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?
253 /* OOM-Killer disable */
254 int oom_kill_disable;
256 /* set when res.limit == memsw.limit */
257 bool memsw_is_minimum;
259 /* protect arrays of thresholds */
260 struct mutex thresholds_lock;
262 /* thresholds for memory usage. RCU-protected */
263 struct mem_cgroup_thresholds thresholds;
265 /* thresholds for mem+swap usage. RCU-protected */
266 struct mem_cgroup_thresholds memsw_thresholds;
268 /* For oom notifier event fd */
269 struct list_head oom_notify;
272 * Should we move charges of a task when a task is moved into this
273 * mem_cgroup ? And what type of charges should we move ?
275 unsigned long move_charge_at_immigrate;
279 struct mem_cgroup_stat_cpu *stat;
281 * used when a cpu is offlined or other synchronizations
282 * See mem_cgroup_read_stat().
284 struct mem_cgroup_stat_cpu nocpu_base;
285 spinlock_t pcp_counter_lock;
288 /* Stuffs for move charges at task migration. */
290 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
291 * left-shifted bitmap of these types.
294 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
295 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
299 /* "mc" and its members are protected by cgroup_mutex */
300 static struct move_charge_struct {
301 spinlock_t lock; /* for from, to */
302 struct mem_cgroup *from;
303 struct mem_cgroup *to;
304 unsigned long precharge;
305 unsigned long moved_charge;
306 unsigned long moved_swap;
307 struct task_struct *moving_task; /* a task moving charges */
308 wait_queue_head_t waitq; /* a waitq for other context */
310 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
311 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
314 static bool move_anon(void)
316 return test_bit(MOVE_CHARGE_TYPE_ANON,
317 &mc.to->move_charge_at_immigrate);
320 static bool move_file(void)
322 return test_bit(MOVE_CHARGE_TYPE_FILE,
323 &mc.to->move_charge_at_immigrate);
327 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
328 * limit reclaim to prevent infinite loops, if they ever occur.
330 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
331 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
334 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
335 MEM_CGROUP_CHARGE_TYPE_MAPPED,
336 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
337 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
338 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
339 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
343 /* for encoding cft->private value on file */
346 #define _OOM_TYPE (2)
347 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
348 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
349 #define MEMFILE_ATTR(val) ((val) & 0xffff)
350 /* Used for OOM nofiier */
351 #define OOM_CONTROL (0)
354 * Reclaim flags for mem_cgroup_hierarchical_reclaim
356 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
357 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
358 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
359 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
360 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
361 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
363 static void mem_cgroup_get(struct mem_cgroup *mem);
364 static void mem_cgroup_put(struct mem_cgroup *mem);
365 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
366 static void drain_all_stock_async(struct mem_cgroup *mem);
368 static struct mem_cgroup_per_zone *
369 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
371 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
374 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
379 static struct mem_cgroup_per_zone *
380 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
382 int nid = page_to_nid(page);
383 int zid = page_zonenum(page);
385 return mem_cgroup_zoneinfo(mem, nid, zid);
388 static struct mem_cgroup_tree_per_zone *
389 soft_limit_tree_node_zone(int nid, int zid)
391 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
394 static struct mem_cgroup_tree_per_zone *
395 soft_limit_tree_from_page(struct page *page)
397 int nid = page_to_nid(page);
398 int zid = page_zonenum(page);
400 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
404 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
405 struct mem_cgroup_per_zone *mz,
406 struct mem_cgroup_tree_per_zone *mctz,
407 unsigned long long new_usage_in_excess)
409 struct rb_node **p = &mctz->rb_root.rb_node;
410 struct rb_node *parent = NULL;
411 struct mem_cgroup_per_zone *mz_node;
416 mz->usage_in_excess = new_usage_in_excess;
417 if (!mz->usage_in_excess)
421 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
423 if (mz->usage_in_excess < mz_node->usage_in_excess)
426 * We can't avoid mem cgroups that are over their soft
427 * limit by the same amount
429 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
432 rb_link_node(&mz->tree_node, parent, p);
433 rb_insert_color(&mz->tree_node, &mctz->rb_root);
438 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
439 struct mem_cgroup_per_zone *mz,
440 struct mem_cgroup_tree_per_zone *mctz)
444 rb_erase(&mz->tree_node, &mctz->rb_root);
449 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
450 struct mem_cgroup_per_zone *mz,
451 struct mem_cgroup_tree_per_zone *mctz)
453 spin_lock(&mctz->lock);
454 __mem_cgroup_remove_exceeded(mem, mz, mctz);
455 spin_unlock(&mctz->lock);
459 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
461 unsigned long long excess;
462 struct mem_cgroup_per_zone *mz;
463 struct mem_cgroup_tree_per_zone *mctz;
464 int nid = page_to_nid(page);
465 int zid = page_zonenum(page);
466 mctz = soft_limit_tree_from_page(page);
469 * Necessary to update all ancestors when hierarchy is used.
470 * because their event counter is not touched.
472 for (; mem; mem = parent_mem_cgroup(mem)) {
473 mz = mem_cgroup_zoneinfo(mem, nid, zid);
474 excess = res_counter_soft_limit_excess(&mem->res);
476 * We have to update the tree if mz is on RB-tree or
477 * mem is over its softlimit.
479 if (excess || mz->on_tree) {
480 spin_lock(&mctz->lock);
481 /* if on-tree, remove it */
483 __mem_cgroup_remove_exceeded(mem, mz, mctz);
485 * Insert again. mz->usage_in_excess will be updated.
486 * If excess is 0, no tree ops.
488 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
489 spin_unlock(&mctz->lock);
494 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
497 struct mem_cgroup_per_zone *mz;
498 struct mem_cgroup_tree_per_zone *mctz;
500 for_each_node_state(node, N_POSSIBLE) {
501 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
502 mz = mem_cgroup_zoneinfo(mem, node, zone);
503 mctz = soft_limit_tree_node_zone(node, zone);
504 mem_cgroup_remove_exceeded(mem, mz, mctz);
509 static struct mem_cgroup_per_zone *
510 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
512 struct rb_node *rightmost = NULL;
513 struct mem_cgroup_per_zone *mz;
517 rightmost = rb_last(&mctz->rb_root);
519 goto done; /* Nothing to reclaim from */
521 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
523 * Remove the node now but someone else can add it back,
524 * we will to add it back at the end of reclaim to its correct
525 * position in the tree.
527 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
528 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
529 !css_tryget(&mz->mem->css))
535 static struct mem_cgroup_per_zone *
536 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
538 struct mem_cgroup_per_zone *mz;
540 spin_lock(&mctz->lock);
541 mz = __mem_cgroup_largest_soft_limit_node(mctz);
542 spin_unlock(&mctz->lock);
547 * Implementation Note: reading percpu statistics for memcg.
549 * Both of vmstat[] and percpu_counter has threshold and do periodic
550 * synchronization to implement "quick" read. There are trade-off between
551 * reading cost and precision of value. Then, we may have a chance to implement
552 * a periodic synchronizion of counter in memcg's counter.
554 * But this _read() function is used for user interface now. The user accounts
555 * memory usage by memory cgroup and he _always_ requires exact value because
556 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
557 * have to visit all online cpus and make sum. So, for now, unnecessary
558 * synchronization is not implemented. (just implemented for cpu hotplug)
560 * If there are kernel internal actions which can make use of some not-exact
561 * value, and reading all cpu value can be performance bottleneck in some
562 * common workload, threashold and synchonization as vmstat[] should be
565 static long mem_cgroup_read_stat(struct mem_cgroup *mem,
566 enum mem_cgroup_stat_index idx)
572 for_each_online_cpu(cpu)
573 val += per_cpu(mem->stat->count[idx], cpu);
574 #ifdef CONFIG_HOTPLUG_CPU
575 spin_lock(&mem->pcp_counter_lock);
576 val += mem->nocpu_base.count[idx];
577 spin_unlock(&mem->pcp_counter_lock);
583 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
586 int val = (charge) ? 1 : -1;
587 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
590 void mem_cgroup_pgfault(struct mem_cgroup *mem, int val)
592 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
595 void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val)
597 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
600 static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
601 enum mem_cgroup_events_index idx)
603 unsigned long val = 0;
606 for_each_online_cpu(cpu)
607 val += per_cpu(mem->stat->events[idx], cpu);
608 #ifdef CONFIG_HOTPLUG_CPU
609 spin_lock(&mem->pcp_counter_lock);
610 val += mem->nocpu_base.events[idx];
611 spin_unlock(&mem->pcp_counter_lock);
616 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
617 bool file, int nr_pages)
622 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
624 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
626 /* pagein of a big page is an event. So, ignore page size */
628 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
630 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
631 nr_pages = -nr_pages; /* for event */
634 __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
640 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *mem, int nid, int zid,
641 unsigned int lru_mask)
643 struct mem_cgroup_per_zone *mz;
645 unsigned long ret = 0;
647 mz = mem_cgroup_zoneinfo(mem, nid, zid);
650 if (BIT(l) & lru_mask)
651 ret += MEM_CGROUP_ZSTAT(mz, l);
657 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *mem,
658 int nid, unsigned int lru_mask)
663 for (zid = 0; zid < MAX_NR_ZONES; zid++)
664 total += mem_cgroup_zone_nr_lru_pages(mem, nid, zid, lru_mask);
669 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *mem,
670 unsigned int lru_mask)
675 for_each_node_state(nid, N_HIGH_MEMORY)
676 total += mem_cgroup_node_nr_lru_pages(mem, nid, lru_mask);
680 static bool __memcg_event_check(struct mem_cgroup *mem, int target)
682 unsigned long val, next;
684 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
685 next = this_cpu_read(mem->stat->targets[target]);
686 /* from time_after() in jiffies.h */
687 return ((long)next - (long)val < 0);
690 static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target)
692 unsigned long val, next;
694 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
697 case MEM_CGROUP_TARGET_THRESH:
698 next = val + THRESHOLDS_EVENTS_TARGET;
700 case MEM_CGROUP_TARGET_SOFTLIMIT:
701 next = val + SOFTLIMIT_EVENTS_TARGET;
703 case MEM_CGROUP_TARGET_NUMAINFO:
704 next = val + NUMAINFO_EVENTS_TARGET;
710 this_cpu_write(mem->stat->targets[target], next);
714 * Check events in order.
717 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
719 /* threshold event is triggered in finer grain than soft limit */
720 if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) {
721 mem_cgroup_threshold(mem);
722 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH);
723 if (unlikely(__memcg_event_check(mem,
724 MEM_CGROUP_TARGET_SOFTLIMIT))) {
725 mem_cgroup_update_tree(mem, page);
726 __mem_cgroup_target_update(mem,
727 MEM_CGROUP_TARGET_SOFTLIMIT);
730 if (unlikely(__memcg_event_check(mem,
731 MEM_CGROUP_TARGET_NUMAINFO))) {
732 atomic_inc(&mem->numainfo_events);
733 __mem_cgroup_target_update(mem,
734 MEM_CGROUP_TARGET_NUMAINFO);
740 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
742 return container_of(cgroup_subsys_state(cont,
743 mem_cgroup_subsys_id), struct mem_cgroup,
747 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
750 * mm_update_next_owner() may clear mm->owner to NULL
751 * if it races with swapoff, page migration, etc.
752 * So this can be called with p == NULL.
757 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
758 struct mem_cgroup, css);
761 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
763 struct mem_cgroup *mem = NULL;
768 * Because we have no locks, mm->owner's may be being moved to other
769 * cgroup. We use css_tryget() here even if this looks
770 * pessimistic (rather than adding locks here).
774 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
777 } while (!css_tryget(&mem->css));
782 /* The caller has to guarantee "mem" exists before calling this */
783 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
785 struct cgroup_subsys_state *css;
788 if (!mem) /* ROOT cgroup has the smallest ID */
789 return root_mem_cgroup; /*css_put/get against root is ignored*/
790 if (!mem->use_hierarchy) {
791 if (css_tryget(&mem->css))
797 * searching a memory cgroup which has the smallest ID under given
798 * ROOT cgroup. (ID >= 1)
800 css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
801 if (css && css_tryget(css))
802 mem = container_of(css, struct mem_cgroup, css);
809 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
810 struct mem_cgroup *root,
813 int nextid = css_id(&iter->css) + 1;
816 struct cgroup_subsys_state *css;
818 hierarchy_used = iter->use_hierarchy;
821 /* If no ROOT, walk all, ignore hierarchy */
822 if (!cond || (root && !hierarchy_used))
826 root = root_mem_cgroup;
832 css = css_get_next(&mem_cgroup_subsys, nextid,
834 if (css && css_tryget(css))
835 iter = container_of(css, struct mem_cgroup, css);
837 /* If css is NULL, no more cgroups will be found */
839 } while (css && !iter);
844 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
845 * be careful that "break" loop is not allowed. We have reference count.
846 * Instead of that modify "cond" to be false and "continue" to exit the loop.
848 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
849 for (iter = mem_cgroup_start_loop(root);\
851 iter = mem_cgroup_get_next(iter, root, cond))
853 #define for_each_mem_cgroup_tree(iter, root) \
854 for_each_mem_cgroup_tree_cond(iter, root, true)
856 #define for_each_mem_cgroup_all(iter) \
857 for_each_mem_cgroup_tree_cond(iter, NULL, true)
860 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
862 return (mem == root_mem_cgroup);
865 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
867 struct mem_cgroup *mem;
873 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
879 mem_cgroup_pgmajfault(mem, 1);
882 mem_cgroup_pgfault(mem, 1);
890 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
893 * Following LRU functions are allowed to be used without PCG_LOCK.
894 * Operations are called by routine of global LRU independently from memcg.
895 * What we have to take care of here is validness of pc->mem_cgroup.
897 * Changes to pc->mem_cgroup happens when
900 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
901 * It is added to LRU before charge.
902 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
903 * When moving account, the page is not on LRU. It's isolated.
906 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
908 struct page_cgroup *pc;
909 struct mem_cgroup_per_zone *mz;
911 if (mem_cgroup_disabled())
913 pc = lookup_page_cgroup(page);
914 /* can happen while we handle swapcache. */
915 if (!TestClearPageCgroupAcctLRU(pc))
917 VM_BUG_ON(!pc->mem_cgroup);
919 * We don't check PCG_USED bit. It's cleared when the "page" is finally
920 * removed from global LRU.
