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,
113 #define THRESHOLDS_EVENTS_TARGET (128)
114 #define SOFTLIMIT_EVENTS_TARGET (1024)
116 struct mem_cgroup_stat_cpu {
117 long count[MEM_CGROUP_STAT_NSTATS];
118 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
119 unsigned long targets[MEM_CGROUP_NTARGETS];
123 * per-zone information in memory controller.
125 struct mem_cgroup_per_zone {
127 * spin_lock to protect the per cgroup LRU
129 struct list_head lists[NR_LRU_LISTS];
130 unsigned long count[NR_LRU_LISTS];
132 struct zone_reclaim_stat reclaim_stat;
133 struct rb_node tree_node; /* RB tree node */
134 unsigned long long usage_in_excess;/* Set to the value by which */
135 /* the soft limit is exceeded*/
137 struct mem_cgroup *mem; /* Back pointer, we cannot */
138 /* use container_of */
140 /* Macro for accessing counter */
141 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
143 struct mem_cgroup_per_node {
144 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
147 struct mem_cgroup_lru_info {
148 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
152 * Cgroups above their limits are maintained in a RB-Tree, independent of
153 * their hierarchy representation
156 struct mem_cgroup_tree_per_zone {
157 struct rb_root rb_root;
161 struct mem_cgroup_tree_per_node {
162 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
165 struct mem_cgroup_tree {
166 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
169 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
171 struct mem_cgroup_threshold {
172 struct eventfd_ctx *eventfd;
177 struct mem_cgroup_threshold_ary {
178 /* An array index points to threshold just below usage. */
179 int current_threshold;
180 /* Size of entries[] */
182 /* Array of thresholds */
183 struct mem_cgroup_threshold entries[0];
186 struct mem_cgroup_thresholds {
187 /* Primary thresholds array */
188 struct mem_cgroup_threshold_ary *primary;
190 * Spare threshold array.
191 * This is needed to make mem_cgroup_unregister_event() "never fail".
192 * It must be able to store at least primary->size - 1 entries.
194 struct mem_cgroup_threshold_ary *spare;
198 struct mem_cgroup_eventfd_list {
199 struct list_head list;
200 struct eventfd_ctx *eventfd;
203 static void mem_cgroup_threshold(struct mem_cgroup *mem);
204 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
207 * The memory controller data structure. The memory controller controls both
208 * page cache and RSS per cgroup. We would eventually like to provide
209 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
210 * to help the administrator determine what knobs to tune.
212 * TODO: Add a water mark for the memory controller. Reclaim will begin when
213 * we hit the water mark. May be even add a low water mark, such that
214 * no reclaim occurs from a cgroup at it's low water mark, this is
215 * a feature that will be implemented much later in the future.
218 struct cgroup_subsys_state css;
220 * the counter to account for memory usage
222 struct res_counter res;
224 * the counter to account for mem+swap usage.
226 struct res_counter memsw;
228 * Per cgroup active and inactive list, similar to the
229 * per zone LRU lists.
231 struct mem_cgroup_lru_info info;
233 * While reclaiming in a hierarchy, we cache the last child we
236 int last_scanned_child;
237 int last_scanned_node;
239 nodemask_t scan_nodes;
240 unsigned long next_scan_node_update;
243 * Should the accounting and control be hierarchical, per subtree?
249 unsigned int swappiness;
250 /* OOM-Killer disable */
251 int oom_kill_disable;
253 /* set when res.limit == memsw.limit */
254 bool memsw_is_minimum;
256 /* protect arrays of thresholds */
257 struct mutex thresholds_lock;
259 /* thresholds for memory usage. RCU-protected */
260 struct mem_cgroup_thresholds thresholds;
262 /* thresholds for mem+swap usage. RCU-protected */
263 struct mem_cgroup_thresholds memsw_thresholds;
265 /* For oom notifier event fd */
266 struct list_head oom_notify;
269 * Should we move charges of a task when a task is moved into this
270 * mem_cgroup ? And what type of charges should we move ?
272 unsigned long move_charge_at_immigrate;
276 struct mem_cgroup_stat_cpu *stat;
278 * used when a cpu is offlined or other synchronizations
279 * See mem_cgroup_read_stat().
281 struct mem_cgroup_stat_cpu nocpu_base;
282 spinlock_t pcp_counter_lock;
285 /* Stuffs for move charges at task migration. */
287 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
288 * left-shifted bitmap of these types.
291 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
292 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
296 /* "mc" and its members are protected by cgroup_mutex */
297 static struct move_charge_struct {
298 spinlock_t lock; /* for from, to */
299 struct mem_cgroup *from;
300 struct mem_cgroup *to;
301 unsigned long precharge;
302 unsigned long moved_charge;
303 unsigned long moved_swap;
304 struct task_struct *moving_task; /* a task moving charges */
305 wait_queue_head_t waitq; /* a waitq for other context */
307 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
308 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
311 static bool move_anon(void)
313 return test_bit(MOVE_CHARGE_TYPE_ANON,
314 &mc.to->move_charge_at_immigrate);
317 static bool move_file(void)
319 return test_bit(MOVE_CHARGE_TYPE_FILE,
320 &mc.to->move_charge_at_immigrate);
324 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
325 * limit reclaim to prevent infinite loops, if they ever occur.
327 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
328 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
331 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
332 MEM_CGROUP_CHARGE_TYPE_MAPPED,
333 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
334 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
335 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
336 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
340 /* for encoding cft->private value on file */
343 #define _OOM_TYPE (2)
344 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
345 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
346 #define MEMFILE_ATTR(val) ((val) & 0xffff)
347 /* Used for OOM nofiier */
348 #define OOM_CONTROL (0)
351 * Reclaim flags for mem_cgroup_hierarchical_reclaim
353 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
354 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
355 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
356 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
357 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
358 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
360 static void mem_cgroup_get(struct mem_cgroup *mem);
361 static void mem_cgroup_put(struct mem_cgroup *mem);
362 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
363 static void drain_all_stock_async(struct mem_cgroup *mem);
365 static struct mem_cgroup_per_zone *
366 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
368 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
371 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
376 static struct mem_cgroup_per_zone *
377 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
379 int nid = page_to_nid(page);
380 int zid = page_zonenum(page);
382 return mem_cgroup_zoneinfo(mem, nid, zid);
385 static struct mem_cgroup_tree_per_zone *
386 soft_limit_tree_node_zone(int nid, int zid)
388 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
391 static struct mem_cgroup_tree_per_zone *
392 soft_limit_tree_from_page(struct page *page)
394 int nid = page_to_nid(page);
395 int zid = page_zonenum(page);
397 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
401 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
402 struct mem_cgroup_per_zone *mz,
403 struct mem_cgroup_tree_per_zone *mctz,
404 unsigned long long new_usage_in_excess)
406 struct rb_node **p = &mctz->rb_root.rb_node;
407 struct rb_node *parent = NULL;
408 struct mem_cgroup_per_zone *mz_node;
413 mz->usage_in_excess = new_usage_in_excess;
414 if (!mz->usage_in_excess)
418 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
420 if (mz->usage_in_excess < mz_node->usage_in_excess)
423 * We can't avoid mem cgroups that are over their soft
424 * limit by the same amount
426 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
429 rb_link_node(&mz->tree_node, parent, p);
430 rb_insert_color(&mz->tree_node, &mctz->rb_root);
435 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
436 struct mem_cgroup_per_zone *mz,
437 struct mem_cgroup_tree_per_zone *mctz)
441 rb_erase(&mz->tree_node, &mctz->rb_root);
446 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
447 struct mem_cgroup_per_zone *mz,
448 struct mem_cgroup_tree_per_zone *mctz)
450 spin_lock(&mctz->lock);
451 __mem_cgroup_remove_exceeded(mem, mz, mctz);
452 spin_unlock(&mctz->lock);
456 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
458 unsigned long long excess;
459 struct mem_cgroup_per_zone *mz;
460 struct mem_cgroup_tree_per_zone *mctz;
461 int nid = page_to_nid(page);
462 int zid = page_zonenum(page);
463 mctz = soft_limit_tree_from_page(page);
466 * Necessary to update all ancestors when hierarchy is used.
467 * because their event counter is not touched.
469 for (; mem; mem = parent_mem_cgroup(mem)) {
470 mz = mem_cgroup_zoneinfo(mem, nid, zid);
471 excess = res_counter_soft_limit_excess(&mem->res);
473 * We have to update the tree if mz is on RB-tree or
474 * mem is over its softlimit.
476 if (excess || mz->on_tree) {
477 spin_lock(&mctz->lock);
478 /* if on-tree, remove it */
480 __mem_cgroup_remove_exceeded(mem, mz, mctz);
482 * Insert again. mz->usage_in_excess will be updated.
483 * If excess is 0, no tree ops.
485 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
486 spin_unlock(&mctz->lock);
491 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
494 struct mem_cgroup_per_zone *mz;
495 struct mem_cgroup_tree_per_zone *mctz;
497 for_each_node_state(node, N_POSSIBLE) {
498 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
499 mz = mem_cgroup_zoneinfo(mem, node, zone);
500 mctz = soft_limit_tree_node_zone(node, zone);
501 mem_cgroup_remove_exceeded(mem, mz, mctz);
506 static struct mem_cgroup_per_zone *
507 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
509 struct rb_node *rightmost = NULL;
510 struct mem_cgroup_per_zone *mz;
514 rightmost = rb_last(&mctz->rb_root);
516 goto done; /* Nothing to reclaim from */
518 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
520 * Remove the node now but someone else can add it back,
521 * we will to add it back at the end of reclaim to its correct
522 * position in the tree.
524 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
525 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
526 !css_tryget(&mz->mem->css))
532 static struct mem_cgroup_per_zone *
533 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
535 struct mem_cgroup_per_zone *mz;
537 spin_lock(&mctz->lock);
538 mz = __mem_cgroup_largest_soft_limit_node(mctz);
539 spin_unlock(&mctz->lock);
544 * Implementation Note: reading percpu statistics for memcg.
546 * Both of vmstat[] and percpu_counter has threshold and do periodic
547 * synchronization to implement "quick" read. There are trade-off between
548 * reading cost and precision of value. Then, we may have a chance to implement
549 * a periodic synchronizion of counter in memcg's counter.
551 * But this _read() function is used for user interface now. The user accounts
552 * memory usage by memory cgroup and he _always_ requires exact value because
553 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
554 * have to visit all online cpus and make sum. So, for now, unnecessary
555 * synchronization is not implemented. (just implemented for cpu hotplug)
557 * If there are kernel internal actions which can make use of some not-exact
558 * value, and reading all cpu value can be performance bottleneck in some
559 * common workload, threashold and synchonization as vmstat[] should be
562 static long mem_cgroup_read_stat(struct mem_cgroup *mem,
563 enum mem_cgroup_stat_index idx)
569 for_each_online_cpu(cpu)
570 val += per_cpu(mem->stat->count[idx], cpu);
571 #ifdef CONFIG_HOTPLUG_CPU
572 spin_lock(&mem->pcp_counter_lock);
573 val += mem->nocpu_base.count[idx];
574 spin_unlock(&mem->pcp_counter_lock);
580 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
583 int val = (charge) ? 1 : -1;
584 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
587 void mem_cgroup_pgfault(struct mem_cgroup *mem, int val)
589 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
592 void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val)
594 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
597 static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
598 enum mem_cgroup_events_index idx)
600 unsigned long val = 0;
603 for_each_online_cpu(cpu)
604 val += per_cpu(mem->stat->events[idx], cpu);
605 #ifdef CONFIG_HOTPLUG_CPU
606 spin_lock(&mem->pcp_counter_lock);
607 val += mem->nocpu_base.events[idx];
608 spin_unlock(&mem->pcp_counter_lock);
613 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
614 bool file, int nr_pages)
619 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
621 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
623 /* pagein of a big page is an event. So, ignore page size */
625 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
627 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
628 nr_pages = -nr_pages; /* for event */
631 __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
637 mem_cgroup_get_zonestat_node(struct mem_cgroup *mem, int nid, enum lru_list idx)
639 struct mem_cgroup_per_zone *mz;
643 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
644 mz = mem_cgroup_zoneinfo(mem, nid, zid);
645 total += MEM_CGROUP_ZSTAT(mz, idx);
649 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
655 for_each_online_node(nid)
656 total += mem_cgroup_get_zonestat_node(mem, nid, idx);
660 static bool __memcg_event_check(struct mem_cgroup *mem, int target)
662 unsigned long val, next;
664 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
665 next = this_cpu_read(mem->stat->targets[target]);
666 /* from time_after() in jiffies.h */
667 return ((long)next - (long)val < 0);
670 static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target)
672 unsigned long val, next;
674 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
677 case MEM_CGROUP_TARGET_THRESH:
678 next = val + THRESHOLDS_EVENTS_TARGET;
680 case MEM_CGROUP_TARGET_SOFTLIMIT:
681 next = val + SOFTLIMIT_EVENTS_TARGET;
687 this_cpu_write(mem->stat->targets[target], next);
691 * Check events in order.
694 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
696 /* threshold event is triggered in finer grain than soft limit */
697 if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) {
698 mem_cgroup_threshold(mem);
699 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH);
700 if (unlikely(__memcg_event_check(mem,
701 MEM_CGROUP_TARGET_SOFTLIMIT))){
702 mem_cgroup_update_tree(mem, page);
703 __mem_cgroup_target_update(mem,
704 MEM_CGROUP_TARGET_SOFTLIMIT);
709 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
711 return container_of(cgroup_subsys_state(cont,
712 mem_cgroup_subsys_id), struct mem_cgroup,
716 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
719 * mm_update_next_owner() may clear mm->owner to NULL
720 * if it races with swapoff, page migration, etc.
721 * So this can be called with p == NULL.
726 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
727 struct mem_cgroup, css);
730 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
732 struct mem_cgroup *mem = NULL;
737 * Because we have no locks, mm->owner's may be being moved to other
738 * cgroup. We use css_tryget() here even if this looks
739 * pessimistic (rather than adding locks here).
743 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
746 } while (!css_tryget(&mem->css));
751 /* The caller has to guarantee "mem" exists before calling this */
752 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
754 struct cgroup_subsys_state *css;
757 if (!mem) /* ROOT cgroup has the smallest ID */
758 return root_mem_cgroup; /*css_put/get against root is ignored*/
759 if (!mem->use_hierarchy) {
760 if (css_tryget(&mem->css))
766 * searching a memory cgroup which has the smallest ID under given
767 * ROOT cgroup. (ID >= 1)
769 css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
770 if (css && css_tryget(css))
771 mem = container_of(css, struct mem_cgroup, css);
778 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
779 struct mem_cgroup *root,
782 int nextid = css_id(&iter->css) + 1;
785 struct cgroup_subsys_state *css;
787 hierarchy_used = iter->use_hierarchy;
790 /* If no ROOT, walk all, ignore hierarchy */
791 if (!cond || (root && !hierarchy_used))
795 root = root_mem_cgroup;
801 css = css_get_next(&mem_cgroup_subsys, nextid,
803 if (css && css_tryget(css))
804 iter = container_of(css, struct mem_cgroup, css);
806 /* If css is NULL, no more cgroups will be found */
808 } while (css && !iter);
813 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
814 * be careful that "break" loop is not allowed. We have reference count.
