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)
726 /* threshold event is triggered in finer grain than soft limit */
727 if (unlikely(__memcg_event_check(memcg, MEM_CGROUP_TARGET_THRESH))) {
728 mem_cgroup_threshold(memcg);
729 __mem_cgroup_target_update(memcg, MEM_CGROUP_TARGET_THRESH);
730 if (unlikely(__memcg_event_check(memcg,
731 MEM_CGROUP_TARGET_SOFTLIMIT))) {
732 mem_cgroup_update_tree(memcg, page);
733 __mem_cgroup_target_update(memcg,
734 MEM_CGROUP_TARGET_SOFTLIMIT);
737 if (unlikely(__memcg_event_check(memcg,
738 MEM_CGROUP_TARGET_NUMAINFO))) {
739 atomic_inc(&memcg->numainfo_events);
740 __mem_cgroup_target_update(memcg,
741 MEM_CGROUP_TARGET_NUMAINFO);
748 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
750 return container_of(cgroup_subsys_state(cont,
751 mem_cgroup_subsys_id), struct mem_cgroup,
755 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
758 * mm_update_next_owner() may clear mm->owner to NULL
759 * if it races with swapoff, page migration, etc.
760 * So this can be called with p == NULL.
765 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
766 struct mem_cgroup, css);
769 struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
771 struct mem_cgroup *memcg = NULL;
776 * Because we have no locks, mm->owner's may be being moved to other
777 * cgroup. We use css_tryget() here even if this looks
778 * pessimistic (rather than adding locks here).
782 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
783 if (unlikely(!memcg))
785 } while (!css_tryget(&memcg->css));
790 /* The caller has to guarantee "mem" exists before calling this */
791 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *memcg)
793 struct cgroup_subsys_state *css;
796 if (!memcg) /* ROOT cgroup has the smallest ID */
797 return root_mem_cgroup; /*css_put/get against root is ignored*/
798 if (!memcg->use_hierarchy) {
799 if (css_tryget(&memcg->css))
805 * searching a memory cgroup which has the smallest ID under given
806 * ROOT cgroup. (ID >= 1)
808 css = css_get_next(&mem_cgroup_subsys, 1, &memcg->css, &found);
809 if (css && css_tryget(css))
810 memcg = container_of(css, struct mem_cgroup, css);
817 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
818 struct mem_cgroup *root,
821 int nextid = css_id(&iter->css) + 1;
824 struct cgroup_subsys_state *css;
826 hierarchy_used = iter->use_hierarchy;
829 /* If no ROOT, walk all, ignore hierarchy */
830 if (!cond || (root && !hierarchy_used))
834 root = root_mem_cgroup;
840 css = css_get_next(&mem_cgroup_subsys, nextid,
842 if (css && css_tryget(css))
843 iter = container_of(css, struct mem_cgroup, css);
845 /* If css is NULL, no more cgroups will be found */
847 } while (css && !iter);
852 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
853 * be careful that "break" loop is not allowed. We have reference count.
854 * Instead of that modify "cond" to be false and "continue" to exit the loop.
856 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
857 for (iter = mem_cgroup_start_loop(root);\
859 iter = mem_cgroup_get_next(iter, root, cond))
861 #define for_each_mem_cgroup_tree(iter, root) \
862 for_each_mem_cgroup_tree_cond(iter, root, true)
864 #define for_each_mem_cgroup_all(iter) \
865 for_each_mem_cgroup_tree_cond(iter, NULL, true)
868 static inline bool mem_cgroup_is_root(struct mem_cgroup *memcg)
870 return (memcg == root_mem_cgroup);
873 void mem_cgroup_count_vm_event(struct mm_struct *mm, enum vm_event_item idx)
875 struct mem_cgroup *memcg;
881 memcg = mem_cgroup_from_task(rcu_dereference(mm->owner));
882 if (unlikely(!memcg))
887 mem_cgroup_pgmajfault(memcg, 1);
890 mem_cgroup_pgfault(memcg, 1);
898 EXPORT_SYMBOL(mem_cgroup_count_vm_event);
901 * Following LRU functions are allowed to be used without PCG_LOCK.
902 * Operations are called by routine of global LRU independently from memcg.
903 * What we have to take care of here is validness of pc->mem_cgroup.
905 * Changes to pc->mem_cgroup happens when
908 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
909 * It is added to LRU before charge.
910 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
911 * When moving account, the page is not on LRU. It's isolated.
914 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
916 struct page_cgroup *pc;
917 struct mem_cgroup_per_zone *mz;
919 if (mem_cgroup_disabled())
921 pc = lookup_page_cgroup(page);
922 /* can happen while we handle swapcache. */
923 if (!TestClearPageCgroupAcctLRU(pc))
925 VM_BUG_ON(!pc->mem_cgroup);
927 * We don't check PCG_USED bit. It's cleared when the "page" is finally
928 * removed from global LRU.
930 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
931 /* huge page split is done under lru_lock. so, we have no races. */
932 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
933 if (mem_cgroup_is_root(pc->mem_cgroup))
935 VM_BUG_ON(list_empty(&pc->lru));
936 list_del_init(&pc->lru);
939 void mem_cgroup_del_lru(struct page *page)
941 mem_cgroup_del_lru_list(page, page_lru(page));
945 * Writeback is about to end against a page which has been marked for immediate
946 * reclaim. If it still appears to be reclaimable, move it to the tail of the
949 void mem_cgroup_rotate_reclaimable_page(struct page *page)
951 struct mem_cgroup_per_zone *mz;
952 struct page_cgroup *pc;
953 enum lru_list lru = page_lru(page);
955 if (mem_cgroup_disabled())
958 pc = lookup_page_cgroup(page);
959 /* unused or root page is not rotated. */
960 if (!PageCgroupUsed(pc))
962 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
964 if (mem_cgroup_is_root(pc->mem_cgroup))
966 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
967 list_move_tail(&pc->lru, &mz->lists[lru]);
970 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
972 struct mem_cgroup_per_zone *mz;
973 struct page_cgroup *pc;
975 if (mem_cgroup_disabled())
978 pc = lookup_page_cgroup(page);
979 /* unused or root page is not rotated. */
980 if (!PageCgroupUsed(pc))
982 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
984 if (mem_cgroup_is_root(pc->mem_cgroup))
986 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
987 list_move(&pc->lru, &mz->lists[lru]);
990 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
992 struct page_cgroup *pc;
993 struct mem_cgroup_per_zone *mz;
995 if (mem_cgroup_disabled())
997 pc = lookup_page_cgroup(page);
998 VM_BUG_ON(PageCgroupAcctLRU(pc));
1001 * SetPageLRU SetPageCgroupUsed
1003 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1005 * Ensure that one of the two sides adds the page to the memcg
1006 * LRU during a race.
1009 if (!PageCgroupUsed(pc))
1011 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1013 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1014 /* huge page split is done under lru_lock. so, we have no races. */
1015 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
1016 SetPageCgroupAcctLRU(pc);
1017 if (mem_cgroup_is_root(pc->mem_cgroup))
1019 list_add(&pc->lru, &mz->lists[lru]);
1023 * At handling SwapCache and other FUSE stuff, pc->mem_cgroup may be changed
1024 * while it's linked to lru because the page may be reused after it's fully
1025 * uncharged. To handle that, unlink page_cgroup from LRU when charge it again.
1026 * It's done under lock_page and expected that zone->lru_lock isnever held.
1028 static void mem_cgroup_lru_del_before_commit(struct page *page)
1030 unsigned long flags;
1031 struct zone *zone = page_zone(page);
1032 struct page_cgroup *pc = lookup_page_cgroup(page);
1035 * Doing this check without taking ->lru_lock seems wrong but this
1036 * is safe. Because if page_cgroup's USED bit is unset, the page
1037 * will not be added to any memcg's LRU. If page_cgroup's USED bit is
1038 * set, the commit after this will fail, anyway.
1039 * This all charge/uncharge is done under some mutual execustion.
1040 * So, we don't need to taking care of changes in USED bit.
1042 if (likely(!PageLRU(page)))
1045 spin_lock_irqsave(&zone->lru_lock, flags);
1047 * Forget old LRU when this page_cgroup is *not* used. This Used bit
1048 * is guarded by lock_page() because the page is SwapCache.
1050 if (!PageCgroupUsed(pc))
1051 mem_cgroup_del_lru_list(page, page_lru(page));
1052 spin_unlock_irqrestore(&zone->lru_lock, flags);
1055 static void mem_cgroup_lru_add_after_commit(struct page *page)
1057 unsigned long flags;
1058 struct zone *zone = page_zone(page);
1059 struct page_cgroup *pc = lookup_page_cgroup(page);
1062 * SetPageLRU SetPageCgroupUsed
1064 * PageCgroupUsed && add to memcg LRU PageLRU && add to memcg LRU
1066 * Ensure that one of the two sides adds the page to the memcg
1067 * LRU during a race.
1070 /* taking care of that the page is added to LRU while we commit it */
1071 if (likely(!PageLRU(page)))
1073 spin_lock_irqsave(&zone->lru_lock, flags);
1074 /* link when the page is linked to LRU but page_cgroup isn't */
1075 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
1076 mem_cgroup_add_lru_list(page, page_lru(page));
1077 spin_unlock_irqrestore(&zone->lru_lock, flags);
1081 void mem_cgroup_move_lists(struct page *page,
1082 enum lru_list from, enum lru_list to)
1084 if (mem_cgroup_disabled())
1086 mem_cgroup_del_lru_list(page, from);
1087 mem_cgroup_add_lru_list(page, to);
1091 * Checks whether given mem is same or in the root_mem_cgroup's
1094 static bool mem_cgroup_same_or_subtree(const struct mem_cgroup *root_memcg,
1095 struct mem_cgroup *memcg)
1097 if (root_memcg != memcg) {
1098 return (root_memcg->use_hierarchy &&
1099 css_is_ancestor(&memcg->css, &root_memcg->css));
1105 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *memcg)
1108 struct mem_cgroup *curr = NULL;
1109 struct task_struct *p;
1111 p = find_lock_task_mm(task);
1114 curr = try_get_mem_cgroup_from_mm(p->mm);
1119 * We should check use_hierarchy of "memcg" not "curr". Because checking
1120 * use_hierarchy of "curr" here make this function true if hierarchy is
1121 * enabled in "curr" and "curr" is a child of "memcg" in *cgroup*
1122 * hierarchy(even if use_hierarchy is disabled in "memcg").
1124 ret = mem_cgroup_same_or_subtree(memcg, curr);
1125 css_put(&curr->css);
1129 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg, struct zone *zone)
1131 unsigned long inactive_ratio;
1132 int nid = zone_to_nid(zone);
1133 int zid = zone_idx(zone);
1134 unsigned long inactive;
1135 unsigned long active;
1138 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1139 BIT(LRU_INACTIVE_ANON));
1140 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1141 BIT(LRU_ACTIVE_ANON));
1143 gb = (inactive + active) >> (30 - PAGE_SHIFT);
1145 inactive_ratio = int_sqrt(10 * gb);
1149 return inactive * inactive_ratio < active;
1152 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg, struct zone *zone)
1154 unsigned long active;
1155 unsigned long inactive;
1156 int zid = zone_idx(zone);
1157 int nid = zone_to_nid(zone);
1159 inactive = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1160 BIT(LRU_INACTIVE_FILE));
1161 active = mem_cgroup_zone_nr_lru_pages(memcg, nid, zid,
1162 BIT(LRU_ACTIVE_FILE));
1164 return (active > inactive);
1167 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1170 int nid = zone_to_nid(zone);
1171 int zid = zone_idx(zone);
1172 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1174 return &mz->reclaim_stat;
1177 struct zone_reclaim_stat *
1178 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1180 struct page_cgroup *pc;
1181 struct mem_cgroup_per_zone *mz;
1183 if (mem_cgroup_disabled())
1186 pc = lookup_page_cgroup(page);
1187 if (!PageCgroupUsed(pc))
1189 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1191 mz = page_cgroup_zoneinfo(pc->mem_cgroup, page);
1192 return &mz->reclaim_stat;
1195 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1196 struct list_head *dst,
1197 unsigned long *scanned, int order,
1198 isolate_mode_t mode,
1200 struct mem_cgroup *mem_cont,
1201 int active, int file)
1203 unsigned long nr_taken = 0;
1207 struct list_head *src;
1208 struct page_cgroup *pc, *tmp;
1209 int nid = zone_to_nid(z);
1210 int zid = zone_idx(z);
1211 struct mem_cgroup_per_zone *mz;
1212 int lru = LRU_FILE * file + active;
1216 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1217 src = &mz->lists[lru];
1220 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1221 if (scan >= nr_to_scan)
1224 if (unlikely(!PageCgroupUsed(pc)))
1227 page = lookup_cgroup_page(pc);
1229 if (unlikely(!PageLRU(page)))
1233 ret = __isolate_lru_page(page, mode, file);
1236 list_move(&page->lru, dst);
1237 mem_cgroup_del_lru(page);
1238 nr_taken += hpage_nr_pages(page);
1241 /* we don't affect global LRU but rotate in our LRU */
1242 mem_cgroup_rotate_lru_list(page, page_lru(page));
1251 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1257 #define mem_cgroup_from_res_counter(counter, member) \
1258 container_of(counter, struct mem_cgroup, member)
1261 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1262 * @mem: the memory cgroup
1264 * Returns the maximum amount of memory @mem can be charged with, in
1267 static unsigned long mem_cgroup_margin(struct mem_cgroup *memcg)
1269 unsigned long long margin;
1271 margin = res_counter_margin(&memcg->res);
1272 if (do_swap_account)
1273 margin = min(margin, res_counter_margin(&memcg->memsw));
1274 return margin >> PAGE_SHIFT;
1277 int mem_cgroup_swappiness(struct mem_cgroup *memcg)
1279 struct cgroup *cgrp = memcg->css.cgroup;
1282 if (cgrp->parent == NULL)
1283 return vm_swappiness;
1285 return memcg->swappiness;
1288 static void mem_cgroup_start_move(struct mem_cgroup *memcg)
1293 spin_lock(&memcg->pcp_counter_lock);
1294 for_each_online_cpu(cpu)
1295 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1296 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1297 spin_unlock(&memcg->pcp_counter_lock);
1303 static void mem_cgroup_end_move(struct mem_cgroup *memcg)
1310 spin_lock(&memcg->pcp_counter_lock);
1311 for_each_online_cpu(cpu)
1312 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1313 memcg->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1314 spin_unlock(&memcg->pcp_counter_lock);
1318 * 2 routines for checking "mem" is under move_account() or not.
