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];
125 * per-zone information in memory controller.
127 struct mem_cgroup_per_zone {
129 * spin_lock to protect the per cgroup LRU
131 struct list_head lists[NR_LRU_LISTS];
132 unsigned long count[NR_LRU_LISTS];
134 struct zone_reclaim_stat reclaim_stat;
135 struct rb_node tree_node; /* RB tree node */
136 unsigned long long usage_in_excess;/* Set to the value by which */
137 /* the soft limit is exceeded*/
139 struct mem_cgroup *mem; /* Back pointer, we cannot */
140 /* use container_of */
142 /* Macro for accessing counter */
143 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
145 struct mem_cgroup_per_node {
146 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
149 struct mem_cgroup_lru_info {
150 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
154 * Cgroups above their limits are maintained in a RB-Tree, independent of
155 * their hierarchy representation
158 struct mem_cgroup_tree_per_zone {
159 struct rb_root rb_root;
163 struct mem_cgroup_tree_per_node {
164 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
167 struct mem_cgroup_tree {
168 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
171 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
173 struct mem_cgroup_threshold {
174 struct eventfd_ctx *eventfd;
179 struct mem_cgroup_threshold_ary {
180 /* An array index points to threshold just below usage. */
181 int current_threshold;
182 /* Size of entries[] */
184 /* Array of thresholds */
185 struct mem_cgroup_threshold entries[0];
188 struct mem_cgroup_thresholds {
189 /* Primary thresholds array */
190 struct mem_cgroup_threshold_ary *primary;
192 * Spare threshold array.
193 * This is needed to make mem_cgroup_unregister_event() "never fail".
194 * It must be able to store at least primary->size - 1 entries.
196 struct mem_cgroup_threshold_ary *spare;
200 struct mem_cgroup_eventfd_list {
201 struct list_head list;
202 struct eventfd_ctx *eventfd;
205 static void mem_cgroup_threshold(struct mem_cgroup *memcg);
206 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg);
209 * The memory controller data structure. The memory controller controls both
210 * page cache and RSS per cgroup. We would eventually like to provide
211 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
212 * to help the administrator determine what knobs to tune.
214 * TODO: Add a water mark for the memory controller. Reclaim will begin when
215 * we hit the water mark. May be even add a low water mark, such that
216 * no reclaim occurs from a cgroup at it's low water mark, this is
217 * a feature that will be implemented much later in the future.
220 struct cgroup_subsys_state css;
222 * the counter to account for memory usage
224 struct res_counter res;
226 * the counter to account for mem+swap usage.
228 struct res_counter memsw;
230 * Per cgroup active and inactive list, similar to the
231 * per zone LRU lists.
233 struct mem_cgroup_lru_info info;
235 * While reclaiming in a hierarchy, we cache the last child we
238 int last_scanned_child;
239 int last_scanned_node;
241 nodemask_t scan_nodes;
242 atomic_t numainfo_events;
243 atomic_t numainfo_updating;
246 * Should the accounting and control be hierarchical, per subtree?
256 /* OOM-Killer disable */
257 int oom_kill_disable;
259 /* set when res.limit == memsw.limit */
260 bool memsw_is_minimum;
262 /* protect arrays of thresholds */
263 struct mutex thresholds_lock;
265 /* thresholds for memory usage. RCU-protected */
266 struct mem_cgroup_thresholds thresholds;
268 /* thresholds for mem+swap usage. RCU-protected */
269 struct mem_cgroup_thresholds memsw_thresholds;
271 /* For oom notifier event fd */
272 struct list_head oom_notify;
275 * Should we move charges of a task when a task is moved into this
276 * mem_cgroup ? And what type of charges should we move ?
278 unsigned long move_charge_at_immigrate;
282 struct mem_cgroup_stat_cpu *stat;
284 * used when a cpu is offlined or other synchronizations
285 * See mem_cgroup_read_stat().
287 struct mem_cgroup_stat_cpu nocpu_base;
288 spinlock_t pcp_counter_lock;
291 /* Stuffs for move charges at task migration. */
293 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
294 * left-shifted bitmap of these types.
297 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
298 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
302 /* "mc" and its members are protected by cgroup_mutex */
303 static struct move_charge_struct {
304 spinlock_t lock; /* for from, to */
305 struct mem_cgroup *from;
306 struct mem_cgroup *to;
307 unsigned long precharge;
308 unsigned long moved_charge;
309 unsigned long moved_swap;
310 struct task_struct *moving_task; /* a task moving charges */
311 wait_queue_head_t waitq; /* a waitq for other context */
313 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
314 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
317 static bool move_anon(void)
319 return test_bit(MOVE_CHARGE_TYPE_ANON,
320 &mc.to->move_charge_at_immigrate);
323 static bool move_file(void)
325 return test_bit(MOVE_CHARGE_TYPE_FILE,
326 &mc.to->move_charge_at_immigrate);
330 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
331 * limit reclaim to prevent infinite loops, if they ever occur.
333 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
334 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
337 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
338 MEM_CGROUP_CHARGE_TYPE_MAPPED,
339 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
340 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
341 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
342 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
346 /* for encoding cft->private value on file */
349 #define _OOM_TYPE (2)
350 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
351 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
352 #define MEMFILE_ATTR(val) ((val) & 0xffff)
353 /* Used for OOM nofiier */
354 #define OOM_CONTROL (0)
357 * Reclaim flags for mem_cgroup_hierarchical_reclaim
359 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
360 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
361 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
362 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
363 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
364 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
366 static void mem_cgroup_get(struct mem_cgroup *memcg);
367 static void mem_cgroup_put(struct mem_cgroup *memcg);
368 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg);
369 static void drain_all_stock_async(struct mem_cgroup *memcg);
371 static struct mem_cgroup_per_zone *
372 mem_cgroup_zoneinfo(struct mem_cgroup *memcg, int nid, int zid)
374 return &memcg->info.nodeinfo[nid]->zoneinfo[zid];
377 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *memcg)
382 static struct mem_cgroup_per_zone *
383 page_cgroup_zoneinfo(struct mem_cgroup *memcg, struct page *page)
385 int nid = page_to_nid(page);
386 int zid = page_zonenum(page);
388 return mem_cgroup_zoneinfo(memcg, nid, zid);
391 static struct mem_cgroup_tree_per_zone *
392 soft_limit_tree_node_zone(int nid, int zid)
394 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
397 static struct mem_cgroup_tree_per_zone *
398 soft_limit_tree_from_page(struct page *page)
400 int nid = page_to_nid(page);
401 int zid = page_zonenum(page);
403 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
407 __mem_cgroup_insert_exceeded(struct mem_cgroup *memcg,
408 struct mem_cgroup_per_zone *mz,
409 struct mem_cgroup_tree_per_zone *mctz,
410 unsigned long long new_usage_in_excess)
412 struct rb_node **p = &mctz->rb_root.rb_node;
413 struct rb_node *parent = NULL;
414 struct mem_cgroup_per_zone *mz_node;
419 mz->usage_in_excess = new_usage_in_excess;
420 if (!mz->usage_in_excess)
424 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
426 if (mz->usage_in_excess < mz_node->usage_in_excess)
429 * We can't avoid mem cgroups that are over their soft
430 * limit by the same amount
432 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
435 rb_link_node(&mz->tree_node, parent, p);
436 rb_insert_color(&mz->tree_node, &mctz->rb_root);
441 __mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
442 struct mem_cgroup_per_zone *mz,
443 struct mem_cgroup_tree_per_zone *mctz)
447 rb_erase(&mz->tree_node, &mctz->rb_root);
452 mem_cgroup_remove_exceeded(struct mem_cgroup *memcg,
453 struct mem_cgroup_per_zone *mz,
454 struct mem_cgroup_tree_per_zone *mctz)
456 spin_lock(&mctz->lock);
457 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
458 spin_unlock(&mctz->lock);
462 static void mem_cgroup_update_tree(struct mem_cgroup *memcg, struct page *page)
464 unsigned long long excess;
465 struct mem_cgroup_per_zone *mz;
466 struct mem_cgroup_tree_per_zone *mctz;
467 int nid = page_to_nid(page);
468 int zid = page_zonenum(page);
469 mctz = soft_limit_tree_from_page(page);
472 * Necessary to update all ancestors when hierarchy is used.
473 * because their event counter is not touched.
475 for (; memcg; memcg = parent_mem_cgroup(memcg)) {
476 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
477 excess = res_counter_soft_limit_excess(&memcg->res);
479 * We have to update the tree if mz is on RB-tree or
480 * mem is over its softlimit.
482 if (excess || mz->on_tree) {
483 spin_lock(&mctz->lock);
484 /* if on-tree, remove it */
486 __mem_cgroup_remove_exceeded(memcg, mz, mctz);
488 * Insert again. mz->usage_in_excess will be updated.
489 * If excess is 0, no tree ops.
491 __mem_cgroup_insert_exceeded(memcg, mz, mctz, excess);
492 spin_unlock(&mctz->lock);
497 static void mem_cgroup_remove_from_trees(struct mem_cgroup *memcg)
500 struct mem_cgroup_per_zone *mz;
501 struct mem_cgroup_tree_per_zone *mctz;
503 for_each_node_state(node, N_POSSIBLE) {
504 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
505 mz = mem_cgroup_zoneinfo(memcg, node, zone);
506 mctz = soft_limit_tree_node_zone(node, zone);
507 mem_cgroup_remove_exceeded(memcg, mz, mctz);
512 static struct mem_cgroup_per_zone *
513 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
515 struct rb_node *rightmost = NULL;
516 struct mem_cgroup_per_zone *mz;
520 rightmost = rb_last(&mctz->rb_root);
522 goto done; /* Nothing to reclaim from */
524 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
526 * Remove the node now but someone else can add it back,
527 * we will to add it back at the end of reclaim to its correct
528 * position in the tree.
530 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
531 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
532 !css_tryget(&mz->mem->css))
538 static struct mem_cgroup_per_zone *
539 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
541 struct mem_cgroup_per_zone *mz;
543 spin_lock(&mctz->lock);
544 mz = __mem_cgroup_largest_soft_limit_node(mctz);
545 spin_unlock(&mctz->lock);
550 * Implementation Note: reading percpu statistics for memcg.
552 * Both of vmstat[] and percpu_counter has threshold and do periodic
553 * synchronization to implement "quick" read. There are trade-off between
554 * reading cost and precision of value. Then, we may have a chance to implement
555 * a periodic synchronizion of counter in memcg's counter.
557 * But this _read() function is used for user interface now. The user accounts
558 * memory usage by memory cgroup and he _always_ requires exact value because
559 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
560 * have to visit all online cpus and make sum. So, for now, unnecessary
561 * synchronization is not implemented. (just implemented for cpu hotplug)
563 * If there are kernel internal actions which can make use of some not-exact
564 * value, and reading all cpu value can be performance bottleneck in some
565 * common workload, threashold and synchonization as vmstat[] should be
568 static long mem_cgroup_read_stat(struct mem_cgroup *memcg,
569 enum mem_cgroup_stat_index idx)
575 for_each_online_cpu(cpu)
576 val += per_cpu(memcg->stat->count[idx], cpu);
577 #ifdef CONFIG_HOTPLUG_CPU
578 spin_lock(&memcg->pcp_counter_lock);
579 val += memcg->nocpu_base.count[idx];
580 spin_unlock(&memcg->pcp_counter_lock);
586 static void mem_cgroup_swap_statistics(struct mem_cgroup *memcg,
589 int val = (charge) ? 1 : -1;
590 this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
593 void mem_cgroup_pgfault(struct mem_cgroup *memcg, int val)
595 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGFAULT], val);
598 void mem_cgroup_pgmajfault(struct mem_cgroup *memcg, int val)
600 this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_PGMAJFAULT], val);
603 static unsigned long mem_cgroup_read_events(struct mem_cgroup *memcg,
604 enum mem_cgroup_events_index idx)
606 unsigned long val = 0;
609 for_each_online_cpu(cpu)
610 val += per_cpu(memcg->stat->events[idx], cpu);
611 #ifdef CONFIG_HOTPLUG_CPU
612 spin_lock(&memcg->pcp_counter_lock);
613 val += memcg->nocpu_base.events[idx];
614 spin_unlock(&memcg->pcp_counter_lock);
619 static void mem_cgroup_charge_statistics(struct mem_cgroup *memcg,
620 bool file, int nr_pages)
625 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_CACHE],
628 __this_cpu_add(memcg->stat->count[MEM_CGROUP_STAT_RSS],
631 /* pagein of a big page is an event. So, ignore page size */
633 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGIN]);
635 __this_cpu_inc(memcg->stat->events[MEM_CGROUP_EVENTS_PGPGOUT]);
636 nr_pages = -nr_pages; /* for event */
639 __this_cpu_add(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT], nr_pages);
645 mem_cgroup_zone_nr_lru_pages(struct mem_cgroup *memcg, int nid, int zid,
646 unsigned int lru_mask)
648 struct mem_cgroup_per_zone *mz;
650 unsigned long ret = 0;
652 mz = mem_cgroup_zoneinfo(memcg, nid, zid);
655 if (BIT(l) & lru_mask)
656 ret += MEM_CGROUP_ZSTAT(mz, l);
662 mem_cgroup_node_nr_lru_pages(struct mem_cgroup *memcg,
663 int nid, unsigned int lru_mask)
668 for (zid = 0; zid < MAX_NR_ZONES; zid++)
669 total += mem_cgroup_zone_nr_lru_pages(memcg,
675 static unsigned long mem_cgroup_nr_lru_pages(struct mem_cgroup *memcg,
676 unsigned int lru_mask)
681 for_each_node_state(nid, N_HIGH_MEMORY)
682 total += mem_cgroup_node_nr_lru_pages(memcg, nid, lru_mask);
686 static bool __memcg_event_check(struct mem_cgroup *memcg, int target)
688 unsigned long val, next;
690 val = this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
691 next = this_cpu_read(memcg->stat->targets[target]);
692 /* from time_after() in jiffies.h */
693 return ((long)next - (long)val < 0);
696 static void __mem_cgroup_target_update(struct mem_cgroup *memcg, int target)
698 unsigned long val, next;
700 val = this_cpu_read(memcg->stat->events[MEM_CGROUP_EVENTS_COUNT]);
703 case MEM_CGROUP_TARGET_THRESH:
704 next = val + THRESHOLDS_EVENTS_TARGET;
706 case MEM_CGROUP_TARGET_SOFTLIMIT:
707 next = val + SOFTLIMIT_EVENTS_TARGET;
709 case MEM_CGROUP_TARGET_NUMAINFO:
710 next = val + NUMAINFO_EVENTS_TARGET;
716 this_cpu_write(memcg->stat->targets[target], next);
720 * Check events in order.
