if the
2400 * thread group leader migrates. It's possible that mm is not
2401 * set, if so charge the init_mm (happens for pagecache usage).
2406 if (*memcg) { /* css should be a valid one */
2408 VM_BUG_ON(css_is_removed(&mem->css));
2409 if (mem_cgroup_is_root(mem))
2411 if (nr_pages == 1 && consume_stock(mem))
2415 struct task_struct *p;
2418 p = rcu_dereference(mm->owner);
2420 * Because we don't have task_lock(), "p" can exit.
2421 * In that case, "mem" can point to root or p can be NULL with
2422 * race with swapoff. Then, we have small risk of mis-accouning.
2423 * But such kind of mis-account by race always happens because
2424 * we don't have cgroup_mutex(). It's overkill and we allo that
2426 * (*) swapoff at el will charge against mm-struct not against
2427 * task-struct. So, mm->owner can be NULL.
2429 mem = mem_cgroup_from_task(p);
2430 if (!mem || mem_cgroup_is_root(mem)) {
2434 if (nr_pages == 1 && consume_stock(mem)) {
2436 * It seems dagerous to access memcg without css_get().
2437 * But considering how consume_stok works, it's not
2438 * necessary. If consume_stock success, some charges
2439 * from this memcg are cached on this cpu. So, we
2440 * don't need to call css_get()/css_tryget() before
2441 * calling consume_stock().
2446 /* after here, we may be blocked. we need to get refcnt */
2447 if (!css_tryget(&mem->css)) {
2457 /* If killed, bypass charge */
2458 if (fatal_signal_pending(current)) {
2464 if (oom && !nr_oom_retries) {
2466 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2469 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2473 case CHARGE_RETRY: /* not in OOM situation but retry */
2478 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2481 case CHARGE_NOMEM: /* OOM routine works */
2486 /* If oom, we never return -ENOMEM */
2489 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2493 } while (ret != CHARGE_OK);
2495 if (batch > nr_pages)
2496 refill_stock(mem, batch - nr_pages);
2510 * Somemtimes we have to undo a charge we got by try_charge().
2511 * This function is for that and do uncharge, put css's refcnt.
2512 * gotten by try_charge().
2514 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2515 unsigned int nr_pages)
2517 if (!mem_cgroup_is_root(mem)) {
2518 unsigned long bytes = nr_pages * PAGE_SIZE;
2520 res_counter_uncharge(&mem->res, bytes);
2521 if (do_swap_account)
2522 res_counter_uncharge(&mem->memsw, bytes);
2527 * A helper function to get mem_cgroup from ID. must be called under
2528 * rcu_read_lock(). The caller must check css_is_removed() or some if
2529 * it's concern. (dropping refcnt from swap can be called against removed
2532 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2534 struct cgroup_subsys_state *css;
2536 /* ID 0 is unused ID */
2539 css = css_lookup(&mem_cgroup_subsys, id);
2542 return container_of(css, struct mem_cgroup, css);
2545 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2547 struct mem_cgroup *mem = NULL;
2548 struct page_cgroup *pc;
2552 VM_BUG_ON(!PageLocked(page));
2554 pc = lookup_page_cgroup(page);
2555 lock_page_cgroup(pc);
2556 if (PageCgroupUsed(pc)) {
2557 mem = pc->mem_cgroup;
2558 if (mem && !css_tryget(&mem->css))
2560 } else if (PageSwapCache(page)) {
2561 ent.val = page_private(page);
2562 id = lookup_swap_cgroup(ent);
2564 mem = mem_cgroup_lookup(id);
2565 if (mem && !css_tryget(&mem->css))
2569 unlock_page_cgroup(pc);
2573 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2575 unsigned int nr_pages,
2576 struct page_cgroup *pc,
2577 enum charge_type ctype)
2579 lock_page_cgroup(pc);
2580 if (unlikely(PageCgroupUsed(pc))) {
2581 unlock_page_cgroup(pc);
2582 __mem_cgroup_cancel_charge(mem, nr_pages);
2586 * we don't need page_cgroup_lock about tail pages, becase they are not
2587 * accessed by any other context at this point.
2589 pc->mem_cgroup = mem;
2591 * We access a page_cgroup asynchronously without lock_page_cgroup().
2592 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2593 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2594 * before USED bit, we need memory barrier here.
2595 * See mem_cgroup_add_lru_list(), etc.
2599 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2600 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2601 SetPageCgroupCache(pc);
2602 SetPageCgroupUsed(pc);
2604 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2605 ClearPageCgroupCache(pc);
2606 SetPageCgroupUsed(pc);
2612 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2613 unlock_page_cgroup(pc);
2615 * "charge_statistics" updated event counter. Then, check it.
2616 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2617 * if they exceeds softlimit.
2619 memcg_check_events(mem, page);
2622 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2624 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2625 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2627 * Because tail pages are not marked as "used", set it. We're under
2628 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2630 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2632 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2633 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2634 unsigned long flags;
2636 if (mem_cgroup_disabled())
2639 * We have no races with charge/uncharge but will have races with
2640 * page state accounting.
2642 move_lock_page_cgroup(head_pc, &flags);
2644 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2645 smp_wmb(); /* see __commit_charge() */
2646 if (PageCgroupAcctLRU(head_pc)) {
2648 struct mem_cgroup_per_zone *mz;
2651 * LRU flags cannot be copied because we need to add tail
2652 *.page to LRU by generic call and our hook will be called.
2653 * We hold lru_lock, then, reduce counter directly.
2655 lru = page_lru(head);
2656 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2657 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2659 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2660 move_unlock_page_cgroup(head_pc, &flags);
2665 * mem_cgroup_move_account - move account of the page
2667 * @nr_pages: number of regular pages (>1 for huge pages)
2668 * @pc: page_cgroup of the page.
2669 * @from: mem_cgroup which the page is moved from.
2670 * @to: mem_cgroup which the page is moved to. @from != @to.
2671 * @uncharge: whether we should call uncharge and css_put against @from.
2673 * The caller must confirm following.
2674 * - page is not on LRU (isolate_page() is useful.)
2675 * - compound_lock is held when nr_pages > 1
2677 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2678 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2679 * true, this function does "uncharge" from old cgroup, but it doesn't if
2680 * @uncharge is false, so a caller should do "uncharge".
2682 static int mem_cgroup_move_account(struct page *page,
2683 unsigned int nr_pages,
2684 struct page_cgroup *pc,
2685 struct mem_cgroup *from,
2686 struct mem_cgroup *to,
2689 unsigned long flags;
2692 VM_BUG_ON(from == to);
2693 VM_BUG_ON(PageLRU(page));
2695 * The page is isolated from LRU. So, collapse function
2696 * will not handle this page. But page splitting can happen.
2697 * Do this check under compound_page_lock(). The caller should
2701 if (nr_pages > 1 && !PageTransHuge(page))
2704 lock_page_cgroup(pc);
2707 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2710 move_lock_page_cgroup(pc, &flags);
2712 if (PageCgroupFileMapped(pc)) {
2713 /* Update mapped_file data for mem_cgroup */
2715 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2716 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2719 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2721 /* This is not "cancel", but cancel_charge does all we need. */
2722 __mem_cgroup_cancel_charge(from, nr_pages);
2724 /* caller should have done css_get */
2725 pc->mem_cgroup = to;
2726 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2728 * We charges against "to" which may not have any tasks. Then, "to"
2729 * can be under rmdir(). But in current implementation, caller of
2730 * this function is just force_empty() and move charge, so it's
2731 * guaranteed that "to" is never removed. So, we don't check rmdir
2734 move_unlock_page_cgroup(pc, &flags);
2737 unlock_page_cgroup(pc);
2741 memcg_check_events(to, page);
2742 memcg_check_events(from, page);
2748 * move charges to its parent.
2751 static int mem_cgroup_move_parent(struct page *page,
2752 struct page_cgroup *pc,
2753 struct mem_cgroup *child,
2756 struct cgroup *cg = child->css.cgroup;
2757 struct cgroup *pcg = cg->parent;
2758 struct mem_cgroup *parent;
2759 unsigned int nr_pages;
2760 unsigned long uninitialized_var(flags);
2768 if (!get_page_unless_zero(page))
2770 if (isolate_lru_page(page))
2773 nr_pages = hpage_nr_pages(page);
2775 parent = mem_cgroup_from_cont(pcg);
2776 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2781 flags = compound_lock_irqsave(page);
2783 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2785 __mem_cgroup_cancel_charge(parent, nr_pages);
2788 compound_unlock_irqrestore(page, flags);
2790 putback_lru_page(page);
2798 * Charge the memory controller for page usage.
2800 * 0 if the charge was successful
2801 * < 0 if the cgroup is over its limit
2803 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2804 gfp_t gfp_mask, enum charge_type ctype)
2806 struct mem_cgroup *mem = NULL;
2807 unsigned int nr_pages = 1;
2808 struct page_cgroup *pc;
2812 if (PageTransHuge(page)) {
2813 nr_pages <<= compound_order(page);
2814 VM_BUG_ON(!PageTransHuge(page));
2816 * Never OOM-kill a process for a huge page. The
2817 * fault handler will fall back to regular pages.
