1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/mutex.h>
37 #include <linux/rbtree.h>
38 #include <linux/shmem_fs.h>
39 #include <linux/slab.h>
40 #include <linux/swap.h>
41 #include <linux/swapops.h>
42 #include <linux/spinlock.h>
43 #include <linux/eventfd.h>
44 #include <linux/sort.h>
46 #include <linux/seq_file.h>
47 #include <linux/vmalloc.h>
48 #include <linux/mm_inline.h>
49 #include <linux/page_cgroup.h>
50 #include <linux/cpu.h>
51 #include <linux/oom.h>
54 #include <asm/uaccess.h>
56 #include <trace/events/vmscan.h>
58 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
59 #define MEM_CGROUP_RECLAIM_RETRIES 5
60 struct mem_cgroup *root_mem_cgroup __read_mostly;
62 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
63 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
64 int do_swap_account __read_mostly;
66 /* for remember boot option*/
67 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
68 static int really_do_swap_account __initdata = 1;
70 static int really_do_swap_account __initdata = 0;
74 #define do_swap_account (0)
79 * Statistics for memory cgroup.
81 enum mem_cgroup_stat_index {
83 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
85 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
86 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
87 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
88 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
89 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
90 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */
91 MEM_CGROUP_STAT_NSTATS,
94 enum mem_cgroup_events_index {
95 MEM_CGROUP_EVENTS_PGPGIN, /* # of pages paged in */
96 MEM_CGROUP_EVENTS_PGPGOUT, /* # of pages paged out */
97 MEM_CGROUP_EVENTS_COUNT, /* # of pages paged in/out */
98 MEM_CGROUP_EVENTS_PGFAULT, /* # of page-faults */
99 MEM_CGROUP_EVENTS_PGMAJFAULT, /* # of major page-faults */
100 MEM_CGROUP_EVENTS_NSTATS,
103 * Per memcg event counter is incremented at every pagein/pageout. With THP,
104 * it will be incremated by the number of pages. This counter is used for
105 * for trigger some periodic events. This is straightforward and better
106 * than using jiffies etc. to handle periodic memcg event.
108 enum mem_cgroup_events_target {
109 MEM_CGROUP_TARGET_THRESH,
110 MEM_CGROUP_TARGET_SOFTLIMIT,
111 MEM_CGROUP_TARGET_NUMAINFO,
114 #define THRESHOLDS_EVENTS_TARGET (128)
115 #define SOFTLIMIT_EVENTS_TARGET (1024)
116 #define NUMAINFO_EVENTS_TARGET (1024)
118 struct mem_cgroup_stat_cpu {
119 long count[MEM_CGROUP_STAT_NSTATS];
120 unsigned long events[MEM_CGROUP_EVENTS_NSTATS];
121 unsigned long targets[MEM_CGROUP_NTARGETS];
125 * per-zone information in memory controller.
127 struct mem_cgroup_per_zone {
129 * spin_lock to protect the per cgroup LRU
131 struct list_head lists[NR_LRU_LISTS];
132 unsigned long count[NR_LRU_LISTS];
134 struct zone_reclaim_stat reclaim_stat;
135 struct rb_node tree_node; /* RB tree node */
136 unsigned long long usage_in_excess;/* Set to the value by which */
137 /* the soft limit is exceeded*/
139 struct mem_cgroup *mem; /* Back pointer, we cannot */
140 /* use container_of */
142 /* Macro for accessing counter */
143 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
145 struct mem_cgroup_per_node {
146 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
149 struct mem_cgroup_lru_info {
150 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
154 * Cgroups above their limits are maintained in a RB-Tree, independent of
155 * their hierarchy representation
158 struct mem_cgroup_tree_per_zone {
159 struct rb_root rb_root;
163 struct mem_cgroup_tree_per_node {
164 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
167 struct mem_cgroup_tree {
168 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
171 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
173 struct mem_cgroup_threshold {
174 struct eventfd_ctx *eventfd;
179 struct mem_cgroup_threshold_ary {
180 /* An array index points to threshold just below usage. */
181 int current_threshold;
182 /* Size of entries[] */
184 /* Array of thresholds */
185 struct mem_cgroup_threshold entries[0];
188 struct mem_cgroup_thresholds {
189 /* Primary thresholds array */
190 struct mem_cgroup_threshold_ary *primary;
192 * Spare threshold array.
193 * This is needed to make mem_cgroup_unregister_event() "never fail".
194 * It must be able to store at least primary->size - 1 entries.
196 struct mem_cgroup_threshold_ary *spare;
200 struct mem_cgroup_eventfd_list {
201 struct list_head list;
202 struct eventfd_ctx *eventfd;
205 static void mem_cgroup_threshold(struct mem_cgroup *mem);
206 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
209 * The memory controller data structure. The memory controller controls both
210 * page cache and RSS per cgroup. We would eventually like to provide
211 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
212 * to help the administrator determine what knobs to tune.
214 * TODO: Add a water mark for the memory controller. Reclaim will begin when
215 * we hit the water mark. May be even add a low water mark, such that
216 * no reclaim occurs from a cgroup at it's low water mark, this is
217 * a feature that will be implemented much later in the future.
220 struct cgroup_subsys_state css;
222 * the counter to account for memory usage
224 struct res_counter res;
226 * the counter to account for mem+swap usage.
228 struct res_counter memsw;
230 * Per cgroup active and inactive list, similar to the
231 * per zone LRU lists.
233 struct mem_cgroup_lru_info info;
235 * While reclaiming in a hierarchy, we cache the last child we
238 int last_scanned_child;
239 int last_scanned_node;
241 nodemask_t scan_nodes;
242 atomic_t numainfo_events;
243 atomic_t numainfo_updating;
246 * Should the accounting and control be hierarchical, per subtree?
252 unsigned int swappiness;
253 /* OOM-Killer disable */
254 int oom_kill_disable;
256 /* set when res.limit == memsw.limit */
257 bool memsw_is_minimum;
259 /* protect arrays of thresholds */
260 struct mutex thresholds_lock;
262 /* thresholds for memory usage. RCU-protected */
263 struct mem_cgroup_thresholds thresholds;
265 /* thresholds for mem+swap usage. RCU-protected */
266 struct mem_cgroup_thresholds memsw_thresholds;
268 /* For oom notifier event fd */
269 struct list_head oom_notify;
272 * Should we move charges of a task when a task is moved into this
273 * mem_cgroup ? And what type of charges should we move ?
275 unsigned long move_charge_at_immigrate;
279 struct mem_cgroup_stat_cpu *stat;
281 * used when a cpu is offlined or other synchronizations
282 * See mem_cgroup_read_stat().
284 struct mem_cgroup_stat_cpu nocpu_base;
285 spinlock_t pcp_counter_lock;
288 /* Stuffs for move charges at task migration. */
290 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
291 * left-shifted bitmap of these types.
294 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
295 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
299 /* "mc" and its members are protected by cgroup_mutex */
300 static struct move_charge_struct {
301 spinlock_t lock; /* for from, to */
302 struct mem_cgroup *from;
303 struct mem_cgroup *to;
304 unsigned long precharge;
305 unsigned long moved_charge;
306 unsigned long moved_swap;
307 struct task_struct *moving_task; /* a task moving charges */
308 wait_queue_head_t waitq; /* a waitq for other context */
310 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
311 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
314 static bool move_anon(void)
316 return test_bit(MOVE_CHARGE_TYPE_ANON,
317 &mc.to->move_charge_at_immigrate);
320 static bool move_file(void)
322 return test_bit(MOVE_CHARGE_TYPE_FILE,
323 &mc.to->move_charge_at_immigrate);
327 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
328 * limit reclaim to prevent infinite loops, if they ever occur.
330 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
331 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
334 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
335 MEM_CGROUP_CHARGE_TYPE_MAPPED,
336 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
337 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
338 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
339 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
343 /* for encoding cft->private value on file */
346 #define _OOM_TYPE (2)
347 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
348 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
349 #define MEMFILE_ATTR(val) ((val) & 0xffff)
350 /* Used for OOM nofiier */
351 #define OOM_CONTROL (0)
354 * Reclaim flags for mem_cgroup_hierarchical_reclaim
356 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
357 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
358 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
359 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
360 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
361 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
363 static void mem_cgroup_get(struct mem_cgroup *mem);
364 static void mem_cgroup_put(struct mem_cgroup *mem);
365 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
366 static void drain_all_stock_async(struct mem_cgroup *mem);
368 static struct mem_cgroup_per_zone *
369 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
371 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
374 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
379 static struct mem_cgroup_per_zone *
380 page_cgroup_zoneinfo(struct mem_cgroup *mem, struct page *page)
382 int nid = page_to_nid(page);
383 int zid = page_zonenum(page);
385 return mem_cgroup_zoneinfo(mem, nid, zid);
388 static struct mem_cgroup_tree_per_zone *
389 soft_limit_tree_node_zone(int nid, int zid)
391 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
394 static struct mem_cgroup_tree_per_zone *
395 soft_limit_tree_from_page(struct page *page)
397 int nid = page_to_nid(page);
398 int zid = page_zonenum(page);
400 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
404 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
405 struct mem_cgroup_per_zone *mz,
406 struct mem_cgroup_tree_per_zone *mctz,
407 unsigned long long new_usage_in_excess)
409 struct rb_node **p = &mctz->rb_root.rb_node;
410 struct rb_node *parent = NULL;
411 struct mem_cgroup_per_zone *mz_node;
416 mz->usage_in_excess = new_usage_in_excess;
417 if (!mz->usage_in_excess)
421 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
423 if (mz->usage_in_excess < mz_node->usage_in_excess)
426 * We can't avoid mem cgroups that are over their soft
427 * limit by the same amount
429 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
432 rb_link_node(&mz->tree_node, parent, p);
433 rb_insert_color(&mz->tree_node, &mctz->rb_root);
438 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
439 struct mem_cgroup_per_zone *mz,
440 struct mem_cgroup_tree_per_zone *mctz)
444 rb_erase(&mz->tree_node, &mctz->rb_root);
449 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
450 struct mem_cgroup_per_zone *mz,
451 struct mem_cgroup_tree_per_zone *mctz)
453 spin_lock(&mctz->lock);
454 __mem_cgroup_remove_exceeded(mem, mz, mctz);
455 spin_unlock(&mctz->lock);
459 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
461 unsigned long long excess;
462 struct mem_cgroup_per_zone *mz;
463 struct mem_cgroup_tree_per_zone *mctz;
464 int nid = page_to_nid(page);
465 int zid = page_zonenum(page);
466 mctz = soft_limit_tree_from_page(page);
469 * Necessary to update all ancestors when hierarchy is used.
470 * because their event counter is not touched.
472 for (; mem; mem = parent_mem_cgroup(mem)) {
473 mz = mem_cgroup_zoneinfo(mem, nid, zid);
474 excess = res_counter_soft_limit_excess(&mem->res);
476 * We have to update the tree if mz is on RB-tree or
477 * mem is over its softlimit.
479 if (excess || mz->on_tree) {
480 spin_lock(&mctz->lock);
481 /* if on-tree, remove it */
483 __mem_cgroup_remove_exceeded(mem, mz, mctz);
485 * Insert again. mz->usage_in_excess will be updated.
486 * If excess is 0, no tree ops.
488 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
489 spin_unlock(&mctz->lock);
494 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
497 struct mem_cgroup_per_zone *mz;
498 struct mem_cgroup_tree_per_zone *mctz;
500 for_each_node_state(node, N_POSSIBLE) {
501 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
502 mz = mem_cgroup_zoneinfo(mem, node, zone);
503 mctz = soft_limit_tree_node_zone(node, zone);
504 mem_cgroup_remove_exceeded(mem, mz, mctz);
509 static struct mem_cgroup_per_zone *
510 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
512 struct rb_node *rightmost = NULL;
513 struct mem_cgroup_per_zone *mz;
517 rightmost = rb_last(&mctz->rb_root);
519 goto done; /* Nothing to reclaim from */
521 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
523 * Remove the node now but someone else can add it back,
524 * we will to add it back at the end of reclaim to its correct
525 * position in the tree.
527 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
528 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
529 !css_tryget(&mz->mem->css))
535 static struct mem_cgroup_per_zone *
536 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
538 struct mem_cgroup_per_zone *mz;
540 spin_lock(&mctz->lock);
541 mz = __mem_cgroup_largest_soft_limit_node(mctz);
542 spin_unlock(&mctz->lock);
547 * Implementation Note: reading percpu statistics for memcg.
549 * Both of vmstat[] and percpu_counter has threshold and do periodic
550 * synchronization to implement "quick" read. There are trade-off between
551 * reading cost and precision of value. Then, we may have a chance to implement
552 * a periodic synchronizion of counter in memcg's counter.
554 * But this _read() function is used for user interface now. The user accounts
555 * memory usage by memory cgroup and he _always_ requires exact value because
556 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
557 * have to visit all online cpus and make sum. So, for now, unnecessary
558 * synchronization is not implemented. (just implemented for cpu hotplug)
560 * If there are kernel internal actions which can make use of some not-exact
561 * value, and reading all cpu value can be performance bottleneck in some
562 * common workload, threashold and synchonization as vmstat[] should be
565 static long mem_cgroup_read_stat(struct mem_cgroup *mem,
566 enum mem_cgroup_stat_index idx)
572 for_each_online_cpu(cpu)
573 val += per_cpu(mem->stat->count[idx], cpu);
574 #ifdef CONFIG_HOTPLUG_CPU
575 spin_lock(&mem->pcp_counter_lock);
576 val += mem->nocpu_base.count[idx];
577 spin_unlock(&mem->pcp_counter_lock);
583 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
586 int val = (charge) ? 1 : -1;
587 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
590 void mem_cgroup_pgfault(struct mem_cgroup *mem, int val)
592 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
595 void mem_cgroup_pgmajfault(struct mem_cgroup *mem, int val)
597 this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
600 static unsigned long mem_cgroup_read_events(struct mem_cgroup *mem,
601 enum mem_cgroup_events_index idx)
603 unsigned long val = 0;
606 for_each_online_cpu(cpu)
607 val += per_cpu(mem->stat->events[idx], cpu);
608 #ifdef CONFIG_HOTPLUG_CPU
609 spin_lock(&mem->pcp_counter_lock);
610 val += mem->nocpu_base.events[idx];
611 spin_unlock(&mem->pcp_counter_lock);
616 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
617 bool file, int nr_pages)
622 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
624 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
626 /* pagein of a big page is an event. So, ignore page size */
628 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
630 __this_cpu_inc(mem->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
631 nr_pages = -nr_pages; /* for event */
634 __this_cpu_add(mem->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
640 mem_cgroup_get_zonestat_node(struct mem_cgroup *mem, int nid, enum lru_list idx)
642 struct mem_cgroup_per_zone *mz;
646 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
647 mz = mem_cgroup_zoneinfo(mem, nid, zid);
648 total += MEM_CGROUP_ZSTAT(mz, idx);
652 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
658 for_each_online_node(nid)
659 total += mem_cgroup_get_zonestat_node(mem, nid, idx);
663 static bool __memcg_event_check(struct mem_cgroup *mem, int target)
665 unsigned long val, next;
667 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
668 next = this_cpu_read(mem->stat->targets[target]);
669 /* from time_after() in jiffies.h */
670 return ((long)next - (long)val < 0);
673 static void __mem_cgroup_target_update(struct mem_cgroup *mem, int target)
675 unsigned long val, next;
677 val = this_cpu_read(mem->stat->events[MEM_CGROUP_EVENTS_COUNT]);
680 case MEM_CGROUP_TARGET_THRESH:
681 next = val + THRESHOLDS_EVENTS_TARGET;
683 case MEM_CGROUP_TARGET_SOFTLIMIT:
684 next = val + SOFTLIMIT_EVENTS_TARGET;
686 case MEM_CGROUP_TARGET_NUMAINFO:
687 next = val + NUMAINFO_EVENTS_TARGET;
693 this_cpu_write(mem->stat->targets[target], next);
697 * Check events in order.