815 * Instead of that modify "cond" to be false and "continue" to exit the loop.
817 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
818 for (iter = mem_cgroup_start_loop(root);\
820 iter = mem_cgroup_get_next(iter, root, cond))
822 #define for_each_mem_cgroup_tree(iter, root) \
823 for_each_mem_cgroup_tree_cond(iter, root, true)
825 #define for_each_mem_cgroup_all(iter) \
826 for_each_mem_cgroup_tree_cond(iter, NULL, true)
829 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
831 return (mem == root_mem_cgroup);
834 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
836 struct mem_cgroup *mem;
842 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
848 mem_cgroup_pgmajfault(mem, 1);
851 mem_cgroup_pgfault(mem, 1);
859 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
862 * Following LRU functions are allowed to be used without PCG_LOCK.
863 * Operations are called by routine of global LRU independently from memcg.
864 * What we have to take care of here is validness of pc->mem_cgroup.
866 * Changes to pc->mem_cgroup happens when
869 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
870 * It is added to LRU before charge.
871 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
872 * When moving account, the page is not on LRU. It's isolated.
875 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
877 struct page_cgroup *pc;
878 struct mem_cgroup_per_zone *mz;
880 if (mem_cgroup_disabled())
882 pc = lookup_page_cgroup(page);
883 /* can happen while we handle swapcache. */
884 if (!TestClearPageCgroupAcctLRU(pc))
886 VM_BUG_ON(!pc->mem_cgroup);
888 * We don't check PCG_USED bit. It's cleared when the "page" is finally
889 * removed from global LRU.
891 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
892 /* huge page split is done under lru_lock. so, we have no races. */
893 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
894 if (mem_cgroup_is_root(pc->mem_cgroup))
896 VM_BUG_ON(list_empty(&pc->lru));
897 list_del_init(&pc->lru);
900 void mem_cgroup_del_lru(struct page *page)
902 mem_cgroup_del_lru_list(page, page_lru(page));
906 * Writeback is about to end against a page which has been marked for immediate
907 * reclaim. If it still appears to be reclaimable, move it to the tail of the
910 void mem_cgroup_rotate_reclaimable_page(struct page *page)
912 struct mem_cgroup_per_zone *mz;
913 struct page_cgroup *pc;
914 enum lru_list lru = page_lru(page);
916 if (mem_cgroup_disabled())
919 pc = lookup_page_cgroup(page);
920 /* unused or root page is not rotated. */
921 if (!PageCgroupUsed(pc))
923 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
925 if (mem_cgroup_is_root(pc->mem_cgroup))
927 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
928 list_move_tail(&pc->lru, &mz->lists[lru]);
931 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
933 struct mem_cgroup_per_zone *mz;
934 struct page_cgroup *pc;
936 if (mem_cgroup_disabled())
939 pc = lookup_page_cgroup(page);
940 /* unused or root page is not rotated. */
941 if (!PageCgroupUsed(pc))
943 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
945 if (mem_cgroup_is_root(pc->mem_cgroup))
947 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
948 list_move(&pc->lru, &mz->lists[lru]);
951 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
953 struct page_cgroup *pc;
954 struct mem_cgroup_per_zone *mz;
956 if (mem_cgroup_disabled())
958 pc = lookup_page_cgroup(page);
959 VM_BUG_ON(PageCgroupAcctLRU(pc));
960 if (!PageCgroupUsed(pc))
962 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
964 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
965 /* huge page split is done under lru_lock. so, we have no races. */
966 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
967 SetPageCgroupAcctLRU(pc);
968 if (mem_cgroup_is_root(pc->mem_cgroup))
970 list_add(&pc->lru, &mz->lists[lru]);
974 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
975 * while it's linked to lru because the page may be reused after it's fully
976 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
977 * It's done under lock_page and expected that zone->lru_lock isnever held.
979 static void mem_cgroup_lru_del_before_commit(struct page *page)
982 struct zone *zone = page_zone(page);
983 struct page_cgroup *pc = lookup_page_cgroup(page);
986 * Doing this check without taking ->lru_lock seems wrong but this
987 * is safe. Because if page_cgroup's USED bit is unset, the page
988 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
989 * set, the commit after this will fail, anyway.
990 * This all charge/uncharge is done under some mutual execustion.
991 * So, we don't need to taking care of changes in USED bit.
993 if (likely(!PageLRU(page)))
996 spin_lock_irqsave(&zone->lru_lock, flags);
998 * Forget old LRU when this page_cgroup is *not* used. This Used bit
999 * is guarded by lock_page() because the page is SwapCache.
1001 if (!PageCgroupUsed(pc))
1002 mem_cgroup_del_lru_list(page, page_lru(page));
1003 spin_unlock_irqrestore(&zone->lru_lock, flags);
1006 static void mem_cgroup_lru_add_after_commit(struct page *page)
1008 unsigned long flags;
1009 struct zone *zone = page_zone(page);
1010 struct page_cgroup *pc = lookup_page_cgroup(page);
1012 /* taking care of that the page is added to LRU while we commit it */
1013 if (likely(!PageLRU(page)))
1015 spin_lock_irqsave(&zone->lru_lock, flags);
1016 /* link when the page is linked to LRU but page_cgroup isn't */
1017 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1018 mem_cgroup_add_lru_list(page, page_lru(page));
1019 spin_unlock_irqrestore(&zone->lru_lock, flags);
1023 void mem_cgroup_move_lists(struct page *page,
1024 enum lru_list from, enum lru_list to)
1026 if (mem_cgroup_disabled())
1028 mem_cgroup_del_lru_list(page, from);
1029 mem_cgroup_add_lru_list(page, to);
1032 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
1035 struct mem_cgroup *curr = NULL;
1036 struct task_struct *p;
1038 p = find_lock_task_mm(task);
1041 curr = try_get_mem_cgroup_from_mm(p->mm);
1046 * We should check use_hierarchy of "mem" not "curr". Because checking
1047 * use_hierarchy of "curr" here make this function true if hierarchy is
1048 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
1049 * hierarchy(even if use_hierarchy is disabled in "mem").
1051 if (mem->use_hierarchy)
1052 ret = css_is_ancestor(&curr->css, &mem->css);
1054 ret = (curr == mem);
1055 css_put(&curr->css);
1059 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1061 unsigned long active;
1062 unsigned long inactive;
1064 unsigned long inactive_ratio;
1066 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
1067 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
1069 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1071 inactive_ratio = int_sqrt(10 * gb);
1075 if (present_pages) {
1076 present_pages[0] = inactive;
1077 present_pages[1] = active;
1080 return inactive_ratio;
1083 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1085 unsigned long active;
1086 unsigned long inactive;
1087 unsigned long present_pages[2];
1088 unsigned long inactive_ratio;
1090 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1092 inactive = present_pages[0];
1093 active = present_pages[1];
1095 if (inactive * inactive_ratio < active)
1101 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1103 unsigned long active;
1104 unsigned long inactive;
1106 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
1107 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
1109 return (active > inactive);
1112 unsigned long mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg,
1116 int nid = zone_to_nid(zone);
1117 int zid = zone_idx(zone);
1118 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1120 return MEM_CGROUP_ZSTAT(mz, lru);
1123 static unsigned long mem_cgroup_node_nr_file_lru_pages(struct mem_cgroup *memcg,
1128 ret = mem_cgroup_get_zonestat_node(memcg, nid, LRU_INACTIVE_FILE) +
1129 mem_cgroup_get_zonestat_node(memcg, nid, LRU_ACTIVE_FILE);
1134 static unsigned long mem_cgroup_node_nr_anon_lru_pages(struct mem_cgroup *memcg,
1139 ret = mem_cgroup_get_zonestat_node(memcg, nid, LRU_INACTIVE_ANON) +
1140 mem_cgroup_get_zonestat_node(memcg, nid, LRU_ACTIVE_ANON);
1144 #if MAX_NUMNODES > 1
1145 static unsigned long mem_cgroup_nr_file_lru_pages(struct mem_cgroup *memcg)
1150 for_each_node_state(nid, N_HIGH_MEMORY)
1151 total += mem_cgroup_node_nr_file_lru_pages(memcg, nid);
1156 static unsigned long mem_cgroup_nr_anon_lru_pages(struct mem_cgroup *memcg)
1161 for_each_node_state(nid, N_HIGH_MEMORY)
1162 total += mem_cgroup_node_nr_anon_lru_pages(memcg, nid);
1167 static unsigned long
1168 mem_cgroup_node_nr_unevictable_lru_pages(struct mem_cgroup *memcg, int nid)
1170 return mem_cgroup_get_zonestat_node(memcg, nid, LRU_UNEVICTABLE);
1173 static unsigned long
1174 mem_cgroup_nr_unevictable_lru_pages(struct mem_cgroup *memcg)
1179 for_each_node_state(nid, N_HIGH_MEMORY)
1180 total += mem_cgroup_node_nr_unevictable_lru_pages(memcg, nid);
1185 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
1192 total += mem_cgroup_get_zonestat_node(memcg, nid, l);
1197 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg)
1202 for_each_node_state(nid, N_HIGH_MEMORY)
1203 total += mem_cgroup_node_nr_lru_pages(memcg, nid);
1207 #endif /* CONFIG_NUMA */
1209 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1212 int nid = zone_to_nid(zone);
1213 int zid = zone_idx(zone);
1214 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1216 return &mz->reclaim_stat;
1219 struct zone_reclaim_stat *
1220 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1222 struct page_cgroup *pc;
1223 struct mem_cgroup_per_zone *mz;
1225 if (mem_cgroup_disabled())
1228 pc = lookup_page_cgroup(page);
1229 if (!PageCgroupUsed(pc))
1231 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1233 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1234 return &mz->reclaim_stat;
1237 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1238 struct list_head *dst,
1239 unsigned long *scanned, int order,
1240 int mode, struct zone *z,
1241 struct mem_cgroup *mem_cont,
1242 int active, int file)
1244 unsigned long nr_taken = 0;
1248 struct list_head *src;
1249 struct page_cgroup *pc, *tmp;
1250 int nid = zone_to_nid(z);
1251 int zid = zone_idx(z);
1252 struct mem_cgroup_per_zone *mz;
1253 int lru = LRU_FILE * file + active;
1257 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1258 src = &mz->lists[lru];
1261 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1262 if (scan >= nr_to_scan)
1265 if (unlikely(!PageCgroupUsed(pc)))
1268 page = lookup_cgroup_page(pc);
1270 if (unlikely(!PageLRU(page)))
1274 ret = __isolate_lru_page(page, mode, file);
1277 list_move(&page->lru, dst);
1278 mem_cgroup_del_lru(page);
1279 nr_taken += hpage_nr_pages(page);
1282 /* we don't affect global LRU but rotate in our LRU */
1283 mem_cgroup_rotate_lru_list(page, page_lru(page));
1292 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1298 #define mem_cgroup_from_res_counter(counter, member) \
1299 container_of(counter, struct mem_cgroup, member)
1302 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1303 * @mem: the memory cgroup
1305 * Returns the maximum amount of memory @mem can be charged with, in
1308 static unsigned long mem_cgroup_margin(struct mem_cgroup *mem)
1310 unsigned long long margin;
1312 margin = res_counter_margin(&mem->res);
1313 if (do_swap_account)
1314 margin = min(margin, res_counter_margin(&mem->memsw));
1315 return margin >> PAGE_SHIFT;
1318 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1320 struct cgroup *cgrp = memcg->css.cgroup;
1323 if (cgrp->parent == NULL)
1324 return vm_swappiness;
1326 return memcg->swappiness;
1329 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1334 spin_lock(&mem->pcp_counter_lock);
1335 for_each_online_cpu(cpu)
1336 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1337 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1338 spin_unlock(&mem->pcp_counter_lock);
1344 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1351 spin_lock(&mem->pcp_counter_lock);
1352 for_each_online_cpu(cpu)
1353 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1354 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1355 spin_unlock(&mem->pcp_counter_lock);
1359 * 2 routines for checking "mem" is under move_account() or not.
1361 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1362 * for avoiding race in accounting. If true,
1363 * pc->mem_cgroup may be overwritten.
1365 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1366 * under hierarchy of moving cgroups. This is for
1367 * waiting at hith-memory prressure caused by "move".
1370 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1372 VM_BUG_ON(!rcu_read_lock_held());
1373 return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1376 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1378 struct mem_cgroup *from;
1379 struct mem_cgroup *to;
1382 * Unlike task_move routines, we access mc.to, mc.from not under
1383 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1385 spin_lock(&mc.lock);
1390 if (from == mem || to == mem
1391 || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1392 || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1395 spin_unlock(&mc.lock);
1399 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1401 if (mc.moving_task && current != mc.moving_task) {
1402 if (mem_cgroup_under_move(mem)) {
1404 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1405 /* moving charge context might have finished. */
1408 finish_wait(&mc.waitq, &wait);
1416 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1417 * @memcg: The memory cgroup that went over limit
1418 * @p: Task that is going to be killed
1420 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1423 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1425 struct cgroup *task_cgrp;
1426 struct cgroup *mem_cgrp;
1428 * Need a buffer in BSS, can't rely on allocations. The code relies
1429 * on the assumption that OOM is serialized for memory controller.
1430 * If this assumption is broken, revisit this code.
1432 static char memcg_name[PATH_MAX];
1441 mem_cgrp = memcg->css.cgroup;
1442 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1444 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1447 * Unfortunately, we are unable to convert to a useful name
1448 * But we'll still print out the usage information
1455 printk(KERN_INFO "Task in %s killed", memcg_name);
1458 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1466 * Continues from above, so we don't need an KERN_ level
1468 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1471 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1472 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1473 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1474 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1475 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1477 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1478 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1479 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1483 * This function returns the number of memcg under hierarchy tree. Returns
1484 * 1(self count) if no children.
1486 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1489 struct mem_cgroup *iter;
1491 for_each_mem_cgroup_tree(iter, mem)
1497 * Return the memory (and swap, if configured) limit for a memcg.
1499 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1504 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1505 limit += total_swap_pages << PAGE_SHIFT;
1507 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1509 * If memsw is finite and limits the amount of swap space available
1510 * to this memcg, return that limit.
1512 return min(limit, memsw);
1516 * Visit the first child (need not be the first child as per the ordering
1517 * of the cgroup list, since we track last_scanned_child) of @mem and use
1518 * that to reclaim free pages from.
