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,
112 #define THRESHOLDS_EVENTS_TARGET (128)
113 #define SOFTLIMIT_EVENTS_TARGET (1024)
115 struct mem_cgroup_stat_cpu {
116 long count[MEM_CGROUP_STAT_NSTATS];
117 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
118 unsigned long targets[MEM_CGROUP_NTARGETS];
122 * per-zone information in memory controller.
124 struct mem_cgroup_per_zone {
126 * spin_lock to protect the per cgroup LRU
128 struct list_head lists[NR_LRU_LISTS];
129 unsigned long count[NR_LRU_LISTS];
131 struct zone_reclaim_stat reclaim_stat;
132 struct rb_node tree_node; /* RB tree node */
133 unsigned long long usage_in_excess;/* Set to the value by which */
134 /* the soft limit is exceeded*/
136 struct mem_cgroup *mem; /* Back pointer, we cannot */
137 /* use container_of */
139 /* Macro for accessing counter */
140 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
142 struct mem_cgroup_per_node {
143 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
146 struct mem_cgroup_lru_info {
147 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
151 * Cgroups above their limits are maintained in a RB-Tree, independent of
152 * their hierarchy representation
155 struct mem_cgroup_tree_per_zone {
156 struct rb_root rb_root;
160 struct mem_cgroup_tree_per_node {
161 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
164 struct mem_cgroup_tree {
165 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
168 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
170 struct mem_cgroup_threshold {
171 struct eventfd_ctx *eventfd;
176 struct mem_cgroup_threshold_ary {
177 /* An array index points to threshold just below usage. */
178 int current_threshold;
179 /* Size of entries[] */
181 /* Array of thresholds */
182 struct mem_cgroup_threshold entries[0];
185 struct mem_cgroup_thresholds {
186 /* Primary thresholds array */
187 struct mem_cgroup_threshold_ary *primary;
189 * Spare threshold array.
190 * This is needed to make mem_cgroup_unregister_event() "never fail".
191 * It must be able to store at least primary->size - 1 entries.
193 struct mem_cgroup_threshold_ary *spare;
197 struct mem_cgroup_eventfd_list {
198 struct list_head list;
199 struct eventfd_ctx *eventfd;
202 static void mem_cgroup_threshold(struct mem_cgroup *mem);
203 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
206 * The memory controller data structure. The memory controller controls both
207 * page cache and RSS per cgroup. We would eventually like to provide
208 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
209 * to help the administrator determine what knobs to tune.
211 * TODO: Add a water mark for the memory controller. Reclaim will begin when
212 * we hit the water mark. May be even add a low water mark, such that
213 * no reclaim occurs from a cgroup at it's low water mark, this is
214 * a feature that will be implemented much later in the future.
217 struct cgroup_subsys_state css;
219 * the counter to account for memory usage
221 struct res_counter res;
223 * the counter to account for mem+swap usage.
225 struct res_counter memsw;
227 * Per cgroup active and inactive list, similar to the
228 * per zone LRU lists.
230 struct mem_cgroup_lru_info info;
232 * While reclaiming in a hierarchy, we cache the last child we
235 int last_scanned_child;
236 int last_scanned_node;
238 nodemask_t scan_nodes;
239 unsigned long next_scan_node_update;
242 * Should the accounting and control be hierarchical, per subtree?
248 unsigned int swappiness;
249 /* OOM-Killer disable */
250 int oom_kill_disable;
252 /* set when res.limit == memsw.limit */
253 bool memsw_is_minimum;
255 /* protect arrays of thresholds */
256 struct mutex thresholds_lock;
258 /* thresholds for memory usage. RCU-protected */
259 struct mem_cgroup_thresholds thresholds;
261 /* thresholds for mem+swap usage. RCU-protected */
262 struct mem_cgroup_thresholds memsw_thresholds;
264 /* For oom notifier event fd */
265 struct list_head oom_notify;
268 * Should we move charges of a task when a task is moved into this
269 * mem_cgroup ? And what type of charges should we move ?
271 unsigned long move_charge_at_immigrate;
275 struct mem_cgroup_stat_cpu *stat;
277 * used when a cpu is offlined or other synchronizations
278 * See mem_cgroup_read_stat().
280 struct mem_cgroup_stat_cpu nocpu_base;
281 spinlock_t pcp_counter_lock;
284 /* Stuffs for move charges at task migration. */
286 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
287 * left-shifted bitmap of these types.
290 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
291 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
295 /* "mc" and its members are protected by cgroup_mutex */
296 static struct move_charge_struct {
297 spinlock_t lock; /* for from, to */
298 struct mem_cgroup *from;
299 struct mem_cgroup *to;
300 unsigned long precharge;
301 unsigned long moved_charge;
302 unsigned long moved_swap;
303 struct task_struct *moving_task; /* a task moving charges */
304 wait_queue_head_t waitq; /* a waitq for other context */
306 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
307 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
310 static bool move_anon(void)
312 return test_bit(MOVE_CHARGE_TYPE_ANON,
313 &mc.to->move_charge_at_immigrate);
316 static bool move_file(void)
318 return test_bit(MOVE_CHARGE_TYPE_FILE,
319 &mc.to->move_charge_at_immigrate);
323 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
324 * limit reclaim to prevent infinite loops, if they ever occur.
326 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
327 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
330 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
331 MEM_CGROUP_CHARGE_TYPE_MAPPED,
332 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
333 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
334 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
335 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
339 /* for encoding cft->private value on file */
342 #define _OOM_TYPE (2)
343 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
344 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
345 #define MEMFILE_ATTR(val) ((val) & 0xffff)
346 /* Used for OOM nofiier */
347 #define OOM_CONTROL (0)
350 * Reclaim flags for mem_cgroup_hierarchical_reclaim
352 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
353 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
354 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
355 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
356 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
357 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
359 static void mem_cgroup_get(struct mem_cgroup *mem);
360 static void mem_cgroup_put(struct mem_cgroup *mem);
361 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
362 static void drain_all_stock_async(struct mem_cgroup *mem);
364 static struct mem_cgroup_per_zone *
365 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
367 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
370 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
375 static struct mem_cgroup_per_zone *
376 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
378 int nid = page_to_nid(page);
379 int zid = page_zonenum(page);
381 return mem_cgroup_zoneinfo(mem, nid, zid);
384 static struct mem_cgroup_tree_per_zone *
385 soft_limit_tree_node_zone(int nid, int zid)
387 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
390 static struct mem_cgroup_tree_per_zone *
391 soft_limit_tree_from_page(struct page *page)
393 int nid = page_to_nid(page);
394 int zid = page_zonenum(page);
396 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
400 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
401 struct mem_cgroup_per_zone *mz,
402 struct mem_cgroup_tree_per_zone *mctz,
403 unsigned long long new_usage_in_excess)
405 struct rb_node **p = &mctz->rb_root.rb_node;
406 struct rb_node *parent = NULL;
407 struct mem_cgroup_per_zone *mz_node;
412 mz->usage_in_excess = new_usage_in_excess;
413 if (!mz->usage_in_excess)
417 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
419 if (mz->usage_in_excess < mz_node->usage_in_excess)
422 * We can't avoid mem cgroups that are over their soft
423 * limit by the same amount
425 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
428 rb_link_node(&mz->tree_node, parent, p);
429 rb_insert_color(&mz->tree_node, &mctz->rb_root);
434 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
435 struct mem_cgroup_per_zone *mz,
436 struct mem_cgroup_tree_per_zone *mctz)
440 rb_erase(&mz->tree_node, &mctz->rb_root);
445 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
446 struct mem_cgroup_per_zone *mz,
447 struct mem_cgroup_tree_per_zone *mctz)
449 spin_lock(&mctz->lock);
450 __mem_cgroup_remove_exceeded(mem, mz, mctz);
451 spin_unlock(&mctz->lock);
455 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
457 unsigned long long excess;
458 struct mem_cgroup_per_zone *mz;
459 struct mem_cgroup_tree_per_zone *mctz;
460 int nid = page_to_nid(page);
461 int zid = page_zonenum(page);
462 mctz = soft_limit_tree_from_page(page);
465 * Necessary to update all ancestors when hierarchy is used.
466 * because their event counter is not touched.
468 for (; mem; mem = parent_mem_cgroup(mem)) {
469 mz = mem_cgroup_zoneinfo(mem, nid, zid);
470 excess = res_counter_soft_limit_excess(&mem->res);
472 * We have to update the tree if mz is on RB-tree or
473 * mem is over its softlimit.
475 if (excess || mz->on_tree) {
476 spin_lock(&mctz->lock);
477 /* if on-tree, remove it */
479 __mem_cgroup_remove_exceeded(mem, mz, mctz);
481 * Insert again. mz->usage_in_excess will be updated.
482 * If excess is 0, no tree ops.
484 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
485 spin_unlock(&mctz->lock);
490 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
493 struct mem_cgroup_per_zone *mz;
494 struct mem_cgroup_tree_per_zone *mctz;
496 for_each_node_state(node, N_POSSIBLE) {
497 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
498 mz = mem_cgroup_zoneinfo(mem, node, zone);
499 mctz = soft_limit_tree_node_zone(node, zone);
500 mem_cgroup_remove_exceeded(mem, mz, mctz);
505 static struct mem_cgroup_per_zone *
506 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
508 struct rb_node *rightmost = NULL;
509 struct mem_cgroup_per_zone *mz;
513 rightmost = rb_last(&mctz->rb_root);
515 goto done; /* Nothing to reclaim from */
517 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
519 * Remove the node now but someone else can add it back,
520 * we will to add it back at the end of reclaim to its correct
521 * position in the tree.
523 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
524 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
525 !css_tryget(&mz->mem->css))
531 static struct mem_cgroup_per_zone *
532 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
534 struct mem_cgroup_per_zone *mz;
536 spin_lock(&mctz->lock);
537 mz = __mem_cgroup_largest_soft_limit_node(mctz);
538 spin_unlock(&mctz->lock);
543 * Implementation Note: reading percpu statistics for memcg.
545 * Both of vmstat[] and percpu_counter has threshold and do periodic
546 * synchronization to implement "quick" read. There are trade-off between
547 * reading cost and precision of value. Then, we may have a chance to implement
548 * a periodic synchronizion of counter in memcg's counter.
550 * But this _read() function is used for user interface now. The user accounts
551 * memory usage by memory cgroup and he _always_ requires exact value because
552 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
553 * have to visit all online cpus and make sum. So, for now, unnecessary
554 * synchronization is not implemented. (just implemented for cpu hotplug)
556 * If there are kernel internal actions which can make use of some not-exact
557 * value, and reading all cpu value can be performance bottleneck in some
558 * common workload, threashold and synchonization as vmstat[] should be
561 static long mem_cgroup_read_stat(struct mem_cgroup *mem,
562 enum mem_cgroup_stat_index idx)
568 for_each_online_cpu(cpu)
569 val += per_cpu(mem->stat->count[idx], cpu);
570 #ifdef CONFIG_HOTPLUG_CPU
571 spin_lock(&mem->pcp_counter_lock);
572 val += mem->nocpu_base.count[idx];
573 spin_unlock(&mem->pcp_counter_lock);
579 static long mem_cgroup_local_usage(struct mem_cgroup *mem)
583 ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
584 ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
588 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
591 int val = (charge) ? 1 : -1;
592 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
595 void mem_cgroup_pgfault(struct mem_cgroup *mem, int val)
597 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
600 void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val)
602 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
605 static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
606 enum mem_cgroup_events_index idx)
608 unsigned long val = 0;
611 for_each_online_cpu(cpu)
612 val += per_cpu(mem->stat->events[idx], cpu);
613 #ifdef CONFIG_HOTPLUG_CPU
614 spin_lock(&mem->pcp_counter_lock);
615 val += mem->nocpu_base.events[idx];
616 spin_unlock(&mem->pcp_counter_lock);
621 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
622 bool file, int nr_pages)
627 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
629 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
631 /* pagein of a big page is an event. So, ignore page size */
633 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
635 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
636 nr_pages = -nr_pages; /* for event */
639 __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
645 mem_cgroup_get_zonestat_node(struct mem_cgroup *mem, int nid, enum lru_list idx)
647 struct mem_cgroup_per_zone *mz;
651 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
652 mz = mem_cgroup_zoneinfo(mem, nid, zid);
653 total += MEM_CGROUP_ZSTAT(mz, idx);
657 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
663 for_each_online_node(nid)
664 total += mem_cgroup_get_zonestat_node(mem, nid, idx);
668 static bool __memcg_event_check(struct mem_cgroup *mem, int target)
670 unsigned long val, next;
672 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
673 next = this_cpu_read(mem->stat->targets[target]);
674 /* from time_after() in jiffies.h */
675 return ((long)next - (long)val < 0);
678 static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target)
680 unsigned long val, next;
682 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
685 case MEM_CGROUP_TARGET_THRESH:
686 next = val + THRESHOLDS_EVENTS_TARGET;
688 case MEM_CGROUP_TARGET_SOFTLIMIT:
689 next = val + SOFTLIMIT_EVENTS_TARGET;
695 this_cpu_write(mem->stat->targets[target], next);
699 * Check events in order.
702 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
704 /* threshold event is triggered in finer grain than soft limit */
705 if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) {
706 mem_cgroup_threshold(mem);
707 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH);
708 if (unlikely(__memcg_event_check(mem,
709 MEM_CGROUP_TARGET_SOFTLIMIT))){
710 mem_cgroup_update_tree(mem, page);
711 __mem_cgroup_target_update(mem,
712 MEM_CGROUP_TARGET_SOFTLIMIT);
717 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
719 return container_of(cgroup_subsys_state(cont,
720 mem_cgroup_subsys_id), struct mem_cgroup,
724 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
727 * mm_update_next_owner() may clear mm->owner to NULL
728 * if it races with swapoff, page migration, etc.
729 * So this can be called with p == NULL.
734 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
735 struct mem_cgroup, css);
738 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
740 struct mem_cgroup *mem = NULL;
745 * Because we have no locks, mm->owner's may be being moved to other
746 * cgroup. We use css_tryget() here even if this looks
747 * pessimistic (rather than adding locks here).
