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>
50 #include <linux/oom.h>
53 #include <asm/uaccess.h>
55 #include <trace/events/vmscan.h>
57 struct cgroup_subsys mem_cgroup_subsys __read_mostly;
58 #define MEM_CGROUP_RECLAIM_RETRIES 5
59 struct mem_cgroup *root_mem_cgroup __read_mostly;
61 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
62 /* Turned on only when memory cgroup is enabled && really_do_swap_account = 1 */
63 int do_swap_account __read_mostly;
65 /* for remember boot option*/
66 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP_ENABLED
67 static int really_do_swap_account __initdata = 1;
69 static int really_do_swap_account __initdata = 0;
73 #define do_swap_account (0)
77 * Per memcg event counter is incremented at every pagein/pageout. This counter
78 * is used for trigger some periodic events. This is straightforward and better
79 * than using jiffies etc. to handle periodic memcg event.
81 * These values will be used as !((event) & ((1 <<(thresh)) - 1))
83 #define THRESHOLDS_EVENTS_THRESH (7) /* once in 128 */
84 #define SOFTLIMIT_EVENTS_THRESH (10) /* once in 1024 */
87 * Statistics for memory cgroup.
89 enum mem_cgroup_stat_index {
91 * For MEM_CONTAINER_TYPE_ALL, usage = pagecache + rss.
93 MEM_CGROUP_STAT_CACHE, /* # of pages charged as cache */
94 MEM_CGROUP_STAT_RSS, /* # of pages charged as anon rss */
95 MEM_CGROUP_STAT_FILE_MAPPED, /* # of pages charged as file rss */
96 MEM_CGROUP_STAT_PGPGIN_COUNT, /* # of pages paged in */
97 MEM_CGROUP_STAT_PGPGOUT_COUNT, /* # of pages paged out */
98 MEM_CGROUP_STAT_SWAPOUT, /* # of pages, swapped out */
99 MEM_CGROUP_STAT_DATA, /* end of data requires synchronization */
100 /* incremented at every pagein/pageout */
101 MEM_CGROUP_EVENTS = MEM_CGROUP_STAT_DATA,
102 MEM_CGROUP_ON_MOVE, /* someone is moving account between groups */
104 MEM_CGROUP_STAT_NSTATS,
107 struct mem_cgroup_stat_cpu {
108 s64 count[MEM_CGROUP_STAT_NSTATS];
112 * per-zone information in memory controller.
114 struct mem_cgroup_per_zone {
116 * spin_lock to protect the per cgroup LRU
118 struct list_head lists[NR_LRU_LISTS];
119 unsigned long count[NR_LRU_LISTS];
121 struct zone_reclaim_stat reclaim_stat;
122 struct rb_node tree_node; /* RB tree node */
123 unsigned long long usage_in_excess;/* Set to the value by which */
124 /* the soft limit is exceeded*/
126 struct mem_cgroup *mem; /* Back pointer, we cannot */
127 /* use container_of */
129 /* Macro for accessing counter */
130 #define MEM_CGROUP_ZSTAT(mz, idx) ((mz)->count[(idx)])
132 struct mem_cgroup_per_node {
133 struct mem_cgroup_per_zone zoneinfo[MAX_NR_ZONES];
136 struct mem_cgroup_lru_info {
137 struct mem_cgroup_per_node *nodeinfo[MAX_NUMNODES];
141 * Cgroups above their limits are maintained in a RB-Tree, independent of
142 * their hierarchy representation
145 struct mem_cgroup_tree_per_zone {
146 struct rb_root rb_root;
150 struct mem_cgroup_tree_per_node {
151 struct mem_cgroup_tree_per_zone rb_tree_per_zone[MAX_NR_ZONES];
154 struct mem_cgroup_tree {
155 struct mem_cgroup_tree_per_node *rb_tree_per_node[MAX_NUMNODES];
158 static struct mem_cgroup_tree soft_limit_tree __read_mostly;
160 struct mem_cgroup_threshold {
161 struct eventfd_ctx *eventfd;
166 struct mem_cgroup_threshold_ary {
167 /* An array index points to threshold just below usage. */
168 int current_threshold;
169 /* Size of entries[] */
171 /* Array of thresholds */
172 struct mem_cgroup_threshold entries[0];
175 struct mem_cgroup_thresholds {
176 /* Primary thresholds array */
177 struct mem_cgroup_threshold_ary *primary;
179 * Spare threshold array.
180 * This is needed to make mem_cgroup_unregister_event() "never fail".
181 * It must be able to store at least primary->size - 1 entries.
183 struct mem_cgroup_threshold_ary *spare;
187 struct mem_cgroup_eventfd_list {
188 struct list_head list;
189 struct eventfd_ctx *eventfd;
192 static void mem_cgroup_threshold(struct mem_cgroup *mem);
193 static void mem_cgroup_oom_notify(struct mem_cgroup *mem);
196 * The memory controller data structure. The memory controller controls both
197 * page cache and RSS per cgroup. We would eventually like to provide
198 * statistics based on the statistics developed by Rik Van Riel for clock-pro,
199 * to help the administrator determine what knobs to tune.
201 * TODO: Add a water mark for the memory controller. Reclaim will begin when
202 * we hit the water mark. May be even add a low water mark, such that
203 * no reclaim occurs from a cgroup at it's low water mark, this is
204 * a feature that will be implemented much later in the future.
207 struct cgroup_subsys_state css;
209 * the counter to account for memory usage
211 struct res_counter res;
213 * the counter to account for mem+swap usage.
215 struct res_counter memsw;
217 * Per cgroup active and inactive list, similar to the
218 * per zone LRU lists.
220 struct mem_cgroup_lru_info info;
223 protect against reclaim related member.
225 spinlock_t reclaim_param_lock;
228 * While reclaiming in a hierarchy, we cache the last child we
231 int last_scanned_child;
233 * Should the accounting and control be hierarchical, per subtree?
239 unsigned int swappiness;
240 /* OOM-Killer disable */
241 int oom_kill_disable;
243 /* set when res.limit == memsw.limit */
244 bool memsw_is_minimum;
246 /* protect arrays of thresholds */
247 struct mutex thresholds_lock;
249 /* thresholds for memory usage. RCU-protected */
250 struct mem_cgroup_thresholds thresholds;
252 /* thresholds for mem+swap usage. RCU-protected */
253 struct mem_cgroup_thresholds memsw_thresholds;
255 /* For oom notifier event fd */
256 struct list_head oom_notify;
259 * Should we move charges of a task when a task is moved into this
260 * mem_cgroup ? And what type of charges should we move ?
262 unsigned long move_charge_at_immigrate;
266 struct mem_cgroup_stat_cpu *stat;
268 * used when a cpu is offlined or other synchronizations
269 * See mem_cgroup_read_stat().
271 struct mem_cgroup_stat_cpu nocpu_base;
272 spinlock_t pcp_counter_lock;
275 /* Stuffs for move charges at task migration. */
277 * Types of charges to be moved. "move_charge_at_immitgrate" is treated as a
278 * left-shifted bitmap of these types.
281 MOVE_CHARGE_TYPE_ANON, /* private anonymous page and swap of it */
282 MOVE_CHARGE_TYPE_FILE, /* file page(including tmpfs) and swap of it */
286 /* "mc" and its members are protected by cgroup_mutex */
287 static struct move_charge_struct {
288 spinlock_t lock; /* for from, to */
289 struct mem_cgroup *from;
290 struct mem_cgroup *to;
291 unsigned long precharge;
292 unsigned long moved_charge;
293 unsigned long moved_swap;
294 struct task_struct *moving_task; /* a task moving charges */
295 wait_queue_head_t waitq; /* a waitq for other context */
297 .lock = __SPIN_LOCK_UNLOCKED(mc.lock),
298 .waitq = __WAIT_QUEUE_HEAD_INITIALIZER(mc.waitq),
301 static bool move_anon(void)
303 return test_bit(MOVE_CHARGE_TYPE_ANON,
304 &mc.to->move_charge_at_immigrate);
307 static bool move_file(void)
309 return test_bit(MOVE_CHARGE_TYPE_FILE,
310 &mc.to->move_charge_at_immigrate);
314 * Maximum loops in mem_cgroup_hierarchical_reclaim(), used for soft
315 * limit reclaim to prevent infinite loops, if they ever occur.
317 #define MEM_CGROUP_MAX_RECLAIM_LOOPS (100)
318 #define MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS (2)
321 MEM_CGROUP_CHARGE_TYPE_CACHE = 0,
322 MEM_CGROUP_CHARGE_TYPE_MAPPED,
323 MEM_CGROUP_CHARGE_TYPE_SHMEM, /* used by page migration of shmem */
324 MEM_CGROUP_CHARGE_TYPE_FORCE, /* used by force_empty */
325 MEM_CGROUP_CHARGE_TYPE_SWAPOUT, /* for accounting swapcache */
326 MEM_CGROUP_CHARGE_TYPE_DROP, /* a page was unused swap cache */
330 /* for encoding cft->private value on file */
333 #define _OOM_TYPE (2)
334 #define MEMFILE_PRIVATE(x, val) (((x) << 16) | (val))
335 #define MEMFILE_TYPE(val) (((val) >> 16) & 0xffff)
336 #define MEMFILE_ATTR(val) ((val) & 0xffff)
337 /* Used for OOM nofiier */
338 #define OOM_CONTROL (0)
341 * Reclaim flags for mem_cgroup_hierarchical_reclaim
343 #define MEM_CGROUP_RECLAIM_NOSWAP_BIT 0x0
344 #define MEM_CGROUP_RECLAIM_NOSWAP (1 << MEM_CGROUP_RECLAIM_NOSWAP_BIT)
345 #define MEM_CGROUP_RECLAIM_SHRINK_BIT 0x1
346 #define MEM_CGROUP_RECLAIM_SHRINK (1 << MEM_CGROUP_RECLAIM_SHRINK_BIT)
347 #define MEM_CGROUP_RECLAIM_SOFT_BIT 0x2
348 #define MEM_CGROUP_RECLAIM_SOFT (1 << MEM_CGROUP_RECLAIM_SOFT_BIT)
350 static void mem_cgroup_get(struct mem_cgroup *mem);
351 static void mem_cgroup_put(struct mem_cgroup *mem);
352 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem);
353 static void drain_all_stock_async(void);
355 static struct mem_cgroup_per_zone *
356 mem_cgroup_zoneinfo(struct mem_cgroup *mem, int nid, int zid)
358 return &mem->info.nodeinfo[nid]->zoneinfo[zid];
361 struct cgroup_subsys_state *mem_cgroup_css(struct mem_cgroup *mem)
366 static struct mem_cgroup_per_zone *
367 page_cgroup_zoneinfo(struct page_cgroup *pc)
369 struct mem_cgroup *mem = pc->mem_cgroup;
370 int nid = page_cgroup_nid(pc);
371 int zid = page_cgroup_zid(pc);
373 return mem_cgroup_zoneinfo(mem, nid, zid);
376 static struct mem_cgroup_tree_per_zone *
377 soft_limit_tree_node_zone(int nid, int zid)
379 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
382 static struct mem_cgroup_tree_per_zone *
383 soft_limit_tree_from_page(struct page *page)
385 int nid = page_to_nid(page);
386 int zid = page_zonenum(page);
388 return &soft_limit_tree.rb_tree_per_node[nid]->rb_tree_per_zone[zid];
392 __mem_cgroup_insert_exceeded(struct mem_cgroup *mem,
393 struct mem_cgroup_per_zone *mz,
394 struct mem_cgroup_tree_per_zone *mctz,
395 unsigned long long new_usage_in_excess)
397 struct rb_node **p = &mctz->rb_root.rb_node;
398 struct rb_node *parent = NULL;
399 struct mem_cgroup_per_zone *mz_node;
404 mz->usage_in_excess = new_usage_in_excess;
405 if (!mz->usage_in_excess)
409 mz_node = rb_entry(parent, struct mem_cgroup_per_zone,
411 if (mz->usage_in_excess < mz_node->usage_in_excess)
414 * We can't avoid mem cgroups that are over their soft
415 * limit by the same amount
417 else if (mz->usage_in_excess >= mz_node->usage_in_excess)
420 rb_link_node(&mz->tree_node, parent, p);
421 rb_insert_color(&mz->tree_node, &mctz->rb_root);
426 __mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
427 struct mem_cgroup_per_zone *mz,
428 struct mem_cgroup_tree_per_zone *mctz)
432 rb_erase(&mz->tree_node, &mctz->rb_root);
437 mem_cgroup_remove_exceeded(struct mem_cgroup *mem,
438 struct mem_cgroup_per_zone *mz,
439 struct mem_cgroup_tree_per_zone *mctz)
441 spin_lock(&mctz->lock);
442 __mem_cgroup_remove_exceeded(mem, mz, mctz);
443 spin_unlock(&mctz->lock);
447 static void mem_cgroup_update_tree(struct mem_cgroup *mem, struct page *page)
449 unsigned long long excess;
450 struct mem_cgroup_per_zone *mz;
451 struct mem_cgroup_tree_per_zone *mctz;
452 int nid = page_to_nid(page);
453 int zid = page_zonenum(page);
454 mctz = soft_limit_tree_from_page(page);
457 * Necessary to update all ancestors when hierarchy is used.
458 * because their event counter is not touched.
460 for (; mem; mem = parent_mem_cgroup(mem)) {
461 mz = mem_cgroup_zoneinfo(mem, nid, zid);
462 excess = res_counter_soft_limit_excess(&mem->res);
464 * We have to update the tree if mz is on RB-tree or
465 * mem is over its softlimit.
467 if (excess || mz->on_tree) {
468 spin_lock(&mctz->lock);
469 /* if on-tree, remove it */
471 __mem_cgroup_remove_exceeded(mem, mz, mctz);
473 * Insert again. mz->usage_in_excess will be updated.
474 * If excess is 0, no tree ops.
476 __mem_cgroup_insert_exceeded(mem, mz, mctz, excess);
477 spin_unlock(&mctz->lock);
482 static void mem_cgroup_remove_from_trees(struct mem_cgroup *mem)
485 struct mem_cgroup_per_zone *mz;
486 struct mem_cgroup_tree_per_zone *mctz;
488 for_each_node_state(node, N_POSSIBLE) {
489 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
490 mz = mem_cgroup_zoneinfo(mem, node, zone);
491 mctz = soft_limit_tree_node_zone(node, zone);
492 mem_cgroup_remove_exceeded(mem, mz, mctz);
497 static struct mem_cgroup_per_zone *
498 __mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
500 struct rb_node *rightmost = NULL;
501 struct mem_cgroup_per_zone *mz;
505 rightmost = rb_last(&mctz->rb_root);
507 goto done; /* Nothing to reclaim from */
509 mz = rb_entry(rightmost, struct mem_cgroup_per_zone, tree_node);
511 * Remove the node now but someone else can add it back,
512 * we will to add it back at the end of reclaim to its correct
513 * position in the tree.
515 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
516 if (!res_counter_soft_limit_excess(&mz->mem->res) ||
517 !css_tryget(&mz->mem->css))
523 static struct mem_cgroup_per_zone *
524 mem_cgroup_largest_soft_limit_node(struct mem_cgroup_tree_per_zone *mctz)
526 struct mem_cgroup_per_zone *mz;
528 spin_lock(&mctz->lock);
529 mz = __mem_cgroup_largest_soft_limit_node(mctz);
530 spin_unlock(&mctz->lock);
535 * Implementation Note: reading percpu statistics for memcg.
537 * Both of vmstat[] and percpu_counter has threshold and do periodic
538 * synchronization to implement "quick" read. There are trade-off between
539 * reading cost and precision of value. Then, we may have a chance to implement
540 * a periodic synchronizion of counter in memcg's counter.
542 * But this _read() function is used for user interface now. The user accounts
543 * memory usage by memory cgroup and he _always_ requires exact value because
544 * he accounts memory. Even if we provide quick-and-fuzzy read, we always
545 * have to visit all online cpus and make sum. So, for now, unnecessary
546 * synchronization is not implemented. (just implemented for cpu hotplug)
548 * If there are kernel internal actions which can make use of some not-exact
549 * value, and reading all cpu value can be performance bottleneck in some
550 * common workload, threashold and synchonization as vmstat[] should be
553 static s64 mem_cgroup_read_stat(struct mem_cgroup *mem,
554 enum mem_cgroup_stat_index idx)
560 for_each_online_cpu(cpu)
561 val += per_cpu(mem->stat->count[idx], cpu);
562 #ifdef CONFIG_HOTPLUG_CPU
563 spin_lock(&mem->pcp_counter_lock);
564 val += mem->nocpu_base.count[idx];
565 spin_unlock(&mem->pcp_counter_lock);
571 static s64 mem_cgroup_local_usage(struct mem_cgroup *mem)
575 ret = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
576 ret += mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
580 static void mem_cgroup_swap_statistics(struct mem_cgroup *mem,
583 int val = (charge) ? 1 : -1;
584 this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_SWAPOUT], val);
587 static void mem_cgroup_charge_statistics(struct mem_cgroup *mem,
588 bool file, int nr_pages)
593 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_CACHE], nr_pages);
595 __this_cpu_add(mem->stat->count[MEM_CGROUP_STAT_RSS], nr_pages);
597 /* pagein of a big page is an event. So, ignore page size */
599 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGIN_COUNT]);
601 __this_cpu_inc(mem->stat->count[MEM_CGROUP_STAT_PGPGOUT_COUNT]);
602 nr_pages = -nr_pages; /* for event */
605 __this_cpu_add(mem->stat->count[MEM_CGROUP_EVENTS], nr_pages);
610 static unsigned long mem_cgroup_get_local_zonestat(struct mem_cgroup *mem,
614 struct mem_cgroup_per_zone *mz;
617 for_each_online_node(nid)
618 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
619 mz = mem_cgroup_zoneinfo(mem, nid, zid);
620 total += MEM_CGROUP_ZSTAT(mz, idx);
625 static bool __memcg_event_check(struct mem_cgroup *mem, int event_mask_shift)
629 val = this_cpu_read(mem->stat->count[MEM_CGROUP_EVENTS]);
631 return !(val & ((1 << event_mask_shift) - 1));
635 * Check events in order.