700 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
702 /* threshold event is triggered in finer grain than soft limit */
703 if (unlikely(__memcg_event_check(mem, MEM_CGROUP_TARGET_THRESH))) {
704 mem_cgroup_threshold(mem);
705 __mem_cgroup_target_update(mem, MEM_CGROUP_TARGET_THRESH);
706 if (unlikely(__memcg_event_check(mem,
707 MEM_CGROUP_TARGET_SOFTLIMIT))) {
708 mem_cgroup_update_tree(mem, page);
709 __mem_cgroup_target_update(mem,
710 MEM_CGROUP_TARGET_SOFTLIMIT);
713 if (unlikely(__memcg_event_check(mem,
714 MEM_CGROUP_TARGET_NUMAINFO))) {
715 atomic_inc(&mem->numainfo_events);
716 __mem_cgroup_target_update(mem,
717 MEM_CGROUP_TARGET_NUMAINFO);
723 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
725 return container_of(cgroup_subsys_state(cont,
726 mem_cgroup_subsys_id), struct mem_cgroup,
730 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
733 * mm_update_next_owner() may clear mm->owner to NULL
734 * if it races with swapoff, page migration, etc.
735 * So this can be called with p == NULL.
740 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
741 struct mem_cgroup, css);
744 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
746 struct mem_cgroup *mem = NULL;
751 * Because we have no locks, mm->owner's may be being moved to other
752 * cgroup. We use css_tryget() here even if this looks
753 * pessimistic (rather than adding locks here).
757 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
760 } while (!css_tryget(&mem->css));
765 /* The caller has to guarantee "mem" exists before calling this */
766 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
768 struct cgroup_subsys_state *css;
771 if (!mem) /* ROOT cgroup has the smallest ID */
772 return root_mem_cgroup; /*css_put/get against root is ignored*/
773 if (!mem->use_hierarchy) {
774 if (css_tryget(&mem->css))
780 * searching a memory cgroup which has the smallest ID under given
781 * ROOT cgroup. (ID >= 1)
783 css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
784 if (css && css_tryget(css))
785 mem = container_of(css, struct mem_cgroup, css);
792 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
793 struct mem_cgroup *root,
796 int nextid = css_id(&iter->css) + 1;
799 struct cgroup_subsys_state *css;
801 hierarchy_used = iter->use_hierarchy;
804 /* If no ROOT, walk all, ignore hierarchy */
805 if (!cond || (root && !hierarchy_used))
809 root = root_mem_cgroup;
815 css = css_get_next(&mem_cgroup_subsys, nextid,
817 if (css && css_tryget(css))
818 iter = container_of(css, struct mem_cgroup, css);
820 /* If css is NULL, no more cgroups will be found */
822 } while (css && !iter);
827 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
828 * be careful that "break" loop is not allowed. We have reference count.
829 * Instead of that modify "cond" to be false and "continue" to exit the loop.
831 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
832 for (iter = mem_cgroup_start_loop(root);\
834 iter = mem_cgroup_get_next(iter, root, cond))
836 #define for_each_mem_cgroup_tree(iter, root) \
837 for_each_mem_cgroup_tree_cond(iter, root, true)
839 #define for_each_mem_cgroup_all(iter) \
840 for_each_mem_cgroup_tree_cond(iter, NULL, true)
843 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
845 return (mem == root_mem_cgroup);
848 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
850 struct mem_cgroup *mem;
856 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
862 mem_cgroup_pgmajfault(mem, 1);
865 mem_cgroup_pgfault(mem, 1);
873 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
876 * Following LRU functions are allowed to be used without PCG_LOCK.
877 * Operations are called by routine of global LRU independently from memcg.
878 * What we have to take care of here is validness of pc->mem_cgroup.
880 * Changes to pc->mem_cgroup happens when
883 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
884 * It is added to LRU before charge.
885 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
886 * When moving account, the page is not on LRU. It's isolated.
889 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
891 struct page_cgroup *pc;
892 struct mem_cgroup_per_zone *mz;
894 if (mem_cgroup_disabled())
896 pc = lookup_page_cgroup(page);
897 /* can happen while we handle swapcache. */
898 if (!TestClearPageCgroupAcctLRU(pc))
900 VM_BUG_ON(!pc->mem_cgroup);
902 * We don't check PCG_USED bit. It's cleared when the "page" is finally
903 * removed from global LRU.
905 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
906 /* huge page split is done under lru_lock. so, we have no races. */
907 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
908 if (mem_cgroup_is_root(pc->mem_cgroup))
910 VM_BUG_ON(list_empty(&pc->lru));
911 list_del_init(&pc->lru);
914 void mem_cgroup_del_lru(struct page *page)
916 mem_cgroup_del_lru_list(page, page_lru(page));
920 * Writeback is about to end against a page which has been marked for immediate
921 * reclaim. If it still appears to be reclaimable, move it to the tail of the
924 void mem_cgroup_rotate_reclaimable_page(struct page *page)
926 struct mem_cgroup_per_zone *mz;
927 struct page_cgroup *pc;
928 enum lru_list lru = page_lru(page);
930 if (mem_cgroup_disabled())
933 pc = lookup_page_cgroup(page);
934 /* unused or root page is not rotated. */
935 if (!PageCgroupUsed(pc))
937 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
939 if (mem_cgroup_is_root(pc->mem_cgroup))
941 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
942 list_move_tail(&pc->lru, &mz->lists[lru]);
945 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
947 struct mem_cgroup_per_zone *mz;
948 struct page_cgroup *pc;
950 if (mem_cgroup_disabled())
953 pc = lookup_page_cgroup(page);
954 /* unused or root page is not rotated. */
955 if (!PageCgroupUsed(pc))
957 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
959 if (mem_cgroup_is_root(pc->mem_cgroup))
961 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
962 list_move(&pc->lru, &mz->lists[lru]);
965 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
967 struct page_cgroup *pc;
968 struct mem_cgroup_per_zone *mz;
970 if (mem_cgroup_disabled())
972 pc = lookup_page_cgroup(page);
973 VM_BUG_ON(PageCgroupAcctLRU(pc));
974 if (!PageCgroupUsed(pc))
976 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
978 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
979 /* huge page split is done under lru_lock. so, we have no races. */
980 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
981 SetPageCgroupAcctLRU(pc);
982 if (mem_cgroup_is_root(pc->mem_cgroup))
984 list_add(&pc->lru, &mz->lists[lru]);
988 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
989 * while it's linked to lru because the page may be reused after it's fully
990 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
991 * It's done under lock_page and expected that zone->lru_lock isnever held.
993 static void mem_cgroup_lru_del_before_commit(struct page *page)
996 struct zone *zone = page_zone(page);
997 struct page_cgroup *pc = lookup_page_cgroup(page);
1000 * Doing this check without taking ->lru_lock seems wrong but this
1001 * is safe. Because if page_cgroup's USED bit is unset, the page
1002 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1003 * set, the commit after this will fail, anyway.
1004 * This all charge/uncharge is done under some mutual execustion.
1005 * So, we don't need to taking care of changes in USED bit.
1007 if (likely(!PageLRU(page)))
1010 spin_lock_irqsave(&zone->lru_lock, flags);
1012 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1013 * is guarded by lock_page() because the page is SwapCache.
1015 if (!PageCgroupUsed(pc))
1016 mem_cgroup_del_lru_list(page, page_lru(page));
1017 spin_unlock_irqrestore(&zone->lru_lock, flags);
1020 static void mem_cgroup_lru_add_after_commit(struct page *page)
1022 unsigned long flags;
1023 struct zone *zone = page_zone(page);
1024 struct page_cgroup *pc = lookup_page_cgroup(page);
1026 /* taking care of that the page is added to LRU while we commit it */
1027 if (likely(!PageLRU(page)))
1029 spin_lock_irqsave(&zone->lru_lock, flags);
1030 /* link when the page is linked to LRU but page_cgroup isn't */
1031 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1032 mem_cgroup_add_lru_list(page, page_lru(page));
1033 spin_unlock_irqrestore(&zone->lru_lock, flags);
1037 void mem_cgroup_move_lists(struct page *page,
1038 enum lru_list from, enum lru_list to)
1040 if (mem_cgroup_disabled())
1042 mem_cgroup_del_lru_list(page, from);
1043 mem_cgroup_add_lru_list(page, to);
1046 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
1049 struct mem_cgroup *curr = NULL;
1050 struct task_struct *p;
1052 p = find_lock_task_mm(task);
1055 curr = try_get_mem_cgroup_from_mm(p->mm);
1060 * We should check use_hierarchy of "mem" not "curr". Because checking
1061 * use_hierarchy of "curr" here make this function true if hierarchy is
1062 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
1063 * hierarchy(even if use_hierarchy is disabled in "mem").
1065 if (mem->use_hierarchy)
1066 ret = css_is_ancestor(&curr->css, &mem->css);
1068 ret = (curr == mem);
1069 css_put(&curr->css);
1073 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1075 unsigned long active;
1076 unsigned long inactive;
1078 unsigned long inactive_ratio;
1080 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
1081 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
1083 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1085 inactive_ratio = int_sqrt(10 * gb);
1089 if (present_pages) {
1090 present_pages[0] = inactive;
1091 present_pages[1] = active;
1094 return inactive_ratio;
1097 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1099 unsigned long active;
1100 unsigned long inactive;
1101 unsigned long present_pages[2];
1102 unsigned long inactive_ratio;
1104 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1106 inactive = present_pages[0];
1107 active = present_pages[1];
1109 if (inactive * inactive_ratio < active)
1115 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1117 unsigned long active;
1118 unsigned long inactive;
1120 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
1121 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
1123 return (active > inactive);
1126 unsigned long mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg,
1130 int nid = zone_to_nid(zone);
1131 int zid = zone_idx(zone);
1132 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1134 return MEM_CGROUP_ZSTAT(mz, lru);
1137 static unsigned long mem_cgroup_node_nr_file_lru_pages(struct mem_cgroup *memcg,
1142 ret = mem_cgroup_get_zonestat_node(memcg, nid, LRU_INACTIVE_FILE) +
1143 mem_cgroup_get_zonestat_node(memcg, nid, LRU_ACTIVE_FILE);
1148 static unsigned long mem_cgroup_node_nr_anon_lru_pages(struct mem_cgroup *memcg,
1153 ret = mem_cgroup_get_zonestat_node(memcg, nid, LRU_INACTIVE_ANON) +
1154 mem_cgroup_get_zonestat_node(memcg, nid, LRU_ACTIVE_ANON);
1158 #if MAX_NUMNODES > 1
1159 static unsigned long mem_cgroup_nr_file_lru_pages(struct mem_cgroup *memcg)
1164 for_each_node_state(nid, N_HIGH_MEMORY)
1165 total += mem_cgroup_node_nr_file_lru_pages(memcg, nid);
1170 static unsigned long mem_cgroup_nr_anon_lru_pages(struct mem_cgroup *memcg)
1175 for_each_node_state(nid, N_HIGH_MEMORY)
1176 total += mem_cgroup_node_nr_anon_lru_pages(memcg, nid);
1181 static unsigned long
1182 mem_cgroup_node_nr_unevictable_lru_pages(struct mem_cgroup *memcg, int nid)
1184 return mem_cgroup_get_zonestat_node(memcg, nid, LRU_UNEVICTABLE);
1187 static unsigned long
1188 mem_cgroup_nr_unevictable_lru_pages(struct mem_cgroup *memcg)
1193 for_each_node_state(nid, N_HIGH_MEMORY)
1194 total += mem_cgroup_node_nr_unevictable_lru_pages(memcg, nid);
1199 static unsigned long mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
1206 total += mem_cgroup_get_zonestat_node(memcg, nid, l);
1211 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg)
1216 for_each_node_state(nid, N_HIGH_MEMORY)
1217 total += mem_cgroup_node_nr_lru_pages(memcg, nid);
1221 #endif /* CONFIG_NUMA */
1223 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1226 int nid = zone_to_nid(zone);
1227 int zid = zone_idx(zone);
1228 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1230 return &mz->reclaim_stat;
1233 struct zone_reclaim_stat *
1234 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1236 struct page_cgroup *pc;
1237 struct mem_cgroup_per_zone *mz;
1239 if (mem_cgroup_disabled())
1242 pc = lookup_page_cgroup(page);
1243 if (!PageCgroupUsed(pc))
1245 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1247 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1248 return &mz->reclaim_stat;
1251 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1252 struct list_head *dst,
1253 unsigned long *scanned, int order,
1254 isolate_mode_t mode,
1256 struct mem_cgroup *mem_cont,
1257 int active, int file)
1259 unsigned long nr_taken = 0;
1263 struct list_head *src;
1264 struct page_cgroup *pc, *tmp;
1265 int nid = zone_to_nid(z);
1266 int zid = zone_idx(z);
1267 struct mem_cgroup_per_zone *mz;
1268 int lru = LRU_FILE * file + active;
1272 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1273 src = &mz->lists[lru];
1276 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1277 if (scan >= nr_to_scan)
1280 if (unlikely(!PageCgroupUsed(pc)))
1283 page = lookup_cgroup_page(pc);
1285 if (unlikely(!PageLRU(page)))
1289 ret = __isolate_lru_page(page, mode, file);
1292 list_move(&page->lru, dst);
1293 mem_cgroup_del_lru(page);
1294 nr_taken += hpage_nr_pages(page);
1297 /* we don't affect global LRU but rotate in our LRU */
1298 mem_cgroup_rotate_lru_list(page, page_lru(page));
1307 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1313 #define mem_cgroup_from_res_counter(counter, member) \
1314 container_of(counter, struct mem_cgroup, member)
1317 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1318 * @mem: the memory cgroup
1320 * Returns the maximum amount of memory @mem can be charged with, in
1323 static unsigned long mem_cgroup_margin(struct mem_cgroup *mem)
1325 unsigned long long margin;
1327 margin = res_counter_margin(&mem->res);
1328 if (do_swap_account)
1329 margin = min(margin, res_counter_margin(&mem->memsw));
1330 return margin >> PAGE_SHIFT;
1333 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1335 struct cgroup *cgrp = memcg->css.cgroup;
1338 if (cgrp->parent == NULL)
1339 return vm_swappiness;
1341 return memcg->swappiness;
1344 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1349 spin_lock(&mem->pcp_counter_lock);
1350 for_each_online_cpu(cpu)
1351 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1352 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1353 spin_unlock(&mem->pcp_counter_lock);
1359 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1366 spin_lock(&mem->pcp_counter_lock);
1367 for_each_online_cpu(cpu)
1368 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1369 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1370 spin_unlock(&mem->pcp_counter_lock);
1374 * 2 routines for checking "mem" is under move_account() or not.
1376 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1377 * for avoiding race in accounting. If true,
1378 * pc->mem_cgroup may be overwritten.
1380 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1381 * under hierarchy of moving cgroups. This is for
1382 * waiting at hith-memory prressure caused by "move".