723 static void memcg_check_events(struct mem_cgroup *memcg, struct page *page)
725 /* threshold event is triggered in finer grain than soft limit */
726 if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
727 mem_cgroup_threshold(memcg);
728 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
729 if (unlikely(__memcg_event_check(memcg,
730 MEM_CGROUP_TARGET_SOFTLIMIT))) {
731 mem_cgroup_update_tree(memcg, page);
732 __mem_cgroup_target_update(memcg,
733 MEM_CGROUP_TARGET_SOFTLIMIT);
736 if (unlikely(__memcg_event_check(memcg,
737 MEM_CGROUP_TARGET_NUMAINFO))) {
738 atomic_inc(&memcg->numainfo_events);
739 __mem_cgroup_target_update(memcg,
740 MEM_CGROUP_TARGET_NUMAINFO);
746 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
748 return container_of(cgroup_subsys_state(cont,
749 mem_cgroup_subsys_id), struct mem_cgroup,
753 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
756 * mm_update_next_owner() may clear mm->owner to NULL
757 * if it races with swapoff, page migration, etc.
758 * So this can be called with p == NULL.
763 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
764 struct mem_cgroup, css);
767 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
769 struct mem_cgroup *memcg = NULL;
774 * Because we have no locks, mm->owner's may be being moved to other
775 * cgroup. We use css_tryget() here even if this looks
776 * pessimistic (rather than adding locks here).
780 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
781 if (unlikely(!memcg))
783 } while (!css_tryget(&memcg->css));
788 /* The caller has to guarantee "mem" exists before calling this */
789 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *memcg)
791 struct cgroup_subsys_state *css;
794 if (!memcg) /* ROOT cgroup has the smallest ID */
795 return root_mem_cgroup; /*css_put/get against root is ignored*/
796 if (!memcg->use_hierarchy) {
797 if (css_tryget(&memcg->css))
803 * searching a memory cgroup which has the smallest ID under given
804 * ROOT cgroup. (ID >= 1)
806 css = css_get_next(&mem_cgroup_subsys, 1, &memcg->css, &found);
807 if (css && css_tryget(css))
808 memcg = container_of(css, struct mem_cgroup, css);
815 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
816 struct mem_cgroup *root,
819 int nextid = css_id(&iter->css) + 1;
822 struct cgroup_subsys_state *css;
824 hierarchy_used = iter->use_hierarchy;
827 /* If no ROOT, walk all, ignore hierarchy */
828 if (!cond || (root && !hierarchy_used))
832 root = root_mem_cgroup;
838 css = css_get_next(&mem_cgroup_subsys, nextid,
840 if (css && css_tryget(css))
841 iter = container_of(css, struct mem_cgroup, css);
843 /* If css is NULL, no more cgroups will be found */
845 } while (css && !iter);
850 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
851 * be careful that "break" loop is not allowed. We have reference count.
852 * Instead of that modify "cond" to be false and "continue" to exit the loop.
854 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
855 for (iter = mem_cgroup_start_loop(root);\
857 iter = mem_cgroup_get_next(iter, root, cond))
859 #define for_each_mem_cgroup_tree(iter, root) \
860 for_each_mem_cgroup_tree_cond(iter, root, true)
862 #define for_each_mem_cgroup_all(iter) \
863 for_each_mem_cgroup_tree_cond(iter, NULL, true)
866 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
868 return (memcg == root_mem_cgroup);
871 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
873 struct mem_cgroup *memcg;
879 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
880 if (unlikely(!memcg))
885 mem_cgroup_pgmajfault(memcg, 1);
888 mem_cgroup_pgfault(memcg, 1);
896 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
899 * Following LRU functions are allowed to be used without PCG_LOCK.
900 * Operations are called by routine of global LRU independently from memcg.
901 * What we have to take care of here is validness of pc->mem_cgroup.
903 * Changes to pc->mem_cgroup happens when
906 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
907 * It is added to LRU before charge.
908 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
909 * When moving account, the page is not on LRU. It's isolated.
912 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
914 struct page_cgroup *pc;
915 struct mem_cgroup_per_zone *mz;
917 if (mem_cgroup_disabled())
919 pc = lookup_page_cgroup(page);
920 /* can happen while we handle swapcache. */
921 if (!TestClearPageCgroupAcctLRU(pc))
923 VM_BUG_ON(!pc->mem_cgroup);
925 * We don't check PCG_USED bit. It's cleared when the "page" is finally
926 * removed from global LRU.
928 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
929 /* huge page split is done under lru_lock. so, we have no races. */
930 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
931 if (mem_cgroup_is_root(pc->mem_cgroup))
933 VM_BUG_ON(list_empty(&pc->lru));
934 list_del_init(&pc->lru);
937 void mem_cgroup_del_lru(struct page *page)
939 mem_cgroup_del_lru_list(page, page_lru(page));
943 * Writeback is about to end against a page which has been marked for immediate
944 * reclaim. If it still appears to be reclaimable, move it to the tail of the
947 void mem_cgroup_rotate_reclaimable_page(struct page *page)
949 struct mem_cgroup_per_zone *mz;
950 struct page_cgroup *pc;
951 enum lru_list lru = page_lru(page);
953 if (mem_cgroup_disabled())
956 pc = lookup_page_cgroup(page);
957 /* unused or root page is not rotated. */
958 if (!PageCgroupUsed(pc))
960 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
962 if (mem_cgroup_is_root(pc->mem_cgroup))
964 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
965 list_move_tail(&pc->lru, &mz->lists[lru]);
968 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
970 struct mem_cgroup_per_zone *mz;
971 struct page_cgroup *pc;
973 if (mem_cgroup_disabled())
976 pc = lookup_page_cgroup(page);
977 /* unused or root page is not rotated. */
978 if (!PageCgroupUsed(pc))
980 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
982 if (mem_cgroup_is_root(pc->mem_cgroup))
984 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
985 list_move(&pc->lru, &mz->lists[lru]);
988 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
990 struct page_cgroup *pc;
991 struct mem_cgroup_per_zone *mz;
993 if (mem_cgroup_disabled())
995 pc = lookup_page_cgroup(page);
996 VM_BUG_ON(PageCgroupAcctLRU(pc));
997 if (!PageCgroupUsed(pc))
999 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1001 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1002 /* huge page split is done under lru_lock. so, we have no races. */
1003 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1004 SetPageCgroupAcctLRU(pc);
1005 if (mem_cgroup_is_root(pc->mem_cgroup))
1007 list_add(&pc->lru, &mz->lists[lru]);
1011 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1012 * while it's linked to lru because the page may be reused after it's fully
1013 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1014 * It's done under lock_page and expected that zone->lru_lock isnever held.
1016 static void mem_cgroup_lru_del_before_commit(struct page *page)
1018 unsigned long flags;
1019 struct zone *zone = page_zone(page);
1020 struct page_cgroup *pc = lookup_page_cgroup(page);
1023 * Doing this check without taking ->lru_lock seems wrong but this
1024 * is safe. Because if page_cgroup's USED bit is unset, the page
1025 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1026 * set, the commit after this will fail, anyway.
1027 * This all charge/uncharge is done under some mutual execustion.
1028 * So, we don't need to taking care of changes in USED bit.
1030 if (likely(!PageLRU(page)))
1033 spin_lock_irqsave(&zone->lru_lock, flags);
1035 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1036 * is guarded by lock_page() because the page is SwapCache.
1038 if (!PageCgroupUsed(pc))
1039 mem_cgroup_del_lru_list(page, page_lru(page));
1040 spin_unlock_irqrestore(&zone->lru_lock, flags);
1043 static void mem_cgroup_lru_add_after_commit(struct page *page)
1045 unsigned long flags;
1046 struct zone *zone = page_zone(page);
1047 struct page_cgroup *pc = lookup_page_cgroup(page);
1049 /* taking care of that the page is added to LRU while we commit it */
1050 if (likely(!PageLRU(page)))
1052 spin_lock_irqsave(&zone->lru_lock, flags);
1053 /* link when the page is linked to LRU but page_cgroup isn't */
1054 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1055 mem_cgroup_add_lru_list(page, page_lru(page));
1056 spin_unlock_irqrestore(&zone->lru_lock, flags);
1060 void mem_cgroup_move_lists(struct page *page,
1061 enum lru_list from, enum lru_list to)
1063 if (mem_cgroup_disabled())
1065 mem_cgroup_del_lru_list(page, from);
1066 mem_cgroup_add_lru_list(page, to);
1070 * Checks whether given mem is same or in the root_mem's
1073 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1074 struct mem_cgroup *memcg)
1076 if (root_memcg != memcg) {
1077 return (root_memcg->use_hierarchy &&
1078 css_is_ancestor(&memcg->css, &root_memcg->css));
1084 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1087 struct mem_cgroup *curr = NULL;
1088 struct task_struct *p;
1090 p = find_lock_task_mm(task);
1093 curr = try_get_mem_cgroup_from_mm(p->mm);
1098 * We should check use_hierarchy of "memcg" not "curr". Because checking
1099 * use_hierarchy of "curr" here make this function true if hierarchy is
1100 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1101 * hierarchy(even if use_hierarchy is disabled in "memcg").
1103 ret = mem_cgroup_same_or_subtree(memcg, curr);
1104 css_put(&curr->css);
1108 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
1110 unsigned long active;
1111 unsigned long inactive;
1113 unsigned long inactive_ratio;
1115 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
1116 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
1118 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1120 inactive_ratio = int_sqrt(10 * gb);
1124 if (present_pages) {
1125 present_pages[0] = inactive;
1126 present_pages[1] = active;
1129 return inactive_ratio;
1132 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
1134 unsigned long active;
1135 unsigned long inactive;
1136 unsigned long present_pages[2];
1137 unsigned long inactive_ratio;
1139 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
1141 inactive = present_pages[0];
1142 active = present_pages[1];
1144 if (inactive * inactive_ratio < active)
1150 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1152 unsigned long active;
1153 unsigned long inactive;
1155 inactive = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
1156 active = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
1158 return (active > inactive);
1161 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1164 int nid = zone_to_nid(zone);
1165 int zid = zone_idx(zone);
1166 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1168 return &mz->reclaim_stat;
1171 struct zone_reclaim_stat *
1172 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1174 struct page_cgroup *pc;
1175 struct mem_cgroup_per_zone *mz;
1177 if (mem_cgroup_disabled())
1180 pc = lookup_page_cgroup(page);
1181 if (!PageCgroupUsed(pc))
1183 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1185 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1186 return &mz->reclaim_stat;
1189 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1190 struct list_head *dst,
1191 unsigned long *scanned, int order,
1192 isolate_mode_t mode,
1194 struct mem_cgroup *mem_cont,
1195 int active, int file)
1197 unsigned long nr_taken = 0;
1201 struct list_head *src;
1202 struct page_cgroup *pc, *tmp;
1203 int nid = zone_to_nid(z);
1204 int zid = zone_idx(z);
1205 struct mem_cgroup_per_zone *mz;
1206 int lru = LRU_FILE * file + active;
1210 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1211 src = &mz->lists[lru];
1214 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1215 if (scan >= nr_to_scan)
1218 if (unlikely(!PageCgroupUsed(pc)))
1221 page = lookup_cgroup_page(pc);
1223 if (unlikely(!PageLRU(page)))
1227 ret = __isolate_lru_page(page, mode, file);
1230 list_move(&page->lru, dst);
1231 mem_cgroup_del_lru(page);
1232 nr_taken += hpage_nr_pages(page);
1235 /* we don't affect global LRU but rotate in our LRU */
1236 mem_cgroup_rotate_lru_list(page, page_lru(page));
1245 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1251 #define mem_cgroup_from_res_counter(counter, member) \
1252 container_of(counter, struct mem_cgroup, member)
1255 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1256 * @mem: the memory cgroup
1258 * Returns the maximum amount of memory @mem can be charged with, in
1261 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1263 unsigned long long margin;
1265 margin = res_counter_margin(&memcg->res);
1266 if (do_swap_account)
1267 margin = min(margin, res_counter_margin(&memcg->memsw));
1268 return margin >> PAGE_SHIFT;
1271 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1273 struct cgroup *cgrp = memcg->css.cgroup;
1276 if (cgrp->parent == NULL)
1277 return vm_swappiness;
1279 return memcg->swappiness;
1282 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1287 spin_lock(&memcg->pcp_counter_lock);
1288 for_each_online_cpu(cpu)
1289 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1290 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1291 spin_unlock(&memcg->pcp_counter_lock);
1297 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1304 spin_lock(&memcg->pcp_counter_lock);
1305 for_each_online_cpu(cpu)
1306 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1307 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1308 spin_unlock(&memcg->pcp_counter_lock);
1312 * 2 routines for checking "mem" is under move_account() or not.
1314 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1315 * for avoiding race in accounting. If true,
1316 * pc->mem_cgroup may be overwritten.
1318 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1319 * under hierarchy of moving cgroups. This is for
1320 * waiting at hith-memory prressure caused by "move".
1323 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1325 VM_BUG_ON(!rcu_read_lock_held());
1326 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1329 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1331 struct mem_cgroup *from;
1332 struct mem_cgroup *to;
1335 * Unlike task_move routines, we access mc.to, mc.from not under
1336 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1338 spin_lock(&mc.lock);
1344 ret = mem_cgroup_same_or_subtree(memcg, from)
1345 || mem_cgroup_same_or_subtree(memcg, to);
1347 spin_unlock(&mc.lock);
1351 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1353 if (mc.moving_task && current != mc.moving_task) {
1354 if (mem_cgroup_under_move(memcg)) {
1356 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1357 /* moving charge context might have finished. */
1360 finish_wait(&mc.waitq, &wait);
1368 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1369 * @memcg: The memory cgroup that went over limit
1370 * @p: Task that is going to be killed
1372 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1375 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1377 struct cgroup *task_cgrp;
1378 struct cgroup *mem_cgrp;
1380 * Need a buffer in BSS, can't rely on allocations. The code relies
1381 * on the assumption that OOM is serialized for memory controller.
1382 * If this assumption is broken, revisit this code.
1384 static char memcg_name[PATH_MAX];
1393 mem_cgrp = memcg->css.cgroup;
1394 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1396 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1399 * Unfortunately, we are unable to convert to a useful name
1400 * But we'll still print out the usage information
1407 printk(KERN_INFO "Task in %s killed", memcg_name);
1410 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1418 * Continues from above, so we don't need an KERN_ level
1420 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1423 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1424 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1425 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1426 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1427 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1429 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1430 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1431 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1435 * This function returns the number of memcg under hierarchy tree. Returns
1436 * 1(self count) if no children.
