1 /* memcontrol.c - Memory Controller
3 * Copyright IBM Corporation, 2007
4 * Author Balbir Singh <balbir@linux.vnet.ibm.com>
6 * Copyright 2007 OpenVZ SWsoft Inc
7 * Author: Pavel Emelianov <xemul@openvz.org>
10 * Copyright (C) 2009 Nokia Corporation
11 * Author: Kirill A. Shutemov
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License as published by
15 * the Free Software Foundation; either version 2 of the License, or
16 * (at your option) any later version.
18 * This program is distributed in the hope that it will be useful,
19 * but WITHOUT ANY WARRANTY; without even the implied warranty of
20 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
21 * GNU General Public License for more details.
24 #include <linux/res_counter.h>
25 #include <linux/memcontrol.h>
26 #include <linux/cgroup.h>
28 #include <linux/hugetlb.h>
29 #include <linux/pagemap.h>
30 #include <linux/smp.h>
31 #include <linux/page-flags.h>
32 #include <linux/backing-dev.h>
33 #include <linux/bit_spinlock.h>
34 #include <linux/rcupdate.h>
35 #include <linux/limits.h>
36 #include <linux/mutex.h>
37 #include <linux/rbtree.h>
38 #include <linux/slab.h>
39 #include <linux/swap.h>
40 #include <linux/swapops.h>
41 #include <linux/spinlock.h>
42 #include <linux/eventfd.h>
43 #include <linux/sort.h>
45 #include <linux/seq_file.h>
46 #include <linux/vmalloc.h>
47 #include <linux/mm_inline.h>
48 #include <linux/page_cgroup.h>
49 #include <linux/cpu.h>
52 #include <asm/uaccess.h>
54 #include <trace/events/vmscan.h>
56 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
57 #define MEM_CGROUP_RECLAIM_RETRIES 5
58 struct mem_cgroup *root_mem_cgroup __read_mostly;
60 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
61 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
62 int do_swap_account __read_mostly;
63 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
65 #define do_swap_account (0)
69 * Per memcg event counter is incremented at every pagein/pageout. This counter
70 * is used for trigger some periodic events. This is straightforward and better
71 * than using jiffies etc. to handle periodic memcg event.
73 * These values will be used as !((event) & ((1 <<(thresh)) - 1))
75 #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
76 #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
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_PGPGIN_COUNT, /* # of pages paged in */
89 MEM_CGROUP_STAT_PGPGOUT_COUNT, /* # of pages paged out */
90 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
91 MEM_CGROUP_EVENTS, /* incremented at every pagein/pageout */
93 MEM_CGROUP_STAT_NSTATS,
96 struct mem_cgroup_stat_cpu {
97 s64 count[MEM_CGROUP_STAT_NSTATS];
101 * per-zone information in memory controller.
103 struct mem_cgroup_per_zone {
105 * spin_lock to protect the per cgroup LRU
107 struct list_head lists[NR_LRU_LISTS];
108 unsigned long count[NR_LRU_LISTS];
110 struct zone_reclaim_stat reclaim_stat;
111 struct rb_node tree_node; /* RB tree node */
112 unsigned long long usage_in_excess;/* Set to the value by which */
113 /* the soft limit is exceeded*/
115 struct mem_cgroup *mem; /* Back pointer, we cannot */
116 /* use container_of */
118 /* Macro for accessing counter */
119 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
121 struct mem_cgroup_per_node {
122 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
125 struct mem_cgroup_lru_info {
126 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
130 * Cgroups above their limits are maintained in a RB-Tree, independent of
131 * their hierarchy representation
134 struct mem_cgroup_tree_per_zone {
135 struct rb_root rb_root;
139 struct mem_cgroup_tree_per_node {
140 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
143 struct mem_cgroup_tree {
144 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
147 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
149 struct mem_cgroup_threshold {
150 struct eventfd_ctx *eventfd;
155 struct mem_cgroup_threshold_ary {
156 /* An array index points to threshold just below usage. */
157 int current_threshold;
158 /* Size of entries[] */
160 /* Array of thresholds */
161 struct mem_cgroup_threshold entries[0];
164 struct mem_cgroup_thresholds {
165 /* Primary thresholds array */
166 struct mem_cgroup_threshold_ary *primary;
168 * Spare threshold array.
169 * This is needed to make mem_cgroup_unregister_event() "never fail".
170 * It must be able to store at least primary->size - 1 entries.
172 struct mem_cgroup_threshold_ary *spare;
176 struct mem_cgroup_eventfd_list {
177 struct list_head list;
178 struct eventfd_ctx *eventfd;
181 static void mem_cgroup_threshold(struct mem_cgroup *mem);
182 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
185 * The memory controller data structure. The memory controller controls both
186 * page cache and RSS per cgroup. We would eventually like to provide
187 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
188 * to help the administrator determine what knobs to tune.
190 * TODO: Add a water mark for the memory controller. Reclaim will begin when
191 * we hit the water mark. May be even add a low water mark, such that
192 * no reclaim occurs from a cgroup at it's low water mark, this is
193 * a feature that will be implemented much later in the future.
196 struct cgroup_subsys_state css;
198 * the counter to account for memory usage
200 struct res_counter res;
202 * the counter to account for mem+swap usage.
204 struct res_counter memsw;
206 * Per cgroup active and inactive list, similar to the
207 * per zone LRU lists.
209 struct mem_cgroup_lru_info info;
212 protect against reclaim related member.
214 spinlock_t reclaim_param_lock;
217 * While reclaiming in a hierarchy, we cache the last child we
220 int last_scanned_child;
222 * Should the accounting and control be hierarchical, per subtree?
228 unsigned int swappiness;
229 /* OOM-Killer disable */
230 int oom_kill_disable;
232 /* set when res.limit == memsw.limit */
233 bool memsw_is_minimum;
235 /* protect arrays of thresholds */
236 struct mutex thresholds_lock;
238 /* thresholds for memory usage. RCU-protected */
239 struct mem_cgroup_thresholds thresholds;
241 /* thresholds for mem+swap usage. RCU-protected */
242 struct mem_cgroup_thresholds memsw_thresholds;
244 /* For oom notifier event fd */
245 struct list_head oom_notify;
248 * Should we move charges of a task when a task is moved into this
249 * mem_cgroup ? And what type of charges should we move ?
251 unsigned long move_charge_at_immigrate;
255 struct mem_cgroup_stat_cpu *stat;
258 /* Stuffs for move charges at task migration. */
260 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
261 * left-shifted bitmap of these types.
264 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
265 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
269 /* "mc" and its members are protected by cgroup_mutex */
270 static struct move_charge_struct {
271 spinlock_t lock; /* for from, to, moving_task */
272 struct mem_cgroup *from;
273 struct mem_cgroup *to;
274 unsigned long precharge;
275 unsigned long moved_charge;
276 unsigned long moved_swap;
277 struct task_struct *moving_task; /* a task moving charges */
278 wait_queue_head_t waitq; /* a waitq for other context */
280 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
281 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
284 static bool move_anon(void)
286 return test_bit(MOVE_CHARGE_TYPE_ANON,
287 &mc.to->move_charge_at_immigrate);
290 static bool move_file(void)
292 return test_bit(MOVE_CHARGE_TYPE_FILE,
293 &mc.to->move_charge_at_immigrate);
297 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
298 * limit reclaim to prevent infinite loops, if they ever occur.
300 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
301 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
304 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
305 MEM_CGROUP_CHARGE_TYPE_MAPPED,
306 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
307 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
308 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
309 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
313 /* only for here (for easy reading.) */
314 #define PCGF_CACHE (1UL << PCG_CACHE)
315 #define PCGF_USED (1UL << PCG_USED)
316 #define PCGF_LOCK (1UL << PCG_LOCK)
317 /* Not used, but added here for completeness */
318 #define PCGF_ACCT (1UL << PCG_ACCT)
320 /* for encoding cft->private value on file */
323 #define _OOM_TYPE (2)
324 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
325 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
326 #define MEMFILE_ATTR(val) ((val) & 0xffff)
327 /* Used for OOM nofiier */
328 #define OOM_CONTROL (0)
331 * Reclaim flags for mem_cgroup_hierarchical_reclaim
333 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
334 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
335 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
336 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
337 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
338 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
340 static void mem_cgroup_get(struct mem_cgroup *mem);
341 static void mem_cgroup_put(struct mem_cgroup *mem);
342 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
343 static void drain_all_stock_async(void);
345 static struct mem_cgroup_per_zone *
346 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
348 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
351 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
356 static struct mem_cgroup_per_zone *
357 page_cgroup_zoneinfo(struct page_cgroup *pc)
359 struct mem_cgroup *mem = pc->mem_cgroup;
360 int nid = page_cgroup_nid(pc);
361 int zid = page_cgroup_zid(pc);
366 return mem_cgroup_zoneinfo(mem, nid, zid);
369 static struct mem_cgroup_tree_per_zone *
370 soft_limit_tree_node_zone(int nid, int zid)
372 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
375 static struct mem_cgroup_tree_per_zone *
376 soft_limit_tree_from_page(struct page *page)
378 int nid = page_to_nid(page);
379 int zid = page_zonenum(page);
381 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
385 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
386 struct mem_cgroup_per_zone *mz,
387 struct mem_cgroup_tree_per_zone *mctz,
388 unsigned long long new_usage_in_excess)
390 struct rb_node **p = &mctz->rb_root.rb_node;
391 struct rb_node *parent = NULL;
392 struct mem_cgroup_per_zone *mz_node;
397 mz->usage_in_excess = new_usage_in_excess;
398 if (!mz->usage_in_excess)
402 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
404 if (mz->usage_in_excess < mz_node->usage_in_excess)
407 * We can't avoid mem cgroups that are over their soft
408 * limit by the same amount
410 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
413 rb_link_node(&mz->tree_node, parent, p);
414 rb_insert_color(&mz->tree_node, &mctz->rb_root);
419 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
420 struct mem_cgroup_per_zone *mz,
421 struct mem_cgroup_tree_per_zone *mctz)
425 rb_erase(&mz->tree_node, &mctz->rb_root);
430 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
431 struct mem_cgroup_per_zone *mz,
432 struct mem_cgroup_tree_per_zone *mctz)
434 spin_lock(&mctz->lock);
435 __mem_cgroup_remove_exceeded(mem, mz, mctz);
436 spin_unlock(&mctz->lock);
440 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
442 unsigned long long excess;
443 struct mem_cgroup_per_zone *mz;
444 struct mem_cgroup_tree_per_zone *mctz;
445 int nid = page_to_nid(page);
446 int zid = page_zonenum(page);
447 mctz = soft_limit_tree_from_page(page);
450 * Necessary to update all ancestors when hierarchy is used.
451 * because their event counter is not touched.
453 for (; mem; mem = parent_mem_cgroup(mem)) {
454 mz = mem_cgroup_zoneinfo(mem, nid, zid);
455 excess = res_counter_soft_limit_excess(&mem->res);
457 * We have to update the tree if mz is on RB-tree or
458 * mem is over its softlimit.
460 if (excess || mz->on_tree) {
461 spin_lock(&mctz->lock);
462 /* if on-tree, remove it */
464 __mem_cgroup_remove_exceeded(mem, mz, mctz);
466 * Insert again. mz->usage_in_excess will be updated.
467 * If excess is 0, no tree ops.
469 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
470 spin_unlock(&mctz->lock);
475 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
478 struct mem_cgroup_per_zone *mz;
479 struct mem_cgroup_tree_per_zone *mctz;
481 for_each_node_state(node, N_POSSIBLE) {
482 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
483 mz = mem_cgroup_zoneinfo(mem, node, zone);
484 mctz = soft_limit_tree_node_zone(node, zone);
485 mem_cgroup_remove_exceeded(mem, mz, mctz);
490 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
492 return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
495 static struct mem_cgroup_per_zone *
496 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
498 struct rb_node *rightmost = NULL;
499 struct mem_cgroup_per_zone *mz;
503 rightmost = rb_last(&mctz->rb_root);
505 goto done; /* Nothing to reclaim from */
507 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
509 * Remove the node now but someone else can add it back,
510 * we will to add it back at the end of reclaim to its correct
511 * position in the tree.
