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 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1109 unsigned long active;
1110 unsigned long inactive;
1112 unsigned long inactive_ratio;
1114 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
1115 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
1117 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1119 inactive_ratio = int_sqrt(10 * gb);
1123 if (present_pages) {
1124 present_pages[0] = inactive;
1125 present_pages[1] = active;
1128 return inactive_ratio;
1131 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1133 unsigned long active;
1134 unsigned long inactive;
1135 unsigned long present_pages[2];
1136 unsigned long inactive_ratio;
1138 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1140 inactive = present_pages[0];
1141 active = present_pages[1];
1143 if (inactive * inactive_ratio < active)
1149 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1151 unsigned long active;
1152 unsigned long inactive;
1154 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
1155 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
1157 return (active > inactive);
1160 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1163 int nid = zone_to_nid(zone);
1164 int zid = zone_idx(zone);
1165 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1167 return &mz->reclaim_stat;
1170 struct zone_reclaim_stat *
1171 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1173 struct page_cgroup *pc;
1174 struct mem_cgroup_per_zone *mz;
1176 if (mem_cgroup_disabled())
1179 pc = lookup_page_cgroup(page);
1180 if (!PageCgroupUsed(pc))
1182 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1184 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1185 return &mz->reclaim_stat;
1188 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1189 struct list_head *dst,
1190 unsigned long *scanned, int order,
1191 isolate_mode_t mode,
1193 struct mem_cgroup *mem_cont,
1194 int active, int file)
1196 unsigned long nr_taken = 0;
1200 struct list_head *src;
1201 struct page_cgroup *pc, *tmp;
1202 int nid = zone_to_nid(z);
1203 int zid = zone_idx(z);
1204 struct mem_cgroup_per_zone *mz;
1205 int lru = LRU_FILE * file + active;
1209 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1210 src = &mz->lists[lru];
1213 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1214 if (scan >= nr_to_scan)
1217 if (unlikely(!PageCgroupUsed(pc)))
1220 page = lookup_cgroup_page(pc);
1222 if (unlikely(!PageLRU(page)))
1226 ret = __isolate_lru_page(page, mode, file);
1229 list_move(&page->lru, dst);
1230 mem_cgroup_del_lru(page);
1231 nr_taken += hpage_nr_pages(page);
1234 /* we don't affect global LRU but rotate in our LRU */
1235 mem_cgroup_rotate_lru_list(page, page_lru(page));
1244 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1250 #define mem_cgroup_from_res_counter(counter, member) \
1251 container_of(counter, struct mem_cgroup, member)
1254 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1255 * @mem: the memory cgroup
1257 * Returns the maximum amount of memory @mem can be charged with, in
1260 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1262 unsigned long long margin;
1264 margin = res_counter_margin(&memcg->res);
1265 if (do_swap_account)
1266 margin = min(margin, res_counter_margin(&memcg->memsw));
1267 return margin >> PAGE_SHIFT;
1270 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1272 struct cgroup *cgrp = memcg->css.cgroup;
1275 if (cgrp->parent == NULL)
1276 return vm_swappiness;
1278 return memcg->swappiness;
1281 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1286 spin_lock(&memcg->pcp_counter_lock);
1287 for_each_online_cpu(cpu)
1288 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1289 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1290 spin_unlock(&memcg->pcp_counter_lock);
1296 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1303 spin_lock(&memcg->pcp_counter_lock);
1304 for_each_online_cpu(cpu)
1305 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1306 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1307 spin_unlock(&memcg->pcp_counter_lock);
1311 * 2 routines for checking "mem" is under move_account() or not.
1313 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1314 * for avoiding race in accounting. If true,
1315 * pc->mem_cgroup may be overwritten.
1317 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1318 * under hierarchy of moving cgroups. This is for
1319 * waiting at hith-memory prressure caused by "move".
1322 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1324 VM_BUG_ON(!rcu_read_lock_held());
1325 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1328 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1330 struct mem_cgroup *from;
1331 struct mem_cgroup *to;
1334 * Unlike task_move routines, we access mc.to, mc.from not under
1335 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1337 spin_lock(&mc.lock);
1343 ret = mem_cgroup_same_or_subtree(memcg, from)
1344 || mem_cgroup_same_or_subtree(memcg, to);
1346 spin_unlock(&mc.lock);
1350 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1352 if (mc.moving_task && current != mc.moving_task) {
1353 if (mem_cgroup_under_move(memcg)) {
1355 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1356 /* moving charge context might have finished. */
1359 finish_wait(&mc.waitq, &wait);
1367 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1368 * @memcg: The memory cgroup that went over limit
1369 * @p: Task that is going to be killed
1371 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1374 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1376 struct cgroup *task_cgrp;
1377 struct cgroup *mem_cgrp;
1379 * Need a buffer in BSS, can't rely on allocations. The code relies
1380 * on the assumption that OOM is serialized for memory controller.
1381 * If this assumption is broken, revisit this code.
1383 static char memcg_name[PATH_MAX];
1392 mem_cgrp = memcg->css.cgroup;
1393 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1395 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1398 * Unfortunately, we are unable to convert to a useful name
1399 * But we'll still print out the usage information
1406 printk(KERN_INFO "Task in %s killed", memcg_name);
1409 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1417 * Continues from above, so we don't need an KERN_ level
1419 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1422 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1423 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1424 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1425 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1426 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1428 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1429 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1430 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1434 * This function returns the number of memcg under hierarchy tree. Returns
1435 * 1(self count) if no children.
1437 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1440 struct mem_cgroup *iter;
1442 for_each_mem_cgroup_tree(iter, memcg)
1448 * Return the memory (and swap, if configured) limit for a memcg.
1450 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1455 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1456 limit += total_swap_pages << PAGE_SHIFT;
1458 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1460 * If memsw is finite and limits the amount of swap space available
1461 * to this memcg, return that limit.
1463 return min(limit, memsw);
1467 * Visit the first child (need not be the first child as per the ordering
1468 * of the cgroup list, since we track last_scanned_child) of @mem and use
1469 * that to reclaim free pages from.
1471 static struct mem_cgroup *
1472 mem_cgroup_select_victim(struct mem_cgroup *root_memcg)
1474 struct mem_cgroup *ret = NULL;
1475 struct cgroup_subsys_state *css;
1478 if (!root_memcg->use_hierarchy) {
1479 css_get(&root_memcg->css);
1485 nextid = root_memcg->last_scanned_child + 1;
1486 css = css_get_next(&mem_cgroup_subsys, nextid, &root_memcg->css,
1488 if (css && css_tryget(css))
1489 ret = container_of(css, struct mem_cgroup, css);
1492 /* Updates scanning parameter */
1494 /* this means start scan from ID:1 */
1495 root_memcg->last_scanned_child = 0;
1497 root_memcg->last_scanned_child = found;
1504 * test_mem_cgroup_node_reclaimable
1505 * @mem: the target memcg
1506 * @nid: the node ID to be checked.
1507 * @noswap : specify true here if the user wants flle only information.
1509 * This function returns whether the specified memcg contains any
1510 * reclaimable pages on a node. Returns true if there are any reclaimable
1511 * pages in the node.
1513 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1514 int nid, bool noswap)
1516 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1518 if (noswap || !total_swap_pages)
1520 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1525 #if MAX_NUMNODES > 1
1528 * Always updating the nodemask is not very good - even if we have an empty
1529 * list or the wrong list here, we can start from some node and traverse all
1530 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1533 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1537 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1538 * pagein/pageout changes since the last update.
1540 if (!atomic_read(&memcg->numainfo_events))
1542 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1545 /* make a nodemask where this memcg uses memory from */
1546 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1548 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1550 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1551 node_clear(nid, memcg->scan_nodes);
1554 atomic_set(&memcg->numainfo_events, 0);
1555 atomic_set(&memcg->numainfo_updating, 0);
1559 * Selecting a node where we start reclaim from. Because what we need is just
1560 * reducing usage counter, start from anywhere is O,K. Considering
1561 * memory reclaim from current node, there are pros. and cons.
1563 * Freeing memory from current node means freeing memory from a node which
1564 * we'll use or we've used. So, it may make LRU bad. And if several threads
1565 * hit limits, it will see a contention on a node. But freeing from remote
1566 * node means more costs for memory reclaim because of memory latency.
1568 * Now, we use round-robin. Better algorithm is welcomed.
1570 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1574 mem_cgroup_may_update_nodemask(memcg);
1575 node = memcg->last_scanned_node;
1577 node = next_node(node, memcg->scan_nodes);
1578 if (node == MAX_NUMNODES)
1579 node = first_node(memcg->scan_nodes);
1581 * We call this when we hit limit, not when pages are added to LRU.
1582 * No LRU may hold pages because all pages are UNEVICTABLE or
1583 * memcg is too small and all pages are not on LRU. In that case,
1584 * we use curret node.
1586 if (unlikely(node == MAX_NUMNODES))
1587 node = numa_node_id();
1589 memcg->last_scanned_node = node;
1594 * Check all nodes whether it contains reclaimable pages or not.
1595 * For quick scan, we make use of scan_nodes. This will allow us to skip
1596 * unused nodes. But scan_nodes is lazily updated and may not cotain
1597 * enough new information. We need to do double check.
1599 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1604 * quick check...making use of scan_node.
1605 * We can skip unused nodes.
1607 if (!nodes_empty(memcg->scan_nodes)) {
1608 for (nid = first_node(memcg->scan_nodes);
1610 nid = next_node(nid, memcg->scan_nodes)) {
1612 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1617 * Check rest of nodes.
1619 for_each_node_state(nid, N_HIGH_MEMORY) {
1620 if (node_isset(nid, memcg->scan_nodes))
1622 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1629 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1634 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1636 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1641 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1642 * we reclaimed from, so that we don't end up penalizing one child extensively
1643 * based on its position in the children list.
1645 * root_memcg is the original ancestor that we've been reclaim from.
1647 * We give up and return to the caller when we visit root_memcg twice.
1648 * (other groups can be removed while we're walking....)
1650 * If shrink==true, for avoiding to free too much, this returns immedieately.
1652 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_memcg,
1655 unsigned long reclaim_options,
1656 unsigned long *total_scanned)
1658 struct mem_cgroup *victim;
1661 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1662 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1663 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1664 unsigned long excess;
1665 unsigned long nr_scanned;
1667 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1669 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1670 if (!check_soft && !shrink && root_memcg->memsw_is_minimum)
1674 victim = mem_cgroup_select_victim(root_memcg);
1675 if (victim == root_memcg) {
1678 * We are not draining per cpu cached charges during
1679 * soft limit reclaim because global reclaim doesn't
1680 * care about charges. It tries to free some memory and
1681 * charges will not give any.
1683 if (!check_soft && loop >= 1)
1684 drain_all_stock_async(root_memcg);
1687 * If we have not been able to reclaim
1688 * anything, it might because there are
1689 * no reclaimable pages under this hierarchy
1691 if (!check_soft || !total) {
1692 css_put(&victim->css);
1696 * We want to do more targeted reclaim.
1697 * excess >> 2 is not to excessive so as to
1698 * reclaim too much, nor too less that we keep
1699 * coming back to reclaim from this cgroup
1701 if (total >= (excess >> 2) ||
1702 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1703 css_put(&victim->css);
1708 if (!mem_cgroup_reclaimable(victim, noswap)) {
1709 /* this cgroup's local usage == 0 */
1710 css_put(&victim->css);
1713 /* we use swappiness of local cgroup */
1715 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1716 noswap, zone, &nr_scanned);
1717 *total_scanned += nr_scanned;
1719 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1721 css_put(&victim->css);
1723 * At shrinking usage, we can't check we should stop here or
1724 * reclaim more. It's depends on callers. last_scanned_child
1725 * will work enough for keeping fairness under tree.