1438 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1441 struct mem_cgroup *iter;
1443 for_each_mem_cgroup_tree(iter, memcg)
1449 * Return the memory (and swap, if configured) limit for a memcg.
1451 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1456 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1457 limit += total_swap_pages << PAGE_SHIFT;
1459 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1461 * If memsw is finite and limits the amount of swap space available
1462 * to this memcg, return that limit.
1464 return min(limit, memsw);
1468 * Visit the first child (need not be the first child as per the ordering
1469 * of the cgroup list, since we track last_scanned_child) of @mem and use
1470 * that to reclaim free pages from.
1472 static struct mem_cgroup *
1473 mem_cgroup_select_victim(struct mem_cgroup *root_memcg)
1475 struct mem_cgroup *ret = NULL;
1476 struct cgroup_subsys_state *css;
1479 if (!root_memcg->use_hierarchy) {
1480 css_get(&root_memcg->css);
1486 nextid = root_memcg->last_scanned_child + 1;
1487 css = css_get_next(&mem_cgroup_subsys, nextid, &root_memcg->css,
1489 if (css && css_tryget(css))
1490 ret = container_of(css, struct mem_cgroup, css);
1493 /* Updates scanning parameter */
1495 /* this means start scan from ID:1 */
1496 root_memcg->last_scanned_child = 0;
1498 root_memcg->last_scanned_child = found;
1505 * test_mem_cgroup_node_reclaimable
1506 * @mem: the target memcg
1507 * @nid: the node ID to be checked.
1508 * @noswap : specify true here if the user wants flle only information.
1510 * This function returns whether the specified memcg contains any
1511 * reclaimable pages on a node. Returns true if there are any reclaimable
1512 * pages in the node.
1514 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1515 int nid, bool noswap)
1517 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1519 if (noswap || !total_swap_pages)
1521 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1526 #if MAX_NUMNODES > 1
1529 * Always updating the nodemask is not very good - even if we have an empty
1530 * list or the wrong list here, we can start from some node and traverse all
1531 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1534 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1538 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1539 * pagein/pageout changes since the last update.
1541 if (!atomic_read(&memcg->numainfo_events))
1543 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1546 /* make a nodemask where this memcg uses memory from */
1547 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1549 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1551 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1552 node_clear(nid, memcg->scan_nodes);
1555 atomic_set(&memcg->numainfo_events, 0);
1556 atomic_set(&memcg->numainfo_updating, 0);
1560 * Selecting a node where we start reclaim from. Because what we need is just
1561 * reducing usage counter, start from anywhere is O,K. Considering
1562 * memory reclaim from current node, there are pros. and cons.
1564 * Freeing memory from current node means freeing memory from a node which
1565 * we'll use or we've used. So, it may make LRU bad. And if several threads
1566 * hit limits, it will see a contention on a node. But freeing from remote
1567 * node means more costs for memory reclaim because of memory latency.
1569 * Now, we use round-robin. Better algorithm is welcomed.
1571 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1575 mem_cgroup_may_update_nodemask(memcg);
1576 node = memcg->last_scanned_node;
1578 node = next_node(node, memcg->scan_nodes);
1579 if (node == MAX_NUMNODES)
1580 node = first_node(memcg->scan_nodes);
1582 * We call this when we hit limit, not when pages are added to LRU.
1583 * No LRU may hold pages because all pages are UNEVICTABLE or
1584 * memcg is too small and all pages are not on LRU. In that case,
1585 * we use curret node.
1587 if (unlikely(node == MAX_NUMNODES))
1588 node = numa_node_id();
1590 memcg->last_scanned_node = node;
1595 * Check all nodes whether it contains reclaimable pages or not.
1596 * For quick scan, we make use of scan_nodes. This will allow us to skip
1597 * unused nodes. But scan_nodes is lazily updated and may not cotain
1598 * enough new information. We need to do double check.
1600 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1605 * quick check...making use of scan_node.
1606 * We can skip unused nodes.
1608 if (!nodes_empty(memcg->scan_nodes)) {
1609 for (nid = first_node(memcg->scan_nodes);
1611 nid = next_node(nid, memcg->scan_nodes)) {
1613 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1618 * Check rest of nodes.
1620 for_each_node_state(nid, N_HIGH_MEMORY) {
1621 if (node_isset(nid, memcg->scan_nodes))
1623 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1630 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1635 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1637 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1642 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1643 * we reclaimed from, so that we don't end up penalizing one child extensively
1644 * based on its position in the children list.
1646 * root_memcg is the original ancestor that we've been reclaim from.
1648 * We give up and return to the caller when we visit root_memcg twice.
1649 * (other groups can be removed while we're walking....)
1651 * If shrink==true, for avoiding to free too much, this returns immedieately.
1653 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_memcg,
1656 unsigned long reclaim_options,
1657 unsigned long *total_scanned)
1659 struct mem_cgroup *victim;
1662 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1663 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1664 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1665 unsigned long excess;
1666 unsigned long nr_scanned;
1668 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1670 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1671 if (!check_soft && !shrink && root_memcg->memsw_is_minimum)
1675 victim = mem_cgroup_select_victim(root_memcg);
1676 if (victim == root_memcg) {
1679 * We are not draining per cpu cached charges during
1680 * soft limit reclaim because global reclaim doesn't
1681 * care about charges. It tries to free some memory and
1682 * charges will not give any.
1684 if (!check_soft && loop >= 1)
1685 drain_all_stock_async(root_memcg);
1688 * If we have not been able to reclaim
1689 * anything, it might because there are
1690 * no reclaimable pages under this hierarchy
1692 if (!check_soft || !total) {
1693 css_put(&victim->css);
1697 * We want to do more targeted reclaim.
1698 * excess >> 2 is not to excessive so as to
1699 * reclaim too much, nor too less that we keep
1700 * coming back to reclaim from this cgroup
1702 if (total >= (excess >> 2) ||
1703 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1704 css_put(&victim->css);
1709 if (!mem_cgroup_reclaimable(victim, noswap)) {
1710 /* this cgroup's local usage == 0 */
1711 css_put(&victim->css);
1714 /* we use swappiness of local cgroup */
1716 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1717 noswap, zone, &nr_scanned);
1718 *total_scanned += nr_scanned;
1720 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1722 css_put(&victim->css);
1724 * At shrinking usage, we can't check we should stop here or
1725 * reclaim more. It's depends on callers. last_scanned_child
1726 * will work enough for keeping fairness under tree.
1732 if (!res_counter_soft_limit_excess(&root_memcg->res))
1734 } else if (mem_cgroup_margin(root_memcg))
1741 * Check OOM-Killer is already running under our hierarchy.
1742 * If someone is running, return false.
1743 * Has to be called with memcg_oom_lock
1745 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1747 struct mem_cgroup *iter, *failed = NULL;
1750 for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1751 if (iter->oom_lock) {
1753 * this subtree of our hierarchy is already locked
1754 * so we cannot give a lock.
1759 iter->oom_lock = true;
1766 * OK, we failed to lock the whole subtree so we have to clean up
1767 * what we set up to the failing subtree
1770 for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1771 if (iter == failed) {
1775 iter->oom_lock = false;
1781 * Has to be called with memcg_oom_lock
1783 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1785 struct mem_cgroup *iter;
1787 for_each_mem_cgroup_tree(iter, memcg)
1788 iter->oom_lock = false;
1792 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1794 struct mem_cgroup *iter;
1796 for_each_mem_cgroup_tree(iter, memcg)
1797 atomic_inc(&iter->under_oom);
1800 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1802 struct mem_cgroup *iter;
1805 * When a new child is created while the hierarchy is under oom,
1806 * mem_cgroup_oom_lock() may not be called. We have to use
1807 * atomic_add_unless() here.
1809 for_each_mem_cgroup_tree(iter, memcg)
1810 atomic_add_unless(&iter->under_oom, -1, 0);
1813 static DEFINE_SPINLOCK(memcg_oom_lock);
1814 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1816 struct oom_wait_info {
1817 struct mem_cgroup *mem;
1821 static int memcg_oom_wake_function(wait_queue_t *wait,
1822 unsigned mode, int sync, void *arg)
1824 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1826 struct oom_wait_info *oom_wait_info;
1828 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1829 oom_wait_memcg = oom_wait_info->mem;
1832 * Both of oom_wait_info->mem and wake_mem are stable under us.
1833 * Then we can use css_is_ancestor without taking care of RCU.
1835 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1836 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1838 return autoremove_wake_function(wait, mode, sync, arg);
1841 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1843 /* for filtering, pass "memcg" as argument. */
1844 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1847 static void memcg_oom_recover(struct mem_cgroup *memcg)
1849 if (memcg && atomic_read(&memcg->under_oom))
1850 memcg_wakeup_oom(memcg);
1854 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1856 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1858 struct oom_wait_info owait;
1859 bool locked, need_to_kill;
1862 owait.wait.flags = 0;
1863 owait.wait.func = memcg_oom_wake_function;
1864 owait.wait.private = current;
1865 INIT_LIST_HEAD(&owait.wait.task_list);
1866 need_to_kill = true;
1867 mem_cgroup_mark_under_oom(memcg);
1869 /* At first, try to OOM lock hierarchy under memcg.*/
1870 spin_lock(&memcg_oom_lock);
1871 locked = mem_cgroup_oom_lock(memcg);
1873 * Even if signal_pending(), we can't quit charge() loop without
1874 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1875 * under OOM is always welcomed, use TASK_KILLABLE here.
1877 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1878 if (!locked || memcg->oom_kill_disable)
1879 need_to_kill = false;
1881 mem_cgroup_oom_notify(memcg);
1882 spin_unlock(&memcg_oom_lock);
1885 finish_wait(&memcg_oom_waitq, &owait.wait);
1886 mem_cgroup_out_of_memory(memcg, mask);
1889 finish_wait(&memcg_oom_waitq, &owait.wait);
1891 spin_lock(&memcg_oom_lock);
1893 mem_cgroup_oom_unlock(memcg);
1894 memcg_wakeup_oom(memcg);
1895 spin_unlock(&memcg_oom_lock);
1897 mem_cgroup_unmark_under_oom(memcg);
1899 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1901 /* Give chance to dying process */
1902 schedule_timeout_uninterruptible(1);
1907 * Currently used to update mapped file statistics, but the routine can be
1908 * generalized to update other statistics as well.
1910 * Notes: Race condition
1912 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1913 * it tends to be costly. But considering some conditions, we doesn't need
1914 * to do so _always_.
1916 * Considering "charge", lock_page_cgroup() is not required because all
1917 * file-stat operations happen after a page is attached to radix-tree. There
1918 * are no race with "charge".
1920 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1921 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1922 * if there are race with "uncharge". Statistics itself is properly handled
1925 * Considering "move", this is an only case we see a race. To make the race
1926 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1927 * possibility of race condition. If there is, we take a lock.
1930 void mem_cgroup_update_page_stat(struct page *page,
1931 enum mem_cgroup_page_stat_item idx, int val)
1933 struct mem_cgroup *memcg;
1934 struct page_cgroup *pc = lookup_page_cgroup(page);
1935 bool need_unlock = false;
1936 unsigned long uninitialized_var(flags);
1942 memcg = pc->mem_cgroup;
1943 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1945 /* pc->mem_cgroup is unstable ? */
1946 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
1947 /* take a lock against to access pc->mem_cgroup */
1948 move_lock_page_cgroup(pc, &flags);
1950 memcg = pc->mem_cgroup;
1951 if (!memcg || !PageCgroupUsed(pc))
1956 case MEMCG_NR_FILE_MAPPED:
1958 SetPageCgroupFileMapped(pc);
1959 else if (!page_mapped(page))
1960 ClearPageCgroupFileMapped(pc);
1961 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1967 this_cpu_add(memcg->stat->count[idx], val);
1970 if (unlikely(need_unlock))
1971 move_unlock_page_cgroup(pc, &flags);
1975 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1978 * size of first charge trial. "32" comes from vmscan.c's magic value.
1979 * TODO: maybe necessary to use big numbers in big irons.
1981 #define CHARGE_BATCH 32U
1982 struct memcg_stock_pcp {
1983 struct mem_cgroup *cached; /* this never be root cgroup */
1984 unsigned int nr_pages;
1985 struct work_struct work;
1986 unsigned long flags;
1987 #define FLUSHING_CACHED_CHARGE (0)
1989 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1990 static DEFINE_MUTEX(percpu_charge_mutex);
1993 * Try to consume stocked charge on this cpu. If success, one page is consumed
1994 * from local stock and true is returned. If the stock is 0 or charges from a
1995 * cgroup which is not current target, returns false. This stock will be
1998 static bool consume_stock(struct mem_cgroup *memcg)
2000 struct memcg_stock_pcp *stock;
2003 stock = &get_cpu_var(memcg_stock);
2004 if (memcg == stock->cached && stock->nr_pages)
2006 else /* need to call res_counter_charge */
2008 put_cpu_var(memcg_stock);
2013 * Returns stocks cached in percpu to res_counter and reset cached information.
2015 static void drain_stock(struct memcg_stock_pcp *stock)
2017 struct mem_cgroup *old = stock->cached;
2019 if (stock->nr_pages) {
2020 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2022 res_counter_uncharge(&old->res, bytes);
2023 if (do_swap_account)
2024 res_counter_uncharge(&old->memsw, bytes);
2025 stock->nr_pages = 0;
2027 stock->cached = NULL;
2031 * This must be called under preempt disabled or must be called by
2032 * a thread which is pinned to local cpu.
2034 static void drain_local_stock(struct work_struct *dummy)
2036 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2038 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2042 * Cache charges(val) which is from res_counter, to local per_cpu area.
2043 * This will be consumed by consume_stock() function, later.
2045 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2047 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2049 if (stock->cached != memcg) { /* reset if necessary */
2051 stock->cached = memcg;
2053 stock->nr_pages += nr_pages;
2054 put_cpu_var(memcg_stock);
2058 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2059 * of the hierarchy under it. sync flag says whether we should block
2060 * until the work is done.
2062 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2066 /* Notify other cpus that system-wide "drain" is running */
2069 for_each_online_cpu(cpu) {
2070 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2071 struct mem_cgroup *memcg;
2073 memcg = stock->cached;
2074 if (!memcg || !stock->nr_pages)
2076 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2078 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2080 drain_local_stock(&stock->work);
2082 schedule_work_on(cpu, &stock->work);
2090 for_each_online_cpu(cpu) {
2091 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2092 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2093 flush_work(&stock->work);
2100 * Tries to drain stocked charges in other cpus. This function is asynchronous
2101 * and just put a work per cpu for draining localy on each cpu. Caller can
2102 * expects some charges will be back to res_counter later but cannot wait for
2105 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2108 * If someone calls draining, avoid adding more kworker runs.
