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/export.h>
37 #include <linux/mutex.h>
38 #include <linux/rbtree.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];
124 struct mem_cgroup_reclaim_iter {
125 /* css_id of the last scanned hierarchy member */
127 /* scan generation, increased every round-trip */
128 unsigned int generation;
132 * per-zone information in memory controller.
134 struct mem_cgroup_per_zone {
136 * spin_lock to protect the per cgroup LRU
138 struct list_head lists[NR_LRU_LISTS];
139 unsigned long count[NR_LRU_LISTS];
141 struct mem_cgroup_reclaim_iter reclaim_iter[DEF_PRIORITY + 1];
143 struct zone_reclaim_stat reclaim_stat;
144 struct rb_node tree_node; /* RB tree node */
145 unsigned long long usage_in_excess;/* Set to the value by which */
146 /* the soft limit is exceeded*/
148 struct mem_cgroup *mem; /* Back pointer, we cannot */
149 /* use container_of */
151 /* Macro for accessing counter */
152 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
154 struct mem_cgroup_per_node {
155 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
158 struct mem_cgroup_lru_info {
159 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
163 * Cgroups above their limits are maintained in a RB-Tree, independent of
164 * their hierarchy representation
167 struct mem_cgroup_tree_per_zone {
168 struct rb_root rb_root;
172 struct mem_cgroup_tree_per_node {
173 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
176 struct mem_cgroup_tree {
177 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
180 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
182 struct mem_cgroup_threshold {
183 struct eventfd_ctx *eventfd;
188 struct mem_cgroup_threshold_ary {
189 /* An array index points to threshold just below usage. */
190 int current_threshold;
191 /* Size of entries[] */
193 /* Array of thresholds */
194 struct mem_cgroup_threshold entries[0];
197 struct mem_cgroup_thresholds {
198 /* Primary thresholds array */
199 struct mem_cgroup_threshold_ary *primary;
201 * Spare threshold array.
202 * This is needed to make mem_cgroup_unregister_event() "never fail".
203 * It must be able to store at least primary->size - 1 entries.
205 struct mem_cgroup_threshold_ary *spare;
209 struct mem_cgroup_eventfd_list {
210 struct list_head list;
211 struct eventfd_ctx *eventfd;
214 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
215 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
218 * The memory controller data structure. The memory controller controls both
219 * page cache and RSS per cgroup. We would eventually like to provide
220 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
221 * to help the administrator determine what knobs to tune.
223 * TODO: Add a water mark for the memory controller. Reclaim will begin when
224 * we hit the water mark. May be even add a low water mark, such that
225 * no reclaim occurs from a cgroup at it's low water mark, this is
226 * a feature that will be implemented much later in the future.
229 struct cgroup_subsys_state css;
231 * the counter to account for memory usage
233 struct res_counter res;
235 * the counter to account for mem+swap usage.
237 struct res_counter memsw;
239 * Per cgroup active and inactive list, similar to the
240 * per zone LRU lists.
242 struct mem_cgroup_lru_info info;
243 int last_scanned_node;
245 nodemask_t scan_nodes;
246 atomic_t numainfo_events;
247 atomic_t numainfo_updating;
250 * Should the accounting and control be hierarchical, per subtree?
260 /* OOM-Killer disable */
261 int oom_kill_disable;
263 /* set when res.limit == memsw.limit */
264 bool memsw_is_minimum;
266 /* protect arrays of thresholds */
267 struct mutex thresholds_lock;
269 /* thresholds for memory usage. RCU-protected */
270 struct mem_cgroup_thresholds thresholds;
272 /* thresholds for mem+swap usage. RCU-protected */
273 struct mem_cgroup_thresholds memsw_thresholds;
275 /* For oom notifier event fd */
276 struct list_head oom_notify;
279 * Should we move charges of a task when a task is moved into this
280 * mem_cgroup ? And what type of charges should we move ?
282 unsigned long move_charge_at_immigrate;
286 struct mem_cgroup_stat_cpu *stat;
288 * used when a cpu is offlined or other synchronizations
289 * See mem_cgroup_read_stat().
291 struct mem_cgroup_stat_cpu nocpu_base;
292 spinlock_t pcp_counter_lock;
295 /* Stuffs for move charges at task migration. */
297 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
298 * left-shifted bitmap of these types.
301 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
302 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
306 /* "mc" and its members are protected by cgroup_mutex */
307 static struct move_charge_struct {
308 spinlock_t lock; /* for from, to */
309 struct mem_cgroup *from;
310 struct mem_cgroup *to;
311 unsigned long precharge;
312 unsigned long moved_charge;
313 unsigned long moved_swap;
314 struct task_struct *moving_task; /* a task moving charges */
315 wait_queue_head_t waitq; /* a waitq for other context */
317 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
318 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
321 static bool move_anon(void)
323 return test_bit(MOVE_CHARGE_TYPE_ANON,
324 &mc.to->move_charge_at_immigrate);
327 static bool move_file(void)
329 return test_bit(MOVE_CHARGE_TYPE_FILE,
330 &mc.to->move_charge_at_immigrate);
334 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
335 * limit reclaim to prevent infinite loops, if they ever occur.
337 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
338 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
341 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
342 MEM_CGROUP_CHARGE_TYPE_MAPPED,
343 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
344 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
345 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
346 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
350 /* for encoding cft->private value on file */
353 #define _OOM_TYPE (2)
354 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
355 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
356 #define MEMFILE_ATTR(val) ((val) & 0xffff)
357 /* Used for OOM nofiier */
358 #define OOM_CONTROL (0)
361 * Reclaim flags for mem_cgroup_hierarchical_reclaim
363 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
364 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
365 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
366 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
368 static void mem_cgroup_get(struct mem_cgroup *memcg);
369 static void mem_cgroup_put(struct mem_cgroup *memcg);
370 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg);
371 static void drain_all_stock_async(struct mem_cgroup *memcg);
373 static struct mem_cgroup_per_zone *
374 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
376 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
379 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
384 static struct mem_cgroup_per_zone *
385 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
387 int nid = page_to_nid(page);
388 int zid = page_zonenum(page);
390 return mem_cgroup_zoneinfo(memcg, nid, zid);
393 static struct mem_cgroup_tree_per_zone *
394 soft_limit_tree_node_zone(int nid, int zid)
396 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
399 static struct mem_cgroup_tree_per_zone *
400 soft_limit_tree_from_page(struct page *page)
402 int nid = page_to_nid(page);
403 int zid = page_zonenum(page);
405 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
409 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
410 struct mem_cgroup_per_zone *mz,
411 struct mem_cgroup_tree_per_zone *mctz,
412 unsigned long long new_usage_in_excess)
414 struct rb_node **p = &mctz->rb_root.rb_node;
415 struct rb_node *parent = NULL;
416 struct mem_cgroup_per_zone *mz_node;
421 mz->usage_in_excess = new_usage_in_excess;
422 if (!mz->usage_in_excess)
426 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
428 if (mz->usage_in_excess < mz_node->usage_in_excess)
431 * We can't avoid mem cgroups that are over their soft
432 * limit by the same amount
434 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
437 rb_link_node(&mz->tree_node, parent, p);
438 rb_insert_color(&mz->tree_node, &mctz->rb_root);
443 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
444 struct mem_cgroup_per_zone *mz,
445 struct mem_cgroup_tree_per_zone *mctz)
449 rb_erase(&mz->tree_node, &mctz->rb_root);
454 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
455 struct mem_cgroup_per_zone *mz,
456 struct mem_cgroup_tree_per_zone *mctz)
458 spin_lock(&mctz->lock);
459 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
460 spin_unlock(&mctz->lock);
464 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
466 unsigned long long excess;
467 struct mem_cgroup_per_zone *mz;
468 struct mem_cgroup_tree_per_zone *mctz;
469 int nid = page_to_nid(page);
470 int zid = page_zonenum(page);
471 mctz = soft_limit_tree_from_page(page);
474 * Necessary to update all ancestors when hierarchy is used.
475 * because their event counter is not touched.
477 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
478 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
479 excess = res_counter_soft_limit_excess(&memcg->res);
481 * We have to update the tree if mz is on RB-tree or
482 * mem is over its softlimit.
484 if (excess || mz->on_tree) {
485 spin_lock(&mctz->lock);
486 /* if on-tree, remove it */
488 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
490 * Insert again. mz->usage_in_excess will be updated.
491 * If excess is 0, no tree ops.
493 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
494 spin_unlock(&mctz->lock);
499 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
502 struct mem_cgroup_per_zone *mz;
503 struct mem_cgroup_tree_per_zone *mctz;
505 for_each_node_state(node, N_POSSIBLE) {
506 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
507 mz = mem_cgroup_zoneinfo(memcg, node, zone);
508 mctz = soft_limit_tree_node_zone(node, zone);
509 mem_cgroup_remove_exceeded(memcg, mz, mctz);
514 static struct mem_cgroup_per_zone *
515 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
517 struct rb_node *rightmost = NULL;
518 struct mem_cgroup_per_zone *mz;
522 rightmost = rb_last(&mctz->rb_root);
524 goto done; /* Nothing to reclaim from */
526 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
528 * Remove the node now but someone else can add it back,
529 * we will to add it back at the end of reclaim to its correct
530 * position in the tree.
532 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
533 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
534 !css_tryget(&mz->mem->css))
540 static struct mem_cgroup_per_zone *
541 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
543 struct mem_cgroup_per_zone *mz;
545 spin_lock(&mctz->lock);
546 mz = __mem_cgroup_largest_soft_limit_node(mctz);
547 spin_unlock(&mctz->lock);
552 * Implementation Note: reading percpu statistics for memcg.
554 * Both of vmstat[] and percpu_counter has threshold and do periodic
555 * synchronization to implement "quick" read. There are trade-off between
556 * reading cost and precision of value. Then, we may have a chance to implement
557 * a periodic synchronizion of counter in memcg's counter.
559 * But this _read() function is used for user interface now. The user accounts
560 * memory usage by memory cgroup and he _always_ requires exact value because
561 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
562 * have to visit all online cpus and make sum. So, for now, unnecessary
563 * synchronization is not implemented. (just implemented for cpu hotplug)
565 * If there are kernel internal actions which can make use of some not-exact
566 * value, and reading all cpu value can be performance bottleneck in some
567 * common workload, threashold and synchonization as vmstat[] should be
570 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
571 enum mem_cgroup_stat_index idx)
577 for_each_online_cpu(cpu)
578 val += per_cpu(memcg->stat->count[idx], cpu);
579 #ifdef CONFIG_HOTPLUG_CPU
580 spin_lock(&memcg->pcp_counter_lock);
581 val += memcg->nocpu_base.count[idx];
582 spin_unlock(&memcg->pcp_counter_lock);
588 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
591 int val = (charge) ? 1 : -1;
592 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
595 void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val)
597 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
600 void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val)
602 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
605 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
606 enum mem_cgroup_events_index idx)
608 unsigned long val = 0;
611 for_each_online_cpu(cpu)
612 val += per_cpu(memcg->stat->events[idx], cpu);
613 #ifdef CONFIG_HOTPLUG_CPU
614 spin_lock(&memcg->pcp_counter_lock);
615 val += memcg->nocpu_base.events[idx];
616 spin_unlock(&memcg->pcp_counter_lock);
621 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
622 bool file, int nr_pages)
627 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
630 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
633 /* pagein of a big page is an event. So, ignore page size */
635 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
637 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
638 nr_pages = -nr_pages; /* for event */
641 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
647 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
648 unsigned int lru_mask)
650 struct mem_cgroup_per_zone *mz;
652 unsigned long ret = 0;
654 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
657 if (BIT(l) & lru_mask)
658 ret += MEM_CGROUP_ZSTAT(mz, l);
664 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
665 int nid, unsigned int lru_mask)
670 for (zid = 0; zid < MAX_NR_ZONES; zid++)
671 total += mem_cgroup_zone_nr_lru_pages(memcg,
677 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
678 unsigned int lru_mask)
683 for_each_node_state(nid, N_HIGH_MEMORY)
684 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
688 static bool __memcg_event_check(struct mem_cgroup *memcg, int target)
690 unsigned long val, next;
692 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
693 next = __this_cpu_read(memcg->stat->targets[target]);
694 /* from time_after() in jiffies.h */
695 return ((long)next - (long)val < 0);
698 static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
700 unsigned long val, next;
702 val = __this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
705 case MEM_CGROUP_TARGET_THRESH:
706 next = val + THRESHOLDS_EVENTS_TARGET;
708 case MEM_CGROUP_TARGET_SOFTLIMIT:
709 next = val + SOFTLIMIT_EVENTS_TARGET;
711 case MEM_CGROUP_TARGET_NUMAINFO:
712 next = val + NUMAINFO_EVENTS_TARGET;
718 __this_cpu_write(memcg->stat->targets[target], next);
722 * Check events in order.
725 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
728 /* threshold event is triggered in finer grain than soft limit */
729 if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
730 mem_cgroup_threshold(memcg);
731 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
732 if (unlikely(__memcg_event_check(memcg,
733 MEM_CGROUP_TARGET_SOFTLIMIT))) {
734 mem_cgroup_update_tree(memcg, page);
735 __mem_cgroup_target_update(memcg,
736 MEM_CGROUP_TARGET_SOFTLIMIT);
739 if (unlikely(__memcg_event_check(memcg,
740 MEM_CGROUP_TARGET_NUMAINFO))) {
741 atomic_inc(&memcg->numainfo_events);
742 __mem_cgroup_target_update(memcg,
743 MEM_CGROUP_TARGET_NUMAINFO);
750 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
752 return container_of(cgroup_subsys_state(cont,
753 mem_cgroup_subsys_id), struct mem_cgroup,
757 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
760 * mm_update_next_owner() may clear mm->owner to NULL
761 * if it races with swapoff, page migration, etc.
762 * So this can be called with p == NULL.
767 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
768 struct mem_cgroup, css);
771 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
773 struct mem_cgroup *memcg = NULL;
778 * Because we have no locks, mm->owner's may be being moved to other
779 * cgroup. We use css_tryget() here even if this looks
780 * pessimistic (rather than adding locks here).
784 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
785 if (unlikely(!memcg))
787 } while (!css_tryget(&memcg->css));
793 * mem_cgroup_iter - iterate over memory cgroup hierarchy
794 * @root: hierarchy root
795 * @prev: previously returned memcg, NULL on first invocation
796 * @reclaim: cookie for shared reclaim walks, NULL for full walks
798 * Returns references to children of the hierarchy below @root, or
799 * @root itself, or %NULL after a full round-trip.
