1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/mutex.h>
37 #include <linux/rbtree.h>
38 #include <linux/shmem_fs.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
54 #include <asm/uaccess.h>
56 #include <trace/events/vmscan.h>
58 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
59 #define MEM_CGROUP_RECLAIM_RETRIES 5
60 struct mem_cgroup *root_mem_cgroup __read_mostly;
62 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
63 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
64 int do_swap_account __read_mostly;
66 /* for remember boot option*/
67 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
68 static int really_do_swap_account __initdata = 1;
70 static int really_do_swap_account __initdata = 0;
74 #define do_swap_account (0)
79 * Statistics for memory cgroup.
81 enum mem_cgroup_stat_index {
83 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
85 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
86 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
87 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
88 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
89 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
90 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */
91 MEM_CGROUP_STAT_NSTATS,
94 enum mem_cgroup_events_index {
95 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
96 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
97 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
98 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
99 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
100 MEM_CGROUP_EVENTS_NSTATS,
103 * Per memcg event counter is incremented at every pagein/pageout. With THP,
104 * it will be incremated by the number of pages. This counter is used for
105 * for trigger some periodic events. This is straightforward and better
106 * than using jiffies etc. to handle periodic memcg event.
108 enum mem_cgroup_events_target {
109 MEM_CGROUP_TARGET_THRESH,
110 MEM_CGROUP_TARGET_SOFTLIMIT,
111 MEM_CGROUP_TARGET_NUMAINFO,
114 #define THRESHOLDS_EVENTS_TARGET (128)
115 #define SOFTLIMIT_EVENTS_TARGET (1024)
116 #define NUMAINFO_EVENTS_TARGET (1024)
118 struct mem_cgroup_stat_cpu {
119 long count[MEM_CGROUP_STAT_NSTATS];
120 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
121 unsigned long targets[MEM_CGROUP_NTARGETS];
125 * per-zone information in memory controller.
127 struct mem_cgroup_per_zone {
129 * spin_lock to protect the per cgroup LRU
131 struct list_head lists[NR_LRU_LISTS];
132 unsigned long count[NR_LRU_LISTS];
134 struct zone_reclaim_stat reclaim_stat;
135 struct rb_node tree_node; /* RB tree node */
136 unsigned long long usage_in_excess;/* Set to the value by which */
137 /* the soft limit is exceeded*/
139 struct mem_cgroup *mem; /* Back pointer, we cannot */
140 /* use container_of */
142 /* Macro for accessing counter */
143 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
145 struct mem_cgroup_per_node {
146 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
149 struct mem_cgroup_lru_info {
150 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
154 * Cgroups above their limits are maintained in a RB-Tree, independent of
155 * their hierarchy representation
158 struct mem_cgroup_tree_per_zone {
159 struct rb_root rb_root;
163 struct mem_cgroup_tree_per_node {
164 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
167 struct mem_cgroup_tree {
168 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
171 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
173 struct mem_cgroup_threshold {
174 struct eventfd_ctx *eventfd;
179 struct mem_cgroup_threshold_ary {
180 /* An array index points to threshold just below usage. */
181 int current_threshold;
182 /* Size of entries[] */
184 /* Array of thresholds */
185 struct mem_cgroup_threshold entries[0];
188 struct mem_cgroup_thresholds {
189 /* Primary thresholds array */
190 struct mem_cgroup_threshold_ary *primary;
192 * Spare threshold array.
193 * This is needed to make mem_cgroup_unregister_event() "never fail".
194 * It must be able to store at least primary->size - 1 entries.
196 struct mem_cgroup_threshold_ary *spare;
200 struct mem_cgroup_eventfd_list {
201 struct list_head list;
202 struct eventfd_ctx *eventfd;
205 static void mem_cgroup_threshold(struct mem_cgroup *mem);
206 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
209 * The memory controller data structure. The memory controller controls both
210 * page cache and RSS per cgroup. We would eventually like to provide
211 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
212 * to help the administrator determine what knobs to tune.
214 * TODO: Add a water mark for the memory controller. Reclaim will begin when
215 * we hit the water mark. May be even add a low water mark, such that
216 * no reclaim occurs from a cgroup at it's low water mark, this is
217 * a feature that will be implemented much later in the future.
220 struct cgroup_subsys_state css;
222 * the counter to account for memory usage
224 struct res_counter res;
226 * the counter to account for mem+swap usage.
228 struct res_counter memsw;
230 * Per cgroup active and inactive list, similar to the
231 * per zone LRU lists.
233 struct mem_cgroup_lru_info info;
235 * While reclaiming in a hierarchy, we cache the last child we
238 int last_scanned_child;
239 int last_scanned_node;
241 nodemask_t scan_nodes;
242 atomic_t numainfo_events;
243 atomic_t numainfo_updating;
246 * Should the accounting and control be hierarchical, per subtree?
256 /* OOM-Killer disable */
257 int oom_kill_disable;
259 /* set when res.limit == memsw.limit */
260 bool memsw_is_minimum;
262 /* protect arrays of thresholds */
263 struct mutex thresholds_lock;
265 /* thresholds for memory usage. RCU-protected */
266 struct mem_cgroup_thresholds thresholds;
268 /* thresholds for mem+swap usage. RCU-protected */
269 struct mem_cgroup_thresholds memsw_thresholds;
271 /* For oom notifier event fd */
272 struct list_head oom_notify;
275 * Should we move charges of a task when a task is moved into this
276 * mem_cgroup ? And what type of charges should we move ?
278 unsigned long move_charge_at_immigrate;
282 struct mem_cgroup_stat_cpu *stat;
284 * used when a cpu is offlined or other synchronizations
285 * See mem_cgroup_read_stat().
287 struct mem_cgroup_stat_cpu nocpu_base;
288 spinlock_t pcp_counter_lock;
291 /* Stuffs for move charges at task migration. */
293 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
294 * left-shifted bitmap of these types.
297 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
298 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
302 /* "mc" and its members are protected by cgroup_mutex */
303 static struct move_charge_struct {
304 spinlock_t lock; /* for from, to */
305 struct mem_cgroup *from;
306 struct mem_cgroup *to;
307 unsigned long precharge;
308 unsigned long moved_charge;
309 unsigned long moved_swap;
310 struct task_struct *moving_task; /* a task moving charges */
311 wait_queue_head_t waitq; /* a waitq for other context */
313 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
314 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
317 static bool move_anon(void)
319 return test_bit(MOVE_CHARGE_TYPE_ANON,
320 &mc.to->move_charge_at_immigrate);
323 static bool move_file(void)
325 return test_bit(MOVE_CHARGE_TYPE_FILE,
326 &mc.to->move_charge_at_immigrate);
330 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
331 * limit reclaim to prevent infinite loops, if they ever occur.
333 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
334 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
337 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
338 MEM_CGROUP_CHARGE_TYPE_MAPPED,
339 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
340 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
341 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
342 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
346 /* for encoding cft->private value on file */
349 #define _OOM_TYPE (2)
350 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
351 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
352 #define MEMFILE_ATTR(val) ((val) & 0xffff)
353 /* Used for OOM nofiier */
354 #define OOM_CONTROL (0)
357 * Reclaim flags for mem_cgroup_hierarchical_reclaim
359 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
360 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
361 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
362 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
363 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
364 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
366 static void mem_cgroup_get(struct mem_cgroup *mem);
367 static void mem_cgroup_put(struct mem_cgroup *mem);
368 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
369 static void drain_all_stock_async(struct mem_cgroup *mem);
371 static struct mem_cgroup_per_zone *
372 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
374 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
377 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
382 static struct mem_cgroup_per_zone *
383 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
385 int nid = page_to_nid(page);
386 int zid = page_zonenum(page);
388 return mem_cgroup_zoneinfo(mem, nid, zid);
391 static struct mem_cgroup_tree_per_zone *
392 soft_limit_tree_node_zone(int nid, int zid)
394 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
397 static struct mem_cgroup_tree_per_zone *
398 soft_limit_tree_from_page(struct page *page)
400 int nid = page_to_nid(page);
401 int zid = page_zonenum(page);
403 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
407 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
408 struct mem_cgroup_per_zone *mz,
409 struct mem_cgroup_tree_per_zone *mctz,
410 unsigned long long new_usage_in_excess)
412 struct rb_node **p = &mctz->rb_root.rb_node;
413 struct rb_node *parent = NULL;
414 struct mem_cgroup_per_zone *mz_node;
419 mz->usage_in_excess = new_usage_in_excess;
420 if (!mz->usage_in_excess)
424 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
426 if (mz->usage_in_excess < mz_node->usage_in_excess)
429 * We can't avoid mem cgroups that are over their soft
430 * limit by the same amount
432 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
435 rb_link_node(&mz->tree_node, parent, p);
436 rb_insert_color(&mz->tree_node, &mctz->rb_root);
441 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
442 struct mem_cgroup_per_zone *mz,
443 struct mem_cgroup_tree_per_zone *mctz)
447 rb_erase(&mz->tree_node, &mctz->rb_root);
452 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
453 struct mem_cgroup_per_zone *mz,
454 struct mem_cgroup_tree_per_zone *mctz)
456 spin_lock(&mctz->lock);
457 __mem_cgroup_remove_exceeded(mem, mz, mctz);
458 spin_unlock(&mctz->lock);
462 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
464 unsigned long long excess;
465 struct mem_cgroup_per_zone *mz;
466 struct mem_cgroup_tree_per_zone *mctz;
467 int nid = page_to_nid(page);
468 int zid = page_zonenum(page);
469 mctz = soft_limit_tree_from_page(page);
472 * Necessary to update all ancestors when hierarchy is used.
473 * because their event counter is not touched.
475 for (; mem; mem = parent_mem_cgroup(mem)) {
476 mz = mem_cgroup_zoneinfo(mem, nid, zid);
477 excess = res_counter_soft_limit_excess(&mem->res);
479 * We have to update the tree if mz is on RB-tree or
480 * mem is over its softlimit.
482 if (excess || mz->on_tree) {
483 spin_lock(&mctz->lock);
484 /* if on-tree, remove it */
486 __mem_cgroup_remove_exceeded(mem, mz, mctz);
488 * Insert again. mz->usage_in_excess will be updated.
489 * If excess is 0, no tree ops.
491 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
492 spin_unlock(&mctz->lock);
497 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
500 struct mem_cgroup_per_zone *mz;
501 struct mem_cgroup_tree_per_zone *mctz;
503 for_each_node_state(node, N_POSSIBLE) {
504 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
505 mz = mem_cgroup_zoneinfo(mem, node, zone);
506 mctz = soft_limit_tree_node_zone(node, zone);
507 mem_cgroup_remove_exceeded(mem, mz, mctz);
512 static struct mem_cgroup_per_zone *
513 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
515 struct rb_node *rightmost = NULL;
516 struct mem_cgroup_per_zone *mz;
520 rightmost = rb_last(&mctz->rb_root);
522 goto done; /* Nothing to reclaim from */
524 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
526 * Remove the node now but someone else can add it back,
527 * we will to add it back at the end of reclaim to its correct
528 * position in the tree.
530 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
531 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
532 !css_tryget(&mz->mem->css))
538 static struct mem_cgroup_per_zone *
539 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
541 struct mem_cgroup_per_zone *mz;
543 spin_lock(&mctz->lock);
544 mz = __mem_cgroup_largest_soft_limit_node(mctz);
545 spin_unlock(&mctz->lock);
550 * Implementation Note: reading percpu statistics for memcg.
552 * Both of vmstat[] and percpu_counter has threshold and do periodic
553 * synchronization to implement "quick" read. There are trade-off between
554 * reading cost and precision of value. Then, we may have a chance to implement
555 * a periodic synchronizion of counter in memcg's counter.
557 * But this _read() function is used for user interface now. The user accounts
558 * memory usage by memory cgroup and he _always_ requires exact value because
559 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
560 * have to visit all online cpus and make sum. So, for now, unnecessary
561 * synchronization is not implemented. (just implemented for cpu hotplug)
563 * If there are kernel internal actions which can make use of some not-exact
564 * value, and reading all cpu value can be performance bottleneck in some
565 * common workload, threashold and synchonization as vmstat[] should be
568 static long mem_cgroup_read_stat(struct mem_cgroup *mem,
569 enum mem_cgroup_stat_index idx)
575 for_each_online_cpu(cpu)
576 val += per_cpu(mem->stat->count[idx], cpu);
577 #ifdef CONFIG_HOTPLUG_CPU
578 spin_lock(&mem->pcp_counter_lock);
579 val += mem->nocpu_base.count[idx];
580 spin_unlock(&mem->pcp_counter_lock);
586 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
589 int val = (charge) ? 1 : -1;
590 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
593 void mem_cgroup_pgfault(struct mem_cgroup *mem, int val)
595 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
598 void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val)
600 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
603 static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
604 enum mem_cgroup_events_index idx)
606 unsigned long val = 0;
609 for_each_online_cpu(cpu)
610 val += per_cpu(mem->stat->events[idx], cpu);
611 #ifdef CONFIG_HOTPLUG_CPU
612 spin_lock(&mem->pcp_counter_lock);
613 val += mem->nocpu_base.events[idx];
614 spin_unlock(&mem->pcp_counter_lock);
619 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
620 bool file, int nr_pages)
625 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
627 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
629 /* pagein of a big page is an event. So, ignore page size */
631 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
633 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
634 nr_pages = -nr_pages; /* for event */
637 __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
643 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *mem, int nid, int zid,
644 unsigned int lru_mask)
646 struct mem_cgroup_per_zone *mz;
648 unsigned long ret = 0;
650 mz = mem_cgroup_zoneinfo(mem, nid, zid);
653 if (BIT(l) & lru_mask)
654 ret += MEM_CGROUP_ZSTAT(mz, l);
660 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *mem,
661 int nid, unsigned int lru_mask)
666 for (zid = 0; zid < MAX_NR_ZONES; zid++)
667 total += mem_cgroup_zone_nr_lru_pages(mem, nid, zid, lru_mask);
672 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *mem,
673 unsigned int lru_mask)
678 for_each_node_state(nid, N_HIGH_MEMORY)
679 total += mem_cgroup_node_nr_lru_pages(mem, nid, lru_mask);
683 static bool __memcg_event_check(struct mem_cgroup *mem, int target)
685 unsigned long val, next;
687 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
688 next = this_cpu_read(mem->stat->targets[target]);
689 /* from time_after() in jiffies.h */
690 return ((long)next - (long)val < 0);
693 static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target)
695 unsigned long val, next;
697 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
700 case MEM_CGROUP_TARGET_THRESH:
701 next = val + THRESHOLDS_EVENTS_TARGET;
703 case MEM_CGROUP_TARGET_SOFTLIMIT:
704 next = val + SOFTLIMIT_EVENTS_TARGET;
706 case MEM_CGROUP_TARGET_NUMAINFO:
707 next = val + NUMAINFO_EVENTS_TARGET;
713 this_cpu_write(mem->stat->targets[target], next);
717 * Check events in order.