638 static void memcg_check_events(struct mem_cgroup *mem, struct page *page)
640 /* threshold event is triggered in finer grain than soft limit */
641 if (unlikely(__memcg_event_check(mem, THRESHOLDS_EVENTS_THRESH))) {
642 mem_cgroup_threshold(mem);
643 if (unlikely(__memcg_event_check(mem, SOFTLIMIT_EVENTS_THRESH)))
644 mem_cgroup_update_tree(mem, page);
648 static struct mem_cgroup *mem_cgroup_from_cont(struct cgroup *cont)
650 return container_of(cgroup_subsys_state(cont,
651 mem_cgroup_subsys_id), struct mem_cgroup,
655 struct mem_cgroup *mem_cgroup_from_task(struct task_struct *p)
658 * mm_update_next_owner() may clear mm->owner to NULL
659 * if it races with swapoff, page migration, etc.
660 * So this can be called with p == NULL.
665 return container_of(task_subsys_state(p, mem_cgroup_subsys_id),
666 struct mem_cgroup, css);
669 static struct mem_cgroup *try_get_mem_cgroup_from_mm(struct mm_struct *mm)
671 struct mem_cgroup *mem = NULL;
676 * Because we have no locks, mm->owner's may be being moved to other
677 * cgroup. We use css_tryget() here even if this looks
678 * pessimistic (rather than adding locks here).
682 mem = mem_cgroup_from_task(rcu_dereference(mm->owner));
685 } while (!css_tryget(&mem->css));
690 /* The caller has to guarantee "mem" exists before calling this */
691 static struct mem_cgroup *mem_cgroup_start_loop(struct mem_cgroup *mem)
693 struct cgroup_subsys_state *css;
696 if (!mem) /* ROOT cgroup has the smallest ID */
697 return root_mem_cgroup; /*css_put/get against root is ignored*/
698 if (!mem->use_hierarchy) {
699 if (css_tryget(&mem->css))
705 * searching a memory cgroup which has the smallest ID under given
706 * ROOT cgroup. (ID >= 1)
708 css = css_get_next(&mem_cgroup_subsys, 1, &mem->css, &found);
709 if (css && css_tryget(css))
710 mem = container_of(css, struct mem_cgroup, css);
717 static struct mem_cgroup *mem_cgroup_get_next(struct mem_cgroup *iter,
718 struct mem_cgroup *root,
721 int nextid = css_id(&iter->css) + 1;
724 struct cgroup_subsys_state *css;
726 hierarchy_used = iter->use_hierarchy;
729 /* If no ROOT, walk all, ignore hierarchy */
730 if (!cond || (root && !hierarchy_used))
734 root = root_mem_cgroup;
740 css = css_get_next(&mem_cgroup_subsys, nextid,
742 if (css && css_tryget(css))
743 iter = container_of(css, struct mem_cgroup, css);
745 /* If css is NULL, no more cgroups will be found */
747 } while (css && !iter);
752 * for_eacn_mem_cgroup_tree() for visiting all cgroup under tree. Please
753 * be careful that "break" loop is not allowed. We have reference count.
754 * Instead of that modify "cond" to be false and "continue" to exit the loop.
756 #define for_each_mem_cgroup_tree_cond(iter, root, cond) \
757 for (iter = mem_cgroup_start_loop(root);\
759 iter = mem_cgroup_get_next(iter, root, cond))
761 #define for_each_mem_cgroup_tree(iter, root) \
762 for_each_mem_cgroup_tree_cond(iter, root, true)
764 #define for_each_mem_cgroup_all(iter) \
765 for_each_mem_cgroup_tree_cond(iter, NULL, true)
768 static inline bool mem_cgroup_is_root(struct mem_cgroup *mem)
770 return (mem == root_mem_cgroup);
774 * Following LRU functions are allowed to be used without PCG_LOCK.
775 * Operations are called by routine of global LRU independently from memcg.
776 * What we have to take care of here is validness of pc->mem_cgroup.
778 * Changes to pc->mem_cgroup happens when
781 * In typical case, "charge" is done before add-to-lru. Exception is SwapCache.
782 * It is added to LRU before charge.
783 * If PCG_USED bit is not set, page_cgroup is not added to this private LRU.
784 * When moving account, the page is not on LRU. It's isolated.
787 void mem_cgroup_del_lru_list(struct page *page, enum lru_list lru)
789 struct page_cgroup *pc;
790 struct mem_cgroup_per_zone *mz;
792 if (mem_cgroup_disabled())
794 pc = lookup_page_cgroup(page);
795 /* can happen while we handle swapcache. */
796 if (!TestClearPageCgroupAcctLRU(pc))
798 VM_BUG_ON(!pc->mem_cgroup);
800 * We don't check PCG_USED bit. It's cleared when the "page" is finally
801 * removed from global LRU.
803 mz = page_cgroup_zoneinfo(pc);
804 /* huge page split is done under lru_lock. so, we have no races. */
805 MEM_CGROUP_ZSTAT(mz, lru) -= 1 << compound_order(page);
806 if (mem_cgroup_is_root(pc->mem_cgroup))
808 VM_BUG_ON(list_empty(&pc->lru));
809 list_del_init(&pc->lru);
812 void mem_cgroup_del_lru(struct page *page)
814 mem_cgroup_del_lru_list(page, page_lru(page));
818 * Writeback is about to end against a page which has been marked for immediate
819 * reclaim. If it still appears to be reclaimable, move it to the tail of the
822 void mem_cgroup_rotate_reclaimable_page(struct page *page)
824 struct mem_cgroup_per_zone *mz;
825 struct page_cgroup *pc;
826 enum lru_list lru = page_lru(page);
828 if (mem_cgroup_disabled())
831 pc = lookup_page_cgroup(page);
832 /* unused or root page is not rotated. */
833 if (!PageCgroupUsed(pc))
835 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
837 if (mem_cgroup_is_root(pc->mem_cgroup))
839 mz = page_cgroup_zoneinfo(pc);
840 list_move_tail(&pc->lru, &mz->lists[lru]);
843 void mem_cgroup_rotate_lru_list(struct page *page, enum lru_list lru)
845 struct mem_cgroup_per_zone *mz;
846 struct page_cgroup *pc;
848 if (mem_cgroup_disabled())
851 pc = lookup_page_cgroup(page);
852 /* unused or root page is not rotated. */
853 if (!PageCgroupUsed(pc))
855 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
857 if (mem_cgroup_is_root(pc->mem_cgroup))
859 mz = page_cgroup_zoneinfo(pc);
860 list_move(&pc->lru, &mz->lists[lru]);
863 void mem_cgroup_add_lru_list(struct page *page, enum lru_list lru)
865 struct page_cgroup *pc;
866 struct mem_cgroup_per_zone *mz;
868 if (mem_cgroup_disabled())
870 pc = lookup_page_cgroup(page);
871 VM_BUG_ON(PageCgroupAcctLRU(pc));
872 if (!PageCgroupUsed(pc))
874 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
876 mz = page_cgroup_zoneinfo(pc);
877 /* huge page split is done under lru_lock. so, we have no races. */
878 MEM_CGROUP_ZSTAT(mz, lru) += 1 << compound_order(page);
879 SetPageCgroupAcctLRU(pc);
880 if (mem_cgroup_is_root(pc->mem_cgroup))
882 list_add(&pc->lru, &mz->lists[lru]);
886 * At handling SwapCache, pc->mem_cgroup may be changed while it's linked to
887 * lru because the page may.be reused after it's fully uncharged (because of
888 * SwapCache behavior).To handle that, unlink page_cgroup from LRU when charge
889 * it again. This function is only used to charge SwapCache. It's done under
890 * lock_page and expected that zone->lru_lock is never held.
892 static void mem_cgroup_lru_del_before_commit_swapcache(struct page *page)
895 struct zone *zone = page_zone(page);
896 struct page_cgroup *pc = lookup_page_cgroup(page);
898 spin_lock_irqsave(&zone->lru_lock, flags);
900 * Forget old LRU when this page_cgroup is *not* used. This Used bit
901 * is guarded by lock_page() because the page is SwapCache.
903 if (!PageCgroupUsed(pc))
904 mem_cgroup_del_lru_list(page, page_lru(page));
905 spin_unlock_irqrestore(&zone->lru_lock, flags);
908 static void mem_cgroup_lru_add_after_commit_swapcache(struct page *page)
911 struct zone *zone = page_zone(page);
912 struct page_cgroup *pc = lookup_page_cgroup(page);
914 spin_lock_irqsave(&zone->lru_lock, flags);
915 /* link when the page is linked to LRU but page_cgroup isn't */
916 if (PageLRU(page) && !PageCgroupAcctLRU(pc))
917 mem_cgroup_add_lru_list(page, page_lru(page));
918 spin_unlock_irqrestore(&zone->lru_lock, flags);
922 void mem_cgroup_move_lists(struct page *page,
923 enum lru_list from, enum lru_list to)
925 if (mem_cgroup_disabled())
927 mem_cgroup_del_lru_list(page, from);
928 mem_cgroup_add_lru_list(page, to);
931 int task_in_mem_cgroup(struct task_struct *task, const struct mem_cgroup *mem)
934 struct mem_cgroup *curr = NULL;
935 struct task_struct *p;
937 p = find_lock_task_mm(task);
940 curr = try_get_mem_cgroup_from_mm(p->mm);
945 * We should check use_hierarchy of "mem" not "curr". Because checking
946 * use_hierarchy of "curr" here make this function true if hierarchy is
947 * enabled in "curr" and "curr" is a child of "mem" in *cgroup*
948 * hierarchy(even if use_hierarchy is disabled in "mem").
950 if (mem->use_hierarchy)
951 ret = css_is_ancestor(&curr->css, &mem->css);
958 static int calc_inactive_ratio(struct mem_cgroup *memcg, unsigned long *present_pages)
960 unsigned long active;
961 unsigned long inactive;
963 unsigned long inactive_ratio;
965 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_ANON);
966 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_ANON);
968 gb = (inactive + active) >> (30 - PAGE_SHIFT);
970 inactive_ratio = int_sqrt(10 * gb);
975 present_pages[0] = inactive;
976 present_pages[1] = active;
979 return inactive_ratio;
982 int mem_cgroup_inactive_anon_is_low(struct mem_cgroup *memcg)
984 unsigned long active;
985 unsigned long inactive;
986 unsigned long present_pages[2];
987 unsigned long inactive_ratio;
989 inactive_ratio = calc_inactive_ratio(memcg, present_pages);
991 inactive = present_pages[0];
992 active = present_pages[1];
994 if (inactive * inactive_ratio < active)
1000 int mem_cgroup_inactive_file_is_low(struct mem_cgroup *memcg)
1002 unsigned long active;
1003 unsigned long inactive;
1005 inactive = mem_cgroup_get_local_zonestat(memcg, LRU_INACTIVE_FILE);
1006 active = mem_cgroup_get_local_zonestat(memcg, LRU_ACTIVE_FILE);
1008 return (active > inactive);
1011 unsigned long mem_cgroup_zone_nr_pages(struct mem_cgroup *memcg,
1015 int nid = zone_to_nid(zone);
1016 int zid = zone_idx(zone);
1017 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1019 return MEM_CGROUP_ZSTAT(mz, lru);
1022 struct zone_reclaim_stat *mem_cgroup_get_reclaim_stat(struct mem_cgroup *memcg,
1025 int nid = zone_to_nid(zone);
1026 int zid = zone_idx(zone);
1027 struct mem_cgroup_per_zone *mz = mem_cgroup_zoneinfo(memcg, nid, zid);
1029 return &mz->reclaim_stat;
1032 struct zone_reclaim_stat *
1033 mem_cgroup_get_reclaim_stat_from_page(struct page *page)
1035 struct page_cgroup *pc;
1036 struct mem_cgroup_per_zone *mz;
1038 if (mem_cgroup_disabled())
1041 pc = lookup_page_cgroup(page);
1042 if (!PageCgroupUsed(pc))
1044 /* Ensure pc->mem_cgroup is visible after reading PCG_USED. */
1046 mz = page_cgroup_zoneinfo(pc);
1050 return &mz->reclaim_stat;
1053 unsigned long mem_cgroup_isolate_pages(unsigned long nr_to_scan,
1054 struct list_head *dst,
1055 unsigned long *scanned, int order,
1056 int mode, struct zone *z,
1057 struct mem_cgroup *mem_cont,
1058 int active, int file)
1060 unsigned long nr_taken = 0;
1064 struct list_head *src;
1065 struct page_cgroup *pc, *tmp;
1066 int nid = zone_to_nid(z);
1067 int zid = zone_idx(z);
1068 struct mem_cgroup_per_zone *mz;
1069 int lru = LRU_FILE * file + active;
1073 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
1074 src = &mz->lists[lru];
1077 list_for_each_entry_safe_reverse(pc, tmp, src, lru) {
1078 if (scan >= nr_to_scan)
1082 if (unlikely(!PageCgroupUsed(pc)))
1084 if (unlikely(!PageLRU(page)))
1088 ret = __isolate_lru_page(page, mode, file);
1091 list_move(&page->lru, dst);
1092 mem_cgroup_del_lru(page);
1093 nr_taken += hpage_nr_pages(page);
1096 /* we don't affect global LRU but rotate in our LRU */
1097 mem_cgroup_rotate_lru_list(page, page_lru(page));
1106 trace_mm_vmscan_memcg_isolate(0, nr_to_scan, scan, nr_taken,
1112 #define mem_cgroup_from_res_counter(counter, member) \
1113 container_of(counter, struct mem_cgroup, member)
1116 * mem_cgroup_margin - calculate chargeable space of a memory cgroup
1117 * @mem: the memory cgroup
1119 * Returns the maximum amount of memory @mem can be charged with, in
1122 static unsigned long long mem_cgroup_margin(struct mem_cgroup *mem)
1124 unsigned long long margin;
1126 margin = res_counter_margin(&mem->res);
1127 if (do_swap_account)
1128 margin = min(margin, res_counter_margin(&mem->memsw));
1132 static unsigned int get_swappiness(struct mem_cgroup *memcg)
1134 struct cgroup *cgrp = memcg->css.cgroup;
1135 unsigned int swappiness;
1138 if (cgrp->parent == NULL)
1139 return vm_swappiness;
1141 spin_lock(&memcg->reclaim_param_lock);
1142 swappiness = memcg->swappiness;
1143 spin_unlock(&memcg->reclaim_param_lock);
1148 static void mem_cgroup_start_move(struct mem_cgroup *mem)
1153 spin_lock(&mem->pcp_counter_lock);
1154 for_each_online_cpu(cpu)
1155 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) += 1;
1156 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] += 1;
1157 spin_unlock(&mem->pcp_counter_lock);
1163 static void mem_cgroup_end_move(struct mem_cgroup *mem)
1170 spin_lock(&mem->pcp_counter_lock);
1171 for_each_online_cpu(cpu)
1172 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) -= 1;
1173 mem->nocpu_base.count[MEM_CGROUP_ON_MOVE] -= 1;
1174 spin_unlock(&mem->pcp_counter_lock);
1178 * 2 routines for checking "mem" is under move_account() or not.
1180 * mem_cgroup_stealed() - checking a cgroup is mc.from or not. This is used
1181 * for avoiding race in accounting. If true,
1182 * pc->mem_cgroup may be overwritten.
1184 * mem_cgroup_under_move() - checking a cgroup is mc.from or mc.to or
1185 * under hierarchy of moving cgroups. This is for
1186 * waiting at hith-memory prressure caused by "move".
1189 static bool mem_cgroup_stealed(struct mem_cgroup *mem)
1191 VM_BUG_ON(!rcu_read_lock_held());
1192 return this_cpu_read(mem->stat->count[MEM_CGROUP_ON_MOVE]) > 0;
1195 static bool mem_cgroup_under_move(struct mem_cgroup *mem)
1197 struct mem_cgroup *from;
1198 struct mem_cgroup *to;
1201 * Unlike task_move routines, we access mc.to, mc.from not under
1202 * mutual exclusion by cgroup_mutex. Here, we take spinlock instead.
1204 spin_lock(&mc.lock);
1209 if (from == mem || to == mem
1210 || (mem->use_hierarchy && css_is_ancestor(&from->css, &mem->css))
1211 || (mem->use_hierarchy && css_is_ancestor(&to->css, &mem->css)))
1214 spin_unlock(&mc.lock);
1218 static bool mem_cgroup_wait_acct_move(struct mem_cgroup *mem)
1220 if (mc.moving_task && current != mc.moving_task) {
1221 if (mem_cgroup_under_move(mem)) {
1223 prepare_to_wait(&mc.waitq, &wait, TASK_INTERRUPTIBLE);
1224 /* moving charge context might have finished. */
1227 finish_wait(&mc.waitq, &wait);
1235 * mem_cgroup_print_oom_info: Called from OOM with tasklist_lock held in read mode.
1236 * @memcg: The memory cgroup that went over limit
1237 * @p: Task that is going to be killed
1239 * NOTE: @memcg and @p's mem_cgroup can be different when hierarchy is
1242 void mem_cgroup_print_oom_info(struct mem_cgroup *memcg, struct task_struct *p)
1244 struct cgroup *task_cgrp;
1245 struct cgroup *mem_cgrp;
1247 * Need a buffer in BSS, can't rely on allocations. The code relies
1248 * on the assumption that OOM is serialized for memory controller.
1249 * If this assumption is broken, revisit this code.
1251 static char memcg_name[PATH_MAX];
1260 mem_cgrp = memcg->css.cgroup;
1261 task_cgrp = task_cgroup(p, mem_cgroup_subsys_id);
1263 ret = cgroup_path(task_cgrp, memcg_name, PATH_MAX);
1266 * Unfortunately, we are unable to convert to a useful name
1267 * But we'll still print out the usage information
1274 printk(KERN_INFO "Task in %s killed", memcg_name);
1277 ret = cgroup_path(mem_cgrp, memcg_name, PATH_MAX);
1285 * Continues from above, so we don't need an KERN_ level
1287 printk(KERN_CONT " as a result of limit of %s\n", memcg_name);
1290 printk(KERN_INFO "memory: usage %llukB, limit %llukB, failcnt %llu\n",
1291 res_counter_read_u64(&memcg->res, RES_USAGE) >> 10,
1292 res_counter_read_u64(&memcg->res, RES_LIMIT) >> 10,
1293 res_counter_read_u64(&memcg->res, RES_FAILCNT));
1294 printk(KERN_INFO "memory+swap: usage %llukB, limit %llukB, "
1296 res_counter_read_u64(&memcg->memsw, RES_USAGE) >> 10,
1297 res_counter_read_u64(&memcg->memsw, RES_LIMIT) >> 10,
1298 res_counter_read_u64(&memcg->memsw, RES_FAILCNT));
1302 * This function returns the number of memcg under hierarchy tree. Returns
1303 * 1(self count) if no children.