801 * Caller must pass the return value in @prev on subsequent
802 * invocations for reference counting, or use mem_cgroup_iter_break()
803 * to cancel a hierarchy walk before the round-trip is complete.
805 * Reclaimers can specify a zone and a priority level in @reclaim to
806 * divide up the memcgs in the hierarchy among all concurrent
807 * reclaimers operating on the same zone and priority.
809 struct mem_cgroup *mem_cgroup_iter(struct mem_cgroup *root,
810 struct mem_cgroup *prev,
811 struct mem_cgroup_reclaim_cookie *reclaim)
813 struct mem_cgroup *memcg = NULL;
816 if (mem_cgroup_disabled())
820 root = root_mem_cgroup;
822 if (prev && !reclaim)
823 id = css_id(&prev->css);
825 if (prev && prev != root)
828 if (!root->use_hierarchy && root != root_mem_cgroup) {
835 struct mem_cgroup_reclaim_iter *uninitialized_var(iter);
836 struct cgroup_subsys_state *css;
839 int nid = zone_to_nid(reclaim->zone);
840 int zid = zone_idx(reclaim->zone);
841 struct mem_cgroup_per_zone *mz;
843 mz = mem_cgroup_zoneinfo(root, nid, zid);
844 iter = &mz->reclaim_iter[reclaim->priority];
845 if (prev && reclaim->generation != iter->generation)
851 css = css_get_next(&mem_cgroup_subsys, id + 1, &root->css, &id);
853 if (css == &root->css || css_tryget(css))
854 memcg = container_of(css,
855 struct mem_cgroup, css);
864 else if (!prev && memcg)
865 reclaim->generation = iter->generation;
875 * mem_cgroup_iter_break - abort a hierarchy walk prematurely
876 * @root: hierarchy root
877 * @prev: last visited hierarchy member as returned by mem_cgroup_iter()
879 void mem_cgroup_iter_break(struct mem_cgroup *root,
880 struct mem_cgroup *prev)
883 root = root_mem_cgroup;
884 if (prev && prev != root)
889 * Iteration constructs for visiting all cgroups (under a tree). If
890 * loops are exited prematurely (break), mem_cgroup_iter_break() must
891 * be used for reference counting.
893 #define for_each_mem_cgroup_tree(iter, root) \
894 for (iter = mem_cgroup_iter(root, NULL, NULL); \
896 iter = mem_cgroup_iter(root, iter, NULL))
898 #define for_each_mem_cgroup(iter) \
899 for (iter = mem_cgroup_iter(NULL, NULL, NULL); \
901 iter = mem_cgroup_iter(NULL, iter, NULL))
903 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
905 return (memcg == root_mem_cgroup);
908 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
910 struct mem_cgroup *memcg;
916 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
917 if (unlikely(!memcg))
922 mem_cgroup_pgmajfault(memcg, 1);
925 mem_cgroup_pgfault(memcg, 1);
933 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
936 * Following LRU functions are allowed to be used without PCG_LOCK.
937 * Operations are called by routine of global LRU independently from memcg.
938 * What we have to take care of here is validness of pc->mem_cgroup.
940 * Changes to pc->mem_cgroup happens when
943 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
944 * It is added to LRU before charge.
945 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
946 * When moving account, the page is not on LRU. It's isolated.
949 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
951 struct page_cgroup *pc;
952 struct mem_cgroup_per_zone *mz;
954 if (mem_cgroup_disabled())
956 pc = lookup_page_cgroup(page);
957 /* can happen while we handle swapcache. */
958 if (!TestClearPageCgroupAcctLRU(pc))
960 VM_BUG_ON(!pc->mem_cgroup);
962 * We don't check PCG_USED bit. It's cleared when the "page" is finally
963 * removed from global LRU.
965 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
966 /* huge page split is done under lru_lock. so, we have no races. */
967 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
968 VM_BUG_ON(list_empty(&pc->lru));
969 list_del_init(&pc->lru);
972 void mem_cgroup_del_lru(struct page *page)
974 mem_cgroup_del_lru_list(page, page_lru(page));
978 * Writeback is about to end against a page which has been marked for immediate
979 * reclaim. If it still appears to be reclaimable, move it to the tail of the
982 void mem_cgroup_rotate_reclaimable_page(struct page *page)
984 struct mem_cgroup_per_zone *mz;
985 struct page_cgroup *pc;
986 enum lru_list lru = page_lru(page);
988 if (mem_cgroup_disabled())
991 pc = lookup_page_cgroup(page);
992 /* unused page is not rotated. */
993 if (!PageCgroupUsed(pc))
995 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
997 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
998 list_move_tail(&pc->lru, &mz->lists[lru]);
1001 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
1003 struct mem_cgroup_per_zone *mz;
1004 struct page_cgroup *pc;
1006 if (mem_cgroup_disabled())
1009 pc = lookup_page_cgroup(page);
1010 /* unused page is not rotated. */
1011 if (!PageCgroupUsed(pc))
1013 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1015 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1016 list_move(&pc->lru, &mz->lists[lru]);
1019 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
1021 struct page_cgroup *pc;
1022 struct mem_cgroup_per_zone *mz;
1024 if (mem_cgroup_disabled())
1026 pc = lookup_page_cgroup(page);
1027 VM_BUG_ON(PageCgroupAcctLRU(pc));
1030 * SetPageLRU SetPageCgroupUsed
1032 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1034 * Ensure that one of the two sides adds the page to the memcg
1035 * LRU during a race.
1038 if (!PageCgroupUsed(pc))
1040 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1042 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1043 /* huge page split is done under lru_lock. so, we have no races. */
1044 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1045 SetPageCgroupAcctLRU(pc);
1046 list_add(&pc->lru, &mz->lists[lru]);
1050 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1051 * while it's linked to lru because the page may be reused after it's fully
1052 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1053 * It's done under lock_page and expected that zone->lru_lock isnever held.
1055 static void mem_cgroup_lru_del_before_commit(struct page *page)
1057 unsigned long flags;
1058 struct zone *zone = page_zone(page);
1059 struct page_cgroup *pc = lookup_page_cgroup(page);
1062 * Doing this check without taking ->lru_lock seems wrong but this
1063 * is safe. Because if page_cgroup's USED bit is unset, the page
1064 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1065 * set, the commit after this will fail, anyway.
1066 * This all charge/uncharge is done under some mutual execustion.
1067 * So, we don't need to taking care of changes in USED bit.
1069 if (likely(!PageLRU(page)))
1072 spin_lock_irqsave(&zone->lru_lock, flags);
1074 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1075 * is guarded by lock_page() because the page is SwapCache.
1077 if (!PageCgroupUsed(pc))
1078 mem_cgroup_del_lru_list(page, page_lru(page));
1079 spin_unlock_irqrestore(&zone->lru_lock, flags);
1082 static void mem_cgroup_lru_add_after_commit(struct page *page)
1084 unsigned long flags;
1085 struct zone *zone = page_zone(page);
1086 struct page_cgroup *pc = lookup_page_cgroup(page);
1089 * SetPageLRU SetPageCgroupUsed
1091 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1093 * Ensure that one of the two sides adds the page to the memcg
1094 * LRU during a race.
1097 /* taking care of that the page is added to LRU while we commit it */
1098 if (likely(!PageLRU(page)))
1100 spin_lock_irqsave(&zone->lru_lock, flags);
1101 /* link when the page is linked to LRU but page_cgroup isn't */
1102 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1103 mem_cgroup_add_lru_list(page, page_lru(page));
1104 spin_unlock_irqrestore(&zone->lru_lock, flags);
1108 void mem_cgroup_move_lists(struct page *page,
1109 enum lru_list from, enum lru_list to)
1111 if (mem_cgroup_disabled())
1113 mem_cgroup_del_lru_list(page, from);
1114 mem_cgroup_add_lru_list(page, to);
1118 * Checks whether given mem is same or in the root_mem_cgroup's
1121 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1122 struct mem_cgroup *memcg)
1124 if (root_memcg != memcg) {
1125 return (root_memcg->use_hierarchy &&
1126 css_is_ancestor(&memcg->css, &root_memcg->css));
1132 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1135 struct mem_cgroup *curr = NULL;
1136 struct task_struct *p;
1138 p = find_lock_task_mm(task);
1141 curr = try_get_mem_cgroup_from_mm(p->mm);
1146 * We should check use_hierarchy of "memcg" not "curr". Because checking
1147 * use_hierarchy of "curr" here make this function true if hierarchy is
1148 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1149 * hierarchy(even if use_hierarchy is disabled in "memcg").
1151 ret = mem_cgroup_same_or_subtree(memcg, curr);
1152 css_put(&curr->css);
1156 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1158 unsigned long inactive_ratio;
1159 int nid = zone_to_nid(zone);
1160 int zid = zone_idx(zone);
1161 unsigned long inactive;
1162 unsigned long active;
1165 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1166 BIT(LRU_INACTIVE_ANON));
1167 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1168 BIT(LRU_ACTIVE_ANON));
1170 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1172 inactive_ratio = int_sqrt(10 * gb);
1176 return inactive * inactive_ratio < active;
1179 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1181 unsigned long active;
1182 unsigned long inactive;
1183 int zid = zone_idx(zone);
1184 int nid = zone_to_nid(zone);
1186 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1187 BIT(LRU_INACTIVE_FILE));
1188 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1189 BIT(LRU_ACTIVE_FILE));
1191 return (active > inactive);
1194 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1197 int nid = zone_to_nid(zone);
1198 int zid = zone_idx(zone);
1199 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1201 return &mz->reclaim_stat;
1204 struct zone_reclaim_stat *
1205 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1207 struct page_cgroup *pc;
1208 struct mem_cgroup_per_zone *mz;
1210 if (mem_cgroup_disabled())
1213 pc = lookup_page_cgroup(page);
1214 if (!PageCgroupUsed(pc))
1216 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1218 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1219 return &mz->reclaim_stat;
1222 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1223 struct list_head *dst,
1224 unsigned long *scanned, int order,
1225 isolate_mode_t mode,
1227 struct mem_cgroup *mem_cont,
1228 int active, int file)
1230 unsigned long nr_taken = 0;
1234 struct list_head *src;
1235 struct page_cgroup *pc, *tmp;
1236 int nid = zone_to_nid(z);
1237 int zid = zone_idx(z);
1238 struct mem_cgroup_per_zone *mz;
1239 int lru = LRU_FILE * file + active;
1243 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1244 src = &mz->lists[lru];
1247 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1248 if (scan >= nr_to_scan)
1251 if (unlikely(!PageCgroupUsed(pc)))
1254 page = lookup_cgroup_page(pc);
1256 if (unlikely(!PageLRU(page)))
1260 ret = __isolate_lru_page(page, mode, file);
1263 list_move(&page->lru, dst);
1264 mem_cgroup_del_lru(page);
1265 nr_taken += hpage_nr_pages(page);
1268 /* we don't affect global LRU but rotate in our LRU */
1269 mem_cgroup_rotate_lru_list(page, page_lru(page));
1278 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1284 #define mem_cgroup_from_res_counter(counter, member) \
1285 container_of(counter, struct mem_cgroup, member)
1288 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1289 * @mem: the memory cgroup
1291 * Returns the maximum amount of memory @mem can be charged with, in
1294 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1296 unsigned long long margin;
1298 margin = res_counter_margin(&memcg->res);
1299 if (do_swap_account)
1300 margin = min(margin, res_counter_margin(&memcg->memsw));
1301 return margin >> PAGE_SHIFT;
1304 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1306 struct cgroup *cgrp = memcg->css.cgroup;
1309 if (cgrp->parent == NULL)
1310 return vm_swappiness;
1312 return memcg->swappiness;
1315 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1320 spin_lock(&memcg->pcp_counter_lock);
1321 for_each_online_cpu(cpu)
1322 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1323 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1324 spin_unlock(&memcg->pcp_counter_lock);
1330 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1337 spin_lock(&memcg->pcp_counter_lock);
1338 for_each_online_cpu(cpu)
1339 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1340 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1341 spin_unlock(&memcg->pcp_counter_lock);
1345 * 2 routines for checking "mem" is under move_account() or not.
1347 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1348 * for avoiding race in accounting. If true,
1349 * pc->mem_cgroup may be overwritten.
1351 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1352 * under hierarchy of moving cgroups. This is for
1353 * waiting at hith-memory prressure caused by "move".
1356 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1358 VM_BUG_ON(!rcu_read_lock_held());
1359 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1362 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1364 struct mem_cgroup *from;
1365 struct mem_cgroup *to;
1368 * Unlike task_move routines, we access mc.to, mc.from not under
1369 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1371 spin_lock(&mc.lock);
1377 ret = mem_cgroup_same_or_subtree(memcg, from)
1378 || mem_cgroup_same_or_subtree(memcg, to);
1380 spin_unlock(&mc.lock);
1384 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1386 if (mc.moving_task && current != mc.moving_task) {
1387 if (mem_cgroup_under_move(memcg)) {
1389 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1390 /* moving charge context might have finished. */
1393 finish_wait(&mc.waitq, &wait);
1401 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1402 * @memcg: The memory cgroup that went over limit
1403 * @p: Task that is going to be killed
1405 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1408 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1410 struct cgroup *task_cgrp;
1411 struct cgroup *mem_cgrp;
1413 * Need a buffer in BSS, can't rely on allocations. The code relies
1414 * on the assumption that OOM is serialized for memory controller.
1415 * If this assumption is broken, revisit this code.
1417 static char memcg_name[PATH_MAX];
1426 mem_cgrp = memcg->css.cgroup;
1427 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1429 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1432 * Unfortunately, we are unable to convert to a useful name
1433 * But we'll still print out the usage information
1440 printk(KERN_INFO "Task in %s killed", memcg_name);
1443 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1451 * Continues from above, so we don't need an KERN_ level
1453 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1456 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1457 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1458 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1459 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1460 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1462 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1463 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1464 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1468 * This function returns the number of memcg under hierarchy tree. Returns
1469 * 1(self count) if no children.
1471 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1474 struct mem_cgroup *iter;
1476 for_each_mem_cgroup_tree(iter, memcg)
1482 * Return the memory (and swap, if configured) limit for a memcg.
1484 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1489 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1490 limit += total_swap_pages << PAGE_SHIFT;
1492 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1494 * If memsw is finite and limits the amount of swap space available
1495 * to this memcg, return that limit.