1320 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1321 * for avoiding race in accounting. If true,
1322 * pc->mem_cgroup may be overwritten.
1324 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1325 * under hierarchy of moving cgroups. This is for
1326 * waiting at hith-memory prressure caused by "move".
1329 static bool mem_cgroup_stealed(struct mem_cgroup *memcg)
1331 VM_BUG_ON(!rcu_read_lock_held());
1332 return this_cpu_read(memcg->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1335 static bool mem_cgroup_under_move(struct mem_cgroup *memcg)
1337 struct mem_cgroup *from;
1338 struct mem_cgroup *to;
1341 * Unlike task_move routines, we access mc.to, mc.from not under
1342 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1344 spin_lock(&mc.lock);
1350 ret = mem_cgroup_same_or_subtree(memcg, from)
1351 || mem_cgroup_same_or_subtree(memcg, to);
1353 spin_unlock(&mc.lock);
1357 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *memcg)
1359 if (mc.moving_task && current != mc.moving_task) {
1360 if (mem_cgroup_under_move(memcg)) {
1362 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1363 /* moving charge context might have finished. */
1366 finish_wait(&mc.waitq, &wait);
1374 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1375 * @memcg: The memory cgroup that went over limit
1376 * @p: Task that is going to be killed
1378 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1381 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1383 struct cgroup *task_cgrp;
1384 struct cgroup *mem_cgrp;
1386 * Need a buffer in BSS, can't rely on allocations. The code relies
1387 * on the assumption that OOM is serialized for memory controller.
1388 * If this assumption is broken, revisit this code.
1390 static char memcg_name[PATH_MAX];
1399 mem_cgrp = memcg->css.cgroup;
1400 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1402 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1405 * Unfortunately, we are unable to convert to a useful name
1406 * But we'll still print out the usage information
1413 printk(KERN_INFO "Task in %s killed", memcg_name);
1416 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1424 * Continues from above, so we don't need an KERN_ level
1426 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1429 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1430 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1431 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1432 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1433 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1435 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1436 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1437 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1441 * This function returns the number of memcg under hierarchy tree. Returns
1442 * 1(self count) if no children.
1444 static int mem_cgroup_count_children(struct mem_cgroup *memcg)
1447 struct mem_cgroup *iter;
1449 for_each_mem_cgroup_tree(iter, memcg)
1455 * Return the memory (and swap, if configured) limit for a memcg.
1457 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1461 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1464 * Do not consider swap space if we cannot swap due to swappiness
1466 if (mem_cgroup_swappiness(memcg)) {
1469 limit += total_swap_pages << PAGE_SHIFT;
1470 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1473 * If memsw is finite and limits the amount of swap space
1474 * available to this memcg, return that limit.
1476 limit = min(limit, memsw);
1483 * Visit the first child (need not be the first child as per the ordering
1484 * of the cgroup list, since we track last_scanned_child) of @mem and use
1485 * that to reclaim free pages from.
1487 static struct mem_cgroup *
1488 mem_cgroup_select_victim(struct mem_cgroup *root_memcg)
1490 struct mem_cgroup *ret = NULL;
1491 struct cgroup_subsys_state *css;
1494 if (!root_memcg->use_hierarchy) {
1495 css_get(&root_memcg->css);
1501 nextid = root_memcg->last_scanned_child + 1;
1502 css = css_get_next(&mem_cgroup_subsys, nextid, &root_memcg->css,
1504 if (css && css_tryget(css))
1505 ret = container_of(css, struct mem_cgroup, css);
1508 /* Updates scanning parameter */
1510 /* this means start scan from ID:1 */
1511 root_memcg->last_scanned_child = 0;
1513 root_memcg->last_scanned_child = found;
1520 * test_mem_cgroup_node_reclaimable
1521 * @mem: the target memcg
1522 * @nid: the node ID to be checked.
1523 * @noswap : specify true here if the user wants flle only information.
1525 * This function returns whether the specified memcg contains any
1526 * reclaimable pages on a node. Returns true if there are any reclaimable
1527 * pages in the node.
1529 static bool test_mem_cgroup_node_reclaimable(struct mem_cgroup *memcg,
1530 int nid, bool noswap)
1532 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_FILE))
1534 if (noswap || !total_swap_pages)
1536 if (mem_cgroup_node_nr_lru_pages(memcg, nid, LRU_ALL_ANON))
1541 #if MAX_NUMNODES > 1
1544 * Always updating the nodemask is not very good - even if we have an empty
1545 * list or the wrong list here, we can start from some node and traverse all
1546 * nodes based on the zonelist. So update the list loosely once per 10 secs.
1549 static void mem_cgroup_may_update_nodemask(struct mem_cgroup *memcg)
1553 * numainfo_events > 0 means there was at least NUMAINFO_EVENTS_TARGET
1554 * pagein/pageout changes since the last update.
1556 if (!atomic_read(&memcg->numainfo_events))
1558 if (atomic_inc_return(&memcg->numainfo_updating) > 1)
1561 /* make a nodemask where this memcg uses memory from */
1562 memcg->scan_nodes = node_states[N_HIGH_MEMORY];
1564 for_each_node_mask(nid, node_states[N_HIGH_MEMORY]) {
1566 if (!test_mem_cgroup_node_reclaimable(memcg, nid, false))
1567 node_clear(nid, memcg->scan_nodes);
1570 atomic_set(&memcg->numainfo_events, 0);
1571 atomic_set(&memcg->numainfo_updating, 0);
1575 * Selecting a node where we start reclaim from. Because what we need is just
1576 * reducing usage counter, start from anywhere is O,K. Considering
1577 * memory reclaim from current node, there are pros. and cons.
1579 * Freeing memory from current node means freeing memory from a node which
1580 * we'll use or we've used. So, it may make LRU bad. And if several threads
1581 * hit limits, it will see a contention on a node. But freeing from remote
1582 * node means more costs for memory reclaim because of memory latency.
1584 * Now, we use round-robin. Better algorithm is welcomed.
1586 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1590 mem_cgroup_may_update_nodemask(memcg);
1591 node = memcg->last_scanned_node;
1593 node = next_node(node, memcg->scan_nodes);
1594 if (node == MAX_NUMNODES)
1595 node = first_node(memcg->scan_nodes);
1597 * We call this when we hit limit, not when pages are added to LRU.
1598 * No LRU may hold pages because all pages are UNEVICTABLE or
1599 * memcg is too small and all pages are not on LRU. In that case,
1600 * we use curret node.
1602 if (unlikely(node == MAX_NUMNODES))
1603 node = numa_node_id();
1605 memcg->last_scanned_node = node;
1610 * Check all nodes whether it contains reclaimable pages or not.
1611 * For quick scan, we make use of scan_nodes. This will allow us to skip
1612 * unused nodes. But scan_nodes is lazily updated and may not cotain
1613 * enough new information. We need to do double check.
1615 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1620 * quick check...making use of scan_node.
1621 * We can skip unused nodes.
1623 if (!nodes_empty(memcg->scan_nodes)) {
1624 for (nid = first_node(memcg->scan_nodes);
1626 nid = next_node(nid, memcg->scan_nodes)) {
1628 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1633 * Check rest of nodes.
1635 for_each_node_state(nid, N_HIGH_MEMORY) {
1636 if (node_isset(nid, memcg->scan_nodes))
1638 if (test_mem_cgroup_node_reclaimable(memcg, nid, noswap))
1645 int mem_cgroup_select_victim_node(struct mem_cgroup *memcg)
1650 bool mem_cgroup_reclaimable(struct mem_cgroup *memcg, bool noswap)
1652 return test_mem_cgroup_node_reclaimable(memcg, 0, noswap);
1657 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1658 * we reclaimed from, so that we don't end up penalizing one child extensively
1659 * based on its position in the children list.
1661 * root_memcg is the original ancestor that we've been reclaim from.
1663 * We give up and return to the caller when we visit root_memcg twice.
1664 * (other groups can be removed while we're walking....)
1666 * If shrink==true, for avoiding to free too much, this returns immedieately.
1668 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_memcg,
1671 unsigned long reclaim_options,
1672 unsigned long *total_scanned)
1674 struct mem_cgroup *victim;
1677 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1678 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1679 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1680 unsigned long excess;
1681 unsigned long nr_scanned;
1683 excess = res_counter_soft_limit_excess(&root_memcg->res) >> PAGE_SHIFT;
1685 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1686 if (!check_soft && !shrink && root_memcg->memsw_is_minimum)
1690 victim = mem_cgroup_select_victim(root_memcg);
1691 if (victim == root_memcg) {
1694 * We are not draining per cpu cached charges during
1695 * soft limit reclaim because global reclaim doesn't
1696 * care about charges. It tries to free some memory and
1697 * charges will not give any.
1699 if (!check_soft && loop >= 1)
1700 drain_all_stock_async(root_memcg);
1703 * If we have not been able to reclaim
1704 * anything, it might because there are
1705 * no reclaimable pages under this hierarchy
1707 if (!check_soft || !total) {
1708 css_put(&victim->css);
1712 * We want to do more targeted reclaim.
1713 * excess >> 2 is not to excessive so as to
1714 * reclaim too much, nor too less that we keep
1715 * coming back to reclaim from this cgroup
1717 if (total >= (excess >> 2) ||
1718 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1719 css_put(&victim->css);
1724 if (!mem_cgroup_reclaimable(victim, noswap)) {
1725 /* this cgroup's local usage == 0 */
1726 css_put(&victim->css);
1729 /* we use swappiness of local cgroup */
1731 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1732 noswap, zone, &nr_scanned);
1733 *total_scanned += nr_scanned;
1735 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1737 css_put(&victim->css);
1739 * At shrinking usage, we can't check we should stop here or
1740 * reclaim more. It's depends on callers. last_scanned_child
1741 * will work enough for keeping fairness under tree.
1747 if (!res_counter_soft_limit_excess(&root_memcg->res))
1749 } else if (mem_cgroup_margin(root_memcg))
1756 * Check OOM-Killer is already running under our hierarchy.
1757 * If someone is running, return false.
1758 * Has to be called with memcg_oom_lock
1760 static bool mem_cgroup_oom_lock(struct mem_cgroup *memcg)
1762 struct mem_cgroup *iter, *failed = NULL;
1765 for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1766 if (iter->oom_lock) {
1768 * this subtree of our hierarchy is already locked
1769 * so we cannot give a lock.
1774 iter->oom_lock = true;
1781 * OK, we failed to lock the whole subtree so we have to clean up
1782 * what we set up to the failing subtree
1785 for_each_mem_cgroup_tree_cond(iter, memcg, cond) {
1786 if (iter == failed) {
1790 iter->oom_lock = false;
1796 * Has to be called with memcg_oom_lock
1798 static int mem_cgroup_oom_unlock(struct mem_cgroup *memcg)
1800 struct mem_cgroup *iter;
1802 for_each_mem_cgroup_tree(iter, memcg)
1803 iter->oom_lock = false;
1807 static void mem_cgroup_mark_under_oom(struct mem_cgroup *memcg)
1809 struct mem_cgroup *iter;
1811 for_each_mem_cgroup_tree(iter, memcg)
1812 atomic_inc(&iter->under_oom);
1815 static void mem_cgroup_unmark_under_oom(struct mem_cgroup *memcg)
1817 struct mem_cgroup *iter;
1820 * When a new child is created while the hierarchy is under oom,
1821 * mem_cgroup_oom_lock() may not be called. We have to use
1822 * atomic_add_unless() here.
1824 for_each_mem_cgroup_tree(iter, memcg)
1825 atomic_add_unless(&iter->under_oom, -1, 0);
1828 static DEFINE_SPINLOCK(memcg_oom_lock);
1829 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1831 struct oom_wait_info {
1832 struct mem_cgroup *mem;
1836 static int memcg_oom_wake_function(wait_queue_t *wait,
1837 unsigned mode, int sync, void *arg)
1839 struct mem_cgroup *wake_memcg = (struct mem_cgroup *)arg,
1841 struct oom_wait_info *oom_wait_info;
1843 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1844 oom_wait_memcg = oom_wait_info->mem;
1847 * Both of oom_wait_info->mem and wake_mem are stable under us.
1848 * Then we can use css_is_ancestor without taking care of RCU.
1850 if (!mem_cgroup_same_or_subtree(oom_wait_memcg, wake_memcg)
1851 && !mem_cgroup_same_or_subtree(wake_memcg, oom_wait_memcg))
1853 return autoremove_wake_function(wait, mode, sync, arg);
1856 static void memcg_wakeup_oom(struct mem_cgroup *memcg)
1858 /* for filtering, pass "memcg" as argument. */
1859 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, memcg);
1862 static void memcg_oom_recover(struct mem_cgroup *memcg)
1864 if (memcg && atomic_read(&memcg->under_oom))
1865 memcg_wakeup_oom(memcg);
1869 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1871 bool mem_cgroup_handle_oom(struct mem_cgroup *memcg, gfp_t mask)
1873 struct oom_wait_info owait;
1874 bool locked, need_to_kill;
1877 owait.wait.flags = 0;
1878 owait.wait.func = memcg_oom_wake_function;
1879 owait.wait.private = current;
1880 INIT_LIST_HEAD(&owait.wait.task_list);
1881 need_to_kill = true;
1882 mem_cgroup_mark_under_oom(memcg);
1884 /* At first, try to OOM lock hierarchy under memcg.*/
1885 spin_lock(&memcg_oom_lock);
1886 locked = mem_cgroup_oom_lock(memcg);
1888 * Even if signal_pending(), we can't quit charge() loop without
1889 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1890 * under OOM is always welcomed, use TASK_KILLABLE here.
1892 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1893 if (!locked || memcg->oom_kill_disable)
1894 need_to_kill = false;
1896 mem_cgroup_oom_notify(memcg);
1897 spin_unlock(&memcg_oom_lock);
1900 finish_wait(&memcg_oom_waitq, &owait.wait);
1901 mem_cgroup_out_of_memory(memcg, mask);
1904 finish_wait(&memcg_oom_waitq, &owait.wait);
1906 spin_lock(&memcg_oom_lock);
1908 mem_cgroup_oom_unlock(memcg);
1909 memcg_wakeup_oom(memcg);
1910 spin_unlock(&memcg_oom_lock);
1912 mem_cgroup_unmark_under_oom(memcg);
1914 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1916 /* Give chance to dying process */
1917 schedule_timeout_uninterruptible(1);
1922 * Currently used to update mapped file statistics, but the routine can be
1923 * generalized to update other statistics as well.
1925 * Notes: Race condition
1927 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1928 * it tends to be costly. But considering some conditions, we doesn't need
1929 * to do so _always_.
