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/slab.h>
39 #include <linux/swap.h>
40 #include <linux/swapops.h>
41 #include <linux/spinlock.h>
42 #include <linux/eventfd.h>
43 #include <linux/sort.h>
45 #include <linux/seq_file.h>
46 #include <linux/vmalloc.h>
47 #include <linux/mm_inline.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/cpu.h>
50 #include <linux/oom.h>
53 #include <asm/uaccess.h>
55 #include <trace/events/vmscan.h>
57 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
58 #define MEM_CGROUP_RECLAIM_RETRIES 5
59 struct mem_cgroup *root_mem_cgroup __read_mostly;
61 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
62 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
63 int do_swap_account __read_mostly;
65 /* for remember boot option*/
66 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
67 static int really_do_swap_account __initdata = 1;
69 static int really_do_swap_account __initdata = 0;
73 #define do_swap_account (0)
78 * Statistics for memory cgroup.
80 enum mem_cgroup_stat_index {
82 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
84 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
85 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
86 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
87 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
88 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
89 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */
90 MEM_CGROUP_STAT_NSTATS,
93 enum mem_cgroup_events_index {
94 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
95 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
96 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
97 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
98 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
99 MEM_CGROUP_EVENTS_NSTATS,
102 * Per memcg event counter is incremented at every pagein/pageout. With THP,
103 * it will be incremated by the number of pages. This counter is used for
104 * for trigger some periodic events. This is straightforward and better
105 * than using jiffies etc. to handle periodic memcg event.
107 enum mem_cgroup_events_target {
108 MEM_CGROUP_TARGET_THRESH,
109 MEM_CGROUP_TARGET_SOFTLIMIT,
110 MEM_CGROUP_TARGET_NUMAINFO,
113 #define THRESHOLDS_EVENTS_TARGET (128)
114 #define SOFTLIMIT_EVENTS_TARGET (1024)
115 #define NUMAINFO_EVENTS_TARGET (1024)
117 struct mem_cgroup_stat_cpu {
118 long count[MEM_CGROUP_STAT_NSTATS];
119 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
120 unsigned long targets[MEM_CGROUP_NTARGETS];
124 * per-zone information in memory controller.
126 struct mem_cgroup_per_zone {
128 * spin_lock to protect the per cgroup LRU
130 struct list_head lists[NR_LRU_LISTS];
131 unsigned long count[NR_LRU_LISTS];
133 struct zone_reclaim_stat reclaim_stat;
134 struct rb_node tree_node; /* RB tree node */
135 unsigned long long usage_in_excess;/* Set to the value by which */
136 /* the soft limit is exceeded*/
138 struct mem_cgroup *mem; /* Back pointer, we cannot */
139 /* use container_of */
141 /* Macro for accessing counter */
142 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
144 struct mem_cgroup_per_node {
145 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
148 struct mem_cgroup_lru_info {
149 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
153 * Cgroups above their limits are maintained in a RB-Tree, independent of
154 * their hierarchy representation
157 struct mem_cgroup_tree_per_zone {
158 struct rb_root rb_root;
162 struct mem_cgroup_tree_per_node {
163 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
166 struct mem_cgroup_tree {
167 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
170 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
172 struct mem_cgroup_threshold {
173 struct eventfd_ctx *eventfd;
178 struct mem_cgroup_threshold_ary {
179 /* An array index points to threshold just below usage. */
180 int current_threshold;
181 /* Size of entries[] */
183 /* Array of thresholds */
184 struct mem_cgroup_threshold entries[0];
187 struct mem_cgroup_thresholds {
188 /* Primary thresholds array */
189 struct mem_cgroup_threshold_ary *primary;
191 * Spare threshold array.
192 * This is needed to make mem_cgroup_unregister_event() "never fail".
193 * It must be able to store at least primary->size - 1 entries.
195 struct mem_cgroup_threshold_ary *spare;
199 struct mem_cgroup_eventfd_list {
200 struct list_head list;
201 struct eventfd_ctx *eventfd;
204 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
205 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
208 * The memory controller data structure. The memory controller controls both
209 * page cache and RSS per cgroup. We would eventually like to provide
210 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
211 * to help the administrator determine what knobs to tune.
213 * TODO: Add a water mark for the memory controller. Reclaim will begin when
214 * we hit the water mark. May be even add a low water mark, such that
215 * no reclaim occurs from a cgroup at it's low water mark, this is
216 * a feature that will be implemented much later in the future.
219 struct cgroup_subsys_state css;
221 * the counter to account for memory usage
223 struct res_counter res;
225 * the counter to account for mem+swap usage.
227 struct res_counter memsw;
229 * Per cgroup active and inactive list, similar to the
230 * per zone LRU lists.
232 struct mem_cgroup_lru_info info;
234 * While reclaiming in a hierarchy, we cache the last child we
237 int last_scanned_child;
238 int last_scanned_node;
240 nodemask_t scan_nodes;
241 atomic_t numainfo_events;
242 atomic_t numainfo_updating;
245 * Should the accounting and control be hierarchical, per subtree?
255 /* OOM-Killer disable */
256 int oom_kill_disable;
258 /* set when res.limit == memsw.limit */
259 bool memsw_is_minimum;
261 /* protect arrays of thresholds */
262 struct mutex thresholds_lock;
264 /* thresholds for memory usage. RCU-protected */
265 struct mem_cgroup_thresholds thresholds;
267 /* thresholds for mem+swap usage. RCU-protected */
268 struct mem_cgroup_thresholds memsw_thresholds;
270 /* For oom notifier event fd */
271 struct list_head oom_notify;
274 * Should we move charges of a task when a task is moved into this
275 * mem_cgroup ? And what type of charges should we move ?
277 unsigned long move_charge_at_immigrate;
281 struct mem_cgroup_stat_cpu *stat;
283 * used when a cpu is offlined or other synchronizations
284 * See mem_cgroup_read_stat().
286 struct mem_cgroup_stat_cpu nocpu_base;
287 spinlock_t pcp_counter_lock;
290 /* Stuffs for move charges at task migration. */
292 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
293 * left-shifted bitmap of these types.
296 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
297 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
301 /* "mc" and its members are protected by cgroup_mutex */
302 static struct move_charge_struct {
303 spinlock_t lock; /* for from, to */
304 struct mem_cgroup *from;
305 struct mem_cgroup *to;
306 unsigned long precharge;
307 unsigned long moved_charge;
308 unsigned long moved_swap;
309 struct task_struct *moving_task; /* a task moving charges */
310 wait_queue_head_t waitq; /* a waitq for other context */
312 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
313 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
316 static bool move_anon(void)
318 return test_bit(MOVE_CHARGE_TYPE_ANON,
319 &mc.to->move_charge_at_immigrate);
322 static bool move_file(void)
324 return test_bit(MOVE_CHARGE_TYPE_FILE,
325 &mc.to->move_charge_at_immigrate);
329 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
330 * limit reclaim to prevent infinite loops, if they ever occur.
332 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
333 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
336 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
337 MEM_CGROUP_CHARGE_TYPE_MAPPED,
338 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
339 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
340 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
341 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
345 /* for encoding cft->private value on file */
348 #define _OOM_TYPE (2)
349 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
350 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
351 #define MEMFILE_ATTR(val) ((val) & 0xffff)
352 /* Used for OOM nofiier */
353 #define OOM_CONTROL (0)
356 * Reclaim flags for mem_cgroup_hierarchical_reclaim
358 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
359 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
360 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
361 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
362 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
363 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
365 static void mem_cgroup_get(struct mem_cgroup *memcg);
366 static void mem_cgroup_put(struct mem_cgroup *memcg);
367 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg);
368 static void drain_all_stock_async(struct mem_cgroup *memcg);
370 static struct mem_cgroup_per_zone *
371 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
373 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
376 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
381 static struct mem_cgroup_per_zone *
382 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
384 int nid = page_to_nid(page);
385 int zid = page_zonenum(page);
387 return mem_cgroup_zoneinfo(memcg, nid, zid);
390 static struct mem_cgroup_tree_per_zone *
391 soft_limit_tree_node_zone(int nid, int zid)
393 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
396 static struct mem_cgroup_tree_per_zone *
397 soft_limit_tree_from_page(struct page *page)
399 int nid = page_to_nid(page);
400 int zid = page_zonenum(page);
402 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
406 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
407 struct mem_cgroup_per_zone *mz,
408 struct mem_cgroup_tree_per_zone *mctz,
409 unsigned long long new_usage_in_excess)
411 struct rb_node **p = &mctz->rb_root.rb_node;
412 struct rb_node *parent = NULL;
413 struct mem_cgroup_per_zone *mz_node;
418 mz->usage_in_excess = new_usage_in_excess;
419 if (!mz->usage_in_excess)
423 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
425 if (mz->usage_in_excess < mz_node->usage_in_excess)
428 * We can't avoid mem cgroups that are over their soft
429 * limit by the same amount
431 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
434 rb_link_node(&mz->tree_node, parent, p);
435 rb_insert_color(&mz->tree_node, &mctz->rb_root);
440 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
441 struct mem_cgroup_per_zone *mz,
442 struct mem_cgroup_tree_per_zone *mctz)
446 rb_erase(&mz->tree_node, &mctz->rb_root);
451 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
452 struct mem_cgroup_per_zone *mz,
453 struct mem_cgroup_tree_per_zone *mctz)
455 spin_lock(&mctz->lock);
456 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
457 spin_unlock(&mctz->lock);
461 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
463 unsigned long long excess;
464 struct mem_cgroup_per_zone *mz;
465 struct mem_cgroup_tree_per_zone *mctz;
466 int nid = page_to_nid(page);
467 int zid = page_zonenum(page);
468 mctz = soft_limit_tree_from_page(page);
471 * Necessary to update all ancestors when hierarchy is used.
472 * because their event counter is not touched.
474 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
475 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
476 excess = res_counter_soft_limit_excess(&memcg->res);
478 * We have to update the tree if mz is on RB-tree or
479 * mem is over its softlimit.
481 if (excess || mz->on_tree) {
482 spin_lock(&mctz->lock);
483 /* if on-tree, remove it */
485 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
487 * Insert again. mz->usage_in_excess will be updated.
488 * If excess is 0, no tree ops.
490 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
491 spin_unlock(&mctz->lock);
496 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
499 struct mem_cgroup_per_zone *mz;
500 struct mem_cgroup_tree_per_zone *mctz;
502 for_each_node_state(node, N_POSSIBLE) {
503 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
504 mz = mem_cgroup_zoneinfo(memcg, node, zone);
505 mctz = soft_limit_tree_node_zone(node, zone);
506 mem_cgroup_remove_exceeded(memcg, mz, mctz);
511 static struct mem_cgroup_per_zone *
512 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
514 struct rb_node *rightmost = NULL;
515 struct mem_cgroup_per_zone *mz;
519 rightmost = rb_last(&mctz->rb_root);
521 goto done; /* Nothing to reclaim from */
523 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
525 * Remove the node now but someone else can add it back,
526 * we will to add it back at the end of reclaim to its correct
527 * position in the tree.
529 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
530 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
531 !css_tryget(&mz->mem->css))
537 static struct mem_cgroup_per_zone *
538 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
540 struct mem_cgroup_per_zone *mz;
542 spin_lock(&mctz->lock);
543 mz = __mem_cgroup_largest_soft_limit_node(mctz);
544 spin_unlock(&mctz->lock);
549 * Implementation Note: reading percpu statistics for memcg.
551 * Both of vmstat[] and percpu_counter has threshold and do periodic
552 * synchronization to implement "quick" read. There are trade-off between
553 * reading cost and precision of value. Then, we may have a chance to implement
554 * a periodic synchronizion of counter in memcg's counter.
556 * But this _read() function is used for user interface now. The user accounts
557 * memory usage by memory cgroup and he _always_ requires exact value because
558 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
559 * have to visit all online cpus and make sum. So, for now, unnecessary
560 * synchronization is not implemented. (just implemented for cpu hotplug)
562 * If there are kernel internal actions which can make use of some not-exact
563 * value, and reading all cpu value can be performance bottleneck in some
564 * common workload, threashold and synchonization as vmstat[] should be
567 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
568 enum mem_cgroup_stat_index idx)
574 for_each_online_cpu(cpu)
575 val += per_cpu(memcg->stat->count[idx], cpu);
576 #ifdef CONFIG_HOTPLUG_CPU
577 spin_lock(&memcg->pcp_counter_lock);
578 val += memcg->nocpu_base.count[idx];
579 spin_unlock(&memcg->pcp_counter_lock);
585 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
588 int val = (charge) ? 1 : -1;
589 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
592 void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val)
594 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
597 void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val)
599 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
602 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
603 enum mem_cgroup_events_index idx)
605 unsigned long val = 0;
608 for_each_online_cpu(cpu)
609 val += per_cpu(memcg->stat->events[idx], cpu);
610 #ifdef CONFIG_HOTPLUG_CPU
611 spin_lock(&memcg->pcp_counter_lock);
612 val += memcg->nocpu_base.events[idx];
613 spin_unlock(&memcg->pcp_counter_lock);
618 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
619 bool file, int nr_pages)
624 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
627 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
630 /* pagein of a big page is an event. So, ignore page size */
632 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
634 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
635 nr_pages = -nr_pages; /* for event */
638 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
644 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
645 unsigned int lru_mask)
647 struct mem_cgroup_per_zone *mz;
649 unsigned long ret = 0;
651 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
654 if (BIT(l) & lru_mask)
655 ret += MEM_CGROUP_ZSTAT(mz, l);
661 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
662 int nid, unsigned int lru_mask)
667 for (zid = 0; zid < MAX_NR_ZONES; zid++)
668 total += mem_cgroup_zone_nr_lru_pages(memcg,
674 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
675 unsigned int lru_mask)
680 for_each_node_state(nid, N_HIGH_MEMORY)
681 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
685 static bool __memcg_event_check(struct mem_cgroup *memcg, int target)
687 unsigned long val, next;
689 val = this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
690 next = this_cpu_read(memcg->stat->targets[target]);
691 /* from time_after() in jiffies.h */
692 return ((long)next - (long)val < 0);
695 static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
697 unsigned long val, next;
699 val = this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
702 case MEM_CGROUP_TARGET_THRESH:
703 next = val + THRESHOLDS_EVENTS_TARGET;
705 case MEM_CGROUP_TARGET_SOFTLIMIT:
706 next = val + SOFTLIMIT_EVENTS_TARGET;
708 case MEM_CGROUP_TARGET_NUMAINFO:
709 next = val + NUMAINFO_EVENTS_TARGET;
715 this_cpu_write(memcg->stat->targets[target], next);
719 * Check events in order.