1385 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1387 VM_BUG_ON(!rcu_read_lock_held());
1388 return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1391 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1393 struct mem_cgroup *from;
1394 struct mem_cgroup *to;
1397 * Unlike task_move routines, we access mc.to, mc.from not under
1398 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1400 spin_lock(&mc.lock);
1405 if (from == mem || to == mem
1406 || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1407 || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1410 spin_unlock(&mc.lock);
1414 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1416 if (mc.moving_task && current != mc.moving_task) {
1417 if (mem_cgroup_under_move(mem)) {
1419 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1420 /* moving charge context might have finished. */
1423 finish_wait(&mc.waitq, &wait);
1431 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1432 * @memcg: The memory cgroup that went over limit
1433 * @p: Task that is going to be killed
1435 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1438 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1440 struct cgroup *task_cgrp;
1441 struct cgroup *mem_cgrp;
1443 * Need a buffer in BSS, can't rely on allocations. The code relies
1444 * on the assumption that OOM is serialized for memory controller.
1445 * If this assumption is broken, revisit this code.
1447 static char memcg_name[PATH_MAX];
1456 mem_cgrp = memcg->css.cgroup;
1457 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1459 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1462 * Unfortunately, we are unable to convert to a useful name
1463 * But we'll still print out the usage information
1470 printk(KERN_INFO "Task in %s killed", memcg_name);
1473 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1481 * Continues from above, so we don't need an KERN_ level
1483 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1486 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1487 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1488 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1489 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1490 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1492 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1493 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1494 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1498 * This function returns the number of memcg under hierarchy tree. Returns
1499 * 1(self count) if no children.
1501 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1504 struct mem_cgroup *iter;
1506 for_each_mem_cgroup_tree(iter, mem)
1512 * Return the memory (and swap, if configured) limit for a memcg.
1514 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1519 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1520 limit += total_swap_pages << PAGE_SHIFT;
1522 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1524 * If memsw is finite and limits the amount of swap space available
1525 * to this memcg, return that limit.
1527 return min(limit, memsw);
1531 * Visit the first child (need not be the first child as per the ordering
1532 * of the cgroup list, since we track last_scanned_child) of @mem and use
1533 * that to reclaim free pages from.
1535 static struct mem_cgroup *
1536 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1538 struct mem_cgroup *ret = NULL;
1539 struct cgroup_subsys_state *css;
1542 if (!root_mem->use_hierarchy) {
1543 css_get(&root_mem->css);
1549 nextid = root_mem->last_scanned_child + 1;
1550 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1552 if (css && css_tryget(css))
1553 ret = container_of(css, struct mem_cgroup, css);
1556 /* Updates scanning parameter */
1558 /* this means start scan from ID:1 */
1559 root_mem->last_scanned_child = 0;
1561 root_mem->last_scanned_child = found;
1568 * test_mem_cgroup_node_reclaimable
1569 * @mem: the target memcg
1570 * @nid: the node ID to be checked.
1571 * @noswap : specify true here if the user wants flle only information.
1573 * This function returns whether the specified memcg contains any
1574 * reclaimable pages on a node. Returns true if there are any reclaimable
1575 * pages in the node.
1577 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *mem,
1578 int nid, bool noswap)
1580 if (mem_cgroup_node_nr_file_lru_pages(mem, nid))
1582 if (noswap || !total_swap_pages)
1584 if (mem_cgroup_node_nr_anon_lru_pages(mem, nid))
1589 #if MAX_NUMNODES > 1
1592 * Always updating the nodemask is not very good - even if we have an empty
1593 * list or the wrong list here, we can start from some node and traverse all
1594 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1597 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *mem)
1601 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1602 * pagein/pageout changes since the last update.
1604 if (!atomic_read(&mem->numainfo_events))
1606 if (atomic_inc_return(&mem->numainfo_updating) > 1)
1609 /* make a nodemask where this memcg uses memory from */
1610 mem->scan_nodes = node_states[N_HIGH_MEMORY];
1612 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1614 if (!test_mem_cgroup_node_reclaimable(mem, nid, false))
1615 node_clear(nid, mem->scan_nodes);
1618 atomic_set(&mem->numainfo_events, 0);
1619 atomic_set(&mem->numainfo_updating, 0);
1623 * Selecting a node where we start reclaim from. Because what we need is just
1624 * reducing usage counter, start from anywhere is O,K. Considering
1625 * memory reclaim from current node, there are pros. and cons.
1627 * Freeing memory from current node means freeing memory from a node which
1628 * we'll use or we've used. So, it may make LRU bad. And if several threads
1629 * hit limits, it will see a contention on a node. But freeing from remote
1630 * node means more costs for memory reclaim because of memory latency.
1632 * Now, we use round-robin. Better algorithm is welcomed.
1634 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1638 mem_cgroup_may_update_nodemask(mem);
1639 node = mem->last_scanned_node;
1641 node = next_node(node, mem->scan_nodes);
1642 if (node == MAX_NUMNODES)
1643 node = first_node(mem->scan_nodes);
1645 * We call this when we hit limit, not when pages are added to LRU.
1646 * No LRU may hold pages because all pages are UNEVICTABLE or
1647 * memcg is too small and all pages are not on LRU. In that case,
1648 * we use curret node.
1650 if (unlikely(node == MAX_NUMNODES))
1651 node = numa_node_id();
1653 mem->last_scanned_node = node;
1658 * Check all nodes whether it contains reclaimable pages or not.
1659 * For quick scan, we make use of scan_nodes. This will allow us to skip
1660 * unused nodes. But scan_nodes is lazily updated and may not cotain
1661 * enough new information. We need to do double check.
1663 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1668 * quick check...making use of scan_node.
1669 * We can skip unused nodes.
1671 if (!nodes_empty(mem->scan_nodes)) {
1672 for (nid = first_node(mem->scan_nodes);
1674 nid = next_node(nid, mem->scan_nodes)) {
1676 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1681 * Check rest of nodes.
1683 for_each_node_state(nid, N_HIGH_MEMORY) {
1684 if (node_isset(nid, mem->scan_nodes))
1686 if (test_mem_cgroup_node_reclaimable(mem, nid, noswap))
1693 int mem_cgroup_select_victim_node(struct mem_cgroup *mem)
1698 bool mem_cgroup_reclaimable(struct mem_cgroup *mem, bool noswap)
1700 return test_mem_cgroup_node_reclaimable(mem, 0, noswap);
1705 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1706 * we reclaimed from, so that we don't end up penalizing one child extensively
1707 * based on its position in the children list.
1709 * root_mem is the original ancestor that we've been reclaim from.
1711 * We give up and return to the caller when we visit root_mem twice.
1712 * (other groups can be removed while we're walking....)
1714 * If shrink==true, for avoiding to free too much, this returns immedieately.
1716 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1719 unsigned long reclaim_options,
1720 unsigned long *total_scanned)
1722 struct mem_cgroup *victim;
1725 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1726 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1727 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1728 unsigned long excess;
1729 unsigned long nr_scanned;
1731 excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1733 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1734 if (!check_soft && !shrink && root_mem->memsw_is_minimum)
1738 victim = mem_cgroup_select_victim(root_mem);
1739 if (victim == root_mem) {
1742 * We are not draining per cpu cached charges during
1743 * soft limit reclaim because global reclaim doesn't
1744 * care about charges. It tries to free some memory and
1745 * charges will not give any.
1747 if (!check_soft && loop >= 1)
1748 drain_all_stock_async(root_mem);
1751 * If we have not been able to reclaim
1752 * anything, it might because there are
1753 * no reclaimable pages under this hierarchy
1755 if (!check_soft || !total) {
1756 css_put(&victim->css);
1760 * We want to do more targeted reclaim.
1761 * excess >> 2 is not to excessive so as to
1762 * reclaim too much, nor too less that we keep
1763 * coming back to reclaim from this cgroup
1765 if (total >= (excess >> 2) ||
1766 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1767 css_put(&victim->css);
1772 if (!mem_cgroup_reclaimable(victim, noswap)) {
1773 /* this cgroup's local usage == 0 */
1774 css_put(&victim->css);
1777 /* we use swappiness of local cgroup */
1779 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1780 noswap, get_swappiness(victim), zone,
1782 *total_scanned += nr_scanned;
1784 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1785 noswap, get_swappiness(victim));
1786 css_put(&victim->css);
1788 * At shrinking usage, we can't check we should stop here or
1789 * reclaim more. It's depends on callers. last_scanned_child
1790 * will work enough for keeping fairness under tree.
1796 if (!res_counter_soft_limit_excess(&root_mem->res))
1798 } else if (mem_cgroup_margin(root_mem))
1805 * Check OOM-Killer is already running under our hierarchy.
1806 * If someone is running, return false.
1808 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1810 int x, lock_count = 0;
1811 struct mem_cgroup *iter;
1813 for_each_mem_cgroup_tree(iter, mem) {
1814 x = atomic_inc_return(&iter->oom_lock);
1815 lock_count = max(x, lock_count);
1818 if (lock_count == 1)
1823 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1825 struct mem_cgroup *iter;
1828 * When a new child is created while the hierarchy is under oom,
1829 * mem_cgroup_oom_lock() may not be called. We have to use
1830 * atomic_add_unless() here.
1832 for_each_mem_cgroup_tree(iter, mem)
1833 atomic_add_unless(&iter->oom_lock, -1, 0);
1838 static DEFINE_MUTEX(memcg_oom_mutex);
1839 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1841 struct oom_wait_info {
1842 struct mem_cgroup *mem;
1846 static int memcg_oom_wake_function(wait_queue_t *wait,
1847 unsigned mode, int sync, void *arg)
1849 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1850 struct oom_wait_info *oom_wait_info;
1852 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1854 if (oom_wait_info->mem == wake_mem)
1856 /* if no hierarchy, no match */
1857 if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1860 * Both of oom_wait_info->mem and wake_mem are stable under us.
1861 * Then we can use css_is_ancestor without taking care of RCU.
1863 if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1864 !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1868 return autoremove_wake_function(wait, mode, sync, arg);
1871 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1873 /* for filtering, pass "mem" as argument. */
1874 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1877 static void memcg_oom_recover(struct mem_cgroup *mem)
1879 if (mem && atomic_read(&mem->oom_lock))
1880 memcg_wakeup_oom(mem);
1884 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1886 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1888 struct oom_wait_info owait;
1889 bool locked, need_to_kill;
1892 owait.wait.flags = 0;
1893 owait.wait.func = memcg_oom_wake_function;
1894 owait.wait.private = current;
1895 INIT_LIST_HEAD(&owait.wait.task_list);
1896 need_to_kill = true;
1897 /* At first, try to OOM lock hierarchy under mem.*/
1898 mutex_lock(&memcg_oom_mutex);
1899 locked = mem_cgroup_oom_lock(mem);
1901 * Even if signal_pending(), we can't quit charge() loop without
1902 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1903 * under OOM is always welcomed, use TASK_KILLABLE here.
1905 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1906 if (!locked || mem->oom_kill_disable)
1907 need_to_kill = false;
1909 mem_cgroup_oom_notify(mem);
1910 mutex_unlock(&memcg_oom_mutex);
1913 finish_wait(&memcg_oom_waitq, &owait.wait);
1914 mem_cgroup_out_of_memory(mem, mask);
1917 finish_wait(&memcg_oom_waitq, &owait.wait);
1919 mutex_lock(&memcg_oom_mutex);
1920 mem_cgroup_oom_unlock(mem);
1921 memcg_wakeup_oom(mem);
1922 mutex_unlock(&memcg_oom_mutex);
1924 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1926 /* Give chance to dying process */
1927 schedule_timeout(1);
1932 * Currently used to update mapped file statistics, but the routine can be
1933 * generalized to update other statistics as well.
1935 * Notes: Race condition
1937 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1938 * it tends to be costly. But considering some conditions, we doesn't need
1939 * to do so _always_.
1941 * Considering "charge", lock_page_cgroup() is not required because all
1942 * file-stat operations happen after a page is attached to radix-tree. There
1943 * are no race with "charge".
1945 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1946 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1947 * if there are race with "uncharge". Statistics itself is properly handled
1950 * Considering "move", this is an only case we see a race. To make the race
1951 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1952 * possibility of race condition. If there is, we take a lock.
1955 void mem_cgroup_update_page_stat(struct page *page,
1956 enum mem_cgroup_page_stat_item idx, int val)
1958 struct mem_cgroup *mem;
1959 struct page_cgroup *pc = lookup_page_cgroup(page);
1960 bool need_unlock = false;
1961 unsigned long uninitialized_var(flags);
1967 mem = pc->mem_cgroup;
1968 if (unlikely(!mem || !PageCgroupUsed(pc)))
1970 /* pc->mem_cgroup is unstable ? */
1971 if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1972 /* take a lock against to access pc->mem_cgroup */
1973 move_lock_page_cgroup(pc, &flags);
1975 mem = pc->mem_cgroup;
1976 if (!mem || !PageCgroupUsed(pc))
1981 case MEMCG_NR_FILE_MAPPED:
1983 SetPageCgroupFileMapped(pc);
1984 else if (!page_mapped(page))
1985 ClearPageCgroupFileMapped(pc);
1986 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1992 this_cpu_add(mem->stat->count[idx], val);
1995 if (unlikely(need_unlock))
1996 move_unlock_page_cgroup(pc, &flags);
2000 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
2003 * size of first charge trial. "32" comes from vmscan.c's magic value.
2004 * TODO: maybe necessary to use big numbers in big irons.
2006 #define CHARGE_BATCH 32U
2007 struct memcg_stock_pcp {
2008 struct mem_cgroup *cached; /* this never be root cgroup */
2009 unsigned int nr_pages;
2010 struct work_struct work;
2011 unsigned long flags;
2012 #define FLUSHING_CACHED_CHARGE (0)
2014 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2015 static DEFINE_MUTEX(percpu_charge_mutex);
2018 * Try to consume stocked charge on this cpu. If success, one page is consumed
2019 * from local stock and true is returned. If the stock is 0 or charges from a
2020 * cgroup which is not current target, returns false. This stock will be
2023 static bool consume_stock(struct mem_cgroup *mem)
2025 struct memcg_stock_pcp *stock;
2028 stock = &get_cpu_var(memcg_stock);
2029 if (mem == stock->cached && stock->nr_pages)
2031 else /* need to call res_counter_charge */
2033 put_cpu_var(memcg_stock);
2038 * Returns stocks cached in percpu to res_counter and reset cached information.
2040 static void drain_stock(struct memcg_stock_pcp *stock)
2042 struct mem_cgroup *old = stock->cached;
2044 if (stock->nr_pages) {
2045 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2047 res_counter_uncharge(&old->res, bytes);
2048 if (do_swap_account)
2049 res_counter_uncharge(&old->memsw, bytes);
2050 stock->nr_pages = 0;
2052 stock->cached = NULL;
2056 * This must be called under preempt disabled or must be called by
2057 * a thread which is pinned to local cpu.
2059 static void drain_local_stock(struct work_struct *dummy)
2061 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2063 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2067 * Cache charges(val) which is from res_counter, to local per_cpu area.
2068 * This will be consumed by consume_stock() function, later.
2070 static void refill_stock(struct mem_cgroup *mem, unsigned int nr_pages)
2072 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2074 if (stock->cached != mem) { /* reset if necessary */
2076 stock->cached = mem;
2078 stock->nr_pages += nr_pages;
2079 put_cpu_var(memcg_stock);
2083 * Tries to drain stocked charges in other cpus. This function is asynchronous
2084 * and just put a work per cpu for draining localy on each cpu. Caller can
2085 * expects some charges will be back to res_counter later but cannot wait for
2088 static void drain_all_stock_async(struct mem_cgroup *root_mem)
2092 * If someone calls draining, avoid adding more kworker runs.
2094 if (!mutex_trylock(&percpu_charge_mutex))
2096 /* Notify other cpus that system-wide "drain" is running */
2099 * Get a hint for avoiding draining charges on the current cpu,
2100 * which must be exhausted by our charging. It is not required that
2101 * this be a precise check, so we use raw_smp_processor_id() instead of
2102 * getcpu()/putcpu().