1931 * Considering "charge", lock_page_cgroup() is not required because all
1932 * file-stat operations happen after a page is attached to radix-tree. There
1933 * are no race with "charge".
1935 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1936 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1937 * if there are race with "uncharge". Statistics itself is properly handled
1940 * Considering "move", this is an only case we see a race. To make the race
1941 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1942 * possibility of race condition. If there is, we take a lock.
1945 void mem_cgroup_update_page_stat(struct page *page,
1946 enum mem_cgroup_page_stat_item idx, int val)
1948 struct mem_cgroup *memcg;
1949 struct page_cgroup *pc = lookup_page_cgroup(page);
1950 bool need_unlock = false;
1951 unsigned long uninitialized_var(flags);
1957 memcg = pc->mem_cgroup;
1958 if (unlikely(!memcg || !PageCgroupUsed(pc)))
1960 /* pc->mem_cgroup is unstable ? */
1961 if (unlikely(mem_cgroup_stealed(memcg)) || PageTransHuge(page)) {
1962 /* take a lock against to access pc->mem_cgroup */
1963 move_lock_page_cgroup(pc, &flags);
1965 memcg = pc->mem_cgroup;
1966 if (!memcg || !PageCgroupUsed(pc))
1971 case MEMCG_NR_FILE_MAPPED:
1973 SetPageCgroupFileMapped(pc);
1974 else if (!page_mapped(page))
1975 ClearPageCgroupFileMapped(pc);
1976 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1982 this_cpu_add(memcg->stat->count[idx], val);
1985 if (unlikely(need_unlock))
1986 move_unlock_page_cgroup(pc, &flags);
1990 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1993 * size of first charge trial. "32" comes from vmscan.c's magic value.
1994 * TODO: maybe necessary to use big numbers in big irons.
1996 #define CHARGE_BATCH 32U
1997 struct memcg_stock_pcp {
1998 struct mem_cgroup *cached; /* this never be root cgroup */
1999 unsigned int nr_pages;
2000 struct work_struct work;
2001 unsigned long flags;
2002 #define FLUSHING_CACHED_CHARGE (0)
2004 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
2005 static DEFINE_MUTEX(percpu_charge_mutex);
2008 * Try to consume stocked charge on this cpu. If success, one page is consumed
2009 * from local stock and true is returned. If the stock is 0 or charges from a
2010 * cgroup which is not current target, returns false. This stock will be
2013 static bool consume_stock(struct mem_cgroup *memcg)
2015 struct memcg_stock_pcp *stock;
2018 stock = &get_cpu_var(memcg_stock);
2019 if (memcg == stock->cached && stock->nr_pages)
2021 else /* need to call res_counter_charge */
2023 put_cpu_var(memcg_stock);
2028 * Returns stocks cached in percpu to res_counter and reset cached information.
2030 static void drain_stock(struct memcg_stock_pcp *stock)
2032 struct mem_cgroup *old = stock->cached;
2034 if (stock->nr_pages) {
2035 unsigned long bytes = stock->nr_pages * PAGE_SIZE;
2037 res_counter_uncharge(&old->res, bytes);
2038 if (do_swap_account)
2039 res_counter_uncharge(&old->memsw, bytes);
2040 stock->nr_pages = 0;
2042 stock->cached = NULL;
2046 * This must be called under preempt disabled or must be called by
2047 * a thread which is pinned to local cpu.
2049 static void drain_local_stock(struct work_struct *dummy)
2051 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
2053 clear_bit(FLUSHING_CACHED_CHARGE, &stock->flags);
2057 * Cache charges(val) which is from res_counter, to local per_cpu area.
2058 * This will be consumed by consume_stock() function, later.
2060 static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
2062 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
2064 if (stock->cached != memcg) { /* reset if necessary */
2066 stock->cached = memcg;
2068 stock->nr_pages += nr_pages;
2069 put_cpu_var(memcg_stock);
2073 * Drains all per-CPU charge caches for given root_memcg resp. subtree
2074 * of the hierarchy under it. sync flag says whether we should block
2075 * until the work is done.
2077 static void drain_all_stock(struct mem_cgroup *root_memcg, bool sync)
2081 /* Notify other cpus that system-wide "drain" is running */
2084 for_each_online_cpu(cpu) {
2085 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2086 struct mem_cgroup *memcg;
2088 memcg = stock->cached;
2089 if (!memcg || !stock->nr_pages)
2091 if (!mem_cgroup_same_or_subtree(root_memcg, memcg))
2093 if (!test_and_set_bit(FLUSHING_CACHED_CHARGE, &stock->flags)) {
2095 drain_local_stock(&stock->work);
2097 schedule_work_on(cpu, &stock->work);
2105 for_each_online_cpu(cpu) {
2106 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
2107 if (test_bit(FLUSHING_CACHED_CHARGE, &stock->flags))
2108 flush_work(&stock->work);
2115 * Tries to drain stocked charges in other cpus. This function is asynchronous
2116 * and just put a work per cpu for draining localy on each cpu. Caller can
2117 * expects some charges will be back to res_counter later but cannot wait for
2120 static void drain_all_stock_async(struct mem_cgroup *root_memcg)
2123 * If someone calls draining, avoid adding more kworker runs.
2125 if (!mutex_trylock(&percpu_charge_mutex))
2127 drain_all_stock(root_memcg, false);
2128 mutex_unlock(&percpu_charge_mutex);
2131 /* This is a synchronous drain interface. */
2132 static void drain_all_stock_sync(struct mem_cgroup *root_memcg)
2134 /* called when force_empty is called */
2135 mutex_lock(&percpu_charge_mutex);
2136 drain_all_stock(root_memcg, true);
2137 mutex_unlock(&percpu_charge_mutex);
2141 * This function drains percpu counter value from DEAD cpu and
2142 * move it to local cpu. Note that this function can be preempted.
2144 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *memcg, int cpu)
2148 spin_lock(&memcg->pcp_counter_lock);
2149 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
2150 long x = per_cpu(memcg->stat->count[i], cpu);
2152 per_cpu(memcg->stat->count[i], cpu) = 0;
2153 memcg->nocpu_base.count[i] += x;
2155 for (i = 0; i < MEM_CGROUP_EVENTS_NSTATS; i++) {
2156 unsigned long x = per_cpu(memcg->stat->events[i], cpu);
2158 per_cpu(memcg->stat->events[i], cpu) = 0;
2159 memcg->nocpu_base.events[i] += x;
2161 /* need to clear ON_MOVE value, works as a kind of lock. */
2162 per_cpu(memcg->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
2163 spin_unlock(&memcg->pcp_counter_lock);
2166 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *memcg, int cpu)
2168 int idx = MEM_CGROUP_ON_MOVE;
2170 spin_lock(&memcg->pcp_counter_lock);
2171 per_cpu(memcg->stat->count[idx], cpu) = memcg->nocpu_base.count[idx];
2172 spin_unlock(&memcg->pcp_counter_lock);
2175 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
2176 unsigned long action,
2179 int cpu = (unsigned long)hcpu;
2180 struct memcg_stock_pcp *stock;
2181 struct mem_cgroup *iter;
2183 if ((action == CPU_ONLINE)) {
2184 for_each_mem_cgroup_all(iter)
2185 synchronize_mem_cgroup_on_move(iter, cpu);
2189 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
2192 for_each_mem_cgroup_all(iter)
2193 mem_cgroup_drain_pcp_counter(iter, cpu);
2195 stock = &per_cpu(memcg_stock, cpu);
2201 /* See __mem_cgroup_try_charge() for details */
2203 CHARGE_OK, /* success */
2204 CHARGE_RETRY, /* need to retry but retry is not bad */
2205 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
2206 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
2207 CHARGE_OOM_DIE, /* the current is killed because of OOM */
2210 static int mem_cgroup_do_charge(struct mem_cgroup *memcg, gfp_t gfp_mask,
2211 unsigned int nr_pages, bool oom_check)
2213 unsigned long csize = nr_pages * PAGE_SIZE;
2214 struct mem_cgroup *mem_over_limit;
2215 struct res_counter *fail_res;
2216 unsigned long flags = 0;
2219 ret = res_counter_charge(&memcg->res, csize, &fail_res);
2222 if (!do_swap_account)
2224 ret = res_counter_charge(&memcg->memsw, csize, &fail_res);
2228 res_counter_uncharge(&memcg->res, csize);
2229 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
2230 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
2232 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
2234 * nr_pages can be either a huge page (HPAGE_PMD_NR), a batch
2235 * of regular pages (CHARGE_BATCH), or a single regular page (1).
2237 * Never reclaim on behalf of optional batching, retry with a
2238 * single page instead.
2240 if (nr_pages == CHARGE_BATCH)
2241 return CHARGE_RETRY;
2243 if (!(gfp_mask & __GFP_WAIT))
2244 return CHARGE_WOULDBLOCK;
2246 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
2247 gfp_mask, flags, NULL);
2248 if (mem_cgroup_margin(mem_over_limit) >= nr_pages)
2249 return CHARGE_RETRY;
2251 * Even though the limit is exceeded at this point, reclaim
2252 * may have been able to free some pages. Retry the charge
2253 * before killing the task.
2255 * Only for regular pages, though: huge pages are rather
2256 * unlikely to succeed so close to the limit, and we fall back
2257 * to regular pages anyway in case of failure.
2259 if (nr_pages == 1 && ret)
2260 return CHARGE_RETRY;
2263 * At task move, charge accounts can be doubly counted. So, it's
2264 * better to wait until the end of task_move if something is going on.
2266 if (mem_cgroup_wait_acct_move(mem_over_limit))
2267 return CHARGE_RETRY;
2269 /* If we don't need to call oom-killer at el, return immediately */
2271 return CHARGE_NOMEM;
2273 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
2274 return CHARGE_OOM_DIE;
2276 return CHARGE_RETRY;
2280 * Unlike exported interface, "oom" parameter is added. if oom==true,
2281 * oom-killer can be invoked.
2283 static int __mem_cgroup_try_charge(struct mm_struct *mm,
2285 unsigned int nr_pages,
2286 struct mem_cgroup **ptr,
2289 unsigned int batch = max(CHARGE_BATCH, nr_pages);
2290 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2291 struct mem_cgroup *memcg = NULL;
2295 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
2296 * in system level. So, allow to go ahead dying process in addition to
2299 if (unlikely(test_thread_flag(TIF_MEMDIE)
2300 || fatal_signal_pending(current)))
2304 * We always charge the cgroup the mm_struct belongs to.
2305 * The mm_struct's mem_cgroup changes on task migration if the
2306 * thread group leader migrates. It's possible that mm is not
2307 * set, if so charge the init_mm (happens for pagecache usage).
2312 if (*ptr) { /* css should be a valid one */
2314 VM_BUG_ON(css_is_removed(&memcg->css));
2315 if (mem_cgroup_is_root(memcg))
2317 if (nr_pages == 1 && consume_stock(memcg))
2319 css_get(&memcg->css);
2321 struct task_struct *p;
2324 p = rcu_dereference(mm->owner);
2326 * Because we don't have task_lock(), "p" can exit.
2327 * In that case, "memcg" can point to root or p can be NULL with
2328 * race with swapoff. Then, we have small risk of mis-accouning.
2329 * But such kind of mis-account by race always happens because
2330 * we don't have cgroup_mutex(). It's overkill and we allo that
2332 * (*) swapoff at el will charge against mm-struct not against
2333 * task-struct. So, mm->owner can be NULL.
2335 memcg = mem_cgroup_from_task(p);
2336 if (!memcg || mem_cgroup_is_root(memcg)) {
2340 if (nr_pages == 1 && consume_stock(memcg)) {
2342 * It seems dagerous to access memcg without css_get().
2343 * But considering how consume_stok works, it's not
2344 * necessary. If consume_stock success, some charges
2345 * from this memcg are cached on this cpu. So, we
2346 * don't need to call css_get()/css_tryget() before
2347 * calling consume_stock().
2352 /* after here, we may be blocked. we need to get refcnt */
2353 if (!css_tryget(&memcg->css)) {
2363 /* If killed, bypass charge */
2364 if (fatal_signal_pending(current)) {
2365 css_put(&memcg->css);
2370 if (oom && !nr_oom_retries) {
2372 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2375 ret = mem_cgroup_do_charge(memcg, gfp_mask, batch, oom_check);
2379 case CHARGE_RETRY: /* not in OOM situation but retry */
2381 css_put(&memcg->css);
2384 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2385 css_put(&memcg->css);
2387 case CHARGE_NOMEM: /* OOM routine works */
2389 css_put(&memcg->css);
2392 /* If oom, we never return -ENOMEM */
2395 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2396 css_put(&memcg->css);
2399 } while (ret != CHARGE_OK);
2401 if (batch > nr_pages)
2402 refill_stock(memcg, batch - nr_pages);
2403 css_put(&memcg->css);
2416 * Somemtimes we have to undo a charge we got by try_charge().
2417 * This function is for that and do uncharge, put css's refcnt.
2418 * gotten by try_charge().
2420 static void __mem_cgroup_cancel_charge(struct mem_cgroup *memcg,
2421 unsigned int nr_pages)
2423 if (!mem_cgroup_is_root(memcg)) {
2424 unsigned long bytes = nr_pages * PAGE_SIZE;
2426 res_counter_uncharge(&memcg->res, bytes);
2427 if (do_swap_account)
2428 res_counter_uncharge(&memcg->memsw, bytes);
2433 * A helper function to get mem_cgroup from ID. must be called under
2434 * rcu_read_lock(). The caller must check css_is_removed() or some if
2435 * it's concern. (dropping refcnt from swap can be called against removed
2438 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2440 struct cgroup_subsys_state *css;
2442 /* ID 0 is unused ID */
2445 css = css_lookup(&mem_cgroup_subsys, id);
2448 return container_of(css, struct mem_cgroup, css);
2451 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2453 struct mem_cgroup *memcg = NULL;
2454 struct page_cgroup *pc;
2458 VM_BUG_ON(!PageLocked(page));
2460 pc = lookup_page_cgroup(page);
2461 lock_page_cgroup(pc);
2462 if (PageCgroupUsed(pc)) {
2463 memcg = pc->mem_cgroup;
2464 if (memcg && !css_tryget(&memcg->css))
2466 } else if (PageSwapCache(page)) {
2467 ent.val = page_private(page);
2468 id = lookup_swap_cgroup(ent);
2470 memcg = mem_cgroup_lookup(id);
2471 if (memcg && !css_tryget(&memcg->css))
2475 unlock_page_cgroup(pc);
2479 static void __mem_cgroup_commit_charge(struct mem_cgroup *memcg,
2481 unsigned int nr_pages,
2482 struct page_cgroup *pc,
2483 enum charge_type ctype)
2485 lock_page_cgroup(pc);
2486 if (unlikely(PageCgroupUsed(pc))) {
2487 unlock_page_cgroup(pc);
2488 __mem_cgroup_cancel_charge(memcg, nr_pages);
2492 * we don't need page_cgroup_lock about tail pages, becase they are not
2493 * accessed by any other context at this point.
