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 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
55 #define MEM_CGROUP_RECLAIM_RETRIES 5
56 struct mem_cgroup *root_mem_cgroup __read_mostly;
58 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
59 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
60 int do_swap_account __read_mostly;
61 static int really_do_swap_account __initdata = 1; /* for remember boot option*/
63 #define do_swap_account (0)
67 * Per memcg event counter is incremented at every pagein/pageout. This counter
68 * is used for trigger some periodic events. This is straightforward and better
69 * than using jiffies etc. to handle periodic memcg event.
71 * These values will be used as !((event) & ((1 <<(thresh)) - 1))
73 #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
74 #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
77 * Statistics for memory cgroup.
79 enum mem_cgroup_stat_index {
81 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
83 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
84 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
85 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
86 MEM_CGROUP_STAT_PGPGIN_COUNT, /* # of pages paged in */
87 MEM_CGROUP_STAT_PGPGOUT_COUNT, /* # of pages paged out */
88 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
89 MEM_CGROUP_EVENTS, /* incremented at every pagein/pageout */
91 MEM_CGROUP_STAT_NSTATS,
94 struct mem_cgroup_stat_cpu {
95 s64 count[MEM_CGROUP_STAT_NSTATS];
99 * per-zone information in memory controller.
101 struct mem_cgroup_per_zone {
103 * spin_lock to protect the per cgroup LRU
105 struct list_head lists[NR_LRU_LISTS];
106 unsigned long count[NR_LRU_LISTS];
108 struct zone_reclaim_stat reclaim_stat;
109 struct rb_node tree_node; /* RB tree node */
110 unsigned long long usage_in_excess;/* Set to the value by which */
111 /* the soft limit is exceeded*/
113 struct mem_cgroup *mem; /* Back pointer, we cannot */
114 /* use container_of */
116 /* Macro for accessing counter */
117 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
119 struct mem_cgroup_per_node {
120 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
123 struct mem_cgroup_lru_info {
124 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
128 * Cgroups above their limits are maintained in a RB-Tree, independent of
129 * their hierarchy representation
132 struct mem_cgroup_tree_per_zone {
133 struct rb_root rb_root;
137 struct mem_cgroup_tree_per_node {
138 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
141 struct mem_cgroup_tree {
142 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
145 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
147 struct mem_cgroup_threshold {
148 struct eventfd_ctx *eventfd;
153 struct mem_cgroup_threshold_ary {
154 /* An array index points to threshold just below usage. */
155 int current_threshold;
156 /* Size of entries[] */
158 /* Array of thresholds */
159 struct mem_cgroup_threshold entries[0];
162 struct mem_cgroup_eventfd_list {
163 struct list_head list;
164 struct eventfd_ctx *eventfd;
167 static void mem_cgroup_threshold(struct mem_cgroup *mem);
168 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
171 * The memory controller data structure. The memory controller controls both
172 * page cache and RSS per cgroup. We would eventually like to provide
173 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
174 * to help the administrator determine what knobs to tune.
176 * TODO: Add a water mark for the memory controller. Reclaim will begin when
177 * we hit the water mark. May be even add a low water mark, such that
178 * no reclaim occurs from a cgroup at it's low water mark, this is
179 * a feature that will be implemented much later in the future.
182 struct cgroup_subsys_state css;
184 * the counter to account for memory usage
186 struct res_counter res;
188 * the counter to account for mem+swap usage.
190 struct res_counter memsw;
192 * Per cgroup active and inactive list, similar to the
193 * per zone LRU lists.
195 struct mem_cgroup_lru_info info;
198 protect against reclaim related member.
200 spinlock_t reclaim_param_lock;
202 int prev_priority; /* for recording reclaim priority */
205 * While reclaiming in a hierarchy, we cache the last child we
208 int last_scanned_child;
210 * Should the accounting and control be hierarchical, per subtree?
216 unsigned int swappiness;
217 /* OOM-Killer disable */
218 int oom_kill_disable;
220 /* set when res.limit == memsw.limit */
221 bool memsw_is_minimum;
223 /* protect arrays of thresholds */
224 struct mutex thresholds_lock;
226 /* thresholds for memory usage. RCU-protected */
227 struct mem_cgroup_threshold_ary *thresholds;
229 /* thresholds for mem+swap usage. RCU-protected */
230 struct mem_cgroup_threshold_ary *memsw_thresholds;
232 /* For oom notifier event fd */
233 struct list_head oom_notify;
236 * Should we move charges of a task when a task is moved into this
237 * mem_cgroup ? And what type of charges should we move ?
239 unsigned long move_charge_at_immigrate;
243 struct mem_cgroup_stat_cpu *stat;
246 /* Stuffs for move charges at task migration. */
248 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
249 * left-shifted bitmap of these types.
252 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
253 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
257 /* "mc" and its members are protected by cgroup_mutex */
258 static struct move_charge_struct {
259 struct mem_cgroup *from;
260 struct mem_cgroup *to;
261 unsigned long precharge;
262 unsigned long moved_charge;
263 unsigned long moved_swap;
264 struct task_struct *moving_task; /* a task moving charges */
265 wait_queue_head_t waitq; /* a waitq for other context */
267 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
270 static bool move_anon(void)
272 return test_bit(MOVE_CHARGE_TYPE_ANON,
273 &mc.to->move_charge_at_immigrate);
276 static bool move_file(void)
278 return test_bit(MOVE_CHARGE_TYPE_FILE,
279 &mc.to->move_charge_at_immigrate);
283 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
284 * limit reclaim to prevent infinite loops, if they ever occur.
286 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
287 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
290 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
291 MEM_CGROUP_CHARGE_TYPE_MAPPED,
292 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
293 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
294 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
295 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
299 /* only for here (for easy reading.) */
300 #define PCGF_CACHE (1UL << PCG_CACHE)
301 #define PCGF_USED (1UL << PCG_USED)
302 #define PCGF_LOCK (1UL << PCG_LOCK)
303 /* Not used, but added here for completeness */
304 #define PCGF_ACCT (1UL << PCG_ACCT)
306 /* for encoding cft->private value on file */
309 #define _OOM_TYPE (2)
310 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
311 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
312 #define MEMFILE_ATTR(val) ((val) & 0xffff)
313 /* Used for OOM nofiier */
314 #define OOM_CONTROL (0)
317 * Reclaim flags for mem_cgroup_hierarchical_reclaim
319 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
320 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
321 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
322 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
323 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
324 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
326 static void mem_cgroup_get(struct mem_cgroup *mem);
327 static void mem_cgroup_put(struct mem_cgroup *mem);
328 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
329 static void drain_all_stock_async(void);
331 static struct mem_cgroup_per_zone *
332 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
334 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
337 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
342 static struct mem_cgroup_per_zone *
343 page_cgroup_zoneinfo(struct page_cgroup *pc)
345 struct mem_cgroup *mem = pc->mem_cgroup;
346 int nid = page_cgroup_nid(pc);
347 int zid = page_cgroup_zid(pc);
352 return mem_cgroup_zoneinfo(mem, nid, zid);
355 static struct mem_cgroup_tree_per_zone *
356 soft_limit_tree_node_zone(int nid, int zid)
358 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
361 static struct mem_cgroup_tree_per_zone *
362 soft_limit_tree_from_page(struct page *page)
364 int nid = page_to_nid(page);
365 int zid = page_zonenum(page);
367 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
371 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
372 struct mem_cgroup_per_zone *mz,
373 struct mem_cgroup_tree_per_zone *mctz,
374 unsigned long long new_usage_in_excess)
376 struct rb_node **p = &mctz->rb_root.rb_node;
377 struct rb_node *parent = NULL;
378 struct mem_cgroup_per_zone *mz_node;
383 mz->usage_in_excess = new_usage_in_excess;
384 if (!mz->usage_in_excess)
388 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
390 if (mz->usage_in_excess < mz_node->usage_in_excess)
393 * We can't avoid mem cgroups that are over their soft
394 * limit by the same amount
396 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
399 rb_link_node(&mz->tree_node, parent, p);
400 rb_insert_color(&mz->tree_node, &mctz->rb_root);
405 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
406 struct mem_cgroup_per_zone *mz,
407 struct mem_cgroup_tree_per_zone *mctz)
411 rb_erase(&mz->tree_node, &mctz->rb_root);
416 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
417 struct mem_cgroup_per_zone *mz,
418 struct mem_cgroup_tree_per_zone *mctz)
420 spin_lock(&mctz->lock);
421 __mem_cgroup_remove_exceeded(mem, mz, mctz);
422 spin_unlock(&mctz->lock);
426 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
428 unsigned long long excess;
429 struct mem_cgroup_per_zone *mz;
430 struct mem_cgroup_tree_per_zone *mctz;
431 int nid = page_to_nid(page);
432 int zid = page_zonenum(page);
433 mctz = soft_limit_tree_from_page(page);
436 * Necessary to update all ancestors when hierarchy is used.
437 * because their event counter is not touched.
439 for (; mem; mem = parent_mem_cgroup(mem)) {
440 mz = mem_cgroup_zoneinfo(mem, nid, zid);
441 excess = res_counter_soft_limit_excess(&mem->res);
443 * We have to update the tree if mz is on RB-tree or
444 * mem is over its softlimit.
446 if (excess || mz->on_tree) {
447 spin_lock(&mctz->lock);
448 /* if on-tree, remove it */
450 __mem_cgroup_remove_exceeded(mem, mz, mctz);
452 * Insert again. mz->usage_in_excess will be updated.
453 * If excess is 0, no tree ops.
455 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
456 spin_unlock(&mctz->lock);
461 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
464 struct mem_cgroup_per_zone *mz;
465 struct mem_cgroup_tree_per_zone *mctz;
467 for_each_node_state(node, N_POSSIBLE) {
468 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
469 mz = mem_cgroup_zoneinfo(mem, node, zone);
470 mctz = soft_limit_tree_node_zone(node, zone);
471 mem_cgroup_remove_exceeded(mem, mz, mctz);
476 static inline unsigned long mem_cgroup_get_excess(struct mem_cgroup *mem)
478 return res_counter_soft_limit_excess(&mem->res) >> PAGE_SHIFT;
481 static struct mem_cgroup_per_zone *
482 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
484 struct rb_node *rightmost = NULL;
485 struct mem_cgroup_per_zone *mz;
489 rightmost = rb_last(&mctz->rb_root);
491 goto done; /* Nothing to reclaim from */
493 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
495 * Remove the node now but someone else can add it back,
496 * we will to add it back at the end of reclaim to its correct
497 * position in the tree.
499 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
500 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
501 !css_tryget(&mz->mem->css))
507 static struct mem_cgroup_per_zone *
508 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
510 struct mem_cgroup_per_zone *mz;
512 spin_lock(&mctz->lock);
513 mz = __mem_cgroup_largest_soft_limit_node(mctz);
514 spin_unlock(&mctz->lock);
518 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
519 enum mem_cgroup_stat_index idx)
524 for_each_possible_cpu(cpu)
525 val += per_cpu(mem->stat->count[idx], cpu);
529 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
533 ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
534 ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
538 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
541 int val = (charge) ? 1 : -1;
542 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
545 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
546 struct page_cgroup *pc,
549 int val = (charge) ? 1 : -1;
553 if (PageCgroupCache(pc))
554 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], val);
556 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], val);
559 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
561 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
562 __this_cpu_inc(mem->stat->count[MEM_CGROUP_EVENTS]);
567 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
571 struct mem_cgroup_per_zone *mz;
574 for_each_online_node(nid)
575 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
576 mz = mem_cgroup_zoneinfo(mem, nid, zid);
577 total += MEM_CGROUP_ZSTAT(mz, idx);
582 static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
586 val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
588 return !(val & ((1 << event_mask_shift) - 1));
592 * Check events in order.
595 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
597 /* threshold event is triggered in finer grain than soft limit */
598 if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
599 mem_cgroup_threshold(mem);
600 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
601 mem_cgroup_update_tree(mem, page);
605 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
607 return container_of(cgroup_subsys_state(cont,
608 mem_cgroup_subsys_id), struct mem_cgroup,
612 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
615 * mm_update_next_owner() may clear mm->owner to NULL
616 * if it races with swapoff, page migration, etc.
617 * So this can be called with p == NULL.
622 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
623 struct mem_cgroup, css);
626 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
628 struct mem_cgroup *mem = NULL;
633 * Because we have no locks, mm->owner's may be being moved to other
634 * cgroup. We use css_tryget() here even if this looks
635 * pessimistic (rather than adding locks here).
639 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
642 } while (!css_tryget(&mem->css));
648 * Call callback function against all cgroup under hierarchy tree.