2822 pc = lookup_page_cgroup(page);
2823 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2825 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2829 __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2833 int mem_cgroup_newpage_charge(struct page *page,
2834 struct mm_struct *mm, gfp_t gfp_mask)
2836 if (mem_cgroup_disabled())
2839 * If already mapped, we don't have to account.
2840 * If page cache, page->mapping has address_space.
2841 * But page->mapping may have out-of-use anon_vma pointer,
2842 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2845 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2849 return mem_cgroup_charge_common(page, mm, gfp_mask,
2850 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2854 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2855 enum charge_type ctype);
2858 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2859 enum charge_type ctype)
2861 struct page_cgroup *pc = lookup_page_cgroup(page);
2863 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2864 * is already on LRU. It means the page may on some other page_cgroup's
2865 * LRU. Take care of it.
2867 mem_cgroup_lru_del_before_commit(page);
2868 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2869 mem_cgroup_lru_add_after_commit(page);
2873 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2876 struct mem_cgroup *mem = NULL;
2879 if (mem_cgroup_disabled())
2881 if (PageCompound(page))
2884 * Corner case handling. This is called from add_to_page_cache()
2885 * in usual. But some FS (shmem) precharges this page before calling it
2886 * and call add_to_page_cache() with GFP_NOWAIT.
2888 * For GFP_NOWAIT case, the page may be pre-charged before calling
2889 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2890 * charge twice. (It works but has to pay a bit larger cost.)
2891 * And when the page is SwapCache, it should take swap information
2892 * into account. This is under lock_page() now.
2894 if (!(gfp_mask & __GFP_WAIT)) {
2895 struct page_cgroup *pc;
2897 pc = lookup_page_cgroup(page);
2900 lock_page_cgroup(pc);
2901 if (PageCgroupUsed(pc)) {
2902 unlock_page_cgroup(pc);
2905 unlock_page_cgroup(pc);
2911 if (page_is_file_cache(page)) {
2912 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2917 * FUSE reuses pages without going through the final
2918 * put that would remove them from the LRU list, make
2919 * sure that they get relinked properly.
2921 __mem_cgroup_commit_charge_lrucare(page, mem,
2922 MEM_CGROUP_CHARGE_TYPE_CACHE);
2926 if (PageSwapCache(page)) {
2927 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2929 __mem_cgroup_commit_charge_swapin(page, mem,
2930 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2932 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2933 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2939 * While swap-in, try_charge -> commit or cancel, the page is locked.
2940 * And when try_charge() successfully returns, one refcnt to memcg without
2941 * struct page_cgroup is acquired. This refcnt will be consumed by
2942 * "commit()" or removed by "cancel()"
2944 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2946 gfp_t mask, struct mem_cgroup **ptr)
2948 struct mem_cgroup *mem;
2953 if (mem_cgroup_disabled())
2956 if (!do_swap_account)
2959 * A racing thread's fault, or swapoff, may have already updated
2960 * the pte, and even removed page from swap cache: in those cases
2961 * do_swap_page()'s pte_same() test will fail; but there's also a
2962 * KSM case which does need to charge the page.
2964 if (!PageSwapCache(page))
2966 mem = try_get_mem_cgroup_from_page(page);
2970 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2976 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2980 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2981 enum charge_type ctype)
2983 if (mem_cgroup_disabled())
2987 cgroup_exclude_rmdir(&ptr->css);
2989 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2991 * Now swap is on-memory. This means this page may be
2992 * counted both as mem and swap....double count.
2993 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2994 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2995 * may call delete_from_swap_cache() before reach here.
2997 if (do_swap_account && PageSwapCache(page)) {
2998 swp_entry_t ent = {.val = page_private(page)};
3000 struct mem_cgroup *memcg;
3002 id = swap_cgroup_record(ent, 0);
3004 memcg = mem_cgroup_lookup(id);
3007 * This recorded memcg can be obsolete one. So, avoid
3008 * calling css_tryget
3010 if (!mem_cgroup_is_root(memcg))
3011 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3012 mem_cgroup_swap_statistics(memcg, false);
3013 mem_cgroup_put(memcg);
3018 * At swapin, we may charge account against cgroup which has no tasks.
3019 * So, rmdir()->pre_destroy() can be called while we do this charge.
3020 * In that case, we need to call pre_destroy() again. check it here.
3022 cgroup_release_and_wakeup_rmdir(&ptr->css);
3025 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
3027 __mem_cgroup_commit_charge_swapin(page, ptr,
3028 MEM_CGROUP_CHARGE_TYPE_MAPPED);
3031 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
3033 if (mem_cgroup_disabled())
3037 __mem_cgroup_cancel_charge(mem, 1);
3040 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
3041 unsigned int nr_pages,
3042 const enum charge_type ctype)
3044 struct memcg_batch_info *batch = NULL;
3045 bool uncharge_memsw = true;
3047 /* If swapout, usage of swap doesn't decrease */
3048 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
3049 uncharge_memsw = false;
3051 batch = ¤t->memcg_batch;
3053 * In usual, we do css_get() when we remember memcg pointer.
3054 * But in this case, we keep res->usage until end of a series of
3055 * uncharges. Then, it's ok to ignore memcg's refcnt.
3060 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
3061 * In those cases, all pages freed continuously can be expected to be in
3062 * the same cgroup and we have chance to coalesce uncharges.
3063 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
3064 * because we want to do uncharge as soon as possible.
3067 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
3068 goto direct_uncharge;
3071 goto direct_uncharge;
3074 * In typical case, batch->memcg == mem. This means we can
3075 * merge a series of uncharges to an uncharge of res_counter.
3076 * If not, we uncharge res_counter ony by one.
3078 if (batch->memcg != mem)
3079 goto direct_uncharge;
3080 /* remember freed charge and uncharge it later */
3083 batch->memsw_nr_pages++;
3086 res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
3088 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
3089 if (unlikely(batch->memcg != mem))
3090 memcg_oom_recover(mem);
3095 * uncharge if !page_mapped(page)
3097 static struct mem_cgroup *
3098 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3100 struct mem_cgroup *mem = NULL;
3101 unsigned int nr_pages = 1;
3102 struct page_cgroup *pc;
3104 if (mem_cgroup_disabled())
3107 if (PageSwapCache(page))
3110 if (PageTransHuge(page)) {
3111 nr_pages <<= compound_order(page);
3112 VM_BUG_ON(!PageTransHuge(page));
3115 * Check if our page_cgroup is valid
3117 pc = lookup_page_cgroup(page);
3118 if (unlikely(!pc || !PageCgroupUsed(pc)))
3121 lock_page_cgroup(pc);
3123 mem = pc->mem_cgroup;
3125 if (!PageCgroupUsed(pc))
3129 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3130 case MEM_CGROUP_CHARGE_TYPE_DROP:
3131 /* See mem_cgroup_prepare_migration() */
3132 if (page_mapped(page) || PageCgroupMigration(pc))
3135 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3136 if (!PageAnon(page)) { /* Shared memory */
3137 if (page->mapping && !page_is_file_cache(page))
3139 } else if (page_mapped(page)) /* Anon */
3146 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
3148 ClearPageCgroupUsed(pc);
3150 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3151 * freed from LRU. This is safe because uncharged page is expected not
3152 * to be reused (freed soon). Exception is SwapCache, it's handled by
3153 * special functions.
3156 unlock_page_cgroup(pc);
3158 * even after unlock, we have mem->res.usage here and this memcg
3159 * will never be freed.
3161 memcg_check_events(mem, page);
3162 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3163 mem_cgroup_swap_statistics(mem, true);
3164 mem_cgroup_get(mem);
3166 if (!mem_cgroup_is_root(mem))
3167 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
3172 unlock_page_cgroup(pc);
3176 void mem_cgroup_uncharge_page(struct page *page)
3179 if (page_mapped(page))
3181 if (page->mapping && !PageAnon(page))
3183 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3186 void mem_cgroup_uncharge_cache_page(struct page *page)
3188 VM_BUG_ON(page_mapped(page));
3189 VM_BUG_ON(page->mapping);
3190 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3194 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3195 * In that cases, pages are freed continuously and we can expect pages
3196 * are in the same memcg. All these calls itself limits the number of
3197 * pages freed at once, then uncharge_start/end() is called properly.
3198 * This may be called prural(2) times in a context,
3201 void mem_cgroup_uncharge_start(void)
3203 current->memcg_batch.do_batch++;
3204 /* We can do nest. */
3205 if (current->memcg_batch.do_batch == 1) {
3206 current->memcg_batch.memcg = NULL;
3207 current->memcg_batch.nr_pages = 0;
3208 current->memcg_batch.memsw_nr_pages = 0;
3212 void mem_cgroup_uncharge_end(void)
3214 struct memcg_batch_info *batch = ¤t->memcg_batch;
3216 if (!batch->do_batch)
3220 if (batch->do_batch) /* If stacked, do nothing. */
3226 * This "batch->memcg" is valid without any css_get/put etc...
3227 * bacause we hide charges behind us.