2104 curcpu = raw_smp_processor_id();
2105 for_each_online_cpu(cpu) {
2106 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2107 struct mem_cgroup *mem;
2112 mem = stock->cached;
2115 if (mem != root_mem) {
2116 if (!root_mem->use_hierarchy)
2118 /* check whether "mem" is under tree of "root_mem" */
2119 if (!css_is_ancestor(&mem->css, &root_mem->css))
2122 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2123 schedule_work_on(cpu, &stock->work);
2126 mutex_unlock(&percpu_charge_mutex);
2127 /* We don't wait for flush_work */
2130 /* This is a synchronous drain interface. */
2131 static void drain_all_stock_sync(void)
2133 /* called when force_empty is called */
2134 mutex_lock(&percpu_charge_mutex);
2135 schedule_on_each_cpu(drain_local_stock);
2136 mutex_unlock(&percpu_charge_mutex);
2140 * This function drains percpu counter value from DEAD cpu and
2141 * move it to local cpu. Note that this function can be preempted.
2143 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
2147 spin_lock(&mem->pcp_counter_lock);
2148 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2149 long x = per_cpu(mem->stat->count[i], cpu);
2151 per_cpu(mem->stat->count[i], cpu) = 0;
2152 mem->nocpu_base.count[i] += x;
2154 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2155 unsigned long x = per_cpu(mem->stat->events[i], cpu);
2157 per_cpu(mem->stat->events[i], cpu) = 0;
2158 mem->nocpu_base.events[i] += x;
2160 /* need to clear ON_MOVE value, works as a kind of lock. */
2161 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2162 spin_unlock(&mem->pcp_counter_lock);
2165 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
2167 int idx = MEM_CGROUP_ON_MOVE;
2169 spin_lock(&mem->pcp_counter_lock);
2170 per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
2171 spin_unlock(&mem->pcp_counter_lock);
2174 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2175 unsigned long action,
2178 int cpu = (unsigned long)hcpu;
2179 struct memcg_stock_pcp *stock;
2180 struct mem_cgroup *iter;
2182 if ((action == CPU_ONLINE)) {
2183 for_each_mem_cgroup_all(iter)
2184 synchronize_mem_cgroup_on_move(iter, cpu);
2188 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2191 for_each_mem_cgroup_all(iter)
2192 mem_cgroup_drain_pcp_counter(iter, cpu);
2194 stock = &per_cpu(memcg_stock, cpu);
2200 /* See __mem_cgroup_try_charge() for details */
2202 CHARGE_OK, /* success */
2203 CHARGE_RETRY, /* need to retry but retry is not bad */
2204 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2205 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2206 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2209 static int mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
2210 unsigned int nr_pages, bool oom_check)
2212 unsigned long csize = nr_pages * PAGE_SIZE;
2213 struct mem_cgroup *mem_over_limit;
2214 struct res_counter *fail_res;
2215 unsigned long flags = 0;
2218 ret = res_counter_charge(&mem->res, csize, &fail_res);
2221 if (!do_swap_account)
2223 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
2227 res_counter_uncharge(&mem->res, csize);
2228 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2229 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2231 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2233 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2234 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2236 * Never reclaim on behalf of optional batching, retry with a
2237 * single page instead.
2239 if (nr_pages == CHARGE_BATCH)
2240 return CHARGE_RETRY;
2242 if (!(gfp_mask & __GFP_WAIT))
2243 return CHARGE_WOULDBLOCK;
2245 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2246 gfp_mask, flags, NULL);
2247 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2248 return CHARGE_RETRY;
2250 * Even though the limit is exceeded at this point, reclaim
2251 * may have been able to free some pages. Retry the charge
2252 * before killing the task.
2254 * Only for regular pages, though: huge pages are rather
2255 * unlikely to succeed so close to the limit, and we fall back
2256 * to regular pages anyway in case of failure.
2258 if (nr_pages == 1 && ret)
2259 return CHARGE_RETRY;
2262 * At task move, charge accounts can be doubly counted. So, it's
2263 * better to wait until the end of task_move if something is going on.
2265 if (mem_cgroup_wait_acct_move(mem_over_limit))
2266 return CHARGE_RETRY;
2268 /* If we don't need to call oom-killer at el, return immediately */
2270 return CHARGE_NOMEM;
2272 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2273 return CHARGE_OOM_DIE;
2275 return CHARGE_RETRY;
2279 * Unlike exported interface, "oom" parameter is added. if oom==true,
2280 * oom-killer can be invoked.
2282 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2284 unsigned int nr_pages,
2285 struct mem_cgroup **memcg,
2288 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2289 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2290 struct mem_cgroup *mem = NULL;
2294 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2295 * in system level. So, allow to go ahead dying process in addition to
2298 if (unlikely(test_thread_flag(TIF_MEMDIE)
2299 || fatal_signal_pending(current)))
2303 * We always charge the cgroup the mm_struct belongs to.
2304 * The mm_struct's mem_cgroup changes on task migration if the
2305 * thread group leader migrates. It's possible that mm is not
2306 * set, if so charge the init_mm (happens for pagecache usage).
2311 if (*memcg) { /* css should be a valid one */
2313 VM_BUG_ON(css_is_removed(&mem->css));
2314 if (mem_cgroup_is_root(mem))
2316 if (nr_pages == 1 && consume_stock(mem))
2320 struct task_struct *p;
2323 p = rcu_dereference(mm->owner);
2325 * Because we don't have task_lock(), "p" can exit.
2326 * In that case, "mem" can point to root or p can be NULL with
2327 * race with swapoff. Then, we have small risk of mis-accouning.
2328 * But such kind of mis-account by race always happens because
2329 * we don't have cgroup_mutex(). It's overkill and we allo that
2331 * (*) swapoff at el will charge against mm-struct not against
2332 * task-struct. So, mm->owner can be NULL.
2334 mem = mem_cgroup_from_task(p);
2335 if (!mem || mem_cgroup_is_root(mem)) {
2339 if (nr_pages == 1 && consume_stock(mem)) {
2341 * It seems dagerous to access memcg without css_get().
2342 * But considering how consume_stok works, it's not
2343 * necessary. If consume_stock success, some charges
2344 * from this memcg are cached on this cpu. So, we
2345 * don't need to call css_get()/css_tryget() before
2346 * calling consume_stock().
2351 /* after here, we may be blocked. we need to get refcnt */
2352 if (!css_tryget(&mem->css)) {
2362 /* If killed, bypass charge */
2363 if (fatal_signal_pending(current)) {
2369 if (oom && !nr_oom_retries) {
2371 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2374 ret = mem_cgroup_do_charge(mem, gfp_mask, batch, oom_check);
2378 case CHARGE_RETRY: /* not in OOM situation but retry */
2383 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2386 case CHARGE_NOMEM: /* OOM routine works */
2391 /* If oom, we never return -ENOMEM */
2394 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2398 } while (ret != CHARGE_OK);
2400 if (batch > nr_pages)
2401 refill_stock(mem, batch - nr_pages);
2415 * Somemtimes we have to undo a charge we got by try_charge().
2416 * This function is for that and do uncharge, put css's refcnt.
2417 * gotten by try_charge().
2419 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2420 unsigned int nr_pages)
2422 if (!mem_cgroup_is_root(mem)) {
2423 unsigned long bytes = nr_pages * PAGE_SIZE;
2425 res_counter_uncharge(&mem->res, bytes);
2426 if (do_swap_account)
2427 res_counter_uncharge(&mem->memsw, bytes);
2432 * A helper function to get mem_cgroup from ID. must be called under
2433 * rcu_read_lock(). The caller must check css_is_removed() or some if
2434 * it's concern. (dropping refcnt from swap can be called against removed
2437 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2439 struct cgroup_subsys_state *css;
2441 /* ID 0 is unused ID */
2444 css = css_lookup(&mem_cgroup_subsys, id);
2447 return container_of(css, struct mem_cgroup, css);
2450 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2452 struct mem_cgroup *mem = NULL;
2453 struct page_cgroup *pc;
2457 VM_BUG_ON(!PageLocked(page));
2459 pc = lookup_page_cgroup(page);
2460 lock_page_cgroup(pc);
2461 if (PageCgroupUsed(pc)) {
2462 mem = pc->mem_cgroup;
2463 if (mem && !css_tryget(&mem->css))
2465 } else if (PageSwapCache(page)) {
2466 ent.val = page_private(page);
2467 id = lookup_swap_cgroup(ent);
2469 mem = mem_cgroup_lookup(id);
2470 if (mem && !css_tryget(&mem->css))
2474 unlock_page_cgroup(pc);
2478 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2480 unsigned int nr_pages,
2481 struct page_cgroup *pc,
2482 enum charge_type ctype)
2484 lock_page_cgroup(pc);
2485 if (unlikely(PageCgroupUsed(pc))) {
2486 unlock_page_cgroup(pc);
2487 __mem_cgroup_cancel_charge(mem, nr_pages);
2491 * we don't need page_cgroup_lock about tail pages, becase they are not
2492 * accessed by any other context at this point.
2494 pc->mem_cgroup = mem;
2496 * We access a page_cgroup asynchronously without lock_page_cgroup().
2497 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2498 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2499 * before USED bit, we need memory barrier here.
2500 * See mem_cgroup_add_lru_list(), etc.
2504 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2505 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2506 SetPageCgroupCache(pc);
2507 SetPageCgroupUsed(pc);
2509 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2510 ClearPageCgroupCache(pc);
2511 SetPageCgroupUsed(pc);
2517 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2518 unlock_page_cgroup(pc);
2520 * "charge_statistics" updated event counter. Then, check it.
2521 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2522 * if they exceeds softlimit.
2524 memcg_check_events(mem, page);
2527 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2529 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2530 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2532 * Because tail pages are not marked as "used", set it. We're under
2533 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2535 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2537 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2538 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2539 unsigned long flags;
2541 if (mem_cgroup_disabled())
2544 * We have no races with charge/uncharge but will have races with
2545 * page state accounting.
2547 move_lock_page_cgroup(head_pc, &flags);
2549 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2550 smp_wmb(); /* see __commit_charge() */
2551 if (PageCgroupAcctLRU(head_pc)) {
2553 struct mem_cgroup_per_zone *mz;
2556 * LRU flags cannot be copied because we need to add tail
2557 *.page to LRU by generic call and our hook will be called.
2558 * We hold lru_lock, then, reduce counter directly.
2560 lru = page_lru(head);
2561 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2562 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2564 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2565 move_unlock_page_cgroup(head_pc, &flags);
2570 * mem_cgroup_move_account - move account of the page
2572 * @nr_pages: number of regular pages (>1 for huge pages)
2573 * @pc: page_cgroup of the page.
2574 * @from: mem_cgroup which the page is moved from.
2575 * @to: mem_cgroup which the page is moved to. @from != @to.
2576 * @uncharge: whether we should call uncharge and css_put against @from.
2578 * The caller must confirm following.
2579 * - page is not on LRU (isolate_page() is useful.)
2580 * - compound_lock is held when nr_pages > 1
2582 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2583 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2584 * true, this function does "uncharge" from old cgroup, but it doesn't if
2585 * @uncharge is false, so a caller should do "uncharge".
2587 static int mem_cgroup_move_account(struct page *page,
2588 unsigned int nr_pages,
2589 struct page_cgroup *pc,
2590 struct mem_cgroup *from,
2591 struct mem_cgroup *to,
2594 unsigned long flags;
2597 VM_BUG_ON(from == to);
2598 VM_BUG_ON(PageLRU(page));
2600 * The page is isolated from LRU. So, collapse function
2601 * will not handle this page. But page splitting can happen.
2602 * Do this check under compound_page_lock(). The caller should
2606 if (nr_pages > 1 && !PageTransHuge(page))
2609 lock_page_cgroup(pc);
2612 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2615 move_lock_page_cgroup(pc, &flags);
2617 if (PageCgroupFileMapped(pc)) {
2618 /* Update mapped_file data for mem_cgroup */
2620 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2621 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2624 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2626 /* This is not "cancel", but cancel_charge does all we need. */
2627 __mem_cgroup_cancel_charge(from, nr_pages);
2629 /* caller should have done css_get */
2630 pc->mem_cgroup = to;
2631 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2633 * We charges against "to" which may not have any tasks. Then, "to"
2634 * can be under rmdir(). But in current implementation, caller of
2635 * this function is just force_empty() and move charge, so it's
2636 * guaranteed that "to" is never removed. So, we don't check rmdir
2639 move_unlock_page_cgroup(pc, &flags);
2642 unlock_page_cgroup(pc);
2646 memcg_check_events(to, page);
2647 memcg_check_events(from, page);
2653 * move charges to its parent.
2656 static int mem_cgroup_move_parent(struct page *page,
2657 struct page_cgroup *pc,
2658 struct mem_cgroup *child,
2661 struct cgroup *cg = child->css.cgroup;
2662 struct cgroup *pcg = cg->parent;
2663 struct mem_cgroup *parent;
2664 unsigned int nr_pages;
2665 unsigned long uninitialized_var(flags);
2673 if (!get_page_unless_zero(page))
2675 if (isolate_lru_page(page))
2678 nr_pages = hpage_nr_pages(page);
2680 parent = mem_cgroup_from_cont(pcg);
2681 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2686 flags = compound_lock_irqsave(page);
2688 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2690 __mem_cgroup_cancel_charge(parent, nr_pages);
2693 compound_unlock_irqrestore(page, flags);
2695 putback_lru_page(page);
2703 * Charge the memory controller for page usage.
2705 * 0 if the charge was successful
2706 * < 0 if the cgroup is over its limit
2708 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2709 gfp_t gfp_mask, enum charge_type ctype)
2711 struct mem_cgroup *mem = NULL;
2712 unsigned int nr_pages = 1;
2713 struct page_cgroup *pc;
2717 if (PageTransHuge(page)) {
2718 nr_pages <<= compound_order(page);
2719 VM_BUG_ON(!PageTransHuge(page));
2721 * Never OOM-kill a process for a huge page. The
2722 * fault handler will fall back to regular pages.
2727 pc = lookup_page_cgroup(page);
2728 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2730 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &mem, oom);
2734 __mem_cgroup_commit_charge(mem, page, nr_pages, pc, ctype);
2738 int mem_cgroup_newpage_charge(struct page *page,
2739 struct mm_struct *mm, gfp_t gfp_mask)
2741 if (mem_cgroup_disabled())
2744 * If already mapped, we don't have to account.
2745 * If page cache, page->mapping has address_space.
2746 * But page->mapping may have out-of-use anon_vma pointer,
2747 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2750 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2754 return mem_cgroup_charge_common(page, mm, gfp_mask,
2755 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2759 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2760 enum charge_type ctype);
2763 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *mem,
2764 enum charge_type ctype)
2766 struct page_cgroup *pc = lookup_page_cgroup(page);
2768 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2769 * is already on LRU. It means the page may on some other page_cgroup's
2770 * LRU. Take care of it.
2772 mem_cgroup_lru_del_before_commit(page);
2773 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
2774 mem_cgroup_lru_add_after_commit(page);
2778 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2781 struct mem_cgroup *mem = NULL;
2784 if (mem_cgroup_disabled())
2786 if (PageCompound(page))
2789 * Corner case handling. This is called from add_to_page_cache()
2790 * in usual. But some FS (shmem) precharges this page before calling it
2791 * and call add_to_page_cache() with GFP_NOWAIT.
2793 * For GFP_NOWAIT case, the page may be pre-charged before calling
2794 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2795 * charge twice. (It works but has to pay a bit larger cost.)