650 static int mem_cgroup_walk_tree(struct mem_cgroup *root, void *data,
651 int (*func)(struct mem_cgroup *, void *))
653 int found, ret, nextid;
654 struct cgroup_subsys_state *css;
655 struct mem_cgroup *mem;
657 if (!root->use_hierarchy)
658 return (*func)(root, data);
666 css = css_get_next(&mem_cgroup_subsys, nextid, &root->css,
668 if (css && css_tryget(css))
669 mem = container_of(css, struct mem_cgroup, css);
673 ret = (*func)(mem, data);
677 } while (!ret && css);
682 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
684 return (mem == root_mem_cgroup);
688 * Following LRU functions are allowed to be used without PCG_LOCK.
689 * Operations are called by routine of global LRU independently from memcg.
690 * What we have to take care of here is validness of pc->mem_cgroup.
692 * Changes to pc->mem_cgroup happens when
695 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
696 * It is added to LRU before charge.
697 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
698 * When moving account, the page is not on LRU. It's isolated.
701 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
703 struct page_cgroup *pc;
704 struct mem_cgroup_per_zone *mz;
706 if (mem_cgroup_disabled())
708 pc = lookup_page_cgroup(page);
709 /* can happen while we handle swapcache. */
710 if (!TestClearPageCgroupAcctLRU(pc))
712 VM_BUG_ON(!pc->mem_cgroup);
714 * We don't check PCG_USED bit. It's cleared when the "page" is finally
715 * removed from global LRU.
717 mz = page_cgroup_zoneinfo(pc);
718 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
719 if (mem_cgroup_is_root(pc->mem_cgroup))
721 VM_BUG_ON(list_empty(&pc->lru));
722 list_del_init(&pc->lru);
726 void mem_cgroup_del_lru(struct page *page)
728 mem_cgroup_del_lru_list(page, page_lru(page));
731 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
733 struct mem_cgroup_per_zone *mz;
734 struct page_cgroup *pc;
736 if (mem_cgroup_disabled())
739 pc = lookup_page_cgroup(page);
741 * Used bit is set without atomic ops but after smp_wmb().
742 * For making pc->mem_cgroup visible, insert smp_rmb() here.
745 /* unused or root page is not rotated. */
746 if (!PageCgroupUsed(pc) || mem_cgroup_is_root(pc->mem_cgroup))
748 mz = page_cgroup_zoneinfo(pc);
749 list_move(&pc->lru, &mz->lists[lru]);
752 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
754 struct page_cgroup *pc;
755 struct mem_cgroup_per_zone *mz;
757 if (mem_cgroup_disabled())
759 pc = lookup_page_cgroup(page);
760 VM_BUG_ON(PageCgroupAcctLRU(pc));
762 * Used bit is set without atomic ops but after smp_wmb().
763 * For making pc->mem_cgroup visible, insert smp_rmb() here.
766 if (!PageCgroupUsed(pc))
769 mz = page_cgroup_zoneinfo(pc);
770 MEM_CGROUP_ZSTAT(mz, lru) += 1;
771 SetPageCgroupAcctLRU(pc);
772 if (mem_cgroup_is_root(pc->mem_cgroup))
774 list_add(&pc->lru, &mz->lists[lru]);
778 * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
779 * lru because the page may.be reused after it's fully uncharged (because of
780 * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
781 * it again. This function is only used to charge SwapCache. It's done under
782 * lock_page and expected that zone->lru_lock is never held.
784 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
787 struct zone *zone = page_zone(page);
788 struct page_cgroup *pc = lookup_page_cgroup(page);
790 spin_lock_irqsave(&zone->lru_lock, flags);
792 * Forget old LRU when this page_cgroup is *not* used. This Used bit
793 * is guarded by lock_page() because the page is SwapCache.
795 if (!PageCgroupUsed(pc))
796 mem_cgroup_del_lru_list(page, page_lru(page));
797 spin_unlock_irqrestore(&zone->lru_lock, flags);
800 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
803 struct zone *zone = page_zone(page);
804 struct page_cgroup *pc = lookup_page_cgroup(page);
806 spin_lock_irqsave(&zone->lru_lock, flags);
807 /* link when the page is linked to LRU but page_cgroup isn't */
808 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
809 mem_cgroup_add_lru_list(page, page_lru(page));
810 spin_unlock_irqrestore(&zone->lru_lock, flags);
814 void mem_cgroup_move_lists(struct page *page,
815 enum lru_list from, enum lru_list to)
817 if (mem_cgroup_disabled())
819 mem_cgroup_del_lru_list(page, from);
820 mem_cgroup_add_lru_list(page, to);
823 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
826 struct mem_cgroup *curr = NULL;
830 curr = try_get_mem_cgroup_from_mm(task->mm);
836 * We should check use_hierarchy of "mem" not "curr". Because checking
837 * use_hierarchy of "curr" here make this function true if hierarchy is
838 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
839 * hierarchy(even if use_hierarchy is disabled in "mem").
841 if (mem->use_hierarchy)
842 ret = css_is_ancestor(&curr->css, &mem->css);
850 * prev_priority control...this will be used in memory reclaim path.
852 int mem_cgroup_get_reclaim_priority(struct mem_cgroup *mem)
856 spin_lock(&mem->reclaim_param_lock);
857 prev_priority = mem->prev_priority;
858 spin_unlock(&mem->reclaim_param_lock);
860 return prev_priority;
863 void mem_cgroup_note_reclaim_priority(struct mem_cgroup *mem, int priority)
865 spin_lock(&mem->reclaim_param_lock);
866 if (priority < mem->prev_priority)
867 mem->prev_priority = priority;
868 spin_unlock(&mem->reclaim_param_lock);
871 void mem_cgroup_record_reclaim_priority(struct mem_cgroup *mem, int priority)
873 spin_lock(&mem->reclaim_param_lock);
874 mem->prev_priority = priority;
875 spin_unlock(&mem->reclaim_param_lock);
878 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
880 unsigned long active;
881 unsigned long inactive;
883 unsigned long inactive_ratio;
885 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
886 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
888 gb = (inactive + active) >> (30 - PAGE_SHIFT);
890 inactive_ratio = int_sqrt(10 * gb);
895 present_pages[0] = inactive;
896 present_pages[1] = active;
899 return inactive_ratio;
902 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
904 unsigned long active;
905 unsigned long inactive;
906 unsigned long present_pages[2];
907 unsigned long inactive_ratio;
909 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
911 inactive = present_pages[0];
912 active = present_pages[1];
914 if (inactive * inactive_ratio < active)
920 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
922 unsigned long active;
923 unsigned long inactive;
925 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
926 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
928 return (active > inactive);
931 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
935 int nid = zone->zone_pgdat->node_id;
936 int zid = zone_idx(zone);
937 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
939 return MEM_CGROUP_ZSTAT(mz, lru);
942 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
945 int nid = zone->zone_pgdat->node_id;
946 int zid = zone_idx(zone);
947 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
949 return &mz->reclaim_stat;
952 struct zone_reclaim_stat *
953 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
955 struct page_cgroup *pc;
956 struct mem_cgroup_per_zone *mz;
958 if (mem_cgroup_disabled())
961 pc = lookup_page_cgroup(page);
963 * Used bit is set without atomic ops but after smp_wmb().
964 * For making pc->mem_cgroup visible, insert smp_rmb() here.
967 if (!PageCgroupUsed(pc))
970 mz = page_cgroup_zoneinfo(pc);
974 return &mz->reclaim_stat;
977 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
978 struct list_head *dst,
979 unsigned long *scanned, int order,
980 int mode, struct zone *z,
981 struct mem_cgroup *mem_cont,
982 int active, int file)
984 unsigned long nr_taken = 0;
988 struct list_head *src;
989 struct page_cgroup *pc, *tmp;
990 int nid = z->zone_pgdat->node_id;
991 int zid = zone_idx(z);
992 struct mem_cgroup_per_zone *mz;
993 int lru = LRU_FILE * file + active;
997 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
998 src = &mz->lists[lru];
1001 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1002 if (scan >= nr_to_scan)
1006 if (unlikely(!PageCgroupUsed(pc)))
1008 if (unlikely(!PageLRU(page)))
1012 ret = __isolate_lru_page(page, mode, file);
1015 list_move(&page->lru, dst);
1016 mem_cgroup_del_lru(page);
1020 /* we don't affect global LRU but rotate in our LRU */
1021 mem_cgroup_rotate_lru_list(page, page_lru(page));
1032 #define mem_cgroup_from_res_counter(counter, member) \
1033 container_of(counter, struct mem_cgroup, member)
1035 static bool mem_cgroup_check_under_limit(struct mem_cgroup *mem)
1037 if (do_swap_account) {
1038 if (res_counter_check_under_limit(&mem->res) &&
1039 res_counter_check_under_limit(&mem->memsw))
1042 if (res_counter_check_under_limit(&mem->res))
1047 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1049 struct cgroup *cgrp = memcg->css.cgroup;
1050 unsigned int swappiness;
1053 if (cgrp->parent == NULL)
1054 return vm_swappiness;
1056 spin_lock(&memcg->reclaim_param_lock);
1057 swappiness = memcg->swappiness;
1058 spin_unlock(&memcg->reclaim_param_lock);
1063 static int mem_cgroup_count_children_cb(struct mem_cgroup *mem, void *data)
1071 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1072 * @memcg: The memory cgroup that went over limit
1073 * @p: Task that is going to be killed
1075 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1078 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1080 struct cgroup *task_cgrp;
1081 struct cgroup *mem_cgrp;
1083 * Need a buffer in BSS, can't rely on allocations. The code relies
1084 * on the assumption that OOM is serialized for memory controller.
1085 * If this assumption is broken, revisit this code.
1087 static char memcg_name[PATH_MAX];
1096 mem_cgrp = memcg->css.cgroup;
1097 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1099 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1102 * Unfortunately, we are unable to convert to a useful name
1103 * But we'll still print out the usage information
1110 printk(KERN_INFO "Task in %s killed", memcg_name);
1113 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1121 * Continues from above, so we don't need an KERN_ level
1123 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1126 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1127 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1128 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1129 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1130 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1132 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1133 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1134 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1138 * This function returns the number of memcg under hierarchy tree. Returns
1139 * 1(self count) if no children.
1141 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1144 mem_cgroup_walk_tree(mem, &num, mem_cgroup_count_children_cb);
1149 * Visit the first child (need not be the first child as per the ordering
1150 * of the cgroup list, since we track last_scanned_child) of @mem and use
1151 * that to reclaim free pages from.
1153 static struct mem_cgroup *
1154 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1156 struct mem_cgroup *ret = NULL;
1157 struct cgroup_subsys_state *css;
1160 if (!root_mem->use_hierarchy) {
1161 css_get(&root_mem->css);
1167 nextid = root_mem->last_scanned_child + 1;
1168 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1170 if (css && css_tryget(css))
1171 ret = container_of(css, struct mem_cgroup, css);
1174 /* Updates scanning parameter */
1175 spin_lock(&root_mem->reclaim_param_lock);
1177 /* this means start scan from ID:1 */
1178 root_mem->last_scanned_child = 0;
1180 root_mem->last_scanned_child = found;
1181 spin_unlock(&root_mem->reclaim_param_lock);
1188 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1189 * we reclaimed from, so that we don't end up penalizing one child extensively
1190 * based on its position in the children list.
1192 * root_mem is the original ancestor that we've been reclaim from.
1194 * We give up and return to the caller when we visit root_mem twice.
1195 * (other groups can be removed while we're walking....)
1197 * If shrink==true, for avoiding to free too much, this returns immedieately.
1199 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1202 unsigned long reclaim_options)
1204 struct mem_cgroup *victim;
1207 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1208 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1209 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1210 unsigned long excess = mem_cgroup_get_excess(root_mem);
1212 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1213 if (root_mem->memsw_is_minimum)
1217 victim = mem_cgroup_select_victim(root_mem);
1218 if (victim == root_mem) {
1221 drain_all_stock_async();
1224 * If we have not been able to reclaim
1225 * anything, it might because there are
1226 * no reclaimable pages under this hierarchy
1228 if (!check_soft || !total) {
1229 css_put(&victim->css);
1233 * We want to do more targetted reclaim.