3229 if (batch->nr_pages)
3230 res_counter_uncharge(&batch->memcg->res,
3231 batch->nr_pages * PAGE_SIZE);
3232 if (batch->memsw_nr_pages)
3233 res_counter_uncharge(&batch->memcg->memsw,
3234 batch->memsw_nr_pages * PAGE_SIZE);
3235 memcg_oom_recover(batch->memcg);
3236 /* forget this pointer (for sanity check) */
3237 batch->memcg = NULL;
3242 * called after __delete_from_swap_cache() and drop "page" account.
3243 * memcg information is recorded to swap_cgroup of "ent"
3246 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3248 struct mem_cgroup *memcg;
3249 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3251 if (!swapout) /* this was a swap cache but the swap is unused ! */
3252 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3254 memcg = __mem_cgroup_uncharge_common(page, ctype);
3257 * record memcg information, if swapout && memcg != NULL,
3258 * mem_cgroup_get() was called in uncharge().
3260 if (do_swap_account && swapout && memcg)
3261 swap_cgroup_record(ent, css_id(&memcg->css));
3265 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3267 * called from swap_entry_free(). remove record in swap_cgroup and
3268 * uncharge "memsw" account.
3270 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3272 struct mem_cgroup *memcg;
3275 if (!do_swap_account)
3278 id = swap_cgroup_record(ent, 0);
3280 memcg = mem_cgroup_lookup(id);
3283 * We uncharge this because swap is freed.
3284 * This memcg can be obsolete one. We avoid calling css_tryget
3286 if (!mem_cgroup_is_root(memcg))
3287 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3288 mem_cgroup_swap_statistics(memcg, false);
3289 mem_cgroup_put(memcg);
3295 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3296 * @entry: swap entry to be moved
3297 * @from: mem_cgroup which the entry is moved from
3298 * @to: mem_cgroup which the entry is moved to
3299 * @need_fixup: whether we should fixup res_counters and refcounts.
3301 * It succeeds only when the swap_cgroup's record for this entry is the same
3302 * as the mem_cgroup's id of @from.
3304 * Returns 0 on success, -EINVAL on failure.
3306 * The caller must have charged to @to, IOW, called res_counter_charge() about
3307 * both res and memsw, and called css_get().
3309 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3310 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3312 unsigned short old_id, new_id;
3314 old_id = css_id(&from->css);
3315 new_id = css_id(&to->css);
3317 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3318 mem_cgroup_swap_statistics(from, false);
3319 mem_cgroup_swap_statistics(to, true);
3321 * This function is only called from task migration context now.
3322 * It postpones res_counter and refcount handling till the end
3323 * of task migration(mem_cgroup_clear_mc()) for performance
3324 * improvement. But we cannot postpone mem_cgroup_get(to)
3325 * because if the process that has been moved to @to does
3326 * swap-in, the refcount of @to might be decreased to 0.
3330 if (!mem_cgroup_is_root(from))
3331 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3332 mem_cgroup_put(from);
3334 * we charged both to->res and to->memsw, so we should
3337 if (!mem_cgroup_is_root(to))
3338 res_counter_uncharge(&to->res, PAGE_SIZE);
3345 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3346 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3353 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3356 int mem_cgroup_prepare_migration(struct page *page,
3357 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3359 struct mem_cgroup *mem = NULL;
3360 struct page_cgroup *pc;
3361 enum charge_type ctype;
3366 VM_BUG_ON(PageTransHuge(page));
3367 if (mem_cgroup_disabled())
3370 pc = lookup_page_cgroup(page);
3371 lock_page_cgroup(pc);
3372 if (PageCgroupUsed(pc)) {
3373 mem = pc->mem_cgroup;
3376 * At migrating an anonymous page, its mapcount goes down
3377 * to 0 and uncharge() will be called. But, even if it's fully
3378 * unmapped, migration may fail and this page has to be
3379 * charged again. We set MIGRATION flag here and delay uncharge
3380 * until end_migration() is called
3382 * Corner Case Thinking
3384 * When the old page was mapped as Anon and it's unmap-and-freed
3385 * while migration was ongoing.
3386 * If unmap finds the old page, uncharge() of it will be delayed
3387 * until end_migration(). If unmap finds a new page, it's
3388 * uncharged when it make mapcount to be 1->0. If unmap code
3389 * finds swap_migration_entry, the new page will not be mapped
3390 * and end_migration() will find it(mapcount==0).
3393 * When the old page was mapped but migraion fails, the kernel
3394 * remaps it. A charge for it is kept by MIGRATION flag even
3395 * if mapcount goes down to 0. We can do remap successfully
3396 * without charging it again.
3399 * The "old" page is under lock_page() until the end of
3400 * migration, so, the old page itself will not be swapped-out.
3401 * If the new page is swapped out before end_migraton, our
3402 * hook to usual swap-out path will catch the event.
3405 SetPageCgroupMigration(pc);
3407 unlock_page_cgroup(pc);
3409 * If the page is not charged at this point,
3416 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3417 css_put(&mem->css);/* drop extra refcnt */
3418 if (ret || *ptr == NULL) {
3419 if (PageAnon(page)) {
3420 lock_page_cgroup(pc);
3421 ClearPageCgroupMigration(pc);
3422 unlock_page_cgroup(pc);
3424 * The old page may be fully unmapped while we kept it.
3426 mem_cgroup_uncharge_page(page);
3431 * We charge new page before it's used/mapped. So, even if unlock_page()
3432 * is called before end_migration, we can catch all events on this new
3433 * page. In the case new page is migrated but not remapped, new page's
3434 * mapcount will be finally 0 and we call uncharge in end_migration().
3436 pc = lookup_page_cgroup(newpage);
3438 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3439 else if (page_is_file_cache(page))
3440 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3442 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3443 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3447 /* remove redundant charge if migration failed*/
3448 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3449 struct page *oldpage, struct page *newpage, bool migration_ok)
3451 struct page *used, *unused;
3452 struct page_cgroup *pc;
3456 /* blocks rmdir() */
3457 cgroup_exclude_rmdir(&mem->css);
3458 if (!migration_ok) {
3466 * We disallowed uncharge of pages under migration because mapcount
3467 * of the page goes down to zero, temporarly.
3468 * Clear the flag and check the page should be charged.
3470 pc = lookup_page_cgroup(oldpage);
3471 lock_page_cgroup(pc);
3472 ClearPageCgroupMigration(pc);
3473 unlock_page_cgroup(pc);
3475 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3478 * If a page is a file cache, radix-tree replacement is very atomic
3479 * and we can skip this check. When it was an Anon page, its mapcount
3480 * goes down to 0. But because we added MIGRATION flage, it's not
3481 * uncharged yet. There are several case but page->mapcount check
3482 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3483 * check. (see prepare_charge() also)
3486 mem_cgroup_uncharge_page(used);
3488 * At migration, we may charge account against cgroup which has no
3490 * So, rmdir()->pre_destroy() can be called while we do this charge.
3491 * In that case, we need to call pre_destroy() again. check it here.
3493 cgroup_release_and_wakeup_rmdir(&mem->css);
3497 * A call to try to shrink memory usage on charge failure at shmem's swapin.
3498 * Calling hierarchical_reclaim is not enough because we should update
3499 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3500 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3501 * not from the memcg which this page would be charged to.
3502 * try_charge_swapin does all of these works properly.
3504 int mem_cgroup_shmem_charge_fallback(struct page *page,
3505 struct mm_struct *mm,
3508 struct mem_cgroup *mem;
3511 if (mem_cgroup_disabled())
3514 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3516 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3521 #ifdef CONFIG_DEBUG_VM
3522 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3524 struct page_cgroup *pc;
3526 pc = lookup_page_cgroup(page);
3527 if (likely(pc) && PageCgroupUsed(pc))
3532 bool mem_cgroup_bad_page_check(struct page *page)
3534 if (mem_cgroup_disabled())
3537 return lookup_page_cgroup_used(page) != NULL;
3540 void mem_cgroup_print_bad_page(struct page *page)
3542 struct page_cgroup *pc;
3544 pc = lookup_page_cgroup_used(page);
3549 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3550 pc, pc->flags, pc->mem_cgroup);
3552 path = kmalloc(PATH_MAX, GFP_KERNEL);
3555 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3560 printk(KERN_CONT "(%s)\n",
3561 (ret < 0) ? "cannot get the path" : path);
3567 static DEFINE_MUTEX(set_limit_mutex);
3569 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3570 unsigned long long val)
3573 u64 memswlimit, memlimit;
3575 int children = mem_cgroup_count_children(memcg);
3576 u64 curusage, oldusage;
3580 * For keeping hierarchical_reclaim simple, how long we should retry
3581 * is depends on callers. We set our retry-count to be function
3582 * of # of children which we should visit in this loop.
3584 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3586 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3589 while (retry_count) {
3590 if (signal_pending(current)) {
3595 * Rather than hide all in some function, I do this in
3596 * open coded manner. You see what this really does.
3597 * We have to guarantee mem->res.limit < mem->memsw.limit.