722 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
724 /* threshold event is triggered in finer grain than soft limit */
725 if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
726 mem_cgroup_threshold(memcg);
727 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
728 if (unlikely(__memcg_event_check(memcg,
729 MEM_CGROUP_TARGET_SOFTLIMIT))) {
730 mem_cgroup_update_tree(memcg, page);
731 __mem_cgroup_target_update(memcg,
732 MEM_CGROUP_TARGET_SOFTLIMIT);
735 if (unlikely(__memcg_event_check(memcg,
736 MEM_CGROUP_TARGET_NUMAINFO))) {
737 atomic_inc(&memcg->numainfo_events);
738 __mem_cgroup_target_update(memcg,
739 MEM_CGROUP_TARGET_NUMAINFO);
745 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
747 return container_of(cgroup_subsys_state(cont,
748 mem_cgroup_subsys_id), struct mem_cgroup,
752 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
755 * mm_update_next_owner() may clear mm->owner to NULL
756 * if it races with swapoff, page migration, etc.
757 * So this can be called with p == NULL.
762 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
763 struct mem_cgroup, css);
766 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
768 struct mem_cgroup *memcg = NULL;
773 * Because we have no locks, mm->owner's may be being moved to other
774 * cgroup. We use css_tryget() here even if this looks
775 * pessimistic (rather than adding locks here).
779 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
780 if (unlikely(!memcg))
782 } while (!css_tryget(&memcg->css));
787 /* The caller has to guarantee "mem" exists before calling this */
788 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *memcg)
790 struct cgroup_subsys_state *css;
793 if (!memcg) /* ROOT cgroup has the smallest ID */
794 return root_mem_cgroup; /*css_put/get against root is ignored*/
795 if (!memcg->use_hierarchy) {
796 if (css_tryget(&memcg->css))
802 * searching a memory cgroup which has the smallest ID under given
803 * ROOT cgroup. (ID >= 1)
805 css = css_get_next(&mem_cgroup_subsys, 1, &memcg->css, &found);
806 if (css && css_tryget(css))
807 memcg = container_of(css, struct mem_cgroup, css);
814 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
815 struct mem_cgroup *root,
818 int nextid = css_id(&iter->css) + 1;
821 struct cgroup_subsys_state *css;
823 hierarchy_used = iter->use_hierarchy;
826 /* If no ROOT, walk all, ignore hierarchy */
827 if (!cond || (root && !hierarchy_used))
831 root = root_mem_cgroup;
837 css = css_get_next(&mem_cgroup_subsys, nextid,
839 if (css && css_tryget(css))
840 iter = container_of(css, struct mem_cgroup, css);
842 /* If css is NULL, no more cgroups will be found */
844 } while (css && !iter);
849 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
850 * be careful that "break" loop is not allowed. We have reference count.
851 * Instead of that modify "cond" to be false and "continue" to exit the loop.
853 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
854 for (iter = mem_cgroup_start_loop(root);\
856 iter = mem_cgroup_get_next(iter, root, cond))
858 #define for_each_mem_cgroup_tree(iter, root) \
859 for_each_mem_cgroup_tree_cond(iter, root, true)
861 #define for_each_mem_cgroup_all(iter) \
862 for_each_mem_cgroup_tree_cond(iter, NULL, true)
865 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
867 return (memcg == root_mem_cgroup);
870 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
872 struct mem_cgroup *memcg;
878 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
879 if (unlikely(!memcg))
884 mem_cgroup_pgmajfault(memcg, 1);
887 mem_cgroup_pgfault(memcg, 1);
895 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
898 * Following LRU functions are allowed to be used without PCG_LOCK.
899 * Operations are called by routine of global LRU independently from memcg.
900 * What we have to take care of here is validness of pc->mem_cgroup.
902 * Changes to pc->mem_cgroup happens when
905 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
906 * It is added to LRU before charge.
907 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
908 * When moving account, the page is not on LRU. It's isolated.
911 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
913 struct page_cgroup *pc;
914 struct mem_cgroup_per_zone *mz;
916 if (mem_cgroup_disabled())
918 pc = lookup_page_cgroup(page);
919 /* can happen while we handle swapcache. */
920 if (!TestClearPageCgroupAcctLRU(pc))
922 VM_BUG_ON(!pc->mem_cgroup);
924 * We don't check PCG_USED bit. It's cleared when the "page" is finally
925 * removed from global LRU.
927 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
928 /* huge page split is done under lru_lock. so, we have no races. */
929 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
930 if (mem_cgroup_is_root(pc->mem_cgroup))
932 VM_BUG_ON(list_empty(&pc->lru));
933 list_del_init(&pc->lru);
936 void mem_cgroup_del_lru(struct page *page)
938 mem_cgroup_del_lru_list(page, page_lru(page));
942 * Writeback is about to end against a page which has been marked for immediate
943 * reclaim. If it still appears to be reclaimable, move it to the tail of the
946 void mem_cgroup_rotate_reclaimable_page(struct page *page)
948 struct mem_cgroup_per_zone *mz;
949 struct page_cgroup *pc;
950 enum lru_list lru = page_lru(page);
952 if (mem_cgroup_disabled())
955 pc = lookup_page_cgroup(page);
956 /* unused or root page is not rotated. */
957 if (!PageCgroupUsed(pc))
959 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
961 if (mem_cgroup_is_root(pc->mem_cgroup))
963 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
964 list_move_tail(&pc->lru, &mz->lists[lru]);
967 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
969 struct mem_cgroup_per_zone *mz;
970 struct page_cgroup *pc;
972 if (mem_cgroup_disabled())
975 pc = lookup_page_cgroup(page);
976 /* unused or root page is not rotated. */
977 if (!PageCgroupUsed(pc))
979 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
981 if (mem_cgroup_is_root(pc->mem_cgroup))
983 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
984 list_move(&pc->lru, &mz->lists[lru]);
987 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
989 struct page_cgroup *pc;
990 struct mem_cgroup_per_zone *mz;
992 if (mem_cgroup_disabled())
994 pc = lookup_page_cgroup(page);
995 VM_BUG_ON(PageCgroupAcctLRU(pc));
996 if (!PageCgroupUsed(pc))
998 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1000 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1001 /* huge page split is done under lru_lock. so, we have no races. */
1002 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1003 SetPageCgroupAcctLRU(pc);
1004 if (mem_cgroup_is_root(pc->mem_cgroup))
1006 list_add(&pc->lru, &mz->lists[lru]);
1010 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1011 * while it's linked to lru because the page may be reused after it's fully
1012 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1013 * It's done under lock_page and expected that zone->lru_lock isnever held.
1015 static void mem_cgroup_lru_del_before_commit(struct page *page)
1017 unsigned long flags;
1018 struct zone *zone = page_zone(page);
1019 struct page_cgroup *pc = lookup_page_cgroup(page);
1022 * Doing this check without taking ->lru_lock seems wrong but this
1023 * is safe. Because if page_cgroup's USED bit is unset, the page
1024 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1025 * set, the commit after this will fail, anyway.
1026 * This all charge/uncharge is done under some mutual execustion.
1027 * So, we don't need to taking care of changes in USED bit.
1029 if (likely(!PageLRU(page)))
1032 spin_lock_irqsave(&zone->lru_lock, flags);
1034 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1035 * is guarded by lock_page() because the page is SwapCache.
1037 if (!PageCgroupUsed(pc))
1038 mem_cgroup_del_lru_list(page, page_lru(page));
1039 spin_unlock_irqrestore(&zone->lru_lock, flags);
1042 static void mem_cgroup_lru_add_after_commit(struct page *page)
1044 unsigned long flags;
1045 struct zone *zone = page_zone(page);
1046 struct page_cgroup *pc = lookup_page_cgroup(page);
1048 /* taking care of that the page is added to LRU while we commit it */
1049 if (likely(!PageLRU(page)))
1051 spin_lock_irqsave(&zone->lru_lock, flags);
1052 /* link when the page is linked to LRU but page_cgroup isn't */
1053 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1054 mem_cgroup_add_lru_list(page, page_lru(page));
1055 spin_unlock_irqrestore(&zone->lru_lock, flags);
1059 void mem_cgroup_move_lists(struct page *page,
1060 enum lru_list from, enum lru_list to)
1062 if (mem_cgroup_disabled())
1064 mem_cgroup_del_lru_list(page, from);
1065 mem_cgroup_add_lru_list(page, to);
1069 * Checks whether given mem is same or in the root_mem_cgroup's
1072 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1073 struct mem_cgroup *memcg)
1075 if (root_memcg != memcg) {
1076 return (root_memcg->use_hierarchy &&
1077 css_is_ancestor(&memcg->css, &root_memcg->css));
1083 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1086 struct mem_cgroup *curr = NULL;
1087 struct task_struct *p;
1089 p = find_lock_task_mm(task);
1092 curr = try_get_mem_cgroup_from_mm(p->mm);
1097 * We should check use_hierarchy of "memcg" not "curr". Because checking
1098 * use_hierarchy of "curr" here make this function true if hierarchy is
1099 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1100 * hierarchy(even if use_hierarchy is disabled in "memcg").
1102 ret = mem_cgroup_same_or_subtree(memcg, curr);
1103 css_put(&curr->css);
1107 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1109 unsigned long inactive_ratio;
1110 int nid = zone_to_nid(zone);
1111 int zid = zone_idx(zone);
1112 unsigned long inactive;
1113 unsigned long active;
1116 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1117 BIT(LRU_INACTIVE_ANON));
1118 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1119 BIT(LRU_ACTIVE_ANON));
1121 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1123 inactive_ratio = int_sqrt(10 * gb);
1127 return inactive * inactive_ratio < active;
1130 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1132 unsigned long active;
1133 unsigned long inactive;
1134 int zid = zone_idx(zone);
1135 int nid = zone_to_nid(zone);
1137 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1138 BIT(LRU_INACTIVE_FILE));
1139 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1140 BIT(LRU_ACTIVE_FILE));
1142 return (active > inactive);
1145 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1148 int nid = zone_to_nid(zone);
1149 int zid = zone_idx(zone);
1150 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1152 return &mz->reclaim_stat;
1155 struct zone_reclaim_stat *
1156 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1158 struct page_cgroup *pc;
1159 struct mem_cgroup_per_zone *mz;
1161 if (mem_cgroup_disabled())
1164 pc = lookup_page_cgroup(page);
1165 if (!PageCgroupUsed(pc))
1167 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1169 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1170 return &mz->reclaim_stat;
1173 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1174 struct list_head *dst,
1175 unsigned long *scanned, int order,
1176 isolate_mode_t mode,
1178 struct mem_cgroup *mem_cont,
1179 int active, int file)
1181 unsigned long nr_taken = 0;
1185 struct list_head *src;
1186 struct page_cgroup *pc, *tmp;
1187 int nid = zone_to_nid(z);
1188 int zid = zone_idx(z);
1189 struct mem_cgroup_per_zone *mz;
1190 int lru = LRU_FILE * file + active;
1194 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1195 src = &mz->lists[lru];
1198 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1199 if (scan >= nr_to_scan)
1202 if (unlikely(!PageCgroupUsed(pc)))
1205 page = lookup_cgroup_page(pc);
1207 if (unlikely(!PageLRU(page)))
1211 ret = __isolate_lru_page(page, mode, file);
1214 list_move(&page->lru, dst);
1215 mem_cgroup_del_lru(page);
1216 nr_taken += hpage_nr_pages(page);
1219 /* we don't affect global LRU but rotate in our LRU */
1220 mem_cgroup_rotate_lru_list(page, page_lru(page));
1229 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1235 #define mem_cgroup_from_res_counter(counter, member) \
1236 container_of(counter, struct mem_cgroup, member)
1239 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1240 * @mem: the memory cgroup
1242 * Returns the maximum amount of memory @mem can be charged with, in
1245 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1247 unsigned long long margin;
1249 margin = res_counter_margin(&memcg->res);
1250 if (do_swap_account)
1251 margin = min(margin, res_counter_margin(&memcg->memsw));
1252 return margin >> PAGE_SHIFT;
1255 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1257 struct cgroup *cgrp = memcg->css.cgroup;
1260 if (cgrp->parent == NULL)
1261 return vm_swappiness;
1263 return memcg->swappiness;
1266 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1271 spin_lock(&memcg->pcp_counter_lock);
1272 for_each_online_cpu(cpu)
1273 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1274 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1275 spin_unlock(&memcg->pcp_counter_lock);
1281 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1288 spin_lock(&memcg->pcp_counter_lock);
1289 for_each_online_cpu(cpu)
1290 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1291 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1292 spin_unlock(&memcg->pcp_counter_lock);
1296 * 2 routines for checking "mem" is under move_account() or not.
1298 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1299 * for avoiding race in accounting. If true,
1300 * pc->mem_cgroup may be overwritten.
1302 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1303 * under hierarchy of moving cgroups. This is for
1304 * waiting at hith-memory prressure caused by "move".
1307 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1309 VM_BUG_ON(!rcu_read_lock_held());
1310 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1313 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1315 struct mem_cgroup *from;
1316 struct mem_cgroup *to;
1319 * Unlike task_move routines, we access mc.to, mc.from not under
1320 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1322 spin_lock(&mc.lock);
1328 ret = mem_cgroup_same_or_subtree(memcg, from)
1329 || mem_cgroup_same_or_subtree(memcg, to);
1331 spin_unlock(&mc.lock);
1335 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1337 if (mc.moving_task && current != mc.moving_task) {
1338 if (mem_cgroup_under_move(memcg)) {
1340 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1341 /* moving charge context might have finished. */
1344 finish_wait(&mc.waitq, &wait);
1352 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1353 * @memcg: The memory cgroup that went over limit
1354 * @p: Task that is going to be killed
1356 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1359 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1361 struct cgroup *task_cgrp;
1362 struct cgroup *mem_cgrp;
1364 * Need a buffer in BSS, can't rely on allocations. The code relies
1365 * on the assumption that OOM is serialized for memory controller.
1366 * If this assumption is broken, revisit this code.
1368 static char memcg_name[PATH_MAX];
1377 mem_cgrp = memcg->css.cgroup;
1378 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1380 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1383 * Unfortunately, we are unable to convert to a useful name
1384 * But we'll still print out the usage information
1391 printk(KERN_INFO "Task in %s killed", memcg_name);
1394 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1402 * Continues from above, so we don't need an KERN_ level
1404 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1407 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1408 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1409 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1410 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1411 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1413 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1414 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1415 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1419 * This function returns the number of memcg under hierarchy tree. Returns
1420 * 1(self count) if no children.