1305 static int mem_cgroup_count_children(struct mem_cgroup *mem)
1308 struct mem_cgroup *iter;
1310 for_each_mem_cgroup_tree(iter, mem)
1316 * Return the memory (and swap, if configured) limit for a memcg.
1318 u64 mem_cgroup_get_limit(struct mem_cgroup *memcg)
1323 limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
1324 limit += total_swap_pages << PAGE_SHIFT;
1326 memsw = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
1328 * If memsw is finite and limits the amount of swap space available
1329 * to this memcg, return that limit.
1331 return min(limit, memsw);
1335 * Visit the first child (need not be the first child as per the ordering
1336 * of the cgroup list, since we track last_scanned_child) of @mem and use
1337 * that to reclaim free pages from.
1339 static struct mem_cgroup *
1340 mem_cgroup_select_victim(struct mem_cgroup *root_mem)
1342 struct mem_cgroup *ret = NULL;
1343 struct cgroup_subsys_state *css;
1346 if (!root_mem->use_hierarchy) {
1347 css_get(&root_mem->css);
1353 nextid = root_mem->last_scanned_child + 1;
1354 css = css_get_next(&mem_cgroup_subsys, nextid, &root_mem->css,
1356 if (css && css_tryget(css))
1357 ret = container_of(css, struct mem_cgroup, css);
1360 /* Updates scanning parameter */
1361 spin_lock(&root_mem->reclaim_param_lock);
1363 /* this means start scan from ID:1 */
1364 root_mem->last_scanned_child = 0;
1366 root_mem->last_scanned_child = found;
1367 spin_unlock(&root_mem->reclaim_param_lock);
1374 * Scan the hierarchy if needed to reclaim memory. We remember the last child
1375 * we reclaimed from, so that we don't end up penalizing one child extensively
1376 * based on its position in the children list.
1378 * root_mem is the original ancestor that we've been reclaim from.
1380 * We give up and return to the caller when we visit root_mem twice.
1381 * (other groups can be removed while we're walking....)
1383 * If shrink==true, for avoiding to free too much, this returns immedieately.
1385 static int mem_cgroup_hierarchical_reclaim(struct mem_cgroup *root_mem,
1388 unsigned long reclaim_options)
1390 struct mem_cgroup *victim;
1393 bool noswap = reclaim_options & MEM_CGROUP_RECLAIM_NOSWAP;
1394 bool shrink = reclaim_options & MEM_CGROUP_RECLAIM_SHRINK;
1395 bool check_soft = reclaim_options & MEM_CGROUP_RECLAIM_SOFT;
1396 unsigned long excess;
1398 excess = res_counter_soft_limit_excess(&root_mem->res) >> PAGE_SHIFT;
1400 /* If memsw_is_minimum==1, swap-out is of-no-use. */
1401 if (root_mem->memsw_is_minimum)
1405 victim = mem_cgroup_select_victim(root_mem);
1406 if (victim == root_mem) {
1409 drain_all_stock_async();
1412 * If we have not been able to reclaim
1413 * anything, it might because there are
1414 * no reclaimable pages under this hierarchy
1416 if (!check_soft || !total) {
1417 css_put(&victim->css);
1421 * We want to do more targetted reclaim.
1422 * excess >> 2 is not to excessive so as to
1423 * reclaim too much, nor too less that we keep
1424 * coming back to reclaim from this cgroup
1426 if (total >= (excess >> 2) ||
1427 (loop > MEM_CGROUP_MAX_RECLAIM_LOOPS)) {
1428 css_put(&victim->css);
1433 if (!mem_cgroup_local_usage(victim)) {
1434 /* this cgroup's local usage == 0 */
1435 css_put(&victim->css);
1438 /* we use swappiness of local cgroup */
1440 ret = mem_cgroup_shrink_node_zone(victim, gfp_mask,
1441 noswap, get_swappiness(victim), zone);
1443 ret = try_to_free_mem_cgroup_pages(victim, gfp_mask,
1444 noswap, get_swappiness(victim));
1445 css_put(&victim->css);
1447 * At shrinking usage, we can't check we should stop here or
1448 * reclaim more. It's depends on callers. last_scanned_child
1449 * will work enough for keeping fairness under tree.
1455 if (!res_counter_soft_limit_excess(&root_mem->res))
1457 } else if (mem_cgroup_margin(root_mem))
1464 * Check OOM-Killer is already running under our hierarchy.
1465 * If someone is running, return false.
1467 static bool mem_cgroup_oom_lock(struct mem_cgroup *mem)
1469 int x, lock_count = 0;
1470 struct mem_cgroup *iter;
1472 for_each_mem_cgroup_tree(iter, mem) {
1473 x = atomic_inc_return(&iter->oom_lock);
1474 lock_count = max(x, lock_count);
1477 if (lock_count == 1)
1482 static int mem_cgroup_oom_unlock(struct mem_cgroup *mem)
1484 struct mem_cgroup *iter;
1487 * When a new child is created while the hierarchy is under oom,
1488 * mem_cgroup_oom_lock() may not be called. We have to use
1489 * atomic_add_unless() here.
1491 for_each_mem_cgroup_tree(iter, mem)
1492 atomic_add_unless(&iter->oom_lock, -1, 0);
1497 static DEFINE_MUTEX(memcg_oom_mutex);
1498 static DECLARE_WAIT_QUEUE_HEAD(memcg_oom_waitq);
1500 struct oom_wait_info {
1501 struct mem_cgroup *mem;
1505 static int memcg_oom_wake_function(wait_queue_t *wait,
1506 unsigned mode, int sync, void *arg)
1508 struct mem_cgroup *wake_mem = (struct mem_cgroup *)arg;
1509 struct oom_wait_info *oom_wait_info;
1511 oom_wait_info = container_of(wait, struct oom_wait_info, wait);
1513 if (oom_wait_info->mem == wake_mem)
1515 /* if no hierarchy, no match */
1516 if (!oom_wait_info->mem->use_hierarchy || !wake_mem->use_hierarchy)
1519 * Both of oom_wait_info->mem and wake_mem are stable under us.
1520 * Then we can use css_is_ancestor without taking care of RCU.
1522 if (!css_is_ancestor(&oom_wait_info->mem->css, &wake_mem->css) &&
1523 !css_is_ancestor(&wake_mem->css, &oom_wait_info->mem->css))
1527 return autoremove_wake_function(wait, mode, sync, arg);
1530 static void memcg_wakeup_oom(struct mem_cgroup *mem)
1532 /* for filtering, pass "mem" as argument. */
1533 __wake_up(&memcg_oom_waitq, TASK_NORMAL, 0, mem);
1536 static void memcg_oom_recover(struct mem_cgroup *mem)
1538 if (mem && atomic_read(&mem->oom_lock))
1539 memcg_wakeup_oom(mem);
1543 * try to call OOM killer. returns false if we should exit memory-reclaim loop.
1545 bool mem_cgroup_handle_oom(struct mem_cgroup *mem, gfp_t mask)
1547 struct oom_wait_info owait;
1548 bool locked, need_to_kill;
1551 owait.wait.flags = 0;
1552 owait.wait.func = memcg_oom_wake_function;
1553 owait.wait.private = current;
1554 INIT_LIST_HEAD(&owait.wait.task_list);
1555 need_to_kill = true;
1556 /* At first, try to OOM lock hierarchy under mem.*/
1557 mutex_lock(&memcg_oom_mutex);
1558 locked = mem_cgroup_oom_lock(mem);
1560 * Even if signal_pending(), we can't quit charge() loop without
1561 * accounting. So, UNINTERRUPTIBLE is appropriate. But SIGKILL
1562 * under OOM is always welcomed, use TASK_KILLABLE here.
1564 prepare_to_wait(&memcg_oom_waitq, &owait.wait, TASK_KILLABLE);
1565 if (!locked || mem->oom_kill_disable)
1566 need_to_kill = false;
1568 mem_cgroup_oom_notify(mem);
1569 mutex_unlock(&memcg_oom_mutex);
1572 finish_wait(&memcg_oom_waitq, &owait.wait);
1573 mem_cgroup_out_of_memory(mem, mask);
1576 finish_wait(&memcg_oom_waitq, &owait.wait);
1578 mutex_lock(&memcg_oom_mutex);
1579 mem_cgroup_oom_unlock(mem);
1580 memcg_wakeup_oom(mem);
1581 mutex_unlock(&memcg_oom_mutex);
1583 if (test_thread_flag(TIF_MEMDIE) || fatal_signal_pending(current))
1585 /* Give chance to dying process */
1586 schedule_timeout(1);
1591 * Currently used to update mapped file statistics, but the routine can be
1592 * generalized to update other statistics as well.
1594 * Notes: Race condition
1596 * We usually use page_cgroup_lock() for accessing page_cgroup member but
1597 * it tends to be costly. But considering some conditions, we doesn't need
1598 * to do so _always_.
1600 * Considering "charge", lock_page_cgroup() is not required because all
1601 * file-stat operations happen after a page is attached to radix-tree. There
1602 * are no race with "charge".
1604 * Considering "uncharge", we know that memcg doesn't clear pc->mem_cgroup
1605 * at "uncharge" intentionally. So, we always see valid pc->mem_cgroup even
1606 * if there are race with "uncharge". Statistics itself is properly handled
1609 * Considering "move", this is an only case we see a race. To make the race
1610 * small, we check MEM_CGROUP_ON_MOVE percpu value and detect there are
1611 * possibility of race condition. If there is, we take a lock.
1614 void mem_cgroup_update_page_stat(struct page *page,
1615 enum mem_cgroup_page_stat_item idx, int val)
1617 struct mem_cgroup *mem;
1618 struct page_cgroup *pc = lookup_page_cgroup(page);
1619 bool need_unlock = false;
1620 unsigned long uninitialized_var(flags);
1626 mem = pc->mem_cgroup;
1627 if (unlikely(!mem || !PageCgroupUsed(pc)))
1629 /* pc->mem_cgroup is unstable ? */
1630 if (unlikely(mem_cgroup_stealed(mem)) || PageTransHuge(page)) {
1631 /* take a lock against to access pc->mem_cgroup */
1632 move_lock_page_cgroup(pc, &flags);
1634 mem = pc->mem_cgroup;
1635 if (!mem || !PageCgroupUsed(pc))
1640 case MEMCG_NR_FILE_MAPPED:
1642 SetPageCgroupFileMapped(pc);
1643 else if (!page_mapped(page))
1644 ClearPageCgroupFileMapped(pc);
1645 idx = MEM_CGROUP_STAT_FILE_MAPPED;
1651 this_cpu_add(mem->stat->count[idx], val);
1654 if (unlikely(need_unlock))
1655 move_unlock_page_cgroup(pc, &flags);
1659 EXPORT_SYMBOL(mem_cgroup_update_page_stat);
1662 * size of first charge trial. "32" comes from vmscan.c's magic value.
1663 * TODO: maybe necessary to use big numbers in big irons.
1665 #define CHARGE_SIZE (32 * PAGE_SIZE)
1666 struct memcg_stock_pcp {
1667 struct mem_cgroup *cached; /* this never be root cgroup */
1669 struct work_struct work;
1671 static DEFINE_PER_CPU(struct memcg_stock_pcp, memcg_stock);
1672 static atomic_t memcg_drain_count;
1675 * Try to consume stocked charge on this cpu. If success, PAGE_SIZE is consumed
1676 * from local stock and true is returned. If the stock is 0 or charges from a
1677 * cgroup which is not current target, returns false. This stock will be
1680 static bool consume_stock(struct mem_cgroup *mem)
1682 struct memcg_stock_pcp *stock;
1685 stock = &get_cpu_var(memcg_stock);
1686 if (mem == stock->cached && stock->charge)
1687 stock->charge -= PAGE_SIZE;
1688 else /* need to call res_counter_charge */
1690 put_cpu_var(memcg_stock);
1695 * Returns stocks cached in percpu to res_counter and reset cached information.
1697 static void drain_stock(struct memcg_stock_pcp *stock)
1699 struct mem_cgroup *old = stock->cached;
1701 if (stock->charge) {
1702 res_counter_uncharge(&old->res, stock->charge);
1703 if (do_swap_account)
1704 res_counter_uncharge(&old->memsw, stock->charge);
1706 stock->cached = NULL;
1711 * This must be called under preempt disabled or must be called by
1712 * a thread which is pinned to local cpu.
1714 static void drain_local_stock(struct work_struct *dummy)
1716 struct memcg_stock_pcp *stock = &__get_cpu_var(memcg_stock);
1721 * Cache charges(val) which is from res_counter, to local per_cpu area.
1722 * This will be consumed by consume_stock() function, later.
1724 static void refill_stock(struct mem_cgroup *mem, int val)
1726 struct memcg_stock_pcp *stock = &get_cpu_var(memcg_stock);
1728 if (stock->cached != mem) { /* reset if necessary */
1730 stock->cached = mem;
1732 stock->charge += val;
1733 put_cpu_var(memcg_stock);
1737 * Tries to drain stocked charges in other cpus. This function is asynchronous
1738 * and just put a work per cpu for draining localy on each cpu. Caller can
1739 * expects some charges will be back to res_counter later but cannot wait for
1742 static void drain_all_stock_async(void)
1745 /* This function is for scheduling "drain" in asynchronous way.
1746 * The result of "drain" is not directly handled by callers. Then,
1747 * if someone is calling drain, we don't have to call drain more.
1748 * Anyway, WORK_STRUCT_PENDING check in queue_work_on() will catch if
1749 * there is a race. We just do loose check here.
1751 if (atomic_read(&memcg_drain_count))
1753 /* Notify other cpus that system-wide "drain" is running */
1754 atomic_inc(&memcg_drain_count);
1756 for_each_online_cpu(cpu) {
1757 struct memcg_stock_pcp *stock = &per_cpu(memcg_stock, cpu);
1758 schedule_work_on(cpu, &stock->work);
1761 atomic_dec(&memcg_drain_count);
1762 /* We don't wait for flush_work */
1765 /* This is a synchronous drain interface. */
1766 static void drain_all_stock_sync(void)
1768 /* called when force_empty is called */
1769 atomic_inc(&memcg_drain_count);
1770 schedule_on_each_cpu(drain_local_stock);
1771 atomic_dec(&memcg_drain_count);
1775 * This function drains percpu counter value from DEAD cpu and
1776 * move it to local cpu. Note that this function can be preempted.
1778 static void mem_cgroup_drain_pcp_counter(struct mem_cgroup *mem, int cpu)
1782 spin_lock(&mem->pcp_counter_lock);
1783 for (i = 0; i < MEM_CGROUP_STAT_DATA; i++) {
1784 s64 x = per_cpu(mem->stat->count[i], cpu);
1786 per_cpu(mem->stat->count[i], cpu) = 0;
1787 mem->nocpu_base.count[i] += x;
1789 /* need to clear ON_MOVE value, works as a kind of lock. */
1790 per_cpu(mem->stat->count[MEM_CGROUP_ON_MOVE], cpu) = 0;
1791 spin_unlock(&mem->pcp_counter_lock);
1794 static void synchronize_mem_cgroup_on_move(struct mem_cgroup *mem, int cpu)
1796 int idx = MEM_CGROUP_ON_MOVE;
1798 spin_lock(&mem->pcp_counter_lock);
1799 per_cpu(mem->stat->count[idx], cpu) = mem->nocpu_base.count[idx];
1800 spin_unlock(&mem->pcp_counter_lock);
1803 static int __cpuinit memcg_cpu_hotplug_callback(struct notifier_block *nb,
1804 unsigned long action,
1807 int cpu = (unsigned long)hcpu;
1808 struct memcg_stock_pcp *stock;
1809 struct mem_cgroup *iter;
1811 if ((action == CPU_ONLINE)) {
1812 for_each_mem_cgroup_all(iter)
1813 synchronize_mem_cgroup_on_move(iter, cpu);
1817 if ((action != CPU_DEAD) || action != CPU_DEAD_FROZEN)
1820 for_each_mem_cgroup_all(iter)
1821 mem_cgroup_drain_pcp_counter(iter, cpu);
1823 stock = &per_cpu(memcg_stock, cpu);
1829 /* See __mem_cgroup_try_charge() for details */
1831 CHARGE_OK, /* success */
1832 CHARGE_RETRY, /* need to retry but retry is not bad */
1833 CHARGE_NOMEM, /* we can't do more. return -ENOMEM */
1834 CHARGE_WOULDBLOCK, /* GFP_WAIT wasn't set and no enough res. */
1835 CHARGE_OOM_DIE, /* the current is killed because of OOM */
1838 static int __mem_cgroup_do_charge(struct mem_cgroup *mem, gfp_t gfp_mask,
1839 int csize, bool oom_check)
1841 struct mem_cgroup *mem_over_limit;
1842 struct res_counter *fail_res;
1843 unsigned long flags = 0;
1846 ret = res_counter_charge(&mem->res, csize, &fail_res);
1849 if (!do_swap_account)
1851 ret = res_counter_charge(&mem->memsw, csize, &fail_res);
1855 res_counter_uncharge(&mem->res, csize);
1856 mem_over_limit = mem_cgroup_from_res_counter(fail_res, memsw);
1857 flags |= MEM_CGROUP_RECLAIM_NOSWAP;
1859 mem_over_limit = mem_cgroup_from_res_counter(fail_res, res);
1861 * csize can be either a huge page (HPAGE_SIZE), a batch of
1862 * regular pages (CHARGE_SIZE), or a single regular page
1865 * Never reclaim on behalf of optional batching, retry with a
1866 * single page instead.