1520 static struct mem_cgroup *
1521 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1523 struct mem_cgroup *ret = NULL;
1524 struct cgroup_subsys_state *css;
1527 if (!root_mem->use_hierarchy) {
1528 css_get(&root_mem->css);
1534 nextid = root_mem->last_scanned_child + 1;
1535 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1537 if (css && css_tryget(css))
1538 ret = container_of(css, struct mem_cgroup, css);
1541 /* Updates scanning parameter */
1543 /* this means start scan from ID:1 */
1544 root_mem->last_scanned_child = 0;
1546 root_mem->last_scanned_child = found;
1553 * test_mem_cgroup_node_reclaimable
1554 * @mem: the target memcg
1555 * @nid: the node ID to be checked.
1556 * @noswap : specify true here if the user wants flle only information.
1558 * This function returns whether the specified memcg contains any
1559 * reclaimable pages on a node. Returns true if there are any reclaimable
1560 * pages in the node.
1562 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *mem,
1563 int nid, bool noswap)
1565 if (mem_cgroup_node_nr_file_lru_pages(mem, nid))
1567 if (noswap || !total_swap_pages)
1569 if (mem_cgroup_node_nr_anon_lru_pages(mem, nid))
1574 #if MAX_NUMNODES > 1
1577 * Always updating the nodemask is not very good - even if we have an empty
1578 * list or the wrong list here, we can start from some node and traverse all
1579 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1582 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
1586 if (time_after(mem->next_scan_node_update, jiffies))
1589 mem->next_scan_node_update = jiffies + 10*HZ;
1590 /* make a nodemask where this memcg uses memory from */
1591 mem->scan_nodes = node_states[N_HIGH_MEMORY];
1593 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1595 if (!test_mem_cgroup_node_reclaimable(mem, nid, false))
1596 node_clear(nid, mem->scan_nodes);
1601 * Selecting a node where we start reclaim from. Because what we need is just
1602 * reducing usage counter, start from anywhere is O,K. Considering
1603 * memory reclaim from current node, there are pros. and cons.
1605 * Freeing memory from current node means freeing memory from a node which
1606 * we'll use or we've used. So, it may make LRU bad. And if several threads
1607 * hit limits, it will see a contention on a node. But freeing from remote
1608 * node means more costs for memory reclaim because of memory latency.
1610 * Now, we use round-robin. Better algorithm is welcomed.
1612 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1616 mem_cgroup_may_update_nodemask(mem);
1617 node = mem->last_scanned_node;
1619 node = next_node(node, mem->scan_nodes);
1620 if (node == MAX_NUMNODES)
1621 node = first_node(mem->scan_nodes);
1623 * We call this when we hit limit, not when pages are added to LRU.
1624 * No LRU may hold pages because all pages are UNEVICTABLE or
1625 * memcg is too small and all pages are not on LRU. In that case,
1626 * we use curret node.
1628 if (unlikely(node == MAX_NUMNODES))
1629 node = numa_node_id();
1631 mem->last_scanned_node = node;
1636 * Check all nodes whether it contains reclaimable pages or not.
1637 * For quick scan, we make use of scan_nodes. This will allow us to skip
1638 * unused nodes. But scan_nodes is lazily updated and may not cotain
1639 * enough new information. We need to do double check.
1641 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1646 * quick check...making use of scan_node.
1647 * We can skip unused nodes.
1649 if (!nodes_empty(mem->scan_nodes)) {
1650 for (nid = first_node(mem->scan_nodes);
1652 nid = next_node(nid, mem->scan_nodes)) {
1654 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1659 * Check rest of nodes.
1661 for_each_node_state(nid, N_HIGH_MEMORY) {
1662 if (node_isset(nid, mem->scan_nodes))
1664 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1671 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1676 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1678 return test_mem_cgroup_node_reclaimable(mem, 0, noswap);
1683 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1684 * we reclaimed from, so that we don't end up penalizing one child extensively
1685 * based on its position in the children list.
1687 * root_mem is the original ancestor that we've been reclaim from.
1689 * We give up and return to the caller when we visit root_mem twice.
1690 * (other groups can be removed while we're walking....)
1692 * If shrink==true, for avoiding to free too much, this returns immedieately.
1694 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1697 unsigned long reclaim_options,
1698 unsigned long *total_scanned)
1700 struct mem_cgroup *victim;
1703 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1704 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1705 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1706 unsigned long excess;
1707 unsigned long nr_scanned;
1709 excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1711 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1712 if (!check_soft && root_mem->memsw_is_minimum)
1716 victim = mem_cgroup_select_victim(root_mem);
1717 if (victim == root_mem) {
1720 * We are not draining per cpu cached charges during
1721 * soft limit reclaim because global reclaim doesn't
1722 * care about charges. It tries to free some memory and
1723 * charges will not give any.
1725 if (!check_soft && loop >= 1)
1726 drain_all_stock_async(root_mem);
1729 * If we have not been able to reclaim
1730 * anything, it might because there are
1731 * no reclaimable pages under this hierarchy
1733 if (!check_soft || !total) {
1734 css_put(&victim->css);
1738 * We want to do more targeted reclaim.
1739 * excess >> 2 is not to excessive so as to
1740 * reclaim too much, nor too less that we keep
1741 * coming back to reclaim from this cgroup
1743 if (total >= (excess >> 2) ||
1744 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1745 css_put(&victim->css);
1750 if (!mem_cgroup_reclaimable(victim, noswap)) {
1751 /* this cgroup's local usage == 0 */
1752 css_put(&victim->css);
1755 /* we use swappiness of local cgroup */
1757 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1758 noswap, get_swappiness(victim), zone,
1760 *total_scanned += nr_scanned;
1762 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1763 noswap, get_swappiness(victim));
1764 css_put(&victim->css);
1766 * At shrinking usage, we can't check we should stop here or
1767 * reclaim more. It's depends on callers. last_scanned_child
1768 * will work enough for keeping fairness under tree.
1774 if (!res_counter_soft_limit_excess(&root_mem->res))
1776 } else if (mem_cgroup_margin(root_mem))
1783 * Check OOM-Killer is already running under our hierarchy.
1784 * If someone is running, return false.
1786 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1788 int x, lock_count = 0;
1789 struct mem_cgroup *iter;
1791 for_each_mem_cgroup_tree(iter, mem) {
1792 x = atomic_inc_return(&iter->oom_lock);
1793 lock_count = max(x, lock_count);
1796 if (lock_count == 1)
1801 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1803 struct mem_cgroup *iter;
1806 * When a new child is created while the hierarchy is under oom,
1807 * mem_cgroup_oom_lock() may not be called. We have to use
1808 * atomic_add_unless() here.
1810 for_each_mem_cgroup_tree(iter, mem)
1811 atomic_add_unless(&iter->oom_lock, -1, 0);
1816 static DEFINE_MUTEX(memcg_oom_mutex);
1817 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1819 struct oom_wait_info {
1820 struct mem_cgroup *mem;
1824 static int memcg_oom_wake_function(wait_queue_t *wait,
1825 unsigned mode, int sync, void *arg)
1827 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1828 struct oom_wait_info *oom_wait_info;
1830 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1832 if (oom_wait_info->mem == wake_mem)
1834 /* if no hierarchy, no match */
1835 if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1838 * Both of oom_wait_info->mem and wake_mem are stable under us.
1839 * Then we can use css_is_ancestor without taking care of RCU.
1841 if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1842 !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1846 return autoremove_wake_function(wait, mode, sync, arg);
1849 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1851 /* for filtering, pass "mem" as argument. */
1852 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1855 static void memcg_oom_recover(struct mem_cgroup *mem)
1857 if (mem && atomic_read(&mem->oom_lock))
1858 memcg_wakeup_oom(mem);
1862 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1864 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1866 struct oom_wait_info owait;
1867 bool locked, need_to_kill;
1870 owait.wait.flags = 0;
1871 owait.wait.func = memcg_oom_wake_function;
1872 owait.wait.private = current;
1873 INIT_LIST_HEAD(&owait.wait.task_list);
1874 need_to_kill = true;
1875 /* At first, try to OOM lock hierarchy under mem.*/
1876 mutex_lock(&memcg_oom_mutex);
1877 locked = mem_cgroup_oom_lock(mem);
1879 * Even if signal_pending(), we can't quit charge() loop without
1880 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1881 * under OOM is always welcomed, use TASK_KILLABLE here.
1883 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1884 if (!locked || mem->oom_kill_disable)
1885 need_to_kill = false;
1887 mem_cgroup_oom_notify(mem);
1888 mutex_unlock(&memcg_oom_mutex);
1891 finish_wait(&memcg_oom_waitq, &owait.wait);
1892 mem_cgroup_out_of_memory(mem, mask);
1895 finish_wait(&memcg_oom_waitq, &owait.wait);
1897 mutex_lock(&memcg_oom_mutex);
1898 mem_cgroup_oom_unlock(mem);
1899 memcg_wakeup_oom(mem);
1900 mutex_unlock(&memcg_oom_mutex);
1902 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1904 /* Give chance to dying process */
1905 schedule_timeout(1);
1910 * Currently used to update mapped file statistics, but the routine can be
1911 * generalized to update other statistics as well.
1913 * Notes: Race condition
1915 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1916 * it tends to be costly. But considering some conditions, we doesn't need
1917 * to do so _always_.
1919 * Considering "charge", lock_page_cgroup() is not required because all
1920 * file-stat operations happen after a page is attached to radix-tree. There
1921 * are no race with "charge".
1923 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1924 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1925 * if there are race with "uncharge". Statistics itself is properly handled
1928 * Considering "move", this is an only case we see a race. To make the race
1929 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1930 * possibility of race condition. If there is, we take a lock.
1933 void mem_cgroup_update_page_stat(struct page *page,
1934 enum mem_cgroup_page_stat_item idx, int val)
1936 struct mem_cgroup *mem;
1937 struct page_cgroup *pc = lookup_page_cgroup(page);
1938 bool need_unlock = false;
1939 unsigned long uninitialized_var(flags);
1945 mem = pc->mem_cgroup;
1946 if (unlikely(!mem || !PageCgroupUsed(pc)))
1948 /* pc->mem_cgroup is unstable ? */
1949 if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1950 /* take a lock against to access pc->mem_cgroup */
1951 move_lock_page_cgroup(pc, &flags);
1953 mem = pc->mem_cgroup;
1954 if (!mem || !PageCgroupUsed(pc))
1959 case MEMCG_NR_FILE_MAPPED:
1961 SetPageCgroupFileMapped(pc);
1962 else if (!page_mapped(page))
1963 ClearPageCgroupFileMapped(pc);
1964 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1970 this_cpu_add(mem->stat->count[idx], val);
1973 if (unlikely(need_unlock))
1974 move_unlock_page_cgroup(pc, &flags);
1978 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1981 * size of first charge trial. "32" comes from vmscan.c's magic value.
1982 * TODO: maybe necessary to use big numbers in big irons.
1984 #define CHARGE_BATCH 32U
1985 struct memcg_stock_pcp {
1986 struct mem_cgroup *cached; /* this never be root cgroup */
1987 unsigned int nr_pages;
1988 struct work_struct work;
1989 unsigned long flags;
1990 #define FLUSHING_CACHED_CHARGE (0)
1992 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1993 static DEFINE_MUTEX(percpu_charge_mutex);
1996 * Try to consume stocked charge on this cpu. If success, one page is consumed
1997 * from local stock and true is returned. If the stock is 0 or charges from a
1998 * cgroup which is not current target, returns false. This stock will be
2001 static bool consume_stock(struct mem_cgroup *mem)
2003 struct memcg_stock_pcp *stock;
2006 stock = &get_cpu_var(memcg_stock);
2007 if (mem == stock->cached && stock->nr_pages)
2009 else /* need to call res_counter_charge */
2011 put_cpu_var(memcg_stock);
2016 * Returns stocks cached in percpu to res_counter and reset cached information.
2018 static void drain_stock(struct memcg_stock_pcp *stock)
2020 struct mem_cgroup *old = stock->cached;
2022 if (stock->nr_pages) {
2023 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2025 res_counter_uncharge(&old->res, bytes);
2026 if (do_swap_account)
2027 res_counter_uncharge(&old->memsw, bytes);
2028 stock->nr_pages = 0;
2030 stock->cached = NULL;
2034 * This must be called under preempt disabled or must be called by
2035 * a thread which is pinned to local cpu.
2037 static void drain_local_stock(struct work_struct *dummy)
2039 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2041 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2045 * Cache charges(val) which is from res_counter, to local per_cpu area.
2046 * This will be consumed by consume_stock() function, later.
2048 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
2050 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2052 if (stock->cached != mem) { /* reset if necessary */
2054 stock->cached = mem;
2056 stock->nr_pages += nr_pages;
2057 put_cpu_var(memcg_stock);
2061 * Tries to drain stocked charges in other cpus. This function is asynchronous
2062 * and just put a work per cpu for draining localy on each cpu. Caller can
2063 * expects some charges will be back to res_counter later but cannot wait for
2066 static void drain_all_stock_async(struct mem_cgroup *root_mem)
2070 * If someone calls draining, avoid adding more kworker runs.
2072 if (!mutex_trylock(&percpu_charge_mutex))
2074 /* Notify other cpus that system-wide "drain" is running */
2077 * Get a hint for avoiding draining charges on the current cpu,
2078 * which must be exhausted by our charging. It is not required that
2079 * this be a precise check, so we use raw_smp_processor_id() instead of
2080 * getcpu()/putcpu().
2082 curcpu = raw_smp_processor_id();
2083 for_each_online_cpu(cpu) {
2084 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2085 struct mem_cgroup *mem;
2090 mem = stock->cached;
2093 if (mem != root_mem) {
2094 if (!root_mem->use_hierarchy)
2096 /* check whether "mem" is under tree of "root_mem" */
2097 if (!css_is_ancestor(&mem->css, &root_mem->css))
2100 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2101 schedule_work_on(cpu, &stock->work);
2104 mutex_unlock(&percpu_charge_mutex);
2105 /* We don't wait for flush_work */
2108 /* This is a synchronous drain interface. */
2109 static void drain_all_stock_sync(void)
2111 /* called when force_empty is called */
2112 mutex_lock(&percpu_charge_mutex);
2113 schedule_on_each_cpu(drain_local_stock);
2114 mutex_unlock(&percpu_charge_mutex);
2118 * This function drains percpu counter value from DEAD cpu and
2119 * move it to local cpu. Note that this function can be preempted.