1422 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1425 struct mem_cgroup *iter;
1427 for_each_mem_cgroup_tree(iter, memcg)
1433 * Return the memory (and swap, if configured) limit for a memcg.
1435 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1440 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1441 limit += total_swap_pages << PAGE_SHIFT;
1443 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1445 * If memsw is finite and limits the amount of swap space available
1446 * to this memcg, return that limit.
1448 return min(limit, memsw);
1452 * Visit the first child (need not be the first child as per the ordering
1453 * of the cgroup list, since we track last_scanned_child) of @mem and use
1454 * that to reclaim free pages from.
1456 static struct mem_cgroup *
1457 mem_cgroup_select_victim(struct mem_cgroup *root_memcg)
1459 struct mem_cgroup *ret = NULL;
1460 struct cgroup_subsys_state *css;
1463 if (!root_memcg->use_hierarchy) {
1464 css_get(&root_memcg->css);
1470 nextid = root_memcg->last_scanned_child + 1;
1471 css = css_get_next(&mem_cgroup_subsys, nextid, &root_memcg->css,
1473 if (css && css_tryget(css))
1474 ret = container_of(css, struct mem_cgroup, css);
1477 /* Updates scanning parameter */
1479 /* this means start scan from ID:1 */
1480 root_memcg->last_scanned_child = 0;
1482 root_memcg->last_scanned_child = found;
1489 * test_mem_cgroup_node_reclaimable
1490 * @mem: the target memcg
1491 * @nid: the node ID to be checked.
1492 * @noswap : specify true here if the user wants flle only information.
1494 * This function returns whether the specified memcg contains any
1495 * reclaimable pages on a node. Returns true if there are any reclaimable
1496 * pages in the node.
1498 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1499 int nid, bool noswap)
1501 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1503 if (noswap || !total_swap_pages)
1505 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1510 #if MAX_NUMNODES > 1
1513 * Always updating the nodemask is not very good - even if we have an empty
1514 * list or the wrong list here, we can start from some node and traverse all
1515 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1518 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1522 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1523 * pagein/pageout changes since the last update.
1525 if (!atomic_read(&memcg->numainfo_events))
1527 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1530 /* make a nodemask where this memcg uses memory from */
1531 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1533 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1535 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1536 node_clear(nid, memcg->scan_nodes);
1539 atomic_set(&memcg->numainfo_events, 0);
1540 atomic_set(&memcg->numainfo_updating, 0);
1544 * Selecting a node where we start reclaim from. Because what we need is just
1545 * reducing usage counter, start from anywhere is O,K. Considering
1546 * memory reclaim from current node, there are pros. and cons.
1548 * Freeing memory from current node means freeing memory from a node which
1549 * we'll use or we've used. So, it may make LRU bad. And if several threads
1550 * hit limits, it will see a contention on a node. But freeing from remote
1551 * node means more costs for memory reclaim because of memory latency.
1553 * Now, we use round-robin. Better algorithm is welcomed.
1555 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1559 mem_cgroup_may_update_nodemask(memcg);
1560 node = memcg->last_scanned_node;
1562 node = next_node(node, memcg->scan_nodes);
1563 if (node == MAX_NUMNODES)
1564 node = first_node(memcg->scan_nodes);
1566 * We call this when we hit limit, not when pages are added to LRU.
1567 * No LRU may hold pages because all pages are UNEVICTABLE or
1568 * memcg is too small and all pages are not on LRU. In that case,
1569 * we use curret node.
1571 if (unlikely(node == MAX_NUMNODES))
1572 node = numa_node_id();
1574 memcg->last_scanned_node = node;
1579 * Check all nodes whether it contains reclaimable pages or not.
1580 * For quick scan, we make use of scan_nodes. This will allow us to skip
1581 * unused nodes. But scan_nodes is lazily updated and may not cotain
1582 * enough new information. We need to do double check.
1584 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1589 * quick check...making use of scan_node.
1590 * We can skip unused nodes.
1592 if (!nodes_empty(memcg->scan_nodes)) {
1593 for (nid = first_node(memcg->scan_nodes);
1595 nid = next_node(nid, memcg->scan_nodes)) {
1597 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1602 * Check rest of nodes.
1604 for_each_node_state(nid, N_HIGH_MEMORY) {
1605 if (node_isset(nid, memcg->scan_nodes))
1607 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1614 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1619 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1621 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1626 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1627 * we reclaimed from, so that we don't end up penalizing one child extensively
1628 * based on its position in the children list.
1630 * root_memcg is the original ancestor that we've been reclaim from.
1632 * We give up and return to the caller when we visit root_memcg twice.
1633 * (other groups can be removed while we're walking....)
1635 * If shrink==true, for avoiding to free too much, this returns immedieately.
1637 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_memcg,
1640 unsigned long reclaim_options,
1641 unsigned long *total_scanned)
1643 struct mem_cgroup *victim;
1646 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1647 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1648 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1649 unsigned long excess;
1650 unsigned long nr_scanned;
1652 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1654 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1655 if (!check_soft && !shrink && root_memcg->memsw_is_minimum)
1659 victim = mem_cgroup_select_victim(root_memcg);
1660 if (victim == root_memcg) {
1663 * We are not draining per cpu cached charges during
1664 * soft limit reclaim because global reclaim doesn't
1665 * care about charges. It tries to free some memory and
1666 * charges will not give any.
1668 if (!check_soft && loop >= 1)
1669 drain_all_stock_async(root_memcg);
1672 * If we have not been able to reclaim
1673 * anything, it might because there are
1674 * no reclaimable pages under this hierarchy
1676 if (!check_soft || !total) {
1677 css_put(&victim->css);
1681 * We want to do more targeted reclaim.
1682 * excess >> 2 is not to excessive so as to
1683 * reclaim too much, nor too less that we keep
1684 * coming back to reclaim from this cgroup
1686 if (total >= (excess >> 2) ||
1687 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1688 css_put(&victim->css);
1693 if (!mem_cgroup_reclaimable(victim, noswap)) {
1694 /* this cgroup's local usage == 0 */
1695 css_put(&victim->css);
1698 /* we use swappiness of local cgroup */
1700 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1701 noswap, zone, &nr_scanned);
1702 *total_scanned += nr_scanned;
1704 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1706 css_put(&victim->css);
1708 * At shrinking usage, we can't check we should stop here or
1709 * reclaim more. It's depends on callers. last_scanned_child
1710 * will work enough for keeping fairness under tree.
1716 if (!res_counter_soft_limit_excess(&root_memcg->res))
1718 } else if (mem_cgroup_margin(root_memcg))
1725 * Check OOM-Killer is already running under our hierarchy.
1726 * If someone is running, return false.
1727 * Has to be called with memcg_oom_lock
1729 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1731 struct mem_cgroup *iter, *failed = NULL;
1734 for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1735 if (iter->oom_lock) {
1737 * this subtree of our hierarchy is already locked
1738 * so we cannot give a lock.
1743 iter->oom_lock = true;
1750 * OK, we failed to lock the whole subtree so we have to clean up
1751 * what we set up to the failing subtree
1754 for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1755 if (iter == failed) {
1759 iter->oom_lock = false;
1765 * Has to be called with memcg_oom_lock
1767 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1769 struct mem_cgroup *iter;
1771 for_each_mem_cgroup_tree(iter, memcg)
1772 iter->oom_lock = false;
1776 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1778 struct mem_cgroup *iter;
1780 for_each_mem_cgroup_tree(iter, memcg)
1781 atomic_inc(&iter->under_oom);
1784 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1786 struct mem_cgroup *iter;
1789 * When a new child is created while the hierarchy is under oom,
1790 * mem_cgroup_oom_lock() may not be called. We have to use
1791 * atomic_add_unless() here.
1793 for_each_mem_cgroup_tree(iter, memcg)
1794 atomic_add_unless(&iter->under_oom, -1, 0);
1797 static DEFINE_SPINLOCK(memcg_oom_lock);
1798 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1800 struct oom_wait_info {
1801 struct mem_cgroup *mem;
1805 static int memcg_oom_wake_function(wait_queue_t *wait,
1806 unsigned mode, int sync, void *arg)
1808 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1810 struct oom_wait_info *oom_wait_info;
1812 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1813 oom_wait_memcg = oom_wait_info->mem;
1816 * Both of oom_wait_info->mem and wake_mem are stable under us.
1817 * Then we can use css_is_ancestor without taking care of RCU.
1819 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1820 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1822 return autoremove_wake_function(wait, mode, sync, arg);
1825 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1827 /* for filtering, pass "memcg" as argument. */
1828 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1831 static void memcg_oom_recover(struct mem_cgroup *memcg)
1833 if (memcg && atomic_read(&memcg->under_oom))
1834 memcg_wakeup_oom(memcg);
1838 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1840 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1842 struct oom_wait_info owait;
1843 bool locked, need_to_kill;
1846 owait.wait.flags = 0;
1847 owait.wait.func = memcg_oom_wake_function;
1848 owait.wait.private = current;
1849 INIT_LIST_HEAD(&owait.wait.task_list);
1850 need_to_kill = true;
1851 mem_cgroup_mark_under_oom(memcg);
1853 /* At first, try to OOM lock hierarchy under memcg.*/
1854 spin_lock(&memcg_oom_lock);
1855 locked = mem_cgroup_oom_lock(memcg);
1857 * Even if signal_pending(), we can't quit charge() loop without
1858 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1859 * under OOM is always welcomed, use TASK_KILLABLE here.
1861 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1862 if (!locked || memcg->oom_kill_disable)
1863 need_to_kill = false;
1865 mem_cgroup_oom_notify(memcg);
1866 spin_unlock(&memcg_oom_lock);
1869 finish_wait(&memcg_oom_waitq, &owait.wait);
1870 mem_cgroup_out_of_memory(memcg, mask);
1873 finish_wait(&memcg_oom_waitq, &owait.wait);
1875 spin_lock(&memcg_oom_lock);
1877 mem_cgroup_oom_unlock(memcg);
1878 memcg_wakeup_oom(memcg);
1879 spin_unlock(&memcg_oom_lock);
1881 mem_cgroup_unmark_under_oom(memcg);
1883 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1885 /* Give chance to dying process */
1886 schedule_timeout_uninterruptible(1);
1891 * Currently used to update mapped file statistics, but the routine can be
1892 * generalized to update other statistics as well.
1894 * Notes: Race condition
1896 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1897 * it tends to be costly. But considering some conditions, we doesn't need
1898 * to do so _always_.
1900 * Considering "charge", lock_page_cgroup() is not required because all
1901 * file-stat operations happen after a page is attached to radix-tree. There
1902 * are no race with "charge".
1904 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1905 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1906 * if there are race with "uncharge". Statistics itself is properly handled
1909 * Considering "move", this is an only case we see a race. To make the race
1910 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1911 * possibility of race condition. If there is, we take a lock.
1914 void mem_cgroup_update_page_stat(struct page *page,
1915 enum mem_cgroup_page_stat_item idx, int val)
1917 struct mem_cgroup *memcg;
1918 struct page_cgroup *pc = lookup_page_cgroup(page);
1919 bool need_unlock = false;
1920 unsigned long uninitialized_var(flags);
1926 memcg = pc->mem_cgroup;
1927 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1929 /* pc->mem_cgroup is unstable ? */
1930 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
1931 /* take a lock against to access pc->mem_cgroup */
1932 move_lock_page_cgroup(pc, &flags);
1934 memcg = pc->mem_cgroup;
1935 if (!memcg || !PageCgroupUsed(pc))
1940 case MEMCG_NR_FILE_MAPPED:
1942 SetPageCgroupFileMapped(pc);
1943 else if (!page_mapped(page))
1944 ClearPageCgroupFileMapped(pc);
1945 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1951 this_cpu_add(memcg->stat->count[idx], val);
1954 if (unlikely(need_unlock))
1955 move_unlock_page_cgroup(pc, &flags);
1959 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1962 * size of first charge trial. "32" comes from vmscan.c's magic value.
1963 * TODO: maybe necessary to use big numbers in big irons.
1965 #define CHARGE_BATCH 32U
1966 struct memcg_stock_pcp {
1967 struct mem_cgroup *cached; /* this never be root cgroup */
1968 unsigned int nr_pages;
1969 struct work_struct work;
1970 unsigned long flags;
1971 #define FLUSHING_CACHED_CHARGE (0)
1973 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1974 static DEFINE_MUTEX(percpu_charge_mutex);
1977 * Try to consume stocked charge on this cpu. If success, one page is consumed
1978 * from local stock and true is returned. If the stock is 0 or charges from a
1979 * cgroup which is not current target, returns false. This stock will be
1982 static bool consume_stock(struct mem_cgroup *memcg)
1984 struct memcg_stock_pcp *stock;
1987 stock = &get_cpu_var(memcg_stock);
1988 if (memcg == stock->cached && stock->nr_pages)
1990 else /* need to call res_counter_charge */
1992 put_cpu_var(memcg_stock);
1997 * Returns stocks cached in percpu to res_counter and reset cached information.
1999 static void drain_stock(struct memcg_stock_pcp *stock)
2001 struct mem_cgroup *old = stock->cached;
2003 if (stock->nr_pages) {
2004 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2006 res_counter_uncharge(&old->res, bytes);
2007 if (do_swap_account)
2008 res_counter_uncharge(&old->memsw, bytes);
2009 stock->nr_pages = 0;
2011 stock->cached = NULL;
2015 * This must be called under preempt disabled or must be called by
2016 * a thread which is pinned to local cpu.
2018 static void drain_local_stock(struct work_struct *dummy)
2020 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2022 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2026 * Cache charges(val) which is from res_counter, to local per_cpu area.
2027 * This will be consumed by consume_stock() function, later.
2029 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2031 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2033 if (stock->cached != memcg) { /* reset if necessary */
2035 stock->cached = memcg;
2037 stock->nr_pages += nr_pages;
2038 put_cpu_var(memcg_stock);
2042 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2043 * of the hierarchy under it. sync flag says whether we should block
2044 * until the work is done.
2046 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2050 /* Notify other cpus that system-wide "drain" is running */
2053 for_each_online_cpu(cpu) {
2054 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2055 struct mem_cgroup *memcg;
2057 memcg = stock->cached;
2058 if (!memcg || !stock->nr_pages)
2060 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2062 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2064 drain_local_stock(&stock->work);
2066 schedule_work_on(cpu, &stock->work);
2074 for_each_online_cpu(cpu) {
2075 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2076 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2077 flush_work(&stock->work);
2084 * Tries to drain stocked charges in other cpus. This function is asynchronous
2085 * and just put a work per cpu for draining localy on each cpu. Caller can
2086 * expects some charges will be back to res_counter later but cannot wait for
2089 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2092 * If someone calls draining, avoid adding more kworker runs.