720 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
722 /* threshold event is triggered in finer grain than soft limit */
723 if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) {
724 mem_cgroup_threshold(mem);
725 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH);
726 if (unlikely(__memcg_event_check(mem,
727 MEM_CGROUP_TARGET_SOFTLIMIT))) {
728 mem_cgroup_update_tree(mem, page);
729 __mem_cgroup_target_update(mem,
730 MEM_CGROUP_TARGET_SOFTLIMIT);
733 if (unlikely(__memcg_event_check(mem,
734 MEM_CGROUP_TARGET_NUMAINFO))) {
735 atomic_inc(&mem->numainfo_events);
736 __mem_cgroup_target_update(mem,
737 MEM_CGROUP_TARGET_NUMAINFO);
743 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
745 return container_of(cgroup_subsys_state(cont,
746 mem_cgroup_subsys_id), struct mem_cgroup,
750 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
753 * mm_update_next_owner() may clear mm->owner to NULL
754 * if it races with swapoff, page migration, etc.
755 * So this can be called with p == NULL.
760 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
761 struct mem_cgroup, css);
764 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
766 struct mem_cgroup *mem = NULL;
771 * Because we have no locks, mm->owner's may be being moved to other
772 * cgroup. We use css_tryget() here even if this looks
773 * pessimistic (rather than adding locks here).
777 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
780 } while (!css_tryget(&mem->css));
785 /* The caller has to guarantee "mem" exists before calling this */
786 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
788 struct cgroup_subsys_state *css;
791 if (!mem) /* ROOT cgroup has the smallest ID */
792 return root_mem_cgroup; /*css_put/get against root is ignored*/
793 if (!mem->use_hierarchy) {
794 if (css_tryget(&mem->css))
800 * searching a memory cgroup which has the smallest ID under given
801 * ROOT cgroup. (ID >= 1)
803 css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
804 if (css && css_tryget(css))
805 mem = container_of(css, struct mem_cgroup, css);
812 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
813 struct mem_cgroup *root,
816 int nextid = css_id(&iter->css) + 1;
819 struct cgroup_subsys_state *css;
821 hierarchy_used = iter->use_hierarchy;
824 /* If no ROOT, walk all, ignore hierarchy */
825 if (!cond || (root && !hierarchy_used))
829 root = root_mem_cgroup;
835 css = css_get_next(&mem_cgroup_subsys, nextid,
837 if (css && css_tryget(css))
838 iter = container_of(css, struct mem_cgroup, css);
840 /* If css is NULL, no more cgroups will be found */
842 } while (css && !iter);
847 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
848 * be careful that "break" loop is not allowed. We have reference count.
849 * Instead of that modify "cond" to be false and "continue" to exit the loop.
851 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
852 for (iter = mem_cgroup_start_loop(root);\
854 iter = mem_cgroup_get_next(iter, root, cond))
856 #define for_each_mem_cgroup_tree(iter, root) \
857 for_each_mem_cgroup_tree_cond(iter, root, true)
859 #define for_each_mem_cgroup_all(iter) \
860 for_each_mem_cgroup_tree_cond(iter, NULL, true)
863 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
865 return (mem == root_mem_cgroup);
868 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
870 struct mem_cgroup *mem;
876 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
882 mem_cgroup_pgmajfault(mem, 1);
885 mem_cgroup_pgfault(mem, 1);
893 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
896 * Following LRU functions are allowed to be used without PCG_LOCK.
897 * Operations are called by routine of global LRU independently from memcg.
898 * What we have to take care of here is validness of pc->mem_cgroup.
900 * Changes to pc->mem_cgroup happens when
903 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
904 * It is added to LRU before charge.
905 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
906 * When moving account, the page is not on LRU. It's isolated.
909 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
911 struct page_cgroup *pc;
912 struct mem_cgroup_per_zone *mz;
914 if (mem_cgroup_disabled())
916 pc = lookup_page_cgroup(page);
917 /* can happen while we handle swapcache. */
918 if (!TestClearPageCgroupAcctLRU(pc))
920 VM_BUG_ON(!pc->mem_cgroup);
922 * We don't check PCG_USED bit. It's cleared when the "page" is finally
923 * removed from global LRU.
925 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
926 /* huge page split is done under lru_lock. so, we have no races. */
927 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
928 if (mem_cgroup_is_root(pc->mem_cgroup))
930 VM_BUG_ON(list_empty(&pc->lru));
931 list_del_init(&pc->lru);
934 void mem_cgroup_del_lru(struct page *page)
936 mem_cgroup_del_lru_list(page, page_lru(page));
940 * Writeback is about to end against a page which has been marked for immediate
941 * reclaim. If it still appears to be reclaimable, move it to the tail of the
944 void mem_cgroup_rotate_reclaimable_page(struct page *page)
946 struct mem_cgroup_per_zone *mz;
947 struct page_cgroup *pc;
948 enum lru_list lru = page_lru(page);
950 if (mem_cgroup_disabled())
953 pc = lookup_page_cgroup(page);
954 /* unused or root page is not rotated. */
955 if (!PageCgroupUsed(pc))
957 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
959 if (mem_cgroup_is_root(pc->mem_cgroup))
961 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
962 list_move_tail(&pc->lru, &mz->lists[lru]);
965 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
967 struct mem_cgroup_per_zone *mz;
968 struct page_cgroup *pc;
970 if (mem_cgroup_disabled())
973 pc = lookup_page_cgroup(page);
974 /* unused or root page is not rotated. */
975 if (!PageCgroupUsed(pc))
977 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
979 if (mem_cgroup_is_root(pc->mem_cgroup))
981 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
982 list_move(&pc->lru, &mz->lists[lru]);
985 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
987 struct page_cgroup *pc;
988 struct mem_cgroup_per_zone *mz;
990 if (mem_cgroup_disabled())
992 pc = lookup_page_cgroup(page);
993 VM_BUG_ON(PageCgroupAcctLRU(pc));
994 if (!PageCgroupUsed(pc))
996 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
998 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
999 /* huge page split is done under lru_lock. so, we have no races. */
1000 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1001 SetPageCgroupAcctLRU(pc);
1002 if (mem_cgroup_is_root(pc->mem_cgroup))
1004 list_add(&pc->lru, &mz->lists[lru]);
1008 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1009 * while it's linked to lru because the page may be reused after it's fully
1010 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1011 * It's done under lock_page and expected that zone->lru_lock isnever held.
1013 static void mem_cgroup_lru_del_before_commit(struct page *page)
1015 unsigned long flags;
1016 struct zone *zone = page_zone(page);
1017 struct page_cgroup *pc = lookup_page_cgroup(page);
1020 * Doing this check without taking ->lru_lock seems wrong but this
1021 * is safe. Because if page_cgroup's USED bit is unset, the page
1022 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1023 * set, the commit after this will fail, anyway.
1024 * This all charge/uncharge is done under some mutual execustion.
1025 * So, we don't need to taking care of changes in USED bit.
1027 if (likely(!PageLRU(page)))
1030 spin_lock_irqsave(&zone->lru_lock, flags);
1032 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1033 * is guarded by lock_page() because the page is SwapCache.
1035 if (!PageCgroupUsed(pc))
1036 mem_cgroup_del_lru_list(page, page_lru(page));
1037 spin_unlock_irqrestore(&zone->lru_lock, flags);
1040 static void mem_cgroup_lru_add_after_commit(struct page *page)
1042 unsigned long flags;
1043 struct zone *zone = page_zone(page);
1044 struct page_cgroup *pc = lookup_page_cgroup(page);
1046 /* taking care of that the page is added to LRU while we commit it */
1047 if (likely(!PageLRU(page)))
1049 spin_lock_irqsave(&zone->lru_lock, flags);
1050 /* link when the page is linked to LRU but page_cgroup isn't */
1051 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1052 mem_cgroup_add_lru_list(page, page_lru(page));
1053 spin_unlock_irqrestore(&zone->lru_lock, flags);
1057 void mem_cgroup_move_lists(struct page *page,
1058 enum lru_list from, enum lru_list to)
1060 if (mem_cgroup_disabled())
1062 mem_cgroup_del_lru_list(page, from);
1063 mem_cgroup_add_lru_list(page, to);
1066 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
1069 struct mem_cgroup *curr = NULL;
1070 struct task_struct *p;
1072 p = find_lock_task_mm(task);
1075 curr = try_get_mem_cgroup_from_mm(p->mm);
1080 * We should check use_hierarchy of "mem" not "curr". Because checking
1081 * use_hierarchy of "curr" here make this function true if hierarchy is
1082 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
1083 * hierarchy(even if use_hierarchy is disabled in "mem").
1085 if (mem->use_hierarchy)
1086 ret = css_is_ancestor(&curr->css, &mem->css);
1088 ret = (curr == mem);
1089 css_put(&curr->css);
1093 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1095 unsigned long active;
1096 unsigned long inactive;
1098 unsigned long inactive_ratio;
1100 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
1101 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
1103 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1105 inactive_ratio = int_sqrt(10 * gb);
1109 if (present_pages) {
1110 present_pages[0] = inactive;
1111 present_pages[1] = active;
1114 return inactive_ratio;
1117 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1119 unsigned long active;
1120 unsigned long inactive;
1121 unsigned long present_pages[2];
1122 unsigned long inactive_ratio;
1124 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1126 inactive = present_pages[0];
1127 active = present_pages[1];
1129 if (inactive * inactive_ratio < active)
1135 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1137 unsigned long active;
1138 unsigned long inactive;
1140 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
1141 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
1143 return (active > inactive);
1146 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1149 int nid = zone_to_nid(zone);
1150 int zid = zone_idx(zone);
1151 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1153 return &mz->reclaim_stat;
1156 struct zone_reclaim_stat *
1157 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1159 struct page_cgroup *pc;
1160 struct mem_cgroup_per_zone *mz;
1162 if (mem_cgroup_disabled())
1165 pc = lookup_page_cgroup(page);
1166 if (!PageCgroupUsed(pc))
1168 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1170 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1171 return &mz->reclaim_stat;
1174 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1175 struct list_head *dst,
1176 unsigned long *scanned, int order,
1177 int mode, struct zone *z,
1178 struct mem_cgroup *mem_cont,
1179 int active, int file)
1181 unsigned long nr_taken = 0;
1185 struct list_head *src;
1186 struct page_cgroup *pc, *tmp;
1187 int nid = zone_to_nid(z);
1188 int zid = zone_idx(z);
1189 struct mem_cgroup_per_zone *mz;
1190 int lru = LRU_FILE * file + active;
1194 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1195 src = &mz->lists[lru];
1198 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1199 if (scan >= nr_to_scan)
1202 if (unlikely(!PageCgroupUsed(pc)))
1205 page = lookup_cgroup_page(pc);
1207 if (unlikely(!PageLRU(page)))
1211 ret = __isolate_lru_page(page, mode, file);
1214 list_move(&page->lru, dst);
1215 mem_cgroup_del_lru(page);
1216 nr_taken += hpage_nr_pages(page);
1219 /* we don't affect global LRU but rotate in our LRU */
1220 mem_cgroup_rotate_lru_list(page, page_lru(page));
1229 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1235 #define mem_cgroup_from_res_counter(counter, member) \
1236 container_of(counter, struct mem_cgroup, member)
1239 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1240 * @mem: the memory cgroup
1242 * Returns the maximum amount of memory @mem can be charged with, in
1245 static unsigned long mem_cgroup_margin(struct mem_cgroup *mem)
1247 unsigned long long margin;
1249 margin = res_counter_margin(&mem->res);
1250 if (do_swap_account)
1251 margin = min(margin, res_counter_margin(&mem->memsw));
1252 return margin >> PAGE_SHIFT;
1255 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1257 struct cgroup *cgrp = memcg->css.cgroup;
1260 if (cgrp->parent == NULL)
1261 return vm_swappiness;
1263 return memcg->swappiness;
1266 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1271 spin_lock(&mem->pcp_counter_lock);
1272 for_each_online_cpu(cpu)
1273 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1274 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1275 spin_unlock(&mem->pcp_counter_lock);
1281 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1288 spin_lock(&mem->pcp_counter_lock);
1289 for_each_online_cpu(cpu)
1290 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1291 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1292 spin_unlock(&mem->pcp_counter_lock);
1296 * 2 routines for checking "mem" is under move_account() or not.
1298 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1299 * for avoiding race in accounting. If true,
1300 * pc->mem_cgroup may be overwritten.
1302 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1303 * under hierarchy of moving cgroups. This is for
1304 * waiting at hith-memory prressure caused by "move".
1307 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1309 VM_BUG_ON(!rcu_read_lock_held());
1310 return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1313 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1315 struct mem_cgroup *from;
1316 struct mem_cgroup *to;
1319 * Unlike task_move routines, we access mc.to, mc.from not under
1320 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1322 spin_lock(&mc.lock);
1327 if (from == mem || to == mem
1328 || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1329 || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1332 spin_unlock(&mc.lock);
1336 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1338 if (mc.moving_task && current != mc.moving_task) {
1339 if (mem_cgroup_under_move(mem)) {
1341 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1342 /* moving charge context might have finished. */
1345 finish_wait(&mc.waitq, &wait);
1353 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1354 * @memcg: The memory cgroup that went over limit
1355 * @p: Task that is going to be killed
1357 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1360 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1362 struct cgroup *task_cgrp;
1363 struct cgroup *mem_cgrp;
1365 * Need a buffer in BSS, can't rely on allocations. The code relies
1366 * on the assumption that OOM is serialized for memory controller.
1367 * If this assumption is broken, revisit this code.
1369 static char memcg_name[PATH_MAX];
1378 mem_cgrp = memcg->css.cgroup;
1379 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1381 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1384 * Unfortunately, we are unable to convert to a useful name
1385 * But we'll still print out the usage information
1392 printk(KERN_INFO "Task in %s killed", memcg_name);
1395 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1403 * Continues from above, so we don't need an KERN_ level
1405 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1408 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1409 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1410 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1411 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1412 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1414 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1415 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1416 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1420 * This function returns the number of memcg under hierarchy tree. Returns
1421 * 1(self count) if no children.
1423 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1426 struct mem_cgroup *iter;
1428 for_each_mem_cgroup_tree(iter, mem)
1434 * Return the memory (and swap, if configured) limit for a memcg.
1436 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1441 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1442 limit += total_swap_pages << PAGE_SHIFT;
1444 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1446 * If memsw is finite and limits the amount of swap space available
1447 * to this memcg, return that limit.
1449 return min(limit, memsw);
1453 * Visit the first child (need not be the first child as per the ordering
1454 * of the cgroup list, since we track last_scanned_child) of @mem and use
1455 * that to reclaim free pages from.