1497 return min(limit, memsw);
1500 static unsigned long mem_cgroup_reclaim(struct mem_cgroup *memcg,
1502 unsigned long flags)
1504 unsigned long total = 0;
1505 bool noswap = false;
1508 if (flags & MEM_CGROUP_RECLAIM_NOSWAP)
1510 if (!(flags & MEM_CGROUP_RECLAIM_SHRINK) && memcg->memsw_is_minimum)
1513 for (loop = 0; loop < MEM_CGROUP_MAX_RECLAIM_LOOPS; loop++) {
1515 drain_all_stock_async(memcg);
1516 total += try_to_free_mem_cgroup_pages(memcg, gfp_mask, noswap);
1518 * Allow limit shrinkers, which are triggered directly
1519 * by userspace, to catch signals and stop reclaim
1520 * after minimal progress, regardless of the margin.
1522 if (total && (flags & MEM_CGROUP_RECLAIM_SHRINK))
1524 if (mem_cgroup_margin(memcg))
1527 * If nothing was reclaimed after two attempts, there
1528 * may be no reclaimable pages in this hierarchy.
1537 * test_mem_cgroup_node_reclaimable
1538 * @mem: the target memcg
1539 * @nid: the node ID to be checked.
1540 * @noswap : specify true here if the user wants flle only information.
1542 * This function returns whether the specified memcg contains any
1543 * reclaimable pages on a node. Returns true if there are any reclaimable
1544 * pages in the node.
1546 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1547 int nid, bool noswap)
1549 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1551 if (noswap || !total_swap_pages)
1553 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1558 #if MAX_NUMNODES > 1
1561 * Always updating the nodemask is not very good - even if we have an empty
1562 * list or the wrong list here, we can start from some node and traverse all
1563 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1566 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1570 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1571 * pagein/pageout changes since the last update.
1573 if (!atomic_read(&memcg->numainfo_events))
1575 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1578 /* make a nodemask where this memcg uses memory from */
1579 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1581 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1583 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1584 node_clear(nid, memcg->scan_nodes);
1587 atomic_set(&memcg->numainfo_events, 0);
1588 atomic_set(&memcg->numainfo_updating, 0);
1592 * Selecting a node where we start reclaim from. Because what we need is just
1593 * reducing usage counter, start from anywhere is O,K. Considering
1594 * memory reclaim from current node, there are pros. and cons.
1596 * Freeing memory from current node means freeing memory from a node which
1597 * we'll use or we've used. So, it may make LRU bad. And if several threads
1598 * hit limits, it will see a contention on a node. But freeing from remote
1599 * node means more costs for memory reclaim because of memory latency.
1601 * Now, we use round-robin. Better algorithm is welcomed.
1603 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1607 mem_cgroup_may_update_nodemask(memcg);
1608 node = memcg->last_scanned_node;
1610 node = next_node(node, memcg->scan_nodes);
1611 if (node == MAX_NUMNODES)
1612 node = first_node(memcg->scan_nodes);
1614 * We call this when we hit limit, not when pages are added to LRU.
1615 * No LRU may hold pages because all pages are UNEVICTABLE or
1616 * memcg is too small and all pages are not on LRU. In that case,
1617 * we use curret node.
1619 if (unlikely(node == MAX_NUMNODES))
1620 node = numa_node_id();
1622 memcg->last_scanned_node = node;
1627 * Check all nodes whether it contains reclaimable pages or not.
1628 * For quick scan, we make use of scan_nodes. This will allow us to skip
1629 * unused nodes. But scan_nodes is lazily updated and may not cotain
1630 * enough new information. We need to do double check.
1632 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1637 * quick check...making use of scan_node.
1638 * We can skip unused nodes.
1640 if (!nodes_empty(memcg->scan_nodes)) {
1641 for (nid = first_node(memcg->scan_nodes);
1643 nid = next_node(nid, memcg->scan_nodes)) {
1645 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1650 * Check rest of nodes.
1652 for_each_node_state(nid, N_HIGH_MEMORY) {
1653 if (node_isset(nid, memcg->scan_nodes))
1655 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1662 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1667 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1669 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1673 static int mem_cgroup_soft_reclaim(struct mem_cgroup *root_memcg,
1676 unsigned long *total_scanned)
1678 struct mem_cgroup *victim = NULL;
1681 unsigned long excess;
1682 unsigned long nr_scanned;
1683 struct mem_cgroup_reclaim_cookie reclaim = {
1688 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1691 victim = mem_cgroup_iter(root_memcg, victim, &reclaim);
1696 * If we have not been able to reclaim
1697 * anything, it might because there are
1698 * no reclaimable pages under this hierarchy
1703 * We want to do more targeted reclaim.
1704 * excess >> 2 is not to excessive so as to
1705 * reclaim too much, nor too less that we keep
1706 * coming back to reclaim from this cgroup
1708 if (total >= (excess >> 2) ||
1709 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS))
1714 if (!mem_cgroup_reclaimable(victim, false))
1716 total += mem_cgroup_shrink_node_zone(victim, gfp_mask, false,
1718 *total_scanned += nr_scanned;
1719 if (!res_counter_soft_limit_excess(&root_memcg->res))
1722 mem_cgroup_iter_break(root_memcg, victim);
1727 * Check OOM-Killer is already running under our hierarchy.
1728 * If someone is running, return false.
1729 * Has to be called with memcg_oom_lock
1731 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1733 struct mem_cgroup *iter, *failed = NULL;
1735 for_each_mem_cgroup_tree(iter, memcg) {
1736 if (iter->oom_lock) {
1738 * this subtree of our hierarchy is already locked
1739 * so we cannot give a lock.
1742 mem_cgroup_iter_break(memcg, iter);
1745 iter->oom_lock = true;
1752 * OK, we failed to lock the whole subtree so we have to clean up
1753 * what we set up to the failing subtree
1755 for_each_mem_cgroup_tree(iter, memcg) {
1756 if (iter == failed) {
1757 mem_cgroup_iter_break(memcg, iter);
1760 iter->oom_lock = false;
1766 * Has to be called with memcg_oom_lock
1768 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1770 struct mem_cgroup *iter;
1772 for_each_mem_cgroup_tree(iter, memcg)
1773 iter->oom_lock = false;
1777 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1779 struct mem_cgroup *iter;
1781 for_each_mem_cgroup_tree(iter, memcg)
1782 atomic_inc(&iter->under_oom);
1785 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1787 struct mem_cgroup *iter;
1790 * When a new child is created while the hierarchy is under oom,
1791 * mem_cgroup_oom_lock() may not be called. We have to use
1792 * atomic_add_unless() here.
1794 for_each_mem_cgroup_tree(iter, memcg)
1795 atomic_add_unless(&iter->under_oom, -1, 0);
1798 static DEFINE_SPINLOCK(memcg_oom_lock);
1799 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1801 struct oom_wait_info {
1802 struct mem_cgroup *mem;
1806 static int memcg_oom_wake_function(wait_queue_t *wait,
1807 unsigned mode, int sync, void *arg)
1809 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1811 struct oom_wait_info *oom_wait_info;
1813 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1814 oom_wait_memcg = oom_wait_info->mem;
1817 * Both of oom_wait_info->mem and wake_mem are stable under us.
1818 * Then we can use css_is_ancestor without taking care of RCU.
1820 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1821 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1823 return autoremove_wake_function(wait, mode, sync, arg);
1826 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1828 /* for filtering, pass "memcg" as argument. */
1829 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1832 static void memcg_oom_recover(struct mem_cgroup *memcg)
1834 if (memcg && atomic_read(&memcg->under_oom))
1835 memcg_wakeup_oom(memcg);
1839 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1841 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1843 struct oom_wait_info owait;
1844 bool locked, need_to_kill;
1847 owait.wait.flags = 0;
1848 owait.wait.func = memcg_oom_wake_function;
1849 owait.wait.private = current;
1850 INIT_LIST_HEAD(&owait.wait.task_list);
1851 need_to_kill = true;
1852 mem_cgroup_mark_under_oom(memcg);
1854 /* At first, try to OOM lock hierarchy under memcg.*/
1855 spin_lock(&memcg_oom_lock);
1856 locked = mem_cgroup_oom_lock(memcg);
1858 * Even if signal_pending(), we can't quit charge() loop without
1859 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1860 * under OOM is always welcomed, use TASK_KILLABLE here.
1862 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1863 if (!locked || memcg->oom_kill_disable)
1864 need_to_kill = false;
1866 mem_cgroup_oom_notify(memcg);
1867 spin_unlock(&memcg_oom_lock);
1870 finish_wait(&memcg_oom_waitq, &owait.wait);
1871 mem_cgroup_out_of_memory(memcg, mask);
1874 finish_wait(&memcg_oom_waitq, &owait.wait);
1876 spin_lock(&memcg_oom_lock);
1878 mem_cgroup_oom_unlock(memcg);
1879 memcg_wakeup_oom(memcg);
1880 spin_unlock(&memcg_oom_lock);
1882 mem_cgroup_unmark_under_oom(memcg);
1884 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1886 /* Give chance to dying process */
1887 schedule_timeout_uninterruptible(1);
1892 * Currently used to update mapped file statistics, but the routine can be
1893 * generalized to update other statistics as well.
1895 * Notes: Race condition
1897 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1898 * it tends to be costly. But considering some conditions, we doesn't need
1899 * to do so _always_.
1901 * Considering "charge", lock_page_cgroup() is not required because all
1902 * file-stat operations happen after a page is attached to radix-tree. There
1903 * are no race with "charge".
1905 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1906 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1907 * if there are race with "uncharge". Statistics itself is properly handled
1910 * Considering "move", this is an only case we see a race. To make the race
1911 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1912 * possibility of race condition. If there is, we take a lock.
1915 void mem_cgroup_update_page_stat(struct page *page,
1916 enum mem_cgroup_page_stat_item idx, int val)
1918 struct mem_cgroup *memcg;
1919 struct page_cgroup *pc = lookup_page_cgroup(page);
1920 bool need_unlock = false;
1921 unsigned long uninitialized_var(flags);
1927 memcg = pc->mem_cgroup;
1928 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1930 /* pc->mem_cgroup is unstable ? */
1931 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
1932 /* take a lock against to access pc->mem_cgroup */
1933 move_lock_page_cgroup(pc, &flags);
1935 memcg = pc->mem_cgroup;
1936 if (!memcg || !PageCgroupUsed(pc))
1941 case MEMCG_NR_FILE_MAPPED:
1943 SetPageCgroupFileMapped(pc);
1944 else if (!page_mapped(page))
1945 ClearPageCgroupFileMapped(pc);
1946 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1952 this_cpu_add(memcg->stat->count[idx], val);
1955 if (unlikely(need_unlock))
1956 move_unlock_page_cgroup(pc, &flags);
1960 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1963 * size of first charge trial. "32" comes from vmscan.c's magic value.
1964 * TODO: maybe necessary to use big numbers in big irons.
1966 #define CHARGE_BATCH 32U
1967 struct memcg_stock_pcp {
1968 struct mem_cgroup *cached; /* this never be root cgroup */
1969 unsigned int nr_pages;
1970 struct work_struct work;
1971 unsigned long flags;
1972 #define FLUSHING_CACHED_CHARGE (0)
1974 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1975 static DEFINE_MUTEX(percpu_charge_mutex);
1978 * Try to consume stocked charge on this cpu. If success, one page is consumed
1979 * from local stock and true is returned. If the stock is 0 or charges from a
1980 * cgroup which is not current target, returns false. This stock will be
1983 static bool consume_stock(struct mem_cgroup *memcg)
1985 struct memcg_stock_pcp *stock;
1988 stock = &get_cpu_var(memcg_stock);
1989 if (memcg == stock->cached && stock->nr_pages)
1991 else /* need to call res_counter_charge */
1993 put_cpu_var(memcg_stock);
1998 * Returns stocks cached in percpu to res_counter and reset cached information.
2000 static void drain_stock(struct memcg_stock_pcp *stock)
2002 struct mem_cgroup *old = stock->cached;
2004 if (stock->nr_pages) {
2005 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2007 res_counter_uncharge(&old->res, bytes);
2008 if (do_swap_account)
2009 res_counter_uncharge(&old->memsw, bytes);
2010 stock->nr_pages = 0;
2012 stock->cached = NULL;
2016 * This must be called under preempt disabled or must be called by
2017 * a thread which is pinned to local cpu.
2019 static void drain_local_stock(struct work_struct *dummy)
2021 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2023 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2027 * Cache charges(val) which is from res_counter, to local per_cpu area.
2028 * This will be consumed by consume_stock() function, later.
2030 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2032 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2034 if (stock->cached != memcg) { /* reset if necessary */
2036 stock->cached = memcg;
2038 stock->nr_pages += nr_pages;
2039 put_cpu_var(memcg_stock);
2043 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2044 * of the hierarchy under it. sync flag says whether we should block
2045 * until the work is done.
2047 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2051 /* Notify other cpus that system-wide "drain" is running */
2054 for_each_online_cpu(cpu) {
2055 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2056 struct mem_cgroup *memcg;
2058 memcg = stock->cached;
2059 if (!memcg || !stock->nr_pages)
2061 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2063 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2065 drain_local_stock(&stock->work);
2067 schedule_work_on(cpu, &stock->work);
2075 for_each_online_cpu(cpu) {
2076 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2077 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2078 flush_work(&stock->work);
2085 * Tries to drain stocked charges in other cpus. This function is asynchronous
2086 * and just put a work per cpu for draining localy on each cpu. Caller can
2087 * expects some charges will be back to res_counter later but cannot wait for
2090 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2093 * If someone calls draining, avoid adding more kworker runs.
2095 if (!mutex_trylock(&percpu_charge_mutex))
2097 drain_all_stock(root_memcg, false);
2098 mutex_unlock(&percpu_charge_mutex);
2101 /* This is a synchronous drain interface. */
2102 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2104 /* called when force_empty is called */
2105 mutex_lock(&percpu_charge_mutex);
2106 drain_all_stock(root_memcg, true);
2107 mutex_unlock(&percpu_charge_mutex);
2111 * This function drains percpu counter value from DEAD cpu and
2112 * move it to local cpu. Note that this function can be preempted.