2495 pc->mem_cgroup = memcg;
2497 * We access a page_cgroup asynchronously without lock_page_cgroup().
2498 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2499 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2500 * before USED bit, we need memory barrier here.
2501 * See mem_cgroup_add_lru_list(), etc.
2505 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2506 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2507 SetPageCgroupCache(pc);
2508 SetPageCgroupUsed(pc);
2510 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2511 ClearPageCgroupCache(pc);
2512 SetPageCgroupUsed(pc);
2518 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), nr_pages);
2519 unlock_page_cgroup(pc);
2521 * "charge_statistics" updated event counter. Then, check it.
2522 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2523 * if they exceeds softlimit.
2525 memcg_check_events(memcg, page);
2528 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2530 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2531 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2533 * Because tail pages are not marked as "used", set it. We're under
2534 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2536 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2538 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2539 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2540 unsigned long flags;
2542 if (mem_cgroup_disabled())
2545 * We have no races with charge/uncharge but will have races with
2546 * page state accounting.
2548 move_lock_page_cgroup(head_pc, &flags);
2550 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2551 smp_wmb(); /* see __commit_charge() */
2552 if (PageCgroupAcctLRU(head_pc)) {
2554 struct mem_cgroup_per_zone *mz;
2557 * LRU flags cannot be copied because we need to add tail
2558 *.page to LRU by generic call and our hook will be called.
2559 * We hold lru_lock, then, reduce counter directly.
2561 lru = page_lru(head);
2562 mz = page_cgroup_zoneinfo(head_pc->mem_cgroup, head);
2563 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2565 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2566 move_unlock_page_cgroup(head_pc, &flags);
2571 * mem_cgroup_move_account - move account of the page
2573 * @nr_pages: number of regular pages (>1 for huge pages)
2574 * @pc: page_cgroup of the page.
2575 * @from: mem_cgroup which the page is moved from.
2576 * @to: mem_cgroup which the page is moved to. @from != @to.
2577 * @uncharge: whether we should call uncharge and css_put against @from.
2579 * The caller must confirm following.
2580 * - page is not on LRU (isolate_page() is useful.)
2581 * - compound_lock is held when nr_pages > 1
2583 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2584 * done by a caller(__mem_cgroup_try_charge would be useful). If @uncharge is
2585 * true, this function does "uncharge" from old cgroup, but it doesn't if
2586 * @uncharge is false, so a caller should do "uncharge".
2588 static int mem_cgroup_move_account(struct page *page,
2589 unsigned int nr_pages,
2590 struct page_cgroup *pc,
2591 struct mem_cgroup *from,
2592 struct mem_cgroup *to,
2595 unsigned long flags;
2598 VM_BUG_ON(from == to);
2599 VM_BUG_ON(PageLRU(page));
2601 * The page is isolated from LRU. So, collapse function
2602 * will not handle this page. But page splitting can happen.
2603 * Do this check under compound_page_lock(). The caller should
2607 if (nr_pages > 1 && !PageTransHuge(page))
2610 lock_page_cgroup(pc);
2613 if (!PageCgroupUsed(pc) || pc->mem_cgroup != from)
2616 move_lock_page_cgroup(pc, &flags);
2618 if (PageCgroupFileMapped(pc)) {
2619 /* Update mapped_file data for mem_cgroup */
2621 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2622 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2625 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2627 /* This is not "cancel", but cancel_charge does all we need. */
2628 __mem_cgroup_cancel_charge(from, nr_pages);
2630 /* caller should have done css_get */
2631 pc->mem_cgroup = to;
2632 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2634 * We charges against "to" which may not have any tasks. Then, "to"
2635 * can be under rmdir(). But in current implementation, caller of
2636 * this function is just force_empty() and move charge, so it's
2637 * guaranteed that "to" is never removed. So, we don't check rmdir
2640 move_unlock_page_cgroup(pc, &flags);
2643 unlock_page_cgroup(pc);
2647 memcg_check_events(to, page);
2648 memcg_check_events(from, page);
2654 * move charges to its parent.
2657 static int mem_cgroup_move_parent(struct page *page,
2658 struct page_cgroup *pc,
2659 struct mem_cgroup *child,
2662 struct cgroup *cg = child->css.cgroup;
2663 struct cgroup *pcg = cg->parent;
2664 struct mem_cgroup *parent;
2665 unsigned int nr_pages;
2666 unsigned long uninitialized_var(flags);
2674 if (!get_page_unless_zero(page))
2676 if (isolate_lru_page(page))
2679 nr_pages = hpage_nr_pages(page);
2681 parent = mem_cgroup_from_cont(pcg);
2682 ret = __mem_cgroup_try_charge(NULL, gfp_mask, nr_pages, &parent, false);
2687 flags = compound_lock_irqsave(page);
2689 ret = mem_cgroup_move_account(page, nr_pages, pc, child, parent, true);
2691 __mem_cgroup_cancel_charge(parent, nr_pages);
2694 compound_unlock_irqrestore(page, flags);
2696 putback_lru_page(page);
2704 * Charge the memory controller for page usage.
2706 * 0 if the charge was successful
2707 * < 0 if the cgroup is over its limit
2709 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2710 gfp_t gfp_mask, enum charge_type ctype)
2712 struct mem_cgroup *memcg = NULL;
2713 unsigned int nr_pages = 1;
2714 struct page_cgroup *pc;
2718 if (PageTransHuge(page)) {
2719 nr_pages <<= compound_order(page);
2720 VM_BUG_ON(!PageTransHuge(page));
2722 * Never OOM-kill a process for a huge page. The
2723 * fault handler will fall back to regular pages.
2728 pc = lookup_page_cgroup(page);
2729 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2731 ret = __mem_cgroup_try_charge(mm, gfp_mask, nr_pages, &memcg, oom);
2735 __mem_cgroup_commit_charge(memcg, page, nr_pages, pc, ctype);
2739 int mem_cgroup_newpage_charge(struct page *page,
2740 struct mm_struct *mm, gfp_t gfp_mask)
2742 if (mem_cgroup_disabled())
2745 * If already mapped, we don't have to account.
2746 * If page cache, page->mapping has address_space.
2747 * But page->mapping may have out-of-use anon_vma pointer,
2748 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2751 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2755 return mem_cgroup_charge_common(page, mm, gfp_mask,
2756 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2760 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2761 enum charge_type ctype);
2764 __mem_cgroup_commit_charge_lrucare(struct page *page, struct mem_cgroup *memcg,
2765 enum charge_type ctype)
2767 struct page_cgroup *pc = lookup_page_cgroup(page);
2769 * In some case, SwapCache, FUSE(splice_buf->radixtree), the page
2770 * is already on LRU. It means the page may on some other page_cgroup's
2771 * LRU. Take care of it.
2773 mem_cgroup_lru_del_before_commit(page);
2774 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
2775 mem_cgroup_lru_add_after_commit(page);
2779 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2782 struct mem_cgroup *memcg = NULL;
2785 if (mem_cgroup_disabled())
2787 if (PageCompound(page))
2793 if (page_is_file_cache(page)) {
2794 ret = __mem_cgroup_try_charge(mm, gfp_mask, 1, &memcg, true);
2799 * FUSE reuses pages without going through the final
2800 * put that would remove them from the LRU list, make
2801 * sure that they get relinked properly.
2803 __mem_cgroup_commit_charge_lrucare(page, memcg,
2804 MEM_CGROUP_CHARGE_TYPE_CACHE);
2808 if (PageSwapCache(page)) {
2809 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &memcg);
2811 __mem_cgroup_commit_charge_swapin(page, memcg,
2812 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2814 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2815 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2821 * While swap-in, try_charge -> commit or cancel, the page is locked.
2822 * And when try_charge() successfully returns, one refcnt to memcg without
2823 * struct page_cgroup is acquired. This refcnt will be consumed by
2824 * "commit()" or removed by "cancel()"
2826 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2828 gfp_t mask, struct mem_cgroup **ptr)
2830 struct mem_cgroup *memcg;
2835 if (mem_cgroup_disabled())
2838 if (!do_swap_account)
2841 * A racing thread's fault, or swapoff, may have already updated
2842 * the pte, and even removed page from swap cache: in those cases
2843 * do_swap_page()'s pte_same() test will fail; but there's also a
2844 * KSM case which does need to charge the page.
2846 if (!PageSwapCache(page))
2848 memcg = try_get_mem_cgroup_from_page(page);
2852 ret = __mem_cgroup_try_charge(NULL, mask, 1, ptr, true);
2853 css_put(&memcg->css);
2858 return __mem_cgroup_try_charge(mm, mask, 1, ptr, true);
2862 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2863 enum charge_type ctype)
2865 if (mem_cgroup_disabled())
2869 cgroup_exclude_rmdir(&ptr->css);
2871 __mem_cgroup_commit_charge_lrucare(page, ptr, ctype);
2873 * Now swap is on-memory. This means this page may be
2874 * counted both as mem and swap....double count.
2875 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2876 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2877 * may call delete_from_swap_cache() before reach here.
2879 if (do_swap_account && PageSwapCache(page)) {
2880 swp_entry_t ent = {.val = page_private(page)};
2882 struct mem_cgroup *memcg;
2884 id = swap_cgroup_record(ent, 0);
2886 memcg = mem_cgroup_lookup(id);
2889 * This recorded memcg can be obsolete one. So, avoid
2890 * calling css_tryget
2892 if (!mem_cgroup_is_root(memcg))
2893 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2894 mem_cgroup_swap_statistics(memcg, false);
2895 mem_cgroup_put(memcg);
2900 * At swapin, we may charge account against cgroup which has no tasks.
2901 * So, rmdir()->pre_destroy() can be called while we do this charge.
2902 * In that case, we need to call pre_destroy() again. check it here.
2904 cgroup_release_and_wakeup_rmdir(&ptr->css);
2907 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2909 __mem_cgroup_commit_charge_swapin(page, ptr,
2910 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2913 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *memcg)
2915 if (mem_cgroup_disabled())
2919 __mem_cgroup_cancel_charge(memcg, 1);
2922 static void mem_cgroup_do_uncharge(struct mem_cgroup *memcg,
2923 unsigned int nr_pages,
2924 const enum charge_type ctype)
2926 struct memcg_batch_info *batch = NULL;
2927 bool uncharge_memsw = true;
2929 /* If swapout, usage of swap doesn't decrease */
2930 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2931 uncharge_memsw = false;
2933 batch = ¤t->memcg_batch;
2935 * In usual, we do css_get() when we remember memcg pointer.
2936 * But in this case, we keep res->usage until end of a series of
2937 * uncharges. Then, it's ok to ignore memcg's refcnt.
2940 batch->memcg = memcg;
2942 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2943 * In those cases, all pages freed continuously can be expected to be in
2944 * the same cgroup and we have chance to coalesce uncharges.
2945 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2946 * because we want to do uncharge as soon as possible.
2949 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2950 goto direct_uncharge;
2953 goto direct_uncharge;
2956 * In typical case, batch->memcg == mem. This means we can
2957 * merge a series of uncharges to an uncharge of res_counter.
2958 * If not, we uncharge res_counter ony by one.
2960 if (batch->memcg != memcg)
2961 goto direct_uncharge;
2962 /* remember freed charge and uncharge it later */
2965 batch->memsw_nr_pages++;
2968 res_counter_uncharge(&memcg->res, nr_pages * PAGE_SIZE);
2970 res_counter_uncharge(&memcg->memsw, nr_pages * PAGE_SIZE);
2971 if (unlikely(batch->memcg != memcg))
2972 memcg_oom_recover(memcg);
2977 * uncharge if !page_mapped(page)
2979 static struct mem_cgroup *
2980 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2982 struct mem_cgroup *memcg = NULL;
2983 unsigned int nr_pages = 1;
2984 struct page_cgroup *pc;
2986 if (mem_cgroup_disabled())
2989 if (PageSwapCache(page))
2992 if (PageTransHuge(page)) {
2993 nr_pages <<= compound_order(page);
2994 VM_BUG_ON(!PageTransHuge(page));
2997 * Check if our page_cgroup is valid
2999 pc = lookup_page_cgroup(page);
3000 if (unlikely(!pc || !PageCgroupUsed(pc)))
3003 lock_page_cgroup(pc);
3005 memcg = pc->mem_cgroup;
3007 if (!PageCgroupUsed(pc))
3011 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
3012 case MEM_CGROUP_CHARGE_TYPE_DROP:
3013 /* See mem_cgroup_prepare_migration() */
3014 if (page_mapped(page) || PageCgroupMigration(pc))
3017 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
3018 if (!PageAnon(page)) { /* Shared memory */
3019 if (page->mapping && !page_is_file_cache(page))
3021 } else if (page_mapped(page)) /* Anon */
3028 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -nr_pages);
3030 ClearPageCgroupUsed(pc);
3032 * pc->mem_cgroup is not cleared here. It will be accessed when it's
3033 * freed from LRU. This is safe because uncharged page is expected not
3034 * to be reused (freed soon). Exception is SwapCache, it's handled by
3035 * special functions.
3038 unlock_page_cgroup(pc);
3040 * even after unlock, we have memcg->res.usage here and this memcg
3041 * will never be freed.
3043 memcg_check_events(memcg, page);
3044 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
3045 mem_cgroup_swap_statistics(memcg, true);
3046 mem_cgroup_get(memcg);
3048 if (!mem_cgroup_is_root(memcg))
3049 mem_cgroup_do_uncharge(memcg, nr_pages, ctype);
3054 unlock_page_cgroup(pc);
3058 void mem_cgroup_uncharge_page(struct page *page)
3061 if (page_mapped(page))
3063 if (page->mapping && !PageAnon(page))
3065 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
3068 void mem_cgroup_uncharge_cache_page(struct page *page)
3070 VM_BUG_ON(page_mapped(page));
3071 VM_BUG_ON(page->mapping);
3072 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
3076 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
3077 * In that cases, pages are freed continuously and we can expect pages
3078 * are in the same memcg. All these calls itself limits the number of
3079 * pages freed at once, then uncharge_start/end() is called properly.