513 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
514 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
515 !css_tryget(&mz->mem->css))
521 static struct mem_cgroup_per_zone *
522 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
524 struct mem_cgroup_per_zone *mz;
526 spin_lock(&mctz->lock);
527 mz = __mem_cgroup_largest_soft_limit_node(mctz);
528 spin_unlock(&mctz->lock);
532 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
533 enum mem_cgroup_stat_index idx)
538 for_each_possible_cpu(cpu)
539 val += per_cpu(mem->stat->count[idx], cpu);
543 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
547 ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
548 ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
552 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
555 int val = (charge) ? 1 : -1;
556 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
559 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
560 struct page_cgroup *pc,
563 int val = (charge) ? 1 : -1;
567 if (PageCgroupCache(pc))
568 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], val);
570 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], val);
573 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
575 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
576 __this_cpu_inc(mem->stat->count[MEM_CGROUP_EVENTS]);
581 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
585 struct mem_cgroup_per_zone *mz;
588 for_each_online_node(nid)
589 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
590 mz = mem_cgroup_zoneinfo(mem, nid, zid);
591 total += MEM_CGROUP_ZSTAT(mz, idx);
596 static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
600 val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
602 return !(val & ((1 << event_mask_shift) - 1));
606 * Check events in order.
609 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
611 /* threshold event is triggered in finer grain than soft limit */
612 if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
613 mem_cgroup_threshold(mem);
614 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
615 mem_cgroup_update_tree(mem, page);
619 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
621 return container_of(cgroup_subsys_state(cont,
622 mem_cgroup_subsys_id), struct mem_cgroup,
626 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
629 * mm_update_next_owner() may clear mm->owner to NULL
630 * if it races with swapoff, page migration, etc.
631 * So this can be called with p == NULL.
636 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
637 struct mem_cgroup, css);
640 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
642 struct mem_cgroup *mem = NULL;
647 * Because we have no locks, mm->owner's may be being moved to other
648 * cgroup. We use css_tryget() here even if this looks
649 * pessimistic (rather than adding locks here).
653 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
656 } while (!css_tryget(&mem->css));
662 * Call callback function against all cgroup under hierarchy tree.
664 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
665 int (*func)(struct mem_cgroup *, void *))
667 int found, ret, nextid;
668 struct cgroup_subsys_state *css;
669 struct mem_cgroup *mem;
671 if (!root->use_hierarchy)
672 return (*func)(root, data);
680 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
682 if (css && css_tryget(css))
683 mem = container_of(css, struct mem_cgroup, css);
687 ret = (*func)(mem, data);
691 } while (!ret && css);
696 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
698 return (mem == root_mem_cgroup);
702 * Following LRU functions are allowed to be used without PCG_LOCK.
703 * Operations are called by routine of global LRU independently from memcg.
704 * What we have to take care of here is validness of pc->mem_cgroup.
706 * Changes to pc->mem_cgroup happens when
709 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
710 * It is added to LRU before charge.
711 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
712 * When moving account, the page is not on LRU. It's isolated.
715 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
717 struct page_cgroup *pc;
718 struct mem_cgroup_per_zone *mz;
720 if (mem_cgroup_disabled())
722 pc = lookup_page_cgroup(page);
723 /* can happen while we handle swapcache. */
724 if (!TestClearPageCgroupAcctLRU(pc))
726 VM_BUG_ON(!pc->mem_cgroup);
728 * We don't check PCG_USED bit. It's cleared when the "page" is finally
729 * removed from global LRU.
731 mz = page_cgroup_zoneinfo(pc);
732 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
733 if (mem_cgroup_is_root(pc->mem_cgroup))
735 VM_BUG_ON(list_empty(&pc->lru));
736 list_del_init(&pc->lru);
740 void mem_cgroup_del_lru(struct page *page)
742 mem_cgroup_del_lru_list(page, page_lru(page));
745 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
747 struct mem_cgroup_per_zone *mz;
748 struct page_cgroup *pc;
750 if (mem_cgroup_disabled())
753 pc = lookup_page_cgroup(page);
755 * Used bit is set without atomic ops but after smp_wmb().
756 * For making pc->mem_cgroup visible, insert smp_rmb() here.
759 /* unused or root page is not rotated. */
760 if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
762 mz = page_cgroup_zoneinfo(pc);
763 list_move(&pc->lru, &mz->lists[lru]);
766 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
768 struct page_cgroup *pc;
769 struct mem_cgroup_per_zone *mz;
771 if (mem_cgroup_disabled())
773 pc = lookup_page_cgroup(page);
774 VM_BUG_ON(PageCgroupAcctLRU(pc));
776 * Used bit is set without atomic ops but after smp_wmb().
777 * For making pc->mem_cgroup visible, insert smp_rmb() here.
780 if (!PageCgroupUsed(pc))
783 mz = page_cgroup_zoneinfo(pc);
784 MEM_CGROUP_ZSTAT(mz, lru) += 1;
785 SetPageCgroupAcctLRU(pc);
786 if (mem_cgroup_is_root(pc->mem_cgroup))
788 list_add(&pc->lru, &mz->lists[lru]);
792 * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
793 * lru because the page may.be reused after it's fully uncharged (because of
794 * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
795 * it again. This function is only used to charge SwapCache. It's done under
796 * lock_page and expected that zone->lru_lock is never held.
798 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
801 struct zone *zone = page_zone(page);
802 struct page_cgroup *pc = lookup_page_cgroup(page);
804 spin_lock_irqsave(&zone->lru_lock, flags);
806 * Forget old LRU when this page_cgroup is *not* used. This Used bit
807 * is guarded by lock_page() because the page is SwapCache.
809 if (!PageCgroupUsed(pc))
810 mem_cgroup_del_lru_list(page, page_lru(page));
811 spin_unlock_irqrestore(&zone->lru_lock, flags);
814 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
817 struct zone *zone = page_zone(page);
818 struct page_cgroup *pc = lookup_page_cgroup(page);
820 spin_lock_irqsave(&zone->lru_lock, flags);
821 /* link when the page is linked to LRU but page_cgroup isn't */
822 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
823 mem_cgroup_add_lru_list(page, page_lru(page));
824 spin_unlock_irqrestore(&zone->lru_lock, flags);
828 void mem_cgroup_move_lists(struct page *page,
829 enum lru_list from, enum lru_list to)
831 if (mem_cgroup_disabled())
833 mem_cgroup_del_lru_list(page, from);
834 mem_cgroup_add_lru_list(page, to);
837 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
840 struct mem_cgroup *curr = NULL;
843 curr = try_get_mem_cgroup_from_mm(task->mm);
848 * We should check use_hierarchy of "mem" not "curr". Because checking
849 * use_hierarchy of "curr" here make this function true if hierarchy is
850 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
851 * hierarchy(even if use_hierarchy is disabled in "mem").
853 if (mem->use_hierarchy)
854 ret = css_is_ancestor(&curr->css, &mem->css);
861 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
863 unsigned long active;
864 unsigned long inactive;
866 unsigned long inactive_ratio;
868 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
869 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
871 gb = (inactive + active) >> (30 - PAGE_SHIFT);
873 inactive_ratio = int_sqrt(10 * gb);
878 present_pages[0] = inactive;
879 present_pages[1] = active;
882 return inactive_ratio;
885 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
887 unsigned long active;
888 unsigned long inactive;
889 unsigned long present_pages[2];
890 unsigned long inactive_ratio;
892 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
894 inactive = present_pages[0];
895 active = present_pages[1];
897 if (inactive * inactive_ratio < active)
903 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
905 unsigned long active;
906 unsigned long inactive;
908 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
909 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
911 return (active > inactive);
914 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
918 int nid = zone->zone_pgdat->node_id;
919 int zid = zone_idx(zone);
920 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
922 return MEM_CGROUP_ZSTAT(mz, lru);
925 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
928 int nid = zone->zone_pgdat->node_id;
929 int zid = zone_idx(zone);
930 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
932 return &mz->reclaim_stat;
935 struct zone_reclaim_stat *
936 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
938 struct page_cgroup *pc;
939 struct mem_cgroup_per_zone *mz;
941 if (mem_cgroup_disabled())
944 pc = lookup_page_cgroup(page);
946 * Used bit is set without atomic ops but after smp_wmb().
947 * For making pc->mem_cgroup visible, insert smp_rmb() here.
950 if (!PageCgroupUsed(pc))
953 mz = page_cgroup_zoneinfo(pc);
957 return &mz->reclaim_stat;
960 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
961 struct list_head *dst,
962 unsigned long *scanned, int order,
963 int mode, struct zone *z,
964 struct mem_cgroup *mem_cont,
965 int active, int file)
967 unsigned long nr_taken = 0;
971 struct list_head *src;
972 struct page_cgroup *pc, *tmp;
973 int nid = z->zone_pgdat->node_id;
974 int zid = zone_idx(z);
975 struct mem_cgroup_per_zone *mz;
976 int lru = LRU_FILE * file + active;
980 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
981 src = &mz->lists[lru];
984 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
985 if (scan >= nr_to_scan)
989 if (unlikely(!PageCgroupUsed(pc)))
991 if (unlikely(!PageLRU(page)))
995 ret = __isolate_lru_page(page, mode, file);
998 list_move(&page->lru, dst);
999 mem_cgroup_del_lru(page);
1003 /* we don't affect global LRU but rotate in our LRU */
1004 mem_cgroup_rotate_lru_list(page, page_lru(page));
1013 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1019 #define mem_cgroup_from_res_counter(counter, member) \
1020 container_of(counter, struct mem_cgroup, member)
1022 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
1024 if (do_swap_account) {
1025 if (res_counter_check_under_limit(&mem->res) &&
1026 res_counter_check_under_limit(&mem->memsw))
1029 if (res_counter_check_under_limit(&mem->res))
1034 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1036 struct cgroup *cgrp = memcg->css.cgroup;
1037 unsigned int swappiness;
1040 if (cgrp->parent == NULL)
1041 return vm_swappiness;
1043 spin_lock(&memcg->reclaim_param_lock);
1044 swappiness = memcg->swappiness;
1045 spin_unlock(&memcg->reclaim_param_lock);
1050 /* A routine for testing mem is not under move_account */
1052 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1054 struct mem_cgroup *from;
1055 struct mem_cgroup *to;
1058 * Unlike task_move routines, we access mc.to, mc.from not under
1059 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1061 spin_lock(&mc.lock);
1066 if (from == mem || to == mem
1067 || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1068 || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1071 spin_unlock(&mc.lock);
1075 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1077 if (mc.moving_task && current != mc.moving_task) {
1078 if (mem_cgroup_under_move(mem)) {
1080 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1081 /* moving charge context might have finished. */
1084 finish_wait(&mc.waitq, &wait);
1091 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
1099 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1100 * @memcg: The memory cgroup that went over limit
1101 * @p: Task that is going to be killed
1103 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1106 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1108 struct cgroup *task_cgrp;
1109 struct cgroup *mem_cgrp;
1111 * Need a buffer in BSS, can't rely on allocations. The code relies
1112 * on the assumption that OOM is serialized for memory controller.
1113 * If this assumption is broken, revisit this code.
1115 static char memcg_name[PATH_MAX];
1124 mem_cgrp = memcg->css.cgroup;
1125 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1127 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1130 * Unfortunately, we are unable to convert to a useful name
1131 * But we'll still print out the usage information
1138 printk(KERN_INFO "Task in %s killed", memcg_name);
1141 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1149 * Continues from above, so we don't need an KERN_ level
1151 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1154 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1155 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1156 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1157 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1158 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1160 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1161 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1162 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1166 * This function returns the number of memcg under hierarchy tree. Returns
1167 * 1(self count) if no children.
1169 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1172 mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
1177 * Return the memory (and swap, if configured) limit for a memcg.
1179 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1184 limit = res_counter_read_u64(&memcg->res, RES_LIMIT) +
1186 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1188 * If memsw is finite and limits the amount of swap space available
1189 * to this memcg, return that limit.
1191 return min(limit, memsw);
1195 * Visit the first child (need not be the first child as per the ordering
1196 * of the cgroup list, since we track last_scanned_child) of @mem and use
1197 * that to reclaim free pages from.
1199 static struct mem_cgroup *
1200 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1202 struct mem_cgroup *ret = NULL;
1203 struct cgroup_subsys_state *css;
1206 if (!root_mem->use_hierarchy) {
1207 css_get(&root_mem->css);
1213 nextid = root_mem->last_scanned_child + 1;
1214 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1216 if (css && css_tryget(css))
1217 ret = container_of(css, struct mem_cgroup, css);
1220 /* Updates scanning parameter */
1221 spin_lock(&root_mem->reclaim_param_lock);
1223 /* this means start scan from ID:1 */
1224 root_mem->last_scanned_child = 0;
1226 root_mem->last_scanned_child = found;
1227 spin_unlock(&root_mem->reclaim_param_lock);
1234 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1235 * we reclaimed from, so that we don't end up penalizing one child extensively
1236 * based on its position in the children list.
1238 * root_mem is the original ancestor that we've been reclaim from.
1240 * We give up and return to the caller when we visit root_mem twice.
1241 * (other groups can be removed while we're walking....)
1243 * If shrink==true, for avoiding to free too much, this returns immedieately.