1457 static struct mem_cgroup *
1458 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1460 struct mem_cgroup *ret = NULL;
1461 struct cgroup_subsys_state *css;
1464 if (!root_mem->use_hierarchy) {
1465 css_get(&root_mem->css);
1471 nextid = root_mem->last_scanned_child + 1;
1472 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1474 if (css && css_tryget(css))
1475 ret = container_of(css, struct mem_cgroup, css);
1478 /* Updates scanning parameter */
1480 /* this means start scan from ID:1 */
1481 root_mem->last_scanned_child = 0;
1483 root_mem->last_scanned_child = found;
1490 * test_mem_cgroup_node_reclaimable
1491 * @mem: the target memcg
1492 * @nid: the node ID to be checked.
1493 * @noswap : specify true here if the user wants flle only information.
1495 * This function returns whether the specified memcg contains any
1496 * reclaimable pages on a node. Returns true if there are any reclaimable
1497 * pages in the node.
1499 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *mem,
1500 int nid, bool noswap)
1502 if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_FILE))
1504 if (noswap || !total_swap_pages)
1506 if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_ANON))
1511 #if MAX_NUMNODES > 1
1514 * Always updating the nodemask is not very good - even if we have an empty
1515 * list or the wrong list here, we can start from some node and traverse all
1516 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1519 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
1523 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1524 * pagein/pageout changes since the last update.
1526 if (!atomic_read(&mem->numainfo_events))
1528 if (atomic_inc_return(&mem->numainfo_updating) > 1)
1531 /* make a nodemask where this memcg uses memory from */
1532 mem->scan_nodes = node_states[N_HIGH_MEMORY];
1534 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1536 if (!test_mem_cgroup_node_reclaimable(mem, nid, false))
1537 node_clear(nid, mem->scan_nodes);
1540 atomic_set(&mem->numainfo_events, 0);
1541 atomic_set(&mem->numainfo_updating, 0);
1545 * Selecting a node where we start reclaim from. Because what we need is just
1546 * reducing usage counter, start from anywhere is O,K. Considering
1547 * memory reclaim from current node, there are pros. and cons.
1549 * Freeing memory from current node means freeing memory from a node which
1550 * we'll use or we've used. So, it may make LRU bad. And if several threads
1551 * hit limits, it will see a contention on a node. But freeing from remote
1552 * node means more costs for memory reclaim because of memory latency.
1554 * Now, we use round-robin. Better algorithm is welcomed.
1556 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1560 mem_cgroup_may_update_nodemask(mem);
1561 node = mem->last_scanned_node;
1563 node = next_node(node, mem->scan_nodes);
1564 if (node == MAX_NUMNODES)
1565 node = first_node(mem->scan_nodes);
1567 * We call this when we hit limit, not when pages are added to LRU.
1568 * No LRU may hold pages because all pages are UNEVICTABLE or
1569 * memcg is too small and all pages are not on LRU. In that case,
1570 * we use curret node.
1572 if (unlikely(node == MAX_NUMNODES))
1573 node = numa_node_id();
1575 mem->last_scanned_node = node;
1580 * Check all nodes whether it contains reclaimable pages or not.
1581 * For quick scan, we make use of scan_nodes. This will allow us to skip
1582 * unused nodes. But scan_nodes is lazily updated and may not cotain
1583 * enough new information. We need to do double check.
1585 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1590 * quick check...making use of scan_node.
1591 * We can skip unused nodes.
1593 if (!nodes_empty(mem->scan_nodes)) {
1594 for (nid = first_node(mem->scan_nodes);
1596 nid = next_node(nid, mem->scan_nodes)) {
1598 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1603 * Check rest of nodes.
1605 for_each_node_state(nid, N_HIGH_MEMORY) {
1606 if (node_isset(nid, mem->scan_nodes))
1608 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1615 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1620 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1622 return test_mem_cgroup_node_reclaimable(mem, 0, noswap);
1627 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1628 * we reclaimed from, so that we don't end up penalizing one child extensively
1629 * based on its position in the children list.
1631 * root_mem is the original ancestor that we've been reclaim from.
1633 * We give up and return to the caller when we visit root_mem twice.
1634 * (other groups can be removed while we're walking....)
1636 * If shrink==true, for avoiding to free too much, this returns immedieately.
1638 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1641 unsigned long reclaim_options,
1642 unsigned long *total_scanned)
1644 struct mem_cgroup *victim;
1647 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1648 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1649 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1650 unsigned long excess;
1651 unsigned long nr_scanned;
1653 excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1655 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1656 if (!check_soft && root_mem->memsw_is_minimum)
1660 victim = mem_cgroup_select_victim(root_mem);
1661 if (victim == root_mem) {
1664 * We are not draining per cpu cached charges during
1665 * soft limit reclaim because global reclaim doesn't
1666 * care about charges. It tries to free some memory and
1667 * charges will not give any.
1669 if (!check_soft && loop >= 1)
1670 drain_all_stock_async(root_mem);
1673 * If we have not been able to reclaim
1674 * anything, it might because there are
1675 * no reclaimable pages under this hierarchy
1677 if (!check_soft || !total) {
1678 css_put(&victim->css);
1682 * We want to do more targeted reclaim.
1683 * excess >> 2 is not to excessive so as to
1684 * reclaim too much, nor too less that we keep
1685 * coming back to reclaim from this cgroup
1687 if (total >= (excess >> 2) ||
1688 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1689 css_put(&victim->css);
1694 if (!mem_cgroup_reclaimable(victim, noswap)) {
1695 /* this cgroup's local usage == 0 */
1696 css_put(&victim->css);
1699 /* we use swappiness of local cgroup */
1701 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1702 noswap, zone, &nr_scanned);
1703 *total_scanned += nr_scanned;
1705 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1707 css_put(&victim->css);
1709 * At shrinking usage, we can't check we should stop here or
1710 * reclaim more. It's depends on callers. last_scanned_child
1711 * will work enough for keeping fairness under tree.
1717 if (!res_counter_soft_limit_excess(&root_mem->res))
1719 } else if (mem_cgroup_margin(root_mem))
1726 * Check OOM-Killer is already running under our hierarchy.
1727 * If someone is running, return false.
1728 * Has to be called with memcg_oom_mutex
1730 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1732 int lock_count = -1;
1733 struct mem_cgroup *iter, *failed = NULL;
1736 for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1737 bool locked = iter->oom_lock;
1739 iter->oom_lock = true;
1740 if (lock_count == -1)
1741 lock_count = iter->oom_lock;
1742 else if (lock_count != locked) {
1744 * this subtree of our hierarchy is already locked
1745 * so we cannot give a lock.
1757 * OK, we failed to lock the whole subtree so we have to clean up
1758 * what we set up to the failing subtree
1761 for_each_mem_cgroup_tree_cond(iter, mem, cond) {
1762 if (iter == failed) {
1766 iter->oom_lock = false;
1773 * Has to be called with memcg_oom_mutex
1775 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1777 struct mem_cgroup *iter;
1779 for_each_mem_cgroup_tree(iter, mem)
1780 iter->oom_lock = false;
1784 static void mem_cgroup_mark_under_oom(struct mem_cgroup *mem)
1786 struct mem_cgroup *iter;
1788 for_each_mem_cgroup_tree(iter, mem)
1789 atomic_inc(&iter->under_oom);
1792 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *mem)
1794 struct mem_cgroup *iter;
1797 * When a new child is created while the hierarchy is under oom,
1798 * mem_cgroup_oom_lock() may not be called. We have to use
1799 * atomic_add_unless() here.
1801 for_each_mem_cgroup_tree(iter, mem)
1802 atomic_add_unless(&iter->under_oom, -1, 0);
1805 static DEFINE_MUTEX(memcg_oom_mutex);
1806 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1808 struct oom_wait_info {
1809 struct mem_cgroup *mem;
1813 static int memcg_oom_wake_function(wait_queue_t *wait,
1814 unsigned mode, int sync, void *arg)
1816 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1817 struct oom_wait_info *oom_wait_info;
1819 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1821 if (oom_wait_info->mem == wake_mem)
1823 /* if no hierarchy, no match */
1824 if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1827 * Both of oom_wait_info->mem and wake_mem are stable under us.
1828 * Then we can use css_is_ancestor without taking care of RCU.
1830 if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1831 !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1835 return autoremove_wake_function(wait, mode, sync, arg);
1838 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1840 /* for filtering, pass "mem" as argument. */
1841 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1844 static void memcg_oom_recover(struct mem_cgroup *mem)
1846 if (mem && atomic_read(&mem->under_oom))
1847 memcg_wakeup_oom(mem);
1851 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1853 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1855 struct oom_wait_info owait;
1856 bool locked, need_to_kill;
1859 owait.wait.flags = 0;
1860 owait.wait.func = memcg_oom_wake_function;
1861 owait.wait.private = current;
1862 INIT_LIST_HEAD(&owait.wait.task_list);
1863 need_to_kill = true;
1864 mem_cgroup_mark_under_oom(mem);
1866 /* At first, try to OOM lock hierarchy under mem.*/
1867 mutex_lock(&memcg_oom_mutex);
1868 locked = mem_cgroup_oom_lock(mem);
1870 * Even if signal_pending(), we can't quit charge() loop without
1871 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1872 * under OOM is always welcomed, use TASK_KILLABLE here.
1874 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1875 if (!locked || mem->oom_kill_disable)
1876 need_to_kill = false;
1878 mem_cgroup_oom_notify(mem);
1879 mutex_unlock(&memcg_oom_mutex);
1882 finish_wait(&memcg_oom_waitq, &owait.wait);
1883 mem_cgroup_out_of_memory(mem, mask);
1886 finish_wait(&memcg_oom_waitq, &owait.wait);
1888 mutex_lock(&memcg_oom_mutex);
1890 mem_cgroup_oom_unlock(mem);
1891 memcg_wakeup_oom(mem);
1892 mutex_unlock(&memcg_oom_mutex);
1894 mem_cgroup_unmark_under_oom(mem);
1896 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1898 /* Give chance to dying process */
1899 schedule_timeout(1);
1904 * Currently used to update mapped file statistics, but the routine can be
1905 * generalized to update other statistics as well.
1907 * Notes: Race condition
1909 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1910 * it tends to be costly. But considering some conditions, we doesn't need
1911 * to do so _always_.
1913 * Considering "charge", lock_page_cgroup() is not required because all
1914 * file-stat operations happen after a page is attached to radix-tree. There
1915 * are no race with "charge".
1917 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1918 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1919 * if there are race with "uncharge". Statistics itself is properly handled
1922 * Considering "move", this is an only case we see a race. To make the race
1923 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1924 * possibility of race condition. If there is, we take a lock.
1927 void mem_cgroup_update_page_stat(struct page *page,
1928 enum mem_cgroup_page_stat_item idx, int val)
1930 struct mem_cgroup *mem;
1931 struct page_cgroup *pc = lookup_page_cgroup(page);
1932 bool need_unlock = false;
1933 unsigned long uninitialized_var(flags);
1939 mem = pc->mem_cgroup;
1940 if (unlikely(!mem || !PageCgroupUsed(pc)))
1942 /* pc->mem_cgroup is unstable ? */
1943 if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1944 /* take a lock against to access pc->mem_cgroup */
1945 move_lock_page_cgroup(pc, &flags);
1947 mem = pc->mem_cgroup;
1948 if (!mem || !PageCgroupUsed(pc))
1953 case MEMCG_NR_FILE_MAPPED:
1955 SetPageCgroupFileMapped(pc);
1956 else if (!page_mapped(page))
1957 ClearPageCgroupFileMapped(pc);
1958 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1964 this_cpu_add(mem->stat->count[idx], val);
1967 if (unlikely(need_unlock))
1968 move_unlock_page_cgroup(pc, &flags);
1972 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1975 * size of first charge trial. "32" comes from vmscan.c's magic value.
1976 * TODO: maybe necessary to use big numbers in big irons.
1978 #define CHARGE_BATCH 32U
1979 struct memcg_stock_pcp {
1980 struct mem_cgroup *cached; /* this never be root cgroup */
1981 unsigned int nr_pages;
1982 struct work_struct work;
1983 unsigned long flags;
1984 #define FLUSHING_CACHED_CHARGE (0)
1986 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1987 static DEFINE_MUTEX(percpu_charge_mutex);
1990 * Try to consume stocked charge on this cpu. If success, one page is consumed
1991 * from local stock and true is returned. If the stock is 0 or charges from a
1992 * cgroup which is not current target, returns false. This stock will be
1995 static bool consume_stock(struct mem_cgroup *mem)
1997 struct memcg_stock_pcp *stock;
2000 stock = &get_cpu_var(memcg_stock);
2001 if (mem == stock->cached && stock->nr_pages)
2003 else /* need to call res_counter_charge */
2005 put_cpu_var(memcg_stock);
2010 * Returns stocks cached in percpu to res_counter and reset cached information.
2012 static void drain_stock(struct memcg_stock_pcp *stock)
2014 struct mem_cgroup *old = stock->cached;
2016 if (stock->nr_pages) {
2017 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2019 res_counter_uncharge(&old->res, bytes);
2020 if (do_swap_account)
2021 res_counter_uncharge(&old->memsw, bytes);
2022 stock->nr_pages = 0;
2024 stock->cached = NULL;
2028 * This must be called under preempt disabled or must be called by
2029 * a thread which is pinned to local cpu.
2031 static void drain_local_stock(struct work_struct *dummy)
2033 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2035 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2039 * Cache charges(val) which is from res_counter, to local per_cpu area.
2040 * This will be consumed by consume_stock() function, later.
2042 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
2044 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2046 if (stock->cached != mem) { /* reset if necessary */
2048 stock->cached = mem;
2050 stock->nr_pages += nr_pages;
2051 put_cpu_var(memcg_stock);
2055 * Tries to drain stocked charges in other cpus. This function is asynchronous
2056 * and just put a work per cpu for draining localy on each cpu. Caller can
2057 * expects some charges will be back to res_counter later but cannot wait for
2060 static void drain_all_stock_async(struct mem_cgroup *root_mem)
2064 * If someone calls draining, avoid adding more kworker runs.
2066 if (!mutex_trylock(&percpu_charge_mutex))
2068 /* Notify other cpus that system-wide "drain" is running */
2071 * Get a hint for avoiding draining charges on the current cpu,
2072 * which must be exhausted by our charging. It is not required that
2073 * this be a precise check, so we use raw_smp_processor_id() instead of
2074 * getcpu()/putcpu().