3080 * This may be called prural(2) times in a context,
3083 void mem_cgroup_uncharge_start(void)
3085 current->memcg_batch.do_batch++;
3086 /* We can do nest. */
3087 if (current->memcg_batch.do_batch == 1) {
3088 current->memcg_batch.memcg = NULL;
3089 current->memcg_batch.nr_pages = 0;
3090 current->memcg_batch.memsw_nr_pages = 0;
3094 void mem_cgroup_uncharge_end(void)
3096 struct memcg_batch_info *batch = ¤t->memcg_batch;
3098 if (!batch->do_batch)
3102 if (batch->do_batch) /* If stacked, do nothing. */
3108 * This "batch->memcg" is valid without any css_get/put etc...
3109 * bacause we hide charges behind us.
3111 if (batch->nr_pages)
3112 res_counter_uncharge(&batch->memcg->res,
3113 batch->nr_pages * PAGE_SIZE);
3114 if (batch->memsw_nr_pages)
3115 res_counter_uncharge(&batch->memcg->memsw,
3116 batch->memsw_nr_pages * PAGE_SIZE);
3117 memcg_oom_recover(batch->memcg);
3118 /* forget this pointer (for sanity check) */
3119 batch->memcg = NULL;
3124 * called after __delete_from_swap_cache() and drop "page" account.
3125 * memcg information is recorded to swap_cgroup of "ent"
3128 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
3130 struct mem_cgroup *memcg;
3131 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
3133 if (!swapout) /* this was a swap cache but the swap is unused ! */
3134 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
3136 memcg = __mem_cgroup_uncharge_common(page, ctype);
3139 * record memcg information, if swapout && memcg != NULL,
3140 * mem_cgroup_get() was called in uncharge().
3142 if (do_swap_account && swapout && memcg)
3143 swap_cgroup_record(ent, css_id(&memcg->css));
3147 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3149 * called from swap_entry_free(). remove record in swap_cgroup and
3150 * uncharge "memsw" account.
3152 void mem_cgroup_uncharge_swap(swp_entry_t ent)
3154 struct mem_cgroup *memcg;
3157 if (!do_swap_account)
3160 id = swap_cgroup_record(ent, 0);
3162 memcg = mem_cgroup_lookup(id);
3165 * We uncharge this because swap is freed.
3166 * This memcg can be obsolete one. We avoid calling css_tryget
3168 if (!mem_cgroup_is_root(memcg))
3169 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
3170 mem_cgroup_swap_statistics(memcg, false);
3171 mem_cgroup_put(memcg);
3177 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
3178 * @entry: swap entry to be moved
3179 * @from: mem_cgroup which the entry is moved from
3180 * @to: mem_cgroup which the entry is moved to
3181 * @need_fixup: whether we should fixup res_counters and refcounts.
3183 * It succeeds only when the swap_cgroup's record for this entry is the same
3184 * as the mem_cgroup's id of @from.
3186 * Returns 0 on success, -EINVAL on failure.
3188 * The caller must have charged to @to, IOW, called res_counter_charge() about
3189 * both res and memsw, and called css_get().
3191 static int mem_cgroup_move_swap_account(swp_entry_t entry,
3192 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3194 unsigned short old_id, new_id;
3196 old_id = css_id(&from->css);
3197 new_id = css_id(&to->css);
3199 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
3200 mem_cgroup_swap_statistics(from, false);
3201 mem_cgroup_swap_statistics(to, true);
3203 * This function is only called from task migration context now.
3204 * It postpones res_counter and refcount handling till the end
3205 * of task migration(mem_cgroup_clear_mc()) for performance
3206 * improvement. But we cannot postpone mem_cgroup_get(to)
3207 * because if the process that has been moved to @to does
3208 * swap-in, the refcount of @to might be decreased to 0.
3212 if (!mem_cgroup_is_root(from))
3213 res_counter_uncharge(&from->memsw, PAGE_SIZE);
3214 mem_cgroup_put(from);
3216 * we charged both to->res and to->memsw, so we should
3219 if (!mem_cgroup_is_root(to))
3220 res_counter_uncharge(&to->res, PAGE_SIZE);
3227 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
3228 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
3235 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
3238 int mem_cgroup_prepare_migration(struct page *page,
3239 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
3241 struct mem_cgroup *memcg = NULL;
3242 struct page_cgroup *pc;
3243 enum charge_type ctype;
3248 VM_BUG_ON(PageTransHuge(page));
3249 if (mem_cgroup_disabled())
3252 pc = lookup_page_cgroup(page);
3253 lock_page_cgroup(pc);
3254 if (PageCgroupUsed(pc)) {
3255 memcg = pc->mem_cgroup;
3256 css_get(&memcg->css);
3258 * At migrating an anonymous page, its mapcount goes down
3259 * to 0 and uncharge() will be called. But, even if it's fully
3260 * unmapped, migration may fail and this page has to be
3261 * charged again. We set MIGRATION flag here and delay uncharge
3262 * until end_migration() is called
3264 * Corner Case Thinking
3266 * When the old page was mapped as Anon and it's unmap-and-freed
3267 * while migration was ongoing.
3268 * If unmap finds the old page, uncharge() of it will be delayed
3269 * until end_migration(). If unmap finds a new page, it's
3270 * uncharged when it make mapcount to be 1->0. If unmap code
3271 * finds swap_migration_entry, the new page will not be mapped
3272 * and end_migration() will find it(mapcount==0).
3275 * When the old page was mapped but migraion fails, the kernel
3276 * remaps it. A charge for it is kept by MIGRATION flag even
3277 * if mapcount goes down to 0. We can do remap successfully
3278 * without charging it again.
3281 * The "old" page is under lock_page() until the end of
3282 * migration, so, the old page itself will not be swapped-out.
3283 * If the new page is swapped out before end_migraton, our
3284 * hook to usual swap-out path will catch the event.
3287 SetPageCgroupMigration(pc);
3289 unlock_page_cgroup(pc);
3291 * If the page is not charged at this point,
3298 ret = __mem_cgroup_try_charge(NULL, gfp_mask, 1, ptr, false);
3299 css_put(&memcg->css);/* drop extra refcnt */
3300 if (ret || *ptr == NULL) {
3301 if (PageAnon(page)) {
3302 lock_page_cgroup(pc);
3303 ClearPageCgroupMigration(pc);
3304 unlock_page_cgroup(pc);
3306 * The old page may be fully unmapped while we kept it.
3308 mem_cgroup_uncharge_page(page);
3313 * We charge new page before it's used/mapped. So, even if unlock_page()
3314 * is called before end_migration, we can catch all events on this new
3315 * page. In the case new page is migrated but not remapped, new page's
3316 * mapcount will be finally 0 and we call uncharge in end_migration().
3318 pc = lookup_page_cgroup(newpage);
3320 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
3321 else if (page_is_file_cache(page))
3322 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
3324 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3325 __mem_cgroup_commit_charge(memcg, page, 1, pc, ctype);
3329 /* remove redundant charge if migration failed*/
3330 void mem_cgroup_end_migration(struct mem_cgroup *memcg,
3331 struct page *oldpage, struct page *newpage, bool migration_ok)
3333 struct page *used, *unused;
3334 struct page_cgroup *pc;
3338 /* blocks rmdir() */
3339 cgroup_exclude_rmdir(&memcg->css);
3340 if (!migration_ok) {
3348 * We disallowed uncharge of pages under migration because mapcount
3349 * of the page goes down to zero, temporarly.
3350 * Clear the flag and check the page should be charged.
3352 pc = lookup_page_cgroup(oldpage);
3353 lock_page_cgroup(pc);
3354 ClearPageCgroupMigration(pc);
3355 unlock_page_cgroup(pc);
3357 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3360 * If a page is a file cache, radix-tree replacement is very atomic
3361 * and we can skip this check. When it was an Anon page, its mapcount
3362 * goes down to 0. But because we added MIGRATION flage, it's not
3363 * uncharged yet. There are several case but page->mapcount check
3364 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3365 * check. (see prepare_charge() also)
3368 mem_cgroup_uncharge_page(used);
3370 * At migration, we may charge account against cgroup which has no
3372 * So, rmdir()->pre_destroy() can be called while we do this charge.
3373 * In that case, we need to call pre_destroy() again. check it here.
3375 cgroup_release_and_wakeup_rmdir(&memcg->css);
3379 * At replace page cache, newpage is not under any memcg but it's on
3380 * LRU. So, this function doesn't touch res_counter but handles LRU
3381 * in correct way. Both pages are locked so we cannot race with uncharge.
3383 void mem_cgroup_replace_page_cache(struct page *oldpage,
3384 struct page *newpage)
3386 struct mem_cgroup *memcg;
3387 struct page_cgroup *pc;
3389 enum charge_type type = MEM_CGROUP_CHARGE_TYPE_CACHE;
3390 unsigned long flags;
3392 if (mem_cgroup_disabled())
3395 pc = lookup_page_cgroup(oldpage);
3396 /* fix accounting on old pages */
3397 lock_page_cgroup(pc);
3398 memcg = pc->mem_cgroup;
3399 mem_cgroup_charge_statistics(memcg, PageCgroupCache(pc), -1);
3400 ClearPageCgroupUsed(pc);
3401 unlock_page_cgroup(pc);
3403 if (PageSwapBacked(oldpage))
3404 type = MEM_CGROUP_CHARGE_TYPE_SHMEM;
3406 zone = page_zone(newpage);
3407 pc = lookup_page_cgroup(newpage);
3409 * Even if newpage->mapping was NULL before starting replacement,
3410 * the newpage may be on LRU(or pagevec for LRU) already. We lock
3411 * LRU while we overwrite pc->mem_cgroup.
3413 spin_lock_irqsave(&zone->lru_lock, flags);
3414 if (PageLRU(newpage))
3415 del_page_from_lru_list(zone, newpage, page_lru(newpage));
3416 __mem_cgroup_commit_charge(memcg, newpage, 1, pc, type);
3417 if (PageLRU(newpage))
3418 add_page_to_lru_list(zone, newpage, page_lru(newpage));
3419 spin_unlock_irqrestore(&zone->lru_lock, flags);
3422 #ifdef CONFIG_DEBUG_VM
3423 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3425 struct page_cgroup *pc;
3427 pc = lookup_page_cgroup(page);
3428 if (likely(pc) && PageCgroupUsed(pc))
3433 bool mem_cgroup_bad_page_check(struct page *page)
3435 if (mem_cgroup_disabled())
3438 return lookup_page_cgroup_used(page) != NULL;
3441 void mem_cgroup_print_bad_page(struct page *page)
3443 struct page_cgroup *pc;
3445 pc = lookup_page_cgroup_used(page);
3450 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3451 pc, pc->flags, pc->mem_cgroup);
3453 path = kmalloc(PATH_MAX, GFP_KERNEL);
3456 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3461 printk(KERN_CONT "(%s)\n",
3462 (ret < 0) ? "cannot get the path" : path);
3468 static DEFINE_MUTEX(set_limit_mutex);
3470 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3471 unsigned long long val)
3474 u64 memswlimit, memlimit;
3476 int children = mem_cgroup_count_children(memcg);
3477 u64 curusage, oldusage;
3481 * For keeping hierarchical_reclaim simple, how long we should retry
3482 * is depends on callers. We set our retry-count to be function
3483 * of # of children which we should visit in this loop.
3485 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3487 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3490 while (retry_count) {
3491 if (signal_pending(current)) {
3496 * Rather than hide all in some function, I do this in
3497 * open coded manner. You see what this really does.
3498 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3500 mutex_lock(&set_limit_mutex);
3501 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3502 if (memswlimit < val) {
3504 mutex_unlock(&set_limit_mutex);
3508 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3512 ret = res_counter_set_limit(&memcg->res, val);
3514 if (memswlimit == val)
3515 memcg->memsw_is_minimum = true;
3517 memcg->memsw_is_minimum = false;
3519 mutex_unlock(&set_limit_mutex);
3524 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3525 MEM_CGROUP_RECLAIM_SHRINK,
3527 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3528 /* Usage is reduced ? */
3529 if (curusage >= oldusage)
3532 oldusage = curusage;
3534 if (!ret && enlarge)
3535 memcg_oom_recover(memcg);
3540 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3541 unsigned long long val)
3544 u64 memlimit, memswlimit, oldusage, curusage;
3545 int children = mem_cgroup_count_children(memcg);
3549 /* see mem_cgroup_resize_res_limit */
3550 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3551 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3552 while (retry_count) {
3553 if (signal_pending(current)) {
3558 * Rather than hide all in some function, I do this in
3559 * open coded manner. You see what this really does.
3560 * We have to guarantee memcg->res.limit < memcg->memsw.limit.
3562 mutex_lock(&set_limit_mutex);
3563 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3564 if (memlimit > val) {
3566 mutex_unlock(&set_limit_mutex);
3569 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3570 if (memswlimit < val)
3572 ret = res_counter_set_limit(&memcg->memsw, val);
3574 if (memlimit == val)
3575 memcg->memsw_is_minimum = true;
3577 memcg->memsw_is_minimum = false;
3579 mutex_unlock(&set_limit_mutex);
3584 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3585 MEM_CGROUP_RECLAIM_NOSWAP |
3586 MEM_CGROUP_RECLAIM_SHRINK,
3588 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3589 /* Usage is reduced ? */
3590 if (curusage >= oldusage)
3593 oldusage = curusage;
3595 if (!ret && enlarge)
3596 memcg_oom_recover(memcg);
3600 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3602 unsigned long *total_scanned)
3604 unsigned long nr_reclaimed = 0;
3605 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3606 unsigned long reclaimed;
3608 struct mem_cgroup_tree_per_zone *mctz;
3609 unsigned long long excess;
3610 unsigned long nr_scanned;
3615 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3617 * This loop can run a while, specially if mem_cgroup's continuously
3618 * keep exceeding their soft limit and putting the system under
3625 mz = mem_cgroup_largest_soft_limit_node(mctz);
3630 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3632 MEM_CGROUP_RECLAIM_SOFT,
3634 nr_reclaimed += reclaimed;
3635 *total_scanned += nr_scanned;
3636 spin_lock(&mctz->lock);
3639 * If we failed to reclaim anything from this memory cgroup
3640 * it is time to move on to the next cgroup
3646 * Loop until we find yet another one.
3648 * By the time we get the soft_limit lock
3649 * again, someone might have aded the
3650 * group back on the RB tree. Iterate to
3651 * make sure we get a different mem.