751 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
754 } while (!css_tryget(&mem->css));
759 /* The caller has to guarantee "mem" exists before calling this */
760 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
762 struct cgroup_subsys_state *css;
765 if (!mem) /* ROOT cgroup has the smallest ID */
766 return root_mem_cgroup; /*css_put/get against root is ignored*/
767 if (!mem->use_hierarchy) {
768 if (css_tryget(&mem->css))
774 * searching a memory cgroup which has the smallest ID under given
775 * ROOT cgroup. (ID >= 1)
777 css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
778 if (css && css_tryget(css))
779 mem = container_of(css, struct mem_cgroup, css);
786 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
787 struct mem_cgroup *root,
790 int nextid = css_id(&iter->css) + 1;
793 struct cgroup_subsys_state *css;
795 hierarchy_used = iter->use_hierarchy;
798 /* If no ROOT, walk all, ignore hierarchy */
799 if (!cond || (root && !hierarchy_used))
803 root = root_mem_cgroup;
809 css = css_get_next(&mem_cgroup_subsys, nextid,
811 if (css && css_tryget(css))
812 iter = container_of(css, struct mem_cgroup, css);
814 /* If css is NULL, no more cgroups will be found */
816 } while (css && !iter);
821 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
822 * be careful that "break" loop is not allowed. We have reference count.
823 * Instead of that modify "cond" to be false and "continue" to exit the loop.
825 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
826 for (iter = mem_cgroup_start_loop(root);\
828 iter = mem_cgroup_get_next(iter, root, cond))
830 #define for_each_mem_cgroup_tree(iter, root) \
831 for_each_mem_cgroup_tree_cond(iter, root, true)
833 #define for_each_mem_cgroup_all(iter) \
834 for_each_mem_cgroup_tree_cond(iter, NULL, true)
837 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
839 return (mem == root_mem_cgroup);
842 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
844 struct mem_cgroup *mem;
850 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
856 mem_cgroup_pgmajfault(mem, 1);
859 mem_cgroup_pgfault(mem, 1);
867 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
870 * Following LRU functions are allowed to be used without PCG_LOCK.
871 * Operations are called by routine of global LRU independently from memcg.
872 * What we have to take care of here is validness of pc->mem_cgroup.
874 * Changes to pc->mem_cgroup happens when
877 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
878 * It is added to LRU before charge.
879 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
880 * When moving account, the page is not on LRU. It's isolated.
883 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
885 struct page_cgroup *pc;
886 struct mem_cgroup_per_zone *mz;
888 if (mem_cgroup_disabled())
890 pc = lookup_page_cgroup(page);
891 /* can happen while we handle swapcache. */
892 if (!TestClearPageCgroupAcctLRU(pc))
894 VM_BUG_ON(!pc->mem_cgroup);
896 * We don't check PCG_USED bit. It's cleared when the "page" is finally
897 * removed from global LRU.
899 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
900 /* huge page split is done under lru_lock. so, we have no races. */
901 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
902 if (mem_cgroup_is_root(pc->mem_cgroup))
904 VM_BUG_ON(list_empty(&pc->lru));
905 list_del_init(&pc->lru);
908 void mem_cgroup_del_lru(struct page *page)
910 mem_cgroup_del_lru_list(page, page_lru(page));
914 * Writeback is about to end against a page which has been marked for immediate
915 * reclaim. If it still appears to be reclaimable, move it to the tail of the
918 void mem_cgroup_rotate_reclaimable_page(struct page *page)
920 struct mem_cgroup_per_zone *mz;
921 struct page_cgroup *pc;
922 enum lru_list lru = page_lru(page);
924 if (mem_cgroup_disabled())
927 pc = lookup_page_cgroup(page);
928 /* unused or root page is not rotated. */
929 if (!PageCgroupUsed(pc))
931 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
933 if (mem_cgroup_is_root(pc->mem_cgroup))
935 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
936 list_move_tail(&pc->lru, &mz->lists[lru]);
939 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
941 struct mem_cgroup_per_zone *mz;
942 struct page_cgroup *pc;
944 if (mem_cgroup_disabled())
947 pc = lookup_page_cgroup(page);
948 /* unused or root page is not rotated. */
949 if (!PageCgroupUsed(pc))
951 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
953 if (mem_cgroup_is_root(pc->mem_cgroup))
955 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
956 list_move(&pc->lru, &mz->lists[lru]);
959 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
961 struct page_cgroup *pc;
962 struct mem_cgroup_per_zone *mz;
964 if (mem_cgroup_disabled())
966 pc = lookup_page_cgroup(page);
967 VM_BUG_ON(PageCgroupAcctLRU(pc));
968 if (!PageCgroupUsed(pc))
970 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
972 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
973 /* huge page split is done under lru_lock. so, we have no races. */
974 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
975 SetPageCgroupAcctLRU(pc);
976 if (mem_cgroup_is_root(pc->mem_cgroup))
978 list_add(&pc->lru, &mz->lists[lru]);
982 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
983 * while it's linked to lru because the page may be reused after it's fully
984 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
985 * It's done under lock_page and expected that zone->lru_lock isnever held.
987 static void mem_cgroup_lru_del_before_commit(struct page *page)
990 struct zone *zone = page_zone(page);
991 struct page_cgroup *pc = lookup_page_cgroup(page);
994 * Doing this check without taking ->lru_lock seems wrong but this
995 * is safe. Because if page_cgroup's USED bit is unset, the page
996 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
997 * set, the commit after this will fail, anyway.
998 * This all charge/uncharge is done under some mutual execustion.
999 * So, we don't need to taking care of changes in USED bit.
1001 if (likely(!PageLRU(page)))
1004 spin_lock_irqsave(&zone->lru_lock, flags);
1006 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1007 * is guarded by lock_page() because the page is SwapCache.
1009 if (!PageCgroupUsed(pc))
1010 mem_cgroup_del_lru_list(page, page_lru(page));
1011 spin_unlock_irqrestore(&zone->lru_lock, flags);
1014 static void mem_cgroup_lru_add_after_commit(struct page *page)
1016 unsigned long flags;
1017 struct zone *zone = page_zone(page);
1018 struct page_cgroup *pc = lookup_page_cgroup(page);
1020 /* taking care of that the page is added to LRU while we commit it */
1021 if (likely(!PageLRU(page)))
1023 spin_lock_irqsave(&zone->lru_lock, flags);
1024 /* link when the page is linked to LRU but page_cgroup isn't */
1025 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1026 mem_cgroup_add_lru_list(page, page_lru(page));
1027 spin_unlock_irqrestore(&zone->lru_lock, flags);
1031 void mem_cgroup_move_lists(struct page *page,
1032 enum lru_list from, enum lru_list to)
1034 if (mem_cgroup_disabled())
1036 mem_cgroup_del_lru_list(page, from);
1037 mem_cgroup_add_lru_list(page, to);
1040 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
1043 struct mem_cgroup *curr = NULL;
1044 struct task_struct *p;
1046 p = find_lock_task_mm(task);
1049 curr = try_get_mem_cgroup_from_mm(p->mm);
1054 * We should check use_hierarchy of "mem" not "curr". Because checking
1055 * use_hierarchy of "curr" here make this function true if hierarchy is
1056 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
1057 * hierarchy(even if use_hierarchy is disabled in "mem").
1059 if (mem->use_hierarchy)
1060 ret = css_is_ancestor(&curr->css, &mem->css);
1062 ret = (curr == mem);
1063 css_put(&curr->css);
1067 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1069 unsigned long active;
1070 unsigned long inactive;
1072 unsigned long inactive_ratio;
1074 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
1075 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
1077 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1079 inactive_ratio = int_sqrt(10 * gb);
1083 if (present_pages) {
1084 present_pages[0] = inactive;
1085 present_pages[1] = active;
1088 return inactive_ratio;
1091 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1093 unsigned long active;
1094 unsigned long inactive;
1095 unsigned long present_pages[2];
1096 unsigned long inactive_ratio;
1098 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1100 inactive = present_pages[0];
1101 active = present_pages[1];
1103 if (inactive * inactive_ratio < active)
1109 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1111 unsigned long active;
1112 unsigned long inactive;
1114 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
1115 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
1117 return (active > inactive);
1120 unsigned long mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg,
1124 int nid = zone_to_nid(zone);
1125 int zid = zone_idx(zone);
1126 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1128 return MEM_CGROUP_ZSTAT(mz, lru);
1132 static unsigned long mem_cgroup_node_nr_file_lru_pages(struct mem_cgroup *memcg,
1137 ret = mem_cgroup_get_zonestat_node(memcg, nid, LRU_INACTIVE_FILE) +
1138 mem_cgroup_get_zonestat_node(memcg, nid, LRU_ACTIVE_FILE);
1143 static unsigned long mem_cgroup_nr_file_lru_pages(struct mem_cgroup *memcg)
1148 for_each_node_state(nid, N_HIGH_MEMORY)
1149 total += mem_cgroup_node_nr_file_lru_pages(memcg, nid);
1154 static unsigned long mem_cgroup_node_nr_anon_lru_pages(struct mem_cgroup *memcg,
1159 ret = mem_cgroup_get_zonestat_node(memcg, nid, LRU_INACTIVE_ANON) +
1160 mem_cgroup_get_zonestat_node(memcg, nid, LRU_ACTIVE_ANON);
1165 static unsigned long mem_cgroup_nr_anon_lru_pages(struct mem_cgroup *memcg)
1170 for_each_node_state(nid, N_HIGH_MEMORY)
1171 total += mem_cgroup_node_nr_anon_lru_pages(memcg, nid);
1176 static unsigned long
1177 mem_cgroup_node_nr_unevictable_lru_pages(struct mem_cgroup *memcg, int nid)
1179 return mem_cgroup_get_zonestat_node(memcg, nid, LRU_UNEVICTABLE);
1182 static unsigned long
1183 mem_cgroup_nr_unevictable_lru_pages(struct mem_cgroup *memcg)
1188 for_each_node_state(nid, N_HIGH_MEMORY)
1189 total += mem_cgroup_node_nr_unevictable_lru_pages(memcg, nid);
1194 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
1201 total += mem_cgroup_get_zonestat_node(memcg, nid, l);
1206 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg)
1211 for_each_node_state(nid, N_HIGH_MEMORY)
1212 total += mem_cgroup_node_nr_lru_pages(memcg, nid);
1216 #endif /* CONFIG_NUMA */
1218 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1221 int nid = zone_to_nid(zone);
1222 int zid = zone_idx(zone);
1223 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1225 return &mz->reclaim_stat;
1228 struct zone_reclaim_stat *
1229 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1231 struct page_cgroup *pc;
1232 struct mem_cgroup_per_zone *mz;
1234 if (mem_cgroup_disabled())
1237 pc = lookup_page_cgroup(page);
1238 if (!PageCgroupUsed(pc))
1240 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1242 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1243 return &mz->reclaim_stat;
1246 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1247 struct list_head *dst,
1248 unsigned long *scanned, int order,
1249 int mode, struct zone *z,
1250 struct mem_cgroup *mem_cont,
1251 int active, int file)
1253 unsigned long nr_taken = 0;
1257 struct list_head *src;
1258 struct page_cgroup *pc, *tmp;
1259 int nid = zone_to_nid(z);
1260 int zid = zone_idx(z);
1261 struct mem_cgroup_per_zone *mz;
1262 int lru = LRU_FILE * file + active;
1266 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1267 src = &mz->lists[lru];
1270 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1271 if (scan >= nr_to_scan)
1274 if (unlikely(!PageCgroupUsed(pc)))
1277 page = lookup_cgroup_page(pc);
1279 if (unlikely(!PageLRU(page)))
1283 ret = __isolate_lru_page(page, mode, file);
1286 list_move(&page->lru, dst);
1287 mem_cgroup_del_lru(page);
1288 nr_taken += hpage_nr_pages(page);
1291 /* we don't affect global LRU but rotate in our LRU */
1292 mem_cgroup_rotate_lru_list(page, page_lru(page));
1301 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1307 #define mem_cgroup_from_res_counter(counter, member) \
1308 container_of(counter, struct mem_cgroup, member)
1311 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1312 * @mem: the memory cgroup
1314 * Returns the maximum amount of memory @mem can be charged with, in
1317 static unsigned long mem_cgroup_margin(struct mem_cgroup *mem)
1319 unsigned long long margin;
1321 margin = res_counter_margin(&mem->res);
1322 if (do_swap_account)
1323 margin = min(margin, res_counter_margin(&mem->memsw));
1324 return margin >> PAGE_SHIFT;
1327 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1329 struct cgroup *cgrp = memcg->css.cgroup;
1332 if (cgrp->parent == NULL)
1333 return vm_swappiness;
1335 return memcg->swappiness;
1338 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1343 spin_lock(&mem->pcp_counter_lock);
1344 for_each_online_cpu(cpu)
1345 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1346 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1347 spin_unlock(&mem->pcp_counter_lock);
1353 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1360 spin_lock(&mem->pcp_counter_lock);
1361 for_each_online_cpu(cpu)
1362 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1363 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1364 spin_unlock(&mem->pcp_counter_lock);
1368 * 2 routines for checking "mem" is under move_account() or not.
1370 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1371 * for avoiding race in accounting. If true,
1372 * pc->mem_cgroup may be overwritten.
1374 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1375 * under hierarchy of moving cgroups. This is for
1376 * waiting at hith-memory prressure caused by "move".
1379 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1381 VM_BUG_ON(!rcu_read_lock_held());
1382 return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1385 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1387 struct mem_cgroup *from;
1388 struct mem_cgroup *to;
1391 * Unlike task_move routines, we access mc.to, mc.from not under
1392 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1394 spin_lock(&mc.lock);
1399 if (from == mem || to == mem
1400 || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1401 || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1404 spin_unlock(&mc.lock);
1408 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1410 if (mc.moving_task && current != mc.moving_task) {
1411 if (mem_cgroup_under_move(mem)) {
1413 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1414 /* moving charge context might have finished. */
1417 finish_wait(&mc.waitq, &wait);
1425 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1426 * @memcg: The memory cgroup that went over limit
1427 * @p: Task that is going to be killed
1429 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1432 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1434 struct cgroup *task_cgrp;
1435 struct cgroup *mem_cgrp;
1437 * Need a buffer in BSS, can't rely on allocations. The code relies
1438 * on the assumption that OOM is serialized for memory controller.
1439 * If this assumption is broken, revisit this code.
1441 static char memcg_name[PATH_MAX];
1450 mem_cgrp = memcg->css.cgroup;
1451 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1453 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1456 * Unfortunately, we are unable to convert to a useful name
1457 * But we'll still print out the usage information
1464 printk(KERN_INFO "Task in %s killed", memcg_name);
1467 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1475 * Continues from above, so we don't need an KERN_ level
1477 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1480 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1481 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1482 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1483 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1484 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1486 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1487 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1488 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1492 * This function returns the number of memcg under hierarchy tree. Returns
1493 * 1(self count) if no children.
1495 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1498 struct mem_cgroup *iter;
1500 for_each_mem_cgroup_tree(iter, mem)
1506 * Return the memory (and swap, if configured) limit for a memcg.
1508 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1513 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1514 limit += total_swap_pages << PAGE_SHIFT;
1516 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1518 * If memsw is finite and limits the amount of swap space available
1519 * to this memcg, return that limit.