1868 if (csize == CHARGE_SIZE)
1869 return CHARGE_RETRY;
1871 if (!(gfp_mask & __GFP_WAIT))
1872 return CHARGE_WOULDBLOCK;
1874 ret = mem_cgroup_hierarchical_reclaim(mem_over_limit, NULL,
1876 if (mem_cgroup_margin(mem_over_limit) >= csize)
1877 return CHARGE_RETRY;
1879 * Even though the limit is exceeded at this point, reclaim
1880 * may have been able to free some pages. Retry the charge
1881 * before killing the task.
1883 * Only for regular pages, though: huge pages are rather
1884 * unlikely to succeed so close to the limit, and we fall back
1885 * to regular pages anyway in case of failure.
1887 if (csize == PAGE_SIZE && ret)
1888 return CHARGE_RETRY;
1891 * At task move, charge accounts can be doubly counted. So, it's
1892 * better to wait until the end of task_move if something is going on.
1894 if (mem_cgroup_wait_acct_move(mem_over_limit))
1895 return CHARGE_RETRY;
1897 /* If we don't need to call oom-killer at el, return immediately */
1899 return CHARGE_NOMEM;
1901 if (!mem_cgroup_handle_oom(mem_over_limit, gfp_mask))
1902 return CHARGE_OOM_DIE;
1904 return CHARGE_RETRY;
1908 * Unlike exported interface, "oom" parameter is added. if oom==true,
1909 * oom-killer can be invoked.
1911 static int __mem_cgroup_try_charge(struct mm_struct *mm,
1913 struct mem_cgroup **memcg, bool oom,
1916 int nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
1917 struct mem_cgroup *mem = NULL;
1919 int csize = max(CHARGE_SIZE, (unsigned long) page_size);
1922 * Unlike gloval-vm's OOM-kill, we're not in memory shortage
1923 * in system level. So, allow to go ahead dying process in addition to
1926 if (unlikely(test_thread_flag(TIF_MEMDIE)
1927 || fatal_signal_pending(current)))
1931 * We always charge the cgroup the mm_struct belongs to.
1932 * The mm_struct's mem_cgroup changes on task migration if the
1933 * thread group leader migrates. It's possible that mm is not
1934 * set, if so charge the init_mm (happens for pagecache usage).
1939 if (*memcg) { /* css should be a valid one */
1941 VM_BUG_ON(css_is_removed(&mem->css));
1942 if (mem_cgroup_is_root(mem))
1944 if (page_size == PAGE_SIZE && consume_stock(mem))
1948 struct task_struct *p;
1951 p = rcu_dereference(mm->owner);
1953 * Because we don't have task_lock(), "p" can exit.
1954 * In that case, "mem" can point to root or p can be NULL with
1955 * race with swapoff. Then, we have small risk of mis-accouning.
1956 * But such kind of mis-account by race always happens because
1957 * we don't have cgroup_mutex(). It's overkill and we allo that
1959 * (*) swapoff at el will charge against mm-struct not against
1960 * task-struct. So, mm->owner can be NULL.
1962 mem = mem_cgroup_from_task(p);
1963 if (!mem || mem_cgroup_is_root(mem)) {
1967 if (page_size == PAGE_SIZE && consume_stock(mem)) {
1969 * It seems dagerous to access memcg without css_get().
1970 * But considering how consume_stok works, it's not
1971 * necessary. If consume_stock success, some charges
1972 * from this memcg are cached on this cpu. So, we
1973 * don't need to call css_get()/css_tryget() before
1974 * calling consume_stock().
1979 /* after here, we may be blocked. we need to get refcnt */
1980 if (!css_tryget(&mem->css)) {
1990 /* If killed, bypass charge */
1991 if (fatal_signal_pending(current)) {
1997 if (oom && !nr_oom_retries) {
1999 nr_oom_retries = MEM_CGROUP_RECLAIM_RETRIES;
2002 ret = __mem_cgroup_do_charge(mem, gfp_mask, csize, oom_check);
2007 case CHARGE_RETRY: /* not in OOM situation but retry */
2012 case CHARGE_WOULDBLOCK: /* !__GFP_WAIT */
2015 case CHARGE_NOMEM: /* OOM routine works */
2020 /* If oom, we never return -ENOMEM */
2023 case CHARGE_OOM_DIE: /* Killed by OOM Killer */
2027 } while (ret != CHARGE_OK);
2029 if (csize > page_size)
2030 refill_stock(mem, csize - page_size);
2044 * Somemtimes we have to undo a charge we got by try_charge().
2045 * This function is for that and do uncharge, put css's refcnt.
2046 * gotten by try_charge().
2048 static void __mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2049 unsigned long count)
2051 if (!mem_cgroup_is_root(mem)) {
2052 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
2053 if (do_swap_account)
2054 res_counter_uncharge(&mem->memsw, PAGE_SIZE * count);
2058 static void mem_cgroup_cancel_charge(struct mem_cgroup *mem,
2061 __mem_cgroup_cancel_charge(mem, page_size >> PAGE_SHIFT);
2065 * A helper function to get mem_cgroup from ID. must be called under
2066 * rcu_read_lock(). The caller must check css_is_removed() or some if
2067 * it's concern. (dropping refcnt from swap can be called against removed
2070 static struct mem_cgroup *mem_cgroup_lookup(unsigned short id)
2072 struct cgroup_subsys_state *css;
2074 /* ID 0 is unused ID */
2077 css = css_lookup(&mem_cgroup_subsys, id);
2080 return container_of(css, struct mem_cgroup, css);
2083 struct mem_cgroup *try_get_mem_cgroup_from_page(struct page *page)
2085 struct mem_cgroup *mem = NULL;
2086 struct page_cgroup *pc;
2090 VM_BUG_ON(!PageLocked(page));
2092 pc = lookup_page_cgroup(page);
2093 lock_page_cgroup(pc);
2094 if (PageCgroupUsed(pc)) {
2095 mem = pc->mem_cgroup;
2096 if (mem && !css_tryget(&mem->css))
2098 } else if (PageSwapCache(page)) {
2099 ent.val = page_private(page);
2100 id = lookup_swap_cgroup(ent);
2102 mem = mem_cgroup_lookup(id);
2103 if (mem && !css_tryget(&mem->css))
2107 unlock_page_cgroup(pc);
2111 static void __mem_cgroup_commit_charge(struct mem_cgroup *mem,
2112 struct page_cgroup *pc,
2113 enum charge_type ctype,
2116 int nr_pages = page_size >> PAGE_SHIFT;
2118 lock_page_cgroup(pc);
2119 if (unlikely(PageCgroupUsed(pc))) {
2120 unlock_page_cgroup(pc);
2121 mem_cgroup_cancel_charge(mem, page_size);
2125 * we don't need page_cgroup_lock about tail pages, becase they are not
2126 * accessed by any other context at this point.
2128 pc->mem_cgroup = mem;
2130 * We access a page_cgroup asynchronously without lock_page_cgroup().
2131 * Especially when a page_cgroup is taken from a page, pc->mem_cgroup
2132 * is accessed after testing USED bit. To make pc->mem_cgroup visible
2133 * before USED bit, we need memory barrier here.
2134 * See mem_cgroup_add_lru_list(), etc.
2138 case MEM_CGROUP_CHARGE_TYPE_CACHE:
2139 case MEM_CGROUP_CHARGE_TYPE_SHMEM:
2140 SetPageCgroupCache(pc);
2141 SetPageCgroupUsed(pc);
2143 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2144 ClearPageCgroupCache(pc);
2145 SetPageCgroupUsed(pc);
2151 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), nr_pages);
2152 unlock_page_cgroup(pc);
2154 * "charge_statistics" updated event counter. Then, check it.
2155 * Insert ancestor (and ancestor's ancestors), to softlimit RB-tree.
2156 * if they exceeds softlimit.
2158 memcg_check_events(mem, pc->page);
2161 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2163 #define PCGF_NOCOPY_AT_SPLIT ((1 << PCG_LOCK) | (1 << PCG_MOVE_LOCK) |\
2164 (1 << PCG_ACCT_LRU) | (1 << PCG_MIGRATION))
2166 * Because tail pages are not marked as "used", set it. We're under
2167 * zone->lru_lock, 'splitting on pmd' and compund_lock.
2169 void mem_cgroup_split_huge_fixup(struct page *head, struct page *tail)
2171 struct page_cgroup *head_pc = lookup_page_cgroup(head);
2172 struct page_cgroup *tail_pc = lookup_page_cgroup(tail);
2173 unsigned long flags;
2175 if (mem_cgroup_disabled())
2178 * We have no races with charge/uncharge but will have races with
2179 * page state accounting.
2181 move_lock_page_cgroup(head_pc, &flags);
2183 tail_pc->mem_cgroup = head_pc->mem_cgroup;
2184 smp_wmb(); /* see __commit_charge() */
2185 if (PageCgroupAcctLRU(head_pc)) {
2187 struct mem_cgroup_per_zone *mz;
2190 * LRU flags cannot be copied because we need to add tail
2191 *.page to LRU by generic call and our hook will be called.
2192 * We hold lru_lock, then, reduce counter directly.
2194 lru = page_lru(head);
2195 mz = page_cgroup_zoneinfo(head_pc);
2196 MEM_CGROUP_ZSTAT(mz, lru) -= 1;
2198 tail_pc->flags = head_pc->flags & ~PCGF_NOCOPY_AT_SPLIT;
2199 move_unlock_page_cgroup(head_pc, &flags);
2204 * __mem_cgroup_move_account - move account of the page
2205 * @pc: page_cgroup of the page.
2206 * @from: mem_cgroup which the page is moved from.
2207 * @to: mem_cgroup which the page is moved to. @from != @to.
2208 * @uncharge: whether we should call uncharge and css_put against @from.
2210 * The caller must confirm following.
2211 * - page is not on LRU (isolate_page() is useful.)
2212 * - the pc is locked, used, and ->mem_cgroup points to @from.
2214 * This function doesn't do "charge" nor css_get to new cgroup. It should be
2215 * done by a caller(__mem_cgroup_try_charge would be usefull). If @uncharge is
2216 * true, this function does "uncharge" from old cgroup, but it doesn't if
2217 * @uncharge is false, so a caller should do "uncharge".
2220 static void __mem_cgroup_move_account(struct page_cgroup *pc,
2221 struct mem_cgroup *from, struct mem_cgroup *to, bool uncharge,
2224 int nr_pages = charge_size >> PAGE_SHIFT;
2226 VM_BUG_ON(from == to);
2227 VM_BUG_ON(PageLRU(pc->page));
2228 VM_BUG_ON(!page_is_cgroup_locked(pc));
2229 VM_BUG_ON(!PageCgroupUsed(pc));
2230 VM_BUG_ON(pc->mem_cgroup != from);
2232 if (PageCgroupFileMapped(pc)) {
2233 /* Update mapped_file data for mem_cgroup */
2235 __this_cpu_dec(from->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2236 __this_cpu_inc(to->stat->count[MEM_CGROUP_STAT_FILE_MAPPED]);
2239 mem_cgroup_charge_statistics(from, PageCgroupCache(pc), -nr_pages);
2241 /* This is not "cancel", but cancel_charge does all we need. */
2242 mem_cgroup_cancel_charge(from, charge_size);
2244 /* caller should have done css_get */
2245 pc->mem_cgroup = to;
2246 mem_cgroup_charge_statistics(to, PageCgroupCache(pc), nr_pages);
2248 * We charges against "to" which may not have any tasks. Then, "to"
2249 * can be under rmdir(). But in current implementation, caller of
2250 * this function is just force_empty() and move charge, so it's
2251 * garanteed that "to" is never removed. So, we don't check rmdir
2257 * check whether the @pc is valid for moving account and call
2258 * __mem_cgroup_move_account()
2260 static int mem_cgroup_move_account(struct page_cgroup *pc,
2261 struct mem_cgroup *from, struct mem_cgroup *to,
2262 bool uncharge, int charge_size)
2265 unsigned long flags;
2267 * The page is isolated from LRU. So, collapse function
2268 * will not handle this page. But page splitting can happen.
2269 * Do this check under compound_page_lock(). The caller should
2272 if ((charge_size > PAGE_SIZE) && !PageTransHuge(pc->page))
2275 lock_page_cgroup(pc);
2276 if (PageCgroupUsed(pc) && pc->mem_cgroup == from) {
2277 move_lock_page_cgroup(pc, &flags);
2278 __mem_cgroup_move_account(pc, from, to, uncharge, charge_size);
2279 move_unlock_page_cgroup(pc, &flags);
2282 unlock_page_cgroup(pc);
2286 memcg_check_events(to, pc->page);
2287 memcg_check_events(from, pc->page);
2292 * move charges to its parent.
2295 static int mem_cgroup_move_parent(struct page_cgroup *pc,
2296 struct mem_cgroup *child,
2299 struct page *page = pc->page;
2300 struct cgroup *cg = child->css.cgroup;
2301 struct cgroup *pcg = cg->parent;
2302 struct mem_cgroup *parent;
2303 int page_size = PAGE_SIZE;
2304 unsigned long flags;
2312 if (!get_page_unless_zero(page))
2314 if (isolate_lru_page(page))
2317 if (PageTransHuge(page))
2318 page_size = HPAGE_SIZE;
2320 parent = mem_cgroup_from_cont(pcg);
2321 ret = __mem_cgroup_try_charge(NULL, gfp_mask,
2322 &parent, false, page_size);
2326 if (page_size > PAGE_SIZE)
2327 flags = compound_lock_irqsave(page);
2329 ret = mem_cgroup_move_account(pc, child, parent, true, page_size);
2331 mem_cgroup_cancel_charge(parent, page_size);
2333 if (page_size > PAGE_SIZE)
2334 compound_unlock_irqrestore(page, flags);
2336 putback_lru_page(page);
2344 * Charge the memory controller for page usage.
2346 * 0 if the charge was successful
2347 * < 0 if the cgroup is over its limit
2349 static int mem_cgroup_charge_common(struct page *page, struct mm_struct *mm,
2350 gfp_t gfp_mask, enum charge_type ctype)
2352 struct mem_cgroup *mem = NULL;
2353 int page_size = PAGE_SIZE;
2354 struct page_cgroup *pc;
2358 if (PageTransHuge(page)) {
2359 page_size <<= compound_order(page);
2360 VM_BUG_ON(!PageTransHuge(page));
2362 * Never OOM-kill a process for a huge page. The
2363 * fault handler will fall back to regular pages.
2368 pc = lookup_page_cgroup(page);
2369 BUG_ON(!pc); /* XXX: remove this and move pc lookup into commit */
2371 ret = __mem_cgroup_try_charge(mm, gfp_mask, &mem, oom, page_size);
2375 __mem_cgroup_commit_charge(mem, pc, ctype, page_size);
2379 int mem_cgroup_newpage_charge(struct page *page,
2380 struct mm_struct *mm, gfp_t gfp_mask)
2382 if (mem_cgroup_disabled())
2385 * If already mapped, we don't have to account.
2386 * If page cache, page->mapping has address_space.
2387 * But page->mapping may have out-of-use anon_vma pointer,
2388 * detecit it by PageAnon() check. newly-mapped-anon's page->mapping
2391 if (page_mapped(page) || (page->mapping && !PageAnon(page)))
2395 return mem_cgroup_charge_common(page, mm, gfp_mask,
2396 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2400 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2401 enum charge_type ctype);
2403 int mem_cgroup_cache_charge(struct page *page, struct mm_struct *mm,
2408 if (mem_cgroup_disabled())
2410 if (PageCompound(page))
2413 * Corner case handling. This is called from add_to_page_cache()
2414 * in usual. But some FS (shmem) precharges this page before calling it
2415 * and call add_to_page_cache() with GFP_NOWAIT.
2417 * For GFP_NOWAIT case, the page may be pre-charged before calling
2418 * add_to_page_cache(). (See shmem.c) check it here and avoid to call
2419 * charge twice. (It works but has to pay a bit larger cost.)
2420 * And when the page is SwapCache, it should take swap information
2421 * into account. This is under lock_page() now.
2423 if (!(gfp_mask & __GFP_WAIT)) {
2424 struct page_cgroup *pc;
2426 pc = lookup_page_cgroup(page);
2429 lock_page_cgroup(pc);
2430 if (PageCgroupUsed(pc)) {
2431 unlock_page_cgroup(pc);
2434 unlock_page_cgroup(pc);
2440 if (page_is_file_cache(page))
2441 return mem_cgroup_charge_common(page, mm, gfp_mask,
2442 MEM_CGROUP_CHARGE_TYPE_CACHE);
2445 if (PageSwapCache(page)) {
2446 struct mem_cgroup *mem;
2448 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
2450 __mem_cgroup_commit_charge_swapin(page, mem,
2451 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2453 ret = mem_cgroup_charge_common(page, mm, gfp_mask,
2454 MEM_CGROUP_CHARGE_TYPE_SHMEM);
2460 * While swap-in, try_charge -> commit or cancel, the page is locked.
2461 * And when try_charge() successfully returns, one refcnt to memcg without
2462 * struct page_cgroup is acquired. This refcnt will be consumed by
2463 * "commit()" or removed by "cancel()"
2465 int mem_cgroup_try_charge_swapin(struct mm_struct *mm,
2467 gfp_t mask, struct mem_cgroup **ptr)
2469 struct mem_cgroup *mem;
2474 if (mem_cgroup_disabled())
2477 if (!do_swap_account)
2480 * A racing thread's fault, or swapoff, may have already updated
2481 * the pte, and even removed page from swap cache: in those cases
2482 * do_swap_page()'s pte_same() test will fail; but there's also a
2483 * KSM case which does need to charge the page.