922 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
923 /* huge page split is done under lru_lock. so, we have no races. */
924 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
925 if (mem_cgroup_is_root(pc->mem_cgroup))
927 VM_BUG_ON(list_empty(&pc->lru));
928 list_del_init(&pc->lru);
931 void mem_cgroup_del_lru(struct page *page)
933 mem_cgroup_del_lru_list(page, page_lru(page));
937 * Writeback is about to end against a page which has been marked for immediate
938 * reclaim. If it still appears to be reclaimable, move it to the tail of the
941 void mem_cgroup_rotate_reclaimable_page(struct page *page)
943 struct mem_cgroup_per_zone *mz;
944 struct page_cgroup *pc;
945 enum lru_list lru = page_lru(page);
947 if (mem_cgroup_disabled())
950 pc = lookup_page_cgroup(page);
951 /* unused or root page is not rotated. */
952 if (!PageCgroupUsed(pc))
954 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
956 if (mem_cgroup_is_root(pc->mem_cgroup))
958 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
959 list_move_tail(&pc->lru, &mz->lists[lru]);
962 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
964 struct mem_cgroup_per_zone *mz;
965 struct page_cgroup *pc;
967 if (mem_cgroup_disabled())
970 pc = lookup_page_cgroup(page);
971 /* unused or root page is not rotated. */
972 if (!PageCgroupUsed(pc))
974 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
976 if (mem_cgroup_is_root(pc->mem_cgroup))
978 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
979 list_move(&pc->lru, &mz->lists[lru]);
982 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
984 struct page_cgroup *pc;
985 struct mem_cgroup_per_zone *mz;
987 if (mem_cgroup_disabled())
989 pc = lookup_page_cgroup(page);
990 VM_BUG_ON(PageCgroupAcctLRU(pc));
991 if (!PageCgroupUsed(pc))
993 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
995 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
996 /* huge page split is done under lru_lock. so, we have no races. */
997 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
998 SetPageCgroupAcctLRU(pc);
999 if (mem_cgroup_is_root(pc->mem_cgroup))
1001 list_add(&pc->lru, &mz->lists[lru]);
1005 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1006 * while it's linked to lru because the page may be reused after it's fully
1007 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1008 * It's done under lock_page and expected that zone->lru_lock isnever held.
1010 static void mem_cgroup_lru_del_before_commit(struct page *page)
1012 unsigned long flags;
1013 struct zone *zone = page_zone(page);
1014 struct page_cgroup *pc = lookup_page_cgroup(page);
1017 * Doing this check without taking ->lru_lock seems wrong but this
1018 * is safe. Because if page_cgroup's USED bit is unset, the page
1019 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1020 * set, the commit after this will fail, anyway.
1021 * This all charge/uncharge is done under some mutual execustion.
1022 * So, we don't need to taking care of changes in USED bit.
1024 if (likely(!PageLRU(page)))
1027 spin_lock_irqsave(&zone->lru_lock, flags);
1029 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1030 * is guarded by lock_page() because the page is SwapCache.
1032 if (!PageCgroupUsed(pc))
1033 mem_cgroup_del_lru_list(page, page_lru(page));
1034 spin_unlock_irqrestore(&zone->lru_lock, flags);
1037 static void mem_cgroup_lru_add_after_commit(struct page *page)
1039 unsigned long flags;
1040 struct zone *zone = page_zone(page);
1041 struct page_cgroup *pc = lookup_page_cgroup(page);
1043 /* taking care of that the page is added to LRU while we commit it */
1044 if (likely(!PageLRU(page)))
1046 spin_lock_irqsave(&zone->lru_lock, flags);
1047 /* link when the page is linked to LRU but page_cgroup isn't */
1048 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1049 mem_cgroup_add_lru_list(page, page_lru(page));
1050 spin_unlock_irqrestore(&zone->lru_lock, flags);
1054 void mem_cgroup_move_lists(struct page *page,
1055 enum lru_list from, enum lru_list to)
1057 if (mem_cgroup_disabled())
1059 mem_cgroup_del_lru_list(page, from);
1060 mem_cgroup_add_lru_list(page, to);
1063 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
1066 struct mem_cgroup *curr = NULL;
1067 struct task_struct *p;
1069 p = find_lock_task_mm(task);
1072 curr = try_get_mem_cgroup_from_mm(p->mm);
1077 * We should check use_hierarchy of "mem" not "curr". Because checking
1078 * use_hierarchy of "curr" here make this function true if hierarchy is
1079 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
1080 * hierarchy(even if use_hierarchy is disabled in "mem").
1082 if (mem->use_hierarchy)
1083 ret = css_is_ancestor(&curr->css, &mem->css);
1085 ret = (curr == mem);
1086 css_put(&curr->css);
1090 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1092 unsigned long active;
1093 unsigned long inactive;
1095 unsigned long inactive_ratio;
1097 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
1098 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
1100 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1102 inactive_ratio = int_sqrt(10 * gb);
1106 if (present_pages) {
1107 present_pages[0] = inactive;
1108 present_pages[1] = active;
1111 return inactive_ratio;
1114 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1116 unsigned long active;
1117 unsigned long inactive;
1118 unsigned long present_pages[2];
1119 unsigned long inactive_ratio;
1121 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1123 inactive = present_pages[0];
1124 active = present_pages[1];
1126 if (inactive * inactive_ratio < active)
1132 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1134 unsigned long active;
1135 unsigned long inactive;
1137 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
1138 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
1140 return (active > inactive);
1143 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1146 int nid = zone_to_nid(zone);
1147 int zid = zone_idx(zone);
1148 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1150 return &mz->reclaim_stat;
1153 struct zone_reclaim_stat *
1154 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1156 struct page_cgroup *pc;
1157 struct mem_cgroup_per_zone *mz;
1159 if (mem_cgroup_disabled())
1162 pc = lookup_page_cgroup(page);
1163 if (!PageCgroupUsed(pc))
1165 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1167 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1168 return &mz->reclaim_stat;
1171 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1172 struct list_head *dst,
1173 unsigned long *scanned, int order,
1174 int mode, struct zone *z,
1175 struct mem_cgroup *mem_cont,
1176 int active, int file)
1178 unsigned long nr_taken = 0;
1182 struct list_head *src;
1183 struct page_cgroup *pc, *tmp;
1184 int nid = zone_to_nid(z);
1185 int zid = zone_idx(z);
1186 struct mem_cgroup_per_zone *mz;
1187 int lru = LRU_FILE * file + active;
1191 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1192 src = &mz->lists[lru];
1195 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1196 if (scan >= nr_to_scan)
1199 if (unlikely(!PageCgroupUsed(pc)))
1202 page = lookup_cgroup_page(pc);
1204 if (unlikely(!PageLRU(page)))
1208 ret = __isolate_lru_page(page, mode, file);
1211 list_move(&page->lru, dst);
1212 mem_cgroup_del_lru(page);
1213 nr_taken += hpage_nr_pages(page);
1216 /* we don't affect global LRU but rotate in our LRU */
1217 mem_cgroup_rotate_lru_list(page, page_lru(page));
1226 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1232 #define mem_cgroup_from_res_counter(counter, member) \
1233 container_of(counter, struct mem_cgroup, member)
1236 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1237 * @mem: the memory cgroup
1239 * Returns the maximum amount of memory @mem can be charged with, in
1242 static unsigned long mem_cgroup_margin(struct mem_cgroup *mem)
1244 unsigned long long margin;
1246 margin = res_counter_margin(&mem->res);
1247 if (do_swap_account)
1248 margin = min(margin, res_counter_margin(&mem->memsw));
1249 return margin >> PAGE_SHIFT;
1252 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1254 struct cgroup *cgrp = memcg->css.cgroup;
1257 if (cgrp->parent == NULL)
1258 return vm_swappiness;
1260 return memcg->swappiness;
1263 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1268 spin_lock(&mem->pcp_counter_lock);
1269 for_each_online_cpu(cpu)
1270 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1271 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1272 spin_unlock(&mem->pcp_counter_lock);
1278 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1285 spin_lock(&mem->pcp_counter_lock);
1286 for_each_online_cpu(cpu)
1287 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1288 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1289 spin_unlock(&mem->pcp_counter_lock);
1293 * 2 routines for checking "mem" is under move_account() or not.
1295 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1296 * for avoiding race in accounting. If true,
1297 * pc->mem_cgroup may be overwritten.
1299 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1300 * under hierarchy of moving cgroups. This is for
1301 * waiting at hith-memory prressure caused by "move".
1304 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1306 VM_BUG_ON(!rcu_read_lock_held());
1307 return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1310 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1312 struct mem_cgroup *from;
1313 struct mem_cgroup *to;
1316 * Unlike task_move routines, we access mc.to, mc.from not under
1317 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1319 spin_lock(&mc.lock);
1324 if (from == mem || to == mem
1325 || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1326 || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1329 spin_unlock(&mc.lock);
1333 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1335 if (mc.moving_task && current != mc.moving_task) {
1336 if (mem_cgroup_under_move(mem)) {
1338 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1339 /* moving charge context might have finished. */
1342 finish_wait(&mc.waitq, &wait);
1350 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1351 * @memcg: The memory cgroup that went over limit
1352 * @p: Task that is going to be killed
1354 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1357 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1359 struct cgroup *task_cgrp;
1360 struct cgroup *mem_cgrp;
1362 * Need a buffer in BSS, can't rely on allocations. The code relies
1363 * on the assumption that OOM is serialized for memory controller.
1364 * If this assumption is broken, revisit this code.
1366 static char memcg_name[PATH_MAX];
1375 mem_cgrp = memcg->css.cgroup;
1376 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1378 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1381 * Unfortunately, we are unable to convert to a useful name
1382 * But we'll still print out the usage information
1389 printk(KERN_INFO "Task in %s killed", memcg_name);
1392 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1400 * Continues from above, so we don't need an KERN_ level
1402 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1405 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1406 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1407 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1408 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1409 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1411 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1412 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1413 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1417 * This function returns the number of memcg under hierarchy tree. Returns
1418 * 1(self count) if no children.
1420 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1423 struct mem_cgroup *iter;
1425 for_each_mem_cgroup_tree(iter, mem)
1431 * Return the memory (and swap, if configured) limit for a memcg.
1433 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1438 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1439 limit += total_swap_pages << PAGE_SHIFT;
1441 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1443 * If memsw is finite and limits the amount of swap space available
1444 * to this memcg, return that limit.
1446 return min(limit, memsw);
1450 * Visit the first child (need not be the first child as per the ordering
1451 * of the cgroup list, since we track last_scanned_child) of @mem and use
1452 * that to reclaim free pages from.
1454 static struct mem_cgroup *
1455 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1457 struct mem_cgroup *ret = NULL;
1458 struct cgroup_subsys_state *css;
1461 if (!root_mem->use_hierarchy) {
1462 css_get(&root_mem->css);
1468 nextid = root_mem->last_scanned_child + 1;
1469 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1471 if (css && css_tryget(css))
1472 ret = container_of(css, struct mem_cgroup, css);
1475 /* Updates scanning parameter */
1477 /* this means start scan from ID:1 */
1478 root_mem->last_scanned_child = 0;
1480 root_mem->last_scanned_child = found;
1487 * test_mem_cgroup_node_reclaimable
1488 * @mem: the target memcg
1489 * @nid: the node ID to be checked.
1490 * @noswap : specify true here if the user wants flle only information.
1492 * This function returns whether the specified memcg contains any
1493 * reclaimable pages on a node. Returns true if there are any reclaimable
1494 * pages in the node.
1496 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *mem,
1497 int nid, bool noswap)
1499 if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_FILE))
1501 if (noswap || !total_swap_pages)
1503 if (mem_cgroup_node_nr_lru_pages(mem, nid, LRU_ALL_ANON))
1508 #if MAX_NUMNODES > 1
1511 * Always updating the nodemask is not very good - even if we have an empty
1512 * list or the wrong list here, we can start from some node and traverse all
1513 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1516 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
1520 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1521 * pagein/pageout changes since the last update.
1523 if (!atomic_read(&mem->numainfo_events))
1525 if (atomic_inc_return(&mem->numainfo_updating) > 1)
1528 /* make a nodemask where this memcg uses memory from */
1529 mem->scan_nodes = node_states[N_HIGH_MEMORY];
1531 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1533 if (!test_mem_cgroup_node_reclaimable(mem, nid, false))
1534 node_clear(nid, mem->scan_nodes);
1537 atomic_set(&mem->numainfo_events, 0);
1538 atomic_set(&mem->numainfo_updating, 0);
1542 * Selecting a node where we start reclaim from. Because what we need is just
1543 * reducing usage counter, start from anywhere is O,K. Considering
1544 * memory reclaim from current node, there are pros. and cons.
1546 * Freeing memory from current node means freeing memory from a node which
1547 * we'll use or we've used. So, it may make LRU bad. And if several threads
1548 * hit limits, it will see a contention on a node. But freeing from remote
1549 * node means more costs for memory reclaim because of memory latency.
1551 * Now, we use round-robin. Better algorithm is welcomed.
1553 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1557 mem_cgroup_may_update_nodemask(mem);
1558 node = mem->last_scanned_node;
1560 node = next_node(node, mem->scan_nodes);
1561 if (node == MAX_NUMNODES)
1562 node = first_node(mem->scan_nodes);
1564 * We call this when we hit limit, not when pages are added to LRU.
1565 * No LRU may hold pages because all pages are UNEVICTABLE or
1566 * memcg is too small and all pages are not on LRU. In that case,
1567 * we use curret node.
1569 if (unlikely(node == MAX_NUMNODES))
1570 node = numa_node_id();
1572 mem->last_scanned_node = node;
1577 * Check all nodes whether it contains reclaimable pages or not.
1578 * For quick scan, we make use of scan_nodes. This will allow us to skip
1579 * unused nodes. But scan_nodes is lazily updated and may not cotain
1580 * enough new information. We need to do double check.
1582 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1587 * quick check...making use of scan_node.
1588 * We can skip unused nodes.