2121 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2125 spin_lock(&mem->pcp_counter_lock);
2126 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2127 long x = per_cpu(mem->stat->count[i], cpu);
2129 per_cpu(mem->stat->count[i], cpu) = 0;
2130 mem->nocpu_base.count[i] += x;
2132 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2133 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2135 per_cpu(mem->stat->events[i], cpu) = 0;
2136 mem->nocpu_base.events[i] += x;
2138 /* need to clear ON_MOVE value, works as a kind of lock. */
2139 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2140 spin_unlock(&mem->pcp_counter_lock);
2143 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2145 int idx = MEM_CGROUP_ON_MOVE;
2147 spin_lock(&mem->pcp_counter_lock);
2148 per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2149 spin_unlock(&mem->pcp_counter_lock);
2152 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2153 unsigned long action,
2156 int cpu = (unsigned long)hcpu;
2157 struct memcg_stock_pcp *stock;
2158 struct mem_cgroup *iter;
2160 if ((action == CPU_ONLINE)) {
2161 for_each_mem_cgroup_all(iter)
2162 synchronize_mem_cgroup_on_move(iter, cpu);
2166 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2169 for_each_mem_cgroup_all(iter)
2170 mem_cgroup_drain_pcp_counter(iter, cpu);
2172 stock = &per_cpu(memcg_stock, cpu);
2178 /* See __mem_cgroup_try_charge() for details */
2180 CHARGE_OK, /* success */
2181 CHARGE_RETRY, /* need to retry but retry is not bad */
2182 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2183 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2184 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2187 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2188 unsigned int nr_pages, bool oom_check)
2190 unsigned long csize = nr_pages * PAGE_SIZE;
2191 struct mem_cgroup *mem_over_limit;
2192 struct res_counter *fail_res;
2193 unsigned long flags = 0;
2196 ret = res_counter_charge(&mem->res, csize, &fail_res);
2199 if (!do_swap_account)
2201 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2205 res_counter_uncharge(&mem->res, csize);
2206 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2207 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2209 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2211 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2212 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2214 * Never reclaim on behalf of optional batching, retry with a
2215 * single page instead.
2217 if (nr_pages == CHARGE_BATCH)
2218 return CHARGE_RETRY;
2220 if (!(gfp_mask & __GFP_WAIT))
2221 return CHARGE_WOULDBLOCK;
2223 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2224 gfp_mask, flags, NULL);
2225 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2226 return CHARGE_RETRY;
2228 * Even though the limit is exceeded at this point, reclaim
2229 * may have been able to free some pages. Retry the charge
2230 * before killing the task.
2232 * Only for regular pages, though: huge pages are rather
2233 * unlikely to succeed so close to the limit, and we fall back
2234 * to regular pages anyway in case of failure.
2236 if (nr_pages == 1 && ret)
2237 return CHARGE_RETRY;
2240 * At task move, charge accounts can be doubly counted. So, it's
2241 * better to wait until the end of task_move if something is going on.
2243 if (mem_cgroup_wait_acct_move(mem_over_limit))
2244 return CHARGE_RETRY;
2246 /* If we don't need to call oom-killer at el, return immediately */
2248 return CHARGE_NOMEM;
2250 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2251 return CHARGE_OOM_DIE;
2253 return CHARGE_RETRY;
2257 * Unlike exported interface, "oom" parameter is added. if oom==true,
2258 * oom-killer can be invoked.
2260 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2262 unsigned int nr_pages,
2263 struct mem_cgroup **memcg,
2266 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2267 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2268 struct mem_cgroup *mem = NULL;
2272 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2273 * in system level. So, allow to go ahead dying process in addition to
2276 if (unlikely(test_thread_flag(TIF_MEMDIE)
2277 || fatal_signal_pending(current)))
2281 * We always charge the cgroup the mm_struct belongs to.
2282 * The mm_struct's mem_cgroup changes on task migration if the
2283 * thread group leader migrates. It's possible that mm is not
2284 * set, if so charge the init_mm (happens for pagecache usage).
2289 if (*memcg) { /* css should be a valid one */
2291 VM_BUG_ON(css_is_removed(&mem->css));
2292 if (mem_cgroup_is_root(mem))
2294 if (nr_pages == 1 && consume_stock(mem))
2298 struct task_struct *p;
2301 p = rcu_dereference(mm->owner);
2303 * Because we don't have task_lock(), "p" can exit.
2304 * In that case, "mem" can point to root or p can be NULL with
2305 * race with swapoff. Then, we have small risk of mis-accouning.
2306 * But such kind of mis-account by race always happens because
2307 * we don't have cgroup_mutex(). It's overkill and we allo that
2309 * (*) swapoff at el will charge against mm-struct not against
2310 * task-struct. So, mm->owner can be NULL.
2312 mem = mem_cgroup_from_task(p);
2313 if (!mem || mem_cgroup_is_root(mem)) {
2317 if (nr_pages == 1 && consume_stock(mem)) {
2319 * It seems dagerous to access memcg without css_get().
2320 * But considering how consume_stok works, it's not
2321 * necessary. If consume_stock success, some charges
2322 * from this memcg are cached on this cpu. So, we
2323 * don't need to call css_get()/css_tryget() before
2324 * calling consume_stock().
2329 /* after here, we may be blocked. we need to get refcnt */
2330 if (!css_tryget(&mem->css)) {
2340 /* If killed, bypass charge */
2341 if (fatal_signal_pending(current)) {
2347 if (oom && !nr_oom_retries) {
2349 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2352 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2356 case CHARGE_RETRY: /* not in OOM situation but retry */
2361 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2364 case CHARGE_NOMEM: /* OOM routine works */
2369 /* If oom, we never return -ENOMEM */
2372 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2376 } while (ret != CHARGE_OK);
2378 if (batch > nr_pages)
2379 refill_stock(mem, batch - nr_pages);
2393 * Somemtimes we have to undo a charge we got by try_charge().
2394 * This function is for that and do uncharge, put css's refcnt.
2395 * gotten by try_charge().
2397 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2398 unsigned int nr_pages)
2400 if (!mem_cgroup_is_root(mem)) {
2401 unsigned long bytes = nr_pages * PAGE_SIZE;
2403 res_counter_uncharge(&mem->res, bytes);
2404 if (do_swap_account)
2405 res_counter_uncharge(&mem->memsw, bytes);
2410 * A helper function to get mem_cgroup from ID. must be called under
2411 * rcu_read_lock(). The caller must check css_is_removed() or some if
2412 * it's concern. (dropping refcnt from swap can be called against removed
2415 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2417 struct cgroup_subsys_state *css;
2419 /* ID 0 is unused ID */
2422 css = css_lookup(&mem_cgroup_subsys, id);
2425 return container_of(css, struct mem_cgroup, css);
2428 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2430 struct mem_cgroup *mem = NULL;
2431 struct page_cgroup *pc;
2435 VM_BUG_ON(!PageLocked(page));
2437 pc = lookup_page_cgroup(page);
2438 lock_page_cgroup(pc);
2439 if (PageCgroupUsed(pc)) {
2440 mem = pc->mem_cgroup;
2441 if (mem && !css_tryget(&mem->css))
2443 } else if (PageSwapCache(page)) {
2444 ent.val = page_private(page);
2445 id = lookup_swap_cgroup(ent);
2447 mem = mem_cgroup_lookup(id);
2448 if (mem && !css_tryget(&mem->css))
2452 unlock_page_cgroup(pc);
2456 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2458 unsigned int nr_pages,
2459 struct page_cgroup *pc,
2460 enum charge_type ctype)
2462 lock_page_cgroup(pc);
2463 if (unlikely(PageCgroupUsed(pc))) {
2464 unlock_page_cgroup(pc);
2465 __mem_cgroup_cancel_charge(mem, nr_pages);
2469 * we don't need page_cgroup_lock about tail pages, becase they are not
2470 * accessed by any other context at this point.
2472 pc->mem_cgroup = mem;
2474 * We access a page_cgroup asynchronously without lock_page_cgroup().
2475 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2476 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2477 * before USED bit, we need memory barrier here.
2478 * See mem_cgroup_add_lru_list(), etc.
2482 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2483 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2484 SetPageCgroupCache(pc);
2485 SetPageCgroupUsed(pc);
2487 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2488 ClearPageCgroupCache(pc);
2489 SetPageCgroupUsed(pc);
2495 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2496 unlock_page_cgroup(pc);
2498 * "charge_statistics" updated event counter. Then, check it.
2499 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2500 * if they exceeds softlimit.
2502 memcg_check_events(mem, page);
2505 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2507 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2508 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2510 * Because tail pages are not marked as "used", set it. We're under
2511 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2513 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2515 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2516 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2517 unsigned long flags;
2519 if (mem_cgroup_disabled())
2522 * We have no races with charge/uncharge but will have races with
2523 * page state accounting.
2525 move_lock_page_cgroup(head_pc, &flags);
2527 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2528 smp_wmb(); /* see __commit_charge() */
2529 if (PageCgroupAcctLRU(head_pc)) {
2531 struct mem_cgroup_per_zone *mz;
2534 * LRU flags cannot be copied because we need to add tail
2535 *.page to LRU by generic call and our hook will be called.
2536 * We hold lru_lock, then, reduce counter directly.
2538 lru = page_lru(head);
2539 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2540 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2542 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2543 move_unlock_page_cgroup(head_pc, &flags);
2548 * mem_cgroup_move_account - move account of the page
2550 * @nr_pages: number of regular pages (>1 for huge pages)
2551 * @pc: page_cgroup of the page.
2552 * @from: mem_cgroup which the page is moved from.
2553 * @to: mem_cgroup which the page is moved to. @from != @to.
2554 * @uncharge: whether we should call uncharge and css_put against @from.
2556 * The caller must confirm following.
2557 * - page is not on LRU (isolate_page() is useful.)
2558 * - compound_lock is held when nr_pages > 1
2560 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2561 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2562 * true, this function does "uncharge" from old cgroup, but it doesn't if
2563 * @uncharge is false, so a caller should do "uncharge".
2565 static int mem_cgroup_move_account(struct page *page,
2566 unsigned int nr_pages,
2567 struct page_cgroup *pc,
2568 struct mem_cgroup *from,
2569 struct mem_cgroup *to,
2572 unsigned long flags;
2575 VM_BUG_ON(from == to);
2576 VM_BUG_ON(PageLRU(page));
2578 * The page is isolated from LRU. So, collapse function
2579 * will not handle this page. But page splitting can happen.
2580 * Do this check under compound_page_lock(). The caller should
2584 if (nr_pages > 1 && !PageTransHuge(page))
2587 lock_page_cgroup(pc);
2590 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2593 move_lock_page_cgroup(pc, &flags);
2595 if (PageCgroupFileMapped(pc)) {
2596 /* Update mapped_file data for mem_cgroup */
2598 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2599 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2602 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2604 /* This is not "cancel", but cancel_charge does all we need. */
2605 __mem_cgroup_cancel_charge(from, nr_pages);
2607 /* caller should have done css_get */
2608 pc->mem_cgroup = to;
2609 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2611 * We charges against "to" which may not have any tasks. Then, "to"
2612 * can be under rmdir(). But in current implementation, caller of
2613 * this function is just force_empty() and move charge, so it's
2614 * guaranteed that "to" is never removed. So, we don't check rmdir
2617 move_unlock_page_cgroup(pc, &flags);
2620 unlock_page_cgroup(pc);
2624 memcg_check_events(to, page);
2625 memcg_check_events(from, page);
2631 * move charges to its parent.
2634 static int mem_cgroup_move_parent(struct page *page,
2635 struct page_cgroup *pc,
2636 struct mem_cgroup *child,
2639 struct cgroup *cg = child->css.cgroup;
2640 struct cgroup *pcg = cg->parent;
2641 struct mem_cgroup *parent;
2642 unsigned int nr_pages;
2643 unsigned long uninitialized_var(flags);
2651 if (!get_page_unless_zero(page))
2653 if (isolate_lru_page(page))
2656 nr_pages = hpage_nr_pages(page);
2658 parent = mem_cgroup_from_cont(pcg);
2659 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2664 flags = compound_lock_irqsave(page);
2666 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2668 __mem_cgroup_cancel_charge(parent, nr_pages);
2671 compound_unlock_irqrestore(page, flags);
2673 putback_lru_page(page);
2681 * Charge the memory controller for page usage.
2683 * 0 if the charge was successful
2684 * < 0 if the cgroup is over its limit
2686 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2687 gfp_t gfp_mask, enum charge_type ctype)
2689 struct mem_cgroup *mem = NULL;
2690 unsigned int nr_pages = 1;
2691 struct page_cgroup *pc;
2695 if (PageTransHuge(page)) {
2696 nr_pages <<= compound_order(page);
2697 VM_BUG_ON(!PageTransHuge(page));
2699 * Never OOM-kill a process for a huge page. The
2700 * fault handler will fall back to regular pages.
2705 pc = lookup_page_cgroup(page);
2706 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2708 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2712 __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2716 int mem_cgroup_newpage_charge(struct page *page,
2717 struct mm_struct *mm, gfp_t gfp_mask)
2719 if (mem_cgroup_disabled())
2722 * If already mapped, we don't have to account.
2723 * If page cache, page->mapping has address_space.
2724 * But page->mapping may have out-of-use anon_vma pointer,
2725 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2728 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2732 return mem_cgroup_charge_common(page, mm, gfp_mask,
2733 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2737 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2738 enum charge_type ctype);
2741 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2742 enum charge_type ctype)
2744 struct page_cgroup *pc = lookup_page_cgroup(page);
2746 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2747 * is already on LRU. It means the page may on some other page_cgroup's
2748 * LRU. Take care of it.
2750 mem_cgroup_lru_del_before_commit(page);
2751 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2752 mem_cgroup_lru_add_after_commit(page);
2756 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2759 struct mem_cgroup *mem = NULL;
2762 if (mem_cgroup_disabled())
2764 if (PageCompound(page))
2767 * Corner case handling. This is called from add_to_page_cache()
2768 * in usual. But some FS (shmem) precharges this page before calling it
2769 * and call add_to_page_cache() with GFP_NOWAIT.
2771 * For GFP_NOWAIT case, the page may be pre-charged before calling
2772 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2773 * charge twice. (It works but has to pay a bit larger cost.)
2774 * And when the page is SwapCache, it should take swap information
2775 * into account. This is under lock_page() now.
2777 if (!(gfp_mask & __GFP_WAIT)) {
2778 struct page_cgroup *pc;
2780 pc = lookup_page_cgroup(page);
2783 lock_page_cgroup(pc);
2784 if (PageCgroupUsed(pc)) {
2785 unlock_page_cgroup(pc);
2788 unlock_page_cgroup(pc);
2794 if (page_is_file_cache(page)) {
2795 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2800 * FUSE reuses pages without going through the final
2801 * put that would remove them from the LRU list, make
2802 * sure that they get relinked properly.