3599 mutex_lock(&set_limit_mutex);
3600 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3601 if (memswlimit < val) {
3603 mutex_unlock(&set_limit_mutex);
3607 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3611 ret = res_counter_set_limit(&memcg->res, val);
3613 if (memswlimit == val)
3614 memcg->memsw_is_minimum = true;
3616 memcg->memsw_is_minimum = false;
3618 mutex_unlock(&set_limit_mutex);
3623 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3624 MEM_CGROUP_RECLAIM_SHRINK,
3626 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3627 /* Usage is reduced ? */
3628 if (curusage >= oldusage)
3631 oldusage = curusage;
3633 if (!ret && enlarge)
3634 memcg_oom_recover(memcg);
3639 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3640 unsigned long long val)
3643 u64 memlimit, memswlimit, oldusage, curusage;
3644 int children = mem_cgroup_count_children(memcg);
3648 /* see mem_cgroup_resize_res_limit */
3649 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3650 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3651 while (retry_count) {
3652 if (signal_pending(current)) {
3657 * Rather than hide all in some function, I do this in
3658 * open coded manner. You see what this really does.
3659 * We have to guarantee mem->res.limit < mem->memsw.limit.
3661 mutex_lock(&set_limit_mutex);
3662 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3663 if (memlimit > val) {
3665 mutex_unlock(&set_limit_mutex);
3668 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3669 if (memswlimit < val)
3671 ret = res_counter_set_limit(&memcg->memsw, val);
3673 if (memlimit == val)
3674 memcg->memsw_is_minimum = true;
3676 memcg->memsw_is_minimum = false;
3678 mutex_unlock(&set_limit_mutex);
3683 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3684 MEM_CGROUP_RECLAIM_NOSWAP |
3685 MEM_CGROUP_RECLAIM_SHRINK,
3687 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3688 /* Usage is reduced ? */
3689 if (curusage >= oldusage)
3692 oldusage = curusage;
3694 if (!ret && enlarge)
3695 memcg_oom_recover(memcg);
3699 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3701 unsigned long *total_scanned)
3703 unsigned long nr_reclaimed = 0;
3704 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3705 unsigned long reclaimed;
3707 struct mem_cgroup_tree_per_zone *mctz;
3708 unsigned long long excess;
3709 unsigned long nr_scanned;
3714 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3716 * This loop can run a while, specially if mem_cgroup's continuously
3717 * keep exceeding their soft limit and putting the system under
3724 mz = mem_cgroup_largest_soft_limit_node(mctz);
3729 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3731 MEM_CGROUP_RECLAIM_SOFT,
3733 nr_reclaimed += reclaimed;
3734 *total_scanned += nr_scanned;
3735 spin_lock(&mctz->lock);
3738 * If we failed to reclaim anything from this memory cgroup
3739 * it is time to move on to the next cgroup
3745 * Loop until we find yet another one.
3747 * By the time we get the soft_limit lock
3748 * again, someone might have aded the
3749 * group back on the RB tree. Iterate to
3750 * make sure we get a different mem.
3751 * mem_cgroup_largest_soft_limit_node returns
3752 * NULL if no other cgroup is present on
3756 __mem_cgroup_largest_soft_limit_node(mctz);
3758 css_put(&next_mz->mem->css);
3759 else /* next_mz == NULL or other memcg */
3763 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3764 excess = res_counter_soft_limit_excess(&mz->mem->res);
3766 * One school of thought says that we should not add
3767 * back the node to the tree if reclaim returns 0.
3768 * But our reclaim could return 0, simply because due
3769 * to priority we are exposing a smaller subset of
3770 * memory to reclaim from. Consider this as a longer
3773 /* If excess == 0, no tree ops */
3774 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3775 spin_unlock(&mctz->lock);
3776 css_put(&mz->mem->css);
3779 * Could not reclaim anything and there are no more
3780 * mem cgroups to try or we seem to be looping without
3781 * reclaiming anything.
3783 if (!nr_reclaimed &&
3785 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3787 } while (!nr_reclaimed);
3789 css_put(&next_mz->mem->css);
3790 return nr_reclaimed;
3794 * This routine traverse page_cgroup in given list and drop them all.
3795 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3797 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3798 int node, int zid, enum lru_list lru)
3801 struct mem_cgroup_per_zone *mz;
3802 struct page_cgroup *pc, *busy;
3803 unsigned long flags, loop;
3804 struct list_head *list;
3807 zone = &NODE_DATA(node)->node_zones[zid];
3808 mz = mem_cgroup_zoneinfo(mem, node, zid);
3809 list = &mz->lists[lru];
3811 loop = MEM_CGROUP_ZSTAT(mz, lru);
3812 /* give some margin against EBUSY etc...*/
3819 spin_lock_irqsave(&zone->lru_lock, flags);
3820 if (list_empty(list)) {
3821 spin_unlock_irqrestore(&zone->lru_lock, flags);
3824 pc = list_entry(list->prev, struct page_cgroup, lru);
3826 list_move(&pc->lru, list);
3828 spin_unlock_irqrestore(&zone->lru_lock, flags);
3831 spin_unlock_irqrestore(&zone->lru_lock, flags);
3833 page = lookup_cgroup_page(pc);
3835 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3839 if (ret == -EBUSY || ret == -EINVAL) {
3840 /* found lock contention or "pc" is obsolete. */
3847 if (!ret && !list_empty(list))
3853 * make mem_cgroup's charge to be 0 if there is no task.
3854 * This enables deleting this mem_cgroup.
3856 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3859 int node, zid, shrink;
3860 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3861 struct cgroup *cgrp = mem->css.cgroup;
3866 /* should free all ? */
3872 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3875 if (signal_pending(current))
3877 /* This is for making all *used* pages to be on LRU. */
3878 lru_add_drain_all();
3879 drain_all_stock_sync(mem);
3881 mem_cgroup_start_move(mem);
3882 for_each_node_state(node, N_HIGH_MEMORY) {
3883 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3886 ret = mem_cgroup_force_empty_list(mem,
3895 mem_cgroup_end_move(mem);
3896 memcg_oom_recover(mem);
3897 /* it seems parent cgroup doesn't have enough mem */
3901 /* "ret" should also be checked to ensure all lists are empty. */
3902 } while (mem->res.usage > 0 || ret);
3908 /* returns EBUSY if there is a task or if we come here twice. */
3909 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3913 /* we call try-to-free pages for make this cgroup empty */
3914 lru_add_drain_all();
3915 /* try to free all pages in this cgroup */
3917 while (nr_retries && mem->res.usage > 0) {
3918 struct memcg_scanrecord rec;
3921 if (signal_pending(current)) {
3925 rec.context = SCAN_BY_SHRINK;
3928 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3932 /* maybe some writeback is necessary */
3933 congestion_wait(BLK_RW_ASYNC, HZ/10);
3938 /* try move_account...there may be some *locked* pages. */
3942 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3944 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3948 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3950 return mem_cgroup_from_cont(cont)->use_hierarchy;
3953 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3957 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3958 struct cgroup *parent = cont->parent;
3959 struct mem_cgroup *parent_mem = NULL;
3962 parent_mem = mem_cgroup_from_cont(parent);
3966 * If parent's use_hierarchy is set, we can't make any modifications
3967 * in the child subtrees. If it is unset, then the change can
3968 * occur, provided the current cgroup has no children.
3970 * For the root cgroup, parent_mem is NULL, we allow value to be
3971 * set if there are no children.
3973 if ((!parent_mem || !parent_mem->use_hierarchy) &&
3974 (val == 1 || val == 0)) {
3975 if (list_empty(&cont->children))
3976 mem->use_hierarchy = val;
3987 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3988 enum mem_cgroup_stat_index idx)
3990 struct mem_cgroup *iter;
3993 /* Per-cpu values can be negative, use a signed accumulator */
3994 for_each_mem_cgroup_tree(iter, mem)
3995 val += mem_cgroup_read_stat(iter, idx);
3997 if (val < 0) /* race ? */
4002 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
4006 if (!mem_cgroup_is_root(mem)) {
4008 return res_counter_read_u64(&mem->res, RES_USAGE);
4010 return res_counter_read_u64(&mem->memsw, RES_USAGE);
4013 val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
4014 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
4017 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4019 return val << PAGE_SHIFT;
4022 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
4024 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4028 type = MEMFILE_TYPE(cft->private);
4029 name = MEMFILE_ATTR(cft->private);
4032 if (name == RES_USAGE)
4033 val = mem_cgroup_usage(mem, false);
4035 val = res_counter_read_u64(&mem->res, name);
4038 if (name == RES_USAGE)
4039 val = mem_cgroup_usage(mem, true);
4041 val = res_counter_read_u64(&mem->memsw, name);
4050 * The user of this function is...