2114 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2118 spin_lock(&memcg->pcp_counter_lock);
2119 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2120 long x = per_cpu(memcg->stat->count[i], cpu);
2122 per_cpu(memcg->stat->count[i], cpu) = 0;
2123 memcg->nocpu_base.count[i] += x;
2125 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2126 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2128 per_cpu(memcg->stat->events[i], cpu) = 0;
2129 memcg->nocpu_base.events[i] += x;
2131 /* need to clear ON_MOVE value, works as a kind of lock. */
2132 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2133 spin_unlock(&memcg->pcp_counter_lock);
2136 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2138 int idx = MEM_CGROUP_ON_MOVE;
2140 spin_lock(&memcg->pcp_counter_lock);
2141 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2142 spin_unlock(&memcg->pcp_counter_lock);
2145 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2146 unsigned long action,
2149 int cpu = (unsigned long)hcpu;
2150 struct memcg_stock_pcp *stock;
2151 struct mem_cgroup *iter;
2153 if ((action == CPU_ONLINE)) {
2154 for_each_mem_cgroup(iter)
2155 synchronize_mem_cgroup_on_move(iter, cpu);
2159 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2162 for_each_mem_cgroup(iter)
2163 mem_cgroup_drain_pcp_counter(iter, cpu);
2165 stock = &per_cpu(memcg_stock, cpu);
2171 /* See __mem_cgroup_try_charge() for details */
2173 CHARGE_OK, /* success */
2174 CHARGE_RETRY, /* need to retry but retry is not bad */
2175 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2176 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2177 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2180 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2181 unsigned int nr_pages, bool oom_check)
2183 unsigned long csize = nr_pages * PAGE_SIZE;
2184 struct mem_cgroup *mem_over_limit;
2185 struct res_counter *fail_res;
2186 unsigned long flags = 0;
2189 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2192 if (!do_swap_account)
2194 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2198 res_counter_uncharge(&memcg->res, csize);
2199 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2200 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2202 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2204 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2205 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2207 * Never reclaim on behalf of optional batching, retry with a
2208 * single page instead.
2210 if (nr_pages == CHARGE_BATCH)
2211 return CHARGE_RETRY;
2213 if (!(gfp_mask & __GFP_WAIT))
2214 return CHARGE_WOULDBLOCK;
2216 ret = mem_cgroup_reclaim(mem_over_limit, gfp_mask, flags);
2217 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2218 return CHARGE_RETRY;
2220 * Even though the limit is exceeded at this point, reclaim
2221 * may have been able to free some pages. Retry the charge
2222 * before killing the task.
2224 * Only for regular pages, though: huge pages are rather
2225 * unlikely to succeed so close to the limit, and we fall back
2226 * to regular pages anyway in case of failure.
2228 if (nr_pages == 1 && ret)
2229 return CHARGE_RETRY;
2232 * At task move, charge accounts can be doubly counted. So, it's
2233 * better to wait until the end of task_move if something is going on.
2235 if (mem_cgroup_wait_acct_move(mem_over_limit))
2236 return CHARGE_RETRY;
2238 /* If we don't need to call oom-killer at el, return immediately */
2240 return CHARGE_NOMEM;
2242 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2243 return CHARGE_OOM_DIE;
2245 return CHARGE_RETRY;
2249 * Unlike exported interface, "oom" parameter is added. if oom==true,
2250 * oom-killer can be invoked.
2252 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2254 unsigned int nr_pages,
2255 struct mem_cgroup **ptr,
2258 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2259 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2260 struct mem_cgroup *memcg = NULL;
2264 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2265 * in system level. So, allow to go ahead dying process in addition to
2268 if (unlikely(test_thread_flag(TIF_MEMDIE)
2269 || fatal_signal_pending(current)))
2273 * We always charge the cgroup the mm_struct belongs to.
2274 * The mm_struct's mem_cgroup changes on task migration if the
2275 * thread group leader migrates. It's possible that mm is not
2276 * set, if so charge the init_mm (happens for pagecache usage).
2281 if (*ptr) { /* css should be a valid one */
2283 VM_BUG_ON(css_is_removed(&memcg->css));
2284 if (mem_cgroup_is_root(memcg))
2286 if (nr_pages == 1 && consume_stock(memcg))
2288 css_get(&memcg->css);
2290 struct task_struct *p;
2293 p = rcu_dereference(mm->owner);
2295 * Because we don't have task_lock(), "p" can exit.
2296 * In that case, "memcg" can point to root or p can be NULL with
2297 * race with swapoff. Then, we have small risk of mis-accouning.
2298 * But such kind of mis-account by race always happens because
2299 * we don't have cgroup_mutex(). It's overkill and we allo that
2301 * (*) swapoff at el will charge against mm-struct not against
2302 * task-struct. So, mm->owner can be NULL.
2304 memcg = mem_cgroup_from_task(p);
2305 if (!memcg || mem_cgroup_is_root(memcg)) {
2309 if (nr_pages == 1 && consume_stock(memcg)) {
2311 * It seems dagerous to access memcg without css_get().
2312 * But considering how consume_stok works, it's not
2313 * necessary. If consume_stock success, some charges
2314 * from this memcg are cached on this cpu. So, we
2315 * don't need to call css_get()/css_tryget() before
2316 * calling consume_stock().
2321 /* after here, we may be blocked. we need to get refcnt */
2322 if (!css_tryget(&memcg->css)) {
2332 /* If killed, bypass charge */
2333 if (fatal_signal_pending(current)) {
2334 css_put(&memcg->css);
2339 if (oom && !nr_oom_retries) {
2341 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2344 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2348 case CHARGE_RETRY: /* not in OOM situation but retry */
2350 css_put(&memcg->css);
2353 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2354 css_put(&memcg->css);
2356 case CHARGE_NOMEM: /* OOM routine works */
2358 css_put(&memcg->css);
2361 /* If oom, we never return -ENOMEM */
2364 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2365 css_put(&memcg->css);
2368 } while (ret != CHARGE_OK);
2370 if (batch > nr_pages)
2371 refill_stock(memcg, batch - nr_pages);
2372 css_put(&memcg->css);
2385 * Somemtimes we have to undo a charge we got by try_charge().
2386 * This function is for that and do uncharge, put css's refcnt.
2387 * gotten by try_charge().
2389 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2390 unsigned int nr_pages)
2392 if (!mem_cgroup_is_root(memcg)) {
2393 unsigned long bytes = nr_pages * PAGE_SIZE;
2395 res_counter_uncharge(&memcg->res, bytes);
2396 if (do_swap_account)
2397 res_counter_uncharge(&memcg->memsw, bytes);
2402 * A helper function to get mem_cgroup from ID. must be called under
2403 * rcu_read_lock(). The caller must check css_is_removed() or some if
2404 * it's concern. (dropping refcnt from swap can be called against removed
2407 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2409 struct cgroup_subsys_state *css;
2411 /* ID 0 is unused ID */
2414 css = css_lookup(&mem_cgroup_subsys, id);
2417 return container_of(css, struct mem_cgroup, css);
2420 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2422 struct mem_cgroup *memcg = NULL;
2423 struct page_cgroup *pc;
2427 VM_BUG_ON(!PageLocked(page));
2429 pc = lookup_page_cgroup(page);
2430 lock_page_cgroup(pc);
2431 if (PageCgroupUsed(pc)) {
2432 memcg = pc->mem_cgroup;
2433 if (memcg && !css_tryget(&memcg->css))
2435 } else if (PageSwapCache(page)) {
2436 ent.val = page_private(page);
2437 id = lookup_swap_cgroup(ent);
2439 memcg = mem_cgroup_lookup(id);
2440 if (memcg && !css_tryget(&memcg->css))
2444 unlock_page_cgroup(pc);
2448 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2450 unsigned int nr_pages,
2451 struct page_cgroup *pc,
2452 enum charge_type ctype)
2454 lock_page_cgroup(pc);
2455 if (unlikely(PageCgroupUsed(pc))) {
2456 unlock_page_cgroup(pc);
2457 __mem_cgroup_cancel_charge(memcg, nr_pages);
2461 * we don't need page_cgroup_lock about tail pages, becase they are not
2462 * accessed by any other context at this point.
2464 pc->mem_cgroup = memcg;
2466 * We access a page_cgroup asynchronously without lock_page_cgroup().
2467 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2468 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2469 * before USED bit, we need memory barrier here.
2470 * See mem_cgroup_add_lru_list(), etc.
2474 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2475 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2476 SetPageCgroupCache(pc);
2477 SetPageCgroupUsed(pc);
2479 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2480 ClearPageCgroupCache(pc);
2481 SetPageCgroupUsed(pc);
2487 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2488 unlock_page_cgroup(pc);
2490 * "charge_statistics" updated event counter. Then, check it.
2491 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2492 * if they exceeds softlimit.
2494 memcg_check_events(memcg, page);
2497 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2499 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2500 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2502 * Because tail pages are not marked as "used", set it. We're under
2503 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2505 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2507 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2508 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2509 unsigned long flags;
2511 if (mem_cgroup_disabled())
2514 * We have no races with charge/uncharge but will have races with
2515 * page state accounting.
2517 move_lock_page_cgroup(head_pc, &flags);
2519 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2520 smp_wmb(); /* see __commit_charge() */
2521 if (PageCgroupAcctLRU(head_pc)) {
2523 struct mem_cgroup_per_zone *mz;
2526 * LRU flags cannot be copied because we need to add tail
2527 *.page to LRU by generic call and our hook will be called.
2528 * We hold lru_lock, then, reduce counter directly.
2530 lru = page_lru(head);
2531 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2532 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2534 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2535 move_unlock_page_cgroup(head_pc, &flags);
2540 * mem_cgroup_move_account - move account of the page
2542 * @nr_pages: number of regular pages (>1 for huge pages)
2543 * @pc: page_cgroup of the page.
2544 * @from: mem_cgroup which the page is moved from.
2545 * @to: mem_cgroup which the page is moved to. @from != @to.
2546 * @uncharge: whether we should call uncharge and css_put against @from.
2548 * The caller must confirm following.
2549 * - page is not on LRU (isolate_page() is useful.)
2550 * - compound_lock is held when nr_pages > 1
2552 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2553 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2554 * true, this function does "uncharge" from old cgroup, but it doesn't if
2555 * @uncharge is false, so a caller should do "uncharge".
2557 static int mem_cgroup_move_account(struct page *page,
2558 unsigned int nr_pages,
2559 struct page_cgroup *pc,
2560 struct mem_cgroup *from,
2561 struct mem_cgroup *to,
2564 unsigned long flags;
2567 VM_BUG_ON(from == to);
2568 VM_BUG_ON(PageLRU(page));
2570 * The page is isolated from LRU. So, collapse function
2571 * will not handle this page. But page splitting can happen.
2572 * Do this check under compound_page_lock(). The caller should
2576 if (nr_pages > 1 && !PageTransHuge(page))
2579 lock_page_cgroup(pc);
2582 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2585 move_lock_page_cgroup(pc, &flags);
2587 if (PageCgroupFileMapped(pc)) {
2588 /* Update mapped_file data for mem_cgroup */
2590 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2591 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2594 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2596 /* This is not "cancel", but cancel_charge does all we need. */
2597 __mem_cgroup_cancel_charge(from, nr_pages);
2599 /* caller should have done css_get */
2600 pc->mem_cgroup = to;
2601 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2603 * We charges against "to" which may not have any tasks. Then, "to"
2604 * can be under rmdir(). But in current implementation, caller of
2605 * this function is just force_empty() and move charge, so it's
2606 * guaranteed that "to" is never removed. So, we don't check rmdir
2609 move_unlock_page_cgroup(pc, &flags);
2612 unlock_page_cgroup(pc);
2616 memcg_check_events(to, page);
2617 memcg_check_events(from, page);
2623 * move charges to its parent.
2626 static int mem_cgroup_move_parent(struct page *page,
2627 struct page_cgroup *pc,
2628 struct mem_cgroup *child,
2631 struct cgroup *cg = child->css.cgroup;
2632 struct cgroup *pcg = cg->parent;
2633 struct mem_cgroup *parent;
2634 unsigned int nr_pages;
2635 unsigned long uninitialized_var(flags);
2643 if (!get_page_unless_zero(page))
2645 if (isolate_lru_page(page))
2648 nr_pages = hpage_nr_pages(page);
2650 parent = mem_cgroup_from_cont(pcg);
2651 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2656 flags = compound_lock_irqsave(page);
2658 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2660 __mem_cgroup_cancel_charge(parent, nr_pages);
2663 compound_unlock_irqrestore(page, flags);
2665 putback_lru_page(page);
2673 * Charge the memory controller for page usage.
2675 * 0 if the charge was successful
2676 * < 0 if the cgroup is over its limit
2678 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2679 gfp_t gfp_mask, enum charge_type ctype)
2681 struct mem_cgroup *memcg = NULL;
2682 unsigned int nr_pages = 1;
2683 struct page_cgroup *pc;
2687 if (PageTransHuge(page)) {
2688 nr_pages <<= compound_order(page);
2689 VM_BUG_ON(!PageTransHuge(page));
2691 * Never OOM-kill a process for a huge page. The
2692 * fault handler will fall back to regular pages.
2697 pc = lookup_page_cgroup(page);
2698 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2700 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2704 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2708 int mem_cgroup_newpage_charge(struct page *page,
2709 struct mm_struct *mm, gfp_t gfp_mask)
2711 if (mem_cgroup_disabled())
2714 * If already mapped, we don't have to account.
2715 * If page cache, page->mapping has address_space.
2716 * But page->mapping may have out-of-use anon_vma pointer,
2717 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2720 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2724 return mem_cgroup_charge_common(page, mm, gfp_mask,
2725 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2729 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2730 enum charge_type ctype);
2733 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2734 enum charge_type ctype)
2736 struct page_cgroup *pc = lookup_page_cgroup(page);
2738 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2739 * is already on LRU. It means the page may on some other page_cgroup's
2740 * LRU. Take care of it.
2742 mem_cgroup_lru_del_before_commit(page);
2743 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2744 mem_cgroup_lru_add_after_commit(page);
2748 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2751 struct mem_cgroup *memcg = NULL;
2754 if (mem_cgroup_disabled())
2756 if (PageCompound(page))
2762 if (page_is_file_cache(page)) {
2763 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2768 * FUSE reuses pages without going through the final
2769 * put that would remove them from the LRU list, make
2770 * sure that they get relinked properly.
2772 __mem_cgroup_commit_charge_lrucare(page, memcg,
2773 MEM_CGROUP_CHARGE_TYPE_CACHE);
2777 if (PageSwapCache(page)) {
2778 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2780 __mem_cgroup_commit_charge_swapin(page, memcg,
2781 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2783 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2784 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2790 * While swap-in, try_charge -> commit or cancel, the page is locked.