2094 if (!mutex_trylock(&percpu_charge_mutex))
2096 drain_all_stock(root_memcg, false);
2097 mutex_unlock(&percpu_charge_mutex);
2100 /* This is a synchronous drain interface. */
2101 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2103 /* called when force_empty is called */
2104 mutex_lock(&percpu_charge_mutex);
2105 drain_all_stock(root_memcg, true);
2106 mutex_unlock(&percpu_charge_mutex);
2110 * This function drains percpu counter value from DEAD cpu and
2111 * move it to local cpu. Note that this function can be preempted.
2113 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2117 spin_lock(&memcg->pcp_counter_lock);
2118 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2119 long x = per_cpu(memcg->stat->count[i], cpu);
2121 per_cpu(memcg->stat->count[i], cpu) = 0;
2122 memcg->nocpu_base.count[i] += x;
2124 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2125 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2127 per_cpu(memcg->stat->events[i], cpu) = 0;
2128 memcg->nocpu_base.events[i] += x;
2130 /* need to clear ON_MOVE value, works as a kind of lock. */
2131 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2132 spin_unlock(&memcg->pcp_counter_lock);
2135 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2137 int idx = MEM_CGROUP_ON_MOVE;
2139 spin_lock(&memcg->pcp_counter_lock);
2140 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2141 spin_unlock(&memcg->pcp_counter_lock);
2144 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2145 unsigned long action,
2148 int cpu = (unsigned long)hcpu;
2149 struct memcg_stock_pcp *stock;
2150 struct mem_cgroup *iter;
2152 if ((action == CPU_ONLINE)) {
2153 for_each_mem_cgroup_all(iter)
2154 synchronize_mem_cgroup_on_move(iter, cpu);
2158 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2161 for_each_mem_cgroup_all(iter)
2162 mem_cgroup_drain_pcp_counter(iter, cpu);
2164 stock = &per_cpu(memcg_stock, cpu);
2170 /* See __mem_cgroup_try_charge() for details */
2172 CHARGE_OK, /* success */
2173 CHARGE_RETRY, /* need to retry but retry is not bad */
2174 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2175 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2176 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2179 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2180 unsigned int nr_pages, bool oom_check)
2182 unsigned long csize = nr_pages * PAGE_SIZE;
2183 struct mem_cgroup *mem_over_limit;
2184 struct res_counter *fail_res;
2185 unsigned long flags = 0;
2188 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2191 if (!do_swap_account)
2193 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2197 res_counter_uncharge(&memcg->res, csize);
2198 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2199 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2201 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2203 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2204 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2206 * Never reclaim on behalf of optional batching, retry with a
2207 * single page instead.
2209 if (nr_pages == CHARGE_BATCH)
2210 return CHARGE_RETRY;
2212 if (!(gfp_mask & __GFP_WAIT))
2213 return CHARGE_WOULDBLOCK;
2215 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2216 gfp_mask, flags, NULL);
2217 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2218 return CHARGE_RETRY;
2220 * Even though the limit is exceeded at this point, reclaim
2221 * may have been able to free some pages. Retry the charge
2222 * before killing the task.
2224 * Only for regular pages, though: huge pages are rather
2225 * unlikely to succeed so close to the limit, and we fall back
2226 * to regular pages anyway in case of failure.
2228 if (nr_pages == 1 && ret)
2229 return CHARGE_RETRY;
2232 * At task move, charge accounts can be doubly counted. So, it's
2233 * better to wait until the end of task_move if something is going on.
2235 if (mem_cgroup_wait_acct_move(mem_over_limit))
2236 return CHARGE_RETRY;
2238 /* If we don't need to call oom-killer at el, return immediately */
2240 return CHARGE_NOMEM;
2242 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2243 return CHARGE_OOM_DIE;
2245 return CHARGE_RETRY;
2249 * Unlike exported interface, "oom" parameter is added. if oom==true,
2250 * oom-killer can be invoked.
2252 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2254 unsigned int nr_pages,
2255 struct mem_cgroup **ptr,
2258 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2259 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2260 struct mem_cgroup *memcg = NULL;
2264 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2265 * in system level. So, allow to go ahead dying process in addition to
2268 if (unlikely(test_thread_flag(TIF_MEMDIE)
2269 || fatal_signal_pending(current)))
2273 * We always charge the cgroup the mm_struct belongs to.
2274 * The mm_struct's mem_cgroup changes on task migration if the
2275 * thread group leader migrates. It's possible that mm is not
2276 * set, if so charge the init_mm (happens for pagecache usage).
2281 if (*ptr) { /* css should be a valid one */
2283 VM_BUG_ON(css_is_removed(&memcg->css));
2284 if (mem_cgroup_is_root(memcg))
2286 if (nr_pages == 1 && consume_stock(memcg))
2288 css_get(&memcg->css);
2290 struct task_struct *p;
2293 p = rcu_dereference(mm->owner);
2295 * Because we don't have task_lock(), "p" can exit.
2296 * In that case, "memcg" can point to root or p can be NULL with
2297 * race with swapoff. Then, we have small risk of mis-accouning.
2298 * But such kind of mis-account by race always happens because
2299 * we don't have cgroup_mutex(). It's overkill and we allo that
2301 * (*) swapoff at el will charge against mm-struct not against
2302 * task-struct. So, mm->owner can be NULL.
2304 memcg = mem_cgroup_from_task(p);
2305 if (!memcg || mem_cgroup_is_root(memcg)) {
2309 if (nr_pages == 1 && consume_stock(memcg)) {
2311 * It seems dagerous to access memcg without css_get().
2312 * But considering how consume_stok works, it's not
2313 * necessary. If consume_stock success, some charges
2314 * from this memcg are cached on this cpu. So, we
2315 * don't need to call css_get()/css_tryget() before
2316 * calling consume_stock().
2321 /* after here, we may be blocked. we need to get refcnt */
2322 if (!css_tryget(&memcg->css)) {
2332 /* If killed, bypass charge */
2333 if (fatal_signal_pending(current)) {
2334 css_put(&memcg->css);
2339 if (oom && !nr_oom_retries) {
2341 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2344 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2348 case CHARGE_RETRY: /* not in OOM situation but retry */
2350 css_put(&memcg->css);
2353 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2354 css_put(&memcg->css);
2356 case CHARGE_NOMEM: /* OOM routine works */
2358 css_put(&memcg->css);
2361 /* If oom, we never return -ENOMEM */
2364 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2365 css_put(&memcg->css);
2368 } while (ret != CHARGE_OK);
2370 if (batch > nr_pages)
2371 refill_stock(memcg, batch - nr_pages);
2372 css_put(&memcg->css);
2385 * Somemtimes we have to undo a charge we got by try_charge().
2386 * This function is for that and do uncharge, put css's refcnt.
2387 * gotten by try_charge().
2389 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2390 unsigned int nr_pages)
2392 if (!mem_cgroup_is_root(memcg)) {
2393 unsigned long bytes = nr_pages * PAGE_SIZE;
2395 res_counter_uncharge(&memcg->res, bytes);
2396 if (do_swap_account)
2397 res_counter_uncharge(&memcg->memsw, bytes);
2402 * A helper function to get mem_cgroup from ID. must be called under
2403 * rcu_read_lock(). The caller must check css_is_removed() or some if
2404 * it's concern. (dropping refcnt from swap can be called against removed
2407 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2409 struct cgroup_subsys_state *css;
2411 /* ID 0 is unused ID */
2414 css = css_lookup(&mem_cgroup_subsys, id);
2417 return container_of(css, struct mem_cgroup, css);
2420 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2422 struct mem_cgroup *memcg = NULL;
2423 struct page_cgroup *pc;
2427 VM_BUG_ON(!PageLocked(page));
2429 pc = lookup_page_cgroup(page);
2430 lock_page_cgroup(pc);
2431 if (PageCgroupUsed(pc)) {
2432 memcg = pc->mem_cgroup;
2433 if (memcg && !css_tryget(&memcg->css))
2435 } else if (PageSwapCache(page)) {
2436 ent.val = page_private(page);
2437 id = lookup_swap_cgroup(ent);
2439 memcg = mem_cgroup_lookup(id);
2440 if (memcg && !css_tryget(&memcg->css))
2444 unlock_page_cgroup(pc);
2448 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2450 unsigned int nr_pages,
2451 struct page_cgroup *pc,
2452 enum charge_type ctype)
2454 lock_page_cgroup(pc);
2455 if (unlikely(PageCgroupUsed(pc))) {
2456 unlock_page_cgroup(pc);
2457 __mem_cgroup_cancel_charge(memcg, nr_pages);
2461 * we don't need page_cgroup_lock about tail pages, becase they are not
2462 * accessed by any other context at this point.
2464 pc->mem_cgroup = memcg;
2466 * We access a page_cgroup asynchronously without lock_page_cgroup().
2467 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2468 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2469 * before USED bit, we need memory barrier here.
2470 * See mem_cgroup_add_lru_list(), etc.
2474 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2475 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2476 SetPageCgroupCache(pc);
2477 SetPageCgroupUsed(pc);
2479 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2480 ClearPageCgroupCache(pc);
2481 SetPageCgroupUsed(pc);
2487 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2488 unlock_page_cgroup(pc);
2490 * "charge_statistics" updated event counter. Then, check it.
2491 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2492 * if they exceeds softlimit.
2494 memcg_check_events(memcg, page);
2497 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2499 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2500 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2502 * Because tail pages are not marked as "used", set it. We're under
2503 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2505 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2507 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2508 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2509 unsigned long flags;
2511 if (mem_cgroup_disabled())
2514 * We have no races with charge/uncharge but will have races with
2515 * page state accounting.
2517 move_lock_page_cgroup(head_pc, &flags);
2519 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2520 smp_wmb(); /* see __commit_charge() */
2521 if (PageCgroupAcctLRU(head_pc)) {
2523 struct mem_cgroup_per_zone *mz;
2526 * LRU flags cannot be copied because we need to add tail
2527 *.page to LRU by generic call and our hook will be called.
2528 * We hold lru_lock, then, reduce counter directly.
2530 lru = page_lru(head);
2531 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2532 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2534 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2535 move_unlock_page_cgroup(head_pc, &flags);
2540 * mem_cgroup_move_account - move account of the page
2542 * @nr_pages: number of regular pages (>1 for huge pages)
2543 * @pc: page_cgroup of the page.
2544 * @from: mem_cgroup which the page is moved from.
2545 * @to: mem_cgroup which the page is moved to. @from != @to.
2546 * @uncharge: whether we should call uncharge and css_put against @from.
2548 * The caller must confirm following.
2549 * - page is not on LRU (isolate_page() is useful.)
2550 * - compound_lock is held when nr_pages > 1
2552 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2553 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2554 * true, this function does "uncharge" from old cgroup, but it doesn't if
2555 * @uncharge is false, so a caller should do "uncharge".
2557 static int mem_cgroup_move_account(struct page *page,
2558 unsigned int nr_pages,
2559 struct page_cgroup *pc,
2560 struct mem_cgroup *from,
2561 struct mem_cgroup *to,
2564 unsigned long flags;
2567 VM_BUG_ON(from == to);
2568 VM_BUG_ON(PageLRU(page));
2570 * The page is isolated from LRU. So, collapse function
2571 * will not handle this page. But page splitting can happen.
2572 * Do this check under compound_page_lock(). The caller should
2576 if (nr_pages > 1 && !PageTransHuge(page))
2579 lock_page_cgroup(pc);
2582 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2585 move_lock_page_cgroup(pc, &flags);
2587 if (PageCgroupFileMapped(pc)) {
2588 /* Update mapped_file data for mem_cgroup */
2590 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2591 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2594 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2596 /* This is not "cancel", but cancel_charge does all we need. */
2597 __mem_cgroup_cancel_charge(from, nr_pages);
2599 /* caller should have done css_get */
2600 pc->mem_cgroup = to;
2601 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2603 * We charges against "to" which may not have any tasks. Then, "to"
2604 * can be under rmdir(). But in current implementation, caller of
2605 * this function is just force_empty() and move charge, so it's
2606 * guaranteed that "to" is never removed. So, we don't check rmdir
2609 move_unlock_page_cgroup(pc, &flags);
2612 unlock_page_cgroup(pc);
2616 memcg_check_events(to, page);
2617 memcg_check_events(from, page);
2623 * move charges to its parent.
2626 static int mem_cgroup_move_parent(struct page *page,
2627 struct page_cgroup *pc,
2628 struct mem_cgroup *child,
2631 struct cgroup *cg = child->css.cgroup;
2632 struct cgroup *pcg = cg->parent;
2633 struct mem_cgroup *parent;
2634 unsigned int nr_pages;
2635 unsigned long uninitialized_var(flags);
2643 if (!get_page_unless_zero(page))
2645 if (isolate_lru_page(page))
2648 nr_pages = hpage_nr_pages(page);
2650 parent = mem_cgroup_from_cont(pcg);
2651 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2656 flags = compound_lock_irqsave(page);
2658 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2660 __mem_cgroup_cancel_charge(parent, nr_pages);
2663 compound_unlock_irqrestore(page, flags);
2665 putback_lru_page(page);
2673 * Charge the memory controller for page usage.
2675 * 0 if the charge was successful
2676 * < 0 if the cgroup is over its limit
2678 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2679 gfp_t gfp_mask, enum charge_type ctype)
2681 struct mem_cgroup *memcg = NULL;
2682 unsigned int nr_pages = 1;
2683 struct page_cgroup *pc;
2687 if (PageTransHuge(page)) {
2688 nr_pages <<= compound_order(page);
2689 VM_BUG_ON(!PageTransHuge(page));
2691 * Never OOM-kill a process for a huge page. The
2692 * fault handler will fall back to regular pages.
2697 pc = lookup_page_cgroup(page);
2698 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2700 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2704 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2708 int mem_cgroup_newpage_charge(struct page *page,
2709 struct mm_struct *mm, gfp_t gfp_mask)
2711 if (mem_cgroup_disabled())
2714 * If already mapped, we don't have to account.
2715 * If page cache, page->mapping has address_space.
2716 * But page->mapping may have out-of-use anon_vma pointer,
2717 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2720 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2724 return mem_cgroup_charge_common(page, mm, gfp_mask,
2725 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2729 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2730 enum charge_type ctype);
2733 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2734 enum charge_type ctype)
2736 struct page_cgroup *pc = lookup_page_cgroup(page);
2738 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2739 * is already on LRU. It means the page may on some other page_cgroup's
2740 * LRU. Take care of it.