1590 if (!nodes_empty(mem->scan_nodes)) {
1591 for (nid = first_node(mem->scan_nodes);
1593 nid = next_node(nid, mem->scan_nodes)) {
1595 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1600 * Check rest of nodes.
1602 for_each_node_state(nid, N_HIGH_MEMORY) {
1603 if (node_isset(nid, mem->scan_nodes))
1605 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1612 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1617 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1619 return test_mem_cgroup_node_reclaimable(mem, 0, noswap);
1624 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1625 * we reclaimed from, so that we don't end up penalizing one child extensively
1626 * based on its position in the children list.
1628 * root_mem is the original ancestor that we've been reclaim from.
1630 * We give up and return to the caller when we visit root_mem twice.
1631 * (other groups can be removed while we're walking....)
1633 * If shrink==true, for avoiding to free too much, this returns immedieately.
1635 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1638 unsigned long reclaim_options,
1639 unsigned long *total_scanned)
1641 struct mem_cgroup *victim;
1644 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1645 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1646 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1647 unsigned long excess;
1648 unsigned long nr_scanned;
1650 excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1652 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1653 if (!check_soft && root_mem->memsw_is_minimum)
1657 victim = mem_cgroup_select_victim(root_mem);
1658 if (victim == root_mem) {
1661 * We are not draining per cpu cached charges during
1662 * soft limit reclaim because global reclaim doesn't
1663 * care about charges. It tries to free some memory and
1664 * charges will not give any.
1666 if (!check_soft && loop >= 1)
1667 drain_all_stock_async(root_mem);
1670 * If we have not been able to reclaim
1671 * anything, it might because there are
1672 * no reclaimable pages under this hierarchy
1674 if (!check_soft || !total) {
1675 css_put(&victim->css);
1679 * We want to do more targeted reclaim.
1680 * excess >> 2 is not to excessive so as to
1681 * reclaim too much, nor too less that we keep
1682 * coming back to reclaim from this cgroup
1684 if (total >= (excess >> 2) ||
1685 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1686 css_put(&victim->css);
1691 if (!mem_cgroup_reclaimable(victim, noswap)) {
1692 /* this cgroup's local usage == 0 */
1693 css_put(&victim->css);
1696 /* we use swappiness of local cgroup */
1698 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1699 noswap, zone, &nr_scanned);
1700 *total_scanned += nr_scanned;
1702 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1704 css_put(&victim->css);
1706 * At shrinking usage, we can't check we should stop here or
1707 * reclaim more. It's depends on callers. last_scanned_child
1708 * will work enough for keeping fairness under tree.
1714 if (!res_counter_soft_limit_excess(&root_mem->res))
1716 } else if (mem_cgroup_margin(root_mem))
1723 * Check OOM-Killer is already running under our hierarchy.
1724 * If someone is running, return false.
1726 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1728 int x, lock_count = 0;
1729 struct mem_cgroup *iter;
1731 for_each_mem_cgroup_tree(iter, mem) {
1732 x = atomic_inc_return(&iter->oom_lock);
1733 lock_count = max(x, lock_count);
1736 if (lock_count == 1)
1741 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1743 struct mem_cgroup *iter;
1746 * When a new child is created while the hierarchy is under oom,
1747 * mem_cgroup_oom_lock() may not be called. We have to use
1748 * atomic_add_unless() here.
1750 for_each_mem_cgroup_tree(iter, mem)
1751 atomic_add_unless(&iter->oom_lock, -1, 0);
1756 static DEFINE_MUTEX(memcg_oom_mutex);
1757 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1759 struct oom_wait_info {
1760 struct mem_cgroup *mem;
1764 static int memcg_oom_wake_function(wait_queue_t *wait,
1765 unsigned mode, int sync, void *arg)
1767 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1768 struct oom_wait_info *oom_wait_info;
1770 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1772 if (oom_wait_info->mem == wake_mem)
1774 /* if no hierarchy, no match */
1775 if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1778 * Both of oom_wait_info->mem and wake_mem are stable under us.
1779 * Then we can use css_is_ancestor without taking care of RCU.
1781 if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1782 !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1786 return autoremove_wake_function(wait, mode, sync, arg);
1789 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1791 /* for filtering, pass "mem" as argument. */
1792 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1795 static void memcg_oom_recover(struct mem_cgroup *mem)
1797 if (mem && atomic_read(&mem->oom_lock))
1798 memcg_wakeup_oom(mem);
1802 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1804 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1806 struct oom_wait_info owait;
1807 bool locked, need_to_kill;
1810 owait.wait.flags = 0;
1811 owait.wait.func = memcg_oom_wake_function;
1812 owait.wait.private = current;
1813 INIT_LIST_HEAD(&owait.wait.task_list);
1814 need_to_kill = true;
1815 /* At first, try to OOM lock hierarchy under mem.*/
1816 mutex_lock(&memcg_oom_mutex);
1817 locked = mem_cgroup_oom_lock(mem);
1819 * Even if signal_pending(), we can't quit charge() loop without
1820 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1821 * under OOM is always welcomed, use TASK_KILLABLE here.
1823 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1824 if (!locked || mem->oom_kill_disable)
1825 need_to_kill = false;
1827 mem_cgroup_oom_notify(mem);
1828 mutex_unlock(&memcg_oom_mutex);
1831 finish_wait(&memcg_oom_waitq, &owait.wait);
1832 mem_cgroup_out_of_memory(mem, mask);
1835 finish_wait(&memcg_oom_waitq, &owait.wait);
1837 mutex_lock(&memcg_oom_mutex);
1838 mem_cgroup_oom_unlock(mem);
1839 memcg_wakeup_oom(mem);
1840 mutex_unlock(&memcg_oom_mutex);
1842 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1844 /* Give chance to dying process */
1845 schedule_timeout(1);
1850 * Currently used to update mapped file statistics, but the routine can be
1851 * generalized to update other statistics as well.
1853 * Notes: Race condition
1855 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1856 * it tends to be costly. But considering some conditions, we doesn't need
1857 * to do so _always_.
1859 * Considering "charge", lock_page_cgroup() is not required because all
1860 * file-stat operations happen after a page is attached to radix-tree. There
1861 * are no race with "charge".
1863 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1864 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1865 * if there are race with "uncharge". Statistics itself is properly handled
1868 * Considering "move", this is an only case we see a race. To make the race
1869 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1870 * possibility of race condition. If there is, we take a lock.
1873 void mem_cgroup_update_page_stat(struct page *page,
1874 enum mem_cgroup_page_stat_item idx, int val)
1876 struct mem_cgroup *mem;
1877 struct page_cgroup *pc = lookup_page_cgroup(page);
1878 bool need_unlock = false;
1879 unsigned long uninitialized_var(flags);
1885 mem = pc->mem_cgroup;
1886 if (unlikely(!mem || !PageCgroupUsed(pc)))
1888 /* pc->mem_cgroup is unstable ? */
1889 if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1890 /* take a lock against to access pc->mem_cgroup */
1891 move_lock_page_cgroup(pc, &flags);
1893 mem = pc->mem_cgroup;
1894 if (!mem || !PageCgroupUsed(pc))
1899 case MEMCG_NR_FILE_MAPPED:
1901 SetPageCgroupFileMapped(pc);
1902 else if (!page_mapped(page))
1903 ClearPageCgroupFileMapped(pc);
1904 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1910 this_cpu_add(mem->stat->count[idx], val);
1913 if (unlikely(need_unlock))
1914 move_unlock_page_cgroup(pc, &flags);
1918 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1921 * size of first charge trial. "32" comes from vmscan.c's magic value.
1922 * TODO: maybe necessary to use big numbers in big irons.
1924 #define CHARGE_BATCH 32U
1925 struct memcg_stock_pcp {
1926 struct mem_cgroup *cached; /* this never be root cgroup */
1927 unsigned int nr_pages;
1928 struct work_struct work;
1929 unsigned long flags;
1930 #define FLUSHING_CACHED_CHARGE (0)
1932 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1933 static DEFINE_MUTEX(percpu_charge_mutex);
1936 * Try to consume stocked charge on this cpu. If success, one page is consumed
1937 * from local stock and true is returned. If the stock is 0 or charges from a
1938 * cgroup which is not current target, returns false. This stock will be
1941 static bool consume_stock(struct mem_cgroup *mem)
1943 struct memcg_stock_pcp *stock;
1946 stock = &get_cpu_var(memcg_stock);
1947 if (mem == stock->cached && stock->nr_pages)
1949 else /* need to call res_counter_charge */
1951 put_cpu_var(memcg_stock);
1956 * Returns stocks cached in percpu to res_counter and reset cached information.
1958 static void drain_stock(struct memcg_stock_pcp *stock)
1960 struct mem_cgroup *old = stock->cached;
1962 if (stock->nr_pages) {
1963 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
1965 res_counter_uncharge(&old->res, bytes);
1966 if (do_swap_account)
1967 res_counter_uncharge(&old->memsw, bytes);
1968 stock->nr_pages = 0;
1970 stock->cached = NULL;
1974 * This must be called under preempt disabled or must be called by
1975 * a thread which is pinned to local cpu.
1977 static void drain_local_stock(struct work_struct *dummy)
1979 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1981 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1985 * Cache charges(val) which is from res_counter, to local per_cpu area.
1986 * This will be consumed by consume_stock() function, later.
1988 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
1990 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1992 if (stock->cached != mem) { /* reset if necessary */
1994 stock->cached = mem;
1996 stock->nr_pages += nr_pages;
1997 put_cpu_var(memcg_stock);
2001 * Tries to drain stocked charges in other cpus. This function is asynchronous
2002 * and just put a work per cpu for draining localy on each cpu. Caller can
2003 * expects some charges will be back to res_counter later but cannot wait for
2006 static void drain_all_stock_async(struct mem_cgroup *root_mem)
2010 * If someone calls draining, avoid adding more kworker runs.
2012 if (!mutex_trylock(&percpu_charge_mutex))
2014 /* Notify other cpus that system-wide "drain" is running */
2017 * Get a hint for avoiding draining charges on the current cpu,
2018 * which must be exhausted by our charging. It is not required that
2019 * this be a precise check, so we use raw_smp_processor_id() instead of
2020 * getcpu()/putcpu().
2022 curcpu = raw_smp_processor_id();
2023 for_each_online_cpu(cpu) {
2024 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2025 struct mem_cgroup *mem;
2030 mem = stock->cached;
2033 if (mem != root_mem) {
2034 if (!root_mem->use_hierarchy)
2036 /* check whether "mem" is under tree of "root_mem" */
2037 if (!css_is_ancestor(&mem->css, &root_mem->css))
2040 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2041 schedule_work_on(cpu, &stock->work);
2044 mutex_unlock(&percpu_charge_mutex);
2045 /* We don't wait for flush_work */
2048 /* This is a synchronous drain interface. */
2049 static void drain_all_stock_sync(void)
2051 /* called when force_empty is called */
2052 mutex_lock(&percpu_charge_mutex);
2053 schedule_on_each_cpu(drain_local_stock);
2054 mutex_unlock(&percpu_charge_mutex);
2058 * This function drains percpu counter value from DEAD cpu and
2059 * move it to local cpu. Note that this function can be preempted.
2061 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2065 spin_lock(&mem->pcp_counter_lock);
2066 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2067 long x = per_cpu(mem->stat->count[i], cpu);
2069 per_cpu(mem->stat->count[i], cpu) = 0;
2070 mem->nocpu_base.count[i] += x;
2072 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2073 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2075 per_cpu(mem->stat->events[i], cpu) = 0;
2076 mem->nocpu_base.events[i] += x;
2078 /* need to clear ON_MOVE value, works as a kind of lock. */
2079 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2080 spin_unlock(&mem->pcp_counter_lock);
2083 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2085 int idx = MEM_CGROUP_ON_MOVE;
2087 spin_lock(&mem->pcp_counter_lock);
2088 per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2089 spin_unlock(&mem->pcp_counter_lock);
2092 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2093 unsigned long action,
2096 int cpu = (unsigned long)hcpu;
2097 struct memcg_stock_pcp *stock;
2098 struct mem_cgroup *iter;
2100 if ((action == CPU_ONLINE)) {
2101 for_each_mem_cgroup_all(iter)
2102 synchronize_mem_cgroup_on_move(iter, cpu);
2106 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2109 for_each_mem_cgroup_all(iter)
2110 mem_cgroup_drain_pcp_counter(iter, cpu);
2112 stock = &per_cpu(memcg_stock, cpu);
2118 /* See __mem_cgroup_try_charge() for details */
2120 CHARGE_OK, /* success */
2121 CHARGE_RETRY, /* need to retry but retry is not bad */
2122 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2123 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2124 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2127 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2128 unsigned int nr_pages, bool oom_check)
2130 unsigned long csize = nr_pages * PAGE_SIZE;
2131 struct mem_cgroup *mem_over_limit;
2132 struct res_counter *fail_res;
2133 unsigned long flags = 0;
2136 ret = res_counter_charge(&mem->res, csize, &fail_res);
2139 if (!do_swap_account)
2141 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2145 res_counter_uncharge(&mem->res, csize);
2146 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2147 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2149 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2151 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2152 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2154 * Never reclaim on behalf of optional batching, retry with a
2155 * single page instead.
2157 if (nr_pages == CHARGE_BATCH)
2158 return CHARGE_RETRY;
2160 if (!(gfp_mask & __GFP_WAIT))
2161 return CHARGE_WOULDBLOCK;
2163 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2164 gfp_mask, flags, NULL);
2165 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2166 return CHARGE_RETRY;
2168 * Even though the limit is exceeded at this point, reclaim
2169 * may have been able to free some pages. Retry the charge
2170 * before killing the task.
2172 * Only for regular pages, though: huge pages are rather
2173 * unlikely to succeed so close to the limit, and we fall back
2174 * to regular pages anyway in case of failure.
2176 if (nr_pages == 1 && ret)
2177 return CHARGE_RETRY;
2180 * At task move, charge accounts can be doubly counted. So, it's
2181 * better to wait until the end of task_move if something is going on.