1731 if (!res_counter_soft_limit_excess(&root_memcg->res))
1733 } else if (mem_cgroup_margin(root_memcg))
1740 * Check OOM-Killer is already running under our hierarchy.
1741 * If someone is running, return false.
1742 * Has to be called with memcg_oom_lock
1744 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1746 struct mem_cgroup *iter, *failed = NULL;
1749 for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1750 if (iter->oom_lock) {
1752 * this subtree of our hierarchy is already locked
1753 * so we cannot give a lock.
1758 iter->oom_lock = true;
1765 * OK, we failed to lock the whole subtree so we have to clean up
1766 * what we set up to the failing subtree
1769 for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1770 if (iter == failed) {
1774 iter->oom_lock = false;
1780 * Has to be called with memcg_oom_lock
1782 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1784 struct mem_cgroup *iter;
1786 for_each_mem_cgroup_tree(iter, memcg)
1787 iter->oom_lock = false;
1791 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1793 struct mem_cgroup *iter;
1795 for_each_mem_cgroup_tree(iter, memcg)
1796 atomic_inc(&iter->under_oom);
1799 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1801 struct mem_cgroup *iter;
1804 * When a new child is created while the hierarchy is under oom,
1805 * mem_cgroup_oom_lock() may not be called. We have to use
1806 * atomic_add_unless() here.
1808 for_each_mem_cgroup_tree(iter, memcg)
1809 atomic_add_unless(&iter->under_oom, -1, 0);
1812 static DEFINE_SPINLOCK(memcg_oom_lock);
1813 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1815 struct oom_wait_info {
1816 struct mem_cgroup *mem;
1820 static int memcg_oom_wake_function(wait_queue_t *wait,
1821 unsigned mode, int sync, void *arg)
1823 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1825 struct oom_wait_info *oom_wait_info;
1827 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1828 oom_wait_memcg = oom_wait_info->mem;
1831 * Both of oom_wait_info->mem and wake_mem are stable under us.
1832 * Then we can use css_is_ancestor without taking care of RCU.
1834 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1835 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1837 return autoremove_wake_function(wait, mode, sync, arg);
1840 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1842 /* for filtering, pass "memcg" as argument. */
1843 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1846 static void memcg_oom_recover(struct mem_cgroup *memcg)
1848 if (memcg && atomic_read(&memcg->under_oom))
1849 memcg_wakeup_oom(memcg);
1853 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1855 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1857 struct oom_wait_info owait;
1858 bool locked, need_to_kill;
1861 owait.wait.flags = 0;
1862 owait.wait.func = memcg_oom_wake_function;
1863 owait.wait.private = current;
1864 INIT_LIST_HEAD(&owait.wait.task_list);
1865 need_to_kill = true;
1866 mem_cgroup_mark_under_oom(memcg);
1868 /* At first, try to OOM lock hierarchy under memcg.*/
1869 spin_lock(&memcg_oom_lock);
1870 locked = mem_cgroup_oom_lock(memcg);
1872 * Even if signal_pending(), we can't quit charge() loop without
1873 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1874 * under OOM is always welcomed, use TASK_KILLABLE here.
1876 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1877 if (!locked || memcg->oom_kill_disable)
1878 need_to_kill = false;
1880 mem_cgroup_oom_notify(memcg);
1881 spin_unlock(&memcg_oom_lock);
1884 finish_wait(&memcg_oom_waitq, &owait.wait);
1885 mem_cgroup_out_of_memory(memcg, mask);
1888 finish_wait(&memcg_oom_waitq, &owait.wait);
1890 spin_lock(&memcg_oom_lock);
1892 mem_cgroup_oom_unlock(memcg);
1893 memcg_wakeup_oom(memcg);
1894 spin_unlock(&memcg_oom_lock);
1896 mem_cgroup_unmark_under_oom(memcg);
1898 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1900 /* Give chance to dying process */
1901 schedule_timeout(1);
1906 * Currently used to update mapped file statistics, but the routine can be
1907 * generalized to update other statistics as well.
1909 * Notes: Race condition
1911 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1912 * it tends to be costly. But considering some conditions, we doesn't need
1913 * to do so _always_.
1915 * Considering "charge", lock_page_cgroup() is not required because all
1916 * file-stat operations happen after a page is attached to radix-tree. There
1917 * are no race with "charge".
1919 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1920 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1921 * if there are race with "uncharge". Statistics itself is properly handled
1924 * Considering "move", this is an only case we see a race. To make the race
1925 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1926 * possibility of race condition. If there is, we take a lock.
1929 void mem_cgroup_update_page_stat(struct page *page,
1930 enum mem_cgroup_page_stat_item idx, int val)
1932 struct mem_cgroup *memcg;
1933 struct page_cgroup *pc = lookup_page_cgroup(page);
1934 bool need_unlock = false;
1935 unsigned long uninitialized_var(flags);
1941 memcg = pc->mem_cgroup;
1942 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1944 /* pc->mem_cgroup is unstable ? */
1945 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
1946 /* take a lock against to access pc->mem_cgroup */
1947 move_lock_page_cgroup(pc, &flags);
1949 memcg = pc->mem_cgroup;
1950 if (!memcg || !PageCgroupUsed(pc))
1955 case MEMCG_NR_FILE_MAPPED:
1957 SetPageCgroupFileMapped(pc);
1958 else if (!page_mapped(page))
1959 ClearPageCgroupFileMapped(pc);
1960 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1966 this_cpu_add(memcg->stat->count[idx], val);
1969 if (unlikely(need_unlock))
1970 move_unlock_page_cgroup(pc, &flags);
1974 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1977 * size of first charge trial. "32" comes from vmscan.c's magic value.
1978 * TODO: maybe necessary to use big numbers in big irons.
1980 #define CHARGE_BATCH 32U
1981 struct memcg_stock_pcp {
1982 struct mem_cgroup *cached; /* this never be root cgroup */
1983 unsigned int nr_pages;
1984 struct work_struct work;
1985 unsigned long flags;
1986 #define FLUSHING_CACHED_CHARGE (0)
1988 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1989 static DEFINE_MUTEX(percpu_charge_mutex);
1992 * Try to consume stocked charge on this cpu. If success, one page is consumed
1993 * from local stock and true is returned. If the stock is 0 or charges from a
1994 * cgroup which is not current target, returns false. This stock will be
1997 static bool consume_stock(struct mem_cgroup *memcg)
1999 struct memcg_stock_pcp *stock;
2002 stock = &get_cpu_var(memcg_stock);
2003 if (memcg == stock->cached && stock->nr_pages)
2005 else /* need to call res_counter_charge */
2007 put_cpu_var(memcg_stock);
2012 * Returns stocks cached in percpu to res_counter and reset cached information.
2014 static void drain_stock(struct memcg_stock_pcp *stock)
2016 struct mem_cgroup *old = stock->cached;
2018 if (stock->nr_pages) {
2019 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2021 res_counter_uncharge(&old->res, bytes);
2022 if (do_swap_account)
2023 res_counter_uncharge(&old->memsw, bytes);
2024 stock->nr_pages = 0;
2026 stock->cached = NULL;
2030 * This must be called under preempt disabled or must be called by
2031 * a thread which is pinned to local cpu.
2033 static void drain_local_stock(struct work_struct *dummy)
2035 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2037 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2041 * Cache charges(val) which is from res_counter, to local per_cpu area.
2042 * This will be consumed by consume_stock() function, later.
2044 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2046 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2048 if (stock->cached != memcg) { /* reset if necessary */
2050 stock->cached = memcg;
2052 stock->nr_pages += nr_pages;
2053 put_cpu_var(memcg_stock);
2057 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2058 * of the hierarchy under it. sync flag says whether we should block
2059 * until the work is done.
2061 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2065 /* Notify other cpus that system-wide "drain" is running */
2068 for_each_online_cpu(cpu) {
2069 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2070 struct mem_cgroup *memcg;
2072 memcg = stock->cached;
2073 if (!memcg || !stock->nr_pages)
2075 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2077 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2079 drain_local_stock(&stock->work);
2081 schedule_work_on(cpu, &stock->work);
2089 for_each_online_cpu(cpu) {
2090 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2091 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2092 flush_work(&stock->work);
2099 * Tries to drain stocked charges in other cpus. This function is asynchronous
2100 * and just put a work per cpu for draining localy on each cpu. Caller can
2101 * expects some charges will be back to res_counter later but cannot wait for
2104 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2107 * If someone calls draining, avoid adding more kworker runs.
2109 if (!mutex_trylock(&percpu_charge_mutex))
2111 drain_all_stock(root_memcg, false);
2112 mutex_unlock(&percpu_charge_mutex);
2115 /* This is a synchronous drain interface. */
2116 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2118 /* called when force_empty is called */
2119 mutex_lock(&percpu_charge_mutex);
2120 drain_all_stock(root_memcg, true);
2121 mutex_unlock(&percpu_charge_mutex);
2125 * This function drains percpu counter value from DEAD cpu and
2126 * move it to local cpu. Note that this function can be preempted.
2128 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2132 spin_lock(&memcg->pcp_counter_lock);
2133 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2134 long x = per_cpu(memcg->stat->count[i], cpu);
2136 per_cpu(memcg->stat->count[i], cpu) = 0;
2137 memcg->nocpu_base.count[i] += x;
2139 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2140 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2142 per_cpu(memcg->stat->events[i], cpu) = 0;
2143 memcg->nocpu_base.events[i] += x;
2145 /* need to clear ON_MOVE value, works as a kind of lock. */
2146 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2147 spin_unlock(&memcg->pcp_counter_lock);
2150 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2152 int idx = MEM_CGROUP_ON_MOVE;
2154 spin_lock(&memcg->pcp_counter_lock);
2155 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2156 spin_unlock(&memcg->pcp_counter_lock);
2159 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2160 unsigned long action,
2163 int cpu = (unsigned long)hcpu;
2164 struct memcg_stock_pcp *stock;
2165 struct mem_cgroup *iter;
2167 if ((action == CPU_ONLINE)) {
2168 for_each_mem_cgroup_all(iter)
2169 synchronize_mem_cgroup_on_move(iter, cpu);
2173 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2176 for_each_mem_cgroup_all(iter)
2177 mem_cgroup_drain_pcp_counter(iter, cpu);
2179 stock = &per_cpu(memcg_stock, cpu);
2185 /* See __mem_cgroup_try_charge() for details */
2187 CHARGE_OK, /* success */
2188 CHARGE_RETRY, /* need to retry but retry is not bad */
2189 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2190 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2191 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2194 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2195 unsigned int nr_pages, bool oom_check)
2197 unsigned long csize = nr_pages * PAGE_SIZE;
2198 struct mem_cgroup *mem_over_limit;
2199 struct res_counter *fail_res;
2200 unsigned long flags = 0;
2203 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2206 if (!do_swap_account)
2208 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2212 res_counter_uncharge(&memcg->res, csize);
2213 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2214 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2216 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2218 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2219 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2221 * Never reclaim on behalf of optional batching, retry with a
2222 * single page instead.
2224 if (nr_pages == CHARGE_BATCH)
2225 return CHARGE_RETRY;
2227 if (!(gfp_mask & __GFP_WAIT))
2228 return CHARGE_WOULDBLOCK;
2230 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2231 gfp_mask, flags, NULL);
2232 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2233 return CHARGE_RETRY;
2235 * Even though the limit is exceeded at this point, reclaim
2236 * may have been able to free some pages. Retry the charge
2237 * before killing the task.
2239 * Only for regular pages, though: huge pages are rather
2240 * unlikely to succeed so close to the limit, and we fall back
2241 * to regular pages anyway in case of failure.