1521 return min(limit, memsw);
1525 * Visit the first child (need not be the first child as per the ordering
1526 * of the cgroup list, since we track last_scanned_child) of @mem and use
1527 * that to reclaim free pages from.
1529 static struct mem_cgroup *
1530 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1532 struct mem_cgroup *ret = NULL;
1533 struct cgroup_subsys_state *css;
1536 if (!root_mem->use_hierarchy) {
1537 css_get(&root_mem->css);
1543 nextid = root_mem->last_scanned_child + 1;
1544 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1546 if (css && css_tryget(css))
1547 ret = container_of(css, struct mem_cgroup, css);
1550 /* Updates scanning parameter */
1552 /* this means start scan from ID:1 */
1553 root_mem->last_scanned_child = 0;
1555 root_mem->last_scanned_child = found;
1561 #if MAX_NUMNODES > 1
1564 * Always updating the nodemask is not very good - even if we have an empty
1565 * list or the wrong list here, we can start from some node and traverse all
1566 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1569 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
1573 if (time_after(mem->next_scan_node_update, jiffies))
1576 mem->next_scan_node_update = jiffies + 10*HZ;
1577 /* make a nodemask where this memcg uses memory from */
1578 mem->scan_nodes = node_states[N_HIGH_MEMORY];
1580 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1582 if (mem_cgroup_get_zonestat_node(mem, nid, LRU_INACTIVE_FILE) ||
1583 mem_cgroup_get_zonestat_node(mem, nid, LRU_ACTIVE_FILE))
1586 if (total_swap_pages &&
1587 (mem_cgroup_get_zonestat_node(mem, nid, LRU_INACTIVE_ANON) ||
1588 mem_cgroup_get_zonestat_node(mem, nid, LRU_ACTIVE_ANON)))
1590 node_clear(nid, mem->scan_nodes);
1595 * Selecting a node where we start reclaim from. Because what we need is just
1596 * reducing usage counter, start from anywhere is O,K. Considering
1597 * memory reclaim from current node, there are pros. and cons.
1599 * Freeing memory from current node means freeing memory from a node which
1600 * we'll use or we've used. So, it may make LRU bad. And if several threads
1601 * hit limits, it will see a contention on a node. But freeing from remote
1602 * node means more costs for memory reclaim because of memory latency.
1604 * Now, we use round-robin. Better algorithm is welcomed.
1606 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1610 mem_cgroup_may_update_nodemask(mem);
1611 node = mem->last_scanned_node;
1613 node = next_node(node, mem->scan_nodes);
1614 if (node == MAX_NUMNODES)
1615 node = first_node(mem->scan_nodes);
1617 * We call this when we hit limit, not when pages are added to LRU.
1618 * No LRU may hold pages because all pages are UNEVICTABLE or
1619 * memcg is too small and all pages are not on LRU. In that case,
1620 * we use curret node.
1622 if (unlikely(node == MAX_NUMNODES))
1623 node = numa_node_id();
1625 mem->last_scanned_node = node;
1630 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1637 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1638 * we reclaimed from, so that we don't end up penalizing one child extensively
1639 * based on its position in the children list.
1641 * root_mem is the original ancestor that we've been reclaim from.
1643 * We give up and return to the caller when we visit root_mem twice.
1644 * (other groups can be removed while we're walking....)
1646 * If shrink==true, for avoiding to free too much, this returns immedieately.
1648 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1651 unsigned long reclaim_options,
1652 unsigned long *total_scanned)
1654 struct mem_cgroup *victim;
1657 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1658 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1659 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1660 unsigned long excess;
1661 unsigned long nr_scanned;
1663 excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1665 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1666 if (!check_soft && root_mem->memsw_is_minimum)
1670 victim = mem_cgroup_select_victim(root_mem);
1671 if (victim == root_mem) {
1674 * We are not draining per cpu cached charges during
1675 * soft limit reclaim because global reclaim doesn't
1676 * care about charges. It tries to free some memory and
1677 * charges will not give any.
1679 if (!check_soft && loop >= 1)
1680 drain_all_stock_async(root_mem);
1683 * If we have not been able to reclaim
1684 * anything, it might because there are
1685 * no reclaimable pages under this hierarchy
1687 if (!check_soft || !total) {
1688 css_put(&victim->css);
1692 * We want to do more targeted reclaim.
1693 * excess >> 2 is not to excessive so as to
1694 * reclaim too much, nor too less that we keep
1695 * coming back to reclaim from this cgroup
1697 if (total >= (excess >> 2) ||
1698 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1699 css_put(&victim->css);
1704 if (!mem_cgroup_local_usage(victim)) {
1705 /* this cgroup's local usage == 0 */
1706 css_put(&victim->css);
1709 /* we use swappiness of local cgroup */
1711 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1712 noswap, get_swappiness(victim), zone,
1714 *total_scanned += nr_scanned;
1716 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1717 noswap, get_swappiness(victim));
1718 css_put(&victim->css);
1720 * At shrinking usage, we can't check we should stop here or
1721 * reclaim more. It's depends on callers. last_scanned_child
1722 * will work enough for keeping fairness under tree.
1728 if (!res_counter_soft_limit_excess(&root_mem->res))
1730 } else if (mem_cgroup_margin(root_mem))
1737 * Check OOM-Killer is already running under our hierarchy.
1738 * If someone is running, return false.
1740 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1742 int x, lock_count = 0;
1743 struct mem_cgroup *iter;
1745 for_each_mem_cgroup_tree(iter, mem) {
1746 x = atomic_inc_return(&iter->oom_lock);
1747 lock_count = max(x, lock_count);
1750 if (lock_count == 1)
1755 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1757 struct mem_cgroup *iter;
1760 * When a new child is created while the hierarchy is under oom,
1761 * mem_cgroup_oom_lock() may not be called. We have to use
1762 * atomic_add_unless() here.
1764 for_each_mem_cgroup_tree(iter, mem)
1765 atomic_add_unless(&iter->oom_lock, -1, 0);
1770 static DEFINE_MUTEX(memcg_oom_mutex);
1771 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1773 struct oom_wait_info {
1774 struct mem_cgroup *mem;
1778 static int memcg_oom_wake_function(wait_queue_t *wait,
1779 unsigned mode, int sync, void *arg)
1781 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1782 struct oom_wait_info *oom_wait_info;
1784 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1786 if (oom_wait_info->mem == wake_mem)
1788 /* if no hierarchy, no match */
1789 if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1792 * Both of oom_wait_info->mem and wake_mem are stable under us.
1793 * Then we can use css_is_ancestor without taking care of RCU.
1795 if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1796 !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1800 return autoremove_wake_function(wait, mode, sync, arg);
1803 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1805 /* for filtering, pass "mem" as argument. */
1806 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1809 static void memcg_oom_recover(struct mem_cgroup *mem)
1811 if (mem && atomic_read(&mem->oom_lock))
1812 memcg_wakeup_oom(mem);
1816 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1818 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1820 struct oom_wait_info owait;
1821 bool locked, need_to_kill;
1824 owait.wait.flags = 0;
1825 owait.wait.func = memcg_oom_wake_function;
1826 owait.wait.private = current;
1827 INIT_LIST_HEAD(&owait.wait.task_list);
1828 need_to_kill = true;
1829 /* At first, try to OOM lock hierarchy under mem.*/
1830 mutex_lock(&memcg_oom_mutex);
1831 locked = mem_cgroup_oom_lock(mem);
1833 * Even if signal_pending(), we can't quit charge() loop without
1834 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1835 * under OOM is always welcomed, use TASK_KILLABLE here.
1837 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1838 if (!locked || mem->oom_kill_disable)
1839 need_to_kill = false;
1841 mem_cgroup_oom_notify(mem);
1842 mutex_unlock(&memcg_oom_mutex);
1845 finish_wait(&memcg_oom_waitq, &owait.wait);
1846 mem_cgroup_out_of_memory(mem, mask);
1849 finish_wait(&memcg_oom_waitq, &owait.wait);
1851 mutex_lock(&memcg_oom_mutex);
1852 mem_cgroup_oom_unlock(mem);
1853 memcg_wakeup_oom(mem);
1854 mutex_unlock(&memcg_oom_mutex);
1856 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1858 /* Give chance to dying process */
1859 schedule_timeout(1);
1864 * Currently used to update mapped file statistics, but the routine can be
1865 * generalized to update other statistics as well.
1867 * Notes: Race condition
1869 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1870 * it tends to be costly. But considering some conditions, we doesn't need
1871 * to do so _always_.
1873 * Considering "charge", lock_page_cgroup() is not required because all
1874 * file-stat operations happen after a page is attached to radix-tree. There
1875 * are no race with "charge".
1877 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1878 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1879 * if there are race with "uncharge". Statistics itself is properly handled
1882 * Considering "move", this is an only case we see a race. To make the race
1883 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1884 * possibility of race condition. If there is, we take a lock.
1887 void mem_cgroup_update_page_stat(struct page *page,
1888 enum mem_cgroup_page_stat_item idx, int val)
1890 struct mem_cgroup *mem;
1891 struct page_cgroup *pc = lookup_page_cgroup(page);
1892 bool need_unlock = false;
1893 unsigned long uninitialized_var(flags);
1899 mem = pc->mem_cgroup;
1900 if (unlikely(!mem || !PageCgroupUsed(pc)))
1902 /* pc->mem_cgroup is unstable ? */
1903 if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1904 /* take a lock against to access pc->mem_cgroup */
1905 move_lock_page_cgroup(pc, &flags);
1907 mem = pc->mem_cgroup;
1908 if (!mem || !PageCgroupUsed(pc))
1913 case MEMCG_NR_FILE_MAPPED:
1915 SetPageCgroupFileMapped(pc);
1916 else if (!page_mapped(page))
1917 ClearPageCgroupFileMapped(pc);
1918 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1924 this_cpu_add(mem->stat->count[idx], val);
1927 if (unlikely(need_unlock))
1928 move_unlock_page_cgroup(pc, &flags);
1932 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1935 * size of first charge trial. "32" comes from vmscan.c's magic value.
1936 * TODO: maybe necessary to use big numbers in big irons.
1938 #define CHARGE_BATCH 32U
1939 struct memcg_stock_pcp {
1940 struct mem_cgroup *cached; /* this never be root cgroup */
1941 unsigned int nr_pages;
1942 struct work_struct work;
1943 unsigned long flags;
1944 #define FLUSHING_CACHED_CHARGE (0)
1946 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1947 static DEFINE_MUTEX(percpu_charge_mutex);
1950 * Try to consume stocked charge on this cpu. If success, one page is consumed
1951 * from local stock and true is returned. If the stock is 0 or charges from a
1952 * cgroup which is not current target, returns false. This stock will be
1955 static bool consume_stock(struct mem_cgroup *mem)
1957 struct memcg_stock_pcp *stock;
1960 stock = &get_cpu_var(memcg_stock);
1961 if (mem == stock->cached && stock->nr_pages)
1963 else /* need to call res_counter_charge */
1965 put_cpu_var(memcg_stock);
1970 * Returns stocks cached in percpu to res_counter and reset cached information.
1972 static void drain_stock(struct memcg_stock_pcp *stock)
1974 struct mem_cgroup *old = stock->cached;
1976 if (stock->nr_pages) {
1977 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
1979 res_counter_uncharge(&old->res, bytes);
1980 if (do_swap_account)
1981 res_counter_uncharge(&old->memsw, bytes);
1982 stock->nr_pages = 0;
1984 stock->cached = NULL;
1988 * This must be called under preempt disabled or must be called by
1989 * a thread which is pinned to local cpu.
1991 static void drain_local_stock(struct work_struct *dummy)
1993 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1995 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
1999 * Cache charges(val) which is from res_counter, to local per_cpu area.
2000 * This will be consumed by consume_stock() function, later.
2002 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
2004 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2006 if (stock->cached != mem) { /* reset if necessary */
2008 stock->cached = mem;
2010 stock->nr_pages += nr_pages;
2011 put_cpu_var(memcg_stock);
2015 * Tries to drain stocked charges in other cpus. This function is asynchronous
2016 * and just put a work per cpu for draining localy on each cpu. Caller can
2017 * expects some charges will be back to res_counter later but cannot wait for
2020 static void drain_all_stock_async(struct mem_cgroup *root_mem)
2024 * If someone calls draining, avoid adding more kworker runs.
2026 if (!mutex_trylock(&percpu_charge_mutex))
2028 /* Notify other cpus that system-wide "drain" is running */
2031 * Get a hint for avoiding draining charges on the current cpu,
2032 * which must be exhausted by our charging. It is not required that
2033 * this be a precise check, so we use raw_smp_processor_id() instead of
2034 * getcpu()/putcpu().
2036 curcpu = raw_smp_processor_id();
2037 for_each_online_cpu(cpu) {
2038 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2039 struct mem_cgroup *mem;
2044 mem = stock->cached;
2047 if (mem != root_mem) {
2048 if (!root_mem->use_hierarchy)
2050 /* check whether "mem" is under tree of "root_mem" */
2051 if (!css_is_ancestor(&mem->css, &root_mem->css))
2054 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2055 schedule_work_on(cpu, &stock->work);
2058 mutex_unlock(&percpu_charge_mutex);
2059 /* We don't wait for flush_work */
2062 /* This is a synchronous drain interface. */
2063 static void drain_all_stock_sync(void)
2065 /* called when force_empty is called */
2066 mutex_lock(&percpu_charge_mutex);
2067 schedule_on_each_cpu(drain_local_stock);
2068 mutex_unlock(&percpu_charge_mutex);
2072 * This function drains percpu counter value from DEAD cpu and
2073 * move it to local cpu. Note that this function can be preempted.