2804 __mem_cgroup_commit_charge_lrucare(page, mem,
2805 MEM_CGROUP_CHARGE_TYPE_CACHE);
2809 if (PageSwapCache(page)) {
2810 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2812 __mem_cgroup_commit_charge_swapin(page, mem,
2813 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2815 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2816 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2822 * While swap-in, try_charge -> commit or cancel, the page is locked.
2823 * And when try_charge() successfully returns, one refcnt to memcg without
2824 * struct page_cgroup is acquired. This refcnt will be consumed by
2825 * "commit()" or removed by "cancel()"
2827 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2829 gfp_t mask, struct mem_cgroup **ptr)
2831 struct mem_cgroup *mem;
2836 if (mem_cgroup_disabled())
2839 if (!do_swap_account)
2842 * A racing thread's fault, or swapoff, may have already updated
2843 * the pte, and even removed page from swap cache: in those cases
2844 * do_swap_page()'s pte_same() test will fail; but there's also a
2845 * KSM case which does need to charge the page.
2847 if (!PageSwapCache(page))
2849 mem = try_get_mem_cgroup_from_page(page);
2853 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2859 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2863 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2864 enum charge_type ctype)
2866 if (mem_cgroup_disabled())
2870 cgroup_exclude_rmdir(&ptr->css);
2872 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2874 * Now swap is on-memory. This means this page may be
2875 * counted both as mem and swap....double count.
2876 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2877 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2878 * may call delete_from_swap_cache() before reach here.
2880 if (do_swap_account && PageSwapCache(page)) {
2881 swp_entry_t ent = {.val = page_private(page)};
2883 struct mem_cgroup *memcg;
2885 id = swap_cgroup_record(ent, 0);
2887 memcg = mem_cgroup_lookup(id);
2890 * This recorded memcg can be obsolete one. So, avoid
2891 * calling css_tryget
2893 if (!mem_cgroup_is_root(memcg))
2894 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2895 mem_cgroup_swap_statistics(memcg, false);
2896 mem_cgroup_put(memcg);
2901 * At swapin, we may charge account against cgroup which has no tasks.
2902 * So, rmdir()->pre_destroy() can be called while we do this charge.
2903 * In that case, we need to call pre_destroy() again. check it here.
2905 cgroup_release_and_wakeup_rmdir(&ptr->css);
2908 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2910 __mem_cgroup_commit_charge_swapin(page, ptr,
2911 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2914 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2916 if (mem_cgroup_disabled())
2920 __mem_cgroup_cancel_charge(mem, 1);
2923 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
2924 unsigned int nr_pages,
2925 const enum charge_type ctype)
2927 struct memcg_batch_info *batch = NULL;
2928 bool uncharge_memsw = true;
2930 /* If swapout, usage of swap doesn't decrease */
2931 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2932 uncharge_memsw = false;
2934 batch = ¤t->memcg_batch;
2936 * In usual, we do css_get() when we remember memcg pointer.
2937 * But in this case, we keep res->usage until end of a series of
2938 * uncharges. Then, it's ok to ignore memcg's refcnt.
2943 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2944 * In those cases, all pages freed continuously can be expected to be in
2945 * the same cgroup and we have chance to coalesce uncharges.
2946 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2947 * because we want to do uncharge as soon as possible.
2950 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2951 goto direct_uncharge;
2954 goto direct_uncharge;
2957 * In typical case, batch->memcg == mem. This means we can
2958 * merge a series of uncharges to an uncharge of res_counter.
2959 * If not, we uncharge res_counter ony by one.
2961 if (batch->memcg != mem)
2962 goto direct_uncharge;
2963 /* remember freed charge and uncharge it later */
2966 batch->memsw_nr_pages++;
2969 res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
2971 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
2972 if (unlikely(batch->memcg != mem))
2973 memcg_oom_recover(mem);
2978 * uncharge if !page_mapped(page)
2980 static struct mem_cgroup *
2981 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2983 struct mem_cgroup *mem = NULL;
2984 unsigned int nr_pages = 1;
2985 struct page_cgroup *pc;
2987 if (mem_cgroup_disabled())
2990 if (PageSwapCache(page))
2993 if (PageTransHuge(page)) {
2994 nr_pages <<= compound_order(page);
2995 VM_BUG_ON(!PageTransHuge(page));
2998 * Check if our page_cgroup is valid
3000 pc = lookup_page_cgroup(page);
3001 if (unlikely(!pc || !PageCgroupUsed(pc)))
3004 lock_page_cgroup(pc);
3006 mem = pc->mem_cgroup;
3008 if (!PageCgroupUsed(pc))
3012 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3013 case MEM_CGROUP_CHARGE_TYPE_DROP:
3014 /* See mem_cgroup_prepare_migration() */
3015 if (page_mapped(page) || PageCgroupMigration(pc))
3018 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3019 if (!PageAnon(page)) { /* Shared memory */
3020 if (page->mapping && !page_is_file_cache(page))
3022 } else if (page_mapped(page)) /* Anon */
3029 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
3031 ClearPageCgroupUsed(pc);
3033 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3034 * freed from LRU. This is safe because uncharged page is expected not
3035 * to be reused (freed soon). Exception is SwapCache, it's handled by
3036 * special functions.
3039 unlock_page_cgroup(pc);
3041 * even after unlock, we have mem->res.usage here and this memcg
3042 * will never be freed.
3044 memcg_check_events(mem, page);
3045 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3046 mem_cgroup_swap_statistics(mem, true);
3047 mem_cgroup_get(mem);
3049 if (!mem_cgroup_is_root(mem))
3050 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
3055 unlock_page_cgroup(pc);
3059 void mem_cgroup_uncharge_page(struct page *page)
3062 if (page_mapped(page))
3064 if (page->mapping && !PageAnon(page))
3066 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3069 void mem_cgroup_uncharge_cache_page(struct page *page)
3071 VM_BUG_ON(page_mapped(page));
3072 VM_BUG_ON(page->mapping);
3073 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3077 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3078 * In that cases, pages are freed continuously and we can expect pages
3079 * are in the same memcg. All these calls itself limits the number of
3080 * pages freed at once, then uncharge_start/end() is called properly.
3081 * This may be called prural(2) times in a context,
3084 void mem_cgroup_uncharge_start(void)
3086 current->memcg_batch.do_batch++;
3087 /* We can do nest. */
3088 if (current->memcg_batch.do_batch == 1) {
3089 current->memcg_batch.memcg = NULL;
3090 current->memcg_batch.nr_pages = 0;
3091 current->memcg_batch.memsw_nr_pages = 0;
3095 void mem_cgroup_uncharge_end(void)
3097 struct memcg_batch_info *batch = ¤t->memcg_batch;
3099 if (!batch->do_batch)
3103 if (batch->do_batch) /* If stacked, do nothing. */
3109 * This "batch->memcg" is valid without any css_get/put etc...
3110 * bacause we hide charges behind us.
3112 if (batch->nr_pages)
3113 res_counter_uncharge(&batch->memcg->res,
3114 batch->nr_pages * PAGE_SIZE);
3115 if (batch->memsw_nr_pages)
3116 res_counter_uncharge(&batch->memcg->memsw,
3117 batch->memsw_nr_pages * PAGE_SIZE);
3118 memcg_oom_recover(batch->memcg);
3119 /* forget this pointer (for sanity check) */
3120 batch->memcg = NULL;
3125 * called after __delete_from_swap_cache() and drop "page" account.
3126 * memcg information is recorded to swap_cgroup of "ent"
3129 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3131 struct mem_cgroup *memcg;
3132 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3134 if (!swapout) /* this was a swap cache but the swap is unused ! */
3135 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3137 memcg = __mem_cgroup_uncharge_common(page, ctype);
3140 * record memcg information, if swapout && memcg != NULL,
3141 * mem_cgroup_get() was called in uncharge().
3143 if (do_swap_account && swapout && memcg)
3144 swap_cgroup_record(ent, css_id(&memcg->css));
3148 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3150 * called from swap_entry_free(). remove record in swap_cgroup and
3151 * uncharge "memsw" account.
3153 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3155 struct mem_cgroup *memcg;
3158 if (!do_swap_account)
3161 id = swap_cgroup_record(ent, 0);
3163 memcg = mem_cgroup_lookup(id);
3166 * We uncharge this because swap is freed.
3167 * This memcg can be obsolete one. We avoid calling css_tryget
3169 if (!mem_cgroup_is_root(memcg))
3170 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3171 mem_cgroup_swap_statistics(memcg, false);
3172 mem_cgroup_put(memcg);
3178 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3179 * @entry: swap entry to be moved
3180 * @from: mem_cgroup which the entry is moved from
3181 * @to: mem_cgroup which the entry is moved to
3182 * @need_fixup: whether we should fixup res_counters and refcounts.
3184 * It succeeds only when the swap_cgroup's record for this entry is the same
3185 * as the mem_cgroup's id of @from.
3187 * Returns 0 on success, -EINVAL on failure.
3189 * The caller must have charged to @to, IOW, called res_counter_charge() about
3190 * both res and memsw, and called css_get().
3192 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3193 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3195 unsigned short old_id, new_id;
3197 old_id = css_id(&from->css);
3198 new_id = css_id(&to->css);
3200 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3201 mem_cgroup_swap_statistics(from, false);
3202 mem_cgroup_swap_statistics(to, true);
3204 * This function is only called from task migration context now.
3205 * It postpones res_counter and refcount handling till the end
3206 * of task migration(mem_cgroup_clear_mc()) for performance
3207 * improvement. But we cannot postpone mem_cgroup_get(to)
3208 * because if the process that has been moved to @to does
3209 * swap-in, the refcount of @to might be decreased to 0.
3213 if (!mem_cgroup_is_root(from))
3214 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3215 mem_cgroup_put(from);
3217 * we charged both to->res and to->memsw, so we should
3220 if (!mem_cgroup_is_root(to))
3221 res_counter_uncharge(&to->res, PAGE_SIZE);
3228 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3229 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3236 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3239 int mem_cgroup_prepare_migration(struct page *page,
3240 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3242 struct mem_cgroup *mem = NULL;
3243 struct page_cgroup *pc;
3244 enum charge_type ctype;
3249 VM_BUG_ON(PageTransHuge(page));
3250 if (mem_cgroup_disabled())
3253 pc = lookup_page_cgroup(page);
3254 lock_page_cgroup(pc);
3255 if (PageCgroupUsed(pc)) {
3256 mem = pc->mem_cgroup;
3259 * At migrating an anonymous page, its mapcount goes down
3260 * to 0 and uncharge() will be called. But, even if it's fully
3261 * unmapped, migration may fail and this page has to be
3262 * charged again. We set MIGRATION flag here and delay uncharge
3263 * until end_migration() is called
3265 * Corner Case Thinking
3267 * When the old page was mapped as Anon and it's unmap-and-freed
3268 * while migration was ongoing.
3269 * If unmap finds the old page, uncharge() of it will be delayed
3270 * until end_migration(). If unmap finds a new page, it's
3271 * uncharged when it make mapcount to be 1->0. If unmap code
3272 * finds swap_migration_entry, the new page will not be mapped
3273 * and end_migration() will find it(mapcount==0).
3276 * When the old page was mapped but migraion fails, the kernel
3277 * remaps it. A charge for it is kept by MIGRATION flag even
3278 * if mapcount goes down to 0. We can do remap successfully
3279 * without charging it again.
3282 * The "old" page is under lock_page() until the end of
3283 * migration, so, the old page itself will not be swapped-out.
3284 * If the new page is swapped out before end_migraton, our
3285 * hook to usual swap-out path will catch the event.
3288 SetPageCgroupMigration(pc);
3290 unlock_page_cgroup(pc);
3292 * If the page is not charged at this point,
3299 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3300 css_put(&mem->css);/* drop extra refcnt */
3301 if (ret || *ptr == NULL) {
3302 if (PageAnon(page)) {
3303 lock_page_cgroup(pc);
3304 ClearPageCgroupMigration(pc);
3305 unlock_page_cgroup(pc);
3307 * The old page may be fully unmapped while we kept it.
3309 mem_cgroup_uncharge_page(page);
3314 * We charge new page before it's used/mapped. So, even if unlock_page()
3315 * is called before end_migration, we can catch all events on this new
3316 * page. In the case new page is migrated but not remapped, new page's
3317 * mapcount will be finally 0 and we call uncharge in end_migration().
3319 pc = lookup_page_cgroup(newpage);
3321 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3322 else if (page_is_file_cache(page))
3323 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3325 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3326 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3330 /* remove redundant charge if migration failed*/
3331 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3332 struct page *oldpage, struct page *newpage, bool migration_ok)
3334 struct page *used, *unused;
3335 struct page_cgroup *pc;
3339 /* blocks rmdir() */
3340 cgroup_exclude_rmdir(&mem->css);
3341 if (!migration_ok) {
3349 * We disallowed uncharge of pages under migration because mapcount
3350 * of the page goes down to zero, temporarly.
3351 * Clear the flag and check the page should be charged.
3353 pc = lookup_page_cgroup(oldpage);
3354 lock_page_cgroup(pc);
3355 ClearPageCgroupMigration(pc);
3356 unlock_page_cgroup(pc);
3358 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3361 * If a page is a file cache, radix-tree replacement is very atomic
3362 * and we can skip this check. When it was an Anon page, its mapcount
3363 * goes down to 0. But because we added MIGRATION flage, it's not
3364 * uncharged yet. There are several case but page->mapcount check
3365 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3366 * check. (see prepare_charge() also)
3369 mem_cgroup_uncharge_page(used);
3371 * At migration, we may charge account against cgroup which has no
3373 * So, rmdir()->pre_destroy() can be called while we do this charge.
3374 * In that case, we need to call pre_destroy() again. check it here.
3376 cgroup_release_and_wakeup_rmdir(&mem->css);
3380 * A call to try to shrink memory usage on charge failure at shmem's swapin.
3381 * Calling hierarchical_reclaim is not enough because we should update
3382 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3383 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3384 * not from the memcg which this page would be charged to.
3385 * try_charge_swapin does all of these works properly.
3387 int mem_cgroup_shmem_charge_fallback(struct page *page,
3388 struct mm_struct *mm,
3391 struct mem_cgroup *mem;
3394 if (mem_cgroup_disabled())
3397 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3399 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3404 #ifdef CONFIG_DEBUG_VM
3405 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3407 struct page_cgroup *pc;
3409 pc = lookup_page_cgroup(page);
3410 if (likely(pc) && PageCgroupUsed(pc))
3415 bool mem_cgroup_bad_page_check(struct page *page)
3417 if (mem_cgroup_disabled())
3420 return lookup_page_cgroup_used(page) != NULL;
3423 void mem_cgroup_print_bad_page(struct page *page)
3425 struct page_cgroup *pc;
3427 pc = lookup_page_cgroup_used(page);
3432 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3433 pc, pc->flags, pc->mem_cgroup);
3435 path = kmalloc(PATH_MAX, GFP_KERNEL);
3438 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3443 printk(KERN_CONT "(%s)\n",
3444 (ret < 0) ? "cannot get the path" : path);
3450 static DEFINE_MUTEX(set_limit_mutex);
3452 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3453 unsigned long long val)
3456 u64 memswlimit, memlimit;
3458 int children = mem_cgroup_count_children(memcg);
3459 u64 curusage, oldusage;
3463 * For keeping hierarchical_reclaim simple, how long we should retry
3464 * is depends on callers. We set our retry-count to be function
3465 * of # of children which we should visit in this loop.