4053 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4056 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4058 unsigned long long val;
4061 type = MEMFILE_TYPE(cft->private);
4062 name = MEMFILE_ATTR(cft->private);
4065 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4069 /* This function does all necessary parse...reuse it */
4070 ret = res_counter_memparse_write_strategy(buffer, &val);
4074 ret = mem_cgroup_resize_limit(memcg, val);
4076 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4078 case RES_SOFT_LIMIT:
4079 ret = res_counter_memparse_write_strategy(buffer, &val);
4083 * For memsw, soft limits are hard to implement in terms
4084 * of semantics, for now, we support soft limits for
4085 * control without swap
4088 ret = res_counter_set_soft_limit(&memcg->res, val);
4093 ret = -EINVAL; /* should be BUG() ? */
4099 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4100 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4102 struct cgroup *cgroup;
4103 unsigned long long min_limit, min_memsw_limit, tmp;
4105 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4106 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4107 cgroup = memcg->css.cgroup;
4108 if (!memcg->use_hierarchy)
4111 while (cgroup->parent) {
4112 cgroup = cgroup->parent;
4113 memcg = mem_cgroup_from_cont(cgroup);
4114 if (!memcg->use_hierarchy)
4116 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4117 min_limit = min(min_limit, tmp);
4118 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4119 min_memsw_limit = min(min_memsw_limit, tmp);
4122 *mem_limit = min_limit;
4123 *memsw_limit = min_memsw_limit;
4127 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4129 struct mem_cgroup *mem;
4132 mem = mem_cgroup_from_cont(cont);
4133 type = MEMFILE_TYPE(event);
4134 name = MEMFILE_ATTR(event);
4138 res_counter_reset_max(&mem->res);
4140 res_counter_reset_max(&mem->memsw);
4144 res_counter_reset_failcnt(&mem->res);
4146 res_counter_reset_failcnt(&mem->memsw);
4153 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4156 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4160 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4161 struct cftype *cft, u64 val)
4163 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4165 if (val >= (1 << NR_MOVE_TYPE))
4168 * We check this value several times in both in can_attach() and
4169 * attach(), so we need cgroup lock to prevent this value from being
4173 mem->move_charge_at_immigrate = val;
4179 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4180 struct cftype *cft, u64 val)
4187 /* For read statistics */
4205 struct mcs_total_stat {
4206 s64 stat[NR_MCS_STAT];
4212 } memcg_stat_strings[NR_MCS_STAT] = {
4213 {"cache", "total_cache"},
4214 {"rss", "total_rss"},
4215 {"mapped_file", "total_mapped_file"},
4216 {"pgpgin", "total_pgpgin"},
4217 {"pgpgout", "total_pgpgout"},
4218 {"swap", "total_swap"},
4219 {"pgfault", "total_pgfault"},
4220 {"pgmajfault", "total_pgmajfault"},
4221 {"inactive_anon", "total_inactive_anon"},
4222 {"active_anon", "total_active_anon"},
4223 {"inactive_file", "total_inactive_file"},
4224 {"active_file", "total_active_file"},
4225 {"unevictable", "total_unevictable"}
4230 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4235 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4236 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4237 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4238 s->stat[MCS_RSS] += val * PAGE_SIZE;
4239 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4240 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4241 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4242 s->stat[MCS_PGPGIN] += val;
4243 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4244 s->stat[MCS_PGPGOUT] += val;
4245 if (do_swap_account) {
4246 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4247 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4249 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4250 s->stat[MCS_PGFAULT] += val;
4251 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4252 s->stat[MCS_PGMAJFAULT] += val;
4255 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_ANON));
4256 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4257 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_ANON));
4258 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4259 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_FILE));
4260 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4261 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_FILE));
4262 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4263 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_UNEVICTABLE));
4264 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4268 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4270 struct mem_cgroup *iter;
4272 for_each_mem_cgroup_tree(iter, mem)
4273 mem_cgroup_get_local_stat(iter, s);
4277 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4280 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4281 unsigned long node_nr;
4282 struct cgroup *cont = m->private;
4283 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4285 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4286 seq_printf(m, "total=%lu", total_nr);
4287 for_each_node_state(nid, N_HIGH_MEMORY) {
4288 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4289 seq_printf(m, " N%d=%lu", nid, node_nr);
4293 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4294 seq_printf(m, "file=%lu", file_nr);
4295 for_each_node_state(nid, N_HIGH_MEMORY) {
4296 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4298 seq_printf(m, " N%d=%lu", nid, node_nr);
4302 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4303 seq_printf(m, "anon=%lu", anon_nr);
4304 for_each_node_state(nid, N_HIGH_MEMORY) {
4305 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4307 seq_printf(m, " N%d=%lu", nid, node_nr);
4311 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4312 seq_printf(m, "unevictable=%lu", unevictable_nr);
4313 for_each_node_state(nid, N_HIGH_MEMORY) {
4314 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4315 BIT(LRU_UNEVICTABLE));
4316 seq_printf(m, " N%d=%lu", nid, node_nr);
4321 #endif /* CONFIG_NUMA */
4323 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4324 struct cgroup_map_cb *cb)
4326 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4327 struct mcs_total_stat mystat;
4330 memset(&mystat, 0, sizeof(mystat));
4331 mem_cgroup_get_local_stat(mem_cont, &mystat);
4334 for (i = 0; i < NR_MCS_STAT; i++) {
4335 if (i == MCS_SWAP && !do_swap_account)
4337 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4340 /* Hierarchical information */
4342 unsigned long long limit, memsw_limit;
4343 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4344 cb->fill(cb, "hierarchical_memory_limit", limit);
4345 if (do_swap_account)
4346 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4349 memset(&mystat, 0, sizeof(mystat));
4350 mem_cgroup_get_total_stat(mem_cont, &mystat);
4351 for (i = 0; i < NR_MCS_STAT; i++) {
4352 if (i == MCS_SWAP && !do_swap_account)
4354 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4357 #ifdef CONFIG_DEBUG_VM
4358 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4362 struct mem_cgroup_per_zone *mz;
4363 unsigned long recent_rotated[2] = {0, 0};
4364 unsigned long recent_scanned[2] = {0, 0};
4366 for_each_online_node(nid)
4367 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4368 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4370 recent_rotated[0] +=
4371 mz->reclaim_stat.recent_rotated[0];
4372 recent_rotated[1] +=
4373 mz->reclaim_stat.recent_rotated[1];
4374 recent_scanned[0] +=
4375 mz->reclaim_stat.recent_scanned[0];
4376 recent_scanned[1] +=
4377 mz->reclaim_stat.recent_scanned[1];
4379 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4380 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4381 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4382 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4389 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4391 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4393 return mem_cgroup_swappiness(memcg);
4396 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4399 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4400 struct mem_cgroup *parent;
4405 if (cgrp->parent == NULL)
4408 parent = mem_cgroup_from_cont(cgrp->parent);
4412 /* If under hierarchy, only empty-root can set this value */
4413 if ((parent->use_hierarchy) ||
4414 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4419 memcg->swappiness = val;
4426 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4428 struct mem_cgroup_threshold_ary *t;
4434 t = rcu_dereference(memcg->thresholds.primary);
4436 t = rcu_dereference(memcg->memsw_thresholds.primary);
4441 usage = mem_cgroup_usage(memcg, swap);
4444 * current_threshold points to threshold just below usage.
4445 * If it's not true, a threshold was crossed after last
4446 * call of __mem_cgroup_threshold().
4448 i = t->current_threshold;
4451 * Iterate backward over array of thresholds starting from
4452 * current_threshold and check if a threshold is crossed.
4453 * If none of thresholds below usage is crossed, we read
4454 * only one element of the array here.
4456 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4457 eventfd_signal(t->entries[i].eventfd, 1);
4459 /* i = current_threshold + 1 */
4463 * Iterate forward over array of thresholds starting from
4464 * current_threshold+1 and check if a threshold is crossed.
4465 * If none of thresholds above usage is crossed, we read
4466 * only one element of the array here.