2791 * And when try_charge() successfully returns, one refcnt to memcg without
2792 * struct page_cgroup is acquired. This refcnt will be consumed by
2793 * "commit()" or removed by "cancel()"
2795 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2797 gfp_t mask, struct mem_cgroup **ptr)
2799 struct mem_cgroup *memcg;
2804 if (mem_cgroup_disabled())
2807 if (!do_swap_account)
2810 * A racing thread's fault, or swapoff, may have already updated
2811 * the pte, and even removed page from swap cache: in those cases
2812 * do_swap_page()'s pte_same() test will fail; but there's also a
2813 * KSM case which does need to charge the page.
2815 if (!PageSwapCache(page))
2817 memcg = try_get_mem_cgroup_from_page(page);
2821 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2822 css_put(&memcg->css);
2827 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2831 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2832 enum charge_type ctype)
2834 if (mem_cgroup_disabled())
2838 cgroup_exclude_rmdir(&ptr->css);
2840 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2842 * Now swap is on-memory. This means this page may be
2843 * counted both as mem and swap....double count.
2844 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2845 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2846 * may call delete_from_swap_cache() before reach here.
2848 if (do_swap_account && PageSwapCache(page)) {
2849 swp_entry_t ent = {.val = page_private(page)};
2851 struct mem_cgroup *memcg;
2853 id = swap_cgroup_record(ent, 0);
2855 memcg = mem_cgroup_lookup(id);
2858 * This recorded memcg can be obsolete one. So, avoid
2859 * calling css_tryget
2861 if (!mem_cgroup_is_root(memcg))
2862 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2863 mem_cgroup_swap_statistics(memcg, false);
2864 mem_cgroup_put(memcg);
2869 * At swapin, we may charge account against cgroup which has no tasks.
2870 * So, rmdir()->pre_destroy() can be called while we do this charge.
2871 * In that case, we need to call pre_destroy() again. check it here.
2873 cgroup_release_and_wakeup_rmdir(&ptr->css);
2876 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2878 __mem_cgroup_commit_charge_swapin(page, ptr,
2879 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2882 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2884 if (mem_cgroup_disabled())
2888 __mem_cgroup_cancel_charge(memcg, 1);
2891 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2892 unsigned int nr_pages,
2893 const enum charge_type ctype)
2895 struct memcg_batch_info *batch = NULL;
2896 bool uncharge_memsw = true;
2898 /* If swapout, usage of swap doesn't decrease */
2899 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2900 uncharge_memsw = false;
2902 batch = ¤t->memcg_batch;
2904 * In usual, we do css_get() when we remember memcg pointer.
2905 * But in this case, we keep res->usage until end of a series of
2906 * uncharges. Then, it's ok to ignore memcg's refcnt.
2909 batch->memcg = memcg;
2911 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2912 * In those cases, all pages freed continuously can be expected to be in
2913 * the same cgroup and we have chance to coalesce uncharges.
2914 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2915 * because we want to do uncharge as soon as possible.
2918 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2919 goto direct_uncharge;
2922 goto direct_uncharge;
2925 * In typical case, batch->memcg == mem. This means we can
2926 * merge a series of uncharges to an uncharge of res_counter.
2927 * If not, we uncharge res_counter ony by one.
2929 if (batch->memcg != memcg)
2930 goto direct_uncharge;
2931 /* remember freed charge and uncharge it later */
2934 batch->memsw_nr_pages++;
2937 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2939 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2940 if (unlikely(batch->memcg != memcg))
2941 memcg_oom_recover(memcg);
2946 * uncharge if !page_mapped(page)
2948 static struct mem_cgroup *
2949 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2951 struct mem_cgroup *memcg = NULL;
2952 unsigned int nr_pages = 1;
2953 struct page_cgroup *pc;
2955 if (mem_cgroup_disabled())
2958 if (PageSwapCache(page))
2961 if (PageTransHuge(page)) {
2962 nr_pages <<= compound_order(page);
2963 VM_BUG_ON(!PageTransHuge(page));
2966 * Check if our page_cgroup is valid
2968 pc = lookup_page_cgroup(page);
2969 if (unlikely(!pc || !PageCgroupUsed(pc)))
2972 lock_page_cgroup(pc);
2974 memcg = pc->mem_cgroup;
2976 if (!PageCgroupUsed(pc))
2980 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2981 case MEM_CGROUP_CHARGE_TYPE_DROP:
2982 /* See mem_cgroup_prepare_migration() */
2983 if (page_mapped(page) || PageCgroupMigration(pc))
2986 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2987 if (!PageAnon(page)) { /* Shared memory */
2988 if (page->mapping && !page_is_file_cache(page))
2990 } else if (page_mapped(page)) /* Anon */
2997 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
2999 ClearPageCgroupUsed(pc);
3001 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3002 * freed from LRU. This is safe because uncharged page is expected not
3003 * to be reused (freed soon). Exception is SwapCache, it's handled by
3004 * special functions.
3007 unlock_page_cgroup(pc);
3009 * even after unlock, we have memcg->res.usage here and this memcg
3010 * will never be freed.
3012 memcg_check_events(memcg, page);
3013 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3014 mem_cgroup_swap_statistics(memcg, true);
3015 mem_cgroup_get(memcg);
3017 if (!mem_cgroup_is_root(memcg))
3018 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3023 unlock_page_cgroup(pc);
3027 void mem_cgroup_uncharge_page(struct page *page)
3030 if (page_mapped(page))
3032 if (page->mapping && !PageAnon(page))
3034 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3037 void mem_cgroup_uncharge_cache_page(struct page *page)
3039 VM_BUG_ON(page_mapped(page));
3040 VM_BUG_ON(page->mapping);
3041 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3045 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3046 * In that cases, pages are freed continuously and we can expect pages
3047 * are in the same memcg. All these calls itself limits the number of
3048 * pages freed at once, then uncharge_start/end() is called properly.
3049 * This may be called prural(2) times in a context,
3052 void mem_cgroup_uncharge_start(void)
3054 current->memcg_batch.do_batch++;
3055 /* We can do nest. */
3056 if (current->memcg_batch.do_batch == 1) {
3057 current->memcg_batch.memcg = NULL;
3058 current->memcg_batch.nr_pages = 0;
3059 current->memcg_batch.memsw_nr_pages = 0;
3063 void mem_cgroup_uncharge_end(void)
3065 struct memcg_batch_info *batch = ¤t->memcg_batch;
3067 if (!batch->do_batch)
3071 if (batch->do_batch) /* If stacked, do nothing. */
3077 * This "batch->memcg" is valid without any css_get/put etc...
3078 * bacause we hide charges behind us.
3080 if (batch->nr_pages)
3081 res_counter_uncharge(&batch->memcg->res,
3082 batch->nr_pages * PAGE_SIZE);
3083 if (batch->memsw_nr_pages)
3084 res_counter_uncharge(&batch->memcg->memsw,
3085 batch->memsw_nr_pages * PAGE_SIZE);
3086 memcg_oom_recover(batch->memcg);
3087 /* forget this pointer (for sanity check) */
3088 batch->memcg = NULL;
3093 * called after __delete_from_swap_cache() and drop "page" account.
3094 * memcg information is recorded to swap_cgroup of "ent"
3097 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3099 struct mem_cgroup *memcg;
3100 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3102 if (!swapout) /* this was a swap cache but the swap is unused ! */
3103 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3105 memcg = __mem_cgroup_uncharge_common(page, ctype);
3108 * record memcg information, if swapout && memcg != NULL,
3109 * mem_cgroup_get() was called in uncharge().
3111 if (do_swap_account && swapout && memcg)
3112 swap_cgroup_record(ent, css_id(&memcg->css));
3116 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3118 * called from swap_entry_free(). remove record in swap_cgroup and
3119 * uncharge "memsw" account.
3121 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3123 struct mem_cgroup *memcg;
3126 if (!do_swap_account)
3129 id = swap_cgroup_record(ent, 0);
3131 memcg = mem_cgroup_lookup(id);
3134 * We uncharge this because swap is freed.
3135 * This memcg can be obsolete one. We avoid calling css_tryget
3137 if (!mem_cgroup_is_root(memcg))
3138 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3139 mem_cgroup_swap_statistics(memcg, false);
3140 mem_cgroup_put(memcg);
3146 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3147 * @entry: swap entry to be moved
3148 * @from: mem_cgroup which the entry is moved from
3149 * @to: mem_cgroup which the entry is moved to
3150 * @need_fixup: whether we should fixup res_counters and refcounts.
3152 * It succeeds only when the swap_cgroup's record for this entry is the same
3153 * as the mem_cgroup's id of @from.
3155 * Returns 0 on success, -EINVAL on failure.
3157 * The caller must have charged to @to, IOW, called res_counter_charge() about
3158 * both res and memsw, and called css_get().
3160 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3161 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3163 unsigned short old_id, new_id;
3165 old_id = css_id(&from->css);
3166 new_id = css_id(&to->css);
3168 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3169 mem_cgroup_swap_statistics(from, false);
3170 mem_cgroup_swap_statistics(to, true);
3172 * This function is only called from task migration context now.
3173 * It postpones res_counter and refcount handling till the end
3174 * of task migration(mem_cgroup_clear_mc()) for performance
3175 * improvement. But we cannot postpone mem_cgroup_get(to)
3176 * because if the process that has been moved to @to does
3177 * swap-in, the refcount of @to might be decreased to 0.
3181 if (!mem_cgroup_is_root(from))
3182 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3183 mem_cgroup_put(from);
3185 * we charged both to->res and to->memsw, so we should
3188 if (!mem_cgroup_is_root(to))
3189 res_counter_uncharge(&to->res, PAGE_SIZE);
3196 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3197 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3204 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3207 int mem_cgroup_prepare_migration(struct page *page,
3208 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3210 struct mem_cgroup *memcg = NULL;
3211 struct page_cgroup *pc;
3212 enum charge_type ctype;
3217 VM_BUG_ON(PageTransHuge(page));
3218 if (mem_cgroup_disabled())
3221 pc = lookup_page_cgroup(page);
3222 lock_page_cgroup(pc);
3223 if (PageCgroupUsed(pc)) {
3224 memcg = pc->mem_cgroup;
3225 css_get(&memcg->css);
3227 * At migrating an anonymous page, its mapcount goes down
3228 * to 0 and uncharge() will be called. But, even if it's fully
3229 * unmapped, migration may fail and this page has to be
3230 * charged again. We set MIGRATION flag here and delay uncharge
3231 * until end_migration() is called
3233 * Corner Case Thinking
3235 * When the old page was mapped as Anon and it's unmap-and-freed
3236 * while migration was ongoing.
3237 * If unmap finds the old page, uncharge() of it will be delayed
3238 * until end_migration(). If unmap finds a new page, it's
3239 * uncharged when it make mapcount to be 1->0. If unmap code
3240 * finds swap_migration_entry, the new page will not be mapped
3241 * and end_migration() will find it(mapcount==0).
3244 * When the old page was mapped but migraion fails, the kernel
3245 * remaps it. A charge for it is kept by MIGRATION flag even
3246 * if mapcount goes down to 0. We can do remap successfully
3247 * without charging it again.
3250 * The "old" page is under lock_page() until the end of
3251 * migration, so, the old page itself will not be swapped-out.
3252 * If the new page is swapped out before end_migraton, our
3253 * hook to usual swap-out path will catch the event.
3256 SetPageCgroupMigration(pc);
3258 unlock_page_cgroup(pc);
3260 * If the page is not charged at this point,
3267 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3268 css_put(&memcg->css);/* drop extra refcnt */
3269 if (ret || *ptr == NULL) {
3270 if (PageAnon(page)) {
3271 lock_page_cgroup(pc);
3272 ClearPageCgroupMigration(pc);
3273 unlock_page_cgroup(pc);
3275 * The old page may be fully unmapped while we kept it.
3277 mem_cgroup_uncharge_page(page);
3282 * We charge new page before it's used/mapped. So, even if unlock_page()
3283 * is called before end_migration, we can catch all events on this new
3284 * page. In the case new page is migrated but not remapped, new page's
3285 * mapcount will be finally 0 and we call uncharge in end_migration().
3287 pc = lookup_page_cgroup(newpage);
3289 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3290 else if (page_is_file_cache(page))
3291 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3293 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3294 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3298 /* remove redundant charge if migration failed*/
3299 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3300 struct page *oldpage, struct page *newpage, bool migration_ok)
3302 struct page *used, *unused;
3303 struct page_cgroup *pc;
3307 /* blocks rmdir() */
3308 cgroup_exclude_rmdir(&memcg->css);
3309 if (!migration_ok) {
3317 * We disallowed uncharge of pages under migration because mapcount
3318 * of the page goes down to zero, temporarly.
3319 * Clear the flag and check the page should be charged.
3321 pc = lookup_page_cgroup(oldpage);
3322 lock_page_cgroup(pc);
3323 ClearPageCgroupMigration(pc);
3324 unlock_page_cgroup(pc);
3326 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3329 * If a page is a file cache, radix-tree replacement is very atomic
3330 * and we can skip this check. When it was an Anon page, its mapcount
3331 * goes down to 0. But because we added MIGRATION flage, it's not
3332 * uncharged yet. There are several case but page->mapcount check
3333 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3334 * check. (see prepare_charge() also)
3337 mem_cgroup_uncharge_page(used);
3339 * At migration, we may charge account against cgroup which has no
3341 * So, rmdir()->pre_destroy() can be called while we do this charge.
3342 * In that case, we need to call pre_destroy() again. check it here.
3344 cgroup_release_and_wakeup_rmdir(&memcg->css);
3347 #ifdef CONFIG_DEBUG_VM
3348 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3350 struct page_cgroup *pc;
3352 pc = lookup_page_cgroup(page);
3353 if (likely(pc) && PageCgroupUsed(pc))
3358 bool mem_cgroup_bad_page_check(struct page *page)
3360 if (mem_cgroup_disabled())
3363 return lookup_page_cgroup_used(page) != NULL;
3366 void mem_cgroup_print_bad_page(struct page *page)
3368 struct page_cgroup *pc;
3370 pc = lookup_page_cgroup_used(page);
3375 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3376 pc, pc->flags, pc->mem_cgroup);
3378 path = kmalloc(PATH_MAX, GFP_KERNEL);
3381 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3386 printk(KERN_CONT "(%s)\n",
3387 (ret < 0) ? "cannot get the path" : path);
3393 static DEFINE_MUTEX(set_limit_mutex);
3395 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3396 unsigned long long val)
3399 u64 memswlimit, memlimit;
3401 int children = mem_cgroup_count_children(memcg);
3402 u64 curusage, oldusage;
3406 * For keeping hierarchical_reclaim simple, how long we should retry
3407 * is depends on callers. We set our retry-count to be function
3408 * of # of children which we should visit in this loop.