2110 if (!mutex_trylock(&percpu_charge_mutex))
2112 drain_all_stock(root_memcg, false);
2113 mutex_unlock(&percpu_charge_mutex);
2116 /* This is a synchronous drain interface. */
2117 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2119 /* called when force_empty is called */
2120 mutex_lock(&percpu_charge_mutex);
2121 drain_all_stock(root_memcg, true);
2122 mutex_unlock(&percpu_charge_mutex);
2126 * This function drains percpu counter value from DEAD cpu and
2127 * move it to local cpu. Note that this function can be preempted.
2129 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2133 spin_lock(&memcg->pcp_counter_lock);
2134 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2135 long x = per_cpu(memcg->stat->count[i], cpu);
2137 per_cpu(memcg->stat->count[i], cpu) = 0;
2138 memcg->nocpu_base.count[i] += x;
2140 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2141 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2143 per_cpu(memcg->stat->events[i], cpu) = 0;
2144 memcg->nocpu_base.events[i] += x;
2146 /* need to clear ON_MOVE value, works as a kind of lock. */
2147 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2148 spin_unlock(&memcg->pcp_counter_lock);
2151 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2153 int idx = MEM_CGROUP_ON_MOVE;
2155 spin_lock(&memcg->pcp_counter_lock);
2156 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2157 spin_unlock(&memcg->pcp_counter_lock);
2160 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2161 unsigned long action,
2164 int cpu = (unsigned long)hcpu;
2165 struct memcg_stock_pcp *stock;
2166 struct mem_cgroup *iter;
2168 if ((action == CPU_ONLINE)) {
2169 for_each_mem_cgroup_all(iter)
2170 synchronize_mem_cgroup_on_move(iter, cpu);
2174 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2177 for_each_mem_cgroup_all(iter)
2178 mem_cgroup_drain_pcp_counter(iter, cpu);
2180 stock = &per_cpu(memcg_stock, cpu);
2186 /* See __mem_cgroup_try_charge() for details */
2188 CHARGE_OK, /* success */
2189 CHARGE_RETRY, /* need to retry but retry is not bad */
2190 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2191 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2192 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2195 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2196 unsigned int nr_pages, bool oom_check)
2198 unsigned long csize = nr_pages * PAGE_SIZE;
2199 struct mem_cgroup *mem_over_limit;
2200 struct res_counter *fail_res;
2201 unsigned long flags = 0;
2204 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2207 if (!do_swap_account)
2209 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2213 res_counter_uncharge(&memcg->res, csize);
2214 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2215 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2217 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2219 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2220 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2222 * Never reclaim on behalf of optional batching, retry with a
2223 * single page instead.
2225 if (nr_pages == CHARGE_BATCH)
2226 return CHARGE_RETRY;
2228 if (!(gfp_mask & __GFP_WAIT))
2229 return CHARGE_WOULDBLOCK;
2231 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2232 gfp_mask, flags, NULL);
2233 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2234 return CHARGE_RETRY;
2236 * Even though the limit is exceeded at this point, reclaim
2237 * may have been able to free some pages. Retry the charge
2238 * before killing the task.
2240 * Only for regular pages, though: huge pages are rather
2241 * unlikely to succeed so close to the limit, and we fall back
2242 * to regular pages anyway in case of failure.
2244 if (nr_pages == 1 && ret)
2245 return CHARGE_RETRY;
2248 * At task move, charge accounts can be doubly counted. So, it's
2249 * better to wait until the end of task_move if something is going on.
2251 if (mem_cgroup_wait_acct_move(mem_over_limit))
2252 return CHARGE_RETRY;
2254 /* If we don't need to call oom-killer at el, return immediately */
2256 return CHARGE_NOMEM;
2258 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2259 return CHARGE_OOM_DIE;
2261 return CHARGE_RETRY;
2265 * Unlike exported interface, "oom" parameter is added. if oom==true,
2266 * oom-killer can be invoked.
2268 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2270 unsigned int nr_pages,
2271 struct mem_cgroup **ptr,
2274 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2275 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2276 struct mem_cgroup *memcg = NULL;
2280 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2281 * in system level. So, allow to go ahead dying process in addition to
2284 if (unlikely(test_thread_flag(TIF_MEMDIE)
2285 || fatal_signal_pending(current)))
2289 * We always charge the cgroup the mm_struct belongs to.
2290 * The mm_struct's mem_cgroup changes on task migration if the
2291 * thread group leader migrates. It's possible that mm is not
2292 * set, if so charge the init_mm (happens for pagecache usage).
2297 if (*ptr) { /* css should be a valid one */
2299 VM_BUG_ON(css_is_removed(&memcg->css));
2300 if (mem_cgroup_is_root(memcg))
2302 if (nr_pages == 1 && consume_stock(memcg))
2304 css_get(&memcg->css);
2306 struct task_struct *p;
2309 p = rcu_dereference(mm->owner);
2311 * Because we don't have task_lock(), "p" can exit.
2312 * In that case, "memcg" can point to root or p can be NULL with
2313 * race with swapoff. Then, we have small risk of mis-accouning.
2314 * But such kind of mis-account by race always happens because
2315 * we don't have cgroup_mutex(). It's overkill and we allo that
2317 * (*) swapoff at el will charge against mm-struct not against
2318 * task-struct. So, mm->owner can be NULL.
2320 memcg = mem_cgroup_from_task(p);
2321 if (!memcg || mem_cgroup_is_root(memcg)) {
2325 if (nr_pages == 1 && consume_stock(memcg)) {
2327 * It seems dagerous to access memcg without css_get().
2328 * But considering how consume_stok works, it's not
2329 * necessary. If consume_stock success, some charges
2330 * from this memcg are cached on this cpu. So, we
2331 * don't need to call css_get()/css_tryget() before
2332 * calling consume_stock().
2337 /* after here, we may be blocked. we need to get refcnt */
2338 if (!css_tryget(&memcg->css)) {
2348 /* If killed, bypass charge */
2349 if (fatal_signal_pending(current)) {
2350 css_put(&memcg->css);
2355 if (oom && !nr_oom_retries) {
2357 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2360 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2364 case CHARGE_RETRY: /* not in OOM situation but retry */
2366 css_put(&memcg->css);
2369 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2370 css_put(&memcg->css);
2372 case CHARGE_NOMEM: /* OOM routine works */
2374 css_put(&memcg->css);
2377 /* If oom, we never return -ENOMEM */
2380 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2381 css_put(&memcg->css);
2384 } while (ret != CHARGE_OK);
2386 if (batch > nr_pages)
2387 refill_stock(memcg, batch - nr_pages);
2388 css_put(&memcg->css);
2401 * Somemtimes we have to undo a charge we got by try_charge().
2402 * This function is for that and do uncharge, put css's refcnt.
2403 * gotten by try_charge().
2405 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2406 unsigned int nr_pages)
2408 if (!mem_cgroup_is_root(memcg)) {
2409 unsigned long bytes = nr_pages * PAGE_SIZE;
2411 res_counter_uncharge(&memcg->res, bytes);
2412 if (do_swap_account)
2413 res_counter_uncharge(&memcg->memsw, bytes);
2418 * A helper function to get mem_cgroup from ID. must be called under
2419 * rcu_read_lock(). The caller must check css_is_removed() or some if
2420 * it's concern. (dropping refcnt from swap can be called against removed
2423 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2425 struct cgroup_subsys_state *css;
2427 /* ID 0 is unused ID */
2430 css = css_lookup(&mem_cgroup_subsys, id);
2433 return container_of(css, struct mem_cgroup, css);
2436 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2438 struct mem_cgroup *memcg = NULL;
2439 struct page_cgroup *pc;
2443 VM_BUG_ON(!PageLocked(page));
2445 pc = lookup_page_cgroup(page);
2446 lock_page_cgroup(pc);
2447 if (PageCgroupUsed(pc)) {
2448 memcg = pc->mem_cgroup;
2449 if (memcg && !css_tryget(&memcg->css))
2451 } else if (PageSwapCache(page)) {
2452 ent.val = page_private(page);
2453 id = lookup_swap_cgroup(ent);
2455 memcg = mem_cgroup_lookup(id);
2456 if (memcg && !css_tryget(&memcg->css))
2460 unlock_page_cgroup(pc);
2464 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2466 unsigned int nr_pages,
2467 struct page_cgroup *pc,
2468 enum charge_type ctype)
2470 lock_page_cgroup(pc);
2471 if (unlikely(PageCgroupUsed(pc))) {
2472 unlock_page_cgroup(pc);
2473 __mem_cgroup_cancel_charge(memcg, nr_pages);
2477 * we don't need page_cgroup_lock about tail pages, becase they are not
2478 * accessed by any other context at this point.
2480 pc->mem_cgroup = memcg;
2482 * We access a page_cgroup asynchronously without lock_page_cgroup().
2483 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2484 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2485 * before USED bit, we need memory barrier here.
2486 * See mem_cgroup_add_lru_list(), etc.
2490 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2491 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2492 SetPageCgroupCache(pc);
2493 SetPageCgroupUsed(pc);
2495 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2496 ClearPageCgroupCache(pc);
2497 SetPageCgroupUsed(pc);
2503 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2504 unlock_page_cgroup(pc);
2506 * "charge_statistics" updated event counter. Then, check it.
2507 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2508 * if they exceeds softlimit.
2510 memcg_check_events(memcg, page);
2513 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2515 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2516 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2518 * Because tail pages are not marked as "used", set it. We're under
2519 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2521 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2523 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2524 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2525 unsigned long flags;
2527 if (mem_cgroup_disabled())
2530 * We have no races with charge/uncharge but will have races with
2531 * page state accounting.
2533 move_lock_page_cgroup(head_pc, &flags);
2535 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2536 smp_wmb(); /* see __commit_charge() */
2537 if (PageCgroupAcctLRU(head_pc)) {
2539 struct mem_cgroup_per_zone *mz;
2542 * LRU flags cannot be copied because we need to add tail
2543 *.page to LRU by generic call and our hook will be called.
2544 * We hold lru_lock, then, reduce counter directly.
2546 lru = page_lru(head);
2547 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2548 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2550 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2551 move_unlock_page_cgroup(head_pc, &flags);
2556 * mem_cgroup_move_account - move account of the page
2558 * @nr_pages: number of regular pages (>1 for huge pages)
2559 * @pc: page_cgroup of the page.
2560 * @from: mem_cgroup which the page is moved from.
2561 * @to: mem_cgroup which the page is moved to. @from != @to.
2562 * @uncharge: whether we should call uncharge and css_put against @from.
2564 * The caller must confirm following.
2565 * - page is not on LRU (isolate_page() is useful.)
2566 * - compound_lock is held when nr_pages > 1
2568 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2569 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2570 * true, this function does "uncharge" from old cgroup, but it doesn't if
2571 * @uncharge is false, so a caller should do "uncharge".
2573 static int mem_cgroup_move_account(struct page *page,
2574 unsigned int nr_pages,
2575 struct page_cgroup *pc,
2576 struct mem_cgroup *from,
2577 struct mem_cgroup *to,
2580 unsigned long flags;
2583 VM_BUG_ON(from == to);
2584 VM_BUG_ON(PageLRU(page));
2586 * The page is isolated from LRU. So, collapse function
2587 * will not handle this page. But page splitting can happen.
2588 * Do this check under compound_page_lock(). The caller should
2592 if (nr_pages > 1 && !PageTransHuge(page))
2595 lock_page_cgroup(pc);
2598 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2601 move_lock_page_cgroup(pc, &flags);
2603 if (PageCgroupFileMapped(pc)) {
2604 /* Update mapped_file data for mem_cgroup */
2606 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2607 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2610 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2612 /* This is not "cancel", but cancel_charge does all we need. */
2613 __mem_cgroup_cancel_charge(from, nr_pages);
2615 /* caller should have done css_get */
2616 pc->mem_cgroup = to;
2617 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2619 * We charges against "to" which may not have any tasks. Then, "to"
2620 * can be under rmdir(). But in current implementation, caller of
2621 * this function is just force_empty() and move charge, so it's
2622 * guaranteed that "to" is never removed. So, we don't check rmdir
2625 move_unlock_page_cgroup(pc, &flags);
2628 unlock_page_cgroup(pc);
2632 memcg_check_events(to, page);
2633 memcg_check_events(from, page);
2639 * move charges to its parent.
2642 static int mem_cgroup_move_parent(struct page *page,
2643 struct page_cgroup *pc,
2644 struct mem_cgroup *child,
2647 struct cgroup *cg = child->css.cgroup;
2648 struct cgroup *pcg = cg->parent;
2649 struct mem_cgroup *parent;
2650 unsigned int nr_pages;
2651 unsigned long uninitialized_var(flags);
2659 if (!get_page_unless_zero(page))
2661 if (isolate_lru_page(page))
2664 nr_pages = hpage_nr_pages(page);
2666 parent = mem_cgroup_from_cont(pcg);
2667 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2672 flags = compound_lock_irqsave(page);
2674 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2676 __mem_cgroup_cancel_charge(parent, nr_pages);
2679 compound_unlock_irqrestore(page, flags);
2681 putback_lru_page(page);
2689 * Charge the memory controller for page usage.
2691 * 0 if the charge was successful
2692 * < 0 if the cgroup is over its limit
2694 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2695 gfp_t gfp_mask, enum charge_type ctype)
2697 struct mem_cgroup *memcg = NULL;
2698 unsigned int nr_pages = 1;
2699 struct page_cgroup *pc;
2703 if (PageTransHuge(page)) {
2704 nr_pages <<= compound_order(page);
2705 VM_BUG_ON(!PageTransHuge(page));
2707 * Never OOM-kill a process for a huge page. The
2708 * fault handler will fall back to regular pages.
2713 pc = lookup_page_cgroup(page);
2714 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2716 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2720 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2724 int mem_cgroup_newpage_charge(struct page *page,
2725 struct mm_struct *mm, gfp_t gfp_mask)
2727 if (mem_cgroup_disabled())
2730 * If already mapped, we don't have to account.
2731 * If page cache, page->mapping has address_space.
2732 * But page->mapping may have out-of-use anon_vma pointer,
2733 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2736 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2740 return mem_cgroup_charge_common(page, mm, gfp_mask,
2741 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2745 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2746 enum charge_type ctype);
2749 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2750 enum charge_type ctype)
2752 struct page_cgroup *pc = lookup_page_cgroup(page);
2754 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2755 * is already on LRU. It means the page may on some other page_cgroup's
2756 * LRU. Take care of it.