3652 * mem_cgroup_largest_soft_limit_node returns
3653 * NULL if no other cgroup is present on
3657 __mem_cgroup_largest_soft_limit_node(mctz);
3659 css_put(&next_mz->mem->css);
3660 else /* next_mz == NULL or other memcg */
3664 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3665 excess = res_counter_soft_limit_excess(&mz->mem->res);
3667 * One school of thought says that we should not add
3668 * back the node to the tree if reclaim returns 0.
3669 * But our reclaim could return 0, simply because due
3670 * to priority we are exposing a smaller subset of
3671 * memory to reclaim from. Consider this as a longer
3674 /* If excess == 0, no tree ops */
3675 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3676 spin_unlock(&mctz->lock);
3677 css_put(&mz->mem->css);
3680 * Could not reclaim anything and there are no more
3681 * mem cgroups to try or we seem to be looping without
3682 * reclaiming anything.
3684 if (!nr_reclaimed &&
3686 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3688 } while (!nr_reclaimed);
3690 css_put(&next_mz->mem->css);
3691 return nr_reclaimed;
3695 * This routine traverse page_cgroup in given list and drop them all.
3696 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3698 static int mem_cgroup_force_empty_list(struct mem_cgroup *memcg,
3699 int node, int zid, enum lru_list lru)
3702 struct mem_cgroup_per_zone *mz;
3703 struct page_cgroup *pc, *busy;
3704 unsigned long flags, loop;
3705 struct list_head *list;
3708 zone = &NODE_DATA(node)->node_zones[zid];
3709 mz = mem_cgroup_zoneinfo(memcg, node, zid);
3710 list = &mz->lists[lru];
3712 loop = MEM_CGROUP_ZSTAT(mz, lru);
3713 /* give some margin against EBUSY etc...*/
3720 spin_lock_irqsave(&zone->lru_lock, flags);
3721 if (list_empty(list)) {
3722 spin_unlock_irqrestore(&zone->lru_lock, flags);
3725 pc = list_entry(list->prev, struct page_cgroup, lru);
3727 list_move(&pc->lru, list);
3729 spin_unlock_irqrestore(&zone->lru_lock, flags);
3732 spin_unlock_irqrestore(&zone->lru_lock, flags);
3734 page = lookup_cgroup_page(pc);
3736 ret = mem_cgroup_move_parent(page, pc, memcg, GFP_KERNEL);
3740 if (ret == -EBUSY || ret == -EINVAL) {
3741 /* found lock contention or "pc" is obsolete. */
3748 if (!ret && !list_empty(list))
3754 * make mem_cgroup's charge to be 0 if there is no task.
3755 * This enables deleting this mem_cgroup.
3757 static int mem_cgroup_force_empty(struct mem_cgroup *memcg, bool free_all)
3760 int node, zid, shrink;
3761 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3762 struct cgroup *cgrp = memcg->css.cgroup;
3764 css_get(&memcg->css);
3767 /* should free all ? */
3773 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3776 if (signal_pending(current))
3778 /* This is for making all *used* pages to be on LRU. */
3779 lru_add_drain_all();
3780 drain_all_stock_sync(memcg);
3782 mem_cgroup_start_move(memcg);
3783 for_each_node_state(node, N_HIGH_MEMORY) {
3784 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3787 ret = mem_cgroup_force_empty_list(memcg,
3796 mem_cgroup_end_move(memcg);
3797 memcg_oom_recover(memcg);
3798 /* it seems parent cgroup doesn't have enough mem */
3802 /* "ret" should also be checked to ensure all lists are empty. */
3803 } while (memcg->res.usage > 0 || ret);
3805 css_put(&memcg->css);
3809 /* returns EBUSY if there is a task or if we come here twice. */
3810 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3814 /* we call try-to-free pages for make this cgroup empty */
3815 lru_add_drain_all();
3816 /* try to free all pages in this cgroup */
3818 while (nr_retries && memcg->res.usage > 0) {
3821 if (signal_pending(current)) {
3825 progress = try_to_free_mem_cgroup_pages(memcg, GFP_KERNEL,
3829 /* maybe some writeback is necessary */
3830 congestion_wait(BLK_RW_ASYNC, HZ/10);
3835 /* try move_account...there may be some *locked* pages. */
3839 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3841 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3845 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3847 return mem_cgroup_from_cont(cont)->use_hierarchy;
3850 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3854 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3855 struct cgroup *parent = cont->parent;
3856 struct mem_cgroup *parent_memcg = NULL;
3859 parent_memcg = mem_cgroup_from_cont(parent);
3863 * If parent's use_hierarchy is set, we can't make any modifications
3864 * in the child subtrees. If it is unset, then the change can
3865 * occur, provided the current cgroup has no children.
3867 * For the root cgroup, parent_mem is NULL, we allow value to be
3868 * set if there are no children.
3870 if ((!parent_memcg || !parent_memcg->use_hierarchy) &&
3871 (val == 1 || val == 0)) {
3872 if (list_empty(&cont->children))
3873 memcg->use_hierarchy = val;
3884 static unsigned long mem_cgroup_recursive_stat(struct mem_cgroup *memcg,
3885 enum mem_cgroup_stat_index idx)
3887 struct mem_cgroup *iter;
3890 /* Per-cpu values can be negative, use a signed accumulator */
3891 for_each_mem_cgroup_tree(iter, memcg)
3892 val += mem_cgroup_read_stat(iter, idx);
3894 if (val < 0) /* race ? */
3899 static inline u64 mem_cgroup_usage(struct mem_cgroup *memcg, bool swap)
3903 if (!mem_cgroup_is_root(memcg)) {
3905 return res_counter_read_u64(&memcg->res, RES_USAGE);
3907 return res_counter_read_u64(&memcg->memsw, RES_USAGE);
3910 val = mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_CACHE);
3911 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_RSS);
3914 val += mem_cgroup_recursive_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
3916 return val << PAGE_SHIFT;
3919 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3921 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3925 type = MEMFILE_TYPE(cft->private);
3926 name = MEMFILE_ATTR(cft->private);
3929 if (name == RES_USAGE)
3930 val = mem_cgroup_usage(memcg, false);
3932 val = res_counter_read_u64(&memcg->res, name);
3935 if (name == RES_USAGE)
3936 val = mem_cgroup_usage(memcg, true);
3938 val = res_counter_read_u64(&memcg->memsw, name);
3947 * The user of this function is...
3950 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3953 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3955 unsigned long long val;
3958 type = MEMFILE_TYPE(cft->private);
3959 name = MEMFILE_ATTR(cft->private);
3962 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3966 /* This function does all necessary parse...reuse it */
3967 ret = res_counter_memparse_write_strategy(buffer, &val);
3971 ret = mem_cgroup_resize_limit(memcg, val);
3973 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3975 case RES_SOFT_LIMIT:
3976 ret = res_counter_memparse_write_strategy(buffer, &val);
3980 * For memsw, soft limits are hard to implement in terms
3981 * of semantics, for now, we support soft limits for
3982 * control without swap
3985 ret = res_counter_set_soft_limit(&memcg->res, val);
3990 ret = -EINVAL; /* should be BUG() ? */
3996 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3997 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3999 struct cgroup *cgroup;
4000 unsigned long long min_limit, min_memsw_limit, tmp;
4002 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
4003 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4004 cgroup = memcg->css.cgroup;
4005 if (!memcg->use_hierarchy)
4008 while (cgroup->parent) {
4009 cgroup = cgroup->parent;
4010 memcg = mem_cgroup_from_cont(cgroup);
4011 if (!memcg->use_hierarchy)
4013 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
4014 min_limit = min(min_limit, tmp);
4015 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
4016 min_memsw_limit = min(min_memsw_limit, tmp);
4019 *mem_limit = min_limit;
4020 *memsw_limit = min_memsw_limit;
4024 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
4026 struct mem_cgroup *memcg;
4029 memcg = mem_cgroup_from_cont(cont);
4030 type = MEMFILE_TYPE(event);
4031 name = MEMFILE_ATTR(event);
4035 res_counter_reset_max(&memcg->res);
4037 res_counter_reset_max(&memcg->memsw);
4041 res_counter_reset_failcnt(&memcg->res);
4043 res_counter_reset_failcnt(&memcg->memsw);
4050 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
4053 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
4057 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4058 struct cftype *cft, u64 val)
4060 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4062 if (val >= (1 << NR_MOVE_TYPE))
4065 * We check this value several times in both in can_attach() and
4066 * attach(), so we need cgroup lock to prevent this value from being
4070 memcg->move_charge_at_immigrate = val;
4076 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
4077 struct cftype *cft, u64 val)
4084 /* For read statistics */
4102 struct mcs_total_stat {
4103 s64 stat[NR_MCS_STAT];
4109 } memcg_stat_strings[NR_MCS_STAT] = {
4110 {"cache", "total_cache"},
4111 {"rss", "total_rss"},
4112 {"mapped_file", "total_mapped_file"},
4113 {"pgpgin", "total_pgpgin"},
4114 {"pgpgout", "total_pgpgout"},
4115 {"swap", "total_swap"},
4116 {"pgfault", "total_pgfault"},
4117 {"pgmajfault", "total_pgmajfault"},
4118 {"inactive_anon", "total_inactive_anon"},
4119 {"active_anon", "total_active_anon"},
4120 {"inactive_file", "total_inactive_file"},
4121 {"active_file", "total_active_file"},
4122 {"unevictable", "total_unevictable"}
4127 mem_cgroup_get_local_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4132 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_CACHE);
4133 s->stat[MCS_CACHE] += val * PAGE_SIZE;
4134 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_RSS);
4135 s->stat[MCS_RSS] += val * PAGE_SIZE;
4136 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_FILE_MAPPED);
4137 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
4138 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGIN);
4139 s->stat[MCS_PGPGIN] += val;
4140 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGPGOUT);
4141 s->stat[MCS_PGPGOUT] += val;
4142 if (do_swap_account) {
4143 val = mem_cgroup_read_stat(memcg, MEM_CGROUP_STAT_SWAPOUT);
4144 s->stat[MCS_SWAP] += val * PAGE_SIZE;
4146 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGFAULT);
4147 s->stat[MCS_PGFAULT] += val;
4148 val = mem_cgroup_read_events(memcg, MEM_CGROUP_EVENTS_PGMAJFAULT);
4149 s->stat[MCS_PGMAJFAULT] += val;
4152 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_ANON));
4153 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
4154 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_ANON));
4155 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
4156 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_INACTIVE_FILE));
4157 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
4158 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_ACTIVE_FILE));
4159 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
4160 val = mem_cgroup_nr_lru_pages(memcg, BIT(LRU_UNEVICTABLE));
4161 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
4165 mem_cgroup_get_total_stat(struct mem_cgroup *memcg, struct mcs_total_stat *s)
4167 struct mem_cgroup *iter;
4169 for_each_mem_cgroup_tree(iter, memcg)
4170 mem_cgroup_get_local_stat(iter, s);
4174 static int mem_control_numa_stat_show(struct seq_file *m, void *arg)
4177 unsigned long total_nr, file_nr, anon_nr, unevictable_nr;
4178 unsigned long node_nr;
4179 struct cgroup *cont = m->private;
4180 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4182 total_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL);
4183 seq_printf(m, "total=%lu", total_nr);
4184 for_each_node_state(nid, N_HIGH_MEMORY) {
4185 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid, LRU_ALL);
4186 seq_printf(m, " N%d=%lu", nid, node_nr);
4190 file_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_FILE);
4191 seq_printf(m, "file=%lu", file_nr);
4192 for_each_node_state(nid, N_HIGH_MEMORY) {
4193 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4195 seq_printf(m, " N%d=%lu", nid, node_nr);
4199 anon_nr = mem_cgroup_nr_lru_pages(mem_cont, LRU_ALL_ANON);
4200 seq_printf(m, "anon=%lu", anon_nr);
4201 for_each_node_state(nid, N_HIGH_MEMORY) {
4202 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4204 seq_printf(m, " N%d=%lu", nid, node_nr);
4208 unevictable_nr = mem_cgroup_nr_lru_pages(mem_cont, BIT(LRU_UNEVICTABLE));
4209 seq_printf(m, "unevictable=%lu", unevictable_nr);
4210 for_each_node_state(nid, N_HIGH_MEMORY) {
4211 node_nr = mem_cgroup_node_nr_lru_pages(mem_cont, nid,
4212 BIT(LRU_UNEVICTABLE));
4213 seq_printf(m, " N%d=%lu", nid, node_nr);
4218 #endif /* CONFIG_NUMA */
4220 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
4221 struct cgroup_map_cb *cb)
4223 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
4224 struct mcs_total_stat mystat;
4227 memset(&mystat, 0, sizeof(mystat));
4228 mem_cgroup_get_local_stat(mem_cont, &mystat);
4231 for (i = 0; i < NR_MCS_STAT; i++) {
4232 if (i == MCS_SWAP && !do_swap_account)
4234 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
4237 /* Hierarchical information */
4239 unsigned long long limit, memsw_limit;
4240 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
4241 cb->fill(cb, "hierarchical_memory_limit", limit);
4242 if (do_swap_account)
4243 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
4246 memset(&mystat, 0, sizeof(mystat));
4247 mem_cgroup_get_total_stat(mem_cont, &mystat);
4248 for (i = 0; i < NR_MCS_STAT; i++) {
4249 if (i == MCS_SWAP && !do_swap_account)
4251 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
4254 #ifdef CONFIG_DEBUG_VM
4257 struct mem_cgroup_per_zone *mz;
4258 unsigned long recent_rotated[2] = {0, 0};
4259 unsigned long recent_scanned[2] = {0, 0};
4261 for_each_online_node(nid)
4262 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
4263 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
4265 recent_rotated[0] +=
4266 mz->reclaim_stat.recent_rotated[0];
4267 recent_rotated[1] +=
4268 mz->reclaim_stat.recent_rotated[1];
4269 recent_scanned[0] +=
4270 mz->reclaim_stat.recent_scanned[0];
4271 recent_scanned[1] +=
4272 mz->reclaim_stat.recent_scanned[1];
4274 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
4275 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
4276 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
4277 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
4284 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
4286 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4288 return mem_cgroup_swappiness(memcg);
4291 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
4294 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4295 struct mem_cgroup *parent;
4300 if (cgrp->parent == NULL)
4303 parent = mem_cgroup_from_cont(cgrp->parent);
4307 /* If under hierarchy, only empty-root can set this value */
4308 if ((parent->use_hierarchy) ||
4309 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4314 memcg->swappiness = val;
4321 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
4323 struct mem_cgroup_threshold_ary *t;
4329 t = rcu_dereference(memcg->thresholds.primary);
4331 t = rcu_dereference(memcg->memsw_thresholds.primary);
4336 usage = mem_cgroup_usage(memcg, swap);
4339 * current_threshold points to threshold just below usage.