2796 * And when the page is SwapCache, it should take swap information
2797 * into account. This is under lock_page() now.
2799 if (!(gfp_mask & __GFP_WAIT)) {
2800 struct page_cgroup *pc;
2802 pc = lookup_page_cgroup(page);
2805 lock_page_cgroup(pc);
2806 if (PageCgroupUsed(pc)) {
2807 unlock_page_cgroup(pc);
2810 unlock_page_cgroup(pc);
2816 if (page_is_file_cache(page)) {
2817 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &mem, true);
2822 * FUSE reuses pages without going through the final
2823 * put that would remove them from the LRU list, make
2824 * sure that they get relinked properly.
2826 __mem_cgroup_commit_charge_lrucare(page, mem,
2827 MEM_CGROUP_CHARGE_TYPE_CACHE);
2831 if (PageSwapCache(page)) {
2832 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2834 __mem_cgroup_commit_charge_swapin(page, mem,
2835 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2837 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2838 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2844 * While swap-in, try_charge -> commit or cancel, the page is locked.
2845 * And when try_charge() successfully returns, one refcnt to memcg without
2846 * struct page_cgroup is acquired. This refcnt will be consumed by
2847 * "commit()" or removed by "cancel()"
2849 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2851 gfp_t mask, struct mem_cgroup **ptr)
2853 struct mem_cgroup *mem;
2858 if (mem_cgroup_disabled())
2861 if (!do_swap_account)
2864 * A racing thread's fault, or swapoff, may have already updated
2865 * the pte, and even removed page from swap cache: in those cases
2866 * do_swap_page()'s pte_same() test will fail; but there's also a
2867 * KSM case which does need to charge the page.
2869 if (!PageSwapCache(page))
2871 mem = try_get_mem_cgroup_from_page(page);
2875 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2881 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2885 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2886 enum charge_type ctype)
2888 if (mem_cgroup_disabled())
2892 cgroup_exclude_rmdir(&ptr->css);
2894 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2896 * Now swap is on-memory. This means this page may be
2897 * counted both as mem and swap....double count.
2898 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2899 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2900 * may call delete_from_swap_cache() before reach here.
2902 if (do_swap_account && PageSwapCache(page)) {
2903 swp_entry_t ent = {.val = page_private(page)};
2905 struct mem_cgroup *memcg;
2907 id = swap_cgroup_record(ent, 0);
2909 memcg = mem_cgroup_lookup(id);
2912 * This recorded memcg can be obsolete one. So, avoid
2913 * calling css_tryget
2915 if (!mem_cgroup_is_root(memcg))
2916 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2917 mem_cgroup_swap_statistics(memcg, false);
2918 mem_cgroup_put(memcg);
2923 * At swapin, we may charge account against cgroup which has no tasks.
2924 * So, rmdir()->pre_destroy() can be called while we do this charge.
2925 * In that case, we need to call pre_destroy() again. check it here.
2927 cgroup_release_and_wakeup_rmdir(&ptr->css);
2930 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2932 __mem_cgroup_commit_charge_swapin(page, ptr,
2933 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2936 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2938 if (mem_cgroup_disabled())
2942 __mem_cgroup_cancel_charge(mem, 1);
2945 static void mem_cgroup_do_uncharge(struct mem_cgroup *mem,
2946 unsigned int nr_pages,
2947 const enum charge_type ctype)
2949 struct memcg_batch_info *batch = NULL;
2950 bool uncharge_memsw = true;
2952 /* If swapout, usage of swap doesn't decrease */
2953 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2954 uncharge_memsw = false;
2956 batch = ¤t->memcg_batch;
2958 * In usual, we do css_get() when we remember memcg pointer.
2959 * But in this case, we keep res->usage until end of a series of
2960 * uncharges. Then, it's ok to ignore memcg's refcnt.
2965 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2966 * In those cases, all pages freed continuously can be expected to be in
2967 * the same cgroup and we have chance to coalesce uncharges.
2968 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2969 * because we want to do uncharge as soon as possible.
2972 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2973 goto direct_uncharge;
2976 goto direct_uncharge;
2979 * In typical case, batch->memcg == mem. This means we can
2980 * merge a series of uncharges to an uncharge of res_counter.
2981 * If not, we uncharge res_counter ony by one.
2983 if (batch->memcg != mem)
2984 goto direct_uncharge;
2985 /* remember freed charge and uncharge it later */
2988 batch->memsw_nr_pages++;
2991 res_counter_uncharge(&mem->res, nr_pages * PAGE_SIZE);
2993 res_counter_uncharge(&mem->memsw, nr_pages * PAGE_SIZE);
2994 if (unlikely(batch->memcg != mem))
2995 memcg_oom_recover(mem);
3000 * uncharge if !page_mapped(page)
3002 static struct mem_cgroup *
3003 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
3005 struct mem_cgroup *mem = NULL;
3006 unsigned int nr_pages = 1;
3007 struct page_cgroup *pc;
3009 if (mem_cgroup_disabled())
3012 if (PageSwapCache(page))
3015 if (PageTransHuge(page)) {
3016 nr_pages <<= compound_order(page);
3017 VM_BUG_ON(!PageTransHuge(page));
3020 * Check if our page_cgroup is valid
3022 pc = lookup_page_cgroup(page);
3023 if (unlikely(!pc || !PageCgroupUsed(pc)))
3026 lock_page_cgroup(pc);
3028 mem = pc->mem_cgroup;
3030 if (!PageCgroupUsed(pc))
3034 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3035 case MEM_CGROUP_CHARGE_TYPE_DROP:
3036 /* See mem_cgroup_prepare_migration() */
3037 if (page_mapped(page) || PageCgroupMigration(pc))
3040 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3041 if (!PageAnon(page)) { /* Shared memory */
3042 if (page->mapping && !page_is_file_cache(page))
3044 } else if (page_mapped(page)) /* Anon */
3051 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -nr_pages);
3053 ClearPageCgroupUsed(pc);
3055 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3056 * freed from LRU. This is safe because uncharged page is expected not
3057 * to be reused (freed soon). Exception is SwapCache, it's handled by
3058 * special functions.
3061 unlock_page_cgroup(pc);
3063 * even after unlock, we have mem->res.usage here and this memcg
3064 * will never be freed.
3066 memcg_check_events(mem, page);
3067 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3068 mem_cgroup_swap_statistics(mem, true);
3069 mem_cgroup_get(mem);
3071 if (!mem_cgroup_is_root(mem))
3072 mem_cgroup_do_uncharge(mem, nr_pages, ctype);
3077 unlock_page_cgroup(pc);
3081 void mem_cgroup_uncharge_page(struct page *page)
3084 if (page_mapped(page))
3086 if (page->mapping && !PageAnon(page))
3088 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3091 void mem_cgroup_uncharge_cache_page(struct page *page)
3093 VM_BUG_ON(page_mapped(page));
3094 VM_BUG_ON(page->mapping);
3095 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3099 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3100 * In that cases, pages are freed continuously and we can expect pages
3101 * are in the same memcg. All these calls itself limits the number of
3102 * pages freed at once, then uncharge_start/end() is called properly.
3103 * This may be called prural(2) times in a context,
3106 void mem_cgroup_uncharge_start(void)
3108 current->memcg_batch.do_batch++;
3109 /* We can do nest. */
3110 if (current->memcg_batch.do_batch == 1) {
3111 current->memcg_batch.memcg = NULL;
3112 current->memcg_batch.nr_pages = 0;
3113 current->memcg_batch.memsw_nr_pages = 0;
3117 void mem_cgroup_uncharge_end(void)
3119 struct memcg_batch_info *batch = ¤t->memcg_batch;
3121 if (!batch->do_batch)
3125 if (batch->do_batch) /* If stacked, do nothing. */
3131 * This "batch->memcg" is valid without any css_get/put etc...
3132 * bacause we hide charges behind us.
3134 if (batch->nr_pages)
3135 res_counter_uncharge(&batch->memcg->res,
3136 batch->nr_pages * PAGE_SIZE);
3137 if (batch->memsw_nr_pages)
3138 res_counter_uncharge(&batch->memcg->memsw,
3139 batch->memsw_nr_pages * PAGE_SIZE);
3140 memcg_oom_recover(batch->memcg);
3141 /* forget this pointer (for sanity check) */
3142 batch->memcg = NULL;
3147 * called after __delete_from_swap_cache() and drop "page" account.
3148 * memcg information is recorded to swap_cgroup of "ent"
3151 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3153 struct mem_cgroup *memcg;
3154 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3156 if (!swapout) /* this was a swap cache but the swap is unused ! */
3157 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3159 memcg = __mem_cgroup_uncharge_common(page, ctype);
3162 * record memcg information, if swapout && memcg != NULL,
3163 * mem_cgroup_get() was called in uncharge().
3165 if (do_swap_account && swapout && memcg)
3166 swap_cgroup_record(ent, css_id(&memcg->css));
3170 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3172 * called from swap_entry_free(). remove record in swap_cgroup and
3173 * uncharge "memsw" account.
3175 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3177 struct mem_cgroup *memcg;
3180 if (!do_swap_account)
3183 id = swap_cgroup_record(ent, 0);
3185 memcg = mem_cgroup_lookup(id);
3188 * We uncharge this because swap is freed.
3189 * This memcg can be obsolete one. We avoid calling css_tryget
3191 if (!mem_cgroup_is_root(memcg))
3192 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3193 mem_cgroup_swap_statistics(memcg, false);
3194 mem_cgroup_put(memcg);
3200 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3201 * @entry: swap entry to be moved
3202 * @from: mem_cgroup which the entry is moved from
3203 * @to: mem_cgroup which the entry is moved to
3204 * @need_fixup: whether we should fixup res_counters and refcounts.
3206 * It succeeds only when the swap_cgroup's record for this entry is the same
3207 * as the mem_cgroup's id of @from.
3209 * Returns 0 on success, -EINVAL on failure.
3211 * The caller must have charged to @to, IOW, called res_counter_charge() about
3212 * both res and memsw, and called css_get().
3214 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3215 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3217 unsigned short old_id, new_id;
3219 old_id = css_id(&from->css);
3220 new_id = css_id(&to->css);
3222 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3223 mem_cgroup_swap_statistics(from, false);
3224 mem_cgroup_swap_statistics(to, true);
3226 * This function is only called from task migration context now.
3227 * It postpones res_counter and refcount handling till the end
3228 * of task migration(mem_cgroup_clear_mc()) for performance
3229 * improvement. But we cannot postpone mem_cgroup_get(to)
3230 * because if the process that has been moved to @to does
3231 * swap-in, the refcount of @to might be decreased to 0.
3235 if (!mem_cgroup_is_root(from))
3236 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3237 mem_cgroup_put(from);
3239 * we charged both to->res and to->memsw, so we should
3242 if (!mem_cgroup_is_root(to))
3243 res_counter_uncharge(&to->res, PAGE_SIZE);
3250 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3251 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3258 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3261 int mem_cgroup_prepare_migration(struct page *page,
3262 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3264 struct mem_cgroup *mem = NULL;
3265 struct page_cgroup *pc;
3266 enum charge_type ctype;
3271 VM_BUG_ON(PageTransHuge(page));
3272 if (mem_cgroup_disabled())
3275 pc = lookup_page_cgroup(page);
3276 lock_page_cgroup(pc);
3277 if (PageCgroupUsed(pc)) {
3278 mem = pc->mem_cgroup;
3281 * At migrating an anonymous page, its mapcount goes down
3282 * to 0 and uncharge() will be called. But, even if it's fully
3283 * unmapped, migration may fail and this page has to be
3284 * charged again. We set MIGRATION flag here and delay uncharge
3285 * until end_migration() is called
3287 * Corner Case Thinking
3289 * When the old page was mapped as Anon and it's unmap-and-freed
3290 * while migration was ongoing.
3291 * If unmap finds the old page, uncharge() of it will be delayed
3292 * until end_migration(). If unmap finds a new page, it's
3293 * uncharged when it make mapcount to be 1->0. If unmap code
3294 * finds swap_migration_entry, the new page will not be mapped
3295 * and end_migration() will find it(mapcount==0).
3298 * When the old page was mapped but migraion fails, the kernel
3299 * remaps it. A charge for it is kept by MIGRATION flag even
3300 * if mapcount goes down to 0. We can do remap successfully
3301 * without charging it again.
3304 * The "old" page is under lock_page() until the end of
3305 * migration, so, the old page itself will not be swapped-out.
3306 * If the new page is swapped out before end_migraton, our
3307 * hook to usual swap-out path will catch the event.
3310 SetPageCgroupMigration(pc);
3312 unlock_page_cgroup(pc);
3314 * If the page is not charged at this point,
3321 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3322 css_put(&mem->css);/* drop extra refcnt */
3323 if (ret || *ptr == NULL) {
3324 if (PageAnon(page)) {
3325 lock_page_cgroup(pc);
3326 ClearPageCgroupMigration(pc);
3327 unlock_page_cgroup(pc);
3329 * The old page may be fully unmapped while we kept it.
3331 mem_cgroup_uncharge_page(page);
3336 * We charge new page before it's used/mapped. So, even if unlock_page()
3337 * is called before end_migration, we can catch all events on this new
3338 * page. In the case new page is migrated but not remapped, new page's
3339 * mapcount will be finally 0 and we call uncharge in end_migration().
3341 pc = lookup_page_cgroup(newpage);
3343 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3344 else if (page_is_file_cache(page))
3345 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3347 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3348 __mem_cgroup_commit_charge(mem, page, 1, pc, ctype);
3352 /* remove redundant charge if migration failed*/
3353 void mem_cgroup_end_migration(struct mem_cgroup *mem,
3354 struct page *oldpage, struct page *newpage, bool migration_ok)
3356 struct page *used, *unused;
3357 struct page_cgroup *pc;
3361 /* blocks rmdir() */
3362 cgroup_exclude_rmdir(&mem->css);
3363 if (!migration_ok) {
3371 * We disallowed uncharge of pages under migration because mapcount
3372 * of the page goes down to zero, temporarly.
3373 * Clear the flag and check the page should be charged.
3375 pc = lookup_page_cgroup(oldpage);
3376 lock_page_cgroup(pc);
3377 ClearPageCgroupMigration(pc);
3378 unlock_page_cgroup(pc);
3380 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3383 * If a page is a file cache, radix-tree replacement is very atomic
3384 * and we can skip this check. When it was an Anon page, its mapcount
3385 * goes down to 0. But because we added MIGRATION flage, it's not
3386 * uncharged yet. There are several case but page->mapcount check
3387 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3388 * check. (see prepare_charge() also)
3391 mem_cgroup_uncharge_page(used);
3393 * At migration, we may charge account against cgroup which has no
3395 * So, rmdir()->pre_destroy() can be called while we do this charge.
3396 * In that case, we need to call pre_destroy() again. check it here.
3398 cgroup_release_and_wakeup_rmdir(&mem->css);
3402 * A call to try to shrink memory usage on charge failure at shmem's swapin.
3403 * Calling hierarchical_reclaim is not enough because we should update
3404 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3405 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3406 * not from the memcg which this page would be charged to.
3407 * try_charge_swapin does all of these works properly.
3409 int mem_cgroup_shmem_charge_fallback(struct page *page,
3410 struct mm_struct *mm,
3413 struct mem_cgroup *mem;
3416 if (mem_cgroup_disabled())
3419 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3421 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3427 * At replace page cache, newpage is not under any memcg but it's on
3428 * LRU. So, this function doesn't touch res_counter but handles LRU
3429 * in correct way. Both pages are locked so we cannot race with uncharge.