2243 if (nr_pages == 1 && ret)
2244 return CHARGE_RETRY;
2247 * At task move, charge accounts can be doubly counted. So, it's
2248 * better to wait until the end of task_move if something is going on.
2250 if (mem_cgroup_wait_acct_move(mem_over_limit))
2251 return CHARGE_RETRY;
2253 /* If we don't need to call oom-killer at el, return immediately */
2255 return CHARGE_NOMEM;
2257 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2258 return CHARGE_OOM_DIE;
2260 return CHARGE_RETRY;
2264 * Unlike exported interface, "oom" parameter is added. if oom==true,
2265 * oom-killer can be invoked.
2267 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2269 unsigned int nr_pages,
2270 struct mem_cgroup **ptr,
2273 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2274 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2275 struct mem_cgroup *memcg = NULL;
2279 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2280 * in system level. So, allow to go ahead dying process in addition to
2283 if (unlikely(test_thread_flag(TIF_MEMDIE)
2284 || fatal_signal_pending(current)))
2288 * We always charge the cgroup the mm_struct belongs to.
2289 * The mm_struct's mem_cgroup changes on task migration if the
2290 * thread group leader migrates. It's possible that mm is not
2291 * set, if so charge the init_mm (happens for pagecache usage).
2296 if (*ptr) { /* css should be a valid one */
2298 VM_BUG_ON(css_is_removed(&memcg->css));
2299 if (mem_cgroup_is_root(memcg))
2301 if (nr_pages == 1 && consume_stock(memcg))
2303 css_get(&memcg->css);
2305 struct task_struct *p;
2308 p = rcu_dereference(mm->owner);
2310 * Because we don't have task_lock(), "p" can exit.
2311 * In that case, "memcg" can point to root or p can be NULL with
2312 * race with swapoff. Then, we have small risk of mis-accouning.
2313 * But such kind of mis-account by race always happens because
2314 * we don't have cgroup_mutex(). It's overkill and we allo that
2316 * (*) swapoff at el will charge against mm-struct not against
2317 * task-struct. So, mm->owner can be NULL.
2319 memcg = mem_cgroup_from_task(p);
2320 if (!memcg || mem_cgroup_is_root(memcg)) {
2324 if (nr_pages == 1 && consume_stock(memcg)) {
2326 * It seems dagerous to access memcg without css_get().
2327 * But considering how consume_stok works, it's not
2328 * necessary. If consume_stock success, some charges
2329 * from this memcg are cached on this cpu. So, we
2330 * don't need to call css_get()/css_tryget() before
2331 * calling consume_stock().
2336 /* after here, we may be blocked. we need to get refcnt */
2337 if (!css_tryget(&memcg->css)) {
2347 /* If killed, bypass charge */
2348 if (fatal_signal_pending(current)) {
2349 css_put(&memcg->css);
2354 if (oom && !nr_oom_retries) {
2356 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2359 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2363 case CHARGE_RETRY: /* not in OOM situation but retry */
2365 css_put(&memcg->css);
2368 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2369 css_put(&memcg->css);
2371 case CHARGE_NOMEM: /* OOM routine works */
2373 css_put(&memcg->css);
2376 /* If oom, we never return -ENOMEM */
2379 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2380 css_put(&memcg->css);
2383 } while (ret != CHARGE_OK);
2385 if (batch > nr_pages)
2386 refill_stock(memcg, batch - nr_pages);
2387 css_put(&memcg->css);
2400 * Somemtimes we have to undo a charge we got by try_charge().
2401 * This function is for that and do uncharge, put css's refcnt.
2402 * gotten by try_charge().
2404 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2405 unsigned int nr_pages)
2407 if (!mem_cgroup_is_root(memcg)) {
2408 unsigned long bytes = nr_pages * PAGE_SIZE;
2410 res_counter_uncharge(&memcg->res, bytes);
2411 if (do_swap_account)
2412 res_counter_uncharge(&memcg->memsw, bytes);
2417 * A helper function to get mem_cgroup from ID. must be called under
2418 * rcu_read_lock(). The caller must check css_is_removed() or some if
2419 * it's concern. (dropping refcnt from swap can be called against removed
2422 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2424 struct cgroup_subsys_state *css;
2426 /* ID 0 is unused ID */
2429 css = css_lookup(&mem_cgroup_subsys, id);
2432 return container_of(css, struct mem_cgroup, css);
2435 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2437 struct mem_cgroup *memcg = NULL;
2438 struct page_cgroup *pc;
2442 VM_BUG_ON(!PageLocked(page));
2444 pc = lookup_page_cgroup(page);
2445 lock_page_cgroup(pc);
2446 if (PageCgroupUsed(pc)) {
2447 memcg = pc->mem_cgroup;
2448 if (memcg && !css_tryget(&memcg->css))
2450 } else if (PageSwapCache(page)) {
2451 ent.val = page_private(page);
2452 id = lookup_swap_cgroup(ent);
2454 memcg = mem_cgroup_lookup(id);
2455 if (memcg && !css_tryget(&memcg->css))
2459 unlock_page_cgroup(pc);
2463 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2465 unsigned int nr_pages,
2466 struct page_cgroup *pc,
2467 enum charge_type ctype)
2469 lock_page_cgroup(pc);
2470 if (unlikely(PageCgroupUsed(pc))) {
2471 unlock_page_cgroup(pc);
2472 __mem_cgroup_cancel_charge(memcg, nr_pages);
2476 * we don't need page_cgroup_lock about tail pages, becase they are not
2477 * accessed by any other context at this point.
2479 pc->mem_cgroup = memcg;
2481 * We access a page_cgroup asynchronously without lock_page_cgroup().
2482 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2483 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2484 * before USED bit, we need memory barrier here.
2485 * See mem_cgroup_add_lru_list(), etc.
2489 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2490 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2491 SetPageCgroupCache(pc);
2492 SetPageCgroupUsed(pc);
2494 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2495 ClearPageCgroupCache(pc);
2496 SetPageCgroupUsed(pc);
2502 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2503 unlock_page_cgroup(pc);
2505 * "charge_statistics" updated event counter. Then, check it.
2506 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2507 * if they exceeds softlimit.
2509 memcg_check_events(memcg, page);
2512 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2514 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2515 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2517 * Because tail pages are not marked as "used", set it. We're under
2518 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2520 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2522 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2523 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2524 unsigned long flags;
2526 if (mem_cgroup_disabled())
2529 * We have no races with charge/uncharge but will have races with
2530 * page state accounting.
2532 move_lock_page_cgroup(head_pc, &flags);
2534 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2535 smp_wmb(); /* see __commit_charge() */
2536 if (PageCgroupAcctLRU(head_pc)) {
2538 struct mem_cgroup_per_zone *mz;
2541 * LRU flags cannot be copied because we need to add tail
2542 *.page to LRU by generic call and our hook will be called.
2543 * We hold lru_lock, then, reduce counter directly.
2545 lru = page_lru(head);
2546 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2547 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2549 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2550 move_unlock_page_cgroup(head_pc, &flags);
2555 * mem_cgroup_move_account - move account of the page
2557 * @nr_pages: number of regular pages (>1 for huge pages)
2558 * @pc: page_cgroup of the page.
2559 * @from: mem_cgroup which the page is moved from.
2560 * @to: mem_cgroup which the page is moved to. @from != @to.
2561 * @uncharge: whether we should call uncharge and css_put against @from.
2563 * The caller must confirm following.
2564 * - page is not on LRU (isolate_page() is useful.)
2565 * - compound_lock is held when nr_pages > 1
2567 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2568 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2569 * true, this function does "uncharge" from old cgroup, but it doesn't if
2570 * @uncharge is false, so a caller should do "uncharge".
2572 static int mem_cgroup_move_account(struct page *page,
2573 unsigned int nr_pages,
2574 struct page_cgroup *pc,
2575 struct mem_cgroup *from,
2576 struct mem_cgroup *to,
2579 unsigned long flags;
2582 VM_BUG_ON(from == to);
2583 VM_BUG_ON(PageLRU(page));
2585 * The page is isolated from LRU. So, collapse function
2586 * will not handle this page. But page splitting can happen.
2587 * Do this check under compound_page_lock(). The caller should
2591 if (nr_pages > 1 && !PageTransHuge(page))
2594 lock_page_cgroup(pc);
2597 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2600 move_lock_page_cgroup(pc, &flags);
2602 if (PageCgroupFileMapped(pc)) {
2603 /* Update mapped_file data for mem_cgroup */
2605 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2606 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2609 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2611 /* This is not "cancel", but cancel_charge does all we need. */
2612 __mem_cgroup_cancel_charge(from, nr_pages);
2614 /* caller should have done css_get */
2615 pc->mem_cgroup = to;
2616 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2618 * We charges against "to" which may not have any tasks. Then, "to"
2619 * can be under rmdir(). But in current implementation, caller of
2620 * this function is just force_empty() and move charge, so it's
2621 * guaranteed that "to" is never removed. So, we don't check rmdir
2624 move_unlock_page_cgroup(pc, &flags);
2627 unlock_page_cgroup(pc);
2631 memcg_check_events(to, page);
2632 memcg_check_events(from, page);
2638 * move charges to its parent.
2641 static int mem_cgroup_move_parent(struct page *page,
2642 struct page_cgroup *pc,
2643 struct mem_cgroup *child,
2646 struct cgroup *cg = child->css.cgroup;
2647 struct cgroup *pcg = cg->parent;
2648 struct mem_cgroup *parent;
2649 unsigned int nr_pages;
2650 unsigned long uninitialized_var(flags);
2658 if (!get_page_unless_zero(page))
2660 if (isolate_lru_page(page))
2663 nr_pages = hpage_nr_pages(page);
2665 parent = mem_cgroup_from_cont(pcg);
2666 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2671 flags = compound_lock_irqsave(page);
2673 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2675 __mem_cgroup_cancel_charge(parent, nr_pages);
2678 compound_unlock_irqrestore(page, flags);
2680 putback_lru_page(page);
2688 * Charge the memory controller for page usage.
2690 * 0 if the charge was successful
2691 * < 0 if the cgroup is over its limit
2693 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2694 gfp_t gfp_mask, enum charge_type ctype)
2696 struct mem_cgroup *memcg = NULL;
2697 unsigned int nr_pages = 1;
2698 struct page_cgroup *pc;
2702 if (PageTransHuge(page)) {
2703 nr_pages <<= compound_order(page);
2704 VM_BUG_ON(!PageTransHuge(page));
2706 * Never OOM-kill a process for a huge page. The
2707 * fault handler will fall back to regular pages.
2712 pc = lookup_page_cgroup(page);
2713 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2715 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2719 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2723 int mem_cgroup_newpage_charge(struct page *page,
2724 struct mm_struct *mm, gfp_t gfp_mask)
2726 if (mem_cgroup_disabled())
2729 * If already mapped, we don't have to account.
2730 * If page cache, page->mapping has address_space.
2731 * But page->mapping may have out-of-use anon_vma pointer,
2732 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2735 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2739 return mem_cgroup_charge_common(page, mm, gfp_mask,
2740 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2744 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2745 enum charge_type ctype);
2748 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2749 enum charge_type ctype)
2751 struct page_cgroup *pc = lookup_page_cgroup(page);
2753 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2754 * is already on LRU. It means the page may on some other page_cgroup's
2755 * LRU. Take care of it.
2757 mem_cgroup_lru_del_before_commit(page);
2758 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2759 mem_cgroup_lru_add_after_commit(page);
2763 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2766 struct mem_cgroup *memcg = NULL;
2769 if (mem_cgroup_disabled())
2771 if (PageCompound(page))
2777 if (page_is_file_cache(page)) {
2778 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2783 * FUSE reuses pages without going through the final
2784 * put that would remove them from the LRU list, make
2785 * sure that they get relinked properly.