2075 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2079 spin_lock(&mem->pcp_counter_lock);
2080 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2081 long x = per_cpu(mem->stat->count[i], cpu);
2083 per_cpu(mem->stat->count[i], cpu) = 0;
2084 mem->nocpu_base.count[i] += x;
2086 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2087 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2089 per_cpu(mem->stat->events[i], cpu) = 0;
2090 mem->nocpu_base.events[i] += x;
2092 /* need to clear ON_MOVE value, works as a kind of lock. */
2093 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2094 spin_unlock(&mem->pcp_counter_lock);
2097 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2099 int idx = MEM_CGROUP_ON_MOVE;
2101 spin_lock(&mem->pcp_counter_lock);
2102 per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2103 spin_unlock(&mem->pcp_counter_lock);
2106 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2107 unsigned long action,
2110 int cpu = (unsigned long)hcpu;
2111 struct memcg_stock_pcp *stock;
2112 struct mem_cgroup *iter;
2114 if ((action == CPU_ONLINE)) {
2115 for_each_mem_cgroup_all(iter)
2116 synchronize_mem_cgroup_on_move(iter, cpu);
2120 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2123 for_each_mem_cgroup_all(iter)
2124 mem_cgroup_drain_pcp_counter(iter, cpu);
2126 stock = &per_cpu(memcg_stock, cpu);
2132 /* See __mem_cgroup_try_charge() for details */
2134 CHARGE_OK, /* success */
2135 CHARGE_RETRY, /* need to retry but retry is not bad */
2136 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2137 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2138 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2141 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2142 unsigned int nr_pages, bool oom_check)
2144 unsigned long csize = nr_pages * PAGE_SIZE;
2145 struct mem_cgroup *mem_over_limit;
2146 struct res_counter *fail_res;
2147 unsigned long flags = 0;
2150 ret = res_counter_charge(&mem->res, csize, &fail_res);
2153 if (!do_swap_account)
2155 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2159 res_counter_uncharge(&mem->res, csize);
2160 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2161 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2163 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2165 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2166 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2168 * Never reclaim on behalf of optional batching, retry with a
2169 * single page instead.
2171 if (nr_pages == CHARGE_BATCH)
2172 return CHARGE_RETRY;
2174 if (!(gfp_mask & __GFP_WAIT))
2175 return CHARGE_WOULDBLOCK;
2177 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2178 gfp_mask, flags, NULL);
2179 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2180 return CHARGE_RETRY;
2182 * Even though the limit is exceeded at this point, reclaim
2183 * may have been able to free some pages. Retry the charge
2184 * before killing the task.
2186 * Only for regular pages, though: huge pages are rather
2187 * unlikely to succeed so close to the limit, and we fall back
2188 * to regular pages anyway in case of failure.
2190 if (nr_pages == 1 && ret)
2191 return CHARGE_RETRY;
2194 * At task move, charge accounts can be doubly counted. So, it's
2195 * better to wait until the end of task_move if something is going on.
2197 if (mem_cgroup_wait_acct_move(mem_over_limit))
2198 return CHARGE_RETRY;
2200 /* If we don't need to call oom-killer at el, return immediately */
2202 return CHARGE_NOMEM;
2204 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2205 return CHARGE_OOM_DIE;
2207 return CHARGE_RETRY;
2211 * Unlike exported interface, "oom" parameter is added. if oom==true,
2212 * oom-killer can be invoked.
2214 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2216 unsigned int nr_pages,
2217 struct mem_cgroup **memcg,
2220 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2221 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2222 struct mem_cgroup *mem = NULL;
2226 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2227 * in system level. So, allow to go ahead dying process in addition to
2230 if (unlikely(test_thread_flag(TIF_MEMDIE)
2231 || fatal_signal_pending(current)))
2235 * We always charge the cgroup the mm_struct belongs to.
2236 * The mm_struct's mem_cgroup changes on task migration if the
2237 * thread group leader migrates. It's possible that mm is not
2238 * set, if so charge the init_mm (happens for pagecache usage).
2243 if (*memcg) { /* css should be a valid one */
2245 VM_BUG_ON(css_is_removed(&mem->css));
2246 if (mem_cgroup_is_root(mem))
2248 if (nr_pages == 1 && consume_stock(mem))
2252 struct task_struct *p;
2255 p = rcu_dereference(mm->owner);
2257 * Because we don't have task_lock(), "p" can exit.
2258 * In that case, "mem" can point to root or p can be NULL with
2259 * race with swapoff. Then, we have small risk of mis-accouning.
2260 * But such kind of mis-account by race always happens because
2261 * we don't have cgroup_mutex(). It's overkill and we allo that
2263 * (*) swapoff at el will charge against mm-struct not against
2264 * task-struct. So, mm->owner can be NULL.
2266 mem = mem_cgroup_from_task(p);
2267 if (!mem || mem_cgroup_is_root(mem)) {
2271 if (nr_pages == 1 && consume_stock(mem)) {
2273 * It seems dagerous to access memcg without css_get().
2274 * But considering how consume_stok works, it's not
2275 * necessary. If consume_stock success, some charges
2276 * from this memcg are cached on this cpu. So, we
2277 * don't need to call css_get()/css_tryget() before
2278 * calling consume_stock().
2283 /* after here, we may be blocked. we need to get refcnt */
2284 if (!css_tryget(&mem->css)) {
2294 /* If killed, bypass charge */
2295 if (fatal_signal_pending(current)) {
2301 if (oom && !nr_oom_retries) {
2303 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2306 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2310 case CHARGE_RETRY: /* not in OOM situation but retry */
2315 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2318 case CHARGE_NOMEM: /* OOM routine works */
2323 /* If oom, we never return -ENOMEM */
2326 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2330 } while (ret != CHARGE_OK);
2332 if (batch > nr_pages)
2333 refill_stock(mem, batch - nr_pages);
2347 * Somemtimes we have to undo a charge we got by try_charge().
2348 * This function is for that and do uncharge, put css's refcnt.
2349 * gotten by try_charge().
2351 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2352 unsigned int nr_pages)
2354 if (!mem_cgroup_is_root(mem)) {
2355 unsigned long bytes = nr_pages * PAGE_SIZE;
2357 res_counter_uncharge(&mem->res, bytes);
2358 if (do_swap_account)
2359 res_counter_uncharge(&mem->memsw, bytes);
2364 * A helper function to get mem_cgroup from ID. must be called under
2365 * rcu_read_lock(). The caller must check css_is_removed() or some if
2366 * it's concern. (dropping refcnt from swap can be called against removed
2369 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2371 struct cgroup_subsys_state *css;
2373 /* ID 0 is unused ID */
2376 css = css_lookup(&mem_cgroup_subsys, id);
2379 return container_of(css, struct mem_cgroup, css);
2382 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2384 struct mem_cgroup *mem = NULL;
2385 struct page_cgroup *pc;
2389 VM_BUG_ON(!PageLocked(page));
2391 pc = lookup_page_cgroup(page);
2392 lock_page_cgroup(pc);
2393 if (PageCgroupUsed(pc)) {
2394 mem = pc->mem_cgroup;
2395 if (mem && !css_tryget(&mem->css))
2397 } else if (PageSwapCache(page)) {
2398 ent.val = page_private(page);
2399 id = lookup_swap_cgroup(ent);
2401 mem = mem_cgroup_lookup(id);
2402 if (mem && !css_tryget(&mem->css))
2406 unlock_page_cgroup(pc);
2410 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2412 unsigned int nr_pages,
2413 struct page_cgroup *pc,
2414 enum charge_type ctype)
2416 lock_page_cgroup(pc);
2417 if (unlikely(PageCgroupUsed(pc))) {
2418 unlock_page_cgroup(pc);
2419 __mem_cgroup_cancel_charge(mem, nr_pages);
2423 * we don't need page_cgroup_lock about tail pages, becase they are not
2424 * accessed by any other context at this point.
2426 pc->mem_cgroup = mem;
2428 * We access a page_cgroup asynchronously without lock_page_cgroup().
2429 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2430 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2431 * before USED bit, we need memory barrier here.
2432 * See mem_cgroup_add_lru_list(), etc.
2436 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2437 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2438 SetPageCgroupCache(pc);
2439 SetPageCgroupUsed(pc);
2441 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2442 ClearPageCgroupCache(pc);
2443 SetPageCgroupUsed(pc);
2449 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2450 unlock_page_cgroup(pc);
2452 * "charge_statistics" updated event counter. Then, check it.
2453 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2454 * if they exceeds softlimit.
2456 memcg_check_events(mem, page);
2459 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2461 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2462 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2464 * Because tail pages are not marked as "used", set it. We're under
2465 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2467 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2469 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2470 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2471 unsigned long flags;
2473 if (mem_cgroup_disabled())
2476 * We have no races with charge/uncharge but will have races with
2477 * page state accounting.
2479 move_lock_page_cgroup(head_pc, &flags);
2481 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2482 smp_wmb(); /* see __commit_charge() */
2483 if (PageCgroupAcctLRU(head_pc)) {
2485 struct mem_cgroup_per_zone *mz;
2488 * LRU flags cannot be copied because we need to add tail
2489 *.page to LRU by generic call and our hook will be called.
2490 * We hold lru_lock, then, reduce counter directly.
2492 lru = page_lru(head);
2493 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2494 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2496 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2497 move_unlock_page_cgroup(head_pc, &flags);
2502 * mem_cgroup_move_account - move account of the page
2504 * @nr_pages: number of regular pages (>1 for huge pages)
2505 * @pc: page_cgroup of the page.
2506 * @from: mem_cgroup which the page is moved from.
2507 * @to: mem_cgroup which the page is moved to. @from != @to.
2508 * @uncharge: whether we should call uncharge and css_put against @from.
2510 * The caller must confirm following.
2511 * - page is not on LRU (isolate_page() is useful.)
2512 * - compound_lock is held when nr_pages > 1
2514 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2515 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2516 * true, this function does "uncharge" from old cgroup, but it doesn't if
2517 * @uncharge is false, so a caller should do "uncharge".
2519 static int mem_cgroup_move_account(struct page *page,
2520 unsigned int nr_pages,
2521 struct page_cgroup *pc,
2522 struct mem_cgroup *from,
2523 struct mem_cgroup *to,
2526 unsigned long flags;
2529 VM_BUG_ON(from == to);
2530 VM_BUG_ON(PageLRU(page));
2532 * The page is isolated from LRU. So, collapse function
2533 * will not handle this page. But page splitting can happen.
2534 * Do this check under compound_page_lock(). The caller should
2538 if (nr_pages > 1 && !PageTransHuge(page))
2541 lock_page_cgroup(pc);
2544 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2547 move_lock_page_cgroup(pc, &flags);
2549 if (PageCgroupFileMapped(pc)) {
2550 /* Update mapped_file data for mem_cgroup */
2552 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2553 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2556 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2558 /* This is not "cancel", but cancel_charge does all we need. */
2559 __mem_cgroup_cancel_charge(from, nr_pages);
2561 /* caller should have done css_get */
2562 pc->mem_cgroup = to;
2563 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2565 * We charges against "to" which may not have any tasks. Then, "to"
2566 * can be under rmdir(). But in current implementation, caller of
2567 * this function is just force_empty() and move charge, so it's
2568 * guaranteed that "to" is never removed. So, we don't check rmdir
2571 move_unlock_page_cgroup(pc, &flags);
2574 unlock_page_cgroup(pc);
2578 memcg_check_events(to, page);
2579 memcg_check_events(from, page);
2585 * move charges to its parent.
2588 static int mem_cgroup_move_parent(struct page *page,
2589 struct page_cgroup *pc,
2590 struct mem_cgroup *child,
2593 struct cgroup *cg = child->css.cgroup;
2594 struct cgroup *pcg = cg->parent;
2595 struct mem_cgroup *parent;
2596 unsigned int nr_pages;
2597 unsigned long uninitialized_var(flags);
2605 if (!get_page_unless_zero(page))
2607 if (isolate_lru_page(page))
2610 nr_pages = hpage_nr_pages(page);
2612 parent = mem_cgroup_from_cont(pcg);
2613 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2618 flags = compound_lock_irqsave(page);
2620 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2622 __mem_cgroup_cancel_charge(parent, nr_pages);
2625 compound_unlock_irqrestore(page, flags);
2627 putback_lru_page(page);
2635 * Charge the memory controller for page usage.
2637 * 0 if the charge was successful
2638 * < 0 if the cgroup is over its limit
2640 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2641 gfp_t gfp_mask, enum charge_type ctype)
2643 struct mem_cgroup *mem = NULL;
2644 unsigned int nr_pages = 1;
2645 struct page_cgroup *pc;
2649 if (PageTransHuge(page)) {
2650 nr_pages <<= compound_order(page);
2651 VM_BUG_ON(!PageTransHuge(page));
2653 * Never OOM-kill a process for a huge page. The
2654 * fault handler will fall back to regular pages.
2659 pc = lookup_page_cgroup(page);
2660 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2662 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2666 __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2670 int mem_cgroup_newpage_charge(struct page *page,
2671 struct mm_struct *mm, gfp_t gfp_mask)
2673 if (mem_cgroup_disabled())
2676 * If already mapped, we don't have to account.
2677 * If page cache, page->mapping has address_space.
2678 * But page->mapping may have out-of-use anon_vma pointer,
2679 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2682 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2686 return mem_cgroup_charge_common(page, mm, gfp_mask,
2687 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2691 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2692 enum charge_type ctype);
2695 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2696 enum charge_type ctype)
2698 struct page_cgroup *pc = lookup_page_cgroup(page);
2700 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2701 * is already on LRU. It means the page may on some other page_cgroup's
2702 * LRU. Take care of it.
2704 mem_cgroup_lru_del_before_commit(page);
2705 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2706 mem_cgroup_lru_add_after_commit(page);
2710 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2713 struct mem_cgroup *mem = NULL;
2716 if (mem_cgroup_disabled())
2718 if (PageCompound(page))
2721 * Corner case handling. This is called from add_to_page_cache()
2722 * in usual. But some FS (shmem) precharges this page before calling it
2723 * and call add_to_page_cache() with GFP_NOWAIT.
2725 * For GFP_NOWAIT case, the page may be pre-charged before calling
2726 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2727 * charge twice. (It works but has to pay a bit larger cost.)
2728 * And when the page is SwapCache, it should take swap information
2729 * into account. This is under lock_page() now.
2731 if (!(gfp_mask & __GFP_WAIT)) {
2732 struct page_cgroup *pc;
2734 pc = lookup_page_cgroup(page);
2737 lock_page_cgroup(pc);
2738 if (PageCgroupUsed(pc)) {
2739 unlock_page_cgroup(pc);
2742 unlock_page_cgroup(pc);
2748 if (page_is_file_cache(page)) {
2749 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2754 * FUSE reuses pages without going through the final
2755 * put that would remove them from the LRU list, make
2756 * sure that they get relinked properly.
2758 __mem_cgroup_commit_charge_lrucare(page, mem,
2759 MEM_CGROUP_CHARGE_TYPE_CACHE);
2763 if (PageSwapCache(page)) {
2764 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2766 __mem_cgroup_commit_charge_swapin(page, mem,
2767 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2769 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2770 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2776 * While swap-in, try_charge -> commit or cancel, the page is locked.
2777 * And when try_charge() successfully returns, one refcnt to memcg without
2778 * struct page_cgroup is acquired. This refcnt will be consumed by
2779 * "commit()" or removed by "cancel()"
2781 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2783 gfp_t mask, struct mem_cgroup **ptr)
2785 struct mem_cgroup *mem;
2790 if (mem_cgroup_disabled())
2793 if (!do_swap_account)
2796 * A racing thread's fault, or swapoff, may have already updated
2797 * the pte, and even removed page from swap cache: in those cases
2798 * do_swap_page()'s pte_same() test will fail; but there's also a
2799 * KSM case which does need to charge the page.