1234 * excess >> 2 is not to excessive so as to
1235 * reclaim too much, nor too less that we keep
1236 * coming back to reclaim from this cgroup
1238 if (total >= (excess >> 2) ||
1239 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1240 css_put(&victim->css);
1245 if (!mem_cgroup_local_usage(victim)) {
1246 /* this cgroup's local usage == 0 */
1247 css_put(&victim->css);
1250 /* we use swappiness of local cgroup */
1252 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1253 noswap, get_swappiness(victim), zone,
1254 zone->zone_pgdat->node_id);
1256 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1257 noswap, get_swappiness(victim));
1258 css_put(&victim->css);
1260 * At shrinking usage, we can't check we should stop here or
1261 * reclaim more. It's depends on callers. last_scanned_child
1262 * will work enough for keeping fairness under tree.
1268 if (res_counter_check_under_soft_limit(&root_mem->res))
1270 } else if (mem_cgroup_check_under_limit(root_mem))
1276 static int mem_cgroup_oom_lock_cb(struct mem_cgroup *mem, void *data)
1278 int *val = (int *)data;
1281 * Logically, we can stop scanning immediately when we find
1282 * a memcg is already locked. But condidering unlock ops and
1283 * creation/removal of memcg, scan-all is simple operation.
1285 x = atomic_inc_return(&mem->oom_lock);
1286 *val = max(x, *val);
1290 * Check OOM-Killer is already running under our hierarchy.
1291 * If someone is running, return false.
1293 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1297 mem_cgroup_walk_tree(mem, &lock_count, mem_cgroup_oom_lock_cb);
1299 if (lock_count == 1)
1304 static int mem_cgroup_oom_unlock_cb(struct mem_cgroup *mem, void *data)
1307 * When a new child is created while the hierarchy is under oom,
1308 * mem_cgroup_oom_lock() may not be called. We have to use
1309 * atomic_add_unless() here.
1311 atomic_add_unless(&mem->oom_lock, -1, 0);
1315 static void mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1317 mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_unlock_cb);
1320 static DEFINE_MUTEX(memcg_oom_mutex);
1321 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1323 struct oom_wait_info {
1324 struct mem_cgroup *mem;
1328 static int memcg_oom_wake_function(wait_queue_t *wait,
1329 unsigned mode, int sync, void *arg)
1331 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1332 struct oom_wait_info *oom_wait_info;
1334 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1336 if (oom_wait_info->mem == wake_mem)
1338 /* if no hierarchy, no match */
1339 if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1342 * Both of oom_wait_info->mem and wake_mem are stable under us.
1343 * Then we can use css_is_ancestor without taking care of RCU.
1345 if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1346 !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1350 return autoremove_wake_function(wait, mode, sync, arg);
1353 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1355 /* for filtering, pass "mem" as argument. */
1356 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1359 static void memcg_oom_recover(struct mem_cgroup *mem)
1361 if (mem->oom_kill_disable && atomic_read(&mem->oom_lock))
1362 memcg_wakeup_oom(mem);
1366 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1368 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1370 struct oom_wait_info owait;
1371 bool locked, need_to_kill;
1374 owait.wait.flags = 0;
1375 owait.wait.func = memcg_oom_wake_function;
1376 owait.wait.private = current;
1377 INIT_LIST_HEAD(&owait.wait.task_list);
1378 need_to_kill = true;
1379 /* At first, try to OOM lock hierarchy under mem.*/
1380 mutex_lock(&memcg_oom_mutex);
1381 locked = mem_cgroup_oom_lock(mem);
1383 * Even if signal_pending(), we can't quit charge() loop without
1384 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1385 * under OOM is always welcomed, use TASK_KILLABLE here.
1387 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1388 if (!locked || mem->oom_kill_disable)
1389 need_to_kill = false;
1391 mem_cgroup_oom_notify(mem);
1392 mutex_unlock(&memcg_oom_mutex);
1395 finish_wait(&memcg_oom_waitq, &owait.wait);
1396 mem_cgroup_out_of_memory(mem, mask);
1399 finish_wait(&memcg_oom_waitq, &owait.wait);
1401 mutex_lock(&memcg_oom_mutex);
1402 mem_cgroup_oom_unlock(mem);
1403 memcg_wakeup_oom(mem);
1404 mutex_unlock(&memcg_oom_mutex);
1406 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1408 /* Give chance to dying process */
1409 schedule_timeout(1);
1414 * Currently used to update mapped file statistics, but the routine can be
1415 * generalized to update other statistics as well.
1417 void mem_cgroup_update_file_mapped(struct page *page, int val)
1419 struct mem_cgroup *mem;
1420 struct page_cgroup *pc;
1422 pc = lookup_page_cgroup(page);
1426 lock_page_cgroup(pc);
1427 mem = pc->mem_cgroup;
1428 if (!mem || !PageCgroupUsed(pc))
1432 * Preemption is already disabled. We can use __this_cpu_xxx
1435 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1436 SetPageCgroupFileMapped(pc);
1438 __this_cpu_dec(mem->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1439 ClearPageCgroupFileMapped(pc);
1443 unlock_page_cgroup(pc);
1447 * size of first charge trial. "32" comes from vmscan.c's magic value.
1448 * TODO: maybe necessary to use big numbers in big irons.
1450 #define CHARGE_SIZE (32 * PAGE_SIZE)
1451 struct memcg_stock_pcp {
1452 struct mem_cgroup *cached; /* this never be root cgroup */
1454 struct work_struct work;
1456 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1457 static atomic_t memcg_drain_count;
1460 * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1461 * from local stock and true is returned. If the stock is 0 or charges from a
1462 * cgroup which is not current target, returns false. This stock will be
1465 static bool consume_stock(struct mem_cgroup *mem)
1467 struct memcg_stock_pcp *stock;
1470 stock = &get_cpu_var(memcg_stock);
1471 if (mem == stock->cached && stock->charge)
1472 stock->charge -= PAGE_SIZE;
1473 else /* need to call res_counter_charge */
1475 put_cpu_var(memcg_stock);
1480 * Returns stocks cached in percpu to res_counter and reset cached information.
1482 static void drain_stock(struct memcg_stock_pcp *stock)
1484 struct mem_cgroup *old = stock->cached;
1486 if (stock->charge) {
1487 res_counter_uncharge(&old->res, stock->charge);
1488 if (do_swap_account)
1489 res_counter_uncharge(&old->memsw, stock->charge);
1491 stock->cached = NULL;
1496 * This must be called under preempt disabled or must be called by
1497 * a thread which is pinned to local cpu.
1499 static void drain_local_stock(struct work_struct *dummy)
1501 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1506 * Cache charges(val) which is from res_counter, to local per_cpu area.
1507 * This will be consumed by consume_stock() function, later.
1509 static void refill_stock(struct mem_cgroup *mem, int val)
1511 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1513 if (stock->cached != mem) { /* reset if necessary */
1515 stock->cached = mem;
1517 stock->charge += val;
1518 put_cpu_var(memcg_stock);
1522 * Tries to drain stocked charges in other cpus. This function is asynchronous
1523 * and just put a work per cpu for draining localy on each cpu. Caller can
1524 * expects some charges will be back to res_counter later but cannot wait for
1527 static void drain_all_stock_async(void)
1530 /* This function is for scheduling "drain" in asynchronous way.
1531 * The result of "drain" is not directly handled by callers. Then,
1532 * if someone is calling drain, we don't have to call drain more.
1533 * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1534 * there is a race. We just do loose check here.
1536 if (atomic_read(&memcg_drain_count))
1538 /* Notify other cpus that system-wide "drain" is running */
1539 atomic_inc(&memcg_drain_count);
1541 for_each_online_cpu(cpu) {
1542 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1543 schedule_work_on(cpu, &stock->work);
1546 atomic_dec(&memcg_drain_count);
1547 /* We don't wait for flush_work */
1550 /* This is a synchronous drain interface. */
1551 static void drain_all_stock_sync(void)
1553 /* called when force_empty is called */
1554 atomic_inc(&memcg_drain_count);
1555 schedule_on_each_cpu(drain_local_stock);
1556 atomic_dec(&memcg_drain_count);
1559 static int __cpuinit memcg_stock_cpu_callback(struct notifier_block *nb,
1560 unsigned long action,
1563 int cpu = (unsigned long)hcpu;
1564 struct memcg_stock_pcp *stock;
1566 if (action != CPU_DEAD)
1568 stock = &per_cpu(memcg_stock, cpu);
1574 * Unlike exported interface, "oom" parameter is added. if oom==true,
1575 * oom-killer can be invoked.
1577 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1578 gfp_t gfp_mask, struct mem_cgroup **memcg, bool oom)
1580 struct mem_cgroup *mem, *mem_over_limit;
1581 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1582 struct res_counter *fail_res;
1583 int csize = CHARGE_SIZE;
1586 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1587 * in system level. So, allow to go ahead dying process in addition to
1590 if (unlikely(test_thread_flag(TIF_MEMDIE)
1591 || fatal_signal_pending(current)))
1595 * We always charge the cgroup the mm_struct belongs to.
1596 * The mm_struct's mem_cgroup changes on task migration if the
1597 * thread group leader migrates. It's possible that mm is not
1598 * set, if so charge the init_mm (happens for pagecache usage).
1602 mem = try_get_mem_cgroup_from_mm(mm);
1610 VM_BUG_ON(css_is_removed(&mem->css));
1611 if (mem_cgroup_is_root(mem))
1616 unsigned long flags = 0;
1618 if (consume_stock(mem))
1621 ret = res_counter_charge(&mem->res, csize, &fail_res);
1623 if (!do_swap_account)
1625 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1628 /* mem+swap counter fails */
1629 res_counter_uncharge(&mem->res, csize);
1630 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1631 mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1634 /* mem counter fails */
1635 mem_over_limit = mem_cgroup_from_res_counter(fail_res,
1638 /* reduce request size and retry */
1639 if (csize > PAGE_SIZE) {
1643 if (!(gfp_mask & __GFP_WAIT))
1646 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1652 * try_to_free_mem_cgroup_pages() might not give us a full
1653 * picture of reclaim. Some pages are reclaimed and might be
1654 * moved to swap cache or just unmapped from the cgroup.
1655 * Check the limit again to see if the reclaim reduced the
1656 * current usage of the cgroup before giving up
1659 if (mem_cgroup_check_under_limit(mem_over_limit))
1662 /* try to avoid oom while someone is moving charge */
1663 if (mc.moving_task && current != mc.moving_task) {
1664 struct mem_cgroup *from, *to;
1665 bool do_continue = false;
1667 * There is a small race that "from" or "to" can be
1668 * freed by rmdir, so we use css_tryget().
1672 if (from && css_tryget(&from->css)) {
1673 if (mem_over_limit->use_hierarchy)
1674 do_continue = css_is_ancestor(
1676 &mem_over_limit->css);
1678 do_continue = (from == mem_over_limit);
1679 css_put(&from->css);
1681 if (!do_continue && to && css_tryget(&to->css)) {
1682 if (mem_over_limit->use_hierarchy)
1683 do_continue = css_is_ancestor(
1685 &mem_over_limit->css);
1687 do_continue = (to == mem_over_limit);
1692 prepare_to_wait(&mc.waitq, &wait,
1693 TASK_INTERRUPTIBLE);
1694 /* moving charge context might have finished. */
1697 finish_wait(&mc.waitq, &wait);
1702 if (!nr_retries--) {
1705 if (mem_cgroup_handle_oom(mem_over_limit, gfp_mask)) {
1706 nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
1709 /* When we reach here, current task is dying .*/
1714 if (csize > PAGE_SIZE)
1715 refill_stock(mem, csize - PAGE_SIZE);
1727 * Somemtimes we have to undo a charge we got by try_charge().
1728 * This function is for that and do uncharge, put css's refcnt.
1729 * gotten by try_charge().
1731 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
1732 unsigned long count)
1734 if (!mem_cgroup_is_root(mem)) {
1735 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
1736 if (do_swap_account)
1737 res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
1738 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
1739 WARN_ON_ONCE(count > INT_MAX);
1740 __css_put(&mem->css, (int)count);
1742 /* we don't need css_put for root */
1745 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem)
1747 __mem_cgroup_cancel_charge(mem, 1);
1751 * A helper function to get mem_cgroup from ID. must be called under
1752 * rcu_read_lock(). The caller must check css_is_removed() or some if
1753 * it's concern. (dropping refcnt from swap can be called against removed
1756 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
1758 struct cgroup_subsys_state *css;
1760 /* ID 0 is unused ID */
1763 css = css_lookup(&mem_cgroup_subsys, id);
1766 return container_of(css, struct mem_cgroup, css);
1769 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
1771 struct mem_cgroup *mem = NULL;
1772 struct page_cgroup *pc;
1776 VM_BUG_ON(!PageLocked(page));
1778 pc = lookup_page_cgroup(page);
1779 lock_page_cgroup(pc);
1780 if (PageCgroupUsed(pc)) {
1781 mem = pc->mem_cgroup;
1782 if (mem && !css_tryget(&mem->css))
1784 } else if (PageSwapCache(page)) {
1785 ent.val = page_private(page);
1786 id = lookup_swap_cgroup(ent);
1788 mem = mem_cgroup_lookup(id);
1789 if (mem && !css_tryget(&mem->css))
1793 unlock_page_cgroup(pc);
1798 * commit a charge got by __mem_cgroup_try_charge() and makes page_cgroup to be
1799 * USED state. If already USED, uncharge and return.