2787 __mem_cgroup_commit_charge_lrucare(page, memcg,
2788 MEM_CGROUP_CHARGE_TYPE_CACHE);
2792 if (PageSwapCache(page)) {
2793 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2795 __mem_cgroup_commit_charge_swapin(page, memcg,
2796 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2798 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2799 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2805 * While swap-in, try_charge -> commit or cancel, the page is locked.
2806 * And when try_charge() successfully returns, one refcnt to memcg without
2807 * struct page_cgroup is acquired. This refcnt will be consumed by
2808 * "commit()" or removed by "cancel()"
2810 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2812 gfp_t mask, struct mem_cgroup **ptr)
2814 struct mem_cgroup *memcg;
2819 if (mem_cgroup_disabled())
2822 if (!do_swap_account)
2825 * A racing thread's fault, or swapoff, may have already updated
2826 * the pte, and even removed page from swap cache: in those cases
2827 * do_swap_page()'s pte_same() test will fail; but there's also a
2828 * KSM case which does need to charge the page.
2830 if (!PageSwapCache(page))
2832 memcg = try_get_mem_cgroup_from_page(page);
2836 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2837 css_put(&memcg->css);
2842 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2846 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2847 enum charge_type ctype)
2849 if (mem_cgroup_disabled())
2853 cgroup_exclude_rmdir(&ptr->css);
2855 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2857 * Now swap is on-memory. This means this page may be
2858 * counted both as mem and swap....double count.
2859 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2860 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2861 * may call delete_from_swap_cache() before reach here.
2863 if (do_swap_account && PageSwapCache(page)) {
2864 swp_entry_t ent = {.val = page_private(page)};
2866 struct mem_cgroup *memcg;
2868 id = swap_cgroup_record(ent, 0);
2870 memcg = mem_cgroup_lookup(id);
2873 * This recorded memcg can be obsolete one. So, avoid
2874 * calling css_tryget
2876 if (!mem_cgroup_is_root(memcg))
2877 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2878 mem_cgroup_swap_statistics(memcg, false);
2879 mem_cgroup_put(memcg);
2884 * At swapin, we may charge account against cgroup which has no tasks.
2885 * So, rmdir()->pre_destroy() can be called while we do this charge.
2886 * In that case, we need to call pre_destroy() again. check it here.
2888 cgroup_release_and_wakeup_rmdir(&ptr->css);
2891 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2893 __mem_cgroup_commit_charge_swapin(page, ptr,
2894 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2897 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2899 if (mem_cgroup_disabled())
2903 __mem_cgroup_cancel_charge(memcg, 1);
2906 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2907 unsigned int nr_pages,
2908 const enum charge_type ctype)
2910 struct memcg_batch_info *batch = NULL;
2911 bool uncharge_memsw = true;
2913 /* If swapout, usage of swap doesn't decrease */
2914 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2915 uncharge_memsw = false;
2917 batch = ¤t->memcg_batch;
2919 * In usual, we do css_get() when we remember memcg pointer.
2920 * But in this case, we keep res->usage until end of a series of
2921 * uncharges. Then, it's ok to ignore memcg's refcnt.
2924 batch->memcg = memcg;
2926 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2927 * In those cases, all pages freed continuously can be expected to be in
2928 * the same cgroup and we have chance to coalesce uncharges.
2929 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2930 * because we want to do uncharge as soon as possible.
2933 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2934 goto direct_uncharge;
2937 goto direct_uncharge;
2940 * In typical case, batch->memcg == mem. This means we can
2941 * merge a series of uncharges to an uncharge of res_counter.
2942 * If not, we uncharge res_counter ony by one.
2944 if (batch->memcg != memcg)
2945 goto direct_uncharge;
2946 /* remember freed charge and uncharge it later */
2949 batch->memsw_nr_pages++;
2952 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2954 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2955 if (unlikely(batch->memcg != memcg))
2956 memcg_oom_recover(memcg);
2961 * uncharge if !page_mapped(page)
2963 static struct mem_cgroup *
2964 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2966 struct mem_cgroup *memcg = NULL;
2967 unsigned int nr_pages = 1;
2968 struct page_cgroup *pc;
2970 if (mem_cgroup_disabled())
2973 if (PageSwapCache(page))
2976 if (PageTransHuge(page)) {
2977 nr_pages <<= compound_order(page);
2978 VM_BUG_ON(!PageTransHuge(page));
2981 * Check if our page_cgroup is valid
2983 pc = lookup_page_cgroup(page);
2984 if (unlikely(!pc || !PageCgroupUsed(pc)))
2987 lock_page_cgroup(pc);
2989 memcg = pc->mem_cgroup;
2991 if (!PageCgroupUsed(pc))
2995 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2996 case MEM_CGROUP_CHARGE_TYPE_DROP:
2997 /* See mem_cgroup_prepare_migration() */
2998 if (page_mapped(page) || PageCgroupMigration(pc))
3001 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3002 if (!PageAnon(page)) { /* Shared memory */
3003 if (page->mapping && !page_is_file_cache(page))
3005 } else if (page_mapped(page)) /* Anon */
3012 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3014 ClearPageCgroupUsed(pc);
3016 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3017 * freed from LRU. This is safe because uncharged page is expected not
3018 * to be reused (freed soon). Exception is SwapCache, it's handled by
3019 * special functions.
3022 unlock_page_cgroup(pc);
3024 * even after unlock, we have memcg->res.usage here and this memcg
3025 * will never be freed.
3027 memcg_check_events(memcg, page);
3028 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3029 mem_cgroup_swap_statistics(memcg, true);
3030 mem_cgroup_get(memcg);
3032 if (!mem_cgroup_is_root(memcg))
3033 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3038 unlock_page_cgroup(pc);
3042 void mem_cgroup_uncharge_page(struct page *page)
3045 if (page_mapped(page))
3047 if (page->mapping && !PageAnon(page))
3049 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3052 void mem_cgroup_uncharge_cache_page(struct page *page)
3054 VM_BUG_ON(page_mapped(page));
3055 VM_BUG_ON(page->mapping);
3056 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3060 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3061 * In that cases, pages are freed continuously and we can expect pages
3062 * are in the same memcg. All these calls itself limits the number of
3063 * pages freed at once, then uncharge_start/end() is called properly.
3064 * This may be called prural(2) times in a context,
3067 void mem_cgroup_uncharge_start(void)
3069 current->memcg_batch.do_batch++;
3070 /* We can do nest. */
3071 if (current->memcg_batch.do_batch == 1) {
3072 current->memcg_batch.memcg = NULL;
3073 current->memcg_batch.nr_pages = 0;
3074 current->memcg_batch.memsw_nr_pages = 0;
3078 void mem_cgroup_uncharge_end(void)
3080 struct memcg_batch_info *batch = ¤t->memcg_batch;
3082 if (!batch->do_batch)
3086 if (batch->do_batch) /* If stacked, do nothing. */
3092 * This "batch->memcg" is valid without any css_get/put etc...
3093 * bacause we hide charges behind us.
3095 if (batch->nr_pages)
3096 res_counter_uncharge(&batch->memcg->res,
3097 batch->nr_pages * PAGE_SIZE);
3098 if (batch->memsw_nr_pages)
3099 res_counter_uncharge(&batch->memcg->memsw,
3100 batch->memsw_nr_pages * PAGE_SIZE);
3101 memcg_oom_recover(batch->memcg);
3102 /* forget this pointer (for sanity check) */
3103 batch->memcg = NULL;
3108 * called after __delete_from_swap_cache() and drop "page" account.
3109 * memcg information is recorded to swap_cgroup of "ent"
3112 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3114 struct mem_cgroup *memcg;
3115 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3117 if (!swapout) /* this was a swap cache but the swap is unused ! */
3118 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3120 memcg = __mem_cgroup_uncharge_common(page, ctype);
3123 * record memcg information, if swapout && memcg != NULL,
3124 * mem_cgroup_get() was called in uncharge().
3126 if (do_swap_account && swapout && memcg)
3127 swap_cgroup_record(ent, css_id(&memcg->css));
3131 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3133 * called from swap_entry_free(). remove record in swap_cgroup and
3134 * uncharge "memsw" account.
3136 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3138 struct mem_cgroup *memcg;
3141 if (!do_swap_account)
3144 id = swap_cgroup_record(ent, 0);
3146 memcg = mem_cgroup_lookup(id);
3149 * We uncharge this because swap is freed.
3150 * This memcg can be obsolete one. We avoid calling css_tryget
3152 if (!mem_cgroup_is_root(memcg))
3153 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3154 mem_cgroup_swap_statistics(memcg, false);
3155 mem_cgroup_put(memcg);
3161 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3162 * @entry: swap entry to be moved
3163 * @from: mem_cgroup which the entry is moved from
3164 * @to: mem_cgroup which the entry is moved to
3165 * @need_fixup: whether we should fixup res_counters and refcounts.
3167 * It succeeds only when the swap_cgroup's record for this entry is the same
3168 * as the mem_cgroup's id of @from.
3170 * Returns 0 on success, -EINVAL on failure.
3172 * The caller must have charged to @to, IOW, called res_counter_charge() about
3173 * both res and memsw, and called css_get().
3175 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3176 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3178 unsigned short old_id, new_id;
3180 old_id = css_id(&from->css);
3181 new_id = css_id(&to->css);
3183 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3184 mem_cgroup_swap_statistics(from, false);
3185 mem_cgroup_swap_statistics(to, true);
3187 * This function is only called from task migration context now.
3188 * It postpones res_counter and refcount handling till the end
3189 * of task migration(mem_cgroup_clear_mc()) for performance
3190 * improvement. But we cannot postpone mem_cgroup_get(to)
3191 * because if the process that has been moved to @to does
3192 * swap-in, the refcount of @to might be decreased to 0.
3196 if (!mem_cgroup_is_root(from))
3197 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3198 mem_cgroup_put(from);
3200 * we charged both to->res and to->memsw, so we should
3203 if (!mem_cgroup_is_root(to))
3204 res_counter_uncharge(&to->res, PAGE_SIZE);
3211 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3212 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3219 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3222 int mem_cgroup_prepare_migration(struct page *page,
3223 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3225 struct mem_cgroup *memcg = NULL;
3226 struct page_cgroup *pc;
3227 enum charge_type ctype;
3232 VM_BUG_ON(PageTransHuge(page));
3233 if (mem_cgroup_disabled())
3236 pc = lookup_page_cgroup(page);
3237 lock_page_cgroup(pc);
3238 if (PageCgroupUsed(pc)) {
3239 memcg = pc->mem_cgroup;
3240 css_get(&memcg->css);
3242 * At migrating an anonymous page, its mapcount goes down
3243 * to 0 and uncharge() will be called. But, even if it's fully
3244 * unmapped, migration may fail and this page has to be
3245 * charged again. We set MIGRATION flag here and delay uncharge
3246 * until end_migration() is called
3248 * Corner Case Thinking
3250 * When the old page was mapped as Anon and it's unmap-and-freed
3251 * while migration was ongoing.
3252 * If unmap finds the old page, uncharge() of it will be delayed
3253 * until end_migration(). If unmap finds a new page, it's
3254 * uncharged when it make mapcount to be 1->0. If unmap code
3255 * finds swap_migration_entry, the new page will not be mapped
3256 * and end_migration() will find it(mapcount==0).
3259 * When the old page was mapped but migraion fails, the kernel
3260 * remaps it. A charge for it is kept by MIGRATION flag even
3261 * if mapcount goes down to 0. We can do remap successfully
3262 * without charging it again.
3265 * The "old" page is under lock_page() until the end of
3266 * migration, so, the old page itself will not be swapped-out.
3267 * If the new page is swapped out before end_migraton, our
3268 * hook to usual swap-out path will catch the event.