2801 if (!PageSwapCache(page))
2803 mem = try_get_mem_cgroup_from_page(page);
2807 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2813 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2817 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2818 enum charge_type ctype)
2820 if (mem_cgroup_disabled())
2824 cgroup_exclude_rmdir(&ptr->css);
2826 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2828 * Now swap is on-memory. This means this page may be
2829 * counted both as mem and swap....double count.
2830 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2831 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2832 * may call delete_from_swap_cache() before reach here.
2834 if (do_swap_account && PageSwapCache(page)) {
2835 swp_entry_t ent = {.val = page_private(page)};
2837 struct mem_cgroup *memcg;
2839 id = swap_cgroup_record(ent, 0);
2841 memcg = mem_cgroup_lookup(id);
2844 * This recorded memcg can be obsolete one. So, avoid
2845 * calling css_tryget
2847 if (!mem_cgroup_is_root(memcg))
2848 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2849 mem_cgroup_swap_statistics(memcg, false);
2850 mem_cgroup_put(memcg);
2855 * At swapin, we may charge account against cgroup which has no tasks.
2856 * So, rmdir()->pre_destroy() can be called while we do this charge.
2857 * In that case, we need to call pre_destroy() again. check it here.
2859 cgroup_release_and_wakeup_rmdir(&ptr->css);
2862 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2864 __mem_cgroup_commit_charge_swapin(page, ptr,
2865 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2868 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2870 if (mem_cgroup_disabled())
2874 __mem_cgroup_cancel_charge(mem, 1);
2877 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
2878 unsigned int nr_pages,
2879 const enum charge_type ctype)
2881 struct memcg_batch_info *batch = NULL;
2882 bool uncharge_memsw = true;
2884 /* If swapout, usage of swap doesn't decrease */
2885 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2886 uncharge_memsw = false;
2888 batch = ¤t->memcg_batch;
2890 * In usual, we do css_get() when we remember memcg pointer.
2891 * But in this case, we keep res->usage until end of a series of
2892 * uncharges. Then, it's ok to ignore memcg's refcnt.
2897 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2898 * In those cases, all pages freed continuously can be expected to be in
2899 * the same cgroup and we have chance to coalesce uncharges.
2900 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2901 * because we want to do uncharge as soon as possible.
2904 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2905 goto direct_uncharge;
2908 goto direct_uncharge;
2911 * In typical case, batch->memcg == mem. This means we can
2912 * merge a series of uncharges to an uncharge of res_counter.
2913 * If not, we uncharge res_counter ony by one.
2915 if (batch->memcg != mem)
2916 goto direct_uncharge;
2917 /* remember freed charge and uncharge it later */
2920 batch->memsw_nr_pages++;
2923 res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
2925 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
2926 if (unlikely(batch->memcg != mem))
2927 memcg_oom_recover(mem);
2932 * uncharge if !page_mapped(page)
2934 static struct mem_cgroup *
2935 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2937 struct mem_cgroup *mem = NULL;
2938 unsigned int nr_pages = 1;
2939 struct page_cgroup *pc;
2941 if (mem_cgroup_disabled())
2944 if (PageSwapCache(page))
2947 if (PageTransHuge(page)) {
2948 nr_pages <<= compound_order(page);
2949 VM_BUG_ON(!PageTransHuge(page));
2952 * Check if our page_cgroup is valid
2954 pc = lookup_page_cgroup(page);
2955 if (unlikely(!pc || !PageCgroupUsed(pc)))
2958 lock_page_cgroup(pc);
2960 mem = pc->mem_cgroup;
2962 if (!PageCgroupUsed(pc))
2966 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2967 case MEM_CGROUP_CHARGE_TYPE_DROP:
2968 /* See mem_cgroup_prepare_migration() */
2969 if (page_mapped(page) || PageCgroupMigration(pc))
2972 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2973 if (!PageAnon(page)) { /* Shared memory */
2974 if (page->mapping && !page_is_file_cache(page))
2976 } else if (page_mapped(page)) /* Anon */
2983 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
2985 ClearPageCgroupUsed(pc);
2987 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2988 * freed from LRU. This is safe because uncharged page is expected not
2989 * to be reused (freed soon). Exception is SwapCache, it's handled by
2990 * special functions.
2993 unlock_page_cgroup(pc);
2995 * even after unlock, we have mem->res.usage here and this memcg
2996 * will never be freed.
2998 memcg_check_events(mem, page);
2999 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3000 mem_cgroup_swap_statistics(mem, true);
3001 mem_cgroup_get(mem);
3003 if (!mem_cgroup_is_root(mem))
3004 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
3009 unlock_page_cgroup(pc);
3013 void mem_cgroup_uncharge_page(struct page *page)
3016 if (page_mapped(page))
3018 if (page->mapping && !PageAnon(page))
3020 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3023 void mem_cgroup_uncharge_cache_page(struct page *page)
3025 VM_BUG_ON(page_mapped(page));
3026 VM_BUG_ON(page->mapping);
3027 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3031 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3032 * In that cases, pages are freed continuously and we can expect pages
3033 * are in the same memcg. All these calls itself limits the number of
3034 * pages freed at once, then uncharge_start/end() is called properly.
3035 * This may be called prural(2) times in a context,
3038 void mem_cgroup_uncharge_start(void)
3040 current->memcg_batch.do_batch++;
3041 /* We can do nest. */
3042 if (current->memcg_batch.do_batch == 1) {
3043 current->memcg_batch.memcg = NULL;
3044 current->memcg_batch.nr_pages = 0;
3045 current->memcg_batch.memsw_nr_pages = 0;
3049 void mem_cgroup_uncharge_end(void)
3051 struct memcg_batch_info *batch = ¤t->memcg_batch;
3053 if (!batch->do_batch)
3057 if (batch->do_batch) /* If stacked, do nothing. */
3063 * This "batch->memcg" is valid without any css_get/put etc...
3064 * bacause we hide charges behind us.
3066 if (batch->nr_pages)
3067 res_counter_uncharge(&batch->memcg->res,
3068 batch->nr_pages * PAGE_SIZE);
3069 if (batch->memsw_nr_pages)
3070 res_counter_uncharge(&batch->memcg->memsw,
3071 batch->memsw_nr_pages * PAGE_SIZE);
3072 memcg_oom_recover(batch->memcg);
3073 /* forget this pointer (for sanity check) */
3074 batch->memcg = NULL;
3079 * called after __delete_from_swap_cache() and drop "page" account.
3080 * memcg information is recorded to swap_cgroup of "ent"
3083 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3085 struct mem_cgroup *memcg;
3086 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3088 if (!swapout) /* this was a swap cache but the swap is unused ! */
3089 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3091 memcg = __mem_cgroup_uncharge_common(page, ctype);
3094 * record memcg information, if swapout && memcg != NULL,
3095 * mem_cgroup_get() was called in uncharge().
3097 if (do_swap_account && swapout && memcg)
3098 swap_cgroup_record(ent, css_id(&memcg->css));
3102 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3104 * called from swap_entry_free(). remove record in swap_cgroup and
3105 * uncharge "memsw" account.
3107 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3109 struct mem_cgroup *memcg;
3112 if (!do_swap_account)
3115 id = swap_cgroup_record(ent, 0);
3117 memcg = mem_cgroup_lookup(id);
3120 * We uncharge this because swap is freed.
3121 * This memcg can be obsolete one. We avoid calling css_tryget
3123 if (!mem_cgroup_is_root(memcg))
3124 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3125 mem_cgroup_swap_statistics(memcg, false);
3126 mem_cgroup_put(memcg);
3132 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3133 * @entry: swap entry to be moved
3134 * @from: mem_cgroup which the entry is moved from
3135 * @to: mem_cgroup which the entry is moved to
3136 * @need_fixup: whether we should fixup res_counters and refcounts.
3138 * It succeeds only when the swap_cgroup's record for this entry is the same
3139 * as the mem_cgroup's id of @from.
3141 * Returns 0 on success, -EINVAL on failure.
3143 * The caller must have charged to @to, IOW, called res_counter_charge() about
3144 * both res and memsw, and called css_get().
3146 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3147 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3149 unsigned short old_id, new_id;
3151 old_id = css_id(&from->css);
3152 new_id = css_id(&to->css);
3154 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3155 mem_cgroup_swap_statistics(from, false);
3156 mem_cgroup_swap_statistics(to, true);
3158 * This function is only called from task migration context now.
3159 * It postpones res_counter and refcount handling till the end
3160 * of task migration(mem_cgroup_clear_mc()) for performance
3161 * improvement. But we cannot postpone mem_cgroup_get(to)
3162 * because if the process that has been moved to @to does
3163 * swap-in, the refcount of @to might be decreased to 0.
3167 if (!mem_cgroup_is_root(from))
3168 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3169 mem_cgroup_put(from);
3171 * we charged both to->res and to->memsw, so we should
3174 if (!mem_cgroup_is_root(to))
3175 res_counter_uncharge(&to->res, PAGE_SIZE);
3182 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3183 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3190 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3193 int mem_cgroup_prepare_migration(struct page *page,
3194 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3196 struct mem_cgroup *mem = NULL;
3197 struct page_cgroup *pc;
3198 enum charge_type ctype;
3203 VM_BUG_ON(PageTransHuge(page));
3204 if (mem_cgroup_disabled())
3207 pc = lookup_page_cgroup(page);
3208 lock_page_cgroup(pc);
3209 if (PageCgroupUsed(pc)) {
3210 mem = pc->mem_cgroup;
3213 * At migrating an anonymous page, its mapcount goes down
3214 * to 0 and uncharge() will be called. But, even if it's fully
3215 * unmapped, migration may fail and this page has to be
3216 * charged again. We set MIGRATION flag here and delay uncharge
3217 * until end_migration() is called
3219 * Corner Case Thinking
3221 * When the old page was mapped as Anon and it's unmap-and-freed
3222 * while migration was ongoing.
3223 * If unmap finds the old page, uncharge() of it will be delayed
3224 * until end_migration(). If unmap finds a new page, it's
3225 * uncharged when it make mapcount to be 1->0. If unmap code
3226 * finds swap_migration_entry, the new page will not be mapped
3227 * and end_migration() will find it(mapcount==0).
3230 * When the old page was mapped but migraion fails, the kernel
3231 * remaps it. A charge for it is kept by MIGRATION flag even
3232 * if mapcount goes down to 0. We can do remap successfully
3233 * without charging it again.
3236 * The "old" page is under lock_page() until the end of
3237 * migration, so, the old page itself will not be swapped-out.
3238 * If the new page is swapped out before end_migraton, our
3239 * hook to usual swap-out path will catch the event.
3242 SetPageCgroupMigration(pc);
3244 unlock_page_cgroup(pc);
3246 * If the page is not charged at this point,
3253 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3254 css_put(&mem->css);/* drop extra refcnt */
3255 if (ret || *ptr == NULL) {
3256 if (PageAnon(page)) {
3257 lock_page_cgroup(pc);
3258 ClearPageCgroupMigration(pc);
3259 unlock_page_cgroup(pc);
3261 * The old page may be fully unmapped while we kept it.
3263 mem_cgroup_uncharge_page(page);
3268 * We charge new page before it's used/mapped. So, even if unlock_page()
3269 * is called before end_migration, we can catch all events on this new
3270 * page. In the case new page is migrated but not remapped, new page's
3271 * mapcount will be finally 0 and we call uncharge in end_migration().
3273 pc = lookup_page_cgroup(newpage);
3275 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3276 else if (page_is_file_cache(page))
3277 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3279 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3280 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3284 /* remove redundant charge if migration failed*/
3285 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3286 struct page *oldpage, struct page *newpage, bool migration_ok)
3288 struct page *used, *unused;
3289 struct page_cgroup *pc;
3293 /* blocks rmdir() */
3294 cgroup_exclude_rmdir(&mem->css);
3295 if (!migration_ok) {
3303 * We disallowed uncharge of pages under migration because mapcount
3304 * of the page goes down to zero, temporarly.
3305 * Clear the flag and check the page should be charged.
3307 pc = lookup_page_cgroup(oldpage);
3308 lock_page_cgroup(pc);
3309 ClearPageCgroupMigration(pc);
3310 unlock_page_cgroup(pc);
3312 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3315 * If a page is a file cache, radix-tree replacement is very atomic
3316 * and we can skip this check. When it was an Anon page, its mapcount
3317 * goes down to 0. But because we added MIGRATION flage, it's not
3318 * uncharged yet. There are several case but page->mapcount check
3319 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3320 * check. (see prepare_charge() also)
3323 mem_cgroup_uncharge_page(used);
3325 * At migration, we may charge account against cgroup which has no
3327 * So, rmdir()->pre_destroy() can be called while we do this charge.
3328 * In that case, we need to call pre_destroy() again. check it here.
3330 cgroup_release_and_wakeup_rmdir(&mem->css);
3334 * A call to try to shrink memory usage on charge failure at shmem's swapin.
3335 * Calling hierarchical_reclaim is not enough because we should update
3336 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3337 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3338 * not from the memcg which this page would be charged to.
3339 * try_charge_swapin does all of these works properly.
3341 int mem_cgroup_shmem_charge_fallback(struct page *page,
3342 struct mm_struct *mm,
3345 struct mem_cgroup *mem;
3348 if (mem_cgroup_disabled())
3351 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3353 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3358 #ifdef CONFIG_DEBUG_VM
3359 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3361 struct page_cgroup *pc;
3363 pc = lookup_page_cgroup(page);
3364 if (likely(pc) && PageCgroupUsed(pc))
3369 bool mem_cgroup_bad_page_check(struct page *page)
3371 if (mem_cgroup_disabled())
3374 return lookup_page_cgroup_used(page) != NULL;
3377 void mem_cgroup_print_bad_page(struct page *page)
3379 struct page_cgroup *pc;
3381 pc = lookup_page_cgroup_used(page);
3386 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3387 pc, pc->flags, pc->mem_cgroup);
3389 path = kmalloc(PATH_MAX, GFP_KERNEL);
3392 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3397 printk(KERN_CONT "(%s)\n",
3398 (ret < 0) ? "cannot get the path" : path);
3404 static DEFINE_MUTEX(set_limit_mutex);
3406 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3407 unsigned long long val)
3410 u64 memswlimit, memlimit;
3412 int children = mem_cgroup_count_children(memcg);
3413 u64 curusage, oldusage;
3417 * For keeping hierarchical_reclaim simple, how long we should retry
3418 * is depends on callers. We set our retry-count to be function
3419 * of # of children which we should visit in this loop.