2758 mem_cgroup_lru_del_before_commit(page);
2759 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2760 mem_cgroup_lru_add_after_commit(page);
2764 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2767 struct mem_cgroup *memcg = NULL;
2770 if (mem_cgroup_disabled())
2772 if (PageCompound(page))
2778 if (page_is_file_cache(page)) {
2779 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2784 * FUSE reuses pages without going through the final
2785 * put that would remove them from the LRU list, make
2786 * sure that they get relinked properly.
2788 __mem_cgroup_commit_charge_lrucare(page, memcg,
2789 MEM_CGROUP_CHARGE_TYPE_CACHE);
2793 if (PageSwapCache(page)) {
2794 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2796 __mem_cgroup_commit_charge_swapin(page, memcg,
2797 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2799 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2800 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2806 * While swap-in, try_charge -> commit or cancel, the page is locked.
2807 * And when try_charge() successfully returns, one refcnt to memcg without
2808 * struct page_cgroup is acquired. This refcnt will be consumed by
2809 * "commit()" or removed by "cancel()"
2811 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2813 gfp_t mask, struct mem_cgroup **ptr)
2815 struct mem_cgroup *memcg;
2820 if (mem_cgroup_disabled())
2823 if (!do_swap_account)
2826 * A racing thread's fault, or swapoff, may have already updated
2827 * the pte, and even removed page from swap cache: in those cases
2828 * do_swap_page()'s pte_same() test will fail; but there's also a
2829 * KSM case which does need to charge the page.
2831 if (!PageSwapCache(page))
2833 memcg = try_get_mem_cgroup_from_page(page);
2837 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2838 css_put(&memcg->css);
2843 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2847 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2848 enum charge_type ctype)
2850 if (mem_cgroup_disabled())
2854 cgroup_exclude_rmdir(&ptr->css);
2856 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2858 * Now swap is on-memory. This means this page may be
2859 * counted both as mem and swap....double count.
2860 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2861 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2862 * may call delete_from_swap_cache() before reach here.
2864 if (do_swap_account && PageSwapCache(page)) {
2865 swp_entry_t ent = {.val = page_private(page)};
2867 struct mem_cgroup *memcg;
2869 id = swap_cgroup_record(ent, 0);
2871 memcg = mem_cgroup_lookup(id);
2874 * This recorded memcg can be obsolete one. So, avoid
2875 * calling css_tryget
2877 if (!mem_cgroup_is_root(memcg))
2878 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2879 mem_cgroup_swap_statistics(memcg, false);
2880 mem_cgroup_put(memcg);
2885 * At swapin, we may charge account against cgroup which has no tasks.
2886 * So, rmdir()->pre_destroy() can be called while we do this charge.
2887 * In that case, we need to call pre_destroy() again. check it here.
2889 cgroup_release_and_wakeup_rmdir(&ptr->css);
2892 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2894 __mem_cgroup_commit_charge_swapin(page, ptr,
2895 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2898 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2900 if (mem_cgroup_disabled())
2904 __mem_cgroup_cancel_charge(memcg, 1);
2907 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2908 unsigned int nr_pages,
2909 const enum charge_type ctype)
2911 struct memcg_batch_info *batch = NULL;
2912 bool uncharge_memsw = true;
2914 /* If swapout, usage of swap doesn't decrease */
2915 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2916 uncharge_memsw = false;
2918 batch = ¤t->memcg_batch;
2920 * In usual, we do css_get() when we remember memcg pointer.
2921 * But in this case, we keep res->usage until end of a series of
2922 * uncharges. Then, it's ok to ignore memcg's refcnt.
2925 batch->memcg = memcg;
2927 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2928 * In those cases, all pages freed continuously can be expected to be in
2929 * the same cgroup and we have chance to coalesce uncharges.
2930 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2931 * because we want to do uncharge as soon as possible.
2934 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2935 goto direct_uncharge;
2938 goto direct_uncharge;
2941 * In typical case, batch->memcg == mem. This means we can
2942 * merge a series of uncharges to an uncharge of res_counter.
2943 * If not, we uncharge res_counter ony by one.
2945 if (batch->memcg != memcg)
2946 goto direct_uncharge;
2947 /* remember freed charge and uncharge it later */
2950 batch->memsw_nr_pages++;
2953 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2955 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2956 if (unlikely(batch->memcg != memcg))
2957 memcg_oom_recover(memcg);
2962 * uncharge if !page_mapped(page)
2964 static struct mem_cgroup *
2965 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2967 struct mem_cgroup *memcg = NULL;
2968 unsigned int nr_pages = 1;
2969 struct page_cgroup *pc;
2971 if (mem_cgroup_disabled())
2974 if (PageSwapCache(page))
2977 if (PageTransHuge(page)) {
2978 nr_pages <<= compound_order(page);
2979 VM_BUG_ON(!PageTransHuge(page));
2982 * Check if our page_cgroup is valid
2984 pc = lookup_page_cgroup(page);
2985 if (unlikely(!pc || !PageCgroupUsed(pc)))
2988 lock_page_cgroup(pc);
2990 memcg = pc->mem_cgroup;
2992 if (!PageCgroupUsed(pc))
2996 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2997 case MEM_CGROUP_CHARGE_TYPE_DROP:
2998 /* See mem_cgroup_prepare_migration() */
2999 if (page_mapped(page) || PageCgroupMigration(pc))
3002 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3003 if (!PageAnon(page)) { /* Shared memory */
3004 if (page->mapping && !page_is_file_cache(page))
3006 } else if (page_mapped(page)) /* Anon */
3013 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3015 ClearPageCgroupUsed(pc);
3017 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3018 * freed from LRU. This is safe because uncharged page is expected not
3019 * to be reused (freed soon). Exception is SwapCache, it's handled by
3020 * special functions.
3023 unlock_page_cgroup(pc);
3025 * even after unlock, we have mem->res.usage here and this memcg
3026 * will never be freed.
3028 memcg_check_events(memcg, page);
3029 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3030 mem_cgroup_swap_statistics(memcg, true);
3031 mem_cgroup_get(memcg);
3033 if (!mem_cgroup_is_root(memcg))
3034 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3039 unlock_page_cgroup(pc);
3043 void mem_cgroup_uncharge_page(struct page *page)
3046 if (page_mapped(page))
3048 if (page->mapping && !PageAnon(page))
3050 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3053 void mem_cgroup_uncharge_cache_page(struct page *page)
3055 VM_BUG_ON(page_mapped(page));
3056 VM_BUG_ON(page->mapping);
3057 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3061 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3062 * In that cases, pages are freed continuously and we can expect pages
3063 * are in the same memcg. All these calls itself limits the number of
3064 * pages freed at once, then uncharge_start/end() is called properly.
3065 * This may be called prural(2) times in a context,
3068 void mem_cgroup_uncharge_start(void)
3070 current->memcg_batch.do_batch++;
3071 /* We can do nest. */
3072 if (current->memcg_batch.do_batch == 1) {
3073 current->memcg_batch.memcg = NULL;
3074 current->memcg_batch.nr_pages = 0;
3075 current->memcg_batch.memsw_nr_pages = 0;
3079 void mem_cgroup_uncharge_end(void)
3081 struct memcg_batch_info *batch = ¤t->memcg_batch;
3083 if (!batch->do_batch)
3087 if (batch->do_batch) /* If stacked, do nothing. */
3093 * This "batch->memcg" is valid without any css_get/put etc...
3094 * bacause we hide charges behind us.
3096 if (batch->nr_pages)
3097 res_counter_uncharge(&batch->memcg->res,
3098 batch->nr_pages * PAGE_SIZE);
3099 if (batch->memsw_nr_pages)
3100 res_counter_uncharge(&batch->memcg->memsw,
3101 batch->memsw_nr_pages * PAGE_SIZE);
3102 memcg_oom_recover(batch->memcg);
3103 /* forget this pointer (for sanity check) */
3104 batch->memcg = NULL;
3109 * called after __delete_from_swap_cache() and drop "page" account.
3110 * memcg information is recorded to swap_cgroup of "ent"
3113 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3115 struct mem_cgroup *memcg;
3116 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3118 if (!swapout) /* this was a swap cache but the swap is unused ! */
3119 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3121 memcg = __mem_cgroup_uncharge_common(page, ctype);
3124 * record memcg information, if swapout && memcg != NULL,
3125 * mem_cgroup_get() was called in uncharge().
3127 if (do_swap_account && swapout && memcg)
3128 swap_cgroup_record(ent, css_id(&memcg->css));
3132 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3134 * called from swap_entry_free(). remove record in swap_cgroup and
3135 * uncharge "memsw" account.
3137 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3139 struct mem_cgroup *memcg;
3142 if (!do_swap_account)
3145 id = swap_cgroup_record(ent, 0);
3147 memcg = mem_cgroup_lookup(id);
3150 * We uncharge this because swap is freed.
3151 * This memcg can be obsolete one. We avoid calling css_tryget
3153 if (!mem_cgroup_is_root(memcg))
3154 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3155 mem_cgroup_swap_statistics(memcg, false);
3156 mem_cgroup_put(memcg);
3162 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3163 * @entry: swap entry to be moved
3164 * @from: mem_cgroup which the entry is moved from
3165 * @to: mem_cgroup which the entry is moved to
3166 * @need_fixup: whether we should fixup res_counters and refcounts.
3168 * It succeeds only when the swap_cgroup's record for this entry is the same
3169 * as the mem_cgroup's id of @from.
3171 * Returns 0 on success, -EINVAL on failure.
3173 * The caller must have charged to @to, IOW, called res_counter_charge() about
3174 * both res and memsw, and called css_get().
3176 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3177 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3179 unsigned short old_id, new_id;
3181 old_id = css_id(&from->css);
3182 new_id = css_id(&to->css);
3184 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3185 mem_cgroup_swap_statistics(from, false);
3186 mem_cgroup_swap_statistics(to, true);
3188 * This function is only called from task migration context now.
3189 * It postpones res_counter and refcount handling till the end
3190 * of task migration(mem_cgroup_clear_mc()) for performance
3191 * improvement. But we cannot postpone mem_cgroup_get(to)
3192 * because if the process that has been moved to @to does
3193 * swap-in, the refcount of @to might be decreased to 0.
3197 if (!mem_cgroup_is_root(from))
3198 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3199 mem_cgroup_put(from);
3201 * we charged both to->res and to->memsw, so we should
3204 if (!mem_cgroup_is_root(to))
3205 res_counter_uncharge(&to->res, PAGE_SIZE);
3212 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3213 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3220 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3223 int mem_cgroup_prepare_migration(struct page *page,
3224 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3226 struct mem_cgroup *memcg = NULL;
3227 struct page_cgroup *pc;
3228 enum charge_type ctype;
3233 VM_BUG_ON(PageTransHuge(page));
3234 if (mem_cgroup_disabled())
3237 pc = lookup_page_cgroup(page);
3238 lock_page_cgroup(pc);
3239 if (PageCgroupUsed(pc)) {
3240 memcg = pc->mem_cgroup;
3241 css_get(&memcg->css);
3243 * At migrating an anonymous page, its mapcount goes down
3244 * to 0 and uncharge() will be called. But, even if it's fully
3245 * unmapped, migration may fail and this page has to be
3246 * charged again. We set MIGRATION flag here and delay uncharge
3247 * until end_migration() is called
3249 * Corner Case Thinking
3251 * When the old page was mapped as Anon and it's unmap-and-freed
3252 * while migration was ongoing.
3253 * If unmap finds the old page, uncharge() of it will be delayed
3254 * until end_migration(). If unmap finds a new page, it's
3255 * uncharged when it make mapcount to be 1->0. If unmap code
3256 * finds swap_migration_entry, the new page will not be mapped
3257 * and end_migration() will find it(mapcount==0).
3260 * When the old page was mapped but migraion fails, the kernel
3261 * remaps it. A charge for it is kept by MIGRATION flag even
3262 * if mapcount goes down to 0. We can do remap successfully
3263 * without charging it again.
3266 * The "old" page is under lock_page() until the end of
3267 * migration, so, the old page itself will not be swapped-out.
3268 * If the new page is swapped out before end_migraton, our
3269 * hook to usual swap-out path will catch the event.
3272 SetPageCgroupMigration(pc);
3274 unlock_page_cgroup(pc);
3276 * If the page is not charged at this point,
3283 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3284 css_put(&memcg->css);/* drop extra refcnt */
3285 if (ret || *ptr == NULL) {
3286 if (PageAnon(page)) {
3287 lock_page_cgroup(pc);
3288 ClearPageCgroupMigration(pc);
3289 unlock_page_cgroup(pc);
3291 * The old page may be fully unmapped while we kept it.
3293 mem_cgroup_uncharge_page(page);
3298 * We charge new page before it's used/mapped. So, even if unlock_page()
3299 * is called before end_migration, we can catch all events on this new
3300 * page. In the case new page is migrated but not remapped, new page's
3301 * mapcount will be finally 0 and we call uncharge in end_migration().
3303 pc = lookup_page_cgroup(newpage);
3305 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3306 else if (page_is_file_cache(page))
3307 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3309 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3310 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3314 /* remove redundant charge if migration failed*/
3315 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3316 struct page *oldpage, struct page *newpage, bool migration_ok)
3318 struct page *used, *unused;
3319 struct page_cgroup *pc;
3323 /* blocks rmdir() */
3324 cgroup_exclude_rmdir(&memcg->css);
3325 if (!migration_ok) {
3333 * We disallowed uncharge of pages under migration because mapcount
3334 * of the page goes down to zero, temporarly.
3335 * Clear the flag and check the page should be charged.
3337 pc = lookup_page_cgroup(oldpage);
3338 lock_page_cgroup(pc);
3339 ClearPageCgroupMigration(pc);
3340 unlock_page_cgroup(pc);
3342 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3345 * If a page is a file cache, radix-tree replacement is very atomic
3346 * and we can skip this check. When it was an Anon page, its mapcount
3347 * goes down to 0. But because we added MIGRATION flage, it's not
3348 * uncharged yet. There are several case but page->mapcount check
3349 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3350 * check. (see prepare_charge() also)
3353 mem_cgroup_uncharge_page(used);
3355 * At migration, we may charge account against cgroup which has no
3357 * So, rmdir()->pre_destroy() can be called while we do this charge.
3358 * In that case, we need to call pre_destroy() again. check it here.