1802 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
1803 struct page_cgroup *pc,
1804 enum charge_type ctype)
1806 /* try_charge() can return NULL to *memcg, taking care of it. */
1810 lock_page_cgroup(pc);
1811 if (unlikely(PageCgroupUsed(pc))) {
1812 unlock_page_cgroup(pc);
1813 mem_cgroup_cancel_charge(mem);
1817 pc->mem_cgroup = mem;
1819 * We access a page_cgroup asynchronously without lock_page_cgroup().
1820 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
1821 * is accessed after testing USED bit. To make pc->mem_cgroup visible
1822 * before USED bit, we need memory barrier here.
1823 * See mem_cgroup_add_lru_list(), etc.
1827 case MEM_CGROUP_CHARGE_TYPE_CACHE:
1828 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
1829 SetPageCgroupCache(pc);
1830 SetPageCgroupUsed(pc);
1832 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
1833 ClearPageCgroupCache(pc);
1834 SetPageCgroupUsed(pc);
1840 mem_cgroup_charge_statistics(mem, pc, true);
1842 unlock_page_cgroup(pc);
1844 * "charge_statistics" updated event counter. Then, check it.
1845 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
1846 * if they exceeds softlimit.
1848 memcg_check_events(mem, pc->page);
1852 * __mem_cgroup_move_account - move account of the page
1853 * @pc: page_cgroup of the page.
1854 * @from: mem_cgroup which the page is moved from.
1855 * @to: mem_cgroup which the page is moved to. @from != @to.
1856 * @uncharge: whether we should call uncharge and css_put against @from.
1858 * The caller must confirm following.
1859 * - page is not on LRU (isolate_page() is useful.)
1860 * - the pc is locked, used, and ->mem_cgroup points to @from.
1862 * This function doesn't do "charge" nor css_get to new cgroup. It should be
1863 * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
1864 * true, this function does "uncharge" from old cgroup, but it doesn't if
1865 * @uncharge is false, so a caller should do "uncharge".
1868 static void __mem_cgroup_move_account(struct page_cgroup *pc,
1869 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1871 VM_BUG_ON(from == to);
1872 VM_BUG_ON(PageLRU(pc->page));
1873 VM_BUG_ON(!PageCgroupLocked(pc));
1874 VM_BUG_ON(!PageCgroupUsed(pc));
1875 VM_BUG_ON(pc->mem_cgroup != from);
1877 if (PageCgroupFileMapped(pc)) {
1878 /* Update mapped_file data for mem_cgroup */
1880 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1881 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
1884 mem_cgroup_charge_statistics(from, pc, false);
1886 /* This is not "cancel", but cancel_charge does all we need. */
1887 mem_cgroup_cancel_charge(from);
1889 /* caller should have done css_get */
1890 pc->mem_cgroup = to;
1891 mem_cgroup_charge_statistics(to, pc, true);
1893 * We charges against "to" which may not have any tasks. Then, "to"
1894 * can be under rmdir(). But in current implementation, caller of
1895 * this function is just force_empty() and move charge, so it's
1896 * garanteed that "to" is never removed. So, we don't check rmdir
1902 * check whether the @pc is valid for moving account and call
1903 * __mem_cgroup_move_account()
1905 static int mem_cgroup_move_account(struct page_cgroup *pc,
1906 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge)
1909 lock_page_cgroup(pc);
1910 if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
1911 __mem_cgroup_move_account(pc, from, to, uncharge);
1914 unlock_page_cgroup(pc);
1918 memcg_check_events(to, pc->page);
1919 memcg_check_events(from, pc->page);
1924 * move charges to its parent.
1927 static int mem_cgroup_move_parent(struct page_cgroup *pc,
1928 struct mem_cgroup *child,
1931 struct page *page = pc->page;
1932 struct cgroup *cg = child->css.cgroup;
1933 struct cgroup *pcg = cg->parent;
1934 struct mem_cgroup *parent;
1942 if (!get_page_unless_zero(page))
1944 if (isolate_lru_page(page))
1947 parent = mem_cgroup_from_cont(pcg);
1948 ret = __mem_cgroup_try_charge(NULL, gfp_mask, &parent, false);
1952 ret = mem_cgroup_move_account(pc, child, parent, true);
1954 mem_cgroup_cancel_charge(parent);
1956 putback_lru_page(page);
1964 * Charge the memory controller for page usage.
1966 * 0 if the charge was successful
1967 * < 0 if the cgroup is over its limit
1969 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
1970 gfp_t gfp_mask, enum charge_type ctype,
1971 struct mem_cgroup *memcg)
1973 struct mem_cgroup *mem;
1974 struct page_cgroup *pc;
1977 pc = lookup_page_cgroup(page);
1978 /* can happen at boot */
1984 ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, true);
1988 __mem_cgroup_commit_charge(mem, pc, ctype);
1992 int mem_cgroup_newpage_charge(struct page *page,
1993 struct mm_struct *mm, gfp_t gfp_mask)
1995 if (mem_cgroup_disabled())
1997 if (PageCompound(page))
2000 * If already mapped, we don't have to account.
2001 * If page cache, page->mapping has address_space.
2002 * But page->mapping may have out-of-use anon_vma pointer,
2003 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2006 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2010 return mem_cgroup_charge_common(page, mm, gfp_mask,
2011 MEM_CGROUP_CHARGE_TYPE_MAPPED, NULL);
2015 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2016 enum charge_type ctype);
2018 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2021 struct mem_cgroup *mem = NULL;
2024 if (mem_cgroup_disabled())
2026 if (PageCompound(page))
2029 * Corner case handling. This is called from add_to_page_cache()
2030 * in usual. But some FS (shmem) precharges this page before calling it
2031 * and call add_to_page_cache() with GFP_NOWAIT.
2033 * For GFP_NOWAIT case, the page may be pre-charged before calling
2034 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2035 * charge twice. (It works but has to pay a bit larger cost.)
2036 * And when the page is SwapCache, it should take swap information
2037 * into account. This is under lock_page() now.
2039 if (!(gfp_mask & __GFP_WAIT)) {
2040 struct page_cgroup *pc;
2043 pc = lookup_page_cgroup(page);
2046 lock_page_cgroup(pc);
2047 if (PageCgroupUsed(pc)) {
2048 unlock_page_cgroup(pc);
2051 unlock_page_cgroup(pc);
2054 if (unlikely(!mm && !mem))
2057 if (page_is_file_cache(page))
2058 return mem_cgroup_charge_common(page, mm, gfp_mask,
2059 MEM_CGROUP_CHARGE_TYPE_CACHE, NULL);
2062 if (PageSwapCache(page)) {
2063 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2065 __mem_cgroup_commit_charge_swapin(page, mem,
2066 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2068 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2069 MEM_CGROUP_CHARGE_TYPE_SHMEM, mem);
2075 * While swap-in, try_charge -> commit or cancel, the page is locked.
2076 * And when try_charge() successfully returns, one refcnt to memcg without
2077 * struct page_cgroup is acquired. This refcnt will be consumed by
2078 * "commit()" or removed by "cancel()"
2080 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2082 gfp_t mask, struct mem_cgroup **ptr)
2084 struct mem_cgroup *mem;
2087 if (mem_cgroup_disabled())
2090 if (!do_swap_account)
2093 * A racing thread's fault, or swapoff, may have already updated
2094 * the pte, and even removed page from swap cache: in those cases
2095 * do_swap_page()'s pte_same() test will fail; but there's also a
2096 * KSM case which does need to charge the page.
2098 if (!PageSwapCache(page))
2100 mem = try_get_mem_cgroup_from_page(page);
2104 ret = __mem_cgroup_try_charge(NULL, mask, ptr, true);
2105 /* drop extra refcnt from tryget */
2111 return __mem_cgroup_try_charge(mm, mask, ptr, true);
2115 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2116 enum charge_type ctype)
2118 struct page_cgroup *pc;
2120 if (mem_cgroup_disabled())
2124 cgroup_exclude_rmdir(&ptr->css);
2125 pc = lookup_page_cgroup(page);
2126 mem_cgroup_lru_del_before_commit_swapcache(page);
2127 __mem_cgroup_commit_charge(ptr, pc, ctype);
2128 mem_cgroup_lru_add_after_commit_swapcache(page);
2130 * Now swap is on-memory. This means this page may be
2131 * counted both as mem and swap....double count.
2132 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2133 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2134 * may call delete_from_swap_cache() before reach here.
2136 if (do_swap_account && PageSwapCache(page)) {
2137 swp_entry_t ent = {.val = page_private(page)};
2139 struct mem_cgroup *memcg;
2141 id = swap_cgroup_record(ent, 0);
2143 memcg = mem_cgroup_lookup(id);
2146 * This recorded memcg can be obsolete one. So, avoid
2147 * calling css_tryget
2149 if (!mem_cgroup_is_root(memcg))
2150 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2151 mem_cgroup_swap_statistics(memcg, false);
2152 mem_cgroup_put(memcg);
2157 * At swapin, we may charge account against cgroup which has no tasks.
2158 * So, rmdir()->pre_destroy() can be called while we do this charge.
2159 * In that case, we need to call pre_destroy() again. check it here.
2161 cgroup_release_and_wakeup_rmdir(&ptr->css);
2164 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2166 __mem_cgroup_commit_charge_swapin(page, ptr,
2167 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2170 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2172 if (mem_cgroup_disabled())
2176 mem_cgroup_cancel_charge(mem);
2180 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype)
2182 struct memcg_batch_info *batch = NULL;
2183 bool uncharge_memsw = true;
2184 /* If swapout, usage of swap doesn't decrease */
2185 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2186 uncharge_memsw = false;
2188 batch = ¤t->memcg_batch;
2190 * In usual, we do css_get() when we remember memcg pointer.
2191 * But in this case, we keep res->usage until end of a series of
2192 * uncharges. Then, it's ok to ignore memcg's refcnt.
2197 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2198 * In those cases, all pages freed continously can be expected to be in
2199 * the same cgroup and we have chance to coalesce uncharges.
2200 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2201 * because we want to do uncharge as soon as possible.
2204 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2205 goto direct_uncharge;
2208 * In typical case, batch->memcg == mem. This means we can
2209 * merge a series of uncharges to an uncharge of res_counter.
2210 * If not, we uncharge res_counter ony by one.
2212 if (batch->memcg != mem)
2213 goto direct_uncharge;
2214 /* remember freed charge and uncharge it later */
2215 batch->bytes += PAGE_SIZE;
2217 batch->memsw_bytes += PAGE_SIZE;
2220 res_counter_uncharge(&mem->res, PAGE_SIZE);
2222 res_counter_uncharge(&mem->memsw, PAGE_SIZE);
2223 if (unlikely(batch->memcg != mem))
2224 memcg_oom_recover(mem);
2229 * uncharge if !page_mapped(page)
2231 static struct mem_cgroup *
2232 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2234 struct page_cgroup *pc;
2235 struct mem_cgroup *mem = NULL;
2236 struct mem_cgroup_per_zone *mz;
2238 if (mem_cgroup_disabled())
2241 if (PageSwapCache(page))
2245 * Check if our page_cgroup is valid
2247 pc = lookup_page_cgroup(page);
2248 if (unlikely(!pc || !PageCgroupUsed(pc)))
2251 lock_page_cgroup(pc);
2253 mem = pc->mem_cgroup;
2255 if (!PageCgroupUsed(pc))
2259 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2260 case MEM_CGROUP_CHARGE_TYPE_DROP:
2261 if (page_mapped(page))
2264 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2265 if (!PageAnon(page)) { /* Shared memory */
2266 if (page->mapping && !page_is_file_cache(page))
2268 } else if (page_mapped(page)) /* Anon */
2275 if (!mem_cgroup_is_root(mem))
2276 __do_uncharge(mem, ctype);
2277 if (ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2278 mem_cgroup_swap_statistics(mem, true);
2279 mem_cgroup_charge_statistics(mem, pc, false);
2281 ClearPageCgroupUsed(pc);
2283 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2284 * freed from LRU. This is safe because uncharged page is expected not
2285 * to be reused (freed soon). Exception is SwapCache, it's handled by
2286 * special functions.