1245 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1248 unsigned long reclaim_options)
1250 struct mem_cgroup *victim;
1253 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1254 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1255 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1256 unsigned long excess = mem_cgroup_get_excess(root_mem);
1258 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1259 if (root_mem->memsw_is_minimum)
1263 victim = mem_cgroup_select_victim(root_mem);
1264 if (victim == root_mem) {
1267 drain_all_stock_async();
1270 * If we have not been able to reclaim
1271 * anything, it might because there are
1272 * no reclaimable pages under this hierarchy
1274 if (!check_soft || !total) {
1275 css_put(&victim->css);
1279 * We want to do more targetted reclaim.
1280 * excess >> 2 is not to excessive so as to
1281 * reclaim too much, nor too less that we keep
1282 * coming back to reclaim from this cgroup
1284 if (total >= (excess >> 2) ||
1285 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1286 css_put(&victim->css);
1291 if (!mem_cgroup_local_usage(victim)) {
1292 /* this cgroup's local usage == 0 */
1293 css_put(&victim->css);
1296 /* we use swappiness of local cgroup */
1298 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1299 noswap, get_swappiness(victim), zone,
1300 zone->zone_pgdat->node_id);
1302 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1303 noswap, get_swappiness(victim));
1304 css_put(&victim->css);
1306 * At shrinking usage, we can't check we should stop here or
1307 * reclaim more. It's depends on callers. last_scanned_child
1308 * will work enough for keeping fairness under tree.
1314 if (res_counter_check_under_soft_limit(&root_mem->res))
1316 } else if (mem_cgroup_check_under_limit(root_mem))
1322 static int mem_cgroup_oom_lock_cb(struct mem_cgroup *mem, void *data)
1324 int *val = (int *)data;
1327 * Logically, we can stop scanning immediately when we find
1328 * a memcg is already locked. But condidering unlock ops and
1329 * creation/removal of memcg, scan-all is simple operation.
1331 x = atomic_inc_return(&mem->oom_lock);
1332 *val = max(x, *val);
1336 * Check OOM-Killer is already running under our hierarchy.
1337 * If someone is running, return false.
1339 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1343 mem_cgroup_walk_tree(mem, &lock_count, mem_cgroup_oom_lock_cb);
1345 if (lock_count == 1)
1350 static int mem_cgroup_oom_unlock_cb(struct mem_cgroup *mem, void *data)
1353 * When a new child is created while the hierarchy is under oom,
1354 * mem_cgroup_oom_lock() may not be called. We have to use
1355 * atomic_add_unless() here.
1357 atomic_add_unless(&mem->oom_lock, -1, 0);
1361 static void mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1363 mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_unlock_cb);
1366 static DEFINE_MUTEX(memcg_oom_mutex);
1367 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1369 struct oom_wait_info {
1370 struct mem_cgroup *mem;
1374 static int memcg_oom_wake_function(wait_queue_t *wait,
1375 unsigned mode, int sync, void *arg)
1377 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1378 struct oom_wait_info *oom_wait_info;
1380 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1382 if (oom_wait_info->mem == wake_mem)
1384 /* if no hierarchy, no match */
1385 if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1388 * Both of oom_wait_info->mem and wake_mem are stable under us.
1389 * Then we can use css_is_ancestor without taking care of RCU.
1391 if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1392 !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1396 return autoremove_wake_function(wait, mode, sync, arg);
1399 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1401 /* for filtering, pass "mem" as argument. */
1402 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1405 static void memcg_oom_recover(struct mem_cgroup *mem)
1407 if (mem && atomic_read(&mem->oom_lock))
1408 memcg_wakeup_oom(mem);
1412 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1414 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1416 struct oom_wait_info owait;
1417 bool locked, need_to_kill;
1420 owait.wait.flags = 0;
1421 owait.wait.func = memcg_oom_wake_function;
1422 owait.wait.private = current;
1423 INIT_LIST_HEAD(&owait.wait.task_list);
1424 need_to_kill = true;
1425 /* At first, try to OOM lock hierarchy under mem.*/
1426 mutex_lock(&memcg_oom_mutex);
1427 locked = mem_cgroup_oom_lock(mem);
1429 * Even if signal_pending(), we can't quit charge() loop without
1430 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1431 * under OOM is always welcomed, use TASK_KILLABLE here.
1433 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1434 if (!locked || mem->oom_kill_disable)
1435 need_to_kill = false;
1437 mem_cgroup_oom_notify(mem);
1438 mutex_unlock(&memcg_oom_mutex);
1441 finish_wait(&memcg_oom_waitq, &owait.wait);
1442 mem_cgroup_out_of_memory(mem, mask);
1445 finish_wait(&memcg_oom_waitq, &owait.wait);
1447 mutex_lock(&memcg_oom_mutex);
1448 mem_cgroup_oom_unlock(mem);
1449 memcg_wakeup_oom(mem);
1450 mutex_unlock(&memcg_oom_mutex);
1452 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1454 /* Give chance to dying process */
1455 schedule_timeout(1);
1460 * Currently used to update mapped file statistics, but the routine can be
1461 * generalized to update other statistics as well.
1463 void mem_cgroup_update_file_mapped(struct page *page, int val)
1465 struct mem_cgroup *mem;
1466 struct page_cgroup *pc;
1468 pc = lookup_page_cgroup(page);
1472 lock_page_cgroup(pc);
1473 mem = pc->mem_cgroup;
1474 if (!mem || !PageCgroupUsed(pc))
1478 * Preemption is already disabled. We can use __this_cpu_xxx
1481 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1482 SetPageCgroupFileMapped(pc);
1484 __this_cpu_dec(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1485 ClearPageCgroupFileMapped(pc);
1489 unlock_page_cgroup(pc);
1493 * size of first charge trial. "32" comes from vmscan.c's magic value.
1494 * TODO: maybe necessary to use big numbers in big irons.
1496 #define CHARGE_SIZE (32 * PAGE_SIZE)
1497 struct memcg_stock_pcp {
1498 struct mem_cgroup *cached; /* this never be root cgroup */
1500 struct work_struct work;
1502 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1503 static atomic_t memcg_drain_count;
1506 * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1507 * from local stock and true is returned. If the stock is 0 or charges from a
1508 * cgroup which is not current target, returns false. This stock will be
1511 static bool consume_stock(struct mem_cgroup *mem)
1513 struct memcg_stock_pcp *stock;
1516 stock = &get_cpu_var(memcg_stock);
1517 if (mem == stock->cached && stock->charge)
1518 stock->charge -= PAGE_SIZE;
1519 else /* need to call res_counter_charge */
1521 put_cpu_var(memcg_stock);
1526 * Returns stocks cached in percpu to res_counter and reset cached information.
1528 static void drain_stock(struct memcg_stock_pcp *stock)
1530 struct mem_cgroup *old = stock->cached;
1532 if (stock->charge) {
1533 res_counter_uncharge(&old->res, stock->charge);
1534 if (do_swap_account)
1535 res_counter_uncharge(&old->memsw, stock->charge);
1537 stock->cached = NULL;
1542 * This must be called under preempt disabled or must be called by
1543 * a thread which is pinned to local cpu.
1545 static void drain_local_stock(struct work_struct *dummy)
1547 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1552 * Cache charges(val) which is from res_counter, to local per_cpu area.
1553 * This will be consumed by consume_stock() function, later.
1555 static void refill_stock(struct mem_cgroup *mem, int val)
1557 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1559 if (stock->cached != mem) { /* reset if necessary */
1561 stock->cached = mem;
1563 stock->charge += val;
1564 put_cpu_var(memcg_stock);
1568 * Tries to drain stocked charges in other cpus. This function is asynchronous
1569 * and just put a work per cpu for draining localy on each cpu. Caller can
1570 * expects some charges will be back to res_counter later but cannot wait for
1573 static void drain_all_stock_async(void)
1576 /* This function is for scheduling "drain" in asynchronous way.
1577 * The result of "drain" is not directly handled by callers. Then,
1578 * if someone is calling drain, we don't have to call drain more.
1579 * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1580 * there is a race. We just do loose check here.
1582 if (atomic_read(&memcg_drain_count))
1584 /* Notify other cpus that system-wide "drain" is running */
1585 atomic_inc(&memcg_drain_count);
1587 for_each_online_cpu(cpu) {
1588 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1589 schedule_work_on(cpu, &stock->work);
1592 atomic_dec(&memcg_drain_count);
1593 /* We don't wait for flush_work */
1596 /* This is a synchronous drain interface. */
1597 static void drain_all_stock_sync(void)
1599 /* called when force_empty is called */
1600 atomic_inc(&memcg_drain_count);
1601 schedule_on_each_cpu(drain_local_stock);
1602 atomic_dec(&memcg_drain_count);
1605 static int __cpuinit memcg_stock_cpu_callback(struct notifier_block *nb,
1606 unsigned long action,
1609 int cpu = (unsigned long)hcpu;
1610 struct memcg_stock_pcp *stock;
1612 if (action != CPU_DEAD)
1614 stock = &per_cpu(memcg_stock, cpu);
1620 /* See __mem_cgroup_try_charge() for details */
1622 CHARGE_OK, /* success */
1623 CHARGE_RETRY, /* need to retry but retry is not bad */
1624 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
1625 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
1626 CHARGE_OOM_DIE, /* the current is killed because of OOM */
1629 static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
1630 int csize, bool oom_check)
1632 struct mem_cgroup *mem_over_limit;
1633 struct res_counter *fail_res;
1634 unsigned long flags = 0;
1637 ret = res_counter_charge(&mem->res, csize, &fail_res);
1640 if (!do_swap_account)
1642 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1646 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
1647 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1649 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
1651 if (csize > PAGE_SIZE) /* change csize and retry */
1652 return CHARGE_RETRY;
1654 if (!(gfp_mask & __GFP_WAIT))
1655 return CHARGE_WOULDBLOCK;
1657 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1660 * try_to_free_mem_cgroup_pages() might not give us a full
1661 * picture of reclaim. Some pages are reclaimed and might be
1662 * moved to swap cache or just unmapped from the cgroup.
1663 * Check the limit again to see if the reclaim reduced the
1664 * current usage of the cgroup before giving up
1666 if (ret || mem_cgroup_check_under_limit(mem_over_limit))
1667 return CHARGE_RETRY;
1670 * At task move, charge accounts can be doubly counted. So, it's
1671 * better to wait until the end of task_move if something is going on.
1673 if (mem_cgroup_wait_acct_move(mem_over_limit))
1674 return CHARGE_RETRY;
1676 /* If we don't need to call oom-killer at el, return immediately */
1678 return CHARGE_NOMEM;
1680 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
1681 return CHARGE_OOM_DIE;
1683 return CHARGE_RETRY;
1687 * Unlike exported interface, "oom" parameter is added. if oom==true,
1688 * oom-killer can be invoked.
1690 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1691 gfp_t gfp_mask, struct mem_cgroup **memcg, bool oom)
1693 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1694 struct mem_cgroup *mem = NULL;
1696 int csize = CHARGE_SIZE;
1699 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1700 * in system level. So, allow to go ahead dying process in addition to
1703 if (unlikely(test_thread_flag(TIF_MEMDIE)
1704 || fatal_signal_pending(current)))
1708 * We always charge the cgroup the mm_struct belongs to.
1709 * The mm_struct's mem_cgroup changes on task migration if the
1710 * thread group leader migrates. It's possible that mm is not
1711 * set, if so charge the init_mm (happens for pagecache usage).
1717 mem = try_get_mem_cgroup_from_mm(mm);
1723 VM_BUG_ON(css_is_removed(&mem->css));
1724 if (mem_cgroup_is_root(mem))
1730 if (consume_stock(mem))
1731 goto done; /* don't need to fill stock */
1732 /* If killed, bypass charge */
1733 if (fatal_signal_pending(current))
1737 if (oom && !nr_oom_retries) {
1739 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1742 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check);
1747 case CHARGE_RETRY: /* not in OOM situation but retry */
1750 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
1752 case CHARGE_NOMEM: /* OOM routine works */
1755 /* If oom, we never return -ENOMEM */
1758 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
1761 } while (ret != CHARGE_OK);
1763 if (csize > PAGE_SIZE)
1764 refill_stock(mem, csize - PAGE_SIZE);
1778 * Somemtimes we have to undo a charge we got by try_charge().
1779 * This function is for that and do uncharge, put css's refcnt.
1780 * gotten by try_charge().