3410 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3412 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3415 while (retry_count) {
3416 if (signal_pending(current)) {
3421 * Rather than hide all in some function, I do this in
3422 * open coded manner. You see what this really does.
3423 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3425 mutex_lock(&set_limit_mutex);
3426 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3427 if (memswlimit < val) {
3429 mutex_unlock(&set_limit_mutex);
3433 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3437 ret = res_counter_set_limit(&memcg->res, val);
3439 if (memswlimit == val)
3440 memcg->memsw_is_minimum = true;
3442 memcg->memsw_is_minimum = false;
3444 mutex_unlock(&set_limit_mutex);
3449 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3450 MEM_CGROUP_RECLAIM_SHRINK);
3451 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3452 /* Usage is reduced ? */
3453 if (curusage >= oldusage)
3456 oldusage = curusage;
3458 if (!ret && enlarge)
3459 memcg_oom_recover(memcg);
3464 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3465 unsigned long long val)
3468 u64 memlimit, memswlimit, oldusage, curusage;
3469 int children = mem_cgroup_count_children(memcg);
3473 /* see mem_cgroup_resize_res_limit */
3474 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3475 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3476 while (retry_count) {
3477 if (signal_pending(current)) {
3482 * Rather than hide all in some function, I do this in
3483 * open coded manner. You see what this really does.
3484 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3486 mutex_lock(&set_limit_mutex);
3487 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3488 if (memlimit > val) {
3490 mutex_unlock(&set_limit_mutex);
3493 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3494 if (memswlimit < val)
3496 ret = res_counter_set_limit(&memcg->memsw, val);
3498 if (memlimit == val)
3499 memcg->memsw_is_minimum = true;
3501 memcg->memsw_is_minimum = false;
3503 mutex_unlock(&set_limit_mutex);
3508 mem_cgroup_reclaim(memcg, GFP_KERNEL,
3509 MEM_CGROUP_RECLAIM_NOSWAP |
3510 MEM_CGROUP_RECLAIM_SHRINK);
3511 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3512 /* Usage is reduced ? */
3513 if (curusage >= oldusage)
3516 oldusage = curusage;
3518 if (!ret && enlarge)
3519 memcg_oom_recover(memcg);
3523 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3525 unsigned long *total_scanned)
3527 unsigned long nr_reclaimed = 0;
3528 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3529 unsigned long reclaimed;
3531 struct mem_cgroup_tree_per_zone *mctz;
3532 unsigned long long excess;
3533 unsigned long nr_scanned;
3538 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3540 * This loop can run a while, specially if mem_cgroup's continuously
3541 * keep exceeding their soft limit and putting the system under
3548 mz = mem_cgroup_largest_soft_limit_node(mctz);
3553 reclaimed = mem_cgroup_soft_reclaim(mz->mem, zone,
3554 gfp_mask, &nr_scanned);
3555 nr_reclaimed += reclaimed;
3556 *total_scanned += nr_scanned;
3557 spin_lock(&mctz->lock);
3560 * If we failed to reclaim anything from this memory cgroup
3561 * it is time to move on to the next cgroup
3567 * Loop until we find yet another one.
3569 * By the time we get the soft_limit lock
3570 * again, someone might have aded the
3571 * group back on the RB tree. Iterate to
3572 * make sure we get a different mem.
3573 * mem_cgroup_largest_soft_limit_node returns
3574 * NULL if no other cgroup is present on
3578 __mem_cgroup_largest_soft_limit_node(mctz);
3580 css_put(&next_mz->mem->css);
3581 else /* next_mz == NULL or other memcg */
3585 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3586 excess = res_counter_soft_limit_excess(&mz->mem->res);
3588 * One school of thought says that we should not add
3589 * back the node to the tree if reclaim returns 0.
3590 * But our reclaim could return 0, simply because due
3591 * to priority we are exposing a smaller subset of
3592 * memory to reclaim from. Consider this as a longer
3595 /* If excess == 0, no tree ops */
3596 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3597 spin_unlock(&mctz->lock);
3598 css_put(&mz->mem->css);
3601 * Could not reclaim anything and there are no more
3602 * mem cgroups to try or we seem to be looping without
3603 * reclaiming anything.
3605 if (!nr_reclaimed &&
3607 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3609 } while (!nr_reclaimed);
3611 css_put(&next_mz->mem->css);
3612 return nr_reclaimed;
3616 * This routine traverse page_cgroup in given list and drop them all.
3617 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3619 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3620 int node, int zid, enum lru_list lru)
3623 struct mem_cgroup_per_zone *mz;
3624 struct page_cgroup *pc, *busy;
3625 unsigned long flags, loop;
3626 struct list_head *list;
3629 zone = &NODE_DATA(node)->node_zones[zid];
3630 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3631 list = &mz->lists[lru];
3633 loop = MEM_CGROUP_ZSTAT(mz, lru);
3634 /* give some margin against EBUSY etc...*/
3641 spin_lock_irqsave(&zone->lru_lock, flags);
3642 if (list_empty(list)) {
3643 spin_unlock_irqrestore(&zone->lru_lock, flags);
3646 pc = list_entry(list->prev, struct page_cgroup, lru);
3648 list_move(&pc->lru, list);
3650 spin_unlock_irqrestore(&zone->lru_lock, flags);
3653 spin_unlock_irqrestore(&zone->lru_lock, flags);
3655 page = lookup_cgroup_page(pc);
3657 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3661 if (ret == -EBUSY || ret == -EINVAL) {
3662 /* found lock contention or "pc" is obsolete. */
3669 if (!ret && !list_empty(list))
3675 * make mem_cgroup's charge to be 0 if there is no task.
3676 * This enables deleting this mem_cgroup.
3678 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3681 int node, zid, shrink;
3682 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3683 struct cgroup *cgrp = memcg->css.cgroup;
3685 css_get(&memcg->css);
3688 /* should free all ? */
3694 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3697 if (signal_pending(current))
3699 /* This is for making all *used* pages to be on LRU. */
3700 lru_add_drain_all();
3701 drain_all_stock_sync(memcg);
3703 mem_cgroup_start_move(memcg);
3704 for_each_node_state(node, N_HIGH_MEMORY) {
3705 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3708 ret = mem_cgroup_force_empty_list(memcg,
3717 mem_cgroup_end_move(memcg);
3718 memcg_oom_recover(memcg);
3719 /* it seems parent cgroup doesn't have enough mem */
3723 /* "ret" should also be checked to ensure all lists are empty. */
3724 } while (memcg->res.usage > 0 || ret);
3726 css_put(&memcg->css);
3730 /* returns EBUSY if there is a task or if we come here twice. */
3731 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3735 /* we call try-to-free pages for make this cgroup empty */
3736 lru_add_drain_all();
3737 /* try to free all pages in this cgroup */
3739 while (nr_retries && memcg->res.usage > 0) {
3742 if (signal_pending(current)) {
3746 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3750 /* maybe some writeback is necessary */
3751 congestion_wait(BLK_RW_ASYNC, HZ/10);
3756 /* try move_account...there may be some *locked* pages. */
3760 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3762 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3766 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3768 return mem_cgroup_from_cont(cont)->use_hierarchy;
3771 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3775 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3776 struct cgroup *parent = cont->parent;
3777 struct mem_cgroup *parent_memcg = NULL;
3780 parent_memcg = mem_cgroup_from_cont(parent);
3784 * If parent's use_hierarchy is set, we can't make any modifications
3785 * in the child subtrees. If it is unset, then the change can
3786 * occur, provided the current cgroup has no children.
3788 * For the root cgroup, parent_mem is NULL, we allow value to be
3789 * set if there are no children.
3791 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3792 (val == 1 || val == 0)) {
3793 if (list_empty(&cont->children))
3794 memcg->use_hierarchy = val;
3805 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3806 enum mem_cgroup_stat_index idx)
3808 struct mem_cgroup *iter;
3811 /* Per-cpu values can be negative, use a signed accumulator */
3812 for_each_mem_cgroup_tree(iter, memcg)
3813 val += mem_cgroup_read_stat(iter, idx);
3815 if (val < 0) /* race ? */
3820 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3824 if (!mem_cgroup_is_root(memcg)) {
3826 return res_counter_read_u64(&memcg->res, RES_USAGE);
3828 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3831 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3832 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3835 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3837 return val << PAGE_SHIFT;
3840 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3842 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3846 type = MEMFILE_TYPE(cft->private);
3847 name = MEMFILE_ATTR(cft->private);
3850 if (name == RES_USAGE)
3851 val = mem_cgroup_usage(memcg, false);
3853 val = res_counter_read_u64(&memcg->res, name);
3856 if (name == RES_USAGE)
3857 val = mem_cgroup_usage(memcg, true);
3859 val = res_counter_read_u64(&memcg->memsw, name);
3868 * The user of this function is...
3871 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3874 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3876 unsigned long long val;
3879 type = MEMFILE_TYPE(cft->private);
3880 name = MEMFILE_ATTR(cft->private);
3883 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3887 /* This function does all necessary parse...reuse it */
3888 ret = res_counter_memparse_write_strategy(buffer, &val);
3892 ret = mem_cgroup_resize_limit(memcg, val);
3894 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3896 case RES_SOFT_LIMIT:
3897 ret = res_counter_memparse_write_strategy(buffer, &val);
3901 * For memsw, soft limits are hard to implement in terms
3902 * of semantics, for now, we support soft limits for
3903 * control without swap
3906 ret = res_counter_set_soft_limit(&memcg->res, val);
3911 ret = -EINVAL; /* should be BUG() ? */
3917 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3918 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3920 struct cgroup *cgroup;
3921 unsigned long long min_limit, min_memsw_limit, tmp;
3923 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3924 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3925 cgroup = memcg->css.cgroup;
3926 if (!memcg->use_hierarchy)
3929 while (cgroup->parent) {
3930 cgroup = cgroup->parent;
3931 memcg = mem_cgroup_from_cont(cgroup);
3932 if (!memcg->use_hierarchy)
3934 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3935 min_limit = min(min_limit, tmp);
3936 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3937 min_memsw_limit = min(min_memsw_limit, tmp);
3940 *mem_limit = min_limit;
3941 *memsw_limit = min_memsw_limit;
3945 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3947 struct mem_cgroup *memcg;
3950 memcg = mem_cgroup_from_cont(cont);
3951 type = MEMFILE_TYPE(event);
3952 name = MEMFILE_ATTR(event);
3956 res_counter_reset_max(&memcg->res);
3958 res_counter_reset_max(&memcg->memsw);
3962 res_counter_reset_failcnt(&memcg->res);
3964 res_counter_reset_failcnt(&memcg->memsw);
3971 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3974 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3978 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3979 struct cftype *cft, u64 val)
3981 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3983 if (val >= (1 << NR_MOVE_TYPE))
3986 * We check this value several times in both in can_attach() and
3987 * attach(), so we need cgroup lock to prevent this value from being
3991 memcg->move_charge_at_immigrate = val;
3997 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3998 struct cftype *cft, u64 val)
4005 /* For read statistics */
4023 struct mcs_total_stat {
4024 s64 stat[NR_MCS_STAT];
4030 } memcg_stat_strings[NR_MCS_STAT] = {
4031 {"cache", "total_cache"},
4032 {"rss", "total_rss"},
4033 {"mapped_file", "total_mapped_file"},
4034 {"pgpgin", "total_pgpgin"},
4035 {"pgpgout", "total_pgpgout"},
4036 {"swap", "total_swap"},
4037 {"pgfault", "total_pgfault"},
4038 {"pgmajfault", "total_pgmajfault"},
4039 {"inactive_anon", "total_inactive_anon"},
4040 {"active_anon", "total_active_anon"},
4041 {"inactive_file", "total_inactive_file"},
4042 {"active_file", "total_active_file"},
4043 {"unevictable", "total_unevictable"}
4048 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4053 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4054 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4055 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4056 s->stat[MCS_RSS] += val * PAGE_SIZE;
4057 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4058 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4059 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4060 s->stat[MCS_PGPGIN] += val;
4061 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4062 s->stat[MCS_PGPGOUT] += val;
4063 if (do_swap_account) {
4064 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4065 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4067 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4068 s->stat[MCS_PGFAULT] += val;
4069 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4070 s->stat[MCS_PGMAJFAULT] += val;
4073 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4074 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4075 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4076 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4077 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4078 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4079 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4080 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4081 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4082 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4086 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4088 struct mem_cgroup *iter;
4090 for_each_mem_cgroup_tree(iter, memcg)
4091 mem_cgroup_get_local_stat(iter, s);
4095 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4098 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4099 unsigned long node_nr;
4100 struct cgroup *cont = m->private;
4101 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4103 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4104 seq_printf(m, "total=%lu", total_nr);
4105 for_each_node_state(nid, N_HIGH_MEMORY) {
4106 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4107 seq_printf(m, " N%d=%lu", nid, node_nr);
4111 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4112 seq_printf(m, "file=%lu", file_nr);
4113 for_each_node_state(nid, N_HIGH_MEMORY) {
4114 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4116 seq_printf(m, " N%d=%lu", nid, node_nr);
4120 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4121 seq_printf(m, "anon=%lu", anon_nr);
4122 for_each_node_state(nid, N_HIGH_MEMORY) {
4123 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4125 seq_printf(m, " N%d=%lu", nid, node_nr);
4129 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4130 seq_printf(m, "unevictable=%lu", unevictable_nr);
4131 for_each_node_state(nid, N_HIGH_MEMORY) {
4132 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4133 BIT(LRU_UNEVICTABLE));
4134 seq_printf(m, " N%d=%lu", nid, node_nr);
4139 #endif /* CONFIG_NUMA */
4141 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4142 struct cgroup_map_cb *cb)
4144 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4145 struct mcs_total_stat mystat;
4148 memset(&mystat, 0, sizeof(mystat));
4149 mem_cgroup_get_local_stat(mem_cont, &mystat);
4152 for (i = 0; i < NR_MCS_STAT; i++) {
4153 if (i == MCS_SWAP && !do_swap_account)
4155 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4158 /* Hierarchical information */
4160 unsigned long long limit, memsw_limit;
4161 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4162 cb->fill(cb, "hierarchical_memory_limit", limit);
4163 if (do_swap_account)
4164 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4167 memset(&mystat, 0, sizeof(mystat));
4168 mem_cgroup_get_total_stat(mem_cont, &mystat);
4169 for (i = 0; i < NR_MCS_STAT; i++) {
4170 if (i == MCS_SWAP && !do_swap_account)
4172 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4175 #ifdef CONFIG_DEBUG_VM
4178 struct mem_cgroup_per_zone *mz;
4179 unsigned long recent_rotated[2] = {0, 0};
4180 unsigned long recent_scanned[2] = {0, 0};
4182 for_each_online_node(nid)
4183 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4184 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4186 recent_rotated[0] +=
4187 mz->reclaim_stat.recent_rotated[0];
4188 recent_rotated[1] +=
4189 mz->reclaim_stat.recent_rotated[1];
4190 recent_scanned[0] +=
4191 mz->reclaim_stat.recent_scanned[0];
4192 recent_scanned[1] +=
4193 mz->reclaim_stat.recent_scanned[1];
4195 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4196 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4197 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4198 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4205 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4207 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4209 return mem_cgroup_swappiness(memcg);
4212 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4215 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4216 struct mem_cgroup *parent;
4221 if (cgrp->parent == NULL)
4224 parent = mem_cgroup_from_cont(cgrp->parent);
4228 /* If under hierarchy, only empty-root can set this value */
4229 if ((parent->use_hierarchy) ||
4230 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4235 memcg->swappiness = val;
4242 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4244 struct mem_cgroup_threshold_ary *t;
4250 t = rcu_dereference(memcg->thresholds.primary);
4252 t = rcu_dereference(memcg->memsw_thresholds.primary);
4257 usage = mem_cgroup_usage(memcg, swap);
4260 * current_threshold points to threshold just below usage.