3360 cgroup_release_and_wakeup_rmdir(&memcg->css);
3363 #ifdef CONFIG_DEBUG_VM
3364 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3366 struct page_cgroup *pc;
3368 pc = lookup_page_cgroup(page);
3369 if (likely(pc) && PageCgroupUsed(pc))
3374 bool mem_cgroup_bad_page_check(struct page *page)
3376 if (mem_cgroup_disabled())
3379 return lookup_page_cgroup_used(page) != NULL;
3382 void mem_cgroup_print_bad_page(struct page *page)
3384 struct page_cgroup *pc;
3386 pc = lookup_page_cgroup_used(page);
3391 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3392 pc, pc->flags, pc->mem_cgroup);
3394 path = kmalloc(PATH_MAX, GFP_KERNEL);
3397 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3402 printk(KERN_CONT "(%s)\n",
3403 (ret < 0) ? "cannot get the path" : path);
3409 static DEFINE_MUTEX(set_limit_mutex);
3411 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3412 unsigned long long val)
3415 u64 memswlimit, memlimit;
3417 int children = mem_cgroup_count_children(memcg);
3418 u64 curusage, oldusage;
3422 * For keeping hierarchical_reclaim simple, how long we should retry
3423 * is depends on callers. We set our retry-count to be function
3424 * of # of children which we should visit in this loop.
3426 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3428 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3431 while (retry_count) {
3432 if (signal_pending(current)) {
3437 * Rather than hide all in some function, I do this in
3438 * open coded manner. You see what this really does.
3439 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3441 mutex_lock(&set_limit_mutex);
3442 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3443 if (memswlimit < val) {
3445 mutex_unlock(&set_limit_mutex);
3449 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3453 ret = res_counter_set_limit(&memcg->res, val);
3455 if (memswlimit == val)
3456 memcg->memsw_is_minimum = true;
3458 memcg->memsw_is_minimum = false;
3460 mutex_unlock(&set_limit_mutex);
3465 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3466 MEM_CGROUP_RECLAIM_SHRINK,
3468 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3469 /* Usage is reduced ? */
3470 if (curusage >= oldusage)
3473 oldusage = curusage;
3475 if (!ret && enlarge)
3476 memcg_oom_recover(memcg);
3481 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3482 unsigned long long val)
3485 u64 memlimit, memswlimit, oldusage, curusage;
3486 int children = mem_cgroup_count_children(memcg);
3490 /* see mem_cgroup_resize_res_limit */
3491 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3492 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3493 while (retry_count) {
3494 if (signal_pending(current)) {
3499 * Rather than hide all in some function, I do this in
3500 * open coded manner. You see what this really does.
3501 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3503 mutex_lock(&set_limit_mutex);
3504 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3505 if (memlimit > val) {
3507 mutex_unlock(&set_limit_mutex);
3510 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3511 if (memswlimit < val)
3513 ret = res_counter_set_limit(&memcg->memsw, val);
3515 if (memlimit == val)
3516 memcg->memsw_is_minimum = true;
3518 memcg->memsw_is_minimum = false;
3520 mutex_unlock(&set_limit_mutex);
3525 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3526 MEM_CGROUP_RECLAIM_NOSWAP |
3527 MEM_CGROUP_RECLAIM_SHRINK,
3529 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3530 /* Usage is reduced ? */
3531 if (curusage >= oldusage)
3534 oldusage = curusage;
3536 if (!ret && enlarge)
3537 memcg_oom_recover(memcg);
3541 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3543 unsigned long *total_scanned)
3545 unsigned long nr_reclaimed = 0;
3546 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3547 unsigned long reclaimed;
3549 struct mem_cgroup_tree_per_zone *mctz;
3550 unsigned long long excess;
3551 unsigned long nr_scanned;
3556 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3558 * This loop can run a while, specially if mem_cgroup's continuously
3559 * keep exceeding their soft limit and putting the system under
3566 mz = mem_cgroup_largest_soft_limit_node(mctz);
3571 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3573 MEM_CGROUP_RECLAIM_SOFT,
3575 nr_reclaimed += reclaimed;
3576 *total_scanned += nr_scanned;
3577 spin_lock(&mctz->lock);
3580 * If we failed to reclaim anything from this memory cgroup
3581 * it is time to move on to the next cgroup
3587 * Loop until we find yet another one.
3589 * By the time we get the soft_limit lock
3590 * again, someone might have aded the
3591 * group back on the RB tree. Iterate to
3592 * make sure we get a different mem.
3593 * mem_cgroup_largest_soft_limit_node returns
3594 * NULL if no other cgroup is present on
3598 __mem_cgroup_largest_soft_limit_node(mctz);
3600 css_put(&next_mz->mem->css);
3601 else /* next_mz == NULL or other memcg */
3605 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3606 excess = res_counter_soft_limit_excess(&mz->mem->res);
3608 * One school of thought says that we should not add
3609 * back the node to the tree if reclaim returns 0.
3610 * But our reclaim could return 0, simply because due
3611 * to priority we are exposing a smaller subset of
3612 * memory to reclaim from. Consider this as a longer
3615 /* If excess == 0, no tree ops */
3616 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3617 spin_unlock(&mctz->lock);
3618 css_put(&mz->mem->css);
3621 * Could not reclaim anything and there are no more
3622 * mem cgroups to try or we seem to be looping without
3623 * reclaiming anything.
3625 if (!nr_reclaimed &&
3627 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3629 } while (!nr_reclaimed);
3631 css_put(&next_mz->mem->css);
3632 return nr_reclaimed;
3636 * This routine traverse page_cgroup in given list and drop them all.
3637 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3639 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3640 int node, int zid, enum lru_list lru)
3643 struct mem_cgroup_per_zone *mz;
3644 struct page_cgroup *pc, *busy;
3645 unsigned long flags, loop;
3646 struct list_head *list;
3649 zone = &NODE_DATA(node)->node_zones[zid];
3650 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3651 list = &mz->lists[lru];
3653 loop = MEM_CGROUP_ZSTAT(mz, lru);
3654 /* give some margin against EBUSY etc...*/
3661 spin_lock_irqsave(&zone->lru_lock, flags);
3662 if (list_empty(list)) {
3663 spin_unlock_irqrestore(&zone->lru_lock, flags);
3666 pc = list_entry(list->prev, struct page_cgroup, lru);
3668 list_move(&pc->lru, list);
3670 spin_unlock_irqrestore(&zone->lru_lock, flags);
3673 spin_unlock_irqrestore(&zone->lru_lock, flags);
3675 page = lookup_cgroup_page(pc);
3677 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3681 if (ret == -EBUSY || ret == -EINVAL) {
3682 /* found lock contention or "pc" is obsolete. */
3689 if (!ret && !list_empty(list))
3695 * make mem_cgroup's charge to be 0 if there is no task.
3696 * This enables deleting this mem_cgroup.
3698 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3701 int node, zid, shrink;
3702 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3703 struct cgroup *cgrp = memcg->css.cgroup;
3705 css_get(&memcg->css);
3708 /* should free all ? */
3714 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3717 if (signal_pending(current))
3719 /* This is for making all *used* pages to be on LRU. */
3720 lru_add_drain_all();
3721 drain_all_stock_sync(memcg);
3723 mem_cgroup_start_move(memcg);
3724 for_each_node_state(node, N_HIGH_MEMORY) {
3725 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3728 ret = mem_cgroup_force_empty_list(memcg,
3737 mem_cgroup_end_move(memcg);
3738 memcg_oom_recover(memcg);
3739 /* it seems parent cgroup doesn't have enough mem */
3743 /* "ret" should also be checked to ensure all lists are empty. */
3744 } while (memcg->res.usage > 0 || ret);
3746 css_put(&memcg->css);
3750 /* returns EBUSY if there is a task or if we come here twice. */
3751 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3755 /* we call try-to-free pages for make this cgroup empty */
3756 lru_add_drain_all();
3757 /* try to free all pages in this cgroup */
3759 while (nr_retries && memcg->res.usage > 0) {
3762 if (signal_pending(current)) {
3766 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3770 /* maybe some writeback is necessary */
3771 congestion_wait(BLK_RW_ASYNC, HZ/10);
3776 /* try move_account...there may be some *locked* pages. */
3780 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3782 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3786 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3788 return mem_cgroup_from_cont(cont)->use_hierarchy;
3791 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3795 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3796 struct cgroup *parent = cont->parent;
3797 struct mem_cgroup *parent_memcg = NULL;
3800 parent_memcg = mem_cgroup_from_cont(parent);
3804 * If parent's use_hierarchy is set, we can't make any modifications
3805 * in the child subtrees. If it is unset, then the change can
3806 * occur, provided the current cgroup has no children.
3808 * For the root cgroup, parent_mem is NULL, we allow value to be
3809 * set if there are no children.
3811 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3812 (val == 1 || val == 0)) {
3813 if (list_empty(&cont->children))
3814 memcg->use_hierarchy = val;
3825 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3826 enum mem_cgroup_stat_index idx)
3828 struct mem_cgroup *iter;
3831 /* Per-cpu values can be negative, use a signed accumulator */
3832 for_each_mem_cgroup_tree(iter, memcg)
3833 val += mem_cgroup_read_stat(iter, idx);
3835 if (val < 0) /* race ? */
3840 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3844 if (!mem_cgroup_is_root(memcg)) {
3846 return res_counter_read_u64(&memcg->res, RES_USAGE);
3848 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3851 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3852 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3855 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3857 return val << PAGE_SHIFT;
3860 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3862 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3866 type = MEMFILE_TYPE(cft->private);
3867 name = MEMFILE_ATTR(cft->private);
3870 if (name == RES_USAGE)
3871 val = mem_cgroup_usage(memcg, false);
3873 val = res_counter_read_u64(&memcg->res, name);
3876 if (name == RES_USAGE)
3877 val = mem_cgroup_usage(memcg, true);
3879 val = res_counter_read_u64(&memcg->memsw, name);
3888 * The user of this function is...
3891 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3894 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3896 unsigned long long val;
3899 type = MEMFILE_TYPE(cft->private);
3900 name = MEMFILE_ATTR(cft->private);
3903 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3907 /* This function does all necessary parse...reuse it */
3908 ret = res_counter_memparse_write_strategy(buffer, &val);
3912 ret = mem_cgroup_resize_limit(memcg, val);
3914 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3916 case RES_SOFT_LIMIT:
3917 ret = res_counter_memparse_write_strategy(buffer, &val);
3921 * For memsw, soft limits are hard to implement in terms
3922 * of semantics, for now, we support soft limits for
3923 * control without swap
3926 ret = res_counter_set_soft_limit(&memcg->res, val);
3931 ret = -EINVAL; /* should be BUG() ? */
3937 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3938 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3940 struct cgroup *cgroup;
3941 unsigned long long min_limit, min_memsw_limit, tmp;
3943 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3944 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3945 cgroup = memcg->css.cgroup;
3946 if (!memcg->use_hierarchy)
3949 while (cgroup->parent) {
3950 cgroup = cgroup->parent;
3951 memcg = mem_cgroup_from_cont(cgroup);
3952 if (!memcg->use_hierarchy)
3954 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3955 min_limit = min(min_limit, tmp);
3956 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3957 min_memsw_limit = min(min_memsw_limit, tmp);
3960 *mem_limit = min_limit;
3961 *memsw_limit = min_memsw_limit;
3965 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3967 struct mem_cgroup *memcg;
3970 memcg = mem_cgroup_from_cont(cont);
3971 type = MEMFILE_TYPE(event);
3972 name = MEMFILE_ATTR(event);
3976 res_counter_reset_max(&memcg->res);
3978 res_counter_reset_max(&memcg->memsw);
3982 res_counter_reset_failcnt(&memcg->res);
3984 res_counter_reset_failcnt(&memcg->memsw);
3991 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3994 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3998 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3999 struct cftype *cft, u64 val)
4001 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4003 if (val >= (1 << NR_MOVE_TYPE))
4006 * We check this value several times in both in can_attach() and
4007 * attach(), so we need cgroup lock to prevent this value from being
4011 memcg->move_charge_at_immigrate = val;
4017 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4018 struct cftype *cft, u64 val)
4025 /* For read statistics */
4043 struct mcs_total_stat {
4044 s64 stat[NR_MCS_STAT];
4050 } memcg_stat_strings[NR_MCS_STAT] = {
4051 {"cache", "total_cache"},
4052 {"rss", "total_rss"},
4053 {"mapped_file", "total_mapped_file"},
4054 {"pgpgin", "total_pgpgin"},
4055 {"pgpgout", "total_pgpgout"},
4056 {"swap", "total_swap"},
4057 {"pgfault", "total_pgfault"},
4058 {"pgmajfault", "total_pgmajfault"},
4059 {"inactive_anon", "total_inactive_anon"},
4060 {"active_anon", "total_active_anon"},
4061 {"inactive_file", "total_inactive_file"},
4062 {"active_file", "total_active_file"},
4063 {"unevictable", "total_unevictable"}
4068 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4073 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4074 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4075 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4076 s->stat[MCS_RSS] += val * PAGE_SIZE;
4077 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4078 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4079 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4080 s->stat[MCS_PGPGIN] += val;
4081 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4082 s->stat[MCS_PGPGOUT] += val;
4083 if (do_swap_account) {
4084 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4085 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4087 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4088 s->stat[MCS_PGFAULT] += val;
4089 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4090 s->stat[MCS_PGMAJFAULT] += val;
4093 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4094 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4095 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4096 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4097 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4098 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4099 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4100 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4101 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4102 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4106 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4108 struct mem_cgroup *iter;
4110 for_each_mem_cgroup_tree(iter, memcg)
4111 mem_cgroup_get_local_stat(iter, s);
4115 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4118 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4119 unsigned long node_nr;
4120 struct cgroup *cont = m->private;
4121 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4123 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4124 seq_printf(m, "total=%lu", total_nr);
4125 for_each_node_state(nid, N_HIGH_MEMORY) {
4126 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4127 seq_printf(m, " N%d=%lu", nid, node_nr);
4131 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4132 seq_printf(m, "file=%lu", file_nr);
4133 for_each_node_state(nid, N_HIGH_MEMORY) {
4134 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4136 seq_printf(m, " N%d=%lu", nid, node_nr);
4140 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4141 seq_printf(m, "anon=%lu", anon_nr);
4142 for_each_node_state(nid, N_HIGH_MEMORY) {
4143 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4145 seq_printf(m, " N%d=%lu", nid, node_nr);
4149 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4150 seq_printf(m, "unevictable=%lu", unevictable_nr);
4151 for_each_node_state(nid, N_HIGH_MEMORY) {
4152 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4153 BIT(LRU_UNEVICTABLE));
4154 seq_printf(m, " N%d=%lu", nid, node_nr);
4159 #endif /* CONFIG_NUMA */
4161 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4162 struct cgroup_map_cb *cb)
4164 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4165 struct mcs_total_stat mystat;
4168 memset(&mystat, 0, sizeof(mystat));
4169 mem_cgroup_get_local_stat(mem_cont, &mystat);
4172 for (i = 0; i < NR_MCS_STAT; i++) {
4173 if (i == MCS_SWAP && !do_swap_account)
4175 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4178 /* Hierarchical information */
4180 unsigned long long limit, memsw_limit;
4181 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4182 cb->fill(cb, "hierarchical_memory_limit", limit);
4183 if (do_swap_account)
4184 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4187 memset(&mystat, 0, sizeof(mystat));
4188 mem_cgroup_get_total_stat(mem_cont, &mystat);
4189 for (i = 0; i < NR_MCS_STAT; i++) {
4190 if (i == MCS_SWAP && !do_swap_account)
4192 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4195 #ifdef CONFIG_DEBUG_VM
4196 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
4200 struct mem_cgroup_per_zone *mz;
4201 unsigned long recent_rotated[2] = {0, 0};
4202 unsigned long recent_scanned[2] = {0, 0};
4204 for_each_online_node(nid)
4205 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4206 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4208 recent_rotated[0] +=
4209 mz->reclaim_stat.recent_rotated[0];
4210 recent_rotated[1] +=
4211 mz->reclaim_stat.recent_rotated[1];
4212 recent_scanned[0] +=
4213 mz->reclaim_stat.recent_scanned[0];
4214 recent_scanned[1] +=
4215 mz->reclaim_stat.recent_scanned[1];
4217 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4218 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4219 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4220 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4227 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4229 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4231 return mem_cgroup_swappiness(memcg);
4234 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4237 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4238 struct mem_cgroup *parent;
4243 if (cgrp->parent == NULL)
4246 parent = mem_cgroup_from_cont(cgrp->parent);
4250 /* If under hierarchy, only empty-root can set this value */
4251 if ((parent->use_hierarchy) ||
4252 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4257 memcg->swappiness = val;
4264 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4266 struct mem_cgroup_threshold_ary *t;
4272 t = rcu_dereference(memcg->thresholds.primary);
4274 t = rcu_dereference(memcg->memsw_thresholds.primary);
4279 usage = mem_cgroup_usage(memcg, swap);
4282 * current_threshold points to threshold just below usage.