2485 if (!PageSwapCache(page))
2487 mem = try_get_mem_cgroup_from_page(page);
2491 ret = __mem_cgroup_try_charge(NULL, mask, ptr, true, PAGE_SIZE);
2497 return __mem_cgroup_try_charge(mm, mask, ptr, true, PAGE_SIZE);
2501 __mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr,
2502 enum charge_type ctype)
2504 struct page_cgroup *pc;
2506 if (mem_cgroup_disabled())
2510 cgroup_exclude_rmdir(&ptr->css);
2511 pc = lookup_page_cgroup(page);
2512 mem_cgroup_lru_del_before_commit_swapcache(page);
2513 __mem_cgroup_commit_charge(ptr, pc, ctype, PAGE_SIZE);
2514 mem_cgroup_lru_add_after_commit_swapcache(page);
2516 * Now swap is on-memory. This means this page may be
2517 * counted both as mem and swap....double count.
2518 * Fix it by uncharging from memsw. Basically, this SwapCache is stable
2519 * under lock_page(). But in do_swap_page()::memory.c, reuse_swap_page()
2520 * may call delete_from_swap_cache() before reach here.
2522 if (do_swap_account && PageSwapCache(page)) {
2523 swp_entry_t ent = {.val = page_private(page)};
2525 struct mem_cgroup *memcg;
2527 id = swap_cgroup_record(ent, 0);
2529 memcg = mem_cgroup_lookup(id);
2532 * This recorded memcg can be obsolete one. So, avoid
2533 * calling css_tryget
2535 if (!mem_cgroup_is_root(memcg))
2536 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2537 mem_cgroup_swap_statistics(memcg, false);
2538 mem_cgroup_put(memcg);
2543 * At swapin, we may charge account against cgroup which has no tasks.
2544 * So, rmdir()->pre_destroy() can be called while we do this charge.
2545 * In that case, we need to call pre_destroy() again. check it here.
2547 cgroup_release_and_wakeup_rmdir(&ptr->css);
2550 void mem_cgroup_commit_charge_swapin(struct page *page, struct mem_cgroup *ptr)
2552 __mem_cgroup_commit_charge_swapin(page, ptr,
2553 MEM_CGROUP_CHARGE_TYPE_MAPPED);
2556 void mem_cgroup_cancel_charge_swapin(struct mem_cgroup *mem)
2558 if (mem_cgroup_disabled())
2562 mem_cgroup_cancel_charge(mem, PAGE_SIZE);
2566 __do_uncharge(struct mem_cgroup *mem, const enum charge_type ctype,
2569 struct memcg_batch_info *batch = NULL;
2570 bool uncharge_memsw = true;
2571 /* If swapout, usage of swap doesn't decrease */
2572 if (!do_swap_account || ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT)
2573 uncharge_memsw = false;
2575 batch = ¤t->memcg_batch;
2577 * In usual, we do css_get() when we remember memcg pointer.
2578 * But in this case, we keep res->usage until end of a series of
2579 * uncharges. Then, it's ok to ignore memcg's refcnt.
2584 * do_batch > 0 when unmapping pages or inode invalidate/truncate.
2585 * In those cases, all pages freed continously can be expected to be in
2586 * the same cgroup and we have chance to coalesce uncharges.
2587 * But we do uncharge one by one if this is killed by OOM(TIF_MEMDIE)
2588 * because we want to do uncharge as soon as possible.
2591 if (!batch->do_batch || test_thread_flag(TIF_MEMDIE))
2592 goto direct_uncharge;
2594 if (page_size != PAGE_SIZE)
2595 goto direct_uncharge;
2598 * In typical case, batch->memcg == mem. This means we can
2599 * merge a series of uncharges to an uncharge of res_counter.
2600 * If not, we uncharge res_counter ony by one.
2602 if (batch->memcg != mem)
2603 goto direct_uncharge;
2604 /* remember freed charge and uncharge it later */
2605 batch->bytes += PAGE_SIZE;
2607 batch->memsw_bytes += PAGE_SIZE;
2610 res_counter_uncharge(&mem->res, page_size);
2612 res_counter_uncharge(&mem->memsw, page_size);
2613 if (unlikely(batch->memcg != mem))
2614 memcg_oom_recover(mem);
2619 * uncharge if !page_mapped(page)
2621 static struct mem_cgroup *
2622 __mem_cgroup_uncharge_common(struct page *page, enum charge_type ctype)
2625 struct page_cgroup *pc;
2626 struct mem_cgroup *mem = NULL;
2627 int page_size = PAGE_SIZE;
2629 if (mem_cgroup_disabled())
2632 if (PageSwapCache(page))
2635 if (PageTransHuge(page)) {
2636 page_size <<= compound_order(page);
2637 VM_BUG_ON(!PageTransHuge(page));
2640 count = page_size >> PAGE_SHIFT;
2642 * Check if our page_cgroup is valid
2644 pc = lookup_page_cgroup(page);
2645 if (unlikely(!pc || !PageCgroupUsed(pc)))
2648 lock_page_cgroup(pc);
2650 mem = pc->mem_cgroup;
2652 if (!PageCgroupUsed(pc))
2656 case MEM_CGROUP_CHARGE_TYPE_MAPPED:
2657 case MEM_CGROUP_CHARGE_TYPE_DROP:
2658 /* See mem_cgroup_prepare_migration() */
2659 if (page_mapped(page) || PageCgroupMigration(pc))
2662 case MEM_CGROUP_CHARGE_TYPE_SWAPOUT:
2663 if (!PageAnon(page)) { /* Shared memory */
2664 if (page->mapping && !page_is_file_cache(page))
2666 } else if (page_mapped(page)) /* Anon */
2673 mem_cgroup_charge_statistics(mem, PageCgroupCache(pc), -count);
2675 ClearPageCgroupUsed(pc);
2677 * pc->mem_cgroup is not cleared here. It will be accessed when it's
2678 * freed from LRU. This is safe because uncharged page is expected not
2679 * to be reused (freed soon). Exception is SwapCache, it's handled by
2680 * special functions.
2683 unlock_page_cgroup(pc);
2685 * even after unlock, we have mem->res.usage here and this memcg
2686 * will never be freed.
2688 memcg_check_events(mem, page);
2689 if (do_swap_account && ctype == MEM_CGROUP_CHARGE_TYPE_SWAPOUT) {
2690 mem_cgroup_swap_statistics(mem, true);
2691 mem_cgroup_get(mem);
2693 if (!mem_cgroup_is_root(mem))
2694 __do_uncharge(mem, ctype, page_size);
2699 unlock_page_cgroup(pc);
2703 void mem_cgroup_uncharge_page(struct page *page)
2706 if (page_mapped(page))
2708 if (page->mapping && !PageAnon(page))
2710 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_MAPPED);
2713 void mem_cgroup_uncharge_cache_page(struct page *page)
2715 VM_BUG_ON(page_mapped(page));
2716 VM_BUG_ON(page->mapping);
2717 __mem_cgroup_uncharge_common(page, MEM_CGROUP_CHARGE_TYPE_CACHE);
2721 * Batch_start/batch_end is called in unmap_page_range/invlidate/trucate.
2722 * In that cases, pages are freed continuously and we can expect pages
2723 * are in the same memcg. All these calls itself limits the number of
2724 * pages freed at once, then uncharge_start/end() is called properly.
2725 * This may be called prural(2) times in a context,
2728 void mem_cgroup_uncharge_start(void)
2730 current->memcg_batch.do_batch++;
2731 /* We can do nest. */
2732 if (current->memcg_batch.do_batch == 1) {
2733 current->memcg_batch.memcg = NULL;
2734 current->memcg_batch.bytes = 0;
2735 current->memcg_batch.memsw_bytes = 0;
2739 void mem_cgroup_uncharge_end(void)
2741 struct memcg_batch_info *batch = ¤t->memcg_batch;
2743 if (!batch->do_batch)
2747 if (batch->do_batch) /* If stacked, do nothing. */
2753 * This "batch->memcg" is valid without any css_get/put etc...
2754 * bacause we hide charges behind us.
2757 res_counter_uncharge(&batch->memcg->res, batch->bytes);
2758 if (batch->memsw_bytes)
2759 res_counter_uncharge(&batch->memcg->memsw, batch->memsw_bytes);
2760 memcg_oom_recover(batch->memcg);
2761 /* forget this pointer (for sanity check) */
2762 batch->memcg = NULL;
2767 * called after __delete_from_swap_cache() and drop "page" account.
2768 * memcg information is recorded to swap_cgroup of "ent"
2771 mem_cgroup_uncharge_swapcache(struct page *page, swp_entry_t ent, bool swapout)
2773 struct mem_cgroup *memcg;
2774 int ctype = MEM_CGROUP_CHARGE_TYPE_SWAPOUT;
2776 if (!swapout) /* this was a swap cache but the swap is unused ! */
2777 ctype = MEM_CGROUP_CHARGE_TYPE_DROP;
2779 memcg = __mem_cgroup_uncharge_common(page, ctype);
2782 * record memcg information, if swapout && memcg != NULL,
2783 * mem_cgroup_get() was called in uncharge().
2785 if (do_swap_account && swapout && memcg)
2786 swap_cgroup_record(ent, css_id(&memcg->css));
2790 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
2792 * called from swap_entry_free(). remove record in swap_cgroup and
2793 * uncharge "memsw" account.
2795 void mem_cgroup_uncharge_swap(swp_entry_t ent)
2797 struct mem_cgroup *memcg;
2800 if (!do_swap_account)
2803 id = swap_cgroup_record(ent, 0);
2805 memcg = mem_cgroup_lookup(id);
2808 * We uncharge this because swap is freed.
2809 * This memcg can be obsolete one. We avoid calling css_tryget
2811 if (!mem_cgroup_is_root(memcg))
2812 res_counter_uncharge(&memcg->memsw, PAGE_SIZE);
2813 mem_cgroup_swap_statistics(memcg, false);
2814 mem_cgroup_put(memcg);
2820 * mem_cgroup_move_swap_account - move swap charge and swap_cgroup's record.
2821 * @entry: swap entry to be moved
2822 * @from: mem_cgroup which the entry is moved from
2823 * @to: mem_cgroup which the entry is moved to
2824 * @need_fixup: whether we should fixup res_counters and refcounts.
2826 * It succeeds only when the swap_cgroup's record for this entry is the same
2827 * as the mem_cgroup's id of @from.
2829 * Returns 0 on success, -EINVAL on failure.
2831 * The caller must have charged to @to, IOW, called res_counter_charge() about
2832 * both res and memsw, and called css_get().
2834 static int mem_cgroup_move_swap_account(swp_entry_t entry,
2835 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2837 unsigned short old_id, new_id;
2839 old_id = css_id(&from->css);
2840 new_id = css_id(&to->css);
2842 if (swap_cgroup_cmpxchg(entry, old_id, new_id) == old_id) {
2843 mem_cgroup_swap_statistics(from, false);
2844 mem_cgroup_swap_statistics(to, true);
2846 * This function is only called from task migration context now.
2847 * It postpones res_counter and refcount handling till the end
2848 * of task migration(mem_cgroup_clear_mc()) for performance
2849 * improvement. But we cannot postpone mem_cgroup_get(to)
2850 * because if the process that has been moved to @to does
2851 * swap-in, the refcount of @to might be decreased to 0.
2855 if (!mem_cgroup_is_root(from))
2856 res_counter_uncharge(&from->memsw, PAGE_SIZE);
2857 mem_cgroup_put(from);
2859 * we charged both to->res and to->memsw, so we should
2862 if (!mem_cgroup_is_root(to))
2863 res_counter_uncharge(&to->res, PAGE_SIZE);
2870 static inline int mem_cgroup_move_swap_account(swp_entry_t entry,
2871 struct mem_cgroup *from, struct mem_cgroup *to, bool need_fixup)
2878 * Before starting migration, account PAGE_SIZE to mem_cgroup that the old
2881 int mem_cgroup_prepare_migration(struct page *page,
2882 struct page *newpage, struct mem_cgroup **ptr, gfp_t gfp_mask)
2884 struct page_cgroup *pc;
2885 struct mem_cgroup *mem = NULL;
2886 enum charge_type ctype;
2891 VM_BUG_ON(PageTransHuge(page));
2892 if (mem_cgroup_disabled())
2895 pc = lookup_page_cgroup(page);
2896 lock_page_cgroup(pc);
2897 if (PageCgroupUsed(pc)) {
2898 mem = pc->mem_cgroup;
2901 * At migrating an anonymous page, its mapcount goes down
2902 * to 0 and uncharge() will be called. But, even if it's fully
2903 * unmapped, migration may fail and this page has to be
2904 * charged again. We set MIGRATION flag here and delay uncharge
2905 * until end_migration() is called
2907 * Corner Case Thinking
2909 * When the old page was mapped as Anon and it's unmap-and-freed
2910 * while migration was ongoing.
2911 * If unmap finds the old page, uncharge() of it will be delayed
2912 * until end_migration(). If unmap finds a new page, it's
2913 * uncharged when it make mapcount to be 1->0. If unmap code
2914 * finds swap_migration_entry, the new page will not be mapped
2915 * and end_migration() will find it(mapcount==0).
2918 * When the old page was mapped but migraion fails, the kernel
2919 * remaps it. A charge for it is kept by MIGRATION flag even
2920 * if mapcount goes down to 0. We can do remap successfully
2921 * without charging it again.
2924 * The "old" page is under lock_page() until the end of
2925 * migration, so, the old page itself will not be swapped-out.
2926 * If the new page is swapped out before end_migraton, our
2927 * hook to usual swap-out path will catch the event.
2930 SetPageCgroupMigration(pc);
2932 unlock_page_cgroup(pc);
2934 * If the page is not charged at this point,
2941 ret = __mem_cgroup_try_charge(NULL, gfp_mask, ptr, false, PAGE_SIZE);
2942 css_put(&mem->css);/* drop extra refcnt */
2943 if (ret || *ptr == NULL) {
2944 if (PageAnon(page)) {
2945 lock_page_cgroup(pc);
2946 ClearPageCgroupMigration(pc);
2947 unlock_page_cgroup(pc);
2949 * The old page may be fully unmapped while we kept it.
2951 mem_cgroup_uncharge_page(page);
2956 * We charge new page before it's used/mapped. So, even if unlock_page()
2957 * is called before end_migration, we can catch all events on this new
2958 * page. In the case new page is migrated but not remapped, new page's
2959 * mapcount will be finally 0 and we call uncharge in end_migration().
2961 pc = lookup_page_cgroup(newpage);
2963 ctype = MEM_CGROUP_CHARGE_TYPE_MAPPED;
2964 else if (page_is_file_cache(page))
2965 ctype = MEM_CGROUP_CHARGE_TYPE_CACHE;
2967 ctype = MEM_CGROUP_CHARGE_TYPE_SHMEM;
2968 __mem_cgroup_commit_charge(mem, pc, ctype, PAGE_SIZE);
2972 /* remove redundant charge if migration failed*/
2973 void mem_cgroup_end_migration(struct mem_cgroup *mem,
2974 struct page *oldpage, struct page *newpage, bool migration_ok)
2976 struct page *used, *unused;
2977 struct page_cgroup *pc;
2981 /* blocks rmdir() */
2982 cgroup_exclude_rmdir(&mem->css);
2983 if (!migration_ok) {
2991 * We disallowed uncharge of pages under migration because mapcount
2992 * of the page goes down to zero, temporarly.
2993 * Clear the flag and check the page should be charged.
2995 pc = lookup_page_cgroup(oldpage);
2996 lock_page_cgroup(pc);
2997 ClearPageCgroupMigration(pc);
2998 unlock_page_cgroup(pc);
3000 __mem_cgroup_uncharge_common(unused, MEM_CGROUP_CHARGE_TYPE_FORCE);
3003 * If a page is a file cache, radix-tree replacement is very atomic
3004 * and we can skip this check. When it was an Anon page, its mapcount
3005 * goes down to 0. But because we added MIGRATION flage, it's not
3006 * uncharged yet. There are several case but page->mapcount check
3007 * and USED bit check in mem_cgroup_uncharge_page() will do enough
3008 * check. (see prepare_charge() also)
3011 mem_cgroup_uncharge_page(used);
3013 * At migration, we may charge account against cgroup which has no
3015 * So, rmdir()->pre_destroy() can be called while we do this charge.
3016 * In that case, we need to call pre_destroy() again. check it here.
3018 cgroup_release_and_wakeup_rmdir(&mem->css);
3022 * A call to try to shrink memory usage on charge failure at shmem's swapin.
3023 * Calling hierarchical_reclaim is not enough because we should update
3024 * last_oom_jiffies to prevent pagefault_out_of_memory from invoking global OOM.
3025 * Moreover considering hierarchy, we should reclaim from the mem_over_limit,
3026 * not from the memcg which this page would be charged to.
3027 * try_charge_swapin does all of these works properly.
3029 int mem_cgroup_shmem_charge_fallback(struct page *page,
3030 struct mm_struct *mm,
3033 struct mem_cgroup *mem;
3036 if (mem_cgroup_disabled())
3039 ret = mem_cgroup_try_charge_swapin(mm, page, gfp_mask, &mem);
3041 mem_cgroup_cancel_charge_swapin(mem); /* it does !mem check */
3046 #ifdef CONFIG_DEBUG_VM
3047 static struct page_cgroup *lookup_page_cgroup_used(struct page *page)
3049 struct page_cgroup *pc;
3051 pc = lookup_page_cgroup(page);
3052 if (likely(pc) && PageCgroupUsed(pc))
3057 bool mem_cgroup_bad_page_check(struct page *page)
3059 if (mem_cgroup_disabled())
3062 return lookup_page_cgroup_used(page) != NULL;
3065 void mem_cgroup_print_bad_page(struct page *page)
3067 struct page_cgroup *pc;
3069 pc = lookup_page_cgroup_used(page);
3074 printk(KERN_ALERT "pc:%p pc->flags:%lx pc->mem_cgroup:%p",
3075 pc, pc->flags, pc->mem_cgroup);
3077 path = kmalloc(PATH_MAX, GFP_KERNEL);
3080 ret = cgroup_path(pc->mem_cgroup->css.cgroup,
3085 printk(KERN_CONT "(%s)\n",
3086 (ret < 0) ? "cannot get the path" : path);
3092 static DEFINE_MUTEX(set_limit_mutex);
3094 static int mem_cgroup_resize_limit(struct mem_cgroup *memcg,
3095 unsigned long long val)
3098 u64 memswlimit, memlimit;
3100 int children = mem_cgroup_count_children(memcg);
3101 u64 curusage, oldusage;
3105 * For keeping hierarchical_reclaim simple, how long we should retry
3106 * is depends on callers. We set our retry-count to be function
3107 * of # of children which we should visit in this loop.