2076 curcpu = raw_smp_processor_id();
2077 for_each_online_cpu(cpu) {
2078 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2079 struct mem_cgroup *mem;
2084 mem = stock->cached;
2087 if (mem != root_mem) {
2088 if (!root_mem->use_hierarchy)
2090 /* check whether "mem" is under tree of "root_mem" */
2091 if (!css_is_ancestor(&mem->css, &root_mem->css))
2094 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2095 schedule_work_on(cpu, &stock->work);
2098 mutex_unlock(&percpu_charge_mutex);
2099 /* We don't wait for flush_work */
2102 /* This is a synchronous drain interface. */
2103 static void drain_all_stock_sync(void)
2105 /* called when force_empty is called */
2106 mutex_lock(&percpu_charge_mutex);
2107 schedule_on_each_cpu(drain_local_stock);
2108 mutex_unlock(&percpu_charge_mutex);
2112 * This function drains percpu counter value from DEAD cpu and
2113 * move it to local cpu. Note that this function can be preempted.
2115 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2119 spin_lock(&mem->pcp_counter_lock);
2120 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2121 long x = per_cpu(mem->stat->count[i], cpu);
2123 per_cpu(mem->stat->count[i], cpu) = 0;
2124 mem->nocpu_base.count[i] += x;
2126 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2127 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2129 per_cpu(mem->stat->events[i], cpu) = 0;
2130 mem->nocpu_base.events[i] += x;
2132 /* need to clear ON_MOVE value, works as a kind of lock. */
2133 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2134 spin_unlock(&mem->pcp_counter_lock);
2137 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2139 int idx = MEM_CGROUP_ON_MOVE;
2141 spin_lock(&mem->pcp_counter_lock);
2142 per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2143 spin_unlock(&mem->pcp_counter_lock);
2146 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2147 unsigned long action,
2150 int cpu = (unsigned long)hcpu;
2151 struct memcg_stock_pcp *stock;
2152 struct mem_cgroup *iter;
2154 if ((action == CPU_ONLINE)) {
2155 for_each_mem_cgroup_all(iter)
2156 synchronize_mem_cgroup_on_move(iter, cpu);
2160 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2163 for_each_mem_cgroup_all(iter)
2164 mem_cgroup_drain_pcp_counter(iter, cpu);
2166 stock = &per_cpu(memcg_stock, cpu);
2172 /* See __mem_cgroup_try_charge() for details */
2174 CHARGE_OK, /* success */
2175 CHARGE_RETRY, /* need to retry but retry is not bad */
2176 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2177 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2178 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2181 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2182 unsigned int nr_pages, bool oom_check)
2184 unsigned long csize = nr_pages * PAGE_SIZE;
2185 struct mem_cgroup *mem_over_limit;
2186 struct res_counter *fail_res;
2187 unsigned long flags = 0;
2190 ret = res_counter_charge(&mem->res, csize, &fail_res);
2193 if (!do_swap_account)
2195 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2199 res_counter_uncharge(&mem->res, csize);
2200 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2201 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2203 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2205 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2206 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2208 * Never reclaim on behalf of optional batching, retry with a
2209 * single page instead.
2211 if (nr_pages == CHARGE_BATCH)
2212 return CHARGE_RETRY;
2214 if (!(gfp_mask & __GFP_WAIT))
2215 return CHARGE_WOULDBLOCK;
2217 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2218 gfp_mask, flags, NULL);
2219 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2220 return CHARGE_RETRY;
2222 * Even though the limit is exceeded at this point, reclaim
2223 * may have been able to free some pages. Retry the charge
2224 * before killing the task.
2226 * Only for regular pages, though: huge pages are rather
2227 * unlikely to succeed so close to the limit, and we fall back
2228 * to regular pages anyway in case of failure.
2230 if (nr_pages == 1 && ret)
2231 return CHARGE_RETRY;
2234 * At task move, charge accounts can be doubly counted. So, it's
2235 * better to wait until the end of task_move if something is going on.
2237 if (mem_cgroup_wait_acct_move(mem_over_limit))
2238 return CHARGE_RETRY;
2240 /* If we don't need to call oom-killer at el, return immediately */
2242 return CHARGE_NOMEM;
2244 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2245 return CHARGE_OOM_DIE;
2247 return CHARGE_RETRY;
2251 * Unlike exported interface, "oom" parameter is added. if oom==true,
2252 * oom-killer can be invoked.
2254 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2256 unsigned int nr_pages,
2257 struct mem_cgroup **memcg,
2260 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2261 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2262 struct mem_cgroup *mem = NULL;
2266 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2267 * in system level. So, allow to go ahead dying process in addition to
2270 if (unlikely(test_thread_flag(TIF_MEMDIE)
2271 || fatal_signal_pending(current)))
2275 * We always charge the cgroup the mm_struct belongs to.
2276 * The mm_struct's mem_cgroup changes on task migration if the
2277 * thread group leader migrates. It's possible that mm is not
2278 * set, if so charge the init_mm (happens for pagecache usage).
2283 if (*memcg) { /* css should be a valid one */
2285 VM_BUG_ON(css_is_removed(&mem->css));
2286 if (mem_cgroup_is_root(mem))
2288 if (nr_pages == 1 && consume_stock(mem))
2292 struct task_struct *p;
2295 p = rcu_dereference(mm->owner);
2297 * Because we don't have task_lock(), "p" can exit.
2298 * In that case, "mem" can point to root or p can be NULL with
2299 * race with swapoff. Then, we have small risk of mis-accouning.
2300 * But such kind of mis-account by race always happens because
2301 * we don't have cgroup_mutex(). It's overkill and we allo that
2303 * (*) swapoff at el will charge against mm-struct not against
2304 * task-struct. So, mm->owner can be NULL.
2306 mem = mem_cgroup_from_task(p);
2307 if (!mem || mem_cgroup_is_root(mem)) {
2311 if (nr_pages == 1 && consume_stock(mem)) {
2313 * It seems dagerous to access memcg without css_get().
2314 * But considering how consume_stok works, it's not
2315 * necessary. If consume_stock success, some charges
2316 * from this memcg are cached on this cpu. So, we
2317 * don't need to call css_get()/css_tryget() before
2318 * calling consume_stock().
2323 /* after here, we may be blocked. we need to get refcnt */
2324 if (!css_tryget(&mem->css)) {
2334 /* If killed, bypass charge */
2335 if (fatal_signal_pending(current)) {
2341 if (oom && !nr_oom_retries) {
2343 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2346 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2350 case CHARGE_RETRY: /* not in OOM situation but retry */
2355 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2358 case CHARGE_NOMEM: /* OOM routine works */
2363 /* If oom, we never return -ENOMEM */
2366 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2370 } while (ret != CHARGE_OK);
2372 if (batch > nr_pages)
2373 refill_stock(mem, batch - nr_pages);
2387 * Somemtimes we have to undo a charge we got by try_charge().
2388 * This function is for that and do uncharge, put css's refcnt.
2389 * gotten by try_charge().
2391 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2392 unsigned int nr_pages)
2394 if (!mem_cgroup_is_root(mem)) {
2395 unsigned long bytes = nr_pages * PAGE_SIZE;
2397 res_counter_uncharge(&mem->res, bytes);
2398 if (do_swap_account)
2399 res_counter_uncharge(&mem->memsw, bytes);
2404 * A helper function to get mem_cgroup from ID. must be called under
2405 * rcu_read_lock(). The caller must check css_is_removed() or some if
2406 * it's concern. (dropping refcnt from swap can be called against removed
2409 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2411 struct cgroup_subsys_state *css;
2413 /* ID 0 is unused ID */
2416 css = css_lookup(&mem_cgroup_subsys, id);
2419 return container_of(css, struct mem_cgroup, css);
2422 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2424 struct mem_cgroup *mem = NULL;
2425 struct page_cgroup *pc;
2429 VM_BUG_ON(!PageLocked(page));
2431 pc = lookup_page_cgroup(page);
2432 lock_page_cgroup(pc);
2433 if (PageCgroupUsed(pc)) {
2434 mem = pc->mem_cgroup;
2435 if (mem && !css_tryget(&mem->css))
2437 } else if (PageSwapCache(page)) {
2438 ent.val = page_private(page);
2439 id = lookup_swap_cgroup(ent);
2441 mem = mem_cgroup_lookup(id);
2442 if (mem && !css_tryget(&mem->css))
2446 unlock_page_cgroup(pc);
2450 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2452 unsigned int nr_pages,
2453 struct page_cgroup *pc,
2454 enum charge_type ctype)
2456 lock_page_cgroup(pc);
2457 if (unlikely(PageCgroupUsed(pc))) {
2458 unlock_page_cgroup(pc);
2459 __mem_cgroup_cancel_charge(mem, nr_pages);
2463 * we don't need page_cgroup_lock about tail pages, becase they are not
2464 * accessed by any other context at this point.
2466 pc->mem_cgroup = mem;
2468 * We access a page_cgroup asynchronously without lock_page_cgroup().
2469 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2470 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2471 * before USED bit, we need memory barrier here.
2472 * See mem_cgroup_add_lru_list(), etc.
2476 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2477 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2478 SetPageCgroupCache(pc);
2479 SetPageCgroupUsed(pc);
2481 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2482 ClearPageCgroupCache(pc);
2483 SetPageCgroupUsed(pc);
2489 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2490 unlock_page_cgroup(pc);
2492 * "charge_statistics" updated event counter. Then, check it.
2493 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2494 * if they exceeds softlimit.
2496 memcg_check_events(mem, page);
2499 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2501 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2502 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2504 * Because tail pages are not marked as "used", set it. We're under
2505 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2507 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2509 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2510 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2511 unsigned long flags;
2513 if (mem_cgroup_disabled())
2516 * We have no races with charge/uncharge but will have races with
2517 * page state accounting.
2519 move_lock_page_cgroup(head_pc, &flags);
2521 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2522 smp_wmb(); /* see __commit_charge() */
2523 if (PageCgroupAcctLRU(head_pc)) {
2525 struct mem_cgroup_per_zone *mz;
2528 * LRU flags cannot be copied because we need to add tail
2529 *.page to LRU by generic call and our hook will be called.
2530 * We hold lru_lock, then, reduce counter directly.
2532 lru = page_lru(head);
2533 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2534 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2536 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2537 move_unlock_page_cgroup(head_pc, &flags);
2542 * mem_cgroup_move_account - move account of the page
2544 * @nr_pages: number of regular pages (>1 for huge pages)
2545 * @pc: page_cgroup of the page.
2546 * @from: mem_cgroup which the page is moved from.
2547 * @to: mem_cgroup which the page is moved to. @from != @to.
2548 * @uncharge: whether we should call uncharge and css_put against @from.
2550 * The caller must confirm following.
2551 * - page is not on LRU (isolate_page() is useful.)
2552 * - compound_lock is held when nr_pages > 1
2554 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2555 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2556 * true, this function does "uncharge" from old cgroup, but it doesn't if
2557 * @uncharge is false, so a caller should do "uncharge".
2559 static int mem_cgroup_move_account(struct page *page,
2560 unsigned int nr_pages,
2561 struct page_cgroup *pc,
2562 struct mem_cgroup *from,
2563 struct mem_cgroup *to,
2566 unsigned long flags;
2569 VM_BUG_ON(from == to);
2570 VM_BUG_ON(PageLRU(page));
2572 * The page is isolated from LRU. So, collapse function
2573 * will not handle this page. But page splitting can happen.
2574 * Do this check under compound_page_lock(). The caller should
2578 if (nr_pages > 1 && !PageTransHuge(page))
2581 lock_page_cgroup(pc);
2584 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2587 move_lock_page_cgroup(pc, &flags);
2589 if (PageCgroupFileMapped(pc)) {
2590 /* Update mapped_file data for mem_cgroup */
2592 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2593 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2596 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2598 /* This is not "cancel", but cancel_charge does all we need. */
2599 __mem_cgroup_cancel_charge(from, nr_pages);
2601 /* caller should have done css_get */
2602 pc->mem_cgroup = to;
2603 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2605 * We charges against "to" which may not have any tasks. Then, "to"
2606 * can be under rmdir(). But in current implementation, caller of
2607 * this function is just force_empty() and move charge, so it's
2608 * guaranteed that "to" is never removed. So, we don't check rmdir
2611 move_unlock_page_cgroup(pc, &flags);
2614 unlock_page_cgroup(pc);
2618 memcg_check_events(to, page);
2619 memcg_check_events(from, page);
2625 * move charges to its parent.
2628 static int mem_cgroup_move_parent(struct page *page,
2629 struct page_cgroup *pc,
2630 struct mem_cgroup *child,
2633 struct cgroup *cg = child->css.cgroup;
2634 struct cgroup *pcg = cg->parent;
2635 struct mem_cgroup *parent;
2636 unsigned int nr_pages;
2637 unsigned long uninitialized_var(flags);
2645 if (!get_page_unless_zero(page))
2647 if (isolate_lru_page(page))
2650 nr_pages = hpage_nr_pages(page);
2652 parent = mem_cgroup_from_cont(pcg);
2653 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2658 flags = compound_lock_irqsave(page);
2660 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2662 __mem_cgroup_cancel_charge(parent, nr_pages);
2665 compound_unlock_irqrestore(page, flags);
2667 putback_lru_page(page);
2675 * Charge the memory controller for page usage.
2677 * 0 if the charge was successful
2678 * < 0 if the cgroup is over its limit
2680 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2681 gfp_t gfp_mask, enum charge_type ctype)
2683 struct mem_cgroup *mem = NULL;
2684 unsigned int nr_pages = 1;
2685 struct page_cgroup *pc;
2689 if (PageTransHuge(page)) {
2690 nr_pages <<= compound_order(page);
2691 VM_BUG_ON(!PageTransHuge(page));
2693 * Never OOM-kill a process for a huge page. The
2694 * fault handler will fall back to regular pages.
2699 pc = lookup_page_cgroup(page);
2700 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2702 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2706 __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2710 int mem_cgroup_newpage_charge(struct page *page,
2711 struct mm_struct *mm, gfp_t gfp_mask)
2713 if (mem_cgroup_disabled())
2716 * If already mapped, we don't have to account.
2717 * If page cache, page->mapping has address_space.
2718 * But page->mapping may have out-of-use anon_vma pointer,
2719 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2722 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2726 return mem_cgroup_charge_common(page, mm, gfp_mask,
2727 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2731 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2732 enum charge_type ctype);
2735 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2736 enum charge_type ctype)
2738 struct page_cgroup *pc = lookup_page_cgroup(page);
2740 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2741 * is already on LRU. It means the page may on some other page_cgroup's
2742 * LRU. Take care of it.
2744 mem_cgroup_lru_del_before_commit(page);
2745 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2746 mem_cgroup_lru_add_after_commit(page);
2750 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2753 struct mem_cgroup *mem = NULL;
2756 if (mem_cgroup_disabled())
2758 if (PageCompound(page))
2761 * Corner case handling. This is called from add_to_page_cache()
2762 * in usual. But some FS (shmem) precharges this page before calling it
2763 * and call add_to_page_cache() with GFP_NOWAIT.
2765 * For GFP_NOWAIT case, the page may be pre-charged before calling
2766 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2767 * charge twice. (It works but has to pay a bit larger cost.)
2768 * And when the page is SwapCache, it should take swap information
2769 * into account. This is under lock_page() now.