3421 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3423 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3426 while (retry_count) {
3427 if (signal_pending(current)) {
3432 * Rather than hide all in some function, I do this in
3433 * open coded manner. You see what this really does.
3434 * We have to guarantee mem->res.limit < mem->memsw.limit.
3436 mutex_lock(&set_limit_mutex);
3437 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3438 if (memswlimit < val) {
3440 mutex_unlock(&set_limit_mutex);
3444 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3448 ret = res_counter_set_limit(&memcg->res, val);
3450 if (memswlimit == val)
3451 memcg->memsw_is_minimum = true;
3453 memcg->memsw_is_minimum = false;
3455 mutex_unlock(&set_limit_mutex);
3460 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3461 MEM_CGROUP_RECLAIM_SHRINK,
3463 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3464 /* Usage is reduced ? */
3465 if (curusage >= oldusage)
3468 oldusage = curusage;
3470 if (!ret && enlarge)
3471 memcg_oom_recover(memcg);
3476 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3477 unsigned long long val)
3480 u64 memlimit, memswlimit, oldusage, curusage;
3481 int children = mem_cgroup_count_children(memcg);
3485 /* see mem_cgroup_resize_res_limit */
3486 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3487 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3488 while (retry_count) {
3489 if (signal_pending(current)) {
3494 * Rather than hide all in some function, I do this in
3495 * open coded manner. You see what this really does.
3496 * We have to guarantee mem->res.limit < mem->memsw.limit.
3498 mutex_lock(&set_limit_mutex);
3499 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3500 if (memlimit > val) {
3502 mutex_unlock(&set_limit_mutex);
3505 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3506 if (memswlimit < val)
3508 ret = res_counter_set_limit(&memcg->memsw, val);
3510 if (memlimit == val)
3511 memcg->memsw_is_minimum = true;
3513 memcg->memsw_is_minimum = false;
3515 mutex_unlock(&set_limit_mutex);
3520 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3521 MEM_CGROUP_RECLAIM_NOSWAP |
3522 MEM_CGROUP_RECLAIM_SHRINK,
3524 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3525 /* Usage is reduced ? */
3526 if (curusage >= oldusage)
3529 oldusage = curusage;
3531 if (!ret && enlarge)
3532 memcg_oom_recover(memcg);
3536 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3538 unsigned long *total_scanned)
3540 unsigned long nr_reclaimed = 0;
3541 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3542 unsigned long reclaimed;
3544 struct mem_cgroup_tree_per_zone *mctz;
3545 unsigned long long excess;
3546 unsigned long nr_scanned;
3551 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3553 * This loop can run a while, specially if mem_cgroup's continuously
3554 * keep exceeding their soft limit and putting the system under
3561 mz = mem_cgroup_largest_soft_limit_node(mctz);
3566 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3568 MEM_CGROUP_RECLAIM_SOFT,
3570 nr_reclaimed += reclaimed;
3571 *total_scanned += nr_scanned;
3572 spin_lock(&mctz->lock);
3575 * If we failed to reclaim anything from this memory cgroup
3576 * it is time to move on to the next cgroup
3582 * Loop until we find yet another one.
3584 * By the time we get the soft_limit lock
3585 * again, someone might have aded the
3586 * group back on the RB tree. Iterate to
3587 * make sure we get a different mem.
3588 * mem_cgroup_largest_soft_limit_node returns
3589 * NULL if no other cgroup is present on
3593 __mem_cgroup_largest_soft_limit_node(mctz);
3595 css_put(&next_mz->mem->css);
3596 else /* next_mz == NULL or other memcg */
3600 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3601 excess = res_counter_soft_limit_excess(&mz->mem->res);
3603 * One school of thought says that we should not add
3604 * back the node to the tree if reclaim returns 0.
3605 * But our reclaim could return 0, simply because due
3606 * to priority we are exposing a smaller subset of
3607 * memory to reclaim from. Consider this as a longer
3610 /* If excess == 0, no tree ops */
3611 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3612 spin_unlock(&mctz->lock);
3613 css_put(&mz->mem->css);
3616 * Could not reclaim anything and there are no more
3617 * mem cgroups to try or we seem to be looping without
3618 * reclaiming anything.
3620 if (!nr_reclaimed &&
3622 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3624 } while (!nr_reclaimed);
3626 css_put(&next_mz->mem->css);
3627 return nr_reclaimed;
3631 * This routine traverse page_cgroup in given list and drop them all.
3632 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3634 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3635 int node, int zid, enum lru_list lru)
3638 struct mem_cgroup_per_zone *mz;
3639 struct page_cgroup *pc, *busy;
3640 unsigned long flags, loop;
3641 struct list_head *list;
3644 zone = &NODE_DATA(node)->node_zones[zid];
3645 mz = mem_cgroup_zoneinfo(mem, node, zid);
3646 list = &mz->lists[lru];
3648 loop = MEM_CGROUP_ZSTAT(mz, lru);
3649 /* give some margin against EBUSY etc...*/
3656 spin_lock_irqsave(&zone->lru_lock, flags);
3657 if (list_empty(list)) {
3658 spin_unlock_irqrestore(&zone->lru_lock, flags);
3661 pc = list_entry(list->prev, struct page_cgroup, lru);
3663 list_move(&pc->lru, list);
3665 spin_unlock_irqrestore(&zone->lru_lock, flags);
3668 spin_unlock_irqrestore(&zone->lru_lock, flags);
3670 page = lookup_cgroup_page(pc);
3672 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3676 if (ret == -EBUSY || ret == -EINVAL) {
3677 /* found lock contention or "pc" is obsolete. */
3684 if (!ret && !list_empty(list))
3690 * make mem_cgroup's charge to be 0 if there is no task.
3691 * This enables deleting this mem_cgroup.
3693 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3696 int node, zid, shrink;
3697 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3698 struct cgroup *cgrp = mem->css.cgroup;
3703 /* should free all ? */
3709 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3712 if (signal_pending(current))
3714 /* This is for making all *used* pages to be on LRU. */
3715 lru_add_drain_all();
3716 drain_all_stock_sync();
3718 mem_cgroup_start_move(mem);
3719 for_each_node_state(node, N_HIGH_MEMORY) {
3720 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3723 ret = mem_cgroup_force_empty_list(mem,
3732 mem_cgroup_end_move(mem);
3733 memcg_oom_recover(mem);
3734 /* it seems parent cgroup doesn't have enough mem */
3738 /* "ret" should also be checked to ensure all lists are empty. */
3739 } while (mem->res.usage > 0 || ret);
3745 /* returns EBUSY if there is a task or if we come here twice. */
3746 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3750 /* we call try-to-free pages for make this cgroup empty */
3751 lru_add_drain_all();
3752 /* try to free all pages in this cgroup */
3754 while (nr_retries && mem->res.usage > 0) {
3757 if (signal_pending(current)) {
3761 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3762 false, get_swappiness(mem));
3765 /* maybe some writeback is necessary */
3766 congestion_wait(BLK_RW_ASYNC, HZ/10);
3771 /* try move_account...there may be some *locked* pages. */
3775 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3777 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3781 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3783 return mem_cgroup_from_cont(cont)->use_hierarchy;
3786 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3790 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3791 struct cgroup *parent = cont->parent;
3792 struct mem_cgroup *parent_mem = NULL;
3795 parent_mem = mem_cgroup_from_cont(parent);
3799 * If parent's use_hierarchy is set, we can't make any modifications
3800 * in the child subtrees. If it is unset, then the change can
3801 * occur, provided the current cgroup has no children.
3803 * For the root cgroup, parent_mem is NULL, we allow value to be
3804 * set if there are no children.
3806 if ((!parent_mem || !parent_mem->use_hierarchy) &&
3807 (val == 1 || val == 0)) {
3808 if (list_empty(&cont->children))
3809 mem->use_hierarchy = val;
3820 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3821 enum mem_cgroup_stat_index idx)
3823 struct mem_cgroup *iter;
3826 /* Per-cpu values can be negative, use a signed accumulator */
3827 for_each_mem_cgroup_tree(iter, mem)
3828 val += mem_cgroup_read_stat(iter, idx);
3830 if (val < 0) /* race ? */
3835 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3839 if (!mem_cgroup_is_root(mem)) {
3841 return res_counter_read_u64(&mem->res, RES_USAGE);
3843 return res_counter_read_u64(&mem->memsw, RES_USAGE);
3846 val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
3847 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
3850 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3852 return val << PAGE_SHIFT;
3855 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3857 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3861 type = MEMFILE_TYPE(cft->private);
3862 name = MEMFILE_ATTR(cft->private);
3865 if (name == RES_USAGE)
3866 val = mem_cgroup_usage(mem, false);
3868 val = res_counter_read_u64(&mem->res, name);
3871 if (name == RES_USAGE)
3872 val = mem_cgroup_usage(mem, true);
3874 val = res_counter_read_u64(&mem->memsw, name);
3883 * The user of this function is...
3886 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3889 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3891 unsigned long long val;
3894 type = MEMFILE_TYPE(cft->private);
3895 name = MEMFILE_ATTR(cft->private);
3898 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3902 /* This function does all necessary parse...reuse it */
3903 ret = res_counter_memparse_write_strategy(buffer, &val);
3907 ret = mem_cgroup_resize_limit(memcg, val);
3909 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3911 case RES_SOFT_LIMIT:
3912 ret = res_counter_memparse_write_strategy(buffer, &val);
3916 * For memsw, soft limits are hard to implement in terms
3917 * of semantics, for now, we support soft limits for
3918 * control without swap
3921 ret = res_counter_set_soft_limit(&memcg->res, val);
3926 ret = -EINVAL; /* should be BUG() ? */
3932 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3933 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3935 struct cgroup *cgroup;
3936 unsigned long long min_limit, min_memsw_limit, tmp;
3938 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3939 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3940 cgroup = memcg->css.cgroup;
3941 if (!memcg->use_hierarchy)
3944 while (cgroup->parent) {
3945 cgroup = cgroup->parent;
3946 memcg = mem_cgroup_from_cont(cgroup);
3947 if (!memcg->use_hierarchy)
3949 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3950 min_limit = min(min_limit, tmp);
3951 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3952 min_memsw_limit = min(min_memsw_limit, tmp);
3955 *mem_limit = min_limit;
3956 *memsw_limit = min_memsw_limit;
3960 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3962 struct mem_cgroup *mem;
3965 mem = mem_cgroup_from_cont(cont);
3966 type = MEMFILE_TYPE(event);
3967 name = MEMFILE_ATTR(event);
3971 res_counter_reset_max(&mem->res);
3973 res_counter_reset_max(&mem->memsw);
3977 res_counter_reset_failcnt(&mem->res);
3979 res_counter_reset_failcnt(&mem->memsw);
3986 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3989 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3993 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3994 struct cftype *cft, u64 val)
3996 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3998 if (val >= (1 << NR_MOVE_TYPE))
4001 * We check this value several times in both in can_attach() and
4002 * attach(), so we need cgroup lock to prevent this value from being
4006 mem->move_charge_at_immigrate = val;
4012 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4013 struct cftype *cft, u64 val)
4020 /* For read statistics */
4038 struct mcs_total_stat {
4039 s64 stat[NR_MCS_STAT];
4045 } memcg_stat_strings[NR_MCS_STAT] = {
4046 {"cache", "total_cache"},
4047 {"rss", "total_rss"},
4048 {"mapped_file", "total_mapped_file"},
4049 {"pgpgin", "total_pgpgin"},
4050 {"pgpgout", "total_pgpgout"},
4051 {"swap", "total_swap"},
4052 {"pgfault", "total_pgfault"},
4053 {"pgmajfault", "total_pgmajfault"},
4054 {"inactive_anon", "total_inactive_anon"},
4055 {"active_anon", "total_active_anon"},
4056 {"inactive_file", "total_inactive_file"},
4057 {"active_file", "total_active_file"},
4058 {"unevictable", "total_unevictable"}
4063 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4068 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4069 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4070 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4071 s->stat[MCS_RSS] += val * PAGE_SIZE;
4072 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4073 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4074 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4075 s->stat[MCS_PGPGIN] += val;
4076 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4077 s->stat[MCS_PGPGOUT] += val;
4078 if (do_swap_account) {
4079 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4080 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4082 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4083 s->stat[MCS_PGFAULT] += val;
4084 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4085 s->stat[MCS_PGMAJFAULT] += val;
4088 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
4089 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4090 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
4091 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4092 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
4093 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4094 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
4095 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4096 val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
4097 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4101 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4103 struct mem_cgroup *iter;
4105 for_each_mem_cgroup_tree(iter, mem)
4106 mem_cgroup_get_local_stat(iter, s);
4110 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4113 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4114 unsigned long node_nr;
4115 struct cgroup *cont = m->private;
4116 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4118 total_nr = mem_cgroup_nr_lru_pages(mem_cont);
4119 seq_printf(m, "total=%lu", total_nr);
4120 for_each_node_state(nid, N_HIGH_MEMORY) {
4121 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid);
4122 seq_printf(m, " N%d=%lu", nid, node_nr);
4126 file_nr = mem_cgroup_nr_file_lru_pages(mem_cont);
4127 seq_printf(m, "file=%lu", file_nr);
4128 for_each_node_state(nid, N_HIGH_MEMORY) {
4129 node_nr = mem_cgroup_node_nr_file_lru_pages(mem_cont, nid);
4130 seq_printf(m, " N%d=%lu", nid, node_nr);
4134 anon_nr = mem_cgroup_nr_anon_lru_pages(mem_cont);
4135 seq_printf(m, "anon=%lu", anon_nr);
4136 for_each_node_state(nid, N_HIGH_MEMORY) {
4137 node_nr = mem_cgroup_node_nr_anon_lru_pages(mem_cont, nid);
4138 seq_printf(m, " N%d=%lu", nid, node_nr);
4142 unevictable_nr = mem_cgroup_nr_unevictable_lru_pages(mem_cont);
4143 seq_printf(m, "unevictable=%lu", unevictable_nr);
4144 for_each_node_state(nid, N_HIGH_MEMORY) {
4145 node_nr = mem_cgroup_node_nr_unevictable_lru_pages(mem_cont,
4147 seq_printf(m, " N%d=%lu", nid, node_nr);
4152 #endif /* CONFIG_NUMA */
4154 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4155 struct cgroup_map_cb *cb)
4157 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4158 struct mcs_total_stat mystat;
4161 memset(&mystat, 0, sizeof(mystat));
4162 mem_cgroup_get_local_stat(mem_cont, &mystat);
4165 for (i = 0; i < NR_MCS_STAT; i++) {
4166 if (i == MCS_SWAP && !do_swap_account)
4168 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4171 /* Hierarchical information */
4173 unsigned long long limit, memsw_limit;
4174 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4175 cb->fill(cb, "hierarchical_memory_limit", limit);
4176 if (do_swap_account)
4177 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4180 memset(&mystat, 0, sizeof(mystat));
4181 mem_cgroup_get_total_stat(mem_cont, &mystat);
4182 for (i = 0; i < NR_MCS_STAT; i++) {
4183 if (i == MCS_SWAP && !do_swap_account)
4185 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4188 #ifdef CONFIG_DEBUG_VM
4189 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4193 struct mem_cgroup_per_zone *mz;
4194 unsigned long recent_rotated[2] = {0, 0};
4195 unsigned long recent_scanned[2] = {0, 0};
4197 for_each_online_node(nid)
4198 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4199 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4201 recent_rotated[0] +=
4202 mz->reclaim_stat.recent_rotated[0];
4203 recent_rotated[1] +=
4204 mz->reclaim_stat.recent_rotated[1];
4205 recent_scanned[0] +=
4206 mz->reclaim_stat.recent_scanned[0];
4207 recent_scanned[1] +=
4208 mz->reclaim_stat.recent_scanned[1];
4210 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4211 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4212 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4213 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4220 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4222 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4224 return get_swappiness(memcg);
4227 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4230 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4231 struct mem_cgroup *parent;
4236 if (cgrp->parent == NULL)
4239 parent = mem_cgroup_from_cont(cgrp->parent);
4243 /* If under hierarchy, only empty-root can set this value */
4244 if ((parent->use_hierarchy) ||
4245 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4250 memcg->swappiness = val;
4257 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4259 struct mem_cgroup_threshold_ary *t;
4265 t = rcu_dereference(memcg->thresholds.primary);
4267 t = rcu_dereference(memcg->memsw_thresholds.primary);
4272 usage = mem_cgroup_usage(memcg, swap);
4275 * current_threshold points to threshold just below usage.