4468 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4469 eventfd_signal(t->entries[i].eventfd, 1);
4471 /* Update current_threshold */
4472 t->current_threshold = i - 1;
4477 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4480 __mem_cgroup_threshold(memcg, false);
4481 if (do_swap_account)
4482 __mem_cgroup_threshold(memcg, true);
4484 memcg = parent_mem_cgroup(memcg);
4488 static int compare_thresholds(const void *a, const void *b)
4490 const struct mem_cgroup_threshold *_a = a;
4491 const struct mem_cgroup_threshold *_b = b;
4493 return _a->threshold - _b->threshold;
4496 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4498 struct mem_cgroup_eventfd_list *ev;
4500 list_for_each_entry(ev, &mem->oom_notify, list)
4501 eventfd_signal(ev->eventfd, 1);
4505 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4507 struct mem_cgroup *iter;
4509 for_each_mem_cgroup_tree(iter, mem)
4510 mem_cgroup_oom_notify_cb(iter);
4513 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4514 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4516 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4517 struct mem_cgroup_thresholds *thresholds;
4518 struct mem_cgroup_threshold_ary *new;
4519 int type = MEMFILE_TYPE(cft->private);
4520 u64 threshold, usage;
4523 ret = res_counter_memparse_write_strategy(args, &threshold);
4527 mutex_lock(&memcg->thresholds_lock);
4530 thresholds = &memcg->thresholds;
4531 else if (type == _MEMSWAP)
4532 thresholds = &memcg->memsw_thresholds;
4536 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4538 /* Check if a threshold crossed before adding a new one */
4539 if (thresholds->primary)
4540 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4542 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4544 /* Allocate memory for new array of thresholds */
4545 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4553 /* Copy thresholds (if any) to new array */
4554 if (thresholds->primary) {
4555 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4556 sizeof(struct mem_cgroup_threshold));
4559 /* Add new threshold */
4560 new->entries[size - 1].eventfd = eventfd;
4561 new->entries[size - 1].threshold = threshold;
4563 /* Sort thresholds. Registering of new threshold isn't time-critical */
4564 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4565 compare_thresholds, NULL);
4567 /* Find current threshold */
4568 new->current_threshold = -1;
4569 for (i = 0; i < size; i++) {
4570 if (new->entries[i].threshold < usage) {
4572 * new->current_threshold will not be used until
4573 * rcu_assign_pointer(), so it's safe to increment
4576 ++new->current_threshold;
4580 /* Free old spare buffer and save old primary buffer as spare */
4581 kfree(thresholds->spare);
4582 thresholds->spare = thresholds->primary;
4584 rcu_assign_pointer(thresholds->primary, new);
4586 /* To be sure that nobody uses thresholds */
4590 mutex_unlock(&memcg->thresholds_lock);
4595 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4596 struct cftype *cft, struct eventfd_ctx *eventfd)
4598 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4599 struct mem_cgroup_thresholds *thresholds;
4600 struct mem_cgroup_threshold_ary *new;
4601 int type = MEMFILE_TYPE(cft->private);
4605 mutex_lock(&memcg->thresholds_lock);
4607 thresholds = &memcg->thresholds;
4608 else if (type == _MEMSWAP)
4609 thresholds = &memcg->memsw_thresholds;
4614 * Something went wrong if we trying to unregister a threshold
4615 * if we don't have thresholds
4617 BUG_ON(!thresholds);
4619 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4621 /* Check if a threshold crossed before removing */
4622 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4624 /* Calculate new number of threshold */
4626 for (i = 0; i < thresholds->primary->size; i++) {
4627 if (thresholds->primary->entries[i].eventfd != eventfd)
4631 new = thresholds->spare;
4633 /* Set thresholds array to NULL if we don't have thresholds */
4642 /* Copy thresholds and find current threshold */
4643 new->current_threshold = -1;
4644 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4645 if (thresholds->primary->entries[i].eventfd == eventfd)
4648 new->entries[j] = thresholds->primary->entries[i];
4649 if (new->entries[j].threshold < usage) {
4651 * new->current_threshold will not be used
4652 * until rcu_assign_pointer(), so it's safe to increment
4655 ++new->current_threshold;
4661 /* Swap primary and spare array */
4662 thresholds->spare = thresholds->primary;
4663 rcu_assign_pointer(thresholds->primary, new);
4665 /* To be sure that nobody uses thresholds */
4668 mutex_unlock(&memcg->thresholds_lock);
4671 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4672 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4674 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4675 struct mem_cgroup_eventfd_list *event;
4676 int type = MEMFILE_TYPE(cft->private);
4678 BUG_ON(type != _OOM_TYPE);
4679 event = kmalloc(sizeof(*event), GFP_KERNEL);
4683 spin_lock(&memcg_oom_lock);
4685 event->eventfd = eventfd;
4686 list_add(&event->list, &memcg->oom_notify);
4688 /* already in OOM ? */
4689 if (atomic_read(&memcg->under_oom))
4690 eventfd_signal(eventfd, 1);
4691 spin_unlock(&memcg_oom_lock);
4696 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4697 struct cftype *cft, struct eventfd_ctx *eventfd)
4699 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4700 struct mem_cgroup_eventfd_list *ev, *tmp;
4701 int type = MEMFILE_TYPE(cft->private);
4703 BUG_ON(type != _OOM_TYPE);
4705 spin_lock(&memcg_oom_lock);
4707 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4708 if (ev->eventfd == eventfd) {
4709 list_del(&ev->list);
4714 spin_unlock(&memcg_oom_lock);
4717 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4718 struct cftype *cft, struct cgroup_map_cb *cb)
4720 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4722 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4724 if (atomic_read(&mem->under_oom))
4725 cb->fill(cb, "under_oom", 1);
4727 cb->fill(cb, "under_oom", 0);
4731 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4732 struct cftype *cft, u64 val)
4734 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4735 struct mem_cgroup *parent;
4737 /* cannot set to root cgroup and only 0 and 1 are allowed */
4738 if (!cgrp->parent || !((val == 0) || (val == 1)))
4741 parent = mem_cgroup_from_cont(cgrp->parent);
4744 /* oom-kill-disable is a flag for subhierarchy. */
4745 if ((parent->use_hierarchy) ||
4746 (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4750 mem->oom_kill_disable = val;
4752 memcg_oom_recover(mem);
4758 static const struct file_operations mem_control_numa_stat_file_operations = {
4760 .llseek = seq_lseek,
4761 .release = single_release,
4764 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4766 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4768 file->f_op = &mem_control_numa_stat_file_operations;
4769 return single_open(file, mem_control_numa_stat_show, cont);
4771 #endif /* CONFIG_NUMA */
4773 static int mem_cgroup_vmscan_stat_read(struct cgroup *cgrp,
4775 struct cgroup_map_cb *cb)
4777 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4781 for (i = 0; i < NR_SCANSTATS; i++) {
4782 strcpy(string, scanstat_string[i]);
4783 strcat(string, SCANSTAT_WORD_LIMIT);
4784 cb->fill(cb, string, mem->scanstat.stats[SCAN_BY_LIMIT][i]);
4787 for (i = 0; i < NR_SCANSTATS; i++) {
4788 strcpy(string, scanstat_string[i]);
4789 strcat(string, SCANSTAT_WORD_SYSTEM);
4790 cb->fill(cb, string, mem->scanstat.stats[SCAN_BY_SYSTEM][i]);
4793 for (i = 0; i < NR_SCANSTATS; i++) {
4794 strcpy(string, scanstat_string[i]);
4795 strcat(string, SCANSTAT_WORD_LIMIT);
4796 strcat(string, SCANSTAT_WORD_HIERARCHY);
4797 cb->fill(cb, string, mem->scanstat.rootstats[SCAN_BY_LIMIT][i]);
4799 for (i = 0; i < NR_SCANSTATS; i++) {
4800 strcpy(string, scanstat_string[i]);
4801 strcat(string, SCANSTAT_WORD_SYSTEM);
4802 strcat(string, SCANSTAT_WORD_HIERARCHY);
4803 cb->fill(cb, string, mem->scanstat.rootstats[SCAN_BY_SYSTEM][i]);
4808 static int mem_cgroup_reset_vmscan_stat(struct cgroup *cgrp,
4811 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4813 spin_lock(&mem->scanstat.lock);
4814 memset(&mem->scanstat.stats, 0, sizeof(mem->scanstat.stats));
4815 memset(&mem->scanstat.rootstats, 0, sizeof(mem->scanstat.rootstats));
4816 spin_unlock(&mem->scanstat.lock);
4821 static struct cftype mem_cgroup_files[] = {
4823 .name = "usage_in_bytes",
4824 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4825 .read_u64 = mem_cgroup_read,
4826 .register_event = mem_cgroup_usage_register_event,
4827 .unregister_event = mem_cgroup_usage_unregister_event,
4830 .name = "max_usage_in_bytes",
4831 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4832 .trigger = mem_cgroup_reset,
4833 .read_u64 = mem_cgroup_read,
4836 .name = "limit_in_bytes",
4837 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4838 .write_string = mem_cgroup_write,
4839 .read_u64 = mem_cgroup_read,
4842 .name = "soft_limit_in_bytes",
4843 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4844 .write_string = mem_cgroup_write,
4845 .read_u64 = mem_cgroup_read,
4849 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4850 .trigger = mem_cgroup_reset,
4851 .read_u64 = mem_cgroup_read,
4855 .read_map = mem_control_stat_show,
4858 .name = "force_empty",
4859 .trigger = mem_cgroup_force_empty_write,
4862 .name = "use_hierarchy",
4863 .write_u64 = mem_cgroup_hierarchy_write,
4864 .read_u64 = mem_cgroup_hierarchy_read,
4867 .name = "swappiness",
4868 .read_u64 = mem_cgroup_swappiness_read,
4869 .write_u64 = mem_cgroup_swappiness_write,
4872 .name = "move_charge_at_immigrate",
4873 .read_u64 = mem_cgroup_move_charge_read,
4874 .write_u64 = mem_cgroup_move_charge_write,
4877 .name = "oom_control",
4878 .read_map = mem_cgroup_oom_control_read,
4879 .write_u64 = mem_cgroup_oom_control_write,
4880 .register_event = mem_cgroup_oom_register_event,
4881 .unregister_event = mem_cgroup_oom_unregister_event,
4882 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4886 .name = "numa_stat",
4887 .open = mem_control_numa_stat_open,
4892 .name = "vmscan_stat",
4893 .read_map = mem_cgroup_vmscan_stat_read,
4894 .trigger = mem_cgroup_reset_vmscan_stat,
4898 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4899 static struct cftype memsw_cgroup_files[] = {
4901 .name = "memsw.usage_in_bytes",
4902 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4903 .read_u64 = mem_cgroup_read,
4904 .register_event = mem_cgroup_usage_register_event,
4905 .unregister_event = mem_cgroup_usage_unregister_event,
4908 .name = "memsw.max_usage_in_bytes",
4909 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4910 .trigger = mem_cgroup_reset,
4911 .read_u64 = mem_cgroup_read,
4914 .name = "memsw.limit_in_bytes",
4915 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4916 .write_string = mem_cgroup_write,
4917 .read_u64 = mem_cgroup_read,
4920 .name = "memsw.failcnt",
4921 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4922 .trigger = mem_cgroup_reset,
4923 .read_u64 = mem_cgroup_read,
4927 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4929 if (!do_swap_account)
4931 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4932 ARRAY_SIZE(memsw_cgroup_files));
4935 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4941 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4943 struct mem_cgroup_per_node *pn;
4944 struct mem_cgroup_per_zone *mz;
4946 int zone, tmp = node;
4948 * This routine is called against possible nodes.