2771 if (!(gfp_mask & __GFP_WAIT)) {
2772 struct page_cgroup *pc;
2774 pc = lookup_page_cgroup(page);
2777 lock_page_cgroup(pc);
2778 if (PageCgroupUsed(pc)) {
2779 unlock_page_cgroup(pc);
2782 unlock_page_cgroup(pc);
2788 if (page_is_file_cache(page)) {
2789 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2794 * FUSE reuses pages without going through the final
2795 * put that would remove them from the LRU list, make
2796 * sure that they get relinked properly.
2798 __mem_cgroup_commit_charge_lrucare(page, mem,
2799 MEM_CGROUP_CHARGE_TYPE_CACHE);
2803 if (PageSwapCache(page)) {
2804 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2806 __mem_cgroup_commit_charge_swapin(page, mem,
2807 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2809 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2810 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2816 * While swap-in, try_charge -> commit or cancel, the page is locked.
2817 * And when try_charge() successfully returns, one refcnt to memcg without
2818 * struct page_cgroup is acquired. This refcnt will be consumed by
2819 * "commit()" or removed by "cancel()"
2821 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2823 gfp_t mask, struct mem_cgroup **ptr)
2825 struct mem_cgroup *mem;
2830 if (mem_cgroup_disabled())
2833 if (!do_swap_account)
2836 * A racing thread's fault, or swapoff, may have already updated
2837 * the pte, and even removed page from swap cache: in those cases
2838 * do_swap_page()'s pte_same() test will fail; but there's also a
2839 * KSM case which does need to charge the page.
2841 if (!PageSwapCache(page))
2843 mem = try_get_mem_cgroup_from_page(page);
2847 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2853 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2857 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2858 enum charge_type ctype)
2860 if (mem_cgroup_disabled())
2864 cgroup_exclude_rmdir(&ptr->css);
2866 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2868 * Now swap is on-memory. This means this page may be
2869 * counted both as mem and swap....double count.
2870 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2871 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2872 * may call delete_from_swap_cache() before reach here.
2874 if (do_swap_account && PageSwapCache(page)) {
2875 swp_entry_t ent = {.val = page_private(page)};
2877 struct mem_cgroup *memcg;
2879 id = swap_cgroup_record(ent, 0);
2881 memcg = mem_cgroup_lookup(id);
2884 * This recorded memcg can be obsolete one. So, avoid
2885 * calling css_tryget
2887 if (!mem_cgroup_is_root(memcg))
2888 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2889 mem_cgroup_swap_statistics(memcg, false);
2890 mem_cgroup_put(memcg);
2895 * At swapin, we may charge account against cgroup which has no tasks.
2896 * So, rmdir()->pre_destroy() can be called while we do this charge.
2897 * In that case, we need to call pre_destroy() again. check it here.
2899 cgroup_release_and_wakeup_rmdir(&ptr->css);
2902 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2904 __mem_cgroup_commit_charge_swapin(page, ptr,
2905 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2908 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2910 if (mem_cgroup_disabled())
2914 __mem_cgroup_cancel_charge(mem, 1);
2917 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
2918 unsigned int nr_pages,
2919 const enum charge_type ctype)
2921 struct memcg_batch_info *batch = NULL;
2922 bool uncharge_memsw = true;
2924 /* If swapout, usage of swap doesn't decrease */
2925 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2926 uncharge_memsw = false;
2928 batch = ¤t->memcg_batch;
2930 * In usual, we do css_get() when we remember memcg pointer.
2931 * But in this case, we keep res->usage until end of a series of
2932 * uncharges. Then, it's ok to ignore memcg's refcnt.
2937 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2938 * In those cases, all pages freed continuously can be expected to be in
2939 * the same cgroup and we have chance to coalesce uncharges.
2940 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2941 * because we want to do uncharge as soon as possible.
2944 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2945 goto direct_uncharge;
2948 goto direct_uncharge;
2951 * In typical case, batch->memcg == mem. This means we can
2952 * merge a series of uncharges to an uncharge of res_counter.
2953 * If not, we uncharge res_counter ony by one.
2955 if (batch->memcg != mem)
2956 goto direct_uncharge;
2957 /* remember freed charge and uncharge it later */
2960 batch->memsw_nr_pages++;
2963 res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
2965 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
2966 if (unlikely(batch->memcg != mem))
2967 memcg_oom_recover(mem);
2972 * uncharge if !page_mapped(page)
2974 static struct mem_cgroup *
2975 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2977 struct mem_cgroup *mem = NULL;
2978 unsigned int nr_pages = 1;
2979 struct page_cgroup *pc;
2981 if (mem_cgroup_disabled())
2984 if (PageSwapCache(page))
2987 if (PageTransHuge(page)) {
2988 nr_pages <<= compound_order(page);
2989 VM_BUG_ON(!PageTransHuge(page));
2992 * Check if our page_cgroup is valid
2994 pc = lookup_page_cgroup(page);
2995 if (unlikely(!pc || !PageCgroupUsed(pc)))
2998 lock_page_cgroup(pc);
3000 mem = pc->mem_cgroup;
3002 if (!PageCgroupUsed(pc))
3006 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3007 case MEM_CGROUP_CHARGE_TYPE_DROP:
3008 /* See mem_cgroup_prepare_migration() */
3009 if (page_mapped(page) || PageCgroupMigration(pc))
3012 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3013 if (!PageAnon(page)) { /* Shared memory */
3014 if (page->mapping && !page_is_file_cache(page))
3016 } else if (page_mapped(page)) /* Anon */
3023 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
3025 ClearPageCgroupUsed(pc);
3027 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3028 * freed from LRU. This is safe because uncharged page is expected not
3029 * to be reused (freed soon). Exception is SwapCache, it's handled by
3030 * special functions.
3033 unlock_page_cgroup(pc);
3035 * even after unlock, we have mem->res.usage here and this memcg
3036 * will never be freed.
3038 memcg_check_events(mem, page);
3039 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3040 mem_cgroup_swap_statistics(mem, true);
3041 mem_cgroup_get(mem);
3043 if (!mem_cgroup_is_root(mem))
3044 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
3049 unlock_page_cgroup(pc);
3053 void mem_cgroup_uncharge_page(struct page *page)
3056 if (page_mapped(page))
3058 if (page->mapping && !PageAnon(page))
3060 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3063 void mem_cgroup_uncharge_cache_page(struct page *page)
3065 VM_BUG_ON(page_mapped(page));
3066 VM_BUG_ON(page->mapping);
3067 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3071 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3072 * In that cases, pages are freed continuously and we can expect pages
3073 * are in the same memcg. All these calls itself limits the number of
3074 * pages freed at once, then uncharge_start/end() is called properly.
3075 * This may be called prural(2) times in a context,
3078 void mem_cgroup_uncharge_start(void)
3080 current->memcg_batch.do_batch++;
3081 /* We can do nest. */
3082 if (current->memcg_batch.do_batch == 1) {
3083 current->memcg_batch.memcg = NULL;
3084 current->memcg_batch.nr_pages = 0;
3085 current->memcg_batch.memsw_nr_pages = 0;
3089 void mem_cgroup_uncharge_end(void)
3091 struct memcg_batch_info *batch = ¤t->memcg_batch;
3093 if (!batch->do_batch)
3097 if (batch->do_batch) /* If stacked, do nothing. */
3103 * This "batch->memcg" is valid without any css_get/put etc...
3104 * bacause we hide charges behind us.
3106 if (batch->nr_pages)
3107 res_counter_uncharge(&batch->memcg->res,
3108 batch->nr_pages * PAGE_SIZE);
3109 if (batch->memsw_nr_pages)
3110 res_counter_uncharge(&batch->memcg->memsw,
3111 batch->memsw_nr_pages * PAGE_SIZE);
3112 memcg_oom_recover(batch->memcg);
3113 /* forget this pointer (for sanity check) */
3114 batch->memcg = NULL;
3119 * called after __delete_from_swap_cache() and drop "page" account.
3120 * memcg information is recorded to swap_cgroup of "ent"
3123 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3125 struct mem_cgroup *memcg;
3126 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3128 if (!swapout) /* this was a swap cache but the swap is unused ! */
3129 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3131 memcg = __mem_cgroup_uncharge_common(page, ctype);
3134 * record memcg information, if swapout && memcg != NULL,
3135 * mem_cgroup_get() was called in uncharge().
3137 if (do_swap_account && swapout && memcg)
3138 swap_cgroup_record(ent, css_id(&memcg->css));
3142 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3144 * called from swap_entry_free(). remove record in swap_cgroup and
3145 * uncharge "memsw" account.
3147 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3149 struct mem_cgroup *memcg;
3152 if (!do_swap_account)
3155 id = swap_cgroup_record(ent, 0);
3157 memcg = mem_cgroup_lookup(id);
3160 * We uncharge this because swap is freed.
3161 * This memcg can be obsolete one. We avoid calling css_tryget
3163 if (!mem_cgroup_is_root(memcg))
3164 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3165 mem_cgroup_swap_statistics(memcg, false);
3166 mem_cgroup_put(memcg);
3172 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3173 * @entry: swap entry to be moved
3174 * @from: mem_cgroup which the entry is moved from
3175 * @to: mem_cgroup which the entry is moved to
3176 * @need_fixup: whether we should fixup res_counters and refcounts.
3178 * It succeeds only when the swap_cgroup's record for this entry is the same
3179 * as the mem_cgroup's id of @from.
3181 * Returns 0 on success, -EINVAL on failure.
3183 * The caller must have charged to @to, IOW, called res_counter_charge() about
3184 * both res and memsw, and called css_get().
3186 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3187 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3189 unsigned short old_id, new_id;
3191 old_id = css_id(&from->css);
3192 new_id = css_id(&to->css);
3194 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3195 mem_cgroup_swap_statistics(from, false);
3196 mem_cgroup_swap_statistics(to, true);
3198 * This function is only called from task migration context now.
3199 * It postpones res_counter and refcount handling till the end
3200 * of task migration(mem_cgroup_clear_mc()) for performance
3201 * improvement. But we cannot postpone mem_cgroup_get(to)
3202 * because if the process that has been moved to @to does
3203 * swap-in, the refcount of @to might be decreased to 0.
3207 if (!mem_cgroup_is_root(from))
3208 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3209 mem_cgroup_put(from);
3211 * we charged both to->res and to->memsw, so we should
3214 if (!mem_cgroup_is_root(to))
3215 res_counter_uncharge(&to->res, PAGE_SIZE);
3222 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3223 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3230 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3233 int mem_cgroup_prepare_migration(struct page *page,
3234 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3236 struct mem_cgroup *mem = NULL;
3237 struct page_cgroup *pc;
3238 enum charge_type ctype;
3243 VM_BUG_ON(PageTransHuge(page));
3244 if (mem_cgroup_disabled())
3247 pc = lookup_page_cgroup(page);
3248 lock_page_cgroup(pc);
3249 if (PageCgroupUsed(pc)) {
3250 mem = pc->mem_cgroup;
3253 * At migrating an anonymous page, its mapcount goes down
3254 * to 0 and uncharge() will be called. But, even if it's fully
3255 * unmapped, migration may fail and this page has to be
3256 * charged again. We set MIGRATION flag here and delay uncharge
3257 * until end_migration() is called
3259 * Corner Case Thinking
3261 * When the old page was mapped as Anon and it's unmap-and-freed
3262 * while migration was ongoing.
3263 * If unmap finds the old page, uncharge() of it will be delayed
3264 * until end_migration(). If unmap finds a new page, it's
3265 * uncharged when it make mapcount to be 1->0. If unmap code
3266 * finds swap_migration_entry, the new page will not be mapped
3267 * and end_migration() will find it(mapcount==0).
3270 * When the old page was mapped but migraion fails, the kernel
3271 * remaps it. A charge for it is kept by MIGRATION flag even
3272 * if mapcount goes down to 0. We can do remap successfully
3273 * without charging it again.
3276 * The "old" page is under lock_page() until the end of
3277 * migration, so, the old page itself will not be swapped-out.
3278 * If the new page is swapped out before end_migraton, our
3279 * hook to usual swap-out path will catch the event.
3282 SetPageCgroupMigration(pc);
3284 unlock_page_cgroup(pc);
3286 * If the page is not charged at this point,
3293 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3294 css_put(&mem->css);/* drop extra refcnt */
3295 if (ret || *ptr == NULL) {
3296 if (PageAnon(page)) {
3297 lock_page_cgroup(pc);
3298 ClearPageCgroupMigration(pc);
3299 unlock_page_cgroup(pc);
3301 * The old page may be fully unmapped while we kept it.
3303 mem_cgroup_uncharge_page(page);
3308 * We charge new page before it's used/mapped. So, even if unlock_page()
3309 * is called before end_migration, we can catch all events on this new
3310 * page. In the case new page is migrated but not remapped, new page's
3311 * mapcount will be finally 0 and we call uncharge in end_migration().
3313 pc = lookup_page_cgroup(newpage);
3315 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3316 else if (page_is_file_cache(page))
3317 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3319 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3320 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3324 /* remove redundant charge if migration failed*/
3325 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3326 struct page *oldpage, struct page *newpage, bool migration_ok)
3328 struct page *used, *unused;
3329 struct page_cgroup *pc;
3333 /* blocks rmdir() */
3334 cgroup_exclude_rmdir(&mem->css);
3335 if (!migration_ok) {
3343 * We disallowed uncharge of pages under migration because mapcount
3344 * of the page goes down to zero, temporarly.
3345 * Clear the flag and check the page should be charged.
3347 pc = lookup_page_cgroup(oldpage);
3348 lock_page_cgroup(pc);
3349 ClearPageCgroupMigration(pc);
3350 unlock_page_cgroup(pc);
3352 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3355 * If a page is a file cache, radix-tree replacement is very atomic
3356 * and we can skip this check. When it was an Anon page, its mapcount
3357 * goes down to 0. But because we added MIGRATION flage, it's not
3358 * uncharged yet. There are several case but page->mapcount check
3359 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3360 * check. (see prepare_charge() also)
3363 mem_cgroup_uncharge_page(used);
3365 * At migration, we may charge account against cgroup which has no
3367 * So, rmdir()->pre_destroy() can be called while we do this charge.
3368 * In that case, we need to call pre_destroy() again. check it here.
3370 cgroup_release_and_wakeup_rmdir(&mem->css);
3374 * A call to try to shrink memory usage on charge failure at shmem's swapin.
3375 * Calling hierarchical_reclaim is not enough because we should update
3376 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3377 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3378 * not from the memcg which this page would be charged to.
3379 * try_charge_swapin does all of these works properly.