3271 SetPageCgroupMigration(pc);
3273 unlock_page_cgroup(pc);
3275 * If the page is not charged at this point,
3282 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3283 css_put(&memcg->css);/* drop extra refcnt */
3284 if (ret || *ptr == NULL) {
3285 if (PageAnon(page)) {
3286 lock_page_cgroup(pc);
3287 ClearPageCgroupMigration(pc);
3288 unlock_page_cgroup(pc);
3290 * The old page may be fully unmapped while we kept it.
3292 mem_cgroup_uncharge_page(page);
3297 * We charge new page before it's used/mapped. So, even if unlock_page()
3298 * is called before end_migration, we can catch all events on this new
3299 * page. In the case new page is migrated but not remapped, new page's
3300 * mapcount will be finally 0 and we call uncharge in end_migration().
3302 pc = lookup_page_cgroup(newpage);
3304 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3305 else if (page_is_file_cache(page))
3306 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3308 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3309 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3313 /* remove redundant charge if migration failed*/
3314 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3315 struct page *oldpage, struct page *newpage, bool migration_ok)
3317 struct page *used, *unused;
3318 struct page_cgroup *pc;
3322 /* blocks rmdir() */
3323 cgroup_exclude_rmdir(&memcg->css);
3324 if (!migration_ok) {
3332 * We disallowed uncharge of pages under migration because mapcount
3333 * of the page goes down to zero, temporarly.
3334 * Clear the flag and check the page should be charged.
3336 pc = lookup_page_cgroup(oldpage);
3337 lock_page_cgroup(pc);
3338 ClearPageCgroupMigration(pc);
3339 unlock_page_cgroup(pc);
3341 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3344 * If a page is a file cache, radix-tree replacement is very atomic
3345 * and we can skip this check. When it was an Anon page, its mapcount
3346 * goes down to 0. But because we added MIGRATION flage, it's not
3347 * uncharged yet. There are several case but page->mapcount check
3348 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3349 * check. (see prepare_charge() also)
3352 mem_cgroup_uncharge_page(used);
3354 * At migration, we may charge account against cgroup which has no
3356 * So, rmdir()->pre_destroy() can be called while we do this charge.
3357 * In that case, we need to call pre_destroy() again. check it here.
3359 cgroup_release_and_wakeup_rmdir(&memcg->css);
3362 #ifdef CONFIG_DEBUG_VM
3363 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3365 struct page_cgroup *pc;
3367 pc = lookup_page_cgroup(page);
3368 if (likely(pc) && PageCgroupUsed(pc))
3373 bool mem_cgroup_bad_page_check(struct page *page)
3375 if (mem_cgroup_disabled())
3378 return lookup_page_cgroup_used(page) != NULL;
3381 void mem_cgroup_print_bad_page(struct page *page)
3383 struct page_cgroup *pc;
3385 pc = lookup_page_cgroup_used(page);
3390 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3391 pc, pc->flags, pc->mem_cgroup);
3393 path = kmalloc(PATH_MAX, GFP_KERNEL);
3396 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3401 printk(KERN_CONT "(%s)\n",
3402 (ret < 0) ? "cannot get the path" : path);
3408 static DEFINE_MUTEX(set_limit_mutex);
3410 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3411 unsigned long long val)
3414 u64 memswlimit, memlimit;
3416 int children = mem_cgroup_count_children(memcg);
3417 u64 curusage, oldusage;
3421 * For keeping hierarchical_reclaim simple, how long we should retry
3422 * is depends on callers. We set our retry-count to be function
3423 * of # of children which we should visit in this loop.
3425 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3427 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3430 while (retry_count) {
3431 if (signal_pending(current)) {
3436 * Rather than hide all in some function, I do this in
3437 * open coded manner. You see what this really does.
3438 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3440 mutex_lock(&set_limit_mutex);
3441 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3442 if (memswlimit < val) {
3444 mutex_unlock(&set_limit_mutex);
3448 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3452 ret = res_counter_set_limit(&memcg->res, val);
3454 if (memswlimit == val)
3455 memcg->memsw_is_minimum = true;
3457 memcg->memsw_is_minimum = false;
3459 mutex_unlock(&set_limit_mutex);
3464 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3465 MEM_CGROUP_RECLAIM_SHRINK,
3467 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3468 /* Usage is reduced ? */
3469 if (curusage >= oldusage)
3472 oldusage = curusage;
3474 if (!ret && enlarge)
3475 memcg_oom_recover(memcg);
3480 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3481 unsigned long long val)
3484 u64 memlimit, memswlimit, oldusage, curusage;
3485 int children = mem_cgroup_count_children(memcg);
3489 /* see mem_cgroup_resize_res_limit */
3490 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3491 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3492 while (retry_count) {
3493 if (signal_pending(current)) {
3498 * Rather than hide all in some function, I do this in
3499 * open coded manner. You see what this really does.
3500 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3502 mutex_lock(&set_limit_mutex);
3503 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3504 if (memlimit > val) {
3506 mutex_unlock(&set_limit_mutex);
3509 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3510 if (memswlimit < val)
3512 ret = res_counter_set_limit(&memcg->memsw, val);
3514 if (memlimit == val)
3515 memcg->memsw_is_minimum = true;
3517 memcg->memsw_is_minimum = false;
3519 mutex_unlock(&set_limit_mutex);
3524 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3525 MEM_CGROUP_RECLAIM_NOSWAP |
3526 MEM_CGROUP_RECLAIM_SHRINK,
3528 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3529 /* Usage is reduced ? */
3530 if (curusage >= oldusage)
3533 oldusage = curusage;
3535 if (!ret && enlarge)
3536 memcg_oom_recover(memcg);
3540 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3542 unsigned long *total_scanned)
3544 unsigned long nr_reclaimed = 0;
3545 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3546 unsigned long reclaimed;
3548 struct mem_cgroup_tree_per_zone *mctz;
3549 unsigned long long excess;
3550 unsigned long nr_scanned;
3555 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3557 * This loop can run a while, specially if mem_cgroup's continuously
3558 * keep exceeding their soft limit and putting the system under
3565 mz = mem_cgroup_largest_soft_limit_node(mctz);
3570 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3572 MEM_CGROUP_RECLAIM_SOFT,
3574 nr_reclaimed += reclaimed;
3575 *total_scanned += nr_scanned;
3576 spin_lock(&mctz->lock);
3579 * If we failed to reclaim anything from this memory cgroup
3580 * it is time to move on to the next cgroup
3586 * Loop until we find yet another one.
3588 * By the time we get the soft_limit lock
3589 * again, someone might have aded the
3590 * group back on the RB tree. Iterate to
3591 * make sure we get a different mem.
3592 * mem_cgroup_largest_soft_limit_node returns
3593 * NULL if no other cgroup is present on
3597 __mem_cgroup_largest_soft_limit_node(mctz);
3599 css_put(&next_mz->mem->css);
3600 else /* next_mz == NULL or other memcg */
3604 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3605 excess = res_counter_soft_limit_excess(&mz->mem->res);
3607 * One school of thought says that we should not add
3608 * back the node to the tree if reclaim returns 0.
3609 * But our reclaim could return 0, simply because due
3610 * to priority we are exposing a smaller subset of
3611 * memory to reclaim from. Consider this as a longer
3614 /* If excess == 0, no tree ops */
3615 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3616 spin_unlock(&mctz->lock);
3617 css_put(&mz->mem->css);
3620 * Could not reclaim anything and there are no more
3621 * mem cgroups to try or we seem to be looping without
3622 * reclaiming anything.
3624 if (!nr_reclaimed &&
3626 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3628 } while (!nr_reclaimed);
3630 css_put(&next_mz->mem->css);
3631 return nr_reclaimed;
3635 * This routine traverse page_cgroup in given list and drop them all.
3636 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3638 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3639 int node, int zid, enum lru_list lru)
3642 struct mem_cgroup_per_zone *mz;
3643 struct page_cgroup *pc, *busy;
3644 unsigned long flags, loop;
3645 struct list_head *list;
3648 zone = &NODE_DATA(node)->node_zones[zid];
3649 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3650 list = &mz->lists[lru];
3652 loop = MEM_CGROUP_ZSTAT(mz, lru);
3653 /* give some margin against EBUSY etc...*/
3660 spin_lock_irqsave(&zone->lru_lock, flags);
3661 if (list_empty(list)) {
3662 spin_unlock_irqrestore(&zone->lru_lock, flags);
3665 pc = list_entry(list->prev, struct page_cgroup, lru);
3667 list_move(&pc->lru, list);
3669 spin_unlock_irqrestore(&zone->lru_lock, flags);
3672 spin_unlock_irqrestore(&zone->lru_lock, flags);
3674 page = lookup_cgroup_page(pc);
3676 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3680 if (ret == -EBUSY || ret == -EINVAL) {
3681 /* found lock contention or "pc" is obsolete. */
3688 if (!ret && !list_empty(list))
3694 * make mem_cgroup's charge to be 0 if there is no task.
3695 * This enables deleting this mem_cgroup.
3697 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3700 int node, zid, shrink;
3701 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3702 struct cgroup *cgrp = memcg->css.cgroup;
3704 css_get(&memcg->css);
3707 /* should free all ? */
3713 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3716 if (signal_pending(current))
3718 /* This is for making all *used* pages to be on LRU. */
3719 lru_add_drain_all();
3720 drain_all_stock_sync(memcg);
3722 mem_cgroup_start_move(memcg);
3723 for_each_node_state(node, N_HIGH_MEMORY) {
3724 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3727 ret = mem_cgroup_force_empty_list(memcg,
3736 mem_cgroup_end_move(memcg);
3737 memcg_oom_recover(memcg);
3738 /* it seems parent cgroup doesn't have enough mem */
3742 /* "ret" should also be checked to ensure all lists are empty. */
3743 } while (memcg->res.usage > 0 || ret);
3745 css_put(&memcg->css);
3749 /* returns EBUSY if there is a task or if we come here twice. */
3750 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3754 /* we call try-to-free pages for make this cgroup empty */
3755 lru_add_drain_all();
3756 /* try to free all pages in this cgroup */
3758 while (nr_retries && memcg->res.usage > 0) {
3761 if (signal_pending(current)) {
3765 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3769 /* maybe some writeback is necessary */
3770 congestion_wait(BLK_RW_ASYNC, HZ/10);
3775 /* try move_account...there may be some *locked* pages. */
3779 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3781 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3785 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3787 return mem_cgroup_from_cont(cont)->use_hierarchy;
3790 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3794 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3795 struct cgroup *parent = cont->parent;
3796 struct mem_cgroup *parent_memcg = NULL;
3799 parent_memcg = mem_cgroup_from_cont(parent);
3803 * If parent's use_hierarchy is set, we can't make any modifications
3804 * in the child subtrees. If it is unset, then the change can
3805 * occur, provided the current cgroup has no children.
3807 * For the root cgroup, parent_mem is NULL, we allow value to be
3808 * set if there are no children.
3810 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3811 (val == 1 || val == 0)) {
3812 if (list_empty(&cont->children))
3813 memcg->use_hierarchy = val;
3824 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3825 enum mem_cgroup_stat_index idx)
3827 struct mem_cgroup *iter;
3830 /* Per-cpu values can be negative, use a signed accumulator */
3831 for_each_mem_cgroup_tree(iter, memcg)
3832 val += mem_cgroup_read_stat(iter, idx);
3834 if (val < 0) /* race ? */
3839 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3843 if (!mem_cgroup_is_root(memcg)) {
3845 return res_counter_read_u64(&memcg->res, RES_USAGE);
3847 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3850 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3851 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3854 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3856 return val << PAGE_SHIFT;
3859 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3861 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3865 type = MEMFILE_TYPE(cft->private);
3866 name = MEMFILE_ATTR(cft->private);
3869 if (name == RES_USAGE)
3870 val = mem_cgroup_usage(memcg, false);
3872 val = res_counter_read_u64(&memcg->res, name);
3875 if (name == RES_USAGE)
3876 val = mem_cgroup_usage(memcg, true);
3878 val = res_counter_read_u64(&memcg->memsw, name);
3887 * The user of this function is...