4276 * If it's not true, a threshold was crossed after last
4277 * call of __mem_cgroup_threshold().
4279 i = t->current_threshold;
4282 * Iterate backward over array of thresholds starting from
4283 * current_threshold and check if a threshold is crossed.
4284 * If none of thresholds below usage is crossed, we read
4285 * only one element of the array here.
4287 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4288 eventfd_signal(t->entries[i].eventfd, 1);
4290 /* i = current_threshold + 1 */
4294 * Iterate forward over array of thresholds starting from
4295 * current_threshold+1 and check if a threshold is crossed.
4296 * If none of thresholds above usage is crossed, we read
4297 * only one element of the array here.
4299 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4300 eventfd_signal(t->entries[i].eventfd, 1);
4302 /* Update current_threshold */
4303 t->current_threshold = i - 1;
4308 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4311 __mem_cgroup_threshold(memcg, false);
4312 if (do_swap_account)
4313 __mem_cgroup_threshold(memcg, true);
4315 memcg = parent_mem_cgroup(memcg);
4319 static int compare_thresholds(const void *a, const void *b)
4321 const struct mem_cgroup_threshold *_a = a;
4322 const struct mem_cgroup_threshold *_b = b;
4324 return _a->threshold - _b->threshold;
4327 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4329 struct mem_cgroup_eventfd_list *ev;
4331 list_for_each_entry(ev, &mem->oom_notify, list)
4332 eventfd_signal(ev->eventfd, 1);
4336 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4338 struct mem_cgroup *iter;
4340 for_each_mem_cgroup_tree(iter, mem)
4341 mem_cgroup_oom_notify_cb(iter);
4344 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4345 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4347 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4348 struct mem_cgroup_thresholds *thresholds;
4349 struct mem_cgroup_threshold_ary *new;
4350 int type = MEMFILE_TYPE(cft->private);
4351 u64 threshold, usage;
4354 ret = res_counter_memparse_write_strategy(args, &threshold);
4358 mutex_lock(&memcg->thresholds_lock);
4361 thresholds = &memcg->thresholds;
4362 else if (type == _MEMSWAP)
4363 thresholds = &memcg->memsw_thresholds;
4367 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4369 /* Check if a threshold crossed before adding a new one */
4370 if (thresholds->primary)
4371 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4373 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4375 /* Allocate memory for new array of thresholds */
4376 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4384 /* Copy thresholds (if any) to new array */
4385 if (thresholds->primary) {
4386 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4387 sizeof(struct mem_cgroup_threshold));
4390 /* Add new threshold */
4391 new->entries[size - 1].eventfd = eventfd;
4392 new->entries[size - 1].threshold = threshold;
4394 /* Sort thresholds. Registering of new threshold isn't time-critical */
4395 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4396 compare_thresholds, NULL);
4398 /* Find current threshold */
4399 new->current_threshold = -1;
4400 for (i = 0; i < size; i++) {
4401 if (new->entries[i].threshold < usage) {
4403 * new->current_threshold will not be used until
4404 * rcu_assign_pointer(), so it's safe to increment
4407 ++new->current_threshold;
4411 /* Free old spare buffer and save old primary buffer as spare */
4412 kfree(thresholds->spare);
4413 thresholds->spare = thresholds->primary;
4415 rcu_assign_pointer(thresholds->primary, new);
4417 /* To be sure that nobody uses thresholds */
4421 mutex_unlock(&memcg->thresholds_lock);
4426 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4427 struct cftype *cft, struct eventfd_ctx *eventfd)
4429 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4430 struct mem_cgroup_thresholds *thresholds;
4431 struct mem_cgroup_threshold_ary *new;
4432 int type = MEMFILE_TYPE(cft->private);
4436 mutex_lock(&memcg->thresholds_lock);
4438 thresholds = &memcg->thresholds;
4439 else if (type == _MEMSWAP)
4440 thresholds = &memcg->memsw_thresholds;
4445 * Something went wrong if we trying to unregister a threshold
4446 * if we don't have thresholds
4448 BUG_ON(!thresholds);
4450 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4452 /* Check if a threshold crossed before removing */
4453 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4455 /* Calculate new number of threshold */
4457 for (i = 0; i < thresholds->primary->size; i++) {
4458 if (thresholds->primary->entries[i].eventfd != eventfd)
4462 new = thresholds->spare;
4464 /* Set thresholds array to NULL if we don't have thresholds */
4473 /* Copy thresholds and find current threshold */
4474 new->current_threshold = -1;
4475 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4476 if (thresholds->primary->entries[i].eventfd == eventfd)
4479 new->entries[j] = thresholds->primary->entries[i];
4480 if (new->entries[j].threshold < usage) {
4482 * new->current_threshold will not be used
4483 * until rcu_assign_pointer(), so it's safe to increment
4486 ++new->current_threshold;
4492 /* Swap primary and spare array */
4493 thresholds->spare = thresholds->primary;
4494 rcu_assign_pointer(thresholds->primary, new);
4496 /* To be sure that nobody uses thresholds */
4499 mutex_unlock(&memcg->thresholds_lock);
4502 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4503 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4505 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4506 struct mem_cgroup_eventfd_list *event;
4507 int type = MEMFILE_TYPE(cft->private);
4509 BUG_ON(type != _OOM_TYPE);
4510 event = kmalloc(sizeof(*event), GFP_KERNEL);
4514 mutex_lock(&memcg_oom_mutex);
4516 event->eventfd = eventfd;
4517 list_add(&event->list, &memcg->oom_notify);
4519 /* already in OOM ? */
4520 if (atomic_read(&memcg->oom_lock))
4521 eventfd_signal(eventfd, 1);
4522 mutex_unlock(&memcg_oom_mutex);
4527 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4528 struct cftype *cft, struct eventfd_ctx *eventfd)
4530 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4531 struct mem_cgroup_eventfd_list *ev, *tmp;
4532 int type = MEMFILE_TYPE(cft->private);
4534 BUG_ON(type != _OOM_TYPE);
4536 mutex_lock(&memcg_oom_mutex);
4538 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4539 if (ev->eventfd == eventfd) {
4540 list_del(&ev->list);
4545 mutex_unlock(&memcg_oom_mutex);
4548 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4549 struct cftype *cft, struct cgroup_map_cb *cb)
4551 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4553 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4555 if (atomic_read(&mem->oom_lock))
4556 cb->fill(cb, "under_oom", 1);
4558 cb->fill(cb, "under_oom", 0);
4562 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4563 struct cftype *cft, u64 val)
4565 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4566 struct mem_cgroup *parent;
4568 /* cannot set to root cgroup and only 0 and 1 are allowed */
4569 if (!cgrp->parent || !((val == 0) || (val == 1)))
4572 parent = mem_cgroup_from_cont(cgrp->parent);
4575 /* oom-kill-disable is a flag for subhierarchy. */
4576 if ((parent->use_hierarchy) ||
4577 (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4581 mem->oom_kill_disable = val;
4583 memcg_oom_recover(mem);
4589 static const struct file_operations mem_control_numa_stat_file_operations = {
4591 .llseek = seq_lseek,
4592 .release = single_release,
4595 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4597 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4599 file->f_op = &mem_control_numa_stat_file_operations;
4600 return single_open(file, mem_control_numa_stat_show, cont);
4602 #endif /* CONFIG_NUMA */
4604 static struct cftype mem_cgroup_files[] = {
4606 .name = "usage_in_bytes",
4607 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4608 .read_u64 = mem_cgroup_read,
4609 .register_event = mem_cgroup_usage_register_event,
4610 .unregister_event = mem_cgroup_usage_unregister_event,
4613 .name = "max_usage_in_bytes",
4614 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4615 .trigger = mem_cgroup_reset,
4616 .read_u64 = mem_cgroup_read,
4619 .name = "limit_in_bytes",
4620 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4621 .write_string = mem_cgroup_write,
4622 .read_u64 = mem_cgroup_read,
4625 .name = "soft_limit_in_bytes",
4626 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4627 .write_string = mem_cgroup_write,
4628 .read_u64 = mem_cgroup_read,
4632 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4633 .trigger = mem_cgroup_reset,
4634 .read_u64 = mem_cgroup_read,
4638 .read_map = mem_control_stat_show,
4641 .name = "force_empty",
4642 .trigger = mem_cgroup_force_empty_write,
4645 .name = "use_hierarchy",
4646 .write_u64 = mem_cgroup_hierarchy_write,
4647 .read_u64 = mem_cgroup_hierarchy_read,
4650 .name = "swappiness",
4651 .read_u64 = mem_cgroup_swappiness_read,
4652 .write_u64 = mem_cgroup_swappiness_write,
4655 .name = "move_charge_at_immigrate",
4656 .read_u64 = mem_cgroup_move_charge_read,
4657 .write_u64 = mem_cgroup_move_charge_write,
4660 .name = "oom_control",
4661 .read_map = mem_cgroup_oom_control_read,
4662 .write_u64 = mem_cgroup_oom_control_write,
4663 .register_event = mem_cgroup_oom_register_event,
4664 .unregister_event = mem_cgroup_oom_unregister_event,
4665 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4669 .name = "numa_stat",
4670 .open = mem_control_numa_stat_open,
4676 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4677 static struct cftype memsw_cgroup_files[] = {
4679 .name = "memsw.usage_in_bytes",
4680 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4681 .read_u64 = mem_cgroup_read,
4682 .register_event = mem_cgroup_usage_register_event,
4683 .unregister_event = mem_cgroup_usage_unregister_event,
4686 .name = "memsw.max_usage_in_bytes",
4687 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4688 .trigger = mem_cgroup_reset,
4689 .read_u64 = mem_cgroup_read,
4692 .name = "memsw.limit_in_bytes",
4693 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4694 .write_string = mem_cgroup_write,
4695 .read_u64 = mem_cgroup_read,
4698 .name = "memsw.failcnt",
4699 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4700 .trigger = mem_cgroup_reset,
4701 .read_u64 = mem_cgroup_read,
4705 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4707 if (!do_swap_account)
4709 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4710 ARRAY_SIZE(memsw_cgroup_files));
4713 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4719 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4721 struct mem_cgroup_per_node *pn;
4722 struct mem_cgroup_per_zone *mz;
4724 int zone, tmp = node;
4726 * This routine is called against possible nodes.
4727 * But it's BUG to call kmalloc() against offline node.
4729 * TODO: this routine can waste much memory for nodes which will
4730 * never be onlined. It's better to use memory hotplug callback
4733 if (!node_state(node, N_NORMAL_MEMORY))
4735 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4739 mem->info.nodeinfo[node] = pn;
4740 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4741 mz = &pn->zoneinfo[zone];
4743 INIT_LIST_HEAD(&mz->lists[l]);
4744 mz->usage_in_excess = 0;
4745 mz->on_tree = false;
4751 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4753 kfree(mem->info.nodeinfo[node]);
4756 static struct mem_cgroup *mem_cgroup_alloc(void)
4758 struct mem_cgroup *mem;
4759 int size = sizeof(struct mem_cgroup);
4761 /* Can be very big if MAX_NUMNODES is very big */
4762 if (size < PAGE_SIZE)
4763 mem = kzalloc(size, GFP_KERNEL);
4765 mem = vzalloc(size);
4770 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4773 spin_lock_init(&mem->pcp_counter_lock);
4777 if (size < PAGE_SIZE)
4785 * At destroying mem_cgroup, references from swap_cgroup can remain.
4786 * (scanning all at force_empty is too costly...)
4788 * Instead of clearing all references at force_empty, we remember
4789 * the number of reference from swap_cgroup and free mem_cgroup when
4790 * it goes down to 0.
4792 * Removal of cgroup itself succeeds regardless of refs from swap.