1782 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
1783 unsigned long count)
1785 if (!mem_cgroup_is_root(mem)) {
1786 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
1787 if (do_swap_account)
1788 res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
1789 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
1790 WARN_ON_ONCE(count > INT_MAX);
1791 __css_put(&mem->css, (int)count);
1793 /* we don't need css_put for root */
1796 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem)
1798 __mem_cgroup_cancel_charge(mem, 1);
1802 * A helper function to get mem_cgroup from ID. must be called under
1803 * rcu_read_lock(). The caller must check css_is_removed() or some if
1804 * it's concern. (dropping refcnt from swap can be called against removed
1807 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1809 struct cgroup_subsys_state *css;
1811 /* ID 0 is unused ID */
1814 css = css_lookup(&mem_cgroup_subsys, id);
1817 return container_of(css, struct mem_cgroup, css);
1820 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
1822 struct mem_cgroup *mem = NULL;
1823 struct page_cgroup *pc;
1827 VM_BUG_ON(!PageLocked(page));
1829 pc = lookup_page_cgroup(page);
1830 lock_page_cgroup(pc);
1831 if (PageCgroupUsed(pc)) {
1832 mem = pc->mem_cgroup;
1833 if (mem && !css_tryget(&mem->css))
1835 } else if (PageSwapCache(page)) {
1836 ent.val = page_private(page);
1837 id = lookup_swap_cgroup(ent);
1839 mem = mem_cgroup_lookup(id);
1840 if (mem && !css_tryget(&mem->css))
1844 unlock_page_cgroup(pc);
1849 * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1850 * USED state. If already USED, uncharge and return.
1853 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1854 struct page_cgroup *pc,
1855 enum charge_type ctype)
1857 /* try_charge() can return NULL to *memcg, taking care of it. */
1861 lock_page_cgroup(pc);
1862 if (unlikely(PageCgroupUsed(pc))) {
1863 unlock_page_cgroup(pc);
1864 mem_cgroup_cancel_charge(mem);
1868 pc->mem_cgroup = mem;
1870 * We access a page_cgroup asynchronously without lock_page_cgroup().
1871 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
1872 * is accessed after testing USED bit. To make pc->mem_cgroup visible
1873 * before USED bit, we need memory barrier here.
1874 * See mem_cgroup_add_lru_list(), etc.
1878 case MEM_CGROUP_CHARGE_TYPE_CACHE:
1879 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
1880 SetPageCgroupCache(pc);
1881 SetPageCgroupUsed(pc);
1883 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1884 ClearPageCgroupCache(pc);
1885 SetPageCgroupUsed(pc);
1891 mem_cgroup_charge_statistics(mem, pc, true);
1893 unlock_page_cgroup(pc);
1895 * "charge_statistics" updated event counter. Then, check it.
1896 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
1897 * if they exceeds softlimit.
1899 memcg_check_events(mem, pc->page);
1903 * __mem_cgroup_move_account - move account of the page
1904 * @pc: page_cgroup of the page.
1905 * @from: mem_cgroup which the page is moved from.
1906 * @to: mem_cgroup which the page is moved to. @from != @to.
1907 * @uncharge: whether we should call uncharge and css_put against @from.
1909 * The caller must confirm following.
1910 * - page is not on LRU (isolate_page() is useful.)
1911 * - the pc is locked, used, and ->mem_cgroup points to @from.
1913 * This function doesn't do "charge" nor css_get to new cgroup. It should be
1914 * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
1915 * true, this function does "uncharge" from old cgroup, but it doesn't if
1916 * @uncharge is false, so a caller should do "uncharge".
1919 static void __mem_cgroup_move_account(struct page_cgroup *pc,
1920 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1922 VM_BUG_ON(from == to);
1923 VM_BUG_ON(PageLRU(pc->page));
1924 VM_BUG_ON(!PageCgroupLocked(pc));
1925 VM_BUG_ON(!PageCgroupUsed(pc));
1926 VM_BUG_ON(pc->mem_cgroup != from);
1928 if (PageCgroupFileMapped(pc)) {
1929 /* Update mapped_file data for mem_cgroup */
1931 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1932 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1935 mem_cgroup_charge_statistics(from, pc, false);
1937 /* This is not "cancel", but cancel_charge does all we need. */
1938 mem_cgroup_cancel_charge(from);
1940 /* caller should have done css_get */
1941 pc->mem_cgroup = to;
1942 mem_cgroup_charge_statistics(to, pc, true);
1944 * We charges against "to" which may not have any tasks. Then, "to"
1945 * can be under rmdir(). But in current implementation, caller of
1946 * this function is just force_empty() and move charge, so it's
1947 * garanteed that "to" is never removed. So, we don't check rmdir
1953 * check whether the @pc is valid for moving account and call
1954 * __mem_cgroup_move_account()
1956 static int mem_cgroup_move_account(struct page_cgroup *pc,
1957 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1960 lock_page_cgroup(pc);
1961 if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
1962 __mem_cgroup_move_account(pc, from, to, uncharge);
1965 unlock_page_cgroup(pc);
1969 memcg_check_events(to, pc->page);
1970 memcg_check_events(from, pc->page);
1975 * move charges to its parent.
1978 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1979 struct mem_cgroup *child,
1982 struct page *page = pc->page;
1983 struct cgroup *cg = child->css.cgroup;
1984 struct cgroup *pcg = cg->parent;
1985 struct mem_cgroup *parent;
1993 if (!get_page_unless_zero(page))
1995 if (isolate_lru_page(page))
1998 parent = mem_cgroup_from_cont(pcg);
1999 ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
2003 ret = mem_cgroup_move_account(pc, child, parent, true);
2005 mem_cgroup_cancel_charge(parent);
2007 putback_lru_page(page);
2015 * Charge the memory controller for page usage.
2017 * 0 if the charge was successful
2018 * < 0 if the cgroup is over its limit
2020 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2021 gfp_t gfp_mask, enum charge_type ctype)
2023 struct mem_cgroup *mem = NULL;
2024 struct page_cgroup *pc;
2027 pc = lookup_page_cgroup(page);
2028 /* can happen at boot */
2033 ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
2037 __mem_cgroup_commit_charge(mem, pc, ctype);
2041 int mem_cgroup_newpage_charge(struct page *page,
2042 struct mm_struct *mm, gfp_t gfp_mask)
2044 if (mem_cgroup_disabled())
2046 if (PageCompound(page))
2049 * If already mapped, we don't have to account.
2050 * If page cache, page->mapping has address_space.
2051 * But page->mapping may have out-of-use anon_vma pointer,
2052 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2055 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2059 return mem_cgroup_charge_common(page, mm, gfp_mask,
2060 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2064 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2065 enum charge_type ctype);
2067 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2072 if (mem_cgroup_disabled())
2074 if (PageCompound(page))
2077 * Corner case handling. This is called from add_to_page_cache()
2078 * in usual. But some FS (shmem) precharges this page before calling it
2079 * and call add_to_page_cache() with GFP_NOWAIT.
2081 * For GFP_NOWAIT case, the page may be pre-charged before calling
2082 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2083 * charge twice. (It works but has to pay a bit larger cost.)
2084 * And when the page is SwapCache, it should take swap information
2085 * into account. This is under lock_page() now.
2087 if (!(gfp_mask & __GFP_WAIT)) {
2088 struct page_cgroup *pc;
2090 pc = lookup_page_cgroup(page);
2093 lock_page_cgroup(pc);
2094 if (PageCgroupUsed(pc)) {
2095 unlock_page_cgroup(pc);
2098 unlock_page_cgroup(pc);
2104 if (page_is_file_cache(page))
2105 return mem_cgroup_charge_common(page, mm, gfp_mask,
2106 MEM_CGROUP_CHARGE_TYPE_CACHE);
2109 if (PageSwapCache(page)) {
2110 struct mem_cgroup *mem = NULL;
2112 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2114 __mem_cgroup_commit_charge_swapin(page, mem,
2115 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2117 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2118 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2124 * While swap-in, try_charge -> commit or cancel, the page is locked.
2125 * And when try_charge() successfully returns, one refcnt to memcg without
2126 * struct page_cgroup is acquired. This refcnt will be consumed by
2127 * "commit()" or removed by "cancel()"
2129 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2131 gfp_t mask, struct mem_cgroup **ptr)
2133 struct mem_cgroup *mem;
2136 if (mem_cgroup_disabled())
2139 if (!do_swap_account)
2142 * A racing thread's fault, or swapoff, may have already updated
2143 * the pte, and even removed page from swap cache: in those cases
2144 * do_swap_page()'s pte_same() test will fail; but there's also a
2145 * KSM case which does need to charge the page.
2147 if (!PageSwapCache(page))
2149 mem = try_get_mem_cgroup_from_page(page);
2153 ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
2154 /* drop extra refcnt from tryget */
2160 return __mem_cgroup_try_charge(mm, mask, ptr, true);
2164 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2165 enum charge_type ctype)
2167 struct page_cgroup *pc;
2169 if (mem_cgroup_disabled())
2173 cgroup_exclude_rmdir(&ptr->css);
2174 pc = lookup_page_cgroup(page);
2175 mem_cgroup_lru_del_before_commit_swapcache(page);
2176 __mem_cgroup_commit_charge(ptr, pc, ctype);
2177 mem_cgroup_lru_add_after_commit_swapcache(page);
2179 * Now swap is on-memory. This means this page may be
2180 * counted both as mem and swap....double count.
2181 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2182 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2183 * may call delete_from_swap_cache() before reach here.
2185 if (do_swap_account && PageSwapCache(page)) {
2186 swp_entry_t ent = {.val = page_private(page)};
2188 struct mem_cgroup *memcg;
2190 id = swap_cgroup_record(ent, 0);
2192 memcg = mem_cgroup_lookup(id);
2195 * This recorded memcg can be obsolete one. So, avoid
2196 * calling css_tryget
2198 if (!mem_cgroup_is_root(memcg))
2199 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2200 mem_cgroup_swap_statistics(memcg, false);
2201 mem_cgroup_put(memcg);
2206 * At swapin, we may charge account against cgroup which has no tasks.
2207 * So, rmdir()->pre_destroy() can be called while we do this charge.
2208 * In that case, we need to call pre_destroy() again. check it here.
2210 cgroup_release_and_wakeup_rmdir(&ptr->css);
2213 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2215 __mem_cgroup_commit_charge_swapin(page, ptr,
2216 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2219 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2221 if (mem_cgroup_disabled())
2225 mem_cgroup_cancel_charge(mem);
2229 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype)
2231 struct memcg_batch_info *batch = NULL;
2232 bool uncharge_memsw = true;
2233 /* If swapout, usage of swap doesn't decrease */
2234 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2235 uncharge_memsw = false;
2237 batch = ¤t->memcg_batch;
2239 * In usual, we do css_get() when we remember memcg pointer.
2240 * But in this case, we keep res->usage until end of a series of
2241 * uncharges. Then, it's ok to ignore memcg's refcnt.
2246 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2247 * In those cases, all pages freed continously can be expected to be in
2248 * the same cgroup and we have chance to coalesce uncharges.
2249 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2250 * because we want to do uncharge as soon as possible.
2253 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2254 goto direct_uncharge;
2257 * In typical case, batch->memcg == mem. This means we can
2258 * merge a series of uncharges to an uncharge of res_counter.
2259 * If not, we uncharge res_counter ony by one.
2261 if (batch->memcg != mem)
2262 goto direct_uncharge;
2263 /* remember freed charge and uncharge it later */
2264 batch->bytes += PAGE_SIZE;
2266 batch->memsw_bytes += PAGE_SIZE;
2269 res_counter_uncharge(&mem->res, PAGE_SIZE);
2271 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
2272 if (unlikely(batch->memcg != mem))
2273 memcg_oom_recover(mem);
2278 * uncharge if !page_mapped(page)
2280 static struct mem_cgroup *
2281 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2283 struct page_cgroup *pc;
2284 struct mem_cgroup *mem = NULL;
2286 if (mem_cgroup_disabled())
2289 if (PageSwapCache(page))
2293 * Check if our page_cgroup is valid
2295 pc = lookup_page_cgroup(page);
2296 if (unlikely(!pc || !PageCgroupUsed(pc)))
2299 lock_page_cgroup(pc);
2301 mem = pc->mem_cgroup;
2303 if (!PageCgroupUsed(pc))
2307 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2308 case MEM_CGROUP_CHARGE_TYPE_DROP:
2309 /* See mem_cgroup_prepare_migration() */
2310 if (page_mapped(page) || PageCgroupMigration(pc))
2313 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2314 if (!PageAnon(page)) { /* Shared memory */
2315 if (page->mapping && !page_is_file_cache(page))
2317 } else if (page_mapped(page)) /* Anon */
2324 if (!mem_cgroup_is_root(mem))
2325 __do_uncharge(mem, ctype);
2326 if (ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2327 mem_cgroup_swap_statistics(mem, true);
2328 mem_cgroup_charge_statistics(mem, pc, false);
2330 ClearPageCgroupUsed(pc);
2332 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2333 * freed from LRU. This is safe because uncharged page is expected not
2334 * to be reused (freed soon). Exception is SwapCache, it's handled by
2335 * special functions.