4340 * If it's not true, a threshold was crossed after last
4341 * call of __mem_cgroup_threshold().
4343 i = t->current_threshold;
4346 * Iterate backward over array of thresholds starting from
4347 * current_threshold and check if a threshold is crossed.
4348 * If none of thresholds below usage is crossed, we read
4349 * only one element of the array here.
4351 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
4352 eventfd_signal(t->entries[i].eventfd, 1);
4354 /* i = current_threshold + 1 */
4358 * Iterate forward over array of thresholds starting from
4359 * current_threshold+1 and check if a threshold is crossed.
4360 * If none of thresholds above usage is crossed, we read
4361 * only one element of the array here.
4363 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
4364 eventfd_signal(t->entries[i].eventfd, 1);
4366 /* Update current_threshold */
4367 t->current_threshold = i - 1;
4372 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
4375 __mem_cgroup_threshold(memcg, false);
4376 if (do_swap_account)
4377 __mem_cgroup_threshold(memcg, true);
4379 memcg = parent_mem_cgroup(memcg);
4383 static int compare_thresholds(const void *a, const void *b)
4385 const struct mem_cgroup_threshold *_a = a;
4386 const struct mem_cgroup_threshold *_b = b;
4388 if (_a->threshold > _b->threshold)
4391 if (_a->threshold < _b->threshold)
4397 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *memcg)
4399 struct mem_cgroup_eventfd_list *ev;
4401 list_for_each_entry(ev, &memcg->oom_notify, list)
4402 eventfd_signal(ev->eventfd, 1);
4406 static void mem_cgroup_oom_notify(struct mem_cgroup *memcg)
4408 struct mem_cgroup *iter;
4410 for_each_mem_cgroup_tree(iter, memcg)
4411 mem_cgroup_oom_notify_cb(iter);
4414 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
4415 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4417 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4418 struct mem_cgroup_thresholds *thresholds;
4419 struct mem_cgroup_threshold_ary *new;
4420 int type = MEMFILE_TYPE(cft->private);
4421 u64 threshold, usage;
4424 ret = res_counter_memparse_write_strategy(args, &threshold);
4428 mutex_lock(&memcg->thresholds_lock);
4431 thresholds = &memcg->thresholds;
4432 else if (type == _MEMSWAP)
4433 thresholds = &memcg->memsw_thresholds;
4437 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4439 /* Check if a threshold crossed before adding a new one */
4440 if (thresholds->primary)
4441 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4443 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4445 /* Allocate memory for new array of thresholds */
4446 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4454 /* Copy thresholds (if any) to new array */
4455 if (thresholds->primary) {
4456 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4457 sizeof(struct mem_cgroup_threshold));
4460 /* Add new threshold */
4461 new->entries[size - 1].eventfd = eventfd;
4462 new->entries[size - 1].threshold = threshold;
4464 /* Sort thresholds. Registering of new threshold isn't time-critical */
4465 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4466 compare_thresholds, NULL);
4468 /* Find current threshold */
4469 new->current_threshold = -1;
4470 for (i = 0; i < size; i++) {
4471 if (new->entries[i].threshold < usage) {
4473 * new->current_threshold will not be used until
4474 * rcu_assign_pointer(), so it's safe to increment
4477 ++new->current_threshold;
4481 /* Free old spare buffer and save old primary buffer as spare */
4482 kfree(thresholds->spare);
4483 thresholds->spare = thresholds->primary;
4485 rcu_assign_pointer(thresholds->primary, new);
4487 /* To be sure that nobody uses thresholds */
4491 mutex_unlock(&memcg->thresholds_lock);
4496 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4497 struct cftype *cft, struct eventfd_ctx *eventfd)
4499 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4500 struct mem_cgroup_thresholds *thresholds;
4501 struct mem_cgroup_threshold_ary *new;
4502 int type = MEMFILE_TYPE(cft->private);
4506 mutex_lock(&memcg->thresholds_lock);
4508 thresholds = &memcg->thresholds;
4509 else if (type == _MEMSWAP)
4510 thresholds = &memcg->memsw_thresholds;
4515 * Something went wrong if we trying to unregister a threshold
4516 * if we don't have thresholds
4518 BUG_ON(!thresholds);
4520 if (!thresholds->primary)
4523 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4525 /* Check if a threshold crossed before removing */
4526 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4528 /* Calculate new number of threshold */
4530 for (i = 0; i < thresholds->primary->size; i++) {
4531 if (thresholds->primary->entries[i].eventfd != eventfd)
4535 new = thresholds->spare;
4537 /* Set thresholds array to NULL if we don't have thresholds */
4546 /* Copy thresholds and find current threshold */
4547 new->current_threshold = -1;
4548 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4549 if (thresholds->primary->entries[i].eventfd == eventfd)
4552 new->entries[j] = thresholds->primary->entries[i];
4553 if (new->entries[j].threshold < usage) {
4555 * new->current_threshold will not be used
4556 * until rcu_assign_pointer(), so it's safe to increment
4559 ++new->current_threshold;
4565 /* Swap primary and spare array */
4566 thresholds->spare = thresholds->primary;
4568 rcu_assign_pointer(thresholds->primary, new);
4570 /* To be sure that nobody uses thresholds */
4573 /* If all events are unregistered, free the spare array */
4575 kfree(thresholds->spare);
4576 thresholds->spare = NULL;
4579 mutex_unlock(&memcg->thresholds_lock);
4582 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4583 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4585 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4586 struct mem_cgroup_eventfd_list *event;
4587 int type = MEMFILE_TYPE(cft->private);
4589 BUG_ON(type != _OOM_TYPE);
4590 event = kmalloc(sizeof(*event), GFP_KERNEL);
4594 spin_lock(&memcg_oom_lock);
4596 event->eventfd = eventfd;
4597 list_add(&event->list, &memcg->oom_notify);
4599 /* already in OOM ? */
4600 if (atomic_read(&memcg->under_oom))
4601 eventfd_signal(eventfd, 1);
4602 spin_unlock(&memcg_oom_lock);
4607 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4608 struct cftype *cft, struct eventfd_ctx *eventfd)
4610 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4611 struct mem_cgroup_eventfd_list *ev, *tmp;
4612 int type = MEMFILE_TYPE(cft->private);
4614 BUG_ON(type != _OOM_TYPE);
4616 spin_lock(&memcg_oom_lock);
4618 list_for_each_entry_safe(ev, tmp, &memcg->oom_notify, list) {
4619 if (ev->eventfd == eventfd) {
4620 list_del(&ev->list);
4625 spin_unlock(&memcg_oom_lock);
4628 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4629 struct cftype *cft, struct cgroup_map_cb *cb)
4631 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4633 cb->fill(cb, "oom_kill_disable", memcg->oom_kill_disable);
4635 if (atomic_read(&memcg->under_oom))
4636 cb->fill(cb, "under_oom", 1);
4638 cb->fill(cb, "under_oom", 0);
4642 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4643 struct cftype *cft, u64 val)
4645 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4646 struct mem_cgroup *parent;
4648 /* cannot set to root cgroup and only 0 and 1 are allowed */
4649 if (!cgrp->parent || !((val == 0) || (val == 1)))
4652 parent = mem_cgroup_from_cont(cgrp->parent);
4655 /* oom-kill-disable is a flag for subhierarchy. */
4656 if ((parent->use_hierarchy) ||
4657 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
4661 memcg->oom_kill_disable = val;
4663 memcg_oom_recover(memcg);
4669 static const struct file_operations mem_control_numa_stat_file_operations = {
4671 .llseek = seq_lseek,
4672 .release = single_release,
4675 static int mem_control_numa_stat_open(struct inode *unused, struct file *file)
4677 struct cgroup *cont = file->f_dentry->d_parent->d_fsdata;
4679 file->f_op = &mem_control_numa_stat_file_operations;
4680 return single_open(file, mem_control_numa_stat_show, cont);
4682 #endif /* CONFIG_NUMA */
4684 static struct cftype mem_cgroup_files[] = {
4686 .name = "usage_in_bytes",
4687 .private = MEMFILE_PRIVATE(_MEM, 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 = "max_usage_in_bytes",
4694 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4695 .trigger = mem_cgroup_reset,
4696 .read_u64 = mem_cgroup_read,
4699 .name = "limit_in_bytes",
4700 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4701 .write_string = mem_cgroup_write,
4702 .read_u64 = mem_cgroup_read,
4705 .name = "soft_limit_in_bytes",
4706 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4707 .write_string = mem_cgroup_write,
4708 .read_u64 = mem_cgroup_read,
4712 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4713 .trigger = mem_cgroup_reset,
4714 .read_u64 = mem_cgroup_read,
4718 .read_map = mem_control_stat_show,
4721 .name = "force_empty",
4722 .trigger = mem_cgroup_force_empty_write,
4725 .name = "use_hierarchy",
4726 .write_u64 = mem_cgroup_hierarchy_write,
4727 .read_u64 = mem_cgroup_hierarchy_read,
4730 .name = "swappiness",
4731 .read_u64 = mem_cgroup_swappiness_read,
4732 .write_u64 = mem_cgroup_swappiness_write,
4735 .name = "move_charge_at_immigrate",
4736 .read_u64 = mem_cgroup_move_charge_read,
4737 .write_u64 = mem_cgroup_move_charge_write,
4740 .name = "oom_control",
4741 .read_map = mem_cgroup_oom_control_read,
4742 .write_u64 = mem_cgroup_oom_control_write,
4743 .register_event = mem_cgroup_oom_register_event,
4744 .unregister_event = mem_cgroup_oom_unregister_event,
4745 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4749 .name = "numa_stat",
4750 .open = mem_control_numa_stat_open,
4756 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4757 static struct cftype memsw_cgroup_files[] = {
4759 .name = "memsw.usage_in_bytes",
4760 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4761 .read_u64 = mem_cgroup_read,
4762 .register_event = mem_cgroup_usage_register_event,
4763 .unregister_event = mem_cgroup_usage_unregister_event,
4766 .name = "memsw.max_usage_in_bytes",
4767 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4768 .trigger = mem_cgroup_reset,
4769 .read_u64 = mem_cgroup_read,
4772 .name = "memsw.limit_in_bytes",
4773 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4774 .write_string = mem_cgroup_write,
4775 .read_u64 = mem_cgroup_read,
4778 .name = "memsw.failcnt",
4779 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4780 .trigger = mem_cgroup_reset,
4781 .read_u64 = mem_cgroup_read,
4785 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4787 if (!do_swap_account)
4789 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4790 ARRAY_SIZE(memsw_cgroup_files));
4793 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4799 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4801 struct mem_cgroup_per_node *pn;
4802 struct mem_cgroup_per_zone *mz;
4804 int zone, tmp = node;
4806 * This routine is called against possible nodes.
4807 * But it's BUG to call kmalloc() against offline node.
4809 * TODO: this routine can waste much memory for nodes which will
4810 * never be onlined. It's better to use memory hotplug callback
4813 if (!node_state(node, N_NORMAL_MEMORY))
4815 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4819 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4820 mz = &pn->zoneinfo[zone];
4822 INIT_LIST_HEAD(&mz->lists[l]);
4823 mz->usage_in_excess = 0;
4824 mz->on_tree = false;
4827 memcg->info.nodeinfo[node] = pn;
4831 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *memcg, int node)
4833 kfree(memcg->info.nodeinfo[node]);
4836 static struct mem_cgroup *mem_cgroup_alloc(void)
4838 struct mem_cgroup *mem;
4839 int size = sizeof(struct mem_cgroup);
4841 /* Can be very big if MAX_NUMNODES is very big */
4842 if (size < PAGE_SIZE)
4843 mem = kzalloc(size, GFP_KERNEL);
4845 mem = vzalloc(size);
4850 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4853 spin_lock_init(&mem->pcp_counter_lock);
4857 if (size < PAGE_SIZE)
4865 * At destroying mem_cgroup, references from swap_cgroup can remain.
4866 * (scanning all at force_empty is too costly...)
4868 * Instead of clearing all references at force_empty, we remember
4869 * the number of reference from swap_cgroup and free mem_cgroup when
4870 * it goes down to 0.
4872 * Removal of cgroup itself succeeds regardless of refs from swap.
4875 static void __mem_cgroup_free(struct mem_cgroup *memcg)
4879 mem_cgroup_remove_from_trees(memcg);
4880 free_css_id(&mem_cgroup_subsys, &memcg->css);
4882 for_each_node_state(node, N_POSSIBLE)
4883 free_mem_cgroup_per_zone_info(memcg, node);
4885 free_percpu(memcg->stat);
4886 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4892 static void mem_cgroup_get(struct mem_cgroup *memcg)
4894 atomic_inc(&memcg->refcnt);
4897 static void __mem_cgroup_put(struct mem_cgroup *memcg, int count)
4899 if (atomic_sub_and_test(count, &memcg->refcnt)) {
4900 struct mem_cgroup *parent = parent_mem_cgroup(memcg);
4901 __mem_cgroup_free(memcg);
4903 mem_cgroup_put(parent);
4907 static void mem_cgroup_put(struct mem_cgroup *memcg)
4909 __mem_cgroup_put(memcg, 1);
4913 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4915 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *memcg)
4917 if (!memcg->res.parent)
4919 return mem_cgroup_from_res_counter(memcg->res.parent, res);
4922 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4923 static void __init enable_swap_cgroup(void)
4925 if (!mem_cgroup_disabled() && really_do_swap_account)
4926 do_swap_account = 1;
4929 static void __init enable_swap_cgroup(void)
4934 static int mem_cgroup_soft_limit_tree_init(void)
4936 struct mem_cgroup_tree_per_node *rtpn;
4937 struct mem_cgroup_tree_per_zone *rtpz;
4938 int tmp, node, zone;
4940 for_each_node_state(node, N_POSSIBLE) {
4942 if (!node_state(node, N_NORMAL_MEMORY))
4944 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4948 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4950 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4951 rtpz = &rtpn->rb_tree_per_zone[zone];
4952 rtpz->rb_root = RB_ROOT;
4953 spin_lock_init(&rtpz->lock);
4959 static struct cgroup_subsys_state * __ref
4960 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4962 struct mem_cgroup *memcg, *parent;
4963 long error = -ENOMEM;
4966 memcg = mem_cgroup_alloc();
4968 return ERR_PTR(error);
4970 for_each_node_state(node, N_POSSIBLE)
4971 if (alloc_mem_cgroup_per_zone_info(memcg, node))
4975 if (cont->parent == NULL) {
4977 enable_swap_cgroup();
4979 if (mem_cgroup_soft_limit_tree_init())
4981 root_mem_cgroup = memcg;
4982 for_each_possible_cpu(cpu) {
4983 struct memcg_stock_pcp *stock =
4984 &per_cpu(memcg_stock, cpu);
4985 INIT_WORK(&stock->work, drain_local_stock);
4987 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4989 parent = mem_cgroup_from_cont(cont->parent);
4990 memcg->use_hierarchy = parent->use_hierarchy;
4991 memcg->oom_kill_disable = parent->oom_kill_disable;
4994 if (parent && parent->use_hierarchy) {
4995 res_counter_init(&memcg->res, &parent->res);
4996 res_counter_init(&memcg->memsw, &parent->memsw);
4998 * We increment refcnt of the parent to ensure that we can
4999 * safely access it on res_counter_charge/uncharge.