4795 static void __mem_cgroup_free(struct mem_cgroup *mem)
4799 mem_cgroup_remove_from_trees(mem);
4800 free_css_id(&mem_cgroup_subsys, &mem->css);
4802 for_each_node_state(node, N_POSSIBLE)
4803 free_mem_cgroup_per_zone_info(mem, node);
4805 free_percpu(mem->stat);
4806 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4812 static void mem_cgroup_get(struct mem_cgroup *mem)
4814 atomic_inc(&mem->refcnt);
4817 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4819 if (atomic_sub_and_test(count, &mem->refcnt)) {
4820 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4821 __mem_cgroup_free(mem);
4823 mem_cgroup_put(parent);
4827 static void mem_cgroup_put(struct mem_cgroup *mem)
4829 __mem_cgroup_put(mem, 1);
4833 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4835 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4837 if (!mem->res.parent)
4839 return mem_cgroup_from_res_counter(mem->res.parent, res);
4842 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4843 static void __init enable_swap_cgroup(void)
4845 if (!mem_cgroup_disabled() && really_do_swap_account)
4846 do_swap_account = 1;
4849 static void __init enable_swap_cgroup(void)
4854 static int mem_cgroup_soft_limit_tree_init(void)
4856 struct mem_cgroup_tree_per_node *rtpn;
4857 struct mem_cgroup_tree_per_zone *rtpz;
4858 int tmp, node, zone;
4860 for_each_node_state(node, N_POSSIBLE) {
4862 if (!node_state(node, N_NORMAL_MEMORY))
4864 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4868 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4870 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4871 rtpz = &rtpn->rb_tree_per_zone[zone];
4872 rtpz->rb_root = RB_ROOT;
4873 spin_lock_init(&rtpz->lock);
4879 static struct cgroup_subsys_state * __ref
4880 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4882 struct mem_cgroup *mem, *parent;
4883 long error = -ENOMEM;
4886 mem = mem_cgroup_alloc();
4888 return ERR_PTR(error);
4890 for_each_node_state(node, N_POSSIBLE)
4891 if (alloc_mem_cgroup_per_zone_info(mem, node))
4895 if (cont->parent == NULL) {
4897 enable_swap_cgroup();
4899 root_mem_cgroup = mem;
4900 if (mem_cgroup_soft_limit_tree_init())
4902 for_each_possible_cpu(cpu) {
4903 struct memcg_stock_pcp *stock =
4904 &per_cpu(memcg_stock, cpu);
4905 INIT_WORK(&stock->work, drain_local_stock);
4907 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4909 parent = mem_cgroup_from_cont(cont->parent);
4910 mem->use_hierarchy = parent->use_hierarchy;
4911 mem->oom_kill_disable = parent->oom_kill_disable;
4914 if (parent && parent->use_hierarchy) {
4915 res_counter_init(&mem->res, &parent->res);
4916 res_counter_init(&mem->memsw, &parent->memsw);
4918 * We increment refcnt of the parent to ensure that we can
4919 * safely access it on res_counter_charge/uncharge.
4920 * This refcnt will be decremented when freeing this
4921 * mem_cgroup(see mem_cgroup_put).
4923 mem_cgroup_get(parent);
4925 res_counter_init(&mem->res, NULL);
4926 res_counter_init(&mem->memsw, NULL);
4928 mem->last_scanned_child = 0;
4929 mem->last_scanned_node = MAX_NUMNODES;
4930 INIT_LIST_HEAD(&mem->oom_notify);
4933 mem->swappiness = get_swappiness(parent);
4934 atomic_set(&mem->refcnt, 1);
4935 mem->move_charge_at_immigrate = 0;
4936 mutex_init(&mem->thresholds_lock);
4939 __mem_cgroup_free(mem);
4940 root_mem_cgroup = NULL;
4941 return ERR_PTR(error);
4944 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4945 struct cgroup *cont)
4947 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4949 return mem_cgroup_force_empty(mem, false);
4952 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4953 struct cgroup *cont)
4955 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4957 mem_cgroup_put(mem);
4960 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4961 struct cgroup *cont)
4965 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4966 ARRAY_SIZE(mem_cgroup_files));
4969 ret = register_memsw_files(cont, ss);
4974 /* Handlers for move charge at task migration. */
4975 #define PRECHARGE_COUNT_AT_ONCE 256
4976 static int mem_cgroup_do_precharge(unsigned long count)
4979 int batch_count = PRECHARGE_COUNT_AT_ONCE;
4980 struct mem_cgroup *mem = mc.to;
4982 if (mem_cgroup_is_root(mem)) {
4983 mc.precharge += count;
4984 /* we don't need css_get for root */
4987 /* try to charge at once */
4989 struct res_counter *dummy;
4991 * "mem" cannot be under rmdir() because we've already checked
4992 * by cgroup_lock_live_cgroup() that it is not removed and we
4993 * are still under the same cgroup_mutex. So we can postpone
4996 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4998 if (do_swap_account && res_counter_charge(&mem->memsw,
4999 PAGE_SIZE * count, &dummy)) {
5000 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
5003 mc.precharge += count;
5007 /* fall back to one by one charge */
5009 if (signal_pending(current)) {
5013 if (!batch_count--) {
5014 batch_count = PRECHARGE_COUNT_AT_ONCE;
5017 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5019 /* mem_cgroup_clear_mc() will do uncharge later */
5027 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5028 * @vma: the vma the pte to be checked belongs
5029 * @addr: the address corresponding to the pte to be checked
5030 * @ptent: the pte to be checked
5031 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5034 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5035 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5036 * move charge. if @target is not NULL, the page is stored in target->page
5037 * with extra refcnt got(Callers should handle it).
5038 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5039 * target for charge migration. if @target is not NULL, the entry is stored
5042 * Called with pte lock held.
5049 enum mc_target_type {
5050 MC_TARGET_NONE, /* not used */
5055 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5056 unsigned long addr, pte_t ptent)
5058 struct page *page = vm_normal_page(vma, addr, ptent);
5060 if (!page || !page_mapped(page))
5062 if (PageAnon(page)) {
5063 /* we don't move shared anon */
5064 if (!move_anon() || page_mapcount(page) > 2)
5066 } else if (!move_file())
5067 /* we ignore mapcount for file pages */
5069 if (!get_page_unless_zero(page))
5075 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5076 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5079 struct page *page = NULL;
5080 swp_entry_t ent = pte_to_swp_entry(ptent);
5082 if (!move_anon() || non_swap_entry(ent))
5084 usage_count = mem_cgroup_count_swap_user(ent, &page);
5085 if (usage_count > 1) { /* we don't move shared anon */
5090 if (do_swap_account)
5091 entry->val = ent.val;
5096 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5097 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5099 struct page *page = NULL;
5100 struct inode *inode;
5101 struct address_space *mapping;
5104 if (!vma->vm_file) /* anonymous vma */
5109 inode = vma->vm_file->f_path.dentry->d_inode;
5110 mapping = vma->vm_file->f_mapping;
5111 if (pte_none(ptent))
5112 pgoff = linear_page_index(vma, addr);
5113 else /* pte_file(ptent) is true */
5114 pgoff = pte_to_pgoff(ptent);
5116 /* page is moved even if it's not RSS of this task(page-faulted). */
5117 if (!mapping_cap_swap_backed(mapping)) { /* normal file */
5118 page = find_get_page(mapping, pgoff);
5119 } else { /* shmem/tmpfs file. we should take account of swap too. */
5121 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
5122 if (do_swap_account)
5123 entry->val = ent.val;
5129 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5130 unsigned long addr, pte_t ptent, union mc_target *target)
5132 struct page *page = NULL;
5133 struct page_cgroup *pc;
5135 swp_entry_t ent = { .val = 0 };
5137 if (pte_present(ptent))
5138 page = mc_handle_present_pte(vma, addr, ptent);
5139 else if (is_swap_pte(ptent))
5140 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5141 else if (pte_none(ptent) || pte_file(ptent))
5142 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5144 if (!page && !ent.val)
5147 pc = lookup_page_cgroup(page);
5149 * Do only loose check w/o page_cgroup lock.
5150 * mem_cgroup_move_account() checks the pc is valid or not under
5153 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5154 ret = MC_TARGET_PAGE;
5156 target->page = page;
5158 if (!ret || !target)
5161 /* There is a swap entry and a page doesn't exist or isn't charged */
5162 if (ent.val && !ret &&
5163 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5164 ret = MC_TARGET_SWAP;
5171 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5172 unsigned long addr, unsigned long end,
5173 struct mm_walk *walk)
5175 struct vm_area_struct *vma = walk->private;
5179 split_huge_page_pmd(walk->mm, pmd);
5181 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5182 for (; addr != end; pte++, addr += PAGE_SIZE)
5183 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5184 mc.precharge++; /* increment precharge temporarily */
5185 pte_unmap_unlock(pte - 1, ptl);
5191 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5193 unsigned long precharge;
5194 struct vm_area_struct *vma;
5196 down_read(&mm->mmap_sem);
5197 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5198 struct mm_walk mem_cgroup_count_precharge_walk = {
5199 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5203 if (is_vm_hugetlb_page(vma))
5205 walk_page_range(vma->vm_start, vma->vm_end,
5206 &mem_cgroup_count_precharge_walk);
5208 up_read(&mm->mmap_sem);
5210 precharge = mc.precharge;
5216 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5218 unsigned long precharge = mem_cgroup_count_precharge(mm);
5220 VM_BUG_ON(mc.moving_task);
5221 mc.moving_task = current;
5222 return mem_cgroup_do_precharge(precharge);
5225 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5226 static void __mem_cgroup_clear_mc(void)
5228 struct mem_cgroup *from = mc.from;
5229 struct mem_cgroup *to = mc.to;
5231 /* we must uncharge all the leftover precharges from mc.to */
5233 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5237 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5238 * we must uncharge here.
5240 if (mc.moved_charge) {
5241 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5242 mc.moved_charge = 0;
5244 /* we must fixup refcnts and charges */
5245 if (mc.moved_swap) {
5246 /* uncharge swap account from the old cgroup */
5247 if (!mem_cgroup_is_root(mc.from))
5248 res_counter_uncharge(&mc.from->memsw,
5249 PAGE_SIZE * mc.moved_swap);
5250 __mem_cgroup_put(mc.from, mc.moved_swap);
5252 if (!mem_cgroup_is_root(mc.to)) {
5254 * we charged both to->res and to->memsw, so we should
5257 res_counter_uncharge(&mc.to->res,
5258 PAGE_SIZE * mc.moved_swap);
5260 /* we've already done mem_cgroup_get(mc.to) */
5263 memcg_oom_recover(from);
5264 memcg_oom_recover(to);
5265 wake_up_all(&mc.waitq);
5268 static void mem_cgroup_clear_mc(void)
5270 struct mem_cgroup *from = mc.from;
5273 * we must clear moving_task before waking up waiters at the end of
5276 mc.moving_task = NULL;
5277 __mem_cgroup_clear_mc();
5278 spin_lock(&mc.lock);
5281 spin_unlock(&mc.lock);
5282 mem_cgroup_end_move(from);
5285 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5286 struct cgroup *cgroup,
5287 struct task_struct *p)
5290 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5292 if (mem->move_charge_at_immigrate) {
5293 struct mm_struct *mm;
5294 struct mem_cgroup *from = mem_cgroup_from_task(p);
5296 VM_BUG_ON(from == mem);
5298 mm = get_task_mm(p);
5301 /* We move charges only when we move a owner of the mm */
5302 if (mm->owner == p) {
5305 VM_BUG_ON(mc.precharge);
5306 VM_BUG_ON(mc.moved_charge);
5307 VM_BUG_ON(mc.moved_swap);
5308 mem_cgroup_start_move(from);
5309 spin_lock(&mc.lock);
5312 spin_unlock(&mc.lock);
5313 /* We set mc.moving_task later */
5315 ret = mem_cgroup_precharge_mc(mm);
5317 mem_cgroup_clear_mc();
5324 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5325 struct cgroup *cgroup,
5326 struct task_struct *p)
5328 mem_cgroup_clear_mc();
5331 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5332 unsigned long addr, unsigned long end,
5333 struct mm_walk *walk)
5336 struct vm_area_struct *vma = walk->private;
5340 split_huge_page_pmd(walk->mm, pmd);
5342 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5343 for (; addr != end; addr += PAGE_SIZE) {
5344 pte_t ptent = *(pte++);
5345 union mc_target target;
5348 struct page_cgroup *pc;
5354 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5356 case MC_TARGET_PAGE:
5358 if (isolate_lru_page(page))
5360 pc = lookup_page_cgroup(page);
5361 if (!mem_cgroup_move_account(page, 1, pc,
5362 mc.from, mc.to, false)) {
5364 /* we uncharge from mc.from later. */
5367 putback_lru_page(page);
5368 put: /* is_target_pte_for_mc() gets the page */
5371 case MC_TARGET_SWAP:
5373 if (!mem_cgroup_move_swap_account(ent,
5374 mc.from, mc.to, false)) {
5376 /* we fixup refcnts and charges later. */
5384 pte_unmap_unlock(pte - 1, ptl);
5389 * We have consumed all precharges we got in can_attach().
5390 * We try charge one by one, but don't do any additional
5391 * charges to mc.to if we have failed in charge once in attach()
5394 ret = mem_cgroup_do_precharge(1);
5402 static void mem_cgroup_move_charge(struct mm_struct *mm)
5404 struct vm_area_struct *vma;
5406 lru_add_drain_all();
5408 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5410 * Someone who are holding the mmap_sem might be waiting in
5411 * waitq. So we cancel all extra charges, wake up all waiters,
5412 * and retry. Because we cancel precharges, we might not be able
5413 * to move enough charges, but moving charge is a best-effort
5414 * feature anyway, so it wouldn't be a big problem.
5416 __mem_cgroup_clear_mc();
5420 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5422 struct mm_walk mem_cgroup_move_charge_walk = {
5423 .pmd_entry = mem_cgroup_move_charge_pte_range,
5427 if (is_vm_hugetlb_page(vma))
5429 ret = walk_page_range(vma->vm_start, vma->vm_end,
5430 &mem_cgroup_move_charge_walk);
5433 * means we have consumed all precharges and failed in
5434 * doing additional charge. Just abandon here.
5438 up_read(&mm->mmap_sem);
5441 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5442 struct cgroup *cont,
5443 struct cgroup *old_cont,
5444 struct task_struct *p)
5446 struct mm_struct *mm = get_task_mm(p);
5450 mem_cgroup_move_charge(mm);
5455 mem_cgroup_clear_mc();
5457 #else /* !CONFIG_MMU */
5458 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5459 struct cgroup *cgroup,
5460 struct task_struct *p)
5464 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5465 struct cgroup *cgroup,
5466 struct task_struct *p)
5469 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5470 struct cgroup *cont,
5471 struct cgroup *old_cont,
5472 struct task_struct *p)
5477 struct cgroup_subsys mem_cgroup_subsys = {
5479 .subsys_id = mem_cgroup_subsys_id,
5480 .create = mem_cgroup_create,
5481 .pre_destroy = mem_cgroup_pre_destroy,
5482 .destroy = mem_cgroup_destroy,
5483 .populate = mem_cgroup_populate,
5484 .can_attach = mem_cgroup_can_attach,
5485 .cancel_attach = mem_cgroup_cancel_attach,
5486 .attach = mem_cgroup_move_task,
5491 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5492 static int __init enable_swap_account(char *s)
5494 /* consider enabled if no parameter or 1 is given */
5495 if (!strcmp(s, "1"))
5496 really_do_swap_account = 1;
5497 else if (!strcmp(s, "0"))
5498 really_do_swap_account = 0;
5501 __setup("swapaccount=", enable_swap_account);