2338 unlock_page_cgroup(pc);
2340 memcg_check_events(mem, page);
2341 /* at swapout, this memcg will be accessed to record to swap */
2342 if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2348 unlock_page_cgroup(pc);
2352 void mem_cgroup_uncharge_page(struct page *page)
2355 if (page_mapped(page))
2357 if (page->mapping && !PageAnon(page))
2359 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2362 void mem_cgroup_uncharge_cache_page(struct page *page)
2364 VM_BUG_ON(page_mapped(page));
2365 VM_BUG_ON(page->mapping);
2366 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2370 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2371 * In that cases, pages are freed continuously and we can expect pages
2372 * are in the same memcg. All these calls itself limits the number of
2373 * pages freed at once, then uncharge_start/end() is called properly.
2374 * This may be called prural(2) times in a context,
2377 void mem_cgroup_uncharge_start(void)
2379 current->memcg_batch.do_batch++;
2380 /* We can do nest. */
2381 if (current->memcg_batch.do_batch == 1) {
2382 current->memcg_batch.memcg = NULL;
2383 current->memcg_batch.bytes = 0;
2384 current->memcg_batch.memsw_bytes = 0;
2388 void mem_cgroup_uncharge_end(void)
2390 struct memcg_batch_info *batch = ¤t->memcg_batch;
2392 if (!batch->do_batch)
2396 if (batch->do_batch) /* If stacked, do nothing. */
2402 * This "batch->memcg" is valid without any css_get/put etc...
2403 * bacause we hide charges behind us.
2406 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2407 if (batch->memsw_bytes)
2408 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2409 memcg_oom_recover(batch->memcg);
2410 /* forget this pointer (for sanity check) */
2411 batch->memcg = NULL;
2416 * called after __delete_from_swap_cache() and drop "page" account.
2417 * memcg information is recorded to swap_cgroup of "ent"
2420 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2422 struct mem_cgroup *memcg;
2423 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2425 if (!swapout) /* this was a swap cache but the swap is unused ! */
2426 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2428 memcg = __mem_cgroup_uncharge_common(page, ctype);
2430 /* record memcg information */
2431 if (do_swap_account && swapout && memcg) {
2432 swap_cgroup_record(ent, css_id(&memcg->css));
2433 mem_cgroup_get(memcg);
2435 if (swapout && memcg)
2436 css_put(&memcg->css);
2440 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2442 * called from swap_entry_free(). remove record in swap_cgroup and
2443 * uncharge "memsw" account.
2445 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2447 struct mem_cgroup *memcg;
2450 if (!do_swap_account)
2453 id = swap_cgroup_record(ent, 0);
2455 memcg = mem_cgroup_lookup(id);
2458 * We uncharge this because swap is freed.
2459 * This memcg can be obsolete one. We avoid calling css_tryget
2461 if (!mem_cgroup_is_root(memcg))
2462 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2463 mem_cgroup_swap_statistics(memcg, false);
2464 mem_cgroup_put(memcg);
2470 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2471 * @entry: swap entry to be moved
2472 * @from: mem_cgroup which the entry is moved from
2473 * @to: mem_cgroup which the entry is moved to
2474 * @need_fixup: whether we should fixup res_counters and refcounts.
2476 * It succeeds only when the swap_cgroup's record for this entry is the same
2477 * as the mem_cgroup's id of @from.
2479 * Returns 0 on success, -EINVAL on failure.
2481 * The caller must have charged to @to, IOW, called res_counter_charge() about
2482 * both res and memsw, and called css_get().
2484 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2485 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2487 unsigned short old_id, new_id;
2489 old_id = css_id(&from->css);
2490 new_id = css_id(&to->css);
2492 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2493 mem_cgroup_swap_statistics(from, false);
2494 mem_cgroup_swap_statistics(to, true);
2496 * This function is only called from task migration context now.
2497 * It postpones res_counter and refcount handling till the end
2498 * of task migration(mem_cgroup_clear_mc()) for performance
2499 * improvement. But we cannot postpone mem_cgroup_get(to)
2500 * because if the process that has been moved to @to does
2501 * swap-in, the refcount of @to might be decreased to 0.
2505 if (!mem_cgroup_is_root(from))
2506 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2507 mem_cgroup_put(from);
2509 * we charged both to->res and to->memsw, so we should
2512 if (!mem_cgroup_is_root(to))
2513 res_counter_uncharge(&to->res, PAGE_SIZE);
2521 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2522 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2529 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2532 int mem_cgroup_prepare_migration(struct page *page,
2533 struct page *newpage, struct mem_cgroup **ptr)
2535 struct page_cgroup *pc;
2536 struct mem_cgroup *mem = NULL;
2537 enum charge_type ctype;
2540 if (mem_cgroup_disabled())
2543 pc = lookup_page_cgroup(page);
2544 lock_page_cgroup(pc);
2545 if (PageCgroupUsed(pc)) {
2546 mem = pc->mem_cgroup;
2549 * At migrating an anonymous page, its mapcount goes down
2550 * to 0 and uncharge() will be called. But, even if it's fully
2551 * unmapped, migration may fail and this page has to be
2552 * charged again. We set MIGRATION flag here and delay uncharge
2553 * until end_migration() is called
2555 * Corner Case Thinking
2557 * When the old page was mapped as Anon and it's unmap-and-freed
2558 * while migration was ongoing.
2559 * If unmap finds the old page, uncharge() of it will be delayed
2560 * until end_migration(). If unmap finds a new page, it's
2561 * uncharged when it make mapcount to be 1->0. If unmap code
2562 * finds swap_migration_entry, the new page will not be mapped
2563 * and end_migration() will find it(mapcount==0).
2566 * When the old page was mapped but migraion fails, the kernel
2567 * remaps it. A charge for it is kept by MIGRATION flag even
2568 * if mapcount goes down to 0. We can do remap successfully
2569 * without charging it again.
2572 * The "old" page is under lock_page() until the end of
2573 * migration, so, the old page itself will not be swapped-out.
2574 * If the new page is swapped out before end_migraton, our
2575 * hook to usual swap-out path will catch the event.
2578 SetPageCgroupMigration(pc);
2580 unlock_page_cgroup(pc);
2582 * If the page is not charged at this point,
2589 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false);
2590 css_put(&mem->css);/* drop extra refcnt */
2591 if (ret || *ptr == NULL) {
2592 if (PageAnon(page)) {
2593 lock_page_cgroup(pc);
2594 ClearPageCgroupMigration(pc);
2595 unlock_page_cgroup(pc);
2597 * The old page may be fully unmapped while we kept it.
2599 mem_cgroup_uncharge_page(page);
2604 * We charge new page before it's used/mapped. So, even if unlock_page()
2605 * is called before end_migration, we can catch all events on this new
2606 * page. In the case new page is migrated but not remapped, new page's
2607 * mapcount will be finally 0 and we call uncharge in end_migration().
2609 pc = lookup_page_cgroup(newpage);
2611 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2612 else if (page_is_file_cache(page))
2613 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2615 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2616 __mem_cgroup_commit_charge(mem, pc, ctype);
2620 /* remove redundant charge if migration failed*/
2621 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2622 struct page *oldpage, struct page *newpage)
2624 struct page *used, *unused;
2625 struct page_cgroup *pc;
2629 /* blocks rmdir() */
2630 cgroup_exclude_rmdir(&mem->css);
2631 /* at migration success, oldpage->mapping is NULL. */
2632 if (oldpage->mapping) {
2640 * We disallowed uncharge of pages under migration because mapcount
2641 * of the page goes down to zero, temporarly.
2642 * Clear the flag and check the page should be charged.
2644 pc = lookup_page_cgroup(oldpage);
2645 lock_page_cgroup(pc);
2646 ClearPageCgroupMigration(pc);
2647 unlock_page_cgroup(pc);
2649 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
2652 * If a page is a file cache, radix-tree replacement is very atomic
2653 * and we can skip this check. When it was an Anon page, its mapcount
2654 * goes down to 0. But because we added MIGRATION flage, it's not
2655 * uncharged yet. There are several case but page->mapcount check
2656 * and USED bit check in mem_cgroup_uncharge_page() will do enough
2657 * check. (see prepare_charge() also)
2660 mem_cgroup_uncharge_page(used);
2662 * At migration, we may charge account against cgroup which has no
2664 * So, rmdir()->pre_destroy() can be called while we do this charge.
2665 * In that case, we need to call pre_destroy() again. check it here.
2667 cgroup_release_and_wakeup_rmdir(&mem->css);
2671 * A call to try to shrink memory usage on charge failure at shmem's swapin.
2672 * Calling hierarchical_reclaim is not enough because we should update
2673 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2674 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2675 * not from the memcg which this page would be charged to.
2676 * try_charge_swapin does all of these works properly.
2678 int mem_cgroup_shmem_charge_fallback(struct page *page,
2679 struct mm_struct *mm,
2682 struct mem_cgroup *mem = NULL;
2685 if (mem_cgroup_disabled())
2688 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2690 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2695 static DEFINE_MUTEX(set_limit_mutex);
2697 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2698 unsigned long long val)
2701 u64 memswlimit, memlimit;
2703 int children = mem_cgroup_count_children(memcg);
2704 u64 curusage, oldusage;
2708 * For keeping hierarchical_reclaim simple, how long we should retry
2709 * is depends on callers. We set our retry-count to be function
2710 * of # of children which we should visit in this loop.
2712 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2714 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2717 while (retry_count) {
2718 if (signal_pending(current)) {
2723 * Rather than hide all in some function, I do this in
2724 * open coded manner. You see what this really does.
2725 * We have to guarantee mem->res.limit < mem->memsw.limit.
2727 mutex_lock(&set_limit_mutex);
2728 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2729 if (memswlimit < val) {
2731 mutex_unlock(&set_limit_mutex);
2735 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2739 ret = res_counter_set_limit(&memcg->res, val);
2741 if (memswlimit == val)
2742 memcg->memsw_is_minimum = true;
2744 memcg->memsw_is_minimum = false;
2746 mutex_unlock(&set_limit_mutex);
2751 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2752 MEM_CGROUP_RECLAIM_SHRINK);
2753 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2754 /* Usage is reduced ? */
2755 if (curusage >= oldusage)
2758 oldusage = curusage;
2760 if (!ret && enlarge)
2761 memcg_oom_recover(memcg);
2766 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2767 unsigned long long val)
2770 u64 memlimit, memswlimit, oldusage, curusage;
2771 int children = mem_cgroup_count_children(memcg);
2775 /* see mem_cgroup_resize_res_limit */
2776 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
2777 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2778 while (retry_count) {
2779 if (signal_pending(current)) {
2784 * Rather than hide all in some function, I do this in
2785 * open coded manner. You see what this really does.
2786 * We have to guarantee mem->res.limit < mem->memsw.limit.
2788 mutex_lock(&set_limit_mutex);
2789 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2790 if (memlimit > val) {
2792 mutex_unlock(&set_limit_mutex);
2795 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2796 if (memswlimit < val)
2798 ret = res_counter_set_limit(&memcg->memsw, val);
2800 if (memlimit == val)
2801 memcg->memsw_is_minimum = true;
2803 memcg->memsw_is_minimum = false;
2805 mutex_unlock(&set_limit_mutex);
2810 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2811 MEM_CGROUP_RECLAIM_NOSWAP |
2812 MEM_CGROUP_RECLAIM_SHRINK);
2813 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2814 /* Usage is reduced ? */
2815 if (curusage >= oldusage)
2818 oldusage = curusage;
2820 if (!ret && enlarge)
2821 memcg_oom_recover(memcg);
2825 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2826 gfp_t gfp_mask, int nid,
2829 unsigned long nr_reclaimed = 0;
2830 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2831 unsigned long reclaimed;
2833 struct mem_cgroup_tree_per_zone *mctz;
2834 unsigned long long excess;
2839 mctz = soft_limit_tree_node_zone(nid, zid);
2841 * This loop can run a while, specially if mem_cgroup's continuously
2842 * keep exceeding their soft limit and putting the system under
2849 mz = mem_cgroup_largest_soft_limit_node(mctz);
2853 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
2855 MEM_CGROUP_RECLAIM_SOFT);
2856 nr_reclaimed += reclaimed;
2857 spin_lock(&mctz->lock);
2860 * If we failed to reclaim anything from this memory cgroup
2861 * it is time to move on to the next cgroup
2867 * Loop until we find yet another one.
2869 * By the time we get the soft_limit lock
2870 * again, someone might have aded the
2871 * group back on the RB tree. Iterate to
2872 * make sure we get a different mem.