3467 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3469 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3472 while (retry_count) {
3473 if (signal_pending(current)) {
3478 * Rather than hide all in some function, I do this in
3479 * open coded manner. You see what this really does.
3480 * We have to guarantee mem->res.limit < mem->memsw.limit.
3482 mutex_lock(&set_limit_mutex);
3483 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3484 if (memswlimit < val) {
3486 mutex_unlock(&set_limit_mutex);
3490 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3494 ret = res_counter_set_limit(&memcg->res, val);
3496 if (memswlimit == val)
3497 memcg->memsw_is_minimum = true;
3499 memcg->memsw_is_minimum = false;
3501 mutex_unlock(&set_limit_mutex);
3506 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3507 MEM_CGROUP_RECLAIM_SHRINK,
3509 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3510 /* Usage is reduced ? */
3511 if (curusage >= oldusage)
3514 oldusage = curusage;
3516 if (!ret && enlarge)
3517 memcg_oom_recover(memcg);
3522 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3523 unsigned long long val)
3526 u64 memlimit, memswlimit, oldusage, curusage;
3527 int children = mem_cgroup_count_children(memcg);
3531 /* see mem_cgroup_resize_res_limit */
3532 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3533 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3534 while (retry_count) {
3535 if (signal_pending(current)) {
3540 * Rather than hide all in some function, I do this in
3541 * open coded manner. You see what this really does.
3542 * We have to guarantee mem->res.limit < mem->memsw.limit.
3544 mutex_lock(&set_limit_mutex);
3545 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3546 if (memlimit > val) {
3548 mutex_unlock(&set_limit_mutex);
3551 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3552 if (memswlimit < val)
3554 ret = res_counter_set_limit(&memcg->memsw, val);
3556 if (memlimit == val)
3557 memcg->memsw_is_minimum = true;
3559 memcg->memsw_is_minimum = false;
3561 mutex_unlock(&set_limit_mutex);
3566 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3567 MEM_CGROUP_RECLAIM_NOSWAP |
3568 MEM_CGROUP_RECLAIM_SHRINK,
3570 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3571 /* Usage is reduced ? */
3572 if (curusage >= oldusage)
3575 oldusage = curusage;
3577 if (!ret && enlarge)
3578 memcg_oom_recover(memcg);
3582 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3584 unsigned long *total_scanned)
3586 unsigned long nr_reclaimed = 0;
3587 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3588 unsigned long reclaimed;
3590 struct mem_cgroup_tree_per_zone *mctz;
3591 unsigned long long excess;
3592 unsigned long nr_scanned;
3597 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3599 * This loop can run a while, specially if mem_cgroup's continuously
3600 * keep exceeding their soft limit and putting the system under
3607 mz = mem_cgroup_largest_soft_limit_node(mctz);
3612 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3614 MEM_CGROUP_RECLAIM_SOFT,
3616 nr_reclaimed += reclaimed;
3617 *total_scanned += nr_scanned;
3618 spin_lock(&mctz->lock);
3621 * If we failed to reclaim anything from this memory cgroup
3622 * it is time to move on to the next cgroup
3628 * Loop until we find yet another one.
3630 * By the time we get the soft_limit lock
3631 * again, someone might have aded the
3632 * group back on the RB tree. Iterate to
3633 * make sure we get a different mem.
3634 * mem_cgroup_largest_soft_limit_node returns
3635 * NULL if no other cgroup is present on
3639 __mem_cgroup_largest_soft_limit_node(mctz);
3641 css_put(&next_mz->mem->css);
3642 else /* next_mz == NULL or other memcg */
3646 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3647 excess = res_counter_soft_limit_excess(&mz->mem->res);
3649 * One school of thought says that we should not add
3650 * back the node to the tree if reclaim returns 0.
3651 * But our reclaim could return 0, simply because due
3652 * to priority we are exposing a smaller subset of
3653 * memory to reclaim from. Consider this as a longer
3656 /* If excess == 0, no tree ops */
3657 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3658 spin_unlock(&mctz->lock);
3659 css_put(&mz->mem->css);
3662 * Could not reclaim anything and there are no more
3663 * mem cgroups to try or we seem to be looping without
3664 * reclaiming anything.
3666 if (!nr_reclaimed &&
3668 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3670 } while (!nr_reclaimed);
3672 css_put(&next_mz->mem->css);
3673 return nr_reclaimed;
3677 * This routine traverse page_cgroup in given list and drop them all.
3678 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3680 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3681 int node, int zid, enum lru_list lru)
3684 struct mem_cgroup_per_zone *mz;
3685 struct page_cgroup *pc, *busy;
3686 unsigned long flags, loop;
3687 struct list_head *list;
3690 zone = &NODE_DATA(node)->node_zones[zid];
3691 mz = mem_cgroup_zoneinfo(mem, node, zid);
3692 list = &mz->lists[lru];
3694 loop = MEM_CGROUP_ZSTAT(mz, lru);
3695 /* give some margin against EBUSY etc...*/
3702 spin_lock_irqsave(&zone->lru_lock, flags);
3703 if (list_empty(list)) {
3704 spin_unlock_irqrestore(&zone->lru_lock, flags);
3707 pc = list_entry(list->prev, struct page_cgroup, lru);
3709 list_move(&pc->lru, list);
3711 spin_unlock_irqrestore(&zone->lru_lock, flags);
3714 spin_unlock_irqrestore(&zone->lru_lock, flags);
3716 page = lookup_cgroup_page(pc);
3718 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3722 if (ret == -EBUSY || ret == -EINVAL) {
3723 /* found lock contention or "pc" is obsolete. */
3730 if (!ret && !list_empty(list))
3736 * make mem_cgroup's charge to be 0 if there is no task.
3737 * This enables deleting this mem_cgroup.
3739 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3742 int node, zid, shrink;
3743 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3744 struct cgroup *cgrp = mem->css.cgroup;
3749 /* should free all ? */
3755 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3758 if (signal_pending(current))
3760 /* This is for making all *used* pages to be on LRU. */
3761 lru_add_drain_all();
3762 drain_all_stock_sync();
3764 mem_cgroup_start_move(mem);
3765 for_each_node_state(node, N_HIGH_MEMORY) {
3766 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3769 ret = mem_cgroup_force_empty_list(mem,
3778 mem_cgroup_end_move(mem);
3779 memcg_oom_recover(mem);
3780 /* it seems parent cgroup doesn't have enough mem */
3784 /* "ret" should also be checked to ensure all lists are empty. */
3785 } while (mem->res.usage > 0 || ret);
3791 /* returns EBUSY if there is a task or if we come here twice. */
3792 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3796 /* we call try-to-free pages for make this cgroup empty */
3797 lru_add_drain_all();
3798 /* try to free all pages in this cgroup */
3800 while (nr_retries && mem->res.usage > 0) {
3803 if (signal_pending(current)) {
3807 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3808 false, get_swappiness(mem));
3811 /* maybe some writeback is necessary */
3812 congestion_wait(BLK_RW_ASYNC, HZ/10);
3817 /* try move_account...there may be some *locked* pages. */
3821 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3823 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3827 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3829 return mem_cgroup_from_cont(cont)->use_hierarchy;
3832 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3836 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3837 struct cgroup *parent = cont->parent;
3838 struct mem_cgroup *parent_mem = NULL;
3841 parent_mem = mem_cgroup_from_cont(parent);
3845 * If parent's use_hierarchy is set, we can't make any modifications
3846 * in the child subtrees. If it is unset, then the change can
3847 * occur, provided the current cgroup has no children.
3849 * For the root cgroup, parent_mem is NULL, we allow value to be
3850 * set if there are no children.
3852 if ((!parent_mem || !parent_mem->use_hierarchy) &&
3853 (val == 1 || val == 0)) {
3854 if (list_empty(&cont->children))
3855 mem->use_hierarchy = val;
3866 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3867 enum mem_cgroup_stat_index idx)
3869 struct mem_cgroup *iter;
3872 /* Per-cpu values can be negative, use a signed accumulator */
3873 for_each_mem_cgroup_tree(iter, mem)
3874 val += mem_cgroup_read_stat(iter, idx);
3876 if (val < 0) /* race ? */
3881 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3885 if (!mem_cgroup_is_root(mem)) {
3887 return res_counter_read_u64(&mem->res, RES_USAGE);
3889 return res_counter_read_u64(&mem->memsw, RES_USAGE);
3892 val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
3893 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
3896 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3898 return val << PAGE_SHIFT;
3901 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3903 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3907 type = MEMFILE_TYPE(cft->private);
3908 name = MEMFILE_ATTR(cft->private);
3911 if (name == RES_USAGE)
3912 val = mem_cgroup_usage(mem, false);
3914 val = res_counter_read_u64(&mem->res, name);
3917 if (name == RES_USAGE)
3918 val = mem_cgroup_usage(mem, true);
3920 val = res_counter_read_u64(&mem->memsw, name);
3929 * The user of this function is...
3932 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3935 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3937 unsigned long long val;
3940 type = MEMFILE_TYPE(cft->private);
3941 name = MEMFILE_ATTR(cft->private);
3944 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3948 /* This function does all necessary parse...reuse it */
3949 ret = res_counter_memparse_write_strategy(buffer, &val);
3953 ret = mem_cgroup_resize_limit(memcg, val);
3955 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3957 case RES_SOFT_LIMIT:
3958 ret = res_counter_memparse_write_strategy(buffer, &val);
3962 * For memsw, soft limits are hard to implement in terms
3963 * of semantics, for now, we support soft limits for
3964 * control without swap
3967 ret = res_counter_set_soft_limit(&memcg->res, val);
3972 ret = -EINVAL; /* should be BUG() ? */
3978 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3979 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3981 struct cgroup *cgroup;
3982 unsigned long long min_limit, min_memsw_limit, tmp;
3984 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3985 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3986 cgroup = memcg->css.cgroup;
3987 if (!memcg->use_hierarchy)
3990 while (cgroup->parent) {
3991 cgroup = cgroup->parent;
3992 memcg = mem_cgroup_from_cont(cgroup);
3993 if (!memcg->use_hierarchy)
3995 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3996 min_limit = min(min_limit, tmp);
3997 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3998 min_memsw_limit = min(min_memsw_limit, tmp);
4001 *mem_limit = min_limit;
4002 *memsw_limit = min_memsw_limit;
4006 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4008 struct mem_cgroup *mem;
4011 mem = mem_cgroup_from_cont(cont);
4012 type = MEMFILE_TYPE(event);
4013 name = MEMFILE_ATTR(event);
4017 res_counter_reset_max(&mem->res);
4019 res_counter_reset_max(&mem->memsw);
4023 res_counter_reset_failcnt(&mem->res);
4025 res_counter_reset_failcnt(&mem->memsw);
4032 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4035 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4039 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4040 struct cftype *cft, u64 val)
4042 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4044 if (val >= (1 << NR_MOVE_TYPE))
4047 * We check this value several times in both in can_attach() and
4048 * attach(), so we need cgroup lock to prevent this value from being
4052 mem->move_charge_at_immigrate = val;
4058 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4059 struct cftype *cft, u64 val)
4066 /* For read statistics */
4084 struct mcs_total_stat {
4085 s64 stat[NR_MCS_STAT];
4091 } memcg_stat_strings[NR_MCS_STAT] = {
4092 {"cache", "total_cache"},
4093 {"rss", "total_rss"},
4094 {"mapped_file", "total_mapped_file"},
4095 {"pgpgin", "total_pgpgin"},
4096 {"pgpgout", "total_pgpgout"},
4097 {"swap", "total_swap"},
4098 {"pgfault", "total_pgfault"},
4099 {"pgmajfault", "total_pgmajfault"},
4100 {"inactive_anon", "total_inactive_anon"},
4101 {"active_anon", "total_active_anon"},
4102 {"inactive_file", "total_inactive_file"},
4103 {"active_file", "total_active_file"},
4104 {"unevictable", "total_unevictable"}
4109 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4114 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4115 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4116 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4117 s->stat[MCS_RSS] += val * PAGE_SIZE;
4118 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4119 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4120 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4121 s->stat[MCS_PGPGIN] += val;
4122 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4123 s->stat[MCS_PGPGOUT] += val;
4124 if (do_swap_account) {
4125 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4126 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4128 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4129 s->stat[MCS_PGFAULT] += val;
4130 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4131 s->stat[MCS_PGMAJFAULT] += val;
4134 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
4135 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4136 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
4137 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4138 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
4139 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4140 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
4141 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4142 val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
4143 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4147 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4149 struct mem_cgroup *iter;
4151 for_each_mem_cgroup_tree(iter, mem)
4152 mem_cgroup_get_local_stat(iter, s);
4156 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4159 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4160 unsigned long node_nr;
4161 struct cgroup *cont = m->private;
4162 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4164 total_nr = mem_cgroup_nr_lru_pages(mem_cont);
4165 seq_printf(m, "total=%lu", total_nr);
4166 for_each_node_state(nid, N_HIGH_MEMORY) {
4167 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid);
4168 seq_printf(m, " N%d=%lu", nid, node_nr);
4172 file_nr = mem_cgroup_nr_file_lru_pages(mem_cont);
4173 seq_printf(m, "file=%lu", file_nr);
4174 for_each_node_state(nid, N_HIGH_MEMORY) {
4175 node_nr = mem_cgroup_node_nr_file_lru_pages(mem_cont, nid);
4176 seq_printf(m, " N%d=%lu", nid, node_nr);
4180 anon_nr = mem_cgroup_nr_anon_lru_pages(mem_cont);
4181 seq_printf(m, "anon=%lu", anon_nr);
4182 for_each_node_state(nid, N_HIGH_MEMORY) {
4183 node_nr = mem_cgroup_node_nr_anon_lru_pages(mem_cont, nid);
4184 seq_printf(m, " N%d=%lu", nid, node_nr);
4188 unevictable_nr = mem_cgroup_nr_unevictable_lru_pages(mem_cont);
4189 seq_printf(m, "unevictable=%lu", unevictable_nr);
4190 for_each_node_state(nid, N_HIGH_MEMORY) {
4191 node_nr = mem_cgroup_node_nr_unevictable_lru_pages(mem_cont,
4193 seq_printf(m, " N%d=%lu", nid, node_nr);
4198 #endif /* CONFIG_NUMA */
4200 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4201 struct cgroup_map_cb *cb)
4203 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4204 struct mcs_total_stat mystat;
4207 memset(&mystat, 0, sizeof(mystat));
4208 mem_cgroup_get_local_stat(mem_cont, &mystat);
4211 for (i = 0; i < NR_MCS_STAT; i++) {
4212 if (i == MCS_SWAP && !do_swap_account)
4214 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4217 /* Hierarchical information */
4219 unsigned long long limit, memsw_limit;
4220 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4221 cb->fill(cb, "hierarchical_memory_limit", limit);
4222 if (do_swap_account)
4223 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4226 memset(&mystat, 0, sizeof(mystat));
4227 mem_cgroup_get_total_stat(mem_cont, &mystat);
4228 for (i = 0; i < NR_MCS_STAT; i++) {
4229 if (i == MCS_SWAP && !do_swap_account)
4231 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4234 #ifdef CONFIG_DEBUG_VM
4235 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4239 struct mem_cgroup_per_zone *mz;
4240 unsigned long recent_rotated[2] = {0, 0};
4241 unsigned long recent_scanned[2] = {0, 0};
4243 for_each_online_node(nid)
4244 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4245 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4247 recent_rotated[0] +=
4248 mz->reclaim_stat.recent_rotated[0];
4249 recent_rotated[1] +=
4250 mz->reclaim_stat.recent_rotated[1];
4251 recent_scanned[0] +=
4252 mz->reclaim_stat.recent_scanned[0];
4253 recent_scanned[1] +=
4254 mz->reclaim_stat.recent_scanned[1];
4256 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4257 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4258 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4259 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4266 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4268 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4270 return get_swappiness(memcg);
4273 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4276 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4277 struct mem_cgroup *parent;
4282 if (cgrp->parent == NULL)
4285 parent = mem_cgroup_from_cont(cgrp->parent);
4289 /* If under hierarchy, only empty-root can set this value */
4290 if ((parent->use_hierarchy) ||
4291 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4296 memcg->swappiness = val;
4303 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4305 struct mem_cgroup_threshold_ary *t;
4311 t = rcu_dereference(memcg->thresholds.primary);
4313 t = rcu_dereference(memcg->memsw_thresholds.primary);
4318 usage = mem_cgroup_usage(memcg, swap);
4321 * current_threshold points to threshold just below usage.