2742 mem_cgroup_lru_del_before_commit(page);
2743 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2744 mem_cgroup_lru_add_after_commit(page);
2748 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2751 struct mem_cgroup *memcg = NULL;
2754 if (mem_cgroup_disabled())
2756 if (PageCompound(page))
2762 if (page_is_file_cache(page)) {
2763 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2768 * FUSE reuses pages without going through the final
2769 * put that would remove them from the LRU list, make
2770 * sure that they get relinked properly.
2772 __mem_cgroup_commit_charge_lrucare(page, memcg,
2773 MEM_CGROUP_CHARGE_TYPE_CACHE);
2777 if (PageSwapCache(page)) {
2778 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2780 __mem_cgroup_commit_charge_swapin(page, memcg,
2781 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2783 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2784 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2790 * While swap-in, try_charge -> commit or cancel, the page is locked.
2791 * And when try_charge() successfully returns, one refcnt to memcg without
2792 * struct page_cgroup is acquired. This refcnt will be consumed by
2793 * "commit()" or removed by "cancel()"
2795 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2797 gfp_t mask, struct mem_cgroup **ptr)
2799 struct mem_cgroup *memcg;
2804 if (mem_cgroup_disabled())
2807 if (!do_swap_account)
2810 * A racing thread's fault, or swapoff, may have already updated
2811 * the pte, and even removed page from swap cache: in those cases
2812 * do_swap_page()'s pte_same() test will fail; but there's also a
2813 * KSM case which does need to charge the page.
2815 if (!PageSwapCache(page))
2817 memcg = try_get_mem_cgroup_from_page(page);
2821 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2822 css_put(&memcg->css);
2827 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2831 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2832 enum charge_type ctype)
2834 if (mem_cgroup_disabled())
2838 cgroup_exclude_rmdir(&ptr->css);
2840 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2842 * Now swap is on-memory. This means this page may be
2843 * counted both as mem and swap....double count.
2844 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2845 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2846 * may call delete_from_swap_cache() before reach here.
2848 if (do_swap_account && PageSwapCache(page)) {
2849 swp_entry_t ent = {.val = page_private(page)};
2851 struct mem_cgroup *memcg;
2853 id = swap_cgroup_record(ent, 0);
2855 memcg = mem_cgroup_lookup(id);
2858 * This recorded memcg can be obsolete one. So, avoid
2859 * calling css_tryget
2861 if (!mem_cgroup_is_root(memcg))
2862 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2863 mem_cgroup_swap_statistics(memcg, false);
2864 mem_cgroup_put(memcg);
2869 * At swapin, we may charge account against cgroup which has no tasks.
2870 * So, rmdir()->pre_destroy() can be called while we do this charge.
2871 * In that case, we need to call pre_destroy() again. check it here.
2873 cgroup_release_and_wakeup_rmdir(&ptr->css);
2876 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2878 __mem_cgroup_commit_charge_swapin(page, ptr,
2879 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2882 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2884 if (mem_cgroup_disabled())
2888 __mem_cgroup_cancel_charge(memcg, 1);
2891 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2892 unsigned int nr_pages,
2893 const enum charge_type ctype)
2895 struct memcg_batch_info *batch = NULL;
2896 bool uncharge_memsw = true;
2898 /* If swapout, usage of swap doesn't decrease */
2899 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2900 uncharge_memsw = false;
2902 batch = ¤t->memcg_batch;
2904 * In usual, we do css_get() when we remember memcg pointer.
2905 * But in this case, we keep res->usage until end of a series of
2906 * uncharges. Then, it's ok to ignore memcg's refcnt.
2909 batch->memcg = memcg;
2911 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2912 * In those cases, all pages freed continuously can be expected to be in
2913 * the same cgroup and we have chance to coalesce uncharges.
2914 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2915 * because we want to do uncharge as soon as possible.
2918 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2919 goto direct_uncharge;
2922 goto direct_uncharge;
2925 * In typical case, batch->memcg == mem. This means we can
2926 * merge a series of uncharges to an uncharge of res_counter.
2927 * If not, we uncharge res_counter ony by one.
2929 if (batch->memcg != memcg)
2930 goto direct_uncharge;
2931 /* remember freed charge and uncharge it later */
2934 batch->memsw_nr_pages++;
2937 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2939 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2940 if (unlikely(batch->memcg != memcg))
2941 memcg_oom_recover(memcg);
2946 * uncharge if !page_mapped(page)
2948 static struct mem_cgroup *
2949 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2951 struct mem_cgroup *memcg = NULL;
2952 unsigned int nr_pages = 1;
2953 struct page_cgroup *pc;
2955 if (mem_cgroup_disabled())
2958 if (PageSwapCache(page))
2961 if (PageTransHuge(page)) {
2962 nr_pages <<= compound_order(page);
2963 VM_BUG_ON(!PageTransHuge(page));
2966 * Check if our page_cgroup is valid
2968 pc = lookup_page_cgroup(page);
2969 if (unlikely(!pc || !PageCgroupUsed(pc)))
2972 lock_page_cgroup(pc);
2974 memcg = pc->mem_cgroup;
2976 if (!PageCgroupUsed(pc))
2980 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2981 case MEM_CGROUP_CHARGE_TYPE_DROP:
2982 /* See mem_cgroup_prepare_migration() */
2983 if (page_mapped(page) || PageCgroupMigration(pc))
2986 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2987 if (!PageAnon(page)) { /* Shared memory */
2988 if (page->mapping && !page_is_file_cache(page))
2990 } else if (page_mapped(page)) /* Anon */
2997 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
2999 ClearPageCgroupUsed(pc);
3001 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3002 * freed from LRU. This is safe because uncharged page is expected not
3003 * to be reused (freed soon). Exception is SwapCache, it's handled by
3004 * special functions.
3007 unlock_page_cgroup(pc);
3009 * even after unlock, we have memcg->res.usage here and this memcg
3010 * will never be freed.
3012 memcg_check_events(memcg, page);
3013 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3014 mem_cgroup_swap_statistics(memcg, true);
3015 mem_cgroup_get(memcg);
3017 if (!mem_cgroup_is_root(memcg))
3018 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3023 unlock_page_cgroup(pc);
3027 void mem_cgroup_uncharge_page(struct page *page)
3030 if (page_mapped(page))
3032 if (page->mapping && !PageAnon(page))
3034 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3037 void mem_cgroup_uncharge_cache_page(struct page *page)
3039 VM_BUG_ON(page_mapped(page));
3040 VM_BUG_ON(page->mapping);
3041 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3045 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3046 * In that cases, pages are freed continuously and we can expect pages
3047 * are in the same memcg. All these calls itself limits the number of
3048 * pages freed at once, then uncharge_start/end() is called properly.
3049 * This may be called prural(2) times in a context,
3052 void mem_cgroup_uncharge_start(void)
3054 current->memcg_batch.do_batch++;
3055 /* We can do nest. */
3056 if (current->memcg_batch.do_batch == 1) {
3057 current->memcg_batch.memcg = NULL;
3058 current->memcg_batch.nr_pages = 0;
3059 current->memcg_batch.memsw_nr_pages = 0;
3063 void mem_cgroup_uncharge_end(void)
3065 struct memcg_batch_info *batch = ¤t->memcg_batch;
3067 if (!batch->do_batch)
3071 if (batch->do_batch) /* If stacked, do nothing. */
3077 * This "batch->memcg" is valid without any css_get/put etc...
3078 * bacause we hide charges behind us.
3080 if (batch->nr_pages)
3081 res_counter_uncharge(&batch->memcg->res,
3082 batch->nr_pages * PAGE_SIZE);
3083 if (batch->memsw_nr_pages)
3084 res_counter_uncharge(&batch->memcg->memsw,
3085 batch->memsw_nr_pages * PAGE_SIZE);
3086 memcg_oom_recover(batch->memcg);
3087 /* forget this pointer (for sanity check) */
3088 batch->memcg = NULL;
3093 * called after __delete_from_swap_cache() and drop "page" account.
3094 * memcg information is recorded to swap_cgroup of "ent"
3097 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3099 struct mem_cgroup *memcg;
3100 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3102 if (!swapout) /* this was a swap cache but the swap is unused ! */
3103 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3105 memcg = __mem_cgroup_uncharge_common(page, ctype);
3108 * record memcg information, if swapout && memcg != NULL,
3109 * mem_cgroup_get() was called in uncharge().
3111 if (do_swap_account && swapout && memcg)
3112 swap_cgroup_record(ent, css_id(&memcg->css));
3116 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3118 * called from swap_entry_free(). remove record in swap_cgroup and
3119 * uncharge "memsw" account.
3121 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3123 struct mem_cgroup *memcg;
3126 if (!do_swap_account)
3129 id = swap_cgroup_record(ent, 0);
3131 memcg = mem_cgroup_lookup(id);
3134 * We uncharge this because swap is freed.
3135 * This memcg can be obsolete one. We avoid calling css_tryget
3137 if (!mem_cgroup_is_root(memcg))
3138 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3139 mem_cgroup_swap_statistics(memcg, false);
3140 mem_cgroup_put(memcg);
3146 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3147 * @entry: swap entry to be moved
3148 * @from: mem_cgroup which the entry is moved from
3149 * @to: mem_cgroup which the entry is moved to
3150 * @need_fixup: whether we should fixup res_counters and refcounts.
3152 * It succeeds only when the swap_cgroup's record for this entry is the same
3153 * as the mem_cgroup's id of @from.
3155 * Returns 0 on success, -EINVAL on failure.
3157 * The caller must have charged to @to, IOW, called res_counter_charge() about
3158 * both res and memsw, and called css_get().
3160 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3161 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3163 unsigned short old_id, new_id;
3165 old_id = css_id(&from->css);
3166 new_id = css_id(&to->css);
3168 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3169 mem_cgroup_swap_statistics(from, false);
3170 mem_cgroup_swap_statistics(to, true);
3172 * This function is only called from task migration context now.
3173 * It postpones res_counter and refcount handling till the end
3174 * of task migration(mem_cgroup_clear_mc()) for performance
3175 * improvement. But we cannot postpone mem_cgroup_get(to)
3176 * because if the process that has been moved to @to does
3177 * swap-in, the refcount of @to might be decreased to 0.
3181 if (!mem_cgroup_is_root(from))
3182 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3183 mem_cgroup_put(from);
3185 * we charged both to->res and to->memsw, so we should
3188 if (!mem_cgroup_is_root(to))
3189 res_counter_uncharge(&to->res, PAGE_SIZE);
3196 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3197 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3204 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3207 int mem_cgroup_prepare_migration(struct page *page,
3208 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3210 struct mem_cgroup *memcg = NULL;
3211 struct page_cgroup *pc;
3212 enum charge_type ctype;
3217 VM_BUG_ON(PageTransHuge(page));
3218 if (mem_cgroup_disabled())
3221 pc = lookup_page_cgroup(page);
3222 lock_page_cgroup(pc);
3223 if (PageCgroupUsed(pc)) {
3224 memcg = pc->mem_cgroup;
3225 css_get(&memcg->css);
3227 * At migrating an anonymous page, its mapcount goes down
3228 * to 0 and uncharge() will be called. But, even if it's fully
3229 * unmapped, migration may fail and this page has to be
3230 * charged again. We set MIGRATION flag here and delay uncharge
3231 * until end_migration() is called
3233 * Corner Case Thinking
3235 * When the old page was mapped as Anon and it's unmap-and-freed
3236 * while migration was ongoing.
3237 * If unmap finds the old page, uncharge() of it will be delayed
3238 * until end_migration(). If unmap finds a new page, it's
3239 * uncharged when it make mapcount to be 1->0. If unmap code
3240 * finds swap_migration_entry, the new page will not be mapped
3241 * and end_migration() will find it(mapcount==0).
3244 * When the old page was mapped but migraion fails, the kernel
3245 * remaps it. A charge for it is kept by MIGRATION flag even
3246 * if mapcount goes down to 0. We can do remap successfully
3247 * without charging it again.
3250 * The "old" page is under lock_page() until the end of
3251 * migration, so, the old page itself will not be swapped-out.
3252 * If the new page is swapped out before end_migraton, our
3253 * hook to usual swap-out path will catch the event.
3256 SetPageCgroupMigration(pc);
3258 unlock_page_cgroup(pc);
3260 * If the page is not charged at this point,
3267 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3268 css_put(&memcg->css);/* drop extra refcnt */
3269 if (ret || *ptr == NULL) {
3270 if (PageAnon(page)) {
3271 lock_page_cgroup(pc);
3272 ClearPageCgroupMigration(pc);
3273 unlock_page_cgroup(pc);
3275 * The old page may be fully unmapped while we kept it.
3277 mem_cgroup_uncharge_page(page);
3282 * We charge new page before it's used/mapped. So, even if unlock_page()
3283 * is called before end_migration, we can catch all events on this new
3284 * page. In the case new page is migrated but not remapped, new page's
3285 * mapcount will be finally 0 and we call uncharge in end_migration().
3287 pc = lookup_page_cgroup(newpage);
3289 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3290 else if (page_is_file_cache(page))
3291 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3293 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3294 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3298 /* remove redundant charge if migration failed*/
3299 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3300 struct page *oldpage, struct page *newpage, bool migration_ok)
3302 struct page *used, *unused;
3303 struct page_cgroup *pc;
3307 /* blocks rmdir() */
3308 cgroup_exclude_rmdir(&memcg->css);
3309 if (!migration_ok) {
3317 * We disallowed uncharge of pages under migration because mapcount
3318 * of the page goes down to zero, temporarly.
3319 * Clear the flag and check the page should be charged.
3321 pc = lookup_page_cgroup(oldpage);
3322 lock_page_cgroup(pc);
3323 ClearPageCgroupMigration(pc);
3324 unlock_page_cgroup(pc);
3326 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3329 * If a page is a file cache, radix-tree replacement is very atomic
3330 * and we can skip this check. When it was an Anon page, its mapcount
3331 * goes down to 0. But because we added MIGRATION flage, it's not
3332 * uncharged yet. There are several case but page->mapcount check
3333 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3334 * check. (see prepare_charge() also)
3337 mem_cgroup_uncharge_page(used);
3339 * At migration, we may charge account against cgroup which has no
3341 * So, rmdir()->pre_destroy() can be called while we do this charge.
3342 * In that case, we need to call pre_destroy() again. check it here.