3431 void mem_cgroup_replace_page_cache(struct page *oldpage,
3432 struct page *newpage)
3434 struct mem_cgroup *memcg;
3435 struct page_cgroup *pc;
3437 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3438 unsigned long flags;
3440 if (mem_cgroup_disabled())
3443 pc = lookup_page_cgroup(oldpage);
3444 /* fix accounting on old pages */
3445 lock_page_cgroup(pc);
3446 memcg = pc->mem_cgroup;
3447 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1);
3448 ClearPageCgroupUsed(pc);
3449 unlock_page_cgroup(pc);
3451 if (PageSwapBacked(oldpage))
3452 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3454 zone = page_zone(newpage);
3455 pc = lookup_page_cgroup(newpage);
3457 * Even if newpage->mapping was NULL before starting replacement,
3458 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3459 * LRU while we overwrite pc->mem_cgroup.
3461 spin_lock_irqsave(&zone->lru_lock, flags);
3462 if (PageLRU(newpage))
3463 del_page_from_lru_list(zone, newpage, page_lru(newpage));
3464 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type);
3465 if (PageLRU(newpage))
3466 add_page_to_lru_list(zone, newpage, page_lru(newpage));
3467 spin_unlock_irqrestore(&zone->lru_lock, flags);
3470 #ifdef CONFIG_DEBUG_VM
3471 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3473 struct page_cgroup *pc;
3475 pc = lookup_page_cgroup(page);
3476 if (likely(pc) && PageCgroupUsed(pc))
3481 bool mem_cgroup_bad_page_check(struct page *page)
3483 if (mem_cgroup_disabled())
3486 return lookup_page_cgroup_used(page) != NULL;
3489 void mem_cgroup_print_bad_page(struct page *page)
3491 struct page_cgroup *pc;
3493 pc = lookup_page_cgroup_used(page);
3498 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3499 pc, pc->flags, pc->mem_cgroup);
3501 path = kmalloc(PATH_MAX, GFP_KERNEL);
3504 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3509 printk(KERN_CONT "(%s)\n",
3510 (ret < 0) ? "cannot get the path" : path);
3516 static DEFINE_MUTEX(set_limit_mutex);
3518 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3519 unsigned long long val)
3522 u64 memswlimit, memlimit;
3524 int children = mem_cgroup_count_children(memcg);
3525 u64 curusage, oldusage;
3529 * For keeping hierarchical_reclaim simple, how long we should retry
3530 * is depends on callers. We set our retry-count to be function
3531 * of # of children which we should visit in this loop.
3533 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3535 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3538 while (retry_count) {
3539 if (signal_pending(current)) {
3544 * Rather than hide all in some function, I do this in
3545 * open coded manner. You see what this really does.
3546 * We have to guarantee mem->res.limit < mem->memsw.limit.
3548 mutex_lock(&set_limit_mutex);
3549 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3550 if (memswlimit < val) {
3552 mutex_unlock(&set_limit_mutex);
3556 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3560 ret = res_counter_set_limit(&memcg->res, val);
3562 if (memswlimit == val)
3563 memcg->memsw_is_minimum = true;
3565 memcg->memsw_is_minimum = false;
3567 mutex_unlock(&set_limit_mutex);
3572 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3573 MEM_CGROUP_RECLAIM_SHRINK,
3575 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3576 /* Usage is reduced ? */
3577 if (curusage >= oldusage)
3580 oldusage = curusage;
3582 if (!ret && enlarge)
3583 memcg_oom_recover(memcg);
3588 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3589 unsigned long long val)
3592 u64 memlimit, memswlimit, oldusage, curusage;
3593 int children = mem_cgroup_count_children(memcg);
3597 /* see mem_cgroup_resize_res_limit */
3598 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3599 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3600 while (retry_count) {
3601 if (signal_pending(current)) {
3606 * Rather than hide all in some function, I do this in
3607 * open coded manner. You see what this really does.
3608 * We have to guarantee mem->res.limit < mem->memsw.limit.
3610 mutex_lock(&set_limit_mutex);
3611 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3612 if (memlimit > val) {
3614 mutex_unlock(&set_limit_mutex);
3617 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3618 if (memswlimit < val)
3620 ret = res_counter_set_limit(&memcg->memsw, val);
3622 if (memlimit == val)
3623 memcg->memsw_is_minimum = true;
3625 memcg->memsw_is_minimum = false;
3627 mutex_unlock(&set_limit_mutex);
3632 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3633 MEM_CGROUP_RECLAIM_NOSWAP |
3634 MEM_CGROUP_RECLAIM_SHRINK,
3636 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3637 /* Usage is reduced ? */
3638 if (curusage >= oldusage)
3641 oldusage = curusage;
3643 if (!ret && enlarge)
3644 memcg_oom_recover(memcg);
3648 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3650 unsigned long *total_scanned)
3652 unsigned long nr_reclaimed = 0;
3653 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3654 unsigned long reclaimed;
3656 struct mem_cgroup_tree_per_zone *mctz;
3657 unsigned long long excess;
3658 unsigned long nr_scanned;
3663 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3665 * This loop can run a while, specially if mem_cgroup's continuously
3666 * keep exceeding their soft limit and putting the system under
3673 mz = mem_cgroup_largest_soft_limit_node(mctz);
3678 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3680 MEM_CGROUP_RECLAIM_SOFT,
3682 nr_reclaimed += reclaimed;
3683 *total_scanned += nr_scanned;
3684 spin_lock(&mctz->lock);
3687 * If we failed to reclaim anything from this memory cgroup
3688 * it is time to move on to the next cgroup
3694 * Loop until we find yet another one.
3696 * By the time we get the soft_limit lock
3697 * again, someone might have aded the
3698 * group back on the RB tree. Iterate to
3699 * make sure we get a different mem.
3700 * mem_cgroup_largest_soft_limit_node returns
3701 * NULL if no other cgroup is present on
3705 __mem_cgroup_largest_soft_limit_node(mctz);
3707 css_put(&next_mz->mem->css);
3708 else /* next_mz == NULL or other memcg */
3712 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3713 excess = res_counter_soft_limit_excess(&mz->mem->res);
3715 * One school of thought says that we should not add
3716 * back the node to the tree if reclaim returns 0.
3717 * But our reclaim could return 0, simply because due
3718 * to priority we are exposing a smaller subset of
3719 * memory to reclaim from. Consider this as a longer
3722 /* If excess == 0, no tree ops */
3723 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3724 spin_unlock(&mctz->lock);
3725 css_put(&mz->mem->css);
3728 * Could not reclaim anything and there are no more
3729 * mem cgroups to try or we seem to be looping without
3730 * reclaiming anything.
3732 if (!nr_reclaimed &&
3734 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3736 } while (!nr_reclaimed);
3738 css_put(&next_mz->mem->css);
3739 return nr_reclaimed;
3743 * This routine traverse page_cgroup in given list and drop them all.
3744 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3746 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3747 int node, int zid, enum lru_list lru)
3750 struct mem_cgroup_per_zone *mz;
3751 struct page_cgroup *pc, *busy;
3752 unsigned long flags, loop;
3753 struct list_head *list;
3756 zone = &NODE_DATA(node)->node_zones[zid];
3757 mz = mem_cgroup_zoneinfo(mem, node, zid);
3758 list = &mz->lists[lru];
3760 loop = MEM_CGROUP_ZSTAT(mz, lru);
3761 /* give some margin against EBUSY etc...*/
3768 spin_lock_irqsave(&zone->lru_lock, flags);
3769 if (list_empty(list)) {
3770 spin_unlock_irqrestore(&zone->lru_lock, flags);
3773 pc = list_entry(list->prev, struct page_cgroup, lru);
3775 list_move(&pc->lru, list);
3777 spin_unlock_irqrestore(&zone->lru_lock, flags);
3780 spin_unlock_irqrestore(&zone->lru_lock, flags);
3782 page = lookup_cgroup_page(pc);
3784 ret = mem_cgroup_move_parent(page, pc, mem, GFP_KERNEL);
3788 if (ret == -EBUSY || ret == -EINVAL) {
3789 /* found lock contention or "pc" is obsolete. */
3796 if (!ret && !list_empty(list))
3802 * make mem_cgroup's charge to be 0 if there is no task.
3803 * This enables deleting this mem_cgroup.
3805 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3808 int node, zid, shrink;
3809 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3810 struct cgroup *cgrp = mem->css.cgroup;
3815 /* should free all ? */
3821 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3824 if (signal_pending(current))
3826 /* This is for making all *used* pages to be on LRU. */
3827 lru_add_drain_all();
3828 drain_all_stock_sync();
3830 mem_cgroup_start_move(mem);
3831 for_each_node_state(node, N_HIGH_MEMORY) {
3832 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3835 ret = mem_cgroup_force_empty_list(mem,
3844 mem_cgroup_end_move(mem);
3845 memcg_oom_recover(mem);
3846 /* it seems parent cgroup doesn't have enough mem */
3850 /* "ret" should also be checked to ensure all lists are empty. */
3851 } while (mem->res.usage > 0 || ret);
3857 /* returns EBUSY if there is a task or if we come here twice. */
3858 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3862 /* we call try-to-free pages for make this cgroup empty */
3863 lru_add_drain_all();
3864 /* try to free all pages in this cgroup */
3866 while (nr_retries && mem->res.usage > 0) {
3869 if (signal_pending(current)) {
3873 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3874 false, get_swappiness(mem));
3877 /* maybe some writeback is necessary */
3878 congestion_wait(BLK_RW_ASYNC, HZ/10);
3883 /* try move_account...there may be some *locked* pages. */
3887 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3889 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3893 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3895 return mem_cgroup_from_cont(cont)->use_hierarchy;
3898 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3902 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3903 struct cgroup *parent = cont->parent;
3904 struct mem_cgroup *parent_mem = NULL;
3907 parent_mem = mem_cgroup_from_cont(parent);
3911 * If parent's use_hierarchy is set, we can't make any modifications
3912 * in the child subtrees. If it is unset, then the change can
3913 * occur, provided the current cgroup has no children.
3915 * For the root cgroup, parent_mem is NULL, we allow value to be
3916 * set if there are no children.
3918 if ((!parent_mem || !parent_mem->use_hierarchy) &&
3919 (val == 1 || val == 0)) {
3920 if (list_empty(&cont->children))
3921 mem->use_hierarchy = val;
3932 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *mem,
3933 enum mem_cgroup_stat_index idx)
3935 struct mem_cgroup *iter;
3938 /* Per-cpu values can be negative, use a signed accumulator */
3939 for_each_mem_cgroup_tree(iter, mem)
3940 val += mem_cgroup_read_stat(iter, idx);
3942 if (val < 0) /* race ? */
3947 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3951 if (!mem_cgroup_is_root(mem)) {
3953 return res_counter_read_u64(&mem->res, RES_USAGE);
3955 return res_counter_read_u64(&mem->memsw, RES_USAGE);
3958 val = mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_CACHE);
3959 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_RSS);
3962 val += mem_cgroup_recursive_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3964 return val << PAGE_SHIFT;
3967 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3969 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3973 type = MEMFILE_TYPE(cft->private);
3974 name = MEMFILE_ATTR(cft->private);
3977 if (name == RES_USAGE)
3978 val = mem_cgroup_usage(mem, false);
3980 val = res_counter_read_u64(&mem->res, name);
3983 if (name == RES_USAGE)
3984 val = mem_cgroup_usage(mem, true);
3986 val = res_counter_read_u64(&mem->memsw, name);
3995 * The user of this function is...
3998 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
4001 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4003 unsigned long long val;
4006 type = MEMFILE_TYPE(cft->private);
4007 name = MEMFILE_ATTR(cft->private);
4010 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
4014 /* This function does all necessary parse...reuse it */
4015 ret = res_counter_memparse_write_strategy(buffer, &val);
4019 ret = mem_cgroup_resize_limit(memcg, val);
4021 ret = mem_cgroup_resize_memsw_limit(memcg, val);
4023 case RES_SOFT_LIMIT:
4024 ret = res_counter_memparse_write_strategy(buffer, &val);
4028 * For memsw, soft limits are hard to implement in terms
4029 * of semantics, for now, we support soft limits for
4030 * control without swap
4033 ret = res_counter_set_soft_limit(&memcg->res, val);
4038 ret = -EINVAL; /* should be BUG() ? */
4044 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
4045 unsigned long long *mem_limit, unsigned long long *memsw_limit)
4047 struct cgroup *cgroup;
4048 unsigned long long min_limit, min_memsw_limit, tmp;
4050 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4051 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4052 cgroup = memcg->css.cgroup;
4053 if (!memcg->use_hierarchy)
4056 while (cgroup->parent) {
4057 cgroup = cgroup->parent;
4058 memcg = mem_cgroup_from_cont(cgroup);
4059 if (!memcg->use_hierarchy)
4061 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4062 min_limit = min(min_limit, tmp);
4063 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4064 min_memsw_limit = min(min_memsw_limit, tmp);
4067 *mem_limit = min_limit;
4068 *memsw_limit = min_memsw_limit;
4072 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4074 struct mem_cgroup *mem;
4077 mem = mem_cgroup_from_cont(cont);
4078 type = MEMFILE_TYPE(event);
4079 name = MEMFILE_ATTR(event);
4083 res_counter_reset_max(&mem->res);
4085 res_counter_reset_max(&mem->memsw);
4089 res_counter_reset_failcnt(&mem->res);
4091 res_counter_reset_failcnt(&mem->memsw);
4098 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4101 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4105 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4106 struct cftype *cft, u64 val)
4108 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4110 if (val >= (1 << NR_MOVE_TYPE))
4113 * We check this value several times in both in can_attach() and
4114 * attach(), so we need cgroup lock to prevent this value from being
4118 mem->move_charge_at_immigrate = val;
4124 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4125 struct cftype *cft, u64 val)
4132 /* For read statistics */
4150 struct mcs_total_stat {
4151 s64 stat[NR_MCS_STAT];
4157 } memcg_stat_strings[NR_MCS_STAT] = {
4158 {"cache", "total_cache"},
4159 {"rss", "total_rss"},
4160 {"mapped_file", "total_mapped_file"},
4161 {"pgpgin", "total_pgpgin"},
4162 {"pgpgout", "total_pgpgout"},
4163 {"swap", "total_swap"},
4164 {"pgfault", "total_pgfault"},
4165 {"pgmajfault", "total_pgmajfault"},
4166 {"inactive_anon", "total_inactive_anon"},
4167 {"active_anon", "total_active_anon"},
4168 {"inactive_file", "total_inactive_file"},
4169 {"active_file", "total_active_file"},
4170 {"unevictable", "total_unevictable"}
4175 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4180 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
4181 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4182 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
4183 s->stat[MCS_RSS] += val * PAGE_SIZE;
4184 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
4185 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4186 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGIN);
4187 s->stat[MCS_PGPGIN] += val;
4188 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGPGOUT);
4189 s->stat[MCS_PGPGOUT] += val;
4190 if (do_swap_account) {
4191 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
4192 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4194 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGFAULT);
4195 s->stat[MCS_PGFAULT] += val;
4196 val = mem_cgroup_read_events(mem, MEM_CGROUP_EVENTS_PGMAJFAULT);
4197 s->stat[MCS_PGMAJFAULT] += val;
4200 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
4201 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4202 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
4203 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4204 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
4205 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4206 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
4207 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4208 val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
4209 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4213 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
4215 struct mem_cgroup *iter;
4217 for_each_mem_cgroup_tree(iter, mem)
4218 mem_cgroup_get_local_stat(iter, s);
4222 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4225 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4226 unsigned long node_nr;
4227 struct cgroup *cont = m->private;
4228 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4230 total_nr = mem_cgroup_nr_lru_pages(mem_cont);
4231 seq_printf(m, "total=%lu", total_nr);
4232 for_each_node_state(nid, N_HIGH_MEMORY) {
4233 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid);
4234 seq_printf(m, " N%d=%lu", nid, node_nr);
4238 file_nr = mem_cgroup_nr_file_lru_pages(mem_cont);
4239 seq_printf(m, "file=%lu", file_nr);
4240 for_each_node_state(nid, N_HIGH_MEMORY) {
4241 node_nr = mem_cgroup_node_nr_file_lru_pages(mem_cont, nid);
4242 seq_printf(m, " N%d=%lu", nid, node_nr);
4246 anon_nr = mem_cgroup_nr_anon_lru_pages(mem_cont);
4247 seq_printf(m, "anon=%lu", anon_nr);
4248 for_each_node_state(nid, N_HIGH_MEMORY) {
4249 node_nr = mem_cgroup_node_nr_anon_lru_pages(mem_cont, nid);
4250 seq_printf(m, " N%d=%lu", nid, node_nr);
4254 unevictable_nr = mem_cgroup_nr_unevictable_lru_pages(mem_cont);
4255 seq_printf(m, "unevictable=%lu", unevictable_nr);
4256 for_each_node_state(nid, N_HIGH_MEMORY) {
4257 node_nr = mem_cgroup_node_nr_unevictable_lru_pages(mem_cont,
4259 seq_printf(m, " N%d=%lu", nid, node_nr);
4264 #endif /* CONFIG_NUMA */
4266 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4267 struct cgroup_map_cb *cb)
4269 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4270 struct mcs_total_stat mystat;
4273 memset(&mystat, 0, sizeof(mystat));
4274 mem_cgroup_get_local_stat(mem_cont, &mystat);
4277 for (i = 0; i < NR_MCS_STAT; i++) {
4278 if (i == MCS_SWAP && !do_swap_account)
4280 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4283 /* Hierarchical information */
4285 unsigned long long limit, memsw_limit;
4286 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4287 cb->fill(cb, "hierarchical_memory_limit", limit);
4288 if (do_swap_account)
4289 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4292 memset(&mystat, 0, sizeof(mystat));
4293 mem_cgroup_get_total_stat(mem_cont, &mystat);
4294 for (i = 0; i < NR_MCS_STAT; i++) {
4295 if (i == MCS_SWAP && !do_swap_account)
4297 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4300 #ifdef CONFIG_DEBUG_VM
4301 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4305 struct mem_cgroup_per_zone *mz;
4306 unsigned long recent_rotated[2] = {0, 0};
4307 unsigned long recent_scanned[2] = {0, 0};
4309 for_each_online_node(nid)
4310 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4311 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4313 recent_rotated[0] +=
4314 mz->reclaim_stat.recent_rotated[0];
4315 recent_rotated[1] +=
4316 mz->reclaim_stat.recent_rotated[1];
4317 recent_scanned[0] +=
4318 mz->reclaim_stat.recent_scanned[0];
4319 recent_scanned[1] +=
4320 mz->reclaim_stat.recent_scanned[1];
4322 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4323 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4324 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4325 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4332 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4334 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4336 return get_swappiness(memcg);
4339 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4342 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4343 struct mem_cgroup *parent;
4348 if (cgrp->parent == NULL)
4351 parent = mem_cgroup_from_cont(cgrp->parent);
4355 /* If under hierarchy, only empty-root can set this value */
4356 if ((parent->use_hierarchy) ||
4357 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4362 memcg->swappiness = val;
4369 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4371 struct mem_cgroup_threshold_ary *t;
4377 t = rcu_dereference(memcg->thresholds.primary);
4379 t = rcu_dereference(memcg->memsw_thresholds.primary);
4384 usage = mem_cgroup_usage(memcg, swap);
4387 * current_threshold points to threshold just below usage.