4261 * If it's not true, a threshold was crossed after last
4262 * call of __mem_cgroup_threshold().
4264 i = t->current_threshold;
4267 * Iterate backward over array of thresholds starting from
4268 * current_threshold and check if a threshold is crossed.
4269 * If none of thresholds below usage is crossed, we read
4270 * only one element of the array here.
4272 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4273 eventfd_signal(t->entries[i].eventfd, 1);
4275 /* i = current_threshold + 1 */
4279 * Iterate forward over array of thresholds starting from
4280 * current_threshold+1 and check if a threshold is crossed.
4281 * If none of thresholds above usage is crossed, we read
4282 * only one element of the array here.
4284 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4285 eventfd_signal(t->entries[i].eventfd, 1);
4287 /* Update current_threshold */
4288 t->current_threshold = i - 1;
4293 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4296 __mem_cgroup_threshold(memcg, false);
4297 if (do_swap_account)
4298 __mem_cgroup_threshold(memcg, true);
4300 memcg = parent_mem_cgroup(memcg);
4304 static int compare_thresholds(const void *a, const void *b)
4306 const struct mem_cgroup_threshold *_a = a;
4307 const struct mem_cgroup_threshold *_b = b;
4309 return _a->threshold - _b->threshold;
4312 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4314 struct mem_cgroup_eventfd_list *ev;
4316 list_for_each_entry(ev, &memcg->oom_notify, list)
4317 eventfd_signal(ev->eventfd, 1);
4321 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4323 struct mem_cgroup *iter;
4325 for_each_mem_cgroup_tree(iter, memcg)
4326 mem_cgroup_oom_notify_cb(iter);
4329 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4330 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4332 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4333 struct mem_cgroup_thresholds *thresholds;
4334 struct mem_cgroup_threshold_ary *new;
4335 int type = MEMFILE_TYPE(cft->private);
4336 u64 threshold, usage;
4339 ret = res_counter_memparse_write_strategy(args, &threshold);
4343 mutex_lock(&memcg->thresholds_lock);
4346 thresholds = &memcg->thresholds;
4347 else if (type == _MEMSWAP)
4348 thresholds = &memcg->memsw_thresholds;
4352 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4354 /* Check if a threshold crossed before adding a new one */
4355 if (thresholds->primary)
4356 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4358 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4360 /* Allocate memory for new array of thresholds */
4361 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4369 /* Copy thresholds (if any) to new array */
4370 if (thresholds->primary) {
4371 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4372 sizeof(struct mem_cgroup_threshold));
4375 /* Add new threshold */
4376 new->entries[size - 1].eventfd = eventfd;
4377 new->entries[size - 1].threshold = threshold;
4379 /* Sort thresholds. Registering of new threshold isn't time-critical */
4380 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4381 compare_thresholds, NULL);
4383 /* Find current threshold */
4384 new->current_threshold = -1;
4385 for (i = 0; i < size; i++) {
4386 if (new->entries[i].threshold < usage) {
4388 * new->current_threshold will not be used until
4389 * rcu_assign_pointer(), so it's safe to increment
4392 ++new->current_threshold;
4396 /* Free old spare buffer and save old primary buffer as spare */
4397 kfree(thresholds->spare);
4398 thresholds->spare = thresholds->primary;
4400 rcu_assign_pointer(thresholds->primary, new);
4402 /* To be sure that nobody uses thresholds */
4406 mutex_unlock(&memcg->thresholds_lock);
4411 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4412 struct cftype *cft, struct eventfd_ctx *eventfd)
4414 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4415 struct mem_cgroup_thresholds *thresholds;
4416 struct mem_cgroup_threshold_ary *new;
4417 int type = MEMFILE_TYPE(cft->private);
4421 mutex_lock(&memcg->thresholds_lock);
4423 thresholds = &memcg->thresholds;
4424 else if (type == _MEMSWAP)
4425 thresholds = &memcg->memsw_thresholds;
4430 * Something went wrong if we trying to unregister a threshold
4431 * if we don't have thresholds
4433 BUG_ON(!thresholds);
4435 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4437 /* Check if a threshold crossed before removing */
4438 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4440 /* Calculate new number of threshold */
4442 for (i = 0; i < thresholds->primary->size; i++) {
4443 if (thresholds->primary->entries[i].eventfd != eventfd)
4447 new = thresholds->spare;
4449 /* Set thresholds array to NULL if we don't have thresholds */
4458 /* Copy thresholds and find current threshold */
4459 new->current_threshold = -1;
4460 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4461 if (thresholds->primary->entries[i].eventfd == eventfd)
4464 new->entries[j] = thresholds->primary->entries[i];
4465 if (new->entries[j].threshold < usage) {
4467 * new->current_threshold will not be used
4468 * until rcu_assign_pointer(), so it's safe to increment
4471 ++new->current_threshold;
4477 /* Swap primary and spare array */
4478 thresholds->spare = thresholds->primary;
4479 rcu_assign_pointer(thresholds->primary, new);
4481 /* To be sure that nobody uses thresholds */
4484 mutex_unlock(&memcg->thresholds_lock);
4487 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4488 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4490 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4491 struct mem_cgroup_eventfd_list *event;
4492 int type = MEMFILE_TYPE(cft->private);
4494 BUG_ON(type != _OOM_TYPE);
4495 event = kmalloc(sizeof(*event), GFP_KERNEL);
4499 spin_lock(&memcg_oom_lock);
4501 event->eventfd = eventfd;
4502 list_add(&event->list, &memcg->oom_notify);
4504 /* already in OOM ? */
4505 if (atomic_read(&memcg->under_oom))
4506 eventfd_signal(eventfd, 1);
4507 spin_unlock(&memcg_oom_lock);
4512 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4513 struct cftype *cft, struct eventfd_ctx *eventfd)
4515 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4516 struct mem_cgroup_eventfd_list *ev, *tmp;
4517 int type = MEMFILE_TYPE(cft->private);
4519 BUG_ON(type != _OOM_TYPE);
4521 spin_lock(&memcg_oom_lock);
4523 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4524 if (ev->eventfd == eventfd) {
4525 list_del(&ev->list);
4530 spin_unlock(&memcg_oom_lock);
4533 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4534 struct cftype *cft, struct cgroup_map_cb *cb)
4536 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4538 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4540 if (atomic_read(&memcg->under_oom))
4541 cb->fill(cb, "under_oom", 1);
4543 cb->fill(cb, "under_oom", 0);
4547 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4548 struct cftype *cft, u64 val)
4550 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4551 struct mem_cgroup *parent;
4553 /* cannot set to root cgroup and only 0 and 1 are allowed */
4554 if (!cgrp->parent || !((val == 0) || (val == 1)))
4557 parent = mem_cgroup_from_cont(cgrp->parent);
4560 /* oom-kill-disable is a flag for subhierarchy. */
4561 if ((parent->use_hierarchy) ||
4562 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4566 memcg->oom_kill_disable = val;
4568 memcg_oom_recover(memcg);
4574 static const struct file_operations mem_control_numa_stat_file_operations = {
4576 .llseek = seq_lseek,
4577 .release = single_release,
4580 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4582 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4584 file->f_op = &mem_control_numa_stat_file_operations;
4585 return single_open(file, mem_control_numa_stat_show, cont);
4587 #endif /* CONFIG_NUMA */
4589 static struct cftype mem_cgroup_files[] = {
4591 .name = "usage_in_bytes",
4592 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4593 .read_u64 = mem_cgroup_read,
4594 .register_event = mem_cgroup_usage_register_event,
4595 .unregister_event = mem_cgroup_usage_unregister_event,
4598 .name = "max_usage_in_bytes",
4599 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4600 .trigger = mem_cgroup_reset,
4601 .read_u64 = mem_cgroup_read,
4604 .name = "limit_in_bytes",
4605 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4606 .write_string = mem_cgroup_write,
4607 .read_u64 = mem_cgroup_read,
4610 .name = "soft_limit_in_bytes",
4611 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4612 .write_string = mem_cgroup_write,
4613 .read_u64 = mem_cgroup_read,
4617 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4618 .trigger = mem_cgroup_reset,
4619 .read_u64 = mem_cgroup_read,
4623 .read_map = mem_control_stat_show,
4626 .name = "force_empty",
4627 .trigger = mem_cgroup_force_empty_write,
4630 .name = "use_hierarchy",
4631 .write_u64 = mem_cgroup_hierarchy_write,
4632 .read_u64 = mem_cgroup_hierarchy_read,
4635 .name = "swappiness",
4636 .read_u64 = mem_cgroup_swappiness_read,
4637 .write_u64 = mem_cgroup_swappiness_write,
4640 .name = "move_charge_at_immigrate",
4641 .read_u64 = mem_cgroup_move_charge_read,
4642 .write_u64 = mem_cgroup_move_charge_write,
4645 .name = "oom_control",
4646 .read_map = mem_cgroup_oom_control_read,
4647 .write_u64 = mem_cgroup_oom_control_write,
4648 .register_event = mem_cgroup_oom_register_event,
4649 .unregister_event = mem_cgroup_oom_unregister_event,
4650 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4654 .name = "numa_stat",
4655 .open = mem_control_numa_stat_open,
4661 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4662 static struct cftype memsw_cgroup_files[] = {
4664 .name = "memsw.usage_in_bytes",
4665 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4666 .read_u64 = mem_cgroup_read,
4667 .register_event = mem_cgroup_usage_register_event,
4668 .unregister_event = mem_cgroup_usage_unregister_event,
4671 .name = "memsw.max_usage_in_bytes",
4672 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4673 .trigger = mem_cgroup_reset,
4674 .read_u64 = mem_cgroup_read,
4677 .name = "memsw.limit_in_bytes",
4678 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4679 .write_string = mem_cgroup_write,
4680 .read_u64 = mem_cgroup_read,
4683 .name = "memsw.failcnt",
4684 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4685 .trigger = mem_cgroup_reset,
4686 .read_u64 = mem_cgroup_read,
4690 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4692 if (!do_swap_account)
4694 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4695 ARRAY_SIZE(memsw_cgroup_files));
4698 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4704 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4706 struct mem_cgroup_per_node *pn;
4707 struct mem_cgroup_per_zone *mz;
4709 int zone, tmp = node;
4711 * This routine is called against possible nodes.
4712 * But it's BUG to call kmalloc() against offline node.
4714 * TODO: this routine can waste much memory for nodes which will
4715 * never be onlined. It's better to use memory hotplug callback
4718 if (!node_state(node, N_NORMAL_MEMORY))
4720 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4724 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4725 mz = &pn->zoneinfo[zone];
4727 INIT_LIST_HEAD(&mz->lists[l]);
4728 mz->usage_in_excess = 0;
4729 mz->on_tree = false;
4732 memcg->info.nodeinfo[node] = pn;
4736 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4738 kfree(memcg->info.nodeinfo[node]);
4741 static struct mem_cgroup *mem_cgroup_alloc(void)
4743 struct mem_cgroup *mem;
4744 int size = sizeof(struct mem_cgroup);
4746 /* Can be very big if MAX_NUMNODES is very big */
4747 if (size < PAGE_SIZE)
4748 mem = kzalloc(size, GFP_KERNEL);
4750 mem = vzalloc(size);
4755 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4758 spin_lock_init(&mem->pcp_counter_lock);
4762 if (size < PAGE_SIZE)
4770 * At destroying mem_cgroup, references from swap_cgroup can remain.
4771 * (scanning all at force_empty is too costly...)
4773 * Instead of clearing all references at force_empty, we remember
4774 * the number of reference from swap_cgroup and free mem_cgroup when
4775 * it goes down to 0.
4777 * Removal of cgroup itself succeeds regardless of refs from swap.