3344 cgroup_release_and_wakeup_rmdir(&memcg->css);
3347 #ifdef CONFIG_DEBUG_VM
3348 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3350 struct page_cgroup *pc;
3352 pc = lookup_page_cgroup(page);
3353 if (likely(pc) && PageCgroupUsed(pc))
3358 bool mem_cgroup_bad_page_check(struct page *page)
3360 if (mem_cgroup_disabled())
3363 return lookup_page_cgroup_used(page) != NULL;
3366 void mem_cgroup_print_bad_page(struct page *page)
3368 struct page_cgroup *pc;
3370 pc = lookup_page_cgroup_used(page);
3375 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3376 pc, pc->flags, pc->mem_cgroup);
3378 path = kmalloc(PATH_MAX, GFP_KERNEL);
3381 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3386 printk(KERN_CONT "(%s)\n",
3387 (ret < 0) ? "cannot get the path" : path);
3393 static DEFINE_MUTEX(set_limit_mutex);
3395 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3396 unsigned long long val)
3399 u64 memswlimit, memlimit;
3401 int children = mem_cgroup_count_children(memcg);
3402 u64 curusage, oldusage;
3406 * For keeping hierarchical_reclaim simple, how long we should retry
3407 * is depends on callers. We set our retry-count to be function
3408 * of # of children which we should visit in this loop.
3410 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3412 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3415 while (retry_count) {
3416 if (signal_pending(current)) {
3421 * Rather than hide all in some function, I do this in
3422 * open coded manner. You see what this really does.
3423 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3425 mutex_lock(&set_limit_mutex);
3426 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3427 if (memswlimit < val) {
3429 mutex_unlock(&set_limit_mutex);
3433 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3437 ret = res_counter_set_limit(&memcg->res, val);
3439 if (memswlimit == val)
3440 memcg->memsw_is_minimum = true;
3442 memcg->memsw_is_minimum = false;
3444 mutex_unlock(&set_limit_mutex);
3449 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3450 MEM_CGROUP_RECLAIM_SHRINK,
3452 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3453 /* Usage is reduced ? */
3454 if (curusage >= oldusage)
3457 oldusage = curusage;
3459 if (!ret && enlarge)
3460 memcg_oom_recover(memcg);
3465 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3466 unsigned long long val)
3469 u64 memlimit, memswlimit, oldusage, curusage;
3470 int children = mem_cgroup_count_children(memcg);
3474 /* see mem_cgroup_resize_res_limit */
3475 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3476 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3477 while (retry_count) {
3478 if (signal_pending(current)) {
3483 * Rather than hide all in some function, I do this in
3484 * open coded manner. You see what this really does.
3485 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3487 mutex_lock(&set_limit_mutex);
3488 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3489 if (memlimit > val) {
3491 mutex_unlock(&set_limit_mutex);
3494 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3495 if (memswlimit < val)
3497 ret = res_counter_set_limit(&memcg->memsw, val);
3499 if (memlimit == val)
3500 memcg->memsw_is_minimum = true;
3502 memcg->memsw_is_minimum = false;
3504 mutex_unlock(&set_limit_mutex);
3509 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3510 MEM_CGROUP_RECLAIM_NOSWAP |
3511 MEM_CGROUP_RECLAIM_SHRINK,
3513 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3514 /* Usage is reduced ? */
3515 if (curusage >= oldusage)
3518 oldusage = curusage;
3520 if (!ret && enlarge)
3521 memcg_oom_recover(memcg);
3525 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3527 unsigned long *total_scanned)
3529 unsigned long nr_reclaimed = 0;
3530 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3531 unsigned long reclaimed;
3533 struct mem_cgroup_tree_per_zone *mctz;
3534 unsigned long long excess;
3535 unsigned long nr_scanned;
3540 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3542 * This loop can run a while, specially if mem_cgroup's continuously
3543 * keep exceeding their soft limit and putting the system under
3550 mz = mem_cgroup_largest_soft_limit_node(mctz);
3555 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3557 MEM_CGROUP_RECLAIM_SOFT,
3559 nr_reclaimed += reclaimed;
3560 *total_scanned += nr_scanned;
3561 spin_lock(&mctz->lock);
3564 * If we failed to reclaim anything from this memory cgroup
3565 * it is time to move on to the next cgroup
3571 * Loop until we find yet another one.
3573 * By the time we get the soft_limit lock
3574 * again, someone might have aded the
3575 * group back on the RB tree. Iterate to
3576 * make sure we get a different mem.
3577 * mem_cgroup_largest_soft_limit_node returns
3578 * NULL if no other cgroup is present on
3582 __mem_cgroup_largest_soft_limit_node(mctz);
3584 css_put(&next_mz->mem->css);
3585 else /* next_mz == NULL or other memcg */
3589 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3590 excess = res_counter_soft_limit_excess(&mz->mem->res);
3592 * One school of thought says that we should not add
3593 * back the node to the tree if reclaim returns 0.
3594 * But our reclaim could return 0, simply because due
3595 * to priority we are exposing a smaller subset of
3596 * memory to reclaim from. Consider this as a longer
3599 /* If excess == 0, no tree ops */
3600 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3601 spin_unlock(&mctz->lock);
3602 css_put(&mz->mem->css);
3605 * Could not reclaim anything and there are no more
3606 * mem cgroups to try or we seem to be looping without
3607 * reclaiming anything.
3609 if (!nr_reclaimed &&
3611 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3613 } while (!nr_reclaimed);
3615 css_put(&next_mz->mem->css);
3616 return nr_reclaimed;
3620 * This routine traverse page_cgroup in given list and drop them all.
3621 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3623 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3624 int node, int zid, enum lru_list lru)
3627 struct mem_cgroup_per_zone *mz;
3628 struct page_cgroup *pc, *busy;
3629 unsigned long flags, loop;
3630 struct list_head *list;
3633 zone = &NODE_DATA(node)->node_zones[zid];
3634 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3635 list = &mz->lists[lru];
3637 loop = MEM_CGROUP_ZSTAT(mz, lru);
3638 /* give some margin against EBUSY etc...*/
3645 spin_lock_irqsave(&zone->lru_lock, flags);
3646 if (list_empty(list)) {
3647 spin_unlock_irqrestore(&zone->lru_lock, flags);
3650 pc = list_entry(list->prev, struct page_cgroup, lru);
3652 list_move(&pc->lru, list);
3654 spin_unlock_irqrestore(&zone->lru_lock, flags);
3657 spin_unlock_irqrestore(&zone->lru_lock, flags);
3659 page = lookup_cgroup_page(pc);
3661 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3665 if (ret == -EBUSY || ret == -EINVAL) {
3666 /* found lock contention or "pc" is obsolete. */
3673 if (!ret && !list_empty(list))
3679 * make mem_cgroup's charge to be 0 if there is no task.
3680 * This enables deleting this mem_cgroup.
3682 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3685 int node, zid, shrink;
3686 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3687 struct cgroup *cgrp = memcg->css.cgroup;
3689 css_get(&memcg->css);
3692 /* should free all ? */
3698 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3701 if (signal_pending(current))
3703 /* This is for making all *used* pages to be on LRU. */
3704 lru_add_drain_all();
3705 drain_all_stock_sync(memcg);
3707 mem_cgroup_start_move(memcg);
3708 for_each_node_state(node, N_HIGH_MEMORY) {
3709 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3712 ret = mem_cgroup_force_empty_list(memcg,
3721 mem_cgroup_end_move(memcg);
3722 memcg_oom_recover(memcg);
3723 /* it seems parent cgroup doesn't have enough mem */
3727 /* "ret" should also be checked to ensure all lists are empty. */
3728 } while (memcg->res.usage > 0 || ret);
3730 css_put(&memcg->css);
3734 /* returns EBUSY if there is a task or if we come here twice. */
3735 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3739 /* we call try-to-free pages for make this cgroup empty */
3740 lru_add_drain_all();
3741 /* try to free all pages in this cgroup */
3743 while (nr_retries && memcg->res.usage > 0) {
3746 if (signal_pending(current)) {
3750 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3754 /* maybe some writeback is necessary */
3755 congestion_wait(BLK_RW_ASYNC, HZ/10);
3760 /* try move_account...there may be some *locked* pages. */
3764 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3766 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3770 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3772 return mem_cgroup_from_cont(cont)->use_hierarchy;
3775 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3779 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3780 struct cgroup *parent = cont->parent;
3781 struct mem_cgroup *parent_memcg = NULL;
3784 parent_memcg = mem_cgroup_from_cont(parent);
3788 * If parent's use_hierarchy is set, we can't make any modifications
3789 * in the child subtrees. If it is unset, then the change can
3790 * occur, provided the current cgroup has no children.
3792 * For the root cgroup, parent_mem is NULL, we allow value to be
3793 * set if there are no children.
3795 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3796 (val == 1 || val == 0)) {
3797 if (list_empty(&cont->children))
3798 memcg->use_hierarchy = val;
3809 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3810 enum mem_cgroup_stat_index idx)
3812 struct mem_cgroup *iter;
3815 /* Per-cpu values can be negative, use a signed accumulator */
3816 for_each_mem_cgroup_tree(iter, memcg)
3817 val += mem_cgroup_read_stat(iter, idx);
3819 if (val < 0) /* race ? */
3824 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3828 if (!mem_cgroup_is_root(memcg)) {
3830 return res_counter_read_u64(&memcg->res, RES_USAGE);
3832 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3835 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3836 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3839 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3841 return val << PAGE_SHIFT;
3844 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3846 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3850 type = MEMFILE_TYPE(cft->private);
3851 name = MEMFILE_ATTR(cft->private);
3854 if (name == RES_USAGE)
3855 val = mem_cgroup_usage(memcg, false);
3857 val = res_counter_read_u64(&memcg->res, name);
3860 if (name == RES_USAGE)
3861 val = mem_cgroup_usage(memcg, true);
3863 val = res_counter_read_u64(&memcg->memsw, name);
3872 * The user of this function is...
3875 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3878 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3880 unsigned long long val;
3883 type = MEMFILE_TYPE(cft->private);
3884 name = MEMFILE_ATTR(cft->private);
3887 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3891 /* This function does all necessary parse...reuse it */
3892 ret = res_counter_memparse_write_strategy(buffer, &val);
3896 ret = mem_cgroup_resize_limit(memcg, val);
3898 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3900 case RES_SOFT_LIMIT:
3901 ret = res_counter_memparse_write_strategy(buffer, &val);
3905 * For memsw, soft limits are hard to implement in terms
3906 * of semantics, for now, we support soft limits for
3907 * control without swap
3910 ret = res_counter_set_soft_limit(&memcg->res, val);
3915 ret = -EINVAL; /* should be BUG() ? */
3921 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3922 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3924 struct cgroup *cgroup;
3925 unsigned long long min_limit, min_memsw_limit, tmp;
3927 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3928 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3929 cgroup = memcg->css.cgroup;
3930 if (!memcg->use_hierarchy)
3933 while (cgroup->parent) {
3934 cgroup = cgroup->parent;
3935 memcg = mem_cgroup_from_cont(cgroup);
3936 if (!memcg->use_hierarchy)
3938 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3939 min_limit = min(min_limit, tmp);
3940 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3941 min_memsw_limit = min(min_memsw_limit, tmp);
3944 *mem_limit = min_limit;
3945 *memsw_limit = min_memsw_limit;
3949 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3951 struct mem_cgroup *memcg;
3954 memcg = mem_cgroup_from_cont(cont);
3955 type = MEMFILE_TYPE(event);
3956 name = MEMFILE_ATTR(event);
3960 res_counter_reset_max(&memcg->res);
3962 res_counter_reset_max(&memcg->memsw);
3966 res_counter_reset_failcnt(&memcg->res);
3968 res_counter_reset_failcnt(&memcg->memsw);
3975 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3978 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3982 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3983 struct cftype *cft, u64 val)
3985 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3987 if (val >= (1 << NR_MOVE_TYPE))
3990 * We check this value several times in both in can_attach() and
3991 * attach(), so we need cgroup lock to prevent this value from being
3995 memcg->move_charge_at_immigrate = val;
4001 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4002 struct cftype *cft, u64 val)
4009 /* For read statistics */
4027 struct mcs_total_stat {
4028 s64 stat[NR_MCS_STAT];
4034 } memcg_stat_strings[NR_MCS_STAT] = {
4035 {"cache", "total_cache"},
4036 {"rss", "total_rss"},
4037 {"mapped_file", "total_mapped_file"},
4038 {"pgpgin", "total_pgpgin"},
4039 {"pgpgout", "total_pgpgout"},
4040 {"swap", "total_swap"},
4041 {"pgfault", "total_pgfault"},
4042 {"pgmajfault", "total_pgmajfault"},
4043 {"inactive_anon", "total_inactive_anon"},
4044 {"active_anon", "total_active_anon"},
4045 {"inactive_file", "total_inactive_file"},
4046 {"active_file", "total_active_file"},
4047 {"unevictable", "total_unevictable"}
4052 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4057 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4058 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4059 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4060 s->stat[MCS_RSS] += val * PAGE_SIZE;
4061 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4062 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4063 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4064 s->stat[MCS_PGPGIN] += val;
4065 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4066 s->stat[MCS_PGPGOUT] += val;
4067 if (do_swap_account) {
4068 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4069 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4071 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4072 s->stat[MCS_PGFAULT] += val;
4073 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4074 s->stat[MCS_PGMAJFAULT] += val;
4077 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4078 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4079 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4080 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4081 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4082 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4083 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4084 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4085 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4086 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4090 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4092 struct mem_cgroup *iter;
4094 for_each_mem_cgroup_tree(iter, memcg)
4095 mem_cgroup_get_local_stat(iter, s);
4099 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4102 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4103 unsigned long node_nr;
4104 struct cgroup *cont = m->private;
4105 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4107 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4108 seq_printf(m, "total=%lu", total_nr);
4109 for_each_node_state(nid, N_HIGH_MEMORY) {
4110 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4111 seq_printf(m, " N%d=%lu", nid, node_nr);
4115 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4116 seq_printf(m, "file=%lu", file_nr);
4117 for_each_node_state(nid, N_HIGH_MEMORY) {
4118 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4120 seq_printf(m, " N%d=%lu", nid, node_nr);
4124 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4125 seq_printf(m, "anon=%lu", anon_nr);
4126 for_each_node_state(nid, N_HIGH_MEMORY) {
4127 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4129 seq_printf(m, " N%d=%lu", nid, node_nr);
4133 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4134 seq_printf(m, "unevictable=%lu", unevictable_nr);
4135 for_each_node_state(nid, N_HIGH_MEMORY) {
4136 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4137 BIT(LRU_UNEVICTABLE));
4138 seq_printf(m, " N%d=%lu", nid, node_nr);
4143 #endif /* CONFIG_NUMA */
4145 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4146 struct cgroup_map_cb *cb)
4148 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4149 struct mcs_total_stat mystat;
4152 memset(&mystat, 0, sizeof(mystat));
4153 mem_cgroup_get_local_stat(mem_cont, &mystat);
4156 for (i = 0; i < NR_MCS_STAT; i++) {
4157 if (i == MCS_SWAP && !do_swap_account)
4159 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4162 /* Hierarchical information */
4164 unsigned long long limit, memsw_limit;
4165 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4166 cb->fill(cb, "hierarchical_memory_limit", limit);
4167 if (do_swap_account)
4168 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4171 memset(&mystat, 0, sizeof(mystat));
4172 mem_cgroup_get_total_stat(mem_cont, &mystat);
4173 for (i = 0; i < NR_MCS_STAT; i++) {
4174 if (i == MCS_SWAP && !do_swap_account)
4176 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4179 #ifdef CONFIG_DEBUG_VM
4182 struct mem_cgroup_per_zone *mz;
4183 unsigned long recent_rotated[2] = {0, 0};
4184 unsigned long recent_scanned[2] = {0, 0};
4186 for_each_online_node(nid)
4187 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4188 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4190 recent_rotated[0] +=
4191 mz->reclaim_stat.recent_rotated[0];
4192 recent_rotated[1] +=
4193 mz->reclaim_stat.recent_rotated[1];
4194 recent_scanned[0] +=
4195 mz->reclaim_stat.recent_scanned[0];
4196 recent_scanned[1] +=
4197 mz->reclaim_stat.recent_scanned[1];
4199 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4200 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4201 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4202 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4209 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4211 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4213 return mem_cgroup_swappiness(memcg);
4216 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4219 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4220 struct mem_cgroup *parent;
4225 if (cgrp->parent == NULL)
4228 parent = mem_cgroup_from_cont(cgrp->parent);
4232 /* If under hierarchy, only empty-root can set this value */
4233 if ((parent->use_hierarchy) ||
4234 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4239 memcg->swappiness = val;
4246 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4248 struct mem_cgroup_threshold_ary *t;
4254 t = rcu_dereference(memcg->thresholds.primary);
4256 t = rcu_dereference(memcg->memsw_thresholds.primary);
4261 usage = mem_cgroup_usage(memcg, swap);
4264 * current_threshold points to threshold just below usage.