4949 * But it's BUG to call kmalloc() against offline node.
4951 * TODO: this routine can waste much memory for nodes which will
4952 * never be onlined. It's better to use memory hotplug callback
4955 if (!node_state(node, N_NORMAL_MEMORY))
4957 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4961 mem->info.nodeinfo[node] = pn;
4962 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4963 mz = &pn->zoneinfo[zone];
4965 INIT_LIST_HEAD(&mz->lists[l]);
4966 mz->usage_in_excess = 0;
4967 mz->on_tree = false;
4973 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4975 kfree(mem->info.nodeinfo[node]);
4978 static struct mem_cgroup *mem_cgroup_alloc(void)
4980 struct mem_cgroup *mem;
4981 int size = sizeof(struct mem_cgroup);
4983 /* Can be very big if MAX_NUMNODES is very big */
4984 if (size < PAGE_SIZE)
4985 mem = kzalloc(size, GFP_KERNEL);
4987 mem = vzalloc(size);
4992 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4995 spin_lock_init(&mem->pcp_counter_lock);
4999 if (size < PAGE_SIZE)
5007 * At destroying mem_cgroup, references from swap_cgroup can remain.
5008 * (scanning all at force_empty is too costly...)
5010 * Instead of clearing all references at force_empty, we remember
5011 * the number of reference from swap_cgroup and free mem_cgroup when
5012 * it goes down to 0.
5014 * Removal of cgroup itself succeeds regardless of refs from swap.
5017 static void __mem_cgroup_free(struct mem_cgroup *mem)
5021 mem_cgroup_remove_from_trees(mem);
5022 free_css_id(&mem_cgroup_subsys, &mem->css);
5024 for_each_node_state(node, N_POSSIBLE)
5025 free_mem_cgroup_per_zone_info(mem, node);
5027 free_percpu(mem->stat);
5028 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
5034 static void mem_cgroup_get(struct mem_cgroup *mem)
5036 atomic_inc(&mem->refcnt);
5039 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
5041 if (atomic_sub_and_test(count, &mem->refcnt)) {
5042 struct mem_cgroup *parent = parent_mem_cgroup(mem);
5043 __mem_cgroup_free(mem);
5045 mem_cgroup_put(parent);
5049 static void mem_cgroup_put(struct mem_cgroup *mem)
5051 __mem_cgroup_put(mem, 1);
5055 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
5057 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
5059 if (!mem->res.parent)
5061 return mem_cgroup_from_res_counter(mem->res.parent, res);
5064 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5065 static void __init enable_swap_cgroup(void)
5067 if (!mem_cgroup_disabled() && really_do_swap_account)
5068 do_swap_account = 1;
5071 static void __init enable_swap_cgroup(void)
5076 static int mem_cgroup_soft_limit_tree_init(void)
5078 struct mem_cgroup_tree_per_node *rtpn;
5079 struct mem_cgroup_tree_per_zone *rtpz;
5080 int tmp, node, zone;
5082 for_each_node_state(node, N_POSSIBLE) {
5084 if (!node_state(node, N_NORMAL_MEMORY))
5086 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
5090 soft_limit_tree.rb_tree_per_node[node] = rtpn;
5092 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
5093 rtpz = &rtpn->rb_tree_per_zone[zone];
5094 rtpz->rb_root = RB_ROOT;
5095 spin_lock_init(&rtpz->lock);
5101 static struct cgroup_subsys_state * __ref
5102 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5104 struct mem_cgroup *mem, *parent;
5105 long error = -ENOMEM;
5108 mem = mem_cgroup_alloc();
5110 return ERR_PTR(error);
5112 for_each_node_state(node, N_POSSIBLE)
5113 if (alloc_mem_cgroup_per_zone_info(mem, node))
5117 if (cont->parent == NULL) {
5119 enable_swap_cgroup();
5121 root_mem_cgroup = mem;
5122 if (mem_cgroup_soft_limit_tree_init())
5124 for_each_possible_cpu(cpu) {
5125 struct memcg_stock_pcp *stock =
5126 &per_cpu(memcg_stock, cpu);
5127 INIT_WORK(&stock->work, drain_local_stock);
5129 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5131 parent = mem_cgroup_from_cont(cont->parent);
5132 mem->use_hierarchy = parent->use_hierarchy;
5133 mem->oom_kill_disable = parent->oom_kill_disable;
5136 if (parent && parent->use_hierarchy) {
5137 res_counter_init(&mem->res, &parent->res);
5138 res_counter_init(&mem->memsw, &parent->memsw);
5140 * We increment refcnt of the parent to ensure that we can
5141 * safely access it on res_counter_charge/uncharge.
5142 * This refcnt will be decremented when freeing this
5143 * mem_cgroup(see mem_cgroup_put).
5145 mem_cgroup_get(parent);
5147 res_counter_init(&mem->res, NULL);
5148 res_counter_init(&mem->memsw, NULL);
5150 mem->last_scanned_child = 0;
5151 mem->last_scanned_node = MAX_NUMNODES;
5152 INIT_LIST_HEAD(&mem->oom_notify);
5155 mem->swappiness = mem_cgroup_swappiness(parent);
5156 atomic_set(&mem->refcnt, 1);
5157 mem->move_charge_at_immigrate = 0;
5158 mutex_init(&mem->thresholds_lock);
5159 spin_lock_init(&mem->scanstat.lock);
5162 __mem_cgroup_free(mem);
5163 root_mem_cgroup = NULL;
5164 return ERR_PTR(error);
5167 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5168 struct cgroup *cont)
5170 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5172 return mem_cgroup_force_empty(mem, false);
5175 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5176 struct cgroup *cont)
5178 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5180 mem_cgroup_put(mem);
5183 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5184 struct cgroup *cont)
5188 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5189 ARRAY_SIZE(mem_cgroup_files));
5192 ret = register_memsw_files(cont, ss);
5197 /* Handlers for move charge at task migration. */
5198 #define PRECHARGE_COUNT_AT_ONCE 256
5199 static int mem_cgroup_do_precharge(unsigned long count)
5202 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5203 struct mem_cgroup *mem = mc.to;
5205 if (mem_cgroup_is_root(mem)) {
5206 mc.precharge += count;
5207 /* we don't need css_get for root */
5210 /* try to charge at once */
5212 struct res_counter *dummy;
5214 * "mem" cannot be under rmdir() because we've already checked
5215 * by cgroup_lock_live_cgroup() that it is not removed and we
5216 * are still under the same cgroup_mutex. So we can postpone
5219 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
5221 if (do_swap_account && res_counter_charge(&mem->memsw,
5222 PAGE_SIZE * count, &dummy)) {
5223 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
5226 mc.precharge += count;
5230 /* fall back to one by one charge */
5232 if (signal_pending(current)) {
5236 if (!batch_count--) {
5237 batch_count = PRECHARGE_COUNT_AT_ONCE;
5240 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5242 /* mem_cgroup_clear_mc() will do uncharge later */
5250 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5251 * @vma: the vma the pte to be checked belongs
5252 * @addr: the address corresponding to the pte to be checked
5253 * @ptent: the pte to be checked
5254 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5257 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5258 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5259 * move charge. if @target is not NULL, the page is stored in target->page
5260 * with extra refcnt got(Callers should handle it).
5261 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5262 * target for charge migration. if @target is not NULL, the entry is stored
5265 * Called with pte lock held.