3381 int mem_cgroup_shmem_charge_fallback(struct page *page,
3382 struct mm_struct *mm,
3385 struct mem_cgroup *mem;
3388 if (mem_cgroup_disabled())
3391 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3393 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3398 #ifdef CONFIG_DEBUG_VM
3399 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3401 struct page_cgroup *pc;
3403 pc = lookup_page_cgroup(page);
3404 if (likely(pc) && PageCgroupUsed(pc))
3409 bool mem_cgroup_bad_page_check(struct page *page)
3411 if (mem_cgroup_disabled())
3414 return lookup_page_cgroup_used(page) != NULL;
3417 void mem_cgroup_print_bad_page(struct page *page)
3419 struct page_cgroup *pc;
3421 pc = lookup_page_cgroup_used(page);
3426 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3427 pc, pc->flags, pc->mem_cgroup);
3429 path = kmalloc(PATH_MAX, GFP_KERNEL);
3432 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3437 printk(KERN_CONT "(%s)\n",
3438 (ret < 0) ? "cannot get the path" : path);
3444 static DEFINE_MUTEX(set_limit_mutex);
3446 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3447 unsigned long long val)
3450 u64 memswlimit, memlimit;
3452 int children = mem_cgroup_count_children(memcg);
3453 u64 curusage, oldusage;
3457 * For keeping hierarchical_reclaim simple, how long we should retry
3458 * is depends on callers. We set our retry-count to be function
3459 * of # of children which we should visit in this loop.
3461 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3463 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3466 while (retry_count) {
3467 if (signal_pending(current)) {
3472 * Rather than hide all in some function, I do this in
3473 * open coded manner. You see what this really does.
3474 * We have to guarantee mem->res.limit < mem->memsw.limit.
3476 mutex_lock(&set_limit_mutex);
3477 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3478 if (memswlimit < val) {
3480 mutex_unlock(&set_limit_mutex);
3484 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3488 ret = res_counter_set_limit(&memcg->res, val);
3490 if (memswlimit == val)
3491 memcg->memsw_is_minimum = true;
3493 memcg->memsw_is_minimum = false;
3495 mutex_unlock(&set_limit_mutex);
3500 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3501 MEM_CGROUP_RECLAIM_SHRINK,
3503 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3504 /* Usage is reduced ? */
3505 if (curusage >= oldusage)
3508 oldusage = curusage;
3510 if (!ret && enlarge)
3511 memcg_oom_recover(memcg);
3516 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3517 unsigned long long val)
3520 u64 memlimit, memswlimit, oldusage, curusage;
3521 int children = mem_cgroup_count_children(memcg);
3525 /* see mem_cgroup_resize_res_limit */
3526 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3527 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3528 while (retry_count) {
3529 if (signal_pending(current)) {
3534 * Rather than hide all in some function, I do this in
3535 * open coded manner. You see what this really does.
3536 * We have to guarantee mem->res.limit < mem->memsw.limit.
3538 mutex_lock(&set_limit_mutex);
3539 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3540 if (memlimit > val) {
3542 mutex_unlock(&set_limit_mutex);
3545 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3546 if (memswlimit < val)
3548 ret = res_counter_set_limit(&memcg->memsw, val);
3550 if (memlimit == val)
3551 memcg->memsw_is_minimum = true;
3553 memcg->memsw_is_minimum = false;
3555 mutex_unlock(&set_limit_mutex);
3560 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3561 MEM_CGROUP_RECLAIM_NOSWAP |
3562 MEM_CGROUP_RECLAIM_SHRINK,
3564 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3565 /* Usage is reduced ? */
3566 if (curusage >= oldusage)
3569 oldusage = curusage;
3571 if (!ret && enlarge)
3572 memcg_oom_recover(memcg);
3576 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3578 unsigned long *total_scanned)
3580 unsigned long nr_reclaimed = 0;
3581 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3582 unsigned long reclaimed;
3584 struct mem_cgroup_tree_per_zone *mctz;
3585 unsigned long long excess;
3586 unsigned long nr_scanned;
3591 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3593 * This loop can run a while, specially if mem_cgroup's continuously
3594 * keep exceeding their soft limit and putting the system under
3601 mz = mem_cgroup_largest_soft_limit_node(mctz);
3606 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3608 MEM_CGROUP_RECLAIM_SOFT,
3610 nr_reclaimed += reclaimed;
3611 *total_scanned += nr_scanned;
3612 spin_lock(&mctz->lock);
3615 * If we failed to reclaim anything from this memory cgroup
3616 * it is time to move on to the next cgroup
3622 * Loop until we find yet another one.
3624 * By the time we get the soft_limit lock
3625 * again, someone might have aded the
3626 * group back on the RB tree. Iterate to
3627 * make sure we get a different mem.
3628 * mem_cgroup_largest_soft_limit_node returns
3629 * NULL if no other cgroup is present on
3633 __mem_cgroup_largest_soft_limit_node(mctz);
3635 css_put(&next_mz->mem->css);
3636 else /* next_mz == NULL or other memcg */
3640 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3641 excess = res_counter_soft_limit_excess(&mz->mem->res);
3643 * One school of thought says that we should not add
3644 * back the node to the tree if reclaim returns 0.
3645 * But our reclaim could return 0, simply because due
3646 * to priority we are exposing a smaller subset of
3647 * memory to reclaim from. Consider this as a longer
3650 /* If excess == 0, no tree ops */
3651 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3652 spin_unlock(&mctz->lock);
3653 css_put(&mz->mem->css);
3656 * Could not reclaim anything and there are no more
3657 * mem cgroups to try or we seem to be looping without
3658 * reclaiming anything.
3660 if (!nr_reclaimed &&
3662 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3664 } while (!nr_reclaimed);
3666 css_put(&next_mz->mem->css);
3667 return nr_reclaimed;
3671 * This routine traverse page_cgroup in given list and drop them all.
3672 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3674 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3675 int node, int zid, enum lru_list lru)
3678 struct mem_cgroup_per_zone *mz;
3679 struct page_cgroup *pc, *busy;
3680 unsigned long flags, loop;
3681 struct list_head *list;
3684 zone = &NODE_DATA(node)->node_zones[zid];
3685 mz = mem_cgroup_zoneinfo(mem, node, zid);
3686 list = &mz->lists[lru];
3688 loop = MEM_CGROUP_ZSTAT(mz, lru);
3689 /* give some margin against EBUSY etc...*/
3696 spin_lock_irqsave(&zone->lru_lock, flags);
3697 if (list_empty(list)) {
3698 spin_unlock_irqrestore(&zone->lru_lock, flags);
3701 pc = list_entry(list->prev, struct page_cgroup, lru);
3703 list_move(&pc->lru, list);
3705 spin_unlock_irqrestore(&zone->lru_lock, flags);
3708 spin_unlock_irqrestore(&zone->lru_lock, flags);
3710 page = lookup_cgroup_page(pc);
3712 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3716 if (ret == -EBUSY || ret == -EINVAL) {
3717 /* found lock contention or "pc" is obsolete. */
3724 if (!ret && !list_empty(list))
3730 * make mem_cgroup's charge to be 0 if there is no task.
3731 * This enables deleting this mem_cgroup.
3733 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3736 int node, zid, shrink;
3737 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3738 struct cgroup *cgrp = mem->css.cgroup;
3743 /* should free all ? */
3749 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3752 if (signal_pending(current))
3754 /* This is for making all *used* pages to be on LRU. */
3755 lru_add_drain_all();
3756 drain_all_stock_sync();
3758 mem_cgroup_start_move(mem);
3759 for_each_node_state(node, N_HIGH_MEMORY) {
3760 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3763 ret = mem_cgroup_force_empty_list(mem,
3772 mem_cgroup_end_move(mem);
3773 memcg_oom_recover(mem);
3774 /* it seems parent cgroup doesn't have enough mem */
3778 /* "ret" should also be checked to ensure all lists are empty. */
3779 } while (mem->res.usage > 0 || ret);
3785 /* returns EBUSY if there is a task or if we come here twice. */
3786 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3790 /* we call try-to-free pages for make this cgroup empty */
3791 lru_add_drain_all();
3792 /* try to free all pages in this cgroup */
3794 while (nr_retries && mem->res.usage > 0) {
3797 if (signal_pending(current)) {
3801 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3805 /* maybe some writeback is necessary */
3806 congestion_wait(BLK_RW_ASYNC, HZ/10);
3811 /* try move_account...there may be some *locked* pages. */
3815 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3817 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3821 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3823 return mem_cgroup_from_cont(cont)->use_hierarchy;
3826 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3830 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3831 struct cgroup *parent = cont->parent;
3832 struct mem_cgroup *parent_mem = NULL;
3835 parent_mem = mem_cgroup_from_cont(parent);
3839 * If parent's use_hierarchy is set, we can't make any modifications
3840 * in the child subtrees. If it is unset, then the change can
3841 * occur, provided the current cgroup has no children.
3843 * For the root cgroup, parent_mem is NULL, we allow value to be
3844 * set if there are no children.
3846 if ((!parent_mem || !parent_mem->use_hierarchy) &&
3847 (val == 1 || val == 0)) {
3848 if (list_empty(&cont->children))
3849 mem->use_hierarchy = val;
3860 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3861 enum mem_cgroup_stat_index idx)
3863 struct mem_cgroup *iter;
3866 /* Per-cpu values can be negative, use a signed accumulator */
3867 for_each_mem_cgroup_tree(iter, mem)
3868 val += mem_cgroup_read_stat(iter, idx);
3870 if (val < 0) /* race ? */
3875 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3879 if (!mem_cgroup_is_root(mem)) {
3881 return res_counter_read_u64(&mem->res, RES_USAGE);
3883 return res_counter_read_u64(&mem->memsw, RES_USAGE);
3886 val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
3887 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
3890 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3892 return val << PAGE_SHIFT;
3895 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3897 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3901 type = MEMFILE_TYPE(cft->private);
3902 name = MEMFILE_ATTR(cft->private);
3905 if (name == RES_USAGE)
3906 val = mem_cgroup_usage(mem, false);
3908 val = res_counter_read_u64(&mem->res, name);
3911 if (name == RES_USAGE)
3912 val = mem_cgroup_usage(mem, true);
3914 val = res_counter_read_u64(&mem->memsw, name);
3923 * The user of this function is...
3926 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3929 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3931 unsigned long long val;
3934 type = MEMFILE_TYPE(cft->private);
3935 name = MEMFILE_ATTR(cft->private);
3938 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3942 /* This function does all necessary parse...reuse it */
3943 ret = res_counter_memparse_write_strategy(buffer, &val);
3947 ret = mem_cgroup_resize_limit(memcg, val);
3949 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3951 case RES_SOFT_LIMIT:
3952 ret = res_counter_memparse_write_strategy(buffer, &val);
3956 * For memsw, soft limits are hard to implement in terms
3957 * of semantics, for now, we support soft limits for
3958 * control without swap
3961 ret = res_counter_set_soft_limit(&memcg->res, val);
3966 ret = -EINVAL; /* should be BUG() ? */
3972 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3973 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3975 struct cgroup *cgroup;
3976 unsigned long long min_limit, min_memsw_limit, tmp;
3978 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3979 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3980 cgroup = memcg->css.cgroup;
3981 if (!memcg->use_hierarchy)
3984 while (cgroup->parent) {
3985 cgroup = cgroup->parent;
3986 memcg = mem_cgroup_from_cont(cgroup);
3987 if (!memcg->use_hierarchy)
3989 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3990 min_limit = min(min_limit, tmp);
3991 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3992 min_memsw_limit = min(min_memsw_limit, tmp);
3995 *mem_limit = min_limit;
3996 *memsw_limit = min_memsw_limit;
4000 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4002 struct mem_cgroup *mem;
4005 mem = mem_cgroup_from_cont(cont);
4006 type = MEMFILE_TYPE(event);
4007 name = MEMFILE_ATTR(event);
4011 res_counter_reset_max(&mem->res);
4013 res_counter_reset_max(&mem->memsw);
4017 res_counter_reset_failcnt(&mem->res);
4019 res_counter_reset_failcnt(&mem->memsw);
4026 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4029 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4033 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4034 struct cftype *cft, u64 val)
4036 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4038 if (val >= (1 << NR_MOVE_TYPE))
4041 * We check this value several times in both in can_attach() and
4042 * attach(), so we need cgroup lock to prevent this value from being
4046 mem->move_charge_at_immigrate = val;
4052 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4053 struct cftype *cft, u64 val)
4060 /* For read statistics */
4078 struct mcs_total_stat {
4079 s64 stat[NR_MCS_STAT];
4085 } memcg_stat_strings[NR_MCS_STAT] = {
4086 {"cache", "total_cache"},
4087 {"rss", "total_rss"},
4088 {"mapped_file", "total_mapped_file"},
4089 {"pgpgin", "total_pgpgin"},
4090 {"pgpgout", "total_pgpgout"},
4091 {"swap", "total_swap"},
4092 {"pgfault", "total_pgfault"},
4093 {"pgmajfault", "total_pgmajfault"},
4094 {"inactive_anon", "total_inactive_anon"},
4095 {"active_anon", "total_active_anon"},
4096 {"inactive_file", "total_inactive_file"},
4097 {"active_file", "total_active_file"},
4098 {"unevictable", "total_unevictable"}
4103 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4108 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4109 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4110 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4111 s->stat[MCS_RSS] += val * PAGE_SIZE;
4112 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4113 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4114 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4115 s->stat[MCS_PGPGIN] += val;
4116 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4117 s->stat[MCS_PGPGOUT] += val;
4118 if (do_swap_account) {
4119 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4120 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4122 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4123 s->stat[MCS_PGFAULT] += val;
4124 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4125 s->stat[MCS_PGMAJFAULT] += val;
4128 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_ANON));
4129 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4130 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_ANON));
4131 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4132 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_FILE));
4133 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4134 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_FILE));
4135 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4136 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_UNEVICTABLE));
4137 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4141 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4143 struct mem_cgroup *iter;
4145 for_each_mem_cgroup_tree(iter, mem)
4146 mem_cgroup_get_local_stat(iter, s);
4150 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4153 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4154 unsigned long node_nr;
4155 struct cgroup *cont = m->private;
4156 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4158 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4159 seq_printf(m, "total=%lu", total_nr);
4160 for_each_node_state(nid, N_HIGH_MEMORY) {
4161 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4162 seq_printf(m, " N%d=%lu", nid, node_nr);
4166 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4167 seq_printf(m, "file=%lu", file_nr);
4168 for_each_node_state(nid, N_HIGH_MEMORY) {
4169 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4171 seq_printf(m, " N%d=%lu", nid, node_nr);
4175 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4176 seq_printf(m, "anon=%lu", anon_nr);
4177 for_each_node_state(nid, N_HIGH_MEMORY) {
4178 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4180 seq_printf(m, " N%d=%lu", nid, node_nr);
4184 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4185 seq_printf(m, "unevictable=%lu", unevictable_nr);
4186 for_each_node_state(nid, N_HIGH_MEMORY) {
4187 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4188 BIT(LRU_UNEVICTABLE));
4189 seq_printf(m, " N%d=%lu", nid, node_nr);
4194 #endif /* CONFIG_NUMA */
4196 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4197 struct cgroup_map_cb *cb)
4199 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4200 struct mcs_total_stat mystat;
4203 memset(&mystat, 0, sizeof(mystat));
4204 mem_cgroup_get_local_stat(mem_cont, &mystat);
4207 for (i = 0; i < NR_MCS_STAT; i++) {
4208 if (i == MCS_SWAP && !do_swap_account)
4210 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4213 /* Hierarchical information */
4215 unsigned long long limit, memsw_limit;
4216 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4217 cb->fill(cb, "hierarchical_memory_limit", limit);
4218 if (do_swap_account)
4219 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4222 memset(&mystat, 0, sizeof(mystat));
4223 mem_cgroup_get_total_stat(mem_cont, &mystat);
4224 for (i = 0; i < NR_MCS_STAT; i++) {
4225 if (i == MCS_SWAP && !do_swap_account)
4227 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4230 #ifdef CONFIG_DEBUG_VM
4231 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4235 struct mem_cgroup_per_zone *mz;
4236 unsigned long recent_rotated[2] = {0, 0};
4237 unsigned long recent_scanned[2] = {0, 0};
4239 for_each_online_node(nid)
4240 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4241 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4243 recent_rotated[0] +=
4244 mz->reclaim_stat.recent_rotated[0];
4245 recent_rotated[1] +=
4246 mz->reclaim_stat.recent_rotated[1];
4247 recent_scanned[0] +=
4248 mz->reclaim_stat.recent_scanned[0];
4249 recent_scanned[1] +=
4250 mz->reclaim_stat.recent_scanned[1];
4252 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4253 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4254 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4255 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4262 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4264 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4266 return mem_cgroup_swappiness(memcg);
4269 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4272 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4273 struct mem_cgroup *parent;
4278 if (cgrp->parent == NULL)
4281 parent = mem_cgroup_from_cont(cgrp->parent);
4285 /* If under hierarchy, only empty-root can set this value */
4286 if ((parent->use_hierarchy) ||
4287 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4292 memcg->swappiness = val;
4299 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4301 struct mem_cgroup_threshold_ary *t;
4307 t = rcu_dereference(memcg->thresholds.primary);
4309 t = rcu_dereference(memcg->memsw_thresholds.primary);
4314 usage = mem_cgroup_usage(memcg, swap);
4317 * current_threshold points to threshold just below usage.