3109 retry_count = MEM_CGROUP_RECLAIM_RETRIES * children;
3111 oldusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3114 while (retry_count) {
3115 if (signal_pending(current)) {
3120 * Rather than hide all in some function, I do this in
3121 * open coded manner. You see what this really does.
3122 * We have to guarantee mem->res.limit < mem->memsw.limit.
3124 mutex_lock(&set_limit_mutex);
3125 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3126 if (memswlimit < val) {
3128 mutex_unlock(&set_limit_mutex);
3132 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3136 ret = res_counter_set_limit(&memcg->res, val);
3138 if (memswlimit == val)
3139 memcg->memsw_is_minimum = true;
3141 memcg->memsw_is_minimum = false;
3143 mutex_unlock(&set_limit_mutex);
3148 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3149 MEM_CGROUP_RECLAIM_SHRINK);
3150 curusage = res_counter_read_u64(&memcg->res, RES_USAGE);
3151 /* Usage is reduced ? */
3152 if (curusage >= oldusage)
3155 oldusage = curusage;
3157 if (!ret && enlarge)
3158 memcg_oom_recover(memcg);
3163 static int mem_cgroup_resize_memsw_limit(struct mem_cgroup *memcg,
3164 unsigned long long val)
3167 u64 memlimit, memswlimit, oldusage, curusage;
3168 int children = mem_cgroup_count_children(memcg);
3172 /* see mem_cgroup_resize_res_limit */
3173 retry_count = children * MEM_CGROUP_RECLAIM_RETRIES;
3174 oldusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3175 while (retry_count) {
3176 if (signal_pending(current)) {
3181 * Rather than hide all in some function, I do this in
3182 * open coded manner. You see what this really does.
3183 * We have to guarantee mem->res.limit < mem->memsw.limit.
3185 mutex_lock(&set_limit_mutex);
3186 memlimit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3187 if (memlimit > val) {
3189 mutex_unlock(&set_limit_mutex);
3192 memswlimit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3193 if (memswlimit < val)
3195 ret = res_counter_set_limit(&memcg->memsw, val);
3197 if (memlimit == val)
3198 memcg->memsw_is_minimum = true;
3200 memcg->memsw_is_minimum = false;
3202 mutex_unlock(&set_limit_mutex);
3207 mem_cgroup_hierarchical_reclaim(memcg, NULL, GFP_KERNEL,
3208 MEM_CGROUP_RECLAIM_NOSWAP |
3209 MEM_CGROUP_RECLAIM_SHRINK);
3210 curusage = res_counter_read_u64(&memcg->memsw, RES_USAGE);
3211 /* Usage is reduced ? */
3212 if (curusage >= oldusage)
3215 oldusage = curusage;
3217 if (!ret && enlarge)
3218 memcg_oom_recover(memcg);
3222 unsigned long mem_cgroup_soft_limit_reclaim(struct zone *zone, int order,
3225 unsigned long nr_reclaimed = 0;
3226 struct mem_cgroup_per_zone *mz, *next_mz = NULL;
3227 unsigned long reclaimed;
3229 struct mem_cgroup_tree_per_zone *mctz;
3230 unsigned long long excess;
3235 mctz = soft_limit_tree_node_zone(zone_to_nid(zone), zone_idx(zone));
3237 * This loop can run a while, specially if mem_cgroup's continuously
3238 * keep exceeding their soft limit and putting the system under
3245 mz = mem_cgroup_largest_soft_limit_node(mctz);
3249 reclaimed = mem_cgroup_hierarchical_reclaim(mz->mem, zone,
3251 MEM_CGROUP_RECLAIM_SOFT);
3252 nr_reclaimed += reclaimed;
3253 spin_lock(&mctz->lock);
3256 * If we failed to reclaim anything from this memory cgroup
3257 * it is time to move on to the next cgroup
3263 * Loop until we find yet another one.
3265 * By the time we get the soft_limit lock
3266 * again, someone might have aded the
3267 * group back on the RB tree. Iterate to
3268 * make sure we get a different mem.
3269 * mem_cgroup_largest_soft_limit_node returns
3270 * NULL if no other cgroup is present on
3274 __mem_cgroup_largest_soft_limit_node(mctz);
3275 if (next_mz == mz) {
3276 css_put(&next_mz->mem->css);
3278 } else /* next_mz == NULL or other memcg */
3282 __mem_cgroup_remove_exceeded(mz->mem, mz, mctz);
3283 excess = res_counter_soft_limit_excess(&mz->mem->res);
3285 * One school of thought says that we should not add
3286 * back the node to the tree if reclaim returns 0.
3287 * But our reclaim could return 0, simply because due
3288 * to priority we are exposing a smaller subset of
3289 * memory to reclaim from. Consider this as a longer
3292 /* If excess == 0, no tree ops */
3293 __mem_cgroup_insert_exceeded(mz->mem, mz, mctz, excess);
3294 spin_unlock(&mctz->lock);
3295 css_put(&mz->mem->css);
3298 * Could not reclaim anything and there are no more
3299 * mem cgroups to try or we seem to be looping without
3300 * reclaiming anything.
3302 if (!nr_reclaimed &&
3304 loop > MEM_CGROUP_MAX_SOFT_LIMIT_RECLAIM_LOOPS))
3306 } while (!nr_reclaimed);
3308 css_put(&next_mz->mem->css);
3309 return nr_reclaimed;
3313 * This routine traverse page_cgroup in given list and drop them all.
3314 * *And* this routine doesn't reclaim page itself, just removes page_cgroup.
3316 static int mem_cgroup_force_empty_list(struct mem_cgroup *mem,
3317 int node, int zid, enum lru_list lru)
3320 struct mem_cgroup_per_zone *mz;
3321 struct page_cgroup *pc, *busy;
3322 unsigned long flags, loop;
3323 struct list_head *list;
3326 zone = &NODE_DATA(node)->node_zones[zid];
3327 mz = mem_cgroup_zoneinfo(mem, node, zid);
3328 list = &mz->lists[lru];
3330 loop = MEM_CGROUP_ZSTAT(mz, lru);
3331 /* give some margin against EBUSY etc...*/
3336 spin_lock_irqsave(&zone->lru_lock, flags);
3337 if (list_empty(list)) {
3338 spin_unlock_irqrestore(&zone->lru_lock, flags);
3341 pc = list_entry(list->prev, struct page_cgroup, lru);
3343 list_move(&pc->lru, list);
3345 spin_unlock_irqrestore(&zone->lru_lock, flags);
3348 spin_unlock_irqrestore(&zone->lru_lock, flags);
3350 ret = mem_cgroup_move_parent(pc, mem, GFP_KERNEL);
3354 if (ret == -EBUSY || ret == -EINVAL) {
3355 /* found lock contention or "pc" is obsolete. */
3362 if (!ret && !list_empty(list))
3368 * make mem_cgroup's charge to be 0 if there is no task.
3369 * This enables deleting this mem_cgroup.
3371 static int mem_cgroup_force_empty(struct mem_cgroup *mem, bool free_all)
3374 int node, zid, shrink;
3375 int nr_retries = MEM_CGROUP_RECLAIM_RETRIES;
3376 struct cgroup *cgrp = mem->css.cgroup;
3381 /* should free all ? */
3387 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children))
3390 if (signal_pending(current))
3392 /* This is for making all *used* pages to be on LRU. */
3393 lru_add_drain_all();
3394 drain_all_stock_sync();
3396 mem_cgroup_start_move(mem);
3397 for_each_node_state(node, N_HIGH_MEMORY) {
3398 for (zid = 0; !ret && zid < MAX_NR_ZONES; zid++) {
3401 ret = mem_cgroup_force_empty_list(mem,
3410 mem_cgroup_end_move(mem);
3411 memcg_oom_recover(mem);
3412 /* it seems parent cgroup doesn't have enough mem */
3416 /* "ret" should also be checked to ensure all lists are empty. */
3417 } while (mem->res.usage > 0 || ret);
3423 /* returns EBUSY if there is a task or if we come here twice. */
3424 if (cgroup_task_count(cgrp) || !list_empty(&cgrp->children) || shrink) {
3428 /* we call try-to-free pages for make this cgroup empty */
3429 lru_add_drain_all();
3430 /* try to free all pages in this cgroup */
3432 while (nr_retries && mem->res.usage > 0) {
3435 if (signal_pending(current)) {
3439 progress = try_to_free_mem_cgroup_pages(mem, GFP_KERNEL,
3440 false, get_swappiness(mem));
3443 /* maybe some writeback is necessary */
3444 congestion_wait(BLK_RW_ASYNC, HZ/10);
3449 /* try move_account...there may be some *locked* pages. */
3453 int mem_cgroup_force_empty_write(struct cgroup *cont, unsigned int event)
3455 return mem_cgroup_force_empty(mem_cgroup_from_cont(cont), true);
3459 static u64 mem_cgroup_hierarchy_read(struct cgroup *cont, struct cftype *cft)
3461 return mem_cgroup_from_cont(cont)->use_hierarchy;
3464 static int mem_cgroup_hierarchy_write(struct cgroup *cont, struct cftype *cft,
3468 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3469 struct cgroup *parent = cont->parent;
3470 struct mem_cgroup *parent_mem = NULL;
3473 parent_mem = mem_cgroup_from_cont(parent);
3477 * If parent's use_hierarchy is set, we can't make any modifications
3478 * in the child subtrees. If it is unset, then the change can
3479 * occur, provided the current cgroup has no children.
3481 * For the root cgroup, parent_mem is NULL, we allow value to be
3482 * set if there are no children.
3484 if ((!parent_mem || !parent_mem->use_hierarchy) &&
3485 (val == 1 || val == 0)) {
3486 if (list_empty(&cont->children))
3487 mem->use_hierarchy = val;
3498 static u64 mem_cgroup_get_recursive_idx_stat(struct mem_cgroup *mem,
3499 enum mem_cgroup_stat_index idx)
3501 struct mem_cgroup *iter;
3504 /* each per cpu's value can be minus.Then, use s64 */
3505 for_each_mem_cgroup_tree(iter, mem)
3506 val += mem_cgroup_read_stat(iter, idx);
3508 if (val < 0) /* race ? */
3513 static inline u64 mem_cgroup_usage(struct mem_cgroup *mem, bool swap)
3517 if (!mem_cgroup_is_root(mem)) {
3519 return res_counter_read_u64(&mem->res, RES_USAGE);
3521 return res_counter_read_u64(&mem->memsw, RES_USAGE);
3524 val = mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_CACHE);
3525 val += mem_cgroup_get_recursive_idx_stat(mem, MEM_CGROUP_STAT_RSS);
3528 val += mem_cgroup_get_recursive_idx_stat(mem,
3529 MEM_CGROUP_STAT_SWAPOUT);
3531 return val << PAGE_SHIFT;
3534 static u64 mem_cgroup_read(struct cgroup *cont, struct cftype *cft)
3536 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
3540 type = MEMFILE_TYPE(cft->private);
3541 name = MEMFILE_ATTR(cft->private);
3544 if (name == RES_USAGE)
3545 val = mem_cgroup_usage(mem, false);
3547 val = res_counter_read_u64(&mem->res, name);
3550 if (name == RES_USAGE)
3551 val = mem_cgroup_usage(mem, true);
3553 val = res_counter_read_u64(&mem->memsw, name);
3562 * The user of this function is...
3565 static int mem_cgroup_write(struct cgroup *cont, struct cftype *cft,
3568 struct mem_cgroup *memcg = mem_cgroup_from_cont(cont);
3570 unsigned long long val;
3573 type = MEMFILE_TYPE(cft->private);
3574 name = MEMFILE_ATTR(cft->private);
3577 if (mem_cgroup_is_root(memcg)) { /* Can't set limit on root */
3581 /* This function does all necessary parse...reuse it */
3582 ret = res_counter_memparse_write_strategy(buffer, &val);
3586 ret = mem_cgroup_resize_limit(memcg, val);
3588 ret = mem_cgroup_resize_memsw_limit(memcg, val);
3590 case RES_SOFT_LIMIT:
3591 ret = res_counter_memparse_write_strategy(buffer, &val);
3595 * For memsw, soft limits are hard to implement in terms
3596 * of semantics, for now, we support soft limits for
3597 * control without swap
3600 ret = res_counter_set_soft_limit(&memcg->res, val);
3605 ret = -EINVAL; /* should be BUG() ? */
3611 static void memcg_get_hierarchical_limit(struct mem_cgroup *memcg,
3612 unsigned long long *mem_limit, unsigned long long *memsw_limit)
3614 struct cgroup *cgroup;
3615 unsigned long long min_limit, min_memsw_limit, tmp;
3617 min_limit = res_counter_read_u64(&memcg->res, RES_LIMIT);
3618 min_memsw_limit = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3619 cgroup = memcg->css.cgroup;
3620 if (!memcg->use_hierarchy)
3623 while (cgroup->parent) {
3624 cgroup = cgroup->parent;
3625 memcg = mem_cgroup_from_cont(cgroup);
3626 if (!memcg->use_hierarchy)
3628 tmp = res_counter_read_u64(&memcg->res, RES_LIMIT);
3629 min_limit = min(min_limit, tmp);
3630 tmp = res_counter_read_u64(&memcg->memsw, RES_LIMIT);
3631 min_memsw_limit = min(min_memsw_limit, tmp);
3634 *mem_limit = min_limit;
3635 *memsw_limit = min_memsw_limit;
3639 static int mem_cgroup_reset(struct cgroup *cont, unsigned int event)
3641 struct mem_cgroup *mem;
3644 mem = mem_cgroup_from_cont(cont);
3645 type = MEMFILE_TYPE(event);
3646 name = MEMFILE_ATTR(event);
3650 res_counter_reset_max(&mem->res);
3652 res_counter_reset_max(&mem->memsw);
3656 res_counter_reset_failcnt(&mem->res);
3658 res_counter_reset_failcnt(&mem->memsw);
3665 static u64 mem_cgroup_move_charge_read(struct cgroup *cgrp,
3668 return mem_cgroup_from_cont(cgrp)->move_charge_at_immigrate;
3672 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3673 struct cftype *cft, u64 val)
3675 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
3677 if (val >= (1 << NR_MOVE_TYPE))
3680 * We check this value several times in both in can_attach() and
3681 * attach(), so we need cgroup lock to prevent this value from being
3685 mem->move_charge_at_immigrate = val;
3691 static int mem_cgroup_move_charge_write(struct cgroup *cgrp,
3692 struct cftype *cft, u64 val)
3699 /* For read statistics */
3715 struct mcs_total_stat {
3716 s64 stat[NR_MCS_STAT];
3722 } memcg_stat_strings[NR_MCS_STAT] = {
3723 {"cache", "total_cache"},
3724 {"rss", "total_rss"},
3725 {"mapped_file", "total_mapped_file"},
3726 {"pgpgin", "total_pgpgin"},
3727 {"pgpgout", "total_pgpgout"},
3728 {"swap", "total_swap"},
3729 {"inactive_anon", "total_inactive_anon"},
3730 {"active_anon", "total_active_anon"},
3731 {"inactive_file", "total_inactive_file"},
3732 {"active_file", "total_active_file"},
3733 {"unevictable", "total_unevictable"}
3738 mem_cgroup_get_local_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3743 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_CACHE);
3744 s->stat[MCS_CACHE] += val * PAGE_SIZE;
3745 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_RSS);
3746 s->stat[MCS_RSS] += val * PAGE_SIZE;
3747 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_FILE_MAPPED);
3748 s->stat[MCS_FILE_MAPPED] += val * PAGE_SIZE;
3749 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGIN_COUNT);
3750 s->stat[MCS_PGPGIN] += val;
3751 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_PGPGOUT_COUNT);
3752 s->stat[MCS_PGPGOUT] += val;
3753 if (do_swap_account) {
3754 val = mem_cgroup_read_stat(mem, MEM_CGROUP_STAT_SWAPOUT);
3755 s->stat[MCS_SWAP] += val * PAGE_SIZE;
3759 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_ANON);
3760 s->stat[MCS_INACTIVE_ANON] += val * PAGE_SIZE;
3761 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_ANON);
3762 s->stat[MCS_ACTIVE_ANON] += val * PAGE_SIZE;
3763 val = mem_cgroup_get_local_zonestat(mem, LRU_INACTIVE_FILE);
3764 s->stat[MCS_INACTIVE_FILE] += val * PAGE_SIZE;
3765 val = mem_cgroup_get_local_zonestat(mem, LRU_ACTIVE_FILE);
3766 s->stat[MCS_ACTIVE_FILE] += val * PAGE_SIZE;
3767 val = mem_cgroup_get_local_zonestat(mem, LRU_UNEVICTABLE);
3768 s->stat[MCS_UNEVICTABLE] += val * PAGE_SIZE;
3772 mem_cgroup_get_total_stat(struct mem_cgroup *mem, struct mcs_total_stat *s)
3774 struct mem_cgroup *iter;
3776 for_each_mem_cgroup_tree(iter, mem)
3777 mem_cgroup_get_local_stat(iter, s);
3780 static int mem_control_stat_show(struct cgroup *cont, struct cftype *cft,
3781 struct cgroup_map_cb *cb)
3783 struct mem_cgroup *mem_cont = mem_cgroup_from_cont(cont);
3784 struct mcs_total_stat mystat;
3787 memset(&mystat, 0, sizeof(mystat));
3788 mem_cgroup_get_local_stat(mem_cont, &mystat);
3790 for (i = 0; i < NR_MCS_STAT; i++) {
3791 if (i == MCS_SWAP && !do_swap_account)
3793 cb->fill(cb, memcg_stat_strings[i].local_name, mystat.stat[i]);
3796 /* Hierarchical information */
3798 unsigned long long limit, memsw_limit;
3799 memcg_get_hierarchical_limit(mem_cont, &limit, &memsw_limit);
3800 cb->fill(cb, "hierarchical_memory_limit", limit);
3801 if (do_swap_account)
3802 cb->fill(cb, "hierarchical_memsw_limit", memsw_limit);
3805 memset(&mystat, 0, sizeof(mystat));
3806 mem_cgroup_get_total_stat(mem_cont, &mystat);
3807 for (i = 0; i < NR_MCS_STAT; i++) {
3808 if (i == MCS_SWAP && !do_swap_account)
3810 cb->fill(cb, memcg_stat_strings[i].total_name, mystat.stat[i]);
3813 #ifdef CONFIG_DEBUG_VM
3814 cb->fill(cb, "inactive_ratio", calc_inactive_ratio(mem_cont, NULL));
3818 struct mem_cgroup_per_zone *mz;
3819 unsigned long recent_rotated[2] = {0, 0};
3820 unsigned long recent_scanned[2] = {0, 0};
3822 for_each_online_node(nid)
3823 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
3824 mz = mem_cgroup_zoneinfo(mem_cont, nid, zid);
3826 recent_rotated[0] +=
3827 mz->reclaim_stat.recent_rotated[0];
3828 recent_rotated[1] +=
3829 mz->reclaim_stat.recent_rotated[1];
3830 recent_scanned[0] +=
3831 mz->reclaim_stat.recent_scanned[0];
3832 recent_scanned[1] +=
3833 mz->reclaim_stat.recent_scanned[1];
3835 cb->fill(cb, "recent_rotated_anon", recent_rotated[0]);
3836 cb->fill(cb, "recent_rotated_file", recent_rotated[1]);
3837 cb->fill(cb, "recent_scanned_anon", recent_scanned[0]);
3838 cb->fill(cb, "recent_scanned_file", recent_scanned[1]);
3845 static u64 mem_cgroup_swappiness_read(struct cgroup *cgrp, struct cftype *cft)
3847 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3849 return get_swappiness(memcg);
3852 static int mem_cgroup_swappiness_write(struct cgroup *cgrp, struct cftype *cft,
3855 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3856 struct mem_cgroup *parent;
3861 if (cgrp->parent == NULL)
3864 parent = mem_cgroup_from_cont(cgrp->parent);
3868 /* If under hierarchy, only empty-root can set this value */
3869 if ((parent->use_hierarchy) ||
3870 (memcg->use_hierarchy && !list_empty(&cgrp->children))) {
3875 spin_lock(&memcg->reclaim_param_lock);
3876 memcg->swappiness = val;
3877 spin_unlock(&memcg->reclaim_param_lock);
3884 static void __mem_cgroup_threshold(struct mem_cgroup *memcg, bool swap)
3886 struct mem_cgroup_threshold_ary *t;
3892 t = rcu_dereference(memcg->thresholds.primary);
3894 t = rcu_dereference(memcg->memsw_thresholds.primary);
3899 usage = mem_cgroup_usage(memcg, swap);
3902 * current_threshold points to threshold just below usage.