3890 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3893 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3895 unsigned long long val;
3898 type = MEMFILE_TYPE(cft->private);
3899 name = MEMFILE_ATTR(cft->private);
3902 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3906 /* This function does all necessary parse...reuse it */
3907 ret = res_counter_memparse_write_strategy(buffer, &val);
3911 ret = mem_cgroup_resize_limit(memcg, val);
3913 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3915 case RES_SOFT_LIMIT:
3916 ret = res_counter_memparse_write_strategy(buffer, &val);
3920 * For memsw, soft limits are hard to implement in terms
3921 * of semantics, for now, we support soft limits for
3922 * control without swap
3925 ret = res_counter_set_soft_limit(&memcg->res, val);
3930 ret = -EINVAL; /* should be BUG() ? */
3936 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3937 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3939 struct cgroup *cgroup;
3940 unsigned long long min_limit, min_memsw_limit, tmp;
3942 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3943 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3944 cgroup = memcg->css.cgroup;
3945 if (!memcg->use_hierarchy)
3948 while (cgroup->parent) {
3949 cgroup = cgroup->parent;
3950 memcg = mem_cgroup_from_cont(cgroup);
3951 if (!memcg->use_hierarchy)
3953 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3954 min_limit = min(min_limit, tmp);
3955 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3956 min_memsw_limit = min(min_memsw_limit, tmp);
3959 *mem_limit = min_limit;
3960 *memsw_limit = min_memsw_limit;
3964 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3966 struct mem_cgroup *memcg;
3969 memcg = mem_cgroup_from_cont(cont);
3970 type = MEMFILE_TYPE(event);
3971 name = MEMFILE_ATTR(event);
3975 res_counter_reset_max(&memcg->res);
3977 res_counter_reset_max(&memcg->memsw);
3981 res_counter_reset_failcnt(&memcg->res);
3983 res_counter_reset_failcnt(&memcg->memsw);
3990 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3993 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3997 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3998 struct cftype *cft, u64 val)
4000 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4002 if (val >= (1 << NR_MOVE_TYPE))
4005 * We check this value several times in both in can_attach() and
4006 * attach(), so we need cgroup lock to prevent this value from being
4010 memcg->move_charge_at_immigrate = val;
4016 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4017 struct cftype *cft, u64 val)
4024 /* For read statistics */
4042 struct mcs_total_stat {
4043 s64 stat[NR_MCS_STAT];
4049 } memcg_stat_strings[NR_MCS_STAT] = {
4050 {"cache", "total_cache"},
4051 {"rss", "total_rss"},
4052 {"mapped_file", "total_mapped_file"},
4053 {"pgpgin", "total_pgpgin"},
4054 {"pgpgout", "total_pgpgout"},
4055 {"swap", "total_swap"},
4056 {"pgfault", "total_pgfault"},
4057 {"pgmajfault", "total_pgmajfault"},
4058 {"inactive_anon", "total_inactive_anon"},
4059 {"active_anon", "total_active_anon"},
4060 {"inactive_file", "total_inactive_file"},
4061 {"active_file", "total_active_file"},
4062 {"unevictable", "total_unevictable"}
4067 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4072 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4073 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4074 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4075 s->stat[MCS_RSS] += val * PAGE_SIZE;
4076 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4077 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4078 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4079 s->stat[MCS_PGPGIN] += val;
4080 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4081 s->stat[MCS_PGPGOUT] += val;
4082 if (do_swap_account) {
4083 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4084 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4086 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4087 s->stat[MCS_PGFAULT] += val;
4088 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4089 s->stat[MCS_PGMAJFAULT] += val;
4092 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4093 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4094 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4095 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4096 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4097 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4098 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4099 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4100 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4101 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4105 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4107 struct mem_cgroup *iter;
4109 for_each_mem_cgroup_tree(iter, memcg)
4110 mem_cgroup_get_local_stat(iter, s);
4114 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4117 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4118 unsigned long node_nr;
4119 struct cgroup *cont = m->private;
4120 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4122 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4123 seq_printf(m, "total=%lu", total_nr);
4124 for_each_node_state(nid, N_HIGH_MEMORY) {
4125 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4126 seq_printf(m, " N%d=%lu", nid, node_nr);
4130 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4131 seq_printf(m, "file=%lu", file_nr);
4132 for_each_node_state(nid, N_HIGH_MEMORY) {
4133 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4135 seq_printf(m, " N%d=%lu", nid, node_nr);
4139 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4140 seq_printf(m, "anon=%lu", anon_nr);
4141 for_each_node_state(nid, N_HIGH_MEMORY) {
4142 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4144 seq_printf(m, " N%d=%lu", nid, node_nr);
4148 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4149 seq_printf(m, "unevictable=%lu", unevictable_nr);
4150 for_each_node_state(nid, N_HIGH_MEMORY) {
4151 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4152 BIT(LRU_UNEVICTABLE));
4153 seq_printf(m, " N%d=%lu", nid, node_nr);
4158 #endif /* CONFIG_NUMA */
4160 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4161 struct cgroup_map_cb *cb)
4163 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4164 struct mcs_total_stat mystat;
4167 memset(&mystat, 0, sizeof(mystat));
4168 mem_cgroup_get_local_stat(mem_cont, &mystat);
4171 for (i = 0; i < NR_MCS_STAT; i++) {
4172 if (i == MCS_SWAP && !do_swap_account)
4174 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4177 /* Hierarchical information */
4179 unsigned long long limit, memsw_limit;
4180 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4181 cb->fill(cb, "hierarchical_memory_limit", limit);
4182 if (do_swap_account)
4183 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4186 memset(&mystat, 0, sizeof(mystat));
4187 mem_cgroup_get_total_stat(mem_cont, &mystat);
4188 for (i = 0; i < NR_MCS_STAT; i++) {
4189 if (i == MCS_SWAP && !do_swap_account)
4191 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4194 #ifdef CONFIG_DEBUG_VM
4195 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4199 struct mem_cgroup_per_zone *mz;
4200 unsigned long recent_rotated[2] = {0, 0};
4201 unsigned long recent_scanned[2] = {0, 0};
4203 for_each_online_node(nid)
4204 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4205 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4207 recent_rotated[0] +=
4208 mz->reclaim_stat.recent_rotated[0];
4209 recent_rotated[1] +=
4210 mz->reclaim_stat.recent_rotated[1];
4211 recent_scanned[0] +=
4212 mz->reclaim_stat.recent_scanned[0];
4213 recent_scanned[1] +=
4214 mz->reclaim_stat.recent_scanned[1];
4216 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4217 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4218 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4219 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4226 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4228 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4230 return mem_cgroup_swappiness(memcg);
4233 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4236 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4237 struct mem_cgroup *parent;
4242 if (cgrp->parent == NULL)
4245 parent = mem_cgroup_from_cont(cgrp->parent);
4249 /* If under hierarchy, only empty-root can set this value */
4250 if ((parent->use_hierarchy) ||
4251 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4256 memcg->swappiness = val;
4263 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4265 struct mem_cgroup_threshold_ary *t;
4271 t = rcu_dereference(memcg->thresholds.primary);
4273 t = rcu_dereference(memcg->memsw_thresholds.primary);
4278 usage = mem_cgroup_usage(memcg, swap);
4281 * current_threshold points to threshold just below usage.
4282 * If it's not true, a threshold was crossed after last
4283 * call of __mem_cgroup_threshold().
4285 i = t->current_threshold;
4288 * Iterate backward over array of thresholds starting from
4289 * current_threshold and check if a threshold is crossed.
4290 * If none of thresholds below usage is crossed, we read
4291 * only one element of the array here.
4293 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4294 eventfd_signal(t->entries[i].eventfd, 1);
4296 /* i = current_threshold + 1 */
4300 * Iterate forward over array of thresholds starting from
4301 * current_threshold+1 and check if a threshold is crossed.
4302 * If none of thresholds above usage is crossed, we read
4303 * only one element of the array here.