4283 * If it's not true, a threshold was crossed after last
4284 * call of __mem_cgroup_threshold().
4286 i = t->current_threshold;
4289 * Iterate backward over array of thresholds starting from
4290 * current_threshold and check if a threshold is crossed.
4291 * If none of thresholds below usage is crossed, we read
4292 * only one element of the array here.
4294 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4295 eventfd_signal(t->entries[i].eventfd, 1);
4297 /* i = current_threshold + 1 */
4301 * Iterate forward over array of thresholds starting from
4302 * current_threshold+1 and check if a threshold is crossed.
4303 * If none of thresholds above usage is crossed, we read
4304 * only one element of the array here.
4306 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4307 eventfd_signal(t->entries[i].eventfd, 1);
4309 /* Update current_threshold */
4310 t->current_threshold = i - 1;
4315 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4318 __mem_cgroup_threshold(memcg, false);
4319 if (do_swap_account)
4320 __mem_cgroup_threshold(memcg, true);
4322 memcg = parent_mem_cgroup(memcg);
4326 static int compare_thresholds(const void *a, const void *b)
4328 const struct mem_cgroup_threshold *_a = a;
4329 const struct mem_cgroup_threshold *_b = b;
4331 return _a->threshold - _b->threshold;
4334 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4336 struct mem_cgroup_eventfd_list *ev;
4338 list_for_each_entry(ev, &memcg->oom_notify, list)
4339 eventfd_signal(ev->eventfd, 1);
4343 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4345 struct mem_cgroup *iter;
4347 for_each_mem_cgroup_tree(iter, memcg)
4348 mem_cgroup_oom_notify_cb(iter);
4351 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4352 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4354 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4355 struct mem_cgroup_thresholds *thresholds;
4356 struct mem_cgroup_threshold_ary *new;
4357 int type = MEMFILE_TYPE(cft->private);
4358 u64 threshold, usage;
4361 ret = res_counter_memparse_write_strategy(args, &threshold);
4365 mutex_lock(&memcg->thresholds_lock);
4368 thresholds = &memcg->thresholds;
4369 else if (type == _MEMSWAP)
4370 thresholds = &memcg->memsw_thresholds;
4374 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4376 /* Check if a threshold crossed before adding a new one */
4377 if (thresholds->primary)
4378 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4380 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4382 /* Allocate memory for new array of thresholds */
4383 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4391 /* Copy thresholds (if any) to new array */
4392 if (thresholds->primary) {
4393 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4394 sizeof(struct mem_cgroup_threshold));
4397 /* Add new threshold */
4398 new->entries[size - 1].eventfd = eventfd;
4399 new->entries[size - 1].threshold = threshold;
4401 /* Sort thresholds. Registering of new threshold isn't time-critical */
4402 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4403 compare_thresholds, NULL);
4405 /* Find current threshold */
4406 new->current_threshold = -1;
4407 for (i = 0; i < size; i++) {
4408 if (new->entries[i].threshold < usage) {
4410 * new->current_threshold will not be used until
4411 * rcu_assign_pointer(), so it's safe to increment
4414 ++new->current_threshold;
4418 /* Free old spare buffer and save old primary buffer as spare */
4419 kfree(thresholds->spare);
4420 thresholds->spare = thresholds->primary;
4422 rcu_assign_pointer(thresholds->primary, new);
4424 /* To be sure that nobody uses thresholds */
4428 mutex_unlock(&memcg->thresholds_lock);
4433 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4434 struct cftype *cft, struct eventfd_ctx *eventfd)
4436 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4437 struct mem_cgroup_thresholds *thresholds;
4438 struct mem_cgroup_threshold_ary *new;
4439 int type = MEMFILE_TYPE(cft->private);
4443 mutex_lock(&memcg->thresholds_lock);
4445 thresholds = &memcg->thresholds;
4446 else if (type == _MEMSWAP)
4447 thresholds = &memcg->memsw_thresholds;
4452 * Something went wrong if we trying to unregister a threshold
4453 * if we don't have thresholds
4455 BUG_ON(!thresholds);
4457 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4459 /* Check if a threshold crossed before removing */
4460 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4462 /* Calculate new number of threshold */
4464 for (i = 0; i < thresholds->primary->size; i++) {
4465 if (thresholds->primary->entries[i].eventfd != eventfd)
4469 new = thresholds->spare;
4471 /* Set thresholds array to NULL if we don't have thresholds */
4480 /* Copy thresholds and find current threshold */
4481 new->current_threshold = -1;
4482 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4483 if (thresholds->primary->entries[i].eventfd == eventfd)
4486 new->entries[j] = thresholds->primary->entries[i];
4487 if (new->entries[j].threshold < usage) {
4489 * new->current_threshold will not be used
4490 * until rcu_assign_pointer(), so it's safe to increment
4493 ++new->current_threshold;
4499 /* Swap primary and spare array */
4500 thresholds->spare = thresholds->primary;
4501 rcu_assign_pointer(thresholds->primary, new);
4503 /* To be sure that nobody uses thresholds */
4506 mutex_unlock(&memcg->thresholds_lock);
4509 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4510 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4512 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4513 struct mem_cgroup_eventfd_list *event;
4514 int type = MEMFILE_TYPE(cft->private);
4516 BUG_ON(type != _OOM_TYPE);
4517 event = kmalloc(sizeof(*event), GFP_KERNEL);
4521 spin_lock(&memcg_oom_lock);
4523 event->eventfd = eventfd;
4524 list_add(&event->list, &memcg->oom_notify);
4526 /* already in OOM ? */
4527 if (atomic_read(&memcg->under_oom))
4528 eventfd_signal(eventfd, 1);
4529 spin_unlock(&memcg_oom_lock);
4534 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4535 struct cftype *cft, struct eventfd_ctx *eventfd)
4537 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4538 struct mem_cgroup_eventfd_list *ev, *tmp;
4539 int type = MEMFILE_TYPE(cft->private);
4541 BUG_ON(type != _OOM_TYPE);
4543 spin_lock(&memcg_oom_lock);
4545 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4546 if (ev->eventfd == eventfd) {
4547 list_del(&ev->list);
4552 spin_unlock(&memcg_oom_lock);
4555 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4556 struct cftype *cft, struct cgroup_map_cb *cb)
4558 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4560 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4562 if (atomic_read(&memcg->under_oom))
4563 cb->fill(cb, "under_oom", 1);
4565 cb->fill(cb, "under_oom", 0);
4569 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4570 struct cftype *cft, u64 val)
4572 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4573 struct mem_cgroup *parent;
4575 /* cannot set to root cgroup and only 0 and 1 are allowed */
4576 if (!cgrp->parent || !((val == 0) || (val == 1)))
4579 parent = mem_cgroup_from_cont(cgrp->parent);
4582 /* oom-kill-disable is a flag for subhierarchy. */
4583 if ((parent->use_hierarchy) ||
4584 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4588 memcg->oom_kill_disable = val;
4590 memcg_oom_recover(memcg);
4596 static const struct file_operations mem_control_numa_stat_file_operations = {
4598 .llseek = seq_lseek,
4599 .release = single_release,
4602 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4604 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4606 file->f_op = &mem_control_numa_stat_file_operations;
4607 return single_open(file, mem_control_numa_stat_show, cont);
4609 #endif /* CONFIG_NUMA */
4611 static struct cftype mem_cgroup_files[] = {
4613 .name = "usage_in_bytes",
4614 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4615 .read_u64 = mem_cgroup_read,
4616 .register_event = mem_cgroup_usage_register_event,
4617 .unregister_event = mem_cgroup_usage_unregister_event,
4620 .name = "max_usage_in_bytes",
4621 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4622 .trigger = mem_cgroup_reset,
4623 .read_u64 = mem_cgroup_read,
4626 .name = "limit_in_bytes",
4627 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4628 .write_string = mem_cgroup_write,
4629 .read_u64 = mem_cgroup_read,
4632 .name = "soft_limit_in_bytes",
4633 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4634 .write_string = mem_cgroup_write,
4635 .read_u64 = mem_cgroup_read,
4639 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4640 .trigger = mem_cgroup_reset,
4641 .read_u64 = mem_cgroup_read,
4645 .read_map = mem_control_stat_show,
4648 .name = "force_empty",
4649 .trigger = mem_cgroup_force_empty_write,
4652 .name = "use_hierarchy",
4653 .write_u64 = mem_cgroup_hierarchy_write,
4654 .read_u64 = mem_cgroup_hierarchy_read,
4657 .name = "swappiness",
4658 .read_u64 = mem_cgroup_swappiness_read,
4659 .write_u64 = mem_cgroup_swappiness_write,
4662 .name = "move_charge_at_immigrate",
4663 .read_u64 = mem_cgroup_move_charge_read,
4664 .write_u64 = mem_cgroup_move_charge_write,
4667 .name = "oom_control",
4668 .read_map = mem_cgroup_oom_control_read,
4669 .write_u64 = mem_cgroup_oom_control_write,
4670 .register_event = mem_cgroup_oom_register_event,
4671 .unregister_event = mem_cgroup_oom_unregister_event,
4672 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4676 .name = "numa_stat",
4677 .open = mem_control_numa_stat_open,
4683 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4684 static struct cftype memsw_cgroup_files[] = {
4686 .name = "memsw.usage_in_bytes",
4687 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4688 .read_u64 = mem_cgroup_read,
4689 .register_event = mem_cgroup_usage_register_event,
4690 .unregister_event = mem_cgroup_usage_unregister_event,
4693 .name = "memsw.max_usage_in_bytes",
4694 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4695 .trigger = mem_cgroup_reset,
4696 .read_u64 = mem_cgroup_read,
4699 .name = "memsw.limit_in_bytes",
4700 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4701 .write_string = mem_cgroup_write,
4702 .read_u64 = mem_cgroup_read,
4705 .name = "memsw.failcnt",
4706 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4707 .trigger = mem_cgroup_reset,
4708 .read_u64 = mem_cgroup_read,
4712 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4714 if (!do_swap_account)
4716 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4717 ARRAY_SIZE(memsw_cgroup_files));
4720 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4726 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4728 struct mem_cgroup_per_node *pn;
4729 struct mem_cgroup_per_zone *mz;
4731 int zone, tmp = node;
4733 * This routine is called against possible nodes.
4734 * But it's BUG to call kmalloc() against offline node.
4736 * TODO: this routine can waste much memory for nodes which will
4737 * never be onlined. It's better to use memory hotplug callback
4740 if (!node_state(node, N_NORMAL_MEMORY))
4742 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4746 memcg->info.nodeinfo[node] = pn;
4747 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4748 mz = &pn->zoneinfo[zone];
4750 INIT_LIST_HEAD(&mz->lists[l]);
4751 mz->usage_in_excess = 0;
4752 mz->on_tree = false;
4758 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4760 kfree(memcg->info.nodeinfo[node]);
4763 static struct mem_cgroup *mem_cgroup_alloc(void)
4765 struct mem_cgroup *mem;
4766 int size = sizeof(struct mem_cgroup);
4768 /* Can be very big if MAX_NUMNODES is very big */
4769 if (size < PAGE_SIZE)
4770 mem = kzalloc(size, GFP_KERNEL);
4772 mem = vzalloc(size);
4777 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4780 spin_lock_init(&mem->pcp_counter_lock);
4784 if (size < PAGE_SIZE)
4792 * At destroying mem_cgroup, references from swap_cgroup can remain.
4793 * (scanning all at force_empty is too costly...)
4795 * Instead of clearing all references at force_empty, we remember
4796 * the number of reference from swap_cgroup and free mem_cgroup when
4797 * it goes down to 0.
4799 * Removal of cgroup itself succeeds regardless of refs from swap.