3903 * If it's not true, a threshold was crossed after last
3904 * call of __mem_cgroup_threshold().
3906 i = t->current_threshold;
3909 * Iterate backward over array of thresholds starting from
3910 * current_threshold and check if a threshold is crossed.
3911 * If none of thresholds below usage is crossed, we read
3912 * only one element of the array here.
3914 for (; i >= 0 && unlikely(t->entries[i].threshold > usage); i--)
3915 eventfd_signal(t->entries[i].eventfd, 1);
3917 /* i = current_threshold + 1 */
3921 * Iterate forward over array of thresholds starting from
3922 * current_threshold+1 and check if a threshold is crossed.
3923 * If none of thresholds above usage is crossed, we read
3924 * only one element of the array here.
3926 for (; i < t->size && unlikely(t->entries[i].threshold <= usage); i++)
3927 eventfd_signal(t->entries[i].eventfd, 1);
3929 /* Update current_threshold */
3930 t->current_threshold = i - 1;
3935 static void mem_cgroup_threshold(struct mem_cgroup *memcg)
3938 __mem_cgroup_threshold(memcg, false);
3939 if (do_swap_account)
3940 __mem_cgroup_threshold(memcg, true);
3942 memcg = parent_mem_cgroup(memcg);
3946 static int compare_thresholds(const void *a, const void *b)
3948 const struct mem_cgroup_threshold *_a = a;
3949 const struct mem_cgroup_threshold *_b = b;
3951 return _a->threshold - _b->threshold;
3954 static int mem_cgroup_oom_notify_cb(struct mem_cgroup *mem)
3956 struct mem_cgroup_eventfd_list *ev;
3958 list_for_each_entry(ev, &mem->oom_notify, list)
3959 eventfd_signal(ev->eventfd, 1);
3963 static void mem_cgroup_oom_notify(struct mem_cgroup *mem)
3965 struct mem_cgroup *iter;
3967 for_each_mem_cgroup_tree(iter, mem)
3968 mem_cgroup_oom_notify_cb(iter);
3971 static int mem_cgroup_usage_register_event(struct cgroup *cgrp,
3972 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
3974 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
3975 struct mem_cgroup_thresholds *thresholds;
3976 struct mem_cgroup_threshold_ary *new;
3977 int type = MEMFILE_TYPE(cft->private);
3978 u64 threshold, usage;
3981 ret = res_counter_memparse_write_strategy(args, &threshold);
3985 mutex_lock(&memcg->thresholds_lock);
3988 thresholds = &memcg->thresholds;
3989 else if (type == _MEMSWAP)
3990 thresholds = &memcg->memsw_thresholds;
3994 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
3996 /* Check if a threshold crossed before adding a new one */
3997 if (thresholds->primary)
3998 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4000 size = thresholds->primary ? thresholds->primary->size + 1 : 1;
4002 /* Allocate memory for new array of thresholds */
4003 new = kmalloc(sizeof(*new) + size * sizeof(struct mem_cgroup_threshold),
4011 /* Copy thresholds (if any) to new array */
4012 if (thresholds->primary) {
4013 memcpy(new->entries, thresholds->primary->entries, (size - 1) *
4014 sizeof(struct mem_cgroup_threshold));
4017 /* Add new threshold */
4018 new->entries[size - 1].eventfd = eventfd;
4019 new->entries[size - 1].threshold = threshold;
4021 /* Sort thresholds. Registering of new threshold isn't time-critical */
4022 sort(new->entries, size, sizeof(struct mem_cgroup_threshold),
4023 compare_thresholds, NULL);
4025 /* Find current threshold */
4026 new->current_threshold = -1;
4027 for (i = 0; i < size; i++) {
4028 if (new->entries[i].threshold < usage) {
4030 * new->current_threshold will not be used until
4031 * rcu_assign_pointer(), so it's safe to increment
4034 ++new->current_threshold;
4038 /* Free old spare buffer and save old primary buffer as spare */
4039 kfree(thresholds->spare);
4040 thresholds->spare = thresholds->primary;
4042 rcu_assign_pointer(thresholds->primary, new);
4044 /* To be sure that nobody uses thresholds */
4048 mutex_unlock(&memcg->thresholds_lock);
4053 static void mem_cgroup_usage_unregister_event(struct cgroup *cgrp,
4054 struct cftype *cft, struct eventfd_ctx *eventfd)
4056 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4057 struct mem_cgroup_thresholds *thresholds;
4058 struct mem_cgroup_threshold_ary *new;
4059 int type = MEMFILE_TYPE(cft->private);
4063 mutex_lock(&memcg->thresholds_lock);
4065 thresholds = &memcg->thresholds;
4066 else if (type == _MEMSWAP)
4067 thresholds = &memcg->memsw_thresholds;
4072 * Something went wrong if we trying to unregister a threshold
4073 * if we don't have thresholds
4075 BUG_ON(!thresholds);
4077 usage = mem_cgroup_usage(memcg, type == _MEMSWAP);
4079 /* Check if a threshold crossed before removing */
4080 __mem_cgroup_threshold(memcg, type == _MEMSWAP);
4082 /* Calculate new number of threshold */
4084 for (i = 0; i < thresholds->primary->size; i++) {
4085 if (thresholds->primary->entries[i].eventfd != eventfd)
4089 new = thresholds->spare;
4091 /* Set thresholds array to NULL if we don't have thresholds */
4100 /* Copy thresholds and find current threshold */
4101 new->current_threshold = -1;
4102 for (i = 0, j = 0; i < thresholds->primary->size; i++) {
4103 if (thresholds->primary->entries[i].eventfd == eventfd)
4106 new->entries[j] = thresholds->primary->entries[i];
4107 if (new->entries[j].threshold < usage) {
4109 * new->current_threshold will not be used
4110 * until rcu_assign_pointer(), so it's safe to increment
4113 ++new->current_threshold;
4119 /* Swap primary and spare array */
4120 thresholds->spare = thresholds->primary;
4121 rcu_assign_pointer(thresholds->primary, new);
4123 /* To be sure that nobody uses thresholds */
4126 mutex_unlock(&memcg->thresholds_lock);
4129 static int mem_cgroup_oom_register_event(struct cgroup *cgrp,
4130 struct cftype *cft, struct eventfd_ctx *eventfd, const char *args)
4132 struct mem_cgroup *memcg = mem_cgroup_from_cont(cgrp);
4133 struct mem_cgroup_eventfd_list *event;
4134 int type = MEMFILE_TYPE(cft->private);
4136 BUG_ON(type != _OOM_TYPE);
4137 event = kmalloc(sizeof(*event), GFP_KERNEL);
4141 mutex_lock(&memcg_oom_mutex);
4143 event->eventfd = eventfd;
4144 list_add(&event->list, &memcg->oom_notify);
4146 /* already in OOM ? */
4147 if (atomic_read(&memcg->oom_lock))
4148 eventfd_signal(eventfd, 1);
4149 mutex_unlock(&memcg_oom_mutex);
4154 static void mem_cgroup_oom_unregister_event(struct cgroup *cgrp,
4155 struct cftype *cft, struct eventfd_ctx *eventfd)
4157 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4158 struct mem_cgroup_eventfd_list *ev, *tmp;
4159 int type = MEMFILE_TYPE(cft->private);
4161 BUG_ON(type != _OOM_TYPE);
4163 mutex_lock(&memcg_oom_mutex);
4165 list_for_each_entry_safe(ev, tmp, &mem->oom_notify, list) {
4166 if (ev->eventfd == eventfd) {
4167 list_del(&ev->list);
4172 mutex_unlock(&memcg_oom_mutex);
4175 static int mem_cgroup_oom_control_read(struct cgroup *cgrp,
4176 struct cftype *cft, struct cgroup_map_cb *cb)
4178 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4180 cb->fill(cb, "oom_kill_disable", mem->oom_kill_disable);
4182 if (atomic_read(&mem->oom_lock))
4183 cb->fill(cb, "under_oom", 1);
4185 cb->fill(cb, "under_oom", 0);
4189 static int mem_cgroup_oom_control_write(struct cgroup *cgrp,
4190 struct cftype *cft, u64 val)
4192 struct mem_cgroup *mem = mem_cgroup_from_cont(cgrp);
4193 struct mem_cgroup *parent;
4195 /* cannot set to root cgroup and only 0 and 1 are allowed */
4196 if (!cgrp->parent || !((val == 0) || (val == 1)))
4199 parent = mem_cgroup_from_cont(cgrp->parent);
4202 /* oom-kill-disable is a flag for subhierarchy. */
4203 if ((parent->use_hierarchy) ||
4204 (mem->use_hierarchy && !list_empty(&cgrp->children))) {
4208 mem->oom_kill_disable = val;
4210 memcg_oom_recover(mem);
4215 static struct cftype mem_cgroup_files[] = {
4217 .name = "usage_in_bytes",
4218 .private = MEMFILE_PRIVATE(_MEM, RES_USAGE),
4219 .read_u64 = mem_cgroup_read,
4220 .register_event = mem_cgroup_usage_register_event,
4221 .unregister_event = mem_cgroup_usage_unregister_event,
4224 .name = "max_usage_in_bytes",
4225 .private = MEMFILE_PRIVATE(_MEM, RES_MAX_USAGE),
4226 .trigger = mem_cgroup_reset,
4227 .read_u64 = mem_cgroup_read,
4230 .name = "limit_in_bytes",
4231 .private = MEMFILE_PRIVATE(_MEM, RES_LIMIT),
4232 .write_string = mem_cgroup_write,
4233 .read_u64 = mem_cgroup_read,
4236 .name = "soft_limit_in_bytes",
4237 .private = MEMFILE_PRIVATE(_MEM, RES_SOFT_LIMIT),
4238 .write_string = mem_cgroup_write,
4239 .read_u64 = mem_cgroup_read,
4243 .private = MEMFILE_PRIVATE(_MEM, RES_FAILCNT),
4244 .trigger = mem_cgroup_reset,
4245 .read_u64 = mem_cgroup_read,
4249 .read_map = mem_control_stat_show,
4252 .name = "force_empty",
4253 .trigger = mem_cgroup_force_empty_write,
4256 .name = "use_hierarchy",
4257 .write_u64 = mem_cgroup_hierarchy_write,
4258 .read_u64 = mem_cgroup_hierarchy_read,
4261 .name = "swappiness",
4262 .read_u64 = mem_cgroup_swappiness_read,
4263 .write_u64 = mem_cgroup_swappiness_write,
4266 .name = "move_charge_at_immigrate",
4267 .read_u64 = mem_cgroup_move_charge_read,
4268 .write_u64 = mem_cgroup_move_charge_write,
4271 .name = "oom_control",
4272 .read_map = mem_cgroup_oom_control_read,
4273 .write_u64 = mem_cgroup_oom_control_write,
4274 .register_event = mem_cgroup_oom_register_event,
4275 .unregister_event = mem_cgroup_oom_unregister_event,
4276 .private = MEMFILE_PRIVATE(_OOM_TYPE, OOM_CONTROL),
4280 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4281 static struct cftype memsw_cgroup_files[] = {
4283 .name = "memsw.usage_in_bytes",
4284 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_USAGE),
4285 .read_u64 = mem_cgroup_read,
4286 .register_event = mem_cgroup_usage_register_event,
4287 .unregister_event = mem_cgroup_usage_unregister_event,
4290 .name = "memsw.max_usage_in_bytes",
4291 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_MAX_USAGE),
4292 .trigger = mem_cgroup_reset,
4293 .read_u64 = mem_cgroup_read,
4296 .name = "memsw.limit_in_bytes",
4297 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_LIMIT),
4298 .write_string = mem_cgroup_write,
4299 .read_u64 = mem_cgroup_read,
4302 .name = "memsw.failcnt",
4303 .private = MEMFILE_PRIVATE(_MEMSWAP, RES_FAILCNT),
4304 .trigger = mem_cgroup_reset,
4305 .read_u64 = mem_cgroup_read,
4309 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4311 if (!do_swap_account)
4313 return cgroup_add_files(cont, ss, memsw_cgroup_files,
4314 ARRAY_SIZE(memsw_cgroup_files));
4317 static int register_memsw_files(struct cgroup *cont, struct cgroup_subsys *ss)
4323 static int alloc_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4325 struct mem_cgroup_per_node *pn;
4326 struct mem_cgroup_per_zone *mz;
4328 int zone, tmp = node;
4330 * This routine is called against possible nodes.
4331 * But it's BUG to call kmalloc() against offline node.
4333 * TODO: this routine can waste much memory for nodes which will
4334 * never be onlined. It's better to use memory hotplug callback
4337 if (!node_state(node, N_NORMAL_MEMORY))
4339 pn = kzalloc_node(sizeof(*pn), GFP_KERNEL, tmp);
4343 mem->info.nodeinfo[node] = pn;
4344 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4345 mz = &pn->zoneinfo[zone];
4347 INIT_LIST_HEAD(&mz->lists[l]);
4348 mz->usage_in_excess = 0;
4349 mz->on_tree = false;
4355 static void free_mem_cgroup_per_zone_info(struct mem_cgroup *mem, int node)
4357 kfree(mem->info.nodeinfo[node]);
4360 static struct mem_cgroup *mem_cgroup_alloc(void)
4362 struct mem_cgroup *mem;
4363 int size = sizeof(struct mem_cgroup);
4365 /* Can be very big if MAX_NUMNODES is very big */
4366 if (size < PAGE_SIZE)
4367 mem = kzalloc(size, GFP_KERNEL);
4369 mem = vzalloc(size);
4374 mem->stat = alloc_percpu(struct mem_cgroup_stat_cpu);
4377 spin_lock_init(&mem->pcp_counter_lock);
4381 if (size < PAGE_SIZE)
4389 * At destroying mem_cgroup, references from swap_cgroup can remain.
4390 * (scanning all at force_empty is too costly...)
4392 * Instead of clearing all references at force_empty, we remember
4393 * the number of reference from swap_cgroup and free mem_cgroup when
4394 * it goes down to 0.
4396 * Removal of cgroup itself succeeds regardless of refs from swap.
4399 static void __mem_cgroup_free(struct mem_cgroup *mem)
4403 mem_cgroup_remove_from_trees(mem);
4404 free_css_id(&mem_cgroup_subsys, &mem->css);
4406 for_each_node_state(node, N_POSSIBLE)
4407 free_mem_cgroup_per_zone_info(mem, node);
4409 free_percpu(mem->stat);
4410 if (sizeof(struct mem_cgroup) < PAGE_SIZE)
4416 static void mem_cgroup_get(struct mem_cgroup *mem)
4418 atomic_inc(&mem->refcnt);
4421 static void __mem_cgroup_put(struct mem_cgroup *mem, int count)
4423 if (atomic_sub_and_test(count, &mem->refcnt)) {
4424 struct mem_cgroup *parent = parent_mem_cgroup(mem);
4425 __mem_cgroup_free(mem);
4427 mem_cgroup_put(parent);
4431 static void mem_cgroup_put(struct mem_cgroup *mem)
4433 __mem_cgroup_put(mem, 1);
4437 * Returns the parent mem_cgroup in memcgroup hierarchy with hierarchy enabled.