4388 * If it's not true, a threshold was crossed after last
4389 * call of __mem_cgroup_threshold().
4391 i = t->current_threshold;
4394 * Iterate backward over array of thresholds starting from
4395 * current_threshold and check if a threshold is crossed.
4396 * If none of thresholds below usage is crossed, we read
4397 * only one element of the array here.
4399 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4400 eventfd_signal(t->entries[i].eventfd, 1);
4402 /* i = current_threshold + 1 */
4406 * Iterate forward over array of thresholds starting from
4407 * current_threshold+1 and check if a threshold is crossed.
4408 * If none of thresholds above usage is crossed, we read
4409 * only one element of the array here.
4411 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4412 eventfd_signal(t->entries[i].eventfd, 1);
4414 /* Update current_threshold */
4415 t->current_threshold = i - 1;
4420 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4423 __mem_cgroup_threshold(memcg, false);
4424 if (do_swap_account)
4425 __mem_cgroup_threshold(memcg, true);
4427 memcg = parent_mem_cgroup(memcg);
4431 static int compare_thresholds(const void *a, const void *b)
4433 const struct mem_cgroup_threshold *_a = a;
4434 const struct mem_cgroup_threshold *_b = b;
4436 return _a->threshold - _b->threshold;
4439 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
4441 struct mem_cgroup_eventfd_list *ev;
4443 list_for_each_entry(ev, &mem->oom_notify, list)
4444 eventfd_signal(ev->eventfd, 1);
4448 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
4450 struct mem_cgroup *iter;
4452 for_each_mem_cgroup_tree(iter, mem)
4453 mem_cgroup_oom_notify_cb(iter);
4456 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4457 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4459 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4460 struct mem_cgroup_thresholds *thresholds;
4461 struct mem_cgroup_threshold_ary *new;
4462 int type = MEMFILE_TYPE(cft->private);
4463 u64 threshold, usage;
4466 ret = res_counter_memparse_write_strategy(args, &threshold);
4470 mutex_lock(&memcg->thresholds_lock);
4473 thresholds = &memcg->thresholds;
4474 else if (type == _MEMSWAP)
4475 thresholds = &memcg->memsw_thresholds;
4479 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4481 /* Check if a threshold crossed before adding a new one */
4482 if (thresholds->primary)
4483 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4485 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4487 /* Allocate memory for new array of thresholds */
4488 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4496 /* Copy thresholds (if any) to new array */
4497 if (thresholds->primary) {
4498 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4499 sizeof(struct mem_cgroup_threshold));
4502 /* Add new threshold */
4503 new->entries[size - 1].eventfd = eventfd;
4504 new->entries[size - 1].threshold = threshold;
4506 /* Sort thresholds. Registering of new threshold isn't time-critical */
4507 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4508 compare_thresholds, NULL);
4510 /* Find current threshold */
4511 new->current_threshold = -1;
4512 for (i = 0; i < size; i++) {
4513 if (new->entries[i].threshold < usage) {
4515 * new->current_threshold will not be used until
4516 * rcu_assign_pointer(), so it's safe to increment
4519 ++new->current_threshold;
4523 /* Free old spare buffer and save old primary buffer as spare */
4524 kfree(thresholds->spare);
4525 thresholds->spare = thresholds->primary;
4527 rcu_assign_pointer(thresholds->primary, new);
4529 /* To be sure that nobody uses thresholds */
4533 mutex_unlock(&memcg->thresholds_lock);
4538 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4539 struct cftype *cft, struct eventfd_ctx *eventfd)
4541 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4542 struct mem_cgroup_thresholds *thresholds;
4543 struct mem_cgroup_threshold_ary *new;
4544 int type = MEMFILE_TYPE(cft->private);
4548 mutex_lock(&memcg->thresholds_lock);
4550 thresholds = &memcg->thresholds;
4551 else if (type == _MEMSWAP)
4552 thresholds = &memcg->memsw_thresholds;
4557 * Something went wrong if we trying to unregister a threshold
4558 * if we don't have thresholds
4560 BUG_ON(!thresholds);
4562 if (!thresholds->primary)
4565 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4567 /* Check if a threshold crossed before removing */
4568 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4570 /* Calculate new number of threshold */
4572 for (i = 0; i < thresholds->primary->size; i++) {
4573 if (thresholds->primary->entries[i].eventfd != eventfd)
4577 new = thresholds->spare;
4579 /* Set thresholds array to NULL if we don't have thresholds */
4588 /* Copy thresholds and find current threshold */
4589 new->current_threshold = -1;
4590 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4591 if (thresholds->primary->entries[i].eventfd == eventfd)
4594 new->entries[j] = thresholds->primary->entries[i];
4595 if (new->entries[j].threshold < usage) {
4597 * new->current_threshold will not be used
4598 * until rcu_assign_pointer(), so it's safe to increment
4601 ++new->current_threshold;
4607 /* Swap primary and spare array */
4608 thresholds->spare = thresholds->primary;
4609 /* If all events are unregistered, free the spare array */
4611 kfree(thresholds->spare);
4612 thresholds->spare = NULL;
4615 rcu_assign_pointer(thresholds->primary, new);
4617 /* To be sure that nobody uses thresholds */
4620 mutex_unlock(&memcg->thresholds_lock);
4623 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4624 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4626 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4627 struct mem_cgroup_eventfd_list *event;
4628 int type = MEMFILE_TYPE(cft->private);
4630 BUG_ON(type != _OOM_TYPE);
4631 event = kmalloc(sizeof(*event), GFP_KERNEL);
4635 mutex_lock(&memcg_oom_mutex);
4637 event->eventfd = eventfd;
4638 list_add(&event->list, &memcg->oom_notify);
4640 /* already in OOM ? */
4641 if (atomic_read(&memcg->oom_lock))
4642 eventfd_signal(eventfd, 1);
4643 mutex_unlock(&memcg_oom_mutex);
4648 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4649 struct cftype *cft, struct eventfd_ctx *eventfd)
4651 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4652 struct mem_cgroup_eventfd_list *ev, *tmp;
4653 int type = MEMFILE_TYPE(cft->private);
4655 BUG_ON(type != _OOM_TYPE);
4657 mutex_lock(&memcg_oom_mutex);
4659 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4660 if (ev->eventfd == eventfd) {
4661 list_del(&ev->list);
4666 mutex_unlock(&memcg_oom_mutex);
4669 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4670 struct cftype *cft, struct cgroup_map_cb *cb)
4672 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4674 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4676 if (atomic_read(&mem->oom_lock))
4677 cb->fill(cb, "under_oom", 1);
4679 cb->fill(cb, "under_oom", 0);
4683 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4684 struct cftype *cft, u64 val)
4686 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4687 struct mem_cgroup *parent;
4689 /* cannot set to root cgroup and only 0 and 1 are allowed */
4690 if (!cgrp->parent || !((val == 0) || (val == 1)))
4693 parent = mem_cgroup_from_cont(cgrp->parent);
4696 /* oom-kill-disable is a flag for subhierarchy. */
4697 if ((parent->use_hierarchy) ||
4698 (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4702 mem->oom_kill_disable = val;
4704 memcg_oom_recover(mem);
4710 static const struct file_operations mem_control_numa_stat_file_operations = {
4712 .llseek = seq_lseek,
4713 .release = single_release,
4716 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4718 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4720 file->f_op = &mem_control_numa_stat_file_operations;
4721 return single_open(file, mem_control_numa_stat_show, cont);
4723 #endif /* CONFIG_NUMA */
4725 static struct cftype mem_cgroup_files[] = {
4727 .name = "usage_in_bytes",
4728 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4729 .read_u64 = mem_cgroup_read,
4730 .register_event = mem_cgroup_usage_register_event,
4731 .unregister_event = mem_cgroup_usage_unregister_event,
4734 .name = "max_usage_in_bytes",
4735 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4736 .trigger = mem_cgroup_reset,
4737 .read_u64 = mem_cgroup_read,
4740 .name = "limit_in_bytes",
4741 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4742 .write_string = mem_cgroup_write,
4743 .read_u64 = mem_cgroup_read,
4746 .name = "soft_limit_in_bytes",
4747 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4748 .write_string = mem_cgroup_write,
4749 .read_u64 = mem_cgroup_read,
4753 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4754 .trigger = mem_cgroup_reset,
4755 .read_u64 = mem_cgroup_read,
4759 .read_map = mem_control_stat_show,
4762 .name = "force_empty",
4763 .trigger = mem_cgroup_force_empty_write,
4766 .name = "use_hierarchy",
4767 .write_u64 = mem_cgroup_hierarchy_write,
4768 .read_u64 = mem_cgroup_hierarchy_read,
4771 .name = "swappiness",
4772 .read_u64 = mem_cgroup_swappiness_read,
4773 .write_u64 = mem_cgroup_swappiness_write,
4776 .name = "move_charge_at_immigrate",
4777 .read_u64 = mem_cgroup_move_charge_read,
4778 .write_u64 = mem_cgroup_move_charge_write,
4781 .name = "oom_control",
4782 .read_map = mem_cgroup_oom_control_read,
4783 .write_u64 = mem_cgroup_oom_control_write,
4784 .register_event = mem_cgroup_oom_register_event,
4785 .unregister_event = mem_cgroup_oom_unregister_event,
4786 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4790 .name = "numa_stat",
4791 .open = mem_control_numa_stat_open,
4797 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4798 static struct cftype memsw_cgroup_files[] = {
4800 .name = "memsw.usage_in_bytes",
4801 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4802 .read_u64 = mem_cgroup_read,
4803 .register_event = mem_cgroup_usage_register_event,
4804 .unregister_event = mem_cgroup_usage_unregister_event,
4807 .name = "memsw.max_usage_in_bytes",
4808 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4809 .trigger = mem_cgroup_reset,
4810 .read_u64 = mem_cgroup_read,
4813 .name = "memsw.limit_in_bytes",
4814 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4815 .write_string = mem_cgroup_write,
4816 .read_u64 = mem_cgroup_read,
4819 .name = "memsw.failcnt",
4820 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4821 .trigger = mem_cgroup_reset,
4822 .read_u64 = mem_cgroup_read,
4826 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4828 if (!do_swap_account)
4830 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4831 ARRAY_SIZE(memsw_cgroup_files));
4834 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4840 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4842 struct mem_cgroup_per_node *pn;
4843 struct mem_cgroup_per_zone *mz;
4845 int zone, tmp = node;
4847 * This routine is called against possible nodes.
4848 * But it's BUG to call kmalloc() against offline node.
4850 * TODO: this routine can waste much memory for nodes which will
4851 * never be onlined. It's better to use memory hotplug callback
4854 if (!node_state(node, N_NORMAL_MEMORY))
4856 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4860 mem->info.nodeinfo[node] = pn;
4861 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4862 mz = &pn->zoneinfo[zone];
4864 INIT_LIST_HEAD(&mz->lists[l]);
4865 mz->usage_in_excess = 0;
4866 mz->on_tree = false;
4872 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4874 kfree(mem->info.nodeinfo[node]);
4877 static struct mem_cgroup *mem_cgroup_alloc(void)
4879 struct mem_cgroup *mem;
4880 int size = sizeof(struct mem_cgroup);
4882 /* Can be very big if MAX_NUMNODES is very big */
4883 if (size < PAGE_SIZE)
4884 mem = kzalloc(size, GFP_KERNEL);
4886 mem = vzalloc(size);
4891 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4894 spin_lock_init(&mem->pcp_counter_lock);
4898 if (size < PAGE_SIZE)
4906 * At destroying mem_cgroup, references from swap_cgroup can remain.
4907 * (scanning all at force_empty is too costly...)
4909 * Instead of clearing all references at force_empty, we remember
4910 * the number of reference from swap_cgroup and free mem_cgroup when
4911 * it goes down to 0.
4913 * Removal of cgroup itself succeeds regardless of refs from swap.