2873 * mem_cgroup_largest_soft_limit_node returns
2874 * NULL if no other cgroup is present on
2878 __mem_cgroup_largest_soft_limit_node(mctz);
2879 if (next_mz == mz) {
2880 css_put(&next_mz->mem->css);
2882 } else /* next_mz == NULL or other memcg */
2886 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
2887 excess = res_counter_soft_limit_excess(&mz->mem->res);
2889 * One school of thought says that we should not add
2890 * back the node to the tree if reclaim returns 0.
2891 * But our reclaim could return 0, simply because due
2892 * to priority we are exposing a smaller subset of
2893 * memory to reclaim from. Consider this as a longer
2896 /* If excess == 0, no tree ops */
2897 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
2898 spin_unlock(&mctz->lock);
2899 css_put(&mz->mem->css);
2902 * Could not reclaim anything and there are no more
2903 * mem cgroups to try or we seem to be looping without
2904 * reclaiming anything.
2906 if (!nr_reclaimed &&
2908 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2910 } while (!nr_reclaimed);
2912 css_put(&next_mz->mem->css);
2913 return nr_reclaimed;
2917 * This routine traverse page_cgroup in given list and drop them all.
2918 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
2920 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
2921 int node, int zid, enum lru_list lru)
2924 struct mem_cgroup_per_zone *mz;
2925 struct page_cgroup *pc, *busy;
2926 unsigned long flags, loop;
2927 struct list_head *list;
2930 zone = &NODE_DATA(node)->node_zones[zid];
2931 mz = mem_cgroup_zoneinfo(mem, node, zid);
2932 list = &mz->lists[lru];
2934 loop = MEM_CGROUP_ZSTAT(mz, lru);
2935 /* give some margin against EBUSY etc...*/
2940 spin_lock_irqsave(&zone->lru_lock, flags);
2941 if (list_empty(list)) {
2942 spin_unlock_irqrestore(&zone->lru_lock, flags);
2945 pc = list_entry(list->prev, struct page_cgroup, lru);
2947 list_move(&pc->lru, list);
2949 spin_unlock_irqrestore(&zone->lru_lock, flags);
2952 spin_unlock_irqrestore(&zone->lru_lock, flags);
2954 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
2958 if (ret == -EBUSY || ret == -EINVAL) {
2959 /* found lock contention or "pc" is obsolete. */
2966 if (!ret && !list_empty(list))
2972 * make mem_cgroup's charge to be 0 if there is no task.
2973 * This enables deleting this mem_cgroup.
2975 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
2978 int node, zid, shrink;
2979 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2980 struct cgroup *cgrp = mem->css.cgroup;
2985 /* should free all ? */
2991 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
2994 if (signal_pending(current))
2996 /* This is for making all *used* pages to be on LRU. */
2997 lru_add_drain_all();
2998 drain_all_stock_sync();
3000 for_each_node_state(node, N_HIGH_MEMORY) {
3001 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3004 ret = mem_cgroup_force_empty_list(mem,
3013 memcg_oom_recover(mem);
3014 /* it seems parent cgroup doesn't have enough mem */
3018 /* "ret" should also be checked to ensure all lists are empty. */
3019 } while (mem->res.usage > 0 || ret);
3025 /* returns EBUSY if there is a task or if we come here twice. */
3026 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3030 /* we call try-to-free pages for make this cgroup empty */
3031 lru_add_drain_all();
3032 /* try to free all pages in this cgroup */
3034 while (nr_retries && mem->res.usage > 0) {
3037 if (signal_pending(current)) {
3041 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3042 false, get_swappiness(mem));
3045 /* maybe some writeback is necessary */
3046 congestion_wait(BLK_RW_ASYNC, HZ/10);
3051 /* try move_account...there may be some *locked* pages. */
3055 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3057 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3061 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3063 return mem_cgroup_from_cont(cont)->use_hierarchy;
3066 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3070 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3071 struct cgroup *parent = cont->parent;
3072 struct mem_cgroup *parent_mem = NULL;
3075 parent_mem = mem_cgroup_from_cont(parent);
3079 * If parent's use_hierarchy is set, we can't make any modifications
3080 * in the child subtrees. If it is unset, then the change can
3081 * occur, provided the current cgroup has no children.
3083 * For the root cgroup, parent_mem is NULL, we allow value to be
3084 * set if there are no children.
3086 if ((!parent_mem || !parent_mem->use_hierarchy) &&
3087 (val == 1 || val == 0)) {
3088 if (list_empty(&cont->children))
3089 mem->use_hierarchy = val;
3099 struct mem_cgroup_idx_data {
3101 enum mem_cgroup_stat_index idx;
3105 mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data)
3107 struct mem_cgroup_idx_data *d = data;
3108 d->val += mem_cgroup_read_stat(mem, d->idx);
3113 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3114 enum mem_cgroup_stat_index idx, s64 *val)
3116 struct mem_cgroup_idx_data d;
3119 mem_cgroup_walk_tree(mem, &d, mem_cgroup_get_idx_stat);
3123 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3127 if (!mem_cgroup_is_root(mem)) {
3129 return res_counter_read_u64(&mem->res, RES_USAGE);
3131 return res_counter_read_u64(&mem->memsw, RES_USAGE);
3134 mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE, &idx_val);
3136 mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS, &idx_val);
3140 mem_cgroup_get_recursive_idx_stat(mem,
3141 MEM_CGROUP_STAT_SWAPOUT, &idx_val);
3145 return val << PAGE_SHIFT;
3148 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3150 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3154 type = MEMFILE_TYPE(cft->private);
3155 name = MEMFILE_ATTR(cft->private);
3158 if (name == RES_USAGE)
3159 val = mem_cgroup_usage(mem, false);
3161 val = res_counter_read_u64(&mem->res, name);
3164 if (name == RES_USAGE)
3165 val = mem_cgroup_usage(mem, true);
3167 val = res_counter_read_u64(&mem->memsw, name);
3176 * The user of this function is...
3179 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3182 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3184 unsigned long long val;
3187 type = MEMFILE_TYPE(cft->private);
3188 name = MEMFILE_ATTR(cft->private);
3191 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3195 /* This function does all necessary parse...reuse it */
3196 ret = res_counter_memparse_write_strategy(buffer, &val);
3200 ret = mem_cgroup_resize_limit(memcg, val);
3202 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3204 case RES_SOFT_LIMIT:
3205 ret = res_counter_memparse_write_strategy(buffer, &val);
3209 * For memsw, soft limits are hard to implement in terms
3210 * of semantics, for now, we support soft limits for
3211 * control without swap
3214 ret = res_counter_set_soft_limit(&memcg->res, val);
3219 ret = -EINVAL; /* should be BUG() ? */
3225 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3226 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3228 struct cgroup *cgroup;
3229 unsigned long long min_limit, min_memsw_limit, tmp;
3231 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3232 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3233 cgroup = memcg->css.cgroup;
3234 if (!memcg->use_hierarchy)
3237 while (cgroup->parent) {
3238 cgroup = cgroup->parent;
3239 memcg = mem_cgroup_from_cont(cgroup);
3240 if (!memcg->use_hierarchy)
3242 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3243 min_limit = min(min_limit, tmp);
3244 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3245 min_memsw_limit = min(min_memsw_limit, tmp);
3248 *mem_limit = min_limit;
3249 *memsw_limit = min_memsw_limit;
3253 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3255 struct mem_cgroup *mem;
3258 mem = mem_cgroup_from_cont(cont);
3259 type = MEMFILE_TYPE(event);
3260 name = MEMFILE_ATTR(event);
3264 res_counter_reset_max(&mem->res);
3266 res_counter_reset_max(&mem->memsw);
3270 res_counter_reset_failcnt(&mem->res);
3272 res_counter_reset_failcnt(&mem->memsw);
3279 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3282 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3286 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3287 struct cftype *cft, u64 val)
3289 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3291 if (val >= (1 << NR_MOVE_TYPE))
3294 * We check this value several times in both in can_attach() and
3295 * attach(), so we need cgroup lock to prevent this value from being
3299 mem->move_charge_at_immigrate = val;
3305 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3306 struct cftype *cft, u64 val)
3313 /* For read statistics */
3329 struct mcs_total_stat {
3330 s64 stat[NR_MCS_STAT];
3336 } memcg_stat_strings[NR_MCS_STAT] = {
3337 {"cache", "total_cache"},
3338 {"rss", "total_rss"},
3339 {"mapped_file", "total_mapped_file"},
3340 {"pgpgin", "total_pgpgin"},
3341 {"pgpgout", "total_pgpgout"},
3342 {"swap", "total_swap"},
3343 {"inactive_anon", "total_inactive_anon"},
3344 {"active_anon", "total_active_anon"},
3345 {"inactive_file", "total_inactive_file"},
3346 {"active_file", "total_active_file"},
3347 {"unevictable", "total_unevictable"}
3351 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
3353 struct mcs_total_stat *s = data;
3357 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3358 s->stat[MCS_CACHE] += val * PAGE_SIZE;
3359 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3360 s->stat[MCS_RSS] += val * PAGE_SIZE;
3361 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3362 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3363 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3364 s->stat[MCS_PGPGIN] += val;
3365 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3366 s->stat[MCS_PGPGOUT] += val;
3367 if (do_swap_account) {
3368 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3369 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3373 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3374 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3375 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3376 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3377 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3378 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3379 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3380 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3381 val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3382 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3387 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3389 mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
3392 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3393 struct cgroup_map_cb *cb)
3395 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3396 struct mcs_total_stat mystat;
3399 memset(&mystat, 0, sizeof(mystat));
3400 mem_cgroup_get_local_stat(mem_cont, &mystat);
3402 for (i = 0; i < NR_MCS_STAT; i++) {
3403 if (i == MCS_SWAP && !do_swap_account)
3405 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3408 /* Hierarchical information */
3410 unsigned long long limit, memsw_limit;
3411 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3412 cb->fill(cb, "hierarchical_memory_limit", limit);
3413 if (do_swap_account)
3414 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3417 memset(&mystat, 0, sizeof(mystat));
3418 mem_cgroup_get_total_stat(mem_cont, &mystat);
3419 for (i = 0; i < NR_MCS_STAT; i++) {
3420 if (i == MCS_SWAP && !do_swap_account)
3422 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3425 #ifdef CONFIG_DEBUG_VM
3426 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3430 struct mem_cgroup_per_zone *mz;
3431 unsigned long recent_rotated[2] = {0, 0};
3432 unsigned long recent_scanned[2] = {0, 0};
3434 for_each_online_node(nid)
3435 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3436 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3438 recent_rotated[0] +=
3439 mz->reclaim_stat.recent_rotated[0];
3440 recent_rotated[1] +=
3441 mz->reclaim_stat.recent_rotated[1];
3442 recent_scanned[0] +=
3443 mz->reclaim_stat.recent_scanned[0];
3444 recent_scanned[1] +=
3445 mz->reclaim_stat.recent_scanned[1];
3447 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3448 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3449 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3450 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3457 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3459 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3461 return get_swappiness(memcg);
3464 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3467 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3468 struct mem_cgroup *parent;
3473 if (cgrp->parent == NULL)
3476 parent = mem_cgroup_from_cont(cgrp->parent);
3480 /* If under hierarchy, only empty-root can set this value */
3481 if ((parent->use_hierarchy) ||
3482 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3487 spin_lock(&memcg->reclaim_param_lock);
3488 memcg->swappiness = val;
3489 spin_unlock(&memcg->reclaim_param_lock);
3496 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3498 struct mem_cgroup_threshold_ary *t;
3504 t = rcu_dereference(memcg->thresholds.primary);
3506 t = rcu_dereference(memcg->memsw_thresholds.primary);
3511 usage = mem_cgroup_usage(memcg, swap);
3514 * current_threshold points to threshold just below usage.
3515 * If it's not true, a threshold was crossed after last
3516 * call of __mem_cgroup_threshold().
3518 i = t->current_threshold;
3521 * Iterate backward over array of thresholds starting from
3522 * current_threshold and check if a threshold is crossed.
3523 * If none of thresholds below usage is crossed, we read
3524 * only one element of the array here.
3526 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3527 eventfd_signal(t->entries[i].eventfd, 1);
3529 /* i = current_threshold + 1 */
3533 * Iterate forward over array of thresholds starting from
3534 * current_threshold+1 and check if a threshold is crossed.
3535 * If none of thresholds above usage is crossed, we read
3536 * only one element of the array here.