4439 static struct mem_cgroup *parent_mem_cgroup(struct mem_cgroup *mem)
4441 if (!mem->res.parent)
4443 return mem_cgroup_from_res_counter(mem->res.parent, res);
4446 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
4447 static void __init enable_swap_cgroup(void)
4449 if (!mem_cgroup_disabled() && really_do_swap_account)
4450 do_swap_account = 1;
4453 static void __init enable_swap_cgroup(void)
4458 static int mem_cgroup_soft_limit_tree_init(void)
4460 struct mem_cgroup_tree_per_node *rtpn;
4461 struct mem_cgroup_tree_per_zone *rtpz;
4462 int tmp, node, zone;
4464 for_each_node_state(node, N_POSSIBLE) {
4466 if (!node_state(node, N_NORMAL_MEMORY))
4468 rtpn = kzalloc_node(sizeof(*rtpn), GFP_KERNEL, tmp);
4472 soft_limit_tree.rb_tree_per_node[node] = rtpn;
4474 for (zone = 0; zone < MAX_NR_ZONES; zone++) {
4475 rtpz = &rtpn->rb_tree_per_zone[zone];
4476 rtpz->rb_root = RB_ROOT;
4477 spin_lock_init(&rtpz->lock);
4483 static struct cgroup_subsys_state * __ref
4484 mem_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cont)
4486 struct mem_cgroup *mem, *parent;
4487 long error = -ENOMEM;
4490 mem = mem_cgroup_alloc();
4492 return ERR_PTR(error);
4494 for_each_node_state(node, N_POSSIBLE)
4495 if (alloc_mem_cgroup_per_zone_info(mem, node))
4499 if (cont->parent == NULL) {
4501 enable_swap_cgroup();
4503 root_mem_cgroup = mem;
4504 if (mem_cgroup_soft_limit_tree_init())
4506 for_each_possible_cpu(cpu) {
4507 struct memcg_stock_pcp *stock =
4508 &per_cpu(memcg_stock, cpu);
4509 INIT_WORK(&stock->work, drain_local_stock);
4511 hotcpu_notifier(memcg_cpu_hotplug_callback, 0);
4513 parent = mem_cgroup_from_cont(cont->parent);
4514 mem->use_hierarchy = parent->use_hierarchy;
4515 mem->oom_kill_disable = parent->oom_kill_disable;
4518 if (parent && parent->use_hierarchy) {
4519 res_counter_init(&mem->res, &parent->res);
4520 res_counter_init(&mem->memsw, &parent->memsw);
4522 * We increment refcnt of the parent to ensure that we can
4523 * safely access it on res_counter_charge/uncharge.
4524 * This refcnt will be decremented when freeing this
4525 * mem_cgroup(see mem_cgroup_put).
4527 mem_cgroup_get(parent);
4529 res_counter_init(&mem->res, NULL);
4530 res_counter_init(&mem->memsw, NULL);
4532 mem->last_scanned_child = 0;
4533 spin_lock_init(&mem->reclaim_param_lock);
4534 INIT_LIST_HEAD(&mem->oom_notify);
4537 mem->swappiness = get_swappiness(parent);
4538 atomic_set(&mem->refcnt, 1);
4539 mem->move_charge_at_immigrate = 0;
4540 mutex_init(&mem->thresholds_lock);
4543 __mem_cgroup_free(mem);
4544 root_mem_cgroup = NULL;
4545 return ERR_PTR(error);
4548 static int mem_cgroup_pre_destroy(struct cgroup_subsys *ss,
4549 struct cgroup *cont)
4551 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4553 return mem_cgroup_force_empty(mem, false);
4556 static void mem_cgroup_destroy(struct cgroup_subsys *ss,
4557 struct cgroup *cont)
4559 struct mem_cgroup *mem = mem_cgroup_from_cont(cont);
4561 mem_cgroup_put(mem);
4564 static int mem_cgroup_populate(struct cgroup_subsys *ss,
4565 struct cgroup *cont)
4569 ret = cgroup_add_files(cont, ss, mem_cgroup_files,
4570 ARRAY_SIZE(mem_cgroup_files));
4573 ret = register_memsw_files(cont, ss);
4578 /* Handlers for move charge at task migration. */
4579 #define PRECHARGE_COUNT_AT_ONCE 256
4580 static int mem_cgroup_do_precharge(unsigned long count)
4583 int batch_count = PRECHARGE_COUNT_AT_ONCE;
4584 struct mem_cgroup *mem = mc.to;
4586 if (mem_cgroup_is_root(mem)) {
4587 mc.precharge += count;
4588 /* we don't need css_get for root */
4591 /* try to charge at once */
4593 struct res_counter *dummy;
4595 * "mem" cannot be under rmdir() because we've already checked
4596 * by cgroup_lock_live_cgroup() that it is not removed and we
4597 * are still under the same cgroup_mutex. So we can postpone
4600 if (res_counter_charge(&mem->res, PAGE_SIZE * count, &dummy))
4602 if (do_swap_account && res_counter_charge(&mem->memsw,
4603 PAGE_SIZE * count, &dummy)) {
4604 res_counter_uncharge(&mem->res, PAGE_SIZE * count);
4607 mc.precharge += count;
4611 /* fall back to one by one charge */
4613 if (signal_pending(current)) {
4617 if (!batch_count--) {
4618 batch_count = PRECHARGE_COUNT_AT_ONCE;
4621 ret = __mem_cgroup_try_charge(NULL, GFP_KERNEL, &mem, false,
4624 /* mem_cgroup_clear_mc() will do uncharge later */
4632 * is_target_pte_for_mc - check a pte whether it is valid for move charge
4633 * @vma: the vma the pte to be checked belongs
4634 * @addr: the address corresponding to the pte to be checked
4635 * @ptent: the pte to be checked
4636 * @target: the pointer the target page or swap ent will be stored(can be NULL)
4639 * 0(MC_TARGET_NONE): if the pte is not a target for move charge.
4640 * 1(MC_TARGET_PAGE): if the page corresponding to this pte is a target for
4641 * move charge. if @target is not NULL, the page is stored in target->page
4642 * with extra refcnt got(Callers should handle it).
4643 * 2(MC_TARGET_SWAP): if the swap entry corresponding to this pte is a
4644 * target for charge migration. if @target is not NULL, the entry is stored
4647 * Called with pte lock held.
4654 enum mc_target_type {
4655 MC_TARGET_NONE, /* not used */
4660 static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
4661 unsigned long addr, pte_t ptent)
4663 struct page *page = vm_normal_page(vma, addr, ptent);
4665 if (!page || !page_mapped(page))
4667 if (PageAnon(page)) {
4668 /* we don't move shared anon */
4669 if (!move_anon() || page_mapcount(page) > 2)
4671 } else if (!move_file())
4672 /* we ignore mapcount for file pages */
4674 if (!get_page_unless_zero(page))
4680 static struct page *mc_handle_swap_pte(struct vm_area_struct *vma,
4681 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4684 struct page *page = NULL;
4685 swp_entry_t ent = pte_to_swp_entry(ptent);
4687 if (!move_anon() || non_swap_entry(ent))
4689 usage_count = mem_cgroup_count_swap_user(ent, &page);
4690 if (usage_count > 1) { /* we don't move shared anon */
4695 if (do_swap_account)
4696 entry->val = ent.val;
4701 static struct page *mc_handle_file_pte(struct vm_area_struct *vma,
4702 unsigned long addr, pte_t ptent, swp_entry_t *entry)
4704 struct page *page = NULL;
4705 struct inode *inode;
4706 struct address_space *mapping;
4709 if (!vma->vm_file) /* anonymous vma */
4714 inode = vma->vm_file->f_path.dentry->d_inode;
4715 mapping = vma->vm_file->f_mapping;
4716 if (pte_none(ptent))
4717 pgoff = linear_page_index(vma, addr);
4718 else /* pte_file(ptent) is true */
4719 pgoff = pte_to_pgoff(ptent);
4721 /* page is moved even if it's not RSS of this task(page-faulted). */
4722 if (!mapping_cap_swap_backed(mapping)) { /* normal file */
4723 page = find_get_page(mapping, pgoff);
4724 } else { /* shmem/tmpfs file. we should take account of swap too. */
4726 mem_cgroup_get_shmem_target(inode, pgoff, &page, &ent);
4727 if (do_swap_account)
4728 entry->val = ent.val;
4734 static int is_target_pte_for_mc(struct vm_area_struct *vma,
4735 unsigned long addr, pte_t ptent, union mc_target *target)
4737 struct page *page = NULL;
4738 struct page_cgroup *pc;
4740 swp_entry_t ent = { .val = 0 };
4742 if (pte_present(ptent))
4743 page = mc_handle_present_pte(vma, addr, ptent);
4744 else if (is_swap_pte(ptent))
4745 page = mc_handle_swap_pte(vma, addr, ptent, &ent);
4746 else if (pte_none(ptent) || pte_file(ptent))
4747 page = mc_handle_file_pte(vma, addr, ptent, &ent);
4749 if (!page && !ent.val)
4752 pc = lookup_page_cgroup(page);
4754 * Do only loose check w/o page_cgroup lock.
4755 * mem_cgroup_move_account() checks the pc is valid or not under
4758 if (PageCgroupUsed(pc) && pc->mem_cgroup == mc.from) {
4759 ret = MC_TARGET_PAGE;
4761 target->page = page;
4763 if (!ret || !target)
4766 /* There is a swap entry and a page doesn't exist or isn't charged */
4767 if (ent.val && !ret &&
4768 css_id(&mc.from->css) == lookup_swap_cgroup(ent)) {
4769 ret = MC_TARGET_SWAP;
4776 static int mem_cgroup_count_precharge_pte_range(pmd_t *pmd,
4777 unsigned long addr, unsigned long end,
4778 struct mm_walk *walk)
4780 struct vm_area_struct *vma = walk->private;
4784 split_huge_page_pmd(walk->mm, pmd);
4786 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4787 for (; addr != end; pte++, addr += PAGE_SIZE)
4788 if (is_target_pte_for_mc(vma, addr, *pte, NULL))
4789 mc.precharge++; /* increment precharge temporarily */
4790 pte_unmap_unlock(pte - 1, ptl);
4796 static unsigned long mem_cgroup_count_precharge(struct mm_struct *mm)
4798 unsigned long precharge;
4799 struct vm_area_struct *vma;
4801 down_read(&mm->mmap_sem);
4802 for (vma = mm->mmap; vma; vma = vma->vm_next) {
4803 struct mm_walk mem_cgroup_count_precharge_walk = {
4804 .pmd_entry = mem_cgroup_count_precharge_pte_range,
4808 if (is_vm_hugetlb_page(vma))
4810 walk_page_range(vma->vm_start, vma->vm_end,
4811 &mem_cgroup_count_precharge_walk);
4813 up_read(&mm->mmap_sem);
4815 precharge = mc.precharge;
4821 static int mem_cgroup_precharge_mc(struct mm_struct *mm)
4823 unsigned long precharge = mem_cgroup_count_precharge(mm);
4825 VM_BUG_ON(mc.moving_task);
4826 mc.moving_task = current;
4827 return mem_cgroup_do_precharge(precharge);
4830 /* cancels all extra charges on mc.from and mc.to, and wakes up all waiters. */
4831 static void __mem_cgroup_clear_mc(void)
4833 struct mem_cgroup *from = mc.from;
4834 struct mem_cgroup *to = mc.to;
4836 /* we must uncharge all the leftover precharges from mc.to */
4838 __mem_cgroup_cancel_charge(mc.to, mc.precharge);
4842 * we didn't uncharge from mc.from at mem_cgroup_move_account(), so
4843 * we must uncharge here.
4845 if (mc.moved_charge) {
4846 __mem_cgroup_cancel_charge(mc.from, mc.moved_charge);
4847 mc.moved_charge = 0;
4849 /* we must fixup refcnts and charges */
4850 if (mc.moved_swap) {
4851 /* uncharge swap account from the old cgroup */
4852 if (!mem_cgroup_is_root(mc.from))
4853 res_counter_uncharge(&mc.from->memsw,
4854 PAGE_SIZE * mc.moved_swap);
4855 __mem_cgroup_put(mc.from, mc.moved_swap);
4857 if (!mem_cgroup_is_root(mc.to)) {
4859 * we charged both to->res and to->memsw, so we should
4862 res_counter_uncharge(&mc.to->res,
4863 PAGE_SIZE * mc.moved_swap);
4865 /* we've already done mem_cgroup_get(mc.to) */
4868 memcg_oom_recover(from);
4869 memcg_oom_recover(to);
4870 wake_up_all(&mc.waitq);
4873 static void mem_cgroup_clear_mc(void)
4875 struct mem_cgroup *from = mc.from;
4878 * we must clear moving_task before waking up waiters at the end of
4881 mc.moving_task = NULL;
4882 __mem_cgroup_clear_mc();
4883 spin_lock(&mc.lock);
4886 spin_unlock(&mc.lock);
4887 mem_cgroup_end_move(from);
4890 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
4891 struct cgroup *cgroup,
4892 struct task_struct *p,
4896 struct mem_cgroup *mem = mem_cgroup_from_cont(cgroup);
4898 if (mem->move_charge_at_immigrate) {
4899 struct mm_struct *mm;
4900 struct mem_cgroup *from = mem_cgroup_from_task(p);
4902 VM_BUG_ON(from == mem);
4904 mm = get_task_mm(p);
4907 /* We move charges only when we move a owner of the mm */
4908 if (mm->owner == p) {
4911 VM_BUG_ON(mc.precharge);
4912 VM_BUG_ON(mc.moved_charge);
4913 VM_BUG_ON(mc.moved_swap);
4914 mem_cgroup_start_move(from);
4915 spin_lock(&mc.lock);
4918 spin_unlock(&mc.lock);
4919 /* We set mc.moving_task later */
4921 ret = mem_cgroup_precharge_mc(mm);
4923 mem_cgroup_clear_mc();
4930 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
4931 struct cgroup *cgroup,
4932 struct task_struct *p,
4935 mem_cgroup_clear_mc();
4938 static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
4939 unsigned long addr, unsigned long end,
4940 struct mm_walk *walk)
4943 struct vm_area_struct *vma = walk->private;
4947 split_huge_page_pmd(walk->mm, pmd);
4949 pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
4950 for (; addr != end; addr += PAGE_SIZE) {
4951 pte_t ptent = *(pte++);
4952 union mc_target target;
4955 struct page_cgroup *pc;
4961 type = is_target_pte_for_mc(vma, addr, ptent, &target);
4963 case MC_TARGET_PAGE:
4965 if (isolate_lru_page(page))
4967 pc = lookup_page_cgroup(page);
4968 if (!mem_cgroup_move_account(pc,
4969 mc.from, mc.to, false, PAGE_SIZE)) {
4971 /* we uncharge from mc.from later. */
4974 putback_lru_page(page);
4975 put: /* is_target_pte_for_mc() gets the page */
4978 case MC_TARGET_SWAP:
4980 if (!mem_cgroup_move_swap_account(ent,
4981 mc.from, mc.to, false)) {
4983 /* we fixup refcnts and charges later. */
4991 pte_unmap_unlock(pte - 1, ptl);
4996 * We have consumed all precharges we got in can_attach().
4997 * We try charge one by one, but don't do any additional
4998 * charges to mc.to if we have failed in charge once in attach()
5001 ret = mem_cgroup_do_precharge(1);
5009 static void mem_cgroup_move_charge(struct mm_struct *mm)
5011 struct vm_area_struct *vma;
5013 lru_add_drain_all();
5015 if (unlikely(!down_read_trylock(&mm->mmap_sem))) {
5017 * Someone who are holding the mmap_sem might be waiting in
5018 * waitq. So we cancel all extra charges, wake up all waiters,
5019 * and retry. Because we cancel precharges, we might not be able
5020 * to move enough charges, but moving charge is a best-effort
5021 * feature anyway, so it wouldn't be a big problem.
5023 __mem_cgroup_clear_mc();
5027 for (vma = mm->mmap; vma; vma = vma->vm_next) {
5029 struct mm_walk mem_cgroup_move_charge_walk = {
5030 .pmd_entry = mem_cgroup_move_charge_pte_range,
5034 if (is_vm_hugetlb_page(vma))
5036 ret = walk_page_range(vma->vm_start, vma->vm_end,
5037 &mem_cgroup_move_charge_walk);
5040 * means we have consumed all precharges and failed in
5041 * doing additional charge. Just abandon here.
5045 up_read(&mm->mmap_sem);
5048 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5049 struct cgroup *cont,
5050 struct cgroup *old_cont,
5051 struct task_struct *p,
5054 struct mm_struct *mm;
5057 /* no need to move charge */
5060 mm = get_task_mm(p);
5062 mem_cgroup_move_charge(mm);
5065 mem_cgroup_clear_mc();
5067 #else /* !CONFIG_MMU */
5068 static int mem_cgroup_can_attach(struct cgroup_subsys *ss,
5069 struct cgroup *cgroup,
5070 struct task_struct *p,
5075 static void mem_cgroup_cancel_attach(struct cgroup_subsys *ss,
5076 struct cgroup *cgroup,
5077 struct task_struct *p,
5081 static void mem_cgroup_move_task(struct cgroup_subsys *ss,
5082 struct cgroup *cont,
5083 struct cgroup *old_cont,
5084 struct task_struct *p,
5090 struct cgroup_subsys mem_cgroup_subsys = {
5092 .subsys_id = mem_cgroup_subsys_id,
5093 .create = mem_cgroup_create,
5094 .pre_destroy = mem_cgroup_pre_destroy,
5095 .destroy = mem_cgroup_destroy,
5096 .populate = mem_cgroup_populate,
5097 .can_attach = mem_cgroup_can_attach,
5098 .cancel_attach = mem_cgroup_cancel_attach,
5099 .attach = mem_cgroup_move_task,
5104 #ifdef CONFIG_CGROUP_MEM_RES_CTLR_SWAP
5105 static int __init enable_swap_account(char *s)
5107 /* consider enabled if no parameter or 1 is given */
5108 if (!(*s) || !strcmp(s, "=1"))
5109 really_do_swap_account = 1;
5110 else if (!strcmp(s, "=0"))
5111 really_do_swap_account = 0;
5114 __setup("swapaccount", enable_swap_account);
5116 static int __init disable_swap_account(char *s)
5118 printk_once("noswapaccount is deprecated and will be removed in 2.6.40. Use swapaccount=0 instead\n");
5119 enable_swap_account("=0");
5122 __setup("noswapaccount", disable_swap_account);