2289 mz = page_cgroup_zoneinfo(pc);
2290 unlock_page_cgroup(pc);
2292 memcg_check_events(mem, page);
2293 /* at swapout, this memcg will be accessed to record to swap */
2294 if (ctype != MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2300 unlock_page_cgroup(pc);
2304 void mem_cgroup_uncharge_page(struct page *page)
2307 if (page_mapped(page))
2309 if (page->mapping && !PageAnon(page))
2311 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2314 void mem_cgroup_uncharge_cache_page(struct page *page)
2316 VM_BUG_ON(page_mapped(page));
2317 VM_BUG_ON(page->mapping);
2318 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2322 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2323 * In that cases, pages are freed continuously and we can expect pages
2324 * are in the same memcg. All these calls itself limits the number of
2325 * pages freed at once, then uncharge_start/end() is called properly.
2326 * This may be called prural(2) times in a context,
2329 void mem_cgroup_uncharge_start(void)
2331 current->memcg_batch.do_batch++;
2332 /* We can do nest. */
2333 if (current->memcg_batch.do_batch == 1) {
2334 current->memcg_batch.memcg = NULL;
2335 current->memcg_batch.bytes = 0;
2336 current->memcg_batch.memsw_bytes = 0;
2340 void mem_cgroup_uncharge_end(void)
2342 struct memcg_batch_info *batch = ¤t->memcg_batch;
2344 if (!batch->do_batch)
2348 if (batch->do_batch) /* If stacked, do nothing. */
2354 * This "batch->memcg" is valid without any css_get/put etc...
2355 * bacause we hide charges behind us.
2358 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2359 if (batch->memsw_bytes)
2360 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2361 memcg_oom_recover(batch->memcg);
2362 /* forget this pointer (for sanity check) */
2363 batch->memcg = NULL;
2368 * called after __delete_from_swap_cache() and drop "page" account.
2369 * memcg information is recorded to swap_cgroup of "ent"
2372 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2374 struct mem_cgroup *memcg;
2375 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2377 if (!swapout) /* this was a swap cache but the swap is unused ! */
2378 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2380 memcg = __mem_cgroup_uncharge_common(page, ctype);
2382 /* record memcg information */
2383 if (do_swap_account && swapout && memcg) {
2384 swap_cgroup_record(ent, css_id(&memcg->css));
2385 mem_cgroup_get(memcg);
2387 if (swapout && memcg)
2388 css_put(&memcg->css);
2392 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2394 * called from swap_entry_free(). remove record in swap_cgroup and
2395 * uncharge "memsw" account.
2397 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2399 struct mem_cgroup *memcg;
2402 if (!do_swap_account)
2405 id = swap_cgroup_record(ent, 0);
2407 memcg = mem_cgroup_lookup(id);
2410 * We uncharge this because swap is freed.
2411 * This memcg can be obsolete one. We avoid calling css_tryget
2413 if (!mem_cgroup_is_root(memcg))
2414 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2415 mem_cgroup_swap_statistics(memcg, false);
2416 mem_cgroup_put(memcg);
2422 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2423 * @entry: swap entry to be moved
2424 * @from: mem_cgroup which the entry is moved from
2425 * @to: mem_cgroup which the entry is moved to
2426 * @need_fixup: whether we should fixup res_counters and refcounts.
2428 * It succeeds only when the swap_cgroup's record for this entry is the same
2429 * as the mem_cgroup's id of @from.
2431 * Returns 0 on success, -EINVAL on failure.
2433 * The caller must have charged to @to, IOW, called res_counter_charge() about
2434 * both res and memsw, and called css_get().
2436 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2437 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2439 unsigned short old_id, new_id;
2441 old_id = css_id(&from->css);
2442 new_id = css_id(&to->css);
2444 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2445 mem_cgroup_swap_statistics(from, false);
2446 mem_cgroup_swap_statistics(to, true);
2448 * This function is only called from task migration context now.
2449 * It postpones res_counter and refcount handling till the end
2450 * of task migration(mem_cgroup_clear_mc()) for performance
2451 * improvement. But we cannot postpone mem_cgroup_get(to)
2452 * because if the process that has been moved to @to does
2453 * swap-in, the refcount of @to might be decreased to 0.
2457 if (!mem_cgroup_is_root(from))
2458 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2459 mem_cgroup_put(from);
2461 * we charged both to->res and to->memsw, so we should
2464 if (!mem_cgroup_is_root(to))
2465 res_counter_uncharge(&to->res, PAGE_SIZE);
2473 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2474 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2481 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2484 int mem_cgroup_prepare_migration(struct page *page, struct mem_cgroup **ptr)
2486 struct page_cgroup *pc;
2487 struct mem_cgroup *mem = NULL;
2490 if (mem_cgroup_disabled())
2493 pc = lookup_page_cgroup(page);
2494 lock_page_cgroup(pc);
2495 if (PageCgroupUsed(pc)) {
2496 mem = pc->mem_cgroup;
2499 unlock_page_cgroup(pc);
2503 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, ptr, false);
2509 /* remove redundant charge if migration failed*/
2510 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2511 struct page *oldpage, struct page *newpage)
2513 struct page *target, *unused;
2514 struct page_cgroup *pc;
2515 enum charge_type ctype;
2519 cgroup_exclude_rmdir(&mem->css);
2520 /* at migration success, oldpage->mapping is NULL. */
2521 if (oldpage->mapping) {
2529 if (PageAnon(target))
2530 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2531 else if (page_is_file_cache(target))
2532 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2534 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2536 /* unused page is not on radix-tree now. */
2538 __mem_cgroup_uncharge_common(unused, ctype);
2540 pc = lookup_page_cgroup(target);
2542 * __mem_cgroup_commit_charge() check PCG_USED bit of page_cgroup.
2543 * So, double-counting is effectively avoided.
2545 __mem_cgroup_commit_charge(mem, pc, ctype);
2548 * Both of oldpage and newpage are still under lock_page().
2549 * Then, we don't have to care about race in radix-tree.
2550 * But we have to be careful that this page is unmapped or not.
2552 * There is a case for !page_mapped(). At the start of
2553 * migration, oldpage was mapped. But now, it's zapped.
2554 * But we know *target* page is not freed/reused under us.
2555 * mem_cgroup_uncharge_page() does all necessary checks.
2557 if (ctype == MEM_CGROUP_CHARGE_TYPE_MAPPED)
2558 mem_cgroup_uncharge_page(target);
2560 * At migration, we may charge account against cgroup which has no tasks
2561 * So, rmdir()->pre_destroy() can be called while we do this charge.
2562 * In that case, we need to call pre_destroy() again. check it here.
2564 cgroup_release_and_wakeup_rmdir(&mem->css);
2568 * A call to try to shrink memory usage on charge failure at shmem's swapin.
2569 * Calling hierarchical_reclaim is not enough because we should update
2570 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
2571 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
2572 * not from the memcg which this page would be charged to.
2573 * try_charge_swapin does all of these works properly.
2575 int mem_cgroup_shmem_charge_fallback(struct page *page,
2576 struct mm_struct *mm,
2579 struct mem_cgroup *mem = NULL;
2582 if (mem_cgroup_disabled())
2585 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2587 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
2592 static DEFINE_MUTEX(set_limit_mutex);
2594 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
2595 unsigned long long val)
2598 u64 memswlimit, memlimit;
2600 int children = mem_cgroup_count_children(memcg);
2601 u64 curusage, oldusage;
2605 * For keeping hierarchical_reclaim simple, how long we should retry
2606 * is depends on callers. We set our retry-count to be function
2607 * of # of children which we should visit in this loop.
2609 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
2611 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2614 while (retry_count) {
2615 if (signal_pending(current)) {
2620 * Rather than hide all in some function, I do this in
2621 * open coded manner. You see what this really does.
2622 * We have to guarantee mem->res.limit < mem->memsw.limit.
2624 mutex_lock(&set_limit_mutex);
2625 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2626 if (memswlimit < val) {
2628 mutex_unlock(&set_limit_mutex);
2632 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2636 ret = res_counter_set_limit(&memcg->res, val);
2638 if (memswlimit == val)
2639 memcg->memsw_is_minimum = true;
2641 memcg->memsw_is_minimum = false;
2643 mutex_unlock(&set_limit_mutex);
2648 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2649 MEM_CGROUP_RECLAIM_SHRINK);
2650 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
2651 /* Usage is reduced ? */
2652 if (curusage >= oldusage)
2655 oldusage = curusage;
2657 if (!ret && enlarge)
2658 memcg_oom_recover(memcg);
2663 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
2664 unsigned long long val)
2667 u64 memlimit, memswlimit, oldusage, curusage;
2668 int children = mem_cgroup_count_children(memcg);
2672 /* see mem_cgroup_resize_res_limit */
2673 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
2674 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2675 while (retry_count) {
2676 if (signal_pending(current)) {
2681 * Rather than hide all in some function, I do this in
2682 * open coded manner. You see what this really does.
2683 * We have to guarantee mem->res.limit < mem->memsw.limit.
2685 mutex_lock(&set_limit_mutex);
2686 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
2687 if (memlimit > val) {
2689 mutex_unlock(&set_limit_mutex);
2692 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
2693 if (memswlimit < val)
2695 ret = res_counter_set_limit(&memcg->memsw, val);
2697 if (memlimit == val)
2698 memcg->memsw_is_minimum = true;
2700 memcg->memsw_is_minimum = false;
2702 mutex_unlock(&set_limit_mutex);
2707 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
2708 MEM_CGROUP_RECLAIM_NOSWAP |
2709 MEM_CGROUP_RECLAIM_SHRINK);
2710 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
2711 /* Usage is reduced ? */
2712 if (curusage >= oldusage)
2715 oldusage = curusage;
2717 if (!ret && enlarge)
2718 memcg_oom_recover(memcg);
2722 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
2723 gfp_t gfp_mask, int nid,
2726 unsigned long nr_reclaimed = 0;
2727 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
2728 unsigned long reclaimed;
2730 struct mem_cgroup_tree_per_zone *mctz;
2731 unsigned long long excess;
2736 mctz = soft_limit_tree_node_zone(nid, zid);
2738 * This loop can run a while, specially if mem_cgroup's continuously
2739 * keep exceeding their soft limit and putting the system under
2746 mz = mem_cgroup_largest_soft_limit_node(mctz);
2750 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
2752 MEM_CGROUP_RECLAIM_SOFT);
2753 nr_reclaimed += reclaimed;
2754 spin_lock(&mctz->lock);
2757 * If we failed to reclaim anything from this memory cgroup
2758 * it is time to move on to the next cgroup
2764 * Loop until we find yet another one.
2766 * By the time we get the soft_limit lock
2767 * again, someone might have aded the
2768 * group back on the RB tree. Iterate to
2769 * make sure we get a different mem.
2770 * mem_cgroup_largest_soft_limit_node returns
2771 * NULL if no other cgroup is present on
2775 __mem_cgroup_largest_soft_limit_node(mctz);
2776 if (next_mz == mz) {
2777 css_put(&next_mz->mem->css);
2779 } else /* next_mz == NULL or other memcg */
2783 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
2784 excess = res_counter_soft_limit_excess(&mz->mem->res);
2786 * One school of thought says that we should not add
2787 * back the node to the tree if reclaim returns 0.
2788 * But our reclaim could return 0, simply because due
2789 * to priority we are exposing a smaller subset of
2790 * memory to reclaim from. Consider this as a longer
2793 /* If excess == 0, no tree ops */
2794 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
2795 spin_unlock(&mctz->lock);
2796 css_put(&mz->mem->css);
2799 * Could not reclaim anything and there are no more
2800 * mem cgroups to try or we seem to be looping without
2801 * reclaiming anything.
2803 if (!nr_reclaimed &&
2805 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
2807 } while (!nr_reclaimed);
2809 css_put(&next_mz->mem->css);
2810 return nr_reclaimed;
2814 * This routine traverse page_cgroup in given list and drop them all.