3538 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3539 eventfd_signal(t->entries[i].eventfd, 1);
3541 /* Update current_threshold */
3542 t->current_threshold = i - 1;
3547 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3549 __mem_cgroup_threshold(memcg, false);
3550 if (do_swap_account)
3551 __mem_cgroup_threshold(memcg, true);
3554 static int compare_thresholds(const void *a, const void *b)
3556 const struct mem_cgroup_threshold *_a = a;
3557 const struct mem_cgroup_threshold *_b = b;
3559 return _a->threshold - _b->threshold;
3562 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem, void *data)
3564 struct mem_cgroup_eventfd_list *ev;
3566 list_for_each_entry(ev, &mem->oom_notify, list)
3567 eventfd_signal(ev->eventfd, 1);
3571 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3573 mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_notify_cb);
3576 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3577 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3579 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3580 struct mem_cgroup_thresholds *thresholds;
3581 struct mem_cgroup_threshold_ary *new;
3582 int type = MEMFILE_TYPE(cft->private);
3583 u64 threshold, usage;
3586 ret = res_counter_memparse_write_strategy(args, &threshold);
3590 mutex_lock(&memcg->thresholds_lock);
3593 thresholds = &memcg->thresholds;
3594 else if (type == _MEMSWAP)
3595 thresholds = &memcg->memsw_thresholds;
3599 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3601 /* Check if a threshold crossed before adding a new one */
3602 if (thresholds->primary)
3603 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3605 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
3607 /* Allocate memory for new array of thresholds */
3608 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
3616 /* Copy thresholds (if any) to new array */
3617 if (thresholds->primary) {
3618 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
3619 sizeof(struct mem_cgroup_threshold));
3622 /* Add new threshold */
3623 new->entries[size - 1].eventfd = eventfd;
3624 new->entries[size - 1].threshold = threshold;
3626 /* Sort thresholds. Registering of new threshold isn't time-critical */
3627 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
3628 compare_thresholds, NULL);
3630 /* Find current threshold */
3631 new->current_threshold = -1;
3632 for (i = 0; i < size; i++) {
3633 if (new->entries[i].threshold < usage) {
3635 * new->current_threshold will not be used until
3636 * rcu_assign_pointer(), so it's safe to increment
3639 ++new->current_threshold;
3643 /* Free old spare buffer and save old primary buffer as spare */
3644 kfree(thresholds->spare);
3645 thresholds->spare = thresholds->primary;
3647 rcu_assign_pointer(thresholds->primary, new);
3649 /* To be sure that nobody uses thresholds */
3653 mutex_unlock(&memcg->thresholds_lock);
3658 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
3659 struct cftype *cft, struct eventfd_ctx *eventfd)
3661 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3662 struct mem_cgroup_thresholds *thresholds;
3663 struct mem_cgroup_threshold_ary *new;
3664 int type = MEMFILE_TYPE(cft->private);
3668 mutex_lock(&memcg->thresholds_lock);
3670 thresholds = &memcg->thresholds;
3671 else if (type == _MEMSWAP)
3672 thresholds = &memcg->memsw_thresholds;
3677 * Something went wrong if we trying to unregister a threshold
3678 * if we don't have thresholds
3680 BUG_ON(!thresholds);
3682 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3684 /* Check if a threshold crossed before removing */
3685 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3687 /* Calculate new number of threshold */
3689 for (i = 0; i < thresholds->primary->size; i++) {
3690 if (thresholds->primary->entries[i].eventfd != eventfd)
3694 new = thresholds->spare;
3696 /* Set thresholds array to NULL if we don't have thresholds */
3705 /* Copy thresholds and find current threshold */
3706 new->current_threshold = -1;
3707 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
3708 if (thresholds->primary->entries[i].eventfd == eventfd)
3711 new->entries[j] = thresholds->primary->entries[i];
3712 if (new->entries[j].threshold < usage) {
3714 * new->current_threshold will not be used
3715 * until rcu_assign_pointer(), so it's safe to increment
3718 ++new->current_threshold;
3724 /* Swap primary and spare array */
3725 thresholds->spare = thresholds->primary;
3726 rcu_assign_pointer(thresholds->primary, new);
3728 /* To be sure that nobody uses thresholds */
3731 mutex_unlock(&memcg->thresholds_lock);
3734 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
3735 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3737 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3738 struct mem_cgroup_eventfd_list *event;
3739 int type = MEMFILE_TYPE(cft->private);
3741 BUG_ON(type != _OOM_TYPE);
3742 event = kmalloc(sizeof(*event), GFP_KERNEL);
3746 mutex_lock(&memcg_oom_mutex);
3748 event->eventfd = eventfd;
3749 list_add(&event->list, &memcg->oom_notify);
3751 /* already in OOM ? */
3752 if (atomic_read(&memcg->oom_lock))
3753 eventfd_signal(eventfd, 1);
3754 mutex_unlock(&memcg_oom_mutex);
3759 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
3760 struct cftype *cft, struct eventfd_ctx *eventfd)
3762 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3763 struct mem_cgroup_eventfd_list *ev, *tmp;
3764 int type = MEMFILE_TYPE(cft->private);
3766 BUG_ON(type != _OOM_TYPE);
3768 mutex_lock(&memcg_oom_mutex);
3770 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
3771 if (ev->eventfd == eventfd) {
3772 list_del(&ev->list);
3777 mutex_unlock(&memcg_oom_mutex);
3780 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
3781 struct cftype *cft, struct cgroup_map_cb *cb)
3783 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3785 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
3787 if (atomic_read(&mem->oom_lock))
3788 cb->fill(cb, "under_oom", 1);
3790 cb->fill(cb, "under_oom", 0);
3794 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
3795 struct cftype *cft, u64 val)
3797 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3798 struct mem_cgroup *parent;
3800 /* cannot set to root cgroup and only 0 and 1 are allowed */
3801 if (!cgrp->parent || !((val == 0) || (val == 1)))
3804 parent = mem_cgroup_from_cont(cgrp->parent);
3807 /* oom-kill-disable is a flag for subhierarchy. */
3808 if ((parent->use_hierarchy) ||
3809 (mem->use_hierarchy && !list_empty(&cgrp->children))) {
3813 mem->oom_kill_disable = val;
3815 memcg_oom_recover(mem);
3820 static struct cftype mem_cgroup_files[] = {
3822 .name = "usage_in_bytes",
3823 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3824 .read_u64 = mem_cgroup_read,
3825 .register_event = mem_cgroup_usage_register_event,
3826 .unregister_event = mem_cgroup_usage_unregister_event,
3829 .name = "max_usage_in_bytes",
3830 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3831 .trigger = mem_cgroup_reset,
3832 .read_u64 = mem_cgroup_read,
3835 .name = "limit_in_bytes",
3836 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3837 .write_string = mem_cgroup_write,
3838 .read_u64 = mem_cgroup_read,
3841 .name = "soft_limit_in_bytes",
3842 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3843 .write_string = mem_cgroup_write,
3844 .read_u64 = mem_cgroup_read,
3848 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3849 .trigger = mem_cgroup_reset,
3850 .read_u64 = mem_cgroup_read,
3854 .read_map = mem_control_stat_show,
3857 .name = "force_empty",
3858 .trigger = mem_cgroup_force_empty_write,
3861 .name = "use_hierarchy",
3862 .write_u64 = mem_cgroup_hierarchy_write,
3863 .read_u64 = mem_cgroup_hierarchy_read,
3866 .name = "swappiness",
3867 .read_u64 = mem_cgroup_swappiness_read,
3868 .write_u64 = mem_cgroup_swappiness_write,
3871 .name = "move_charge_at_immigrate",
3872 .read_u64 = mem_cgroup_move_charge_read,
3873 .write_u64 = mem_cgroup_move_charge_write,
3876 .name = "oom_control",
3877 .read_map = mem_cgroup_oom_control_read,
3878 .write_u64 = mem_cgroup_oom_control_write,
3879 .register_event = mem_cgroup_oom_register_event,
3880 .unregister_event = mem_cgroup_oom_unregister_event,
3881 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3885 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3886 static struct cftype memsw_cgroup_files[] = {
3888 .name = "memsw.usage_in_bytes",
3889 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
3890 .read_u64 = mem_cgroup_read,
3891 .register_event = mem_cgroup_usage_register_event,
3892 .unregister_event = mem_cgroup_usage_unregister_event,
3895 .name = "memsw.max_usage_in_bytes",
3896 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
3897 .trigger = mem_cgroup_reset,
3898 .read_u64 = mem_cgroup_read,
3901 .name = "memsw.limit_in_bytes",
3902 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
3903 .write_string = mem_cgroup_write,
3904 .read_u64 = mem_cgroup_read,
3907 .name = "memsw.failcnt",
3908 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
3909 .trigger = mem_cgroup_reset,
3910 .read_u64 = mem_cgroup_read,
3914 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3916 if (!do_swap_account)
3918 return cgroup_add_files(cont, ss, memsw_cgroup_files,
3919 ARRAY_SIZE(memsw_cgroup_files));
3922 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3928 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3930 struct mem_cgroup_per_node *pn;
3931 struct mem_cgroup_per_zone *mz;
3933 int zone, tmp = node;
3935 * This routine is called against possible nodes.
3936 * But it's BUG to call kmalloc() against offline node.
3938 * TODO: this routine can waste much memory for nodes which will
3939 * never be onlined. It's better to use memory hotplug callback
3942 if (!node_state(node, N_NORMAL_MEMORY))
3944 pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
3948 mem->info.nodeinfo[node] = pn;
3949 memset(pn, 0, sizeof(*pn));
3951 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3952 mz = &pn->zoneinfo[zone];
3954 INIT_LIST_HEAD(&mz->lists[l]);
3955 mz->usage_in_excess = 0;
3956 mz->on_tree = false;
3962 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3964 kfree(mem->info.nodeinfo[node]);
3967 static struct mem_cgroup *mem_cgroup_alloc(void)
3969 struct mem_cgroup *mem;
3970 int size = sizeof(struct mem_cgroup);
3972 /* Can be very big if MAX_NUMNODES is very big */
3973 if (size < PAGE_SIZE)
3974 mem = kmalloc(size, GFP_KERNEL);
3976 mem = vmalloc(size);
3981 memset(mem, 0, size);
3982 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
3984 if (size < PAGE_SIZE)
3994 * At destroying mem_cgroup, references from swap_cgroup can remain.
3995 * (scanning all at force_empty is too costly...)
3997 * Instead of clearing all references at force_empty, we remember
3998 * the number of reference from swap_cgroup and free mem_cgroup when
3999 * it goes down to 0.
4001 * Removal of cgroup itself succeeds regardless of refs from swap.
4004 static void __mem_cgroup_free(struct mem_cgroup *mem)
4008 mem_cgroup_remove_from_trees(mem);
4009 free_css_id(&mem_cgroup_subsys, &mem->css);
4011 for_each_node_state(node, N_POSSIBLE)
4012 free_mem_cgroup_per_zone_info(mem, node);
4014 free_percpu(mem->stat);
4015 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4021 static void mem_cgroup_get(struct mem_cgroup *mem)
4023 atomic_inc(&mem->refcnt);
4026 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4028 if (atomic_sub_and_test(count, &mem->refcnt)) {
4029 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4030 __mem_cgroup_free(mem);
4032 mem_cgroup_put(parent);
4036 static void mem_cgroup_put(struct mem_cgroup *mem)
4038 __mem_cgroup_put(mem, 1);
4042 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4044 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4046 if (!mem->res.parent)
4048 return mem_cgroup_from_res_counter(mem->res.parent, res);
4051 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4052 static void __init enable_swap_cgroup(void)
4054 if (!mem_cgroup_disabled() && really_do_swap_account)
4055 do_swap_account = 1;
4058 static void __init enable_swap_cgroup(void)
4063 static int mem_cgroup_soft_limit_tree_init(void)
4065 struct mem_cgroup_tree_per_node *rtpn;
4066 struct mem_cgroup_tree_per_zone *rtpz;
4067 int tmp, node, zone;
4069 for_each_node_state(node, N_POSSIBLE) {
4071 if (!node_state(node, N_NORMAL_MEMORY))
4073 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4077 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4079 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4080 rtpz = &rtpn->rb_tree_per_zone[zone];
4081 rtpz->rb_root = RB_ROOT;
4082 spin_lock_init(&rtpz->lock);
4088 static struct cgroup_subsys_state * __ref
4089 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4091 struct mem_cgroup *mem, *parent;
4092 long error = -ENOMEM;
4095 mem = mem_cgroup_alloc();
4097 return ERR_PTR(error);
4099 for_each_node_state(node, N_POSSIBLE)
4100 if (alloc_mem_cgroup_per_zone_info(mem, node))
4104 if (cont->parent == NULL) {
4106 enable_swap_cgroup();
4108 root_mem_cgroup = mem;
4109 if (mem_cgroup_soft_limit_tree_init())
4111 for_each_possible_cpu(cpu) {
4112 struct memcg_stock_pcp *stock =
4113 &per_cpu(memcg_stock, cpu);
4114 INIT_WORK(&stock->work, drain_local_stock);
4116 hotcpu_notifier(memcg_stock_cpu_callback, 0);
4118 parent = mem_cgroup_from_cont(cont->parent);
4119 mem->use_hierarchy = parent->use_hierarchy;
4120 mem->oom_kill_disable = parent->oom_kill_disable;
4123 if (parent && parent->use_hierarchy) {
4124 res_counter_init(&mem->res, &parent->res);
4125 res_counter_init(&mem->memsw, &parent->memsw);
4127 * We increment refcnt of the parent to ensure that we can
4128 * safely access it on res_counter_charge/uncharge.