4305 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4306 eventfd_signal(t->entries[i].eventfd, 1);
4308 /* Update current_threshold */
4309 t->current_threshold = i - 1;
4314 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4317 __mem_cgroup_threshold(memcg, false);
4318 if (do_swap_account)
4319 __mem_cgroup_threshold(memcg, true);
4321 memcg = parent_mem_cgroup(memcg);
4325 static int compare_thresholds(const void *a, const void *b)
4327 const struct mem_cgroup_threshold *_a = a;
4328 const struct mem_cgroup_threshold *_b = b;
4330 return _a->threshold - _b->threshold;
4333 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4335 struct mem_cgroup_eventfd_list *ev;
4337 list_for_each_entry(ev, &memcg->oom_notify, list)
4338 eventfd_signal(ev->eventfd, 1);
4342 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4344 struct mem_cgroup *iter;
4346 for_each_mem_cgroup_tree(iter, memcg)
4347 mem_cgroup_oom_notify_cb(iter);
4350 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4351 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4353 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4354 struct mem_cgroup_thresholds *thresholds;
4355 struct mem_cgroup_threshold_ary *new;
4356 int type = MEMFILE_TYPE(cft->private);
4357 u64 threshold, usage;
4360 ret = res_counter_memparse_write_strategy(args, &threshold);
4364 mutex_lock(&memcg->thresholds_lock);
4367 thresholds = &memcg->thresholds;
4368 else if (type == _MEMSWAP)
4369 thresholds = &memcg->memsw_thresholds;
4373 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4375 /* Check if a threshold crossed before adding a new one */
4376 if (thresholds->primary)
4377 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4379 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4381 /* Allocate memory for new array of thresholds */
4382 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4390 /* Copy thresholds (if any) to new array */
4391 if (thresholds->primary) {
4392 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4393 sizeof(struct mem_cgroup_threshold));
4396 /* Add new threshold */
4397 new->entries[size - 1].eventfd = eventfd;
4398 new->entries[size - 1].threshold = threshold;
4400 /* Sort thresholds. Registering of new threshold isn't time-critical */
4401 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4402 compare_thresholds, NULL);
4404 /* Find current threshold */
4405 new->current_threshold = -1;
4406 for (i = 0; i < size; i++) {
4407 if (new->entries[i].threshold < usage) {
4409 * new->current_threshold will not be used until
4410 * rcu_assign_pointer(), so it's safe to increment
4413 ++new->current_threshold;
4417 /* Free old spare buffer and save old primary buffer as spare */
4418 kfree(thresholds->spare);
4419 thresholds->spare = thresholds->primary;
4421 rcu_assign_pointer(thresholds->primary, new);
4423 /* To be sure that nobody uses thresholds */
4427 mutex_unlock(&memcg->thresholds_lock);
4432 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4433 struct cftype *cft, struct eventfd_ctx *eventfd)
4435 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4436 struct mem_cgroup_thresholds *thresholds;
4437 struct mem_cgroup_threshold_ary *new;
4438 int type = MEMFILE_TYPE(cft->private);
4442 mutex_lock(&memcg->thresholds_lock);
4444 thresholds = &memcg->thresholds;
4445 else if (type == _MEMSWAP)
4446 thresholds = &memcg->memsw_thresholds;
4451 * Something went wrong if we trying to unregister a threshold
4452 * if we don't have thresholds
4454 BUG_ON(!thresholds);
4456 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4458 /* Check if a threshold crossed before removing */
4459 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4461 /* Calculate new number of threshold */
4463 for (i = 0; i < thresholds->primary->size; i++) {
4464 if (thresholds->primary->entries[i].eventfd != eventfd)
4468 new = thresholds->spare;
4470 /* Set thresholds array to NULL if we don't have thresholds */
4479 /* Copy thresholds and find current threshold */
4480 new->current_threshold = -1;
4481 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4482 if (thresholds->primary->entries[i].eventfd == eventfd)
4485 new->entries[j] = thresholds->primary->entries[i];
4486 if (new->entries[j].threshold < usage) {
4488 * new->current_threshold will not be used
4489 * until rcu_assign_pointer(), so it's safe to increment
4492 ++new->current_threshold;
4498 /* Swap primary and spare array */
4499 thresholds->spare = thresholds->primary;
4500 rcu_assign_pointer(thresholds->primary, new);
4502 /* To be sure that nobody uses thresholds */
4505 mutex_unlock(&memcg->thresholds_lock);
4508 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4509 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4511 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4512 struct mem_cgroup_eventfd_list *event;
4513 int type = MEMFILE_TYPE(cft->private);
4515 BUG_ON(type != _OOM_TYPE);
4516 event = kmalloc(sizeof(*event), GFP_KERNEL);
4520 spin_lock(&memcg_oom_lock);
4522 event->eventfd = eventfd;
4523 list_add(&event->list, &memcg->oom_notify);
4525 /* already in OOM ? */
4526 if (atomic_read(&memcg->under_oom))
4527 eventfd_signal(eventfd, 1);
4528 spin_unlock(&memcg_oom_lock);
4533 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4534 struct cftype *cft, struct eventfd_ctx *eventfd)
4536 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4537 struct mem_cgroup_eventfd_list *ev, *tmp;
4538 int type = MEMFILE_TYPE(cft->private);
4540 BUG_ON(type != _OOM_TYPE);
4542 spin_lock(&memcg_oom_lock);
4544 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4545 if (ev->eventfd == eventfd) {
4546 list_del(&ev->list);
4551 spin_unlock(&memcg_oom_lock);
4554 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4555 struct cftype *cft, struct cgroup_map_cb *cb)
4557 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4559 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4561 if (atomic_read(&memcg->under_oom))
4562 cb->fill(cb, "under_oom", 1);
4564 cb->fill(cb, "under_oom", 0);
4568 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4569 struct cftype *cft, u64 val)
4571 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4572 struct mem_cgroup *parent;
4574 /* cannot set to root cgroup and only 0 and 1 are allowed */
4575 if (!cgrp->parent || !((val == 0) || (val == 1)))
4578 parent = mem_cgroup_from_cont(cgrp->parent);
4581 /* oom-kill-disable is a flag for subhierarchy. */
4582 if ((parent->use_hierarchy) ||
4583 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4587 memcg->oom_kill_disable = val;
4589 memcg_oom_recover(memcg);
4595 static const struct file_operations mem_control_numa_stat_file_operations = {
4597 .llseek = seq_lseek,
4598 .release = single_release,
4601 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4603 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4605 file->f_op = &mem_control_numa_stat_file_operations;
4606 return single_open(file, mem_control_numa_stat_show, cont);
4608 #endif /* CONFIG_NUMA */
4610 static struct cftype mem_cgroup_files[] = {
4612 .name = "usage_in_bytes",
4613 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4614 .read_u64 = mem_cgroup_read,
4615 .register_event = mem_cgroup_usage_register_event,
4616 .unregister_event = mem_cgroup_usage_unregister_event,
4619 .name = "max_usage_in_bytes",
4620 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4621 .trigger = mem_cgroup_reset,
4622 .read_u64 = mem_cgroup_read,
4625 .name = "limit_in_bytes",
4626 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4627 .write_string = mem_cgroup_write,
4628 .read_u64 = mem_cgroup_read,
4631 .name = "soft_limit_in_bytes",
4632 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4633 .write_string = mem_cgroup_write,
4634 .read_u64 = mem_cgroup_read,
4638 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4639 .trigger = mem_cgroup_reset,
4640 .read_u64 = mem_cgroup_read,
4644 .read_map = mem_control_stat_show,
4647 .name = "force_empty",
4648 .trigger = mem_cgroup_force_empty_write,
4651 .name = "use_hierarchy",
4652 .write_u64 = mem_cgroup_hierarchy_write,
4653 .read_u64 = mem_cgroup_hierarchy_read,
4656 .name = "swappiness",
4657 .read_u64 = mem_cgroup_swappiness_read,
4658 .write_u64 = mem_cgroup_swappiness_write,
4661 .name = "move_charge_at_immigrate",
4662 .read_u64 = mem_cgroup_move_charge_read,
4663 .write_u64 = mem_cgroup_move_charge_write,
4666 .name = "oom_control",
4667 .read_map = mem_cgroup_oom_control_read,
4668 .write_u64 = mem_cgroup_oom_control_write,
4669 .register_event = mem_cgroup_oom_register_event,
4670 .unregister_event = mem_cgroup_oom_unregister_event,
4671 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4675 .name = "numa_stat",
4676 .open = mem_control_numa_stat_open,
4682 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4683 static struct cftype memsw_cgroup_files[] = {
4685 .name = "memsw.usage_in_bytes",
4686 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4687 .read_u64 = mem_cgroup_read,
4688 .register_event = mem_cgroup_usage_register_event,
4689 .unregister_event = mem_cgroup_usage_unregister_event,
4692 .name = "memsw.max_usage_in_bytes",
4693 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4694 .trigger = mem_cgroup_reset,
4695 .read_u64 = mem_cgroup_read,
4698 .name = "memsw.limit_in_bytes",
4699 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4700 .write_string = mem_cgroup_write,
4701 .read_u64 = mem_cgroup_read,
4704 .name = "memsw.failcnt",
4705 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4706 .trigger = mem_cgroup_reset,
4707 .read_u64 = mem_cgroup_read,
4711 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4713 if (!do_swap_account)
4715 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4716 ARRAY_SIZE(memsw_cgroup_files));
4719 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4725 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4727 struct mem_cgroup_per_node *pn;
4728 struct mem_cgroup_per_zone *mz;
4730 int zone, tmp = node;
4732 * This routine is called against possible nodes.
4733 * But it's BUG to call kmalloc() against offline node.
4735 * TODO: this routine can waste much memory for nodes which will
4736 * never be onlined. It's better to use memory hotplug callback
4739 if (!node_state(node, N_NORMAL_MEMORY))
4741 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4745 memcg->info.nodeinfo[node] = pn;
4746 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4747 mz = &pn->zoneinfo[zone];
4749 INIT_LIST_HEAD(&mz->lists[l]);
4750 mz->usage_in_excess = 0;
4751 mz->on_tree = false;
4757 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4759 kfree(memcg->info.nodeinfo[node]);
4762 static struct mem_cgroup *mem_cgroup_alloc(void)
4764 struct mem_cgroup *mem;
4765 int size = sizeof(struct mem_cgroup);
4767 /* Can be very big if MAX_NUMNODES is very big */
4768 if (size < PAGE_SIZE)
4769 mem = kzalloc(size, GFP_KERNEL);
4771 mem = vzalloc(size);
4776 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4779 spin_lock_init(&mem->pcp_counter_lock);
4783 if (size < PAGE_SIZE)
4791 * At destroying mem_cgroup, references from swap_cgroup can remain.
4792 * (scanning all at force_empty is too costly...)
4794 * Instead of clearing all references at force_empty, we remember
4795 * the number of reference from swap_cgroup and free mem_cgroup when
4796 * it goes down to 0.
4798 * Removal of cgroup itself succeeds regardless of refs from swap.
4801 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4805 mem_cgroup_remove_from_trees(memcg);
4806 free_css_id(&mem_cgroup_subsys, &memcg->css);
4808 for_each_node_state(node, N_POSSIBLE)
4809 free_mem_cgroup_per_zone_info(memcg, node);
4811 free_percpu(memcg->stat);
4812 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4818 static void mem_cgroup_get(struct mem_cgroup *memcg)
4820 atomic_inc(&memcg->refcnt);
4823 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4825 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4826 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4827 __mem_cgroup_free(memcg);
4829 mem_cgroup_put(parent);
4833 static void mem_cgroup_put(struct mem_cgroup *memcg)
4835 __mem_cgroup_put(memcg, 1);
4839 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4841 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4843 if (!memcg->res.parent)
4845 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4848 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4849 static void __init enable_swap_cgroup(void)
4851 if (!mem_cgroup_disabled() && really_do_swap_account)
4852 do_swap_account = 1;
4855 static void __init enable_swap_cgroup(void)
4860 static int mem_cgroup_soft_limit_tree_init(void)
4862 struct mem_cgroup_tree_per_node *rtpn;
4863 struct mem_cgroup_tree_per_zone *rtpz;
4864 int tmp, node, zone;
4866 for_each_node_state(node, N_POSSIBLE) {
4868 if (!node_state(node, N_NORMAL_MEMORY))
4870 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4874 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4876 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4877 rtpz = &rtpn->rb_tree_per_zone[zone];
4878 rtpz->rb_root = RB_ROOT;
4879 spin_lock_init(&rtpz->lock);
4885 static struct cgroup_subsys_state * __ref
4886 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4888 struct mem_cgroup *memcg, *parent;
4889 long error = -ENOMEM;
4892 memcg = mem_cgroup_alloc();
4894 return ERR_PTR(error);
4896 for_each_node_state(node, N_POSSIBLE)
4897 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4901 if (cont->parent == NULL) {
4903 enable_swap_cgroup();
4905 root_mem_cgroup = memcg;
4906 if (mem_cgroup_soft_limit_tree_init())
4908 for_each_possible_cpu(cpu) {
4909 struct memcg_stock_pcp *stock =
4910 &per_cpu(memcg_stock, cpu);
4911 INIT_WORK(&stock->work, drain_local_stock);
4913 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4915 parent = mem_cgroup_from_cont(cont->parent);
4916 memcg->use_hierarchy = parent->use_hierarchy;
4917 memcg->oom_kill_disable = parent->oom_kill_disable;
4920 if (parent && parent->use_hierarchy) {
4921 res_counter_init(&memcg->res, &parent->res);
4922 res_counter_init(&memcg->memsw, &parent->memsw);
4924 * We increment refcnt of the parent to ensure that we can
4925 * safely access it on res_counter_charge/uncharge.
4926 * This refcnt will be decremented when freeing this
4927 * mem_cgroup(see mem_cgroup_put).