2183 if (mem_cgroup_wait_acct_move(mem_over_limit))
2184 return CHARGE_RETRY;
2186 /* If we don't need to call oom-killer at el, return immediately */
2188 return CHARGE_NOMEM;
2190 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2191 return CHARGE_OOM_DIE;
2193 return CHARGE_RETRY;
2197 * Unlike exported interface, "oom" parameter is added. if oom==true,
2198 * oom-killer can be invoked.
2200 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2202 unsigned int nr_pages,
2203 struct mem_cgroup **memcg,
2206 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2207 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2208 struct mem_cgroup *mem = NULL;
2212 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2213 * in system level. So, allow to go ahead dying process in addition to
2216 if (unlikely(test_thread_flag(TIF_MEMDIE)
2217 || fatal_signal_pending(current)))
2221 * We always charge the cgroup the mm_struct belongs to.
2222 * The mm_struct's mem_cgroup changes on task migration if the
2223 * thread group leader migrates. It's possible that mm is not
2224 * set, if so charge the init_mm (happens for pagecache usage).
2229 if (*memcg) { /* css should be a valid one */
2231 VM_BUG_ON(css_is_removed(&mem->css));
2232 if (mem_cgroup_is_root(mem))
2234 if (nr_pages == 1 && consume_stock(mem))
2238 struct task_struct *p;
2241 p = rcu_dereference(mm->owner);
2243 * Because we don't have task_lock(), "p" can exit.
2244 * In that case, "mem" can point to root or p can be NULL with
2245 * race with swapoff. Then, we have small risk of mis-accouning.
2246 * But such kind of mis-account by race always happens because
2247 * we don't have cgroup_mutex(). It's overkill and we allo that
2249 * (*) swapoff at el will charge against mm-struct not against
2250 * task-struct. So, mm->owner can be NULL.
2252 mem = mem_cgroup_from_task(p);
2253 if (!mem || mem_cgroup_is_root(mem)) {
2257 if (nr_pages == 1 && consume_stock(mem)) {
2259 * It seems dagerous to access memcg without css_get().
2260 * But considering how consume_stok works, it's not
2261 * necessary. If consume_stock success, some charges
2262 * from this memcg are cached on this cpu. So, we
2263 * don't need to call css_get()/css_tryget() before
2264 * calling consume_stock().
2269 /* after here, we may be blocked. we need to get refcnt */
2270 if (!css_tryget(&mem->css)) {
2280 /* If killed, bypass charge */
2281 if (fatal_signal_pending(current)) {
2287 if (oom && !nr_oom_retries) {
2289 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2292 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2296 case CHARGE_RETRY: /* not in OOM situation but retry */
2301 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2304 case CHARGE_NOMEM: /* OOM routine works */
2309 /* If oom, we never return -ENOMEM */
2312 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2316 } while (ret != CHARGE_OK);
2318 if (batch > nr_pages)
2319 refill_stock(mem, batch - nr_pages);
2333 * Somemtimes we have to undo a charge we got by try_charge().
2334 * This function is for that and do uncharge, put css's refcnt.
2335 * gotten by try_charge().
2337 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2338 unsigned int nr_pages)
2340 if (!mem_cgroup_is_root(mem)) {
2341 unsigned long bytes = nr_pages * PAGE_SIZE;
2343 res_counter_uncharge(&mem->res, bytes);
2344 if (do_swap_account)
2345 res_counter_uncharge(&mem->memsw, bytes);
2350 * A helper function to get mem_cgroup from ID. must be called under
2351 * rcu_read_lock(). The caller must check css_is_removed() or some if
2352 * it's concern. (dropping refcnt from swap can be called against removed
2355 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2357 struct cgroup_subsys_state *css;
2359 /* ID 0 is unused ID */
2362 css = css_lookup(&mem_cgroup_subsys, id);
2365 return container_of(css, struct mem_cgroup, css);
2368 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2370 struct mem_cgroup *mem = NULL;
2371 struct page_cgroup *pc;
2375 VM_BUG_ON(!PageLocked(page));
2377 pc = lookup_page_cgroup(page);
2378 lock_page_cgroup(pc);
2379 if (PageCgroupUsed(pc)) {
2380 mem = pc->mem_cgroup;
2381 if (mem && !css_tryget(&mem->css))
2383 } else if (PageSwapCache(page)) {
2384 ent.val = page_private(page);
2385 id = lookup_swap_cgroup(ent);
2387 mem = mem_cgroup_lookup(id);
2388 if (mem && !css_tryget(&mem->css))
2392 unlock_page_cgroup(pc);
2396 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2398 unsigned int nr_pages,
2399 struct page_cgroup *pc,
2400 enum charge_type ctype)
2402 lock_page_cgroup(pc);
2403 if (unlikely(PageCgroupUsed(pc))) {
2404 unlock_page_cgroup(pc);
2405 __mem_cgroup_cancel_charge(mem, nr_pages);
2409 * we don't need page_cgroup_lock about tail pages, becase they are not
2410 * accessed by any other context at this point.
2412 pc->mem_cgroup = mem;
2414 * We access a page_cgroup asynchronously without lock_page_cgroup().
2415 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2416 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2417 * before USED bit, we need memory barrier here.
2418 * See mem_cgroup_add_lru_list(), etc.
2422 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2423 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2424 SetPageCgroupCache(pc);
2425 SetPageCgroupUsed(pc);
2427 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2428 ClearPageCgroupCache(pc);
2429 SetPageCgroupUsed(pc);
2435 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2436 unlock_page_cgroup(pc);
2438 * "charge_statistics" updated event counter. Then, check it.
2439 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2440 * if they exceeds softlimit.
2442 memcg_check_events(mem, page);
2445 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2447 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2448 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2450 * Because tail pages are not marked as "used", set it. We're under
2451 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2453 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2455 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2456 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2457 unsigned long flags;
2459 if (mem_cgroup_disabled())
2462 * We have no races with charge/uncharge but will have races with
2463 * page state accounting.
2465 move_lock_page_cgroup(head_pc, &flags);
2467 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2468 smp_wmb(); /* see __commit_charge() */
2469 if (PageCgroupAcctLRU(head_pc)) {
2471 struct mem_cgroup_per_zone *mz;
2474 * LRU flags cannot be copied because we need to add tail
2475 *.page to LRU by generic call and our hook will be called.
2476 * We hold lru_lock, then, reduce counter directly.
2478 lru = page_lru(head);
2479 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2480 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2482 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2483 move_unlock_page_cgroup(head_pc, &flags);
2488 * mem_cgroup_move_account - move account of the page
2490 * @nr_pages: number of regular pages (>1 for huge pages)
2491 * @pc: page_cgroup of the page.
2492 * @from: mem_cgroup which the page is moved from.
2493 * @to: mem_cgroup which the page is moved to. @from != @to.
2494 * @uncharge: whether we should call uncharge and css_put against @from.
2496 * The caller must confirm following.
2497 * - page is not on LRU (isolate_page() is useful.)
2498 * - compound_lock is held when nr_pages > 1
2500 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2501 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2502 * true, this function does "uncharge" from old cgroup, but it doesn't if
2503 * @uncharge is false, so a caller should do "uncharge".
2505 static int mem_cgroup_move_account(struct page *page,
2506 unsigned int nr_pages,
2507 struct page_cgroup *pc,
2508 struct mem_cgroup *from,
2509 struct mem_cgroup *to,
2512 unsigned long flags;
2515 VM_BUG_ON(from == to);
2516 VM_BUG_ON(PageLRU(page));
2518 * The page is isolated from LRU. So, collapse function
2519 * will not handle this page. But page splitting can happen.
2520 * Do this check under compound_page_lock(). The caller should
2524 if (nr_pages > 1 && !PageTransHuge(page))
2527 lock_page_cgroup(pc);
2530 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2533 move_lock_page_cgroup(pc, &flags);
2535 if (PageCgroupFileMapped(pc)) {
2536 /* Update mapped_file data for mem_cgroup */
2538 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2539 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2542 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2544 /* This is not "cancel", but cancel_charge does all we need. */
2545 __mem_cgroup_cancel_charge(from, nr_pages);
2547 /* caller should have done css_get */
2548 pc->mem_cgroup = to;
2549 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2551 * We charges against "to" which may not have any tasks. Then, "to"
2552 * can be under rmdir(). But in current implementation, caller of
2553 * this function is just force_empty() and move charge, so it's
2554 * guaranteed that "to" is never removed. So, we don't check rmdir
2557 move_unlock_page_cgroup(pc, &flags);
2560 unlock_page_cgroup(pc);
2564 memcg_check_events(to, page);
2565 memcg_check_events(from, page);
2571 * move charges to its parent.
2574 static int mem_cgroup_move_parent(struct page *page,
2575 struct page_cgroup *pc,
2576 struct mem_cgroup *child,
2579 struct cgroup *cg = child->css.cgroup;
2580 struct cgroup *pcg = cg->parent;
2581 struct mem_cgroup *parent;
2582 unsigned int nr_pages;
2583 unsigned long uninitialized_var(flags);
2591 if (!get_page_unless_zero(page))
2593 if (isolate_lru_page(page))
2596 nr_pages = hpage_nr_pages(page);
2598 parent = mem_cgroup_from_cont(pcg);
2599 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2604 flags = compound_lock_irqsave(page);
2606 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2608 __mem_cgroup_cancel_charge(parent, nr_pages);
2611 compound_unlock_irqrestore(page, flags);
2613 putback_lru_page(page);
2621 * Charge the memory controller for page usage.
2623 * 0 if the charge was successful
2624 * < 0 if the cgroup is over its limit
2626 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2627 gfp_t gfp_mask, enum charge_type ctype)
2629 struct mem_cgroup *mem = NULL;
2630 unsigned int nr_pages = 1;
2631 struct page_cgroup *pc;
2635 if (PageTransHuge(page)) {
2636 nr_pages <<= compound_order(page);
2637 VM_BUG_ON(!PageTransHuge(page));
2639 * Never OOM-kill a process for a huge page. The
2640 * fault handler will fall back to regular pages.
2645 pc = lookup_page_cgroup(page);
2646 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2648 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2652 __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2656 int mem_cgroup_newpage_charge(struct page *page,
2657 struct mm_struct *mm, gfp_t gfp_mask)
2659 if (mem_cgroup_disabled())
2662 * If already mapped, we don't have to account.
2663 * If page cache, page->mapping has address_space.
2664 * But page->mapping may have out-of-use anon_vma pointer,
2665 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2668 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2672 return mem_cgroup_charge_common(page, mm, gfp_mask,
2673 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2677 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2678 enum charge_type ctype);
2681 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2682 enum charge_type ctype)
2684 struct page_cgroup *pc = lookup_page_cgroup(page);
2686 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2687 * is already on LRU. It means the page may on some other page_cgroup's
2688 * LRU. Take care of it.
2690 mem_cgroup_lru_del_before_commit(page);
2691 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2692 mem_cgroup_lru_add_after_commit(page);
2696 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2699 struct mem_cgroup *mem = NULL;
2702 if (mem_cgroup_disabled())
2704 if (PageCompound(page))
2707 * Corner case handling. This is called from add_to_page_cache()
2708 * in usual. But some FS (shmem) precharges this page before calling it
2709 * and call add_to_page_cache() with GFP_NOWAIT.
2711 * For GFP_NOWAIT case, the page may be pre-charged before calling
2712 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2713 * charge twice. (It works but has to pay a bit larger cost.)
2714 * And when the page is SwapCache, it should take swap information
2715 * into account. This is under lock_page() now.
2717 if (!(gfp_mask & __GFP_WAIT)) {
2718 struct page_cgroup *pc;
2720 pc = lookup_page_cgroup(page);
2723 lock_page_cgroup(pc);
2724 if (PageCgroupUsed(pc)) {
2725 unlock_page_cgroup(pc);
2728 unlock_page_cgroup(pc);
2734 if (page_is_file_cache(page)) {
2735 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2740 * FUSE reuses pages without going through the final
2741 * put that would remove them from the LRU list, make
2742 * sure that they get relinked properly.
2744 __mem_cgroup_commit_charge_lrucare(page, mem,
2745 MEM_CGROUP_CHARGE_TYPE_CACHE);
2749 if (PageSwapCache(page)) {
2750 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2752 __mem_cgroup_commit_charge_swapin(page, mem,
2753 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2755 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2756 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2762 * While swap-in, try_charge -> commit or cancel, the page is locked.
2763 * And when try_charge() successfully returns, one refcnt to memcg without
2764 * struct page_cgroup is acquired. This refcnt will be consumed by
2765 * "commit()" or removed by "cancel()"
2767 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2769 gfp_t mask, struct mem_cgroup **ptr)
2771 struct mem_cgroup *mem;
2776 if (mem_cgroup_disabled())
2779 if (!do_swap_account)
2782 * A racing thread's fault, or swapoff, may have already updated
2783 * the pte, and even removed page from swap cache: in those cases
2784 * do_swap_page()'s pte_same() test will fail; but there's also a
2785 * KSM case which does need to charge the page.
2787 if (!PageSwapCache(page))
2789 mem = try_get_mem_cgroup_from_page(page);
2793 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2799 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2803 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2804 enum charge_type ctype)
2806 if (mem_cgroup_disabled())
2810 cgroup_exclude_rmdir(&ptr->css);
2812 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2814 * Now swap is on-memory. This means this page may be
2815 * counted both as mem and swap....double count.
2816 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2817 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2818 * may call delete_from_swap_cache() before reach here.
2820 if (do_swap_account && PageSwapCache(page)) {
2821 swp_entry_t ent = {.val = page_private(page)};
2823 struct mem_cgroup *memcg;
2825 id = swap_cgroup_record(ent, 0);
2827 memcg = mem_cgroup_lookup(id);
2830 * This recorded memcg can be obsolete one. So, avoid
2831 * calling css_tryget
2833 if (!mem_cgroup_is_root(memcg))
2834 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2835 mem_cgroup_swap_statistics(memcg, false);
2836 mem_cgroup_put(memcg);
2841 * At swapin, we may charge account against cgroup which has no tasks.
2842 * So, rmdir()->pre_destroy() can be called while we do this charge.
2843 * In that case, we need to call pre_destroy() again. check it here.