4129 * This refcnt will be decremented when freeing this
4130 * mem_cgroup(see mem_cgroup_put).
4132 mem_cgroup_get(parent);
4134 res_counter_init(&mem->res, NULL);
4135 res_counter_init(&mem->memsw, NULL);
4137 mem->last_scanned_child = 0;
4138 spin_lock_init(&mem->reclaim_param_lock);
4139 INIT_LIST_HEAD(&mem->oom_notify);
4142 mem->swappiness = get_swappiness(parent);
4143 atomic_set(&mem->refcnt, 1);
4144 mem->move_charge_at_immigrate = 0;
4145 mutex_init(&mem->thresholds_lock);
4148 __mem_cgroup_free(mem);
4149 root_mem_cgroup = NULL;
4150 return ERR_PTR(error);
4153 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4154 struct cgroup *cont)
4156 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4158 return mem_cgroup_force_empty(mem, false);
4161 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4162 struct cgroup *cont)
4164 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4166 mem_cgroup_put(mem);
4169 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4170 struct cgroup *cont)
4174 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4175 ARRAY_SIZE(mem_cgroup_files));
4178 ret = register_memsw_files(cont, ss);
4183 /* Handlers for move charge at task migration. */
4184 #define PRECHARGE_COUNT_AT_ONCE 256
4185 static int mem_cgroup_do_precharge(unsigned long count)
4188 int batch_count = PRECHARGE_COUNT_AT_ONCE;
4189 struct mem_cgroup *mem = mc.to;
4191 if (mem_cgroup_is_root(mem)) {
4192 mc.precharge += count;
4193 /* we don't need css_get for root */
4196 /* try to charge at once */
4198 struct res_counter *dummy;
4200 * "mem" cannot be under rmdir() because we've already checked
4201 * by cgroup_lock_live_cgroup() that it is not removed and we
4202 * are still under the same cgroup_mutex. So we can postpone
4205 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4207 if (do_swap_account && res_counter_charge(&mem->memsw,
4208 PAGE_SIZE * count, &dummy)) {
4209 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4212 mc.precharge += count;
4213 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
4214 WARN_ON_ONCE(count > INT_MAX);
4215 __css_get(&mem->css, (int)count);
4219 /* fall back to one by one charge */
4221 if (signal_pending(current)) {
4225 if (!batch_count--) {
4226 batch_count = PRECHARGE_COUNT_AT_ONCE;
4229 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
4231 /* mem_cgroup_clear_mc() will do uncharge later */
4239 * is_target_pte_for_mc - check a pte whether it is valid for move charge
4240 * @vma: the vma the pte to be checked belongs
4241 * @addr: the address corresponding to the pte to be checked
4242 * @ptent: the pte to be checked
4243 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4246 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4247 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4248 * move charge. if @target is not NULL, the page is stored in target->page
4249 * with extra refcnt got(Callers should handle it).
4250 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4251 * target for charge migration. if @target is not NULL, the entry is stored
4254 * Called with pte lock held.
4261 enum mc_target_type {
4262 MC_TARGET_NONE, /* not used */
4267 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4268 unsigned long addr, pte_t ptent)
4270 struct page *page = vm_normal_page(vma, addr, ptent);
4272 if (!page || !page_mapped(page))
4274 if (PageAnon(page)) {
4275 /* we don't move shared anon */
4276 if (!move_anon() || page_mapcount(page) > 2)
4278 } else if (!move_file())
4279 /* we ignore mapcount for file pages */
4281 if (!get_page_unless_zero(page))
4287 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4288 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4291 struct page *page = NULL;
4292 swp_entry_t ent = pte_to_swp_entry(ptent);
4294 if (!move_anon() || non_swap_entry(ent))
4296 usage_count = mem_cgroup_count_swap_user(ent, &page);
4297 if (usage_count > 1) { /* we don't move shared anon */
4302 if (do_swap_account)
4303 entry->val = ent.val;
4308 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4309 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4311 struct page *page = NULL;
4312 struct inode *inode;
4313 struct address_space *mapping;
4316 if (!vma->vm_file) /* anonymous vma */
4321 inode = vma->vm_file->f_path.dentry->d_inode;
4322 mapping = vma->vm_file->f_mapping;
4323 if (pte_none(ptent))
4324 pgoff = linear_page_index(vma, addr);
4325 else /* pte_file(ptent) is true */
4326 pgoff = pte_to_pgoff(ptent);
4328 /* page is moved even if it's not RSS of this task(page-faulted). */
4329 if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4330 page = find_get_page(mapping, pgoff);
4331 } else { /* shmem/tmpfs file. we should take account of swap too. */
4333 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4334 if (do_swap_account)
4335 entry->val = ent.val;
4341 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4342 unsigned long addr, pte_t ptent, union mc_target *target)
4344 struct page *page = NULL;
4345 struct page_cgroup *pc;
4347 swp_entry_t ent = { .val = 0 };
4349 if (pte_present(ptent))
4350 page = mc_handle_present_pte(vma, addr, ptent);
4351 else if (is_swap_pte(ptent))
4352 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4353 else if (pte_none(ptent) || pte_file(ptent))
4354 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4356 if (!page && !ent.val)
4359 pc = lookup_page_cgroup(page);
4361 * Do only loose check w/o page_cgroup lock.
4362 * mem_cgroup_move_account() checks the pc is valid or not under
4365 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4366 ret = MC_TARGET_PAGE;
4368 target->page = page;
4370 if (!ret || !target)
4373 /* There is a swap entry and a page doesn't exist or isn't charged */
4374 if (ent.val && !ret &&
4375 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4376 ret = MC_TARGET_SWAP;
4383 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4384 unsigned long addr, unsigned long end,
4385 struct mm_walk *walk)
4387 struct vm_area_struct *vma = walk->private;
4391 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4392 for (; addr != end; pte++, addr += PAGE_SIZE)
4393 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4394 mc.precharge++; /* increment precharge temporarily */
4395 pte_unmap_unlock(pte - 1, ptl);
4401 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4403 unsigned long precharge;
4404 struct vm_area_struct *vma;
4406 down_read(&mm->mmap_sem);
4407 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4408 struct mm_walk mem_cgroup_count_precharge_walk = {
4409 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4413 if (is_vm_hugetlb_page(vma))
4415 walk_page_range(vma->vm_start, vma->vm_end,
4416 &mem_cgroup_count_precharge_walk);
4418 up_read(&mm->mmap_sem);
4420 precharge = mc.precharge;
4426 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4428 return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm));
4431 static void mem_cgroup_clear_mc(void)
4433 struct mem_cgroup *from = mc.from;
4434 struct mem_cgroup *to = mc.to;
4436 /* we must uncharge all the leftover precharges from mc.to */
4438 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4442 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4443 * we must uncharge here.
4445 if (mc.moved_charge) {
4446 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4447 mc.moved_charge = 0;
4449 /* we must fixup refcnts and charges */
4450 if (mc.moved_swap) {
4451 WARN_ON_ONCE(mc.moved_swap > INT_MAX);
4452 /* uncharge swap account from the old cgroup */
4453 if (!mem_cgroup_is_root(mc.from))
4454 res_counter_uncharge(&mc.from->memsw,
4455 PAGE_SIZE * mc.moved_swap);
4456 __mem_cgroup_put(mc.from, mc.moved_swap);
4458 if (!mem_cgroup_is_root(mc.to)) {
4460 * we charged both to->res and to->memsw, so we should
4463 res_counter_uncharge(&mc.to->res,
4464 PAGE_SIZE * mc.moved_swap);
4465 VM_BUG_ON(test_bit(CSS_ROOT, &mc.to->css.flags));
4466 __css_put(&mc.to->css, mc.moved_swap);
4468 /* we've already done mem_cgroup_get(mc.to) */
4472 spin_lock(&mc.lock);
4475 mc.moving_task = NULL;
4476 spin_unlock(&mc.lock);
4477 memcg_oom_recover(from);
4478 memcg_oom_recover(to);
4479 wake_up_all(&mc.waitq);
4482 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4483 struct cgroup *cgroup,
4484 struct task_struct *p,
4488 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4490 if (mem->move_charge_at_immigrate) {
4491 struct mm_struct *mm;
4492 struct mem_cgroup *from = mem_cgroup_from_task(p);
4494 VM_BUG_ON(from == mem);
4496 mm = get_task_mm(p);
4499 /* We move charges only when we move a owner of the mm */
4500 if (mm->owner == p) {
4503 VM_BUG_ON(mc.precharge);
4504 VM_BUG_ON(mc.moved_charge);
4505 VM_BUG_ON(mc.moved_swap);
4506 VM_BUG_ON(mc.moving_task);
4507 spin_lock(&mc.lock);
4511 mc.moved_charge = 0;
4513 mc.moving_task = current;
4514 spin_unlock(&mc.lock);
4516 ret = mem_cgroup_precharge_mc(mm);
4518 mem_cgroup_clear_mc();
4525 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4526 struct cgroup *cgroup,
4527 struct task_struct *p,
4530 mem_cgroup_clear_mc();
4533 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4534 unsigned long addr, unsigned long end,
4535 struct mm_walk *walk)
4538 struct vm_area_struct *vma = walk->private;
4543 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4544 for (; addr != end; addr += PAGE_SIZE) {
4545 pte_t ptent = *(pte++);
4546 union mc_target target;
4549 struct page_cgroup *pc;
4555 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4557 case MC_TARGET_PAGE:
4559 if (isolate_lru_page(page))
4561 pc = lookup_page_cgroup(page);
4562 if (!mem_cgroup_move_account(pc,
4563 mc.from, mc.to, false)) {
4565 /* we uncharge from mc.from later. */
4568 putback_lru_page(page);
4569 put: /* is_target_pte_for_mc() gets the page */
4572 case MC_TARGET_SWAP:
4574 if (!mem_cgroup_move_swap_account(ent,
4575 mc.from, mc.to, false)) {
4577 /* we fixup refcnts and charges later. */
4585 pte_unmap_unlock(pte - 1, ptl);
4590 * We have consumed all precharges we got in can_attach().
4591 * We try charge one by one, but don't do any additional
4592 * charges to mc.to if we have failed in charge once in attach()
4595 ret = mem_cgroup_do_precharge(1);
4603 static void mem_cgroup_move_charge(struct mm_struct *mm)
4605 struct vm_area_struct *vma;
4607 lru_add_drain_all();
4608 down_read(&mm->mmap_sem);
4609 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4611 struct mm_walk mem_cgroup_move_charge_walk = {
4612 .pmd_entry = mem_cgroup_move_charge_pte_range,
4616 if (is_vm_hugetlb_page(vma))
4618 ret = walk_page_range(vma->vm_start, vma->vm_end,
4619 &mem_cgroup_move_charge_walk);
4622 * means we have consumed all precharges and failed in
4623 * doing additional charge. Just abandon here.
4627 up_read(&mm->mmap_sem);
4630 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4631 struct cgroup *cont,
4632 struct cgroup *old_cont,
4633 struct task_struct *p,
4636 struct mm_struct *mm;
4639 /* no need to move charge */
4642 mm = get_task_mm(p);
4644 mem_cgroup_move_charge(mm);
4647 mem_cgroup_clear_mc();
4649 #else /* !CONFIG_MMU */
4650 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4651 struct cgroup *cgroup,
4652 struct task_struct *p,
4657 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4658 struct cgroup *cgroup,
4659 struct task_struct *p,
4663 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4664 struct cgroup *cont,
4665 struct cgroup *old_cont,
4666 struct task_struct *p,
4672 struct cgroup_subsys mem_cgroup_subsys = {
4674 .subsys_id = mem_cgroup_subsys_id,
4675 .create = mem_cgroup_create,
4676 .pre_destroy = mem_cgroup_pre_destroy,
4677 .destroy = mem_cgroup_destroy,
4678 .populate = mem_cgroup_populate,
4679 .can_attach = mem_cgroup_can_attach,
4680 .cancel_attach = mem_cgroup_cancel_attach,
4681 .attach = mem_cgroup_move_task,
4686 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4688 static int __init disable_swap_account(char *s)
4690 really_do_swap_account = 0;
4693 __setup("noswapaccount", disable_swap_account);