2815 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
2817 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
2818 int node, int zid, enum lru_list lru)
2821 struct mem_cgroup_per_zone *mz;
2822 struct page_cgroup *pc, *busy;
2823 unsigned long flags, loop;
2824 struct list_head *list;
2827 zone = &NODE_DATA(node)->node_zones[zid];
2828 mz = mem_cgroup_zoneinfo(mem, node, zid);
2829 list = &mz->lists[lru];
2831 loop = MEM_CGROUP_ZSTAT(mz, lru);
2832 /* give some margin against EBUSY etc...*/
2837 spin_lock_irqsave(&zone->lru_lock, flags);
2838 if (list_empty(list)) {
2839 spin_unlock_irqrestore(&zone->lru_lock, flags);
2842 pc = list_entry(list->prev, struct page_cgroup, lru);
2844 list_move(&pc->lru, list);
2846 spin_unlock_irqrestore(&zone->lru_lock, flags);
2849 spin_unlock_irqrestore(&zone->lru_lock, flags);
2851 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
2855 if (ret == -EBUSY || ret == -EINVAL) {
2856 /* found lock contention or "pc" is obsolete. */
2863 if (!ret && !list_empty(list))
2869 * make mem_cgroup's charge to be 0 if there is no task.
2870 * This enables deleting this mem_cgroup.
2872 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
2875 int node, zid, shrink;
2876 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
2877 struct cgroup *cgrp = mem->css.cgroup;
2882 /* should free all ? */
2888 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
2891 if (signal_pending(current))
2893 /* This is for making all *used* pages to be on LRU. */
2894 lru_add_drain_all();
2895 drain_all_stock_sync();
2897 for_each_node_state(node, N_HIGH_MEMORY) {
2898 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
2901 ret = mem_cgroup_force_empty_list(mem,
2910 memcg_oom_recover(mem);
2911 /* it seems parent cgroup doesn't have enough mem */
2915 /* "ret" should also be checked to ensure all lists are empty. */
2916 } while (mem->res.usage > 0 || ret);
2922 /* returns EBUSY if there is a task or if we come here twice. */
2923 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
2927 /* we call try-to-free pages for make this cgroup empty */
2928 lru_add_drain_all();
2929 /* try to free all pages in this cgroup */
2931 while (nr_retries && mem->res.usage > 0) {
2934 if (signal_pending(current)) {
2938 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
2939 false, get_swappiness(mem));
2942 /* maybe some writeback is necessary */
2943 congestion_wait(BLK_RW_ASYNC, HZ/10);
2948 /* try move_account...there may be some *locked* pages. */
2952 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
2954 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
2958 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
2960 return mem_cgroup_from_cont(cont)->use_hierarchy;
2963 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
2967 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
2968 struct cgroup *parent = cont->parent;
2969 struct mem_cgroup *parent_mem = NULL;
2972 parent_mem = mem_cgroup_from_cont(parent);
2976 * If parent's use_hierarchy is set, we can't make any modifications
2977 * in the child subtrees. If it is unset, then the change can
2978 * occur, provided the current cgroup has no children.
2980 * For the root cgroup, parent_mem is NULL, we allow value to be
2981 * set if there are no children.
2983 if ((!parent_mem || !parent_mem->use_hierarchy) &&
2984 (val == 1 || val == 0)) {
2985 if (list_empty(&cont->children))
2986 mem->use_hierarchy = val;
2996 struct mem_cgroup_idx_data {
2998 enum mem_cgroup_stat_index idx;
3002 mem_cgroup_get_idx_stat(struct mem_cgroup *mem, void *data)
3004 struct mem_cgroup_idx_data *d = data;
3005 d->val += mem_cgroup_read_stat(mem, d->idx);
3010 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3011 enum mem_cgroup_stat_index idx, s64 *val)
3013 struct mem_cgroup_idx_data d;
3016 mem_cgroup_walk_tree(mem, &d, mem_cgroup_get_idx_stat);
3020 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3024 if (!mem_cgroup_is_root(mem)) {
3026 return res_counter_read_u64(&mem->res, RES_USAGE);
3028 return res_counter_read_u64(&mem->memsw, RES_USAGE);
3031 mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE, &idx_val);
3033 mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS, &idx_val);
3037 mem_cgroup_get_recursive_idx_stat(mem,
3038 MEM_CGROUP_STAT_SWAPOUT, &idx_val);
3042 return val << PAGE_SHIFT;
3045 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3047 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3051 type = MEMFILE_TYPE(cft->private);
3052 name = MEMFILE_ATTR(cft->private);
3055 if (name == RES_USAGE)
3056 val = mem_cgroup_usage(mem, false);
3058 val = res_counter_read_u64(&mem->res, name);
3061 if (name == RES_USAGE)
3062 val = mem_cgroup_usage(mem, true);
3064 val = res_counter_read_u64(&mem->memsw, name);
3073 * The user of this function is...
3076 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3079 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3081 unsigned long long val;
3084 type = MEMFILE_TYPE(cft->private);
3085 name = MEMFILE_ATTR(cft->private);
3088 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3092 /* This function does all necessary parse...reuse it */
3093 ret = res_counter_memparse_write_strategy(buffer, &val);
3097 ret = mem_cgroup_resize_limit(memcg, val);
3099 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3101 case RES_SOFT_LIMIT:
3102 ret = res_counter_memparse_write_strategy(buffer, &val);
3106 * For memsw, soft limits are hard to implement in terms
3107 * of semantics, for now, we support soft limits for
3108 * control without swap
3111 ret = res_counter_set_soft_limit(&memcg->res, val);
3116 ret = -EINVAL; /* should be BUG() ? */
3122 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3123 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3125 struct cgroup *cgroup;
3126 unsigned long long min_limit, min_memsw_limit, tmp;
3128 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3129 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3130 cgroup = memcg->css.cgroup;
3131 if (!memcg->use_hierarchy)
3134 while (cgroup->parent) {
3135 cgroup = cgroup->parent;
3136 memcg = mem_cgroup_from_cont(cgroup);
3137 if (!memcg->use_hierarchy)
3139 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3140 min_limit = min(min_limit, tmp);
3141 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3142 min_memsw_limit = min(min_memsw_limit, tmp);
3145 *mem_limit = min_limit;
3146 *memsw_limit = min_memsw_limit;
3150 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3152 struct mem_cgroup *mem;
3155 mem = mem_cgroup_from_cont(cont);
3156 type = MEMFILE_TYPE(event);
3157 name = MEMFILE_ATTR(event);
3161 res_counter_reset_max(&mem->res);
3163 res_counter_reset_max(&mem->memsw);
3167 res_counter_reset_failcnt(&mem->res);
3169 res_counter_reset_failcnt(&mem->memsw);
3176 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3179 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3183 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3184 struct cftype *cft, u64 val)
3186 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3188 if (val >= (1 << NR_MOVE_TYPE))
3191 * We check this value several times in both in can_attach() and
3192 * attach(), so we need cgroup lock to prevent this value from being
3196 mem->move_charge_at_immigrate = val;
3202 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3203 struct cftype *cft, u64 val)
3210 /* For read statistics */
3226 struct mcs_total_stat {
3227 s64 stat[NR_MCS_STAT];
3233 } memcg_stat_strings[NR_MCS_STAT] = {
3234 {"cache", "total_cache"},
3235 {"rss", "total_rss"},
3236 {"mapped_file", "total_mapped_file"},
3237 {"pgpgin", "total_pgpgin"},
3238 {"pgpgout", "total_pgpgout"},
3239 {"swap", "total_swap"},
3240 {"inactive_anon", "total_inactive_anon"},
3241 {"active_anon", "total_active_anon"},
3242 {"inactive_file", "total_inactive_file"},
3243 {"active_file", "total_active_file"},
3244 {"unevictable", "total_unevictable"}
3248 static int mem_cgroup_get_local_stat(struct mem_cgroup *mem, void *data)
3250 struct mcs_total_stat *s = data;
3254 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3255 s->stat[MCS_CACHE] += val * PAGE_SIZE;
3256 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3257 s->stat[MCS_RSS] += val * PAGE_SIZE;
3258 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3259 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3260 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3261 s->stat[MCS_PGPGIN] += val;
3262 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3263 s->stat[MCS_PGPGOUT] += val;
3264 if (do_swap_account) {
3265 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3266 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3270 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3271 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3272 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3273 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3274 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3275 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3276 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3277 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3278 val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3279 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3284 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3286 mem_cgroup_walk_tree(mem, s, mem_cgroup_get_local_stat);
3289 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3290 struct cgroup_map_cb *cb)
3292 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3293 struct mcs_total_stat mystat;
3296 memset(&mystat, 0, sizeof(mystat));
3297 mem_cgroup_get_local_stat(mem_cont, &mystat);
3299 for (i = 0; i < NR_MCS_STAT; i++) {
3300 if (i == MCS_SWAP && !do_swap_account)
3302 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3305 /* Hierarchical information */
3307 unsigned long long limit, memsw_limit;
3308 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3309 cb->fill(cb, "hierarchical_memory_limit", limit);
3310 if (do_swap_account)
3311 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3314 memset(&mystat, 0, sizeof(mystat));
3315 mem_cgroup_get_total_stat(mem_cont, &mystat);
3316 for (i = 0; i < NR_MCS_STAT; i++) {
3317 if (i == MCS_SWAP && !do_swap_account)
3319 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3322 #ifdef CONFIG_DEBUG_VM
3323 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3327 struct mem_cgroup_per_zone *mz;
3328 unsigned long recent_rotated[2] = {0, 0};
3329 unsigned long recent_scanned[2] = {0, 0};
3331 for_each_online_node(nid)
3332 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3333 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3335 recent_rotated[0] +=
3336 mz->reclaim_stat.recent_rotated[0];
3337 recent_rotated[1] +=
3338 mz->reclaim_stat.recent_rotated[1];
3339 recent_scanned[0] +=
3340 mz->reclaim_stat.recent_scanned[0];
3341 recent_scanned[1] +=
3342 mz->reclaim_stat.recent_scanned[1];
3344 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3345 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3346 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3347 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3354 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3356 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3358 return get_swappiness(memcg);
3361 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3364 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3365 struct mem_cgroup *parent;
3370 if (cgrp->parent == NULL)
3373 parent = mem_cgroup_from_cont(cgrp->parent);
3377 /* If under hierarchy, only empty-root can set this value */
3378 if ((parent->use_hierarchy) ||
3379 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3384 spin_lock(&memcg->reclaim_param_lock);
3385 memcg->swappiness = val;
3386 spin_unlock(&memcg->reclaim_param_lock);
3393 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3395 struct mem_cgroup_threshold_ary *t;
3401 t = rcu_dereference(memcg->thresholds);
3403 t = rcu_dereference(memcg->memsw_thresholds);
3408 usage = mem_cgroup_usage(memcg, swap);
3411 * current_threshold points to threshold just below usage.
3412 * If it's not true, a threshold was crossed after last
3413 * call of __mem_cgroup_threshold().
3415 i = t->current_threshold;
3418 * Iterate backward over array of thresholds starting from
3419 * current_threshold and check if a threshold is crossed.
3420 * If none of thresholds below usage is crossed, we read
3421 * only one element of the array here.
3423 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3424 eventfd_signal(t->entries[i].eventfd, 1);
3426 /* i = current_threshold + 1 */
3430 * Iterate forward over array of thresholds starting from
3431 * current_threshold+1 and check if a threshold is crossed.
3432 * If none of thresholds above usage is crossed, we read
3433 * only one element of the array here.