2845 cgroup_release_and_wakeup_rmdir(&ptr->css);
2848 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2850 __mem_cgroup_commit_charge_swapin(page, ptr,
2851 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2854 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2856 if (mem_cgroup_disabled())
2860 __mem_cgroup_cancel_charge(mem, 1);
2863 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
2864 unsigned int nr_pages,
2865 const enum charge_type ctype)
2867 struct memcg_batch_info *batch = NULL;
2868 bool uncharge_memsw = true;
2870 /* If swapout, usage of swap doesn't decrease */
2871 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2872 uncharge_memsw = false;
2874 batch = ¤t->memcg_batch;
2876 * In usual, we do css_get() when we remember memcg pointer.
2877 * But in this case, we keep res->usage until end of a series of
2878 * uncharges. Then, it's ok to ignore memcg's refcnt.
2883 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2884 * In those cases, all pages freed continuously can be expected to be in
2885 * the same cgroup and we have chance to coalesce uncharges.
2886 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2887 * because we want to do uncharge as soon as possible.
2890 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2891 goto direct_uncharge;
2894 goto direct_uncharge;
2897 * In typical case, batch->memcg == mem. This means we can
2898 * merge a series of uncharges to an uncharge of res_counter.
2899 * If not, we uncharge res_counter ony by one.
2901 if (batch->memcg != mem)
2902 goto direct_uncharge;
2903 /* remember freed charge and uncharge it later */
2906 batch->memsw_nr_pages++;
2909 res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
2911 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
2912 if (unlikely(batch->memcg != mem))
2913 memcg_oom_recover(mem);
2918 * uncharge if !page_mapped(page)
2920 static struct mem_cgroup *
2921 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2923 struct mem_cgroup *mem = NULL;
2924 unsigned int nr_pages = 1;
2925 struct page_cgroup *pc;
2927 if (mem_cgroup_disabled())
2930 if (PageSwapCache(page))
2933 if (PageTransHuge(page)) {
2934 nr_pages <<= compound_order(page);
2935 VM_BUG_ON(!PageTransHuge(page));
2938 * Check if our page_cgroup is valid
2940 pc = lookup_page_cgroup(page);
2941 if (unlikely(!pc || !PageCgroupUsed(pc)))
2944 lock_page_cgroup(pc);
2946 mem = pc->mem_cgroup;
2948 if (!PageCgroupUsed(pc))
2952 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2953 case MEM_CGROUP_CHARGE_TYPE_DROP:
2954 /* See mem_cgroup_prepare_migration() */
2955 if (page_mapped(page) || PageCgroupMigration(pc))
2958 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2959 if (!PageAnon(page)) { /* Shared memory */
2960 if (page->mapping && !page_is_file_cache(page))
2962 } else if (page_mapped(page)) /* Anon */
2969 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
2971 ClearPageCgroupUsed(pc);
2973 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2974 * freed from LRU. This is safe because uncharged page is expected not
2975 * to be reused (freed soon). Exception is SwapCache, it's handled by
2976 * special functions.
2979 unlock_page_cgroup(pc);
2981 * even after unlock, we have mem->res.usage here and this memcg
2982 * will never be freed.
2984 memcg_check_events(mem, page);
2985 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2986 mem_cgroup_swap_statistics(mem, true);
2987 mem_cgroup_get(mem);
2989 if (!mem_cgroup_is_root(mem))
2990 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
2995 unlock_page_cgroup(pc);
2999 void mem_cgroup_uncharge_page(struct page *page)
3002 if (page_mapped(page))
3004 if (page->mapping && !PageAnon(page))
3006 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3009 void mem_cgroup_uncharge_cache_page(struct page *page)
3011 VM_BUG_ON(page_mapped(page));
3012 VM_BUG_ON(page->mapping);
3013 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3017 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3018 * In that cases, pages are freed continuously and we can expect pages
3019 * are in the same memcg. All these calls itself limits the number of
3020 * pages freed at once, then uncharge_start/end() is called properly.
3021 * This may be called prural(2) times in a context,
3024 void mem_cgroup_uncharge_start(void)
3026 current->memcg_batch.do_batch++;
3027 /* We can do nest. */
3028 if (current->memcg_batch.do_batch == 1) {
3029 current->memcg_batch.memcg = NULL;
3030 current->memcg_batch.nr_pages = 0;
3031 current->memcg_batch.memsw_nr_pages = 0;
3035 void mem_cgroup_uncharge_end(void)
3037 struct memcg_batch_info *batch = ¤t->memcg_batch;
3039 if (!batch->do_batch)
3043 if (batch->do_batch) /* If stacked, do nothing. */
3049 * This "batch->memcg" is valid without any css_get/put etc...
3050 * bacause we hide charges behind us.
3052 if (batch->nr_pages)
3053 res_counter_uncharge(&batch->memcg->res,
3054 batch->nr_pages * PAGE_SIZE);
3055 if (batch->memsw_nr_pages)
3056 res_counter_uncharge(&batch->memcg->memsw,
3057 batch->memsw_nr_pages * PAGE_SIZE);
3058 memcg_oom_recover(batch->memcg);
3059 /* forget this pointer (for sanity check) */
3060 batch->memcg = NULL;
3065 * called after __delete_from_swap_cache() and drop "page" account.
3066 * memcg information is recorded to swap_cgroup of "ent"
3069 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3071 struct mem_cgroup *memcg;
3072 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3074 if (!swapout) /* this was a swap cache but the swap is unused ! */
3075 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3077 memcg = __mem_cgroup_uncharge_common(page, ctype);
3080 * record memcg information, if swapout && memcg != NULL,
3081 * mem_cgroup_get() was called in uncharge().
3083 if (do_swap_account && swapout && memcg)
3084 swap_cgroup_record(ent, css_id(&memcg->css));
3088 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3090 * called from swap_entry_free(). remove record in swap_cgroup and
3091 * uncharge "memsw" account.
3093 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3095 struct mem_cgroup *memcg;
3098 if (!do_swap_account)
3101 id = swap_cgroup_record(ent, 0);
3103 memcg = mem_cgroup_lookup(id);
3106 * We uncharge this because swap is freed.
3107 * This memcg can be obsolete one. We avoid calling css_tryget
3109 if (!mem_cgroup_is_root(memcg))
3110 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3111 mem_cgroup_swap_statistics(memcg, false);
3112 mem_cgroup_put(memcg);
3118 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3119 * @entry: swap entry to be moved
3120 * @from: mem_cgroup which the entry is moved from
3121 * @to: mem_cgroup which the entry is moved to
3122 * @need_fixup: whether we should fixup res_counters and refcounts.
3124 * It succeeds only when the swap_cgroup's record for this entry is the same
3125 * as the mem_cgroup's id of @from.
3127 * Returns 0 on success, -EINVAL on failure.
3129 * The caller must have charged to @to, IOW, called res_counter_charge() about
3130 * both res and memsw, and called css_get().
3132 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3133 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3135 unsigned short old_id, new_id;
3137 old_id = css_id(&from->css);
3138 new_id = css_id(&to->css);
3140 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3141 mem_cgroup_swap_statistics(from, false);
3142 mem_cgroup_swap_statistics(to, true);
3144 * This function is only called from task migration context now.
3145 * It postpones res_counter and refcount handling till the end
3146 * of task migration(mem_cgroup_clear_mc()) for performance
3147 * improvement. But we cannot postpone mem_cgroup_get(to)
3148 * because if the process that has been moved to @to does
3149 * swap-in, the refcount of @to might be decreased to 0.
3153 if (!mem_cgroup_is_root(from))
3154 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3155 mem_cgroup_put(from);
3157 * we charged both to->res and to->memsw, so we should
3160 if (!mem_cgroup_is_root(to))
3161 res_counter_uncharge(&to->res, PAGE_SIZE);
3168 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3169 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3176 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3179 int mem_cgroup_prepare_migration(struct page *page,
3180 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3182 struct mem_cgroup *mem = NULL;
3183 struct page_cgroup *pc;
3184 enum charge_type ctype;
3189 VM_BUG_ON(PageTransHuge(page));
3190 if (mem_cgroup_disabled())
3193 pc = lookup_page_cgroup(page);
3194 lock_page_cgroup(pc);
3195 if (PageCgroupUsed(pc)) {
3196 mem = pc->mem_cgroup;
3199 * At migrating an anonymous page, its mapcount goes down
3200 * to 0 and uncharge() will be called. But, even if it's fully
3201 * unmapped, migration may fail and this page has to be
3202 * charged again. We set MIGRATION flag here and delay uncharge
3203 * until end_migration() is called
3205 * Corner Case Thinking
3207 * When the old page was mapped as Anon and it's unmap-and-freed
3208 * while migration was ongoing.
3209 * If unmap finds the old page, uncharge() of it will be delayed
3210 * until end_migration(). If unmap finds a new page, it's
3211 * uncharged when it make mapcount to be 1->0. If unmap code
3212 * finds swap_migration_entry, the new page will not be mapped
3213 * and end_migration() will find it(mapcount==0).
3216 * When the old page was mapped but migraion fails, the kernel
3217 * remaps it. A charge for it is kept by MIGRATION flag even
3218 * if mapcount goes down to 0. We can do remap successfully
3219 * without charging it again.
3222 * The "old" page is under lock_page() until the end of
3223 * migration, so, the old page itself will not be swapped-out.
3224 * If the new page is swapped out before end_migraton, our
3225 * hook to usual swap-out path will catch the event.
3228 SetPageCgroupMigration(pc);
3230 unlock_page_cgroup(pc);
3232 * If the page is not charged at this point,
3239 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3240 css_put(&mem->css);/* drop extra refcnt */
3241 if (ret || *ptr == NULL) {
3242 if (PageAnon(page)) {
3243 lock_page_cgroup(pc);
3244 ClearPageCgroupMigration(pc);
3245 unlock_page_cgroup(pc);
3247 * The old page may be fully unmapped while we kept it.
3249 mem_cgroup_uncharge_page(page);
3254 * We charge new page before it's used/mapped. So, even if unlock_page()
3255 * is called before end_migration, we can catch all events on this new
3256 * page. In the case new page is migrated but not remapped, new page's
3257 * mapcount will be finally 0 and we call uncharge in end_migration().
3259 pc = lookup_page_cgroup(newpage);
3261 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3262 else if (page_is_file_cache(page))
3263 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3265 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3266 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3270 /* remove redundant charge if migration failed*/
3271 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3272 struct page *oldpage, struct page *newpage, bool migration_ok)
3274 struct page *used, *unused;
3275 struct page_cgroup *pc;
3279 /* blocks rmdir() */
3280 cgroup_exclude_rmdir(&mem->css);
3281 if (!migration_ok) {
3289 * We disallowed uncharge of pages under migration because mapcount
3290 * of the page goes down to zero, temporarly.
3291 * Clear the flag and check the page should be charged.
3293 pc = lookup_page_cgroup(oldpage);
3294 lock_page_cgroup(pc);
3295 ClearPageCgroupMigration(pc);
3296 unlock_page_cgroup(pc);
3298 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3301 * If a page is a file cache, radix-tree replacement is very atomic
3302 * and we can skip this check. When it was an Anon page, its mapcount
3303 * goes down to 0. But because we added MIGRATION flage, it's not
3304 * uncharged yet. There are several case but page->mapcount check
3305 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3306 * check. (see prepare_charge() also)
3309 mem_cgroup_uncharge_page(used);
3311 * At migration, we may charge account against cgroup which has no
3313 * So, rmdir()->pre_destroy() can be called while we do this charge.
3314 * In that case, we need to call pre_destroy() again. check it here.
3316 cgroup_release_and_wakeup_rmdir(&mem->css);
3320 * A call to try to shrink memory usage on charge failure at shmem's swapin.
3321 * Calling hierarchical_reclaim is not enough because we should update
3322 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3323 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3324 * not from the memcg which this page would be charged to.
3325 * try_charge_swapin does all of these works properly.
3327 int mem_cgroup_shmem_charge_fallback(struct page *page,
3328 struct mm_struct *mm,
3331 struct mem_cgroup *mem;
3334 if (mem_cgroup_disabled())
3337 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3339 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3344 #ifdef CONFIG_DEBUG_VM
3345 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3347 struct page_cgroup *pc;
3349 pc = lookup_page_cgroup(page);
3350 if (likely(pc) && PageCgroupUsed(pc))
3355 bool mem_cgroup_bad_page_check(struct page *page)
3357 if (mem_cgroup_disabled())
3360 return lookup_page_cgroup_used(page) != NULL;
3363 void mem_cgroup_print_bad_page(struct page *page)
3365 struct page_cgroup *pc;
3367 pc = lookup_page_cgroup_used(page);
3372 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3373 pc, pc->flags, pc->mem_cgroup);
3375 path = kmalloc(PATH_MAX, GFP_KERNEL);
3378 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3383 printk(KERN_CONT "(%s)\n",
3384 (ret < 0) ? "cannot get the path" : path);
3390 static DEFINE_MUTEX(set_limit_mutex);
3392 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3393 unsigned long long val)
3396 u64 memswlimit, memlimit;
3398 int children = mem_cgroup_count_children(memcg);
3399 u64 curusage, oldusage;
3403 * For keeping hierarchical_reclaim simple, how long we should retry
3404 * is depends on callers. We set our retry-count to be function
3405 * of # of children which we should visit in this loop.
3407 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3409 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3412 while (retry_count) {
3413 if (signal_pending(current)) {
3418 * Rather than hide all in some function, I do this in
3419 * open coded manner. You see what this really does.
3420 * We have to guarantee mem->res.limit < mem->memsw.limit.
3422 mutex_lock(&set_limit_mutex);
3423 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3424 if (memswlimit < val) {
3426 mutex_unlock(&set_limit_mutex);
3430 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3434 ret = res_counter_set_limit(&memcg->res, val);
3436 if (memswlimit == val)
3437 memcg->memsw_is_minimum = true;
3439 memcg->memsw_is_minimum = false;
3441 mutex_unlock(&set_limit_mutex);
3446 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3447 MEM_CGROUP_RECLAIM_SHRINK,
3449 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3450 /* Usage is reduced ? */
3451 if (curusage >= oldusage)
3454 oldusage = curusage;
3456 if (!ret && enlarge)
3457 memcg_oom_recover(memcg);
3462 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3463 unsigned long long val)
3466 u64 memlimit, memswlimit, oldusage, curusage;
3467 int children = mem_cgroup_count_children(memcg);
3471 /* see mem_cgroup_resize_res_limit */
3472 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3473 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3474 while (retry_count) {
3475 if (signal_pending(current)) {
3480 * Rather than hide all in some function, I do this in
3481 * open coded manner. You see what this really does.