4802 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4806 mem_cgroup_remove_from_trees(memcg);
4807 free_css_id(&mem_cgroup_subsys, &memcg->css);
4809 for_each_node_state(node, N_POSSIBLE)
4810 free_mem_cgroup_per_zone_info(memcg, node);
4812 free_percpu(memcg->stat);
4813 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4819 static void mem_cgroup_get(struct mem_cgroup *memcg)
4821 atomic_inc(&memcg->refcnt);
4824 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4826 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4827 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4828 __mem_cgroup_free(memcg);
4830 mem_cgroup_put(parent);
4834 static void mem_cgroup_put(struct mem_cgroup *memcg)
4836 __mem_cgroup_put(memcg, 1);
4840 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4842 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4844 if (!memcg->res.parent)
4846 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4849 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4850 static void __init enable_swap_cgroup(void)
4852 if (!mem_cgroup_disabled() && really_do_swap_account)
4853 do_swap_account = 1;
4856 static void __init enable_swap_cgroup(void)
4861 static int mem_cgroup_soft_limit_tree_init(void)
4863 struct mem_cgroup_tree_per_node *rtpn;
4864 struct mem_cgroup_tree_per_zone *rtpz;
4865 int tmp, node, zone;
4867 for_each_node_state(node, N_POSSIBLE) {
4869 if (!node_state(node, N_NORMAL_MEMORY))
4871 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4875 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4877 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4878 rtpz = &rtpn->rb_tree_per_zone[zone];
4879 rtpz->rb_root = RB_ROOT;
4880 spin_lock_init(&rtpz->lock);
4886 static struct cgroup_subsys_state * __ref
4887 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4889 struct mem_cgroup *memcg, *parent;
4890 long error = -ENOMEM;
4893 memcg = mem_cgroup_alloc();
4895 return ERR_PTR(error);
4897 for_each_node_state(node, N_POSSIBLE)
4898 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4902 if (cont->parent == NULL) {
4904 enable_swap_cgroup();
4906 root_mem_cgroup = memcg;
4907 if (mem_cgroup_soft_limit_tree_init())
4909 for_each_possible_cpu(cpu) {
4910 struct memcg_stock_pcp *stock =
4911 &per_cpu(memcg_stock, cpu);
4912 INIT_WORK(&stock->work, drain_local_stock);
4914 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4916 parent = mem_cgroup_from_cont(cont->parent);
4917 memcg->use_hierarchy = parent->use_hierarchy;
4918 memcg->oom_kill_disable = parent->oom_kill_disable;
4921 if (parent && parent->use_hierarchy) {
4922 res_counter_init(&memcg->res, &parent->res);
4923 res_counter_init(&memcg->memsw, &parent->memsw);
4925 * We increment refcnt of the parent to ensure that we can
4926 * safely access it on res_counter_charge/uncharge.
4927 * This refcnt will be decremented when freeing this
4928 * mem_cgroup(see mem_cgroup_put).
4930 mem_cgroup_get(parent);
4932 res_counter_init(&memcg->res, NULL);
4933 res_counter_init(&memcg->memsw, NULL);
4935 memcg->last_scanned_child = 0;
4936 memcg->last_scanned_node = MAX_NUMNODES;
4937 INIT_LIST_HEAD(&memcg->oom_notify);
4940 memcg->swappiness = mem_cgroup_swappiness(parent);
4941 atomic_set(&memcg->refcnt, 1);
4942 memcg->move_charge_at_immigrate = 0;
4943 mutex_init(&memcg->thresholds_lock);
4946 __mem_cgroup_free(memcg);
4947 root_mem_cgroup = NULL;
4948 return ERR_PTR(error);
4951 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4952 struct cgroup *cont)
4954 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4956 return mem_cgroup_force_empty(memcg, false);
4959 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4960 struct cgroup *cont)
4962 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
4964 mem_cgroup_put(memcg);
4967 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4968 struct cgroup *cont)
4972 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4973 ARRAY_SIZE(mem_cgroup_files));
4976 ret = register_memsw_files(cont, ss);
4981 /* Handlers for move charge at task migration. */
4982 #define PRECHARGE_COUNT_AT_ONCE 256
4983 static int mem_cgroup_do_precharge(unsigned long count)
4986 int batch_count = PRECHARGE_COUNT_AT_ONCE;
4987 struct mem_cgroup *memcg = mc.to;
4989 if (mem_cgroup_is_root(memcg)) {
4990 mc.precharge += count;
4991 /* we don't need css_get for root */
4994 /* try to charge at once */
4996 struct res_counter *dummy;
4998 * "memcg" cannot be under rmdir() because we've already checked
4999 * by cgroup_lock_live_cgroup() that it is not removed and we
5000 * are still under the same cgroup_mutex. So we can postpone
5003 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5005 if (do_swap_account && res_counter_charge(&memcg->memsw,
5006 PAGE_SIZE * count, &dummy)) {
5007 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5010 mc.precharge += count;
5014 /* fall back to one by one charge */
5016 if (signal_pending(current)) {
5020 if (!batch_count--) {
5021 batch_count = PRECHARGE_COUNT_AT_ONCE;
5024 ret = __mem_cgroup_try_charge(NULL,
5025 GFP_KERNEL, 1, &memcg, false);
5027 /* mem_cgroup_clear_mc() will do uncharge later */
5035 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5036 * @vma: the vma the pte to be checked belongs
5037 * @addr: the address corresponding to the pte to be checked
5038 * @ptent: the pte to be checked
5039 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5042 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5043 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5044 * move charge. if @target is not NULL, the page is stored in target->page
5045 * with extra refcnt got(Callers should handle it).
5046 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5047 * target for charge migration. if @target is not NULL, the entry is stored
5050 * Called with pte lock held.
5057 enum mc_target_type {
5058 MC_TARGET_NONE, /* not used */
5063 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5064 unsigned long addr, pte_t ptent)
5066 struct page *page = vm_normal_page(vma, addr, ptent);
5068 if (!page || !page_mapped(page))
5070 if (PageAnon(page)) {
5071 /* we don't move shared anon */
5072 if (!move_anon() || page_mapcount(page) > 2)
5074 } else if (!move_file())
5075 /* we ignore mapcount for file pages */
5077 if (!get_page_unless_zero(page))
5083 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5084 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5087 struct page *page = NULL;
5088 swp_entry_t ent = pte_to_swp_entry(ptent);
5090 if (!move_anon() || non_swap_entry(ent))
5092 usage_count = mem_cgroup_count_swap_user(ent, &page);
5093 if (usage_count > 1) { /* we don't move shared anon */
5098 if (do_swap_account)
5099 entry->val = ent.val;
5104 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5105 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5107 struct page *page = NULL;
5108 struct inode *inode;
5109 struct address_space *mapping;
5112 if (!vma->vm_file) /* anonymous vma */
5117 inode = vma->vm_file->f_path.dentry->d_inode;
5118 mapping = vma->vm_file->f_mapping;
5119 if (pte_none(ptent))
5120 pgoff = linear_page_index(vma, addr);
5121 else /* pte_file(ptent) is true */
5122 pgoff = pte_to_pgoff(ptent);
5124 /* page is moved even if it's not RSS of this task(page-faulted). */
5125 page = find_get_page(mapping, pgoff);
5128 /* shmem/tmpfs may report page out on swap: account for that too. */
5129 if (radix_tree_exceptional_entry(page)) {
5130 swp_entry_t swap = radix_to_swp_entry(page);
5131 if (do_swap_account)
5133 page = find_get_page(&swapper_space, swap.val);
5139 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5140 unsigned long addr, pte_t ptent, union mc_target *target)
5142 struct page *page = NULL;
5143 struct page_cgroup *pc;
5145 swp_entry_t ent = { .val = 0 };
5147 if (pte_present(ptent))
5148 page = mc_handle_present_pte(vma, addr, ptent);
5149 else if (is_swap_pte(ptent))
5150 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5151 else if (pte_none(ptent) || pte_file(ptent))
5152 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5154 if (!page && !ent.val)
5157 pc = lookup_page_cgroup(page);
5159 * Do only loose check w/o page_cgroup lock.
5160 * mem_cgroup_move_account() checks the pc is valid or not under
5163 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5164 ret = MC_TARGET_PAGE;
5166 target->page = page;
5168 if (!ret || !target)
5171 /* There is a swap entry and a page doesn't exist or isn't charged */
5172 if (ent.val && !ret &&
5173 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5174 ret = MC_TARGET_SWAP;
5181 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5182 unsigned long addr, unsigned long end,
5183 struct mm_walk *walk)
5185 struct vm_area_struct *vma = walk->private;
5189 split_huge_page_pmd(walk->mm, pmd);
5191 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5192 for (; addr != end; pte++, addr += PAGE_SIZE)
5193 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5194 mc.precharge++; /* increment precharge temporarily */
5195 pte_unmap_unlock(pte - 1, ptl);
5201 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5203 unsigned long precharge;
5204 struct vm_area_struct *vma;
5206 down_read(&mm->mmap_sem);
5207 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5208 struct mm_walk mem_cgroup_count_precharge_walk = {
5209 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5213 if (is_vm_hugetlb_page(vma))
5215 walk_page_range(vma->vm_start, vma->vm_end,
5216 &mem_cgroup_count_precharge_walk);
5218 up_read(&mm->mmap_sem);
5220 precharge = mc.precharge;
5226 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5228 unsigned long precharge = mem_cgroup_count_precharge(mm);
5230 VM_BUG_ON(mc.moving_task);
5231 mc.moving_task = current;
5232 return mem_cgroup_do_precharge(precharge);
5235 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5236 static void __mem_cgroup_clear_mc(void)
5238 struct mem_cgroup *from = mc.from;
5239 struct mem_cgroup *to = mc.to;
5241 /* we must uncharge all the leftover precharges from mc.to */
5243 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5247 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5248 * we must uncharge here.
5250 if (mc.moved_charge) {
5251 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5252 mc.moved_charge = 0;
5254 /* we must fixup refcnts and charges */
5255 if (mc.moved_swap) {
5256 /* uncharge swap account from the old cgroup */
5257 if (!mem_cgroup_is_root(mc.from))
5258 res_counter_uncharge(&mc.from->memsw,
5259 PAGE_SIZE * mc.moved_swap);
5260 __mem_cgroup_put(mc.from, mc.moved_swap);
5262 if (!mem_cgroup_is_root(mc.to)) {
5264 * we charged both to->res and to->memsw, so we should
5267 res_counter_uncharge(&mc.to->res,
5268 PAGE_SIZE * mc.moved_swap);
5270 /* we've already done mem_cgroup_get(mc.to) */
5273 memcg_oom_recover(from);
5274 memcg_oom_recover(to);
5275 wake_up_all(&mc.waitq);
5278 static void mem_cgroup_clear_mc(void)
5280 struct mem_cgroup *from = mc.from;
5283 * we must clear moving_task before waking up waiters at the end of
5286 mc.moving_task = NULL;
5287 __mem_cgroup_clear_mc();
5288 spin_lock(&mc.lock);
5291 spin_unlock(&mc.lock);
5292 mem_cgroup_end_move(from);
5295 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5296 struct cgroup *cgroup,
5297 struct task_struct *p)
5300 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5302 if (memcg->move_charge_at_immigrate) {
5303 struct mm_struct *mm;
5304 struct mem_cgroup *from = mem_cgroup_from_task(p);
5306 VM_BUG_ON(from == memcg);
5308 mm = get_task_mm(p);
5311 /* We move charges only when we move a owner of the mm */
5312 if (mm->owner == p) {
5315 VM_BUG_ON(mc.precharge);
5316 VM_BUG_ON(mc.moved_charge);
5317 VM_BUG_ON(mc.moved_swap);
5318 mem_cgroup_start_move(from);
5319 spin_lock(&mc.lock);
5322 spin_unlock(&mc.lock);
5323 /* We set mc.moving_task later */
5325 ret = mem_cgroup_precharge_mc(mm);
5327 mem_cgroup_clear_mc();
5334 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5335 struct cgroup *cgroup,
5336 struct task_struct *p)
5338 mem_cgroup_clear_mc();
5341 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5342 unsigned long addr, unsigned long end,
5343 struct mm_walk *walk)
5346 struct vm_area_struct *vma = walk->private;
5350 split_huge_page_pmd(walk->mm, pmd);
5352 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5353 for (; addr != end; addr += PAGE_SIZE) {
5354 pte_t ptent = *(pte++);
5355 union mc_target target;
5358 struct page_cgroup *pc;
5364 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5366 case MC_TARGET_PAGE:
5368 if (isolate_lru_page(page))
5370 pc = lookup_page_cgroup(page);
5371 if (!mem_cgroup_move_account(page, 1, pc,
5372 mc.from, mc.to, false)) {
5374 /* we uncharge from mc.from later. */
5377 putback_lru_page(page);
5378 put: /* is_target_pte_for_mc() gets the page */
5381 case MC_TARGET_SWAP:
5383 if (!mem_cgroup_move_swap_account(ent,
5384 mc.from, mc.to, false)) {
5386 /* we fixup refcnts and charges later. */
5394 pte_unmap_unlock(pte - 1, ptl);
5399 * We have consumed all precharges we got in can_attach().
5400 * We try charge one by one, but don't do any additional
5401 * charges to mc.to if we have failed in charge once in attach()
5404 ret = mem_cgroup_do_precharge(1);
5412 static void mem_cgroup_move_charge(struct mm_struct *mm)
5414 struct vm_area_struct *vma;
5416 lru_add_drain_all();
5418 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5420 * Someone who are holding the mmap_sem might be waiting in
5421 * waitq. So we cancel all extra charges, wake up all waiters,
5422 * and retry. Because we cancel precharges, we might not be able
5423 * to move enough charges, but moving charge is a best-effort
5424 * feature anyway, so it wouldn't be a big problem.
5426 __mem_cgroup_clear_mc();
5430 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5432 struct mm_walk mem_cgroup_move_charge_walk = {
5433 .pmd_entry = mem_cgroup_move_charge_pte_range,
5437 if (is_vm_hugetlb_page(vma))
5439 ret = walk_page_range(vma->vm_start, vma->vm_end,
5440 &mem_cgroup_move_charge_walk);
5443 * means we have consumed all precharges and failed in
5444 * doing additional charge. Just abandon here.
5448 up_read(&mm->mmap_sem);
5451 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5452 struct cgroup *cont,
5453 struct cgroup *old_cont,
5454 struct task_struct *p)
5456 struct mm_struct *mm = get_task_mm(p);
5460 mem_cgroup_move_charge(mm);
5465 mem_cgroup_clear_mc();
5467 #else /* !CONFIG_MMU */
5468 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5469 struct cgroup *cgroup,
5470 struct task_struct *p)
5474 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5475 struct cgroup *cgroup,
5476 struct task_struct *p)
5479 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5480 struct cgroup *cont,
5481 struct cgroup *old_cont,
5482 struct task_struct *p)
5487 struct cgroup_subsys mem_cgroup_subsys = {
5489 .subsys_id = mem_cgroup_subsys_id,
5490 .create = mem_cgroup_create,
5491 .pre_destroy = mem_cgroup_pre_destroy,
5492 .destroy = mem_cgroup_destroy,
5493 .populate = mem_cgroup_populate,
5494 .can_attach = mem_cgroup_can_attach,
5495 .cancel_attach = mem_cgroup_cancel_attach,
5496 .attach = mem_cgroup_move_task,
5501 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5502 static int __init enable_swap_account(char *s)
5504 /* consider enabled if no parameter or 1 is given */
5505 if (!strcmp(s, "1"))
5506 really_do_swap_account = 1;
5507 else if (!strcmp(s, "0"))
5508 really_do_swap_account = 0;
5511 __setup("swapaccount=", enable_swap_account);