4916 static void __mem_cgroup_free(struct mem_cgroup *mem)
4920 mem_cgroup_remove_from_trees(mem);
4921 free_css_id(&mem_cgroup_subsys, &mem->css);
4923 for_each_node_state(node, N_POSSIBLE)
4924 free_mem_cgroup_per_zone_info(mem, node);
4926 free_percpu(mem->stat);
4927 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4933 static void mem_cgroup_get(struct mem_cgroup *mem)
4935 atomic_inc(&mem->refcnt);
4938 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4940 if (atomic_sub_and_test(count, &mem->refcnt)) {
4941 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4942 __mem_cgroup_free(mem);
4944 mem_cgroup_put(parent);
4948 static void mem_cgroup_put(struct mem_cgroup *mem)
4950 __mem_cgroup_put(mem, 1);
4954 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4956 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4958 if (!mem->res.parent)
4960 return mem_cgroup_from_res_counter(mem->res.parent, res);
4963 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4964 static void __init enable_swap_cgroup(void)
4966 if (!mem_cgroup_disabled() && really_do_swap_account)
4967 do_swap_account = 1;
4970 static void __init enable_swap_cgroup(void)
4975 static int mem_cgroup_soft_limit_tree_init(void)
4977 struct mem_cgroup_tree_per_node *rtpn;
4978 struct mem_cgroup_tree_per_zone *rtpz;
4979 int tmp, node, zone;
4981 for_each_node_state(node, N_POSSIBLE) {
4983 if (!node_state(node, N_NORMAL_MEMORY))
4985 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4989 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4991 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4992 rtpz = &rtpn->rb_tree_per_zone[zone];
4993 rtpz->rb_root = RB_ROOT;
4994 spin_lock_init(&rtpz->lock);
5000 static struct cgroup_subsys_state * __ref
5001 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
5003 struct mem_cgroup *mem, *parent;
5004 long error = -ENOMEM;
5007 mem = mem_cgroup_alloc();
5009 return ERR_PTR(error);
5011 for_each_node_state(node, N_POSSIBLE)
5012 if (alloc_mem_cgroup_per_zone_info(mem, node))
5016 if (cont->parent == NULL) {
5018 enable_swap_cgroup();
5020 if (mem_cgroup_soft_limit_tree_init())
5022 root_mem_cgroup = mem;
5023 for_each_possible_cpu(cpu) {
5024 struct memcg_stock_pcp *stock =
5025 &per_cpu(memcg_stock, cpu);
5026 INIT_WORK(&stock->work, drain_local_stock);
5028 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
5030 parent = mem_cgroup_from_cont(cont->parent);
5031 mem->use_hierarchy = parent->use_hierarchy;
5032 mem->oom_kill_disable = parent->oom_kill_disable;
5035 if (parent && parent->use_hierarchy) {
5036 res_counter_init(&mem->res, &parent->res);
5037 res_counter_init(&mem->memsw, &parent->memsw);
5039 * We increment refcnt of the parent to ensure that we can
5040 * safely access it on res_counter_charge/uncharge.
5041 * This refcnt will be decremented when freeing this
5042 * mem_cgroup(see mem_cgroup_put).
5044 mem_cgroup_get(parent);
5046 res_counter_init(&mem->res, NULL);
5047 res_counter_init(&mem->memsw, NULL);
5049 mem->last_scanned_child = 0;
5050 mem->last_scanned_node = MAX_NUMNODES;
5051 INIT_LIST_HEAD(&mem->oom_notify);
5054 mem->swappiness = get_swappiness(parent);
5055 atomic_set(&mem->refcnt, 1);
5056 mem->move_charge_at_immigrate = 0;
5057 mutex_init(&mem->thresholds_lock);
5060 __mem_cgroup_free(mem);
5061 return ERR_PTR(error);
5064 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5065 struct cgroup *cont)
5067 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5069 return mem_cgroup_force_empty(mem, false);
5072 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5073 struct cgroup *cont)
5075 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
5077 mem_cgroup_put(mem);
5080 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5081 struct cgroup *cont)
5085 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5086 ARRAY_SIZE(mem_cgroup_files));
5089 ret = register_memsw_files(cont, ss);
5094 /* Handlers for move charge at task migration. */
5095 #define PRECHARGE_COUNT_AT_ONCE 256
5096 static int mem_cgroup_do_precharge(unsigned long count)
5099 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5100 struct mem_cgroup *mem = mc.to;
5102 if (mem_cgroup_is_root(mem)) {
5103 mc.precharge += count;
5104 /* we don't need css_get for root */
5107 /* try to charge at once */
5109 struct res_counter *dummy;
5111 * "mem" cannot be under rmdir() because we've already checked
5112 * by cgroup_lock_live_cgroup() that it is not removed and we
5113 * are still under the same cgroup_mutex. So we can postpone
5116 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
5118 if (do_swap_account && res_counter_charge(&mem->memsw,
5119 PAGE_SIZE * count, &dummy)) {
5120 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
5123 mc.precharge += count;
5127 /* fall back to one by one charge */
5129 if (signal_pending(current)) {
5133 if (!batch_count--) {
5134 batch_count = PRECHARGE_COUNT_AT_ONCE;
5137 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, 1, &mem, false);
5139 /* mem_cgroup_clear_mc() will do uncharge later */
5147 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5148 * @vma: the vma the pte to be checked belongs
5149 * @addr: the address corresponding to the pte to be checked
5150 * @ptent: the pte to be checked
5151 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5154 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5155 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5156 * move charge. if @target is not NULL, the page is stored in target->page
5157 * with extra refcnt got(Callers should handle it).
5158 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5159 * target for charge migration. if @target is not NULL, the entry is stored
5162 * Called with pte lock held.
5169 enum mc_target_type {
5170 MC_TARGET_NONE, /* not used */
5175 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5176 unsigned long addr, pte_t ptent)
5178 struct page *page = vm_normal_page(vma, addr, ptent);
5180 if (!page || !page_mapped(page))
5182 if (PageAnon(page)) {
5183 /* we don't move shared anon */
5184 if (!move_anon() || page_mapcount(page) > 2)
5186 } else if (!move_file())
5187 /* we ignore mapcount for file pages */
5189 if (!get_page_unless_zero(page))
5195 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5196 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5199 struct page *page = NULL;
5200 swp_entry_t ent = pte_to_swp_entry(ptent);
5202 if (!move_anon() || non_swap_entry(ent))
5204 usage_count = mem_cgroup_count_swap_user(ent, &page);
5205 if (usage_count > 1) { /* we don't move shared anon */
5210 if (do_swap_account)
5211 entry->val = ent.val;
5216 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5217 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5219 struct page *page = NULL;
5220 struct inode *inode;
5221 struct address_space *mapping;
5224 if (!vma->vm_file) /* anonymous vma */
5229 inode = vma->vm_file->f_path.dentry->d_inode;
5230 mapping = vma->vm_file->f_mapping;
5231 if (pte_none(ptent))
5232 pgoff = linear_page_index(vma, addr);
5233 else /* pte_file(ptent) is true */
5234 pgoff = pte_to_pgoff(ptent);
5236 /* page is moved even if it's not RSS of this task(page-faulted). */
5237 if (!mapping_cap_swap_backed(mapping)) { /* normal file */
5238 page = find_get_page(mapping, pgoff);
5239 } else { /* shmem/tmpfs file. we should take account of swap too. */
5241 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
5242 if (do_swap_account)
5243 entry->val = ent.val;
5249 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5250 unsigned long addr, pte_t ptent, union mc_target *target)
5252 struct page *page = NULL;
5253 struct page_cgroup *pc;
5255 swp_entry_t ent = { .val = 0 };
5257 if (pte_present(ptent))
5258 page = mc_handle_present_pte(vma, addr, ptent);
5259 else if (is_swap_pte(ptent))
5260 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5261 else if (pte_none(ptent) || pte_file(ptent))
5262 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5264 if (!page && !ent.val)
5267 pc = lookup_page_cgroup(page);
5269 * Do only loose check w/o page_cgroup lock.
5270 * mem_cgroup_move_account() checks the pc is valid or not under
5273 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5274 ret = MC_TARGET_PAGE;
5276 target->page = page;
5278 if (!ret || !target)
5281 /* There is a swap entry and a page doesn't exist or isn't charged */
5282 if (ent.val && !ret &&
5283 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5284 ret = MC_TARGET_SWAP;
5291 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5292 unsigned long addr, unsigned long end,
5293 struct mm_walk *walk)
5295 struct vm_area_struct *vma = walk->private;
5299 split_huge_page_pmd(walk->mm, pmd);
5300 if (pmd_trans_unstable(pmd))
5303 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5304 for (; addr != end; pte++, addr += PAGE_SIZE)
5305 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5306 mc.precharge++; /* increment precharge temporarily */
5307 pte_unmap_unlock(pte - 1, ptl);
5313 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5315 unsigned long precharge;
5316 struct vm_area_struct *vma;
5318 down_read(&mm->mmap_sem);
5319 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5320 struct mm_walk mem_cgroup_count_precharge_walk = {
5321 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5325 if (is_vm_hugetlb_page(vma))
5327 walk_page_range(vma->vm_start, vma->vm_end,
5328 &mem_cgroup_count_precharge_walk);
5330 up_read(&mm->mmap_sem);
5332 precharge = mc.precharge;
5338 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5340 unsigned long precharge = mem_cgroup_count_precharge(mm);
5342 VM_BUG_ON(mc.moving_task);
5343 mc.moving_task = current;
5344 return mem_cgroup_do_precharge(precharge);
5347 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5348 static void __mem_cgroup_clear_mc(void)
5350 struct mem_cgroup *from = mc.from;
5351 struct mem_cgroup *to = mc.to;
5353 /* we must uncharge all the leftover precharges from mc.to */
5355 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5359 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5360 * we must uncharge here.
5362 if (mc.moved_charge) {
5363 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5364 mc.moved_charge = 0;
5366 /* we must fixup refcnts and charges */
5367 if (mc.moved_swap) {
5368 /* uncharge swap account from the old cgroup */
5369 if (!mem_cgroup_is_root(mc.from))
5370 res_counter_uncharge(&mc.from->memsw,
5371 PAGE_SIZE * mc.moved_swap);
5372 __mem_cgroup_put(mc.from, mc.moved_swap);
5374 if (!mem_cgroup_is_root(mc.to)) {
5376 * we charged both to->res and to->memsw, so we should
5379 res_counter_uncharge(&mc.to->res,
5380 PAGE_SIZE * mc.moved_swap);
5382 /* we've already done mem_cgroup_get(mc.to) */
5385 memcg_oom_recover(from);
5386 memcg_oom_recover(to);
5387 wake_up_all(&mc.waitq);
5390 static void mem_cgroup_clear_mc(void)
5392 struct mem_cgroup *from = mc.from;
5395 * we must clear moving_task before waking up waiters at the end of
5398 mc.moving_task = NULL;
5399 __mem_cgroup_clear_mc();
5400 spin_lock(&mc.lock);
5403 spin_unlock(&mc.lock);
5404 mem_cgroup_end_move(from);
5407 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5408 struct cgroup *cgroup,
5409 struct task_struct *p)
5412 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
5414 if (mem->move_charge_at_immigrate) {
5415 struct mm_struct *mm;
5416 struct mem_cgroup *from = mem_cgroup_from_task(p);
5418 VM_BUG_ON(from == mem);
5420 mm = get_task_mm(p);
5423 /* We move charges only when we move a owner of the mm */
5424 if (mm->owner == p) {
5427 VM_BUG_ON(mc.precharge);
5428 VM_BUG_ON(mc.moved_charge);
5429 VM_BUG_ON(mc.moved_swap);
5430 mem_cgroup_start_move(from);
5431 spin_lock(&mc.lock);
5434 spin_unlock(&mc.lock);
5435 /* We set mc.moving_task later */
5437 ret = mem_cgroup_precharge_mc(mm);
5439 mem_cgroup_clear_mc();
5446 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5447 struct cgroup *cgroup,
5448 struct task_struct *p)
5450 mem_cgroup_clear_mc();
5453 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5454 unsigned long addr, unsigned long end,
5455 struct mm_walk *walk)
5458 struct vm_area_struct *vma = walk->private;
5462 split_huge_page_pmd(walk->mm, pmd);
5463 if (pmd_trans_unstable(pmd))
5466 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5467 for (; addr != end; addr += PAGE_SIZE) {
5468 pte_t ptent = *(pte++);
5469 union mc_target target;
5472 struct page_cgroup *pc;
5478 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5480 case MC_TARGET_PAGE:
5482 if (isolate_lru_page(page))
5484 pc = lookup_page_cgroup(page);
5485 if (!mem_cgroup_move_account(page, 1, pc,
5486 mc.from, mc.to, false)) {
5488 /* we uncharge from mc.from later. */
5491 putback_lru_page(page);
5492 put: /* is_target_pte_for_mc() gets the page */
5495 case MC_TARGET_SWAP:
5497 if (!mem_cgroup_move_swap_account(ent,
5498 mc.from, mc.to, false)) {
5500 /* we fixup refcnts and charges later. */
5508 pte_unmap_unlock(pte - 1, ptl);
5513 * We have consumed all precharges we got in can_attach().
5514 * We try charge one by one, but don't do any additional
5515 * charges to mc.to if we have failed in charge once in attach()
5518 ret = mem_cgroup_do_precharge(1);
5526 static void mem_cgroup_move_charge(struct mm_struct *mm)
5528 struct vm_area_struct *vma;
5530 lru_add_drain_all();
5532 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5534 * Someone who are holding the mmap_sem might be waiting in
5535 * waitq. So we cancel all extra charges, wake up all waiters,
5536 * and retry. Because we cancel precharges, we might not be able
5537 * to move enough charges, but moving charge is a best-effort
5538 * feature anyway, so it wouldn't be a big problem.
5540 __mem_cgroup_clear_mc();
5544 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5546 struct mm_walk mem_cgroup_move_charge_walk = {
5547 .pmd_entry = mem_cgroup_move_charge_pte_range,
5551 if (is_vm_hugetlb_page(vma))
5553 ret = walk_page_range(vma->vm_start, vma->vm_end,
5554 &mem_cgroup_move_charge_walk);
5557 * means we have consumed all precharges and failed in
5558 * doing additional charge. Just abandon here.
5562 up_read(&mm->mmap_sem);
5565 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5566 struct cgroup *cont,
5567 struct cgroup *old_cont,
5568 struct task_struct *p)
5570 struct mm_struct *mm = get_task_mm(p);
5574 mem_cgroup_move_charge(mm);
5579 mem_cgroup_clear_mc();
5581 #else /* !CONFIG_MMU */
5582 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5583 struct cgroup *cgroup,
5584 struct task_struct *p)
5588 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5589 struct cgroup *cgroup,
5590 struct task_struct *p)
5593 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5594 struct cgroup *cont,
5595 struct cgroup *old_cont,
5596 struct task_struct *p)
5601 struct cgroup_subsys mem_cgroup_subsys = {
5603 .subsys_id = mem_cgroup_subsys_id,
5604 .create = mem_cgroup_create,
5605 .pre_destroy = mem_cgroup_pre_destroy,
5606 .destroy = mem_cgroup_destroy,
5607 .populate = mem_cgroup_populate,
5608 .can_attach = mem_cgroup_can_attach,
5609 .cancel_attach = mem_cgroup_cancel_attach,
5610 .attach = mem_cgroup_move_task,
5615 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5616 static int __init enable_swap_account(char *s)
5618 /* consider enabled if no parameter or 1 is given */
5619 if (!strcmp(s, "1"))
5620 really_do_swap_account = 1;
5621 else if (!strcmp(s, "0"))
5622 really_do_swap_account = 0;
5625 __setup("swapaccount=", enable_swap_account);