4265 * If it's not true, a threshold was crossed after last
4266 * call of __mem_cgroup_threshold().
4268 i = t->current_threshold;
4271 * Iterate backward over array of thresholds starting from
4272 * current_threshold and check if a threshold is crossed.
4273 * If none of thresholds below usage is crossed, we read
4274 * only one element of the array here.
4276 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4277 eventfd_signal(t->entries[i].eventfd, 1);
4279 /* i = current_threshold + 1 */
4283 * Iterate forward over array of thresholds starting from
4284 * current_threshold+1 and check if a threshold is crossed.
4285 * If none of thresholds above usage is crossed, we read
4286 * only one element of the array here.
4288 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4289 eventfd_signal(t->entries[i].eventfd, 1);
4291 /* Update current_threshold */
4292 t->current_threshold = i - 1;
4297 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4300 __mem_cgroup_threshold(memcg, false);
4301 if (do_swap_account)
4302 __mem_cgroup_threshold(memcg, true);
4304 memcg = parent_mem_cgroup(memcg);
4308 static int compare_thresholds(const void *a, const void *b)
4310 const struct mem_cgroup_threshold *_a = a;
4311 const struct mem_cgroup_threshold *_b = b;
4313 return _a->threshold - _b->threshold;
4316 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4318 struct mem_cgroup_eventfd_list *ev;
4320 list_for_each_entry(ev, &memcg->oom_notify, list)
4321 eventfd_signal(ev->eventfd, 1);
4325 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4327 struct mem_cgroup *iter;
4329 for_each_mem_cgroup_tree(iter, memcg)
4330 mem_cgroup_oom_notify_cb(iter);
4333 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4334 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4336 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4337 struct mem_cgroup_thresholds *thresholds;
4338 struct mem_cgroup_threshold_ary *new;
4339 int type = MEMFILE_TYPE(cft->private);
4340 u64 threshold, usage;
4343 ret = res_counter_memparse_write_strategy(args, &threshold);
4347 mutex_lock(&memcg->thresholds_lock);
4350 thresholds = &memcg->thresholds;
4351 else if (type == _MEMSWAP)
4352 thresholds = &memcg->memsw_thresholds;
4356 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4358 /* Check if a threshold crossed before adding a new one */
4359 if (thresholds->primary)
4360 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4362 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4364 /* Allocate memory for new array of thresholds */
4365 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4373 /* Copy thresholds (if any) to new array */
4374 if (thresholds->primary) {
4375 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4376 sizeof(struct mem_cgroup_threshold));
4379 /* Add new threshold */
4380 new->entries[size - 1].eventfd = eventfd;
4381 new->entries[size - 1].threshold = threshold;
4383 /* Sort thresholds. Registering of new threshold isn't time-critical */
4384 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4385 compare_thresholds, NULL);
4387 /* Find current threshold */
4388 new->current_threshold = -1;
4389 for (i = 0; i < size; i++) {
4390 if (new->entries[i].threshold < usage) {
4392 * new->current_threshold will not be used until
4393 * rcu_assign_pointer(), so it's safe to increment
4396 ++new->current_threshold;
4400 /* Free old spare buffer and save old primary buffer as spare */
4401 kfree(thresholds->spare);
4402 thresholds->spare = thresholds->primary;
4404 rcu_assign_pointer(thresholds->primary, new);
4406 /* To be sure that nobody uses thresholds */
4410 mutex_unlock(&memcg->thresholds_lock);
4415 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4416 struct cftype *cft, struct eventfd_ctx *eventfd)
4418 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4419 struct mem_cgroup_thresholds *thresholds;
4420 struct mem_cgroup_threshold_ary *new;
4421 int type = MEMFILE_TYPE(cft->private);
4425 mutex_lock(&memcg->thresholds_lock);
4427 thresholds = &memcg->thresholds;
4428 else if (type == _MEMSWAP)
4429 thresholds = &memcg->memsw_thresholds;
4434 * Something went wrong if we trying to unregister a threshold
4435 * if we don't have thresholds
4437 BUG_ON(!thresholds);
4439 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4441 /* Check if a threshold crossed before removing */
4442 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4444 /* Calculate new number of threshold */
4446 for (i = 0; i < thresholds->primary->size; i++) {
4447 if (thresholds->primary->entries[i].eventfd != eventfd)
4451 new = thresholds->spare;
4453 /* Set thresholds array to NULL if we don't have thresholds */
4462 /* Copy thresholds and find current threshold */
4463 new->current_threshold = -1;
4464 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4465 if (thresholds->primary->entries[i].eventfd == eventfd)
4468 new->entries[j] = thresholds->primary->entries[i];
4469 if (new->entries[j].threshold < usage) {
4471 * new->current_threshold will not be used
4472 * until rcu_assign_pointer(), so it's safe to increment
4475 ++new->current_threshold;
4481 /* Swap primary and spare array */
4482 thresholds->spare = thresholds->primary;
4483 rcu_assign_pointer(thresholds->primary, new);
4485 /* To be sure that nobody uses thresholds */
4488 mutex_unlock(&memcg->thresholds_lock);
4491 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4492 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4494 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4495 struct mem_cgroup_eventfd_list *event;
4496 int type = MEMFILE_TYPE(cft->private);
4498 BUG_ON(type != _OOM_TYPE);
4499 event = kmalloc(sizeof(*event), GFP_KERNEL);
4503 spin_lock(&memcg_oom_lock);
4505 event->eventfd = eventfd;
4506 list_add(&event->list, &memcg->oom_notify);
4508 /* already in OOM ? */
4509 if (atomic_read(&memcg->under_oom))
4510 eventfd_signal(eventfd, 1);
4511 spin_unlock(&memcg_oom_lock);
4516 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4517 struct cftype *cft, struct eventfd_ctx *eventfd)
4519 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4520 struct mem_cgroup_eventfd_list *ev, *tmp;
4521 int type = MEMFILE_TYPE(cft->private);
4523 BUG_ON(type != _OOM_TYPE);
4525 spin_lock(&memcg_oom_lock);
4527 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4528 if (ev->eventfd == eventfd) {
4529 list_del(&ev->list);
4534 spin_unlock(&memcg_oom_lock);
4537 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4538 struct cftype *cft, struct cgroup_map_cb *cb)
4540 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4542 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4544 if (atomic_read(&memcg->under_oom))
4545 cb->fill(cb, "under_oom", 1);
4547 cb->fill(cb, "under_oom", 0);
4551 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4552 struct cftype *cft, u64 val)
4554 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4555 struct mem_cgroup *parent;
4557 /* cannot set to root cgroup and only 0 and 1 are allowed */
4558 if (!cgrp->parent || !((val == 0) || (val == 1)))
4561 parent = mem_cgroup_from_cont(cgrp->parent);
4564 /* oom-kill-disable is a flag for subhierarchy. */
4565 if ((parent->use_hierarchy) ||
4566 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4570 memcg->oom_kill_disable = val;
4572 memcg_oom_recover(memcg);
4578 static const struct file_operations mem_control_numa_stat_file_operations = {
4580 .llseek = seq_lseek,
4581 .release = single_release,
4584 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4586 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4588 file->f_op = &mem_control_numa_stat_file_operations;
4589 return single_open(file, mem_control_numa_stat_show, cont);
4591 #endif /* CONFIG_NUMA */
4593 static struct cftype mem_cgroup_files[] = {
4595 .name = "usage_in_bytes",
4596 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4597 .read_u64 = mem_cgroup_read,
4598 .register_event = mem_cgroup_usage_register_event,
4599 .unregister_event = mem_cgroup_usage_unregister_event,
4602 .name = "max_usage_in_bytes",
4603 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4604 .trigger = mem_cgroup_reset,
4605 .read_u64 = mem_cgroup_read,
4608 .name = "limit_in_bytes",
4609 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4610 .write_string = mem_cgroup_write,
4611 .read_u64 = mem_cgroup_read,
4614 .name = "soft_limit_in_bytes",
4615 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4616 .write_string = mem_cgroup_write,
4617 .read_u64 = mem_cgroup_read,
4621 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4622 .trigger = mem_cgroup_reset,
4623 .read_u64 = mem_cgroup_read,
4627 .read_map = mem_control_stat_show,
4630 .name = "force_empty",
4631 .trigger = mem_cgroup_force_empty_write,
4634 .name = "use_hierarchy",
4635 .write_u64 = mem_cgroup_hierarchy_write,
4636 .read_u64 = mem_cgroup_hierarchy_read,
4639 .name = "swappiness",
4640 .read_u64 = mem_cgroup_swappiness_read,
4641 .write_u64 = mem_cgroup_swappiness_write,
4644 .name = "move_charge_at_immigrate",
4645 .read_u64 = mem_cgroup_move_charge_read,
4646 .write_u64 = mem_cgroup_move_charge_write,
4649 .name = "oom_control",
4650 .read_map = mem_cgroup_oom_control_read,
4651 .write_u64 = mem_cgroup_oom_control_write,
4652 .register_event = mem_cgroup_oom_register_event,
4653 .unregister_event = mem_cgroup_oom_unregister_event,
4654 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4658 .name = "numa_stat",
4659 .open = mem_control_numa_stat_open,
4665 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4666 static struct cftype memsw_cgroup_files[] = {
4668 .name = "memsw.usage_in_bytes",
4669 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4670 .read_u64 = mem_cgroup_read,
4671 .register_event = mem_cgroup_usage_register_event,
4672 .unregister_event = mem_cgroup_usage_unregister_event,
4675 .name = "memsw.max_usage_in_bytes",
4676 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4677 .trigger = mem_cgroup_reset,
4678 .read_u64 = mem_cgroup_read,
4681 .name = "memsw.limit_in_bytes",
4682 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4683 .write_string = mem_cgroup_write,
4684 .read_u64 = mem_cgroup_read,
4687 .name = "memsw.failcnt",
4688 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4689 .trigger = mem_cgroup_reset,
4690 .read_u64 = mem_cgroup_read,
4694 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4696 if (!do_swap_account)
4698 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4699 ARRAY_SIZE(memsw_cgroup_files));
4702 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4708 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4710 struct mem_cgroup_per_node *pn;
4711 struct mem_cgroup_per_zone *mz;
4713 int zone, tmp = node;
4715 * This routine is called against possible nodes.
4716 * But it's BUG to call kmalloc() against offline node.
4718 * TODO: this routine can waste much memory for nodes which will
4719 * never be onlined. It's better to use memory hotplug callback
4722 if (!node_state(node, N_NORMAL_MEMORY))
4724 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
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;
4736 memcg->info.nodeinfo[node] = pn;
4740 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4742 kfree(memcg->info.nodeinfo[node]);
4745 static struct mem_cgroup *mem_cgroup_alloc(void)
4747 struct mem_cgroup *mem;
4748 int size = sizeof(struct mem_cgroup);
4750 /* Can be very big if MAX_NUMNODES is very big */
4751 if (size < PAGE_SIZE)
4752 mem = kzalloc(size, GFP_KERNEL);
4754 mem = vzalloc(size);
4759 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4762 spin_lock_init(&mem->pcp_counter_lock);
4766 if (size < PAGE_SIZE)
4774 * At destroying mem_cgroup, references from swap_cgroup can remain.
4775 * (scanning all at force_empty is too costly...)
4777 * Instead of clearing all references at force_empty, we remember
4778 * the number of reference from swap_cgroup and free mem_cgroup when
4779 * it goes down to 0.
4781 * Removal of cgroup itself succeeds regardless of refs from swap.