3482 * We have to guarantee mem->res.limit < mem->memsw.limit.
3484 mutex_lock(&set_limit_mutex);
3485 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3486 if (memlimit > val) {
3488 mutex_unlock(&set_limit_mutex);
3491 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3492 if (memswlimit < val)
3494 ret = res_counter_set_limit(&memcg->memsw, val);
3496 if (memlimit == val)
3497 memcg->memsw_is_minimum = true;
3499 memcg->memsw_is_minimum = false;
3501 mutex_unlock(&set_limit_mutex);
3506 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3507 MEM_CGROUP_RECLAIM_NOSWAP |
3508 MEM_CGROUP_RECLAIM_SHRINK,
3510 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3511 /* Usage is reduced ? */
3512 if (curusage >= oldusage)
3515 oldusage = curusage;
3517 if (!ret && enlarge)
3518 memcg_oom_recover(memcg);
3522 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3524 unsigned long *total_scanned)
3526 unsigned long nr_reclaimed = 0;
3527 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3528 unsigned long reclaimed;
3530 struct mem_cgroup_tree_per_zone *mctz;
3531 unsigned long long excess;
3532 unsigned long nr_scanned;
3537 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3539 * This loop can run a while, specially if mem_cgroup's continuously
3540 * keep exceeding their soft limit and putting the system under
3547 mz = mem_cgroup_largest_soft_limit_node(mctz);
3552 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3554 MEM_CGROUP_RECLAIM_SOFT,
3556 nr_reclaimed += reclaimed;
3557 *total_scanned += nr_scanned;
3558 spin_lock(&mctz->lock);
3561 * If we failed to reclaim anything from this memory cgroup
3562 * it is time to move on to the next cgroup
3568 * Loop until we find yet another one.
3570 * By the time we get the soft_limit lock
3571 * again, someone might have aded the
3572 * group back on the RB tree. Iterate to
3573 * make sure we get a different mem.
3574 * mem_cgroup_largest_soft_limit_node returns
3575 * NULL if no other cgroup is present on
3579 __mem_cgroup_largest_soft_limit_node(mctz);
3581 css_put(&next_mz->mem->css);
3582 else /* next_mz == NULL or other memcg */
3586 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3587 excess = res_counter_soft_limit_excess(&mz->mem->res);
3589 * One school of thought says that we should not add
3590 * back the node to the tree if reclaim returns 0.
3591 * But our reclaim could return 0, simply because due
3592 * to priority we are exposing a smaller subset of
3593 * memory to reclaim from. Consider this as a longer
3596 /* If excess == 0, no tree ops */
3597 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3598 spin_unlock(&mctz->lock);
3599 css_put(&mz->mem->css);
3602 * Could not reclaim anything and there are no more
3603 * mem cgroups to try or we seem to be looping without
3604 * reclaiming anything.
3606 if (!nr_reclaimed &&
3608 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3610 } while (!nr_reclaimed);
3612 css_put(&next_mz->mem->css);
3613 return nr_reclaimed;
3617 * This routine traverse page_cgroup in given list and drop them all.
3618 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3620 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3621 int node, int zid, enum lru_list lru)
3624 struct mem_cgroup_per_zone *mz;
3625 struct page_cgroup *pc, *busy;
3626 unsigned long flags, loop;
3627 struct list_head *list;
3630 zone = &NODE_DATA(node)->node_zones[zid];
3631 mz = mem_cgroup_zoneinfo(mem, node, zid);
3632 list = &mz->lists[lru];
3634 loop = MEM_CGROUP_ZSTAT(mz, lru);
3635 /* give some margin against EBUSY etc...*/
3642 spin_lock_irqsave(&zone->lru_lock, flags);
3643 if (list_empty(list)) {
3644 spin_unlock_irqrestore(&zone->lru_lock, flags);
3647 pc = list_entry(list->prev, struct page_cgroup, lru);
3649 list_move(&pc->lru, list);
3651 spin_unlock_irqrestore(&zone->lru_lock, flags);
3654 spin_unlock_irqrestore(&zone->lru_lock, flags);
3656 page = lookup_cgroup_page(pc);
3658 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3662 if (ret == -EBUSY || ret == -EINVAL) {
3663 /* found lock contention or "pc" is obsolete. */
3670 if (!ret && !list_empty(list))
3676 * make mem_cgroup's charge to be 0 if there is no task.
3677 * This enables deleting this mem_cgroup.
3679 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3682 int node, zid, shrink;
3683 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3684 struct cgroup *cgrp = mem->css.cgroup;
3689 /* should free all ? */
3695 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3698 if (signal_pending(current))
3700 /* This is for making all *used* pages to be on LRU. */
3701 lru_add_drain_all();
3702 drain_all_stock_sync();
3704 mem_cgroup_start_move(mem);
3705 for_each_node_state(node, N_HIGH_MEMORY) {
3706 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3709 ret = mem_cgroup_force_empty_list(mem,
3718 mem_cgroup_end_move(mem);
3719 memcg_oom_recover(mem);
3720 /* it seems parent cgroup doesn't have enough mem */
3724 /* "ret" should also be checked to ensure all lists are empty. */
3725 } while (mem->res.usage > 0 || ret);
3731 /* returns EBUSY if there is a task or if we come here twice. */
3732 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3736 /* we call try-to-free pages for make this cgroup empty */
3737 lru_add_drain_all();
3738 /* try to free all pages in this cgroup */
3740 while (nr_retries && mem->res.usage > 0) {
3743 if (signal_pending(current)) {
3747 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3751 /* maybe some writeback is necessary */
3752 congestion_wait(BLK_RW_ASYNC, HZ/10);
3757 /* try move_account...there may be some *locked* pages. */
3761 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3763 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3767 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3769 return mem_cgroup_from_cont(cont)->use_hierarchy;
3772 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3776 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3777 struct cgroup *parent = cont->parent;
3778 struct mem_cgroup *parent_mem = NULL;
3781 parent_mem = mem_cgroup_from_cont(parent);
3785 * If parent's use_hierarchy is set, we can't make any modifications
3786 * in the child subtrees. If it is unset, then the change can
3787 * occur, provided the current cgroup has no children.
3789 * For the root cgroup, parent_mem is NULL, we allow value to be
3790 * set if there are no children.
3792 if ((!parent_mem || !parent_mem->use_hierarchy) &&
3793 (val == 1 || val == 0)) {
3794 if (list_empty(&cont->children))
3795 mem->use_hierarchy = val;
3806 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3807 enum mem_cgroup_stat_index idx)
3809 struct mem_cgroup *iter;
3812 /* Per-cpu values can be negative, use a signed accumulator */
3813 for_each_mem_cgroup_tree(iter, mem)
3814 val += mem_cgroup_read_stat(iter, idx);
3816 if (val < 0) /* race ? */
3821 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3825 if (!mem_cgroup_is_root(mem)) {
3827 return res_counter_read_u64(&mem->res, RES_USAGE);
3829 return res_counter_read_u64(&mem->memsw, RES_USAGE);
3832 val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
3833 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
3836 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3838 return val << PAGE_SHIFT;
3841 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3843 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3847 type = MEMFILE_TYPE(cft->private);
3848 name = MEMFILE_ATTR(cft->private);
3851 if (name == RES_USAGE)
3852 val = mem_cgroup_usage(mem, false);
3854 val = res_counter_read_u64(&mem->res, name);
3857 if (name == RES_USAGE)
3858 val = mem_cgroup_usage(mem, true);
3860 val = res_counter_read_u64(&mem->memsw, name);
3869 * The user of this function is...
3872 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3875 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3877 unsigned long long val;
3880 type = MEMFILE_TYPE(cft->private);
3881 name = MEMFILE_ATTR(cft->private);
3884 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3888 /* This function does all necessary parse...reuse it */
3889 ret = res_counter_memparse_write_strategy(buffer, &val);
3893 ret = mem_cgroup_resize_limit(memcg, val);
3895 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3897 case RES_SOFT_LIMIT:
3898 ret = res_counter_memparse_write_strategy(buffer, &val);
3902 * For memsw, soft limits are hard to implement in terms
3903 * of semantics, for now, we support soft limits for
3904 * control without swap
3907 ret = res_counter_set_soft_limit(&memcg->res, val);
3912 ret = -EINVAL; /* should be BUG() ? */
3918 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3919 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3921 struct cgroup *cgroup;
3922 unsigned long long min_limit, min_memsw_limit, tmp;
3924 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3925 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3926 cgroup = memcg->css.cgroup;
3927 if (!memcg->use_hierarchy)
3930 while (cgroup->parent) {
3931 cgroup = cgroup->parent;
3932 memcg = mem_cgroup_from_cont(cgroup);
3933 if (!memcg->use_hierarchy)
3935 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3936 min_limit = min(min_limit, tmp);
3937 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3938 min_memsw_limit = min(min_memsw_limit, tmp);
3941 *mem_limit = min_limit;
3942 *memsw_limit = min_memsw_limit;
3946 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3948 struct mem_cgroup *mem;
3951 mem = mem_cgroup_from_cont(cont);
3952 type = MEMFILE_TYPE(event);
3953 name = MEMFILE_ATTR(event);
3957 res_counter_reset_max(&mem->res);
3959 res_counter_reset_max(&mem->memsw);
3963 res_counter_reset_failcnt(&mem->res);
3965 res_counter_reset_failcnt(&mem->memsw);
3972 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3975 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3979 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3980 struct cftype *cft, u64 val)
3982 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3984 if (val >= (1 << NR_MOVE_TYPE))
3987 * We check this value several times in both in can_attach() and
3988 * attach(), so we need cgroup lock to prevent this value from being
3992 mem->move_charge_at_immigrate = val;
3998 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3999 struct cftype *cft, u64 val)
4006 /* For read statistics */
4024 struct mcs_total_stat {
4025 s64 stat[NR_MCS_STAT];
4031 } memcg_stat_strings[NR_MCS_STAT] = {
4032 {"cache", "total_cache"},
4033 {"rss", "total_rss"},
4034 {"mapped_file", "total_mapped_file"},
4035 {"pgpgin", "total_pgpgin"},
4036 {"pgpgout", "total_pgpgout"},
4037 {"swap", "total_swap"},
4038 {"pgfault", "total_pgfault"},
4039 {"pgmajfault", "total_pgmajfault"},
4040 {"inactive_anon", "total_inactive_anon"},
4041 {"active_anon", "total_active_anon"},
4042 {"inactive_file", "total_inactive_file"},
4043 {"active_file", "total_active_file"},
4044 {"unevictable", "total_unevictable"}
4049 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4054 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4055 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4056 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4057 s->stat[MCS_RSS] += val * PAGE_SIZE;
4058 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4059 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4060 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4061 s->stat[MCS_PGPGIN] += val;
4062 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4063 s->stat[MCS_PGPGOUT] += val;
4064 if (do_swap_account) {
4065 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4066 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4068 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4069 s->stat[MCS_PGFAULT] += val;
4070 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4071 s->stat[MCS_PGMAJFAULT] += val;
4074 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_ANON));
4075 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4076 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_ANON));
4077 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4078 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_INACTIVE_FILE));
4079 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4080 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_ACTIVE_FILE));
4081 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4082 val = mem_cgroup_nr_lru_pages(mem, BIT(LRU_UNEVICTABLE));
4083 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4087 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4089 struct mem_cgroup *iter;
4091 for_each_mem_cgroup_tree(iter, mem)
4092 mem_cgroup_get_local_stat(iter, s);
4096 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4099 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4100 unsigned long node_nr;
4101 struct cgroup *cont = m->private;
4102 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4104 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4105 seq_printf(m, "total=%lu", total_nr);
4106 for_each_node_state(nid, N_HIGH_MEMORY) {
4107 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4108 seq_printf(m, " N%d=%lu", nid, node_nr);
4112 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4113 seq_printf(m, "file=%lu", file_nr);
4114 for_each_node_state(nid, N_HIGH_MEMORY) {
4115 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4117 seq_printf(m, " N%d=%lu", nid, node_nr);
4121 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4122 seq_printf(m, "anon=%lu", anon_nr);
4123 for_each_node_state(nid, N_HIGH_MEMORY) {
4124 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4126 seq_printf(m, " N%d=%lu", nid, node_nr);
4130 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4131 seq_printf(m, "unevictable=%lu", unevictable_nr);
4132 for_each_node_state(nid, N_HIGH_MEMORY) {
4133 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4134 BIT(LRU_UNEVICTABLE));
4135 seq_printf(m, " N%d=%lu", nid, node_nr);
4140 #endif /* CONFIG_NUMA */
4142 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4143 struct cgroup_map_cb *cb)
4145 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4146 struct mcs_total_stat mystat;
4149 memset(&mystat, 0, sizeof(mystat));
4150 mem_cgroup_get_local_stat(mem_cont, &mystat);
4153 for (i = 0; i < NR_MCS_STAT; i++) {
4154 if (i == MCS_SWAP && !do_swap_account)
4156 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4159 /* Hierarchical information */
4161 unsigned long long limit, memsw_limit;
4162 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4163 cb->fill(cb, "hierarchical_memory_limit", limit);
4164 if (do_swap_account)
4165 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4168 memset(&mystat, 0, sizeof(mystat));
4169 mem_cgroup_get_total_stat(mem_cont, &mystat);
4170 for (i = 0; i < NR_MCS_STAT; i++) {
4171 if (i == MCS_SWAP && !do_swap_account)
4173 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4176 #ifdef CONFIG_DEBUG_VM
4177 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4181 struct mem_cgroup_per_zone *mz;
4182 unsigned long recent_rotated[2] = {0, 0};
4183 unsigned long recent_scanned[2] = {0, 0};
4185 for_each_online_node(nid)
4186 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4187 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4189 recent_rotated[0] +=
4190 mz->reclaim_stat.recent_rotated[0];
4191 recent_rotated[1] +=
4192 mz->reclaim_stat.recent_rotated[1];
4193 recent_scanned[0] +=
4194 mz->reclaim_stat.recent_scanned[0];
4195 recent_scanned[1] +=
4196 mz->reclaim_stat.recent_scanned[1];
4198 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4199 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4200 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4201 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4208 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4210 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4212 return mem_cgroup_swappiness(memcg);
4215 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4218 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4219 struct mem_cgroup *parent;
4224 if (cgrp->parent == NULL)
4227 parent = mem_cgroup_from_cont(cgrp->parent);
4231 /* If under hierarchy, only empty-root can set this value */
4232 if ((parent->use_hierarchy) ||
4233 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4238 memcg->swappiness = val;
4245 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4247 struct mem_cgroup_threshold_ary *t;
4253 t = rcu_dereference(memcg->thresholds.primary);
4255 t = rcu_dereference(memcg->memsw_thresholds.primary);
4260 usage = mem_cgroup_usage(memcg, swap);
4263 * current_threshold points to threshold just below usage.