4318 * If it's not true, a threshold was crossed after last
4319 * call of __mem_cgroup_threshold().
4321 i = t->current_threshold;
4324 * Iterate backward over array of thresholds starting from
4325 * current_threshold and check if a threshold is crossed.
4326 * If none of thresholds below usage is crossed, we read
4327 * only one element of the array here.
4329 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4330 eventfd_signal(t->entries[i].eventfd, 1);
4332 /* i = current_threshold + 1 */
4336 * Iterate forward over array of thresholds starting from
4337 * current_threshold+1 and check if a threshold is crossed.
4338 * If none of thresholds above usage is crossed, we read
4339 * only one element of the array here.
4341 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4342 eventfd_signal(t->entries[i].eventfd, 1);
4344 /* Update current_threshold */
4345 t->current_threshold = i - 1;
4350 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4353 __mem_cgroup_threshold(memcg, false);
4354 if (do_swap_account)
4355 __mem_cgroup_threshold(memcg, true);
4357 memcg = parent_mem_cgroup(memcg);
4361 static int compare_thresholds(const void *a, const void *b)
4363 const struct mem_cgroup_threshold *_a = a;
4364 const struct mem_cgroup_threshold *_b = b;
4366 return _a->threshold - _b->threshold;
4369 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4371 struct mem_cgroup_eventfd_list *ev;
4373 list_for_each_entry(ev, &mem->oom_notify, list)
4374 eventfd_signal(ev->eventfd, 1);
4378 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4380 struct mem_cgroup *iter;
4382 for_each_mem_cgroup_tree(iter, mem)
4383 mem_cgroup_oom_notify_cb(iter);
4386 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4387 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4389 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4390 struct mem_cgroup_thresholds *thresholds;
4391 struct mem_cgroup_threshold_ary *new;
4392 int type = MEMFILE_TYPE(cft->private);
4393 u64 threshold, usage;
4396 ret = res_counter_memparse_write_strategy(args, &threshold);
4400 mutex_lock(&memcg->thresholds_lock);
4403 thresholds = &memcg->thresholds;
4404 else if (type == _MEMSWAP)
4405 thresholds = &memcg->memsw_thresholds;
4409 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4411 /* Check if a threshold crossed before adding a new one */
4412 if (thresholds->primary)
4413 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4415 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4417 /* Allocate memory for new array of thresholds */
4418 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4426 /* Copy thresholds (if any) to new array */
4427 if (thresholds->primary) {
4428 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4429 sizeof(struct mem_cgroup_threshold));
4432 /* Add new threshold */
4433 new->entries[size - 1].eventfd = eventfd;
4434 new->entries[size - 1].threshold = threshold;
4436 /* Sort thresholds. Registering of new threshold isn't time-critical */
4437 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4438 compare_thresholds, NULL);
4440 /* Find current threshold */
4441 new->current_threshold = -1;
4442 for (i = 0; i < size; i++) {
4443 if (new->entries[i].threshold < usage) {
4445 * new->current_threshold will not be used until
4446 * rcu_assign_pointer(), so it's safe to increment
4449 ++new->current_threshold;
4453 /* Free old spare buffer and save old primary buffer as spare */
4454 kfree(thresholds->spare);
4455 thresholds->spare = thresholds->primary;
4457 rcu_assign_pointer(thresholds->primary, new);
4459 /* To be sure that nobody uses thresholds */
4463 mutex_unlock(&memcg->thresholds_lock);
4468 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4469 struct cftype *cft, struct eventfd_ctx *eventfd)
4471 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4472 struct mem_cgroup_thresholds *thresholds;
4473 struct mem_cgroup_threshold_ary *new;
4474 int type = MEMFILE_TYPE(cft->private);
4478 mutex_lock(&memcg->thresholds_lock);
4480 thresholds = &memcg->thresholds;
4481 else if (type == _MEMSWAP)
4482 thresholds = &memcg->memsw_thresholds;
4487 * Something went wrong if we trying to unregister a threshold
4488 * if we don't have thresholds
4490 BUG_ON(!thresholds);
4492 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4494 /* Check if a threshold crossed before removing */
4495 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4497 /* Calculate new number of threshold */
4499 for (i = 0; i < thresholds->primary->size; i++) {
4500 if (thresholds->primary->entries[i].eventfd != eventfd)
4504 new = thresholds->spare;
4506 /* Set thresholds array to NULL if we don't have thresholds */
4515 /* Copy thresholds and find current threshold */
4516 new->current_threshold = -1;
4517 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4518 if (thresholds->primary->entries[i].eventfd == eventfd)
4521 new->entries[j] = thresholds->primary->entries[i];
4522 if (new->entries[j].threshold < usage) {
4524 * new->current_threshold will not be used
4525 * until rcu_assign_pointer(), so it's safe to increment
4528 ++new->current_threshold;
4534 /* Swap primary and spare array */
4535 thresholds->spare = thresholds->primary;
4536 rcu_assign_pointer(thresholds->primary, new);
4538 /* To be sure that nobody uses thresholds */
4541 mutex_unlock(&memcg->thresholds_lock);
4544 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4545 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4547 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4548 struct mem_cgroup_eventfd_list *event;
4549 int type = MEMFILE_TYPE(cft->private);
4551 BUG_ON(type != _OOM_TYPE);
4552 event = kmalloc(sizeof(*event), GFP_KERNEL);
4556 mutex_lock(&memcg_oom_mutex);
4558 event->eventfd = eventfd;
4559 list_add(&event->list, &memcg->oom_notify);
4561 /* already in OOM ? */
4562 if (atomic_read(&memcg->under_oom))
4563 eventfd_signal(eventfd, 1);
4564 mutex_unlock(&memcg_oom_mutex);
4569 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4570 struct cftype *cft, struct eventfd_ctx *eventfd)
4572 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4573 struct mem_cgroup_eventfd_list *ev, *tmp;
4574 int type = MEMFILE_TYPE(cft->private);
4576 BUG_ON(type != _OOM_TYPE);
4578 mutex_lock(&memcg_oom_mutex);
4580 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4581 if (ev->eventfd == eventfd) {
4582 list_del(&ev->list);
4587 mutex_unlock(&memcg_oom_mutex);
4590 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4591 struct cftype *cft, struct cgroup_map_cb *cb)
4593 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4595 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4597 if (atomic_read(&mem->under_oom))
4598 cb->fill(cb, "under_oom", 1);
4600 cb->fill(cb, "under_oom", 0);
4604 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4605 struct cftype *cft, u64 val)
4607 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4608 struct mem_cgroup *parent;
4610 /* cannot set to root cgroup and only 0 and 1 are allowed */
4611 if (!cgrp->parent || !((val == 0) || (val == 1)))
4614 parent = mem_cgroup_from_cont(cgrp->parent);
4617 /* oom-kill-disable is a flag for subhierarchy. */
4618 if ((parent->use_hierarchy) ||
4619 (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4623 mem->oom_kill_disable = val;
4625 memcg_oom_recover(mem);
4631 static const struct file_operations mem_control_numa_stat_file_operations = {
4633 .llseek = seq_lseek,
4634 .release = single_release,
4637 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4639 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4641 file->f_op = &mem_control_numa_stat_file_operations;
4642 return single_open(file, mem_control_numa_stat_show, cont);
4644 #endif /* CONFIG_NUMA */
4646 static struct cftype mem_cgroup_files[] = {
4648 .name = "usage_in_bytes",
4649 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4650 .read_u64 = mem_cgroup_read,
4651 .register_event = mem_cgroup_usage_register_event,
4652 .unregister_event = mem_cgroup_usage_unregister_event,
4655 .name = "max_usage_in_bytes",
4656 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4657 .trigger = mem_cgroup_reset,
4658 .read_u64 = mem_cgroup_read,
4661 .name = "limit_in_bytes",
4662 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4663 .write_string = mem_cgroup_write,
4664 .read_u64 = mem_cgroup_read,
4667 .name = "soft_limit_in_bytes",
4668 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4669 .write_string = mem_cgroup_write,
4670 .read_u64 = mem_cgroup_read,
4674 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4675 .trigger = mem_cgroup_reset,
4676 .read_u64 = mem_cgroup_read,
4680 .read_map = mem_control_stat_show,
4683 .name = "force_empty",
4684 .trigger = mem_cgroup_force_empty_write,
4687 .name = "use_hierarchy",
4688 .write_u64 = mem_cgroup_hierarchy_write,
4689 .read_u64 = mem_cgroup_hierarchy_read,
4692 .name = "swappiness",
4693 .read_u64 = mem_cgroup_swappiness_read,
4694 .write_u64 = mem_cgroup_swappiness_write,
4697 .name = "move_charge_at_immigrate",
4698 .read_u64 = mem_cgroup_move_charge_read,
4699 .write_u64 = mem_cgroup_move_charge_write,
4702 .name = "oom_control",
4703 .read_map = mem_cgroup_oom_control_read,
4704 .write_u64 = mem_cgroup_oom_control_write,
4705 .register_event = mem_cgroup_oom_register_event,
4706 .unregister_event = mem_cgroup_oom_unregister_event,
4707 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4711 .name = "numa_stat",
4712 .open = mem_control_numa_stat_open,
4718 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4719 static struct cftype memsw_cgroup_files[] = {
4721 .name = "memsw.usage_in_bytes",
4722 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4723 .read_u64 = mem_cgroup_read,
4724 .register_event = mem_cgroup_usage_register_event,
4725 .unregister_event = mem_cgroup_usage_unregister_event,
4728 .name = "memsw.max_usage_in_bytes",
4729 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4730 .trigger = mem_cgroup_reset,
4731 .read_u64 = mem_cgroup_read,
4734 .name = "memsw.limit_in_bytes",
4735 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4736 .write_string = mem_cgroup_write,
4737 .read_u64 = mem_cgroup_read,
4740 .name = "memsw.failcnt",
4741 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4742 .trigger = mem_cgroup_reset,
4743 .read_u64 = mem_cgroup_read,
4747 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4749 if (!do_swap_account)
4751 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4752 ARRAY_SIZE(memsw_cgroup_files));
4755 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4761 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4763 struct mem_cgroup_per_node *pn;
4764 struct mem_cgroup_per_zone *mz;
4766 int zone, tmp = node;
4768 * This routine is called against possible nodes.
4769 * But it's BUG to call kmalloc() against offline node.
4771 * TODO: this routine can waste much memory for nodes which will
4772 * never be onlined. It's better to use memory hotplug callback
4775 if (!node_state(node, N_NORMAL_MEMORY))
4777 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4781 mem->info.nodeinfo[node] = pn;
4782 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4783 mz = &pn->zoneinfo[zone];
4785 INIT_LIST_HEAD(&mz->lists[l]);
4786 mz->usage_in_excess = 0;
4787 mz->on_tree = false;
4793 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4795 kfree(mem->info.nodeinfo[node]);
4798 static struct mem_cgroup *mem_cgroup_alloc(void)
4800 struct mem_cgroup *mem;
4801 int size = sizeof(struct mem_cgroup);
4803 /* Can be very big if MAX_NUMNODES is very big */
4804 if (size < PAGE_SIZE)
4805 mem = kzalloc(size, GFP_KERNEL);
4807 mem = vzalloc(size);
4812 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4815 spin_lock_init(&mem->pcp_counter_lock);
4819 if (size < PAGE_SIZE)
4827 * At destroying mem_cgroup, references from swap_cgroup can remain.
4828 * (scanning all at force_empty is too costly...)
4830 * Instead of clearing all references at force_empty, we remember
4831 * the number of reference from swap_cgroup and free mem_cgroup when
4832 * it goes down to 0.
4834 * Removal of cgroup itself succeeds regardless of refs from swap.