4784 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4788 mem_cgroup_remove_from_trees(memcg);
4789 free_css_id(&mem_cgroup_subsys, &memcg->css);
4791 for_each_node_state(node, N_POSSIBLE)
4792 free_mem_cgroup_per_zone_info(memcg, node);
4794 free_percpu(memcg->stat);
4795 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4801 static void mem_cgroup_get(struct mem_cgroup *memcg)
4803 atomic_inc(&memcg->refcnt);
4806 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4808 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4809 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4810 __mem_cgroup_free(memcg);
4812 mem_cgroup_put(parent);
4816 static void mem_cgroup_put(struct mem_cgroup *memcg)
4818 __mem_cgroup_put(memcg, 1);
4822 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4824 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4826 if (!memcg->res.parent)
4828 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4831 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4832 static void __init enable_swap_cgroup(void)
4834 if (!mem_cgroup_disabled() && really_do_swap_account)
4835 do_swap_account = 1;
4838 static void __init enable_swap_cgroup(void)
4843 static int mem_cgroup_soft_limit_tree_init(void)
4845 struct mem_cgroup_tree_per_node *rtpn;
4846 struct mem_cgroup_tree_per_zone *rtpz;
4847 int tmp, node, zone;
4849 for_each_node_state(node, N_POSSIBLE) {
4851 if (!node_state(node, N_NORMAL_MEMORY))
4853 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4857 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4859 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4860 rtpz = &rtpn->rb_tree_per_zone[zone];
4861 rtpz->rb_root = RB_ROOT;
4862 spin_lock_init(&rtpz->lock);
4868 static struct cgroup_subsys_state * __ref
4869 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4871 struct mem_cgroup *memcg, *parent;
4872 long error = -ENOMEM;
4875 memcg = mem_cgroup_alloc();
4877 return ERR_PTR(error);
4879 for_each_node_state(node, N_POSSIBLE)
4880 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4884 if (cont->parent == NULL) {
4886 enable_swap_cgroup();
4888 root_mem_cgroup = memcg;
4889 if (mem_cgroup_soft_limit_tree_init())
4891 for_each_possible_cpu(cpu) {
4892 struct memcg_stock_pcp *stock =
4893 &per_cpu(memcg_stock, cpu);
4894 INIT_WORK(&stock->work, drain_local_stock);
4896 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4898 parent = mem_cgroup_from_cont(cont->parent);
4899 memcg->use_hierarchy = parent->use_hierarchy;
4900 memcg->oom_kill_disable = parent->oom_kill_disable;
4903 if (parent && parent->use_hierarchy) {
4904 res_counter_init(&memcg->res, &parent->res);
4905 res_counter_init(&memcg->memsw, &parent->memsw);
4907 * We increment refcnt of the parent to ensure that we can
4908 * safely access it on res_counter_charge/uncharge.
4909 * This refcnt will be decremented when freeing this
4910 * mem_cgroup(see mem_cgroup_put).
4912 mem_cgroup_get(parent);
4914 res_counter_init(&memcg->res, NULL);
4915 res_counter_init(&memcg->memsw, NULL);
4917 memcg->last_scanned_child = 0;
4918 memcg->last_scanned_node = MAX_NUMNODES;
4919 INIT_LIST_HEAD(&memcg->oom_notify);
4922 memcg->swappiness = mem_cgroup_swappiness(parent);
4923 atomic_set(&memcg->refcnt, 1);
4924 memcg->move_charge_at_immigrate = 0;
4925 mutex_init(&memcg->thresholds_lock);
4928 __mem_cgroup_free(memcg);
4929 root_mem_cgroup = NULL;
4930 return ERR_PTR(error);
4933 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4934 struct cgroup *cont)
4936 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4938 return mem_cgroup_force_empty(memcg, false);
4941 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4942 struct cgroup *cont)
4944 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4946 mem_cgroup_put(memcg);
4949 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4950 struct cgroup *cont)
4954 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4955 ARRAY_SIZE(mem_cgroup_files));
4958 ret = register_memsw_files(cont, ss);
4963 /* Handlers for move charge at task migration. */
4964 #define PRECHARGE_COUNT_AT_ONCE 256
4965 static int mem_cgroup_do_precharge(unsigned long count)
4968 int batch_count = PRECHARGE_COUNT_AT_ONCE;
4969 struct mem_cgroup *memcg = mc.to;
4971 if (mem_cgroup_is_root(memcg)) {
4972 mc.precharge += count;
4973 /* we don't need css_get for root */
4976 /* try to charge at once */
4978 struct res_counter *dummy;
4980 * "memcg" cannot be under rmdir() because we've already checked
4981 * by cgroup_lock_live_cgroup() that it is not removed and we
4982 * are still under the same cgroup_mutex. So we can postpone
4985 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
4987 if (do_swap_account && res_counter_charge(&memcg->memsw,
4988 PAGE_SIZE * count, &dummy)) {
4989 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
4992 mc.precharge += count;
4996 /* fall back to one by one charge */
4998 if (signal_pending(current)) {
5002 if (!batch_count--) {
5003 batch_count = PRECHARGE_COUNT_AT_ONCE;
5006 ret = __mem_cgroup_try_charge(NULL,
5007 GFP_KERNEL, 1, &memcg, false);
5009 /* mem_cgroup_clear_mc() will do uncharge later */
5017 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5018 * @vma: the vma the pte to be checked belongs
5019 * @addr: the address corresponding to the pte to be checked
5020 * @ptent: the pte to be checked
5021 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5024 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5025 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5026 * move charge. if @target is not NULL, the page is stored in target->page
5027 * with extra refcnt got(Callers should handle it).
5028 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5029 * target for charge migration. if @target is not NULL, the entry is stored
5032 * Called with pte lock held.
5039 enum mc_target_type {
5040 MC_TARGET_NONE, /* not used */
5045 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5046 unsigned long addr, pte_t ptent)
5048 struct page *page = vm_normal_page(vma, addr, ptent);
5050 if (!page || !page_mapped(page))
5052 if (PageAnon(page)) {
5053 /* we don't move shared anon */
5054 if (!move_anon() || page_mapcount(page) > 2)
5056 } else if (!move_file())
5057 /* we ignore mapcount for file pages */
5059 if (!get_page_unless_zero(page))
5065 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5066 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5069 struct page *page = NULL;
5070 swp_entry_t ent = pte_to_swp_entry(ptent);
5072 if (!move_anon() || non_swap_entry(ent))
5074 usage_count = mem_cgroup_count_swap_user(ent, &page);
5075 if (usage_count > 1) { /* we don't move shared anon */
5080 if (do_swap_account)
5081 entry->val = ent.val;
5086 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5087 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5089 struct page *page = NULL;
5090 struct inode *inode;
5091 struct address_space *mapping;
5094 if (!vma->vm_file) /* anonymous vma */
5099 inode = vma->vm_file->f_path.dentry->d_inode;
5100 mapping = vma->vm_file->f_mapping;
5101 if (pte_none(ptent))
5102 pgoff = linear_page_index(vma, addr);
5103 else /* pte_file(ptent) is true */
5104 pgoff = pte_to_pgoff(ptent);
5106 /* page is moved even if it's not RSS of this task(page-faulted). */
5107 page = find_get_page(mapping, pgoff);
5110 /* shmem/tmpfs may report page out on swap: account for that too. */
5111 if (radix_tree_exceptional_entry(page)) {
5112 swp_entry_t swap = radix_to_swp_entry(page);
5113 if (do_swap_account)
5115 page = find_get_page(&swapper_space, swap.val);
5121 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5122 unsigned long addr, pte_t ptent, union mc_target *target)
5124 struct page *page = NULL;
5125 struct page_cgroup *pc;
5127 swp_entry_t ent = { .val = 0 };
5129 if (pte_present(ptent))
5130 page = mc_handle_present_pte(vma, addr, ptent);
5131 else if (is_swap_pte(ptent))
5132 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5133 else if (pte_none(ptent) || pte_file(ptent))
5134 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5136 if (!page && !ent.val)
5139 pc = lookup_page_cgroup(page);
5141 * Do only loose check w/o page_cgroup lock.
5142 * mem_cgroup_move_account() checks the pc is valid or not under
5145 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5146 ret = MC_TARGET_PAGE;
5148 target->page = page;
5150 if (!ret || !target)
5153 /* There is a swap entry and a page doesn't exist or isn't charged */
5154 if (ent.val && !ret &&
5155 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5156 ret = MC_TARGET_SWAP;
5163 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5164 unsigned long addr, unsigned long end,
5165 struct mm_walk *walk)
5167 struct vm_area_struct *vma = walk->private;
5171 split_huge_page_pmd(walk->mm, pmd);
5173 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5174 for (; addr != end; pte++, addr += PAGE_SIZE)
5175 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5176 mc.precharge++; /* increment precharge temporarily */
5177 pte_unmap_unlock(pte - 1, ptl);
5183 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5185 unsigned long precharge;
5186 struct vm_area_struct *vma;
5188 down_read(&mm->mmap_sem);
5189 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5190 struct mm_walk mem_cgroup_count_precharge_walk = {
5191 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5195 if (is_vm_hugetlb_page(vma))
5197 walk_page_range(vma->vm_start, vma->vm_end,
5198 &mem_cgroup_count_precharge_walk);
5200 up_read(&mm->mmap_sem);
5202 precharge = mc.precharge;
5208 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5210 unsigned long precharge = mem_cgroup_count_precharge(mm);
5212 VM_BUG_ON(mc.moving_task);
5213 mc.moving_task = current;
5214 return mem_cgroup_do_precharge(precharge);
5217 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5218 static void __mem_cgroup_clear_mc(void)
5220 struct mem_cgroup *from = mc.from;
5221 struct mem_cgroup *to = mc.to;
5223 /* we must uncharge all the leftover precharges from mc.to */
5225 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5229 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5230 * we must uncharge here.
5232 if (mc.moved_charge) {
5233 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5234 mc.moved_charge = 0;
5236 /* we must fixup refcnts and charges */
5237 if (mc.moved_swap) {
5238 /* uncharge swap account from the old cgroup */
5239 if (!mem_cgroup_is_root(mc.from))
5240 res_counter_uncharge(&mc.from->memsw,
5241 PAGE_SIZE * mc.moved_swap);
5242 __mem_cgroup_put(mc.from, mc.moved_swap);
5244 if (!mem_cgroup_is_root(mc.to)) {
5246 * we charged both to->res and to->memsw, so we should
5249 res_counter_uncharge(&mc.to->res,
5250 PAGE_SIZE * mc.moved_swap);
5252 /* we've already done mem_cgroup_get(mc.to) */
5255 memcg_oom_recover(from);
5256 memcg_oom_recover(to);
5257 wake_up_all(&mc.waitq);
5260 static void mem_cgroup_clear_mc(void)
5262 struct mem_cgroup *from = mc.from;
5265 * we must clear moving_task before waking up waiters at the end of
5268 mc.moving_task = NULL;
5269 __mem_cgroup_clear_mc();
5270 spin_lock(&mc.lock);
5273 spin_unlock(&mc.lock);
5274 mem_cgroup_end_move(from);
5277 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5278 struct cgroup *cgroup,
5279 struct task_struct *p)
5282 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5284 if (memcg->move_charge_at_immigrate) {
5285 struct mm_struct *mm;
5286 struct mem_cgroup *from = mem_cgroup_from_task(p);
5288 VM_BUG_ON(from == memcg);
5290 mm = get_task_mm(p);
5293 /* We move charges only when we move a owner of the mm */
5294 if (mm->owner == p) {
5297 VM_BUG_ON(mc.precharge);
5298 VM_BUG_ON(mc.moved_charge);
5299 VM_BUG_ON(mc.moved_swap);
5300 mem_cgroup_start_move(from);
5301 spin_lock(&mc.lock);
5304 spin_unlock(&mc.lock);
5305 /* We set mc.moving_task later */
5307 ret = mem_cgroup_precharge_mc(mm);
5309 mem_cgroup_clear_mc();
5316 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5317 struct cgroup *cgroup,
5318 struct task_struct *p)
5320 mem_cgroup_clear_mc();
5323 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5324 unsigned long addr, unsigned long end,
5325 struct mm_walk *walk)
5328 struct vm_area_struct *vma = walk->private;
5332 split_huge_page_pmd(walk->mm, pmd);
5334 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5335 for (; addr != end; addr += PAGE_SIZE) {
5336 pte_t ptent = *(pte++);
5337 union mc_target target;
5340 struct page_cgroup *pc;
5346 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5348 case MC_TARGET_PAGE:
5350 if (isolate_lru_page(page))
5352 pc = lookup_page_cgroup(page);
5353 if (!mem_cgroup_move_account(page, 1, pc,
5354 mc.from, mc.to, false)) {
5356 /* we uncharge from mc.from later. */
5359 putback_lru_page(page);
5360 put: /* is_target_pte_for_mc() gets the page */
5363 case MC_TARGET_SWAP:
5365 if (!mem_cgroup_move_swap_account(ent,
5366 mc.from, mc.to, false)) {
5368 /* we fixup refcnts and charges later. */
5376 pte_unmap_unlock(pte - 1, ptl);
5381 * We have consumed all precharges we got in can_attach().
5382 * We try charge one by one, but don't do any additional
5383 * charges to mc.to if we have failed in charge once in attach()
5386 ret = mem_cgroup_do_precharge(1);
5394 static void mem_cgroup_move_charge(struct mm_struct *mm)
5396 struct vm_area_struct *vma;
5398 lru_add_drain_all();
5400 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5402 * Someone who are holding the mmap_sem might be waiting in
5403 * waitq. So we cancel all extra charges, wake up all waiters,
5404 * and retry. Because we cancel precharges, we might not be able
5405 * to move enough charges, but moving charge is a best-effort
5406 * feature anyway, so it wouldn't be a big problem.
5408 __mem_cgroup_clear_mc();
5412 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5414 struct mm_walk mem_cgroup_move_charge_walk = {
5415 .pmd_entry = mem_cgroup_move_charge_pte_range,
5419 if (is_vm_hugetlb_page(vma))
5421 ret = walk_page_range(vma->vm_start, vma->vm_end,
5422 &mem_cgroup_move_charge_walk);
5425 * means we have consumed all precharges and failed in
5426 * doing additional charge. Just abandon here.
5430 up_read(&mm->mmap_sem);
5433 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5434 struct cgroup *cont,
5435 struct cgroup *old_cont,
5436 struct task_struct *p)
5438 struct mm_struct *mm = get_task_mm(p);
5442 mem_cgroup_move_charge(mm);
5447 mem_cgroup_clear_mc();
5449 #else /* !CONFIG_MMU */
5450 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5451 struct cgroup *cgroup,
5452 struct task_struct *p)
5456 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5457 struct cgroup *cgroup,
5458 struct task_struct *p)
5461 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5462 struct cgroup *cont,
5463 struct cgroup *old_cont,
5464 struct task_struct *p)
5469 struct cgroup_subsys mem_cgroup_subsys = {
5471 .subsys_id = mem_cgroup_subsys_id,
5472 .create = mem_cgroup_create,
5473 .pre_destroy = mem_cgroup_pre_destroy,
5474 .destroy = mem_cgroup_destroy,
5475 .populate = mem_cgroup_populate,
5476 .can_attach = mem_cgroup_can_attach,
5477 .cancel_attach = mem_cgroup_cancel_attach,
5478 .attach = mem_cgroup_move_task,
5483 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5484 static int __init enable_swap_account(char *s)
5486 /* consider enabled if no parameter or 1 is given */
5487 if (!strcmp(s, "1"))
5488 really_do_swap_account = 1;
5489 else if (!strcmp(s, "0"))
5490 really_do_swap_account = 0;
5493 __setup("swapaccount=", enable_swap_account);