4264 * If it's not true, a threshold was crossed after last
4265 * call of __mem_cgroup_threshold().
4267 i = t->current_threshold;
4270 * Iterate backward over array of thresholds starting from
4271 * current_threshold and check if a threshold is crossed.
4272 * If none of thresholds below usage is crossed, we read
4273 * only one element of the array here.
4275 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4276 eventfd_signal(t->entries[i].eventfd, 1);
4278 /* i = current_threshold + 1 */
4282 * Iterate forward over array of thresholds starting from
4283 * current_threshold+1 and check if a threshold is crossed.
4284 * If none of thresholds above usage is crossed, we read
4285 * only one element of the array here.
4287 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4288 eventfd_signal(t->entries[i].eventfd, 1);
4290 /* Update current_threshold */
4291 t->current_threshold = i - 1;
4296 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4299 __mem_cgroup_threshold(memcg, false);
4300 if (do_swap_account)
4301 __mem_cgroup_threshold(memcg, true);
4303 memcg = parent_mem_cgroup(memcg);
4307 static int compare_thresholds(const void *a, const void *b)
4309 const struct mem_cgroup_threshold *_a = a;
4310 const struct mem_cgroup_threshold *_b = b;
4312 return _a->threshold - _b->threshold;
4315 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4317 struct mem_cgroup_eventfd_list *ev;
4319 list_for_each_entry(ev, &mem->oom_notify, list)
4320 eventfd_signal(ev->eventfd, 1);
4324 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4326 struct mem_cgroup *iter;
4328 for_each_mem_cgroup_tree(iter, mem)
4329 mem_cgroup_oom_notify_cb(iter);
4332 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4333 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4335 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4336 struct mem_cgroup_thresholds *thresholds;
4337 struct mem_cgroup_threshold_ary *new;
4338 int type = MEMFILE_TYPE(cft->private);
4339 u64 threshold, usage;
4342 ret = res_counter_memparse_write_strategy(args, &threshold);
4346 mutex_lock(&memcg->thresholds_lock);
4349 thresholds = &memcg->thresholds;
4350 else if (type == _MEMSWAP)
4351 thresholds = &memcg->memsw_thresholds;
4355 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4357 /* Check if a threshold crossed before adding a new one */
4358 if (thresholds->primary)
4359 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4361 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4363 /* Allocate memory for new array of thresholds */
4364 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4372 /* Copy thresholds (if any) to new array */
4373 if (thresholds->primary) {
4374 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4375 sizeof(struct mem_cgroup_threshold));
4378 /* Add new threshold */
4379 new->entries[size - 1].eventfd = eventfd;
4380 new->entries[size - 1].threshold = threshold;
4382 /* Sort thresholds. Registering of new threshold isn't time-critical */
4383 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4384 compare_thresholds, NULL);
4386 /* Find current threshold */
4387 new->current_threshold = -1;
4388 for (i = 0; i < size; i++) {
4389 if (new->entries[i].threshold < usage) {
4391 * new->current_threshold will not be used until
4392 * rcu_assign_pointer(), so it's safe to increment
4395 ++new->current_threshold;
4399 /* Free old spare buffer and save old primary buffer as spare */
4400 kfree(thresholds->spare);
4401 thresholds->spare = thresholds->primary;
4403 rcu_assign_pointer(thresholds->primary, new);
4405 /* To be sure that nobody uses thresholds */
4409 mutex_unlock(&memcg->thresholds_lock);
4414 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4415 struct cftype *cft, struct eventfd_ctx *eventfd)
4417 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4418 struct mem_cgroup_thresholds *thresholds;
4419 struct mem_cgroup_threshold_ary *new;
4420 int type = MEMFILE_TYPE(cft->private);
4424 mutex_lock(&memcg->thresholds_lock);
4426 thresholds = &memcg->thresholds;
4427 else if (type == _MEMSWAP)
4428 thresholds = &memcg->memsw_thresholds;
4433 * Something went wrong if we trying to unregister a threshold
4434 * if we don't have thresholds
4436 BUG_ON(!thresholds);
4438 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4440 /* Check if a threshold crossed before removing */
4441 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4443 /* Calculate new number of threshold */
4445 for (i = 0; i < thresholds->primary->size; i++) {
4446 if (thresholds->primary->entries[i].eventfd != eventfd)
4450 new = thresholds->spare;
4452 /* Set thresholds array to NULL if we don't have thresholds */
4461 /* Copy thresholds and find current threshold */
4462 new->current_threshold = -1;
4463 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4464 if (thresholds->primary->entries[i].eventfd == eventfd)
4467 new->entries[j] = thresholds->primary->entries[i];
4468 if (new->entries[j].threshold < usage) {
4470 * new->current_threshold will not be used
4471 * until rcu_assign_pointer(), so it's safe to increment
4474 ++new->current_threshold;
4480 /* Swap primary and spare array */
4481 thresholds->spare = thresholds->primary;
4482 rcu_assign_pointer(thresholds->primary, new);
4484 /* To be sure that nobody uses thresholds */
4487 mutex_unlock(&memcg->thresholds_lock);
4490 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4491 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4493 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4494 struct mem_cgroup_eventfd_list *event;
4495 int type = MEMFILE_TYPE(cft->private);
4497 BUG_ON(type != _OOM_TYPE);
4498 event = kmalloc(sizeof(*event), GFP_KERNEL);
4502 mutex_lock(&memcg_oom_mutex);
4504 event->eventfd = eventfd;
4505 list_add(&event->list, &memcg->oom_notify);
4507 /* already in OOM ? */
4508 if (atomic_read(&memcg->oom_lock))
4509 eventfd_signal(eventfd, 1);
4510 mutex_unlock(&memcg_oom_mutex);
4515 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4516 struct cftype *cft, struct eventfd_ctx *eventfd)
4518 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4519 struct mem_cgroup_eventfd_list *ev, *tmp;
4520 int type = MEMFILE_TYPE(cft->private);
4522 BUG_ON(type != _OOM_TYPE);
4524 mutex_lock(&memcg_oom_mutex);
4526 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4527 if (ev->eventfd == eventfd) {
4528 list_del(&ev->list);
4533 mutex_unlock(&memcg_oom_mutex);
4536 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4537 struct cftype *cft, struct cgroup_map_cb *cb)
4539 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4541 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4543 if (atomic_read(&mem->oom_lock))
4544 cb->fill(cb, "under_oom", 1);
4546 cb->fill(cb, "under_oom", 0);
4550 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4551 struct cftype *cft, u64 val)
4553 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4554 struct mem_cgroup *parent;
4556 /* cannot set to root cgroup and only 0 and 1 are allowed */
4557 if (!cgrp->parent || !((val == 0) || (val == 1)))
4560 parent = mem_cgroup_from_cont(cgrp->parent);
4563 /* oom-kill-disable is a flag for subhierarchy. */
4564 if ((parent->use_hierarchy) ||
4565 (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4569 mem->oom_kill_disable = val;
4571 memcg_oom_recover(mem);
4577 static const struct file_operations mem_control_numa_stat_file_operations = {
4579 .llseek = seq_lseek,
4580 .release = single_release,
4583 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4585 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4587 file->f_op = &mem_control_numa_stat_file_operations;
4588 return single_open(file, mem_control_numa_stat_show, cont);
4590 #endif /* CONFIG_NUMA */
4592 static struct cftype mem_cgroup_files[] = {
4594 .name = "usage_in_bytes",
4595 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4596 .read_u64 = mem_cgroup_read,
4597 .register_event = mem_cgroup_usage_register_event,
4598 .unregister_event = mem_cgroup_usage_unregister_event,
4601 .name = "max_usage_in_bytes",
4602 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4603 .trigger = mem_cgroup_reset,
4604 .read_u64 = mem_cgroup_read,
4607 .name = "limit_in_bytes",
4608 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4609 .write_string = mem_cgroup_write,
4610 .read_u64 = mem_cgroup_read,
4613 .name = "soft_limit_in_bytes",
4614 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4615 .write_string = mem_cgroup_write,
4616 .read_u64 = mem_cgroup_read,
4620 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4621 .trigger = mem_cgroup_reset,
4622 .read_u64 = mem_cgroup_read,
4626 .read_map = mem_control_stat_show,
4629 .name = "force_empty",
4630 .trigger = mem_cgroup_force_empty_write,
4633 .name = "use_hierarchy",
4634 .write_u64 = mem_cgroup_hierarchy_write,
4635 .read_u64 = mem_cgroup_hierarchy_read,
4638 .name = "swappiness",
4639 .read_u64 = mem_cgroup_swappiness_read,
4640 .write_u64 = mem_cgroup_swappiness_write,
4643 .name = "move_charge_at_immigrate",
4644 .read_u64 = mem_cgroup_move_charge_read,
4645 .write_u64 = mem_cgroup_move_charge_write,
4648 .name = "oom_control",
4649 .read_map = mem_cgroup_oom_control_read,
4650 .write_u64 = mem_cgroup_oom_control_write,
4651 .register_event = mem_cgroup_oom_register_event,
4652 .unregister_event = mem_cgroup_oom_unregister_event,
4653 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4657 .name = "numa_stat",
4658 .open = mem_control_numa_stat_open,
4664 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4665 static struct cftype memsw_cgroup_files[] = {
4667 .name = "memsw.usage_in_bytes",
4668 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4669 .read_u64 = mem_cgroup_read,
4670 .register_event = mem_cgroup_usage_register_event,
4671 .unregister_event = mem_cgroup_usage_unregister_event,
4674 .name = "memsw.max_usage_in_bytes",
4675 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4676 .trigger = mem_cgroup_reset,
4677 .read_u64 = mem_cgroup_read,
4680 .name = "memsw.limit_in_bytes",
4681 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4682 .write_string = mem_cgroup_write,
4683 .read_u64 = mem_cgroup_read,
4686 .name = "memsw.failcnt",
4687 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4688 .trigger = mem_cgroup_reset,
4689 .read_u64 = mem_cgroup_read,
4693 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4695 if (!do_swap_account)
4697 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4698 ARRAY_SIZE(memsw_cgroup_files));
4701 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4707 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4709 struct mem_cgroup_per_node *pn;
4710 struct mem_cgroup_per_zone *mz;
4712 int zone, tmp = node;
4714 * This routine is called against possible nodes.
4715 * But it's BUG to call kmalloc() against offline node.
4717 * TODO: this routine can waste much memory for nodes which will
4718 * never be onlined. It's better to use memory hotplug callback
4721 if (!node_state(node, N_NORMAL_MEMORY))
4723 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4727 mem->info.nodeinfo[node] = pn;
4728 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4729 mz = &pn->zoneinfo[zone];
4731 INIT_LIST_HEAD(&mz->lists[l]);
4732 mz->usage_in_excess = 0;
4733 mz->on_tree = false;
4739 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4741 kfree(mem->info.nodeinfo[node]);
4744 static struct mem_cgroup *mem_cgroup_alloc(void)
4746 struct mem_cgroup *mem;
4747 int size = sizeof(struct mem_cgroup);
4749 /* Can be very big if MAX_NUMNODES is very big */
4750 if (size < PAGE_SIZE)
4751 mem = kzalloc(size, GFP_KERNEL);
4753 mem = vzalloc(size);
4758 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4761 spin_lock_init(&mem->pcp_counter_lock);
4765 if (size < PAGE_SIZE)
4773 * At destroying mem_cgroup, references from swap_cgroup can remain.
4774 * (scanning all at force_empty is too costly...)
4776 * Instead of clearing all references at force_empty, we remember
4777 * the number of reference from swap_cgroup and free mem_cgroup when
4778 * it goes down to 0.
4780 * Removal of cgroup itself succeeds regardless of refs from swap.
4783 static void __mem_cgroup_free(struct mem_cgroup *mem)
4787 mem_cgroup_remove_from_trees(mem);
4788 free_css_id(&mem_cgroup_subsys, &mem->css);
4790 for_each_node_state(node, N_POSSIBLE)
4791 free_mem_cgroup_per_zone_info(mem, node);
4793 free_percpu(mem->stat);
4794 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4800 static void mem_cgroup_get(struct mem_cgroup *mem)
4802 atomic_inc(&mem->refcnt);
4805 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4807 if (atomic_sub_and_test(count, &mem->refcnt)) {
4808 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4809 __mem_cgroup_free(mem);
4811 mem_cgroup_put(parent);
4815 static void mem_cgroup_put(struct mem_cgroup *mem)
4817 __mem_cgroup_put(mem, 1);
4821 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4823 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4825 if (!mem->res.parent)
4827 return mem_cgroup_from_res_counter(mem->res.parent, res);
4830 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4831 static void __init enable_swap_cgroup(void)
4833 if (!mem_cgroup_disabled() && really_do_swap_account)
4834 do_swap_account = 1;
4837 static void __init enable_swap_cgroup(void)
4842 static int mem_cgroup_soft_limit_tree_init(void)
4844 struct mem_cgroup_tree_per_node *rtpn;
4845 struct mem_cgroup_tree_per_zone *rtpz;
4846 int tmp, node, zone;
4848 for_each_node_state(node, N_POSSIBLE) {
4850 if (!node_state(node, N_NORMAL_MEMORY))
4852 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4856 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4858 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4859 rtpz = &rtpn->rb_tree_per_zone[zone];
4860 rtpz->rb_root = RB_ROOT;
4861 spin_lock_init(&rtpz->lock);
4867 static struct cgroup_subsys_state * __ref
4868 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4870 struct mem_cgroup *mem, *parent;
4871 long error = -ENOMEM;
4874 mem = mem_cgroup_alloc();
4876 return ERR_PTR(error);
4878 for_each_node_state(node, N_POSSIBLE)
4879 if (alloc_mem_cgroup_per_zone_info(mem, node))
4883 if (cont->parent == NULL) {
4885 enable_swap_cgroup();
4887 root_mem_cgroup = mem;
4888 if (mem_cgroup_soft_limit_tree_init())
4890 for_each_possible_cpu(cpu) {
4891 struct memcg_stock_pcp *stock =
4892 &per_cpu(memcg_stock, cpu);
4893 INIT_WORK(&stock->work, drain_local_stock);
4895 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4897 parent = mem_cgroup_from_cont(cont->parent);
4898 mem->use_hierarchy = parent->use_hierarchy;
4899 mem->oom_kill_disable = parent->oom_kill_disable;
4902 if (parent && parent->use_hierarchy) {
4903 res_counter_init(&mem->res, &parent->res);
4904 res_counter_init(&mem->memsw, &parent->memsw);
4906 * We increment refcnt of the parent to ensure that we can
4907 * safely access it on res_counter_charge/uncharge.