4322 * If it's not true, a threshold was crossed after last
4323 * call of __mem_cgroup_threshold().
4325 i = t->current_threshold;
4328 * Iterate backward over array of thresholds starting from
4329 * current_threshold and check if a threshold is crossed.
4330 * If none of thresholds below usage is crossed, we read
4331 * only one element of the array here.
4333 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4334 eventfd_signal(t->entries[i].eventfd, 1);
4336 /* i = current_threshold + 1 */
4340 * Iterate forward over array of thresholds starting from
4341 * current_threshold+1 and check if a threshold is crossed.
4342 * If none of thresholds above usage is crossed, we read
4343 * only one element of the array here.
4345 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4346 eventfd_signal(t->entries[i].eventfd, 1);
4348 /* Update current_threshold */
4349 t->current_threshold = i - 1;
4354 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4357 __mem_cgroup_threshold(memcg, false);
4358 if (do_swap_account)
4359 __mem_cgroup_threshold(memcg, true);
4361 memcg = parent_mem_cgroup(memcg);
4365 static int compare_thresholds(const void *a, const void *b)
4367 const struct mem_cgroup_threshold *_a = a;
4368 const struct mem_cgroup_threshold *_b = b;
4370 return _a->threshold - _b->threshold;
4373 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4375 struct mem_cgroup_eventfd_list *ev;
4377 list_for_each_entry(ev, &mem->oom_notify, list)
4378 eventfd_signal(ev->eventfd, 1);
4382 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4384 struct mem_cgroup *iter;
4386 for_each_mem_cgroup_tree(iter, mem)
4387 mem_cgroup_oom_notify_cb(iter);
4390 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4391 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4393 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4394 struct mem_cgroup_thresholds *thresholds;
4395 struct mem_cgroup_threshold_ary *new;
4396 int type = MEMFILE_TYPE(cft->private);
4397 u64 threshold, usage;
4400 ret = res_counter_memparse_write_strategy(args, &threshold);
4404 mutex_lock(&memcg->thresholds_lock);
4407 thresholds = &memcg->thresholds;
4408 else if (type == _MEMSWAP)
4409 thresholds = &memcg->memsw_thresholds;
4413 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4415 /* Check if a threshold crossed before adding a new one */
4416 if (thresholds->primary)
4417 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4419 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4421 /* Allocate memory for new array of thresholds */
4422 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4430 /* Copy thresholds (if any) to new array */
4431 if (thresholds->primary) {
4432 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4433 sizeof(struct mem_cgroup_threshold));
4436 /* Add new threshold */
4437 new->entries[size - 1].eventfd = eventfd;
4438 new->entries[size - 1].threshold = threshold;
4440 /* Sort thresholds. Registering of new threshold isn't time-critical */
4441 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4442 compare_thresholds, NULL);
4444 /* Find current threshold */
4445 new->current_threshold = -1;
4446 for (i = 0; i < size; i++) {
4447 if (new->entries[i].threshold < usage) {
4449 * new->current_threshold will not be used until
4450 * rcu_assign_pointer(), so it's safe to increment
4453 ++new->current_threshold;
4457 /* Free old spare buffer and save old primary buffer as spare */
4458 kfree(thresholds->spare);
4459 thresholds->spare = thresholds->primary;
4461 rcu_assign_pointer(thresholds->primary, new);
4463 /* To be sure that nobody uses thresholds */
4467 mutex_unlock(&memcg->thresholds_lock);
4472 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4473 struct cftype *cft, struct eventfd_ctx *eventfd)
4475 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4476 struct mem_cgroup_thresholds *thresholds;
4477 struct mem_cgroup_threshold_ary *new;
4478 int type = MEMFILE_TYPE(cft->private);
4482 mutex_lock(&memcg->thresholds_lock);
4484 thresholds = &memcg->thresholds;
4485 else if (type == _MEMSWAP)
4486 thresholds = &memcg->memsw_thresholds;
4491 * Something went wrong if we trying to unregister a threshold
4492 * if we don't have thresholds
4494 BUG_ON(!thresholds);
4496 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4498 /* Check if a threshold crossed before removing */
4499 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4501 /* Calculate new number of threshold */
4503 for (i = 0; i < thresholds->primary->size; i++) {
4504 if (thresholds->primary->entries[i].eventfd != eventfd)
4508 new = thresholds->spare;
4510 /* Set thresholds array to NULL if we don't have thresholds */
4519 /* Copy thresholds and find current threshold */
4520 new->current_threshold = -1;
4521 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4522 if (thresholds->primary->entries[i].eventfd == eventfd)
4525 new->entries[j] = thresholds->primary->entries[i];
4526 if (new->entries[j].threshold < usage) {
4528 * new->current_threshold will not be used
4529 * until rcu_assign_pointer(), so it's safe to increment
4532 ++new->current_threshold;
4538 /* Swap primary and spare array */
4539 thresholds->spare = thresholds->primary;
4540 rcu_assign_pointer(thresholds->primary, new);
4542 /* To be sure that nobody uses thresholds */
4545 mutex_unlock(&memcg->thresholds_lock);
4548 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4549 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4551 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4552 struct mem_cgroup_eventfd_list *event;
4553 int type = MEMFILE_TYPE(cft->private);
4555 BUG_ON(type != _OOM_TYPE);
4556 event = kmalloc(sizeof(*event), GFP_KERNEL);
4560 mutex_lock(&memcg_oom_mutex);
4562 event->eventfd = eventfd;
4563 list_add(&event->list, &memcg->oom_notify);
4565 /* already in OOM ? */
4566 if (atomic_read(&memcg->oom_lock))
4567 eventfd_signal(eventfd, 1);
4568 mutex_unlock(&memcg_oom_mutex);
4573 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4574 struct cftype *cft, struct eventfd_ctx *eventfd)
4576 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4577 struct mem_cgroup_eventfd_list *ev, *tmp;
4578 int type = MEMFILE_TYPE(cft->private);
4580 BUG_ON(type != _OOM_TYPE);
4582 mutex_lock(&memcg_oom_mutex);
4584 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4585 if (ev->eventfd == eventfd) {
4586 list_del(&ev->list);
4591 mutex_unlock(&memcg_oom_mutex);
4594 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4595 struct cftype *cft, struct cgroup_map_cb *cb)
4597 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4599 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4601 if (atomic_read(&mem->oom_lock))
4602 cb->fill(cb, "under_oom", 1);
4604 cb->fill(cb, "under_oom", 0);
4608 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4609 struct cftype *cft, u64 val)
4611 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4612 struct mem_cgroup *parent;
4614 /* cannot set to root cgroup and only 0 and 1 are allowed */
4615 if (!cgrp->parent || !((val == 0) || (val == 1)))
4618 parent = mem_cgroup_from_cont(cgrp->parent);
4621 /* oom-kill-disable is a flag for subhierarchy. */
4622 if ((parent->use_hierarchy) ||
4623 (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4627 mem->oom_kill_disable = val;
4629 memcg_oom_recover(mem);
4635 static const struct file_operations mem_control_numa_stat_file_operations = {
4637 .llseek = seq_lseek,
4638 .release = single_release,
4641 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4643 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4645 file->f_op = &mem_control_numa_stat_file_operations;
4646 return single_open(file, mem_control_numa_stat_show, cont);
4648 #endif /* CONFIG_NUMA */
4650 static struct cftype mem_cgroup_files[] = {
4652 .name = "usage_in_bytes",
4653 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4654 .read_u64 = mem_cgroup_read,
4655 .register_event = mem_cgroup_usage_register_event,
4656 .unregister_event = mem_cgroup_usage_unregister_event,
4659 .name = "max_usage_in_bytes",
4660 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4661 .trigger = mem_cgroup_reset,
4662 .read_u64 = mem_cgroup_read,
4665 .name = "limit_in_bytes",
4666 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4667 .write_string = mem_cgroup_write,
4668 .read_u64 = mem_cgroup_read,
4671 .name = "soft_limit_in_bytes",
4672 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4673 .write_string = mem_cgroup_write,
4674 .read_u64 = mem_cgroup_read,
4678 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4679 .trigger = mem_cgroup_reset,
4680 .read_u64 = mem_cgroup_read,
4684 .read_map = mem_control_stat_show,
4687 .name = "force_empty",
4688 .trigger = mem_cgroup_force_empty_write,
4691 .name = "use_hierarchy",
4692 .write_u64 = mem_cgroup_hierarchy_write,
4693 .read_u64 = mem_cgroup_hierarchy_read,
4696 .name = "swappiness",
4697 .read_u64 = mem_cgroup_swappiness_read,
4698 .write_u64 = mem_cgroup_swappiness_write,
4701 .name = "move_charge_at_immigrate",
4702 .read_u64 = mem_cgroup_move_charge_read,
4703 .write_u64 = mem_cgroup_move_charge_write,
4706 .name = "oom_control",
4707 .read_map = mem_cgroup_oom_control_read,
4708 .write_u64 = mem_cgroup_oom_control_write,
4709 .register_event = mem_cgroup_oom_register_event,
4710 .unregister_event = mem_cgroup_oom_unregister_event,
4711 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4715 .name = "numa_stat",
4716 .open = mem_control_numa_stat_open,
4722 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4723 static struct cftype memsw_cgroup_files[] = {
4725 .name = "memsw.usage_in_bytes",
4726 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4727 .read_u64 = mem_cgroup_read,
4728 .register_event = mem_cgroup_usage_register_event,
4729 .unregister_event = mem_cgroup_usage_unregister_event,
4732 .name = "memsw.max_usage_in_bytes",
4733 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4734 .trigger = mem_cgroup_reset,
4735 .read_u64 = mem_cgroup_read,
4738 .name = "memsw.limit_in_bytes",
4739 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4740 .write_string = mem_cgroup_write,
4741 .read_u64 = mem_cgroup_read,
4744 .name = "memsw.failcnt",
4745 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4746 .trigger = mem_cgroup_reset,
4747 .read_u64 = mem_cgroup_read,
4751 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4753 if (!do_swap_account)
4755 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4756 ARRAY_SIZE(memsw_cgroup_files));
4759 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4765 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4767 struct mem_cgroup_per_node *pn;
4768 struct mem_cgroup_per_zone *mz;
4770 int zone, tmp = node;
4772 * This routine is called against possible nodes.
4773 * But it's BUG to call kmalloc() against offline node.
4775 * TODO: this routine can waste much memory for nodes which will
4776 * never be onlined. It's better to use memory hotplug callback
4779 if (!node_state(node, N_NORMAL_MEMORY))
4781 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4785 mem->info.nodeinfo[node] = pn;
4786 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4787 mz = &pn->zoneinfo[zone];
4789 INIT_LIST_HEAD(&mz->lists[l]);
4790 mz->usage_in_excess = 0;
4791 mz->on_tree = false;
4797 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4799 kfree(mem->info.nodeinfo[node]);
4802 static struct mem_cgroup *mem_cgroup_alloc(void)
4804 struct mem_cgroup *mem;
4805 int size = sizeof(struct mem_cgroup);
4807 /* Can be very big if MAX_NUMNODES is very big */
4808 if (size < PAGE_SIZE)
4809 mem = kzalloc(size, GFP_KERNEL);
4811 mem = vzalloc(size);
4816 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4819 spin_lock_init(&mem->pcp_counter_lock);
4823 if (size < PAGE_SIZE)
4831 * At destroying mem_cgroup, references from swap_cgroup can remain.
4832 * (scanning all at force_empty is too costly...)
4834 * Instead of clearing all references at force_empty, we remember
4835 * the number of reference from swap_cgroup and free mem_cgroup when
4836 * it goes down to 0.
4838 * Removal of cgroup itself succeeds regardless of refs from swap.