3435 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3436 eventfd_signal(t->entries[i].eventfd, 1);
3438 /* Update current_threshold */
3439 t->current_threshold = i - 1;
3444 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3446 __mem_cgroup_threshold(memcg, false);
3447 if (do_swap_account)
3448 __mem_cgroup_threshold(memcg, true);
3451 static int compare_thresholds(const void *a, const void *b)
3453 const struct mem_cgroup_threshold *_a = a;
3454 const struct mem_cgroup_threshold *_b = b;
3456 return _a->threshold - _b->threshold;
3459 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem, void *data)
3461 struct mem_cgroup_eventfd_list *ev;
3463 list_for_each_entry(ev, &mem->oom_notify, list)
3464 eventfd_signal(ev->eventfd, 1);
3468 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3470 mem_cgroup_walk_tree(mem, NULL, mem_cgroup_oom_notify_cb);
3473 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3474 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3476 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3477 struct mem_cgroup_threshold_ary *thresholds, *thresholds_new;
3478 int type = MEMFILE_TYPE(cft->private);
3479 u64 threshold, usage;
3483 ret = res_counter_memparse_write_strategy(args, &threshold);
3487 mutex_lock(&memcg->thresholds_lock);
3489 thresholds = memcg->thresholds;
3490 else if (type == _MEMSWAP)
3491 thresholds = memcg->memsw_thresholds;
3495 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3497 /* Check if a threshold crossed before adding a new one */
3499 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3502 size = thresholds->size + 1;
3506 /* Allocate memory for new array of thresholds */
3507 thresholds_new = kmalloc(sizeof(*thresholds_new) +
3508 size * sizeof(struct mem_cgroup_threshold),
3510 if (!thresholds_new) {
3514 thresholds_new->size = size;
3516 /* Copy thresholds (if any) to new array */
3518 memcpy(thresholds_new->entries, thresholds->entries,
3520 sizeof(struct mem_cgroup_threshold));
3521 /* Add new threshold */
3522 thresholds_new->entries[size - 1].eventfd = eventfd;
3523 thresholds_new->entries[size - 1].threshold = threshold;
3525 /* Sort thresholds. Registering of new threshold isn't time-critical */
3526 sort(thresholds_new->entries, size,
3527 sizeof(struct mem_cgroup_threshold),
3528 compare_thresholds, NULL);
3530 /* Find current threshold */
3531 thresholds_new->current_threshold = -1;
3532 for (i = 0; i < size; i++) {
3533 if (thresholds_new->entries[i].threshold < usage) {
3535 * thresholds_new->current_threshold will not be used
3536 * until rcu_assign_pointer(), so it's safe to increment
3539 ++thresholds_new->current_threshold;
3544 rcu_assign_pointer(memcg->thresholds, thresholds_new);
3546 rcu_assign_pointer(memcg->memsw_thresholds, thresholds_new);
3548 /* To be sure that nobody uses thresholds before freeing it */
3553 mutex_unlock(&memcg->thresholds_lock);
3558 static int mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
3559 struct cftype *cft, struct eventfd_ctx *eventfd)
3561 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3562 struct mem_cgroup_threshold_ary *thresholds, *thresholds_new;
3563 int type = MEMFILE_TYPE(cft->private);
3568 mutex_lock(&memcg->thresholds_lock);
3570 thresholds = memcg->thresholds;
3571 else if (type == _MEMSWAP)
3572 thresholds = memcg->memsw_thresholds;
3577 * Something went wrong if we trying to unregister a threshold
3578 * if we don't have thresholds
3580 BUG_ON(!thresholds);
3582 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3584 /* Check if a threshold crossed before removing */
3585 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
3587 /* Calculate new number of threshold */
3588 for (i = 0; i < thresholds->size; i++) {
3589 if (thresholds->entries[i].eventfd != eventfd)
3593 /* Set thresholds array to NULL if we don't have thresholds */
3595 thresholds_new = NULL;
3599 /* Allocate memory for new array of thresholds */
3600 thresholds_new = kmalloc(sizeof(*thresholds_new) +
3601 size * sizeof(struct mem_cgroup_threshold),
3603 if (!thresholds_new) {
3607 thresholds_new->size = size;
3609 /* Copy thresholds and find current threshold */
3610 thresholds_new->current_threshold = -1;
3611 for (i = 0, j = 0; i < thresholds->size; i++) {
3612 if (thresholds->entries[i].eventfd == eventfd)
3615 thresholds_new->entries[j] = thresholds->entries[i];
3616 if (thresholds_new->entries[j].threshold < usage) {
3618 * thresholds_new->current_threshold will not be used
3619 * until rcu_assign_pointer(), so it's safe to increment
3622 ++thresholds_new->current_threshold;
3629 rcu_assign_pointer(memcg->thresholds, thresholds_new);
3631 rcu_assign_pointer(memcg->memsw_thresholds, thresholds_new);
3633 /* To be sure that nobody uses thresholds before freeing it */
3638 mutex_unlock(&memcg->thresholds_lock);
3643 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
3644 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3646 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3647 struct mem_cgroup_eventfd_list *event;
3648 int type = MEMFILE_TYPE(cft->private);
3650 BUG_ON(type != _OOM_TYPE);
3651 event = kmalloc(sizeof(*event), GFP_KERNEL);
3655 mutex_lock(&memcg_oom_mutex);
3657 event->eventfd = eventfd;
3658 list_add(&event->list, &memcg->oom_notify);
3660 /* already in OOM ? */
3661 if (atomic_read(&memcg->oom_lock))
3662 eventfd_signal(eventfd, 1);
3663 mutex_unlock(&memcg_oom_mutex);
3668 static int mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
3669 struct cftype *cft, struct eventfd_ctx *eventfd)
3671 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3672 struct mem_cgroup_eventfd_list *ev, *tmp;
3673 int type = MEMFILE_TYPE(cft->private);
3675 BUG_ON(type != _OOM_TYPE);
3677 mutex_lock(&memcg_oom_mutex);
3679 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
3680 if (ev->eventfd == eventfd) {
3681 list_del(&ev->list);
3686 mutex_unlock(&memcg_oom_mutex);
3691 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
3692 struct cftype *cft, struct cgroup_map_cb *cb)
3694 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3696 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
3698 if (atomic_read(&mem->oom_lock))
3699 cb->fill(cb, "under_oom", 1);
3701 cb->fill(cb, "under_oom", 0);
3707 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
3708 struct cftype *cft, u64 val)
3710 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3711 struct mem_cgroup *parent;
3713 /* cannot set to root cgroup and only 0 and 1 are allowed */
3714 if (!cgrp->parent || !((val == 0) || (val == 1)))
3717 parent = mem_cgroup_from_cont(cgrp->parent);
3720 /* oom-kill-disable is a flag for subhierarchy. */
3721 if ((parent->use_hierarchy) ||
3722 (mem->use_hierarchy && !list_empty(&cgrp->children))) {
3726 mem->oom_kill_disable = val;
3731 static struct cftype mem_cgroup_files[] = {
3733 .name = "usage_in_bytes",
3734 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
3735 .read_u64 = mem_cgroup_read,
3736 .register_event = mem_cgroup_usage_register_event,
3737 .unregister_event = mem_cgroup_usage_unregister_event,
3740 .name = "max_usage_in_bytes",
3741 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
3742 .trigger = mem_cgroup_reset,
3743 .read_u64 = mem_cgroup_read,
3746 .name = "limit_in_bytes",
3747 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
3748 .write_string = mem_cgroup_write,
3749 .read_u64 = mem_cgroup_read,
3752 .name = "soft_limit_in_bytes",
3753 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
3754 .write_string = mem_cgroup_write,
3755 .read_u64 = mem_cgroup_read,
3759 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
3760 .trigger = mem_cgroup_reset,
3761 .read_u64 = mem_cgroup_read,
3765 .read_map = mem_control_stat_show,
3768 .name = "force_empty",
3769 .trigger = mem_cgroup_force_empty_write,
3772 .name = "use_hierarchy",
3773 .write_u64 = mem_cgroup_hierarchy_write,
3774 .read_u64 = mem_cgroup_hierarchy_read,
3777 .name = "swappiness",
3778 .read_u64 = mem_cgroup_swappiness_read,
3779 .write_u64 = mem_cgroup_swappiness_write,
3782 .name = "move_charge_at_immigrate",
3783 .read_u64 = mem_cgroup_move_charge_read,
3784 .write_u64 = mem_cgroup_move_charge_write,
3787 .name = "oom_control",
3788 .read_map = mem_cgroup_oom_control_read,
3789 .write_u64 = mem_cgroup_oom_control_write,
3790 .register_event = mem_cgroup_oom_register_event,
3791 .unregister_event = mem_cgroup_oom_unregister_event,
3792 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
3796 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3797 static struct cftype memsw_cgroup_files[] = {
3799 .name = "memsw.usage_in_bytes",
3800 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
3801 .read_u64 = mem_cgroup_read,
3802 .register_event = mem_cgroup_usage_register_event,
3803 .unregister_event = mem_cgroup_usage_unregister_event,
3806 .name = "memsw.max_usage_in_bytes",
3807 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
3808 .trigger = mem_cgroup_reset,
3809 .read_u64 = mem_cgroup_read,
3812 .name = "memsw.limit_in_bytes",
3813 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
3814 .write_string = mem_cgroup_write,
3815 .read_u64 = mem_cgroup_read,
3818 .name = "memsw.failcnt",
3819 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
3820 .trigger = mem_cgroup_reset,
3821 .read_u64 = mem_cgroup_read,
3825 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3827 if (!do_swap_account)
3829 return cgroup_add_files(cont, ss, memsw_cgroup_files,
3830 ARRAY_SIZE(memsw_cgroup_files));
3833 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
3839 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3841 struct mem_cgroup_per_node *pn;
3842 struct mem_cgroup_per_zone *mz;
3844 int zone, tmp = node;
3846 * This routine is called against possible nodes.
3847 * But it's BUG to call kmalloc() against offline node.
3849 * TODO: this routine can waste much memory for nodes which will
3850 * never be onlined. It's better to use memory hotplug callback
3853 if (!node_state(node, N_NORMAL_MEMORY))
3855 pn = kmalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
3859 mem->info.nodeinfo[node] = pn;
3860 memset(pn, 0, sizeof(*pn));
3862 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3863 mz = &pn->zoneinfo[zone];
3865 INIT_LIST_HEAD(&mz->lists[l]);
3866 mz->usage_in_excess = 0;
3867 mz->on_tree = false;
3873 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
3875 kfree(mem->info.nodeinfo[node]);
3878 static struct mem_cgroup *mem_cgroup_alloc(void)
3880 struct mem_cgroup *mem;
3881 int size = sizeof(struct mem_cgroup);
3883 /* Can be very big if MAX_NUMNODES is very big */
3884 if (size < PAGE_SIZE)
3885 mem = kmalloc(size, GFP_KERNEL);
3887 mem = vmalloc(size);
3892 memset(mem, 0, size);
3893 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
3895 if (size < PAGE_SIZE)
3905 * At destroying mem_cgroup, references from swap_cgroup can remain.
3906 * (scanning all at force_empty is too costly...)
3908 * Instead of clearing all references at force_empty, we remember
3909 * the number of reference from swap_cgroup and free mem_cgroup when
3910 * it goes down to 0.
3912 * Removal of cgroup itself succeeds regardless of refs from swap.
3915 static void __mem_cgroup_free(struct mem_cgroup *mem)
3919 mem_cgroup_remove_from_trees(mem);
3920 free_css_id(&mem_cgroup_subsys, &mem->css);
3922 for_each_node_state(node, N_POSSIBLE)
3923 free_mem_cgroup_per_zone_info(mem, node);
3925 free_percpu(mem->stat);
3926 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
3932 static void mem_cgroup_get(struct mem_cgroup *mem)
3934 atomic_inc(&mem->refcnt);
3937 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
3939 if (atomic_sub_and_test(count, &mem->refcnt)) {
3940 struct mem_cgroup *parent = parent_mem_cgroup(mem);
3941 __mem_cgroup_free(mem);
3943 mem_cgroup_put(parent);
3947 static void mem_cgroup_put(struct mem_cgroup *mem)
3949 __mem_cgroup_put(mem, 1);
3953 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
3955 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
3957 if (!mem->res.parent)
3959 return mem_cgroup_from_res_counter(mem->res.parent, res);
3962 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
3963 static void __init enable_swap_cgroup(void)
3965 if (!mem_cgroup_disabled() && really_do_swap_account)
3966 do_swap_account = 1;
3969 static void __init enable_swap_cgroup(void)
3974 static int mem_cgroup_soft_limit_tree_init(void)
3976 struct mem_cgroup_tree_per_node *rtpn;
3977 struct mem_cgroup_tree_per_zone *rtpz;
3978 int tmp, node, zone;
3980 for_each_node_state(node, N_POSSIBLE) {
3982 if (!node_state(node, N_NORMAL_MEMORY))
3984 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
3988 soft_limit_tree.rb_tree_per_node[node] = rtpn;
3990 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
3991 rtpz = &rtpn->rb_tree_per_zone[zone];
3992 rtpz->rb_root = RB_ROOT;
3993 spin_lock_init(&rtpz->lock);
3999 static struct cgroup_subsys_state * __ref
4000 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4002 struct mem_cgroup *mem, *parent;
4003 long error = -ENOMEM;
4006 mem = mem_cgroup_alloc();
4008 return ERR_PTR(error);
4010 for_each_node_state(node, N_POSSIBLE)
4011 if (alloc_mem_cgroup_per_zone_info(mem, node))
4015 if (cont->parent == NULL) {
4017 enable_swap_cgroup();
4019 root_mem_cgroup = mem;
4020 if (mem_cgroup_soft_limit_tree_init())
4022 for_each_possible_cpu(cpu) {
4023 struct memcg_stock_pcp *stock =
4024 &per_cpu(memcg_stock, cpu);
4025 INIT_WORK(&stock->work, drain_local_stock);
4027 hotcpu_notifier(memcg_stock_cpu_callback, 0);
4029 parent = mem_cgroup_from_cont(cont->parent);
4030 mem->use_hierarchy = parent->use_hierarchy;
4031 mem->oom_kill_disable = parent->oom_kill_disable;
4034 if (parent && parent->use_hierarchy) {
4035 res_counter_init(&mem->res, &parent->res);
4036 res_counter_init(&mem->memsw, &parent->memsw);
4038 * We increment refcnt of the parent to ensure that we can
4039 * safely access it on res_counter_charge/uncharge.