4908 * This refcnt will be decremented when freeing this
4909 * mem_cgroup(see mem_cgroup_put).
4911 mem_cgroup_get(parent);
4913 res_counter_init(&mem->res, NULL);
4914 res_counter_init(&mem->memsw, NULL);
4916 mem->last_scanned_child = 0;
4917 mem->last_scanned_node = MAX_NUMNODES;
4918 INIT_LIST_HEAD(&mem->oom_notify);
4921 mem->swappiness = mem_cgroup_swappiness(parent);
4922 atomic_set(&mem->refcnt, 1);
4923 mem->move_charge_at_immigrate = 0;
4924 mutex_init(&mem->thresholds_lock);
4927 __mem_cgroup_free(mem);
4928 root_mem_cgroup = NULL;
4929 return ERR_PTR(error);
4932 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4933 struct cgroup *cont)
4935 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4937 return mem_cgroup_force_empty(mem, false);
4940 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4941 struct cgroup *cont)
4943 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4945 mem_cgroup_put(mem);
4948 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4949 struct cgroup *cont)
4953 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4954 ARRAY_SIZE(mem_cgroup_files));
4957 ret = register_memsw_files(cont, ss);
4962 /* Handlers for move charge at task migration. */
4963 #define PRECHARGE_COUNT_AT_ONCE 256
4964 static int mem_cgroup_do_precharge(unsigned long count)
4967 int batch_count = PRECHARGE_COUNT_AT_ONCE;
4968 struct mem_cgroup *mem = mc.to;
4970 if (mem_cgroup_is_root(mem)) {
4971 mc.precharge += count;
4972 /* we don't need css_get for root */
4975 /* try to charge at once */
4977 struct res_counter *dummy;
4979 * "mem" cannot be under rmdir() because we've already checked
4980 * by cgroup_lock_live_cgroup() that it is not removed and we
4981 * are still under the same cgroup_mutex. So we can postpone
4984 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4986 if (do_swap_account && res_counter_charge(&mem->memsw,
4987 PAGE_SIZE * count, &dummy)) {
4988 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4991 mc.precharge += count;
4995 /* fall back to one by one charge */
4997 if (signal_pending(current)) {
5001 if (!batch_count--) {
5002 batch_count = PRECHARGE_COUNT_AT_ONCE;
5005 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5007 /* mem_cgroup_clear_mc() will do uncharge later */
5015 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5016 * @vma: the vma the pte to be checked belongs
5017 * @addr: the address corresponding to the pte to be checked
5018 * @ptent: the pte to be checked
5019 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5022 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5023 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5024 * move charge. if @target is not NULL, the page is stored in target->page
5025 * with extra refcnt got(Callers should handle it).
5026 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5027 * target for charge migration. if @target is not NULL, the entry is stored
5030 * Called with pte lock held.
5037 enum mc_target_type {
5038 MC_TARGET_NONE, /* not used */
5043 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5044 unsigned long addr, pte_t ptent)
5046 struct page *page = vm_normal_page(vma, addr, ptent);
5048 if (!page || !page_mapped(page))
5050 if (PageAnon(page)) {
5051 /* we don't move shared anon */
5052 if (!move_anon() || page_mapcount(page) > 2)
5054 } else if (!move_file())
5055 /* we ignore mapcount for file pages */
5057 if (!get_page_unless_zero(page))
5063 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5064 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5067 struct page *page = NULL;
5068 swp_entry_t ent = pte_to_swp_entry(ptent);
5070 if (!move_anon() || non_swap_entry(ent))
5072 usage_count = mem_cgroup_count_swap_user(ent, &page);
5073 if (usage_count > 1) { /* we don't move shared anon */
5078 if (do_swap_account)
5079 entry->val = ent.val;
5084 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5085 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5087 struct page *page = NULL;
5088 struct inode *inode;
5089 struct address_space *mapping;
5092 if (!vma->vm_file) /* anonymous vma */
5097 inode = vma->vm_file->f_path.dentry->d_inode;
5098 mapping = vma->vm_file->f_mapping;
5099 if (pte_none(ptent))
5100 pgoff = linear_page_index(vma, addr);
5101 else /* pte_file(ptent) is true */
5102 pgoff = pte_to_pgoff(ptent);
5104 /* page is moved even if it's not RSS of this task(page-faulted). */
5105 if (!mapping_cap_swap_backed(mapping)) { /* normal file */
5106 page = find_get_page(mapping, pgoff);
5107 } else { /* shmem/tmpfs file. we should take account of swap too. */
5109 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
5110 if (do_swap_account)
5111 entry->val = ent.val;
5117 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5118 unsigned long addr, pte_t ptent, union mc_target *target)
5120 struct page *page = NULL;
5121 struct page_cgroup *pc;
5123 swp_entry_t ent = { .val = 0 };
5125 if (pte_present(ptent))
5126 page = mc_handle_present_pte(vma, addr, ptent);
5127 else if (is_swap_pte(ptent))
5128 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5129 else if (pte_none(ptent) || pte_file(ptent))
5130 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5132 if (!page && !ent.val)
5135 pc = lookup_page_cgroup(page);
5137 * Do only loose check w/o page_cgroup lock.
5138 * mem_cgroup_move_account() checks the pc is valid or not under
5141 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5142 ret = MC_TARGET_PAGE;
5144 target->page = page;
5146 if (!ret || !target)
5149 /* There is a swap entry and a page doesn't exist or isn't charged */
5150 if (ent.val && !ret &&
5151 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5152 ret = MC_TARGET_SWAP;
5159 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5160 unsigned long addr, unsigned long end,
5161 struct mm_walk *walk)
5163 struct vm_area_struct *vma = walk->private;
5167 split_huge_page_pmd(walk->mm, pmd);
5169 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5170 for (; addr != end; pte++, addr += PAGE_SIZE)
5171 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5172 mc.precharge++; /* increment precharge temporarily */
5173 pte_unmap_unlock(pte - 1, ptl);
5179 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5181 unsigned long precharge;
5182 struct vm_area_struct *vma;
5184 down_read(&mm->mmap_sem);
5185 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5186 struct mm_walk mem_cgroup_count_precharge_walk = {
5187 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5191 if (is_vm_hugetlb_page(vma))
5193 walk_page_range(vma->vm_start, vma->vm_end,
5194 &mem_cgroup_count_precharge_walk);
5196 up_read(&mm->mmap_sem);
5198 precharge = mc.precharge;
5204 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5206 unsigned long precharge = mem_cgroup_count_precharge(mm);
5208 VM_BUG_ON(mc.moving_task);
5209 mc.moving_task = current;
5210 return mem_cgroup_do_precharge(precharge);
5213 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5214 static void __mem_cgroup_clear_mc(void)
5216 struct mem_cgroup *from = mc.from;
5217 struct mem_cgroup *to = mc.to;
5219 /* we must uncharge all the leftover precharges from mc.to */
5221 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5225 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5226 * we must uncharge here.
5228 if (mc.moved_charge) {
5229 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5230 mc.moved_charge = 0;
5232 /* we must fixup refcnts and charges */
5233 if (mc.moved_swap) {
5234 /* uncharge swap account from the old cgroup */
5235 if (!mem_cgroup_is_root(mc.from))
5236 res_counter_uncharge(&mc.from->memsw,
5237 PAGE_SIZE * mc.moved_swap);
5238 __mem_cgroup_put(mc.from, mc.moved_swap);
5240 if (!mem_cgroup_is_root(mc.to)) {
5242 * we charged both to->res and to->memsw, so we should
5245 res_counter_uncharge(&mc.to->res,
5246 PAGE_SIZE * mc.moved_swap);
5248 /* we've already done mem_cgroup_get(mc.to) */
5251 memcg_oom_recover(from);
5252 memcg_oom_recover(to);
5253 wake_up_all(&mc.waitq);
5256 static void mem_cgroup_clear_mc(void)
5258 struct mem_cgroup *from = mc.from;
5261 * we must clear moving_task before waking up waiters at the end of
5264 mc.moving_task = NULL;
5265 __mem_cgroup_clear_mc();
5266 spin_lock(&mc.lock);
5269 spin_unlock(&mc.lock);
5270 mem_cgroup_end_move(from);
5273 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5274 struct cgroup *cgroup,
5275 struct task_struct *p)
5278 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5280 if (mem->move_charge_at_immigrate) {
5281 struct mm_struct *mm;
5282 struct mem_cgroup *from = mem_cgroup_from_task(p);
5284 VM_BUG_ON(from == mem);
5286 mm = get_task_mm(p);
5289 /* We move charges only when we move a owner of the mm */
5290 if (mm->owner == p) {
5293 VM_BUG_ON(mc.precharge);
5294 VM_BUG_ON(mc.moved_charge);
5295 VM_BUG_ON(mc.moved_swap);
5296 mem_cgroup_start_move(from);
5297 spin_lock(&mc.lock);
5300 spin_unlock(&mc.lock);
5301 /* We set mc.moving_task later */
5303 ret = mem_cgroup_precharge_mc(mm);
5305 mem_cgroup_clear_mc();
5312 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5313 struct cgroup *cgroup,
5314 struct task_struct *p)
5316 mem_cgroup_clear_mc();
5319 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5320 unsigned long addr, unsigned long end,
5321 struct mm_walk *walk)
5324 struct vm_area_struct *vma = walk->private;
5328 split_huge_page_pmd(walk->mm, pmd);
5330 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5331 for (; addr != end; addr += PAGE_SIZE) {
5332 pte_t ptent = *(pte++);
5333 union mc_target target;
5336 struct page_cgroup *pc;
5342 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5344 case MC_TARGET_PAGE:
5346 if (isolate_lru_page(page))
5348 pc = lookup_page_cgroup(page);
5349 if (!mem_cgroup_move_account(page, 1, pc,
5350 mc.from, mc.to, false)) {
5352 /* we uncharge from mc.from later. */
5355 putback_lru_page(page);
5356 put: /* is_target_pte_for_mc() gets the page */
5359 case MC_TARGET_SWAP:
5361 if (!mem_cgroup_move_swap_account(ent,
5362 mc.from, mc.to, false)) {
5364 /* we fixup refcnts and charges later. */
5372 pte_unmap_unlock(pte - 1, ptl);
5377 * We have consumed all precharges we got in can_attach().
5378 * We try charge one by one, but don't do any additional
5379 * charges to mc.to if we have failed in charge once in attach()
5382 ret = mem_cgroup_do_precharge(1);
5390 static void mem_cgroup_move_charge(struct mm_struct *mm)
5392 struct vm_area_struct *vma;
5394 lru_add_drain_all();
5396 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5398 * Someone who are holding the mmap_sem might be waiting in
5399 * waitq. So we cancel all extra charges, wake up all waiters,
5400 * and retry. Because we cancel precharges, we might not be able
5401 * to move enough charges, but moving charge is a best-effort
5402 * feature anyway, so it wouldn't be a big problem.
5404 __mem_cgroup_clear_mc();
5408 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5410 struct mm_walk mem_cgroup_move_charge_walk = {
5411 .pmd_entry = mem_cgroup_move_charge_pte_range,
5415 if (is_vm_hugetlb_page(vma))
5417 ret = walk_page_range(vma->vm_start, vma->vm_end,
5418 &mem_cgroup_move_charge_walk);
5421 * means we have consumed all precharges and failed in
5422 * doing additional charge. Just abandon here.
5426 up_read(&mm->mmap_sem);
5429 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5430 struct cgroup *cont,
5431 struct cgroup *old_cont,
5432 struct task_struct *p)
5434 struct mm_struct *mm = get_task_mm(p);
5438 mem_cgroup_move_charge(mm);
5443 mem_cgroup_clear_mc();
5445 #else /* !CONFIG_MMU */
5446 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5447 struct cgroup *cgroup,
5448 struct task_struct *p)
5452 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5453 struct cgroup *cgroup,
5454 struct task_struct *p)
5457 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5458 struct cgroup *cont,
5459 struct cgroup *old_cont,
5460 struct task_struct *p)
5465 struct cgroup_subsys mem_cgroup_subsys = {
5467 .subsys_id = mem_cgroup_subsys_id,
5468 .create = mem_cgroup_create,
5469 .pre_destroy = mem_cgroup_pre_destroy,
5470 .destroy = mem_cgroup_destroy,
5471 .populate = mem_cgroup_populate,
5472 .can_attach = mem_cgroup_can_attach,
5473 .cancel_attach = mem_cgroup_cancel_attach,
5474 .attach = mem_cgroup_move_task,
5479 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5480 static int __init enable_swap_account(char *s)
5482 /* consider enabled if no parameter or 1 is given */
5483 if (!strcmp(s, "1"))
5484 really_do_swap_account = 1;
5485 else if (!strcmp(s, "0"))
5486 really_do_swap_account = 0;
5489 __setup("swapaccount=", enable_swap_account);