5272 enum mc_target_type {
5273 MC_TARGET_NONE, /* not used */
5278 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5279 unsigned long addr, pte_t ptent)
5281 struct page *page = vm_normal_page(vma, addr, ptent);
5283 if (!page || !page_mapped(page))
5285 if (PageAnon(page)) {
5286 /* we don't move shared anon */
5287 if (!move_anon() || page_mapcount(page) > 2)
5289 } else if (!move_file())
5290 /* we ignore mapcount for file pages */
5292 if (!get_page_unless_zero(page))
5298 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5299 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5302 struct page *page = NULL;
5303 swp_entry_t ent = pte_to_swp_entry(ptent);
5305 if (!move_anon() || non_swap_entry(ent))
5307 usage_count = mem_cgroup_count_swap_user(ent, &page);
5308 if (usage_count > 1) { /* we don't move shared anon */
5313 if (do_swap_account)
5314 entry->val = ent.val;
5319 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5320 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5322 struct page *page = NULL;
5323 struct inode *inode;
5324 struct address_space *mapping;
5327 if (!vma->vm_file) /* anonymous vma */
5332 inode = vma->vm_file->f_path.dentry->d_inode;
5333 mapping = vma->vm_file->f_mapping;
5334 if (pte_none(ptent))
5335 pgoff = linear_page_index(vma, addr);
5336 else /* pte_file(ptent) is true */
5337 pgoff = pte_to_pgoff(ptent);
5339 /* page is moved even if it's not RSS of this task(page-faulted). */
5340 if (!mapping_cap_swap_backed(mapping)) { /* normal file */
5341 page = find_get_page(mapping, pgoff);
5342 } else { /* shmem/tmpfs file. we should take account of swap too. */
5344 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
5345 if (do_swap_account)
5346 entry->val = ent.val;
5352 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5353 unsigned long addr, pte_t ptent, union mc_target *target)
5355 struct page *page = NULL;
5356 struct page_cgroup *pc;
5358 swp_entry_t ent = { .val = 0 };
5360 if (pte_present(ptent))
5361 page = mc_handle_present_pte(vma, addr, ptent);
5362 else if (is_swap_pte(ptent))
5363 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5364 else if (pte_none(ptent) || pte_file(ptent))
5365 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5367 if (!page && !ent.val)
5370 pc = lookup_page_cgroup(page);
5372 * Do only loose check w/o page_cgroup lock.
5373 * mem_cgroup_move_account() checks the pc is valid or not under
5376 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5377 ret = MC_TARGET_PAGE;
5379 target->page = page;
5381 if (!ret || !target)
5384 /* There is a swap entry and a page doesn't exist or isn't charged */
5385 if (ent.val && !ret &&
5386 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5387 ret = MC_TARGET_SWAP;
5394 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5395 unsigned long addr, unsigned long end,
5396 struct mm_walk *walk)
5398 struct vm_area_struct *vma = walk->private;
5402 split_huge_page_pmd(walk->mm, pmd);
5404 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5405 for (; addr != end; pte++, addr += PAGE_SIZE)
5406 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5407 mc.precharge++; /* increment precharge temporarily */
5408 pte_unmap_unlock(pte - 1, ptl);
5414 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5416 unsigned long precharge;
5417 struct vm_area_struct *vma;
5419 down_read(&mm->mmap_sem);
5420 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5421 struct mm_walk mem_cgroup_count_precharge_walk = {
5422 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5426 if (is_vm_hugetlb_page(vma))
5428 walk_page_range(vma->vm_start, vma->vm_end,
5429 &mem_cgroup_count_precharge_walk);
5431 up_read(&mm->mmap_sem);
5433 precharge = mc.precharge;
5439 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5441 unsigned long precharge = mem_cgroup_count_precharge(mm);
5443 VM_BUG_ON(mc.moving_task);
5444 mc.moving_task = current;
5445 return mem_cgroup_do_precharge(precharge);
5448 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5449 static void __mem_cgroup_clear_mc(void)
5451 struct mem_cgroup *from = mc.from;
5452 struct mem_cgroup *to = mc.to;
5454 /* we must uncharge all the leftover precharges from mc.to */
5456 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5460 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5461 * we must uncharge here.
5463 if (mc.moved_charge) {
5464 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5465 mc.moved_charge = 0;
5467 /* we must fixup refcnts and charges */
5468 if (mc.moved_swap) {
5469 /* uncharge swap account from the old cgroup */
5470 if (!mem_cgroup_is_root(mc.from))
5471 res_counter_uncharge(&mc.from->memsw,
5472 PAGE_SIZE * mc.moved_swap);
5473 __mem_cgroup_put(mc.from, mc.moved_swap);
5475 if (!mem_cgroup_is_root(mc.to)) {
5477 * we charged both to->res and to->memsw, so we should
5480 res_counter_uncharge(&mc.to->res,
5481 PAGE_SIZE * mc.moved_swap);
5483 /* we've already done mem_cgroup_get(mc.to) */
5486 memcg_oom_recover(from);
5487 memcg_oom_recover(to);
5488 wake_up_all(&mc.waitq);
5491 static void mem_cgroup_clear_mc(void)
5493 struct mem_cgroup *from = mc.from;
5496 * we must clear moving_task before waking up waiters at the end of
5499 mc.moving_task = NULL;
5500 __mem_cgroup_clear_mc();
5501 spin_lock(&mc.lock);
5504 spin_unlock(&mc.lock);
5505 mem_cgroup_end_move(from);
5508 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5509 struct cgroup *cgroup,
5510 struct task_struct *p)
5513 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5515 if (mem->move_charge_at_immigrate) {
5516 struct mm_struct *mm;
5517 struct mem_cgroup *from = mem_cgroup_from_task(p);
5519 VM_BUG_ON(from == mem);
5521 mm = get_task_mm(p);
5524 /* We move charges only when we move a owner of the mm */
5525 if (mm->owner == p) {
5528 VM_BUG_ON(mc.precharge);
5529 VM_BUG_ON(mc.moved_charge);
5530 VM_BUG_ON(mc.moved_swap);
5531 mem_cgroup_start_move(from);
5532 spin_lock(&mc.lock);
5535 spin_unlock(&mc.lock);
5536 /* We set mc.moving_task later */
5538 ret = mem_cgroup_precharge_mc(mm);
5540 mem_cgroup_clear_mc();
5547 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5548 struct cgroup *cgroup,
5549 struct task_struct *p)
5551 mem_cgroup_clear_mc();
5554 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5555 unsigned long addr, unsigned long end,
5556 struct mm_walk *walk)
5559 struct vm_area_struct *vma = walk->private;
5563 split_huge_page_pmd(walk->mm, pmd);
5565 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5566 for (; addr != end; addr += PAGE_SIZE) {
5567 pte_t ptent = *(pte++);
5568 union mc_target target;
5571 struct page_cgroup *pc;
5577 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5579 case MC_TARGET_PAGE:
5581 if (isolate_lru_page(page))
5583 pc = lookup_page_cgroup(page);
5584 if (!mem_cgroup_move_account(page, 1, pc,
5585 mc.from, mc.to, false)) {
5587 /* we uncharge from mc.from later. */
5590 putback_lru_page(page);
5591 put: /* is_target_pte_for_mc() gets the page */
5594 case MC_TARGET_SWAP:
5596 if (!mem_cgroup_move_swap_account(ent,
5597 mc.from, mc.to, false)) {
5599 /* we fixup refcnts and charges later. */
5607 pte_unmap_unlock(pte - 1, ptl);
5612 * We have consumed all precharges we got in can_attach().
5613 * We try charge one by one, but don't do any additional
5614 * charges to mc.to if we have failed in charge once in attach()
5617 ret = mem_cgroup_do_precharge(1);
5625 static void mem_cgroup_move_charge(struct mm_struct *mm)
5627 struct vm_area_struct *vma;
5629 lru_add_drain_all();
5631 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5633 * Someone who are holding the mmap_sem might be waiting in
5634 * waitq. So we cancel all extra charges, wake up all waiters,
5635 * and retry. Because we cancel precharges, we might not be able
5636 * to move enough charges, but moving charge is a best-effort
5637 * feature anyway, so it wouldn't be a big problem.
5639 __mem_cgroup_clear_mc();
5643 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5645 struct mm_walk mem_cgroup_move_charge_walk = {
5646 .pmd_entry = mem_cgroup_move_charge_pte_range,
5650 if (is_vm_hugetlb_page(vma))
5652 ret = walk_page_range(vma->vm_start, vma->vm_end,
5653 &mem_cgroup_move_charge_walk);
5656 * means we have consumed all precharges and failed in
5657 * doing additional charge. Just abandon here.
5661 up_read(&mm->mmap_sem);
5664 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5665 struct cgroup *cont,
5666 struct cgroup *old_cont,
5667 struct task_struct *p)
5669 struct mm_struct *mm = get_task_mm(p);
5673 mem_cgroup_move_charge(mm);
5678 mem_cgroup_clear_mc();
5680 #else /* !CONFIG_MMU */
5681 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5682 struct cgroup *cgroup,
5683 struct task_struct *p)
5687 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5688 struct cgroup *cgroup,
5689 struct task_struct *p)
5692 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5693 struct cgroup *cont,
5694 struct cgroup *old_cont,
5695 struct task_struct *p)
5700 struct cgroup_subsys mem_cgroup_subsys = {
5702 .subsys_id = mem_cgroup_subsys_id,
5703 .create = mem_cgroup_create,
5704 .pre_destroy = mem_cgroup_pre_destroy,
5705 .destroy = mem_cgroup_destroy,
5706 .populate = mem_cgroup_populate,
5707 .can_attach = mem_cgroup_can_attach,
5708 .cancel_attach = mem_cgroup_cancel_attach,
5709 .attach = mem_cgroup_move_task,
5714 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5715 static int __init enable_swap_account(char *s)
5717 /* consider enabled if no parameter or 1 is given */
5718 if (!strcmp(s, "1"))
5719 really_do_swap_account = 1;
5720 else if (!strcmp(s, "0"))
5721 really_do_swap_account = 0;
5724 __setup("swapaccount=", enable_swap_account);