4040 * This refcnt will be decremented when freeing this
4041 * mem_cgroup(see mem_cgroup_put).
4043 mem_cgroup_get(parent);
4045 res_counter_init(&mem->res, NULL);
4046 res_counter_init(&mem->memsw, NULL);
4048 mem->last_scanned_child = 0;
4049 spin_lock_init(&mem->reclaim_param_lock);
4050 INIT_LIST_HEAD(&mem->oom_notify);
4053 mem->swappiness = get_swappiness(parent);
4054 atomic_set(&mem->refcnt, 1);
4055 mem->move_charge_at_immigrate = 0;
4056 mutex_init(&mem->thresholds_lock);
4059 __mem_cgroup_free(mem);
4060 root_mem_cgroup = NULL;
4061 return ERR_PTR(error);
4064 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4065 struct cgroup *cont)
4067 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4069 return mem_cgroup_force_empty(mem, false);
4072 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4073 struct cgroup *cont)
4075 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4077 mem_cgroup_put(mem);
4080 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4081 struct cgroup *cont)
4085 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4086 ARRAY_SIZE(mem_cgroup_files));
4089 ret = register_memsw_files(cont, ss);
4094 /* Handlers for move charge at task migration. */
4095 #define PRECHARGE_COUNT_AT_ONCE 256
4096 static int mem_cgroup_do_precharge(unsigned long count)
4099 int batch_count = PRECHARGE_COUNT_AT_ONCE;
4100 struct mem_cgroup *mem = mc.to;
4102 if (mem_cgroup_is_root(mem)) {
4103 mc.precharge += count;
4104 /* we don't need css_get for root */
4107 /* try to charge at once */
4109 struct res_counter *dummy;
4111 * "mem" cannot be under rmdir() because we've already checked
4112 * by cgroup_lock_live_cgroup() that it is not removed and we
4113 * are still under the same cgroup_mutex. So we can postpone
4116 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4118 if (do_swap_account && res_counter_charge(&mem->memsw,
4119 PAGE_SIZE * count, &dummy)) {
4120 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4123 mc.precharge += count;
4124 VM_BUG_ON(test_bit(CSS_ROOT, &mem->css.flags));
4125 WARN_ON_ONCE(count > INT_MAX);
4126 __css_get(&mem->css, (int)count);
4130 /* fall back to one by one charge */
4132 if (signal_pending(current)) {
4136 if (!batch_count--) {
4137 batch_count = PRECHARGE_COUNT_AT_ONCE;
4140 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false);
4142 /* mem_cgroup_clear_mc() will do uncharge later */
4150 * is_target_pte_for_mc - check a pte whether it is valid for move charge
4151 * @vma: the vma the pte to be checked belongs
4152 * @addr: the address corresponding to the pte to be checked
4153 * @ptent: the pte to be checked
4154 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4157 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4158 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4159 * move charge. if @target is not NULL, the page is stored in target->page
4160 * with extra refcnt got(Callers should handle it).
4161 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4162 * target for charge migration. if @target is not NULL, the entry is stored
4165 * Called with pte lock held.
4172 enum mc_target_type {
4173 MC_TARGET_NONE, /* not used */
4178 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4179 unsigned long addr, pte_t ptent)
4181 struct page *page = vm_normal_page(vma, addr, ptent);
4183 if (!page || !page_mapped(page))
4185 if (PageAnon(page)) {
4186 /* we don't move shared anon */
4187 if (!move_anon() || page_mapcount(page) > 2)
4189 } else if (!move_file())
4190 /* we ignore mapcount for file pages */
4192 if (!get_page_unless_zero(page))
4198 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4199 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4202 struct page *page = NULL;
4203 swp_entry_t ent = pte_to_swp_entry(ptent);
4205 if (!move_anon() || non_swap_entry(ent))
4207 usage_count = mem_cgroup_count_swap_user(ent, &page);
4208 if (usage_count > 1) { /* we don't move shared anon */
4213 if (do_swap_account)
4214 entry->val = ent.val;
4219 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4220 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4222 struct page *page = NULL;
4223 struct inode *inode;
4224 struct address_space *mapping;
4227 if (!vma->vm_file) /* anonymous vma */
4232 inode = vma->vm_file->f_path.dentry->d_inode;
4233 mapping = vma->vm_file->f_mapping;
4234 if (pte_none(ptent))
4235 pgoff = linear_page_index(vma, addr);
4236 else /* pte_file(ptent) is true */
4237 pgoff = pte_to_pgoff(ptent);
4239 /* page is moved even if it's not RSS of this task(page-faulted). */
4240 if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4241 page = find_get_page(mapping, pgoff);
4242 } else { /* shmem/tmpfs file. we should take account of swap too. */
4244 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4245 if (do_swap_account)
4246 entry->val = ent.val;
4252 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4253 unsigned long addr, pte_t ptent, union mc_target *target)
4255 struct page *page = NULL;
4256 struct page_cgroup *pc;
4258 swp_entry_t ent = { .val = 0 };
4260 if (pte_present(ptent))
4261 page = mc_handle_present_pte(vma, addr, ptent);
4262 else if (is_swap_pte(ptent))
4263 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4264 else if (pte_none(ptent) || pte_file(ptent))
4265 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4267 if (!page && !ent.val)
4270 pc = lookup_page_cgroup(page);
4272 * Do only loose check w/o page_cgroup lock.
4273 * mem_cgroup_move_account() checks the pc is valid or not under
4276 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4277 ret = MC_TARGET_PAGE;
4279 target->page = page;
4281 if (!ret || !target)
4284 /* There is a swap entry and a page doesn't exist or isn't charged */
4285 if (ent.val && !ret &&
4286 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4287 ret = MC_TARGET_SWAP;
4294 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4295 unsigned long addr, unsigned long end,
4296 struct mm_walk *walk)
4298 struct vm_area_struct *vma = walk->private;
4302 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4303 for (; addr != end; pte++, addr += PAGE_SIZE)
4304 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4305 mc.precharge++; /* increment precharge temporarily */
4306 pte_unmap_unlock(pte - 1, ptl);
4312 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4314 unsigned long precharge;
4315 struct vm_area_struct *vma;
4317 down_read(&mm->mmap_sem);
4318 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4319 struct mm_walk mem_cgroup_count_precharge_walk = {
4320 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4324 if (is_vm_hugetlb_page(vma))
4326 walk_page_range(vma->vm_start, vma->vm_end,
4327 &mem_cgroup_count_precharge_walk);
4329 up_read(&mm->mmap_sem);
4331 precharge = mc.precharge;
4337 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4339 return mem_cgroup_do_precharge(mem_cgroup_count_precharge(mm));
4342 static void mem_cgroup_clear_mc(void)
4344 /* we must uncharge all the leftover precharges from mc.to */
4346 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4348 memcg_oom_recover(mc.to);
4351 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4352 * we must uncharge here.
4354 if (mc.moved_charge) {
4355 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4356 mc.moved_charge = 0;
4357 memcg_oom_recover(mc.from);
4359 /* we must fixup refcnts and charges */
4360 if (mc.moved_swap) {
4361 WARN_ON_ONCE(mc.moved_swap > INT_MAX);
4362 /* uncharge swap account from the old cgroup */
4363 if (!mem_cgroup_is_root(mc.from))
4364 res_counter_uncharge(&mc.from->memsw,
4365 PAGE_SIZE * mc.moved_swap);
4366 __mem_cgroup_put(mc.from, mc.moved_swap);
4368 if (!mem_cgroup_is_root(mc.to)) {
4370 * we charged both to->res and to->memsw, so we should
4373 res_counter_uncharge(&mc.to->res,
4374 PAGE_SIZE * mc.moved_swap);
4375 VM_BUG_ON(test_bit(CSS_ROOT, &mc.to->css.flags));
4376 __css_put(&mc.to->css, mc.moved_swap);
4378 /* we've already done mem_cgroup_get(mc.to) */
4384 mc.moving_task = NULL;
4385 wake_up_all(&mc.waitq);
4388 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4389 struct cgroup *cgroup,
4390 struct task_struct *p,
4394 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4396 if (mem->move_charge_at_immigrate) {
4397 struct mm_struct *mm;
4398 struct mem_cgroup *from = mem_cgroup_from_task(p);
4400 VM_BUG_ON(from == mem);
4402 mm = get_task_mm(p);
4405 /* We move charges only when we move a owner of the mm */
4406 if (mm->owner == p) {
4409 VM_BUG_ON(mc.precharge);
4410 VM_BUG_ON(mc.moved_charge);
4411 VM_BUG_ON(mc.moved_swap);
4412 VM_BUG_ON(mc.moving_task);
4416 mc.moved_charge = 0;
4418 mc.moving_task = current;
4420 ret = mem_cgroup_precharge_mc(mm);
4422 mem_cgroup_clear_mc();
4429 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4430 struct cgroup *cgroup,
4431 struct task_struct *p,
4434 mem_cgroup_clear_mc();
4437 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4438 unsigned long addr, unsigned long end,
4439 struct mm_walk *walk)
4442 struct vm_area_struct *vma = walk->private;
4447 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4448 for (; addr != end; addr += PAGE_SIZE) {
4449 pte_t ptent = *(pte++);
4450 union mc_target target;
4453 struct page_cgroup *pc;
4459 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4461 case MC_TARGET_PAGE:
4463 if (isolate_lru_page(page))
4465 pc = lookup_page_cgroup(page);
4466 if (!mem_cgroup_move_account(pc,
4467 mc.from, mc.to, false)) {
4469 /* we uncharge from mc.from later. */
4472 putback_lru_page(page);
4473 put: /* is_target_pte_for_mc() gets the page */
4476 case MC_TARGET_SWAP:
4478 if (!mem_cgroup_move_swap_account(ent,
4479 mc.from, mc.to, false)) {
4481 /* we fixup refcnts and charges later. */
4489 pte_unmap_unlock(pte - 1, ptl);
4494 * We have consumed all precharges we got in can_attach().
4495 * We try charge one by one, but don't do any additional
4496 * charges to mc.to if we have failed in charge once in attach()
4499 ret = mem_cgroup_do_precharge(1);
4507 static void mem_cgroup_move_charge(struct mm_struct *mm)
4509 struct vm_area_struct *vma;
4511 lru_add_drain_all();
4512 down_read(&mm->mmap_sem);
4513 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4515 struct mm_walk mem_cgroup_move_charge_walk = {
4516 .pmd_entry = mem_cgroup_move_charge_pte_range,
4520 if (is_vm_hugetlb_page(vma))
4522 ret = walk_page_range(vma->vm_start, vma->vm_end,
4523 &mem_cgroup_move_charge_walk);
4526 * means we have consumed all precharges and failed in
4527 * doing additional charge. Just abandon here.
4531 up_read(&mm->mmap_sem);
4534 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4535 struct cgroup *cont,
4536 struct cgroup *old_cont,
4537 struct task_struct *p,
4540 struct mm_struct *mm;
4543 /* no need to move charge */
4546 mm = get_task_mm(p);
4548 mem_cgroup_move_charge(mm);
4551 mem_cgroup_clear_mc();
4553 #else /* !CONFIG_MMU */
4554 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4555 struct cgroup *cgroup,
4556 struct task_struct *p,
4561 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4562 struct cgroup *cgroup,
4563 struct task_struct *p,
4567 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
4568 struct cgroup *cont,
4569 struct cgroup *old_cont,
4570 struct task_struct *p,
4576 struct cgroup_subsys mem_cgroup_subsys = {
4578 .subsys_id = mem_cgroup_subsys_id,
4579 .create = mem_cgroup_create,
4580 .pre_destroy = mem_cgroup_pre_destroy,
4581 .destroy = mem_cgroup_destroy,
4582 .populate = mem_cgroup_populate,
4583 .can_attach = mem_cgroup_can_attach,
4584 .cancel_attach = mem_cgroup_cancel_attach,
4585 .attach = mem_cgroup_move_task,
4590 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4592 static int __init disable_swap_account(char *s)
4594 really_do_swap_account = 0;
4597 __setup("noswapaccount", disable_swap_account);