4780 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4784 mem_cgroup_remove_from_trees(memcg);
4785 free_css_id(&mem_cgroup_subsys, &memcg->css);
4787 for_each_node_state(node, N_POSSIBLE)
4788 free_mem_cgroup_per_zone_info(memcg, node);
4790 free_percpu(memcg->stat);
4791 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4797 static void mem_cgroup_get(struct mem_cgroup *memcg)
4799 atomic_inc(&memcg->refcnt);
4802 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4804 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4805 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4806 __mem_cgroup_free(memcg);
4808 mem_cgroup_put(parent);
4812 static void mem_cgroup_put(struct mem_cgroup *memcg)
4814 __mem_cgroup_put(memcg, 1);
4818 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4820 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4822 if (!memcg->res.parent)
4824 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4827 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4828 static void __init enable_swap_cgroup(void)
4830 if (!mem_cgroup_disabled() && really_do_swap_account)
4831 do_swap_account = 1;
4834 static void __init enable_swap_cgroup(void)
4839 static int mem_cgroup_soft_limit_tree_init(void)
4841 struct mem_cgroup_tree_per_node *rtpn;
4842 struct mem_cgroup_tree_per_zone *rtpz;
4843 int tmp, node, zone;
4845 for_each_node_state(node, N_POSSIBLE) {
4847 if (!node_state(node, N_NORMAL_MEMORY))
4849 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4853 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4855 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4856 rtpz = &rtpn->rb_tree_per_zone[zone];
4857 rtpz->rb_root = RB_ROOT;
4858 spin_lock_init(&rtpz->lock);
4864 static struct cgroup_subsys_state * __ref
4865 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4867 struct mem_cgroup *memcg, *parent;
4868 long error = -ENOMEM;
4871 memcg = mem_cgroup_alloc();
4873 return ERR_PTR(error);
4875 for_each_node_state(node, N_POSSIBLE)
4876 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4880 if (cont->parent == NULL) {
4882 enable_swap_cgroup();
4884 root_mem_cgroup = memcg;
4885 if (mem_cgroup_soft_limit_tree_init())
4887 for_each_possible_cpu(cpu) {
4888 struct memcg_stock_pcp *stock =
4889 &per_cpu(memcg_stock, cpu);
4890 INIT_WORK(&stock->work, drain_local_stock);
4892 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4894 parent = mem_cgroup_from_cont(cont->parent);
4895 memcg->use_hierarchy = parent->use_hierarchy;
4896 memcg->oom_kill_disable = parent->oom_kill_disable;
4899 if (parent && parent->use_hierarchy) {
4900 res_counter_init(&memcg->res, &parent->res);
4901 res_counter_init(&memcg->memsw, &parent->memsw);
4903 * We increment refcnt of the parent to ensure that we can
4904 * safely access it on res_counter_charge/uncharge.
4905 * This refcnt will be decremented when freeing this
4906 * mem_cgroup(see mem_cgroup_put).
4908 mem_cgroup_get(parent);
4910 res_counter_init(&memcg->res, NULL);
4911 res_counter_init(&memcg->memsw, NULL);
4913 memcg->last_scanned_node = MAX_NUMNODES;
4914 INIT_LIST_HEAD(&memcg->oom_notify);
4917 memcg->swappiness = mem_cgroup_swappiness(parent);
4918 atomic_set(&memcg->refcnt, 1);
4919 memcg->move_charge_at_immigrate = 0;
4920 mutex_init(&memcg->thresholds_lock);
4923 __mem_cgroup_free(memcg);
4924 root_mem_cgroup = NULL;
4925 return ERR_PTR(error);
4928 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4929 struct cgroup *cont)
4931 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4933 return mem_cgroup_force_empty(memcg, false);
4936 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4937 struct cgroup *cont)
4939 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4941 mem_cgroup_put(memcg);
4944 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4945 struct cgroup *cont)
4949 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4950 ARRAY_SIZE(mem_cgroup_files));
4953 ret = register_memsw_files(cont, ss);
4958 /* Handlers for move charge at task migration. */
4959 #define PRECHARGE_COUNT_AT_ONCE 256
4960 static int mem_cgroup_do_precharge(unsigned long count)
4963 int batch_count = PRECHARGE_COUNT_AT_ONCE;
4964 struct mem_cgroup *memcg = mc.to;
4966 if (mem_cgroup_is_root(memcg)) {
4967 mc.precharge += count;
4968 /* we don't need css_get for root */
4971 /* try to charge at once */
4973 struct res_counter *dummy;
4975 * "memcg" cannot be under rmdir() because we've already checked
4976 * by cgroup_lock_live_cgroup() that it is not removed and we
4977 * are still under the same cgroup_mutex. So we can postpone
4980 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
4982 if (do_swap_account && res_counter_charge(&memcg->memsw,
4983 PAGE_SIZE * count, &dummy)) {
4984 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
4987 mc.precharge += count;
4991 /* fall back to one by one charge */
4993 if (signal_pending(current)) {
4997 if (!batch_count--) {
4998 batch_count = PRECHARGE_COUNT_AT_ONCE;
5001 ret = __mem_cgroup_try_charge(NULL,
5002 GFP_KERNEL, 1, &memcg, false);
5004 /* mem_cgroup_clear_mc() will do uncharge later */
5012 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5013 * @vma: the vma the pte to be checked belongs
5014 * @addr: the address corresponding to the pte to be checked
5015 * @ptent: the pte to be checked
5016 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5019 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5020 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5021 * move charge. if @target is not NULL, the page is stored in target->page
5022 * with extra refcnt got(Callers should handle it).
5023 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5024 * target for charge migration. if @target is not NULL, the entry is stored
5027 * Called with pte lock held.
5034 enum mc_target_type {
5035 MC_TARGET_NONE, /* not used */
5040 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5041 unsigned long addr, pte_t ptent)
5043 struct page *page = vm_normal_page(vma, addr, ptent);
5045 if (!page || !page_mapped(page))
5047 if (PageAnon(page)) {
5048 /* we don't move shared anon */
5049 if (!move_anon() || page_mapcount(page) > 2)
5051 } else if (!move_file())
5052 /* we ignore mapcount for file pages */
5054 if (!get_page_unless_zero(page))
5060 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5061 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5064 struct page *page = NULL;
5065 swp_entry_t ent = pte_to_swp_entry(ptent);
5067 if (!move_anon() || non_swap_entry(ent))
5069 usage_count = mem_cgroup_count_swap_user(ent, &page);
5070 if (usage_count > 1) { /* we don't move shared anon */
5075 if (do_swap_account)
5076 entry->val = ent.val;
5081 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5082 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5084 struct page *page = NULL;
5085 struct inode *inode;
5086 struct address_space *mapping;
5089 if (!vma->vm_file) /* anonymous vma */
5094 inode = vma->vm_file->f_path.dentry->d_inode;
5095 mapping = vma->vm_file->f_mapping;
5096 if (pte_none(ptent))
5097 pgoff = linear_page_index(vma, addr);
5098 else /* pte_file(ptent) is true */
5099 pgoff = pte_to_pgoff(ptent);
5101 /* page is moved even if it's not RSS of this task(page-faulted). */
5102 page = find_get_page(mapping, pgoff);
5105 /* shmem/tmpfs may report page out on swap: account for that too. */
5106 if (radix_tree_exceptional_entry(page)) {
5107 swp_entry_t swap = radix_to_swp_entry(page);
5108 if (do_swap_account)
5110 page = find_get_page(&swapper_space, swap.val);
5116 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5117 unsigned long addr, pte_t ptent, union mc_target *target)
5119 struct page *page = NULL;
5120 struct page_cgroup *pc;
5122 swp_entry_t ent = { .val = 0 };
5124 if (pte_present(ptent))
5125 page = mc_handle_present_pte(vma, addr, ptent);
5126 else if (is_swap_pte(ptent))
5127 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5128 else if (pte_none(ptent) || pte_file(ptent))
5129 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5131 if (!page && !ent.val)
5134 pc = lookup_page_cgroup(page);
5136 * Do only loose check w/o page_cgroup lock.
5137 * mem_cgroup_move_account() checks the pc is valid or not under
5140 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5141 ret = MC_TARGET_PAGE;
5143 target->page = page;
5145 if (!ret || !target)
5148 /* There is a swap entry and a page doesn't exist or isn't charged */
5149 if (ent.val && !ret &&
5150 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5151 ret = MC_TARGET_SWAP;
5158 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5159 unsigned long addr, unsigned long end,
5160 struct mm_walk *walk)
5162 struct vm_area_struct *vma = walk->private;
5166 split_huge_page_pmd(walk->mm, pmd);
5168 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5169 for (; addr != end; pte++, addr += PAGE_SIZE)
5170 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5171 mc.precharge++; /* increment precharge temporarily */
5172 pte_unmap_unlock(pte - 1, ptl);
5178 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5180 unsigned long precharge;
5181 struct vm_area_struct *vma;
5183 down_read(&mm->mmap_sem);
5184 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5185 struct mm_walk mem_cgroup_count_precharge_walk = {
5186 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5190 if (is_vm_hugetlb_page(vma))
5192 walk_page_range(vma->vm_start, vma->vm_end,
5193 &mem_cgroup_count_precharge_walk);
5195 up_read(&mm->mmap_sem);
5197 precharge = mc.precharge;
5203 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5205 unsigned long precharge = mem_cgroup_count_precharge(mm);
5207 VM_BUG_ON(mc.moving_task);
5208 mc.moving_task = current;
5209 return mem_cgroup_do_precharge(precharge);
5212 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5213 static void __mem_cgroup_clear_mc(void)
5215 struct mem_cgroup *from = mc.from;
5216 struct mem_cgroup *to = mc.to;
5218 /* we must uncharge all the leftover precharges from mc.to */
5220 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5224 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5225 * we must uncharge here.
5227 if (mc.moved_charge) {
5228 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5229 mc.moved_charge = 0;
5231 /* we must fixup refcnts and charges */
5232 if (mc.moved_swap) {
5233 /* uncharge swap account from the old cgroup */
5234 if (!mem_cgroup_is_root(mc.from))
5235 res_counter_uncharge(&mc.from->memsw,
5236 PAGE_SIZE * mc.moved_swap);
5237 __mem_cgroup_put(mc.from, mc.moved_swap);
5239 if (!mem_cgroup_is_root(mc.to)) {
5241 * we charged both to->res and to->memsw, so we should
5244 res_counter_uncharge(&mc.to->res,
5245 PAGE_SIZE * mc.moved_swap);
5247 /* we've already done mem_cgroup_get(mc.to) */
5250 memcg_oom_recover(from);
5251 memcg_oom_recover(to);
5252 wake_up_all(&mc.waitq);
5255 static void mem_cgroup_clear_mc(void)
5257 struct mem_cgroup *from = mc.from;
5260 * we must clear moving_task before waking up waiters at the end of
5263 mc.moving_task = NULL;
5264 __mem_cgroup_clear_mc();
5265 spin_lock(&mc.lock);
5268 spin_unlock(&mc.lock);
5269 mem_cgroup_end_move(from);
5272 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5273 struct cgroup *cgroup,
5274 struct task_struct *p)
5277 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5279 if (memcg->move_charge_at_immigrate) {
5280 struct mm_struct *mm;
5281 struct mem_cgroup *from = mem_cgroup_from_task(p);
5283 VM_BUG_ON(from == memcg);
5285 mm = get_task_mm(p);
5288 /* We move charges only when we move a owner of the mm */
5289 if (mm->owner == p) {
5292 VM_BUG_ON(mc.precharge);
5293 VM_BUG_ON(mc.moved_charge);
5294 VM_BUG_ON(mc.moved_swap);
5295 mem_cgroup_start_move(from);
5296 spin_lock(&mc.lock);
5299 spin_unlock(&mc.lock);
5300 /* We set mc.moving_task later */
5302 ret = mem_cgroup_precharge_mc(mm);
5304 mem_cgroup_clear_mc();
5311 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5312 struct cgroup *cgroup,
5313 struct task_struct *p)
5315 mem_cgroup_clear_mc();
5318 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5319 unsigned long addr, unsigned long end,
5320 struct mm_walk *walk)
5323 struct vm_area_struct *vma = walk->private;
5327 split_huge_page_pmd(walk->mm, pmd);
5329 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5330 for (; addr != end; addr += PAGE_SIZE) {
5331 pte_t ptent = *(pte++);
5332 union mc_target target;
5335 struct page_cgroup *pc;
5341 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5343 case MC_TARGET_PAGE:
5345 if (isolate_lru_page(page))
5347 pc = lookup_page_cgroup(page);
5348 if (!mem_cgroup_move_account(page, 1, pc,
5349 mc.from, mc.to, false)) {
5351 /* we uncharge from mc.from later. */
5354 putback_lru_page(page);
5355 put: /* is_target_pte_for_mc() gets the page */
5358 case MC_TARGET_SWAP:
5360 if (!mem_cgroup_move_swap_account(ent,
5361 mc.from, mc.to, false)) {
5363 /* we fixup refcnts and charges later. */
5371 pte_unmap_unlock(pte - 1, ptl);
5376 * We have consumed all precharges we got in can_attach().
5377 * We try charge one by one, but don't do any additional
5378 * charges to mc.to if we have failed in charge once in attach()
5381 ret = mem_cgroup_do_precharge(1);
5389 static void mem_cgroup_move_charge(struct mm_struct *mm)
5391 struct vm_area_struct *vma;
5393 lru_add_drain_all();
5395 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5397 * Someone who are holding the mmap_sem might be waiting in
5398 * waitq. So we cancel all extra charges, wake up all waiters,
5399 * and retry. Because we cancel precharges, we might not be able
5400 * to move enough charges, but moving charge is a best-effort
5401 * feature anyway, so it wouldn't be a big problem.
5403 __mem_cgroup_clear_mc();
5407 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5409 struct mm_walk mem_cgroup_move_charge_walk = {
5410 .pmd_entry = mem_cgroup_move_charge_pte_range,
5414 if (is_vm_hugetlb_page(vma))
5416 ret = walk_page_range(vma->vm_start, vma->vm_end,
5417 &mem_cgroup_move_charge_walk);
5420 * means we have consumed all precharges and failed in
5421 * doing additional charge. Just abandon here.
5425 up_read(&mm->mmap_sem);
5428 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5429 struct cgroup *cont,
5430 struct cgroup *old_cont,
5431 struct task_struct *p)
5433 struct mm_struct *mm = get_task_mm(p);
5437 mem_cgroup_move_charge(mm);
5442 mem_cgroup_clear_mc();
5444 #else /* !CONFIG_MMU */
5445 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5446 struct cgroup *cgroup,
5447 struct task_struct *p)
5451 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5452 struct cgroup *cgroup,
5453 struct task_struct *p)
5456 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5457 struct cgroup *cont,
5458 struct cgroup *old_cont,
5459 struct task_struct *p)
5464 struct cgroup_subsys mem_cgroup_subsys = {
5466 .subsys_id = mem_cgroup_subsys_id,
5467 .create = mem_cgroup_create,
5468 .pre_destroy = mem_cgroup_pre_destroy,
5469 .destroy = mem_cgroup_destroy,
5470 .populate = mem_cgroup_populate,
5471 .can_attach = mem_cgroup_can_attach,
5472 .cancel_attach = mem_cgroup_cancel_attach,
5473 .attach = mem_cgroup_move_task,
5478 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5479 static int __init enable_swap_account(char *s)
5481 /* consider enabled if no parameter or 1 is given */
5482 if (!strcmp(s, "1"))
5483 really_do_swap_account = 1;
5484 else if (!strcmp(s, "0"))
5485 really_do_swap_account = 0;
5488 __setup("swapaccount=", enable_swap_account);