4841 static void __mem_cgroup_free(struct mem_cgroup *mem)
4845 mem_cgroup_remove_from_trees(mem);
4846 free_css_id(&mem_cgroup_subsys, &mem->css);
4848 for_each_node_state(node, N_POSSIBLE)
4849 free_mem_cgroup_per_zone_info(mem, node);
4851 free_percpu(mem->stat);
4852 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4858 static void mem_cgroup_get(struct mem_cgroup *mem)
4860 atomic_inc(&mem->refcnt);
4863 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4865 if (atomic_sub_and_test(count, &mem->refcnt)) {
4866 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4867 __mem_cgroup_free(mem);
4869 mem_cgroup_put(parent);
4873 static void mem_cgroup_put(struct mem_cgroup *mem)
4875 __mem_cgroup_put(mem, 1);
4879 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4881 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4883 if (!mem->res.parent)
4885 return mem_cgroup_from_res_counter(mem->res.parent, res);
4888 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4889 static void __init enable_swap_cgroup(void)
4891 if (!mem_cgroup_disabled() && really_do_swap_account)
4892 do_swap_account = 1;
4895 static void __init enable_swap_cgroup(void)
4900 static int mem_cgroup_soft_limit_tree_init(void)
4902 struct mem_cgroup_tree_per_node *rtpn;
4903 struct mem_cgroup_tree_per_zone *rtpz;
4904 int tmp, node, zone;
4906 for_each_node_state(node, N_POSSIBLE) {
4908 if (!node_state(node, N_NORMAL_MEMORY))
4910 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4914 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4916 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4917 rtpz = &rtpn->rb_tree_per_zone[zone];
4918 rtpz->rb_root = RB_ROOT;
4919 spin_lock_init(&rtpz->lock);
4925 static struct cgroup_subsys_state * __ref
4926 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4928 struct mem_cgroup *mem, *parent;
4929 long error = -ENOMEM;
4932 mem = mem_cgroup_alloc();
4934 return ERR_PTR(error);
4936 for_each_node_state(node, N_POSSIBLE)
4937 if (alloc_mem_cgroup_per_zone_info(mem, node))
4941 if (cont->parent == NULL) {
4943 enable_swap_cgroup();
4945 root_mem_cgroup = mem;
4946 if (mem_cgroup_soft_limit_tree_init())
4948 for_each_possible_cpu(cpu) {
4949 struct memcg_stock_pcp *stock =
4950 &per_cpu(memcg_stock, cpu);
4951 INIT_WORK(&stock->work, drain_local_stock);
4953 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4955 parent = mem_cgroup_from_cont(cont->parent);
4956 mem->use_hierarchy = parent->use_hierarchy;
4957 mem->oom_kill_disable = parent->oom_kill_disable;
4960 if (parent && parent->use_hierarchy) {
4961 res_counter_init(&mem->res, &parent->res);
4962 res_counter_init(&mem->memsw, &parent->memsw);
4964 * We increment refcnt of the parent to ensure that we can
4965 * safely access it on res_counter_charge/uncharge.
4966 * This refcnt will be decremented when freeing this
4967 * mem_cgroup(see mem_cgroup_put).
4969 mem_cgroup_get(parent);
4971 res_counter_init(&mem->res, NULL);
4972 res_counter_init(&mem->memsw, NULL);
4974 mem->last_scanned_child = 0;
4975 mem->last_scanned_node = MAX_NUMNODES;
4976 INIT_LIST_HEAD(&mem->oom_notify);
4979 mem->swappiness = get_swappiness(parent);
4980 atomic_set(&mem->refcnt, 1);
4981 mem->move_charge_at_immigrate = 0;
4982 mutex_init(&mem->thresholds_lock);
4985 __mem_cgroup_free(mem);
4986 root_mem_cgroup = NULL;
4987 return ERR_PTR(error);
4990 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4991 struct cgroup *cont)
4993 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4995 return mem_cgroup_force_empty(mem, false);
4998 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4999 struct cgroup *cont)
5001 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5003 mem_cgroup_put(mem);
5006 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5007 struct cgroup *cont)
5011 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5012 ARRAY_SIZE(mem_cgroup_files));
5015 ret = register_memsw_files(cont, ss);
5020 /* Handlers for move charge at task migration. */
5021 #define PRECHARGE_COUNT_AT_ONCE 256
5022 static int mem_cgroup_do_precharge(unsigned long count)
5025 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5026 struct mem_cgroup *mem = mc.to;
5028 if (mem_cgroup_is_root(mem)) {
5029 mc.precharge += count;
5030 /* we don't need css_get for root */
5033 /* try to charge at once */
5035 struct res_counter *dummy;
5037 * "mem" cannot be under rmdir() because we've already checked
5038 * by cgroup_lock_live_cgroup() that it is not removed and we
5039 * are still under the same cgroup_mutex. So we can postpone
5042 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
5044 if (do_swap_account && res_counter_charge(&mem->memsw,
5045 PAGE_SIZE * count, &dummy)) {
5046 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
5049 mc.precharge += count;
5053 /* fall back to one by one charge */
5055 if (signal_pending(current)) {
5059 if (!batch_count--) {
5060 batch_count = PRECHARGE_COUNT_AT_ONCE;
5063 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5065 /* mem_cgroup_clear_mc() will do uncharge later */
5073 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5074 * @vma: the vma the pte to be checked belongs
5075 * @addr: the address corresponding to the pte to be checked
5076 * @ptent: the pte to be checked
5077 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5080 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5081 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5082 * move charge. if @target is not NULL, the page is stored in target->page
5083 * with extra refcnt got(Callers should handle it).
5084 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5085 * target for charge migration. if @target is not NULL, the entry is stored
5088 * Called with pte lock held.
5095 enum mc_target_type {
5096 MC_TARGET_NONE, /* not used */
5101 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5102 unsigned long addr, pte_t ptent)
5104 struct page *page = vm_normal_page(vma, addr, ptent);
5106 if (!page || !page_mapped(page))
5108 if (PageAnon(page)) {
5109 /* we don't move shared anon */
5110 if (!move_anon() || page_mapcount(page) > 2)
5112 } else if (!move_file())
5113 /* we ignore mapcount for file pages */
5115 if (!get_page_unless_zero(page))
5121 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5122 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5125 struct page *page = NULL;
5126 swp_entry_t ent = pte_to_swp_entry(ptent);
5128 if (!move_anon() || non_swap_entry(ent))
5130 usage_count = mem_cgroup_count_swap_user(ent, &page);
5131 if (usage_count > 1) { /* we don't move shared anon */
5136 if (do_swap_account)
5137 entry->val = ent.val;
5142 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5143 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5145 struct page *page = NULL;
5146 struct inode *inode;
5147 struct address_space *mapping;
5150 if (!vma->vm_file) /* anonymous vma */
5155 inode = vma->vm_file->f_path.dentry->d_inode;
5156 mapping = vma->vm_file->f_mapping;
5157 if (pte_none(ptent))
5158 pgoff = linear_page_index(vma, addr);
5159 else /* pte_file(ptent) is true */
5160 pgoff = pte_to_pgoff(ptent);
5162 /* page is moved even if it's not RSS of this task(page-faulted). */
5163 if (!mapping_cap_swap_backed(mapping)) { /* normal file */
5164 page = find_get_page(mapping, pgoff);
5165 } else { /* shmem/tmpfs file. we should take account of swap too. */
5167 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
5168 if (do_swap_account)
5169 entry->val = ent.val;
5175 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5176 unsigned long addr, pte_t ptent, union mc_target *target)
5178 struct page *page = NULL;
5179 struct page_cgroup *pc;
5181 swp_entry_t ent = { .val = 0 };
5183 if (pte_present(ptent))
5184 page = mc_handle_present_pte(vma, addr, ptent);
5185 else if (is_swap_pte(ptent))
5186 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5187 else if (pte_none(ptent) || pte_file(ptent))
5188 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5190 if (!page && !ent.val)
5193 pc = lookup_page_cgroup(page);
5195 * Do only loose check w/o page_cgroup lock.
5196 * mem_cgroup_move_account() checks the pc is valid or not under
5199 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5200 ret = MC_TARGET_PAGE;
5202 target->page = page;
5204 if (!ret || !target)
5207 /* There is a swap entry and a page doesn't exist or isn't charged */
5208 if (ent.val && !ret &&
5209 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5210 ret = MC_TARGET_SWAP;
5217 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5218 unsigned long addr, unsigned long end,
5219 struct mm_walk *walk)
5221 struct vm_area_struct *vma = walk->private;
5225 split_huge_page_pmd(walk->mm, pmd);
5227 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5228 for (; addr != end; pte++, addr += PAGE_SIZE)
5229 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5230 mc.precharge++; /* increment precharge temporarily */
5231 pte_unmap_unlock(pte - 1, ptl);
5237 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5239 unsigned long precharge;
5240 struct vm_area_struct *vma;
5242 down_read(&mm->mmap_sem);
5243 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5244 struct mm_walk mem_cgroup_count_precharge_walk = {
5245 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5249 if (is_vm_hugetlb_page(vma))
5251 walk_page_range(vma->vm_start, vma->vm_end,
5252 &mem_cgroup_count_precharge_walk);
5254 up_read(&mm->mmap_sem);
5256 precharge = mc.precharge;
5262 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5264 unsigned long precharge = mem_cgroup_count_precharge(mm);
5266 VM_BUG_ON(mc.moving_task);
5267 mc.moving_task = current;
5268 return mem_cgroup_do_precharge(precharge);
5271 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5272 static void __mem_cgroup_clear_mc(void)
5274 struct mem_cgroup *from = mc.from;
5275 struct mem_cgroup *to = mc.to;
5277 /* we must uncharge all the leftover precharges from mc.to */
5279 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5283 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5284 * we must uncharge here.
5286 if (mc.moved_charge) {
5287 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5288 mc.moved_charge = 0;
5290 /* we must fixup refcnts and charges */
5291 if (mc.moved_swap) {
5292 /* uncharge swap account from the old cgroup */
5293 if (!mem_cgroup_is_root(mc.from))
5294 res_counter_uncharge(&mc.from->memsw,
5295 PAGE_SIZE * mc.moved_swap);
5296 __mem_cgroup_put(mc.from, mc.moved_swap);
5298 if (!mem_cgroup_is_root(mc.to)) {
5300 * we charged both to->res and to->memsw, so we should
5303 res_counter_uncharge(&mc.to->res,
5304 PAGE_SIZE * mc.moved_swap);
5306 /* we've already done mem_cgroup_get(mc.to) */
5309 memcg_oom_recover(from);
5310 memcg_oom_recover(to);
5311 wake_up_all(&mc.waitq);
5314 static void mem_cgroup_clear_mc(void)
5316 struct mem_cgroup *from = mc.from;
5319 * we must clear moving_task before waking up waiters at the end of
5322 mc.moving_task = NULL;
5323 __mem_cgroup_clear_mc();
5324 spin_lock(&mc.lock);
5327 spin_unlock(&mc.lock);
5328 mem_cgroup_end_move(from);
5331 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5332 struct cgroup *cgroup,
5333 struct task_struct *p)
5336 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5338 if (mem->move_charge_at_immigrate) {
5339 struct mm_struct *mm;
5340 struct mem_cgroup *from = mem_cgroup_from_task(p);
5342 VM_BUG_ON(from == mem);
5344 mm = get_task_mm(p);
5347 /* We move charges only when we move a owner of the mm */
5348 if (mm->owner == p) {
5351 VM_BUG_ON(mc.precharge);
5352 VM_BUG_ON(mc.moved_charge);
5353 VM_BUG_ON(mc.moved_swap);
5354 mem_cgroup_start_move(from);
5355 spin_lock(&mc.lock);
5358 spin_unlock(&mc.lock);
5359 /* We set mc.moving_task later */
5361 ret = mem_cgroup_precharge_mc(mm);
5363 mem_cgroup_clear_mc();
5370 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5371 struct cgroup *cgroup,
5372 struct task_struct *p)
5374 mem_cgroup_clear_mc();
5377 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5378 unsigned long addr, unsigned long end,
5379 struct mm_walk *walk)
5382 struct vm_area_struct *vma = walk->private;
5386 split_huge_page_pmd(walk->mm, pmd);
5388 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5389 for (; addr != end; addr += PAGE_SIZE) {
5390 pte_t ptent = *(pte++);
5391 union mc_target target;
5394 struct page_cgroup *pc;
5400 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5402 case MC_TARGET_PAGE:
5404 if (isolate_lru_page(page))
5406 pc = lookup_page_cgroup(page);
5407 if (!mem_cgroup_move_account(page, 1, pc,
5408 mc.from, mc.to, false)) {
5410 /* we uncharge from mc.from later. */
5413 putback_lru_page(page);
5414 put: /* is_target_pte_for_mc() gets the page */
5417 case MC_TARGET_SWAP:
5419 if (!mem_cgroup_move_swap_account(ent,
5420 mc.from, mc.to, false)) {
5422 /* we fixup refcnts and charges later. */
5430 pte_unmap_unlock(pte - 1, ptl);
5435 * We have consumed all precharges we got in can_attach().
5436 * We try charge one by one, but don't do any additional
5437 * charges to mc.to if we have failed in charge once in attach()
5440 ret = mem_cgroup_do_precharge(1);
5448 static void mem_cgroup_move_charge(struct mm_struct *mm)
5450 struct vm_area_struct *vma;
5452 lru_add_drain_all();
5454 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5456 * Someone who are holding the mmap_sem might be waiting in
5457 * waitq. So we cancel all extra charges, wake up all waiters,
5458 * and retry. Because we cancel precharges, we might not be able
5459 * to move enough charges, but moving charge is a best-effort
5460 * feature anyway, so it wouldn't be a big problem.
5462 __mem_cgroup_clear_mc();
5466 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5468 struct mm_walk mem_cgroup_move_charge_walk = {
5469 .pmd_entry = mem_cgroup_move_charge_pte_range,
5473 if (is_vm_hugetlb_page(vma))
5475 ret = walk_page_range(vma->vm_start, vma->vm_end,
5476 &mem_cgroup_move_charge_walk);
5479 * means we have consumed all precharges and failed in
5480 * doing additional charge. Just abandon here.
5484 up_read(&mm->mmap_sem);
5487 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5488 struct cgroup *cont,
5489 struct cgroup *old_cont,
5490 struct task_struct *p)
5492 struct mm_struct *mm = get_task_mm(p);
5496 mem_cgroup_move_charge(mm);
5501 mem_cgroup_clear_mc();
5503 #else /* !CONFIG_MMU */
5504 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5505 struct cgroup *cgroup,
5506 struct task_struct *p)
5510 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5511 struct cgroup *cgroup,
5512 struct task_struct *p)
5515 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5516 struct cgroup *cont,
5517 struct cgroup *old_cont,
5518 struct task_struct *p)
5523 struct cgroup_subsys mem_cgroup_subsys = {
5525 .subsys_id = mem_cgroup_subsys_id,
5526 .create = mem_cgroup_create,
5527 .pre_destroy = mem_cgroup_pre_destroy,
5528 .destroy = mem_cgroup_destroy,
5529 .populate = mem_cgroup_populate,
5530 .can_attach = mem_cgroup_can_attach,
5531 .cancel_attach = mem_cgroup_cancel_attach,
5532 .attach = mem_cgroup_move_task,
5537 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5538 static int __init enable_swap_account(char *s)
5540 /* consider enabled if no parameter or 1 is given */
5541 if (!strcmp(s, "1"))
5542 really_do_swap_account = 1;
5543 else if (!strcmp(s, "0"))
5544 really_do_swap_account = 0;
5547 __setup("swapaccount=", enable_swap_account);