4837 static void __mem_cgroup_free(struct mem_cgroup *mem)
4841 mem_cgroup_remove_from_trees(mem);
4842 free_css_id(&mem_cgroup_subsys, &mem->css);
4844 for_each_node_state(node, N_POSSIBLE)
4845 free_mem_cgroup_per_zone_info(mem, node);
4847 free_percpu(mem->stat);
4848 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4854 static void mem_cgroup_get(struct mem_cgroup *mem)
4856 atomic_inc(&mem->refcnt);
4859 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4861 if (atomic_sub_and_test(count, &mem->refcnt)) {
4862 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4863 __mem_cgroup_free(mem);
4865 mem_cgroup_put(parent);
4869 static void mem_cgroup_put(struct mem_cgroup *mem)
4871 __mem_cgroup_put(mem, 1);
4875 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4877 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4879 if (!mem->res.parent)
4881 return mem_cgroup_from_res_counter(mem->res.parent, res);
4884 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4885 static void __init enable_swap_cgroup(void)
4887 if (!mem_cgroup_disabled() && really_do_swap_account)
4888 do_swap_account = 1;
4891 static void __init enable_swap_cgroup(void)
4896 static int mem_cgroup_soft_limit_tree_init(void)
4898 struct mem_cgroup_tree_per_node *rtpn;
4899 struct mem_cgroup_tree_per_zone *rtpz;
4900 int tmp, node, zone;
4902 for_each_node_state(node, N_POSSIBLE) {
4904 if (!node_state(node, N_NORMAL_MEMORY))
4906 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4910 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4912 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4913 rtpz = &rtpn->rb_tree_per_zone[zone];
4914 rtpz->rb_root = RB_ROOT;
4915 spin_lock_init(&rtpz->lock);
4921 static struct cgroup_subsys_state * __ref
4922 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4924 struct mem_cgroup *mem, *parent;
4925 long error = -ENOMEM;
4928 mem = mem_cgroup_alloc();
4930 return ERR_PTR(error);
4932 for_each_node_state(node, N_POSSIBLE)
4933 if (alloc_mem_cgroup_per_zone_info(mem, node))
4937 if (cont->parent == NULL) {
4939 enable_swap_cgroup();
4941 root_mem_cgroup = mem;
4942 if (mem_cgroup_soft_limit_tree_init())
4944 for_each_possible_cpu(cpu) {
4945 struct memcg_stock_pcp *stock =
4946 &per_cpu(memcg_stock, cpu);
4947 INIT_WORK(&stock->work, drain_local_stock);
4949 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4951 parent = mem_cgroup_from_cont(cont->parent);
4952 mem->use_hierarchy = parent->use_hierarchy;
4953 mem->oom_kill_disable = parent->oom_kill_disable;
4956 if (parent && parent->use_hierarchy) {
4957 res_counter_init(&mem->res, &parent->res);
4958 res_counter_init(&mem->memsw, &parent->memsw);
4960 * We increment refcnt of the parent to ensure that we can
4961 * safely access it on res_counter_charge/uncharge.
4962 * This refcnt will be decremented when freeing this
4963 * mem_cgroup(see mem_cgroup_put).
4965 mem_cgroup_get(parent);
4967 res_counter_init(&mem->res, NULL);
4968 res_counter_init(&mem->memsw, NULL);
4970 mem->last_scanned_child = 0;
4971 mem->last_scanned_node = MAX_NUMNODES;
4972 INIT_LIST_HEAD(&mem->oom_notify);
4975 mem->swappiness = mem_cgroup_swappiness(parent);
4976 atomic_set(&mem->refcnt, 1);
4977 mem->move_charge_at_immigrate = 0;
4978 mutex_init(&mem->thresholds_lock);
4981 __mem_cgroup_free(mem);
4982 root_mem_cgroup = NULL;
4983 return ERR_PTR(error);
4986 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4987 struct cgroup *cont)
4989 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4991 return mem_cgroup_force_empty(mem, false);
4994 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4995 struct cgroup *cont)
4997 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4999 mem_cgroup_put(mem);
5002 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5003 struct cgroup *cont)
5007 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5008 ARRAY_SIZE(mem_cgroup_files));
5011 ret = register_memsw_files(cont, ss);
5016 /* Handlers for move charge at task migration. */
5017 #define PRECHARGE_COUNT_AT_ONCE 256
5018 static int mem_cgroup_do_precharge(unsigned long count)
5021 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5022 struct mem_cgroup *mem = mc.to;
5024 if (mem_cgroup_is_root(mem)) {
5025 mc.precharge += count;
5026 /* we don't need css_get for root */
5029 /* try to charge at once */
5031 struct res_counter *dummy;
5033 * "mem" cannot be under rmdir() because we've already checked
5034 * by cgroup_lock_live_cgroup() that it is not removed and we
5035 * are still under the same cgroup_mutex. So we can postpone
5038 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
5040 if (do_swap_account && res_counter_charge(&mem->memsw,
5041 PAGE_SIZE * count, &dummy)) {
5042 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
5045 mc.precharge += count;
5049 /* fall back to one by one charge */
5051 if (signal_pending(current)) {
5055 if (!batch_count--) {
5056 batch_count = PRECHARGE_COUNT_AT_ONCE;
5059 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5061 /* mem_cgroup_clear_mc() will do uncharge later */
5069 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5070 * @vma: the vma the pte to be checked belongs
5071 * @addr: the address corresponding to the pte to be checked
5072 * @ptent: the pte to be checked
5073 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5076 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5077 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5078 * move charge. if @target is not NULL, the page is stored in target->page
5079 * with extra refcnt got(Callers should handle it).
5080 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5081 * target for charge migration. if @target is not NULL, the entry is stored
5084 * Called with pte lock held.
5091 enum mc_target_type {
5092 MC_TARGET_NONE, /* not used */
5097 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5098 unsigned long addr, pte_t ptent)
5100 struct page *page = vm_normal_page(vma, addr, ptent);
5102 if (!page || !page_mapped(page))
5104 if (PageAnon(page)) {
5105 /* we don't move shared anon */
5106 if (!move_anon() || page_mapcount(page) > 2)
5108 } else if (!move_file())
5109 /* we ignore mapcount for file pages */
5111 if (!get_page_unless_zero(page))
5117 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5118 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5121 struct page *page = NULL;
5122 swp_entry_t ent = pte_to_swp_entry(ptent);
5124 if (!move_anon() || non_swap_entry(ent))
5126 usage_count = mem_cgroup_count_swap_user(ent, &page);
5127 if (usage_count > 1) { /* we don't move shared anon */
5132 if (do_swap_account)
5133 entry->val = ent.val;
5138 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5139 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5141 struct page *page = NULL;
5142 struct inode *inode;
5143 struct address_space *mapping;
5146 if (!vma->vm_file) /* anonymous vma */
5151 inode = vma->vm_file->f_path.dentry->d_inode;
5152 mapping = vma->vm_file->f_mapping;
5153 if (pte_none(ptent))
5154 pgoff = linear_page_index(vma, addr);
5155 else /* pte_file(ptent) is true */
5156 pgoff = pte_to_pgoff(ptent);
5158 /* page is moved even if it's not RSS of this task(page-faulted). */
5159 if (!mapping_cap_swap_backed(mapping)) { /* normal file */
5160 page = find_get_page(mapping, pgoff);
5161 } else { /* shmem/tmpfs file. we should take account of swap too. */
5163 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
5164 if (do_swap_account)
5165 entry->val = ent.val;
5171 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5172 unsigned long addr, pte_t ptent, union mc_target *target)
5174 struct page *page = NULL;
5175 struct page_cgroup *pc;
5177 swp_entry_t ent = { .val = 0 };
5179 if (pte_present(ptent))
5180 page = mc_handle_present_pte(vma, addr, ptent);
5181 else if (is_swap_pte(ptent))
5182 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5183 else if (pte_none(ptent) || pte_file(ptent))
5184 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5186 if (!page && !ent.val)
5189 pc = lookup_page_cgroup(page);
5191 * Do only loose check w/o page_cgroup lock.
5192 * mem_cgroup_move_account() checks the pc is valid or not under
5195 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5196 ret = MC_TARGET_PAGE;
5198 target->page = page;
5200 if (!ret || !target)
5203 /* There is a swap entry and a page doesn't exist or isn't charged */
5204 if (ent.val && !ret &&
5205 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5206 ret = MC_TARGET_SWAP;
5213 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5214 unsigned long addr, unsigned long end,
5215 struct mm_walk *walk)
5217 struct vm_area_struct *vma = walk->private;
5221 split_huge_page_pmd(walk->mm, pmd);
5223 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5224 for (; addr != end; pte++, addr += PAGE_SIZE)
5225 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5226 mc.precharge++; /* increment precharge temporarily */
5227 pte_unmap_unlock(pte - 1, ptl);
5233 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5235 unsigned long precharge;
5236 struct vm_area_struct *vma;
5238 down_read(&mm->mmap_sem);
5239 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5240 struct mm_walk mem_cgroup_count_precharge_walk = {
5241 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5245 if (is_vm_hugetlb_page(vma))
5247 walk_page_range(vma->vm_start, vma->vm_end,
5248 &mem_cgroup_count_precharge_walk);
5250 up_read(&mm->mmap_sem);
5252 precharge = mc.precharge;
5258 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5260 unsigned long precharge = mem_cgroup_count_precharge(mm);
5262 VM_BUG_ON(mc.moving_task);
5263 mc.moving_task = current;
5264 return mem_cgroup_do_precharge(precharge);
5267 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5268 static void __mem_cgroup_clear_mc(void)
5270 struct mem_cgroup *from = mc.from;
5271 struct mem_cgroup *to = mc.to;
5273 /* we must uncharge all the leftover precharges from mc.to */
5275 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5279 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5280 * we must uncharge here.
5282 if (mc.moved_charge) {
5283 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5284 mc.moved_charge = 0;
5286 /* we must fixup refcnts and charges */
5287 if (mc.moved_swap) {
5288 /* uncharge swap account from the old cgroup */
5289 if (!mem_cgroup_is_root(mc.from))
5290 res_counter_uncharge(&mc.from->memsw,
5291 PAGE_SIZE * mc.moved_swap);
5292 __mem_cgroup_put(mc.from, mc.moved_swap);
5294 if (!mem_cgroup_is_root(mc.to)) {
5296 * we charged both to->res and to->memsw, so we should
5299 res_counter_uncharge(&mc.to->res,
5300 PAGE_SIZE * mc.moved_swap);
5302 /* we've already done mem_cgroup_get(mc.to) */
5305 memcg_oom_recover(from);
5306 memcg_oom_recover(to);
5307 wake_up_all(&mc.waitq);
5310 static void mem_cgroup_clear_mc(void)
5312 struct mem_cgroup *from = mc.from;
5315 * we must clear moving_task before waking up waiters at the end of
5318 mc.moving_task = NULL;
5319 __mem_cgroup_clear_mc();
5320 spin_lock(&mc.lock);
5323 spin_unlock(&mc.lock);
5324 mem_cgroup_end_move(from);
5327 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5328 struct cgroup *cgroup,
5329 struct task_struct *p)
5332 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5334 if (mem->move_charge_at_immigrate) {
5335 struct mm_struct *mm;
5336 struct mem_cgroup *from = mem_cgroup_from_task(p);
5338 VM_BUG_ON(from == mem);
5340 mm = get_task_mm(p);
5343 /* We move charges only when we move a owner of the mm */
5344 if (mm->owner == p) {
5347 VM_BUG_ON(mc.precharge);
5348 VM_BUG_ON(mc.moved_charge);
5349 VM_BUG_ON(mc.moved_swap);
5350 mem_cgroup_start_move(from);
5351 spin_lock(&mc.lock);
5354 spin_unlock(&mc.lock);
5355 /* We set mc.moving_task later */
5357 ret = mem_cgroup_precharge_mc(mm);
5359 mem_cgroup_clear_mc();
5366 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5367 struct cgroup *cgroup,
5368 struct task_struct *p)
5370 mem_cgroup_clear_mc();
5373 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5374 unsigned long addr, unsigned long end,
5375 struct mm_walk *walk)
5378 struct vm_area_struct *vma = walk->private;
5382 split_huge_page_pmd(walk->mm, pmd);
5384 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5385 for (; addr != end; addr += PAGE_SIZE) {
5386 pte_t ptent = *(pte++);
5387 union mc_target target;
5390 struct page_cgroup *pc;
5396 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5398 case MC_TARGET_PAGE:
5400 if (isolate_lru_page(page))
5402 pc = lookup_page_cgroup(page);
5403 if (!mem_cgroup_move_account(page, 1, pc,
5404 mc.from, mc.to, false)) {
5406 /* we uncharge from mc.from later. */
5409 putback_lru_page(page);
5410 put: /* is_target_pte_for_mc() gets the page */
5413 case MC_TARGET_SWAP:
5415 if (!mem_cgroup_move_swap_account(ent,
5416 mc.from, mc.to, false)) {
5418 /* we fixup refcnts and charges later. */
5426 pte_unmap_unlock(pte - 1, ptl);
5431 * We have consumed all precharges we got in can_attach().
5432 * We try charge one by one, but don't do any additional
5433 * charges to mc.to if we have failed in charge once in attach()
5436 ret = mem_cgroup_do_precharge(1);
5444 static void mem_cgroup_move_charge(struct mm_struct *mm)
5446 struct vm_area_struct *vma;
5448 lru_add_drain_all();
5450 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5452 * Someone who are holding the mmap_sem might be waiting in
5453 * waitq. So we cancel all extra charges, wake up all waiters,
5454 * and retry. Because we cancel precharges, we might not be able
5455 * to move enough charges, but moving charge is a best-effort
5456 * feature anyway, so it wouldn't be a big problem.
5458 __mem_cgroup_clear_mc();
5462 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5464 struct mm_walk mem_cgroup_move_charge_walk = {
5465 .pmd_entry = mem_cgroup_move_charge_pte_range,
5469 if (is_vm_hugetlb_page(vma))
5471 ret = walk_page_range(vma->vm_start, vma->vm_end,
5472 &mem_cgroup_move_charge_walk);
5475 * means we have consumed all precharges and failed in
5476 * doing additional charge. Just abandon here.
5480 up_read(&mm->mmap_sem);
5483 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5484 struct cgroup *cont,
5485 struct cgroup *old_cont,
5486 struct task_struct *p)
5488 struct mm_struct *mm = get_task_mm(p);
5492 mem_cgroup_move_charge(mm);
5497 mem_cgroup_clear_mc();
5499 #else /* !CONFIG_MMU */
5500 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5501 struct cgroup *cgroup,
5502 struct task_struct *p)
5506 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5507 struct cgroup *cgroup,
5508 struct task_struct *p)
5511 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5512 struct cgroup *cont,
5513 struct cgroup *old_cont,
5514 struct task_struct *p)
5519 struct cgroup_subsys mem_cgroup_subsys = {
5521 .subsys_id = mem_cgroup_subsys_id,
5522 .create = mem_cgroup_create,
5523 .pre_destroy = mem_cgroup_pre_destroy,
5524 .destroy = mem_cgroup_destroy,
5525 .populate = mem_cgroup_populate,
5526 .can_attach = mem_cgroup_can_attach,
5527 .cancel_attach = mem_cgroup_cancel_attach,
5528 .attach = mem_cgroup_move_task,
5533 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5534 static int __init enable_swap_account(char *s)
5536 /* consider enabled if no parameter or 1 is given */
5537 if (!strcmp(s, "1"))
5538 really_do_swap_account = 1;
5539 else if (!strcmp(s, "0"))
5540 really_do_swap_account = 0;
5543 __setup("swapaccount=", enable_swap_account);