4929 mem_cgroup_get(parent);
4931 res_counter_init(&memcg->res, NULL);
4932 res_counter_init(&memcg->memsw, NULL);
4934 memcg->last_scanned_child = 0;
4935 memcg->last_scanned_node = MAX_NUMNODES;
4936 INIT_LIST_HEAD(&memcg->oom_notify);
4939 memcg->swappiness = mem_cgroup_swappiness(parent);
4940 atomic_set(&memcg->refcnt, 1);
4941 memcg->move_charge_at_immigrate = 0;
4942 mutex_init(&memcg->thresholds_lock);
4945 __mem_cgroup_free(memcg);
4946 root_mem_cgroup = NULL;
4947 return ERR_PTR(error);
4950 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4951 struct cgroup *cont)
4953 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4955 return mem_cgroup_force_empty(memcg, false);
4958 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4959 struct cgroup *cont)
4961 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4963 mem_cgroup_put(memcg);
4966 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4967 struct cgroup *cont)
4971 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4972 ARRAY_SIZE(mem_cgroup_files));
4975 ret = register_memsw_files(cont, ss);
4980 /* Handlers for move charge at task migration. */
4981 #define PRECHARGE_COUNT_AT_ONCE 256
4982 static int mem_cgroup_do_precharge(unsigned long count)
4985 int batch_count = PRECHARGE_COUNT_AT_ONCE;
4986 struct mem_cgroup *memcg = mc.to;
4988 if (mem_cgroup_is_root(memcg)) {
4989 mc.precharge += count;
4990 /* we don't need css_get for root */
4993 /* try to charge at once */
4995 struct res_counter *dummy;
4997 * "memcg" cannot be under rmdir() because we've already checked
4998 * by cgroup_lock_live_cgroup() that it is not removed and we
4999 * are still under the same cgroup_mutex. So we can postpone
5002 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5004 if (do_swap_account && res_counter_charge(&memcg->memsw,
5005 PAGE_SIZE * count, &dummy)) {
5006 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5009 mc.precharge += count;
5013 /* fall back to one by one charge */
5015 if (signal_pending(current)) {
5019 if (!batch_count--) {
5020 batch_count = PRECHARGE_COUNT_AT_ONCE;
5023 ret = __mem_cgroup_try_charge(NULL,
5024 GFP_KERNEL, 1, &memcg, false);
5026 /* mem_cgroup_clear_mc() will do uncharge later */
5034 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5035 * @vma: the vma the pte to be checked belongs
5036 * @addr: the address corresponding to the pte to be checked
5037 * @ptent: the pte to be checked
5038 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5041 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5042 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5043 * move charge. if @target is not NULL, the page is stored in target->page
5044 * with extra refcnt got(Callers should handle it).
5045 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5046 * target for charge migration. if @target is not NULL, the entry is stored
5049 * Called with pte lock held.
5056 enum mc_target_type {
5057 MC_TARGET_NONE, /* not used */
5062 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5063 unsigned long addr, pte_t ptent)
5065 struct page *page = vm_normal_page(vma, addr, ptent);
5067 if (!page || !page_mapped(page))
5069 if (PageAnon(page)) {
5070 /* we don't move shared anon */
5071 if (!move_anon() || page_mapcount(page) > 2)
5073 } else if (!move_file())
5074 /* we ignore mapcount for file pages */
5076 if (!get_page_unless_zero(page))
5082 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5083 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5086 struct page *page = NULL;
5087 swp_entry_t ent = pte_to_swp_entry(ptent);
5089 if (!move_anon() || non_swap_entry(ent))
5091 usage_count = mem_cgroup_count_swap_user(ent, &page);
5092 if (usage_count > 1) { /* we don't move shared anon */
5097 if (do_swap_account)
5098 entry->val = ent.val;
5103 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5104 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5106 struct page *page = NULL;
5107 struct inode *inode;
5108 struct address_space *mapping;
5111 if (!vma->vm_file) /* anonymous vma */
5116 inode = vma->vm_file->f_path.dentry->d_inode;
5117 mapping = vma->vm_file->f_mapping;
5118 if (pte_none(ptent))
5119 pgoff = linear_page_index(vma, addr);
5120 else /* pte_file(ptent) is true */
5121 pgoff = pte_to_pgoff(ptent);
5123 /* page is moved even if it's not RSS of this task(page-faulted). */
5124 page = find_get_page(mapping, pgoff);
5127 /* shmem/tmpfs may report page out on swap: account for that too. */
5128 if (radix_tree_exceptional_entry(page)) {
5129 swp_entry_t swap = radix_to_swp_entry(page);
5130 if (do_swap_account)
5132 page = find_get_page(&swapper_space, swap.val);
5138 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5139 unsigned long addr, pte_t ptent, union mc_target *target)
5141 struct page *page = NULL;
5142 struct page_cgroup *pc;
5144 swp_entry_t ent = { .val = 0 };
5146 if (pte_present(ptent))
5147 page = mc_handle_present_pte(vma, addr, ptent);
5148 else if (is_swap_pte(ptent))
5149 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5150 else if (pte_none(ptent) || pte_file(ptent))
5151 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5153 if (!page && !ent.val)
5156 pc = lookup_page_cgroup(page);
5158 * Do only loose check w/o page_cgroup lock.
5159 * mem_cgroup_move_account() checks the pc is valid or not under
5162 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5163 ret = MC_TARGET_PAGE;
5165 target->page = page;
5167 if (!ret || !target)
5170 /* There is a swap entry and a page doesn't exist or isn't charged */
5171 if (ent.val && !ret &&
5172 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5173 ret = MC_TARGET_SWAP;
5180 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5181 unsigned long addr, unsigned long end,
5182 struct mm_walk *walk)
5184 struct vm_area_struct *vma = walk->private;
5188 split_huge_page_pmd(walk->mm, pmd);
5190 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5191 for (; addr != end; pte++, addr += PAGE_SIZE)
5192 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5193 mc.precharge++; /* increment precharge temporarily */
5194 pte_unmap_unlock(pte - 1, ptl);
5200 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5202 unsigned long precharge;
5203 struct vm_area_struct *vma;
5205 down_read(&mm->mmap_sem);
5206 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5207 struct mm_walk mem_cgroup_count_precharge_walk = {
5208 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5212 if (is_vm_hugetlb_page(vma))
5214 walk_page_range(vma->vm_start, vma->vm_end,
5215 &mem_cgroup_count_precharge_walk);
5217 up_read(&mm->mmap_sem);
5219 precharge = mc.precharge;
5225 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5227 unsigned long precharge = mem_cgroup_count_precharge(mm);
5229 VM_BUG_ON(mc.moving_task);
5230 mc.moving_task = current;
5231 return mem_cgroup_do_precharge(precharge);
5234 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5235 static void __mem_cgroup_clear_mc(void)
5237 struct mem_cgroup *from = mc.from;
5238 struct mem_cgroup *to = mc.to;
5240 /* we must uncharge all the leftover precharges from mc.to */
5242 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5246 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5247 * we must uncharge here.
5249 if (mc.moved_charge) {
5250 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5251 mc.moved_charge = 0;
5253 /* we must fixup refcnts and charges */
5254 if (mc.moved_swap) {
5255 /* uncharge swap account from the old cgroup */
5256 if (!mem_cgroup_is_root(mc.from))
5257 res_counter_uncharge(&mc.from->memsw,
5258 PAGE_SIZE * mc.moved_swap);
5259 __mem_cgroup_put(mc.from, mc.moved_swap);
5261 if (!mem_cgroup_is_root(mc.to)) {
5263 * we charged both to->res and to->memsw, so we should
5266 res_counter_uncharge(&mc.to->res,
5267 PAGE_SIZE * mc.moved_swap);
5269 /* we've already done mem_cgroup_get(mc.to) */
5272 memcg_oom_recover(from);
5273 memcg_oom_recover(to);
5274 wake_up_all(&mc.waitq);
5277 static void mem_cgroup_clear_mc(void)
5279 struct mem_cgroup *from = mc.from;
5282 * we must clear moving_task before waking up waiters at the end of
5285 mc.moving_task = NULL;
5286 __mem_cgroup_clear_mc();
5287 spin_lock(&mc.lock);
5290 spin_unlock(&mc.lock);
5291 mem_cgroup_end_move(from);
5294 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5295 struct cgroup *cgroup,
5296 struct task_struct *p)
5299 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5301 if (memcg->move_charge_at_immigrate) {
5302 struct mm_struct *mm;
5303 struct mem_cgroup *from = mem_cgroup_from_task(p);
5305 VM_BUG_ON(from == memcg);
5307 mm = get_task_mm(p);
5310 /* We move charges only when we move a owner of the mm */
5311 if (mm->owner == p) {
5314 VM_BUG_ON(mc.precharge);
5315 VM_BUG_ON(mc.moved_charge);
5316 VM_BUG_ON(mc.moved_swap);
5317 mem_cgroup_start_move(from);
5318 spin_lock(&mc.lock);
5321 spin_unlock(&mc.lock);
5322 /* We set mc.moving_task later */
5324 ret = mem_cgroup_precharge_mc(mm);
5326 mem_cgroup_clear_mc();
5333 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5334 struct cgroup *cgroup,
5335 struct task_struct *p)
5337 mem_cgroup_clear_mc();
5340 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5341 unsigned long addr, unsigned long end,
5342 struct mm_walk *walk)
5345 struct vm_area_struct *vma = walk->private;
5349 split_huge_page_pmd(walk->mm, pmd);
5351 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5352 for (; addr != end; addr += PAGE_SIZE) {
5353 pte_t ptent = *(pte++);
5354 union mc_target target;
5357 struct page_cgroup *pc;
5363 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5365 case MC_TARGET_PAGE:
5367 if (isolate_lru_page(page))
5369 pc = lookup_page_cgroup(page);
5370 if (!mem_cgroup_move_account(page, 1, pc,
5371 mc.from, mc.to, false)) {
5373 /* we uncharge from mc.from later. */
5376 putback_lru_page(page);
5377 put: /* is_target_pte_for_mc() gets the page */
5380 case MC_TARGET_SWAP:
5382 if (!mem_cgroup_move_swap_account(ent,
5383 mc.from, mc.to, false)) {
5385 /* we fixup refcnts and charges later. */
5393 pte_unmap_unlock(pte - 1, ptl);
5398 * We have consumed all precharges we got in can_attach().
5399 * We try charge one by one, but don't do any additional
5400 * charges to mc.to if we have failed in charge once in attach()
5403 ret = mem_cgroup_do_precharge(1);
5411 static void mem_cgroup_move_charge(struct mm_struct *mm)
5413 struct vm_area_struct *vma;
5415 lru_add_drain_all();
5417 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5419 * Someone who are holding the mmap_sem might be waiting in
5420 * waitq. So we cancel all extra charges, wake up all waiters,
5421 * and retry. Because we cancel precharges, we might not be able
5422 * to move enough charges, but moving charge is a best-effort
5423 * feature anyway, so it wouldn't be a big problem.
5425 __mem_cgroup_clear_mc();
5429 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5431 struct mm_walk mem_cgroup_move_charge_walk = {
5432 .pmd_entry = mem_cgroup_move_charge_pte_range,
5436 if (is_vm_hugetlb_page(vma))
5438 ret = walk_page_range(vma->vm_start, vma->vm_end,
5439 &mem_cgroup_move_charge_walk);
5442 * means we have consumed all precharges and failed in
5443 * doing additional charge. Just abandon here.
5447 up_read(&mm->mmap_sem);
5450 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5451 struct cgroup *cont,
5452 struct cgroup *old_cont,
5453 struct task_struct *p)
5455 struct mm_struct *mm = get_task_mm(p);
5459 mem_cgroup_move_charge(mm);
5464 mem_cgroup_clear_mc();
5466 #else /* !CONFIG_MMU */
5467 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5468 struct cgroup *cgroup,
5469 struct task_struct *p)
5473 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5474 struct cgroup *cgroup,
5475 struct task_struct *p)
5478 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5479 struct cgroup *cont,
5480 struct cgroup *old_cont,
5481 struct task_struct *p)
5486 struct cgroup_subsys mem_cgroup_subsys = {
5488 .subsys_id = mem_cgroup_subsys_id,
5489 .create = mem_cgroup_create,
5490 .pre_destroy = mem_cgroup_pre_destroy,
5491 .destroy = mem_cgroup_destroy,
5492 .populate = mem_cgroup_populate,
5493 .can_attach = mem_cgroup_can_attach,
5494 .cancel_attach = mem_cgroup_cancel_attach,
5495 .attach = mem_cgroup_move_task,
5500 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5501 static int __init enable_swap_account(char *s)
5503 /* consider enabled if no parameter or 1 is given */
5504 if (!strcmp(s, "1"))
5505 really_do_swap_account = 1;
5506 else if (!strcmp(s, "0"))
5507 really_do_swap_account = 0;
5510 __setup("swapaccount=", enable_swap_account);