5000 * This refcnt will be decremented when freeing this
5001 * mem_cgroup(see mem_cgroup_put).
5003 mem_cgroup_get(parent);
5005 res_counter_init(&memcg->res, NULL);
5006 res_counter_init(&memcg->memsw, NULL);
5008 memcg->last_scanned_child = 0;
5009 memcg->last_scanned_node = MAX_NUMNODES;
5010 INIT_LIST_HEAD(&memcg->oom_notify);
5013 memcg->swappiness = mem_cgroup_swappiness(parent);
5014 atomic_set(&memcg->refcnt, 1);
5015 memcg->move_charge_at_immigrate = 0;
5016 mutex_init(&memcg->thresholds_lock);
5019 __mem_cgroup_free(memcg);
5020 return ERR_PTR(error);
5023 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
5024 struct cgroup *cont)
5026 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5028 return mem_cgroup_force_empty(memcg, false);
5031 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
5032 struct cgroup *cont)
5034 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
5036 mem_cgroup_put(memcg);
5039 static int mem_cgroup_populate(struct cgroup_subsys *ss,
5040 struct cgroup *cont)
5044 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
5045 ARRAY_SIZE(mem_cgroup_files));
5048 ret = register_memsw_files(cont, ss);
5053 /* Handlers for move charge at task migration. */
5054 #define PRECHARGE_COUNT_AT_ONCE 256
5055 static int mem_cgroup_do_precharge(unsigned long count)
5058 int batch_count = PRECHARGE_COUNT_AT_ONCE;
5059 struct mem_cgroup *memcg = mc.to;
5061 if (mem_cgroup_is_root(memcg)) {
5062 mc.precharge += count;
5063 /* we don't need css_get for root */
5066 /* try to charge at once */
5068 struct res_counter *dummy;
5070 * "memcg" cannot be under rmdir() because we've already checked
5071 * by cgroup_lock_live_cgroup() that it is not removed and we
5072 * are still under the same cgroup_mutex. So we can postpone
5075 if (res_counter_charge(&memcg->res, PAGE_SIZE * count, &dummy))
5077 if (do_swap_account && res_counter_charge(&memcg->memsw,
5078 PAGE_SIZE * count, &dummy)) {
5079 res_counter_uncharge(&memcg->res, PAGE_SIZE * count);
5082 mc.precharge += count;
5086 /* fall back to one by one charge */
5088 if (signal_pending(current)) {
5092 if (!batch_count--) {
5093 batch_count = PRECHARGE_COUNT_AT_ONCE;
5096 ret = __mem_cgroup_try_charge(NULL,
5097 GFP_KERNEL, 1, &memcg, false);
5099 /* mem_cgroup_clear_mc() will do uncharge later */
5107 * is_target_pte_for_mc - check a pte whether it is valid for move charge
5108 * @vma: the vma the pte to be checked belongs
5109 * @addr: the address corresponding to the pte to be checked
5110 * @ptent: the pte to be checked
5111 * @target: the pointer the target page or swap ent will be stored(can be NULL)
5114 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
5115 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
5116 * move charge. if @target is not NULL, the page is stored in target->page
5117 * with extra refcnt got(Callers should handle it).
5118 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
5119 * target for charge migration. if @target is not NULL, the entry is stored
5122 * Called with pte lock held.
5129 enum mc_target_type {
5130 MC_TARGET_NONE, /* not used */
5135 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
5136 unsigned long addr, pte_t ptent)
5138 struct page *page = vm_normal_page(vma, addr, ptent);
5140 if (!page || !page_mapped(page))
5142 if (PageAnon(page)) {
5143 /* we don't move shared anon */
5144 if (!move_anon() || page_mapcount(page) > 2)
5146 } else if (!move_file())
5147 /* we ignore mapcount for file pages */
5149 if (!get_page_unless_zero(page))
5155 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
5156 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5159 struct page *page = NULL;
5160 swp_entry_t ent = pte_to_swp_entry(ptent);
5162 if (!move_anon() || non_swap_entry(ent))
5164 usage_count = mem_cgroup_count_swap_user(ent, &page);
5165 if (usage_count > 1) { /* we don't move shared anon */
5170 if (do_swap_account)
5171 entry->val = ent.val;
5176 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
5177 unsigned long addr, pte_t ptent, swp_entry_t *entry)
5179 struct page *page = NULL;
5180 struct inode *inode;
5181 struct address_space *mapping;
5184 if (!vma->vm_file) /* anonymous vma */
5189 inode = vma->vm_file->f_path.dentry->d_inode;
5190 mapping = vma->vm_file->f_mapping;
5191 if (pte_none(ptent))
5192 pgoff = linear_page_index(vma, addr);
5193 else /* pte_file(ptent) is true */
5194 pgoff = pte_to_pgoff(ptent);
5196 /* page is moved even if it's not RSS of this task(page-faulted). */
5197 page = find_get_page(mapping, pgoff);
5200 /* shmem/tmpfs may report page out on swap: account for that too. */
5201 if (radix_tree_exceptional_entry(page)) {
5202 swp_entry_t swap = radix_to_swp_entry(page);
5203 if (do_swap_account)
5205 page = find_get_page(&swapper_space, swap.val);
5211 static int is_target_pte_for_mc(struct vm_area_struct *vma,
5212 unsigned long addr, pte_t ptent, union mc_target *target)
5214 struct page *page = NULL;
5215 struct page_cgroup *pc;
5217 swp_entry_t ent = { .val = 0 };
5219 if (pte_present(ptent))
5220 page = mc_handle_present_pte(vma, addr, ptent);
5221 else if (is_swap_pte(ptent))
5222 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
5223 else if (pte_none(ptent) || pte_file(ptent))
5224 page = mc_handle_file_pte(vma, addr, ptent, &ent);
5226 if (!page && !ent.val)
5229 pc = lookup_page_cgroup(page);
5231 * Do only loose check w/o page_cgroup lock.
5232 * mem_cgroup_move_account() checks the pc is valid or not under
5235 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
5236 ret = MC_TARGET_PAGE;
5238 target->page = page;
5240 if (!ret || !target)
5243 /* There is a swap entry and a page doesn't exist or isn't charged */
5244 if (ent.val && !ret &&
5245 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
5246 ret = MC_TARGET_SWAP;
5253 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
5254 unsigned long addr, unsigned long end,
5255 struct mm_walk *walk)
5257 struct vm_area_struct *vma = walk->private;
5261 split_huge_page_pmd(walk->mm, pmd);
5262 if (pmd_trans_unstable(pmd))
5265 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5266 for (; addr != end; pte++, addr += PAGE_SIZE)
5267 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
5268 mc.precharge++; /* increment precharge temporarily */
5269 pte_unmap_unlock(pte - 1, ptl);
5275 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
5277 unsigned long precharge;
5278 struct vm_area_struct *vma;
5280 down_read(&mm->mmap_sem);
5281 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5282 struct mm_walk mem_cgroup_count_precharge_walk = {
5283 .pmd_entry = mem_cgroup_count_precharge_pte_range,
5287 if (is_vm_hugetlb_page(vma))
5289 walk_page_range(vma->vm_start, vma->vm_end,
5290 &mem_cgroup_count_precharge_walk);
5292 up_read(&mm->mmap_sem);
5294 precharge = mc.precharge;
5300 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
5302 unsigned long precharge = mem_cgroup_count_precharge(mm);
5304 VM_BUG_ON(mc.moving_task);
5305 mc.moving_task = current;
5306 return mem_cgroup_do_precharge(precharge);
5309 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
5310 static void __mem_cgroup_clear_mc(void)
5312 struct mem_cgroup *from = mc.from;
5313 struct mem_cgroup *to = mc.to;
5315 /* we must uncharge all the leftover precharges from mc.to */
5317 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
5321 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
5322 * we must uncharge here.
5324 if (mc.moved_charge) {
5325 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
5326 mc.moved_charge = 0;
5328 /* we must fixup refcnts and charges */
5329 if (mc.moved_swap) {
5330 /* uncharge swap account from the old cgroup */
5331 if (!mem_cgroup_is_root(mc.from))
5332 res_counter_uncharge(&mc.from->memsw,
5333 PAGE_SIZE * mc.moved_swap);
5334 __mem_cgroup_put(mc.from, mc.moved_swap);
5336 if (!mem_cgroup_is_root(mc.to)) {
5338 * we charged both to->res and to->memsw, so we should
5341 res_counter_uncharge(&mc.to->res,
5342 PAGE_SIZE * mc.moved_swap);
5344 /* we've already done mem_cgroup_get(mc.to) */
5347 memcg_oom_recover(from);
5348 memcg_oom_recover(to);
5349 wake_up_all(&mc.waitq);
5352 static void mem_cgroup_clear_mc(void)
5354 struct mem_cgroup *from = mc.from;
5357 * we must clear moving_task before waking up waiters at the end of
5360 mc.moving_task = NULL;
5361 __mem_cgroup_clear_mc();
5362 spin_lock(&mc.lock);
5365 spin_unlock(&mc.lock);
5366 mem_cgroup_end_move(from);
5369 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5370 struct cgroup *cgroup,
5371 struct task_struct *p)
5374 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgroup);
5376 if (memcg->move_charge_at_immigrate) {
5377 struct mm_struct *mm;
5378 struct mem_cgroup *from = mem_cgroup_from_task(p);
5380 VM_BUG_ON(from == memcg);
5382 mm = get_task_mm(p);
5385 /* We move charges only when we move a owner of the mm */
5386 if (mm->owner == p) {
5389 VM_BUG_ON(mc.precharge);
5390 VM_BUG_ON(mc.moved_charge);
5391 VM_BUG_ON(mc.moved_swap);
5392 mem_cgroup_start_move(from);
5393 spin_lock(&mc.lock);
5396 spin_unlock(&mc.lock);
5397 /* We set mc.moving_task later */
5399 ret = mem_cgroup_precharge_mc(mm);
5401 mem_cgroup_clear_mc();
5408 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5409 struct cgroup *cgroup,
5410 struct task_struct *p)
5412 mem_cgroup_clear_mc();
5415 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
5416 unsigned long addr, unsigned long end,
5417 struct mm_walk *walk)
5420 struct vm_area_struct *vma = walk->private;
5424 split_huge_page_pmd(walk->mm, pmd);
5425 if (pmd_trans_unstable(pmd))
5428 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
5429 for (; addr != end; addr += PAGE_SIZE) {
5430 pte_t ptent = *(pte++);
5431 union mc_target target;
5434 struct page_cgroup *pc;
5440 type = is_target_pte_for_mc(vma, addr, ptent, &target);
5442 case MC_TARGET_PAGE:
5444 if (isolate_lru_page(page))
5446 pc = lookup_page_cgroup(page);
5447 if (!mem_cgroup_move_account(page, 1, pc,
5448 mc.from, mc.to, false)) {
5450 /* we uncharge from mc.from later. */
5453 putback_lru_page(page);
5454 put: /* is_target_pte_for_mc() gets the page */
5457 case MC_TARGET_SWAP:
5459 if (!mem_cgroup_move_swap_account(ent,
5460 mc.from, mc.to, false)) {
5462 /* we fixup refcnts and charges later. */
5470 pte_unmap_unlock(pte - 1, ptl);
5475 * We have consumed all precharges we got in can_attach().
5476 * We try charge one by one, but don't do any additional
5477 * charges to mc.to if we have failed in charge once in attach()
5480 ret = mem_cgroup_do_precharge(1);
5488 static void mem_cgroup_move_charge(struct mm_struct *mm)
5490 struct vm_area_struct *vma;
5492 lru_add_drain_all();
5494 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5496 * Someone who are holding the mmap_sem might be waiting in
5497 * waitq. So we cancel all extra charges, wake up all waiters,
5498 * and retry. Because we cancel precharges, we might not be able
5499 * to move enough charges, but moving charge is a best-effort
5500 * feature anyway, so it wouldn't be a big problem.
5502 __mem_cgroup_clear_mc();
5506 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5508 struct mm_walk mem_cgroup_move_charge_walk = {
5509 .pmd_entry = mem_cgroup_move_charge_pte_range,
5513 if (is_vm_hugetlb_page(vma))
5515 ret = walk_page_range(vma->vm_start, vma->vm_end,
5516 &mem_cgroup_move_charge_walk);
5519 * means we have consumed all precharges and failed in
5520 * doing additional charge. Just abandon here.
5524 up_read(&mm->mmap_sem);
5527 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5528 struct cgroup *cont,
5529 struct cgroup *old_cont,
5530 struct task_struct *p)
5532 struct mm_struct *mm = get_task_mm(p);
5536 mem_cgroup_move_charge(mm);
5541 mem_cgroup_clear_mc();
5543 #else /* !CONFIG_MMU */
5544 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5545 struct cgroup *cgroup,
5546 struct task_struct *p)
5550 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5551 struct cgroup *cgroup,
5552 struct task_struct *p)
5555 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5556 struct cgroup *cont,
5557 struct cgroup *old_cont,
5558 struct task_struct *p)
5563 struct cgroup_subsys mem_cgroup_subsys = {
5565 .subsys_id = mem_cgroup_subsys_id,
5566 .create = mem_cgroup_create,
5567 .pre_destroy = mem_cgroup_pre_destroy,
5568 .destroy = mem_cgroup_destroy,
5569 .populate = mem_cgroup_populate,
5570 .can_attach = mem_cgroup_can_attach,
5571 .cancel_attach = mem_cgroup_cancel_attach,
5572 .attach = mem_cgroup_move_task,
5577 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5578 static int __init enable_swap_account(char *s)
5580 /* consider enabled if no parameter or 1 is given */
5581 if (!strcmp(s, "1"))
5582 really_do_swap_account = 1;
5583 else if (!strcmp(s, "0"))
5584 really_do_swap_account = 0;
5587 __setup("swapaccount=", enable_swap_account);